1.5$). Another group of approaches, including Neural Markov Logic Networks (NMLN) and Relational Neural Machines (RNM) extend MLNs, allowing factors of exponential distributions to be implemented as neural architectures. Finally, compute ground atoms scores as dot products between relation and entity embeddings; implication rules are then translated into a logical loss through a continuous relaxation of the implication operator.\n\t\\begin{boxedexample}[label = ex:sl]{Semantic Loss}\n\t\t\\noindent The Semantic Loss is an example of an undirected model where (probabilistic) logic is exploited as a \\textit{regularization} term in training a neural model.\n\t\t\\noindent Let $p = [p_1,\\dots,p_n]$ be a vector of probabilities for a list of propositional variables $X = [X_1,\\dots,X_n]$. In particular, $p_i$ denotes the probability of variable $X_i$ being \\textit{True} and corresponds to a single output of a neural net having $n$ outputs.\n\t\tLet $\\alpha$ be a logic sentence defined over $X$.\n\t\t\\noindent Then, the \\textit{semantic loss} between $\\alpha$ and $p$ is:\n\t\t\\begin{equation*}\n\t\tLoss(\\alpha,p) \\propto - \\log \\,\\, \\sum_{x \\models \\alpha} \\,\\,\\, \\prod_{i: x \\models X_i} p_i \\,\\,\\, \\prod_{i: x \\models \\neg X_i} (1-p_i).\n\t\t\\end{equation*}\n\t\t\\noindent The authors provide the intuition behind this loss: \n\t\t\\begin{quote}\n\t\t\t\\textit{The semantic loss is proportional to the negative logarithm of the probability of generating a state that satisfies the constraint when sampling values according to $p$.}\n\t\t\\end{quote} \n\t\t\\noindent Suppose you want to solve a multi-class classification task (example adapted from ), where each input example must be assigned to a single class. Then, one would like to enforce \\textit{mutual exclusivity} among the classes. This can be easily done on supervised examples, by coupling a softmax activation layer with a cross entropy loss. However, there is no standard way to impose this constraint for unlabeled data, which can be useful in a semi-supervised setting.\n\t\t\\noindent The solution provided by the Semantic Loss framework is to encode mutual exclusivity into the propositional constraint $\\beta$:\n\t\t\\begin{equation*}\n\t\t\\beta = (X_1 \\land \\neg X_2 \\land \\neg X_3) \\lor\n\t\t(\\neg X_1 \\land X_2 \\land \\neg X_3) \\lor\n\t\t(\\neg X_1 \\land \\neg X_2 \\land X_3)\n\t\t\\end{equation*}\n\t\tConsider a neural network classifier with three outputs $ p =[ p_1, p_2, p_3]$. Then, for each input example (whether labeled or unlabeled), we can build the semantic loss term:\n\t\t\\begin{equation*}\n\t\tL(\\beta,p) = p_1(1 - p_2)(1 - p_3) +\n\t\t(1 - p_1)p_2(1 - p_3) +\n\t\t(1 - p_1)(1 - p_2)p_3\n\t\t\\end{equation*}\n\t\t\\noindent It can be summed up to the standard cross-entropy term for the labeled examples.\n\t\tUnlike for directed methods such as KBANN (Example \\ref{ex:kbann}) and TensorLog, \n the logic is turned into a loss-function that is used during training. The function constrains the underlying probabilities, but there are no directed or causal relationships among them.\n\tMoreover, during inference only the probabilities $p$ are used while the logic formula $\\beta$ is not used anymore. On the contrary, in KBANN, the logic is compiled into the architecture of the network and, therefore, it it is also exploited at evaluation time. \n\t\\end{boxedexample}\n\t\\rev{To conclude, let us stress a key difference between \n the two classes of NeSy systems w.r.t. \\textit{the way they incorporate the knowledge expressed in the logical clauses}. Proof-based, directed models use logic to define the architecture of a neural symbolic network. \n Thus, logic is part of the inference of the model and acts as a structural constraint. The designer has full control of where and how the logic is used inside the network. Thus, logical knowledge can easily be extended or modified at test-time, without the need to retrain, leading to a high degree of modularity and out-of-distribution generalization . On the other hand, when logic is only encoded in an objective function, the neural net learns to (approximately) satisfy it. Therefore, the knowledge is only latently encoded in the weights of the network, which leads to a loss of control and interpretability. However, the latter techniques are often much more scalable, especially at inference time. The balance between control and interpretability, on the one hand, and scalability, on the other hand, is an open and important research question in the \\nesy{} community. }","id":"dfea86d8-fb1c-484d-a9d2-d17c03c4d644","level":"subsection","origin_cites_number":22,"parent_id":"71d65d84-75cd-4c92-b14d-bddcaa016dd0","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Proof- vs Model-theoretic View of Logic"],["subsection","Implications for NeSy"]],"subsections":[],"title":"Implications for NeSy"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:syntax}\n\tIn Section \\ref{sec:proof_vs_model}, we have introduced clausal logic, without paying much attention to the structure of the atoms and literals. This structure and its consequences for StarAI and \\nesy{} models are the topic of the present section. Consider the following example: \n\t\\begin{boxedexample}[label = ex:prop_clausal_logic]{Propositional Clausal logic}\n\t\tConsider the following set of clauses. \n\t\t\\begin{lstlisting}\n\t\tmortal_socrates $\\leftarrow$ human_socrates.\n\t\tmortal_aristotle $\\leftarrow$ human_aristotle.\n\t\thuman_socrates.\n\t\thuman_aristotle.\n\t\t\\end{lstlisting}\n\t\\end{boxedexample} \n\tHere, the literals do not possess any internal structure. \\rev{They are {\\em propositions}, which are atoms that we can only assign the value $True$ or $False$. We say that we are working in \\textit{propositional logic}. }\n\tThis contrasts with\n\t\\textit{first-order} logic in which the literals take the form $p(t_1, ... , t_m)$, with $p$ a predicate of arity $m$ and the $t_i$ terms, that is, constants, variables, or structured terms of the form $f(t_1, ..., t_q)$, where $f$ is a functor and the $t_i$ are again terms. \n\tIntuitively, constants represent objects or entities, functors represent functions, variables make abstraction of specific objects, and predicates specify relationships amongst objects.\n\tThe subset of first order logic where there are no functors is called {\\em relational logic}. \n\t\\begin{boxedexample}[label=ex:fol_clausal_logic]{First order clausal logic}\n\t\tIn contrast to the previous example, we now\n\t\twrite the theory in a more compact manner using \\textit{first order logic}. By convention, constants start with a lowercase letter, while variables start with an uppercase.\n\t\tEssential is the use of the variable \\textit{X}, which is implicitly universally quantified.\n\t\t\\begin{lstlisting}\n\t\tmortal(X) $\\leftarrow$ human(X). \n\t\thuman(socrates).\n\t\thuman(aristotle).\n\t\t\\end{lstlisting}\n\t\\end{boxedexample} \n\t\\rev{It is interesting to understand the connection between propositional, relational and first-order logic. To this end, we introduce the concept of grounding.} \n\tWhen an expression (i.e, clause, atom or term) does not contain any variable it is called {\\em ground}. \n\tA substitution $\\theta$ is an expression of the form $\\{X_1/c_1, ..., X_k/c_k \\}$ with the $X_i$ different variables, the $c_i$ terms. Applying a substitution $\\theta$ to an expression $e$ (term, atom or clause) yields the instantiated expression $e\\theta$ where all variables $X_i$ in $e$ have been simultaneously replaced by their corresponding terms $c_i$ in $\\theta$. We can take for instance the clause $mortal(X) \\leftarrow human(X)$ and apply the substitution $\\{X/socrates\\}$ to yield $mortal(socrates) \\leftarrow human(socrates)$. Grounding is the process whereby all possible substitutions that ground the clauses are applied. Notice that grounding a first order logical theory may result in an infinite set of ground clauses (when there are functors), and a polynomially larger set of clauses (when working with finite domains).\n\t\\rev{Finite domains are the focus in both StarAI and \\nesy{}. In such domains,} any problem expressed in first-order logic can be equivalently expressed in relational logic and any problem expressed in relational logic can likewise be expressed in propositional logic by grounding out the clauses .","id":"394ef4e2-3d6b-46dc-9354-ff1b43665302","level":"section","origin_cites_number":2,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Syntax"]],"subsections":["3524cff6-f045-4e2a-be1c-cdf237557b34","7f592cdb-5aaa-4d78-a9de-739ab8dc18a6"],"title":"Logic - Syntax"},{"cite_extract_rate":0,"cites":[],"content":"StarAI typically focus on first order logic ~. \\rev{In Section \\ref{sec:proof_vs_model}, we have seen how StarAI models can be easily interpreted in terms of probabilistic graphical models (PGM). Here, we want to show that FOL is a powerful tool for building such models.} \n\t\\begin{evidencebox}\n\t\t\\rev{FOL allows for knowledge in the form of logical rules to be interpreted as a template for defining the graphical models. Grounding the theory then corresponds to \\textit{unrolling} the template. At the same time, first order logic has also an important statistical and learning advantage: a FOL rule leads to parameter sharing in the model as the parameters of a single FOL rule are tied to all its groundings. Parameter sharing compresses the representation of the corresponding probabilistic model, resulting in more efficient learning and better generalization.}\n\t\\end{evidencebox}\n\t\\rev{These properties are reminiscent of plate notation for probabilistic graphical models, bringing logical reasoning into the picture . In Example \\ref{ex:mln}, we have used only two first order rules but we obtained a larger graphical model with six factors (see Figure \\ref{fig:mln}) by grounding (i.e. unrolling) the rules over the domain. All factors corresponding to the same rule share the same weight.}","id":"3524cff6-f045-4e2a-be1c-cdf237557b34","level":"subsection","origin_cites_number":4,"parent_id":"394ef4e2-3d6b-46dc-9354-ff1b43665302","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Syntax"],["subsection","Implications for StarAI"]],"subsections":[],"title":"Implications for StarAI"},{"cite_extract_rate":0.75,"cites":[5714,3741,5712,5710,5707,7304,4323,5709,5713,8943,3726,3727],"content":"}\n\t\\rev{NeSy exploits the internal structure of literals, resulting in many relational and first-order systems. System exploiting propositional logic are Semantic Loss (SL) ~ and DL2~. Relational logic-based systems are DiffLog~, $\\theta$ILP , Lifted Relational Neural Networks (LRNN) , Neural Theorem Provers (NTP) \n\tand NeurASP~.\n\t\tFinally, many systems are based on first-order logic or first-order logic programs, such as \n\t\tDeepProbLog , NLog , NLProlog , \n\t\tDeepStochLog ,\n\t\tLogic Tensor Networks , Semantic Based Regularization , Relational Neural Machines and Logical Neural Networks .}\n\t\\rev{The focus in \\nesy{} on structured terms is strongly related to that in StarAI and plays a fundamental role in \\nesy{}. \n\t In fact, grounding a relational or first-order theory can often be seen as unrolling either the architecture (e.g., DeepStochLog, LRNN ) or the loss function (e.g., SBR , LTN ) of the corresponding neural model. Unrolling fixed modules over multiple elements of a complex data structure is fundamental to neural networks on sequences (recurrent nets, RNN), trees (recursive nets, RvNN) and graphs (graph nets, GNN). \\nesy{} can be regarded as unrolling more complex logical structures, with similar benefits in terms of model capacity, modularization and generalization, and strong control due to the formal semantics.}\n\t\\rev{Moreover, first-order \\nesy{} models can explicitly deal with how subsymbolic data (e.g. images or audio) are fed to the neural components of the system.\n In fact, \\nesy{} systems often use subsymbolic data samples as elements of the domain of discourse. For example, the element \\textit{mary} can be used to refer to an image, e.g. $mary = \\smallimg{mary.jpeg}$. Feeding such samples as input to a neural network can then be naturally encoded as grounding a predicate over the domain of interest. When the internal structure of the literals is absent, as in SL, this mapping must be handled outside the logical framework.}\nWhile both relational logic and first-order logic have their advantages, there is a noteworthy distinction in the latter. \nFirst-order logic allows representing real valued functions through the use of functors. For example, segmentation can be modeled as a functor returning the bounding box of an object inside an image, e.g. \\textit{location(mary,image)} . Therefore, FOL-based systems can address regression tasks, diverging from the conventional classification tasks associated with relational logic systems.","id":"7f592cdb-5aaa-4d78-a9de-739ab8dc18a6","level":"subsection","origin_cites_number":16,"parent_id":"394ef4e2-3d6b-46dc-9354-ff1b43665302","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Syntax"],["subsection","Implications for \\nesy{"]],"subsections":[],"title":"Implications for \\nesy{"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:semantics}","id":"3f30c5c1-1c7c-4ae0-b5f0-88220243574e","level":"section","origin_cites_number":0,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Semantics"]],"subsections":["8d5d9e2a-e3dc-4e1e-ba7c-32324338d1ca","5422e096-fdd4-4f74-a7c2-b0a811af5aa2","09725807-0a82-437e-980b-ae2733823d35","0c28445b-24b7-4df7-a61f-1e0537492ed8"],"title":"Logic - Semantics"},{"cite_extract_rate":0,"cites":[],"content":"\\rev{The semantics of logical, probabilistic logical and neurosymbolic systems is defined in terms of a model theoretic semantics.}\n\\rev{In the present section, we will restrict our attention to Herbrand interpretations and models as is usual in logic programming and statistical relational AI (see Section \\ref{sec:proof_vs_model})}.\n\\rev{We can distinguish three different levels of semantics, which are also closely tied to the used syntax of the underlying logic.}\n\\rev{First, when the logical theory consists of definite clauses only, the semantics is given by the least Herbrand model. The least Herbrand model of a definite clause theory\nis unique and it is the smallest w.r.t. set inclusion. It contains all ground facts (from the Herbrand domain) that are logically entailed by the theory.\nFor instance, considering the facts $a$ and $b$ and the rules $d \\leftarrow a, b$ and $b \\leftarrow c$ would give the least Herbrand model $\\{a, b, d\\}$.}\n\\rev{Second, when the logical theory can contain any set of clauses, the semantics is given by the set of all possible Herbrand models. For instance, considering the clause $a \\vee b $\nyields the models $\\{a\\}, \\{b\\}$ and $\\{a, b\\}$. So there is not necessarily a unique model, not even when considering only minimal models, where\nwe have $\\{a\\}, \\{b\\}$.} \n\\rev{Third, while Horn-clauses are the basis for \"pure\" Prolog and logic programs, there exist several extensions to this formalism to accommodate\nnegated literals in the condition part of rules or disjunction in the head. A popular framework in this regard is answer set programming (ASP). \nIn ASP the clause $a \\vee b$ could be represented by two clauses $a \\leftarrow \\neg b$ and $b \\leftarrow \\neg a$ which would have two stable models $\\{a\\}$ and $\\{b\\}$.}","id":"8d5d9e2a-e3dc-4e1e-ba7c-32324338d1ca","level":"subsection","origin_cites_number":0,"parent_id":"3f30c5c1-1c7c-4ae0-b5f0-88220243574e","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Semantics"],["subsection","Model-theoretic semantics"]],"subsections":[],"title":"Model-theoretic semantics"},{"cite_extract_rate":0,"cites":[],"content":"\\rev{The previous three levels of semantics are based on Boolean models, i.e. models where each atom is either present (i.e. \\textit{True}) or absent (i.e. \\textit{False})}. Differently, \\textit{fuzzy logic}, and in particular t-norm fuzzy logic, assigns a truth value to atoms in the \\textit{continuous} real interval $[0,1]$. Logical operators are then turned into real-valued functions, mathematically grounded in the t-norm theory. A \\textit{t-norm} $t(x, y)$ is a real function $t:[0,1] \\times [0,1] \\to [0,1]$ that models the logical AND and from which the other operators can be derived. Table~\\ref{tab:t-norms} shows well-known t-norms and the functions corresponding to their connectives. A fuzzy logic formula is mapped to a real valued function of its input atoms, as we show in Example~\\ref{ex:fuzzy}. Fuzzy logic generalizes Boolean logic to continuous values. All the different t-norms are coherent with Boolean logic in the endpoints of the interval $[0,1]$, which correspond to \\textit{completely true} and \\textit{completely false} values. The concept of model in fuzzy logic can be easily recovered from an extension of the model-theoretic semantics of the Boolean logic. Any \\textit{fuzzy} interpretation is a model of a formula if the formula evaluates to 1.\n\\begin{boxedexample}[label=ex:fuzzy]{Fuzzy logic}\n\t\tLet us consider the following propositions: \\textit{alarm}, \\textit{burglary} and \\textit{earthquake}. Defining a fuzzy semantics for this language requires one to assign truth degrees to each of the propositions and selecting a particular t-norm to implement the connectives. \n\t\tLet us consider the \\L ukasiewicz t-norm and the following interpretation of the language:\n\t\t\\begin{minipage}{0.40\\linewidth}\n\t\t\t\\begin{align*}\n\t\t\t\\mathcal{I} = \\{&\\texttt{alarm} = 0.7, \\\\\n\t\t\t&\\texttt{burglary}=0.6, \\\\ &\\texttt{earthquake}=0.3\\}\n\t\t\t\\end{align*}\n\t\t\\end{minipage}\n\t\t\\begin{minipage}{0.60\\linewidth}\n\t\t\t\\begin{align*}\n\t\t\tt_{\\vee}(x,y) & = \\min ( 1, x + y)\\\\\n\t\t\tt_{\\rightarrow}(x,y) & = \\min(1, 1 - x + y)\n\t\t\t\\end{align*}\n\t\t\\end{minipage}\n\t\t\\newline\n\t\tOnce we have defined the semantics of the language, we can evaluate logic sentences, e.g.:\n\t\t\\begin{align*}\n\t\t& \\texttt{alarm} \\leftarrow (\\texttt{burglary} \\vee \\texttt{earthquake}) = \\\\\n\t\t& \\min(1,1 - \\min(1, \\texttt{burglary} + \\texttt{earthquake}) +\\texttt{alarm}) = 0.8\n\t\t\\end{align*}\n\t\tThis evaluation can be performed automatically by parsing the logical sentence in the corresponding \\textit{expression tree} and then compiling the expression tree using the corresponding t-norm operation: \n\t\t\\begin{center}\n\t\t\t\\include{to_include/expression_tree_fuzzy}\n\t\t\\end{center}\n\t\tThe resulting circuit represents a differentiable function and the truth degree of the sentence is computed by evaluating the circuit bottom-up.\n\t\\end{boxedexample}\n\t\\begin{table}[tb]\n\t\t\\centering\n\t\t\\begin{tabular}{c|c|c|c}\n\t\t\t& Product & \\L ukasiewicz & G\\\"{o}del \\\\ \n\t\t\t\\hline\n\t\t\t$x \\land y$ & $x \\cdot y$ & $\\max(0, x + y -1$) & $\\min(x, y$) \\\\\n\t\t\t\\hline\n\t\t\t$x \\lor y$ & $x + y - x \\cdot y$ & $\\min(1, x + y)$ & $\\max(x, y)$ \\\\\n\t\t\t\\hline\n\t\t\t$\\lnot x$ & $1 - x$ & $1 - x$ & $1 - x$ \\\\\n\t\t\t\\hline\n\t\t\t$x \\Rightarrow y \\;\\; (x>y)$ & $y/x$ & $\\min(1,1-x+y)$ & $y$ \\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\t\\caption{Logical connectives on the inputs $x,y$ when using the fundamental t-norms.}\n\t\t\\label{tab:t-norms}\n\t\\end{table}","id":"5422e096-fdd4-4f74-a7c2-b0a811af5aa2","level":"subsection","origin_cites_number":0,"parent_id":"3f30c5c1-1c7c-4ae0-b5f0-88220243574e","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Semantics"],["subsection","Fuzzy semantics"]],"subsections":[],"title":"Fuzzy semantics"},{"cite_extract_rate":0.33333333333333304,"cites":[8685,5706],"content":"Statistical Relational AI has extended the previous semantics by defining probability distributions $p(\\omega)$ over models, or \\textit{possible worlds}\\footnote{In this paper, we use the distribution semantics as representative of the probabilistic approach to logic. While this is the most common solution in StarAI, many other solutions exist , whose description is out of the scope of the current survey. A detailed overview of the different flavours of formal reasoning about uncertainty can be found in .}.\nThe goal is to reason about the uncertainty of logical statements. In particular, the probability that a certain formula $\\alpha$ holds is computed as the sum of the probabilities of the possible worlds that are models of $\\alpha$ (i.e. where $\\alpha$ is True):\n\t\\begin{equation}\n\t\\label{eq:wmc}\n\tp(\\alpha) = \\sum_{\\omega \\models \\alpha} p(\\omega) \n\t\\end{equation}\n\tThis is an instance of the Weighted Model Counting (WMC) problem. In fact, we are counting how many worlds are models of $\\alpha$ and we are weighting each of them by its probability according to the distribution $p(\\omega)$.\n\t\\begin{table}\n\t\t\\centering\n\t\t\\begin{tabular}{cccc|l}\n\t\t\tB & E & J & M & $p(\\omega)$ \\\\\n\t\t\t\\hline\n\t\t\tF&F&F&F& 0.2394 \\\\\n\t\t\tF&F&F&T& 0.1026 \\\\\n\t\t\tF&F&T&F& 0.3591 \\\\\n\t\t\tF&F&T&T& 0.1539 \\\\\n\t\t\tF&T&F&F& 0.0126 \\\\\n\t\t\tF&T&F&T& 0.0054 \\\\\n\t\t\tF&T&T&F& 0.0189 \\\\\n\t\t\tF&T&T&T& 0.0081 \\\\\n\t\t\tT&F&F&F& 0.0266 \\\\\n\t\t\tT&F&F&T& 0.0114 \\\\\n\t\t\tT&F&T&F& 0.0399 \\\\\n\t\t\tT&F&T&T& 0.0171 \\\\\n\t\t\tT&T&F&F& 0.0014 * \\\\\n\t\t\tT&T&F&T& 0.0006 * \\\\\n\t\t\tT&T&T&F& 0.0021 * \\\\\n\t\t\tT&T&T&T& 0.0009 * \n\t\t\\end{tabular}\n\t\t\\caption{A distribution over possible worlds for the four propositional variables $burglary$ (B), $earthquake$ (E), $hears\\_alarm\\_john$ (J) and $hears\\_alarm\\_mary$ (M). The $*$ indicates those worlds where $burglary \\wedge earthquake$ is \\textit{True}.}\n\t\t\\label{tab:distribution_semantics}\n\t\\end{table}\n\t\\begin{boxedexample}[label = ex:distr_semantics]{Probabilistic Logic}\n\t\tLet us consider the following set of propositions $B = burglary$, $E = earthquake$, $J = hears\\_alarm\\_john$ and $M = hears\\_alarm\\_mary$. In probabilistic logic, a probability distribution over all the possible worlds is defined. For example, Table \\ref{tab:distribution_semantics} represents a valid distribution.\n\t\tSuppose we want to compute the probability of the formula $burglary \\wedge earthquake$. This is done by summing up the probabilities of all the worlds where both $burglary$ and $earthquake$ are \\textit{True} (indicated by a $*$ in Table \\ref{tab:distribution_semantics}). \n\t\\end{boxedexample}\n The StarAI community has provided several formalisms to define such probability distributions over possible worlds using labeled logic theories. Probabilistic Logic Programs (cf. Example \\ref{ex:problog}) and Markov logic networks (cf. Example \\ref{ex:mln}) are two prototypical frameworks. For example, the distribution in Table~\\ref{tab:distribution_semantics} is the one modeled by the ProbLog program in Example \\ref{ex:problog}.\n\\rev{It is interesting to compare Markov Logic (Example \\ref{ex:mln}) to ProbLog (Example \\ref{ex:problog}) in terms of their model-theoretic semantics. Markov Logic is defined as a set of weighted full clauses, i.e. as an unnormalized probability distribution over full clausal theories.\nThis means that, given any subset of the theory, there can be many possible models. For instance, the theory $a \\vee b$, has three possible models. To obtain a probability distribution over models, Markov Logic needs to distribute the probability mass over its models. \nTo do this, the maximum entropy principle is used, which results in equal distributions of the probability mass.} \n\\rev{Conversely, ProbLog defines a probability distribution over definite clause theories, each obtained as subsets of the provided probabilistic facts. However, since each of these theories has a unique least Herbrand model, the probability mass corresponding to the selected facts is assigned to the corresponding unique Herbrand model. This means that when working with definite clauses only, there is no need to distribute the probability mass to multiple models and, therefore, no extra assumptions such as maximum entropy are necessary.} \nProbabilistic inference (i.e. weighted model counting) is generally intractable. That is why, in StarAI, techniques such as \\textit{knowledge compilation} (KC)~ are used.\n\tKnowledge compilation transforms a logical formula $\\alpha$ into a new representation in an offline step, which can be computationally expensive. Using this new representation a particular set of queries can be answered efficiently (i.e. in poly-time in the size of the new representation). \n\tFrom a probabilistic point of view, this translation solves the disjoint-sum problem, which states that one cannot simply sum up the probability \n of two disjuncts but also has to subtract the probability of the intersection.\n After the translation, the probabilities of any conjunction and of any disjunction can be simply computed by multiplying, resp. summing up, the probabilities of their operands. Thus a logical formula $\\alpha$ can be compiled into an arithmetic circuit $ac(\\alpha)$. The weighted model count of the query formula can then simply be computed by evaluating the corresponding arithmetic circuit bottom up; i.e. $p(\\alpha) = ac(\\alpha)$. \n\t\\begin{figure}[t]\n\t\t\\centering\n\t\t\\noindent\n\t\t\\begin{subfigure}{0.4\\linewidth}\n\t\t\t\\include{to_include/dDNNF}\n\t\t\\end{subfigure}\n\t\t\\hfill\n\t\t\\begin{subfigure}{0.4\\linewidth}\n\t\t\t\\include{to_include/arithmetic_circuit}\n\t\t\\end{subfigure}\n\t\t\\caption{dDNNF (left) and arithmetic circuit (right) corresponding to the ProbLog program in Example \\ref{ex:problog}}\n\t\t\\label{fig:kc}\n\t\\end{figure}\n\t\\begin{boxedexample}[label = ex:kc]{Knowledge Compilation}\n\t\tLet us consider the ProbLog program in Example \\ref{ex:problog} and the corresponding tabular representation in Table \\ref{tab:distribution_semantics}. Let us consider the query $q = calls(mary)$. Now we can use Equation~\\ref{eq:wmc} to compute the probability $p(q)$. To do this, we iterate over the table and we sum all the probabilities of the worlds where $calls(mary)$ is True, which we know from Example \\ref{ex:logic_program} are those where either $burglary=T$ or $earthquake=T$ and where $hears\\_alarm(mary)=T$. This yields $p(q) = 0.0435$. This method would require us to iterate over $2^N$ terms (where $N$ is the number of probabilistic facts).\n\t\tKnowledge compilation compiles $\\alpha$ into some normal form that is logically equivalent. In Figure \\ref{fig:kc}, the target representation is a decomposable, deterministic negative normal form (d-DNNF)~, for which weighted model counting is poly-time in the size of the formula. Decomposability means that, for every conjunction, the two conjuncts do not share any variables. Deterministic means that, for every disjunction, the two disjuncts are \\rev{mutually exclusive, i.e., only one of the disjuncts can be true at the same time}. \n\t\tThe formula in d-DNNF can then be straightforwardly turned into an arithmetic circuit by substituting AND nodes with multiplication and OR nodes by summation. \n In Figure \\ref{fig:kc}, we show the d-DNNF and the arithmetic circuit of the distribution defined by the ProbLog program in Example \\ref{ex:problog}. The bottom-up evaluation of this arithmetic circuit computes the correct marginal probability $p(\\alpha)$ much more efficiently than the naive iterative sum that we have computed before. \n\t\\end{boxedexample}\n\tEven though probabilistic Boolean logic is the most common choice in StarAI, some approaches use probabilistic fuzzy logic. The most prominent approach is Probabilistic Soft Logic (PSL) , illustrated in Example~\\ref{ex:psl}. Similarly to Markov logic networks, Probabilistic Soft Logic (PSL) defines log linear models where features are represented by ground clauses. However, PSL uses a fuzzy semantics of the logical theory. Therefore, atoms are mapped to real valued variables and ground clauses to real valued factors.\n\t\\begin{boxedexample}[label = ex:psl]{Probabilistic Soft Logic}\n\t\tLet us consider the logical rule $\\alpha = smokes(X) \\leftarrow stress(X)$ with weight $\\beta$.\n\t\tAs we have seen in Example~\\ref{ex:mln}, Markov Logic translates the formula into a discrete factor by using the indicator functions $\\mathbbm{1}(\\omega, \\alpha\\theta)$:\n\t\t\\begin{equation*}\n\t\t\\phi^{MLN}(\\omega, \\alpha) = \\beta \\mathbbm{1}(\\omega, \\alpha\\{X/mary\\}) + \\beta \\mathbbm{1}(\\omega, \\alpha\\{X/john\\})\n\t\t\\end{equation*}\n\t\tInstead of discrete indicator functions, PSL translates the formula into a continuous t-norm based function:\n\t\t\\begin{equation*}\n\t\tt(\\omega, \\alpha) = \\min(1, 1 - stress(X) + smokes(X))\n\t\t\\end{equation*}\n\t\tand the corresponding potential is then translated into the continuous and differentiable function:\n\t\t\\begin{equation*}\n\t\t\\phi^{PSL}(\\omega, \\alpha) = \\beta t(\\omega, \\alpha\\{X/mary\\}) + \\beta t(\\omega, \\alpha\\{X/john\\})\n\t\t\\end{equation*}\n\t\tAnother important task in StarAI is MAP inference. In MAP inference, given the distribution $p(\\omega)$, one is interested in finding the interpretation $\\omega^\\star$ where $p$ is maximal, i.e.\n\t\t\\begin{equation}\n\t\t\\label{eq:map}\n\t\t\\omega^\\star = \\text{arg}\\max_\\omega p(\\omega)\n\t\t\\end{equation}\n\t\tWhen the $\\omega$ is a boolean interpretation, i.e. $\\omega \\in \\{0,1\\}^n$, like in ProbLog or MLNs, this problem is \\rev{related to maxSAT, which is NP-hard}. However, in PSL, $\\omega$ is a fuzzy interpretation, i.e. $\\omega \\in [0,1]^n$ and $p(\\omega) \\propto \\exp\\big(\\sum_i \\beta_i \\phi(\\omega, \\alpha_i)\\big)$ is a continuous and differentiable function. The MAP inference problem can thus be \\textit{approximated} more efficiently than its boolean counterpart using gradient-based techniques.\n\t\\end{boxedexample}","id":"09725807-0a82-437e-980b-ae2733823d35","level":"subsection","origin_cites_number":6,"parent_id":"3f30c5c1-1c7c-4ae0-b5f0-88220243574e","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Semantics"],["subsection","Implications for StarAI"]],"subsections":[],"title":"Implications for StarAI"},{"cite_extract_rate":0.695652173913043,"cites":[5712,5707,3726,5718,5717,5715,681,5716,3727,5719,5721,5710,5711,5720,4323,3672],"content":"We have seen that in StarAI, one can turn inference tasks into the evaluation (as in KC) or gradient-based optimization (as in PSL) of a differentiable parametric circuit. The parameters are scalar values (e.g. probabilities or truth degrees) that are attached to basic elements of a logical theory (facts or clauses).\n\tA natural way of carrying over the StarAI approach to NeSy is the reparameterization method. Reparameterization substitutes the scalar values assigned to facts or formulas with the output of a neural network. One can interpret this substitution in terms of a different parameterization of the original model. Many probabilistic methods parameterize the underlying distribution in terms of neural components. In particular, as we show in Example \\ref{ex:deepproblog}, DeepProbLog exploits neural predicates to compute the probabilities of probabilistic facts as the output of neural computations over vectorial representations of the constants, which is similar to SL in the propositional counterpart (see Example \\ref{ex:sl}). NeurASP also inherits the concept of a neural predicate from DeepProbLog.\n\t\\begin{boxedexample}[label = ex:deepproblog]{Probabilistic semantics reparameterization in DeepProbLog}\n\t\tDeepProbLog is a neural extension of the probabilistic logic programming language ProbLog. DeepProbLog allows images or other subsymbolic representations as terms of the program.\n\t\tLet us consider a possible neural extension of the program in Example~\\ref{ex:problog}. We could extend the predicate $calls(X)$ with two extra inputs, i.e. $calls(B,E,X)$. $B$ is supposed to contain an image of a security camera, while $E$ is supposed to contain the time-series of a seismic sensor. We would like to answer queries like $calls(\\smallimg{burglary1.png},\\smallimg{earthquake1.png},mary)$, i.e. what is the probability that $mary$ calls, given that the security camera has captured the image $\\smallimg{burglary1.png}$ and the sensor the signal $\\smallimg{earthquake1.png}$ .\n\t\tDeepProbLog can answer this query using the following program:\n\t\t\\begin{lstlisting}\n\t\tnn(nn_burglary, [B]) :: burglary(B).\n\t\tnn(nn_earthquake, [E]) :: earthquake(E).\n\t\t0.3::hears_alarm(mary). \n\t\t0.6::hears_alarm(john). \n\t\talarm(B,_) <- burglary(B).\n\t\talarm(_,E) <- earthquake(E).\n\t\tcalls(B,E, X) <- alarm(B,E), hears_alarm(X).\n\t\t\\end{lstlisting}\n\t\tHere, the program has been extended in two ways. First, new arguments (i.e. $B$ and $E$) have been introduced in order to deal with the subsymbolic inputs. Second, the probabilistic facts $burglary$ and $earthquake$ have been turned into \\textit{neural predicates}. Neural predicates are special probabilistic facts that are annotated by neural networks instead of by scalar probabilities.\n\t\tInference in DeepProbLog mimics that of ProbLog. Given the query and the program, knowledge compilation is used to build the arithmetic circuit in Figure \\ref{fig:deepproblog}.\n\t\tSince the program is structurally identical to the purely symbolic one in Example \\ref{ex:kc}, the arithmetic circuit is exactly the same. The only only difference is that some leaves of the tree (i.e. capturing probabilities of facts) can now also be neural networks.\n\t\tGiven a set of queries that are \\textit{True}, i.e.:\n\t\t\\begin{align*}\n\t\t\\mathcal{D} = \\{&calls(\\smallimg{burglary1.png},\\smallimg{earthquake1.png},mary),\\\\ &calls(\\smallimg{burglary2.png},\\smallimg{earthquake2.png},john), \\\\ &calls(\\smallimg{burglary3.png},\\smallimg{earthquake3.png},mary), ...\\},\n\t\t\\end{align*} we can train the parameters $\\theta$ of the DeepProbLog program (both neural networks and scalar probabilities) by maximizing the log-likelihood of the training queries using gradient descent:\n\t\t\\begin{equation*}\n\t\t\\max_{\\theta} \\sum_{q \\in \\mathcal{D}} \\log p(q)\n\t\t\\end{equation*}\n\t\\end{boxedexample}\n\t\\begin{figure}[t]\n\t\t\\centering\n\t\t\\include{to_include/arithmetic_circuit_dpl}\n\t\t\\caption{A neural reparametrization of the arithmetic circuit in Example \\ref{ex:kc} as done by DeepProbLog (cf. Example \\ref{ex:deepproblog}). Dashed lines indicate a negative output, i.e 1 - x. \\rev{We use a different notation for negation than in Figure \\ref{fig:kc} to stress that both leaves are parameterized by the same neural network}.}\n\t\t\\label{fig:deepproblog}\n\t\\end{figure}\n\tSimilarly to DeepProbLog, NMLNs and RNMs use neural networks to parameterize the factors (or the weights) of a Markov Logic Network.\n\t computes marginal probabilities as logistic functions over similarity measures between embeddings of entities and relations. An alternative solution to exploit a probabilistic semantics is to use knowledge graphs (see also \\ref{sec:kge_gnn}) to define probabilistic priors to neural network predictions, as done in .\n\tSBR and LTN reparametrize fuzzy atoms using neural networks that take as inputs the feature representation of the constants and return the corresponding truth value, as shown in Example~\\ref{ex:sbr}. Logical rules are then relaxed into soft constraints using fuzzy logic. Many other systems exploit fuzzy logic to inject knowledge into neural models . These methods can be regarded as variants of a unique conceptual framework as the differences are often minor and in the implementation details. \n\t\\begin{boxedexample}[label = ex:sbr]{Semantic-Based Regularization}\n\t\t\\noindent Semantic-Based Regularization (SBR) is an example of an undirected model where fuzzy logic is exploited as a \\textit{regularization} term when training a neural model. \n\t\tLet us consider a possible grounding for the rule in Example \\ref{ex:psl}:\n\t\t\\begin{lstlisting}\n\t\tsmokes(mary) $\\leftarrow$ stress(mary)\n\t\t\\end{lstlisting}\n\t\tFor each grounded rule $r$, SBR builds a regularization loss term $L(r)$ in the following way. First, it maps each constant $c$ (e.g. \\textit{mary}) to a set of (perceptual) features $x_c$ (e.g. a tensor of pixel intensities $x_\\texttt{mary}$). Each relation $r$ (e.g. \\textit{smokes, stress}) is then mapped to a neural network $f_r(x)$, where $x$ is the tensor of features of the input constants and the output is a truth degree in $[0,1]$. For example, the atom \\textit{smokes(mary)} is mapped to the function call $f_\\texttt{smokes}(x_\\texttt{mary})$. \n\t\tThen, a fuzzy logic t-norm is selected and logic connectives are mapped to the corresponding real valued functions. For example, when the \\L ukasiewicz t-norm is selected, the implication is mapped to the binary real function $f(x,y) = \\min(1, 1 - x + y)$.\n\t\tFor the rule above, the Semantic-Based Regularization loss term is (for the \\L ukasiewicz t-norm):\n\t\t\\begin{equation*}\n\t\tL^{\\text{\\L}}(r) = \\min \\Big(1, 1 - f_\\texttt{stress}(x_\\texttt{mary}) + f_\\texttt{smokes}(x_\\texttt{mary}) \\Big)\n\t\t\\end{equation*}\n\t\tThe aim of Semantic-Based Regularization is to use the regularization term together a with classical loss function for supervised learning to learn the functions associated to the relations (here $f_\\texttt{stress}$ and $f_\\texttt{smokes}$).\n\t\tIt is worth comparing this method with the Semantic Loss (Example \\ref{ex:sl}). Both methods turn a logic formula (either propositional or first-order) to a real valued function that is used as a regularization term. However, because of the different semantics, these two methods have different properties. On the one hand, SL preserves the original logical semantics, by using probabilistic logic. However, due to the probabilistic assumption, the input formula cannot be compiled directly into a differentiable loss but needs to be first translated, i.e. compiled, into an equivalent deterministic and decomposable formula. While this step is necessary for the probabilistic model to be sound, the size of the resulting formula can be exponential in the size of the grounded theory. On the other hand, in SBR, the formula can be compiled directly into a differentiable loss, whose size is linear in the size of the grounded theory. However, in order to do so, the semantics of logic is altered, by turning it into fuzzy logic.\n\t\\end{boxedexample}\n\tFuzzy logic can also be used to relax rules. For example, in LRNN, $\\partial$ILP, DiffLog and the approach of , the scores of the proofs are computed using fuzzy logic connectives. \n\tThe theory t-norms has identifying parameterized (i.e. weighted) classes of t-norms that are very close to standard neural computation patterns (e.g. ReLU or sigmoidal layers). This creates an interesting, still not fully understood, connection between soft logical inference and inference in neural networks. A large class of methods relaxes logical statements numerically, without explicitly defining a specific semantics.\n\tUsually, the atoms are assigned scores in $\\mathbb{R}$ computed by a neural scoring function over embeddings.\n\tNumerical approximations are then applied either to combine these scores according to logical formulas or to aggregate proofs scores.\n\tThe resulting neural architecture is usually differentiable and, thus, trained end-to-end.\n\tSome NeSy methods, such as PSL, have used mixed probabilistic and fuzzy semantics. In particular, Deep Logic Models (DLM) extend PSL by adding neurally parameterized factors to the Markov field, while uses fuzzy logic to train posterior regularizers for standard deep networks using knowledge distillation . \n \\rev{The semantics of computational logic has also been explored and extended along other directions that have also been used within AI, for example, \\textit{modal} and \\textit{temporal} logics . While their analysis is out of the scope of the paper, it is worth mentioning that also such formalisms have been investigation from a neurosymbolic perspective .}","id":"0c28445b-24b7-4df7-a61f-1e0537492ed8","level":"subsection","origin_cites_number":23,"parent_id":"3f30c5c1-1c7c-4ae0-b5f0-88220243574e","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic - Semantics"],["subsection","Implications for NeSy"]],"subsections":[],"title":"Implications for NeSy"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:struct}\n\tLearning approaches in StarAI and NeSy are usually distinguished as to whether the structure~ or the parameters of the model are learned~.\n\tIn structure learning, the learning task is to discover the logical theory, i.e., a set of logical clauses and their corresponding probabilities or weights that reliably explains the examples. \n\tWhat \\textit{explaining the examples} exactly means depends on the learning setting.\n\tIn discriminative learning, we are interested in learning a theory that explains, or predicts, a specific target relation given background knowledge.\n\tIn generative learning, there is no specific target relation; instead, we are interested in a theory that explains the interactions between all relations in a dataset.\n\tIn contrast to structure learning, parameter learning starts with a given logical theory and only learns the corresponding probabilities or weights.\n\tStructure learning is an inherently NP-complete problem of searching for the right combinatorial structure, whereas parameter learning can be achieved with any curve fitting technique, such as gradient descent or least-squares.\n\tWhile parameter learning is, in principle, an easier problem to solve, it comes with a strong dependency on the provided user input. If the provided clauses are of low quality, the resulting model will also be of low quality.\n\tStructure learning, on the other hand, is less dependent on the user provided input, but is an inherently more difficult problem.","id":"3cf78a3e-98ed-45de-848f-3a61063e51b3","level":"section","origin_cites_number":3,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Structure versus Parameter Learning"]],"subsections":["88f64637-a474-4f43-9b9a-5ccb928849a4","9af5d6bb-eaff-4088-a25c-009703783367"],"title":"Structure versus Parameter Learning"},{"cite_extract_rate":0,"cites":[],"content":"Both structure and parameter learning are common in StarAI.\n\tStructure learning in StarA is an instance of learning by search~ and is closely connected to program synthesis.\n\tThe existing techniques are typically extensions of techniques originating in inductive logic programming (ILP) ~, which learn deterministic logical theories, and probabilistic graphical models (PGMs), which learn Bayesian or Markov networks from data.\n\tBeing an instance of learning by search, the central components of a learning framework are a space of valid structures and a search procedure.\n\tIn ILP, valid structures are logical theories; for Bayesian networks, valid structures are DAGs capturing their graph structure.\n\tThe resulting search space is then traversed with generic search procedures.\n\t StarAI structure learning techniques suffer from a combinatorial explosion.\n\tThat is especially the case with ILP techniques, in which the search space consists of programs containing several clauses. \n\tTherefore, it is necessary to limit the search space to make learning tractable.\n\tThe most common way to do this is to impose a \\textit{language bias} -- a set of instructions on how to construct the search space, such that it is narrowed down to a subset of the space of all logical theories.\n\tThough language bias can make the problem more tractable, it requires special care: too many restrictions might eliminate the target theory, while too few restrictions make the search space too large to traverse.\n\tAnother strategy is to leverage the compositionality of logic programs: \n adding an additional clause to a program increases its coverage and does not affect the prediction of examples covered by the initial program.\n\tThat is, we can learn a single clause at a time instead of simultaneously searching over theories containing multiple clauses.\n\tLearning clauses and their probabilities is usually treated as a two stage process.\n\tILP-based StarAI learning techniques first identify useful (deterministic) clauses, and then learn the corresponding probabilities or weights via parameter learning. Similarly, StarAI methods grounded primarily in PGMs, such as MLNs, search for frequently occurring cliques in data~, lift them into logical clauses, and then learn the weights or probabilities.\n\tParameter learning techniques are often also extensions of well known statistical approaches such as least-squares regression~, gradient descent~, and expectation maximisation~.\n\t\\begin{boxedexample}[label = ex:probfoil]{Structure learning with ProbFoil}\n\t\tAs an illustration of structure learning techniques, we will focus on ProbFoil~.\n\t\tAssume that we are interested in learning the definition of \\textit{grandparent} from akinship data.\n\t\tThat is, we are given a set of examples of grandparent relations\n\t\t\\begin{lstlisting}\n\t\tgrandparent(abe,lisa).\n\t\tgrandparent(abe,bart).\n\t\tgrandparent(jacqueline,lisa).\n\t\tgrandparent(jacqueline,bart).\n\t\t\\end{lstlisting}\n\t\tand background knowledge containing the following facts:\n \\begin{lstlisting}\n\tfather(homer,lisa). \n father(homer,bart). \n father(abe,homer).\n\tmother(jacqueline,marge). \n mother(marge,bart).\n\tmother(marge,lisa).\n\t\t\\end{lstlisting}\n\t\tProbFoil iteratively searches for a single clause that covers as many examples as possible, until all examples are covered or it adding more clauses does not improve the results. \n\t\tWhile searching for the best clause, it starts from the most general one, \\textit{grandparent(X,Y)}, which effectively states that every pair of people forms a grandparent relationship.\n\t\tThen it gradually specialises the clause by adding literals to the body. For instance, extending \\textit{grandparent(X,Y)} with a \\textit{mother/2} predicate results in the following clauses\n\t\t\\begin{lstlisting}\n\t\tgrandparent(X,Y) <- mother(X,Y).\n\t\tgrandparent(X,Y) <- mother(X,X).\n\t\tgrandparent(X,Y) <- mother(Y,X).\n\t\tgrandparent(X,Y) <- mother(Y,Y).\n\t\t\\end{lstlisting}\n\t\tExtending the initial clause with the \\textit{father/2} results in similar clauses.\n\t\tHaving the new candidate clauses, ProbFoil scores each candidate by counting how many positive and negative examples are covered.\n\t\tThese candidate clauses would not cover any examples and ProbFoil continues to refine the candidates by adding another literal to the body.\n\t\tThis would result in clauses of the following form:\n\t\t\\begin{lstlisting}\n\t\tgrandparent(X,Z) <- mother(X,Y), father(Y,Z).\n\t\tgrandparent(X,Z) <- mother(X,Y), mother(Y,Z).\n\t\tgrandparent(X,Z) <- father(X,Y), mother(Y,Z).\n\t\t...\n\t\\end{lstlisting}\n\t\tSome of the new candidates will cover only positive examples, such as \n \\begin{lstlisting}\n grandparent(X,Z) <- mother(X,Y), mother(Y,Z)\n \\end{lstlisting}\n that covers both examples\n \\begin{lstlisting}\n grandparent(jacqueline,lisa).\n grandparent(jacqueline,bart).\n \\end{lstlisting}\n\t\tHaving found one clause, ProbFoil learns the corresponding probability labels and adds the clause to the theory.\n\t\tProbFoil then repeats the same procedure, starting with the most general clause, to cover the remaining examples.\n\t\\end{boxedexample}","id":"88f64637-a474-4f43-9b9a-5ccb928849a4","level":"subsection","origin_cites_number":8,"parent_id":"3cf78a3e-98ed-45de-848f-3a61063e51b3","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Structure versus Parameter Learning"],["subsection","Implications for StarAI"]],"subsections":[],"title":"Implications for StarAI"},{"cite_extract_rate":0.6000000000000001,"cites":[5724,8942,2665,5710,5722,7989,3726,5723,3727],"content":"While StarAI learning techniques are categorised exclusively as either structure or parameter learning, \\nesy{} learning techniques combine both.\n\tWe will now discuss four groups of NeSy learning approaches: neurally-guided search, structure learning via parameter learning, program sketching, and implicitly structure learning.\n\t\\textit{Neurally guided structure search} is the \\nesy{} paradigm most similar to structure learning in StarAI.\n\tIt addresses one of the major weaknesses of StarAI structure learning methods - uninformed search over valid theories.\n\tInstead, neurally guided search relies on a \\textit{recognition model}, typically a neural network, to prioritise parts of the symbolic search space so that the target model can be found faster.\n\tGenerally speaking, the recognition model predicts the probability of a certain structure, e.g. a predicate or an entire clause, to be a part of the target model.\n\tFor instance, Deepcoder~ uses input-output examples to predict the probability of each predicate appearing in the target model.\n\tTherefore, Deepcoder turns a systematic search into an informed one by introducing a ranking over predicates in the search space.\n\tLikewise, EC$^{2}$~ derives the probability of a program solving the task at hand.\n\tSeveral approaches push this direction further and explore the idea of replacing an explicit symbolic model space with an implicit generative model over symbolic models ~. \n\tFor instance, in , the authors learn a generative model over grammar rules, conditioned on the examples.\n\tStructure learning is then performed by sampling grammar rules from the generative model, according to their probability, and evaluating them symbolically on the provided examples. \n\tThese approaches clearly show how symbolic search can be made tractable by introducing various forms of guidance via neural models.\n\tThese guidance-based approaches reduce, to a large extent, the most important weakness of symbolic structure learning approaches - the generation of many useless clauses or models.\n\tOn the other hand, these approaches often need large amounts of data for training, sometimes millions of examples even though creating data is relatively easy by enumerating random model structures and sampling examples from them~.\n\t\\begin{boxedexample}[label = ex:ngps]{Neurally-guided structure learning}\n\t\t\\rev{To illustrate neurally-guided search, we use the approach of Zhang et al. .\n StarAI techniques for structure learning typically perform a systematic search, which results in many useless models being tested.\n Given $N$ atoms, we can construct $N^l$ clauses of length $l$; this is an enormous space that is difficult to search efficiently.} \\\\\n \\rev{Zhang et al. sidestep the systematic search by introducing a neural network that chooses which programs to explore next.\n This search space is made of clauses such that an empty clause is at the top and children are extensions of the empty clause with all possible predicates; their children are further extensions with all individual atoms.} \\\\\n \\rev{The approach follows a top-down search strategy, exploring shorter clauses before longer ones, with a twist: instead of following a predefined order, the approach uses a neural network to decide which child to expand next. \n The approach can be viewed as a best-first search with a heuristic function implemented by a neural model. To this end, the network (1) encodes all literals in each clause separately, (2) scores all literals, (3) pools the scores of each literal per candidate, and (4) chooses the best candidate based on their scores.\n Ordering the search space in this way leads to substantial improvements in computation time, typically several orders of magnitude.}\n\t\\end{boxedexample}\n\tAn alternative way to reduce the combinatorial complexity of learning is to learn only a part of the program.\n\tThis is known as \\textit{program sketching:} a user provides an almost complete target model with certain parts being unspecified (known as \\textit{holes}). \n\tFor instance, when learning a model in the form of a (logic) program for sorting numbers or strings, the user might leave the comparison operator unspecified and provide the rest of the program.\n\tThe learning task is then to fill in the holes.\n\tExamples of \\nesy{} systems based on sketching are DeepProbLog and $\\partial$4, which fill in the holes in a (symbolic) program via neural networks.\n\tThe advantage of sketching is that it provides a nice interface for \\nesy{} systems, as the holes can be filled either symbolically or neurally.\n\tHoles provide a clear interface in terms of inputs and outputs and are agnostic to the specific implementation.\n\tThe disadvantage of sketching is that the user still needs to know, at least approximatively, the structure of the program.\n\tThe provided structure, the sketch, acts as a strong bias.\n\tDeciding which functionality is left as a hole is a non-trivial issue: as the sketch becomes less strict, the search space becomes larger. \n\t\\textit{Structure learning via parameter learning} (Example \\ref{ex:difflog}) is arguably the most prominent learning paradigm in \\nesy{}, positioned in between the two StarAI learning paradigms.\n\tStructure learning via parameter learning is technically equivalent to parameter learning in that the learning tasks consists of learning the probabilities of a fixed set of clauses.\n\tHowever, in contrast to StarAI in which the user carefully selects the informative clauses, the clauses are typically enumerated from user-provided templates of predefined complexity.\n\tConstructed in this way, the majority of clauses are noisy and erroneous and are of little use.\n\tThey would receive very low, but non-zero, probabilities.\n\t Approaches that follow this learning principle include NTPs , $\\partial$ILP , DeepProbLog, NeuralLP and DiffLog .\n\tThe advantage of structure learning via parameter learning is that it removes the combinatorial search from the learning.\n\tHowever, the number of clauses that needs to be considered is still extremely large, which leads to difficult optimisation problems (cf. ). Furthermore, irrelevant clauses are never removed from the model and are thus always considered during inference.\n\tThis can lead to spurious interactions even when low probabilities are associated to irrelevant clauses: as the number of irrelevant clauses is extremely large, their cumulative effect can be substantial.\n\t\\begin{boxedexample}[label = ex:difflog]{Structure learning via parameter learning}\n\t\tAs an illustration of structure learning via parameter learning, we focus on DiffLog .\n\t\tDiffLog expects the candidate clauses to be provided by the user.\n\t\tThe user can either provide the rules she knows are useful or construct them by using a clause template and instantiating it .. \n\t\tGiven a set of positive examples, DiffLog proceeds by constructing \\textit{derivation trees} for each example.\n\t\tConsider the problem of learning the connectivity relation over a graph.\n\t\tThe input tuples (background knowledge in StarAI terminology) specify edges in a graph\n\t\t\\begin{lstlisting}\n\tedge(a,b). edge(b,c). edge(b,d). edge(d,e). edge(c,f).\n\t\t\\end{lstlisting}\n\t\tThe examples indicate the connectivity relations among the nodes in the graph (for simplicity, consider only the following two examples)\n\t\t\\begin{lstlisting}\n\t\tconnected(a,b). connected(a,c). \n\t\t\\end{lstlisting}\n\t\tAlso assume that the candidate clause set contains the following clauses (with $p_1$ and $p_2$ their weights):\n\t\t\\begin{lstlisting}\n\t\t$p_1$::connected(X,Y) <- edge(X,Y). \n\t\t$p_2$::connected(X,Y) <- edge(X,Z), connected(Z,Y).\n\t\t\\end{lstlisting}\n\t\tDerivation trees are essentially proofs of individual examples that correspond to branches in the SLD-tree .\n\t\tFor instance, the example \\textit{connected(a,b)} can be proven using the first clause, whereas the example \\textit{connected(a,c)} can be proven by chaining the two clauses ($connected(a,c) \\leftarrow edge(a,b), connected(b,c)$ and $connected(b,c) \\leftarrow edge(b,c)$).\n\t\tDiffLog uses derivation trees to formulate the learning problem as numerical optimisation over the weights associated with the rules.\n\t\tMore precisely, DiffLog defines the probability of deriving an example as the product of the weights associated to the clauses used in the derivation tree of the corresponding example.\n\t\tFor instance, DiffLog would formulate the learning problem for the two examples as follows\n\t\t$$\\min_{p_1, p_2} \\ \\underbrace{(1 - p_1)}_{\\small \\tt connected(a,b)} + \\underbrace{(1 - p_1\\times p_2)}_{\\small \\tt connected(a,c)}.$$\n\t\\end{boxedexample}\n\tThe last group of approaches learns the structure of a program only \\textit{implicitly}.\n\tFor instance, Neural Markov Logic Networks (NMLN) , a generalisation of MLNs, extract structural features from relational data.\n\tWhereas MLNs define potentials only over cliques defined by the structure (logical formulas) of a model, NMLNs add potentials over \\textit{fragments} of data (projected over a subset of constants).\n\tNMLNs thus do not necessarily depend on the symbolic structure of the model, be it learned or provided by a user, but can still learn to exploit relational patterns present in data.\n\tMoreover, NMLNs can incorporate embeddings of constants.\n\tThe benefit of this approach is that it removes combinatorial search from learning and performs learning via more scalable gradient-based methods.\n\tHowever, one loses the ability to inspect and interpret the discovered structure.\n\tAdditionally, to retain tractability, NMLNs limit the size of fragments which imposes limits on the complexity of the discovered relational structure.","id":"9af5d6bb-eaff-4088-a25c-009703783367","level":"subsection","origin_cites_number":15,"parent_id":"3cf78a3e-98ed-45de-848f-3a61063e51b3","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Structure versus Parameter Learning"],["subsection","Implications for NeSy"]],"subsections":[],"title":"Implications for NeSy"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:symb_vs_subsymb}\n\tIn neurosymbolic artificial intelligence, approaches can be characterized by the way they represents entities and relationships in two classes: symbolic methods, where entities are represented using symbols such as strings and natural numbers, and subsymbolic methods, where entities are represented using numerical or distributed representations.\n\tSymbolic representations include constants $(an, bob)$, numbers $(4, -3.5)$, variables $(X, Y)$ and structured terms $f(t_1, ... ,t_n)$ where $f$ is a functor and the $t_i$ are \\rev{constants}, variables or structured terms. \n \\rev{Structured terms are a powerful construct that can represent arbitrary structures over entities, such as relations, lists or trees.}\n\tsubsymbolic AI systems, such as neural networks, require that entities are represented numerically using vectors, matrices or tensors. Throughout this this section, we will call these subsymbolic representations or subsymbols. \n\tsubsymbolic AI systems usually require that these representations have a fixed size and dimensionality. Exceptions require special architectures and are still the subject of active research (e.g. RNNs for list-like inputs or GCNs~ for graph-type inputs).","id":"bd238fce-3df5-4531-ba35-f4316c49d81b","level":"section","origin_cites_number":1,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Symbolic vs subsymbolic representations"]],"subsections":["9758e29f-afe8-4348-b5ab-e226dfae12d2","a1474eee-df30-4fdc-964c-c5b1bb43d21d"],"title":"Symbolic vs subsymbolic representations"},{"cite_extract_rate":0,"cites":[],"content":"A powerful and elegant mechanism for reasoning with symbols in logic is {\\em unification}. Essentially, it calculates the most general substitution that makes two symbols syntactically equal, if it exists. \\rev{This does not allow one to compare two different entities, but allows one to find what two structured terms have in common.}\n\tFor instance, the terms $p(a,Y)$ and $p(X,b)$ can be unified using the substitution $\\{X=a, Y=b\\}$. \n\tConversely, due to their numerical nature, calculating the similarity between subsymbols is straightforward. \n\tSimilarity metrics such as the radial-basis function or distance metrics such as the L1 and L2 norm can be used. However, it is not clear when to decide that two subsymbolically represented entities are the same.","id":"9758e29f-afe8-4348-b5ab-e226dfae12d2","level":"paragraph","origin_cites_number":0,"parent_id":"bd238fce-3df5-4531-ba35-f4316c49d81b","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Symbolic vs subsymbolic representations"],["paragraph","Comparing representations"]],"subsections":["a5362428-8a2e-441c-a432-e33fed8ed15b"],"title":"Comparing representations"},{"cite_extract_rate":0.6666666666666661,"cites":[1166,1168],"content":"Many systems need to translate back and forth between symbolic and subsymbolic representations. In fact, a lot of research on deep learning is devoted to efficiently representing symbols so that neural networks can properly leverage them.\n\tA straightforward example is to translate symbols to a subsymbolic representation that can serve as input for a neural network. Generally, these symbols are replaced by a one-hot encoding or by learned embeddings. \n \\rev{Note, however, that this does not imply that the system can perform symbolic manipulation on this input. Rather, it serves as an index to a set of learned, latent embeddings.}\n\tA more interesting example is encoding relations in subsymbolic space. The wide variety of methods~ developed for this purpose indicates that this is far from a solved problem. \n\tDifferent encodings have different benefits. For example, TransE~ encodes relations as vector translations from subject to object embeddings. A disadvantage is that symmetric relations are represented by the null vector, and entities in symmetric relations are pushed towards each other.\n\tMore complex structures are even harder to represent. For example, there is currently a lot of research in how to utilize graph-structured data in neural networks (cf. \\ref{sec:kge_gnn}).\n\tTranslating from a subsymbolic representation back to a symbolic one happens, for example, at the end of a neural network classifier. Here, a subsymbolic vector needs to be translated to discrete classes. Generally, this happens through the use of a final layer with a soft-max activation function which then models the confidence scores of these classes as a categorical distribution. However, other options are possible. For example, some methods are only interested in the most likely class, and will use an arg-max instead. Alternatively, a Gumbel-softmax activation can be used as a differentiable approximation of sampling from the categorical distribution.","id":"a5362428-8a2e-441c-a432-e33fed8ed15b","level":"paragraph","origin_cites_number":3,"parent_id":"9758e29f-afe8-4348-b5ab-e226dfae12d2","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Symbolic vs subsymbolic representations"],["paragraph","Comparing representations"],["paragraph","Translating between representations"]],"subsections":[],"title":"Translating between representations"},{"cite_extract_rate":0.7272727272727271,"cites":[5708,5707,5711,5726,7304,8943,5725,3727],"content":"In StarAI systems, the input, intermediate and output representations are all using the same symbolic representations. \\rev{Although there are StarAI systems that can support numerical values, these are still treated as symbols, which is different than a latent, subsymbolic representation.}\n\tIn neural systems, the input and intermediate representations are subsymbolic. The output representation can be either symbolic (e.g. classifiers) or subsymbolic (e.g. auto-encoders, GANs).\n\tThe most important aspect of neurosymbolic systems is that they combine symbolic and subsymbolic representations.\n\tNeSy systems can be categorized by how they do this. \n We distinguish several approaches.\n\tIn the first approach, the inputs are symbolic, but they are translated to subsymbols in a single translation step, after which the intermediate representations used during reasoning are purely subsymbolic. This approach is followed by the majority of NeSy methods. Some examples include Logic Tensor Networks~, Semantic-based Regularization~, Neural Logic Machines~ and TensorLog~.\n\t\\begin{boxedexample}[label = ex:ltn]{Logic Tensor Networks}\n\t\tLogic tensor networks~ make this translation step explicit. The authors introduce the concept of a \\textit{grounding} (not to be confused with the term grounding used in logic). Here, a grounding is a mapping of all symbolic entities onto their subsymbolic counterpart. More formally, the authors define a grounding as a mapping $\\mathcal{G}$ where:\n\t\t\\begin{itemize}\n\t\t\t\\item $\\mathcal{G}(c) \\in \\mathbb{R}^n$ for every constant symbol $c$\n\t\t\t\\item $\\mathcal{G}(f) \\in \\mathbb{R}^{n.m} \\rightarrow \\mathbb{R}^n$ for every function $f$ of arity $n$\n\t\t\t\\item $\\mathcal{G}(p) \\in \\mathbb{R}^{n.m} \\rightarrow [0,1]$ for every predicate $p$ of arity $n$\n\t\t\\end{itemize}\n\t\tThe grounding of a clause is then performed by combining the aforementioned groundings using a t-norm.\n\t\\end{boxedexample}\n\tIn the second approach, intermediate representations are both symbolic and subsymbolic, but not simultaneously. This means that some parts of the reasoning work on the subsymbolic representation, and other parts deal with the symbolic representation, but not at the same time.\n\tThis is indicative of NeSy methods that implement an interface between the logic and neural aspect.\n\tThis approach is more natural for systems that originate from a logical framework such as DeepProbLog~, NeurASP~), ABL~ and NLog~.\n\t\\begin{boxedexample}[label = ex:abl]\n{ABL}\n\t\tIn ABL~ there are three components that function in an alternating fashion. There is a perception model, a consistency checking component and an abductive reasoning component.\n\t\tTake for example the task where there are 3 MNIST images that need to be recognized such that the last is the result of applying an operation on the first two (e.g. $\\digit{3}+\\digit{5}=\\digit{8}$). The structure of the expression is given as background knowledge, but the exact operation (addition) needs to be abduced.\n\t\tFirst, the perception model classifies the images into pseudo-labels, using the most likely prediction (i.e. arg-max). \n\t\tThe abductive reasoning component then tries to abduce a logically consistent hypothesis. \n\t\tFor example, if the digits are correctly classified as $3$, $5$ and $8$, the only logically consistent hypothesis is that the operation is an addition.\n\t\tIf this is not possible, there is an error in the pseudo-labels. \n\t\tA heuristic function is then used to determine which pseudo-labels are wrong.\n\t\tThe reasoning module then searches for logically consistent pseudo-labels. These revised pseudo-labels are then used to retrain the perception model.\n\t\\end{boxedexample}\n\tIn the final approach, intermediate representations are considered simultaneously as symbolic and subsymbolic by the reasoning mechanism.\n This is implemented in only a few methods, such as the NTP and the CTP. \n\t\\begin{boxedexample}[label = ex:ntp]{Neural Theorem Prover}\n\t\tIn the Neural Theorem Prover, two entities can be unified if they are similar, and not just if they are identical. As such, the NTP interweaves both symbols and subsymbols during inference. For each symbol $S$, there is a learnable subsymbol $T_S$.\n\t\tSoft-unification happens by applying the normal unification procedure where possible. However, if two symbols $S_1$ and $S_2$ can not be unified, the comparison is assigned a score based on the similarity between $T_{S_1}$ and $T_{S_2}$. The similarity is calculated using a radial basis function $\\varphi(||x-y||_2)$.\n\t\tFor example, to unify \\texttt{mother(an,bob)} and \\texttt{parent(X,bob)}, soft-unification proceeds as follows:\n\t\t\\begin{align*}\n\t\t\\{\\mathtt{mother(an,bob)} &= \\mathtt{parent(X,bob)}\\} \\\\\n\t\t&\\Downarrow \\quad \\varphi(\\mathtt{mother}, \\mathtt{parent}) \\\\\n\t\t\\{\\mathtt{an} = \\mathtt{X}&, \\mathtt{bob} = \\mathtt{bob}\\} \\\\\n\t\t&\\Downarrow \\quad X = an \\\\\n\t\t\\{\\mathtt{bob} &= \\mathtt{bob}\\} \\\\\n\t\t&\\Downarrow\\\\\n\t\t&\\{~\\}\n\t\t\\end{align*}\n\t\tSoft-unification is not only used to learn which constants and predicates are similar, but can also be used to perform rule learning. By adding new, parameterized rules with unique predicates, soft-unification allows these new predicates to become very similar to other predicates and as such behave as newly introduced rules.\n\t\tFor example, consider the program consisting of the fact \\textit{mother(an,bob)} and a single parameterized rule $r1(X,Y) \\leftarrow r2(Y,X)$. The Neural Theorem Prover can answer the query \\textit{child(bob,an)} as follows: \n\t\t\\begin{center}\n\\tikzset{every picture/.style={line width=0.75pt}} \n\\begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1]\n\\draw (240,60) -- (140,100) ;\n\\draw (240,60) -- (340,100) ;\n\\draw (340,140) -- (340,180) ;\n\\draw (201,32) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\mathtt{child(bob,an)}$};\n\\draw (299,111) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\mathtt{r2(ann,bob)}$};\n\\draw (335,191) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\mathtt{T}$ };\n\\draw (131,112) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\mathtt{T}$ };\n\\draw (151,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(1)};\n\\draw (299,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(2)};\n\\draw (349,151) node [anchor=north west][inner sep=0.75pt] [align=left] {(3)};\n\\end{tikzpicture}\n\t\t\\end{center}\n\t\t\\begin{align*}\n\t\t(1) ~& \\mathtt{child} = \\mathtt{mother} &\\varphi(||T_{child}-T_{mother}||_2)\\\\\n\t\t& \\mathtt{bob} = \\mathtt{an} &\\varphi(||T_{an}-T_{bob}||_2)\\\\\n\t\t(2)~ & \\mathtt{child} = \\mathtt{r1} &\\varphi(||T_{child}-T_{r1}||_2)\\\\\n\t\t(3) ~& \\mathtt{child} = \\mathtt{r2} &\\varphi(||T_{r2}-T_{mother}||_2)\n\t\t\\end{align*}\n\t\tThe figure above shows the two possible derivations the neural theorem prover can make to infer \\textit{child(bob, an)}. One the one hand, it can soft-unify with the fact \\textit{mother(an, bob)}, where \\textit{mother} unifies with \\textit{child} and \\textit{an} with \\textit{bob}. On the other hand, it can use the parameterized rule which encodes an inverse relation. In that case, \\textit{mother} unifies with \\textit{r1} and \\textit{r2} with \\textit{child}. If we optimize the subsymbolic embeddings for the latter, this will be equivalent to learning the rule $mother(X,Y) \\leftarrow child(Y,X)$. This example also shows that soft-unification potentially adds a lot of different proofs, which can result in computational problems. This problem was solved in later iterations of the system .\n\t\\end{boxedexample}\n\\tikzset{every picture/.style={line width=0.75pt}} \n\\begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1]\n\\draw (240,60) -- (140,100) ;\n\\draw (240,60) -- (340,100) ;\n\\draw (340,140) -- (340,180) ;\n\\draw (201,32) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\mathtt{child(bob,an)}$};\n\\draw (299,111) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\mathtt{r2(ann,bob)}$};\n\\draw (335,191) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\mathtt{T}$ };\n\\draw (131,112) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\mathtt{T}$ };\n\\draw (151,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(1)};\n\\draw (299,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(2)};\n\\draw (349,151) node [anchor=north west][inner sep=0.75pt] [align=left] {(3)};\n\\end{tikzpicture}","id":"a1474eee-df30-4fdc-964c-c5b1bb43d21d","level":"subsection","origin_cites_number":11,"parent_id":"bd238fce-3df5-4531-ba35-f4316c49d81b","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Symbolic vs subsymbolic representations"],["subsection","Implications for StarAI and NeSy"]],"subsections":[],"title":"Implications for StarAI and NeSy"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:paradigms}\n\tWhen two or more paradigms are integrated, examining which of the base paradigms are preserved, and to which extent, tells us a lot about the strengths and weaknesses of the resulting paradigm.\n\tIt has been argued that when combining different perspectives in one model or framework, such as logic, probabilistic and neural ones, it is desirable to have the original paradigms as a special case.\n\t\\rev{In this section, we analyze to which extent different models in StarAI and NeSy preserve the three basic paradigms. Intuitively, with preserving we mean to which extent one can exactly replicate the model and inference algorithm of the original paradigm. We will use the capital letters \\textit{L, P} and \\textit{N} to label systems where the logic, probability and neural paradigms can be recovered in full. We will use lowercase letters (i.e. \\textit{l, p} and \\textit{n}) when a method only partially recovers these paradigms, i.e. retain some but not all of the features.} The absence of a letter means that the paradigm is not considered by an approach.","id":"322e754b-6517-46c9-9006-90e863943e5b","level":"section","origin_cites_number":1,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic vs Probability vs Neural"]],"subsections":["64c621c6-ac4f-48ef-b44f-303dd6942485","b873315e-f300-4fef-8037-818fbe2be872"],"title":"Logic vs Probability vs Neural"},{"cite_extract_rate":0,"cites":[],"content":"Traditionally, StarAI focused on the integration of logic and probability.\n\t\\textbf{\\textit{lP:}}\n\tThe classical knowledge-based model construction approach uses logic only to generate a probabilistic graphical model. \n\tThus the graphical model can be used to define the semantics of the model and also to perform inference.\n\tThis can make it harder to understand the effects of applying logical inference rules to the model. \n\tFor instance, in MLNs, the addition of the resolvent of two weighted rules makes it hard to predict the effect on the distribution. \n\t\\textbf{\\textit{Lp}:} On the other hand, the opposite holds for probabilistic logic programs (PLPs) and their variants. \n\tWhile the effect of a logical operation is clear, it is harder to identify and exploit properties such as conditional or contextual independencies, that are needed for efficient probabilistic inference.","id":"64c621c6-ac4f-48ef-b44f-303dd6942485","level":"subsection","origin_cites_number":0,"parent_id":"322e754b-6517-46c9-9006-90e863943e5b","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic vs Probability vs Neural"],["subsection","StarAI: Logic + Probability"]],"subsections":[],"title":"StarAI: Logic + Probability"},{"cite_extract_rate":0.625,"cites":[5727,5710,5711,5719,3726],"content":"In NeSy, we consider a third paradigm: neural computation.\n\tWith neural computation, we refer mainly to the set of models and techniques that allows for exploiting (deep) latent spaces to learn intermediate representations. This includes dealing with perceptual inputs and also dealing directly with embeddings of symbols. \n\t\\textbf{\\textit{lN}}: Many \\nesy{} approaches focus on the neural aspect (i.e., they originated as a neural method to which logical components have been added). For example, LTNs and SBRs turn the logic into a regularization function to provide a penalty whenever the logical constraints are violated. At test time the logical loss component is dropped and only the network is used to make predictions. Moreover, by using fuzzy logic, these methods do not integrate the probabilistic paradigm. \n\t\\textbf{\\textit{Ln}}: Another class of \\nesy{} methods does retain the focus on logic. \n\tThese methods usually expand an existing logical framework into a differentiable version. Examples include LRNNs , TensorLog , DiffLog , $\\partial$ILP , $\\partial$4 and NTPs .\n\tThe key inference concepts are mapped onto an analogous concept that behaves identically for the edge cases but is continuous and differentiable in non-deterministic cases. \n\tAs described in the previous sections, many such systems cast logical inference as forward or backward chaining. The focus on logic is clear if one considers that logical inference is performed symbolically to build the network and the semantics is relaxed only in a subsequent stage to learn the parameters. \\rev{While the architecture mimics the logical reasoning, it is often far from the deep-stacked architecture of neural networks.}\n \\textbf{\\textit{LN}}: \\rev{It is worth mentioning a later iteration of LRNN, where the framework has been extended to allow for tensorial weights on atoms and custom aggregation functions . In that framework, it is shown how specifying logic rules can be regarded as specifying the layers of a deep architecture. This provides a nice and complete integration between forward-chaining logical reasoning and neural networks\n that is able to implement any existing neural architecture.\n }\n\t\\textbf{\\textit{lPN} and \\textit{LpN}} There are two final classes of methods that start from existing StarAI methods, \\textit{lP} and \\textit{Lp} respectively, and extend them with primitives that can be interfaced with neural networks and allow for differentiable operations.\n\tIn the \\textit{lPN} class, NeSy methods such as SL, RNMs and NMLNs follow the knowledge-based model construction paradigm. In the \\textit{LpN} class, methods such as DeepProbLog and NeurASP extend PLP. \nThere is usually a trade-off that one must make: \t\nsystems in the \\textit{lN} or \\textit{Ln} classes are usually more scalable but \\textit{(i)} do not model a probability distribution and \\textit{(ii)} often relax the logic. On the contrary, \\textit{LpN} or \\textit{lPN} systems preserve the original paradigms but at the cost of more complex inference (e.g. they usually resort to exact probabilistic inference).\n\tAn aspect that significantly aids in developing a common framework, and analysing its properties, is the development of an intermediate representation language that can serve as a kind of \\emph{assembly language} . \n\tOne such idea concerns performing probabilistic inference by mapping it onto a weighted model counting (WMC) problem. \n\tThis can then in turn be solved by compiling it into a structure (e.g. an arithmetic circuit) that allows for efficient inference.\n\tThis has the added benefit that this structure is differentiable, which facilitates the integration between logic based systems and neural networks. \n\t StarAI based systems often use this approach.","id":"b873315e-f300-4fef-8037-818fbe2be872","level":"subsection","origin_cites_number":8,"parent_id":"322e754b-6517-46c9-9006-90e863943e5b","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Logic vs Probability vs Neural"],["subsection","NeSy: Logic + Probability + Neural"]],"subsections":[],"title":"NeSy: Logic + Probability + Neural"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:tasks}\n\t\\rev{In this Section, we analyze the learning tasks to which the \\nesy{} models considered in this paper have been applied.}","id":"528b4897-8064-4ac1-a40f-dc1e7d30614f","level":"section","origin_cites_number":0,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Tasks"]],"subsections":["d6105bd2-5f38-4f97-a7bd-ae4de168c174"],"title":"Tasks"},{"cite_extract_rate":0.5714285714285711,"cites":[181,7304,1806,3727],"content":"\\rev{A classical task in NeSy is to use logic as distant supervision for a learning model. Here, input $X$ is paired with label $y$. However, instead of using a single model to map $X$ to $y$, the input $X$ is firstly mapped to a set of intermediate concepts $C$ by a (set of) neural networks. Then, these concepts are used to compute $y$ in a symbolic way. Logic programs are usually exploited to map the concepts $C$, represented as logical atoms, to the label $y$, which represents the logical query. Therefore, the neural networks are not directly supervised (on $C$) but they are only distantly supervised through the label $y$ and the knowledge contained in the logic program. The intuition is that when the label $y$ is only weakly linked to the input, it is more convenient to break the task in several easier subtasks and then compose them using background knowledge in the form of a logic program. Notice that the logic program is fundamental for the inference. Without the program, the networks will not be able to solve their subtasks, as there is no direct supervision. Moreover, by splitting the task into subtasks, the inference done by the composite system (neural + logic) is far more explainable than a corresponding end-to-end neural network. A classical example is the MNIST addition , shown in Example \\ref{ex:mnist_addition}. Distant supervision tasks are very common in prototypical systems such as DeepProbLog, DeepStochLog, NLog, NeurASP, SATNet . A downside of such tasks is that, to enable learning of untrained neural subtasks, the logic has to consider all possible combinations of concepts that are compatible with the label $y$, even though only few (or one) are correct. The challenge is to balance the exploration of multiple combinations with a greedy strategy for scaling to larger problems . Other problems falling in this category are scene parsing, image segmentation and semantic image interpretation }\n \\begin{boxedexample}[label = ex:mnist_addition]{MNIST Addition} \n Given the classical MNIST dataset, $\\mathcal{D} = \\{(x_i,y_i)\\}$, with $x_i$ an MNIST image, and $y_i$ its numeric label, the MNIST addition dataset is built by mapping pairs of images to the label representing their sum. In particular, $\\mathcal{D}_{\\text{add}} = \\{(x_i,x_j,z_{ij}) : z_{ij} = y_i + y_j \\land (x_i,y_i),(x_j,y_j) \\in \\mathcal{D}\\}$. The idea is to learn to classify the digits without direct supervision on their labels, but only using distant supervision about sums of such images. \n The task is often also coupled to background knowledge of what addition is, e.g. in Prolog syntax:\n \\begin{lstlisting}\naddition(X1, X2, Z) <- digit(X1,Y1), digit(X2,Y2), Z is Y1 + Y2. \n \\end{lstlisting}\n Such knowledge is used to reason about the (most-likely) pairs \\texttt{Y1,Y2} that sum to the provided label \\texttt{Z}. Logic is then used to link the actual outputs of the learning model \\texttt{Y1,Y2} to the distant supervision \\texttt{Z}.\n \\end{boxedexample}","id":"d6105bd2-5f38-4f97-a7bd-ae4de168c174","level":"paragraph","origin_cites_number":7,"parent_id":"528b4897-8064-4ac1-a40f-dc1e7d30614f","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Tasks"],["paragraph","Distant Supervision"]],"subsections":["d9eb4e04-ce78-4d16-9f72-94154473deea"],"title":"Distant Supervision"},{"cite_extract_rate":0.5714285714285711,"cites":[5714,3741,5720,5709],"content":"\\rev{A related class of tasks is semi-supervised classification } with knowledge. Here, the starting point is a standard classification task, where a set of inputs $X$ is mapped by a neural model to a set of labels $C$. However, we are also provided with some additional knowledge $y$ related to the labels $C$ of the inputs. This knowledge is often expressed in terms of logical rules and programs. The setting is very similar to distant supervision, where we have three levels: inputs $X$, concepts $C$ and additional labels $y$. However, in this case, we have also access to supervision for some (usually few) concepts $C$. Although this task could be tackled in a purely supervised way by discarding the information contained in $y$, \\nesy{} approaches can improve the predictions of several input patterns using the external knowledge. When the external knowledge is relating concepts $C$ of multiple patterns, the task is called collective classification , as one can improve the accuracy on multiple patterns by collectively predicting their classes. A classical example in this setting is document classification in citation networks, cf. Example \\ref{ex:citation}. By treating the information contained in $y$ as extra knowledge, these tasks are often tackled using regularization based systems, like SBR, DLM, RNM or Semantic Loss. However, logic programs can also be used to simulate a label-passing scheme along the citation network, as done in DeepStochLog . A characteristic of this class of tasks is that the additional information $y$ is often very noisy (e.g. the manifold rule in the citation network is not always valid). While this task is closely related to distant supervision, there is an important difference: in semi-supervised classification, the additional knowledge $y$ is meant to provide an additional signal, which, however, would not suffice in the absence of direct supervision on the concepts $C$. \n \\begin{boxedexample}[label = ex:citation]{Document classification in citation networks}\n In document classification in citation networks, we are provided with both labelled and unlabelled scientific papers. A label is often the domain area of the paper (e.g. Machine Learning, Artificial Intelligence, Databases, etc.). \n However, a network of citations between papers is also provided, linking papers between domains.\n The idea of the document classification task is that in many domains, a paper cited by other papers with a certain label is likely to belong to the same domain. When classifying a document, one has to balance the signal coming from the features of the document (i.e. words) and that coming from neighbors in the citation network to provide a \\textit{collective} prediction.\n In \\nesy{} systems, this is usually done by coupling the subsymbolic model with a rule of the following type:\n \\begin{lstlisting}\n w:: domain(X,Y) <- cite(X,X1), domain(X1,Y). \n \\end{lstlisting}\n The rules get a different weight according to the domain to account for the differences between them. \n \\end{boxedexample}","id":"d9eb4e04-ce78-4d16-9f72-94154473deea","level":"paragraph","origin_cites_number":7,"parent_id":"d6105bd2-5f38-4f97-a7bd-ae4de168c174","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Tasks"],["paragraph","Distant Supervision"],["paragraph","Semi-supervised Classification"]],"subsections":["c47d6a00-5a1b-40c6-9278-09edc1c79d9b","d7c376c1-f6e7-41a6-9b6e-28a3f939810e","f4d6de92-316d-4daa-a0ba-2c746822d5de"],"title":"Semi-supervised Classification"},{"cite_extract_rate":0.5,"cites":[8942,5710,5711,5720],"content":"\\rev{Another common task in \\nesy{} is knowledge graph completion (KGC) or link prediction. A knowledge graph (KG) is a pair of $(E,R)$, where $N$ is the set of entities and $R$ the set of edges. In a KG, an edge is a triple $(e_1, r, e_2)$, where $e_1$ and $e_2$ are the head and tail of the edge and $r$ is the relation between them. In a KGC task, the goal is to predict missing edges in the input graph. Link prediction has been one of the key tasks in StarAI , and more recently also in NeSy as NeSy allows to merge symbolic reasoning (from StarAI) with the recent geometric deep learning approaches based on Knowledge Graph Embeddings (KGE) and Graph Neural Networks . \\nesy{} systems focusing on this task include NTPs , NMLN , DLM , DiffLog , TensorLog .}","id":"c47d6a00-5a1b-40c6-9278-09edc1c79d9b","level":"paragraph","origin_cites_number":8,"parent_id":"d9eb4e04-ce78-4d16-9f72-94154473deea","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Tasks"],["paragraph","Distant Supervision"],["paragraph","Semi-supervised Classification"],["paragraph","Knowledge Graph Completion"]],"subsections":[],"title":"Knowledge Graph Completion"},{"cite_extract_rate":0.6666666666666661,"cites":[8942,7988],"content":"Most previously mentioned tasks can be described as classification\\footnote{Even though, many of them use a generative model to tackle the classification task, instead of a conditional one.}. \\nesy{} has recently also focused on tasks concerned with modeling the input data distribution as accurately as possible. The goal is then to sample new patterns from the learned distribution. The idea behind \\nesy{} generative approaches is that one can learn important features from data using deep generative models (e.g. variational auto-encoders or Markov Chain Monte Carlo methods). Combining symbolic features with logic reasoning can be used to control, stratify and simplify the inference. The generative modeling can either refer to the relational structure, e.g. molecule generation in NMLNs , or to the subsymbolic space, e.g. image generation in VAEL or .","id":"d7c376c1-f6e7-41a6-9b6e-28a3f939810e","level":"paragraph","origin_cites_number":3,"parent_id":"d9eb4e04-ce78-4d16-9f72-94154473deea","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Tasks"],["paragraph","Distant Supervision"],["paragraph","Semi-supervised Classification"],["paragraph","Generative Tasks"]],"subsections":[],"title":"Generative Tasks"},{"cite_extract_rate":0.7142857142857141,"cites":[2665,5710,5728,3726,3727],"content":"\\rev{Rather than exploiting symbolic knowledge predictive tasks, one can also induce \n \\textit{symbolic knowledge}. In all previous tasks, symbolic knowledge is provided by the user as part of the input. However, as explored in Section \\ref{sec:struct}, we can still apply several neurosymbolic techniques by learning the symbolic knowledge when this is not the case. The unknown symbolic knowledge is then the actual target to be learned. A classical example is \\textit{program synthesis}, where the goal is to learn the program from positive and negative examples of the desired input-output behaviour. Ideally, all positive pairs and none of the negatives should be covered. Many systems learn logic programs, i.e. NTPs , $\\partial$ILP , DeepProbLog, NeuralLP, DiffLog, DeepCoder.\n Sometimes, the input-output pairs are not part of the training dataset, but are actually generated by a black-box neural model. The induced programs then explain the behaviour of the model, which relates NeSy to the domain of \\textit{explainability} .}","id":"f4d6de92-316d-4daa-a0ba-2c746822d5de","level":"paragraph","origin_cites_number":7,"parent_id":"d9eb4e04-ce78-4d16-9f72-94154473deea","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Tasks"],["paragraph","Distant Supervision"],["paragraph","Semi-supervised Classification"],["paragraph","Knowledge Induction"]],"subsections":[],"title":"Knowledge Induction"},{"cite_extract_rate":0,"cites":[],"content":"To conclude, we list some interesting challenges for \\nesy{}.","id":"35970a41-0514-46a0-8d9d-63d0769ff87a","level":"section","origin_cites_number":0,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"]],"subsections":["a1d98b58-9672-4eb5-80a4-d72f0adf272e"],"title":"Open Challenges"},{"cite_extract_rate":0,"cites":[],"content":"The statistical relational AI and probabilistic graphical model communities have devoted a lot of attention to the semantics of its models. This has resulted in several clear choices (such as directed vs. undirected, trace-based vs. possible world ), with corresponding strengths and weaknesses that clarify the relationships between the different models. Workshops have been held on this topic\\footnote{For instance, \\url{https://pps2018.luddy.indiana.edu/}}. Furthermore, some researchers have investigated how to transform one type of model into another . At the same time, weighted model counting has emerged as a common assembly language for inference. The situation in neurosymbolic computation today is very much like that of the early days in statistical relational learning, in which there were many competing formalisms, sometimes characterized as the statistical relational learning alphabet soup. It would be great to get more insight into the semantics of neurosymbolic approaches and their relationships. This survey hopes to contribute towards this goal.","id":"a1d98b58-9672-4eb5-80a4-d72f0adf272e","level":"paragraph","origin_cites_number":2,"parent_id":"35970a41-0514-46a0-8d9d-63d0769ff87a","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"],["paragraph","Semantics"]],"subsections":["6f37ac7d-158b-4d60-89b1-c3be4d76632d"],"title":"Semantics"},{"cite_extract_rate":0,"cites":[],"content":"Although relatively few methods explore the integration of logical and neural methods from a probabilistic perspective, we believe that a probabilistic approach is a very natural way to integrate the two, since it has been shown how one can recover the single methods as special cases. However, many open questions remain. Probabilistic inference is computationally expensive, usually requiring approximations. It would be interesting to determine exactly how probabilistic approximate inference compares with other approximations based on relaxations of the logic, like fuzzy logic.","id":"6f37ac7d-158b-4d60-89b1-c3be4d76632d","level":"paragraph","origin_cites_number":1,"parent_id":"a1d98b58-9672-4eb5-80a4-d72f0adf272e","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"],["paragraph","Semantics"],["paragraph","Probabilistic reasoning"]],"subsections":["2225ccc8-0707-44bd-8f81-0bdebfb52224","422c662c-f7da-4490-a64b-1d55a3d7423a","588fd0c9-1da5-47a9-91a7-ef4be62e8832","9539d3a1-1ce6-40b0-8897-7d2fd89a6849","986b4d82-d655-4355-b6bd-d14bee40b72f"],"title":"Probabilistic reasoning"},{"cite_extract_rate":0,"cites":[],"content":"The selection of the t-norm fuzzy logic and the corresponding translation of the connectives is very heterogeneous in the literature. It is often unclear which properties of Boolean logic a model is preserving, while there is a tendency to consider fuzzy logic as a continuous surrogate of Boolean logic without considering implications for the semantics. There is a clear need for further studies in this field. On the one hand, one may want to define new models which are natively fuzzy, thus not requiring a translation from Boolean logic. On the other hand, an interesting research direction concerns the characterisation of what are appropriate fuzzy approximations of Boolean logic relative to a set of properties that one wants to preserve (see Section \\ref{sec:semantics}).","id":"2225ccc8-0707-44bd-8f81-0bdebfb52224","level":"paragraph","origin_cites_number":0,"parent_id":"6f37ac7d-158b-4d60-89b1-c3be4d76632d","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"],["paragraph","Semantics"],["paragraph","Probabilistic reasoning"],["paragraph","Fuzzy semantics"]],"subsections":[],"title":"Fuzzy semantics"},{"cite_extract_rate":0,"cites":[],"content":"While significant progress has been made on learning the structure of purely relational models (without probabilities), learning StarAI models remains a major challenge due to the complexity of inference and the combinatorial nature of the problem.\n\tIncorporating neural aspects complicates the problem even more.\n\t\\nesy{} methods have certainly shown potential for addressing this problem (Section \\ref{sec:struct}), but the existing methods are still limited and mostly domain-specific which impedes their wide application.\n\tFor instance, the current systems that support structure learning require user effort to specify the clause templates or write a sketch of a model.","id":"422c662c-f7da-4490-a64b-1d55a3d7423a","level":"paragraph","origin_cites_number":0,"parent_id":"6f37ac7d-158b-4d60-89b1-c3be4d76632d","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"],["paragraph","Semantics"],["paragraph","Probabilistic reasoning"],["paragraph","Structure learning"]],"subsections":[],"title":"Structure learning"},{"cite_extract_rate":0,"cites":[],"content":"Scalable inference is a major challenge for StarAI and therefore also for \\nesy{} approaches with an explicit logical or probabilistic reasoning component.\n\tInvestigating to what extent neural methods can help with this challenge by means of lifted (exploiting symmetries in models) or approximate inference, as well as reasoning from intermediate representations , are promising future research directions.","id":"588fd0c9-1da5-47a9-91a7-ef4be62e8832","level":"paragraph","origin_cites_number":1,"parent_id":"6f37ac7d-158b-4d60-89b1-c3be4d76632d","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"],["paragraph","Semantics"],["paragraph","Probabilistic reasoning"],["paragraph","Scaling inference"]],"subsections":[],"title":"Scaling inference"},{"cite_extract_rate":0,"cites":[],"content":"A major advantage of StarAI methods, as compared to neural ones, is their data efficiency -- StarAI methods can efficiently learn from small amounts of data, whereas neural methods are data hungry.\n\tOn the other hand, StarAI methods do not scale to big data sets, while neural methods can easily handle them.\n\tWe believe that understanding how these methods can help each other to overcome their complementary weaknesses, is a promising research direction.","id":"9539d3a1-1ce6-40b0-8897-7d2fd89a6849","level":"paragraph","origin_cites_number":0,"parent_id":"6f37ac7d-158b-4d60-89b1-c3be4d76632d","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"],["paragraph","Semantics"],["paragraph","Probabilistic reasoning"],["paragraph","Data efficiency"]],"subsections":[],"title":"Data efficiency"},{"cite_extract_rate":1,"cites":[5729,5730],"content":"The effectiveness of deep learning comes from the ability to change the representation of the data so that the target task becomes easier to solve.\n\tThe ability to change the representation also at the symbolic level would significantly increase the capabilities of \\nesy{} systems.\n\tThis is a major open challenge for which neurally inspired methods could help achieve progress .\n\t\\section*{Acknowledgements}\n\tThis work has received funding from the Research Foundation-Flanders (FWO)\n\t(G. Marra: 1239422N, S. Dumančić: 12ZE520N, R. Manhaeve: 1S61718N). \n\tLuc De Raedt has received funding from the Flemish Government (AI Research Program), from the FWO, from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 694980 SYNTH: Synthesising Inductive Data Models) and the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. This work was also supported by TAILOR, a project funded by EU Horizon 2020 research and innovation programme under GA No 952215.\n\t\\bibliographystyle{plain}\n\t\\bibliography{main}\n\t\\appendix","id":"986b4d82-d655-4355-b6bd-d14bee40b72f","level":"paragraph","origin_cites_number":2,"parent_id":"6f37ac7d-158b-4d60-89b1-c3be4d76632d","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Open Challenges"],["paragraph","Semantics"],["paragraph","Probabilistic reasoning"],["paragraph","Symbolic representation learning"]],"subsections":[],"title":"Symbolic representation learning"},{"cite_extract_rate":1,"cites":[7233],"content":"\\label{sec:kge_gnn}\n\tIn this appendix, we introduce two approaches, namely Knowledge Graph Embeddings and Graph Neural Networks, which are commonly used in relational tasks in the deep learning community. We analyze them in the spirit of the seven dimensions introduced in the paper. In fact, it turns out that they share many features with neurosymbolic systems. In this way, we would like to suggest that NeSy can also be found at the intersection of statistical relational and geometric deep learning approaches.\n\tOne of the most popular relational representations is that of a knowledge graph (KG). A KG is\n\ta multi-relational graph composed of entities (i.e. nodes) and relations (i.e. edges). It is common to represent an edge as a triple of the form: \\textit{(head entity, relation, tail entity)}, e.g. ($homer$, $fatherOf$, $bart$).\n\tStarAI has been extensively used to solve many tasks on KGs: prediction of missing relationships (i.e. knowledge graph completion), prediction of properties of entities, or clustering entities based on their connectivity patterns. StarAI is particularly well suited to reasoning with knowledge graphs since it models explicitly the probabilistic dependencies among different relationships. \n\tHowever, in order to scale to larger knowledge graphs, the probabilistic dependencies of StarAI models have been relaxed to give rise to a new class of scalable models based on latent features, which are particularly interesting from a neurosymbolic viewpoint. The key intuition behind relational latent feature\n\tmodels is that the relationships between entities can be more efficiently predicted by modeling simpler interactions in a latent feature space. Knowledge Graph Embeddings and Graph Neural Networks represent two of the ways to encode such latent representations.","id":"0c29e7f0-ea36-44aa-a1f0-0b0c69b42bfb","level":"section","origin_cites_number":1,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI"]],"subsections":["525771a4-9400-42b9-9e72-981457921a8b","128f81ba-b844-472a-86b3-8ca1662bba8a"],"title":"Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI"},{"cite_extract_rate":0.4,"cites":[5732,3741,3672,5731],"content":"Knowledge Graph Embedding (KGE) models assume that triples (i.e. relations) are conditionally independent given a set of global latent variables, called embeddings, for entities and relations. Therefore, the existence of a triple can be predicted by a scoring function $f(e_h,e_r,e_t)$, where $e_i$ is the embedding of the corresponding object.\n\tKGE models mainly differ in terms of the scoring function and the embedding space they use. Translation or distance-based models use a scoring function measuring to what extent the tail of a triple can be obtained by a relation-specific translation of the head, i.e. $f(e_h,e_r,e_t) = || e_h + e_r - e_t||$. Semantic matching methods instead exploit similarity-based scoring\n\tfunctions, such as $f(e_h,W_r,e_t) = || e_hW_re_t^\\top||$. It is interesting to note that standard multi-layer perceptrons are often used to learn this similarity measure, i.e. $f(e_h,W_r,e_t) = nn([e_h, e_t]; W_r)$. This is similar to neural interfaces typical of $lPN$ and $LpN$ models of Section \\ref{sec:paradigms}, cf. the neural predicates of DeepProbLog in Example~\\ref{ex:deepproblog}.\n\tKGE learn from ground relations only. They learn the embeddings of entities and relations as to maximize the score for a set of known \\textit{True} triples. However, some methods incorporate higher level information, like first-order logical clauses or logical queries , bridging KGE and multiple NeSy systems (like SL or SBR). \n\tIt is interesting to analyze KGE methods in terms of the dimensions we described in our paper. KGE methods mostly work as model-based, undirected methods. They constrain the embeddings to be coherent with the logical facts and rules, which are no longer used after learning. They are heavily based on subsymbolic representations. These models learn the correct parameters (i.e. embeddings) for the task at hand. Even if no explicit semantics needs to be given to the scoring function $f(e_h,W_r,e_t)$, a fuzzy logic interpretation is often used when injecting logical rules ","id":"525771a4-9400-42b9-9e72-981457921a8b","level":"subsection","origin_cites_number":10,"parent_id":"0c29e7f0-ea36-44aa-a1f0-0b0c69b42bfb","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI"],["subsection","Knowledge Graph Embeddings"]],"subsections":[],"title":"Knowledge Graph Embeddings"},{"cite_extract_rate":0.7333333333333331,"cites":[259,5733,281,208,7332,8475,5734,5735,7045,216,1059],"content":"Graph Neural Networks (GNN) () are deep learning models for dealing with graphs as input. Inference in these models can be cast as \\textit{message passing} : at each inference step (i.e. GNN layer), a node sends a message to all its outgoing neighbors and updates its state by aggregating all the messages coming from its ingoing neighbors. Messages are computed by standard neural networks. Each inference step computes a transformation of the representations of the nodes. In the last layer, these representations are used to classify the nodes or are aggregated to classify the entire graph. \n\tThere are many connections between GNNs and neurosymbolic computation since both of them apply neural methods to relational data: graphs for GNNs, and logic representations for NeSy. While many works try to close the gap between these two representations , the application of GNN techniques to logic-based graph structures is still very limited. The most related line of work is about using GNNs on knowledge graphs . The underlying idea is to differentiate the messages exchanged by two nodes if they are related by different kinds of edges. For example, given two nodes \\texttt{homer} and \\texttt{bart}, the message corresponding to the edge \\texttt{fatherOf(homer,bart)} will be different from the one corresponding to \\texttt{sameFamily(homer,bart)}. However, these models were intended as classifiers of known relational structures (nodes and edges) and not to reason about the knowledge graph itself (e.g. determining whether an edge exists between two nodes). To perform relational reasoning, GNN-based models rely on techniques from the KGE community on top of the representations extracted by the GNN. This often takes the shape of an auto-encoding scheme: a GNN encodes an input graph in a latent representation and a KGE-based factorization technique is used to reconstruct the whole graph . \n\tAn important characteristic of GNNs is that they rely exclusively on neural computation to perform inference (i.e. to compute messages) and there is no clear direction on how to inject external knowledge about inference, e.g. as logical rules. This contrasts with NeSy, where this is one of the main goals. \n\tThere are also some interesting connections between GNNs and StarAI models. In , GNNs based on knowledge-graphs are not used as a modeling choice but rather to approximate inference in Markov Logic Networks, which is somewhat similar to regularization based methods (see Section \\ref{sec:proof_vs_model}). Similarly, in GNNs are used to encode logical formulae expressed as graphs to approximate a weighted model counting problem.\n\tFinally, it is interesting to analyse GNNs in the spirit of some of the dimensions of NeSy. GNNs act as directed models with a proof-based inference scheme: they perform a series of inference steps to compute the final answer. In the original version of GNNs , the node states are updated until a fixed point is reached, which resembles forward-chaining in logic programming. The representation of nodes belongs to a subsymbolic numerical space. Finally, GNNs can be considered as implicit structure learners: inference rules are learned through the learning of the neural message passing functions. \n\tGraph Neural Networks have recently received a lot of attention from many different communities, thanks to the representation power of neural networks and the capability of learning in complex relational settings. It is no surprise that people have started to study the expressivity of this class of models. One of the most interesting analyses from a neurosymbolic viewpoint is measuring the expressivity of GNNs in terms of variable counting logics. Recently, showed that GNNs are as expressive as 2-variable counting logic\n\t$C_2$. This fragment of first order logic admits formulas with at most two variables extended with counting quantifiers. The expressivity of this fragment is limited compared to many neurosymbolic models, especially those based on logic. However, GNNs learn the logical structure of the problem implicitly as part of the message passing learning scheme and they rely neither on expert-provided knowledge nor on heavy combinatorial search strategies to structure learning (see Section \\ref{sec:struct}). An open and challenging question that unites the GNN and NeSy communities is how to bring the expressivity to higher-order fragments , like in NeSy and StarAI, while keeping both the learning and the inference tractable, like in GNNs.","id":"128f81ba-b844-472a-86b3-8ca1662bba8a","level":"subsection","origin_cites_number":15,"parent_id":"0c29e7f0-ea36-44aa-a1f0-0b0c69b42bfb","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI"],["subsection","Graph Neural Networks"]],"subsections":[],"title":"Graph Neural Networks"},{"cite_extract_rate":0.5714285714285711,"cites":[5736,5737,5707,5706],"content":"Fuzzy logic, as many-valued extension over Boolean logic, has a very long tradition . However, the use of fuzzy logic in StarAI and NeSy is not dictated by the need of dealing with vagueness, but by the advantageous computational properties of t-norms. Indeed, a common use case is to have an initial theory defined in Boolean logic which is \\textit{fuzzyfied}. Inference is then carried out with the fuzzyfied theory and the answers are eventually discretized back to Boolean values (usually using a threshold at 0.5). \n\tThe reason for this approach is that one would like to exploit the differentiability of t-norms to address logical inference of FOL theories in a more scalable way than standard combinatorial optimization algorithms (e.g. SAT solvers). This is particularly important in undirected and regularization-based methods (such as PSL and LTN ). In fact, it has been shown that there are fragments of fuzzy logic that can even provide convex inference problems. Example \\ref{ex:bool_vs_fuzzy}, however, shows that naively approaching logical inference through a fuzzy relaxation and gradient-based optimization can introduce unexpected behaviours.\n\t\\begin{boxedexample}[label = ex:bool_vs_fuzzy]{Fuzzyfication and soft-Satisfability}\n\t\tLet us consider a disjunction, like $A \\vee B \\vee C$. In Boolean logic, if we state that the disjunction is \\textit{satisfied} (i.e. \\textit{True}), then we expect at least one among the three variables to be \\textit{True}. \n\t\tSuppose we want to find a truth assignment for all the variables that satisfies the disjunction above. The approach of the majority of NeSy fuzzy approaches is the following. First, the rule is relaxed into a fuzzy real function. For example, using the \\L ukasiewicz t-norm, $F_\\oplus(A,B,C)= min(1, A+B+C)$. Secondly, a gradient-based algorithm (e.g. backpropagation with Adam ) is used to maximize the value of the formula with respect to the \\textit{fuzzy} truth degree of the three variables. Finally, the obtained \\textit{fuzzy} solution $A^\\star, B^\\star, C^\\star$ is translated back into a Boolean assignment using a $0.5$ threshold.\n\t\tLet us consider a possible optimal fuzzy solution, like $(A^\\star, B^\\star, C^\\star) = (0.34, 0.34, 0.34)$ and its discretized version $(A^\\star, B^\\star, C^\\star) = (False, False,$ $False)$, using a threshold at $0.5$. The discretized solution does not satisfy the initial Boolean formula, even though it is a global optimum in the fuzzyfied problem. \n\t\\end{boxedexample}\n\tSimilarly, shows that, while it is very common to reason about universally quantified formulae in the form of $\\forall x: A(x) \\to B(x)$, like `all humans are mortal', using gradients and fuzzy logic to make inference can be extremely counterintuitive, especially with specific t-norms such as the product t-norm. It is unclear whether there exists a generally accepted subset of properties of Boolean logic that one wants to preserve and whether one can define a t-norm that guarantees such properties.","id":"df605b7c-7a16-4a8a-9a34-34eff8f50d0f","level":"section","origin_cites_number":7,"parent_id":"31670094-45b4-4854-914d-94aaf932ce65","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Fuzzy logic, fuzzyfication and soft-Satisfability"]],"subsections":["c6063ac6-99de-4318-925a-eab4859602f1"],"title":"Fuzzy logic, fuzzyfication and soft-Satisfability"},{"cite_extract_rate":0,"cites":[],"content":"Another common reason for using fuzzy logic is to exploit a differentiable semantics. Then, gradient-based methods can be used to train the parameters of a weighted logical theory such as in LRNNs . This contrasts with gradient-based training of the parameters of probabilistic logics based on the distribution semantics. Possible worlds in probabilistic logic are defined as possible assignments of truth values to all the ground atoms of a logical theory. The assignments of truth values specify the \\textit{semantics} of the logic. On the contrary, fuzzy logic assigns continuous truth degrees to formulas or proofs, which are \\textit{syntactic} structures. As a consequence, while the probability of an atom will always be equal to the sum of the probabilities of the worlds in which it is \\textit{True} (cf. Equation \\ref{eq:wmc}), the fuzzy degree of an atom may vary depending on how that atom has been proven or defined, as shown in Example \\ref{ex:prob_vs_fuzzy}.\n\t\\begin{boxedexample}[label = ex:prob_vs_fuzzy]{Distribution semantics vs fuzzy logic semantics}\n\t\tConsider the following annotated program (adapted from ).\n\t\t\\begin{lstlisting}\n\t\t0.3::b.\n\t\ta1 <- b.\n\t\ta2 <- b,b.\n\t\t\\end{lstlisting}\n\t\tHere, $0.3$ is the label of $\\mathtt{b}$ without any particular semantics yet. \n\t\tGiven the idempotency of the Boolean conjunction, we would expect the scores of both $\\mathtt{a1}$ and $\\mathtt{a2}$ to be identical since both $\\mathtt{b}$ and $\\mathtt{b} \\wedge \\mathtt{b}$ are \\textit{True} when $\\mathtt{b}$ is \\textit{True}. In a probabilistic approach, the score is interpreted as a probability, i.e. $p(\\mathtt{b}) = 0.3$. The probability of any atom is the sum of the probabilities of all the worlds where that atom is \\textit{True}. It is easy to see that, for both $\\mathtt{a1}$ and $\\mathtt{a2}$, these are the worlds where $\\mathtt{b}$ is \\textit{True}, thus $p(\\mathtt{a1}) = p(\\mathtt{a2}) = p(\\mathtt{b}) = 0.3$. On the other hand, in the fuzzy setting, the score of $\\mathtt{b}$ is interpreted as its truth degree, i.e. $t(\\mathtt{b}) = 0.3$. Let us consider the product t-norm, where $t(x \\wedge y) = t(x)t(y)$, then $t(\\mathtt{a1}) = t(\\mathtt{b}) = 0.3$ while $t(\\mathtt{a2}) = t(\\mathtt{b})t(\\mathtt{b}) = 0.09$. While this issue could be solved by choosing a different t-norm (e.g. the minimum t-norm), similar issues arise in different definitions.\n\t\\end{boxedexample}\n\tThe differences between the two semantics are not due to the probabilistic or the fuzzy semantics, but more to the distinction between semantics based on possible worlds and semantics based on proofs or derivations. In fact, a similar behaviour is observed in Stochastic Logic Programs under the name of \\textit{memoization}.\n\\end{document}","id":"c6063ac6-99de-4318-925a-eab4859602f1","level":"subsection","origin_cites_number":2,"parent_id":"df605b7c-7a16-4a8a-9a34-34eff8f50d0f","prefix_titles":[["title","From Statistical Relational to Neurosymbolic \\\\ Artificial Intelligence: a Survey."],["section","Fuzzy logic, fuzzyfication and soft-Satisfability"],["subsection","Distribution semantics and fuzzy logic semantics"]],"subsections":[],"title":"Distribution semantics and fuzzy logic semantics"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ]\n ],\n \"subsections\": [\n \"504b42a2-053a-4d14-a0e9-a9517b24369f\",\n \"71d65d84-75cd-4c92-b14d-bddcaa016dd0\",\n \"394ef4e2-3d6b-46dc-9354-ff1b43665302\",\n \"3f30c5c1-1c7c-4ae0-b5f0-88220243574e\",\n \"3cf78a3e-98ed-45de-848f-3a61063e51b3\",\n \"bd238fce-3df5-4531-ba35-f4316c49d81b\",\n \"322e754b-6517-46c9-9006-90e863943e5b\",\n \"528b4897-8064-4ac1-a40f-dc1e7d30614f\",\n \"35970a41-0514-46a0-8d9d-63d0769ff87a\",\n \"0c29e7f0-ea36-44aa-a1f0-0b0c69b42bfb\",\n \"df605b7c-7a16-4a8a-9a34-34eff8f50d0f\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.14285714285714202,\n \"cites\": [\n 7601,\n 8475\n ],\n \"content\": \"The integration of learning and reasoning is a key challenge in artificial intelligence and machine learning today. Various communities are addressing it, especially\\n\\t the field of neurosymbolic artificial intelligence (NeSy) .\\n\\tNeSy's goal is to integrate symbolic reasoning with neural networks.\\n\\t\\\\nesy{} already has a long tradition, and it has recently attracted a lot of attention.\\nIndeed, the topic has been addressed by prominent researchers such as Y. Bengio and H. Kautz in their keynotes at AAAI 2020, by Y. Bengia and G. Marcus in the AI Debate and Hochreiter has recently stated that \\\\nesy{} is ``the most promising approach to a broad AI''.\\n\\tAnother domain with a rich tradition in integrating learning and reasoning is that of statistical relational learning and artificial intelligence (StarAI) . StarAI focuses on integrating logical and probabilistic reasoning.\\nHistorically, these two endeavours have adopted different learning paradigms, probabilistic versus neural, for integrating logic into machine learning. This in turn has resulted in two different subcommunities. StarAI has focused on probabilistic logics, their semantics and making inference more tractable, while learning is usually based on parameter learning techniques from probabilistic graphical models.\\nOn the other hand, \\\\nesy{} extends neural networks with symbolic knowledge, focusing on scalable approximate models, paying less attention to semantical issues. In particular, \\\\nesy{} techniques can often be characterized by a clear parameterization in terms of neural networks, i.e. layered structures of latent representations, and by resorting to the gradient-based backpropagation paradigm for learning.\\n\\\\rev{Despite a different focus and approach, the two domains want to achieve the same goal, that is, to integrate learning and reasoning. \\nIt is therefore surprising that there has been relatively little interaction between the two domains, but see }.\\n\\tThis discrepancy is the key motivation behind this survey: it aims at pointing out the similarities between these two endeavours and, in this way, it wants to stimulate \\n\\tcross-fertilization. \\n\\tWe start from the literature on StarAI, following the key concepts, \\\\rev{definitions} and techniques outlined in several textbooks and tutorials such as , because it turns out that the same issues and techniques that arise in StarAI apply to \\\\nesy{} as well. \\n\\t\\\\rev{The key contributions of this paper are:}\\n\\t\\\\begin{enumerate}\\n\\t\\t\\\\item \\n\\t\\t{\\\\em We identify seven dimensions that these fields have in common and that can be used to categorize both StarAI and \\\\nesy{} approaches}. \\n\\t\\tThese seven dimensions are concerned with (1) model vs proof-based inference, \\\\rev{(2) logic syntax,} (3) semantics, (4) learning parameters or structure, (5) representing entities as symbols or subsymbols, (6) integrating logic with probabilistic and/or neural concepts, and \\\\rev{(7) learning tasks}.\\n\\t\\t\\\\item We provide evidence for our claim by positioning a wide range of StarAI and \\\\nesy{} systems along these dimensions and pointing out analogies between them. \\n\\t\\tThis provides not only new insights into the relationships between StarAI and NeSy, but it also allows one to carry over and adapt techniques from one field to another. These insights provide opportunities for cross-fertilization between StarAI and NeSy, by focusing on those dimensions that have not been fully exploited yet.\\n\\t\\\\item \\\\rev{We gently introduce key logical concepts and techniques inherited from StarAI. In this way, the paper also provides a gentle introduction to symbolic AI and StarAI techniques for the interested ``connectionist'' practitioner.} \\n\\t\\t\\\\item We illustrate each dimension using existing methods, \\n and in this way, also present an intuitive and concrete overview of the research field. \\n\\t\\\\end{enumerate}\\n\\tUnlike some other perspectives on neurosymbolic computation , the present survey limits itself to a logical perspective and to developments in neurosymbolic computation that are consistent with this perspective. \\\\rev{Therefore, we usually refer to symbols and symbolic algorithms as synonyms for logical representations and logical reasoning.}\\n\\tFurthermore, the survey focuses on representative and prototypical systems rather than aiming at completeness (which would not be possible given the fast developments in the field). \\n\\tSeveral other surveys about neurosymbolic AI have been proposed. An early overview of neurosymbolic computation is that of . Unlike the present survey it focuses very much on a logical and a reasoning perspective. Today, the focus has shifted very much to learning. More recently, analysed the intersection between NeSy and graph neural networks (GNN). described neurosymbolic systems in terms of the composition of blocks described by few patterns, concerning processes and exchanged data. In contrast, this survey is more focused on the underlying principles that govern such a composition. Finally, exploits a neural network viewpoint by investigating in which components (i.e. input, loss or structure) symbolic knowledge is injected.\",\n \"id\": \"504b42a2-053a-4d14-a0e9-a9517b24369f\",\n \"level\": \"section\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [\n \"baca8b6b-5536-41e9-a0f9-bc4e2294ec62\"\n ],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\rev{The next seven sections each describe one dimension by first introducing the main underlying concepts, either based on logic, probability or machine learning, and then showing how they are incorporated in StarAI and NeSy systems. Section \\\\ref{sec:proof_vs_model} presents how to use logic for inference by distinguishing between proof-based and model-based systems, while Section \\\\ref{sec:syntax} introduces logic at the syntac level, in particular, propositional, relational and first-order logic. Section \\\\ref{sec:semantics} then introduces the semantics of logic and shows how to extend it to a continuous semantics, using fuzzy and probabilistic logics. Section \\\\ref{sec:struct} discusses the dimension of learning, distinguishing parameter learning from structure learning. Section \\\\ref{sec:symb_vs_subsymb} focuses on the representational level and to what extent neurosymbolic models use symbolic and/or subsymbolic features. Section \\\\ref{sec:paradigms} positions neurosymbolic approaches along the spectrum of three main paradigms, i.e. logic, probability and neural networks. \\\\rev{Section \\\\ref{sec:tasks} describes general classes of learning tasks to which neurosymbolic systems are usually applied.} Finally, in Section 10, we conclude by introducing open challenges in the neurosymbolic landscape.}\\n\\tWe summarize various neurosymbolic approaches along these dimensions in Table \\\\ref{tab:big_table}.\\n\\t\\\\include{big_table}\",\n \"id\": \"baca8b6b-5536-41e9-a0f9-bc4e2294ec62\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"504b42a2-053a-4d14-a0e9-a9517b24369f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"paragraph\",\n \"Structure of the paper\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Structure of the paper\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:proof_vs_model}\\n\\tIn this paper, we focus on clausal logic as it is a standard form to which any first order logical formula can be converted. In clausal logic, theories are represented in terms of {clauses}. \\n\\tMore formally, a \\\\textit{clause} is an expression of the form $ h_1 \\\\vee ... \\\\vee h_k \\\\leftarrow b_1 \\\\wedge ... \\\\wedge b_n$.\\n\\tThe\\n\\t$h_k$ are \\\\textit{head literals} or conclusions, while the $b_i$ are \\\\textit{body literals} or conditions. Clauses with no conditions ($n=0$) and one conclusion ($k=1$) are \\\\textit{facts}.\\n\\tClauses with only one conclusion ($k=1$) are \\\\textit{definite clauses}.\\n\\tThe question we want to answer in this section is how to use such clausal theories to reason? And, how to infer new facts from the known clauses? Along this first dimension, we will investigate the two fundamental ways to view logical inference and determine the implications for StarAI and NeSy systems. In one view, we want to find \\\\textit{proofs} for a certain query, which leads to the proof-theoretic approach to logic. In the other view, we want to find \\\\textit{models} (that is, truth assignments to the logical atoms) that satisfy a given theory. This leads to the model-theoretic approach to logic.\",\n \"id\": \"71d65d84-75cd-4c92-b14d-bddcaa016dd0\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Proof- vs Model-theoretic View of Logic\"\n ]\n ],\n \"subsections\": [\n \"88cc4046-116e-4648-b8c1-799e686a9f88\",\n \"de2541d6-848e-4cbb-aea2-6b2aa605f9f0\",\n \"dfea86d8-fb1c-484d-a9d2-d17c03c4d644\"\n ],\n \"title\": \"Proof- vs Model-theoretic View of Logic\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\n\\tThe proof-theoretic approach finds proofs for a query in a logic theory. While this approach to inference is applicable to any logic theory, we focus on logic programs in this paper. Syntactically, a logic program is a \\\\textit{definite} clause theory, which is a theory where all the clauses are definite (i.e. only one conclusion).\\n\\tIn logic programs, definite clauses are interpreted as if-then \\\\textbf{rules} ($h$ is true if $b_1, ..., b_n$ are true). \\n \\\\rev{A proof for a query $q$ is a sequence of logical inference steps that demonstrates the truth of a query based on the given program. A compact way of representing the set of all proofs in a logic program uses an AND/OR tree, which consists of AND and OR nodes and edges amongst them. Each node represents a goal. An AND node branches into one or more outgoing edges, each representing one of the sub-goals that need to be simultaneously satisfied for the goal in the AND node to be true. An OR node represents choices or alternatives between multiple clauses that can be used to prove a particular sub-goal. The OR node branches into multiple outgoing edges, each representing one of these possible choices. Leaf nodes in the AND/OR tree represent true facts.}\\n\\tTypically, forward or backward chaining inference is used to search for proofs for queries. \\n\\tWe illustrate this in Example \\\\ref{ex:logic_program}.\\n\\t\\\\begin{boxedexample}[label=ex:logic_program]{Logic programs and proofs}\\n\\t\\tConsider the following logic program (adapted from ): \\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tburglary.\\n\\t\\thears_alarm_mary. \\n\\t\\tearthquake.\\n\\t\\thears_alarm_john. \\n\\t\\talarm <- earthquake. \\n\\t\\talarm <- burglary.\\n\\t\\tcalls_mary <- alarm,hears_alarm_mary.\\n\\t\\tcalls_john <- alarm,hears_alarm_john. \\n\\t\\t\\\\end{lstlisting}\\n The rules for \\\\textit{alarm} state that there will be an \\\\textit{alarm} if there is a \\\\textit{burglary} or an \\\\textit{earthquake}.\\n\\\\noindent The set of proofs for the query \\\\textit{calls\\\\_mary} can be represented compactly as an AND/OR tree.\\n \\\\include{to_include/and_or_tree}\\n\\\\end{boxedexample}\",\n \"id\": \"88cc4046-116e-4648-b8c1-799e686a9f88\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"71d65d84-75cd-4c92-b14d-bddcaa016dd0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Proof- vs Model-theoretic View of Logic\"\n ],\n [\n \"paragraph\",\n \"\\\\rev{Proof-theoretic logic\"\n ]\n ],\n \"subsections\": [\n \"565706e8-0b20-4fdb-b6e9-e14ec63f057e\"\n ],\n \"title\": \"\\\\rev{Proof-theoretic logic\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"} On the other hand, the model theoretic perspective on logic is to find a model or truth assignment to the logical atoms that satisfy a given logic theory. \\n\\tAn \\\\textit{interpretation}, or possible world, is a truth-assignment to the propositions (or ground atoms) of the language, and\\n\\tcan be uniquely identified with the set of propositions it assigns $True$ (thus considering all the other \\\\textit{False}). \\n\\tAn interpretation is a \\\\textit{model} of a clause $ h_1 \\\\vee ... \\\\vee h_k \\\\leftarrow b_1 \\\\wedge ... \\\\wedge b_n$ if at least one of the $h_i$ is in the interpretation when all the $b_1 \\\\wedge ... \\\\wedge b_n$ are in the interpretation as well. An interpretation $I$ is a model of a theory $T$, and we write $I \\\\models T$, if it is a model of all clauses in the theory. We say that the theory is \\\\textit{satisfiable} if it has a model. The satisfiability problem, that is, deciding whether a theory has a model, is one of the most fundamental ones in computer science (cf. the SAT problem for propositional logic). \\n\\t\\\\begin{boxedexample}[label = ex:model_based]{Model-theoretic}\\n\\t\\tConsider the following theory (adopted from ): \\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tsmokes_mary $\\\\leftarrow$ smokes_john, influences_john_mary.\\n\\t\\tsmokes_john $\\\\leftarrow$ smokes_mary, influences_mary_john.\\n\\t\\tsmokes_mary $\\\\leftarrow$ stress_mary.\\n\\t\\tsmokes_john $\\\\leftarrow$ stress_john.\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tA model of the previous theory is the set:\\n\\t\\t$$M = \\\\{\\\\mathtt{stress\\\\_john, smokes\\\\_john}\\\\}$$ \\n\\t\\tBy considering all the elements of this set \\\\textit{True} and all others \\\\textit{False}, the four clauses are satisfied.\\n\\t\\\\end{boxedexample}\\n\\tIn the model-theoretic perspective, one uses the logic theory as a set of \\n\\t{\\\\em constraints} on the propositions, that is, the propositions are related to one another, without imposing a directed inference relationship between them as in forward or backward chaining. More details on these connections can be found in .\",\n \"id\": \"565706e8-0b20-4fdb-b6e9-e14ec63f057e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"88cc4046-116e-4648-b8c1-799e686a9f88\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Proof- vs Model-theoretic View of Logic\"\n ],\n [\n \"paragraph\",\n \"\\\\rev{Proof-theoretic logic\"\n ],\n [\n \"paragraph\",\n \"\\\\rev{Model-theoretic logic\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"\\\\rev{Model-theoretic logic\"\n },\n {\n \"cite_extract_rate\": 0.1,\n \"cites\": [\n 5706\n ],\n \"content\": \"Statistical Relational AI's focus is on unifying logical and probabilistic graphical models (PGMs). A PGM is a graphical model that compactly\\n\\trepresents a (joint) probability distribution $P(X_1, ... , X_n)$ over $n$ discrete or continuous random variables $X_1, ... ,X_n$.\\n\\tThe key idea is that the joint factorizes over some factors $f_i$ specified over subsets $X^i$ of the variables $\\\\{X_1, ... ,X_n\\\\}$.\\n\\t$$P(X_1, ... , X_n) = \\\\frac{1}{Z} f_1(X^1) \\\\times ... \\\\times f_k(X^k) $$\\n\\tThe random variables correspond to the nodes in the graphical structure, and the factorization is determined\\n\\tby the edges in the graph. \\n\\tThere are two classes of graphical models: \\\\textit{directed} ones, or Bayesian networks, and \\\\textit{undirected} ones, or Markov Networks.\\n\\tIn Bayesian networks, the underlying graph structure is a directed acyclic graph,\\n\\tand the factors $f^i(X_i | parents(X_i))$ correspond to the conditional probabilities $P(X_i | parents(X_i))$, where $parents(X_i)$\\n\\tdenotes the set of random variables that are a parent of $X_i$ in the graph.\\n\\tIn Markov networks, the graph is undirected and the factors $f^i(X^i)$ correspond to the set of nodes $X^i$ that form (maximal) cliques in the graph. Furthermore, the factors are non-negative and $Z$ is a normalisation constant.\\n\\t\\\\begin{evidencebox}\\n\\t\\t\\\\rev{The distinction between directed and undirected graphical models is parallel to the proof- vs model-theoretic view of logic. This parallel is at the very core of StarAI. In fact, by viewing each variable $X_i$ (or proposition) \\\\textit{at the same time} as a random and as a logical variable , clausal theories can be extended to define probabilistic models. Clauses can then be translated into binary valued factors by labeling them with weights (or probabilities), thus parameterizing the corresponding factors.}\\n\\t\\\\end{evidencebox}\\n\\tIn the remainder of this section, we will show how StarAI has used this parallel to define two types of systems .\\n\\tThe first type of StarAI system generalizes directed models and resembles Bayesian networks. It includes well-known representations such as plate notation , probabilistic relational models (PRMs) , probabilistic logic programs (PLPs) , and Bayesian logic programs (BLPs) . \\n\\tToday the most typical and popular representatives of this category are the probabilistic (logic) programs. \\n\\tProbabilistic logic programs were introduced by \\n\\tPoole~ and the first learning algorithm is due to Sato~. \\n\\tProbabilistic logic programs are essentially definite clause programs where every fact is annotated with the probability that it is \\\\textit{True}. \\n\\tThis then results in a possible world semantics.\\n\\tThe reason why probabilistic logic programs are viewed as directed models is clear when looking at the derivations\\n\\tfor a query, cf. Example \\\\ref{ex:logic_program}. At the top of the AND-OR tree, there is the query that one wants to prove and the structure of the tree\\n\\tis that of a directed graph (even though it need not be acyclic). One can straightforwardly map\\n\\tdirected graphical models, that is, Bayesian networks,\\n\\tonto such probabilistic logic programs by associating one definite clause to every entry in the conditional probability tables,\\n\\tyielding factors of the form $P(X | Y_1, ... , Y_n)$. Assuming boolean random variables, each entry ${x,y_1, ...,y_n}$ with parameter value $v$ can be represented using the definite clause $X(x) \\\\leftarrow Y_1 (y_1) \\\\wedge ... \\\\wedge Y_n(y_n) \\\\wedge p_{x,y_1, ...,y_n}$ and \\n\\tprobabilistic fact $v::p_{x,y_1, ...,y_n}$. \\n\\tA probabilistic version of Example~\\\\ref{ex:logic_program} is shown in Example~\\\\ref{ex:problog} using the syntax of ProbLog .\\n\\t\\\\begin{figure}[t]\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1]\\n\\\\draw (160,105) .. controls (160,96.72) and (177.91,90) .. (200,90) .. controls (222.09,90) and (240,96.72) .. (240,105) .. controls (240,113.28) and (222.09,120) .. (200,120) .. controls (177.91,120) and (160,113.28) .. (160,105) -- cycle ;\\n\\\\draw (110,40) .. controls (110,28.95) and (125.67,20) .. (145,20) .. controls (164.33,20) and (180,28.95) .. (180,40) .. controls (180,51.05) and (164.33,60) .. (145,60) .. controls (125.67,60) and (110,51.05) .. (110,40) -- cycle ;\\n\\\\draw (220,40) .. controls (220,28.95) and (239.7,20) .. (264,20) .. controls (288.3,20) and (308,28.95) .. (308,40) .. controls (308,51.05) and (288.3,60) .. (264,60) .. controls (239.7,60) and (220,51.05) .. (220,40) -- cycle ;\\n\\\\draw (271.04,105) .. controls (271.04,96.72) and (299.91,90) .. (335.52,90) .. controls (371.13,90) and (400,96.72) .. (400,105) .. controls (400,113.28) and (371.13,120) .. (335.52,120) .. controls (299.91,120) and (271.04,113.28) .. (271.04,105) -- cycle ;\\n\\\\draw (100,175) .. controls (100,166.72) and (120.15,160) .. (145,160) .. controls (169.85,160) and (190,166.72) .. (190,175) .. controls (190,183.28) and (169.85,190) .. (145,190) .. controls (120.15,190) and (100,183.28) .. (100,175) -- cycle ;\\n\\\\draw (200,120) -- (252.57,158.24) ;\\n\\\\draw [shift={(255,160)}, rotate = 216.03] [fill={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.08] [draw opacity=0] (8.93,-4.29) -- (0,0) -- (8.93,4.29) -- cycle ;\\n\\\\draw (200,120) -- (147.43,158.24) ;\\n\\\\draw [shift={(145,160)}, rotate = 323.97] [fill={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.08] [draw opacity=0] (8.93,-4.29) -- (0,0) -- (8.93,4.29) -- cycle ;\\n\\\\draw (335.52,120) -- (257.69,158.67) ;\\n\\\\draw [shift={(255,160)}, rotate = 333.58000000000004] [fill={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.08] [draw opacity=0] (8.93,-4.29) -- (0,0) -- (8.93,4.29) -- cycle ;\\n\\\\draw (60,120) -- (142.29,158.72) ;\\n\\\\draw [shift={(145,160)}, rotate = 205.2] [fill={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.08] [draw opacity=0] (8.93,-4.29) -- (0,0) -- (8.93,4.29) -- cycle ;\\n\\\\draw (260,60) -- (202.74,85.77) ;\\n\\\\draw [shift={(200,87)}, rotate = 335.77] [fill={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.08] [draw opacity=0] (8.93,-4.29) -- (0,0) -- (8.93,4.29) -- cycle ;\\n\\\\draw (150,60) -- (197.36,85.57) ;\\n\\\\draw [shift={(200,87)}, rotate = 208.37] [fill={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.08] [draw opacity=0] (8.93,-4.29) -- (0,0) -- (8.93,4.29) -- cycle ;\\n\\\\draw (1.04,105) .. controls (1.04,96.72) and (29.91,90) .. (65.52,90) .. controls (101.13,90) and (130,96.72) .. (130,105) .. controls (130,113.28) and (101.13,120) .. (65.52,120) .. controls (29.91,120) and (1.04,113.28) .. (1.04,105) -- cycle ;\\n\\\\draw (210,175) .. controls (210,166.72) and (230.15,160) .. (255,160) .. controls (279.85,160) and (300,166.72) .. (300,175) .. controls (300,183.28) and (279.85,190) .. (255,190) .. controls (230.15,190) and (210,183.28) .. (210,175) -- cycle ;\\n\\\\draw (119,31) node [anchor=north west][inner sep=0.75pt] [align=left] {\\\\texttt{burglary}};\\n\\\\draw (228,31) node [anchor=north west][inner sep=0.75pt] [align=left] {\\\\texttt{earthquake}};\\n\\\\draw (181.14,95.38) node [anchor=north west][inner sep=0.75pt] [align=left] {\\\\texttt{alarm}};\\n\\\\draw (276,96.13) node [anchor=north west][inner sep=0.75pt] [align=left] {\\\\texttt{ hears\\\\_alarm\\\\_mary}};\\n\\\\draw (106.92,166.13) node [anchor=north west][inner sep=0.75pt] [align=left] {\\\\texttt{calls\\\\_john}};\\n\\\\draw (13,96.13) node [anchor=north west][inner sep=0.75pt] [align=left] {\\\\texttt{hears\\\\_alarm\\\\_john}};\\n\\\\draw (216.92,166.13) node [anchor=north west][inner sep=0.75pt] [align=left] {\\\\texttt{calls\\\\_mary}};\\n\\\\end{tikzpicture}\\n\\t\\t\\\\caption{The Bayesian network corresponding to the ProbLog program in Example \\\\ref{ex:problog}}\\n\\t\\t\\\\label{fig:bayes_net_problog}\\n\\t\\\\end{figure}\\n\\t\\\\begin{boxedexample}\\n [label = ex:problog]{ProbLog}\\n\\t\\tWe show a probabilistic extension of the alarm program using ProbLog.\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\t0.1::burglary.\\n\\t\\t0.3::hears_alarm_mary. \\n\\t\\t0.05::earthquake.\\n\\t\\t0.6::hears_alarm_john.\\n\\t\\talarm <- earthquake. \\n\\t\\talarm <- burglary.\\n\\t\\tcalls_mary <- alarm,hears_alarm_mary.\\n\\t\\tcalls_john <- alarm,hears_alarm_john. \\n\\t\\t\\\\end{lstlisting}\\n\\t\\tThis program can be mapped to the Bayesian network in Figure~\\\\ref{fig:bayes_net_problog}\\n\\t\\tThis probabilistic logic program defines a distribution $p$ over possible worlds $\\\\omega$. Let $P$ be a Problog program and $F = \\\\{p_1::c_1, \\\\cdots, p_n::c_n\\\\}$ be the set of ground probabilistic facts $c_i$ of the program and $p_i$ their corresponding probabilities.\\n\\t\\tProbLog defines a probability distribution over $\\\\omega$ in the following way:\\n\\t\\t\\\\begin{equation*}\\n\\t\\tp(\\\\omega) = \\\\begin{cases} 0, &\\\\mbox{if } \\\\omega \\\\not\\\\models P \\\\\\\\\\n\\t\\t\\\\displaystyle \\\\prod_{c_i \\\\in \\\\omega: c_i = T} p_i \\\\cdot \\\\prod_{c_j \\\\in \\\\omega: c_j = F} (1 - p_j), & \\\\mbox{if } \\\\omega \\\\models P \\\\end{cases}\\n\\t\\t\\\\end{equation*}\\n\\t\\\\end{boxedexample}\\n\\tThe second type of StarAI system generalizes undirected graphical models such as Markov networks or random fields.\\n\\tThe prototypical example is Markov Logic Networks (MLNs) , and also Probabilistic Soft Logic (PSL) follows this idea.\\n\\tUndirected StarAI models consist of a set of weighted clauses $w:h_1 \\\\vee ... \\\\vee h_k \\\\leftarrow b_1 \\\\wedge ... \\\\wedge b_m$\\n that become soft constraints. The higher the weight of a ground clause, the less likely possible worlds that violate these constraints are. In the limit, when the weight is $+\\\\infty$ the constraint must be satisfied and becomes a purely logical constraint, a hard constraint.\\n\\tThe weighted clauses specify a more general relationship between the conclusion and the condition than the definite clauses of directed models. \\n\\tWhile clauses of undirected models can still be used in (resolution) theorem provers, they are commonly viewed as constraints that relate these two sets of atoms.\\n\\tSuch undirected StarAI models can be mapped to an undirected probabilistic graphical model in which there is a one-to-one correspondence between grounded weighted clauses and factors, as we show in Example \\\\ref{ex:mln}.\\n\\t\\\\begin{figure}[t]\\n\\t\\t\\\\centering\\n\\t\\t\\\\include{to_include/markov_field}\\n\\t\\t\\\\caption{The Markov Field corresponding to the Markov logic network in Example \\\\ref{ex:mln}}\\n\\t\\t\\\\label{fig:mln}\\n\\t\\\\end{figure}\\n\\t\\\\begin{boxedexample}[label = ex:mln]{Markov Logic Networks}\\n\\t\\tWe show a probabilistic extension (adapted from ) of the theory in Example~\\\\ref{ex:model_based} using the formalism of Markov Logic Networks. We use a First Order language with domain $D = \\\\{\\\\textit{john},\\\\textit{mary}\\\\}$ and weighted clauses $\\\\alpha_1$ and $\\\\alpha_2$, i.e.:\\n\\t\\t\\\\begin{align*}\\n\\t\\t\\\\alpha_1: \\\\quad &\\\\mathtt{\\n\\t\\t\\t2.0::smokes(Y) \\\\leftarrow smokes(X), influences(X,Y)} \\\\\\\\\\n\\t\\t\\\\alpha_2: \\\\quad &\\\\mathtt{0.5::smokes(X) \\\\leftarrow stress(X)}\\n\\t\\t\\\\end{align*}\\n\\t\\tIn Figure~\\\\ref{fig:mln}, we show the corresponding Markov field.\\n\\t\\tA Markov Logic Network defines a probability distribution over possible worlds as follows.\\n\\t\\tLet $ A = [\\\\alpha_1, \\\\cdots, \\\\alpha_n]$ be a set of logical clauses and let $B = [\\\\beta_1, \\\\cdots, \\\\beta_n]$ the corresponding positive weights. Let $\\\\theta_j$ be a possible assignment of constants (from the domain $D$) to the variables (e.g. $\\\\mathtt{X,Y}$) of the clause $\\\\alpha_i$, that is, a substitution. Let $\\\\alpha_i\\\\theta_j$ the grounded clause where all variables in $\\\\theta_j$ have been replaced by their corresponding constants. Finally, let $\\\\mathbbm{1}(\\\\omega,\\\\alpha_i\\\\theta_i)$ be an indicator function, evaluating to 1 if the ground clause is \\\\textit{True} in $\\\\omega$, 0 otherwise. \\n\\t\\tThe probabilistic semantics of Markov Logic is the distribution (with $Z$ the normalization constant):\\n\\t\\t\\\\begin{equation*}\\n\\t\\tp(\\\\omega) = \\\\frac{1}{Z} \\\\exp \\\\big(\\\\sum_i \\\\beta_i \\\\sum_{j} \\\\mathbbm{1}(\\\\omega, \\\\alpha_i\\\\theta_j) \\\\big)\\n\\t\\t\\\\end{equation*}\\n\\t\\tIntuitively, in MLNs, a world is more probable if it makes many of its ground instances \\\\textit{True}. Notice that MLNs are usually defined on first-order clause theories, with variables and domains. We will further investigate this issue in Section \\\\ref{sec:syntax}.\\n\\t\\\\end{boxedexample}\",\n \"id\": \"de2541d6-848e-4cbb-aea2-6b2aa605f9f0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"71d65d84-75cd-4c92-b14d-bddcaa016dd0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Proof- vs Model-theoretic View of Logic\"\n ],\n [\n \"subsection\",\n \"Implications for StarAI\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for StarAI\"\n },\n {\n \"cite_extract_rate\": 0.636363636363636,\n \"cites\": [\n 2665,\n 5707,\n 3726,\n 5709,\n 3727,\n 3741,\n 8942,\n 7988,\n 5710,\n 5711,\n 4323,\n 7304,\n 3672,\n 5708\n ],\n \"content\": \"\\\\rev{The distinction between proof vs models and inference rules vs constraints, turns out to be fundamental for neurosymbolic systems as well.}\\n\\t\\\\begin{evidencebox}\\n\\t\\t\\\\rev{In neurosymbolic AI, weighted clauses are not used to construct a probabilistic graphical model, but they are likewise used to construct neural models. More specifically, NeSy systems that exploit a proof-theoretic approach use the proofs to build the \\\\textit{architecture} of the neural net. On the other side of the spectrum, NeSy systems that exploit a model-theoretic approach use the constraints to build a \\\\textit{loss function} for the neural net. }\\n\\t\\\\end{evidencebox}\\n\\t\\\\rev{Both choices are extremely natural.\\n Proof trees capture the structure of the inference process in a graphical representation. Therefore, they can be used as the structure of the neural network computation, which corresponds to their architecture. On the other hand, the desired behaviour of the variables is expressed in terms of constraints. \\n Loss functions are the de facto standard to enforce desired behaviours on the output variables of a neural network.}\\n\\\\rev{First, we survey proof-based NeSy models, which use theorem proving for logical inference and proofs to template the neural architecture. In particular, when proving a specific query atom, they keep track of all the used rules in a proof tree, such as the one shown in Example \\\\ref{ex:logic_program}.\\tWeights on facts and rules are then used to label leaves or edges of the tree, respectively, while real valued activation functions are used to label the AND and OR nodes. The result is a computational graph that can be executed (or evaluated) bottom-up, starting from the leaves up to the root. Generally speaking, the output of the computational graph is a \\\\textit{score} for the query atom. Different semantics can be exploited in building the computational graph, ranging from relaxations of truth values (such as in fuzzy logic) to probabilities (see Section \\\\ref{sec:semantics}). The connection between the proof tree and the neural network suggests schemes for learning the parameters of these models. Indeed, the obtained computational graph is always differentiable. Thus, given a set of atoms that are known to be \\\\textit{True} (resp. \\\\textit{False}), one can maximize (resp. minimize) their score using the corresponding computational graphs.\\n\\tInference in these models is turned into \\\\textit{evaluation} of the computational graph. The direction of the rules indicates the direction of the evaluation, in the same way as it indicates the direction of inference in logic programming. \\n\\tAmong this category are} systems based on Prolog or Datalog, such as TensorLog , Neural Theorem Provers (NTPs) , NLProlog , DeepProbLog , NLog and DiffLog .\\n\\tLifted Relational Neural Networks (LRNNs) and $\\\\partial$ILP are other examples of non-probabilistic directed models, where weighted definite clauses are compiled into a neural network architecture in a forward chaining fashion.\\n\\tThe systems that imitate logical reasoning with tensor calculus, Neural Logic Programming (NeuralLP) and Neural Logic Machines (NLM) , are likewise instances of directed logic. An example of a proof-based NeSy model is given in Example \\\\ref{ex:kbann}. \\n\\t\\\\begin{figure}[t]\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{subfigure}[b]{0.55\\\\textwidth}\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\begin{prolog}\\nalarm <- earthquake. \\nalarm <- burglary.\\ncalls_mary <- alarm, \\n hears_alarm_mary.\\n\\t\\t\\t\\\\end{prolog}\\n\\t\\t\\t\\\\caption{}\\n\\t\\t\\\\end{subfigure}\\n\\t\\t\\\\begin{subfigure}[b]{0.43\\\\textwidth}\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\include{to_include/and_or_simple}\\n\\t\\t\\t\\\\caption{}\\n\\t\\t\\\\end{subfigure}\\n\\t\\t\\\\begin{subfigure}[b]{0.48\\\\textwidth}\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\include{to_include/and_or_nn}\\n\\t\\t\\t\\\\caption{}\\n\\t\\t\\\\end{subfigure}\\n\\t\\t\\\\begin{subfigure}[b]{0.48\\\\textwidth}\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\include{to_include/and_or_hidden}\\n\\t\\t\\t\\\\caption{}\\n\\t\\t\\\\end{subfigure}\\n\\t\\t\\\\begin{subfigure}[b]{0.48\\\\textwidth}\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\include{to_include/and_or_fully}\\n\\t\\t\\t\\\\caption{}\\n\\t\\t\\\\end{subfigure}\\n\\t\\t\\\\caption{Knowledge-Based Artificial Neural Network. Network creation process. (1) the initial logic program; (2) the AND-OR tree for the query \\\\textit{calls\\\\_mary}; (3) mapping the tree into a neural network; (4) adding hidden neurons, (5) adding interlayer connections.}\\n\\t\\t\\\\label{fig:kbann}\\n\\t\\\\end{figure}\\n\\\\begin{boxedexample}[label = ex:kbann]{Knowledge-Based Artificial Neural Networks}\\n\\t\\\\noindent Knowledge-Based Artificial Neural Networks (KBANN) is the first method to use definite clausal logic and theorem proving to template the architecture of a neural network. \\n\\t\\\\noindent KBANN turns a program into a neural network in several steps:\\n\\t\\\\begin{enumerate}\\n\\t\\t\\\\item KBANN starts from a definite clause program and a set of queries.\\n\\t\\t\\\\item The program is turned into an AND-OR tree using the proofs for the queries.\\n\\t\\t\\\\item The AND-OR tree is turned into a neural network with a similar structure. Nodes are divided into layers. The weights and the biases are set such that evaluating the network returns the same outcome of querying the program.\\n\\t\\t\\\\item New hidden units are added. Hidden units play the role of unknown rules that need to be learned. They are initialized with zero weights; i.e. they are inactive.\\n\\t\\t\\\\item New links are added from each layer to the next one, obtaining the final neural network. \\n\\t\\\\end{enumerate}\\n\\t\\\\noindent An example of this process is shown in Figure~\\\\ref{fig:kbann}. KBANN needs some restrictions over the kind of rules. In particular, the rules are assumed to be conjunctive, non-recursive, and variable-free (or propositional). Many of these restrictions are removed by more recent systems.\\n\\\\end{boxedexample}\\n\\\\rev{We now survey the second class of NeSy systems, the model-based ones. These systems use logic to define a loss function (usually a regularization term) for neural networks. The networks compute scores for the set of atoms that correspond to the output neurons. At each training step, the logic-based loss function determines the degree to which the assigned scores violate the logical theory and uses this to determine the penalty. \\nLogical inference is turned into a learning problem (i.e. ``learning to satisfy'') and it is usually cast in a variational optimization scheme.\\\\footnote{This is reminiscent of the variational approach to probabilistic inference in probabilistic graphical models and constitutes a further parallel between the fields.}\\nAs a consequence, in constraint-based models, the neural network has to solve two tasks at the same time: solving a subsymbolic learning problem (e.g. perception) as well as approximating the logical inference process .}\\nA large group of NeSy approaches, including Semantic Based Regularization (SBR) , Logic Tensor Networks (LTN) , Semantic Loss (SL) and DL2 , exploits logical knowledge as a soft regularization constraint that favours solutions that satisfy the logical constraints. SBR and LTN compute atom (fuzzy) truth assignments as the output of the neural network and translate the provided logical formulas into a real valued regularization loss term using fuzzy logic. SL uses marginal probabilities of the target atoms to define the regularization term and relies on arithmetic circuits to evaluate it efficiently, as detailed in Example \\\\ref{ex:sl}. DL2 defines a numerical loss providing no specific fuzzy or probabilistic semantics, which allows for including numerical variables in the formulas (e.g. by using a logical term $x > 1.5$). Another group of approaches, including Neural Markov Logic Networks (NMLN) and Relational Neural Machines (RNM) extend MLNs, allowing factors of exponential distributions to be implemented as neural architectures. Finally, compute ground atoms scores as dot products between relation and entity embeddings; implication rules are then translated into a logical loss through a continuous relaxation of the implication operator.\\n\\t\\\\begin{boxedexample}[label = ex:sl]{Semantic Loss}\\n\\t\\t\\\\noindent The Semantic Loss is an example of an undirected model where (probabilistic) logic is exploited as a \\\\textit{regularization} term in training a neural model.\\n\\t\\t\\\\noindent Let $p = [p_1,\\\\dots,p_n]$ be a vector of probabilities for a list of propositional variables $X = [X_1,\\\\dots,X_n]$. In particular, $p_i$ denotes the probability of variable $X_i$ being \\\\textit{True} and corresponds to a single output of a neural net having $n$ outputs.\\n\\t\\tLet $\\\\alpha$ be a logic sentence defined over $X$.\\n\\t\\t\\\\noindent Then, the \\\\textit{semantic loss} between $\\\\alpha$ and $p$ is:\\n\\t\\t\\\\begin{equation*}\\n\\t\\tLoss(\\\\alpha,p) \\\\propto - \\\\log \\\\,\\\\, \\\\sum_{x \\\\models \\\\alpha} \\\\,\\\\,\\\\, \\\\prod_{i: x \\\\models X_i} p_i \\\\,\\\\,\\\\, \\\\prod_{i: x \\\\models \\\\neg X_i} (1-p_i).\\n\\t\\t\\\\end{equation*}\\n\\t\\t\\\\noindent The authors provide the intuition behind this loss: \\n\\t\\t\\\\begin{quote}\\n\\t\\t\\t\\\\textit{The semantic loss is proportional to the negative logarithm of the probability of generating a state that satisfies the constraint when sampling values according to $p$.}\\n\\t\\t\\\\end{quote} \\n\\t\\t\\\\noindent Suppose you want to solve a multi-class classification task (example adapted from ), where each input example must be assigned to a single class. Then, one would like to enforce \\\\textit{mutual exclusivity} among the classes. This can be easily done on supervised examples, by coupling a softmax activation layer with a cross entropy loss. However, there is no standard way to impose this constraint for unlabeled data, which can be useful in a semi-supervised setting.\\n\\t\\t\\\\noindent The solution provided by the Semantic Loss framework is to encode mutual exclusivity into the propositional constraint $\\\\beta$:\\n\\t\\t\\\\begin{equation*}\\n\\t\\t\\\\beta = (X_1 \\\\land \\\\neg X_2 \\\\land \\\\neg X_3) \\\\lor\\n\\t\\t(\\\\neg X_1 \\\\land X_2 \\\\land \\\\neg X_3) \\\\lor\\n\\t\\t(\\\\neg X_1 \\\\land \\\\neg X_2 \\\\land X_3)\\n\\t\\t\\\\end{equation*}\\n\\t\\tConsider a neural network classifier with three outputs $ p =[ p_1, p_2, p_3]$. Then, for each input example (whether labeled or unlabeled), we can build the semantic loss term:\\n\\t\\t\\\\begin{equation*}\\n\\t\\tL(\\\\beta,p) = p_1(1 - p_2)(1 - p_3) +\\n\\t\\t(1 - p_1)p_2(1 - p_3) +\\n\\t\\t(1 - p_1)(1 - p_2)p_3\\n\\t\\t\\\\end{equation*}\\n\\t\\t\\\\noindent It can be summed up to the standard cross-entropy term for the labeled examples.\\n\\t\\tUnlike for directed methods such as KBANN (Example \\\\ref{ex:kbann}) and TensorLog, \\n the logic is turned into a loss-function that is used during training. The function constrains the underlying probabilities, but there are no directed or causal relationships among them.\\n\\tMoreover, during inference only the probabilities $p$ are used while the logic formula $\\\\beta$ is not used anymore. On the contrary, in KBANN, the logic is compiled into the architecture of the network and, therefore, it it is also exploited at evaluation time. \\n\\t\\\\end{boxedexample}\\n\\t\\\\rev{To conclude, let us stress a key difference between \\n the two classes of NeSy systems w.r.t. \\\\textit{the way they incorporate the knowledge expressed in the logical clauses}. Proof-based, directed models use logic to define the architecture of a neural symbolic network. \\n Thus, logic is part of the inference of the model and acts as a structural constraint. The designer has full control of where and how the logic is used inside the network. Thus, logical knowledge can easily be extended or modified at test-time, without the need to retrain, leading to a high degree of modularity and out-of-distribution generalization . On the other hand, when logic is only encoded in an objective function, the neural net learns to (approximately) satisfy it. Therefore, the knowledge is only latently encoded in the weights of the network, which leads to a loss of control and interpretability. However, the latter techniques are often much more scalable, especially at inference time. The balance between control and interpretability, on the one hand, and scalability, on the other hand, is an open and important research question in the \\\\nesy{} community. }\",\n \"id\": \"dfea86d8-fb1c-484d-a9d2-d17c03c4d644\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"71d65d84-75cd-4c92-b14d-bddcaa016dd0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Proof- vs Model-theoretic View of Logic\"\n ],\n [\n \"subsection\",\n \"Implications for NeSy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for NeSy\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:syntax}\\n\\tIn Section \\\\ref{sec:proof_vs_model}, we have introduced clausal logic, without paying much attention to the structure of the atoms and literals. This structure and its consequences for StarAI and \\\\nesy{} models are the topic of the present section. Consider the following example: \\n\\t\\\\begin{boxedexample}[label = ex:prop_clausal_logic]{Propositional Clausal logic}\\n\\t\\tConsider the following set of clauses. \\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tmortal_socrates $\\\\leftarrow$ human_socrates.\\n\\t\\tmortal_aristotle $\\\\leftarrow$ human_aristotle.\\n\\t\\thuman_socrates.\\n\\t\\thuman_aristotle.\\n\\t\\t\\\\end{lstlisting}\\n\\t\\\\end{boxedexample} \\n\\tHere, the literals do not possess any internal structure. \\\\rev{They are {\\\\em propositions}, which are atoms that we can only assign the value $True$ or $False$. We say that we are working in \\\\textit{propositional logic}. }\\n\\tThis contrasts with\\n\\t\\\\textit{first-order} logic in which the literals take the form $p(t_1, ... , t_m)$, with $p$ a predicate of arity $m$ and the $t_i$ terms, that is, constants, variables, or structured terms of the form $f(t_1, ..., t_q)$, where $f$ is a functor and the $t_i$ are again terms. \\n\\tIntuitively, constants represent objects or entities, functors represent functions, variables make abstraction of specific objects, and predicates specify relationships amongst objects.\\n\\tThe subset of first order logic where there are no functors is called {\\\\em relational logic}. \\n\\t\\\\begin{boxedexample}[label=ex:fol_clausal_logic]{First order clausal logic}\\n\\t\\tIn contrast to the previous example, we now\\n\\t\\twrite the theory in a more compact manner using \\\\textit{first order logic}. By convention, constants start with a lowercase letter, while variables start with an uppercase.\\n\\t\\tEssential is the use of the variable \\\\textit{X}, which is implicitly universally quantified.\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tmortal(X) $\\\\leftarrow$ human(X). \\n\\t\\thuman(socrates).\\n\\t\\thuman(aristotle).\\n\\t\\t\\\\end{lstlisting}\\n\\t\\\\end{boxedexample} \\n\\t\\\\rev{It is interesting to understand the connection between propositional, relational and first-order logic. To this end, we introduce the concept of grounding.} \\n\\tWhen an expression (i.e, clause, atom or term) does not contain any variable it is called {\\\\em ground}. \\n\\tA substitution $\\\\theta$ is an expression of the form $\\\\{X_1/c_1, ..., X_k/c_k \\\\}$ with the $X_i$ different variables, the $c_i$ terms. Applying a substitution $\\\\theta$ to an expression $e$ (term, atom or clause) yields the instantiated expression $e\\\\theta$ where all variables $X_i$ in $e$ have been simultaneously replaced by their corresponding terms $c_i$ in $\\\\theta$. We can take for instance the clause $mortal(X) \\\\leftarrow human(X)$ and apply the substitution $\\\\{X/socrates\\\\}$ to yield $mortal(socrates) \\\\leftarrow human(socrates)$. Grounding is the process whereby all possible substitutions that ground the clauses are applied. Notice that grounding a first order logical theory may result in an infinite set of ground clauses (when there are functors), and a polynomially larger set of clauses (when working with finite domains).\\n\\t\\\\rev{Finite domains are the focus in both StarAI and \\\\nesy{}. In such domains,} any problem expressed in first-order logic can be equivalently expressed in relational logic and any problem expressed in relational logic can likewise be expressed in propositional logic by grounding out the clauses .\",\n \"id\": \"394ef4e2-3d6b-46dc-9354-ff1b43665302\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Syntax\"\n ]\n ],\n \"subsections\": [\n \"3524cff6-f045-4e2a-be1c-cdf237557b34\",\n \"7f592cdb-5aaa-4d78-a9de-739ab8dc18a6\"\n ],\n \"title\": \"Logic - Syntax\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"StarAI typically focus on first order logic ~. \\\\rev{In Section \\\\ref{sec:proof_vs_model}, we have seen how StarAI models can be easily interpreted in terms of probabilistic graphical models (PGM). Here, we want to show that FOL is a powerful tool for building such models.} \\n\\t\\\\begin{evidencebox}\\n\\t\\t\\\\rev{FOL allows for knowledge in the form of logical rules to be interpreted as a template for defining the graphical models. Grounding the theory then corresponds to \\\\textit{unrolling} the template. At the same time, first order logic has also an important statistical and learning advantage: a FOL rule leads to parameter sharing in the model as the parameters of a single FOL rule are tied to all its groundings. Parameter sharing compresses the representation of the corresponding probabilistic model, resulting in more efficient learning and better generalization.}\\n\\t\\\\end{evidencebox}\\n\\t\\\\rev{These properties are reminiscent of plate notation for probabilistic graphical models, bringing logical reasoning into the picture . In Example \\\\ref{ex:mln}, we have used only two first order rules but we obtained a larger graphical model with six factors (see Figure \\\\ref{fig:mln}) by grounding (i.e. unrolling) the rules over the domain. All factors corresponding to the same rule share the same weight.}\",\n \"id\": \"3524cff6-f045-4e2a-be1c-cdf237557b34\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"394ef4e2-3d6b-46dc-9354-ff1b43665302\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Syntax\"\n ],\n [\n \"subsection\",\n \"Implications for StarAI\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for StarAI\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 5714,\n 3741,\n 5712,\n 5710,\n 5707,\n 7304,\n 4323,\n 5709,\n 5713,\n 8943,\n 3726,\n 3727\n ],\n \"content\": \"}\\n\\t\\\\rev{NeSy exploits the internal structure of literals, resulting in many relational and first-order systems. System exploiting propositional logic are Semantic Loss (SL) ~ and DL2~. Relational logic-based systems are DiffLog~, $\\\\theta$ILP , Lifted Relational Neural Networks (LRNN) , Neural Theorem Provers (NTP) \\n\\tand NeurASP~.\\n\\t\\tFinally, many systems are based on first-order logic or first-order logic programs, such as \\n\\t\\tDeepProbLog , NLog , NLProlog , \\n\\t\\tDeepStochLog ,\\n\\t\\tLogic Tensor Networks , Semantic Based Regularization , Relational Neural Machines and Logical Neural Networks .}\\n\\t\\\\rev{The focus in \\\\nesy{} on structured terms is strongly related to that in StarAI and plays a fundamental role in \\\\nesy{}. \\n\\t In fact, grounding a relational or first-order theory can often be seen as unrolling either the architecture (e.g., DeepStochLog, LRNN ) or the loss function (e.g., SBR , LTN ) of the corresponding neural model. Unrolling fixed modules over multiple elements of a complex data structure is fundamental to neural networks on sequences (recurrent nets, RNN), trees (recursive nets, RvNN) and graphs (graph nets, GNN). \\\\nesy{} can be regarded as unrolling more complex logical structures, with similar benefits in terms of model capacity, modularization and generalization, and strong control due to the formal semantics.}\\n\\t\\\\rev{Moreover, first-order \\\\nesy{} models can explicitly deal with how subsymbolic data (e.g. images or audio) are fed to the neural components of the system.\\n In fact, \\\\nesy{} systems often use subsymbolic data samples as elements of the domain of discourse. For example, the element \\\\textit{mary} can be used to refer to an image, e.g. $mary = \\\\smallimg{mary.jpeg}$. Feeding such samples as input to a neural network can then be naturally encoded as grounding a predicate over the domain of interest. When the internal structure of the literals is absent, as in SL, this mapping must be handled outside the logical framework.}\\nWhile both relational logic and first-order logic have their advantages, there is a noteworthy distinction in the latter. \\nFirst-order logic allows representing real valued functions through the use of functors. For example, segmentation can be modeled as a functor returning the bounding box of an object inside an image, e.g. \\\\textit{location(mary,image)} . Therefore, FOL-based systems can address regression tasks, diverging from the conventional classification tasks associated with relational logic systems.\",\n \"id\": \"7f592cdb-5aaa-4d78-a9de-739ab8dc18a6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"394ef4e2-3d6b-46dc-9354-ff1b43665302\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Syntax\"\n ],\n [\n \"subsection\",\n \"Implications for \\\\nesy{\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for \\\\nesy{\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:semantics}\",\n \"id\": \"3f30c5c1-1c7c-4ae0-b5f0-88220243574e\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Semantics\"\n ]\n ],\n \"subsections\": [\n \"8d5d9e2a-e3dc-4e1e-ba7c-32324338d1ca\",\n \"5422e096-fdd4-4f74-a7c2-b0a811af5aa2\",\n \"09725807-0a82-437e-980b-ae2733823d35\",\n \"0c28445b-24b7-4df7-a61f-1e0537492ed8\"\n ],\n \"title\": \"Logic - Semantics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\rev{The semantics of logical, probabilistic logical and neurosymbolic systems is defined in terms of a model theoretic semantics.}\\n\\\\rev{In the present section, we will restrict our attention to Herbrand interpretations and models as is usual in logic programming and statistical relational AI (see Section \\\\ref{sec:proof_vs_model})}.\\n\\\\rev{We can distinguish three different levels of semantics, which are also closely tied to the used syntax of the underlying logic.}\\n\\\\rev{First, when the logical theory consists of definite clauses only, the semantics is given by the least Herbrand model. The least Herbrand model of a definite clause theory\\nis unique and it is the smallest w.r.t. set inclusion. It contains all ground facts (from the Herbrand domain) that are logically entailed by the theory.\\nFor instance, considering the facts $a$ and $b$ and the rules $d \\\\leftarrow a, b$ and $b \\\\leftarrow c$ would give the least Herbrand model $\\\\{a, b, d\\\\}$.}\\n\\\\rev{Second, when the logical theory can contain any set of clauses, the semantics is given by the set of all possible Herbrand models. For instance, considering the clause $a \\\\vee b $\\nyields the models $\\\\{a\\\\}, \\\\{b\\\\}$ and $\\\\{a, b\\\\}$. So there is not necessarily a unique model, not even when considering only minimal models, where\\nwe have $\\\\{a\\\\}, \\\\{b\\\\}$.} \\n\\\\rev{Third, while Horn-clauses are the basis for \\\"pure\\\" Prolog and logic programs, there exist several extensions to this formalism to accommodate\\nnegated literals in the condition part of rules or disjunction in the head. A popular framework in this regard is answer set programming (ASP). \\nIn ASP the clause $a \\\\vee b$ could be represented by two clauses $a \\\\leftarrow \\\\neg b$ and $b \\\\leftarrow \\\\neg a$ which would have two stable models $\\\\{a\\\\}$ and $\\\\{b\\\\}$.}\",\n \"id\": \"8d5d9e2a-e3dc-4e1e-ba7c-32324338d1ca\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3f30c5c1-1c7c-4ae0-b5f0-88220243574e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Semantics\"\n ],\n [\n \"subsection\",\n \"Model-theoretic semantics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model-theoretic semantics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\rev{The previous three levels of semantics are based on Boolean models, i.e. models where each atom is either present (i.e. \\\\textit{True}) or absent (i.e. \\\\textit{False})}. Differently, \\\\textit{fuzzy logic}, and in particular t-norm fuzzy logic, assigns a truth value to atoms in the \\\\textit{continuous} real interval $[0,1]$. Logical operators are then turned into real-valued functions, mathematically grounded in the t-norm theory. A \\\\textit{t-norm} $t(x, y)$ is a real function $t:[0,1] \\\\times [0,1] \\\\to [0,1]$ that models the logical AND and from which the other operators can be derived. Table~\\\\ref{tab:t-norms} shows well-known t-norms and the functions corresponding to their connectives. A fuzzy logic formula is mapped to a real valued function of its input atoms, as we show in Example~\\\\ref{ex:fuzzy}. Fuzzy logic generalizes Boolean logic to continuous values. All the different t-norms are coherent with Boolean logic in the endpoints of the interval $[0,1]$, which correspond to \\\\textit{completely true} and \\\\textit{completely false} values. The concept of model in fuzzy logic can be easily recovered from an extension of the model-theoretic semantics of the Boolean logic. Any \\\\textit{fuzzy} interpretation is a model of a formula if the formula evaluates to 1.\\n\\\\begin{boxedexample}[label=ex:fuzzy]{Fuzzy logic}\\n\\t\\tLet us consider the following propositions: \\\\textit{alarm}, \\\\textit{burglary} and \\\\textit{earthquake}. Defining a fuzzy semantics for this language requires one to assign truth degrees to each of the propositions and selecting a particular t-norm to implement the connectives. \\n\\t\\tLet us consider the \\\\L ukasiewicz t-norm and the following interpretation of the language:\\n\\t\\t\\\\begin{minipage}{0.40\\\\linewidth}\\n\\t\\t\\t\\\\begin{align*}\\n\\t\\t\\t\\\\mathcal{I} = \\\\{&\\\\texttt{alarm} = 0.7, \\\\\\\\\\n\\t\\t\\t&\\\\texttt{burglary}=0.6, \\\\\\\\ &\\\\texttt{earthquake}=0.3\\\\}\\n\\t\\t\\t\\\\end{align*}\\n\\t\\t\\\\end{minipage}\\n\\t\\t\\\\begin{minipage}{0.60\\\\linewidth}\\n\\t\\t\\t\\\\begin{align*}\\n\\t\\t\\tt_{\\\\vee}(x,y) & = \\\\min ( 1, x + y)\\\\\\\\\\n\\t\\t\\tt_{\\\\rightarrow}(x,y) & = \\\\min(1, 1 - x + y)\\n\\t\\t\\t\\\\end{align*}\\n\\t\\t\\\\end{minipage}\\n\\t\\t\\\\newline\\n\\t\\tOnce we have defined the semantics of the language, we can evaluate logic sentences, e.g.:\\n\\t\\t\\\\begin{align*}\\n\\t\\t& \\\\texttt{alarm} \\\\leftarrow (\\\\texttt{burglary} \\\\vee \\\\texttt{earthquake}) = \\\\\\\\\\n\\t\\t& \\\\min(1,1 - \\\\min(1, \\\\texttt{burglary} + \\\\texttt{earthquake}) +\\\\texttt{alarm}) = 0.8\\n\\t\\t\\\\end{align*}\\n\\t\\tThis evaluation can be performed automatically by parsing the logical sentence in the corresponding \\\\textit{expression tree} and then compiling the expression tree using the corresponding t-norm operation: \\n\\t\\t\\\\begin{center}\\n\\t\\t\\t\\\\include{to_include/expression_tree_fuzzy}\\n\\t\\t\\\\end{center}\\n\\t\\tThe resulting circuit represents a differentiable function and the truth degree of the sentence is computed by evaluating the circuit bottom-up.\\n\\t\\\\end{boxedexample}\\n\\t\\\\begin{table}[tb]\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{c|c|c|c}\\n\\t\\t\\t& Product & \\\\L ukasiewicz & G\\\\\\\"{o}del \\\\\\\\ \\n\\t\\t\\t\\\\hline\\n\\t\\t\\t$x \\\\land y$ & $x \\\\cdot y$ & $\\\\max(0, x + y -1$) & $\\\\min(x, y$) \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t$x \\\\lor y$ & $x + y - x \\\\cdot y$ & $\\\\min(1, x + y)$ & $\\\\max(x, y)$ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t$\\\\lnot x$ & $1 - x$ & $1 - x$ & $1 - x$ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t$x \\\\Rightarrow y \\\\;\\\\; (x>y)$ & $y/x$ & $\\\\min(1,1-x+y)$ & $y$ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\t\\\\caption{Logical connectives on the inputs $x,y$ when using the fundamental t-norms.}\\n\\t\\t\\\\label{tab:t-norms}\\n\\t\\\\end{table}\",\n \"id\": \"5422e096-fdd4-4f74-a7c2-b0a811af5aa2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3f30c5c1-1c7c-4ae0-b5f0-88220243574e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Semantics\"\n ],\n [\n \"subsection\",\n \"Fuzzy semantics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Fuzzy semantics\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 8685,\n 5706\n ],\n \"content\": \"Statistical Relational AI has extended the previous semantics by defining probability distributions $p(\\\\omega)$ over models, or \\\\textit{possible worlds}\\\\footnote{In this paper, we use the distribution semantics as representative of the probabilistic approach to logic. While this is the most common solution in StarAI, many other solutions exist , whose description is out of the scope of the current survey. A detailed overview of the different flavours of formal reasoning about uncertainty can be found in .}.\\nThe goal is to reason about the uncertainty of logical statements. In particular, the probability that a certain formula $\\\\alpha$ holds is computed as the sum of the probabilities of the possible worlds that are models of $\\\\alpha$ (i.e. where $\\\\alpha$ is True):\\n\\t\\\\begin{equation}\\n\\t\\\\label{eq:wmc}\\n\\tp(\\\\alpha) = \\\\sum_{\\\\omega \\\\models \\\\alpha} p(\\\\omega) \\n\\t\\\\end{equation}\\n\\tThis is an instance of the Weighted Model Counting (WMC) problem. In fact, we are counting how many worlds are models of $\\\\alpha$ and we are weighting each of them by its probability according to the distribution $p(\\\\omega)$.\\n\\t\\\\begin{table}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{cccc|l}\\n\\t\\t\\tB & E & J & M & $p(\\\\omega)$ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tF&F&F&F& 0.2394 \\\\\\\\\\n\\t\\t\\tF&F&F&T& 0.1026 \\\\\\\\\\n\\t\\t\\tF&F&T&F& 0.3591 \\\\\\\\\\n\\t\\t\\tF&F&T&T& 0.1539 \\\\\\\\\\n\\t\\t\\tF&T&F&F& 0.0126 \\\\\\\\\\n\\t\\t\\tF&T&F&T& 0.0054 \\\\\\\\\\n\\t\\t\\tF&T&T&F& 0.0189 \\\\\\\\\\n\\t\\t\\tF&T&T&T& 0.0081 \\\\\\\\\\n\\t\\t\\tT&F&F&F& 0.0266 \\\\\\\\\\n\\t\\t\\tT&F&F&T& 0.0114 \\\\\\\\\\n\\t\\t\\tT&F&T&F& 0.0399 \\\\\\\\\\n\\t\\t\\tT&F&T&T& 0.0171 \\\\\\\\\\n\\t\\t\\tT&T&F&F& 0.0014 * \\\\\\\\\\n\\t\\t\\tT&T&F&T& 0.0006 * \\\\\\\\\\n\\t\\t\\tT&T&T&F& 0.0021 * \\\\\\\\\\n\\t\\t\\tT&T&T&T& 0.0009 * \\n\\t\\t\\\\end{tabular}\\n\\t\\t\\\\caption{A distribution over possible worlds for the four propositional variables $burglary$ (B), $earthquake$ (E), $hears\\\\_alarm\\\\_john$ (J) and $hears\\\\_alarm\\\\_mary$ (M). The $*$ indicates those worlds where $burglary \\\\wedge earthquake$ is \\\\textit{True}.}\\n\\t\\t\\\\label{tab:distribution_semantics}\\n\\t\\\\end{table}\\n\\t\\\\begin{boxedexample}[label = ex:distr_semantics]{Probabilistic Logic}\\n\\t\\tLet us consider the following set of propositions $B = burglary$, $E = earthquake$, $J = hears\\\\_alarm\\\\_john$ and $M = hears\\\\_alarm\\\\_mary$. In probabilistic logic, a probability distribution over all the possible worlds is defined. For example, Table \\\\ref{tab:distribution_semantics} represents a valid distribution.\\n\\t\\tSuppose we want to compute the probability of the formula $burglary \\\\wedge earthquake$. This is done by summing up the probabilities of all the worlds where both $burglary$ and $earthquake$ are \\\\textit{True} (indicated by a $*$ in Table \\\\ref{tab:distribution_semantics}). \\n\\t\\\\end{boxedexample}\\n The StarAI community has provided several formalisms to define such probability distributions over possible worlds using labeled logic theories. Probabilistic Logic Programs (cf. Example \\\\ref{ex:problog}) and Markov logic networks (cf. Example \\\\ref{ex:mln}) are two prototypical frameworks. For example, the distribution in Table~\\\\ref{tab:distribution_semantics} is the one modeled by the ProbLog program in Example \\\\ref{ex:problog}.\\n\\\\rev{It is interesting to compare Markov Logic (Example \\\\ref{ex:mln}) to ProbLog (Example \\\\ref{ex:problog}) in terms of their model-theoretic semantics. Markov Logic is defined as a set of weighted full clauses, i.e. as an unnormalized probability distribution over full clausal theories.\\nThis means that, given any subset of the theory, there can be many possible models. For instance, the theory $a \\\\vee b$, has three possible models. To obtain a probability distribution over models, Markov Logic needs to distribute the probability mass over its models. \\nTo do this, the maximum entropy principle is used, which results in equal distributions of the probability mass.} \\n\\\\rev{Conversely, ProbLog defines a probability distribution over definite clause theories, each obtained as subsets of the provided probabilistic facts. However, since each of these theories has a unique least Herbrand model, the probability mass corresponding to the selected facts is assigned to the corresponding unique Herbrand model. This means that when working with definite clauses only, there is no need to distribute the probability mass to multiple models and, therefore, no extra assumptions such as maximum entropy are necessary.} \\nProbabilistic inference (i.e. weighted model counting) is generally intractable. That is why, in StarAI, techniques such as \\\\textit{knowledge compilation} (KC)~ are used.\\n\\tKnowledge compilation transforms a logical formula $\\\\alpha$ into a new representation in an offline step, which can be computationally expensive. Using this new representation a particular set of queries can be answered efficiently (i.e. in poly-time in the size of the new representation). \\n\\tFrom a probabilistic point of view, this translation solves the disjoint-sum problem, which states that one cannot simply sum up the probability \\n of two disjuncts but also has to subtract the probability of the intersection.\\n After the translation, the probabilities of any conjunction and of any disjunction can be simply computed by multiplying, resp. summing up, the probabilities of their operands. Thus a logical formula $\\\\alpha$ can be compiled into an arithmetic circuit $ac(\\\\alpha)$. The weighted model count of the query formula can then simply be computed by evaluating the corresponding arithmetic circuit bottom up; i.e. $p(\\\\alpha) = ac(\\\\alpha)$. \\n\\t\\\\begin{figure}[t]\\n\\t\\t\\\\centering\\n\\t\\t\\\\noindent\\n\\t\\t\\\\begin{subfigure}{0.4\\\\linewidth}\\n\\t\\t\\t\\\\include{to_include/dDNNF}\\n\\t\\t\\\\end{subfigure}\\n\\t\\t\\\\hfill\\n\\t\\t\\\\begin{subfigure}{0.4\\\\linewidth}\\n\\t\\t\\t\\\\include{to_include/arithmetic_circuit}\\n\\t\\t\\\\end{subfigure}\\n\\t\\t\\\\caption{dDNNF (left) and arithmetic circuit (right) corresponding to the ProbLog program in Example \\\\ref{ex:problog}}\\n\\t\\t\\\\label{fig:kc}\\n\\t\\\\end{figure}\\n\\t\\\\begin{boxedexample}[label = ex:kc]{Knowledge Compilation}\\n\\t\\tLet us consider the ProbLog program in Example \\\\ref{ex:problog} and the corresponding tabular representation in Table \\\\ref{tab:distribution_semantics}. Let us consider the query $q = calls(mary)$. Now we can use Equation~\\\\ref{eq:wmc} to compute the probability $p(q)$. To do this, we iterate over the table and we sum all the probabilities of the worlds where $calls(mary)$ is True, which we know from Example \\\\ref{ex:logic_program} are those where either $burglary=T$ or $earthquake=T$ and where $hears\\\\_alarm(mary)=T$. This yields $p(q) = 0.0435$. This method would require us to iterate over $2^N$ terms (where $N$ is the number of probabilistic facts).\\n\\t\\tKnowledge compilation compiles $\\\\alpha$ into some normal form that is logically equivalent. In Figure \\\\ref{fig:kc}, the target representation is a decomposable, deterministic negative normal form (d-DNNF)~, for which weighted model counting is poly-time in the size of the formula. Decomposability means that, for every conjunction, the two conjuncts do not share any variables. Deterministic means that, for every disjunction, the two disjuncts are \\\\rev{mutually exclusive, i.e., only one of the disjuncts can be true at the same time}. \\n\\t\\tThe formula in d-DNNF can then be straightforwardly turned into an arithmetic circuit by substituting AND nodes with multiplication and OR nodes by summation. \\n In Figure \\\\ref{fig:kc}, we show the d-DNNF and the arithmetic circuit of the distribution defined by the ProbLog program in Example \\\\ref{ex:problog}. The bottom-up evaluation of this arithmetic circuit computes the correct marginal probability $p(\\\\alpha)$ much more efficiently than the naive iterative sum that we have computed before. \\n\\t\\\\end{boxedexample}\\n\\tEven though probabilistic Boolean logic is the most common choice in StarAI, some approaches use probabilistic fuzzy logic. The most prominent approach is Probabilistic Soft Logic (PSL) , illustrated in Example~\\\\ref{ex:psl}. Similarly to Markov logic networks, Probabilistic Soft Logic (PSL) defines log linear models where features are represented by ground clauses. However, PSL uses a fuzzy semantics of the logical theory. Therefore, atoms are mapped to real valued variables and ground clauses to real valued factors.\\n\\t\\\\begin{boxedexample}[label = ex:psl]{Probabilistic Soft Logic}\\n\\t\\tLet us consider the logical rule $\\\\alpha = smokes(X) \\\\leftarrow stress(X)$ with weight $\\\\beta$.\\n\\t\\tAs we have seen in Example~\\\\ref{ex:mln}, Markov Logic translates the formula into a discrete factor by using the indicator functions $\\\\mathbbm{1}(\\\\omega, \\\\alpha\\\\theta)$:\\n\\t\\t\\\\begin{equation*}\\n\\t\\t\\\\phi^{MLN}(\\\\omega, \\\\alpha) = \\\\beta \\\\mathbbm{1}(\\\\omega, \\\\alpha\\\\{X/mary\\\\}) + \\\\beta \\\\mathbbm{1}(\\\\omega, \\\\alpha\\\\{X/john\\\\})\\n\\t\\t\\\\end{equation*}\\n\\t\\tInstead of discrete indicator functions, PSL translates the formula into a continuous t-norm based function:\\n\\t\\t\\\\begin{equation*}\\n\\t\\tt(\\\\omega, \\\\alpha) = \\\\min(1, 1 - stress(X) + smokes(X))\\n\\t\\t\\\\end{equation*}\\n\\t\\tand the corresponding potential is then translated into the continuous and differentiable function:\\n\\t\\t\\\\begin{equation*}\\n\\t\\t\\\\phi^{PSL}(\\\\omega, \\\\alpha) = \\\\beta t(\\\\omega, \\\\alpha\\\\{X/mary\\\\}) + \\\\beta t(\\\\omega, \\\\alpha\\\\{X/john\\\\})\\n\\t\\t\\\\end{equation*}\\n\\t\\tAnother important task in StarAI is MAP inference. In MAP inference, given the distribution $p(\\\\omega)$, one is interested in finding the interpretation $\\\\omega^\\\\star$ where $p$ is maximal, i.e.\\n\\t\\t\\\\begin{equation}\\n\\t\\t\\\\label{eq:map}\\n\\t\\t\\\\omega^\\\\star = \\\\text{arg}\\\\max_\\\\omega p(\\\\omega)\\n\\t\\t\\\\end{equation}\\n\\t\\tWhen the $\\\\omega$ is a boolean interpretation, i.e. $\\\\omega \\\\in \\\\{0,1\\\\}^n$, like in ProbLog or MLNs, this problem is \\\\rev{related to maxSAT, which is NP-hard}. However, in PSL, $\\\\omega$ is a fuzzy interpretation, i.e. $\\\\omega \\\\in [0,1]^n$ and $p(\\\\omega) \\\\propto \\\\exp\\\\big(\\\\sum_i \\\\beta_i \\\\phi(\\\\omega, \\\\alpha_i)\\\\big)$ is a continuous and differentiable function. The MAP inference problem can thus be \\\\textit{approximated} more efficiently than its boolean counterpart using gradient-based techniques.\\n\\t\\\\end{boxedexample}\",\n \"id\": \"09725807-0a82-437e-980b-ae2733823d35\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"3f30c5c1-1c7c-4ae0-b5f0-88220243574e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Semantics\"\n ],\n [\n \"subsection\",\n \"Implications for StarAI\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for StarAI\"\n },\n {\n \"cite_extract_rate\": 0.695652173913043,\n \"cites\": [\n 5712,\n 5707,\n 3726,\n 5718,\n 5717,\n 5715,\n 681,\n 5716,\n 3727,\n 5719,\n 5721,\n 5710,\n 5711,\n 5720,\n 4323,\n 3672\n ],\n \"content\": \"We have seen that in StarAI, one can turn inference tasks into the evaluation (as in KC) or gradient-based optimization (as in PSL) of a differentiable parametric circuit. The parameters are scalar values (e.g. probabilities or truth degrees) that are attached to basic elements of a logical theory (facts or clauses).\\n\\tA natural way of carrying over the StarAI approach to NeSy is the reparameterization method. Reparameterization substitutes the scalar values assigned to facts or formulas with the output of a neural network. One can interpret this substitution in terms of a different parameterization of the original model. Many probabilistic methods parameterize the underlying distribution in terms of neural components. In particular, as we show in Example \\\\ref{ex:deepproblog}, DeepProbLog exploits neural predicates to compute the probabilities of probabilistic facts as the output of neural computations over vectorial representations of the constants, which is similar to SL in the propositional counterpart (see Example \\\\ref{ex:sl}). NeurASP also inherits the concept of a neural predicate from DeepProbLog.\\n\\t\\\\begin{boxedexample}[label = ex:deepproblog]{Probabilistic semantics reparameterization in DeepProbLog}\\n\\t\\tDeepProbLog is a neural extension of the probabilistic logic programming language ProbLog. DeepProbLog allows images or other subsymbolic representations as terms of the program.\\n\\t\\tLet us consider a possible neural extension of the program in Example~\\\\ref{ex:problog}. We could extend the predicate $calls(X)$ with two extra inputs, i.e. $calls(B,E,X)$. $B$ is supposed to contain an image of a security camera, while $E$ is supposed to contain the time-series of a seismic sensor. We would like to answer queries like $calls(\\\\smallimg{burglary1.png},\\\\smallimg{earthquake1.png},mary)$, i.e. what is the probability that $mary$ calls, given that the security camera has captured the image $\\\\smallimg{burglary1.png}$ and the sensor the signal $\\\\smallimg{earthquake1.png}$ .\\n\\t\\tDeepProbLog can answer this query using the following program:\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tnn(nn_burglary, [B]) :: burglary(B).\\n\\t\\tnn(nn_earthquake, [E]) :: earthquake(E).\\n\\t\\t0.3::hears_alarm(mary). \\n\\t\\t0.6::hears_alarm(john). \\n\\t\\talarm(B,_) <- burglary(B).\\n\\t\\talarm(_,E) <- earthquake(E).\\n\\t\\tcalls(B,E, X) <- alarm(B,E), hears_alarm(X).\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tHere, the program has been extended in two ways. First, new arguments (i.e. $B$ and $E$) have been introduced in order to deal with the subsymbolic inputs. Second, the probabilistic facts $burglary$ and $earthquake$ have been turned into \\\\textit{neural predicates}. Neural predicates are special probabilistic facts that are annotated by neural networks instead of by scalar probabilities.\\n\\t\\tInference in DeepProbLog mimics that of ProbLog. Given the query and the program, knowledge compilation is used to build the arithmetic circuit in Figure \\\\ref{fig:deepproblog}.\\n\\t\\tSince the program is structurally identical to the purely symbolic one in Example \\\\ref{ex:kc}, the arithmetic circuit is exactly the same. The only only difference is that some leaves of the tree (i.e. capturing probabilities of facts) can now also be neural networks.\\n\\t\\tGiven a set of queries that are \\\\textit{True}, i.e.:\\n\\t\\t\\\\begin{align*}\\n\\t\\t\\\\mathcal{D} = \\\\{&calls(\\\\smallimg{burglary1.png},\\\\smallimg{earthquake1.png},mary),\\\\\\\\ &calls(\\\\smallimg{burglary2.png},\\\\smallimg{earthquake2.png},john), \\\\\\\\ &calls(\\\\smallimg{burglary3.png},\\\\smallimg{earthquake3.png},mary), ...\\\\},\\n\\t\\t\\\\end{align*} we can train the parameters $\\\\theta$ of the DeepProbLog program (both neural networks and scalar probabilities) by maximizing the log-likelihood of the training queries using gradient descent:\\n\\t\\t\\\\begin{equation*}\\n\\t\\t\\\\max_{\\\\theta} \\\\sum_{q \\\\in \\\\mathcal{D}} \\\\log p(q)\\n\\t\\t\\\\end{equation*}\\n\\t\\\\end{boxedexample}\\n\\t\\\\begin{figure}[t]\\n\\t\\t\\\\centering\\n\\t\\t\\\\include{to_include/arithmetic_circuit_dpl}\\n\\t\\t\\\\caption{A neural reparametrization of the arithmetic circuit in Example \\\\ref{ex:kc} as done by DeepProbLog (cf. Example \\\\ref{ex:deepproblog}). Dashed lines indicate a negative output, i.e 1 - x. \\\\rev{We use a different notation for negation than in Figure \\\\ref{fig:kc} to stress that both leaves are parameterized by the same neural network}.}\\n\\t\\t\\\\label{fig:deepproblog}\\n\\t\\\\end{figure}\\n\\tSimilarly to DeepProbLog, NMLNs and RNMs use neural networks to parameterize the factors (or the weights) of a Markov Logic Network.\\n\\t computes marginal probabilities as logistic functions over similarity measures between embeddings of entities and relations. An alternative solution to exploit a probabilistic semantics is to use knowledge graphs (see also \\\\ref{sec:kge_gnn}) to define probabilistic priors to neural network predictions, as done in .\\n\\tSBR and LTN reparametrize fuzzy atoms using neural networks that take as inputs the feature representation of the constants and return the corresponding truth value, as shown in Example~\\\\ref{ex:sbr}. Logical rules are then relaxed into soft constraints using fuzzy logic. Many other systems exploit fuzzy logic to inject knowledge into neural models . These methods can be regarded as variants of a unique conceptual framework as the differences are often minor and in the implementation details. \\n\\t\\\\begin{boxedexample}[label = ex:sbr]{Semantic-Based Regularization}\\n\\t\\t\\\\noindent Semantic-Based Regularization (SBR) is an example of an undirected model where fuzzy logic is exploited as a \\\\textit{regularization} term when training a neural model. \\n\\t\\tLet us consider a possible grounding for the rule in Example \\\\ref{ex:psl}:\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tsmokes(mary) $\\\\leftarrow$ stress(mary)\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tFor each grounded rule $r$, SBR builds a regularization loss term $L(r)$ in the following way. First, it maps each constant $c$ (e.g. \\\\textit{mary}) to a set of (perceptual) features $x_c$ (e.g. a tensor of pixel intensities $x_\\\\texttt{mary}$). Each relation $r$ (e.g. \\\\textit{smokes, stress}) is then mapped to a neural network $f_r(x)$, where $x$ is the tensor of features of the input constants and the output is a truth degree in $[0,1]$. For example, the atom \\\\textit{smokes(mary)} is mapped to the function call $f_\\\\texttt{smokes}(x_\\\\texttt{mary})$. \\n\\t\\tThen, a fuzzy logic t-norm is selected and logic connectives are mapped to the corresponding real valued functions. For example, when the \\\\L ukasiewicz t-norm is selected, the implication is mapped to the binary real function $f(x,y) = \\\\min(1, 1 - x + y)$.\\n\\t\\tFor the rule above, the Semantic-Based Regularization loss term is (for the \\\\L ukasiewicz t-norm):\\n\\t\\t\\\\begin{equation*}\\n\\t\\tL^{\\\\text{\\\\L}}(r) = \\\\min \\\\Big(1, 1 - f_\\\\texttt{stress}(x_\\\\texttt{mary}) + f_\\\\texttt{smokes}(x_\\\\texttt{mary}) \\\\Big)\\n\\t\\t\\\\end{equation*}\\n\\t\\tThe aim of Semantic-Based Regularization is to use the regularization term together a with classical loss function for supervised learning to learn the functions associated to the relations (here $f_\\\\texttt{stress}$ and $f_\\\\texttt{smokes}$).\\n\\t\\tIt is worth comparing this method with the Semantic Loss (Example \\\\ref{ex:sl}). Both methods turn a logic formula (either propositional or first-order) to a real valued function that is used as a regularization term. However, because of the different semantics, these two methods have different properties. On the one hand, SL preserves the original logical semantics, by using probabilistic logic. However, due to the probabilistic assumption, the input formula cannot be compiled directly into a differentiable loss but needs to be first translated, i.e. compiled, into an equivalent deterministic and decomposable formula. While this step is necessary for the probabilistic model to be sound, the size of the resulting formula can be exponential in the size of the grounded theory. On the other hand, in SBR, the formula can be compiled directly into a differentiable loss, whose size is linear in the size of the grounded theory. However, in order to do so, the semantics of logic is altered, by turning it into fuzzy logic.\\n\\t\\\\end{boxedexample}\\n\\tFuzzy logic can also be used to relax rules. For example, in LRNN, $\\\\partial$ILP, DiffLog and the approach of , the scores of the proofs are computed using fuzzy logic connectives. \\n\\tThe theory t-norms has identifying parameterized (i.e. weighted) classes of t-norms that are very close to standard neural computation patterns (e.g. ReLU or sigmoidal layers). This creates an interesting, still not fully understood, connection between soft logical inference and inference in neural networks. A large class of methods relaxes logical statements numerically, without explicitly defining a specific semantics.\\n\\tUsually, the atoms are assigned scores in $\\\\mathbb{R}$ computed by a neural scoring function over embeddings.\\n\\tNumerical approximations are then applied either to combine these scores according to logical formulas or to aggregate proofs scores.\\n\\tThe resulting neural architecture is usually differentiable and, thus, trained end-to-end.\\n\\tSome NeSy methods, such as PSL, have used mixed probabilistic and fuzzy semantics. In particular, Deep Logic Models (DLM) extend PSL by adding neurally parameterized factors to the Markov field, while uses fuzzy logic to train posterior regularizers for standard deep networks using knowledge distillation . \\n \\\\rev{The semantics of computational logic has also been explored and extended along other directions that have also been used within AI, for example, \\\\textit{modal} and \\\\textit{temporal} logics . While their analysis is out of the scope of the paper, it is worth mentioning that also such formalisms have been investigation from a neurosymbolic perspective .}\",\n \"id\": \"0c28445b-24b7-4df7-a61f-1e0537492ed8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"3f30c5c1-1c7c-4ae0-b5f0-88220243574e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic - Semantics\"\n ],\n [\n \"subsection\",\n \"Implications for NeSy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for NeSy\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:struct}\\n\\tLearning approaches in StarAI and NeSy are usually distinguished as to whether the structure~ or the parameters of the model are learned~.\\n\\tIn structure learning, the learning task is to discover the logical theory, i.e., a set of logical clauses and their corresponding probabilities or weights that reliably explains the examples. \\n\\tWhat \\\\textit{explaining the examples} exactly means depends on the learning setting.\\n\\tIn discriminative learning, we are interested in learning a theory that explains, or predicts, a specific target relation given background knowledge.\\n\\tIn generative learning, there is no specific target relation; instead, we are interested in a theory that explains the interactions between all relations in a dataset.\\n\\tIn contrast to structure learning, parameter learning starts with a given logical theory and only learns the corresponding probabilities or weights.\\n\\tStructure learning is an inherently NP-complete problem of searching for the right combinatorial structure, whereas parameter learning can be achieved with any curve fitting technique, such as gradient descent or least-squares.\\n\\tWhile parameter learning is, in principle, an easier problem to solve, it comes with a strong dependency on the provided user input. If the provided clauses are of low quality, the resulting model will also be of low quality.\\n\\tStructure learning, on the other hand, is less dependent on the user provided input, but is an inherently more difficult problem.\",\n \"id\": \"3cf78a3e-98ed-45de-848f-3a61063e51b3\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Structure versus Parameter Learning\"\n ]\n ],\n \"subsections\": [\n \"88f64637-a474-4f43-9b9a-5ccb928849a4\",\n \"9af5d6bb-eaff-4088-a25c-009703783367\"\n ],\n \"title\": \"Structure versus Parameter Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Both structure and parameter learning are common in StarAI.\\n\\tStructure learning in StarA is an instance of learning by search~ and is closely connected to program synthesis.\\n\\tThe existing techniques are typically extensions of techniques originating in inductive logic programming (ILP) ~, which learn deterministic logical theories, and probabilistic graphical models (PGMs), which learn Bayesian or Markov networks from data.\\n\\tBeing an instance of learning by search, the central components of a learning framework are a space of valid structures and a search procedure.\\n\\tIn ILP, valid structures are logical theories; for Bayesian networks, valid structures are DAGs capturing their graph structure.\\n\\tThe resulting search space is then traversed with generic search procedures.\\n\\t StarAI structure learning techniques suffer from a combinatorial explosion.\\n\\tThat is especially the case with ILP techniques, in which the search space consists of programs containing several clauses. \\n\\tTherefore, it is necessary to limit the search space to make learning tractable.\\n\\tThe most common way to do this is to impose a \\\\textit{language bias} -- a set of instructions on how to construct the search space, such that it is narrowed down to a subset of the space of all logical theories.\\n\\tThough language bias can make the problem more tractable, it requires special care: too many restrictions might eliminate the target theory, while too few restrictions make the search space too large to traverse.\\n\\tAnother strategy is to leverage the compositionality of logic programs: \\n adding an additional clause to a program increases its coverage and does not affect the prediction of examples covered by the initial program.\\n\\tThat is, we can learn a single clause at a time instead of simultaneously searching over theories containing multiple clauses.\\n\\tLearning clauses and their probabilities is usually treated as a two stage process.\\n\\tILP-based StarAI learning techniques first identify useful (deterministic) clauses, and then learn the corresponding probabilities or weights via parameter learning. Similarly, StarAI methods grounded primarily in PGMs, such as MLNs, search for frequently occurring cliques in data~, lift them into logical clauses, and then learn the weights or probabilities.\\n\\tParameter learning techniques are often also extensions of well known statistical approaches such as least-squares regression~, gradient descent~, and expectation maximisation~.\\n\\t\\\\begin{boxedexample}[label = ex:probfoil]{Structure learning with ProbFoil}\\n\\t\\tAs an illustration of structure learning techniques, we will focus on ProbFoil~.\\n\\t\\tAssume that we are interested in learning the definition of \\\\textit{grandparent} from akinship data.\\n\\t\\tThat is, we are given a set of examples of grandparent relations\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tgrandparent(abe,lisa).\\n\\t\\tgrandparent(abe,bart).\\n\\t\\tgrandparent(jacqueline,lisa).\\n\\t\\tgrandparent(jacqueline,bart).\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tand background knowledge containing the following facts:\\n \\\\begin{lstlisting}\\n\\tfather(homer,lisa). \\n father(homer,bart). \\n father(abe,homer).\\n\\tmother(jacqueline,marge). \\n mother(marge,bart).\\n\\tmother(marge,lisa).\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tProbFoil iteratively searches for a single clause that covers as many examples as possible, until all examples are covered or it adding more clauses does not improve the results. \\n\\t\\tWhile searching for the best clause, it starts from the most general one, \\\\textit{grandparent(X,Y)}, which effectively states that every pair of people forms a grandparent relationship.\\n\\t\\tThen it gradually specialises the clause by adding literals to the body. For instance, extending \\\\textit{grandparent(X,Y)} with a \\\\textit{mother/2} predicate results in the following clauses\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tgrandparent(X,Y) <- mother(X,Y).\\n\\t\\tgrandparent(X,Y) <- mother(X,X).\\n\\t\\tgrandparent(X,Y) <- mother(Y,X).\\n\\t\\tgrandparent(X,Y) <- mother(Y,Y).\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tExtending the initial clause with the \\\\textit{father/2} results in similar clauses.\\n\\t\\tHaving the new candidate clauses, ProbFoil scores each candidate by counting how many positive and negative examples are covered.\\n\\t\\tThese candidate clauses would not cover any examples and ProbFoil continues to refine the candidates by adding another literal to the body.\\n\\t\\tThis would result in clauses of the following form:\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tgrandparent(X,Z) <- mother(X,Y), father(Y,Z).\\n\\t\\tgrandparent(X,Z) <- mother(X,Y), mother(Y,Z).\\n\\t\\tgrandparent(X,Z) <- father(X,Y), mother(Y,Z).\\n\\t\\t...\\n\\t\\\\end{lstlisting}\\n\\t\\tSome of the new candidates will cover only positive examples, such as \\n \\\\begin{lstlisting}\\n grandparent(X,Z) <- mother(X,Y), mother(Y,Z)\\n \\\\end{lstlisting}\\n that covers both examples\\n \\\\begin{lstlisting}\\n grandparent(jacqueline,lisa).\\n grandparent(jacqueline,bart).\\n \\\\end{lstlisting}\\n\\t\\tHaving found one clause, ProbFoil learns the corresponding probability labels and adds the clause to the theory.\\n\\t\\tProbFoil then repeats the same procedure, starting with the most general clause, to cover the remaining examples.\\n\\t\\\\end{boxedexample}\",\n \"id\": \"88f64637-a474-4f43-9b9a-5ccb928849a4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"3cf78a3e-98ed-45de-848f-3a61063e51b3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Structure versus Parameter Learning\"\n ],\n [\n \"subsection\",\n \"Implications for StarAI\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for StarAI\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 5724,\n 8942,\n 2665,\n 5710,\n 5722,\n 7989,\n 3726,\n 5723,\n 3727\n ],\n \"content\": \"While StarAI learning techniques are categorised exclusively as either structure or parameter learning, \\\\nesy{} learning techniques combine both.\\n\\tWe will now discuss four groups of NeSy learning approaches: neurally-guided search, structure learning via parameter learning, program sketching, and implicitly structure learning.\\n\\t\\\\textit{Neurally guided structure search} is the \\\\nesy{} paradigm most similar to structure learning in StarAI.\\n\\tIt addresses one of the major weaknesses of StarAI structure learning methods - uninformed search over valid theories.\\n\\tInstead, neurally guided search relies on a \\\\textit{recognition model}, typically a neural network, to prioritise parts of the symbolic search space so that the target model can be found faster.\\n\\tGenerally speaking, the recognition model predicts the probability of a certain structure, e.g. a predicate or an entire clause, to be a part of the target model.\\n\\tFor instance, Deepcoder~ uses input-output examples to predict the probability of each predicate appearing in the target model.\\n\\tTherefore, Deepcoder turns a systematic search into an informed one by introducing a ranking over predicates in the search space.\\n\\tLikewise, EC$^{2}$~ derives the probability of a program solving the task at hand.\\n\\tSeveral approaches push this direction further and explore the idea of replacing an explicit symbolic model space with an implicit generative model over symbolic models ~. \\n\\tFor instance, in , the authors learn a generative model over grammar rules, conditioned on the examples.\\n\\tStructure learning is then performed by sampling grammar rules from the generative model, according to their probability, and evaluating them symbolically on the provided examples. \\n\\tThese approaches clearly show how symbolic search can be made tractable by introducing various forms of guidance via neural models.\\n\\tThese guidance-based approaches reduce, to a large extent, the most important weakness of symbolic structure learning approaches - the generation of many useless clauses or models.\\n\\tOn the other hand, these approaches often need large amounts of data for training, sometimes millions of examples even though creating data is relatively easy by enumerating random model structures and sampling examples from them~.\\n\\t\\\\begin{boxedexample}[label = ex:ngps]{Neurally-guided structure learning}\\n\\t\\t\\\\rev{To illustrate neurally-guided search, we use the approach of Zhang et al. .\\n StarAI techniques for structure learning typically perform a systematic search, which results in many useless models being tested.\\n Given $N$ atoms, we can construct $N^l$ clauses of length $l$; this is an enormous space that is difficult to search efficiently.} \\\\\\\\\\n \\\\rev{Zhang et al. sidestep the systematic search by introducing a neural network that chooses which programs to explore next.\\n This search space is made of clauses such that an empty clause is at the top and children are extensions of the empty clause with all possible predicates; their children are further extensions with all individual atoms.} \\\\\\\\\\n \\\\rev{The approach follows a top-down search strategy, exploring shorter clauses before longer ones, with a twist: instead of following a predefined order, the approach uses a neural network to decide which child to expand next. \\n The approach can be viewed as a best-first search with a heuristic function implemented by a neural model. To this end, the network (1) encodes all literals in each clause separately, (2) scores all literals, (3) pools the scores of each literal per candidate, and (4) chooses the best candidate based on their scores.\\n Ordering the search space in this way leads to substantial improvements in computation time, typically several orders of magnitude.}\\n\\t\\\\end{boxedexample}\\n\\tAn alternative way to reduce the combinatorial complexity of learning is to learn only a part of the program.\\n\\tThis is known as \\\\textit{program sketching:} a user provides an almost complete target model with certain parts being unspecified (known as \\\\textit{holes}). \\n\\tFor instance, when learning a model in the form of a (logic) program for sorting numbers or strings, the user might leave the comparison operator unspecified and provide the rest of the program.\\n\\tThe learning task is then to fill in the holes.\\n\\tExamples of \\\\nesy{} systems based on sketching are DeepProbLog and $\\\\partial$4, which fill in the holes in a (symbolic) program via neural networks.\\n\\tThe advantage of sketching is that it provides a nice interface for \\\\nesy{} systems, as the holes can be filled either symbolically or neurally.\\n\\tHoles provide a clear interface in terms of inputs and outputs and are agnostic to the specific implementation.\\n\\tThe disadvantage of sketching is that the user still needs to know, at least approximatively, the structure of the program.\\n\\tThe provided structure, the sketch, acts as a strong bias.\\n\\tDeciding which functionality is left as a hole is a non-trivial issue: as the sketch becomes less strict, the search space becomes larger. \\n\\t\\\\textit{Structure learning via parameter learning} (Example \\\\ref{ex:difflog}) is arguably the most prominent learning paradigm in \\\\nesy{}, positioned in between the two StarAI learning paradigms.\\n\\tStructure learning via parameter learning is technically equivalent to parameter learning in that the learning tasks consists of learning the probabilities of a fixed set of clauses.\\n\\tHowever, in contrast to StarAI in which the user carefully selects the informative clauses, the clauses are typically enumerated from user-provided templates of predefined complexity.\\n\\tConstructed in this way, the majority of clauses are noisy and erroneous and are of little use.\\n\\tThey would receive very low, but non-zero, probabilities.\\n\\t Approaches that follow this learning principle include NTPs , $\\\\partial$ILP , DeepProbLog, NeuralLP and DiffLog .\\n\\tThe advantage of structure learning via parameter learning is that it removes the combinatorial search from the learning.\\n\\tHowever, the number of clauses that needs to be considered is still extremely large, which leads to difficult optimisation problems (cf. ). Furthermore, irrelevant clauses are never removed from the model and are thus always considered during inference.\\n\\tThis can lead to spurious interactions even when low probabilities are associated to irrelevant clauses: as the number of irrelevant clauses is extremely large, their cumulative effect can be substantial.\\n\\t\\\\begin{boxedexample}[label = ex:difflog]{Structure learning via parameter learning}\\n\\t\\tAs an illustration of structure learning via parameter learning, we focus on DiffLog .\\n\\t\\tDiffLog expects the candidate clauses to be provided by the user.\\n\\t\\tThe user can either provide the rules she knows are useful or construct them by using a clause template and instantiating it .. \\n\\t\\tGiven a set of positive examples, DiffLog proceeds by constructing \\\\textit{derivation trees} for each example.\\n\\t\\tConsider the problem of learning the connectivity relation over a graph.\\n\\t\\tThe input tuples (background knowledge in StarAI terminology) specify edges in a graph\\n\\t\\t\\\\begin{lstlisting}\\n\\tedge(a,b). edge(b,c). edge(b,d). edge(d,e). edge(c,f).\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tThe examples indicate the connectivity relations among the nodes in the graph (for simplicity, consider only the following two examples)\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\tconnected(a,b). connected(a,c). \\n\\t\\t\\\\end{lstlisting}\\n\\t\\tAlso assume that the candidate clause set contains the following clauses (with $p_1$ and $p_2$ their weights):\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\t$p_1$::connected(X,Y) <- edge(X,Y). \\n\\t\\t$p_2$::connected(X,Y) <- edge(X,Z), connected(Z,Y).\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tDerivation trees are essentially proofs of individual examples that correspond to branches in the SLD-tree .\\n\\t\\tFor instance, the example \\\\textit{connected(a,b)} can be proven using the first clause, whereas the example \\\\textit{connected(a,c)} can be proven by chaining the two clauses ($connected(a,c) \\\\leftarrow edge(a,b), connected(b,c)$ and $connected(b,c) \\\\leftarrow edge(b,c)$).\\n\\t\\tDiffLog uses derivation trees to formulate the learning problem as numerical optimisation over the weights associated with the rules.\\n\\t\\tMore precisely, DiffLog defines the probability of deriving an example as the product of the weights associated to the clauses used in the derivation tree of the corresponding example.\\n\\t\\tFor instance, DiffLog would formulate the learning problem for the two examples as follows\\n\\t\\t$$\\\\min_{p_1, p_2} \\\\ \\\\underbrace{(1 - p_1)}_{\\\\small \\\\tt connected(a,b)} + \\\\underbrace{(1 - p_1\\\\times p_2)}_{\\\\small \\\\tt connected(a,c)}.$$\\n\\t\\\\end{boxedexample}\\n\\tThe last group of approaches learns the structure of a program only \\\\textit{implicitly}.\\n\\tFor instance, Neural Markov Logic Networks (NMLN) , a generalisation of MLNs, extract structural features from relational data.\\n\\tWhereas MLNs define potentials only over cliques defined by the structure (logical formulas) of a model, NMLNs add potentials over \\\\textit{fragments} of data (projected over a subset of constants).\\n\\tNMLNs thus do not necessarily depend on the symbolic structure of the model, be it learned or provided by a user, but can still learn to exploit relational patterns present in data.\\n\\tMoreover, NMLNs can incorporate embeddings of constants.\\n\\tThe benefit of this approach is that it removes combinatorial search from learning and performs learning via more scalable gradient-based methods.\\n\\tHowever, one loses the ability to inspect and interpret the discovered structure.\\n\\tAdditionally, to retain tractability, NMLNs limit the size of fragments which imposes limits on the complexity of the discovered relational structure.\",\n \"id\": \"9af5d6bb-eaff-4088-a25c-009703783367\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"3cf78a3e-98ed-45de-848f-3a61063e51b3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Structure versus Parameter Learning\"\n ],\n [\n \"subsection\",\n \"Implications for NeSy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for NeSy\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:symb_vs_subsymb}\\n\\tIn neurosymbolic artificial intelligence, approaches can be characterized by the way they represents entities and relationships in two classes: symbolic methods, where entities are represented using symbols such as strings and natural numbers, and subsymbolic methods, where entities are represented using numerical or distributed representations.\\n\\tSymbolic representations include constants $(an, bob)$, numbers $(4, -3.5)$, variables $(X, Y)$ and structured terms $f(t_1, ... ,t_n)$ where $f$ is a functor and the $t_i$ are \\\\rev{constants}, variables or structured terms. \\n \\\\rev{Structured terms are a powerful construct that can represent arbitrary structures over entities, such as relations, lists or trees.}\\n\\tsubsymbolic AI systems, such as neural networks, require that entities are represented numerically using vectors, matrices or tensors. Throughout this this section, we will call these subsymbolic representations or subsymbols. \\n\\tsubsymbolic AI systems usually require that these representations have a fixed size and dimensionality. Exceptions require special architectures and are still the subject of active research (e.g. RNNs for list-like inputs or GCNs~ for graph-type inputs).\",\n \"id\": \"bd238fce-3df5-4531-ba35-f4316c49d81b\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Symbolic vs subsymbolic representations\"\n ]\n ],\n \"subsections\": [\n \"9758e29f-afe8-4348-b5ab-e226dfae12d2\",\n \"a1474eee-df30-4fdc-964c-c5b1bb43d21d\"\n ],\n \"title\": \"Symbolic vs subsymbolic representations\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"A powerful and elegant mechanism for reasoning with symbols in logic is {\\\\em unification}. Essentially, it calculates the most general substitution that makes two symbols syntactically equal, if it exists. \\\\rev{This does not allow one to compare two different entities, but allows one to find what two structured terms have in common.}\\n\\tFor instance, the terms $p(a,Y)$ and $p(X,b)$ can be unified using the substitution $\\\\{X=a, Y=b\\\\}$. \\n\\tConversely, due to their numerical nature, calculating the similarity between subsymbols is straightforward. \\n\\tSimilarity metrics such as the radial-basis function or distance metrics such as the L1 and L2 norm can be used. However, it is not clear when to decide that two subsymbolically represented entities are the same.\",\n \"id\": \"9758e29f-afe8-4348-b5ab-e226dfae12d2\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"bd238fce-3df5-4531-ba35-f4316c49d81b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Symbolic vs subsymbolic representations\"\n ],\n [\n \"paragraph\",\n \"Comparing representations\"\n ]\n ],\n \"subsections\": [\n \"a5362428-8a2e-441c-a432-e33fed8ed15b\"\n ],\n \"title\": \"Comparing representations\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 1166,\n 1168\n ],\n \"content\": \"Many systems need to translate back and forth between symbolic and subsymbolic representations. In fact, a lot of research on deep learning is devoted to efficiently representing symbols so that neural networks can properly leverage them.\\n\\tA straightforward example is to translate symbols to a subsymbolic representation that can serve as input for a neural network. Generally, these symbols are replaced by a one-hot encoding or by learned embeddings. \\n \\\\rev{Note, however, that this does not imply that the system can perform symbolic manipulation on this input. Rather, it serves as an index to a set of learned, latent embeddings.}\\n\\tA more interesting example is encoding relations in subsymbolic space. The wide variety of methods~ developed for this purpose indicates that this is far from a solved problem. \\n\\tDifferent encodings have different benefits. For example, TransE~ encodes relations as vector translations from subject to object embeddings. A disadvantage is that symmetric relations are represented by the null vector, and entities in symmetric relations are pushed towards each other.\\n\\tMore complex structures are even harder to represent. For example, there is currently a lot of research in how to utilize graph-structured data in neural networks (cf. \\\\ref{sec:kge_gnn}).\\n\\tTranslating from a subsymbolic representation back to a symbolic one happens, for example, at the end of a neural network classifier. Here, a subsymbolic vector needs to be translated to discrete classes. Generally, this happens through the use of a final layer with a soft-max activation function which then models the confidence scores of these classes as a categorical distribution. However, other options are possible. For example, some methods are only interested in the most likely class, and will use an arg-max instead. Alternatively, a Gumbel-softmax activation can be used as a differentiable approximation of sampling from the categorical distribution.\",\n \"id\": \"a5362428-8a2e-441c-a432-e33fed8ed15b\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"9758e29f-afe8-4348-b5ab-e226dfae12d2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Symbolic vs subsymbolic representations\"\n ],\n [\n \"paragraph\",\n \"Comparing representations\"\n ],\n [\n \"paragraph\",\n \"Translating between representations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Translating between representations\"\n },\n {\n \"cite_extract_rate\": 0.7272727272727271,\n \"cites\": [\n 5708,\n 5707,\n 5711,\n 5726,\n 7304,\n 8943,\n 5725,\n 3727\n ],\n \"content\": \"In StarAI systems, the input, intermediate and output representations are all using the same symbolic representations. \\\\rev{Although there are StarAI systems that can support numerical values, these are still treated as symbols, which is different than a latent, subsymbolic representation.}\\n\\tIn neural systems, the input and intermediate representations are subsymbolic. The output representation can be either symbolic (e.g. classifiers) or subsymbolic (e.g. auto-encoders, GANs).\\n\\tThe most important aspect of neurosymbolic systems is that they combine symbolic and subsymbolic representations.\\n\\tNeSy systems can be categorized by how they do this. \\n We distinguish several approaches.\\n\\tIn the first approach, the inputs are symbolic, but they are translated to subsymbols in a single translation step, after which the intermediate representations used during reasoning are purely subsymbolic. This approach is followed by the majority of NeSy methods. Some examples include Logic Tensor Networks~, Semantic-based Regularization~, Neural Logic Machines~ and TensorLog~.\\n\\t\\\\begin{boxedexample}[label = ex:ltn]{Logic Tensor Networks}\\n\\t\\tLogic tensor networks~ make this translation step explicit. The authors introduce the concept of a \\\\textit{grounding} (not to be confused with the term grounding used in logic). Here, a grounding is a mapping of all symbolic entities onto their subsymbolic counterpart. More formally, the authors define a grounding as a mapping $\\\\mathcal{G}$ where:\\n\\t\\t\\\\begin{itemize}\\n\\t\\t\\t\\\\item $\\\\mathcal{G}(c) \\\\in \\\\mathbb{R}^n$ for every constant symbol $c$\\n\\t\\t\\t\\\\item $\\\\mathcal{G}(f) \\\\in \\\\mathbb{R}^{n.m} \\\\rightarrow \\\\mathbb{R}^n$ for every function $f$ of arity $n$\\n\\t\\t\\t\\\\item $\\\\mathcal{G}(p) \\\\in \\\\mathbb{R}^{n.m} \\\\rightarrow [0,1]$ for every predicate $p$ of arity $n$\\n\\t\\t\\\\end{itemize}\\n\\t\\tThe grounding of a clause is then performed by combining the aforementioned groundings using a t-norm.\\n\\t\\\\end{boxedexample}\\n\\tIn the second approach, intermediate representations are both symbolic and subsymbolic, but not simultaneously. This means that some parts of the reasoning work on the subsymbolic representation, and other parts deal with the symbolic representation, but not at the same time.\\n\\tThis is indicative of NeSy methods that implement an interface between the logic and neural aspect.\\n\\tThis approach is more natural for systems that originate from a logical framework such as DeepProbLog~, NeurASP~), ABL~ and NLog~.\\n\\t\\\\begin{boxedexample}[label = ex:abl]\\n{ABL}\\n\\t\\tIn ABL~ there are three components that function in an alternating fashion. There is a perception model, a consistency checking component and an abductive reasoning component.\\n\\t\\tTake for example the task where there are 3 MNIST images that need to be recognized such that the last is the result of applying an operation on the first two (e.g. $\\\\digit{3}+\\\\digit{5}=\\\\digit{8}$). The structure of the expression is given as background knowledge, but the exact operation (addition) needs to be abduced.\\n\\t\\tFirst, the perception model classifies the images into pseudo-labels, using the most likely prediction (i.e. arg-max). \\n\\t\\tThe abductive reasoning component then tries to abduce a logically consistent hypothesis. \\n\\t\\tFor example, if the digits are correctly classified as $3$, $5$ and $8$, the only logically consistent hypothesis is that the operation is an addition.\\n\\t\\tIf this is not possible, there is an error in the pseudo-labels. \\n\\t\\tA heuristic function is then used to determine which pseudo-labels are wrong.\\n\\t\\tThe reasoning module then searches for logically consistent pseudo-labels. These revised pseudo-labels are then used to retrain the perception model.\\n\\t\\\\end{boxedexample}\\n\\tIn the final approach, intermediate representations are considered simultaneously as symbolic and subsymbolic by the reasoning mechanism.\\n This is implemented in only a few methods, such as the NTP and the CTP. \\n\\t\\\\begin{boxedexample}[label = ex:ntp]{Neural Theorem Prover}\\n\\t\\tIn the Neural Theorem Prover, two entities can be unified if they are similar, and not just if they are identical. As such, the NTP interweaves both symbols and subsymbols during inference. For each symbol $S$, there is a learnable subsymbol $T_S$.\\n\\t\\tSoft-unification happens by applying the normal unification procedure where possible. However, if two symbols $S_1$ and $S_2$ can not be unified, the comparison is assigned a score based on the similarity between $T_{S_1}$ and $T_{S_2}$. The similarity is calculated using a radial basis function $\\\\varphi(||x-y||_2)$.\\n\\t\\tFor example, to unify \\\\texttt{mother(an,bob)} and \\\\texttt{parent(X,bob)}, soft-unification proceeds as follows:\\n\\t\\t\\\\begin{align*}\\n\\t\\t\\\\{\\\\mathtt{mother(an,bob)} &= \\\\mathtt{parent(X,bob)}\\\\} \\\\\\\\\\n\\t\\t&\\\\Downarrow \\\\quad \\\\varphi(\\\\mathtt{mother}, \\\\mathtt{parent}) \\\\\\\\\\n\\t\\t\\\\{\\\\mathtt{an} = \\\\mathtt{X}&, \\\\mathtt{bob} = \\\\mathtt{bob}\\\\} \\\\\\\\\\n\\t\\t&\\\\Downarrow \\\\quad X = an \\\\\\\\\\n\\t\\t\\\\{\\\\mathtt{bob} &= \\\\mathtt{bob}\\\\} \\\\\\\\\\n\\t\\t&\\\\Downarrow\\\\\\\\\\n\\t\\t&\\\\{~\\\\}\\n\\t\\t\\\\end{align*}\\n\\t\\tSoft-unification is not only used to learn which constants and predicates are similar, but can also be used to perform rule learning. By adding new, parameterized rules with unique predicates, soft-unification allows these new predicates to become very similar to other predicates and as such behave as newly introduced rules.\\n\\t\\tFor example, consider the program consisting of the fact \\\\textit{mother(an,bob)} and a single parameterized rule $r1(X,Y) \\\\leftarrow r2(Y,X)$. The Neural Theorem Prover can answer the query \\\\textit{child(bob,an)} as follows: \\n\\t\\t\\\\begin{center}\\n\\\\tikzset{every picture/.style={line width=0.75pt}} \\n\\\\begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1]\\n\\\\draw (240,60) -- (140,100) ;\\n\\\\draw (240,60) -- (340,100) ;\\n\\\\draw (340,140) -- (340,180) ;\\n\\\\draw (201,32) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\\\mathtt{child(bob,an)}$};\\n\\\\draw (299,111) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\\\mathtt{r2(ann,bob)}$};\\n\\\\draw (335,191) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\\\mathtt{T}$ };\\n\\\\draw (131,112) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\\\mathtt{T}$ };\\n\\\\draw (151,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(1)};\\n\\\\draw (299,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(2)};\\n\\\\draw (349,151) node [anchor=north west][inner sep=0.75pt] [align=left] {(3)};\\n\\\\end{tikzpicture}\\n\\t\\t\\\\end{center}\\n\\t\\t\\\\begin{align*}\\n\\t\\t(1) ~& \\\\mathtt{child} = \\\\mathtt{mother} &\\\\varphi(||T_{child}-T_{mother}||_2)\\\\\\\\\\n\\t\\t& \\\\mathtt{bob} = \\\\mathtt{an} &\\\\varphi(||T_{an}-T_{bob}||_2)\\\\\\\\\\n\\t\\t(2)~ & \\\\mathtt{child} = \\\\mathtt{r1} &\\\\varphi(||T_{child}-T_{r1}||_2)\\\\\\\\\\n\\t\\t(3) ~& \\\\mathtt{child} = \\\\mathtt{r2} &\\\\varphi(||T_{r2}-T_{mother}||_2)\\n\\t\\t\\\\end{align*}\\n\\t\\tThe figure above shows the two possible derivations the neural theorem prover can make to infer \\\\textit{child(bob, an)}. One the one hand, it can soft-unify with the fact \\\\textit{mother(an, bob)}, where \\\\textit{mother} unifies with \\\\textit{child} and \\\\textit{an} with \\\\textit{bob}. On the other hand, it can use the parameterized rule which encodes an inverse relation. In that case, \\\\textit{mother} unifies with \\\\textit{r1} and \\\\textit{r2} with \\\\textit{child}. If we optimize the subsymbolic embeddings for the latter, this will be equivalent to learning the rule $mother(X,Y) \\\\leftarrow child(Y,X)$. This example also shows that soft-unification potentially adds a lot of different proofs, which can result in computational problems. This problem was solved in later iterations of the system .\\n\\t\\\\end{boxedexample}\\n\\\\tikzset{every picture/.style={line width=0.75pt}} \\n\\\\begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1]\\n\\\\draw (240,60) -- (140,100) ;\\n\\\\draw (240,60) -- (340,100) ;\\n\\\\draw (340,140) -- (340,180) ;\\n\\\\draw (201,32) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\\\mathtt{child(bob,an)}$};\\n\\\\draw (299,111) node [anchor=north west][inner sep=0.75pt] [align=left] {$\\\\mathtt{r2(ann,bob)}$};\\n\\\\draw (335,191) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\\\mathtt{T}$ };\\n\\\\draw (131,112) node [anchor=north west][inner sep=0.75pt] [align=left] { $\\\\mathtt{T}$ };\\n\\\\draw (151,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(1)};\\n\\\\draw (299,62) node [anchor=north west][inner sep=0.75pt] [align=left] {(2)};\\n\\\\draw (349,151) node [anchor=north west][inner sep=0.75pt] [align=left] {(3)};\\n\\\\end{tikzpicture}\",\n \"id\": \"a1474eee-df30-4fdc-964c-c5b1bb43d21d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"bd238fce-3df5-4531-ba35-f4316c49d81b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Symbolic vs subsymbolic representations\"\n ],\n [\n \"subsection\",\n \"Implications for StarAI and NeSy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implications for StarAI and NeSy\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:paradigms}\\n\\tWhen two or more paradigms are integrated, examining which of the base paradigms are preserved, and to which extent, tells us a lot about the strengths and weaknesses of the resulting paradigm.\\n\\tIt has been argued that when combining different perspectives in one model or framework, such as logic, probabilistic and neural ones, it is desirable to have the original paradigms as a special case.\\n\\t\\\\rev{In this section, we analyze to which extent different models in StarAI and NeSy preserve the three basic paradigms. Intuitively, with preserving we mean to which extent one can exactly replicate the model and inference algorithm of the original paradigm. We will use the capital letters \\\\textit{L, P} and \\\\textit{N} to label systems where the logic, probability and neural paradigms can be recovered in full. We will use lowercase letters (i.e. \\\\textit{l, p} and \\\\textit{n}) when a method only partially recovers these paradigms, i.e. retain some but not all of the features.} The absence of a letter means that the paradigm is not considered by an approach.\",\n \"id\": \"322e754b-6517-46c9-9006-90e863943e5b\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic vs Probability vs Neural\"\n ]\n ],\n \"subsections\": [\n \"64c621c6-ac4f-48ef-b44f-303dd6942485\",\n \"b873315e-f300-4fef-8037-818fbe2be872\"\n ],\n \"title\": \"Logic vs Probability vs Neural\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Traditionally, StarAI focused on the integration of logic and probability.\\n\\t\\\\textbf{\\\\textit{lP:}}\\n\\tThe classical knowledge-based model construction approach uses logic only to generate a probabilistic graphical model. \\n\\tThus the graphical model can be used to define the semantics of the model and also to perform inference.\\n\\tThis can make it harder to understand the effects of applying logical inference rules to the model. \\n\\tFor instance, in MLNs, the addition of the resolvent of two weighted rules makes it hard to predict the effect on the distribution. \\n\\t\\\\textbf{\\\\textit{Lp}:} On the other hand, the opposite holds for probabilistic logic programs (PLPs) and their variants. \\n\\tWhile the effect of a logical operation is clear, it is harder to identify and exploit properties such as conditional or contextual independencies, that are needed for efficient probabilistic inference.\",\n \"id\": \"64c621c6-ac4f-48ef-b44f-303dd6942485\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"322e754b-6517-46c9-9006-90e863943e5b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic vs Probability vs Neural\"\n ],\n [\n \"subsection\",\n \"StarAI: Logic + Probability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"StarAI: Logic + Probability\"\n },\n {\n \"cite_extract_rate\": 0.625,\n \"cites\": [\n 5727,\n 5710,\n 5711,\n 5719,\n 3726\n ],\n \"content\": \"In NeSy, we consider a third paradigm: neural computation.\\n\\tWith neural computation, we refer mainly to the set of models and techniques that allows for exploiting (deep) latent spaces to learn intermediate representations. This includes dealing with perceptual inputs and also dealing directly with embeddings of symbols. \\n\\t\\\\textbf{\\\\textit{lN}}: Many \\\\nesy{} approaches focus on the neural aspect (i.e., they originated as a neural method to which logical components have been added). For example, LTNs and SBRs turn the logic into a regularization function to provide a penalty whenever the logical constraints are violated. At test time the logical loss component is dropped and only the network is used to make predictions. Moreover, by using fuzzy logic, these methods do not integrate the probabilistic paradigm. \\n\\t\\\\textbf{\\\\textit{Ln}}: Another class of \\\\nesy{} methods does retain the focus on logic. \\n\\tThese methods usually expand an existing logical framework into a differentiable version. Examples include LRNNs , TensorLog , DiffLog , $\\\\partial$ILP , $\\\\partial$4 and NTPs .\\n\\tThe key inference concepts are mapped onto an analogous concept that behaves identically for the edge cases but is continuous and differentiable in non-deterministic cases. \\n\\tAs described in the previous sections, many such systems cast logical inference as forward or backward chaining. The focus on logic is clear if one considers that logical inference is performed symbolically to build the network and the semantics is relaxed only in a subsequent stage to learn the parameters. \\\\rev{While the architecture mimics the logical reasoning, it is often far from the deep-stacked architecture of neural networks.}\\n \\\\textbf{\\\\textit{LN}}: \\\\rev{It is worth mentioning a later iteration of LRNN, where the framework has been extended to allow for tensorial weights on atoms and custom aggregation functions . In that framework, it is shown how specifying logic rules can be regarded as specifying the layers of a deep architecture. This provides a nice and complete integration between forward-chaining logical reasoning and neural networks\\n that is able to implement any existing neural architecture.\\n }\\n\\t\\\\textbf{\\\\textit{lPN} and \\\\textit{LpN}} There are two final classes of methods that start from existing StarAI methods, \\\\textit{lP} and \\\\textit{Lp} respectively, and extend them with primitives that can be interfaced with neural networks and allow for differentiable operations.\\n\\tIn the \\\\textit{lPN} class, NeSy methods such as SL, RNMs and NMLNs follow the knowledge-based model construction paradigm. In the \\\\textit{LpN} class, methods such as DeepProbLog and NeurASP extend PLP. \\nThere is usually a trade-off that one must make: \\t\\nsystems in the \\\\textit{lN} or \\\\textit{Ln} classes are usually more scalable but \\\\textit{(i)} do not model a probability distribution and \\\\textit{(ii)} often relax the logic. On the contrary, \\\\textit{LpN} or \\\\textit{lPN} systems preserve the original paradigms but at the cost of more complex inference (e.g. they usually resort to exact probabilistic inference).\\n\\tAn aspect that significantly aids in developing a common framework, and analysing its properties, is the development of an intermediate representation language that can serve as a kind of \\\\emph{assembly language} . \\n\\tOne such idea concerns performing probabilistic inference by mapping it onto a weighted model counting (WMC) problem. \\n\\tThis can then in turn be solved by compiling it into a structure (e.g. an arithmetic circuit) that allows for efficient inference.\\n\\tThis has the added benefit that this structure is differentiable, which facilitates the integration between logic based systems and neural networks. \\n\\t StarAI based systems often use this approach.\",\n \"id\": \"b873315e-f300-4fef-8037-818fbe2be872\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"322e754b-6517-46c9-9006-90e863943e5b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Logic vs Probability vs Neural\"\n ],\n [\n \"subsection\",\n \"NeSy: Logic + Probability + Neural\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"NeSy: Logic + Probability + Neural\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:tasks}\\n\\t\\\\rev{In this Section, we analyze the learning tasks to which the \\\\nesy{} models considered in this paper have been applied.}\",\n \"id\": \"528b4897-8064-4ac1-a40f-dc1e7d30614f\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Tasks\"\n ]\n ],\n \"subsections\": [\n \"d6105bd2-5f38-4f97-a7bd-ae4de168c174\"\n ],\n \"title\": \"Tasks\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 181,\n 7304,\n 1806,\n 3727\n ],\n \"content\": \"\\\\rev{A classical task in NeSy is to use logic as distant supervision for a learning model. Here, input $X$ is paired with label $y$. However, instead of using a single model to map $X$ to $y$, the input $X$ is firstly mapped to a set of intermediate concepts $C$ by a (set of) neural networks. Then, these concepts are used to compute $y$ in a symbolic way. Logic programs are usually exploited to map the concepts $C$, represented as logical atoms, to the label $y$, which represents the logical query. Therefore, the neural networks are not directly supervised (on $C$) but they are only distantly supervised through the label $y$ and the knowledge contained in the logic program. The intuition is that when the label $y$ is only weakly linked to the input, it is more convenient to break the task in several easier subtasks and then compose them using background knowledge in the form of a logic program. Notice that the logic program is fundamental for the inference. Without the program, the networks will not be able to solve their subtasks, as there is no direct supervision. Moreover, by splitting the task into subtasks, the inference done by the composite system (neural + logic) is far more explainable than a corresponding end-to-end neural network. A classical example is the MNIST addition , shown in Example \\\\ref{ex:mnist_addition}. Distant supervision tasks are very common in prototypical systems such as DeepProbLog, DeepStochLog, NLog, NeurASP, SATNet . A downside of such tasks is that, to enable learning of untrained neural subtasks, the logic has to consider all possible combinations of concepts that are compatible with the label $y$, even though only few (or one) are correct. The challenge is to balance the exploration of multiple combinations with a greedy strategy for scaling to larger problems . Other problems falling in this category are scene parsing, image segmentation and semantic image interpretation }\\n \\\\begin{boxedexample}[label = ex:mnist_addition]{MNIST Addition} \\n Given the classical MNIST dataset, $\\\\mathcal{D} = \\\\{(x_i,y_i)\\\\}$, with $x_i$ an MNIST image, and $y_i$ its numeric label, the MNIST addition dataset is built by mapping pairs of images to the label representing their sum. In particular, $\\\\mathcal{D}_{\\\\text{add}} = \\\\{(x_i,x_j,z_{ij}) : z_{ij} = y_i + y_j \\\\land (x_i,y_i),(x_j,y_j) \\\\in \\\\mathcal{D}\\\\}$. The idea is to learn to classify the digits without direct supervision on their labels, but only using distant supervision about sums of such images. \\n The task is often also coupled to background knowledge of what addition is, e.g. in Prolog syntax:\\n \\\\begin{lstlisting}\\naddition(X1, X2, Z) <- digit(X1,Y1), digit(X2,Y2), Z is Y1 + Y2. \\n \\\\end{lstlisting}\\n Such knowledge is used to reason about the (most-likely) pairs \\\\texttt{Y1,Y2} that sum to the provided label \\\\texttt{Z}. Logic is then used to link the actual outputs of the learning model \\\\texttt{Y1,Y2} to the distant supervision \\\\texttt{Z}.\\n \\\\end{boxedexample}\",\n \"id\": \"d6105bd2-5f38-4f97-a7bd-ae4de168c174\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"528b4897-8064-4ac1-a40f-dc1e7d30614f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Tasks\"\n ],\n [\n \"paragraph\",\n \"Distant Supervision\"\n ]\n ],\n \"subsections\": [\n \"d9eb4e04-ce78-4d16-9f72-94154473deea\"\n ],\n \"title\": \"Distant Supervision\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 5714,\n 3741,\n 5720,\n 5709\n ],\n \"content\": \"\\\\rev{A related class of tasks is semi-supervised classification } with knowledge. Here, the starting point is a standard classification task, where a set of inputs $X$ is mapped by a neural model to a set of labels $C$. However, we are also provided with some additional knowledge $y$ related to the labels $C$ of the inputs. This knowledge is often expressed in terms of logical rules and programs. The setting is very similar to distant supervision, where we have three levels: inputs $X$, concepts $C$ and additional labels $y$. However, in this case, we have also access to supervision for some (usually few) concepts $C$. Although this task could be tackled in a purely supervised way by discarding the information contained in $y$, \\\\nesy{} approaches can improve the predictions of several input patterns using the external knowledge. When the external knowledge is relating concepts $C$ of multiple patterns, the task is called collective classification , as one can improve the accuracy on multiple patterns by collectively predicting their classes. A classical example in this setting is document classification in citation networks, cf. Example \\\\ref{ex:citation}. By treating the information contained in $y$ as extra knowledge, these tasks are often tackled using regularization based systems, like SBR, DLM, RNM or Semantic Loss. However, logic programs can also be used to simulate a label-passing scheme along the citation network, as done in DeepStochLog . A characteristic of this class of tasks is that the additional information $y$ is often very noisy (e.g. the manifold rule in the citation network is not always valid). While this task is closely related to distant supervision, there is an important difference: in semi-supervised classification, the additional knowledge $y$ is meant to provide an additional signal, which, however, would not suffice in the absence of direct supervision on the concepts $C$. \\n \\\\begin{boxedexample}[label = ex:citation]{Document classification in citation networks}\\n In document classification in citation networks, we are provided with both labelled and unlabelled scientific papers. A label is often the domain area of the paper (e.g. Machine Learning, Artificial Intelligence, Databases, etc.). \\n However, a network of citations between papers is also provided, linking papers between domains.\\n The idea of the document classification task is that in many domains, a paper cited by other papers with a certain label is likely to belong to the same domain. When classifying a document, one has to balance the signal coming from the features of the document (i.e. words) and that coming from neighbors in the citation network to provide a \\\\textit{collective} prediction.\\n In \\\\nesy{} systems, this is usually done by coupling the subsymbolic model with a rule of the following type:\\n \\\\begin{lstlisting}\\n w:: domain(X,Y) <- cite(X,X1), domain(X1,Y). \\n \\\\end{lstlisting}\\n The rules get a different weight according to the domain to account for the differences between them. \\n \\\\end{boxedexample}\",\n \"id\": \"d9eb4e04-ce78-4d16-9f72-94154473deea\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"d6105bd2-5f38-4f97-a7bd-ae4de168c174\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Tasks\"\n ],\n [\n \"paragraph\",\n \"Distant Supervision\"\n ],\n [\n \"paragraph\",\n \"Semi-supervised Classification\"\n ]\n ],\n \"subsections\": [\n \"c47d6a00-5a1b-40c6-9278-09edc1c79d9b\",\n \"d7c376c1-f6e7-41a6-9b6e-28a3f939810e\",\n \"f4d6de92-316d-4daa-a0ba-2c746822d5de\"\n ],\n \"title\": \"Semi-supervised Classification\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8942,\n 5710,\n 5711,\n 5720\n ],\n \"content\": \"\\\\rev{Another common task in \\\\nesy{} is knowledge graph completion (KGC) or link prediction. A knowledge graph (KG) is a pair of $(E,R)$, where $N$ is the set of entities and $R$ the set of edges. In a KG, an edge is a triple $(e_1, r, e_2)$, where $e_1$ and $e_2$ are the head and tail of the edge and $r$ is the relation between them. In a KGC task, the goal is to predict missing edges in the input graph. Link prediction has been one of the key tasks in StarAI , and more recently also in NeSy as NeSy allows to merge symbolic reasoning (from StarAI) with the recent geometric deep learning approaches based on Knowledge Graph Embeddings (KGE) and Graph Neural Networks . \\\\nesy{} systems focusing on this task include NTPs , NMLN , DLM , DiffLog , TensorLog .}\",\n \"id\": \"c47d6a00-5a1b-40c6-9278-09edc1c79d9b\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"d9eb4e04-ce78-4d16-9f72-94154473deea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Tasks\"\n ],\n [\n \"paragraph\",\n \"Distant Supervision\"\n ],\n [\n \"paragraph\",\n \"Semi-supervised Classification\"\n ],\n [\n \"paragraph\",\n \"Knowledge Graph Completion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Graph Completion\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 8942,\n 7988\n ],\n \"content\": \"Most previously mentioned tasks can be described as classification\\\\footnote{Even though, many of them use a generative model to tackle the classification task, instead of a conditional one.}. \\\\nesy{} has recently also focused on tasks concerned with modeling the input data distribution as accurately as possible. The goal is then to sample new patterns from the learned distribution. The idea behind \\\\nesy{} generative approaches is that one can learn important features from data using deep generative models (e.g. variational auto-encoders or Markov Chain Monte Carlo methods). Combining symbolic features with logic reasoning can be used to control, stratify and simplify the inference. The generative modeling can either refer to the relational structure, e.g. molecule generation in NMLNs , or to the subsymbolic space, e.g. image generation in VAEL or .\",\n \"id\": \"d7c376c1-f6e7-41a6-9b6e-28a3f939810e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"d9eb4e04-ce78-4d16-9f72-94154473deea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Tasks\"\n ],\n [\n \"paragraph\",\n \"Distant Supervision\"\n ],\n [\n \"paragraph\",\n \"Semi-supervised Classification\"\n ],\n [\n \"paragraph\",\n \"Generative Tasks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Generative Tasks\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 2665,\n 5710,\n 5728,\n 3726,\n 3727\n ],\n \"content\": \"\\\\rev{Rather than exploiting symbolic knowledge predictive tasks, one can also induce \\n \\\\textit{symbolic knowledge}. In all previous tasks, symbolic knowledge is provided by the user as part of the input. However, as explored in Section \\\\ref{sec:struct}, we can still apply several neurosymbolic techniques by learning the symbolic knowledge when this is not the case. The unknown symbolic knowledge is then the actual target to be learned. A classical example is \\\\textit{program synthesis}, where the goal is to learn the program from positive and negative examples of the desired input-output behaviour. Ideally, all positive pairs and none of the negatives should be covered. Many systems learn logic programs, i.e. NTPs , $\\\\partial$ILP , DeepProbLog, NeuralLP, DiffLog, DeepCoder.\\n Sometimes, the input-output pairs are not part of the training dataset, but are actually generated by a black-box neural model. The induced programs then explain the behaviour of the model, which relates NeSy to the domain of \\\\textit{explainability} .}\",\n \"id\": \"f4d6de92-316d-4daa-a0ba-2c746822d5de\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"d9eb4e04-ce78-4d16-9f72-94154473deea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Tasks\"\n ],\n [\n \"paragraph\",\n \"Distant Supervision\"\n ],\n [\n \"paragraph\",\n \"Semi-supervised Classification\"\n ],\n [\n \"paragraph\",\n \"Knowledge Induction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Induction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To conclude, we list some interesting challenges for \\\\nesy{}.\",\n \"id\": \"35970a41-0514-46a0-8d9d-63d0769ff87a\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ]\n ],\n \"subsections\": [\n \"a1d98b58-9672-4eb5-80a4-d72f0adf272e\"\n ],\n \"title\": \"Open Challenges\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The statistical relational AI and probabilistic graphical model communities have devoted a lot of attention to the semantics of its models. This has resulted in several clear choices (such as directed vs. undirected, trace-based vs. possible world ), with corresponding strengths and weaknesses that clarify the relationships between the different models. Workshops have been held on this topic\\\\footnote{For instance, \\\\url{https://pps2018.luddy.indiana.edu/}}. Furthermore, some researchers have investigated how to transform one type of model into another . At the same time, weighted model counting has emerged as a common assembly language for inference. The situation in neurosymbolic computation today is very much like that of the early days in statistical relational learning, in which there were many competing formalisms, sometimes characterized as the statistical relational learning alphabet soup. It would be great to get more insight into the semantics of neurosymbolic approaches and their relationships. This survey hopes to contribute towards this goal.\",\n \"id\": \"a1d98b58-9672-4eb5-80a4-d72f0adf272e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"35970a41-0514-46a0-8d9d-63d0769ff87a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ],\n [\n \"paragraph\",\n \"Semantics\"\n ]\n ],\n \"subsections\": [\n \"6f37ac7d-158b-4d60-89b1-c3be4d76632d\"\n ],\n \"title\": \"Semantics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Although relatively few methods explore the integration of logical and neural methods from a probabilistic perspective, we believe that a probabilistic approach is a very natural way to integrate the two, since it has been shown how one can recover the single methods as special cases. However, many open questions remain. Probabilistic inference is computationally expensive, usually requiring approximations. It would be interesting to determine exactly how probabilistic approximate inference compares with other approximations based on relaxations of the logic, like fuzzy logic.\",\n \"id\": \"6f37ac7d-158b-4d60-89b1-c3be4d76632d\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"a1d98b58-9672-4eb5-80a4-d72f0adf272e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ],\n [\n \"paragraph\",\n \"Semantics\"\n ],\n [\n \"paragraph\",\n \"Probabilistic reasoning\"\n ]\n ],\n \"subsections\": [\n \"2225ccc8-0707-44bd-8f81-0bdebfb52224\",\n \"422c662c-f7da-4490-a64b-1d55a3d7423a\",\n \"588fd0c9-1da5-47a9-91a7-ef4be62e8832\",\n \"9539d3a1-1ce6-40b0-8897-7d2fd89a6849\",\n \"986b4d82-d655-4355-b6bd-d14bee40b72f\"\n ],\n \"title\": \"Probabilistic reasoning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The selection of the t-norm fuzzy logic and the corresponding translation of the connectives is very heterogeneous in the literature. It is often unclear which properties of Boolean logic a model is preserving, while there is a tendency to consider fuzzy logic as a continuous surrogate of Boolean logic without considering implications for the semantics. There is a clear need for further studies in this field. On the one hand, one may want to define new models which are natively fuzzy, thus not requiring a translation from Boolean logic. On the other hand, an interesting research direction concerns the characterisation of what are appropriate fuzzy approximations of Boolean logic relative to a set of properties that one wants to preserve (see Section \\\\ref{sec:semantics}).\",\n \"id\": \"2225ccc8-0707-44bd-8f81-0bdebfb52224\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"6f37ac7d-158b-4d60-89b1-c3be4d76632d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ],\n [\n \"paragraph\",\n \"Semantics\"\n ],\n [\n \"paragraph\",\n \"Probabilistic reasoning\"\n ],\n [\n \"paragraph\",\n \"Fuzzy semantics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Fuzzy semantics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"While significant progress has been made on learning the structure of purely relational models (without probabilities), learning StarAI models remains a major challenge due to the complexity of inference and the combinatorial nature of the problem.\\n\\tIncorporating neural aspects complicates the problem even more.\\n\\t\\\\nesy{} methods have certainly shown potential for addressing this problem (Section \\\\ref{sec:struct}), but the existing methods are still limited and mostly domain-specific which impedes their wide application.\\n\\tFor instance, the current systems that support structure learning require user effort to specify the clause templates or write a sketch of a model.\",\n \"id\": \"422c662c-f7da-4490-a64b-1d55a3d7423a\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"6f37ac7d-158b-4d60-89b1-c3be4d76632d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ],\n [\n \"paragraph\",\n \"Semantics\"\n ],\n [\n \"paragraph\",\n \"Probabilistic reasoning\"\n ],\n [\n \"paragraph\",\n \"Structure learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Structure learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Scalable inference is a major challenge for StarAI and therefore also for \\\\nesy{} approaches with an explicit logical or probabilistic reasoning component.\\n\\tInvestigating to what extent neural methods can help with this challenge by means of lifted (exploiting symmetries in models) or approximate inference, as well as reasoning from intermediate representations , are promising future research directions.\",\n \"id\": \"588fd0c9-1da5-47a9-91a7-ef4be62e8832\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"6f37ac7d-158b-4d60-89b1-c3be4d76632d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ],\n [\n \"paragraph\",\n \"Semantics\"\n ],\n [\n \"paragraph\",\n \"Probabilistic reasoning\"\n ],\n [\n \"paragraph\",\n \"Scaling inference\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Scaling inference\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"A major advantage of StarAI methods, as compared to neural ones, is their data efficiency -- StarAI methods can efficiently learn from small amounts of data, whereas neural methods are data hungry.\\n\\tOn the other hand, StarAI methods do not scale to big data sets, while neural methods can easily handle them.\\n\\tWe believe that understanding how these methods can help each other to overcome their complementary weaknesses, is a promising research direction.\",\n \"id\": \"9539d3a1-1ce6-40b0-8897-7d2fd89a6849\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"6f37ac7d-158b-4d60-89b1-c3be4d76632d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ],\n [\n \"paragraph\",\n \"Semantics\"\n ],\n [\n \"paragraph\",\n \"Probabilistic reasoning\"\n ],\n [\n \"paragraph\",\n \"Data efficiency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data efficiency\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 5729,\n 5730\n ],\n \"content\": \"The effectiveness of deep learning comes from the ability to change the representation of the data so that the target task becomes easier to solve.\\n\\tThe ability to change the representation also at the symbolic level would significantly increase the capabilities of \\\\nesy{} systems.\\n\\tThis is a major open challenge for which neurally inspired methods could help achieve progress .\\n\\t\\\\section*{Acknowledgements}\\n\\tThis work has received funding from the Research Foundation-Flanders (FWO)\\n\\t(G. Marra: 1239422N, S. Dumančić: 12ZE520N, R. Manhaeve: 1S61718N). \\n\\tLuc De Raedt has received funding from the Flemish Government (AI Research Program), from the FWO, from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 694980 SYNTH: Synthesising Inductive Data Models) and the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. This work was also supported by TAILOR, a project funded by EU Horizon 2020 research and innovation programme under GA No 952215.\\n\\t\\\\bibliographystyle{plain}\\n\\t\\\\bibliography{main}\\n\\t\\\\appendix\",\n \"id\": \"986b4d82-d655-4355-b6bd-d14bee40b72f\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"6f37ac7d-158b-4d60-89b1-c3be4d76632d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Open Challenges\"\n ],\n [\n \"paragraph\",\n \"Semantics\"\n ],\n [\n \"paragraph\",\n \"Probabilistic reasoning\"\n ],\n [\n \"paragraph\",\n \"Symbolic representation learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Symbolic representation learning\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7233\n ],\n \"content\": \"\\\\label{sec:kge_gnn}\\n\\tIn this appendix, we introduce two approaches, namely Knowledge Graph Embeddings and Graph Neural Networks, which are commonly used in relational tasks in the deep learning community. We analyze them in the spirit of the seven dimensions introduced in the paper. In fact, it turns out that they share many features with neurosymbolic systems. In this way, we would like to suggest that NeSy can also be found at the intersection of statistical relational and geometric deep learning approaches.\\n\\tOne of the most popular relational representations is that of a knowledge graph (KG). A KG is\\n\\ta multi-relational graph composed of entities (i.e. nodes) and relations (i.e. edges). It is common to represent an edge as a triple of the form: \\\\textit{(head entity, relation, tail entity)}, e.g. ($homer$, $fatherOf$, $bart$).\\n\\tStarAI has been extensively used to solve many tasks on KGs: prediction of missing relationships (i.e. knowledge graph completion), prediction of properties of entities, or clustering entities based on their connectivity patterns. StarAI is particularly well suited to reasoning with knowledge graphs since it models explicitly the probabilistic dependencies among different relationships. \\n\\tHowever, in order to scale to larger knowledge graphs, the probabilistic dependencies of StarAI models have been relaxed to give rise to a new class of scalable models based on latent features, which are particularly interesting from a neurosymbolic viewpoint. The key intuition behind relational latent feature\\n\\tmodels is that the relationships between entities can be more efficiently predicted by modeling simpler interactions in a latent feature space. Knowledge Graph Embeddings and Graph Neural Networks represent two of the ways to encode such latent representations.\",\n \"id\": \"0c29e7f0-ea36-44aa-a1f0-0b0c69b42bfb\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI\"\n ]\n ],\n \"subsections\": [\n \"525771a4-9400-42b9-9e72-981457921a8b\",\n \"128f81ba-b844-472a-86b3-8ca1662bba8a\"\n ],\n \"title\": \"Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 5732,\n 3741,\n 3672,\n 5731\n ],\n \"content\": \"Knowledge Graph Embedding (KGE) models assume that triples (i.e. relations) are conditionally independent given a set of global latent variables, called embeddings, for entities and relations. Therefore, the existence of a triple can be predicted by a scoring function $f(e_h,e_r,e_t)$, where $e_i$ is the embedding of the corresponding object.\\n\\tKGE models mainly differ in terms of the scoring function and the embedding space they use. Translation or distance-based models use a scoring function measuring to what extent the tail of a triple can be obtained by a relation-specific translation of the head, i.e. $f(e_h,e_r,e_t) = || e_h + e_r - e_t||$. Semantic matching methods instead exploit similarity-based scoring\\n\\tfunctions, such as $f(e_h,W_r,e_t) = || e_hW_re_t^\\\\top||$. It is interesting to note that standard multi-layer perceptrons are often used to learn this similarity measure, i.e. $f(e_h,W_r,e_t) = nn([e_h, e_t]; W_r)$. This is similar to neural interfaces typical of $lPN$ and $LpN$ models of Section \\\\ref{sec:paradigms}, cf. the neural predicates of DeepProbLog in Example~\\\\ref{ex:deepproblog}.\\n\\tKGE learn from ground relations only. They learn the embeddings of entities and relations as to maximize the score for a set of known \\\\textit{True} triples. However, some methods incorporate higher level information, like first-order logical clauses or logical queries , bridging KGE and multiple NeSy systems (like SL or SBR). \\n\\tIt is interesting to analyze KGE methods in terms of the dimensions we described in our paper. KGE methods mostly work as model-based, undirected methods. They constrain the embeddings to be coherent with the logical facts and rules, which are no longer used after learning. They are heavily based on subsymbolic representations. These models learn the correct parameters (i.e. embeddings) for the task at hand. Even if no explicit semantics needs to be given to the scoring function $f(e_h,W_r,e_t)$, a fuzzy logic interpretation is often used when injecting logical rules \",\n \"id\": \"525771a4-9400-42b9-9e72-981457921a8b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"0c29e7f0-ea36-44aa-a1f0-0b0c69b42bfb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI\"\n ],\n [\n \"subsection\",\n \"Knowledge Graph Embeddings\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Graph Embeddings\"\n },\n {\n \"cite_extract_rate\": 0.7333333333333331,\n \"cites\": [\n 259,\n 5733,\n 281,\n 208,\n 7332,\n 8475,\n 5734,\n 5735,\n 7045,\n 216,\n 1059\n ],\n \"content\": \"Graph Neural Networks (GNN) () are deep learning models for dealing with graphs as input. Inference in these models can be cast as \\\\textit{message passing} : at each inference step (i.e. GNN layer), a node sends a message to all its outgoing neighbors and updates its state by aggregating all the messages coming from its ingoing neighbors. Messages are computed by standard neural networks. Each inference step computes a transformation of the representations of the nodes. In the last layer, these representations are used to classify the nodes or are aggregated to classify the entire graph. \\n\\tThere are many connections between GNNs and neurosymbolic computation since both of them apply neural methods to relational data: graphs for GNNs, and logic representations for NeSy. While many works try to close the gap between these two representations , the application of GNN techniques to logic-based graph structures is still very limited. The most related line of work is about using GNNs on knowledge graphs . The underlying idea is to differentiate the messages exchanged by two nodes if they are related by different kinds of edges. For example, given two nodes \\\\texttt{homer} and \\\\texttt{bart}, the message corresponding to the edge \\\\texttt{fatherOf(homer,bart)} will be different from the one corresponding to \\\\texttt{sameFamily(homer,bart)}. However, these models were intended as classifiers of known relational structures (nodes and edges) and not to reason about the knowledge graph itself (e.g. determining whether an edge exists between two nodes). To perform relational reasoning, GNN-based models rely on techniques from the KGE community on top of the representations extracted by the GNN. This often takes the shape of an auto-encoding scheme: a GNN encodes an input graph in a latent representation and a KGE-based factorization technique is used to reconstruct the whole graph . \\n\\tAn important characteristic of GNNs is that they rely exclusively on neural computation to perform inference (i.e. to compute messages) and there is no clear direction on how to inject external knowledge about inference, e.g. as logical rules. This contrasts with NeSy, where this is one of the main goals. \\n\\tThere are also some interesting connections between GNNs and StarAI models. In , GNNs based on knowledge-graphs are not used as a modeling choice but rather to approximate inference in Markov Logic Networks, which is somewhat similar to regularization based methods (see Section \\\\ref{sec:proof_vs_model}). Similarly, in GNNs are used to encode logical formulae expressed as graphs to approximate a weighted model counting problem.\\n\\tFinally, it is interesting to analyse GNNs in the spirit of some of the dimensions of NeSy. GNNs act as directed models with a proof-based inference scheme: they perform a series of inference steps to compute the final answer. In the original version of GNNs , the node states are updated until a fixed point is reached, which resembles forward-chaining in logic programming. The representation of nodes belongs to a subsymbolic numerical space. Finally, GNNs can be considered as implicit structure learners: inference rules are learned through the learning of the neural message passing functions. \\n\\tGraph Neural Networks have recently received a lot of attention from many different communities, thanks to the representation power of neural networks and the capability of learning in complex relational settings. It is no surprise that people have started to study the expressivity of this class of models. One of the most interesting analyses from a neurosymbolic viewpoint is measuring the expressivity of GNNs in terms of variable counting logics. Recently, showed that GNNs are as expressive as 2-variable counting logic\\n\\t$C_2$. This fragment of first order logic admits formulas with at most two variables extended with counting quantifiers. The expressivity of this fragment is limited compared to many neurosymbolic models, especially those based on logic. However, GNNs learn the logical structure of the problem implicitly as part of the message passing learning scheme and they rely neither on expert-provided knowledge nor on heavy combinatorial search strategies to structure learning (see Section \\\\ref{sec:struct}). An open and challenging question that unites the GNN and NeSy communities is how to bring the expressivity to higher-order fragments , like in NeSy and StarAI, while keeping both the learning and the inference tractable, like in GNNs.\",\n \"id\": \"128f81ba-b844-472a-86b3-8ca1662bba8a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"0c29e7f0-ea36-44aa-a1f0-0b0c69b42bfb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Knowledge graphs embeddings and Graph Neural Networks for neurosymbolic AI\"\n ],\n [\n \"subsection\",\n \"Graph Neural Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Graph Neural Networks\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 5736,\n 5737,\n 5707,\n 5706\n ],\n \"content\": \"Fuzzy logic, as many-valued extension over Boolean logic, has a very long tradition . However, the use of fuzzy logic in StarAI and NeSy is not dictated by the need of dealing with vagueness, but by the advantageous computational properties of t-norms. Indeed, a common use case is to have an initial theory defined in Boolean logic which is \\\\textit{fuzzyfied}. Inference is then carried out with the fuzzyfied theory and the answers are eventually discretized back to Boolean values (usually using a threshold at 0.5). \\n\\tThe reason for this approach is that one would like to exploit the differentiability of t-norms to address logical inference of FOL theories in a more scalable way than standard combinatorial optimization algorithms (e.g. SAT solvers). This is particularly important in undirected and regularization-based methods (such as PSL and LTN ). In fact, it has been shown that there are fragments of fuzzy logic that can even provide convex inference problems. Example \\\\ref{ex:bool_vs_fuzzy}, however, shows that naively approaching logical inference through a fuzzy relaxation and gradient-based optimization can introduce unexpected behaviours.\\n\\t\\\\begin{boxedexample}[label = ex:bool_vs_fuzzy]{Fuzzyfication and soft-Satisfability}\\n\\t\\tLet us consider a disjunction, like $A \\\\vee B \\\\vee C$. In Boolean logic, if we state that the disjunction is \\\\textit{satisfied} (i.e. \\\\textit{True}), then we expect at least one among the three variables to be \\\\textit{True}. \\n\\t\\tSuppose we want to find a truth assignment for all the variables that satisfies the disjunction above. The approach of the majority of NeSy fuzzy approaches is the following. First, the rule is relaxed into a fuzzy real function. For example, using the \\\\L ukasiewicz t-norm, $F_\\\\oplus(A,B,C)= min(1, A+B+C)$. Secondly, a gradient-based algorithm (e.g. backpropagation with Adam ) is used to maximize the value of the formula with respect to the \\\\textit{fuzzy} truth degree of the three variables. Finally, the obtained \\\\textit{fuzzy} solution $A^\\\\star, B^\\\\star, C^\\\\star$ is translated back into a Boolean assignment using a $0.5$ threshold.\\n\\t\\tLet us consider a possible optimal fuzzy solution, like $(A^\\\\star, B^\\\\star, C^\\\\star) = (0.34, 0.34, 0.34)$ and its discretized version $(A^\\\\star, B^\\\\star, C^\\\\star) = (False, False,$ $False)$, using a threshold at $0.5$. The discretized solution does not satisfy the initial Boolean formula, even though it is a global optimum in the fuzzyfied problem. \\n\\t\\\\end{boxedexample}\\n\\tSimilarly, shows that, while it is very common to reason about universally quantified formulae in the form of $\\\\forall x: A(x) \\\\to B(x)$, like `all humans are mortal', using gradients and fuzzy logic to make inference can be extremely counterintuitive, especially with specific t-norms such as the product t-norm. It is unclear whether there exists a generally accepted subset of properties of Boolean logic that one wants to preserve and whether one can define a t-norm that guarantees such properties.\",\n \"id\": \"df605b7c-7a16-4a8a-9a34-34eff8f50d0f\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"31670094-45b4-4854-914d-94aaf932ce65\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Fuzzy logic, fuzzyfication and soft-Satisfability\"\n ]\n ],\n \"subsections\": [\n \"c6063ac6-99de-4318-925a-eab4859602f1\"\n ],\n \"title\": \"Fuzzy logic, fuzzyfication and soft-Satisfability\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Another common reason for using fuzzy logic is to exploit a differentiable semantics. Then, gradient-based methods can be used to train the parameters of a weighted logical theory such as in LRNNs . This contrasts with gradient-based training of the parameters of probabilistic logics based on the distribution semantics. Possible worlds in probabilistic logic are defined as possible assignments of truth values to all the ground atoms of a logical theory. The assignments of truth values specify the \\\\textit{semantics} of the logic. On the contrary, fuzzy logic assigns continuous truth degrees to formulas or proofs, which are \\\\textit{syntactic} structures. As a consequence, while the probability of an atom will always be equal to the sum of the probabilities of the worlds in which it is \\\\textit{True} (cf. Equation \\\\ref{eq:wmc}), the fuzzy degree of an atom may vary depending on how that atom has been proven or defined, as shown in Example \\\\ref{ex:prob_vs_fuzzy}.\\n\\t\\\\begin{boxedexample}[label = ex:prob_vs_fuzzy]{Distribution semantics vs fuzzy logic semantics}\\n\\t\\tConsider the following annotated program (adapted from ).\\n\\t\\t\\\\begin{lstlisting}\\n\\t\\t0.3::b.\\n\\t\\ta1 <- b.\\n\\t\\ta2 <- b,b.\\n\\t\\t\\\\end{lstlisting}\\n\\t\\tHere, $0.3$ is the label of $\\\\mathtt{b}$ without any particular semantics yet. \\n\\t\\tGiven the idempotency of the Boolean conjunction, we would expect the scores of both $\\\\mathtt{a1}$ and $\\\\mathtt{a2}$ to be identical since both $\\\\mathtt{b}$ and $\\\\mathtt{b} \\\\wedge \\\\mathtt{b}$ are \\\\textit{True} when $\\\\mathtt{b}$ is \\\\textit{True}. In a probabilistic approach, the score is interpreted as a probability, i.e. $p(\\\\mathtt{b}) = 0.3$. The probability of any atom is the sum of the probabilities of all the worlds where that atom is \\\\textit{True}. It is easy to see that, for both $\\\\mathtt{a1}$ and $\\\\mathtt{a2}$, these are the worlds where $\\\\mathtt{b}$ is \\\\textit{True}, thus $p(\\\\mathtt{a1}) = p(\\\\mathtt{a2}) = p(\\\\mathtt{b}) = 0.3$. On the other hand, in the fuzzy setting, the score of $\\\\mathtt{b}$ is interpreted as its truth degree, i.e. $t(\\\\mathtt{b}) = 0.3$. Let us consider the product t-norm, where $t(x \\\\wedge y) = t(x)t(y)$, then $t(\\\\mathtt{a1}) = t(\\\\mathtt{b}) = 0.3$ while $t(\\\\mathtt{a2}) = t(\\\\mathtt{b})t(\\\\mathtt{b}) = 0.09$. While this issue could be solved by choosing a different t-norm (e.g. the minimum t-norm), similar issues arise in different definitions.\\n\\t\\\\end{boxedexample}\\n\\tThe differences between the two semantics are not due to the probabilistic or the fuzzy semantics, but more to the distinction between semantics based on possible worlds and semantics based on proofs or derivations. In fact, a similar behaviour is observed in Stochastic Logic Programs under the name of \\\\textit{memoization}.\\n\\\\end{document}\",\n \"id\": \"c6063ac6-99de-4318-925a-eab4859602f1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"df605b7c-7a16-4a8a-9a34-34eff8f50d0f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Statistical Relational to Neurosymbolic \\\\\\\\ Artificial Intelligence: a Survey.\"\n ],\n [\n \"section\",\n \"Fuzzy logic, fuzzyfication and soft-Satisfability\"\n ],\n [\n \"subsection\",\n \"Distribution semantics and fuzzy logic semantics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Distribution semantics and fuzzy logic semantics\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":69,"string":"69"},"all_cites":{"kind":"list like","value":[7601,8475,5706,2665,5707,3726,5709,3727,3741,8942,7988,5710,5711,4323,7304,3672,5708,5714,5712,5713,8943,8685,5718,5717,5715,681,5716,5719,5721,5720,5724,5722,7989,5723,1166,1168,5726,5725,5727,181,1806,5728,5729,5730,7233,5732,5731,259,5733,281,208,7332,5734,5735,7045,216,1059,5736,5737],"string":"[\n 7601,\n 8475,\n 5706,\n 2665,\n 5707,\n 3726,\n 5709,\n 3727,\n 3741,\n 8942,\n 7988,\n 5710,\n 5711,\n 4323,\n 7304,\n 3672,\n 5708,\n 5714,\n 5712,\n 5713,\n 8943,\n 8685,\n 5718,\n 5717,\n 5715,\n 681,\n 5716,\n 5719,\n 5721,\n 5720,\n 5724,\n 5722,\n 7989,\n 5723,\n 1166,\n 1168,\n 5726,\n 5725,\n 5727,\n 181,\n 1806,\n 5728,\n 5729,\n 5730,\n 7233,\n 5732,\n 5731,\n 259,\n 5733,\n 281,\n 208,\n 7332,\n 5734,\n 5735,\n 7045,\n 216,\n 1059,\n 5736,\n 5737\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.10903035530303,"string":"1.10903"}}},{"rowIdx":34,"cells":{"authors":{"kind":"list like","value":["Vaishak Belle"],"string":"[\n \"Vaishak Belle\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Symbolic Logic meets Machine Learning: \\\\ A Brief Survey in Infinite Domains"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-06-15T15:29:49Z","string":"2020-06-15T15:29:49Z"},"category":{"kind":"string","value":"cs.AI"},"abstract":{"kind":"string","value":" The tension between deduction and induction is perhaps the most fundamental issue in areas such as philosophy, cognition and artificial intelligence (AI). The deduction camp concerns itself with questions about the expressiveness of formal languages for capturing knowledge about the world, together with proof systems for reasoning from such knowledge bases. The learning camp attempts to generalize from examples about partial descriptions about the world. In AI, historically, these camps have loosely divided the development of the field, but advances in cross-over areas such as statistical relational learning, neuro-symbolic systems, and high-level control have illustrated that the dichotomy is not very constructive, and perhaps even ill-formed. In this article, we survey work that provides further evidence for the connections between logic and learning. Our narrative is structured in terms of three strands: logic versus learning, machine learning for logic, and logic for machine learning, but naturally, there is considerable overlap. We place an emphasis on the following ``sore\" point: there is a common misconception that logic is for discrete properties, whereas probability theory and machine learning, more generally, is for continuous properties. We report on results that challenge this view on the limitations of logic, and expose the role that logic can play for learning in infinite domains. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"3179fec8-9fa2-4590-ac34-0d3a7530e454","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Symbolic Logic meets Machine Learning: \\\\ A Brief Survey in Infinite Domains"]],"subsections":["fd4766f5-f661-4c93-82d9-4297dbbf91fe","dc3d8034-6633-4d75-aa80-d0f0633c74e0","a715b355-5229-4cca-872a-673dcade22b2","66dfe73d-6adf-4a8f-bcd1-7c3121a3fbec","1cda66fd-50ab-40fa-8bc2-8a5604130ee9"],"title":"root"},{"cite_extract_rate":0.4,"cites":[1177,3728,3726,8475,3725,3727],"content":"The tension between \\textit{deduction} and \\textit{induction} is perhaps the most fundamental issue in areas such as philosophy, cognition and artificial intelligence (AI). The deduction camp concerns itself with questions about the expressiveness of formal languages for capturing knowledge about the world, together with proof systems for reasoning from such knowledge bases. The learning camp attempts to generalize from examples about partial descriptions about the world. In AI, historically, these camps have loosely divided the development of the field, but advances in cross-over areas such as \\textit{statistical relational learning} , \\textit{neuro-symbolic systems} , and \\textit{high-level control} have illustrated that the dichotomy is not very constructive, and perhaps even ill-formed. Indeed, logic emphasizes high-level reasoning, and encourages structuring the world in terms of objects, properties, and relations. In contrast, much of the inductive machinery assume random variables to be independent and identically distributed, which can be problematic when attempting to exploit symmetries and causal dependencies between groups of objects. \nBut the threads connecting logic and learning go deeper, far beyond the apparent flexibility that logic offers for modeling relations and hierarchies in noisy domains. At a conceptual level, for example, although there is much debate about what precisely commonsense knowledge might look like, it is widely acknowledged that concepts such as time, space, abstraction and causality are essential . In that regard, (classical, or perhaps non-classical) logic can provide the formal machinery to reason about such concepts in a rigorous way. At a pragmatic level, despite the success of methods such as deep learning, it is now increasingly recognized that owing to a number of reasons, including model re-use, transferability, causal understanding, relational abstraction, explainability and data efficiency, those methods need to be further augmented with logical, symbolic and/or programmatic artifacts . Finally, for building intelligent agents, it is recognized that low-level, data-intensive, reactive computations needs to be tightly integrated with high-level, deliberative computations , the latter possibly also engaging in hypothetical and counterfactual reasoning. \nHere, a parallel is often drawn to Kahneman's so-called \\textit{System 1} versus \\textit{System 2} processing in human cognition , in the sense that experiential and reactive processing (learned behavior) needs to be coupled with cogitative processing (reasoning, deliberation and introspection) for sophisticated machine intelligence. \nThe purpose of this article is not to resolve this debate, but rather provide further evidence for the connections between logic and learning. In particular, our narrative is inspired by a recent symposium on logic and learning , where the landscape was structured in terms of three strands: {\\it\n\\begin{enumerate}\n\\item \\textbf{Logic vs. Machine Learning}, including the study of problems that can be solved using either logic-based techniques or via machine learning, $\\ldots$; \n\\item \\textbf{Machine Learning for Logic}, including the learning of logical artifacts, such as formulas, logic programs, $\\ldots$; and \n\\item \\textbf{Logic for Machine Learning}, including the role of logics in delineating the boundary between tractable and intractable learning problems, $\\ldots,$ and the use of logic as a declarative framework for expressing machine learning constructs.\n\\end{enumerate}\n}{} \nIn this article, we particularly focus on the following ``sore\" point: there is a common misconception that logic is for discrete properties, whereas probability theory and machine learning, more generally, is for continuous properties. It is true that logical formulas are discrete structures, but they can very easily also express properties about countably infinite or even uncountably many objects. Consequently, in this article we survey some recent results that tackle the integration of logic and learning in infinite domains. In particular, in the context of the above three strands, we report on the following developments. On (1), we discuss approaches for logic-based probabilistic inference in continuous domains. On (2), we cover approaches for learning logic programs in continuous domains, as well as learning formulas that represent countably infinite sets of objects. Finally, on (3), we discuss attempts to use logic as a declarative framework for common tasks in machine learning over discrete and continuous features, as well as using logic as a meta-theory to consider notions such as the \\textit{abstraction} of a probabilistic model. \n{We remark that this survey is undoubtedly a biased view, as the area of research is large, but we do attempt to briefly cover the major threads. Readers are encouraged to refer to discussions in , among others, to get a sense of the breadth of the area.\n}","id":"fd4766f5-f661-4c93-82d9-4297dbbf91fe","level":"section","origin_cites_number":15,"parent_id":"3179fec8-9fa2-4590-ac34-0d3a7530e454","prefix_titles":[["title","Symbolic Logic meets Machine Learning: \\\\ A Brief Survey in Infinite Domains"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0.23333333333333303,"cites":[7747,3732,3729,3731,8683,8684,3730],"content":"\\label{sec:logic_vs_machine_learning}\nTo appreciate the role and impact of logic-based solvers for machine learning systems, it is perhaps useful to consider the core computational problem underlying (probabilistic) machine learning: the problem of inference, including evaluating the partition function (or conditional probabilities) of a probabilistic graphical model such as a Bayesian network. \nWhen leveraging Bayesian networks for machine learning tasks , the networks are often learned using local search to maximize a likelihood or a Bayesian quantity. For example, given data \\( \\D \\) and the current guess for the network \\( \\N \\), we might estimate the ``goodness'' of the guess by means of a score: \\( {\\it score}(\\N ,\\D) \\propto \\log \\Pr(\\D\\mid \\N) - {\\it size}(\\N) \\). That is, we want to maximize the fit of the data wrt the current guess, but we would like to penalize the model complexity, to avoid overfitting. Then, we would opt for a second guess \\( \\N' \\) only if \\( {\\it score}(\\N',\\D) \\gt {\\it score}(\\N,\\D) \\). Needless to say, even with a reasonable local search procedure, the most significant computational effort here is that of probabilistic inference. \nReasoning in such networks becomes especially challenging with logical syntax. The prevalence of large-scale social networks, machine reading domains, and other types of relational knowledge bases has led to numerous formalisms that borrow the syntax of predicate logic for probabilistic modeling . This has led to a large family of solvers for the \\emph{weighted model counting} (WMC) problem . The idea is this: given a Bayesian network, a relational Bayesian network, a factor graph, or a probabilistic program , one considers an encoding of the formalism as a \\emph{weighted propositional theory}, consisting of a propositional theory \\( \\Delta \\) and a weight function \\( w \\) that maps atoms in \\( \\Delta \\) to \\( \\real\\su + \\). Recall that SAT is the problem of finding an assignment to such a \\( \\Delta, \\) whereas \\#SAT counts the number of assignments for \\( \\Delta. \\) WMC extends \\#SAT by computing the sum of the weights of all assignments: that is, given a set of models \\( \\M(\\Delta) = \\set{ M \\mid M \\models \\Delta} \\), \nwe evaluate the quantity \\(\n\tW(\\Delta) = \\sum\\sub {M \\in \\M(\\Delta)} w(M)\n\\) \nwhere \\( w(M) \\) is factorized in terms of the atoms true at \\( M. \\) To obtain the conditional probability of a query \\( q \\) against evidence \\( e \\) (wrt the theory \\( \\Delta \\)), we define \\( \\Pr(q\\mid e) = W(\\Delta\\land q \\land e) / W(\\Delta\\land e). \\)\nThe popularity of WMC can be explained as follows. Its formulation elegantly decouples the logical or symbolic representation from the numeric representation, which is encapsulated in the weight function. When building solvers, this allows us to reason about logical equivalence and reuse SAT solving technology (such as constraint propagation and clause learning). WMC also makes it more natural to reason about deterministic, hard constraints in a probabilistic context . Both exact solvers, based on knowledge compilation , as well as approximate solvers have emerged in the recent years, as have lifted techniques that exploit the relational syntax during inference (but in a finite domain setting). For ideas on generating such representations randomly to assess scalability and compare inference algorithms, see , for example. \nOn the point of modelling finite vs infinite properties, \nnote that owing to the underlying propositional language, the formulation is limited to discrete random variables. A similar observation can be made for SAT, which for the longest time could only be applied in discrete domains. This changed with the increasing popularity of \\emph{satisfiability modulo theories} (SMT) , which enable us to, for example, reason about the satisfiability of linear constraints over the rationals. Extending earlier insights on piecewise-polynomial weight functions , the formulation of \\emph{weighted model integration} (WMI) was proposed in . WMI extends WMC by leveraging the idea that SMT theories can represent mixtures of Boolean and continuous variables: for example, a formula such as \\( p \\land (x\\gt 5) \\) denotes the logical conjunction of a Boolean variable \\( p \\) and a real-valued variable \\( x \\) taking values greater than 5. For every assignment to the Boolean and continuous variables, the WMI problem defines a weight. The total WMI is computed by integrating these weights over the domain of solutions to \\( \\Delta \\), which is a mixed discrete-continuous (or simply \\textit{hybrid}) space. Consider, for example, the special case when \\( \\Delta \\) has no Boolean variables, and the weight of every model is 1. Then, the WMI simplifies to computing the volume of the polytope encoded in \\( \\Delta \\). When we additionally allow for Boolean variables in \\( \\Delta \\), this special case becomes the hybrid version of \\#SAT, known as \\#SMT . Since that proposal, numerous advances have been made on building efficient WMI solvers (e.g., ) including the development of compilation targets . \nNote that WMI proposes an extension of WMC for uncountably infinite (i.e., continuous) domains. What about countably infinite domains? The latter type is particularly useful for reasoning in (general) first-order settings, where we may say that a property such as \\( \\forall x,y,z ({\\it parent}(x,y) \\land {\\it parent}(y,z) \\supset {\\it grandparent}(x,z)) \\) applies to every possible \\( x, y\\) and $z$. Of course, in the absence of the finite domain assumption, reasoning in the first-order setting suffers from undecidability properties, and so various strategies have emerged for reasoning about an \\textit{open universe} . One popular approach is to perform \\emph{forward reasoning}, where samples needed for probability estimation are obtained from the facts and declarations in the probabilistic model . Each such sample corresponds to a possible world. But there may be (countably or uncountably) infinitely many worlds, and so exact inference is usually sacrificed. A second approach is to restrict the model wrt the query and evidence atoms and define estimation from the resulting finite sub-model , which may also be substantiated with exact inference in special cases . \nGiven the successes of logic-based solvers for inference and probability estimation, one might wonder whether such solvers would also be applicable to learning tasks in models with relational features and hard, deterministic constraints? \n These, in addition to other topics, are considered in the next section.","id":"dc3d8034-6633-4d75-aa80-d0f0633c74e0","level":"section","origin_cites_number":30,"parent_id":"3179fec8-9fa2-4590-ac34-0d3a7530e454","prefix_titles":[["title","Symbolic Logic meets Machine Learning: \\\\ A Brief Survey in Infinite Domains"],["section","Logic vs. Machine Learning"]],"subsections":[],"title":"Logic vs. Machine Learning"},{"cite_extract_rate":0.30769230769230704,"cites":[8685,3726,3737,3735,3736,8686,3733,3734],"content":"\\label{sec:machine_learning_for_logic}\nAt least since the time of Socrates, inductive reasoning has been a core issue for the logical worldview, as we need a mechanism for obtaining axiomatic knowledge. In that regard, the learning of logical and symbolic artifacts is an important issue in AI, and computer science more generally . There is a considerable body of work on learning propositional and relational formulas, and in context of probabilistic information, learning weighted formulas . Approaches can be broadly lumped together as follows. \\begin{enumerate}\n\t\\item \\emph{Entailment-based scoring:} Given a logical language \\( \\L, \\) background knowledge \\( \\B \\subset \\L, \\) examples \\( \\D \\) (usually a set of \\( \\L \\)-atoms), \n\t\tfind a hypothesis \\( \\H \\in {\\overline \\H}, \\H \\subset \\L \\) such that \\( \\B \\cup \\H \\) entail the instances in \\( \\D. \\)\n\tHere, the set \\( {\\overline \\H} \\) places restrictions of the syntax of \\( \\H \\) so as to control model complexity and generalization. (For example, \\( \\H = \\D \\) is a trivial hypothesis that satisfies the entailment stipulation.)\n\t\\item \\emph{Likelihood-based scoring:} Given \\( \\L,\\B \\) and \\( \\D \\) as defined above, find \\( \\H\\subset \\L \\) such that \\( {\\it score}(\\H, \\D) \\gt {\\it score}(\\H', \\D) \\) for every \\( \\H' \\neq \\H. \\) As discussed before, we might define \\( {\\it score}(\\H ,\\D) \\propto \\log \\Pr(\\D\\mid \\H) - {\\it size}(\\H) \\). Here, like \\( {\\overline \\H} \\) above, \\( {\\it size}(\\H) \\) attempts to the control model complexity and generalization. \n\\end{enumerate}\nMany recipes based on these schemes are possible. For example, we may use entailment-based inductive synthesis for an initial estimate of the hypothesis, and then resort to Bayesian scoring models . The synthesis step might invoke neural machinery . We might not require that the hypothesis entails every example in \\( \\D \\) but only the largest consistent subset, which is sensible when we expect the examples to be noisy . We might compile \\( \\B \\) to an efficient data structure, and perform likelihood-based scoring on that structure , and so \\( \\B \\) could be seen as deterministic domain-specific constraints. Finally, we might stipulate the conditions under which a ``correct'' hypothesis may be inferred wrt unknown ground truth, only a subset of which is provided in \\( \\D. \\) This is perhaps best represented by the (probably approximately correct) PAC-semantics that captures the quality possessed by the output of learning algorithm whilst costing for the number of examples that need to be observed . (But other formulations are also possible, e.g., .) \nThis discussion pertained to finite domains. \nWhat about continuous spaces? By means of arithmetic fragments and \nformulations like WMI, it should be clear that it now becomes possible to extend the above schemes to learn continuous properties. For example, one could learn linear expressions from data . For an account that also tries to evaluate a hypothesis that is correct wrt unknown ground truth, see . \nIf the overall objective is to obtain a distribution of the data, other possibilities present themselves. \nIn , for example, real-valued data points are first lumped together to obtain \natomic continuous random variables. From these, relational formulas are constructed so as to yield hybrid probabilistic programs. The learning is based on likelihood scoring. In , the real-valued data points are first intervalized, and polynomials are learned for those intervals based on likelihood scoring. These weighted atoms are then used for learning clauses by entailment judgements . \nSuch ideas can also be extended to data structures inspired by knowledge compilation, often referred to as \\textit{circuits} . Knowledge compilation arose as a way to represent logical theories in a manner where certain kinds of computations (e.g., checking satisfiability) is significantly more effective, often polynomial in the size of the circuit. In the context of probabilistic inference, the idea was to then position probability estimation to also be computable in time polynomial in the size of the circuit . Consequently, (say) by means of likelihood-based scoring, the learning of circuits is particularly attractive because once learned, the bottleneck of inference is alleviated . In \n, along the lines of the work above on learning logical formulas in continuous domains, it is shown that the learning of circuits can also be coupled with WMI. \nWhat about countably infinite domains? In most pragmatic instances of learning logical artifacts, the difference between the uncountable and countably infinite setting is this: in the former, we see finitely many real-valued samples as being drawn from an (unknown) interval, and we could inspect these samples to crudely infer a lower and upper bound. In the latter, based on finitely many relational atoms, we would need to infer a universally quantified clause, such as \\( \\forall x,y,z ({\\it parent}(x,y) \\land {\\it parent}(y,z) \\supset {\\it grandparent}(x,z)) \\). If we are after a hypothesis that is simply guaranteed to be consistent wrt the observed examples, then standard rule induction strategies would suffice , and we could interpret the rules as quantifying over a countably infinite domain. But this is somewhat unsatisfactory, as there is no distinction between the rules learned in the standard finite setting and its supposed applicability to the infinite setting. What is really needed is an analysis of what rule learning would mean wrt the infinitely many examples that have \\emph{not} been observed. This was recently considered via the PAC-semantics in , by appealing to ideas on reasoning with open universes discussed earlier . \nBefore concluding this section, it is worth noting that although the above discussion is primarily related to the learning of logical artifacts, it can equivalently be seen as a class of machine learning methods that leverage symbolic domain knowledge . \nIndeed, logic-based probabilistic inference over deterministic constraints, and entailment-based induction augmented with background knowledge are instances of such a class. \nAnalogously, the automated construction of relational and statistical knowledge bases by combining background knowledge with extracted tuples (obtained, for example, by applying natural language processing techniques to large textual data) is another instance of such a class. \nIn the next section, we will consider yet another way in which logical and symbolic artifacts can influence learning: we will see how such artifacts are useful to enable tractability, correctness, modularity and compositionality.","id":"a715b355-5229-4cca-872a-673dcade22b2","level":"section","origin_cites_number":26,"parent_id":"3179fec8-9fa2-4590-ac34-0d3a7530e454","prefix_titles":[["title","Symbolic Logic meets Machine Learning: \\\\ A Brief Survey in Infinite Domains"],["section","Machine Learning for Logic"]],"subsections":[],"title":"Machine Learning for Logic"},{"cite_extract_rate":0.268292682926829,"cites":[1613,3742,3743,3738,3735,3744,8686,3727,3740,3739,3741],"content":"\\label{sec:logic_for_machine_learning}\nThere are two obvious ways in which a logical framework can provide insights on machine learning theory. First, consider that computational tractability is of central concern when applying logic in computer science, knowledge representation, database theory and search . Thus, the natural question to wonder is whether these ideas would carry over to probabilistic machine learning. On the one hand, probabilistic extensions to tractable knowledge representation frameworks could be considered . But on the other, as discussed previously, ideas from knowledge compilation, and the use of circuits, in particular, are proving very effective for designing tractable paradigms for machine learning. While there has always been an interest in capturing tractable distributions by means of low tree-width models , knowledge compilation has provided a way to also represent high tree-width models and enable exact inference for a range of queries . See for a comprehensive view on the use of knowledge compilation for machine learning. \nThe other obvious way logic can provide insights on machine learning theory is by offering a formal apparatus to reason about \\textit{context}. \nMachine learning problems are often positioned as atomic tasks, such as a classification task where regions of images need to be labeled as cats or dogs. However, even in that limited context, we imagine the resulting classification system as being deployed as part of a larger system, which includes various modules that communicate or interface with the classification system. We imagine an implicit accountability to the labelling task in that the detected object is either a cat or a dog, but not both. If there is information available that all the entities surrounding the object of interest have been labelled as lions, we would want to accord a high probability to the object being a cat, possibly a wild cat. There is a very low chance of the object being a dog, then. If this is part of a vision system on a robot, we should ensure that the robot never tramples on the object, regardless of whether it is a type of cat or a dog. To inspect such patterns, and provide meta-theory for machine learning, it can be shown that symbolic, programmatic and logical artifacts are enormously useful. We will specifically consider correctness, modularity and compositionality to explore the claim.\nOn the topic of correctness, the classical framework in computer science is \\emph{verification}: can we provide a formal specification of what is desired, and \ncan the system be checked against that specification? In a machine learning context, we might ask whether the system, during or after training, satisfies a specification. \nThe specification here might mean constraints about the physical laws of the domain, or notions of perturbation in the input space while ensuring that the labels do not change, or insisting that the prediction does not label an object as being both a cat and a dog, or otherwise ensuring that outcomes are not subject to, say, gender bias. Although there is a broad body of work on such issues, touching more generally on \\textit{trust} , we discuss approaches closer to the thrust of this article. For example, show that a trained neural network can be verified by means of an SMT encoding of the network. In recent work, show that the loss function of deep learning systems can be adjusted to logical constraints by insisting that the distribution on the predictions is proportional to the weighted model count of those constraints. In , prior (logical) constraints are compiled to a circuit to be used for probability estimation. \nIn , circuits are shown to be amenable to training against probabilistic and causal prior constraints, including assertions about fairness, for example. \nIn , a somewhat different approach to respecting domain constraints is taken: the low-level prediction is obtained as usual from a machine learning module, which is then interfaced with a probabilistic relational language and its symbolic engine. That is, the reasoning is positioned to be tackled directly by the symbolic engine. In a sense, such approaches cut across the three strands: the symbolic engine uses weighted model counting, the formulas in the language could be obtained by (say) entailment-based scoring, and the resulting language supports modularity and compositionality (discussed below). \nWhile there is not much to be said about the distinction between finite vs infinite wrt correctness, many of these ideas are likely amenable to extensions to an infinite setting in the ways discussed in the previous sections (e.g., considering constraints of a continuous or a countably infinite nature). \nOn the topic of modularity, recall that the general idea is to reduce, simplify or otherwise abstract a (probabilistic) computation as an atomic entity, which is then to be referenced in another, possibly more complex, entity. In standard programming languages, this might mean the compartmentalization and interrelation of computational entities. For machine learning, approaches such as probabilistic programming support probabilistic primitives in the language, with the intention of making learning modules re-usable and modular. It can be shown, for example, that the computational semantics of some of these languages reduce to WMC . Thus, in the infinite case, a corresponding reduction to WMI follows . \nA second dimension to modularity is the notion of \\emph{abstraction}. Here, we seek to model, reason and explain the behavior of systems in a more tractable search space, by omitting irrelevant details. The idea is widely used in natural and social sciences. Think of understanding the political dynamics of elections by studying micro level phenomena (say, voter grievances in counties) versus macro level events (e.g., television advertisements, gerrymandering). In particular, in computer science, it is often understood as the process of mapping one representation onto a simpler representation by suppressing irrelevant information. \nIn fact, integrating low-level behavior with high-level reasoning, exploiting relational representations to reduce the number of inference computations, and many other search space reduction techniques can all loosely be seen as instances of abstraction . \nWhile there has been significant work on abstraction in deterministic systems , for machine learning, however, a probabilistic variant is clearly needed. In , an account of abstraction for loop-free propositional probabilistic programs is provided, where certain parts of the program (possibly involving continuous properties) can be reduced to a Bernoulli random variable. For example, suppose every occurrence of the continuous random variable $x$, drawn uniformly on the interval [0,1], in a program is either of the form $x\\leq 7$ or of the form $x\\gt 7$. Then, we could use a discrete random variable $b$ with a 0.7 probability of being true to capture $x\\leq 7$; and analogously, $\\neg b$ to capture $x\\gt 7$. The resulting program is likely to be simpler. \nIn , an account of abstraction for probabilistic relational models is considered, where the notion of abstraction also extends to deterministic constraints and complex formulas. For example, a single probabilistic variable in the abstracted model could denote a complex logical formula in the original model. Moreover, the logical properties that enable verifying and inducing abstractions are also considered, and it is shown how WMC is sufficient for the computability of these properties (also see ). \nIncidentally, abstraction brings to light a reduction between finite vs infinite: it is shown in that the modelling of piecewise densities as weighted propositions, which is leveraged in WMI , is a simple case of the more general account. Therefore, it is worthwhile to investigate whether this or other accounts of abstraction could emerge as general-purpose tools that allow us to inspect the conditions under which infinitary statements reduce to finite computations. \nA broader point here is the role abstraction might play in generating explanations . For example, a user's understanding of the domain is likely to be different from the low-level data that a machine learning system interfaces with~, and so, abstractions can capture these two levels in a formal way. \nFinally, we turn to the topic of compositionality, which, of course, is closely related to modularity in that we want to distinct modules to come together to form a complex composition. \nNot surprisingly, this is of great concern in AI, as it is widely acknowledged that most AI systems will involve heterogeneous components, some of which may involve learning from data, and others reasoning, search and symbol manipulation . In continuation with the above discussion, probabilistic programming is one such endeavor that purports to tackle this challenge by allowing modular components to be composed over programming and/or logical connectives . (See for ideas in deterministic systems.) However, probabilistic programming only composes probabilistic computations, but does not offer an obvious means to capture other types of search-based computations, such as SAT, and integer and convex programming. \nRecall that the computational semantics of probabilistic programs reduces to WMC . Following works such as , an interesting observation made in is that by appealing to a sum of products computation over different semiring structures, we can realize a large number of tasks such as satisfiability, unweighted model counting, sensitivity analysis, gradient computations, in addition to WMC. \nIt was then shown in that the idea could be generalized further for infinite domains: by defining a measure on first-order models, WMI and convex optimization can also be captured. As the underlying language is a logical one, composition can already be defined using logical connectives. But an additional, more involved, notion of composition is also proposed, where a sum of products over different semirings can be concatenated. To reiterate, the general idea behind these proposals \nis to arrive at a principled paradigm that allows us to interface learned modules with other types of search and optimization computations for the compositional building of AI systems. See also for analogous discussions, but where a different type of coupling for the underlying computations is suggested. Overall, we observed that a formal apparatus (symbolic, programmatic and logical artifacts) help us define such compositional constructions by providing a meta-theory.","id":"66dfe73d-6adf-4a8f-bcd1-7c3121a3fbec","level":"section","origin_cites_number":41,"parent_id":"3179fec8-9fa2-4590-ac34-0d3a7530e454","prefix_titles":[["title","Symbolic Logic meets Machine Learning: \\\\ A Brief Survey in Infinite Domains"],["section","Logic for Machine Learning"]],"subsections":[],"title":"Logic for Machine Learning"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclusions}\nIn this article, we surveyed work that provides further evidence for the connections between logic and learning. \nOur narrative was structured in terms of three strands: logic versus learning, machine learning for logic, and logic for machine learning, but naturally, there was considerable overlap. \nWe covered a large body of work on what these connections look like, including, for example, pragmatic concerns such as the use of hard, domain-specific constraints and background knowledge, all of which considerably eases the requirement that all of the agent's knowledge should be derived from observations alone. (See discussions in on the limitations of learned behavior, for example.) \nWhere applicable, we placed an emphasis on how extensions to infinite domains are possible. In the very least, logical artifacts can help in constraining, simplifying and/or composing machine learning entities, and in providing a principled way to study the underlying representational and computational issues. \nIn general, this type of work could help us move beyond the narrow focus of the current learning literature so as to deal with time, space, abstraction, causality, quantified generalizations, relational abstractions, unknown domains, unforeseen examples, among other things, in a principled fashion. In fact, what is being advocated is the tackling of problems that symbolic logic and machine learning might struggle to address individually. One could even think of the need for a recursive combination of strands 2 and 3: purely reactive components interact with purely cogitative elements, but then those reactive components are learned against domain constraints, and the cogitative elements are induced from data, and so on. More broadly, making progress towards a formal realization of \\emph{System 1} versus \\emph{System 2} processing might also contribute to our understanding of human intelligence, or at least capture human-like intelligence in automated systems. \n\\bibliographystyle{abbrv}\n\\bibliography{group}\n\\end{document}","id":"1cda66fd-50ab-40fa-8bc2-8a5604130ee9","level":"section","origin_cites_number":1,"parent_id":"3179fec8-9fa2-4590-ac34-0d3a7530e454","prefix_titles":[["title","Symbolic Logic meets Machine Learning: \\\\ A Brief Survey in Infinite Domains"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"3179fec8-9fa2-4590-ac34-0d3a7530e454\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Symbolic Logic meets Machine Learning: \\\\\\\\ A Brief Survey in Infinite Domains\"\n ]\n ],\n \"subsections\": [\n \"fd4766f5-f661-4c93-82d9-4297dbbf91fe\",\n \"dc3d8034-6633-4d75-aa80-d0f0633c74e0\",\n \"a715b355-5229-4cca-872a-673dcade22b2\",\n \"66dfe73d-6adf-4a8f-bcd1-7c3121a3fbec\",\n \"1cda66fd-50ab-40fa-8bc2-8a5604130ee9\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 1177,\n 3728,\n 3726,\n 8475,\n 3725,\n 3727\n ],\n \"content\": \"The tension between \\\\textit{deduction} and \\\\textit{induction} is perhaps the most fundamental issue in areas such as philosophy, cognition and artificial intelligence (AI). The deduction camp concerns itself with questions about the expressiveness of formal languages for capturing knowledge about the world, together with proof systems for reasoning from such knowledge bases. The learning camp attempts to generalize from examples about partial descriptions about the world. In AI, historically, these camps have loosely divided the development of the field, but advances in cross-over areas such as \\\\textit{statistical relational learning} , \\\\textit{neuro-symbolic systems} , and \\\\textit{high-level control} have illustrated that the dichotomy is not very constructive, and perhaps even ill-formed. Indeed, logic emphasizes high-level reasoning, and encourages structuring the world in terms of objects, properties, and relations. In contrast, much of the inductive machinery assume random variables to be independent and identically distributed, which can be problematic when attempting to exploit symmetries and causal dependencies between groups of objects. \\nBut the threads connecting logic and learning go deeper, far beyond the apparent flexibility that logic offers for modeling relations and hierarchies in noisy domains. At a conceptual level, for example, although there is much debate about what precisely commonsense knowledge might look like, it is widely acknowledged that concepts such as time, space, abstraction and causality are essential . In that regard, (classical, or perhaps non-classical) logic can provide the formal machinery to reason about such concepts in a rigorous way. At a pragmatic level, despite the success of methods such as deep learning, it is now increasingly recognized that owing to a number of reasons, including model re-use, transferability, causal understanding, relational abstraction, explainability and data efficiency, those methods need to be further augmented with logical, symbolic and/or programmatic artifacts . Finally, for building intelligent agents, it is recognized that low-level, data-intensive, reactive computations needs to be tightly integrated with high-level, deliberative computations , the latter possibly also engaging in hypothetical and counterfactual reasoning. \\nHere, a parallel is often drawn to Kahneman's so-called \\\\textit{System 1} versus \\\\textit{System 2} processing in human cognition , in the sense that experiential and reactive processing (learned behavior) needs to be coupled with cogitative processing (reasoning, deliberation and introspection) for sophisticated machine intelligence. \\nThe purpose of this article is not to resolve this debate, but rather provide further evidence for the connections between logic and learning. In particular, our narrative is inspired by a recent symposium on logic and learning , where the landscape was structured in terms of three strands: {\\\\it\\n\\\\begin{enumerate}\\n\\\\item \\\\textbf{Logic vs. Machine Learning}, including the study of problems that can be solved using either logic-based techniques or via machine learning, $\\\\ldots$; \\n\\\\item \\\\textbf{Machine Learning for Logic}, including the learning of logical artifacts, such as formulas, logic programs, $\\\\ldots$; and \\n\\\\item \\\\textbf{Logic for Machine Learning}, including the role of logics in delineating the boundary between tractable and intractable learning problems, $\\\\ldots,$ and the use of logic as a declarative framework for expressing machine learning constructs.\\n\\\\end{enumerate}\\n}{} \\nIn this article, we particularly focus on the following ``sore\\\" point: there is a common misconception that logic is for discrete properties, whereas probability theory and machine learning, more generally, is for continuous properties. It is true that logical formulas are discrete structures, but they can very easily also express properties about countably infinite or even uncountably many objects. Consequently, in this article we survey some recent results that tackle the integration of logic and learning in infinite domains. In particular, in the context of the above three strands, we report on the following developments. On (1), we discuss approaches for logic-based probabilistic inference in continuous domains. On (2), we cover approaches for learning logic programs in continuous domains, as well as learning formulas that represent countably infinite sets of objects. Finally, on (3), we discuss attempts to use logic as a declarative framework for common tasks in machine learning over discrete and continuous features, as well as using logic as a meta-theory to consider notions such as the \\\\textit{abstraction} of a probabilistic model. \\n{We remark that this survey is undoubtedly a biased view, as the area of research is large, but we do attempt to briefly cover the major threads. Readers are encouraged to refer to discussions in , among others, to get a sense of the breadth of the area.\\n}\",\n \"id\": \"fd4766f5-f661-4c93-82d9-4297dbbf91fe\",\n \"level\": \"section\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"3179fec8-9fa2-4590-ac34-0d3a7530e454\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Symbolic Logic meets Machine Learning: \\\\\\\\ A Brief Survey in Infinite Domains\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.23333333333333303,\n \"cites\": [\n 7747,\n 3732,\n 3729,\n 3731,\n 8683,\n 8684,\n 3730\n ],\n \"content\": \"\\\\label{sec:logic_vs_machine_learning}\\nTo appreciate the role and impact of logic-based solvers for machine learning systems, it is perhaps useful to consider the core computational problem underlying (probabilistic) machine learning: the problem of inference, including evaluating the partition function (or conditional probabilities) of a probabilistic graphical model such as a Bayesian network. \\nWhen leveraging Bayesian networks for machine learning tasks , the networks are often learned using local search to maximize a likelihood or a Bayesian quantity. For example, given data \\\\( \\\\D \\\\) and the current guess for the network \\\\( \\\\N \\\\), we might estimate the ``goodness'' of the guess by means of a score: \\\\( {\\\\it score}(\\\\N ,\\\\D) \\\\propto \\\\log \\\\Pr(\\\\D\\\\mid \\\\N) - {\\\\it size}(\\\\N) \\\\). That is, we want to maximize the fit of the data wrt the current guess, but we would like to penalize the model complexity, to avoid overfitting. Then, we would opt for a second guess \\\\( \\\\N' \\\\) only if \\\\( {\\\\it score}(\\\\N',\\\\D) \\\\gt {\\\\it score}(\\\\N,\\\\D) \\\\). Needless to say, even with a reasonable local search procedure, the most significant computational effort here is that of probabilistic inference. \\nReasoning in such networks becomes especially challenging with logical syntax. The prevalence of large-scale social networks, machine reading domains, and other types of relational knowledge bases has led to numerous formalisms that borrow the syntax of predicate logic for probabilistic modeling . This has led to a large family of solvers for the \\\\emph{weighted model counting} (WMC) problem . The idea is this: given a Bayesian network, a relational Bayesian network, a factor graph, or a probabilistic program , one considers an encoding of the formalism as a \\\\emph{weighted propositional theory}, consisting of a propositional theory \\\\( \\\\Delta \\\\) and a weight function \\\\( w \\\\) that maps atoms in \\\\( \\\\Delta \\\\) to \\\\( \\\\real\\\\su + \\\\). Recall that SAT is the problem of finding an assignment to such a \\\\( \\\\Delta, \\\\) whereas \\\\#SAT counts the number of assignments for \\\\( \\\\Delta. \\\\) WMC extends \\\\#SAT by computing the sum of the weights of all assignments: that is, given a set of models \\\\( \\\\M(\\\\Delta) = \\\\set{ M \\\\mid M \\\\models \\\\Delta} \\\\), \\nwe evaluate the quantity \\\\(\\n\\tW(\\\\Delta) = \\\\sum\\\\sub {M \\\\in \\\\M(\\\\Delta)} w(M)\\n\\\\) \\nwhere \\\\( w(M) \\\\) is factorized in terms of the atoms true at \\\\( M. \\\\) To obtain the conditional probability of a query \\\\( q \\\\) against evidence \\\\( e \\\\) (wrt the theory \\\\( \\\\Delta \\\\)), we define \\\\( \\\\Pr(q\\\\mid e) = W(\\\\Delta\\\\land q \\\\land e) / W(\\\\Delta\\\\land e). \\\\)\\nThe popularity of WMC can be explained as follows. Its formulation elegantly decouples the logical or symbolic representation from the numeric representation, which is encapsulated in the weight function. When building solvers, this allows us to reason about logical equivalence and reuse SAT solving technology (such as constraint propagation and clause learning). WMC also makes it more natural to reason about deterministic, hard constraints in a probabilistic context . Both exact solvers, based on knowledge compilation , as well as approximate solvers have emerged in the recent years, as have lifted techniques that exploit the relational syntax during inference (but in a finite domain setting). For ideas on generating such representations randomly to assess scalability and compare inference algorithms, see , for example. \\nOn the point of modelling finite vs infinite properties, \\nnote that owing to the underlying propositional language, the formulation is limited to discrete random variables. A similar observation can be made for SAT, which for the longest time could only be applied in discrete domains. This changed with the increasing popularity of \\\\emph{satisfiability modulo theories} (SMT) , which enable us to, for example, reason about the satisfiability of linear constraints over the rationals. Extending earlier insights on piecewise-polynomial weight functions , the formulation of \\\\emph{weighted model integration} (WMI) was proposed in . WMI extends WMC by leveraging the idea that SMT theories can represent mixtures of Boolean and continuous variables: for example, a formula such as \\\\( p \\\\land (x\\\\gt 5) \\\\) denotes the logical conjunction of a Boolean variable \\\\( p \\\\) and a real-valued variable \\\\( x \\\\) taking values greater than 5. For every assignment to the Boolean and continuous variables, the WMI problem defines a weight. The total WMI is computed by integrating these weights over the domain of solutions to \\\\( \\\\Delta \\\\), which is a mixed discrete-continuous (or simply \\\\textit{hybrid}) space. Consider, for example, the special case when \\\\( \\\\Delta \\\\) has no Boolean variables, and the weight of every model is 1. Then, the WMI simplifies to computing the volume of the polytope encoded in \\\\( \\\\Delta \\\\). When we additionally allow for Boolean variables in \\\\( \\\\Delta \\\\), this special case becomes the hybrid version of \\\\#SAT, known as \\\\#SMT . Since that proposal, numerous advances have been made on building efficient WMI solvers (e.g., ) including the development of compilation targets . \\nNote that WMI proposes an extension of WMC for uncountably infinite (i.e., continuous) domains. What about countably infinite domains? The latter type is particularly useful for reasoning in (general) first-order settings, where we may say that a property such as \\\\( \\\\forall x,y,z ({\\\\it parent}(x,y) \\\\land {\\\\it parent}(y,z) \\\\supset {\\\\it grandparent}(x,z)) \\\\) applies to every possible \\\\( x, y\\\\) and $z$. Of course, in the absence of the finite domain assumption, reasoning in the first-order setting suffers from undecidability properties, and so various strategies have emerged for reasoning about an \\\\textit{open universe} . One popular approach is to perform \\\\emph{forward reasoning}, where samples needed for probability estimation are obtained from the facts and declarations in the probabilistic model . Each such sample corresponds to a possible world. But there may be (countably or uncountably) infinitely many worlds, and so exact inference is usually sacrificed. A second approach is to restrict the model wrt the query and evidence atoms and define estimation from the resulting finite sub-model , which may also be substantiated with exact inference in special cases . \\nGiven the successes of logic-based solvers for inference and probability estimation, one might wonder whether such solvers would also be applicable to learning tasks in models with relational features and hard, deterministic constraints? \\n These, in addition to other topics, are considered in the next section.\",\n \"id\": \"dc3d8034-6633-4d75-aa80-d0f0633c74e0\",\n \"level\": \"section\",\n \"origin_cites_number\": 30,\n \"parent_id\": \"3179fec8-9fa2-4590-ac34-0d3a7530e454\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Symbolic Logic meets Machine Learning: \\\\\\\\ A Brief Survey in Infinite Domains\"\n ],\n [\n \"section\",\n \"Logic vs. Machine Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Logic vs. Machine Learning\"\n },\n {\n \"cite_extract_rate\": 0.30769230769230704,\n \"cites\": [\n 8685,\n 3726,\n 3737,\n 3735,\n 3736,\n 8686,\n 3733,\n 3734\n ],\n \"content\": \"\\\\label{sec:machine_learning_for_logic}\\nAt least since the time of Socrates, inductive reasoning has been a core issue for the logical worldview, as we need a mechanism for obtaining axiomatic knowledge. In that regard, the learning of logical and symbolic artifacts is an important issue in AI, and computer science more generally . There is a considerable body of work on learning propositional and relational formulas, and in context of probabilistic information, learning weighted formulas . Approaches can be broadly lumped together as follows. \\\\begin{enumerate}\\n\\t\\\\item \\\\emph{Entailment-based scoring:} Given a logical language \\\\( \\\\L, \\\\) background knowledge \\\\( \\\\B \\\\subset \\\\L, \\\\) examples \\\\( \\\\D \\\\) (usually a set of \\\\( \\\\L \\\\)-atoms), \\n\\t\\tfind a hypothesis \\\\( \\\\H \\\\in {\\\\overline \\\\H}, \\\\H \\\\subset \\\\L \\\\) such that \\\\( \\\\B \\\\cup \\\\H \\\\) entail the instances in \\\\( \\\\D. \\\\)\\n\\tHere, the set \\\\( {\\\\overline \\\\H} \\\\) places restrictions of the syntax of \\\\( \\\\H \\\\) so as to control model complexity and generalization. (For example, \\\\( \\\\H = \\\\D \\\\) is a trivial hypothesis that satisfies the entailment stipulation.)\\n\\t\\\\item \\\\emph{Likelihood-based scoring:} Given \\\\( \\\\L,\\\\B \\\\) and \\\\( \\\\D \\\\) as defined above, find \\\\( \\\\H\\\\subset \\\\L \\\\) such that \\\\( {\\\\it score}(\\\\H, \\\\D) \\\\gt {\\\\it score}(\\\\H', \\\\D) \\\\) for every \\\\( \\\\H' \\\\neq \\\\H. \\\\) As discussed before, we might define \\\\( {\\\\it score}(\\\\H ,\\\\D) \\\\propto \\\\log \\\\Pr(\\\\D\\\\mid \\\\H) - {\\\\it size}(\\\\H) \\\\). Here, like \\\\( {\\\\overline \\\\H} \\\\) above, \\\\( {\\\\it size}(\\\\H) \\\\) attempts to the control model complexity and generalization. \\n\\\\end{enumerate}\\nMany recipes based on these schemes are possible. For example, we may use entailment-based inductive synthesis for an initial estimate of the hypothesis, and then resort to Bayesian scoring models . The synthesis step might invoke neural machinery . We might not require that the hypothesis entails every example in \\\\( \\\\D \\\\) but only the largest consistent subset, which is sensible when we expect the examples to be noisy . We might compile \\\\( \\\\B \\\\) to an efficient data structure, and perform likelihood-based scoring on that structure , and so \\\\( \\\\B \\\\) could be seen as deterministic domain-specific constraints. Finally, we might stipulate the conditions under which a ``correct'' hypothesis may be inferred wrt unknown ground truth, only a subset of which is provided in \\\\( \\\\D. \\\\) This is perhaps best represented by the (probably approximately correct) PAC-semantics that captures the quality possessed by the output of learning algorithm whilst costing for the number of examples that need to be observed . (But other formulations are also possible, e.g., .) \\nThis discussion pertained to finite domains. \\nWhat about continuous spaces? By means of arithmetic fragments and \\nformulations like WMI, it should be clear that it now becomes possible to extend the above schemes to learn continuous properties. For example, one could learn linear expressions from data . For an account that also tries to evaluate a hypothesis that is correct wrt unknown ground truth, see . \\nIf the overall objective is to obtain a distribution of the data, other possibilities present themselves. \\nIn , for example, real-valued data points are first lumped together to obtain \\natomic continuous random variables. From these, relational formulas are constructed so as to yield hybrid probabilistic programs. The learning is based on likelihood scoring. In , the real-valued data points are first intervalized, and polynomials are learned for those intervals based on likelihood scoring. These weighted atoms are then used for learning clauses by entailment judgements . \\nSuch ideas can also be extended to data structures inspired by knowledge compilation, often referred to as \\\\textit{circuits} . Knowledge compilation arose as a way to represent logical theories in a manner where certain kinds of computations (e.g., checking satisfiability) is significantly more effective, often polynomial in the size of the circuit. In the context of probabilistic inference, the idea was to then position probability estimation to also be computable in time polynomial in the size of the circuit . Consequently, (say) by means of likelihood-based scoring, the learning of circuits is particularly attractive because once learned, the bottleneck of inference is alleviated . In \\n, along the lines of the work above on learning logical formulas in continuous domains, it is shown that the learning of circuits can also be coupled with WMI. \\nWhat about countably infinite domains? In most pragmatic instances of learning logical artifacts, the difference between the uncountable and countably infinite setting is this: in the former, we see finitely many real-valued samples as being drawn from an (unknown) interval, and we could inspect these samples to crudely infer a lower and upper bound. In the latter, based on finitely many relational atoms, we would need to infer a universally quantified clause, such as \\\\( \\\\forall x,y,z ({\\\\it parent}(x,y) \\\\land {\\\\it parent}(y,z) \\\\supset {\\\\it grandparent}(x,z)) \\\\). If we are after a hypothesis that is simply guaranteed to be consistent wrt the observed examples, then standard rule induction strategies would suffice , and we could interpret the rules as quantifying over a countably infinite domain. But this is somewhat unsatisfactory, as there is no distinction between the rules learned in the standard finite setting and its supposed applicability to the infinite setting. What is really needed is an analysis of what rule learning would mean wrt the infinitely many examples that have \\\\emph{not} been observed. This was recently considered via the PAC-semantics in , by appealing to ideas on reasoning with open universes discussed earlier . \\nBefore concluding this section, it is worth noting that although the above discussion is primarily related to the learning of logical artifacts, it can equivalently be seen as a class of machine learning methods that leverage symbolic domain knowledge . \\nIndeed, logic-based probabilistic inference over deterministic constraints, and entailment-based induction augmented with background knowledge are instances of such a class. \\nAnalogously, the automated construction of relational and statistical knowledge bases by combining background knowledge with extracted tuples (obtained, for example, by applying natural language processing techniques to large textual data) is another instance of such a class. \\nIn the next section, we will consider yet another way in which logical and symbolic artifacts can influence learning: we will see how such artifacts are useful to enable tractability, correctness, modularity and compositionality.\",\n \"id\": \"a715b355-5229-4cca-872a-673dcade22b2\",\n \"level\": \"section\",\n \"origin_cites_number\": 26,\n \"parent_id\": \"3179fec8-9fa2-4590-ac34-0d3a7530e454\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Symbolic Logic meets Machine Learning: \\\\\\\\ A Brief Survey in Infinite Domains\"\n ],\n [\n \"section\",\n \"Machine Learning for Logic\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Machine Learning for Logic\"\n },\n {\n \"cite_extract_rate\": 0.268292682926829,\n \"cites\": [\n 1613,\n 3742,\n 3743,\n 3738,\n 3735,\n 3744,\n 8686,\n 3727,\n 3740,\n 3739,\n 3741\n ],\n \"content\": \"\\\\label{sec:logic_for_machine_learning}\\nThere are two obvious ways in which a logical framework can provide insights on machine learning theory. First, consider that computational tractability is of central concern when applying logic in computer science, knowledge representation, database theory and search . Thus, the natural question to wonder is whether these ideas would carry over to probabilistic machine learning. On the one hand, probabilistic extensions to tractable knowledge representation frameworks could be considered . But on the other, as discussed previously, ideas from knowledge compilation, and the use of circuits, in particular, are proving very effective for designing tractable paradigms for machine learning. While there has always been an interest in capturing tractable distributions by means of low tree-width models , knowledge compilation has provided a way to also represent high tree-width models and enable exact inference for a range of queries . See for a comprehensive view on the use of knowledge compilation for machine learning. \\nThe other obvious way logic can provide insights on machine learning theory is by offering a formal apparatus to reason about \\\\textit{context}. \\nMachine learning problems are often positioned as atomic tasks, such as a classification task where regions of images need to be labeled as cats or dogs. However, even in that limited context, we imagine the resulting classification system as being deployed as part of a larger system, which includes various modules that communicate or interface with the classification system. We imagine an implicit accountability to the labelling task in that the detected object is either a cat or a dog, but not both. If there is information available that all the entities surrounding the object of interest have been labelled as lions, we would want to accord a high probability to the object being a cat, possibly a wild cat. There is a very low chance of the object being a dog, then. If this is part of a vision system on a robot, we should ensure that the robot never tramples on the object, regardless of whether it is a type of cat or a dog. To inspect such patterns, and provide meta-theory for machine learning, it can be shown that symbolic, programmatic and logical artifacts are enormously useful. We will specifically consider correctness, modularity and compositionality to explore the claim.\\nOn the topic of correctness, the classical framework in computer science is \\\\emph{verification}: can we provide a formal specification of what is desired, and \\ncan the system be checked against that specification? In a machine learning context, we might ask whether the system, during or after training, satisfies a specification. \\nThe specification here might mean constraints about the physical laws of the domain, or notions of perturbation in the input space while ensuring that the labels do not change, or insisting that the prediction does not label an object as being both a cat and a dog, or otherwise ensuring that outcomes are not subject to, say, gender bias. Although there is a broad body of work on such issues, touching more generally on \\\\textit{trust} , we discuss approaches closer to the thrust of this article. For example, show that a trained neural network can be verified by means of an SMT encoding of the network. In recent work, show that the loss function of deep learning systems can be adjusted to logical constraints by insisting that the distribution on the predictions is proportional to the weighted model count of those constraints. In , prior (logical) constraints are compiled to a circuit to be used for probability estimation. \\nIn , circuits are shown to be amenable to training against probabilistic and causal prior constraints, including assertions about fairness, for example. \\nIn , a somewhat different approach to respecting domain constraints is taken: the low-level prediction is obtained as usual from a machine learning module, which is then interfaced with a probabilistic relational language and its symbolic engine. That is, the reasoning is positioned to be tackled directly by the symbolic engine. In a sense, such approaches cut across the three strands: the symbolic engine uses weighted model counting, the formulas in the language could be obtained by (say) entailment-based scoring, and the resulting language supports modularity and compositionality (discussed below). \\nWhile there is not much to be said about the distinction between finite vs infinite wrt correctness, many of these ideas are likely amenable to extensions to an infinite setting in the ways discussed in the previous sections (e.g., considering constraints of a continuous or a countably infinite nature). \\nOn the topic of modularity, recall that the general idea is to reduce, simplify or otherwise abstract a (probabilistic) computation as an atomic entity, which is then to be referenced in another, possibly more complex, entity. In standard programming languages, this might mean the compartmentalization and interrelation of computational entities. For machine learning, approaches such as probabilistic programming support probabilistic primitives in the language, with the intention of making learning modules re-usable and modular. It can be shown, for example, that the computational semantics of some of these languages reduce to WMC . Thus, in the infinite case, a corresponding reduction to WMI follows . \\nA second dimension to modularity is the notion of \\\\emph{abstraction}. Here, we seek to model, reason and explain the behavior of systems in a more tractable search space, by omitting irrelevant details. The idea is widely used in natural and social sciences. Think of understanding the political dynamics of elections by studying micro level phenomena (say, voter grievances in counties) versus macro level events (e.g., television advertisements, gerrymandering). In particular, in computer science, it is often understood as the process of mapping one representation onto a simpler representation by suppressing irrelevant information. \\nIn fact, integrating low-level behavior with high-level reasoning, exploiting relational representations to reduce the number of inference computations, and many other search space reduction techniques can all loosely be seen as instances of abstraction . \\nWhile there has been significant work on abstraction in deterministic systems , for machine learning, however, a probabilistic variant is clearly needed. In , an account of abstraction for loop-free propositional probabilistic programs is provided, where certain parts of the program (possibly involving continuous properties) can be reduced to a Bernoulli random variable. For example, suppose every occurrence of the continuous random variable $x$, drawn uniformly on the interval [0,1], in a program is either of the form $x\\\\leq 7$ or of the form $x\\\\gt 7$. Then, we could use a discrete random variable $b$ with a 0.7 probability of being true to capture $x\\\\leq 7$; and analogously, $\\\\neg b$ to capture $x\\\\gt 7$. The resulting program is likely to be simpler. \\nIn , an account of abstraction for probabilistic relational models is considered, where the notion of abstraction also extends to deterministic constraints and complex formulas. For example, a single probabilistic variable in the abstracted model could denote a complex logical formula in the original model. Moreover, the logical properties that enable verifying and inducing abstractions are also considered, and it is shown how WMC is sufficient for the computability of these properties (also see ). \\nIncidentally, abstraction brings to light a reduction between finite vs infinite: it is shown in that the modelling of piecewise densities as weighted propositions, which is leveraged in WMI , is a simple case of the more general account. Therefore, it is worthwhile to investigate whether this or other accounts of abstraction could emerge as general-purpose tools that allow us to inspect the conditions under which infinitary statements reduce to finite computations. \\nA broader point here is the role abstraction might play in generating explanations . For example, a user's understanding of the domain is likely to be different from the low-level data that a machine learning system interfaces with~, and so, abstractions can capture these two levels in a formal way. \\nFinally, we turn to the topic of compositionality, which, of course, is closely related to modularity in that we want to distinct modules to come together to form a complex composition. \\nNot surprisingly, this is of great concern in AI, as it is widely acknowledged that most AI systems will involve heterogeneous components, some of which may involve learning from data, and others reasoning, search and symbol manipulation . In continuation with the above discussion, probabilistic programming is one such endeavor that purports to tackle this challenge by allowing modular components to be composed over programming and/or logical connectives . (See for ideas in deterministic systems.) However, probabilistic programming only composes probabilistic computations, but does not offer an obvious means to capture other types of search-based computations, such as SAT, and integer and convex programming. \\nRecall that the computational semantics of probabilistic programs reduces to WMC . Following works such as , an interesting observation made in is that by appealing to a sum of products computation over different semiring structures, we can realize a large number of tasks such as satisfiability, unweighted model counting, sensitivity analysis, gradient computations, in addition to WMC. \\nIt was then shown in that the idea could be generalized further for infinite domains: by defining a measure on first-order models, WMI and convex optimization can also be captured. As the underlying language is a logical one, composition can already be defined using logical connectives. But an additional, more involved, notion of composition is also proposed, where a sum of products over different semirings can be concatenated. To reiterate, the general idea behind these proposals \\nis to arrive at a principled paradigm that allows us to interface learned modules with other types of search and optimization computations for the compositional building of AI systems. See also for analogous discussions, but where a different type of coupling for the underlying computations is suggested. Overall, we observed that a formal apparatus (symbolic, programmatic and logical artifacts) help us define such compositional constructions by providing a meta-theory.\",\n \"id\": \"66dfe73d-6adf-4a8f-bcd1-7c3121a3fbec\",\n \"level\": \"section\",\n \"origin_cites_number\": 41,\n \"parent_id\": \"3179fec8-9fa2-4590-ac34-0d3a7530e454\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Symbolic Logic meets Machine Learning: \\\\\\\\ A Brief Survey in Infinite Domains\"\n ],\n [\n \"section\",\n \"Logic for Machine Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Logic for Machine Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclusions}\\nIn this article, we surveyed work that provides further evidence for the connections between logic and learning. \\nOur narrative was structured in terms of three strands: logic versus learning, machine learning for logic, and logic for machine learning, but naturally, there was considerable overlap. \\nWe covered a large body of work on what these connections look like, including, for example, pragmatic concerns such as the use of hard, domain-specific constraints and background knowledge, all of which considerably eases the requirement that all of the agent's knowledge should be derived from observations alone. (See discussions in on the limitations of learned behavior, for example.) \\nWhere applicable, we placed an emphasis on how extensions to infinite domains are possible. In the very least, logical artifacts can help in constraining, simplifying and/or composing machine learning entities, and in providing a principled way to study the underlying representational and computational issues. \\nIn general, this type of work could help us move beyond the narrow focus of the current learning literature so as to deal with time, space, abstraction, causality, quantified generalizations, relational abstractions, unknown domains, unforeseen examples, among other things, in a principled fashion. In fact, what is being advocated is the tackling of problems that symbolic logic and machine learning might struggle to address individually. One could even think of the need for a recursive combination of strands 2 and 3: purely reactive components interact with purely cogitative elements, but then those reactive components are learned against domain constraints, and the cogitative elements are induced from data, and so on. More broadly, making progress towards a formal realization of \\\\emph{System 1} versus \\\\emph{System 2} processing might also contribute to our understanding of human intelligence, or at least capture human-like intelligence in automated systems. \\n\\\\bibliographystyle{abbrv}\\n\\\\bibliography{group}\\n\\\\end{document}\",\n \"id\": \"1cda66fd-50ab-40fa-8bc2-8a5604130ee9\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"3179fec8-9fa2-4590-ac34-0d3a7530e454\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Symbolic Logic meets Machine Learning: \\\\\\\\ A Brief Survey in Infinite Domains\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":70,"string":"70"},"all_cites":{"kind":"list like","value":[1177,3728,3726,8475,3725,3727,7747,3732,3729,3731,8683,8684,3730,8685,3737,3735,3736,8686,3733,3734,1613,3742,3743,3738,3744,3740,3739,3741],"string":"[\n 1177,\n 3728,\n 3726,\n 8475,\n 3725,\n 3727,\n 7747,\n 3732,\n 3729,\n 3731,\n 8683,\n 8684,\n 3730,\n 8685,\n 3737,\n 3735,\n 3736,\n 8686,\n 3733,\n 3734,\n 1613,\n 3742,\n 3743,\n 3738,\n 3744,\n 3740,\n 3739,\n 3741\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.403267899091801,"string":"1.403268"}}},{"rowIdx":35,"cells":{"authors":{"kind":"list like","value":["Zhuang Li","Lizhen Qu","Gholamreza Haffari"],"string":"[\n \"Zhuang Li\",\n \"Lizhen Qu\",\n \"Gholamreza Haffari\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Context Dependent Semantic Parsing: A Survey"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-11-02T07:51:05Z","string":"2020-11-02T07:51:05Z"},"category":{"kind":"string","value":"cs.CL"},"abstract":{"kind":"string","value":" Semantic parsing is the task of translating natural language utterances into machine-readable meaning representations. Currently, most semantic parsing methods are not able to utilize contextual information (e.g. dialogue and comments history), which has a great potential to boost semantic parsing performance. To address this issue, context dependent semantic parsing has recently drawn a lot of attention. In this survey, we investigate progress on the methods for the context dependent semantic parsing, together with the current datasets and tasks. We then point out open problems and challenges for future research in this area. The collected resources for this topic are available at: \\url{https://github.com/zhuang-li/Contextual-Semantic-Parsing-Paper-List}. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"321e790e-018a-41e3-8473-e7e9883a1392","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"]],"subsections":["b8456ffc-88d1-4dce-9af7-5cb66e55fd57","a541cf51-c35e-4bd7-b0ec-f7b7038dbad0","0888dee2-d5ac-4bd8-807b-f7439f676fed","37d74331-727a-46c3-844f-3da36f456e56","117c71f7-624f-4f92-9105-d8b545d244ba"],"title":"root"},{"cite_extract_rate":0.7142857142857141,"cites":[6899,6900,6898,6897,6896],"content":"\\blfootnote{\n \\hspace{-0.65cm} \n This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: \\url{http://creativecommons.org/licenses/by/4.0/}.\n}\nSemantic parsing is concerned with mapping natural language (NL) utterances into machine-readable structured \\textit{meaning representations} (\\MRs). These representations are in the formats of formal languages, e.g. Prolog, SQL, and Python. A formal language is typically defined by means of a formal \\textit{grammar}, which consists of a set of rules. Following the convention of the chosen formal language, \\MRs are also referred to as logical forms or programs. An \\MR is often executable in a (programming) environment to yield a result (e.g. results of SQL queries) enabling automated reasoning~.\nMost research work on semantic parsing treats each NL utterance as an \\emph{independent} input, ignoring the text surrounding them~, such as interaction histories in dialogues. \nThe surrounding text varies significantly across different application scenarios. In a piece of free text, we refer to the surrounding text of a current utterance as its \\textit{context}. The context is different with respect to different utterances. In our sequel, we differentiate between context \\textit{independent} semantic parsing (\\CiSP) and context \\textit{dependent} semantic parsing (\\CdSP) by whether a corresponding parser utilizes context information. A knowledge base or a database (on which a \\MR is executed for the purpose of question answering) can be considered as context as well~.\nThis type of context does not change with respect to the utterances. In this survey, we only consider the former kind of context which does vary with different utterances.\n\\input{tab-example-pets.tex}\nThe utilization of context in semantic parsing imposes both challenges and opportunities. \nAs shown in Table \\ref{tab:CoSQL}, one challenge is to resolve references, such as \\textit{those} in ``For each of those, what is the maximum age''. This example shows also another challenge caused by elliptical (incomplete) utterances. The sentence ``What about the average age?'' alone misses information about the database table and the column \\textit{pettype}. The incomplete meaning needs to be complemented by the discourse context. \nCompared with \\CiSP, which usually assumes that the information within the utterance is complete, \\CdSP is expected to tackle challenges posed by involving context in the parsing process \n. In addition, tackling the above challenges provides us with more opportunities to inspect the linguistic phenomena which could influence semantic parsing. Our survey on \\CdSP fills the gap in the literature, as the recent surveys in the semantic parsing research mainly focus on \\CiSP~. \nThis paper is organised as follows. We start with providing a brief and fundamental understanding of \\CiSP in \\S2. We then present a comprehensive organization of the recent advances in \\CdSP in \\S3. We discuss current \\CiSP tasks, datasets, and resources in \\S4. Finally, we cover open research problems in \\S5, and conclude by providing a roadmap for future research in this area.","id":"b8456ffc-88d1-4dce-9af7-5cb66e55fd57","level":"section","origin_cites_number":7,"parent_id":"321e790e-018a-41e3-8473-e7e9883a1392","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0.5,"cites":[6900],"content":"\\CiSP aims to learn a mapping $\\pi_{\\theta} : \\mathcal{X} \\rightarrow \\mathcal{Y}$, which translates an NL utterance $x \\in \\mathcal{X}$ into an \\MR $y \\in \\mathcal{Y}$. An \\MR $y$ can be executed in a programming environment (e.g. databases, knowledge graphs, etc.) to yield a result $z$, namely denotation. The structure of an \\MR takes a form of either a tree or graph, depending on its underlying formal language. The languages of \\MRs are categorized into three types of formalism\n: logic based (e.g. first order logic), graph based (e.g. AMR~), and programming languages (e.g. Java, Python)~. Some semantic parsers explicitly apply a production grammar to yield \\MRs from utterances. Such a grammar consists of a set of production rules, which define a list of candidate derivations for each NL utterance. Each derivation deterministically produces a grammatically valid \\MR.","id":"a541cf51-c35e-4bd7-b0ec-f7b7038dbad0","level":"section","origin_cites_number":2,"parent_id":"321e790e-018a-41e3-8473-e7e9883a1392","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Background"]],"subsections":["d1388aa1-f859-436a-97f9-1580b1fa7dad","0accbdad-eb7e-426a-a0ca-4f0b3266d1fb"],"title":"Background"},{"cite_extract_rate":1,"cites":[6900],"content":"Given an utterance $x \\in \\mathcal{X}$ and its paired \\MR $y \\in \\mathcal{Y}$, a \\CiSP model can form a \\textit{conditional} distribution $p(y | x)$.\nThe model learning can be supervised by either utterance-\\MR pairs or merely utterance-denotation pairs. If only denotations are available, a widely used approach~ is to marginalize over all possible \\MRs for a denotation $z$, which leads to a \\textit{marginal} distribution $p(z | x) = \\sum_{y} p(z, y| x)$. A parsing algorithm aims to find the optimal \\MR in the combinatorially large search space.\nWe coarsely categorize the existing models into: symbolic approaches, neural approaches, and neural-symbolic approaches based on the category of machine learning methodology and whether any production grammars are explicitly used in models.","id":"d1388aa1-f859-436a-97f9-1580b1fa7dad","level":"subsection","origin_cites_number":1,"parent_id":"a541cf51-c35e-4bd7-b0ec-f7b7038dbad0","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Background"],["subsection","Semantic Parsing Models"]],"subsections":["e551a263-fa5f-4543-97d2-4b3bf2aa80fe"],"title":"Semantic Parsing Models"},{"cite_extract_rate":0.5,"cites":[6900,6897,6901],"content":"\\label{sec:pipe}\nA symbolic semantic parser employs production grammars to generate candidate derivations and find the most probable one via a scoring model. The scoring model is a statistical or machine learning model. \n Each derivation is represented by handcrafted features extracted from utterances or partial \\MRs. Let $\\Phi(x, d)$ denote the features of a pair of utterance and derivation, and $G(x)$ be the set of candidate derivations based on $x$. A widely used scoring model is the log linear model~. \n\\begin{equation}\n p( d | x) = \\frac{\\exp(\\bm{\\theta}\\Phi(x, d))}{\\sum_{d' \\in G(x)} \\exp(\\bm{\\theta}\\Phi(x, d'))}\n\\end{equation}\nwhere $\\bm{\\theta}$ denotes the model parameters. If only utterance-denotation pairs are provided at training time, a model marginalizes over all possible derivations yielding the same denotations by $p(z | x) = \\sum_{d} p(z, d| x)$~. Those corresponding parsers further differentiate between graph-based parsers~ and shift-reduce parsers~ due to the adopted parsing algorithms and the ways to generate derivations. From a machine learning perspective, these approaches are also linked to a structured prediction problem.","id":"e551a263-fa5f-4543-97d2-4b3bf2aa80fe","level":"paragraph","origin_cites_number":6,"parent_id":"d1388aa1-f859-436a-97f9-1580b1fa7dad","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Background"],["subsection","Semantic Parsing Models"],["paragraph","Symbolic Approaches"]],"subsections":["b4beea59-21ae-4c9c-befd-dc1675e2c2f8","f6a9856b-4e78-4007-961d-d1003c3a47c0"],"title":"Symbolic Approaches"},{"cite_extract_rate":1,"cites":[166,2401,6902,3605,38],"content":"\\label{sec:end}\nNeural approaches apply neural networks to translate NL utterances into \\MRs without using production grammars. These approaches formulate semantic parsing as a machine translation problem by viewing NL as the source language and the formal language of \\MRs as the target language. \nMost work in this category adopts \\Seq as the backbone architecture, which consists of an encoder and a decoder. The encoder projects NL utterances into hidden representations, whereas the decoder generates linearized \\MRs sequentially.\n Both encoders and decoders employ either recurrent neural networks (RNN)~ or Transformers . Note that, these methods do not apply any production grammars to filter out syntactically invalid \\MRs.\nThe variants of the \\Seq based models also explore structural information of \\MRs. \\textsc{Seq2Tree} utilizes a tree-structured RNN as the decoder, which constrains generated \\MRs to take syntactically valid tree structures. The \\textsc{Coarse2Fine} model~ adopts a two-stage generation for the task. In the first stage, a \\Seq model is applied to generate \\MR templates, which replace entities in \\MRs by slot variables for a high-level generalization. In the second stage, another \\Seq model is applied to fill the slot variables with the corresponding entities.","id":"b4beea59-21ae-4c9c-befd-dc1675e2c2f8","level":"paragraph","origin_cites_number":5,"parent_id":"e551a263-fa5f-4543-97d2-4b3bf2aa80fe","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Background"],["subsection","Semantic Parsing Models"],["paragraph","Symbolic Approaches"],["paragraph","Neural Approaches"]],"subsections":[],"title":"Neural Approaches"},{"cite_extract_rate":1,"cites":[6904,9083,4279,6903],"content":"In order to ensure the generated \\MRs to be syntactically valid without compromising the generalization power of neural networks, neural-symbolic approaches fuse both symbolic and neural approaches by applying production grammars to the generated \\MRs; then the derivations are scored by neural networks. \nThe majority of these methods linearize derivations such that they are able to leverage \\Seq~. At each time step, the decoder of these methods emits either a parse action or a production rule, leading to a grammatically valid \\MR at the end. these works produce derivations by varying grammars. \\textsc{NSM} uses a subset of Lisp syntax. \\textsc{TranX}~ defines the grammars in Abstract Syntax Description Language, while \\textsc{IRNet} considers the context-free grammar of a language called SemQL. \nThere are also neural-symbolic approaches adopting neural architectures other than \\Seq. One of such examples is , which adopts a dynamic neural module network (DNMN) to generate \\MRs.","id":"f6a9856b-4e78-4007-961d-d1003c3a47c0","level":"paragraph","origin_cites_number":4,"parent_id":"e551a263-fa5f-4543-97d2-4b3bf2aa80fe","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Background"],["subsection","Semantic Parsing Models"],["paragraph","Symbolic Approaches"],["paragraph","Neural-Symbolic Approaches"]],"subsections":[],"title":"Neural-Symbolic Approaches"},{"cite_extract_rate":0,"cites":[],"content":"In semantic parsing, \\textit{exact match accuracy} is the most commonly used evaluation metric. With \\textit{exact match accuracy}, the parsing results are considered correct only when the output \\MR/denotations exactly match the string of the ground truth \\MR/denotations. One flaw of the evaluation metric is that some types of MRs (e.g., SQL) do not hold order constraints. \\newcite{yu2018spider} proposed a metric \\textit{set match accuracy} to evaluate the semantic parsing performance over SQLs, which treats each SQL statement as a set of clauses and ignore their orders.\nDue to the variety of domains and languages over different datasets, it is difficult to measure all semantic parsing methods in a unified framework. To address this issue, \\newcite{yu2018spider}, \\newcite{yu2019sparc} and \\newcite{yu2019cosql} built different shared-task platforms with leaderboard for semantic parsing evaluation on the common datasets and consistent evaluation metrics.","id":"0accbdad-eb7e-426a-a0ca-4f0b3266d1fb","level":"subsection","origin_cites_number":0,"parent_id":"a541cf51-c35e-4bd7-b0ec-f7b7038dbad0","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Background"],["subsection","Evaluation"]],"subsections":[],"title":"Evaluation"},{"cite_extract_rate":0.6666666666666661,"cites":[6905,6906],"content":"Context dependent semantic parsing involves modelling of context in the parsing process. For each current NL utterance, we define its \\textit{context} as the information beyond this utterance. With this definition, there are two types of context for semantic parsing, \\textit{local} context and \\textit{global} context. The \\textit{local} context for an utterance is the text and multimedia content surrounding it, which is meaningful only for this utterance. In plain texts, the concept of local context is also quite close to discourse, which is defined as a group of collocated, structured, coherent sentences~. In contrast, its \\textit{global} context is the information accessible to more than one utterance, including databases and external text corpora, images or class environment~. The content of local context varies for each NL utterance while the global context is always static. The work in our survey is only concerned with local context. Therefore, we always refer to \"local context\" as \"context\" in the following sections. \nContext provides additional information to resolve ambiguity and vagueness in current utterances. For semantic parsing, one type of ambiguity is caused by references in current utterances, which need to be resolved to previously mentioned objects and relations. References may include explicit or implicit lexical triggers, such as \\textit{those} in \"For each of those, ...\" in our introductory example (Table \\ref{tab:CoSQL}). Another ambiguity illustrated by the same example is resulted by ellipsis. The previous context provides constraints to restrict the scope of possible \\MRs indicated by current utterances. In addition, context provides information to disambiguate word senses and entities, and link them to knowledge bases to enable complex reasoning. However, semantic parsing literature largely neglects word sense disambiguation, which is regarded as an AI complete problem~. Last but not least, context allows to exploit discourse coherence for semantic parsing. Coherence relations characterize structural relationships between sentences, thus limit the search space of parse candidates for the following utterances of current ones.\nFormally, a context dependent parser takes both an input utterance $x_i$ and its context $C_i$, where $C_i$ could include a broad range of multimedia content. And we consider a group of inter-related utterances with the union set of their context as one \\textit{interaction}, $I = (\\mathbf{x}, \\mathbf{C})$, where $\\mathbf{x} = [x_1, ...,x_i,...,x_T]$ and $\\mathbf{C} = \\cup_{i=1}^{T}C_{i}$. Currently, most \\CdSP work focus on the research problems of context $C_i$ regarding the history utterances, \\MRs, denotations. Such a parser learns a mapping from a current utterance $x_{i}$ to an \\MR $y_i$ by $\\pi_{\\theta}(x_{i}, C_i)$.","id":"0888dee2-d5ac-4bd8-807b-f7439f676fed","level":"section","origin_cites_number":3,"parent_id":"321e790e-018a-41e3-8473-e7e9883a1392","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"]],"subsections":["a7325d38-7cf4-41b3-bf34-266fb7445dcc","4f2c4e8f-03a9-4d2a-875a-0025d4eaf479","a42c11dd-cba0-4190-ab4f-e20c9525e938","640e3315-aedb-4aa7-abfb-05ad7be7ec1b","6d868391-5ac7-479e-9463-c4e87430bb7a"],"title":"Context Dependent Semantic Parsing"},{"cite_extract_rate":0.5,"cites":[6907],"content":"Existing symbolic approaches formulate \\CdSP as a structured prediction problem by including contextual information into their feature models. Their models capture $ p( d_i | x_i, C_i)$ by including context as a condition. Both \\newcite{zettlemoyer2009learning} and \\newcite{srivastava2017parsing} divide the parsing process into two steps: i) generate initial parses using \\CiSP; ii) complete initial parses using contextual information. In contrast, \\newcite{long2016simpler} parses a sequence of utterances in one step. In all those work, symbolic features are used to represent contexts. \nIn two-step approaches, \\newcite{zettlemoyer2009learning} and \\newcite{srivastava2017parsing} differ in the details of individual steps. In the first step, \\newcite{zettlemoyer2009learning} extends \\MRs with predicates representing references, while \\newcite{srivastava2017parsing} generates a set of context independent parses for each utterance. In the second step, \\newcite{zettlemoyer2009learning} collects possible derivations by applying three heuristic rules to replace references with entities in context and extend initial \\LFs with constraints, then finds the best derivations according to a linear model. In~, their model expands the initial parse set with parses selected from context using heuristic rules, then finds the best parses in the expanded set. Their feature model includes a multinomial random variable indicating the current hidden state of discourse.\nThe shift-reduce parser in~ generates derivations for a whole utterance sequence. This method stores the previously generated derivations in a stack, performs a sequence of \\textit{shift} and \\textit{build} operations to generate \\LFs. In its feature model, a context is represented by a sequence of past \\LFs and a random variable denoting the current world state.\nUtterances and \\MRs histories form a context of a \\CdSP parser. The common practice is to extract handcrafted features from both utterances and \\MRs to represent contexts. Some typical feature patterns are as follows:","id":"a7325d38-7cf4-41b3-bf34-266fb7445dcc","level":"subsection","origin_cites_number":2,"parent_id":"0888dee2-d5ac-4bd8-807b-f7439f676fed","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Symbolic Approaches"]],"subsections":["6aaa20f2-9158-4b75-92a8-7292af7a0e72"],"title":"Symbolic Approaches"},{"cite_extract_rate":0,"cites":[],"content":"In~, they consider indicator features of lexical triggers, whether the current utterance is repeated, as well as the position of the current utterance in an interaction.","id":"6aaa20f2-9158-4b75-92a8-7292af7a0e72","level":"paragraph","origin_cites_number":1,"parent_id":"a7325d38-7cf4-41b3-bf34-266fb7445dcc","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Symbolic Approaches"],["paragraph","Utterance"]],"subsections":["4ff929d2-dc36-402d-8f59-7bed1be0ef5b"],"title":"Utterance"},{"cite_extract_rate":0,"cites":[],"content":"In , there is a feature indicating if the predicates exist in the history \\LFs. Such feature allows the model to learn to copy the segments from the context that contains the expected predicates. \\newcite{long2016simpler} adopts the feature indicating if the argument in current \\MR is one of arguments in the last \\MR. \\newcite{srivastava2017parsing} uses the combinations of predicates in successive turns as the indicator features.","id":"4ff929d2-dc36-402d-8f59-7bed1be0ef5b","level":"paragraph","origin_cites_number":1,"parent_id":"6aaa20f2-9158-4b75-92a8-7292af7a0e72","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Symbolic Approaches"],["paragraph","Utterance"],["paragraph","Meaning Representations"]],"subsections":[],"title":"Meaning Representations"},{"cite_extract_rate":1,"cites":[6908],"content":"\\begin{figure}\n \\centering\n \\includegraphics[width=1\\textwidth]{figs/SQL_example-cropped.pdf}\n \\label{fig:semantic_tree}\n \\caption{Coreference resolution architecture of . Considering the example in Table \\ref{tab:CoSQL}, \\newcite{chen2019context} firstly generates a \\MR template for $Q_2$ as \"SELECT max(petage), \\textit{REF} FROM pets GROUP BY \\textit{REF}\". The \\textit{REF} tokens would then be replaced with the \"pettype\" from the precedent \\MR. }\n \\label{fig:sql_example}\n\\end{figure}\nExisting neural \\CdSP methods extend the \\Seq architecture to incorporate contextual information in two ways. The first approach is to build context-aware encoders to encode historical utterances or \\MRs into neural representations, which provide decoders contextual information to resolve ambiguity in current utterances. As previously predicted \\MRs provide the constraints and information missed in current utterances, the second approach is to utilize context-aware decoders to reuse or revise those predicted \\MRs for generating current \\MRs.","id":"4f2c4e8f-03a9-4d2a-875a-0025d4eaf479","level":"subsection","origin_cites_number":1,"parent_id":"0888dee2-d5ac-4bd8-807b-f7439f676fed","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Neural Approaches"]],"subsections":["28692668-fe8c-48f8-82c5-52ec47621cd5"],"title":"Neural Approaches"},{"cite_extract_rate":0.818181818181818,"cites":[6899,6898,6909,1888,6911,6910,6896,168,1883],"content":"Encoders of \\CdSP methods differentiate between utterance encoders and \\MR encoders. Utterance encoders construct neural representations for both current and historical utterances, while \\MR encoders build neural representations based on on historical \\MRs.\nUtterance encoders aim to embed rich information hidden in utterances into fixed-length representations, which provide contextual information in addition to current utterances for decoders. They apply first an RNN to map each utterance into a continuous vector of fixed-size. Then there are three ways to encode utterances in context into a fixed-size neural representation.\n\\begin{itemize}\n \\item \n For each utterance in a dialogue, a straightforward method is to concatenate its previous $k - 1$ utterances with current utterance in order and encode them with the RNN~.\n As a result, decoders have access to information in at most $k$ utterances. However, this method fails to access information beyond the $k$ utterances. In addition, it is computationally expensive because if an utterance belongs to multiple contexts, it would be repeatedly encoded for modelling all the contexts.\n \\item To overcome the above weakness, an alternative method is to treat a sub-sequence of utterances up to time $t$ as a sequence of vectors, and project them into a \\textit{discourse state} vector by using a turn-level RNN~. In another word, those models apply hierarchical RNNs to map each context into a fixed-size vector. In this method, each utterance is encoded only once and reused for modelling different contexts. However, this approach often leads to significant information loss~ due to the challenges imposed by encoding sequences of utterances into single vectors. \n \\item In order to focus on history utterances most relevant to current decoder states or utterances, soft attention~ is applied to construct context vectors. The query vectors are either the hidden state of an decoder~ or an utterance vector~. To differentiate between positional information, token embeddings of history utterance are concatenated with their position embeddings~, which encode the positions of history utterances relative to the current utterances. This method reflects the observation that similar utterances tend to share relevant information, such as references of the same entities. Both discourse states and attended representations are also widely used by the neural dialogue models~, thus suffer from the same problems caused by composition complexity. As a result, the trained models are found insensitive to utterance order and word order in context~.\n\\end{itemize}\n\\MR encoders construct a neural context representation at time $t$ based on the \\MRs predicted before $t$. As \\MRs are expressed in a formal language, \\MR encoders also apply RNNs to encode each \\MR or segments of \\MRs into embedding vectors. Then \\MR encoders build context representations of historical \\MRs in the same spirit as utterance encoders. In~, they only concatenate the embeddings of $k$ most recent history \\MR tokens as they assume current \\MR is always an extension of previous \\MRs. In~, a bidirectional RNN is applied to construct a vector for each segment, which is extracted from historical \\MRs. Soft attention is also applied in~ for building context vectors, which uses the current hidden state of their decoder as the query vector to attend over the token embeddings of the previous \\MR.","id":"28692668-fe8c-48f8-82c5-52ec47621cd5","level":"paragraph","origin_cites_number":11,"parent_id":"4f2c4e8f-03a9-4d2a-875a-0025d4eaf479","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Neural Approaches"],["paragraph","Context-aware Encoders"]],"subsections":["8186c75f-b8c3-4431-b509-a482cbf52d9e"],"title":"Context-aware Encoders"},{"cite_extract_rate":1,"cites":[6899,6898,167,1888,6908,6896,6912],"content":"Decoders in \\CdSP models produce \\MRs based on the neural representations provided by their encoders. Such a decoder yields an \\MR by generating a sequence of \\MR tokens according to model distribution $P(\\mathbf{y}| \\mathbf{x}, \\mathbf{C})$, where $\\mathbf{C}$ denotes context information. There are three major ways to utilize context information.\nOne key problem of \\CdSP is incomplete information in current utterances. The straightforward way is to take neural context representations $\\mathbf{C}$ as additional input of decoders, which are yielded by context-aware encoders. Those context representations contains information from previous utterances, historical \\MRs, or both. The decoders take them as input by concatenating them with the ones from current utterances at each decoding step~. Thus, the quality of decoding depends tightly on the quality of contextual encoding, which is still a challenging problem~.\n\\MRs of current utterances often contain segments from previous \\MRs~. The shared parts are references to previously mentioned entities or constraints implied by context. Reuse of \\MR segments is realized by a designated \\textit{copy} component, which selects a segment to copy when the probability of copying is high. As decoders in \\Seq produce a sequence of decisions for each input, the corresponding model generates a sequence of mixed decisions, including both \\textit{copy} of segments and generation of new \\MR tokens. In a similar manner, copying of \\MR tokens from previous \\MR is proposed in~.\nCoreference resolution is explicitly addressed in~\\newcite{chen2019context}. As illustrated by the example in Figure \\ref{fig:sql_example}, a special token \\textit{REF} is introduced in the output vocabulary for denoting if an entity in the preceding \\MR is referred in that utterance. If that is the case, the corresponding entity token is copied from the previous \\MR to replace the \\textit{REF} token via a pointer network module~.","id":"8186c75f-b8c3-4431-b509-a482cbf52d9e","level":"paragraph","origin_cites_number":7,"parent_id":"28692668-fe8c-48f8-82c5-52ec47621cd5","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Neural Approaches"],["paragraph","Context-aware Encoders"],["paragraph","Context-aware Decoders"]],"subsections":[],"title":"Context-aware Decoders"},{"cite_extract_rate":0.6666666666666661,"cites":[6898,6912,6913,7744],"content":"\\begin{figure}\n \\centering\n \\includegraphics[width=1\\textwidth]{figs/copy_action.pdf}\n \\caption{The symbolic memory architecture of . Considering the example in Table \\ref{tab:CoSQL}, \\newcite{guo2018dialog} defines different types of actions, $A_{ac}$ and $A_{ec}$, to copy action sequence $A_{10}, A_{12}, A_{13}$ and the entity \\textit{pettype} from the symbolic memory, respectively.}\n \\label{fig:copy_action}\n\\end{figure}\nNeural-symbolic approaches introduce grammar into the decoding process or utilize symbolic representations as intermediate representations, while applying neural nets for representation learning. They take advantages from both the good context representation obtained by neural nets and reduced complexity of decoding due to the constraints introduced by grammars. In existing work, those approaches regard the generation of an \\MR as the prediction of a sequence of actions. Neural-symbolic methods normally take the same methods as the neural approaches to encode the contextual information. What differentiate them is the neural-symbolic could handle context by i) designing specific actions, and ii) utilizing symbolic context representations.\nThe context specific actions proposed in~ adopt \\textit{copy} mechanism to reuse the previous \\MRs. \\textsc{CAMP}~ include three actions to copy three different SQL clauses from precedent queries. \n\\newcite{liu2020FarAwayContextModelingSP} allows copying of any actions or subtrees from precedent SQL queries. The \\textit{subsequent} action in~ adds SQL conditions from the previous query into the current semantic parse to address the ellipsis problem. Different from other approaches, \\newcite{iyyer2017search} uses a \\textsc{DynSP}, which is in a similar neural network structure as the DNMN, instead of the \\Seq to generate the action sequences.\nProduction rules are also used to explicitly address the coreference resolution. In~, the authors defined fours actions to instantiate the entities, predicates, types and numbers. Then the pointer network is utilized to find mentions of the four entry semantic categories in the current and history utterances. The entities in utterances are later mapped to entities in knowledge bases by using their entity linking tool.\nInstead of directly copying from previous \\MRs, the parser \\textsc{Dialog2Action}~ incorporates a dialogue memory, which maintains \\textit{symbolic} representations of entities, predicates and action subsequences from an interaction history (Figure \\ref{fig:copy_action}). That parser defines three types of designated actions to copy entities, predicates and action subsequences from the memory respectively. Instead of decisively copying from memory, each type of action probabilistically selects the corresponding segments conditioning on the \\textit{symbolic} representations, which are later integrated into the generated action sequences. \n\\newcite{guo2019coupling} employs the same neural-symbolic models as in to capture contextual information. Different from other approaches, \\newcite{guo2019coupling} adopts the meta-learning approach to improve the generation ability of \\CdSP models. Inspired by , \\newcite{guo2019coupling} utilize the context from other interactions to guide the learning of \\CdSP over utterances within current interactions via MAML. \\newcite{guo2019coupling} considers an input utterance $x_i$ and its context $C_i$ as an instance. A context-aware retriever would retrieve instances, which are semantically close to the current instances, from other interactions. When learning model parameters, the retrieved instances and the current instances are considered as the support set and test set, respectively, and grouped as tasks as in the common MAML paradigm.","id":"a42c11dd-cba0-4190-ab4f-e20c9525e938","level":"subsection","origin_cites_number":6,"parent_id":"0888dee2-d5ac-4bd8-807b-f7439f676fed","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Neural-Symbolic Approaches"]],"subsections":[],"title":"Neural-Symbolic Approaches"},{"cite_extract_rate":1,"cites":[6898],"content":"In~, 13 different context modeling methods for both neural and neural-symbolic \\CdSP parsers were evaluated on two benchmark datasets. None of those methods achieve consistent superior results over the others in all experimental settings. Among them, concatenation of $k$ recent utterances for decoders and copy of parse actions from precedent \\MRs are the top performing ones in most settings. ~\\newcite{liu2020FarAwayContextModelingSP} defines 12 fined-grained types summarized with multiple hierarchies according to the contextual linguistic phenomena, and inspects how different linguistic phenomena influence the model behavior. One interesting conclusion is that the methods in their experiments all perform poorly on the instances involving coreference problems that require complex inference. But note that, those methods in that study were not compared with the ones with explicit coreference resolution. Another interesting finding is that all the models perform better on the utterances which only augment the semantics of previous sentences than on the utterances which substitute the partial semantics of the precedent utterances.","id":"640e3315-aedb-4aa7-abfb-05ad7be7ec1b","level":"subsection","origin_cites_number":1,"parent_id":"0888dee2-d5ac-4bd8-807b-f7439f676fed","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Comparison between Different \\CdSP Approaches"]],"subsections":[],"title":"Comparison between Different \\CdSP Approaches"},{"cite_extract_rate":1,"cites":[6899,6914,3063,6916,6915],"content":"Feedback/Interactive Semantic Parsing is another line of research in semantic parsing that utilizes context to refine \\MRs in an iterative manner. Most Feedback Semantic Parsing systems~ start with using an \\CiSP parser to parse a given utterance into an initial \\MR. Then the \\MR is interpreted in natural language and sent to a user. The user provides feedback, based on which the systems revise the initial parse. The process repeats till convergence. Therefore, in Feedback Semantic Parsing, interaction histories are only used to revise the parses. In contrast, \\CdSP focuses on modelling the dependencies between the utterances. \\newcite{elgohary2020speak} empirically compares \\CdSP with Feedback Semantic Parsing. They train a \\CdSP model, EditSQL~, on two \\CdSP datasets, \\Sparc and \\Cosql, and evaluate it on the test set of a feedback semantic parsing dataset, \\SPLASH. The performance is merely 3.4\\% and 3.2\\% in terms of accuracy, indicating that the two tasks are distinct by addressing different aspects of context.","id":"6d868391-5ac7-479e-9463-c4e87430bb7a","level":"subsection","origin_cites_number":5,"parent_id":"0888dee2-d5ac-4bd8-807b-f7439f676fed","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Context Dependent Semantic Parsing"],["subsection","Comparison between \\CdSP and Feedback Semantic Parsing"]],"subsections":[],"title":"Comparison between \\CdSP and Feedback Semantic Parsing"},{"cite_extract_rate":0,"cites":[],"content":"\\input{tab-example-data.tex}\nTable \\ref{tab:context_data} summarizes the basic properties and statistics of existing \\CdSP datasets. There are two scenarios of the \\CdSP datasets, Single-party Scenarios and Multi-party Scenarios. In the former scenarios, the user utterances are translated into \\MRs to obtain the execution results from the programming environment. In the latter scenarios, there are systems which respond to the users in natural language based on the user utterances and the execution results.\nThe user utterances are manually labeled with different types of annotations, including \\MRs, denotations, and dialogue acts. The system responses are usually annotated with the dialogue acts. We especially highlight those annotations that explicitly reflect contextual dependencies of utterances in the sequel.","id":"37d74331-727a-46c3-844f-3da36f456e56","level":"section","origin_cites_number":0,"parent_id":"321e790e-018a-41e3-8473-e7e9883a1392","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Datasets and Resources"]],"subsections":["f2b8d82a-6953-4c79-b18d-ac672c464017","f428c8ab-b908-460e-bb23-7a56d331e087"],"title":"Datasets and Resources"},{"cite_extract_rate":0,"cites":[],"content":"","id":"f2b8d82a-6953-4c79-b18d-ac672c464017","level":"subsection","origin_cites_number":0,"parent_id":"37d74331-727a-46c3-844f-3da36f456e56","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Datasets and Resources"],["subsection","Scenarios"]],"subsections":["1b5a8458-fd5b-4876-ba50-10c54b23220a"],"title":"Scenarios"},{"cite_extract_rate":0.44444444444444403,"cites":[3063,8303,6908,6907],"content":"In \\Sparc~, \\Sqa~ and \\Atis, the user utterances within each interaction are around a topic described by the provided text. To collect \\Sparc and \\Sqa, crowd-workers are asked to raise questions to obtain the information that answers the questions sampled from other corpora~. But the assumption for \\Sqa is the answers of the current question must be the subset of answers from the last turn. In \\Atis~, crowd-workers raise questions around the detailed scripts describing air travel planning scenarios. \n\\TemporalStructure~ and \\TimeExp~ particularly focused on addressing the temporal-related dependency. In \\TemporalStructure, human users or the simulators raise natural language questions chronologically towards a knowledge base. The facts in the knowledge base are organized in time series. Therefore, the questions in \\TemporalStructure are rich with time expressions. \\TimeExp only annotate temporal mentions (text segments that describe time expressions) instead of complete questions. All the mentions are from the time expression-rich corpora.\n\\Scone~ and \\CSQA~ use semi-automatic approaches to simulate the contextual dependency. Each interaction in \\Scone is merely labeled with an initial denotation and an end denotation. The denotations in \\Scone are regarded as the states that can be manipulated by the programs. Within each interaction, multiple candidate sequences of programs would be automatically generated while only the sequence of programs, which could correctly transit the initial state to the end state, would be kept and described with natural language by the crowd-workers. To create \\CSQA dataset, the crowd-workers are asked to raise questions that can be answered from single fact tuples (e.g. relation: \\textit{CEO}, subject: \\textit{Google}, object: \\textit{Sundar Pichai}) in the knowledge graph or the complex facts which are the composition of multiple tuples. To create coherent dependency among questions, the questions that share the relations or entities are placed next to each other. And crowd-workers would manually modify the questions such that the sequence of questions would include contextual linguistic properties such as ambiguity, underspecification or coreference. It is worth mentioning that, with such method, \\CSQA includes the largest number of interactions until now, which is over 200k.","id":"1b5a8458-fd5b-4876-ba50-10c54b23220a","level":"paragraph","origin_cites_number":9,"parent_id":"f2b8d82a-6953-4c79-b18d-ac672c464017","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Datasets and Resources"],["subsection","Scenarios"],["paragraph","Single-party Scenarios"]],"subsections":["164c4a3e-7159-47d0-82af-d206060db598"],"title":"Single-party Scenarios"},{"cite_extract_rate":0,"cites":[],"content":"Similar to the scenario of \\Sparc, to obtain the answers to the questions sampled from \\Spider~, the conversations in \\Cosql~ are conducted between two human interlocutors, who play the roles of user and system, respectively. The dialogues in the \\SpaceBook~ are under the scenarios formed by the routing requests. One human interlocutor pretends to be a tourist walking around Edinburgh while another interlocutor plays the role of a system responding to the tourist. The conversations in \\EmailDiag are between the human agent and an email assistant instead of two humans.","id":"164c4a3e-7159-47d0-82af-d206060db598","level":"paragraph","origin_cites_number":3,"parent_id":"1b5a8458-fd5b-4876-ba50-10c54b23220a","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Datasets and Resources"],["subsection","Scenarios"],["paragraph","Single-party Scenarios"],["paragraph","Multi-party Scenarios"]],"subsections":[],"title":"Multi-party Scenarios"},{"cite_extract_rate":0,"cites":[],"content":"The contextual linguistic phenomena in the \\CdSP corpora is quite close to the phenomena in the corpora of tasks such as document-level machine translation, question answering, dialogue system, etc.. However, in \\CdSP datasets, the contextual linguistic phenomena has a tight relation with the annotations.\n\\newcite{iyyer2017search}, \\newcite{vlachos2014new} defined specific components in the \\MR languages of \\Sqa and \\SpaceBook to explicitly model the context dependency. \\Sqa introduced a keyword \\textit{subsequent}. All the answers of \\MR statements after \\textit{subsequent} would only be the subset of the answers of the precedent \\MR. In the language of \\SpaceBook, to resolve the coreference problem, a special predicate \\textit{equivalent} could indicate the identical entities across questions at different turns.\nThe context dependency could be reflected by some properties of annotations. \\newcite{yu2019sparc} analyzed semantic changes over turns in \\Sparc by calculating the overlapping percentage of tokens between the SQL annotations at different turns. In \\Sparc, the average overlapping percentage increases at later turns within one interaction, where the users tend to narrow down their topics with turns increasing. Both \\newcite{yu2019sparc} and \\newcite{liu2020FarAwayContextModelingSP} categorized the contextual phenomena in \\Sparc into fine-grained types and calculate their frequency. \\newcite{yu2019sparc} found some SQL representations correspond to certain contextual phenomena types. For instance, in the questions of the \\textit{theme-entity}, which means the current question and precedent question are around the same entities but request for different properties, their corresponding SQL representations have the same \\textit{FROM} and \\textit{WHERE} clauses. But the SQL representations for other types may vary.\nFor the datasets, \\SpaceBook and \\Cosql, \\newcite{yu2019cosql} and \\newcite{vlachos2014new} label utterances with dialogue acts along with \\MRs. Different from , \\newcite{vlachos2014new} integrated the dialogue acts into the \\MRs. The dialogue acts can be considered as the overall functions of the utterances while different dialogue acts reflect different properties of utterances. For example, in \\Cosql, the unanswerable questions that can not be parsed into SQLs are labelled with dialogue acts such as \\textit{NOT\\_RELATED}, \\textit{CANNOT\\_UNDERSTAND}, or \\textit{CANNOT\\_ANSWER}. The ambiguous questions that need to be clarified are labelled with \\textit{AMBIGUOUS}. The following questions are then labeled with \\textit{CLARIFY}. The dialogue acts could provide additional contextual information for \\CdSP.","id":"f428c8ab-b908-460e-bb23-7a56d331e087","level":"subsection","origin_cites_number":1,"parent_id":"37d74331-727a-46c3-844f-3da36f456e56","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Datasets and Resources"],["subsection","Context and Annotations"]],"subsections":[],"title":"Context and Annotations"},{"cite_extract_rate":0,"cites":[],"content":"\\CdSP distinct from \\CiSP by context modelling and utilization of context information in the parsing process to complete missing information in \\MRs. Despite significant progress in recent years, there are still multiple directions worth pursuing.","id":"117c71f7-624f-4f92-9105-d8b545d244ba","level":"section","origin_cites_number":0,"parent_id":"321e790e-018a-41e3-8473-e7e9883a1392","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Challenges and Future Directions"]],"subsections":["9f019292-b9cc-44f9-a0be-f8837ae153df"],"title":"Challenges and Future Directions"},{"cite_extract_rate":0,"cites":[],"content":"\\newcite{yu2018spider} and \\newcite{liu2020FarAwayContextModelingSP} analyzed the influence of different types of contextual information on \\CdSP methods. Despite some empirical results, it still lacks of a thorough understanding of pros and cons of each type of context in relation to the parsing task. For example, in which cases should parsers extract information from \\MRs in context instead of utterances? Apart from ellipsis and coreference resolution, are there other linguistically motivated problems in context the current parsers have not addressed yet?","id":"9f019292-b9cc-44f9-a0be-f8837ae153df","level":"paragraph","origin_cites_number":0,"parent_id":"117c71f7-624f-4f92-9105-d8b545d244ba","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Challenges and Future Directions"],["paragraph","Analysis of Linguistic Phenomena Benefiting from Context"]],"subsections":["73fd17e5-1eb7-4596-820a-f4fb6a6042a2"],"title":"Analysis of Linguistic Phenomena Benefiting from Context"},{"cite_extract_rate":0,"cites":[],"content":"Current \\CdSP approaches fall in the scope of near-side pragmatics, in particular reference resolution, and current \\CdSP datasets (e.g. \\Cosql, \\SpaceBook) consider dialogue acts as merely the overall function of the utterances~. However, \\textit{far-side pragmatics} focuses on what happens beyond saying, including implicatures and communicative intentions etc.. Incorporating far-side pragmatics in semantic parsing will be especially useful towards completely understanding dialogues. Thus, there is a need to create large corpora annotated with rich information about various aspects of pragmatics for both training and evaluation.","id":"73fd17e5-1eb7-4596-820a-f4fb6a6042a2","level":"paragraph","origin_cites_number":1,"parent_id":"9f019292-b9cc-44f9-a0be-f8837ae153df","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Challenges and Future Directions"],["paragraph","Analysis of Linguistic Phenomena Benefiting from Context"],["paragraph","Incorporating Far-side Pragmatics"]],"subsections":["e8038842-2355-4479-9735-294e34339301","f4cff5dd-1eff-4533-adfc-1fcd6b740b89"],"title":"Incorporating Far-side Pragmatics"},{"cite_extract_rate":0.75,"cites":[969,6898,3699],"content":"A key challenge of context based modelling is composition complexity caused by highly varying context. The empirical results in~ show that the SOTA models can capture well nearby context information but it is still challenging to capture long-range dependencies in context. One possible direction is to find out the underlying causal structure~, which should be sparse and explains well which contextual information leads to current utterances. If we can focus only on the key reasons in context that lead to changes of \\MRs, the influence from noisy information and overfitting of models is expected to decrease significantly. Another potential benefit of understanding causal structures in context is to improve robustness of parsers by ignoring non-robust features~.","id":"e8038842-2355-4479-9735-294e34339301","level":"paragraph","origin_cites_number":4,"parent_id":"73fd17e5-1eb7-4596-820a-f4fb6a6042a2","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Challenges and Future Directions"],["paragraph","Analysis of Linguistic Phenomena Benefiting from Context"],["paragraph","Incorporating Far-side Pragmatics"],["paragraph","Causal Structure Discovery in Context"]],"subsections":[],"title":"Causal Structure Discovery in Context"},{"cite_extract_rate":1,"cites":[6917],"content":"Since most \\CdSP datasets are small in terms of the number of utterances and interactions, the direction on addressing the low-resource problem in \\CdSP is quite promising. The meta-learning approaches, such as the MAML \\CdSP in , could be a potential direction to address this issue. The other typical methods to solve low-resource issues, including weakly supervision, data augmentation, semi-supervised learning, self-supervised learning etc., could be further investigated in the scenarios of \\CdSP.\n\\section*{Acknowledgements}\nWe thank Xuanli He and anonymous reviewers for their useful suggestions. \n\\bibliographystyle{coling}\n\\bibliography{coling2020}\n\\end{document}","id":"f4cff5dd-1eff-4533-adfc-1fcd6b740b89","level":"paragraph","origin_cites_number":1,"parent_id":"73fd17e5-1eb7-4596-820a-f4fb6a6042a2","prefix_titles":[["title","Context Dependent Semantic Parsing: A Survey"],["section","Challenges and Future Directions"],["paragraph","Analysis of Linguistic Phenomena Benefiting from Context"],["paragraph","Incorporating Far-side Pragmatics"],["paragraph","Low-resource \\CdSP"]],"subsections":[],"title":"Low-resource \\CdSP"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"321e790e-018a-41e3-8473-e7e9883a1392\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ]\n ],\n \"subsections\": [\n \"b8456ffc-88d1-4dce-9af7-5cb66e55fd57\",\n \"a541cf51-c35e-4bd7-b0ec-f7b7038dbad0\",\n \"0888dee2-d5ac-4bd8-807b-f7439f676fed\",\n \"37d74331-727a-46c3-844f-3da36f456e56\",\n \"117c71f7-624f-4f92-9105-d8b545d244ba\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 6899,\n 6900,\n 6898,\n 6897,\n 6896\n ],\n \"content\": \"\\\\blfootnote{\\n \\\\hspace{-0.65cm} \\n This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: \\\\url{http://creativecommons.org/licenses/by/4.0/}.\\n}\\nSemantic parsing is concerned with mapping natural language (NL) utterances into machine-readable structured \\\\textit{meaning representations} (\\\\MRs). These representations are in the formats of formal languages, e.g. Prolog, SQL, and Python. A formal language is typically defined by means of a formal \\\\textit{grammar}, which consists of a set of rules. Following the convention of the chosen formal language, \\\\MRs are also referred to as logical forms or programs. An \\\\MR is often executable in a (programming) environment to yield a result (e.g. results of SQL queries) enabling automated reasoning~.\\nMost research work on semantic parsing treats each NL utterance as an \\\\emph{independent} input, ignoring the text surrounding them~, such as interaction histories in dialogues. \\nThe surrounding text varies significantly across different application scenarios. In a piece of free text, we refer to the surrounding text of a current utterance as its \\\\textit{context}. The context is different with respect to different utterances. In our sequel, we differentiate between context \\\\textit{independent} semantic parsing (\\\\CiSP) and context \\\\textit{dependent} semantic parsing (\\\\CdSP) by whether a corresponding parser utilizes context information. A knowledge base or a database (on which a \\\\MR is executed for the purpose of question answering) can be considered as context as well~.\\nThis type of context does not change with respect to the utterances. In this survey, we only consider the former kind of context which does vary with different utterances.\\n\\\\input{tab-example-pets.tex}\\nThe utilization of context in semantic parsing imposes both challenges and opportunities. \\nAs shown in Table \\\\ref{tab:CoSQL}, one challenge is to resolve references, such as \\\\textit{those} in ``For each of those, what is the maximum age''. This example shows also another challenge caused by elliptical (incomplete) utterances. The sentence ``What about the average age?'' alone misses information about the database table and the column \\\\textit{pettype}. The incomplete meaning needs to be complemented by the discourse context. \\nCompared with \\\\CiSP, which usually assumes that the information within the utterance is complete, \\\\CdSP is expected to tackle challenges posed by involving context in the parsing process \\n. In addition, tackling the above challenges provides us with more opportunities to inspect the linguistic phenomena which could influence semantic parsing. Our survey on \\\\CdSP fills the gap in the literature, as the recent surveys in the semantic parsing research mainly focus on \\\\CiSP~. \\nThis paper is organised as follows. We start with providing a brief and fundamental understanding of \\\\CiSP in \\\\S2. We then present a comprehensive organization of the recent advances in \\\\CdSP in \\\\S3. We discuss current \\\\CiSP tasks, datasets, and resources in \\\\S4. Finally, we cover open research problems in \\\\S5, and conclude by providing a roadmap for future research in this area.\",\n \"id\": \"b8456ffc-88d1-4dce-9af7-5cb66e55fd57\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"321e790e-018a-41e3-8473-e7e9883a1392\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6900\n ],\n \"content\": \"\\\\CiSP aims to learn a mapping $\\\\pi_{\\\\theta} : \\\\mathcal{X} \\\\rightarrow \\\\mathcal{Y}$, which translates an NL utterance $x \\\\in \\\\mathcal{X}$ into an \\\\MR $y \\\\in \\\\mathcal{Y}$. An \\\\MR $y$ can be executed in a programming environment (e.g. databases, knowledge graphs, etc.) to yield a result $z$, namely denotation. The structure of an \\\\MR takes a form of either a tree or graph, depending on its underlying formal language. The languages of \\\\MRs are categorized into three types of formalism\\n: logic based (e.g. first order logic), graph based (e.g. AMR~), and programming languages (e.g. Java, Python)~. Some semantic parsers explicitly apply a production grammar to yield \\\\MRs from utterances. Such a grammar consists of a set of production rules, which define a list of candidate derivations for each NL utterance. Each derivation deterministically produces a grammatically valid \\\\MR.\",\n \"id\": \"a541cf51-c35e-4bd7-b0ec-f7b7038dbad0\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"321e790e-018a-41e3-8473-e7e9883a1392\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ]\n ],\n \"subsections\": [\n \"d1388aa1-f859-436a-97f9-1580b1fa7dad\",\n \"0accbdad-eb7e-426a-a0ca-4f0b3266d1fb\"\n ],\n \"title\": \"Background\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6900\n ],\n \"content\": \"Given an utterance $x \\\\in \\\\mathcal{X}$ and its paired \\\\MR $y \\\\in \\\\mathcal{Y}$, a \\\\CiSP model can form a \\\\textit{conditional} distribution $p(y | x)$.\\nThe model learning can be supervised by either utterance-\\\\MR pairs or merely utterance-denotation pairs. If only denotations are available, a widely used approach~ is to marginalize over all possible \\\\MRs for a denotation $z$, which leads to a \\\\textit{marginal} distribution $p(z | x) = \\\\sum_{y} p(z, y| x)$. A parsing algorithm aims to find the optimal \\\\MR in the combinatorially large search space.\\nWe coarsely categorize the existing models into: symbolic approaches, neural approaches, and neural-symbolic approaches based on the category of machine learning methodology and whether any production grammars are explicitly used in models.\",\n \"id\": \"d1388aa1-f859-436a-97f9-1580b1fa7dad\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"a541cf51-c35e-4bd7-b0ec-f7b7038dbad0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Semantic Parsing Models\"\n ]\n ],\n \"subsections\": [\n \"e551a263-fa5f-4543-97d2-4b3bf2aa80fe\"\n ],\n \"title\": \"Semantic Parsing Models\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6900,\n 6897,\n 6901\n ],\n \"content\": \"\\\\label{sec:pipe}\\nA symbolic semantic parser employs production grammars to generate candidate derivations and find the most probable one via a scoring model. The scoring model is a statistical or machine learning model. \\n Each derivation is represented by handcrafted features extracted from utterances or partial \\\\MRs. Let $\\\\Phi(x, d)$ denote the features of a pair of utterance and derivation, and $G(x)$ be the set of candidate derivations based on $x$. A widely used scoring model is the log linear model~. \\n\\\\begin{equation}\\n p( d | x) = \\\\frac{\\\\exp(\\\\bm{\\\\theta}\\\\Phi(x, d))}{\\\\sum_{d' \\\\in G(x)} \\\\exp(\\\\bm{\\\\theta}\\\\Phi(x, d'))}\\n\\\\end{equation}\\nwhere $\\\\bm{\\\\theta}$ denotes the model parameters. If only utterance-denotation pairs are provided at training time, a model marginalizes over all possible derivations yielding the same denotations by $p(z | x) = \\\\sum_{d} p(z, d| x)$~. Those corresponding parsers further differentiate between graph-based parsers~ and shift-reduce parsers~ due to the adopted parsing algorithms and the ways to generate derivations. From a machine learning perspective, these approaches are also linked to a structured prediction problem.\",\n \"id\": \"e551a263-fa5f-4543-97d2-4b3bf2aa80fe\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"d1388aa1-f859-436a-97f9-1580b1fa7dad\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Semantic Parsing Models\"\n ],\n [\n \"paragraph\",\n \"Symbolic Approaches\"\n ]\n ],\n \"subsections\": [\n \"b4beea59-21ae-4c9c-befd-dc1675e2c2f8\",\n \"f6a9856b-4e78-4007-961d-d1003c3a47c0\"\n ],\n \"title\": \"Symbolic Approaches\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 166,\n 2401,\n 6902,\n 3605,\n 38\n ],\n \"content\": \"\\\\label{sec:end}\\nNeural approaches apply neural networks to translate NL utterances into \\\\MRs without using production grammars. These approaches formulate semantic parsing as a machine translation problem by viewing NL as the source language and the formal language of \\\\MRs as the target language. \\nMost work in this category adopts \\\\Seq as the backbone architecture, which consists of an encoder and a decoder. The encoder projects NL utterances into hidden representations, whereas the decoder generates linearized \\\\MRs sequentially.\\n Both encoders and decoders employ either recurrent neural networks (RNN)~ or Transformers . Note that, these methods do not apply any production grammars to filter out syntactically invalid \\\\MRs.\\nThe variants of the \\\\Seq based models also explore structural information of \\\\MRs. \\\\textsc{Seq2Tree} utilizes a tree-structured RNN as the decoder, which constrains generated \\\\MRs to take syntactically valid tree structures. The \\\\textsc{Coarse2Fine} model~ adopts a two-stage generation for the task. In the first stage, a \\\\Seq model is applied to generate \\\\MR templates, which replace entities in \\\\MRs by slot variables for a high-level generalization. In the second stage, another \\\\Seq model is applied to fill the slot variables with the corresponding entities.\",\n \"id\": \"b4beea59-21ae-4c9c-befd-dc1675e2c2f8\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"e551a263-fa5f-4543-97d2-4b3bf2aa80fe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Semantic Parsing Models\"\n ],\n [\n \"paragraph\",\n \"Symbolic Approaches\"\n ],\n [\n \"paragraph\",\n \"Neural Approaches\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Neural Approaches\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6904,\n 9083,\n 4279,\n 6903\n ],\n \"content\": \"In order to ensure the generated \\\\MRs to be syntactically valid without compromising the generalization power of neural networks, neural-symbolic approaches fuse both symbolic and neural approaches by applying production grammars to the generated \\\\MRs; then the derivations are scored by neural networks. \\nThe majority of these methods linearize derivations such that they are able to leverage \\\\Seq~. At each time step, the decoder of these methods emits either a parse action or a production rule, leading to a grammatically valid \\\\MR at the end. these works produce derivations by varying grammars. \\\\textsc{NSM} uses a subset of Lisp syntax. \\\\textsc{TranX}~ defines the grammars in Abstract Syntax Description Language, while \\\\textsc{IRNet} considers the context-free grammar of a language called SemQL. \\nThere are also neural-symbolic approaches adopting neural architectures other than \\\\Seq. One of such examples is , which adopts a dynamic neural module network (DNMN) to generate \\\\MRs.\",\n \"id\": \"f6a9856b-4e78-4007-961d-d1003c3a47c0\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"e551a263-fa5f-4543-97d2-4b3bf2aa80fe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Semantic Parsing Models\"\n ],\n [\n \"paragraph\",\n \"Symbolic Approaches\"\n ],\n [\n \"paragraph\",\n \"Neural-Symbolic Approaches\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Neural-Symbolic Approaches\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In semantic parsing, \\\\textit{exact match accuracy} is the most commonly used evaluation metric. With \\\\textit{exact match accuracy}, the parsing results are considered correct only when the output \\\\MR/denotations exactly match the string of the ground truth \\\\MR/denotations. One flaw of the evaluation metric is that some types of MRs (e.g., SQL) do not hold order constraints. \\\\newcite{yu2018spider} proposed a metric \\\\textit{set match accuracy} to evaluate the semantic parsing performance over SQLs, which treats each SQL statement as a set of clauses and ignore their orders.\\nDue to the variety of domains and languages over different datasets, it is difficult to measure all semantic parsing methods in a unified framework. To address this issue, \\\\newcite{yu2018spider}, \\\\newcite{yu2019sparc} and \\\\newcite{yu2019cosql} built different shared-task platforms with leaderboard for semantic parsing evaluation on the common datasets and consistent evaluation metrics.\",\n \"id\": \"0accbdad-eb7e-426a-a0ca-4f0b3266d1fb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a541cf51-c35e-4bd7-b0ec-f7b7038dbad0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Evaluation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Evaluation\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 6905,\n 6906\n ],\n \"content\": \"Context dependent semantic parsing involves modelling of context in the parsing process. For each current NL utterance, we define its \\\\textit{context} as the information beyond this utterance. With this definition, there are two types of context for semantic parsing, \\\\textit{local} context and \\\\textit{global} context. The \\\\textit{local} context for an utterance is the text and multimedia content surrounding it, which is meaningful only for this utterance. In plain texts, the concept of local context is also quite close to discourse, which is defined as a group of collocated, structured, coherent sentences~. In contrast, its \\\\textit{global} context is the information accessible to more than one utterance, including databases and external text corpora, images or class environment~. The content of local context varies for each NL utterance while the global context is always static. The work in our survey is only concerned with local context. Therefore, we always refer to \\\"local context\\\" as \\\"context\\\" in the following sections. \\nContext provides additional information to resolve ambiguity and vagueness in current utterances. For semantic parsing, one type of ambiguity is caused by references in current utterances, which need to be resolved to previously mentioned objects and relations. References may include explicit or implicit lexical triggers, such as \\\\textit{those} in \\\"For each of those, ...\\\" in our introductory example (Table \\\\ref{tab:CoSQL}). Another ambiguity illustrated by the same example is resulted by ellipsis. The previous context provides constraints to restrict the scope of possible \\\\MRs indicated by current utterances. In addition, context provides information to disambiguate word senses and entities, and link them to knowledge bases to enable complex reasoning. However, semantic parsing literature largely neglects word sense disambiguation, which is regarded as an AI complete problem~. Last but not least, context allows to exploit discourse coherence for semantic parsing. Coherence relations characterize structural relationships between sentences, thus limit the search space of parse candidates for the following utterances of current ones.\\nFormally, a context dependent parser takes both an input utterance $x_i$ and its context $C_i$, where $C_i$ could include a broad range of multimedia content. And we consider a group of inter-related utterances with the union set of their context as one \\\\textit{interaction}, $I = (\\\\mathbf{x}, \\\\mathbf{C})$, where $\\\\mathbf{x} = [x_1, ...,x_i,...,x_T]$ and $\\\\mathbf{C} = \\\\cup_{i=1}^{T}C_{i}$. Currently, most \\\\CdSP work focus on the research problems of context $C_i$ regarding the history utterances, \\\\MRs, denotations. Such a parser learns a mapping from a current utterance $x_{i}$ to an \\\\MR $y_i$ by $\\\\pi_{\\\\theta}(x_{i}, C_i)$.\",\n \"id\": \"0888dee2-d5ac-4bd8-807b-f7439f676fed\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"321e790e-018a-41e3-8473-e7e9883a1392\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ]\n ],\n \"subsections\": [\n \"a7325d38-7cf4-41b3-bf34-266fb7445dcc\",\n \"4f2c4e8f-03a9-4d2a-875a-0025d4eaf479\",\n \"a42c11dd-cba0-4190-ab4f-e20c9525e938\",\n \"640e3315-aedb-4aa7-abfb-05ad7be7ec1b\",\n \"6d868391-5ac7-479e-9463-c4e87430bb7a\"\n ],\n \"title\": \"Context Dependent Semantic Parsing\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6907\n ],\n \"content\": \"Existing symbolic approaches formulate \\\\CdSP as a structured prediction problem by including contextual information into their feature models. Their models capture $ p( d_i | x_i, C_i)$ by including context as a condition. Both \\\\newcite{zettlemoyer2009learning} and \\\\newcite{srivastava2017parsing} divide the parsing process into two steps: i) generate initial parses using \\\\CiSP; ii) complete initial parses using contextual information. In contrast, \\\\newcite{long2016simpler} parses a sequence of utterances in one step. In all those work, symbolic features are used to represent contexts. \\nIn two-step approaches, \\\\newcite{zettlemoyer2009learning} and \\\\newcite{srivastava2017parsing} differ in the details of individual steps. In the first step, \\\\newcite{zettlemoyer2009learning} extends \\\\MRs with predicates representing references, while \\\\newcite{srivastava2017parsing} generates a set of context independent parses for each utterance. In the second step, \\\\newcite{zettlemoyer2009learning} collects possible derivations by applying three heuristic rules to replace references with entities in context and extend initial \\\\LFs with constraints, then finds the best derivations according to a linear model. In~, their model expands the initial parse set with parses selected from context using heuristic rules, then finds the best parses in the expanded set. Their feature model includes a multinomial random variable indicating the current hidden state of discourse.\\nThe shift-reduce parser in~ generates derivations for a whole utterance sequence. This method stores the previously generated derivations in a stack, performs a sequence of \\\\textit{shift} and \\\\textit{build} operations to generate \\\\LFs. In its feature model, a context is represented by a sequence of past \\\\LFs and a random variable denoting the current world state.\\nUtterances and \\\\MRs histories form a context of a \\\\CdSP parser. The common practice is to extract handcrafted features from both utterances and \\\\MRs to represent contexts. Some typical feature patterns are as follows:\",\n \"id\": \"a7325d38-7cf4-41b3-bf34-266fb7445dcc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"0888dee2-d5ac-4bd8-807b-f7439f676fed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Symbolic Approaches\"\n ]\n ],\n \"subsections\": [\n \"6aaa20f2-9158-4b75-92a8-7292af7a0e72\"\n ],\n \"title\": \"Symbolic Approaches\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In~, they consider indicator features of lexical triggers, whether the current utterance is repeated, as well as the position of the current utterance in an interaction.\",\n \"id\": \"6aaa20f2-9158-4b75-92a8-7292af7a0e72\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"a7325d38-7cf4-41b3-bf34-266fb7445dcc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Symbolic Approaches\"\n ],\n [\n \"paragraph\",\n \"Utterance\"\n ]\n ],\n \"subsections\": [\n \"4ff929d2-dc36-402d-8f59-7bed1be0ef5b\"\n ],\n \"title\": \"Utterance\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In , there is a feature indicating if the predicates exist in the history \\\\LFs. Such feature allows the model to learn to copy the segments from the context that contains the expected predicates. \\\\newcite{long2016simpler} adopts the feature indicating if the argument in current \\\\MR is one of arguments in the last \\\\MR. \\\\newcite{srivastava2017parsing} uses the combinations of predicates in successive turns as the indicator features.\",\n \"id\": \"4ff929d2-dc36-402d-8f59-7bed1be0ef5b\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"6aaa20f2-9158-4b75-92a8-7292af7a0e72\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Symbolic Approaches\"\n ],\n [\n \"paragraph\",\n \"Utterance\"\n ],\n [\n \"paragraph\",\n \"Meaning Representations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Meaning Representations\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6908\n ],\n \"content\": \"\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=1\\\\textwidth]{figs/SQL_example-cropped.pdf}\\n \\\\label{fig:semantic_tree}\\n \\\\caption{Coreference resolution architecture of . Considering the example in Table \\\\ref{tab:CoSQL}, \\\\newcite{chen2019context} firstly generates a \\\\MR template for $Q_2$ as \\\"SELECT max(petage), \\\\textit{REF} FROM pets GROUP BY \\\\textit{REF}\\\". The \\\\textit{REF} tokens would then be replaced with the \\\"pettype\\\" from the precedent \\\\MR. }\\n \\\\label{fig:sql_example}\\n\\\\end{figure}\\nExisting neural \\\\CdSP methods extend the \\\\Seq architecture to incorporate contextual information in two ways. The first approach is to build context-aware encoders to encode historical utterances or \\\\MRs into neural representations, which provide decoders contextual information to resolve ambiguity in current utterances. As previously predicted \\\\MRs provide the constraints and information missed in current utterances, the second approach is to utilize context-aware decoders to reuse or revise those predicted \\\\MRs for generating current \\\\MRs.\",\n \"id\": \"4f2c4e8f-03a9-4d2a-875a-0025d4eaf479\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"0888dee2-d5ac-4bd8-807b-f7439f676fed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Neural Approaches\"\n ]\n ],\n \"subsections\": [\n \"28692668-fe8c-48f8-82c5-52ec47621cd5\"\n ],\n \"title\": \"Neural Approaches\"\n },\n {\n \"cite_extract_rate\": 0.818181818181818,\n \"cites\": [\n 6899,\n 6898,\n 6909,\n 1888,\n 6911,\n 6910,\n 6896,\n 168,\n 1883\n ],\n \"content\": \"Encoders of \\\\CdSP methods differentiate between utterance encoders and \\\\MR encoders. Utterance encoders construct neural representations for both current and historical utterances, while \\\\MR encoders build neural representations based on on historical \\\\MRs.\\nUtterance encoders aim to embed rich information hidden in utterances into fixed-length representations, which provide contextual information in addition to current utterances for decoders. They apply first an RNN to map each utterance into a continuous vector of fixed-size. Then there are three ways to encode utterances in context into a fixed-size neural representation.\\n\\\\begin{itemize}\\n \\\\item \\n For each utterance in a dialogue, a straightforward method is to concatenate its previous $k - 1$ utterances with current utterance in order and encode them with the RNN~.\\n As a result, decoders have access to information in at most $k$ utterances. However, this method fails to access information beyond the $k$ utterances. In addition, it is computationally expensive because if an utterance belongs to multiple contexts, it would be repeatedly encoded for modelling all the contexts.\\n \\\\item To overcome the above weakness, an alternative method is to treat a sub-sequence of utterances up to time $t$ as a sequence of vectors, and project them into a \\\\textit{discourse state} vector by using a turn-level RNN~. In another word, those models apply hierarchical RNNs to map each context into a fixed-size vector. In this method, each utterance is encoded only once and reused for modelling different contexts. However, this approach often leads to significant information loss~ due to the challenges imposed by encoding sequences of utterances into single vectors. \\n \\\\item In order to focus on history utterances most relevant to current decoder states or utterances, soft attention~ is applied to construct context vectors. The query vectors are either the hidden state of an decoder~ or an utterance vector~. To differentiate between positional information, token embeddings of history utterance are concatenated with their position embeddings~, which encode the positions of history utterances relative to the current utterances. This method reflects the observation that similar utterances tend to share relevant information, such as references of the same entities. Both discourse states and attended representations are also widely used by the neural dialogue models~, thus suffer from the same problems caused by composition complexity. As a result, the trained models are found insensitive to utterance order and word order in context~.\\n\\\\end{itemize}\\n\\\\MR encoders construct a neural context representation at time $t$ based on the \\\\MRs predicted before $t$. As \\\\MRs are expressed in a formal language, \\\\MR encoders also apply RNNs to encode each \\\\MR or segments of \\\\MRs into embedding vectors. Then \\\\MR encoders build context representations of historical \\\\MRs in the same spirit as utterance encoders. In~, they only concatenate the embeddings of $k$ most recent history \\\\MR tokens as they assume current \\\\MR is always an extension of previous \\\\MRs. In~, a bidirectional RNN is applied to construct a vector for each segment, which is extracted from historical \\\\MRs. Soft attention is also applied in~ for building context vectors, which uses the current hidden state of their decoder as the query vector to attend over the token embeddings of the previous \\\\MR.\",\n \"id\": \"28692668-fe8c-48f8-82c5-52ec47621cd5\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"4f2c4e8f-03a9-4d2a-875a-0025d4eaf479\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Neural Approaches\"\n ],\n [\n \"paragraph\",\n \"Context-aware Encoders\"\n ]\n ],\n \"subsections\": [\n \"8186c75f-b8c3-4431-b509-a482cbf52d9e\"\n ],\n \"title\": \"Context-aware Encoders\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6899,\n 6898,\n 167,\n 1888,\n 6908,\n 6896,\n 6912\n ],\n \"content\": \"Decoders in \\\\CdSP models produce \\\\MRs based on the neural representations provided by their encoders. Such a decoder yields an \\\\MR by generating a sequence of \\\\MR tokens according to model distribution $P(\\\\mathbf{y}| \\\\mathbf{x}, \\\\mathbf{C})$, where $\\\\mathbf{C}$ denotes context information. There are three major ways to utilize context information.\\nOne key problem of \\\\CdSP is incomplete information in current utterances. The straightforward way is to take neural context representations $\\\\mathbf{C}$ as additional input of decoders, which are yielded by context-aware encoders. Those context representations contains information from previous utterances, historical \\\\MRs, or both. The decoders take them as input by concatenating them with the ones from current utterances at each decoding step~. Thus, the quality of decoding depends tightly on the quality of contextual encoding, which is still a challenging problem~.\\n\\\\MRs of current utterances often contain segments from previous \\\\MRs~. The shared parts are references to previously mentioned entities or constraints implied by context. Reuse of \\\\MR segments is realized by a designated \\\\textit{copy} component, which selects a segment to copy when the probability of copying is high. As decoders in \\\\Seq produce a sequence of decisions for each input, the corresponding model generates a sequence of mixed decisions, including both \\\\textit{copy} of segments and generation of new \\\\MR tokens. In a similar manner, copying of \\\\MR tokens from previous \\\\MR is proposed in~.\\nCoreference resolution is explicitly addressed in~\\\\newcite{chen2019context}. As illustrated by the example in Figure \\\\ref{fig:sql_example}, a special token \\\\textit{REF} is introduced in the output vocabulary for denoting if an entity in the preceding \\\\MR is referred in that utterance. If that is the case, the corresponding entity token is copied from the previous \\\\MR to replace the \\\\textit{REF} token via a pointer network module~.\",\n \"id\": \"8186c75f-b8c3-4431-b509-a482cbf52d9e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"28692668-fe8c-48f8-82c5-52ec47621cd5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Neural Approaches\"\n ],\n [\n \"paragraph\",\n \"Context-aware Encoders\"\n ],\n [\n \"paragraph\",\n \"Context-aware Decoders\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Context-aware Decoders\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 6898,\n 6912,\n 6913,\n 7744\n ],\n \"content\": \"\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=1\\\\textwidth]{figs/copy_action.pdf}\\n \\\\caption{The symbolic memory architecture of . Considering the example in Table \\\\ref{tab:CoSQL}, \\\\newcite{guo2018dialog} defines different types of actions, $A_{ac}$ and $A_{ec}$, to copy action sequence $A_{10}, A_{12}, A_{13}$ and the entity \\\\textit{pettype} from the symbolic memory, respectively.}\\n \\\\label{fig:copy_action}\\n\\\\end{figure}\\nNeural-symbolic approaches introduce grammar into the decoding process or utilize symbolic representations as intermediate representations, while applying neural nets for representation learning. They take advantages from both the good context representation obtained by neural nets and reduced complexity of decoding due to the constraints introduced by grammars. In existing work, those approaches regard the generation of an \\\\MR as the prediction of a sequence of actions. Neural-symbolic methods normally take the same methods as the neural approaches to encode the contextual information. What differentiate them is the neural-symbolic could handle context by i) designing specific actions, and ii) utilizing symbolic context representations.\\nThe context specific actions proposed in~ adopt \\\\textit{copy} mechanism to reuse the previous \\\\MRs. \\\\textsc{CAMP}~ include three actions to copy three different SQL clauses from precedent queries. \\n\\\\newcite{liu2020FarAwayContextModelingSP} allows copying of any actions or subtrees from precedent SQL queries. The \\\\textit{subsequent} action in~ adds SQL conditions from the previous query into the current semantic parse to address the ellipsis problem. Different from other approaches, \\\\newcite{iyyer2017search} uses a \\\\textsc{DynSP}, which is in a similar neural network structure as the DNMN, instead of the \\\\Seq to generate the action sequences.\\nProduction rules are also used to explicitly address the coreference resolution. In~, the authors defined fours actions to instantiate the entities, predicates, types and numbers. Then the pointer network is utilized to find mentions of the four entry semantic categories in the current and history utterances. The entities in utterances are later mapped to entities in knowledge bases by using their entity linking tool.\\nInstead of directly copying from previous \\\\MRs, the parser \\\\textsc{Dialog2Action}~ incorporates a dialogue memory, which maintains \\\\textit{symbolic} representations of entities, predicates and action subsequences from an interaction history (Figure \\\\ref{fig:copy_action}). That parser defines three types of designated actions to copy entities, predicates and action subsequences from the memory respectively. Instead of decisively copying from memory, each type of action probabilistically selects the corresponding segments conditioning on the \\\\textit{symbolic} representations, which are later integrated into the generated action sequences. \\n\\\\newcite{guo2019coupling} employs the same neural-symbolic models as in to capture contextual information. Different from other approaches, \\\\newcite{guo2019coupling} adopts the meta-learning approach to improve the generation ability of \\\\CdSP models. Inspired by , \\\\newcite{guo2019coupling} utilize the context from other interactions to guide the learning of \\\\CdSP over utterances within current interactions via MAML. \\\\newcite{guo2019coupling} considers an input utterance $x_i$ and its context $C_i$ as an instance. A context-aware retriever would retrieve instances, which are semantically close to the current instances, from other interactions. When learning model parameters, the retrieved instances and the current instances are considered as the support set and test set, respectively, and grouped as tasks as in the common MAML paradigm.\",\n \"id\": \"a42c11dd-cba0-4190-ab4f-e20c9525e938\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"0888dee2-d5ac-4bd8-807b-f7439f676fed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Neural-Symbolic Approaches\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Neural-Symbolic Approaches\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6898\n ],\n \"content\": \"In~, 13 different context modeling methods for both neural and neural-symbolic \\\\CdSP parsers were evaluated on two benchmark datasets. None of those methods achieve consistent superior results over the others in all experimental settings. Among them, concatenation of $k$ recent utterances for decoders and copy of parse actions from precedent \\\\MRs are the top performing ones in most settings. ~\\\\newcite{liu2020FarAwayContextModelingSP} defines 12 fined-grained types summarized with multiple hierarchies according to the contextual linguistic phenomena, and inspects how different linguistic phenomena influence the model behavior. One interesting conclusion is that the methods in their experiments all perform poorly on the instances involving coreference problems that require complex inference. But note that, those methods in that study were not compared with the ones with explicit coreference resolution. Another interesting finding is that all the models perform better on the utterances which only augment the semantics of previous sentences than on the utterances which substitute the partial semantics of the precedent utterances.\",\n \"id\": \"640e3315-aedb-4aa7-abfb-05ad7be7ec1b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"0888dee2-d5ac-4bd8-807b-f7439f676fed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Comparison between Different \\\\CdSP Approaches\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Comparison between Different \\\\CdSP Approaches\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6899,\n 6914,\n 3063,\n 6916,\n 6915\n ],\n \"content\": \"Feedback/Interactive Semantic Parsing is another line of research in semantic parsing that utilizes context to refine \\\\MRs in an iterative manner. Most Feedback Semantic Parsing systems~ start with using an \\\\CiSP parser to parse a given utterance into an initial \\\\MR. Then the \\\\MR is interpreted in natural language and sent to a user. The user provides feedback, based on which the systems revise the initial parse. The process repeats till convergence. Therefore, in Feedback Semantic Parsing, interaction histories are only used to revise the parses. In contrast, \\\\CdSP focuses on modelling the dependencies between the utterances. \\\\newcite{elgohary2020speak} empirically compares \\\\CdSP with Feedback Semantic Parsing. They train a \\\\CdSP model, EditSQL~, on two \\\\CdSP datasets, \\\\Sparc and \\\\Cosql, and evaluate it on the test set of a feedback semantic parsing dataset, \\\\SPLASH. The performance is merely 3.4\\\\% and 3.2\\\\% in terms of accuracy, indicating that the two tasks are distinct by addressing different aspects of context.\",\n \"id\": \"6d868391-5ac7-479e-9463-c4e87430bb7a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"0888dee2-d5ac-4bd8-807b-f7439f676fed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Context Dependent Semantic Parsing\"\n ],\n [\n \"subsection\",\n \"Comparison between \\\\CdSP and Feedback Semantic Parsing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Comparison between \\\\CdSP and Feedback Semantic Parsing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\input{tab-example-data.tex}\\nTable \\\\ref{tab:context_data} summarizes the basic properties and statistics of existing \\\\CdSP datasets. There are two scenarios of the \\\\CdSP datasets, Single-party Scenarios and Multi-party Scenarios. In the former scenarios, the user utterances are translated into \\\\MRs to obtain the execution results from the programming environment. In the latter scenarios, there are systems which respond to the users in natural language based on the user utterances and the execution results.\\nThe user utterances are manually labeled with different types of annotations, including \\\\MRs, denotations, and dialogue acts. The system responses are usually annotated with the dialogue acts. We especially highlight those annotations that explicitly reflect contextual dependencies of utterances in the sequel.\",\n \"id\": \"37d74331-727a-46c3-844f-3da36f456e56\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"321e790e-018a-41e3-8473-e7e9883a1392\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Datasets and Resources\"\n ]\n ],\n \"subsections\": [\n \"f2b8d82a-6953-4c79-b18d-ac672c464017\",\n \"f428c8ab-b908-460e-bb23-7a56d331e087\"\n ],\n \"title\": \"Datasets and Resources\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"f2b8d82a-6953-4c79-b18d-ac672c464017\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"37d74331-727a-46c3-844f-3da36f456e56\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Datasets and Resources\"\n ],\n [\n \"subsection\",\n \"Scenarios\"\n ]\n ],\n \"subsections\": [\n \"1b5a8458-fd5b-4876-ba50-10c54b23220a\"\n ],\n \"title\": \"Scenarios\"\n },\n {\n \"cite_extract_rate\": 0.44444444444444403,\n \"cites\": [\n 3063,\n 8303,\n 6908,\n 6907\n ],\n \"content\": \"In \\\\Sparc~, \\\\Sqa~ and \\\\Atis, the user utterances within each interaction are around a topic described by the provided text. To collect \\\\Sparc and \\\\Sqa, crowd-workers are asked to raise questions to obtain the information that answers the questions sampled from other corpora~. But the assumption for \\\\Sqa is the answers of the current question must be the subset of answers from the last turn. In \\\\Atis~, crowd-workers raise questions around the detailed scripts describing air travel planning scenarios. \\n\\\\TemporalStructure~ and \\\\TimeExp~ particularly focused on addressing the temporal-related dependency. In \\\\TemporalStructure, human users or the simulators raise natural language questions chronologically towards a knowledge base. The facts in the knowledge base are organized in time series. Therefore, the questions in \\\\TemporalStructure are rich with time expressions. \\\\TimeExp only annotate temporal mentions (text segments that describe time expressions) instead of complete questions. All the mentions are from the time expression-rich corpora.\\n\\\\Scone~ and \\\\CSQA~ use semi-automatic approaches to simulate the contextual dependency. Each interaction in \\\\Scone is merely labeled with an initial denotation and an end denotation. The denotations in \\\\Scone are regarded as the states that can be manipulated by the programs. Within each interaction, multiple candidate sequences of programs would be automatically generated while only the sequence of programs, which could correctly transit the initial state to the end state, would be kept and described with natural language by the crowd-workers. To create \\\\CSQA dataset, the crowd-workers are asked to raise questions that can be answered from single fact tuples (e.g. relation: \\\\textit{CEO}, subject: \\\\textit{Google}, object: \\\\textit{Sundar Pichai}) in the knowledge graph or the complex facts which are the composition of multiple tuples. To create coherent dependency among questions, the questions that share the relations or entities are placed next to each other. And crowd-workers would manually modify the questions such that the sequence of questions would include contextual linguistic properties such as ambiguity, underspecification or coreference. It is worth mentioning that, with such method, \\\\CSQA includes the largest number of interactions until now, which is over 200k.\",\n \"id\": \"1b5a8458-fd5b-4876-ba50-10c54b23220a\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"f2b8d82a-6953-4c79-b18d-ac672c464017\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Datasets and Resources\"\n ],\n [\n \"subsection\",\n \"Scenarios\"\n ],\n [\n \"paragraph\",\n \"Single-party Scenarios\"\n ]\n ],\n \"subsections\": [\n \"164c4a3e-7159-47d0-82af-d206060db598\"\n ],\n \"title\": \"Single-party Scenarios\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Similar to the scenario of \\\\Sparc, to obtain the answers to the questions sampled from \\\\Spider~, the conversations in \\\\Cosql~ are conducted between two human interlocutors, who play the roles of user and system, respectively. The dialogues in the \\\\SpaceBook~ are under the scenarios formed by the routing requests. One human interlocutor pretends to be a tourist walking around Edinburgh while another interlocutor plays the role of a system responding to the tourist. The conversations in \\\\EmailDiag are between the human agent and an email assistant instead of two humans.\",\n \"id\": \"164c4a3e-7159-47d0-82af-d206060db598\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"1b5a8458-fd5b-4876-ba50-10c54b23220a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Datasets and Resources\"\n ],\n [\n \"subsection\",\n \"Scenarios\"\n ],\n [\n \"paragraph\",\n \"Single-party Scenarios\"\n ],\n [\n \"paragraph\",\n \"Multi-party Scenarios\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-party Scenarios\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The contextual linguistic phenomena in the \\\\CdSP corpora is quite close to the phenomena in the corpora of tasks such as document-level machine translation, question answering, dialogue system, etc.. However, in \\\\CdSP datasets, the contextual linguistic phenomena has a tight relation with the annotations.\\n\\\\newcite{iyyer2017search}, \\\\newcite{vlachos2014new} defined specific components in the \\\\MR languages of \\\\Sqa and \\\\SpaceBook to explicitly model the context dependency. \\\\Sqa introduced a keyword \\\\textit{subsequent}. All the answers of \\\\MR statements after \\\\textit{subsequent} would only be the subset of the answers of the precedent \\\\MR. In the language of \\\\SpaceBook, to resolve the coreference problem, a special predicate \\\\textit{equivalent} could indicate the identical entities across questions at different turns.\\nThe context dependency could be reflected by some properties of annotations. \\\\newcite{yu2019sparc} analyzed semantic changes over turns in \\\\Sparc by calculating the overlapping percentage of tokens between the SQL annotations at different turns. In \\\\Sparc, the average overlapping percentage increases at later turns within one interaction, where the users tend to narrow down their topics with turns increasing. Both \\\\newcite{yu2019sparc} and \\\\newcite{liu2020FarAwayContextModelingSP} categorized the contextual phenomena in \\\\Sparc into fine-grained types and calculate their frequency. \\\\newcite{yu2019sparc} found some SQL representations correspond to certain contextual phenomena types. For instance, in the questions of the \\\\textit{theme-entity}, which means the current question and precedent question are around the same entities but request for different properties, their corresponding SQL representations have the same \\\\textit{FROM} and \\\\textit{WHERE} clauses. But the SQL representations for other types may vary.\\nFor the datasets, \\\\SpaceBook and \\\\Cosql, \\\\newcite{yu2019cosql} and \\\\newcite{vlachos2014new} label utterances with dialogue acts along with \\\\MRs. Different from , \\\\newcite{vlachos2014new} integrated the dialogue acts into the \\\\MRs. The dialogue acts can be considered as the overall functions of the utterances while different dialogue acts reflect different properties of utterances. For example, in \\\\Cosql, the unanswerable questions that can not be parsed into SQLs are labelled with dialogue acts such as \\\\textit{NOT\\\\_RELATED}, \\\\textit{CANNOT\\\\_UNDERSTAND}, or \\\\textit{CANNOT\\\\_ANSWER}. The ambiguous questions that need to be clarified are labelled with \\\\textit{AMBIGUOUS}. The following questions are then labeled with \\\\textit{CLARIFY}. The dialogue acts could provide additional contextual information for \\\\CdSP.\",\n \"id\": \"f428c8ab-b908-460e-bb23-7a56d331e087\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"37d74331-727a-46c3-844f-3da36f456e56\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Datasets and Resources\"\n ],\n [\n \"subsection\",\n \"Context and Annotations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Context and Annotations\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\CdSP distinct from \\\\CiSP by context modelling and utilization of context information in the parsing process to complete missing information in \\\\MRs. Despite significant progress in recent years, there are still multiple directions worth pursuing.\",\n \"id\": \"117c71f7-624f-4f92-9105-d8b545d244ba\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"321e790e-018a-41e3-8473-e7e9883a1392\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Challenges and Future Directions\"\n ]\n ],\n \"subsections\": [\n \"9f019292-b9cc-44f9-a0be-f8837ae153df\"\n ],\n \"title\": \"Challenges and Future Directions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\newcite{yu2018spider} and \\\\newcite{liu2020FarAwayContextModelingSP} analyzed the influence of different types of contextual information on \\\\CdSP methods. Despite some empirical results, it still lacks of a thorough understanding of pros and cons of each type of context in relation to the parsing task. For example, in which cases should parsers extract information from \\\\MRs in context instead of utterances? Apart from ellipsis and coreference resolution, are there other linguistically motivated problems in context the current parsers have not addressed yet?\",\n \"id\": \"9f019292-b9cc-44f9-a0be-f8837ae153df\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"117c71f7-624f-4f92-9105-d8b545d244ba\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Challenges and Future Directions\"\n ],\n [\n \"paragraph\",\n \"Analysis of Linguistic Phenomena Benefiting from Context\"\n ]\n ],\n \"subsections\": [\n \"73fd17e5-1eb7-4596-820a-f4fb6a6042a2\"\n ],\n \"title\": \"Analysis of Linguistic Phenomena Benefiting from Context\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Current \\\\CdSP approaches fall in the scope of near-side pragmatics, in particular reference resolution, and current \\\\CdSP datasets (e.g. \\\\Cosql, \\\\SpaceBook) consider dialogue acts as merely the overall function of the utterances~. However, \\\\textit{far-side pragmatics} focuses on what happens beyond saying, including implicatures and communicative intentions etc.. Incorporating far-side pragmatics in semantic parsing will be especially useful towards completely understanding dialogues. Thus, there is a need to create large corpora annotated with rich information about various aspects of pragmatics for both training and evaluation.\",\n \"id\": \"73fd17e5-1eb7-4596-820a-f4fb6a6042a2\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"9f019292-b9cc-44f9-a0be-f8837ae153df\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Challenges and Future Directions\"\n ],\n [\n \"paragraph\",\n \"Analysis of Linguistic Phenomena Benefiting from Context\"\n ],\n [\n \"paragraph\",\n \"Incorporating Far-side Pragmatics\"\n ]\n ],\n \"subsections\": [\n \"e8038842-2355-4479-9735-294e34339301\",\n \"f4cff5dd-1eff-4533-adfc-1fcd6b740b89\"\n ],\n \"title\": \"Incorporating Far-side Pragmatics\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 969,\n 6898,\n 3699\n ],\n \"content\": \"A key challenge of context based modelling is composition complexity caused by highly varying context. The empirical results in~ show that the SOTA models can capture well nearby context information but it is still challenging to capture long-range dependencies in context. One possible direction is to find out the underlying causal structure~, which should be sparse and explains well which contextual information leads to current utterances. If we can focus only on the key reasons in context that lead to changes of \\\\MRs, the influence from noisy information and overfitting of models is expected to decrease significantly. Another potential benefit of understanding causal structures in context is to improve robustness of parsers by ignoring non-robust features~.\",\n \"id\": \"e8038842-2355-4479-9735-294e34339301\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"73fd17e5-1eb7-4596-820a-f4fb6a6042a2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Challenges and Future Directions\"\n ],\n [\n \"paragraph\",\n \"Analysis of Linguistic Phenomena Benefiting from Context\"\n ],\n [\n \"paragraph\",\n \"Incorporating Far-side Pragmatics\"\n ],\n [\n \"paragraph\",\n \"Causal Structure Discovery in Context\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Causal Structure Discovery in Context\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6917\n ],\n \"content\": \"Since most \\\\CdSP datasets are small in terms of the number of utterances and interactions, the direction on addressing the low-resource problem in \\\\CdSP is quite promising. The meta-learning approaches, such as the MAML \\\\CdSP in , could be a potential direction to address this issue. The other typical methods to solve low-resource issues, including weakly supervision, data augmentation, semi-supervised learning, self-supervised learning etc., could be further investigated in the scenarios of \\\\CdSP.\\n\\\\section*{Acknowledgements}\\nWe thank Xuanli He and anonymous reviewers for their useful suggestions. \\n\\\\bibliographystyle{coling}\\n\\\\bibliography{coling2020}\\n\\\\end{document}\",\n \"id\": \"f4cff5dd-1eff-4533-adfc-1fcd6b740b89\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"73fd17e5-1eb7-4596-820a-f4fb6a6042a2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Context Dependent Semantic Parsing: A Survey\"\n ],\n [\n \"section\",\n \"Challenges and Future Directions\"\n ],\n [\n \"paragraph\",\n \"Analysis of Linguistic Phenomena Benefiting from Context\"\n ],\n [\n \"paragraph\",\n \"Incorporating Far-side Pragmatics\"\n ],\n [\n \"paragraph\",\n \"Low-resource \\\\CdSP\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Low-resource \\\\CdSP\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":71,"string":"71"},"all_cites":{"kind":"list like","value":[6899,6900,6898,6897,6896,6901,166,2401,6902,3605,38,6904,9083,4279,6903,6905,6906,6907,6908,6909,1888,6911,6910,168,1883,167,6912,6913,7744,6914,3063,6916,6915,8303,969,3699,6917],"string":"[\n 6899,\n 6900,\n 6898,\n 6897,\n 6896,\n 6901,\n 166,\n 2401,\n 6902,\n 3605,\n 38,\n 6904,\n 9083,\n 4279,\n 6903,\n 6905,\n 6906,\n 6907,\n 6908,\n 6909,\n 1888,\n 6911,\n 6910,\n 168,\n 1883,\n 167,\n 6912,\n 6913,\n 7744,\n 6914,\n 3063,\n 6916,\n 6915,\n 8303,\n 969,\n 3699,\n 6917\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.188364941129946,"string":"1.188365"}}},{"rowIdx":36,"cells":{"authors":{"kind":"list like","value":["Devis Tuia","Michele Volpi","Loris Copa","Mikhail Kanevski","Jordi Munoz-Mari"],"string":"[\n \"Devis Tuia\",\n \"Michele Volpi\",\n \"Loris Copa\",\n \"Mikhail Kanevski\",\n \"Jordi Munoz-Mari\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A survey of active learning algorithms for supervised remote sensing image classification"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-04-15T21:36:59Z","string":"2021-04-15T21:36:59Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" \\textbf{This is the pre-acceptance version, to read the final version published in 2011 in the IEEE Journal of Selected Topics in Signal Processing (IEEE JSTSP), please go to: \\href{https://doi.org/10.1109/JSTSP.2011.2139193}{10.1109/JSTSP.2011.2139193}}\\\\ Defining an efficient training set is one of the most delicate phases for the success of remote sensing image classification routines. The complexity of the problem, the limited temporal and financial resources, as well as the high intraclass variance can make an algorithm fail if it is trained with a suboptimal dataset. Active learning aims at building efficient training sets by iteratively improving the model performance through sampling. A user-defined heuristic ranks the unlabeled pixels according to a function of the uncertainty of their class membership and then the user is asked to provide labels for the most uncertain pixels. This paper reviews and tests the main families of active learning algorithms: committee, large margin and posterior probability-based. For each of them, the most recent advances in the remote sensing community are discussed and some heuristics are detailed and tested. Several challenging remote sensing scenarios are considered, including very high spatial resolution and hyperspectral image classification. Finally, guidelines for choosing the good architecture are provided for new and/or unexperienced user. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"]],"subsections":["9f8d3c3d-dd73-4be2-949f-6388307ef8c8","0c24f63b-6c0b-4f7c-a826-45cfbfb05ce5","9c33e713-e531-4890-ab4a-b33425a0f7dd","4355cda7-2c05-483f-8ba7-6a88643cc3e2","544044ff-dee5-45d4-b4e9-335bae0209f4","fc8866e1-005e-4ac4-be0e-43641ad50928","16c4e1fd-cca0-4c03-98cb-9399b4dc0f51","fba5b779-cfe8-4a33-833a-07b334d324ec","132e66fa-8f0c-4248-b231-33f0f600a11b"],"title":"root"},{"cite_extract_rate":0.05,"cites":[8400],"content":"\\PARstart{N}{owadays}, the recourse to statistical learning models~ is a common practice for remote sensing data users; models such as Support Vector Machines (SVM,~) or neural networks~ are considered as state of the art algorithms for the classification of landuse using new generation satellite imagery~. Applications of such models to very high spatial~ or spectral~ resolution have proven their efficiency for handling remote sensing data.\nHowever, the performances of supervised algorithms strongly depend on the representativeness of the data used to train the classifier~. This constraint makes the generation of an appropriate training set a difficult and expensive task requiring extensive manual analysis of the image. This is usually done by visual inspection of the scene or by field surveys and successive labeling of each sample.\nIn the case of field surveys, which is usual for medium resolution, hyperspectral or SAR images, the discovery of a new label is expensive -- both in terms of time and money -- because it involves terrain campaigns. Therefore, there is a limit to the number of pixels that can be acquired. For this reason, compact and informative training sets are needed. \nIn the case of visual inspection or photo-interpretation, more common in VHR imagery, it is easier to collect data samples, since the labeling can be done directly on the image. However, the labeling is often done by mass selection on screen and several neighboring pixels carrying the same information are included. As a consequence, the training set is highly redundant. Such a redundancy considerably slows down the training phase of the model. \\mitch{, because several pixels carrying the same information are evaluated.}{} Moreover, the inclusion of noisy pixels may result in a wrong representation of the class statistics, which may lead to poor classification performances and/or overfitting~. For these reasons, a training set built by photointerpretation should also be kept as small as possible and focused on those pixels effectively improving the performance of the model.\nSumming up, besides being small, a desirable training set must be constructed in a smart way, meaning it must represent correctly the class \\mitch{statistics and covering the entirety of the data variability}{boundaries by sampling discriminative pixels}. This is particularly critical in very high spatial and spectral resolution image classification, which deal with large \\mitch{}{and/or} complex features spaces using limited training information only~. \nIn the machine learning literature this approach to sampling is known as \\emph{active learning}. The leading idea is that a classifier trained on a small set of well-chosen examples can perform as well as a classifier trained on a larger number of randomly chosen examples, while it is computationally cheaper~. Active learning focuses on the interaction between the user and the classifier. In other words, the model returns to the user the pixels whose classification outcome is the most uncertain. After accurate labeling by the user, pixels are included into the training set in order to reinforce the model. This way, the model is optimized on well-chosen difficult examples, maximizing its generalization capabilities.\nThe active learning framework has demonstrated its effectiveness when applied to large datasets needing accurate selection of examples~. This is suitable for remote sensing applications, where the number of pixels among which the search is performed is large and manual definition is - as stated above - redundant and time consuming.\nAs a consequence, active learning algorithms gain an increasing interest in remote sensing image processing and several approaches have been proposed to solve image classification tasks. This paper presents the general framework of active learning and reviews some of the methods that have been proposed in remote sensing literature. Note that this survey only covers remote sensing application of active learning principles: for a general introduction and survey of the most recent developments in the machine learning community, please refer to~.\nThe remainder of the paper is organized as follows: Section~\\ref{sec:AL} presents the general framework of active learning and the families of methods that will be detailed in \\red{Sections~\\ref{sec:committee} to~\\ref{sec:poster}}, as well as the references to specific methods. Section~\\ref{sec:data} presents the datasets considered in the experiments. Section~\\ref{sec:res} compares the different approaches numerically. Section~\\ref{sec:disc} gives an overview and guidelines for potential users. Section~\\ref{sec:concl} concludes the paper.","id":"9f8d3c3d-dd73-4be2-949f-6388307ef8c8","level":"section","origin_cites_number":20,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:AL}\nLet $X~=~\\{\\x_i,y_i\\}_{i=1}^l$ be a training set of labeled samples, with $\\x_i \\in \\mathcal{X}$ and $y_i = \\{1, ..., N\\}$. $\\mathcal{X}$ is the $d$-dimensional input space $\\in \\mathbb{R}^d$. Let also $U~=~\\{\\x_i\\}_{i=l+1}^{l+u} \\in \\mathcal{X}$, with $u \\gg l$ be the set of unlabeled pixels to be sampled, or the \\emph{pool of candidates}. \nActive learning algorithms are iterative sampling schemes, where a classification model is adapted regularly by feeding it with new labeled pixels corresponding to the ones that are most beneficial for the improvement of the model performance. These pixels are usually found in the areas of \\emph{uncertainty} of the model and their inclusion in the training set forces the model to solve the \\mitch{uncertainties}{regions of low confidence}. For a given iteration $\\epsilon$, the algorithm selects from the pool $U^\\epsilon$ the $q$ candidates that will at the same time maximize the gain in performance and reduce the uncertainty of the model when added to the current training set $X^\\epsilon$. Once the batch of pixels $S^\\epsilon = \\{\\x_m\\}_{m=1}^q \\subset U$ has been selected, it is labeled by the user, i.e. the labels $\\{y_m\\}_{m=1}^q$ are discovered. Finally, the set $S^\\epsilon$ is both added to the current training set ($X^{\\epsilon+1} = X^\\epsilon \\cup S^\\epsilon$) and removed from the pool ($U^{\\epsilon+1} = U^\\epsilon \\backslash S^\\epsilon$). The process is iterated until a stopping criterion is met. Algorithm~\\ref{alg:AL} summarizes the active selection process. From now on, the iteration index $\\epsilon$ will be omitted in order to ease notation. \n\\begin{algorithm}[!t]\n\\caption{General active learning algorithm}\n\\label{alg:AL}\n\\vspace{0.2cm}\n\\textbf{Inputs}\\\\\n- Initial training set $X^\\epsilon = \\{\\x_i,y_i\\}_{i=1}^l$ ($X \\in \\mathcal{X}$, $\\epsilon=1$).\\\\\n- Pool of candidates $U^\\epsilon = \\{\\x_i\\}_{i=l+1}^{l+u}$ ($U \\in \\mathcal{X}$, $\\epsilon=1$).\\\\\n- Number of pixels $q$ to add at each iteration (defining the batch of selected pixels $S$).\n\\vspace{0.2cm}\n\\begin{algorithmic}[1]\n\\REPEAT\n\t\\STATE Train a model with current training set $X^\\epsilon$.\n\t\\FOR{each candidate in $U^\\epsilon$}\n\t\\STATE Evaluate a user-defined \\emph{heuristic}\n\t\\ENDFOR\n\t\\STATE Rank the candidates in $U^\\epsilon$ according to the score of the heuristic\n\t\\STATE Select the $q$ most interesting pixels. $S^\\epsilon = \\{\\x_k\\}_{k=1}^q$\n\t\\STATE The user assigns a label to the selected pixels. $S^\\epsilon = \\{\\x_k,y_k\\}_{k=1}^q$\n\t\\STATE Add the batch to the training set $X^{\\epsilon+1}=X^{\\epsilon} \\cup S^\\epsilon$.\n\t\\STATE Remove the batch from the pool of candidates $U^{\\epsilon+1}=U^{\\epsilon} \\backslash S^\\epsilon$\n\t\\STATE $\\epsilon = \\epsilon + 1$\n\\UNTIL{a stopping criterion is met.}\n\\end{algorithmic}\n\\vspace{0.2cm}\n\\end{algorithm}\nAn active learning process requires interaction between the user and the model: the first provides the labeled information and the knowledge about the desired \\blue{classes}, while the latter provides both its own interpretation of the distribution of the classes and the most relevant pixels that are needed in order to solve the discrepancies encountered. This point is crucial for the success of an active learning algorithm: the machine needs a strategy to rank the pixels in the pool $U$. \nThese strategies, or \\emph{heuristics}, differentiate the algorithms proposed in the next sections and can be divided into three main families~: \n\\begin{itemize}\n\t\\item[1 -] \\emph{Committee}-based heuristics (Section~\\ref{sec:committee})\n\t\\item[2 -] \\emph{Large margin}-based heuristics (Section~\\ref{sec:lm})\n\t\\item[3 -] \\emph{Posterior probability}-based heuristics (Section~\\ref{sec:poster})\n\\end{itemize}\nA last family of active learning heuristics, the \\emph{cluster}-based, has recently being proposed in the community~: cluster-based heuristics aim at pruning a hierarchical clustering tree until the resulting clusters are consistent with the labels provided by the user. Therefore, these strategies rely on an unsupervised model, rather than on a predictive model. Since the aim of these heuristics is different form that of the other families presented, they will not be detailed in this survey.","id":"0c24f63b-6c0b-4f7c-a826-45cfbfb05ce5","level":"section","origin_cites_number":2,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Active learning: concepts and definitions"]],"subsections":[],"title":"Active learning: concepts and definitions"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:committee}\nThe first family of active learning methods quantifies the uncertainty of a pixel by considering a committee of learners~. Each member of the committee exploits different hypotheses about the classification problem and consequently labels the pixels in the pool of candidates. The algorithm then selects the samples showing maximal disagreement between the different classification models in the committee. \nTo limit computational complexity, heuristics based on multiple classifier systems have been proposed in machine learning literature. In~, methods based on boosting and bagging are proposed in this sense for binary classification only. In~, results obtained by query-by-boosting and query-by-bagging are compared using several batch datasets showing excellent performance of the heuristics proposed. Methods of this family have the advantage to be applicable to any kind of model or combination of models. \nIn the remote sensing community, committee-based approaches to active learning have been proposed exploiting two types of uncertainty: first, committees varying the pixels members have been considered in the query-by-bagging heuristic~. Then, committees based on subsets of the feature space available have been presented in~Di and Crawford~. The next two sections present the algorithms proposed in these papers.","id":"9c33e713-e531-4890-ab4a-b33425a0f7dd","level":"section","origin_cites_number":7,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Committee based active learning"]],"subsections":["dd9d8542-ab43-4865-9ca1-5532a86beb4d","d0e71651-ef54-4cae-9092-e430dfc27eb8"],"title":"Committee based active learning"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:EQB}\nIn the implementations of , bagging is proposed to build the committee: first, $k$ training sets built on a draw with replacement of the original data are defined. These draws account for a part of the available labeled pixels only. Then, each set is used to train a classifier and to predict the $u$ labels of the candidates. At the end of the procedure, $k$ labelings are provided for each candidate pixel $\\x_i \\in U$.\nIn~, the entropy $H^{\\text{BAG}}$ of the distribution of the predictions \\mitch{provided by the $k$ classifiers for each pixel $\\x_i$ in $U$}{} is used as heuristic. In~, this measure has been subsequently normalized in order to bound it with respect to the number of classes predicted by the committee and avoid hot spots of the value of uncertainty in regions where several classes overlap. The \\emph{normalized entropy query-by-bagging} heuristic can be stated as follows:\n\\begin{equation}\n\t\\hat{\\x}^{\\text{\\emph{n}EQB}} = \\arg\\max_{\\x_i \\in U}\\Big\\{\\frac{H^{\\text{BAG}}(\\x_i)}{ \\mbox{log}(N_i)}\\Big\\}\n\\label{eq:neqb}\n\\end{equation}\nwhere\n\\begin{align}\n\t&H^{\\text{BAG}}(\\x_i) = -\\sum_{\\omega=1}^{N_i} p^{\\text{BAG}}(y_i^* = \\omega|\\x_i) \\mbox{log}\\left[p^{\\text{BAG}}(y^*_i= \\omega|\\x_i)\\right]\\\\\n\t\t& \\text{where} \\qquad p^{\\text{BAG}}(y^*_i = \\omega|\\x_i) = \\frac{\\sum_{m=1}^k \\delta(y_{i,m}^{*}, \\omega )}{\\sum_{m=1}^k\\sum_{j = 1}^{N_i} \\delta(y_{i,m}^*, \\omega_j ) } \\nonumber \n\\label{eq:entropy}\n\\end{align}\n$H^{\\text{BAG}}(\\x_i)$ is an empirical measure of entropy, $y_i^*$ is the prediction for the pixel $\\x_i$ and $p^{\\text{BAG}}(y^*_i = \\omega|\\x_i)$ is the observed probability to have the class $\\omega$ predicted using the training set $X$ by the committee of $k$ models for the sample $\\x_i$. $N_i$ is the number of classes predicted for pixel $\\x_i$ \\mitch{by the committee}{}, with $1 \\leq N_i \\leq N$. The $\\delta(y_{i,m}^*,\\omega)$ operator returns the value $1$ if the classifier using the $m$-th bag classifies the sample $\\x_i$ into class $\\omega$ and $0$ otherwise. Entropy maximization gives a naturally multiclass heuristic. A candidate for which all the classifiers in the committee agree is associated with null entropy and its inclusion in the training set does not bring additional information. On the contrary, a candidate with maximum disagreement between the classifiers results in maximum entropy, and its inclusion will be highly beneficial.","id":"dd9d8542-ab43-4865-9ca1-5532a86beb4d","level":"subsection","origin_cites_number":4,"parent_id":"9c33e713-e531-4890-ab4a-b33425a0f7dd","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Committee based active learning"],["subsection","Normalized entropy query-by-bagging (nEQB)"]],"subsections":[],"title":"Normalized entropy query-by-bagging (nEQB)"},{"cite_extract_rate":0.5,"cites":[8401],"content":"When confronted to high dimensional data, it may be relevant to construct the committee by splitting the feature space into a number of subsets, or \\emph{views}~. Di and Crawford~ exploit this principle to generate different views of a hyperspectral image on the basis of the block-diagonal structure of the covariance matrix. By generating views corresponding to the different blocks, independent classifications of the same pixel can be generated and an entropy-based heuristic can be used similarly to $n$EQB.\nGiven a partition of the $d$-dimensional input space into $V$ disjoint views accounting for data subsets $\\x^v$ such that $\\bigcup_{v=1}^V \\x^v = \\x $, the `Adaptive maximum disagreement ' (AMD) heuristic selects candidates according to the following rule:\n\\begin{equation}\n\t\\hat{\\x}^{\\text{AMD}} = \\arg\\max_{\\x_i \\in U} H^{\\text{MV}}(\\x_i)\n\t\\label{eq:AMD}\n\\end{equation}\nwhere the multiview entropy $H^{\\text{MV}}$ is assessed over the predictions of classifiers using a specific view $v$:\n\\begin{align}\n\t&H^{\\text{MV}}(\\x_i) = -\\sum_{\\omega=1}^{N_i} p^{\\text{MV}}(y^*_{i,v} = \\omega|\\x^v_i) \\mbox{log}\\left[p^{\\text{MV}}(y^*_{i,v}= \\omega|\\x^v_i)\\right]\\\\\n\t& \\text{where} \\qquad p^{\\text{MV}}(y^*_i = \\omega|\\x^v_i) = \\frac{\\sum_{v=1}^V W^{\\epsilon-1}(v,\\omega) \\delta(y^*_{i,v}, \\omega )}{\\sum_{v=1}^V\\sum_{j = 1}^{N_i} W^{\\epsilon-1}(v,\\omega)} \\nonumber\n\\label{eq:MVentr}\n\\end{align}\nwhere the $\\delta(y_{i,v}^*,\\omega )$ operator returns the value $1$ if the classifier using the view $v$ classifies the sample $y_i$ into class $\\omega$ and $0$ otherwise. $\\mathbf{W}^{\\epsilon-1}$ is a $N \\times V$ weighting matrix incorporating the abilities of discrimination between the views in the different classes. At each iteration, $\\mathbf{W}^{\\epsilon-1}$ is updated on the basis of the true labels of the pixels sampled at iteration $\\epsilon-1$:\n\\begin{equation}\nW^\\epsilon(v,\\omega) = W^{\\epsilon-1}(v,\\omega)+\\delta(y_{i,v},\\omega ), \\quad \\forall i \\in S\n\\label{eq:W}\n\\end{equation}\nand its columns are normalized to a unitary sum. This matrix weights the confidence of each view to predict a given class. In~, the selection is done on a subset of $U$ containing the candidate pixels maximizing the uncertainty, which are the pixels for which the committee has predicted the highest number of classes. This way, the computational load of the algorithm is reduced.","id":"d0e71651-ef54-4cae-9092-e430dfc27eb8","level":"subsection","origin_cites_number":2,"parent_id":"9c33e713-e531-4890-ab4a-b33425a0f7dd","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Committee based active learning"],["subsection","Adaptive maximum disagreement (AMD)"]],"subsections":[],"title":"Adaptive maximum disagreement (AMD)"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:lm}\nThe second family of methods is specific to margin-based classifiers. \nMethods such as SVM are naturally good base methods for active learning: the distance to the separating hyperplane, that is the absolute value of the decision function without the sign operator, is a straightforward way of estimating the classifier confidence on an unseen sample. Let's first consider a binary problem: the distance of a sample $\\x_i$ from the SVM hyperplane is given by\n\\begin{equation}\nf(\\x_i)=\\sum_{j=1}^n \\alpha_j y_j K(\\x_j,\\x_i)+b\n\\label{eq:decfunct}\n\\end{equation}\nwhere $K(\\x_j, \\x_i)$ is a kernel, which defines the similarity between the candidate $\\x_i$ and the support vectors $\\x_j$, which are the pixels showing non zero $\\alpha_j$ coefficients. The labels $y_j$ of the support vectors are $+1$ for samples of the positive class and $-1$ for those on the negative. For additional information, see SVM literature in~. \nThis evaluation of the distance is the base ingredient of almost all large margin heuristics. Roughly speaking, these heuristics use the intuition that a sample away from the decision boundary (with a high $f(\\x_i)$) has a high confidence about its class assignment and is thus not interesting for future sampling.\nSince SVM rely on a sparse representation of the data, large margin based heuristics aim at finding the pixels in $U$ that are most likely to receive a non-zero $\\alpha_i$ weight if added to $X$. In other words, the points more likely to become support vectors are the ones lying within the margin of the current model~. The heuristic taking advantage of this property is called margin sampling (MS)~. Recent modifications of MS, aiming at minimizing the risk of selecting points that will not become support vectors, can be found in~.\nMS is the most studied active learning algorithm in remote sensing. Its first application can be found in~. Modifications of the MS heuristic have been proposed in~. Later on, since no cross-information among the samples is considered in the MS, the questions of diversity in batches of samples have been considered in~. The next sections present the MS heuristic and subsequent modifications proposed in order to enhance diversity when selecting batches of samples.","id":"4355cda7-2c05-483f-8ba7-6a88643cc3e2","level":"section","origin_cites_number":12,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Large margin based active learning"]],"subsections":["46a14016-6a16-4632-a90d-002c27d86ea9","9dde6920-321f-4866-969e-ae6a532e831e","b4b3fbe3-f9dd-4869-b53d-b3ea1f2f3618","7bae26d1-94ed-4e43-ac7f-09791c25101b"],"title":"Large margin based active learning"},{"cite_extract_rate":0,"cites":[],"content":"As stated above, margin sampling takes advantage of SVM geometrical properties, and in particular of the fact that unbounded support vectors are labeled examples that lie on the margin with a decision function value of exactly one~. Consider the pool of candidates of Fig.~\\ref{fig:dist}(a) referring to a three classes toy problem. In a multiclass one-against-all setting, the distance to each hyperplane is represented by Figs.~\\ref{fig:dist}(d-f). \nThe `margin sampling' (MS) heuristic performs sampling of the candidates by minimizing Eq.~\\eqref{eq:minMarg}:\n\\begin{equation}\n\\hat{\\x}^{\\text{MS}} = \\arg\\min_{\\x_i \\in U } \\Big\\{\\min_{\\omega}|f(\\x_i,\\omega)|\\Big\\}\n\\label{eq:minMarg}\n\\end{equation}\nwhere $f(\\x_i,\\omega)$ is the distance of the sample to the hyperplane defined for class $\\omega$ in a one-against-all setting for multiclass problems. The MS heuristic for the toy problem is reported in~Fig.~\\ref{fig:dist}(b). MS heuristic can be found in the literature under the names of `most ambiguous'~, `binary level uncertainty'~ or SVM$_\\text{SIMPLE}$~.\n\\begin{figure*}\n\\centering\n\\begin{tabular}{|ccc|}\n\\hline\n\\includegraphics[width=4.2cm]{\\figdir Tuia_1a}&\n\\includegraphics[width=4.2cm]{\\figdir distance_tot}&\n\\includegraphics[width=4.2cm]{\\figdir distance_MCLU}\\\\\n(a)&(b) MS, Eq.~\\eqref{eq:minMarg}&(c) MCLU, Eq.~\\eqref{eq:MCLUfct}\\\\\n\\includegraphics[width=4.2cm]{\\figdir distance_cl1}&\n\\includegraphics[width=4.2cm]{\\figdir distance_cl2}&\n\\includegraphics[width=4.2cm]{\\figdir distance_cl3}\\\\\n(d) $|f(\\x,\\omega_0)|$&(e) $|f(\\x,\\omega_1)|$&(f) $|f(\\x,\\omega_2)|$\\\\ \\hline\n\\end{tabular}\n\\caption{Large margin heuristics for a three classes toy example represented in subfigure (a). The color intensity represents the distance from the hyperplane, ranging from black (on the boundary) to white (maximal distance): (b) MS heuristic; (c) MCLU heuristic; areas in black are the areas of maximal uncertainty, minimizing Eq.~\\eqref{eq:minMarg} or Eq.~\\eqref{eq:MCLUfct} respectively. \nBottom row: absolute values of per-class distances (d)-(f). }\n\\label{fig:dist}\n\\end{figure*}","id":"46a14016-6a16-4632-a90d-002c27d86ea9","level":"subsection","origin_cites_number":5,"parent_id":"4355cda7-2c05-483f-8ba7-6a88643cc3e2","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Large margin based active learning"],["subsection","Margin sampling (MS)"]],"subsections":[],"title":"Margin sampling (MS)"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:MCLU}\nIn~, the idea of margin sampling is extended to multiclass uncertainty (see~). Instead of dealing with the most uncertain class of the SVM, the `multiclass level uncertainty' (MCLU) considers the difference between the distance to the margin for the two most probable classes. \n\\begin{align}\n&\\hat{\\x}^{\\text{MCLU}} = \\arg \\min_{\\x_i \\in U} \\Big\\{f(\\x_i)^{\\text{MC}}\\Big\\}\\label{eq:MCLU}\\\\\n&\\text{where} \\qquad f(\\x_i)^{\\text{MC}} = \\max_{\\omega \\in N}|f(\\x_i,\\omega)|-\\max_{\\omega \\in N\\backslash \\omega^+}|f(\\x_i,\\omega)|\\label{eq:MCLUfct}\n\\end{align}\nwhere $\\omega^+$ is the class showing maximal confidence, i.e. the argument of the first term of Eq.~\\eqref{eq:MCLUfct} showing maximal $f(\\x_i)^{\\text{MC}}$. A high value of this criterion corresponds to samples assigned with high certainty to the most confident class, while a small value represents unreliable classification. Fig.~\\ref{fig:dist}(c) illustrates the heuristic in comparison to MS. Although they are very similar, MCLU performs better in the area where the three classes mix, in the top-right area of the feature space: in this area, MCLU returns maximal uncertainty as it is evaluated on the basis of all the per-class decision values, while MS returns an uncertainty slightly lower than on the two-classes boundaries.","id":"9dde6920-321f-4866-969e-ae6a532e831e","level":"subsection","origin_cites_number":2,"parent_id":"4355cda7-2c05-483f-8ba7-6a88643cc3e2","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Large margin based active learning"],["subsection","Multiclass level uncertainty (MCLU)"]],"subsections":[],"title":"Multiclass level uncertainty (MCLU)"},{"cite_extract_rate":0,"cites":[],"content":"In~, instead of using the distance to the hyperplane as a measure of uncertainty, the support vector coefficients are used to convert the multiclass classification problem into a binary support vector detection problem. In the `significance space construction' (SSC) heuristic, the training samples related to support vector coefficients are used to define a second classification function $f(\\x)^{\\text{SSC}}$, where training pixels with $\\alpha_j > 0$ (the support vectors) are classified against training pixels with $\\alpha_j = 0$. Once applied to the pool of candidates, this second classifier predicts which pixels are likely to become support vectors. \n\\begin{equation}\n\\hat{\\x}^{\\text{SSC}} = \\arg_{\\x_i \\in U} f(\\x_i)^{\\text{SSC}} > 0\n\\label{eq:SSC}\n\\end{equation}\nOnce the candidates more likely to become support vectors have been highlighted, a random selection among them is done.","id":"b4b3fbe3-f9dd-4869-b53d-b3ea1f2f3618","level":"subsection","origin_cites_number":1,"parent_id":"4355cda7-2c05-483f-8ba7-6a88643cc3e2","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Large margin based active learning"],["subsection","Significance space construction (SSC)"]],"subsections":[],"title":"Significance space construction (SSC)"},{"cite_extract_rate":0,"cites":[],"content":"In applicative scenarios, diversity among samples~ is highly desirable. Diversity concerns the capability of the model to reject candidates that rank well according to the heuristic, but are redundant among each other. Diversity has been studied extensively for margin-based heuristics, where the base margin sampling heuristic is constrained using a measure of diversity between the candidates (see Algorithm~\\ref{alg:diversity}). \n\\begin{algorithm}[!b]\n\\caption{General diversity based heuristic (for a single iteration)}\n\\label{alg:diversity}\n\\vspace{0.2cm}\n\\textbf{Inputs}\\\\\n- Current training set $X^{\\epsilon} = \\{\\x_i,y_i\\}_{i=1}^l$ ($X \\in \\mathcal{X}$).\\\\\n- Subset of the pool of candidates minimizing Eq.~\\eqref{eq:minMarg} or~\\eqref{eq:MCLU} $U^* = \\{\\x_i\\}$ $ (U^* \\in \\mathcal{X}$ and $U^* \\subset U^{\\epsilon}$).\\\\\n- Number of pixels $q$ to add at each iteration (defining the batch of selected pixels $S$).\n\\vspace{0.2cm}\n\\begin{algorithmic}[1]\n\\STATE Add the pixel minimizing Eq.~\\eqref{eq:minMarg} or~\\eqref{eq:MCLUfct} to $S$.\n\\REPEAT{}\n\t\\STATE Compute the user defined diversity criterion between pixels in $U^*$ and in $S$ (with MAO, cSV or ABD).\n\t\\STATE Select the pixel $\\x_D$ maximizing diversity with current batch.\n\t\\STATE Add $\\x_D$ to current batch $S = S \\cup \\x_D$.\n\t\\STATE Remove $\\x_D$ to current list of cadidates $U^* = U^* \\setminus \\x_D$.\n\\UNTIL{batch $S$ contains $q$ elements.}\n\\STATE The user labels the selected pixels. $S = \\{\\x_k,y_k\\}_{k=1}^q$\n\\STATE Add the batch to the training set $X^{\\epsilon+1}=X^{\\epsilon} \\cup S$.\n\\STATE Remove the batch from the complete pool of candidates $U^{\\epsilon+1}=U^{\\epsilon} \\backslash S$\n\\end{algorithmic}\n\\vspace{0.2cm}\n\\end{algorithm}\nThe first heuristic proposing explicit diversity in remote sensing is found in~, where the margin sampling heuristic is constrained with a measure of the angle between candidates in feature space. This heuristic, called `most ambiguous and orthogonal' (MAO) is iterative: starting from the samples selected by MS, \\mitch{}{$U^{\\text{MS}}\\subset U$}, this heuristic iteratively chooses the samples minimizing the highest values between the candidates list and the samples already included in the batch \\mitch{of samples}{} $S$. For a single iteration, this can be resumed as:\n\\begin{equation}\n\\hat{\\x}^{\\text{MAO}} = \\arg \\min_{\\x_i \\in U^{\\text{MS}} }\\Big\\{\\max_{\\x_j \\in S}K(\\x_i,\\x_j)\\Big\\}\n\\label{eq:MAO}\n\\end{equation}\nIn~, the MAO criterion is combined with the MCLU uncertainty estimation in the `multiclass level uncertainty - angle-based diversity' (MCLU-ABD) heuristic. Te selection is performed among a subset of $U$ maximizing the MCLU criterion. Moreover, the author generalize the MAO heuristic to any type of kernels by including normalization in feature space.\n\\begin{align}\n\\hat{\\x}^{\\text{MCLU-ABD}} = \\arg \\min_{\\x_i \\in U^{\\text{MCLU}} }&\\Bigg\\{\\lambda f(\\x_i)^{\\text{MC}} + \\\\\\nonumber\n&(1-\\lambda)\\max_{\\x_j \\in S}\\frac{K(\\x_i,\\x_j)}{\\sqrt{K(\\x_i,\\x_i)K(\\x_j,\\x_j)}}\\Bigg\\}\\\\\n\\label{eq:MCLU-ABD}\n\\end{align}\nwhere $f(\\x_i)^{\\text{MC}}$ is the multiclass uncertainty function defined by Eq.~\\eqref{eq:MCLUfct}.\nIn~, the diversity of the chosen candidates is enforced by constraining the MS solution to pixels associated to different closest support vectors. This approach ensures a certain degree of diversification in the MS heuristic, by dividing the margin in the feature space as a function of the geometrical distribution of the support vectors. Compared to the previously presented heuristics, this approach has the advantage of ensuring diversity with respect to the current model, but does not guarantee diversity of the samples between each other (since two close samples can be associated to different support vectors).\n\\begin{equation}\n\\hat{\\x}^{\\text{cSV}} = \\arg \\min_{\\x_i \\in U^{\\text{MS}} } \\Big\\{\\lvert f(\\x_i,\\omega)\\rvert \\ \\Big\\vert \\ cSV_i \\not\\in cSV_\\theta\\Big\\}\n\\label{eq:cSV}\n\\end{equation}\nwhere $\\theta = [1, \\ldots, q-1]$ are the indices of the already selected candidates and $cSV$ is the set of selected closest support vectors.\nFinally, diversity can be ensured using clustering in the feature space. In~, kernel $k$-means~ was used to cluster the samples selected by MCLU and select diverse batches. After partitioning the $U^{\\text{MCLU}}$ set into $q$ clusters with kernel $k$-means, the `multiclass level uncertainty - enhanced cluster based diversity (MCLU-ECBD)' selects a single pixel per cluster, minimizing the following query function:\n\\begin{equation}\n\\hat{\\x}^{\\text{MCLU-ECBD}} = \\arg \\min_{\\x_i \\in c_m}\\Big\\{ f(\\x_i)^{\\text{MC}}\\Big\\}\\text{,} \\quad m = [1, ... q], \\x_i \\in U^{\\text{MCLU}}\n\\label{eq:MCLU-ECBD}\n\\end{equation}\nwhere $c_m$ is one among the $q$ clusters defined with kernel $k$-means. \nIn~, a hierarchical extension of this principle is proposed to exclude from the selected batch the pixels more likely to become bounded support vectors. This way, the redundancy affecting samples close to each other in the feature space among different iterations is controlled along with the maximization of the informativeness of each pixel.\nIn the `informative hierarchical margin cluster sampling' (hMCS-i), a dataset composed by i) a subset of the pool of candidates optimizing the MCLU criterion ($U^{\\text{MCLU}}$) and ii) the bounded support vectors sampled at the previous iteration, is iteratively partitioned in a binary way. The partitioning always considers the biggest current cluster found and continues until $q$ clusters not containing a bounded support vector are found. Once the $q$ clusters have been defined, a search among the candidates falling in these $q$ clusters is performed.\n\\begin{equation}\n\\hat{\\x}^{\\text{hMCS-i}} = \\arg \\min_{\\x_i \\in c_m } \\Big\\{f(\\x_i)^{\\text{MC}} \\Big\\}\\text{,}\\quad m = [1, ..., q \\vert n^{bSV}_{c_m} = 0], \\x_i \\in U^{\\text{MCLU}}\n\\label{eq:hMCS}\n\\end{equation}","id":"7bae26d1-94ed-4e43-ac7f-09791c25101b","level":"subsection","origin_cites_number":8,"parent_id":"4355cda7-2c05-483f-8ba7-6a88643cc3e2","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Large margin based active learning"],["subsection","On the need for a diversity criterion"]],"subsections":[],"title":"On the need for a diversity criterion"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:poster}\nThe third class of methods uses the estimation of posterior probabilities of class membership (i.e. $p( y | \\x)$) to rank the candidates. The posterior probability gives an idea of the confidence of the class assignment (which is usually done by maximizing it over all the possible classes): by considering the change on the overall posterior distribution or the per-class distribution for each candidate, these heuristics use these probability estimates to focus sampling in uncertain areas. This section details two heuristics, the KL-max and the Breaking ties.","id":"544044ff-dee5-45d4-b4e9-335bae0209f4","level":"section","origin_cites_number":0,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Posterior probability based active learning"]],"subsections":["e1f6b836-abb9-4d52-abf9-04c6538aa022","e604938a-7c05-4eae-a00c-0f7093b23169"],"title":"Posterior probability based active learning"},{"cite_extract_rate":0,"cites":[],"content":"The first idea is to sample the pixels whose inclusion in the training set would maximize the changes in the posterior distribution. An application of these methods can be found in ~, where the heuristic maximizes the Kullback-Leibler divergence between the distributions before and after adding the candidate. In remote sensing, a probabilistic method based on this strategy and using a Maximum Likelihood classifier can be found in~. In this setting, each candidate is removed from $U$ and it is included in the training set with the label maximizing its posterior probability. The Kullbach-Leibler divergence KL is then computed between the posterior distributions of the models with and without the candidate. After computing this measure for all candidates, the pixel maximizing the following is chosen:\n{\\small\n\\begin{align}\n&\\hat{\\x}^{\\text{KL-max}} = \\\\\\nonumber &= \\arg \\max_{\\x_i \\in U } \\Bigg\\{\\sum_{\\omega \\in N} \\frac{1}{(u-1)} \\text{KL}\\Big(p^+(\\omega|\\x)\\Big\\vert\\Big\\vert p(\\omega|\\x)\\Big)p(y^*_i = \\omega|\\x_i)\\Bigg\\}\\label{eq:Kmax}\n\\end{align} }\nwhere\n{\\small\n\\begin{align}\n \\text{KL}\\Big(p^+( \\omega|\\x)\\Big\\vert\\Big\\vert p( \\omega|\\x)\\Big) = \\sum_{\\x_j \\in U \\setminus \\x_i} p^+(y^*_j = \\omega|\\x_j)\\log\\frac{p^+(y^*_j = \\omega|\\x_j)}{p(y^*_j = \\omega|\\x_j)}\n\\end{align} }\nand $p^+(\\omega|\\x)$ is the posterior distribution for class $\\omega$ and pixel $\\x$, estimated using the increased training set $X^+ = [X,(\\x_i,y^*_i)]$, where $y_i^*$ is the class maximizing the posterior probability. Recently, the authors of~ extended this approach, proposing to use boosting to weight pixels that were previously selected, but were no longer relevant to the current classifier. These heuristics are useful when used with classifiers with small computational cost: since each iteration implies to train $u+1$ models, this type of heuristics is hardly applicable with computationally demanding methods as SVM. Moreover, a selection of batches of pixels is not possible.","id":"e1f6b836-abb9-4d52-abf9-04c6538aa022","level":"subsection","origin_cites_number":3,"parent_id":"544044ff-dee5-45d4-b4e9-335bae0209f4","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Posterior probability based active learning"],["subsection","KL-max"]],"subsections":[],"title":"KL-max"},{"cite_extract_rate":0,"cites":[],"content":"Another strategy, closer to the idea of EQB presented in Section \\ref{sec:EQB}, consists of building a heuristic exploiting the conditional probability of predicting a given label $p(y_i^* = \\omega|\\x_i)$ for each candidate $\\x_i \\in U$. In this case, note that the predictions for the single candidates $y_i^* = \\arg\\max_{\\omega \\in N}f(\\x_i,\\omega)$ are used. Such estimates are provided by several methods, including probabilistic neural networks or maximum likelihood classifiers. A possibility to obtain posterior probabilities from SVM outputs\\footnote{Even though they are not posterior probabilities from a Bayesian point of view} is to use Platt's estimation~. \nIn this case, the per-class posterior probability is assessed fitting a sigmoid function to the SVM decision function~:\n\\begin{equation}\n\tp(y_i^* = \\omega|\\x_i) =\\frac{1}{1 + e^{(Af(\\x_i,\\omega)+B)}}\n\\end{equation}\n\\mitch{where $\\omega = \\{1,... ,N\\}$ is the class considered and $f(\\x_i,\\omega)$ is the binary decision function of the SVM without the sign operator (see Eq.~\\eqref{eq:decfunct}).}{} where $A$ and $B$ are parameters to be estimated (for details, see~). \nOnce the posterior probabilities are obtained, it is possible to assess the uncertainty of the class membership for each candidate in a direct way. In this case the heuristic choses candidates showing a near uniform probability of belonging to each class, i.e. $p( y_i^* = \\omega | \\x_i) = 1/N, \\ \\forall \\omega \\in N$.\nThe `Breaking ties' (BT) heuristic for a binary problem relies on the smallest difference of the posterior probabilities for each sample . In a multi-class setting, this reciprocity can still be confirmed and used, since independently from the number of classes $N$, the difference between the two highest probabilities can be indicative of the way an example is handled by the classifier\\mitch{(see a similar interpretation for the MCLU heuristic in Section~\\ref{sec:MCLU})}{}. When the two highest probabilities are close (``on a tie''), the classifier's confidence is the lowest. The BT heuristic can thus be formulated as:\n\\begin{equation}\n\\hat{\\x}^{\\text{BT}} = \\arg \\min_{\\x_i \\in U } \\Big\\{\\max_{\\omega \\in N} \\big\\{p(y_i^* = \\omega|\\x)\\big\\} - \\max_{\\omega \\in \\mitch{}{N}\\backslash \\omega^+} \\big\\{p(y_i^* = \\omega|\\x)\\big\\}\\Big\\}\n\\label{eq:BT}\n\\end{equation}\nwhere $\\omega^+$ is the class showing maximal probability, i.e. the argument of the first term of Eq.~\\eqref{eq:BT}. By comparing Eq.~\\eqref{eq:MCLU} with Eq.~\\eqref{eq:BT}, it is clear that the link between BT and the MCLU heuristic when using SVM classifiers (see Section~\\ref{sec:MCLU}) is really strong.","id":"e604938a-7c05-4eae-a00c-0f7093b23169","level":"subsection","origin_cites_number":2,"parent_id":"544044ff-dee5-45d4-b4e9-335bae0209f4","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Posterior probability based active learning"],["subsection","Breaking ties (BT)"]],"subsections":[],"title":"Breaking ties (BT)"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:data}\nThis Section details the datasets considered and the setup of the experiments performed.","id":"fc8866e1-005e-4ac4-be0e-43641ad50928","level":"section","origin_cites_number":0,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Datasets and setup"]],"subsections":["6e70c940-20c9-4ab7-ad99-8a4761d0a98e","bbc1da7f-38e6-4e9a-a5cf-edcfa29f4d5e"],"title":"Datasets and setup"},{"cite_extract_rate":0,"cites":[],"content":"Active heuristics have been tested on three challenging remote sensing classification scenarios (Fig.~\\ref{fig:data}), whose data distributions are detailed in Fig.~\\ref{fig:Distr}. \n\\begin{figure}[!b]\n\\centering\n\t\\begin{tabular}{cc}\n\t\\multicolumn{2}{c}{\\includegraphics[width=8.4cm]{\\figdir Pavia}}\\\\\n\t\\multicolumn{2}{c}{\\includegraphics[width=8.4cm]{\\figdir PaviaGT}}\\\\\n\t\\multicolumn{2}{c}{ROSIS Pavia}\\\\\n\t\\includegraphics[height=4cm]{\\figdir Pines_40_30_20}&\n\t\\includegraphics[height=4cm]{\\figdir pinesGT}\\\\\n\t\\multicolumn{2}{c}{AVIRIS Indian Pines}\\\\\n\\includegraphics[height=4cm]{\\figdir imageZH_MR}&\n\\includegraphics[height=4cm]{\\figdir ZH_MR_GT}\\\\\n\t\\multicolumn{2}{c}{QuickBird Zurich}\\\\\n\t\\end{tabular}\n\t\\caption{Images considered in the experiments: (top) ROSIS image of the city of Pavia, Italy (bands $[56-31-6]$ and corresponding ground survey); (middle) AVIRIS Indian Pines hyperspectral data (bands $[40-30-20]$ and corresponding ground survey); (bottom) QuickBird multispectral image of a suburb of the city of Zurich, Switzerland (bands $[3-2-1]$ and corresponding ground survey).}\n\t\\label{fig:data}\n\\end{figure}\n\\begin{figure}\n\\centering\n\\begin{tabular}{ccc}\n\\includegraphics[width=4cm]{\\figdir MeanSpectraPavia}&\n\\includegraphics[width=4cm]{\\figdir testLabels}\\\\\n(a)&(b)\\\\\n\\includegraphics[width=4cm]{\\figdir MeanSpectraPines}&\n\\includegraphics[width=4cm]{\\figdir manifold_52_102_208}\n\\\\\n(c)&(d)\\\\\n\\includegraphics[width=4cm]{\\figdir ZHspectrum_ext2}&\n\\includegraphics[width=4cm]{\\figdir ZHmanifold}\n\\\\\n(e) & (f)\\\\\n\\end{tabular}\n\\caption{Data distribution of the three images considered. First row: ROSIS image of Pavia: (a) mean spectral profiles; (b) example of data manifold in bands 55 (Red) and 77 (Near infrared). Middle row: AVIRIS Indian Pines: (c) mean spectral profiles; (d) example of data manifold in bands 52, 102 and 208. Bottom row: Zurich QuickBird: (e) mean spectral profiles; (f) data manifold in bands 2 (G), 3 (R) and 4 (NIR).}\n\\label{fig:Distr}\n\\end{figure}","id":"6e70c940-20c9-4ab7-ad99-8a4761d0a98e","level":"subsection","origin_cites_number":0,"parent_id":"fc8866e1-005e-4ac4-be0e-43641ad50928","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Datasets and setup"],["subsection","Datasets"]],"subsections":["0d621b05-e50b-4902-ba8d-1510e19e9756","5744c67c-bb8c-4a84-a58d-895f062fc480","2fba1c13-bf38-48d4-a59c-6a88de2989ac"],"title":"Datasets"},{"cite_extract_rate":0,"cites":[],"content":"the two top rows of Fig.~\\ref{fig:data} show a hyperspectral 1.3m spatial resolution image of the city of Pavia (Italy) taken by the airborne ROSIS-03 optical sensor~. The image consists of 102 spectral bands of size ($1400 \\times 512$) pixels with a spectral coverage ranging from 0.43 to 0.86 $\\mu$m. $5$ classes of interest (Buildings, Roads, Water, Vegetation and Shadows) have been selected and a labeled dataset of 206`009 pixels has been extracted by visual inspection. Among the available pixels, 20`000 have been used for the training set $X$ and candidate set $U$. Ten independent experiments have been performed, starting with $5 \\times 5 = 25$ labeled pixels ($5$ per class) in $X$ and the remaining pixels in $U$. When using LDA, 150 pixels ($30$ per class) have been included in the starting set. The higher number of starting training pixels used for LDA is justified by the requirements of the model ($n$ must be greater than the dimensionality of the data). In each experiment, 80`000 randomly selected pixels have been used to test the generalization capabilities of the heuristics.\nThe data distribution of the five classes is illustrated in the first row of Fig.~\\ref{fig:Distr}: from the mean spectra (Fig.~\\ref{fig:Distr}(a)) the classes are well distinguished and separable with the sole spectral information and the resulting data manifold (Fig.~\\ref{fig:Distr}(b)) shows a data distribution which can be handled by most linear and non linear models.","id":"0d621b05-e50b-4902-ba8d-1510e19e9756","level":"subsubsection","origin_cites_number":1,"parent_id":"6e70c940-20c9-4ab7-ad99-8a4761d0a98e","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Datasets and setup"],["subsection","Datasets"],["subsubsection","Hyperspectral VHR"]],"subsections":[],"title":"Hyperspectral VHR"},{"cite_extract_rate":0,"cites":[],"content":"the second dataset, illustrated in the second row of Fig.~\\ref{fig:data}, is a 220-bands AVIRIS image taken over Indiana's Indian Pine test site in June 1992~. The image is $145 \\times 145$ pixels, contains 16 classes representing different crops, and a total of 10`366 labeled pixels. This image is a classical benchmark to validate model accuracy and constitutes a very challenging classification problem because of the strong mixture of the class signatures. Twenty water absorption channels were removed prior to analysis\n. In the experiments, classes with less than 100 labeled pixels were removed, resulting thus in a 12 classes classification problem with 10`171 labeled pixels (see the ground truth pixels in Fig.~\\ref{fig:data}). Among the available labeled pixels, 7`000 were used for the $X$ and $U$ sets. Each experiment starts with $5 \\times 12 =60$ pixels (5 per class). As for the previous image, the remaining 3`171 pixels have been used to test the generalization capabilities.\nVisualization of the spectral mean profiles and of the data manifold (second row of Fig.~\\ref{fig:Distr}) illustrates a completely different situation with respect to the previous image: high nonlinearity and strongly overlapping classes characterize this dataset. Therefore, linear classifiers do not perform well on this dataset and will not be considered in the experiments.","id":"5744c67c-bb8c-4a84-a58d-895f062fc480","level":"subsubsection","origin_cites_number":1,"parent_id":"6e70c940-20c9-4ab7-ad99-8a4761d0a98e","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Datasets and setup"],["subsection","Datasets"],["subsubsection","Hyperspectral MR"]],"subsections":[],"title":"Hyperspectral MR"},{"cite_extract_rate":0,"cites":[],"content":"The third image, illustrated in the last row of Fig.~\\ref{fig:data}, is a 4-bands QuickBird scene of a suburb of the city of Zurich (Switzerland) taken in 2002. The image is $329 \\times 347$ pixels with a spatial resolution of 2.4m. Nine classes of interest have been extracted by careful visual inspection, namely Residential buildings, Commercial buildings, Trees, Vegetation, Harvested fields, Bare soil, Roads, Parking lots and Water. Since some of the classes to be separated are of landuse and have very similar responses in the spectral domain (see, for instance, the residential and commercial buildings, or roads and parking lots), 16 contextual bands, extracted using opening (8) and closing (8) morphological operators (see~), have been added to the four spectral bands, resulting in a 20-dimensional dataset. As for the Pavia dataset, 20`000 pixels have been extracted for the $X$ and $U$ sets. Each experiment starts with $5 \\times 9 =45$ pixels ($5$ per class). \nThe complexity of this third dataset is confirmed by both the spectra and the manifold illustrated in the bottom row of Fig.~\\ref{fig:Distr}. Strong overlaps between the asphalt and the soil classes are observed, which is also confirmed by the similarity between the spectral profiles. However, the spatial features added improve the differentiation of the classes (see, for instance, the opening features for the vegetation classes and the closing features for the asphalt classes).","id":"2fba1c13-bf38-48d4-a59c-6a88de2989ac","level":"subsubsection","origin_cites_number":1,"parent_id":"6e70c940-20c9-4ab7-ad99-8a4761d0a98e","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Datasets and setup"],["subsection","Datasets"],["subsubsection","Multispectral VHR"]],"subsections":[],"title":"Multispectral VHR"},{"cite_extract_rate":0,"cites":[],"content":"In the experiments, SVM classifiers with RBF kernel and LDA classifiers have been considered for the experiments. When using SVM, free parameters have been optimized by 5-fold cross validation optimizing an accuracy criterion. The active learning algorithms have been run in two settings, adding $N+5$ and $N+20$ pixels per iteration. To reach convergence, $70$ ($40$ in the case $N + 20$) iterations have been executed for the first image, $100$ ($50$) for the second and $80$ ($50$) for the third. $n$EQB has been run with committees of $7$ models using $75\\%$ of the available training data. For the experiments using LDA, $40$ (20) iterations have been performed and $n$EQB using $12$ models and $85\\%$ of the data have been used. \nAn upper bound on the classification accuracy has been computed by considering a model trained on the whole $X \\cup U$ set (`Standard SVM/LDA' hereafter). The lower bound on performance has been considered by assessing a model using an increasing training set, randomly adding the same number of pixels at each epoch (`Random Sampling' hereafter). \nEach heuristic has been run ten times with different initial training sets. All the graphics report mean and standard deviation of the ten experiments.","id":"bbc1da7f-38e6-4e9a-a5cf-edcfa29f4d5e","level":"subsection","origin_cites_number":0,"parent_id":"fc8866e1-005e-4ac4-be0e-43641ad50928","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Datasets and setup"],["subsection","Experimental setup"]],"subsections":[],"title":"Experimental setup"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:res}\nIn this section, some of the heuristics presented are compared on the three datasets presented above. \nThe experiments do not aim at defining which heuristic is best, since they respond unequally well to different data architectures. Rather, it attempts to illustrate the strengths and weaknesses of the different families of methods and to help the user in selecting the methodology that will perform best depending on the characteristics of the problem at hand. \nThe heuristics studied are the following: $n$EQB, MS, MCLU, MCLU-ABD and BT. Their comparison with Random sampling (RS), the base learner (Standard SVM/LDA) and between each other will show the main differences among the active learning architectures presented. \n\\begin{figure}[!t]\n\\begin{tabular}{ccc}\n& $N+5$ & $N+20$\\\\\n\\rotatebox{90}{ Pavia ROSIS}\n&\n\\includegraphics[width=3.8cm]{./images/Pavia_comparison_10}\n&\n\\includegraphics[width=3.8cm]{./images/Pavia_comparison_25}\\\\\n&(a)&(d)\\\\\n\\rotatebox{90}{ Indian Pines AVIRIS}\n&\n\\includegraphics[width=3.8cm]{./images/Pines_comparison_17}\n&\n\\includegraphics[width=3.8cm]{./images/Pines_comparison_30}\n\\\\\n&(b)&(e)\\\\\n\\rotatebox{90}{ Zurich QuickBird}\n&\n\\includegraphics[width=3.8cm]{./images/ZH_MR_comparison_14}\n&\n\\includegraphics[width=3.8cm]{./images/ZH_MR_comparison_25}\n\\\\\n&(c)&(f)\\\\\n\\end{tabular}\n\\caption{The three families of heuristics trained with SVMs (RS = Random Sampling).}\n\\label{fig:comp}\n\\end{figure}\nFigure~\\ref{fig:comp} compares a heuristic for each family presented, when using SVM classifiers. In general, MS performs better than the two other families. This is expected, since MS ranks the candidates directly using the SVM decision function without further estimations: the slightly inferior performances of the $n$EQB and the BT algorithms are mainly due to the small size of the initial training set, which does not allow these criteria to perform optimally. $n$EQB uses too small bags of training pixels and BT cannot estimate the posterior probabilities correctly, because the fit of Platt's probabilities is dependent on the number of samples used. As a consequence, the performances of these two heuristics in the first iterations is similar to random sampling, a behavior already observed in~. Summarizing, when using SVMs, the most logical choice among the families seems to be a large margin heuristic.\n\\begin{figure}[!b]\n\\begin{tabular}{ccc}\n& $N+5$ & $N+20$\\\\\n\\rotatebox{90}{ Pavia ROSIS}\n&\n\\includegraphics[width=3.8cm]{./images/Pavia_MS_MCLU_10}\n&\n\\includegraphics[width=3.8cm]{./images/Pavia_MS_MCLU_25} \\\\\n&(a)&(d)\\\\\n\\rotatebox{90}{ Indian Pines AVIRIS}\n&\n\\includegraphics[width=3.8cm]{./images/Pini_MS_MCLU_17}\n&\n\\includegraphics[width=3.8cm]{./images/Pini_MS_MCLU_40} \\\\\n&(b)&(e)\\\\\n\\rotatebox{90}{ Zurich QuickBird}\n&\n\\includegraphics[width=3.8cm]{./images/ZH_MR_MS_MCLU_14}\n&\n\\includegraphics[width=3.8cm]{./images/ZH_MR_MS_MCLU_30} \\\\\n&(c)&(f)\\\\\n\\end{tabular}\n\\caption{Large margin active learning without diversity criterion. An example comparing MS and MCLU (RS = Random Sampling).}\n\\label{fig:largeComp}\n\\end{figure}\nRegarding this family, Figs.~\\ref{fig:largeComp} and~\\ref{fig:diver} illustrate two concepts regarding the two stages of large margin heuristics: the uncertainty and diversity criteria. Figures~\\ref{fig:largeComp} compares the MS and MCLU criteria and shows that both describe the uncertainty of the candidates in a similar way. Therefore, both can be used for efficient active learning. The use of a diversity criterion seems to slightly improve the quality of the results (Fig.~\\ref{fig:diver}): except for the AVIRIS image -- well known for the high degree of mixture of its spectral signatures -- a spectral diversity criterion such as MCLU-ABD efficiently increases performances with little added computational cost. None of the solutions obtained with the inclusion of the diversity criterion degrade the ones relying on the uncertainty assumption only.\n\\begin{figure}[!t]\n\\begin{tabular}{ccc}\n& $N+5$ & $N+20$\\\\\n\\rotatebox{90}{ Pavia ROSIS}\n&\n\\includegraphics[width=3.8cm]{./images/Pavia_MCLU_ABD_10} \n&\n\\includegraphics[width=3.8cm]{./images/Pavia_MCLU_ABD_25} \\\\\n&(a)&(d)\\\\\n\\rotatebox{90}{ Indian Pines AVIRIS}\n&\n\\includegraphics[width=3.8cm]{./images/Pines_MCLU_ABD_17} \n&\n\\includegraphics[width=3.8cm]{./images/Pines_MCLU_ABD_30}\\\\\n&(b)&(e)\\\\\n\\rotatebox{90}{ Zurich QuickBird}\n&\n\\includegraphics[width=3.8cm]{./images/ZH_MR_MCLU_ABD_17}\n&\n\\includegraphics[width=3.8cm]{./images/ZH_MR_MCLU_ABD_30} \\\\\n&(c)&(f)\\\\\n\\end{tabular}\n\\caption{Effect of diversity criteria on large margin active learning. An example comparing MCLU and MCLU-ABD (RS = Random Sampling).}\n\\label{fig:diver}\n\\end{figure}\nAs stated above, other heuristics must be used for other classifiers. Figure~\\ref{fig:LDA} compares the $n$EQB and BT heuristics applied to the Pavia image using LDA. \nIn this case, both heuristics perform similarly and show a very interesting behavior: active sampling helps the LDA to estimate the subspace that better separates data. In fact, when sampling randomly, noise and outliers can make the estimation of the Fisher's ratio biased, resulting in a suboptimal linear combination of variables in the decision function. Sampling and assigning correct labels to the pixels returned by these heuristics help estimating the correct \\emph{per-class} extent (covariance) and position (mean). From 330 pixels up, the standard LDA result is improved by the active learning training sets, providing more harmonious solutions that allow a better generalization. \nSumming up, when using methods other than large margin-based algorithms, performances of the heuristics are similar and the choice must be driven by the specific constraints of time and number of iteration allowed. We will come back to these issues in the next section.\n\\begin{figure}[!t]\n\\centering\n\\begin{tabular}{cc}\n\\includegraphics[width=3.8cm]{./images/Pavia_comparisonLDA_10} &\n\\includegraphics[width=3.8cm]{./images/Pavia_comparisonLDA_25} \\\\\n(a) $N+5$ & (b) $N+20$\\\\\n\\end{tabular}\n\\caption{Committee-based and posterior probability heuristics trained with LDA classifiers on the Pavia ROSIS image (RS = Random Sampling).}\n\\label{fig:LDA}\n\\end{figure}","id":"16c4e1fd-cca0-4c03-98cb-9399b4dc0f51","level":"section","origin_cites_number":1,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Numerical results"]],"subsections":[],"title":"Numerical results"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:disc}\nThroughout the experiments presented, nearly all the algorithms compared showed fast convergence to the upper bounds represented by the Standard SVM/LDA. At convergence, all the heuristics outperformed the random selection of pixels. The interest of using multiclass uncertainty or adding a criterion of diversity has been demonstrated by the experiments above. Table~\\ref{tab:sum} summarizes the points raised in this paper and gives a general overview of the methods reviewed. \nLarge margin-based methods with diversity criterion seem the most appropriate when using SVM, while committee-based heuristic leave freedom to the user to exploit the model he is most confident with. Moreover, the user can build ensembles of classifiers exploiting the intrinsic advantages of specific classifiers for a given type of image. Weighted committees or boosting candidates are also possible (see Section \\ref{sec:committee} for some references). Probabilistic heuristics have the advantage of speed, but cannot always provide batches of samples and, if the classifier does not return such estimates naturally, must rely on further approximations of the posterior probabilities.\nHowever, it would not be correct to base the choice of the heuristic in a model-based fashion only. The choice of the best heuristic is problem-oriented and depends on the needs of the user in terms of time, complexity and size of the batch to be provided at each iteration. \nThis section draws some guidelines to select the most appropriate heuristic. \nA first distinction could be done depending on the type of the images considered: \n\\begin{itemize}\n\\item[-] when dealing with hyperspectral images, which are typically high dimensional, strategies taking direct advantage of the data structure should be preferred: typically, multi-view heuristics such as the AMD or the ECBD-ABD are particularly well-suited to this type of data. The first exploits cross-informations directly in the space of the spectral bands, while the second selects the samples according to spectral angles among the candidates.\n\\item[-] when the initial training set is very small, heuristics based on posterior probabilities should be avoided, since such estimation strongly depends on the quality of the estimation of the class statistics (typically in the case of LDA). The same holds for committees-based on bagging, especially if the bags contain a small share of the original samples.\n\\item[-] when dealing with complex manifolds, in which redundancy can greatly affect the quality of the sampling and of the resulting training set, approaches based on modeling of the relationships among samples in the feature space can be strongly beneficial to select pixels reflecting such complexity. The use of kernel $k$-means in the MCLU-ECBD, in hMCS or the distance to support vectors in the cSV heuristic provide solutions in this sense.\n\\end{itemize}\nA second distinction, more related to operational requirements, is based on the type of sampling to be performed by the user~, \n\\begin{itemize}\n\\item[-] when working by photointerpretation (typically in VHR imaging), sampling can be done on-screen directly by the user. This allows for large amounts of iterations and can thus be solved by small batches of samples. In this case complex heuristics favoring spectral diversity are to be preferred, since the complexity of the heuristics enforcing diversity strongly increases with the size of the batch considered. \n\\item[-] on the contrary, when sampling is to be done on the field (typically in hyperspectral or mid-resolution images), only a few iterations with large batches are allowed. In this case, all the heuristics seem to provide the same type of convergence and the user should prefer simple heuristics such as MCLU, BT or EQB, depending on the model used. In this case, the spatial location of samples seems to be much more important than the heuristic used: a pioneering work in this sense can be found in ~, where MS and BT are exploited with spatially adaptive cost and the sampling is optimized with respect to the spatial distance among the samples chosen.\n\\item[-] when sampling is done with moving sensors and the samples are acquired sequentially by an automatic device, batches of samples are not necessary. In this case models with small computational cost should be preferred, as they can update fast and almost instantly provide the next location to be sampled. In this case, BT and KL-max are most valuable. \n\\end{itemize}\n\\begin{table*}[!t]\n\\caption{Summary of active learning algorithms (B : binary, M : multiclass, $q$ : number of candidates, $k$ : members of the committee of learners, $S$ : batch, $SVs$ : support vectors, $\\checkmark$ : yes, $\\times$ : no).}\n\\label{tab:sum}\n\\begin{center}\n\\setlength{\\tabcolsep}{1pt}\n\\begin{tabular}{p{1.8cm}|p{1.6cm}|c|c|c|c|c c|p{4cm}}\n\\hline\nFamily&Heuristic& Reference &Batches&Uncer-& Classifier& \\multicolumn{2}{c|}{Diversity}&{Models to train}\\\\\n&&&& tainty&&$S$&$SV$s&\\\\\n\\hline\\hline\n\\multirow{2}{*}{Committee}&EQB&&\\checkmark&M&All&$\\times$ &$\\times$&$k$ models\\\\\n&AMD&&\\checkmark&M&All&$\\times$ &$\\times$&$k$ models\\\\\n\\hline\n\\multirow{9}{*}{Large margin}&MS&&\\checkmark&B& SVM &$\\times$ &$\\times$&Single SVM\\\\\n&MCLU&&\\checkmark&M&SVM&$\\times$ &$\\times$&Single SVM\\\\\n&SSC&&\\checkmark&B&SVM&$\\times$ &$\\times$&2 SVMs\\\\\n&cSV&&\\checkmark&B&SVM&$\\checkmark$ &$\\times$&Single SVM + distances to support vectors \\\\\n&MOA&&\\checkmark&B&SVM&\\checkmark &$\\times$&Single SVM + distances to already selected pixels\\\\\n&MCLU-ABD&&\\checkmark&M&SVM&\\checkmark&$\\times$&Single SVM + distances to already selected pixels\\\\\n&MCLU-ECBD&&\\checkmark&M&SVM&\\checkmark &$\\times$&Single SVM + nonlinear clustering of candidates\\\\\n&hMCS-i&&\\checkmark&M&SVM&\\checkmark &\\checkmark&Single SVM + nonlinear clustering of candidates and SVs\\\\\n\\hline\nPosterior&KL-max&&$\\times$&M&$$ $p(y|\\x)$&$\\times$&$\\times$&$(q-1)$ models\\\\\nprobability&BT&&\\checkmark&M& $ p(y|\\x)$&$\\times$&$\\times$&Single model\\\\\n\\hline\n\\end{tabular}\n\\end{center}\n\\end{table*}","id":"fba5b779-cfe8-4a33-833a-07b334d324ec","level":"section","origin_cites_number":10,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:concl}\nIn this paper we presented and compared several state of the art approaches to active learning for the classification of remote sensing images. A series of heuristics have been classified by their characteristics into four families. For each family, some heuristics have been detailed and then applied to {three challenging} remote sensing datasets for multispectral and hyperspectral classification. Advantages and drawbacks of each method have been analyzed in detail and recommendations for further improvement have been worded. \nHowever, this review is not exhaustive and the research in the field is far from being over: \nthere is a healthy and rich research community developing new heuristics for active sampling that have been or will be presented in the remote sensing and signal processing community. \nActive learning has a strong potential for remote sensing data processing. Efficient training sets are needed by the users, especially when dealing with large archives of digital images. New problems are being tackled with active learning algorithms, guaranteeing the renewal of the field. Some recent examples can be found in the active selection of unlabeled pixels for semi-supervised classification~, spatially adaptive heuristics~ or the use of active learning algorithms for model adaptation across domains~. \nFurther steps for active learning methods are the inclusion of contextual information in the heuristics: so far, the heuristics proposed only take advantage of spectral criteria -- or at most include contextual features in the data vector -- but few heuristics directly consider positional information and/or textures. Another crucial issue is the robustness to noise: since they are based on the uncertainty of the pixels, current heuristics are useless for images related to high levels of noise such as SAR. This field remains, at present, totally unexplored.\n\\section*{Acknowledgments}\nThe authors would like to acknowledge prof. Paolo Gamba (Univ. Pavia) who provided the Pavia dataset, as well as the authors in for the Indian Pines data.\n\\bibliographystyle{IEEEbib}\n\\bibliography{refsALreview}\n\\end{document}","id":"132e66fa-8f0c-4248-b231-33f0f600a11b","level":"section","origin_cites_number":5,"parent_id":"32f5c325-ad32-4a99-bf88-0a6a7380d5c1","prefix_titles":[["title","A survey of active learning algorithms for supervised remote sensing image classification"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ]\n ],\n \"subsections\": [\n \"9f8d3c3d-dd73-4be2-949f-6388307ef8c8\",\n \"0c24f63b-6c0b-4f7c-a826-45cfbfb05ce5\",\n \"9c33e713-e531-4890-ab4a-b33425a0f7dd\",\n \"4355cda7-2c05-483f-8ba7-6a88643cc3e2\",\n \"544044ff-dee5-45d4-b4e9-335bae0209f4\",\n \"fc8866e1-005e-4ac4-be0e-43641ad50928\",\n \"16c4e1fd-cca0-4c03-98cb-9399b4dc0f51\",\n \"fba5b779-cfe8-4a33-833a-07b334d324ec\",\n \"132e66fa-8f0c-4248-b231-33f0f600a11b\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.05,\n \"cites\": [\n 8400\n ],\n \"content\": \"\\\\PARstart{N}{owadays}, the recourse to statistical learning models~ is a common practice for remote sensing data users; models such as Support Vector Machines (SVM,~) or neural networks~ are considered as state of the art algorithms for the classification of landuse using new generation satellite imagery~. Applications of such models to very high spatial~ or spectral~ resolution have proven their efficiency for handling remote sensing data.\\nHowever, the performances of supervised algorithms strongly depend on the representativeness of the data used to train the classifier~. This constraint makes the generation of an appropriate training set a difficult and expensive task requiring extensive manual analysis of the image. This is usually done by visual inspection of the scene or by field surveys and successive labeling of each sample.\\nIn the case of field surveys, which is usual for medium resolution, hyperspectral or SAR images, the discovery of a new label is expensive -- both in terms of time and money -- because it involves terrain campaigns. Therefore, there is a limit to the number of pixels that can be acquired. For this reason, compact and informative training sets are needed. \\nIn the case of visual inspection or photo-interpretation, more common in VHR imagery, it is easier to collect data samples, since the labeling can be done directly on the image. However, the labeling is often done by mass selection on screen and several neighboring pixels carrying the same information are included. As a consequence, the training set is highly redundant. Such a redundancy considerably slows down the training phase of the model. \\\\mitch{, because several pixels carrying the same information are evaluated.}{} Moreover, the inclusion of noisy pixels may result in a wrong representation of the class statistics, which may lead to poor classification performances and/or overfitting~. For these reasons, a training set built by photointerpretation should also be kept as small as possible and focused on those pixels effectively improving the performance of the model.\\nSumming up, besides being small, a desirable training set must be constructed in a smart way, meaning it must represent correctly the class \\\\mitch{statistics and covering the entirety of the data variability}{boundaries by sampling discriminative pixels}. This is particularly critical in very high spatial and spectral resolution image classification, which deal with large \\\\mitch{}{and/or} complex features spaces using limited training information only~. \\nIn the machine learning literature this approach to sampling is known as \\\\emph{active learning}. The leading idea is that a classifier trained on a small set of well-chosen examples can perform as well as a classifier trained on a larger number of randomly chosen examples, while it is computationally cheaper~. Active learning focuses on the interaction between the user and the classifier. In other words, the model returns to the user the pixels whose classification outcome is the most uncertain. After accurate labeling by the user, pixels are included into the training set in order to reinforce the model. This way, the model is optimized on well-chosen difficult examples, maximizing its generalization capabilities.\\nThe active learning framework has demonstrated its effectiveness when applied to large datasets needing accurate selection of examples~. This is suitable for remote sensing applications, where the number of pixels among which the search is performed is large and manual definition is - as stated above - redundant and time consuming.\\nAs a consequence, active learning algorithms gain an increasing interest in remote sensing image processing and several approaches have been proposed to solve image classification tasks. This paper presents the general framework of active learning and reviews some of the methods that have been proposed in remote sensing literature. Note that this survey only covers remote sensing application of active learning principles: for a general introduction and survey of the most recent developments in the machine learning community, please refer to~.\\nThe remainder of the paper is organized as follows: Section~\\\\ref{sec:AL} presents the general framework of active learning and the families of methods that will be detailed in \\\\red{Sections~\\\\ref{sec:committee} to~\\\\ref{sec:poster}}, as well as the references to specific methods. Section~\\\\ref{sec:data} presents the datasets considered in the experiments. Section~\\\\ref{sec:res} compares the different approaches numerically. Section~\\\\ref{sec:disc} gives an overview and guidelines for potential users. Section~\\\\ref{sec:concl} concludes the paper.\",\n \"id\": \"9f8d3c3d-dd73-4be2-949f-6388307ef8c8\",\n \"level\": \"section\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:AL}\\nLet $X~=~\\\\{\\\\x_i,y_i\\\\}_{i=1}^l$ be a training set of labeled samples, with $\\\\x_i \\\\in \\\\mathcal{X}$ and $y_i = \\\\{1, ..., N\\\\}$. $\\\\mathcal{X}$ is the $d$-dimensional input space $\\\\in \\\\mathbb{R}^d$. Let also $U~=~\\\\{\\\\x_i\\\\}_{i=l+1}^{l+u} \\\\in \\\\mathcal{X}$, with $u \\\\gg l$ be the set of unlabeled pixels to be sampled, or the \\\\emph{pool of candidates}. \\nActive learning algorithms are iterative sampling schemes, where a classification model is adapted regularly by feeding it with new labeled pixels corresponding to the ones that are most beneficial for the improvement of the model performance. These pixels are usually found in the areas of \\\\emph{uncertainty} of the model and their inclusion in the training set forces the model to solve the \\\\mitch{uncertainties}{regions of low confidence}. For a given iteration $\\\\epsilon$, the algorithm selects from the pool $U^\\\\epsilon$ the $q$ candidates that will at the same time maximize the gain in performance and reduce the uncertainty of the model when added to the current training set $X^\\\\epsilon$. Once the batch of pixels $S^\\\\epsilon = \\\\{\\\\x_m\\\\}_{m=1}^q \\\\subset U$ has been selected, it is labeled by the user, i.e. the labels $\\\\{y_m\\\\}_{m=1}^q$ are discovered. Finally, the set $S^\\\\epsilon$ is both added to the current training set ($X^{\\\\epsilon+1} = X^\\\\epsilon \\\\cup S^\\\\epsilon$) and removed from the pool ($U^{\\\\epsilon+1} = U^\\\\epsilon \\\\backslash S^\\\\epsilon$). The process is iterated until a stopping criterion is met. Algorithm~\\\\ref{alg:AL} summarizes the active selection process. From now on, the iteration index $\\\\epsilon$ will be omitted in order to ease notation. \\n\\\\begin{algorithm}[!t]\\n\\\\caption{General active learning algorithm}\\n\\\\label{alg:AL}\\n\\\\vspace{0.2cm}\\n\\\\textbf{Inputs}\\\\\\\\\\n- Initial training set $X^\\\\epsilon = \\\\{\\\\x_i,y_i\\\\}_{i=1}^l$ ($X \\\\in \\\\mathcal{X}$, $\\\\epsilon=1$).\\\\\\\\\\n- Pool of candidates $U^\\\\epsilon = \\\\{\\\\x_i\\\\}_{i=l+1}^{l+u}$ ($U \\\\in \\\\mathcal{X}$, $\\\\epsilon=1$).\\\\\\\\\\n- Number of pixels $q$ to add at each iteration (defining the batch of selected pixels $S$).\\n\\\\vspace{0.2cm}\\n\\\\begin{algorithmic}[1]\\n\\\\REPEAT\\n\\t\\\\STATE Train a model with current training set $X^\\\\epsilon$.\\n\\t\\\\FOR{each candidate in $U^\\\\epsilon$}\\n\\t\\\\STATE Evaluate a user-defined \\\\emph{heuristic}\\n\\t\\\\ENDFOR\\n\\t\\\\STATE Rank the candidates in $U^\\\\epsilon$ according to the score of the heuristic\\n\\t\\\\STATE Select the $q$ most interesting pixels. $S^\\\\epsilon = \\\\{\\\\x_k\\\\}_{k=1}^q$\\n\\t\\\\STATE The user assigns a label to the selected pixels. $S^\\\\epsilon = \\\\{\\\\x_k,y_k\\\\}_{k=1}^q$\\n\\t\\\\STATE Add the batch to the training set $X^{\\\\epsilon+1}=X^{\\\\epsilon} \\\\cup S^\\\\epsilon$.\\n\\t\\\\STATE Remove the batch from the pool of candidates $U^{\\\\epsilon+1}=U^{\\\\epsilon} \\\\backslash S^\\\\epsilon$\\n\\t\\\\STATE $\\\\epsilon = \\\\epsilon + 1$\\n\\\\UNTIL{a stopping criterion is met.}\\n\\\\end{algorithmic}\\n\\\\vspace{0.2cm}\\n\\\\end{algorithm}\\nAn active learning process requires interaction between the user and the model: the first provides the labeled information and the knowledge about the desired \\\\blue{classes}, while the latter provides both its own interpretation of the distribution of the classes and the most relevant pixels that are needed in order to solve the discrepancies encountered. This point is crucial for the success of an active learning algorithm: the machine needs a strategy to rank the pixels in the pool $U$. \\nThese strategies, or \\\\emph{heuristics}, differentiate the algorithms proposed in the next sections and can be divided into three main families~: \\n\\\\begin{itemize}\\n\\t\\\\item[1 -] \\\\emph{Committee}-based heuristics (Section~\\\\ref{sec:committee})\\n\\t\\\\item[2 -] \\\\emph{Large margin}-based heuristics (Section~\\\\ref{sec:lm})\\n\\t\\\\item[3 -] \\\\emph{Posterior probability}-based heuristics (Section~\\\\ref{sec:poster})\\n\\\\end{itemize}\\nA last family of active learning heuristics, the \\\\emph{cluster}-based, has recently being proposed in the community~: cluster-based heuristics aim at pruning a hierarchical clustering tree until the resulting clusters are consistent with the labels provided by the user. Therefore, these strategies rely on an unsupervised model, rather than on a predictive model. Since the aim of these heuristics is different form that of the other families presented, they will not be detailed in this survey.\",\n \"id\": \"0c24f63b-6c0b-4f7c-a826-45cfbfb05ce5\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Active learning: concepts and definitions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Active learning: concepts and definitions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:committee}\\nThe first family of active learning methods quantifies the uncertainty of a pixel by considering a committee of learners~. Each member of the committee exploits different hypotheses about the classification problem and consequently labels the pixels in the pool of candidates. The algorithm then selects the samples showing maximal disagreement between the different classification models in the committee. \\nTo limit computational complexity, heuristics based on multiple classifier systems have been proposed in machine learning literature. In~, methods based on boosting and bagging are proposed in this sense for binary classification only. In~, results obtained by query-by-boosting and query-by-bagging are compared using several batch datasets showing excellent performance of the heuristics proposed. Methods of this family have the advantage to be applicable to any kind of model or combination of models. \\nIn the remote sensing community, committee-based approaches to active learning have been proposed exploiting two types of uncertainty: first, committees varying the pixels members have been considered in the query-by-bagging heuristic~. Then, committees based on subsets of the feature space available have been presented in~Di and Crawford~. The next two sections present the algorithms proposed in these papers.\",\n \"id\": \"9c33e713-e531-4890-ab4a-b33425a0f7dd\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Committee based active learning\"\n ]\n ],\n \"subsections\": [\n \"dd9d8542-ab43-4865-9ca1-5532a86beb4d\",\n \"d0e71651-ef54-4cae-9092-e430dfc27eb8\"\n ],\n \"title\": \"Committee based active learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:EQB}\\nIn the implementations of , bagging is proposed to build the committee: first, $k$ training sets built on a draw with replacement of the original data are defined. These draws account for a part of the available labeled pixels only. Then, each set is used to train a classifier and to predict the $u$ labels of the candidates. At the end of the procedure, $k$ labelings are provided for each candidate pixel $\\\\x_i \\\\in U$.\\nIn~, the entropy $H^{\\\\text{BAG}}$ of the distribution of the predictions \\\\mitch{provided by the $k$ classifiers for each pixel $\\\\x_i$ in $U$}{} is used as heuristic. In~, this measure has been subsequently normalized in order to bound it with respect to the number of classes predicted by the committee and avoid hot spots of the value of uncertainty in regions where several classes overlap. The \\\\emph{normalized entropy query-by-bagging} heuristic can be stated as follows:\\n\\\\begin{equation}\\n\\t\\\\hat{\\\\x}^{\\\\text{\\\\emph{n}EQB}} = \\\\arg\\\\max_{\\\\x_i \\\\in U}\\\\Big\\\\{\\\\frac{H^{\\\\text{BAG}}(\\\\x_i)}{ \\\\mbox{log}(N_i)}\\\\Big\\\\}\\n\\\\label{eq:neqb}\\n\\\\end{equation}\\nwhere\\n\\\\begin{align}\\n\\t&H^{\\\\text{BAG}}(\\\\x_i) = -\\\\sum_{\\\\omega=1}^{N_i} p^{\\\\text{BAG}}(y_i^* = \\\\omega|\\\\x_i) \\\\mbox{log}\\\\left[p^{\\\\text{BAG}}(y^*_i= \\\\omega|\\\\x_i)\\\\right]\\\\\\\\\\n\\t\\t& \\\\text{where} \\\\qquad p^{\\\\text{BAG}}(y^*_i = \\\\omega|\\\\x_i) = \\\\frac{\\\\sum_{m=1}^k \\\\delta(y_{i,m}^{*}, \\\\omega )}{\\\\sum_{m=1}^k\\\\sum_{j = 1}^{N_i} \\\\delta(y_{i,m}^*, \\\\omega_j ) } \\\\nonumber \\n\\\\label{eq:entropy}\\n\\\\end{align}\\n$H^{\\\\text{BAG}}(\\\\x_i)$ is an empirical measure of entropy, $y_i^*$ is the prediction for the pixel $\\\\x_i$ and $p^{\\\\text{BAG}}(y^*_i = \\\\omega|\\\\x_i)$ is the observed probability to have the class $\\\\omega$ predicted using the training set $X$ by the committee of $k$ models for the sample $\\\\x_i$. $N_i$ is the number of classes predicted for pixel $\\\\x_i$ \\\\mitch{by the committee}{}, with $1 \\\\leq N_i \\\\leq N$. The $\\\\delta(y_{i,m}^*,\\\\omega)$ operator returns the value $1$ if the classifier using the $m$-th bag classifies the sample $\\\\x_i$ into class $\\\\omega$ and $0$ otherwise. Entropy maximization gives a naturally multiclass heuristic. A candidate for which all the classifiers in the committee agree is associated with null entropy and its inclusion in the training set does not bring additional information. On the contrary, a candidate with maximum disagreement between the classifiers results in maximum entropy, and its inclusion will be highly beneficial.\",\n \"id\": \"dd9d8542-ab43-4865-9ca1-5532a86beb4d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"9c33e713-e531-4890-ab4a-b33425a0f7dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Committee based active learning\"\n ],\n [\n \"subsection\",\n \"Normalized entropy query-by-bagging (nEQB)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Normalized entropy query-by-bagging (nEQB)\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8401\n ],\n \"content\": \"When confronted to high dimensional data, it may be relevant to construct the committee by splitting the feature space into a number of subsets, or \\\\emph{views}~. Di and Crawford~ exploit this principle to generate different views of a hyperspectral image on the basis of the block-diagonal structure of the covariance matrix. By generating views corresponding to the different blocks, independent classifications of the same pixel can be generated and an entropy-based heuristic can be used similarly to $n$EQB.\\nGiven a partition of the $d$-dimensional input space into $V$ disjoint views accounting for data subsets $\\\\x^v$ such that $\\\\bigcup_{v=1}^V \\\\x^v = \\\\x $, the `Adaptive maximum disagreement ' (AMD) heuristic selects candidates according to the following rule:\\n\\\\begin{equation}\\n\\t\\\\hat{\\\\x}^{\\\\text{AMD}} = \\\\arg\\\\max_{\\\\x_i \\\\in U} H^{\\\\text{MV}}(\\\\x_i)\\n\\t\\\\label{eq:AMD}\\n\\\\end{equation}\\nwhere the multiview entropy $H^{\\\\text{MV}}$ is assessed over the predictions of classifiers using a specific view $v$:\\n\\\\begin{align}\\n\\t&H^{\\\\text{MV}}(\\\\x_i) = -\\\\sum_{\\\\omega=1}^{N_i} p^{\\\\text{MV}}(y^*_{i,v} = \\\\omega|\\\\x^v_i) \\\\mbox{log}\\\\left[p^{\\\\text{MV}}(y^*_{i,v}= \\\\omega|\\\\x^v_i)\\\\right]\\\\\\\\\\n\\t& \\\\text{where} \\\\qquad p^{\\\\text{MV}}(y^*_i = \\\\omega|\\\\x^v_i) = \\\\frac{\\\\sum_{v=1}^V W^{\\\\epsilon-1}(v,\\\\omega) \\\\delta(y^*_{i,v}, \\\\omega )}{\\\\sum_{v=1}^V\\\\sum_{j = 1}^{N_i} W^{\\\\epsilon-1}(v,\\\\omega)} \\\\nonumber\\n\\\\label{eq:MVentr}\\n\\\\end{align}\\nwhere the $\\\\delta(y_{i,v}^*,\\\\omega )$ operator returns the value $1$ if the classifier using the view $v$ classifies the sample $y_i$ into class $\\\\omega$ and $0$ otherwise. $\\\\mathbf{W}^{\\\\epsilon-1}$ is a $N \\\\times V$ weighting matrix incorporating the abilities of discrimination between the views in the different classes. At each iteration, $\\\\mathbf{W}^{\\\\epsilon-1}$ is updated on the basis of the true labels of the pixels sampled at iteration $\\\\epsilon-1$:\\n\\\\begin{equation}\\nW^\\\\epsilon(v,\\\\omega) = W^{\\\\epsilon-1}(v,\\\\omega)+\\\\delta(y_{i,v},\\\\omega ), \\\\quad \\\\forall i \\\\in S\\n\\\\label{eq:W}\\n\\\\end{equation}\\nand its columns are normalized to a unitary sum. This matrix weights the confidence of each view to predict a given class. In~, the selection is done on a subset of $U$ containing the candidate pixels maximizing the uncertainty, which are the pixels for which the committee has predicted the highest number of classes. This way, the computational load of the algorithm is reduced.\",\n \"id\": \"d0e71651-ef54-4cae-9092-e430dfc27eb8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"9c33e713-e531-4890-ab4a-b33425a0f7dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Committee based active learning\"\n ],\n [\n \"subsection\",\n \"Adaptive maximum disagreement (AMD)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adaptive maximum disagreement (AMD)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:lm}\\nThe second family of methods is specific to margin-based classifiers. \\nMethods such as SVM are naturally good base methods for active learning: the distance to the separating hyperplane, that is the absolute value of the decision function without the sign operator, is a straightforward way of estimating the classifier confidence on an unseen sample. Let's first consider a binary problem: the distance of a sample $\\\\x_i$ from the SVM hyperplane is given by\\n\\\\begin{equation}\\nf(\\\\x_i)=\\\\sum_{j=1}^n \\\\alpha_j y_j K(\\\\x_j,\\\\x_i)+b\\n\\\\label{eq:decfunct}\\n\\\\end{equation}\\nwhere $K(\\\\x_j, \\\\x_i)$ is a kernel, which defines the similarity between the candidate $\\\\x_i$ and the support vectors $\\\\x_j$, which are the pixels showing non zero $\\\\alpha_j$ coefficients. The labels $y_j$ of the support vectors are $+1$ for samples of the positive class and $-1$ for those on the negative. For additional information, see SVM literature in~. \\nThis evaluation of the distance is the base ingredient of almost all large margin heuristics. Roughly speaking, these heuristics use the intuition that a sample away from the decision boundary (with a high $f(\\\\x_i)$) has a high confidence about its class assignment and is thus not interesting for future sampling.\\nSince SVM rely on a sparse representation of the data, large margin based heuristics aim at finding the pixels in $U$ that are most likely to receive a non-zero $\\\\alpha_i$ weight if added to $X$. In other words, the points more likely to become support vectors are the ones lying within the margin of the current model~. The heuristic taking advantage of this property is called margin sampling (MS)~. Recent modifications of MS, aiming at minimizing the risk of selecting points that will not become support vectors, can be found in~.\\nMS is the most studied active learning algorithm in remote sensing. Its first application can be found in~. Modifications of the MS heuristic have been proposed in~. Later on, since no cross-information among the samples is considered in the MS, the questions of diversity in batches of samples have been considered in~. The next sections present the MS heuristic and subsequent modifications proposed in order to enhance diversity when selecting batches of samples.\",\n \"id\": \"4355cda7-2c05-483f-8ba7-6a88643cc3e2\",\n \"level\": \"section\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Large margin based active learning\"\n ]\n ],\n \"subsections\": [\n \"46a14016-6a16-4632-a90d-002c27d86ea9\",\n \"9dde6920-321f-4866-969e-ae6a532e831e\",\n \"b4b3fbe3-f9dd-4869-b53d-b3ea1f2f3618\",\n \"7bae26d1-94ed-4e43-ac7f-09791c25101b\"\n ],\n \"title\": \"Large margin based active learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As stated above, margin sampling takes advantage of SVM geometrical properties, and in particular of the fact that unbounded support vectors are labeled examples that lie on the margin with a decision function value of exactly one~. Consider the pool of candidates of Fig.~\\\\ref{fig:dist}(a) referring to a three classes toy problem. In a multiclass one-against-all setting, the distance to each hyperplane is represented by Figs.~\\\\ref{fig:dist}(d-f). \\nThe `margin sampling' (MS) heuristic performs sampling of the candidates by minimizing Eq.~\\\\eqref{eq:minMarg}:\\n\\\\begin{equation}\\n\\\\hat{\\\\x}^{\\\\text{MS}} = \\\\arg\\\\min_{\\\\x_i \\\\in U } \\\\Big\\\\{\\\\min_{\\\\omega}|f(\\\\x_i,\\\\omega)|\\\\Big\\\\}\\n\\\\label{eq:minMarg}\\n\\\\end{equation}\\nwhere $f(\\\\x_i,\\\\omega)$ is the distance of the sample to the hyperplane defined for class $\\\\omega$ in a one-against-all setting for multiclass problems. The MS heuristic for the toy problem is reported in~Fig.~\\\\ref{fig:dist}(b). MS heuristic can be found in the literature under the names of `most ambiguous'~, `binary level uncertainty'~ or SVM$_\\\\text{SIMPLE}$~.\\n\\\\begin{figure*}\\n\\\\centering\\n\\\\begin{tabular}{|ccc|}\\n\\\\hline\\n\\\\includegraphics[width=4.2cm]{\\\\figdir Tuia_1a}&\\n\\\\includegraphics[width=4.2cm]{\\\\figdir distance_tot}&\\n\\\\includegraphics[width=4.2cm]{\\\\figdir distance_MCLU}\\\\\\\\\\n(a)&(b) MS, Eq.~\\\\eqref{eq:minMarg}&(c) MCLU, Eq.~\\\\eqref{eq:MCLUfct}\\\\\\\\\\n\\\\includegraphics[width=4.2cm]{\\\\figdir distance_cl1}&\\n\\\\includegraphics[width=4.2cm]{\\\\figdir distance_cl2}&\\n\\\\includegraphics[width=4.2cm]{\\\\figdir distance_cl3}\\\\\\\\\\n(d) $|f(\\\\x,\\\\omega_0)|$&(e) $|f(\\\\x,\\\\omega_1)|$&(f) $|f(\\\\x,\\\\omega_2)|$\\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Large margin heuristics for a three classes toy example represented in subfigure (a). The color intensity represents the distance from the hyperplane, ranging from black (on the boundary) to white (maximal distance): (b) MS heuristic; (c) MCLU heuristic; areas in black are the areas of maximal uncertainty, minimizing Eq.~\\\\eqref{eq:minMarg} or Eq.~\\\\eqref{eq:MCLUfct} respectively. \\nBottom row: absolute values of per-class distances (d)-(f). }\\n\\\\label{fig:dist}\\n\\\\end{figure*}\",\n \"id\": \"46a14016-6a16-4632-a90d-002c27d86ea9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"4355cda7-2c05-483f-8ba7-6a88643cc3e2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Large margin based active learning\"\n ],\n [\n \"subsection\",\n \"Margin sampling (MS)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Margin sampling (MS)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:MCLU}\\nIn~, the idea of margin sampling is extended to multiclass uncertainty (see~). Instead of dealing with the most uncertain class of the SVM, the `multiclass level uncertainty' (MCLU) considers the difference between the distance to the margin for the two most probable classes. \\n\\\\begin{align}\\n&\\\\hat{\\\\x}^{\\\\text{MCLU}} = \\\\arg \\\\min_{\\\\x_i \\\\in U} \\\\Big\\\\{f(\\\\x_i)^{\\\\text{MC}}\\\\Big\\\\}\\\\label{eq:MCLU}\\\\\\\\\\n&\\\\text{where} \\\\qquad f(\\\\x_i)^{\\\\text{MC}} = \\\\max_{\\\\omega \\\\in N}|f(\\\\x_i,\\\\omega)|-\\\\max_{\\\\omega \\\\in N\\\\backslash \\\\omega^+}|f(\\\\x_i,\\\\omega)|\\\\label{eq:MCLUfct}\\n\\\\end{align}\\nwhere $\\\\omega^+$ is the class showing maximal confidence, i.e. the argument of the first term of Eq.~\\\\eqref{eq:MCLUfct} showing maximal $f(\\\\x_i)^{\\\\text{MC}}$. A high value of this criterion corresponds to samples assigned with high certainty to the most confident class, while a small value represents unreliable classification. Fig.~\\\\ref{fig:dist}(c) illustrates the heuristic in comparison to MS. Although they are very similar, MCLU performs better in the area where the three classes mix, in the top-right area of the feature space: in this area, MCLU returns maximal uncertainty as it is evaluated on the basis of all the per-class decision values, while MS returns an uncertainty slightly lower than on the two-classes boundaries.\",\n \"id\": \"9dde6920-321f-4866-969e-ae6a532e831e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4355cda7-2c05-483f-8ba7-6a88643cc3e2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Large margin based active learning\"\n ],\n [\n \"subsection\",\n \"Multiclass level uncertainty (MCLU)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multiclass level uncertainty (MCLU)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In~, instead of using the distance to the hyperplane as a measure of uncertainty, the support vector coefficients are used to convert the multiclass classification problem into a binary support vector detection problem. In the `significance space construction' (SSC) heuristic, the training samples related to support vector coefficients are used to define a second classification function $f(\\\\x)^{\\\\text{SSC}}$, where training pixels with $\\\\alpha_j > 0$ (the support vectors) are classified against training pixels with $\\\\alpha_j = 0$. Once applied to the pool of candidates, this second classifier predicts which pixels are likely to become support vectors. \\n\\\\begin{equation}\\n\\\\hat{\\\\x}^{\\\\text{SSC}} = \\\\arg_{\\\\x_i \\\\in U} f(\\\\x_i)^{\\\\text{SSC}} > 0\\n\\\\label{eq:SSC}\\n\\\\end{equation}\\nOnce the candidates more likely to become support vectors have been highlighted, a random selection among them is done.\",\n \"id\": \"b4b3fbe3-f9dd-4869-b53d-b3ea1f2f3618\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"4355cda7-2c05-483f-8ba7-6a88643cc3e2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Large margin based active learning\"\n ],\n [\n \"subsection\",\n \"Significance space construction (SSC)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Significance space construction (SSC)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In applicative scenarios, diversity among samples~ is highly desirable. Diversity concerns the capability of the model to reject candidates that rank well according to the heuristic, but are redundant among each other. Diversity has been studied extensively for margin-based heuristics, where the base margin sampling heuristic is constrained using a measure of diversity between the candidates (see Algorithm~\\\\ref{alg:diversity}). \\n\\\\begin{algorithm}[!b]\\n\\\\caption{General diversity based heuristic (for a single iteration)}\\n\\\\label{alg:diversity}\\n\\\\vspace{0.2cm}\\n\\\\textbf{Inputs}\\\\\\\\\\n- Current training set $X^{\\\\epsilon} = \\\\{\\\\x_i,y_i\\\\}_{i=1}^l$ ($X \\\\in \\\\mathcal{X}$).\\\\\\\\\\n- Subset of the pool of candidates minimizing Eq.~\\\\eqref{eq:minMarg} or~\\\\eqref{eq:MCLU} $U^* = \\\\{\\\\x_i\\\\}$ $ (U^* \\\\in \\\\mathcal{X}$ and $U^* \\\\subset U^{\\\\epsilon}$).\\\\\\\\\\n- Number of pixels $q$ to add at each iteration (defining the batch of selected pixels $S$).\\n\\\\vspace{0.2cm}\\n\\\\begin{algorithmic}[1]\\n\\\\STATE Add the pixel minimizing Eq.~\\\\eqref{eq:minMarg} or~\\\\eqref{eq:MCLUfct} to $S$.\\n\\\\REPEAT{}\\n\\t\\\\STATE Compute the user defined diversity criterion between pixels in $U^*$ and in $S$ (with MAO, cSV or ABD).\\n\\t\\\\STATE Select the pixel $\\\\x_D$ maximizing diversity with current batch.\\n\\t\\\\STATE Add $\\\\x_D$ to current batch $S = S \\\\cup \\\\x_D$.\\n\\t\\\\STATE Remove $\\\\x_D$ to current list of cadidates $U^* = U^* \\\\setminus \\\\x_D$.\\n\\\\UNTIL{batch $S$ contains $q$ elements.}\\n\\\\STATE The user labels the selected pixels. $S = \\\\{\\\\x_k,y_k\\\\}_{k=1}^q$\\n\\\\STATE Add the batch to the training set $X^{\\\\epsilon+1}=X^{\\\\epsilon} \\\\cup S$.\\n\\\\STATE Remove the batch from the complete pool of candidates $U^{\\\\epsilon+1}=U^{\\\\epsilon} \\\\backslash S$\\n\\\\end{algorithmic}\\n\\\\vspace{0.2cm}\\n\\\\end{algorithm}\\nThe first heuristic proposing explicit diversity in remote sensing is found in~, where the margin sampling heuristic is constrained with a measure of the angle between candidates in feature space. This heuristic, called `most ambiguous and orthogonal' (MAO) is iterative: starting from the samples selected by MS, \\\\mitch{}{$U^{\\\\text{MS}}\\\\subset U$}, this heuristic iteratively chooses the samples minimizing the highest values between the candidates list and the samples already included in the batch \\\\mitch{of samples}{} $S$. For a single iteration, this can be resumed as:\\n\\\\begin{equation}\\n\\\\hat{\\\\x}^{\\\\text{MAO}} = \\\\arg \\\\min_{\\\\x_i \\\\in U^{\\\\text{MS}} }\\\\Big\\\\{\\\\max_{\\\\x_j \\\\in S}K(\\\\x_i,\\\\x_j)\\\\Big\\\\}\\n\\\\label{eq:MAO}\\n\\\\end{equation}\\nIn~, the MAO criterion is combined with the MCLU uncertainty estimation in the `multiclass level uncertainty - angle-based diversity' (MCLU-ABD) heuristic. Te selection is performed among a subset of $U$ maximizing the MCLU criterion. Moreover, the author generalize the MAO heuristic to any type of kernels by including normalization in feature space.\\n\\\\begin{align}\\n\\\\hat{\\\\x}^{\\\\text{MCLU-ABD}} = \\\\arg \\\\min_{\\\\x_i \\\\in U^{\\\\text{MCLU}} }&\\\\Bigg\\\\{\\\\lambda f(\\\\x_i)^{\\\\text{MC}} + \\\\\\\\\\\\nonumber\\n&(1-\\\\lambda)\\\\max_{\\\\x_j \\\\in S}\\\\frac{K(\\\\x_i,\\\\x_j)}{\\\\sqrt{K(\\\\x_i,\\\\x_i)K(\\\\x_j,\\\\x_j)}}\\\\Bigg\\\\}\\\\\\\\\\n\\\\label{eq:MCLU-ABD}\\n\\\\end{align}\\nwhere $f(\\\\x_i)^{\\\\text{MC}}$ is the multiclass uncertainty function defined by Eq.~\\\\eqref{eq:MCLUfct}.\\nIn~, the diversity of the chosen candidates is enforced by constraining the MS solution to pixels associated to different closest support vectors. This approach ensures a certain degree of diversification in the MS heuristic, by dividing the margin in the feature space as a function of the geometrical distribution of the support vectors. Compared to the previously presented heuristics, this approach has the advantage of ensuring diversity with respect to the current model, but does not guarantee diversity of the samples between each other (since two close samples can be associated to different support vectors).\\n\\\\begin{equation}\\n\\\\hat{\\\\x}^{\\\\text{cSV}} = \\\\arg \\\\min_{\\\\x_i \\\\in U^{\\\\text{MS}} } \\\\Big\\\\{\\\\lvert f(\\\\x_i,\\\\omega)\\\\rvert \\\\ \\\\Big\\\\vert \\\\ cSV_i \\\\not\\\\in cSV_\\\\theta\\\\Big\\\\}\\n\\\\label{eq:cSV}\\n\\\\end{equation}\\nwhere $\\\\theta = [1, \\\\ldots, q-1]$ are the indices of the already selected candidates and $cSV$ is the set of selected closest support vectors.\\nFinally, diversity can be ensured using clustering in the feature space. In~, kernel $k$-means~ was used to cluster the samples selected by MCLU and select diverse batches. After partitioning the $U^{\\\\text{MCLU}}$ set into $q$ clusters with kernel $k$-means, the `multiclass level uncertainty - enhanced cluster based diversity (MCLU-ECBD)' selects a single pixel per cluster, minimizing the following query function:\\n\\\\begin{equation}\\n\\\\hat{\\\\x}^{\\\\text{MCLU-ECBD}} = \\\\arg \\\\min_{\\\\x_i \\\\in c_m}\\\\Big\\\\{ f(\\\\x_i)^{\\\\text{MC}}\\\\Big\\\\}\\\\text{,} \\\\quad m = [1, ... q], \\\\x_i \\\\in U^{\\\\text{MCLU}}\\n\\\\label{eq:MCLU-ECBD}\\n\\\\end{equation}\\nwhere $c_m$ is one among the $q$ clusters defined with kernel $k$-means. \\nIn~, a hierarchical extension of this principle is proposed to exclude from the selected batch the pixels more likely to become bounded support vectors. This way, the redundancy affecting samples close to each other in the feature space among different iterations is controlled along with the maximization of the informativeness of each pixel.\\nIn the `informative hierarchical margin cluster sampling' (hMCS-i), a dataset composed by i) a subset of the pool of candidates optimizing the MCLU criterion ($U^{\\\\text{MCLU}}$) and ii) the bounded support vectors sampled at the previous iteration, is iteratively partitioned in a binary way. The partitioning always considers the biggest current cluster found and continues until $q$ clusters not containing a bounded support vector are found. Once the $q$ clusters have been defined, a search among the candidates falling in these $q$ clusters is performed.\\n\\\\begin{equation}\\n\\\\hat{\\\\x}^{\\\\text{hMCS-i}} = \\\\arg \\\\min_{\\\\x_i \\\\in c_m } \\\\Big\\\\{f(\\\\x_i)^{\\\\text{MC}} \\\\Big\\\\}\\\\text{,}\\\\quad m = [1, ..., q \\\\vert n^{bSV}_{c_m} = 0], \\\\x_i \\\\in U^{\\\\text{MCLU}}\\n\\\\label{eq:hMCS}\\n\\\\end{equation}\",\n \"id\": \"7bae26d1-94ed-4e43-ac7f-09791c25101b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"4355cda7-2c05-483f-8ba7-6a88643cc3e2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Large margin based active learning\"\n ],\n [\n \"subsection\",\n \"On the need for a diversity criterion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"On the need for a diversity criterion\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:poster}\\nThe third class of methods uses the estimation of posterior probabilities of class membership (i.e. $p( y | \\\\x)$) to rank the candidates. The posterior probability gives an idea of the confidence of the class assignment (which is usually done by maximizing it over all the possible classes): by considering the change on the overall posterior distribution or the per-class distribution for each candidate, these heuristics use these probability estimates to focus sampling in uncertain areas. This section details two heuristics, the KL-max and the Breaking ties.\",\n \"id\": \"544044ff-dee5-45d4-b4e9-335bae0209f4\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Posterior probability based active learning\"\n ]\n ],\n \"subsections\": [\n \"e1f6b836-abb9-4d52-abf9-04c6538aa022\",\n \"e604938a-7c05-4eae-a00c-0f7093b23169\"\n ],\n \"title\": \"Posterior probability based active learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The first idea is to sample the pixels whose inclusion in the training set would maximize the changes in the posterior distribution. An application of these methods can be found in ~, where the heuristic maximizes the Kullback-Leibler divergence between the distributions before and after adding the candidate. In remote sensing, a probabilistic method based on this strategy and using a Maximum Likelihood classifier can be found in~. In this setting, each candidate is removed from $U$ and it is included in the training set with the label maximizing its posterior probability. The Kullbach-Leibler divergence KL is then computed between the posterior distributions of the models with and without the candidate. After computing this measure for all candidates, the pixel maximizing the following is chosen:\\n{\\\\small\\n\\\\begin{align}\\n&\\\\hat{\\\\x}^{\\\\text{KL-max}} = \\\\\\\\\\\\nonumber &= \\\\arg \\\\max_{\\\\x_i \\\\in U } \\\\Bigg\\\\{\\\\sum_{\\\\omega \\\\in N} \\\\frac{1}{(u-1)} \\\\text{KL}\\\\Big(p^+(\\\\omega|\\\\x)\\\\Big\\\\vert\\\\Big\\\\vert p(\\\\omega|\\\\x)\\\\Big)p(y^*_i = \\\\omega|\\\\x_i)\\\\Bigg\\\\}\\\\label{eq:Kmax}\\n\\\\end{align} }\\nwhere\\n{\\\\small\\n\\\\begin{align}\\n \\\\text{KL}\\\\Big(p^+( \\\\omega|\\\\x)\\\\Big\\\\vert\\\\Big\\\\vert p( \\\\omega|\\\\x)\\\\Big) = \\\\sum_{\\\\x_j \\\\in U \\\\setminus \\\\x_i} p^+(y^*_j = \\\\omega|\\\\x_j)\\\\log\\\\frac{p^+(y^*_j = \\\\omega|\\\\x_j)}{p(y^*_j = \\\\omega|\\\\x_j)}\\n\\\\end{align} }\\nand $p^+(\\\\omega|\\\\x)$ is the posterior distribution for class $\\\\omega$ and pixel $\\\\x$, estimated using the increased training set $X^+ = [X,(\\\\x_i,y^*_i)]$, where $y_i^*$ is the class maximizing the posterior probability. Recently, the authors of~ extended this approach, proposing to use boosting to weight pixels that were previously selected, but were no longer relevant to the current classifier. These heuristics are useful when used with classifiers with small computational cost: since each iteration implies to train $u+1$ models, this type of heuristics is hardly applicable with computationally demanding methods as SVM. Moreover, a selection of batches of pixels is not possible.\",\n \"id\": \"e1f6b836-abb9-4d52-abf9-04c6538aa022\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"544044ff-dee5-45d4-b4e9-335bae0209f4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Posterior probability based active learning\"\n ],\n [\n \"subsection\",\n \"KL-max\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"KL-max\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Another strategy, closer to the idea of EQB presented in Section \\\\ref{sec:EQB}, consists of building a heuristic exploiting the conditional probability of predicting a given label $p(y_i^* = \\\\omega|\\\\x_i)$ for each candidate $\\\\x_i \\\\in U$. In this case, note that the predictions for the single candidates $y_i^* = \\\\arg\\\\max_{\\\\omega \\\\in N}f(\\\\x_i,\\\\omega)$ are used. Such estimates are provided by several methods, including probabilistic neural networks or maximum likelihood classifiers. A possibility to obtain posterior probabilities from SVM outputs\\\\footnote{Even though they are not posterior probabilities from a Bayesian point of view} is to use Platt's estimation~. \\nIn this case, the per-class posterior probability is assessed fitting a sigmoid function to the SVM decision function~:\\n\\\\begin{equation}\\n\\tp(y_i^* = \\\\omega|\\\\x_i) =\\\\frac{1}{1 + e^{(Af(\\\\x_i,\\\\omega)+B)}}\\n\\\\end{equation}\\n\\\\mitch{where $\\\\omega = \\\\{1,... ,N\\\\}$ is the class considered and $f(\\\\x_i,\\\\omega)$ is the binary decision function of the SVM without the sign operator (see Eq.~\\\\eqref{eq:decfunct}).}{} where $A$ and $B$ are parameters to be estimated (for details, see~). \\nOnce the posterior probabilities are obtained, it is possible to assess the uncertainty of the class membership for each candidate in a direct way. In this case the heuristic choses candidates showing a near uniform probability of belonging to each class, i.e. $p( y_i^* = \\\\omega | \\\\x_i) = 1/N, \\\\ \\\\forall \\\\omega \\\\in N$.\\nThe `Breaking ties' (BT) heuristic for a binary problem relies on the smallest difference of the posterior probabilities for each sample . In a multi-class setting, this reciprocity can still be confirmed and used, since independently from the number of classes $N$, the difference between the two highest probabilities can be indicative of the way an example is handled by the classifier\\\\mitch{(see a similar interpretation for the MCLU heuristic in Section~\\\\ref{sec:MCLU})}{}. When the two highest probabilities are close (``on a tie''), the classifier's confidence is the lowest. The BT heuristic can thus be formulated as:\\n\\\\begin{equation}\\n\\\\hat{\\\\x}^{\\\\text{BT}} = \\\\arg \\\\min_{\\\\x_i \\\\in U } \\\\Big\\\\{\\\\max_{\\\\omega \\\\in N} \\\\big\\\\{p(y_i^* = \\\\omega|\\\\x)\\\\big\\\\} - \\\\max_{\\\\omega \\\\in \\\\mitch{}{N}\\\\backslash \\\\omega^+} \\\\big\\\\{p(y_i^* = \\\\omega|\\\\x)\\\\big\\\\}\\\\Big\\\\}\\n\\\\label{eq:BT}\\n\\\\end{equation}\\nwhere $\\\\omega^+$ is the class showing maximal probability, i.e. the argument of the first term of Eq.~\\\\eqref{eq:BT}. By comparing Eq.~\\\\eqref{eq:MCLU} with Eq.~\\\\eqref{eq:BT}, it is clear that the link between BT and the MCLU heuristic when using SVM classifiers (see Section~\\\\ref{sec:MCLU}) is really strong.\",\n \"id\": \"e604938a-7c05-4eae-a00c-0f7093b23169\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"544044ff-dee5-45d4-b4e9-335bae0209f4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Posterior probability based active learning\"\n ],\n [\n \"subsection\",\n \"Breaking ties (BT)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Breaking ties (BT)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:data}\\nThis Section details the datasets considered and the setup of the experiments performed.\",\n \"id\": \"fc8866e1-005e-4ac4-be0e-43641ad50928\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Datasets and setup\"\n ]\n ],\n \"subsections\": [\n \"6e70c940-20c9-4ab7-ad99-8a4761d0a98e\",\n \"bbc1da7f-38e6-4e9a-a5cf-edcfa29f4d5e\"\n ],\n \"title\": \"Datasets and setup\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Active heuristics have been tested on three challenging remote sensing classification scenarios (Fig.~\\\\ref{fig:data}), whose data distributions are detailed in Fig.~\\\\ref{fig:Distr}. \\n\\\\begin{figure}[!b]\\n\\\\centering\\n\\t\\\\begin{tabular}{cc}\\n\\t\\\\multicolumn{2}{c}{\\\\includegraphics[width=8.4cm]{\\\\figdir Pavia}}\\\\\\\\\\n\\t\\\\multicolumn{2}{c}{\\\\includegraphics[width=8.4cm]{\\\\figdir PaviaGT}}\\\\\\\\\\n\\t\\\\multicolumn{2}{c}{ROSIS Pavia}\\\\\\\\\\n\\t\\\\includegraphics[height=4cm]{\\\\figdir Pines_40_30_20}&\\n\\t\\\\includegraphics[height=4cm]{\\\\figdir pinesGT}\\\\\\\\\\n\\t\\\\multicolumn{2}{c}{AVIRIS Indian Pines}\\\\\\\\\\n\\\\includegraphics[height=4cm]{\\\\figdir imageZH_MR}&\\n\\\\includegraphics[height=4cm]{\\\\figdir ZH_MR_GT}\\\\\\\\\\n\\t\\\\multicolumn{2}{c}{QuickBird Zurich}\\\\\\\\\\n\\t\\\\end{tabular}\\n\\t\\\\caption{Images considered in the experiments: (top) ROSIS image of the city of Pavia, Italy (bands $[56-31-6]$ and corresponding ground survey); (middle) AVIRIS Indian Pines hyperspectral data (bands $[40-30-20]$ and corresponding ground survey); (bottom) QuickBird multispectral image of a suburb of the city of Zurich, Switzerland (bands $[3-2-1]$ and corresponding ground survey).}\\n\\t\\\\label{fig:data}\\n\\\\end{figure}\\n\\\\begin{figure}\\n\\\\centering\\n\\\\begin{tabular}{ccc}\\n\\\\includegraphics[width=4cm]{\\\\figdir MeanSpectraPavia}&\\n\\\\includegraphics[width=4cm]{\\\\figdir testLabels}\\\\\\\\\\n(a)&(b)\\\\\\\\\\n\\\\includegraphics[width=4cm]{\\\\figdir MeanSpectraPines}&\\n\\\\includegraphics[width=4cm]{\\\\figdir manifold_52_102_208}\\n\\\\\\\\\\n(c)&(d)\\\\\\\\\\n\\\\includegraphics[width=4cm]{\\\\figdir ZHspectrum_ext2}&\\n\\\\includegraphics[width=4cm]{\\\\figdir ZHmanifold}\\n\\\\\\\\\\n(e) & (f)\\\\\\\\\\n\\\\end{tabular}\\n\\\\caption{Data distribution of the three images considered. First row: ROSIS image of Pavia: (a) mean spectral profiles; (b) example of data manifold in bands 55 (Red) and 77 (Near infrared). Middle row: AVIRIS Indian Pines: (c) mean spectral profiles; (d) example of data manifold in bands 52, 102 and 208. Bottom row: Zurich QuickBird: (e) mean spectral profiles; (f) data manifold in bands 2 (G), 3 (R) and 4 (NIR).}\\n\\\\label{fig:Distr}\\n\\\\end{figure}\",\n \"id\": \"6e70c940-20c9-4ab7-ad99-8a4761d0a98e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"fc8866e1-005e-4ac4-be0e-43641ad50928\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Datasets and setup\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [\n \"0d621b05-e50b-4902-ba8d-1510e19e9756\",\n \"5744c67c-bb8c-4a84-a58d-895f062fc480\",\n \"2fba1c13-bf38-48d4-a59c-6a88de2989ac\"\n ],\n \"title\": \"Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"the two top rows of Fig.~\\\\ref{fig:data} show a hyperspectral 1.3m spatial resolution image of the city of Pavia (Italy) taken by the airborne ROSIS-03 optical sensor~. The image consists of 102 spectral bands of size ($1400 \\\\times 512$) pixels with a spectral coverage ranging from 0.43 to 0.86 $\\\\mu$m. $5$ classes of interest (Buildings, Roads, Water, Vegetation and Shadows) have been selected and a labeled dataset of 206`009 pixels has been extracted by visual inspection. Among the available pixels, 20`000 have been used for the training set $X$ and candidate set $U$. Ten independent experiments have been performed, starting with $5 \\\\times 5 = 25$ labeled pixels ($5$ per class) in $X$ and the remaining pixels in $U$. When using LDA, 150 pixels ($30$ per class) have been included in the starting set. The higher number of starting training pixels used for LDA is justified by the requirements of the model ($n$ must be greater than the dimensionality of the data). In each experiment, 80`000 randomly selected pixels have been used to test the generalization capabilities of the heuristics.\\nThe data distribution of the five classes is illustrated in the first row of Fig.~\\\\ref{fig:Distr}: from the mean spectra (Fig.~\\\\ref{fig:Distr}(a)) the classes are well distinguished and separable with the sole spectral information and the resulting data manifold (Fig.~\\\\ref{fig:Distr}(b)) shows a data distribution which can be handled by most linear and non linear models.\",\n \"id\": \"0d621b05-e50b-4902-ba8d-1510e19e9756\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"6e70c940-20c9-4ab7-ad99-8a4761d0a98e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Datasets and setup\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"Hyperspectral VHR\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hyperspectral VHR\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"the second dataset, illustrated in the second row of Fig.~\\\\ref{fig:data}, is a 220-bands AVIRIS image taken over Indiana's Indian Pine test site in June 1992~. The image is $145 \\\\times 145$ pixels, contains 16 classes representing different crops, and a total of 10`366 labeled pixels. This image is a classical benchmark to validate model accuracy and constitutes a very challenging classification problem because of the strong mixture of the class signatures. Twenty water absorption channels were removed prior to analysis\\n. In the experiments, classes with less than 100 labeled pixels were removed, resulting thus in a 12 classes classification problem with 10`171 labeled pixels (see the ground truth pixels in Fig.~\\\\ref{fig:data}). Among the available labeled pixels, 7`000 were used for the $X$ and $U$ sets. Each experiment starts with $5 \\\\times 12 =60$ pixels (5 per class). As for the previous image, the remaining 3`171 pixels have been used to test the generalization capabilities.\\nVisualization of the spectral mean profiles and of the data manifold (second row of Fig.~\\\\ref{fig:Distr}) illustrates a completely different situation with respect to the previous image: high nonlinearity and strongly overlapping classes characterize this dataset. Therefore, linear classifiers do not perform well on this dataset and will not be considered in the experiments.\",\n \"id\": \"5744c67c-bb8c-4a84-a58d-895f062fc480\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"6e70c940-20c9-4ab7-ad99-8a4761d0a98e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Datasets and setup\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"Hyperspectral MR\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hyperspectral MR\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The third image, illustrated in the last row of Fig.~\\\\ref{fig:data}, is a 4-bands QuickBird scene of a suburb of the city of Zurich (Switzerland) taken in 2002. The image is $329 \\\\times 347$ pixels with a spatial resolution of 2.4m. Nine classes of interest have been extracted by careful visual inspection, namely Residential buildings, Commercial buildings, Trees, Vegetation, Harvested fields, Bare soil, Roads, Parking lots and Water. Since some of the classes to be separated are of landuse and have very similar responses in the spectral domain (see, for instance, the residential and commercial buildings, or roads and parking lots), 16 contextual bands, extracted using opening (8) and closing (8) morphological operators (see~), have been added to the four spectral bands, resulting in a 20-dimensional dataset. As for the Pavia dataset, 20`000 pixels have been extracted for the $X$ and $U$ sets. Each experiment starts with $5 \\\\times 9 =45$ pixels ($5$ per class). \\nThe complexity of this third dataset is confirmed by both the spectra and the manifold illustrated in the bottom row of Fig.~\\\\ref{fig:Distr}. Strong overlaps between the asphalt and the soil classes are observed, which is also confirmed by the similarity between the spectral profiles. However, the spatial features added improve the differentiation of the classes (see, for instance, the opening features for the vegetation classes and the closing features for the asphalt classes).\",\n \"id\": \"2fba1c13-bf38-48d4-a59c-6a88de2989ac\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"6e70c940-20c9-4ab7-ad99-8a4761d0a98e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Datasets and setup\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"Multispectral VHR\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multispectral VHR\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In the experiments, SVM classifiers with RBF kernel and LDA classifiers have been considered for the experiments. When using SVM, free parameters have been optimized by 5-fold cross validation optimizing an accuracy criterion. The active learning algorithms have been run in two settings, adding $N+5$ and $N+20$ pixels per iteration. To reach convergence, $70$ ($40$ in the case $N + 20$) iterations have been executed for the first image, $100$ ($50$) for the second and $80$ ($50$) for the third. $n$EQB has been run with committees of $7$ models using $75\\\\%$ of the available training data. For the experiments using LDA, $40$ (20) iterations have been performed and $n$EQB using $12$ models and $85\\\\%$ of the data have been used. \\nAn upper bound on the classification accuracy has been computed by considering a model trained on the whole $X \\\\cup U$ set (`Standard SVM/LDA' hereafter). The lower bound on performance has been considered by assessing a model using an increasing training set, randomly adding the same number of pixels at each epoch (`Random Sampling' hereafter). \\nEach heuristic has been run ten times with different initial training sets. All the graphics report mean and standard deviation of the ten experiments.\",\n \"id\": \"bbc1da7f-38e6-4e9a-a5cf-edcfa29f4d5e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"fc8866e1-005e-4ac4-be0e-43641ad50928\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Datasets and setup\"\n ],\n [\n \"subsection\",\n \"Experimental setup\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Experimental setup\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:res}\\nIn this section, some of the heuristics presented are compared on the three datasets presented above. \\nThe experiments do not aim at defining which heuristic is best, since they respond unequally well to different data architectures. Rather, it attempts to illustrate the strengths and weaknesses of the different families of methods and to help the user in selecting the methodology that will perform best depending on the characteristics of the problem at hand. \\nThe heuristics studied are the following: $n$EQB, MS, MCLU, MCLU-ABD and BT. Their comparison with Random sampling (RS), the base learner (Standard SVM/LDA) and between each other will show the main differences among the active learning architectures presented. \\n\\\\begin{figure}[!t]\\n\\\\begin{tabular}{ccc}\\n& $N+5$ & $N+20$\\\\\\\\\\n\\\\rotatebox{90}{ Pavia ROSIS}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_comparison_10}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_comparison_25}\\\\\\\\\\n&(a)&(d)\\\\\\\\\\n\\\\rotatebox{90}{ Indian Pines AVIRIS}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pines_comparison_17}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pines_comparison_30}\\n\\\\\\\\\\n&(b)&(e)\\\\\\\\\\n\\\\rotatebox{90}{ Zurich QuickBird}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/ZH_MR_comparison_14}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/ZH_MR_comparison_25}\\n\\\\\\\\\\n&(c)&(f)\\\\\\\\\\n\\\\end{tabular}\\n\\\\caption{The three families of heuristics trained with SVMs (RS = Random Sampling).}\\n\\\\label{fig:comp}\\n\\\\end{figure}\\nFigure~\\\\ref{fig:comp} compares a heuristic for each family presented, when using SVM classifiers. In general, MS performs better than the two other families. This is expected, since MS ranks the candidates directly using the SVM decision function without further estimations: the slightly inferior performances of the $n$EQB and the BT algorithms are mainly due to the small size of the initial training set, which does not allow these criteria to perform optimally. $n$EQB uses too small bags of training pixels and BT cannot estimate the posterior probabilities correctly, because the fit of Platt's probabilities is dependent on the number of samples used. As a consequence, the performances of these two heuristics in the first iterations is similar to random sampling, a behavior already observed in~. Summarizing, when using SVMs, the most logical choice among the families seems to be a large margin heuristic.\\n\\\\begin{figure}[!b]\\n\\\\begin{tabular}{ccc}\\n& $N+5$ & $N+20$\\\\\\\\\\n\\\\rotatebox{90}{ Pavia ROSIS}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_MS_MCLU_10}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_MS_MCLU_25} \\\\\\\\\\n&(a)&(d)\\\\\\\\\\n\\\\rotatebox{90}{ Indian Pines AVIRIS}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pini_MS_MCLU_17}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pini_MS_MCLU_40} \\\\\\\\\\n&(b)&(e)\\\\\\\\\\n\\\\rotatebox{90}{ Zurich QuickBird}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/ZH_MR_MS_MCLU_14}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/ZH_MR_MS_MCLU_30} \\\\\\\\\\n&(c)&(f)\\\\\\\\\\n\\\\end{tabular}\\n\\\\caption{Large margin active learning without diversity criterion. An example comparing MS and MCLU (RS = Random Sampling).}\\n\\\\label{fig:largeComp}\\n\\\\end{figure}\\nRegarding this family, Figs.~\\\\ref{fig:largeComp} and~\\\\ref{fig:diver} illustrate two concepts regarding the two stages of large margin heuristics: the uncertainty and diversity criteria. Figures~\\\\ref{fig:largeComp} compares the MS and MCLU criteria and shows that both describe the uncertainty of the candidates in a similar way. Therefore, both can be used for efficient active learning. The use of a diversity criterion seems to slightly improve the quality of the results (Fig.~\\\\ref{fig:diver}): except for the AVIRIS image -- well known for the high degree of mixture of its spectral signatures -- a spectral diversity criterion such as MCLU-ABD efficiently increases performances with little added computational cost. None of the solutions obtained with the inclusion of the diversity criterion degrade the ones relying on the uncertainty assumption only.\\n\\\\begin{figure}[!t]\\n\\\\begin{tabular}{ccc}\\n& $N+5$ & $N+20$\\\\\\\\\\n\\\\rotatebox{90}{ Pavia ROSIS}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_MCLU_ABD_10} \\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_MCLU_ABD_25} \\\\\\\\\\n&(a)&(d)\\\\\\\\\\n\\\\rotatebox{90}{ Indian Pines AVIRIS}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pines_MCLU_ABD_17} \\n&\\n\\\\includegraphics[width=3.8cm]{./images/Pines_MCLU_ABD_30}\\\\\\\\\\n&(b)&(e)\\\\\\\\\\n\\\\rotatebox{90}{ Zurich QuickBird}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/ZH_MR_MCLU_ABD_17}\\n&\\n\\\\includegraphics[width=3.8cm]{./images/ZH_MR_MCLU_ABD_30} \\\\\\\\\\n&(c)&(f)\\\\\\\\\\n\\\\end{tabular}\\n\\\\caption{Effect of diversity criteria on large margin active learning. An example comparing MCLU and MCLU-ABD (RS = Random Sampling).}\\n\\\\label{fig:diver}\\n\\\\end{figure}\\nAs stated above, other heuristics must be used for other classifiers. Figure~\\\\ref{fig:LDA} compares the $n$EQB and BT heuristics applied to the Pavia image using LDA. \\nIn this case, both heuristics perform similarly and show a very interesting behavior: active sampling helps the LDA to estimate the subspace that better separates data. In fact, when sampling randomly, noise and outliers can make the estimation of the Fisher's ratio biased, resulting in a suboptimal linear combination of variables in the decision function. Sampling and assigning correct labels to the pixels returned by these heuristics help estimating the correct \\\\emph{per-class} extent (covariance) and position (mean). From 330 pixels up, the standard LDA result is improved by the active learning training sets, providing more harmonious solutions that allow a better generalization. \\nSumming up, when using methods other than large margin-based algorithms, performances of the heuristics are similar and the choice must be driven by the specific constraints of time and number of iteration allowed. We will come back to these issues in the next section.\\n\\\\begin{figure}[!t]\\n\\\\centering\\n\\\\begin{tabular}{cc}\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_comparisonLDA_10} &\\n\\\\includegraphics[width=3.8cm]{./images/Pavia_comparisonLDA_25} \\\\\\\\\\n(a) $N+5$ & (b) $N+20$\\\\\\\\\\n\\\\end{tabular}\\n\\\\caption{Committee-based and posterior probability heuristics trained with LDA classifiers on the Pavia ROSIS image (RS = Random Sampling).}\\n\\\\label{fig:LDA}\\n\\\\end{figure}\",\n \"id\": \"16c4e1fd-cca0-4c03-98cb-9399b4dc0f51\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Numerical results\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Numerical results\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:disc}\\nThroughout the experiments presented, nearly all the algorithms compared showed fast convergence to the upper bounds represented by the Standard SVM/LDA. At convergence, all the heuristics outperformed the random selection of pixels. The interest of using multiclass uncertainty or adding a criterion of diversity has been demonstrated by the experiments above. Table~\\\\ref{tab:sum} summarizes the points raised in this paper and gives a general overview of the methods reviewed. \\nLarge margin-based methods with diversity criterion seem the most appropriate when using SVM, while committee-based heuristic leave freedom to the user to exploit the model he is most confident with. Moreover, the user can build ensembles of classifiers exploiting the intrinsic advantages of specific classifiers for a given type of image. Weighted committees or boosting candidates are also possible (see Section \\\\ref{sec:committee} for some references). Probabilistic heuristics have the advantage of speed, but cannot always provide batches of samples and, if the classifier does not return such estimates naturally, must rely on further approximations of the posterior probabilities.\\nHowever, it would not be correct to base the choice of the heuristic in a model-based fashion only. The choice of the best heuristic is problem-oriented and depends on the needs of the user in terms of time, complexity and size of the batch to be provided at each iteration. \\nThis section draws some guidelines to select the most appropriate heuristic. \\nA first distinction could be done depending on the type of the images considered: \\n\\\\begin{itemize}\\n\\\\item[-] when dealing with hyperspectral images, which are typically high dimensional, strategies taking direct advantage of the data structure should be preferred: typically, multi-view heuristics such as the AMD or the ECBD-ABD are particularly well-suited to this type of data. The first exploits cross-informations directly in the space of the spectral bands, while the second selects the samples according to spectral angles among the candidates.\\n\\\\item[-] when the initial training set is very small, heuristics based on posterior probabilities should be avoided, since such estimation strongly depends on the quality of the estimation of the class statistics (typically in the case of LDA). The same holds for committees-based on bagging, especially if the bags contain a small share of the original samples.\\n\\\\item[-] when dealing with complex manifolds, in which redundancy can greatly affect the quality of the sampling and of the resulting training set, approaches based on modeling of the relationships among samples in the feature space can be strongly beneficial to select pixels reflecting such complexity. The use of kernel $k$-means in the MCLU-ECBD, in hMCS or the distance to support vectors in the cSV heuristic provide solutions in this sense.\\n\\\\end{itemize}\\nA second distinction, more related to operational requirements, is based on the type of sampling to be performed by the user~, \\n\\\\begin{itemize}\\n\\\\item[-] when working by photointerpretation (typically in VHR imaging), sampling can be done on-screen directly by the user. This allows for large amounts of iterations and can thus be solved by small batches of samples. In this case complex heuristics favoring spectral diversity are to be preferred, since the complexity of the heuristics enforcing diversity strongly increases with the size of the batch considered. \\n\\\\item[-] on the contrary, when sampling is to be done on the field (typically in hyperspectral or mid-resolution images), only a few iterations with large batches are allowed. In this case, all the heuristics seem to provide the same type of convergence and the user should prefer simple heuristics such as MCLU, BT or EQB, depending on the model used. In this case, the spatial location of samples seems to be much more important than the heuristic used: a pioneering work in this sense can be found in ~, where MS and BT are exploited with spatially adaptive cost and the sampling is optimized with respect to the spatial distance among the samples chosen.\\n\\\\item[-] when sampling is done with moving sensors and the samples are acquired sequentially by an automatic device, batches of samples are not necessary. In this case models with small computational cost should be preferred, as they can update fast and almost instantly provide the next location to be sampled. In this case, BT and KL-max are most valuable. \\n\\\\end{itemize}\\n\\\\begin{table*}[!t]\\n\\\\caption{Summary of active learning algorithms (B : binary, M : multiclass, $q$ : number of candidates, $k$ : members of the committee of learners, $S$ : batch, $SVs$ : support vectors, $\\\\checkmark$ : yes, $\\\\times$ : no).}\\n\\\\label{tab:sum}\\n\\\\begin{center}\\n\\\\setlength{\\\\tabcolsep}{1pt}\\n\\\\begin{tabular}{p{1.8cm}|p{1.6cm}|c|c|c|c|c c|p{4cm}}\\n\\\\hline\\nFamily&Heuristic& Reference &Batches&Uncer-& Classifier& \\\\multicolumn{2}{c|}{Diversity}&{Models to train}\\\\\\\\\\n&&&& tainty&&$S$&$SV$s&\\\\\\\\\\n\\\\hline\\\\hline\\n\\\\multirow{2}{*}{Committee}&EQB&&\\\\checkmark&M&All&$\\\\times$ &$\\\\times$&$k$ models\\\\\\\\\\n&AMD&&\\\\checkmark&M&All&$\\\\times$ &$\\\\times$&$k$ models\\\\\\\\\\n\\\\hline\\n\\\\multirow{9}{*}{Large margin}&MS&&\\\\checkmark&B& SVM &$\\\\times$ &$\\\\times$&Single SVM\\\\\\\\\\n&MCLU&&\\\\checkmark&M&SVM&$\\\\times$ &$\\\\times$&Single SVM\\\\\\\\\\n&SSC&&\\\\checkmark&B&SVM&$\\\\times$ &$\\\\times$&2 SVMs\\\\\\\\\\n&cSV&&\\\\checkmark&B&SVM&$\\\\checkmark$ &$\\\\times$&Single SVM + distances to support vectors \\\\\\\\\\n&MOA&&\\\\checkmark&B&SVM&\\\\checkmark &$\\\\times$&Single SVM + distances to already selected pixels\\\\\\\\\\n&MCLU-ABD&&\\\\checkmark&M&SVM&\\\\checkmark&$\\\\times$&Single SVM + distances to already selected pixels\\\\\\\\\\n&MCLU-ECBD&&\\\\checkmark&M&SVM&\\\\checkmark &$\\\\times$&Single SVM + nonlinear clustering of candidates\\\\\\\\\\n&hMCS-i&&\\\\checkmark&M&SVM&\\\\checkmark &\\\\checkmark&Single SVM + nonlinear clustering of candidates and SVs\\\\\\\\\\n\\\\hline\\nPosterior&KL-max&&$\\\\times$&M&$$ $p(y|\\\\x)$&$\\\\times$&$\\\\times$&$(q-1)$ models\\\\\\\\\\nprobability&BT&&\\\\checkmark&M& $ p(y|\\\\x)$&$\\\\times$&$\\\\times$&Single model\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{center}\\n\\\\end{table*}\",\n \"id\": \"fba5b779-cfe8-4a33-833a-07b334d324ec\",\n \"level\": \"section\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:concl}\\nIn this paper we presented and compared several state of the art approaches to active learning for the classification of remote sensing images. A series of heuristics have been classified by their characteristics into four families. For each family, some heuristics have been detailed and then applied to {three challenging} remote sensing datasets for multispectral and hyperspectral classification. Advantages and drawbacks of each method have been analyzed in detail and recommendations for further improvement have been worded. \\nHowever, this review is not exhaustive and the research in the field is far from being over: \\nthere is a healthy and rich research community developing new heuristics for active sampling that have been or will be presented in the remote sensing and signal processing community. \\nActive learning has a strong potential for remote sensing data processing. Efficient training sets are needed by the users, especially when dealing with large archives of digital images. New problems are being tackled with active learning algorithms, guaranteeing the renewal of the field. Some recent examples can be found in the active selection of unlabeled pixels for semi-supervised classification~, spatially adaptive heuristics~ or the use of active learning algorithms for model adaptation across domains~. \\nFurther steps for active learning methods are the inclusion of contextual information in the heuristics: so far, the heuristics proposed only take advantage of spectral criteria -- or at most include contextual features in the data vector -- but few heuristics directly consider positional information and/or textures. Another crucial issue is the robustness to noise: since they are based on the uncertainty of the pixels, current heuristics are useless for images related to high levels of noise such as SAR. This field remains, at present, totally unexplored.\\n\\\\section*{Acknowledgments}\\nThe authors would like to acknowledge prof. Paolo Gamba (Univ. Pavia) who provided the Pavia dataset, as well as the authors in for the Indian Pines data.\\n\\\\bibliographystyle{IEEEbib}\\n\\\\bibliography{refsALreview}\\n\\\\end{document}\",\n \"id\": \"132e66fa-8f0c-4248-b231-33f0f600a11b\",\n \"level\": \"section\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"32f5c325-ad32-4a99-bf88-0a6a7380d5c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A survey of active learning algorithms for supervised remote sensing image classification\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":5,"string":"5"},"all_cites":{"kind":"list like","value":[8400,8401],"string":"[\n 8400,\n 8401\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.49650125766033,"string":"1.496501"}}},{"rowIdx":37,"cells":{"authors":{"kind":"list like","value":["A. Oelen","M. Y. Jaradeh","M. Stocker","S. Auer"],"string":"[\n \"A. Oelen\",\n \"M. Y. Jaradeh\",\n \"M. Stocker\",\n \"S. Auer\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-06-02T16:19:00Z","string":"2020-06-02T16:19:00Z"},"category":{"kind":"string","value":"cs.DL"},"abstract":{"kind":"string","value":" Reviewing scientific literature is a cumbersome, time consuming but crucial activity in research. Leveraging a scholarly knowledge graph, we present a methodology and a system for comparing scholarly literature, in particular \\emph{research contributions} describing the addressed problem, utilized materials, employed methods and yielded results. The system can be used by researchers to quickly get familiar with existing work in a specific research domain (e.g., a concrete research question or hypothesis). Additionally, it can be used to publish literature surveys following the FAIR Data Principles. The methodology to create a research contribution comparison consists of multiple tasks, specifically: (a) finding similar contributions, (b) aligning contribution descriptions, (c) visualizing and finally (d) publishing the comparison. The methodology is implemented within the Open Research Knowledge Graph (ORKG), a scholarly infrastructure that enables researchers to collaboratively describe, find and compare research contributions. We evaluate the implementation using data extracted from published review articles. The evaluation also addresses the FAIRness of comparisons published with the ORKG. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"]],"subsections":["b5fdf9c3-044e-45b3-8bb1-0b7d691c74d6","b95d3fb3-5cd7-4c69-8c96-6cadc3ef7b13","32d3f690-7c93-4310-b607-6553d1915112","a1917066-53fd-432d-9c65-ea4825cda6c7","def65367-1d9c-4420-92cd-cd2f5cd6b277","9c67205a-5ea7-42a6-af06-049ac9b5d00b","f87ef12c-02fb-41f5-9f16-69dee5190666","4c8c53d6-d473-464f-a034-d85b2b90268e"],"title":"root"},{"cite_extract_rate":0.18181818181818102,"cites":[6548,6549],"content":"When conducting scientific research, reviewing the existing literature is an essential activity~. \nFamiliarity with the state-of-the-art is required to effectively contribute to advancing it and do relevant research. Mainly because published scholarly knowledge is unstructured~, it is currently very tedious to review existing literature. Relevant literature has to be found among hundreds and increasingly thousands of PDF articles. This activity is supported by library catalogs and online search engines, such as Scopus or Google Scholar~. Because the search is keyword based, typically large numbers of articles are returned by search engines. Researchers have to manually identify the relevant papers. Having identified the relevant papers, the relevant pieces of information need to be extracted in order to obtain an overview of the literature. Overall, these are manual and time consuming steps. We argue that a key issue is that the scholarly knowledge communicated in the literature does not meet the FAIR Data Principles~. While PDF articles can be found and accessed (assuming Open Access or an institutional subscription), the scholarly literature is insufficiently interoperable and reusable, especially for machines. For units more granular than the PDF article, such as a specific result, findability and accessibility score low even for humans.\nWe present a methodology and its implementation integrated into the Open Research Knowledge Graph (ORKG)~ that can be used to generate and publish literature surveys in form of machine actionable, comparable descriptions of research contributions. Machine actionability of research contributions relates to the ability of machines to access and interpret the contribution data. The benefits for researchers of such an infrastructure are (at least) two-fold. Firstly, it supports researchers in creating state-of-the-art overviews for specific research problems efficiently. Secondly, it supports researchers in publishing literature surveys that adhere to the FAIR principles, thus contributing substantially to reuse of state-of-the-art overviews and therein contained information, for both humans and machines. \nLiterature reviews are articles that focus on analysing existing literature. Among other things, reviews can be used to gain understanding about a research problem or to identify further research directions~. Reviews can be used by authors to quickly obtain an overview of either emerging or mature research topics~. Review papers are important for research fields to develop. When review papers are lacking, the development of a research field is weakened~. Compiling literature review papers is a complicated task~ and is often more time consuming than performing original research~. The structure of such articles often consists of tables that compare published research contributions. Although in the literature the terms ``literature review'' and ``literature survey'' are sometimes used interchangeably, we make the following distinction. We refer to the tables in review articles as \\textit{literature surveys}. Together with a (textual) analysis and explanation, they form the \\textit{literature review}. The state-of-the-art (SoTA) analysis is a special kind of literature review with the objective of comparing the latest and most relevant papers in a specific domain. \nWe implement the presented methodology in the ORKG. The ORKG is a scholarly infrastructure designed to acquire, publish and process \\emph{structured} scholarly knowledge published in the literature~. ORKG is part of a larger research agenda aiming at machine actionable scholarly knowledge that understands the ability to more efficiently compare literature as a key feature. \nWe tackle the following research questions: \n\\begin{enumerate}\n \\item How to generate literature surveys using scholarly knowledge graphs?\n \\item How to ensure that published literature surveys comply with the FAIR principles?\n \\item How to effectively specify and visualize literature surveys in a user interface?\n\\end{enumerate}\nIn support of the first research question, we present a methodology that describes the steps required to generate literature surveys. In support of the second research question, we describe how the FAIRness of the published literature review is ensured. Finally, in support of the third research question, we demonstrate how the methodology is implemented within the ORKG. \nThe paper is structured as follows. Section~\\ref{section:motivating-example} motivates the work. Section \\ref{section:related-literature} reviews related work. Section~\\ref{section:system-design} presents the system design, the underlying methodology and its implementation. Section~\\ref{section:data-collection} explains how the knowledge graph is populated with data. Section~\\ref{section:evaluation} presents the evaluation of the system, specifically system FAIRness and performance. Finally, Section~\\ref{section:discussion} discusses the presented and future work.","id":"b5fdf9c3-044e-45b3-8bb1-0b7d691c74d6","level":"section","origin_cites_number":11,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:motivating-example}\nWe motivate our work by means of two use cases that underscore the usefulness of a literature survey generation system. In the first use case, a researcher wants to obtain an overview on state-of-the-art research addressing a specific problem. The second use case describes how a researcher can publish a FAIR compliant literature review with the ORKG.","id":"b95d3fb3-5cd7-4c69-8c96-6cadc3ef7b13","level":"section","origin_cites_number":0,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Use cases"]],"subsections":["86a170e1-f5ee-40dd-9553-058683d000c4"],"title":"Use cases"},{"cite_extract_rate":0,"cites":[],"content":"A state-of-the-art (SoTA) analysis reviews new and emerging research. They are useful for multiple reasons. Firstly, they provide a broad overview of a research problem and support understanding. Secondly, they juxtapose different approaches for a problem. Thirdly, they can support claims on why certain research is relevant by giving an overview of the breadth of research addressing a problem. \nThe proposed approach enables automated generation of surveys to quickly obtain an overview of state-of-the-art research as well as sharing of surveys for others to reuse.","id":"86a170e1-f5ee-40dd-9553-058683d000c4","level":"paragraph","origin_cites_number":0,"parent_id":"b95d3fb3-5cd7-4c69-8c96-6cadc3ef7b13","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Use cases"],["paragraph","Familiarize with the state-of-the-art"]],"subsections":["b8ac6e76-2ccc-4355-afcd-c2adb9788bf8"],"title":"Familiarize with the state-of-the-art"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:motivating-example:publish-reviews}\nLiterature reviews typically consist of multiple (survey) tables in which different approaches from original papers are compared based on a set of properties. These tables can be seen as the main contribution and most informative part of the review paper, since the tables juxtapose and compare existing work. Comparison tables are published in review papers as static content in PDF documents. This presentational format is generated from datasets that typically contain more (structured) information than what is presented in the published table. However, the additional information is not published. It is ``dark data'' which is not stored or indexed and likely lost over time~. Furthermore, published tables are not machine actionable. Their overall low FAIRness hinders reusability of the published content. With the presented service, it is possible to publish a literature survey with high FAIRness, i.e. that is compliant with the FAIR principles to a high degree. Section~\\ref{section:related-literature} discusses this aspect in more details.","id":"b8ac6e76-2ccc-4355-afcd-c2adb9788bf8","level":"paragraph","origin_cites_number":1,"parent_id":"86a170e1-f5ee-40dd-9553-058683d000c4","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Use cases"],["paragraph","Familiarize with the state-of-the-art"],["paragraph","Publishing of literature reviews"]],"subsections":["34cdcc07-1e15-458e-bb25-8a7ebb70bcd8"],"title":"Publishing of literature reviews"},{"cite_extract_rate":0,"cites":[],"content":"The weaknesses of the current approach to literature review can be summarized as follows:\n\\begin{itemize}\n \\item {\\it Static} -- reviews are static, since once published as PDF they are rarely updated and there are no possibilities or incentives for creating new or updated reviews for considerable time.\n \\item {\\it Lack of machine assistance} -- machine assistance is hardly possible, since the PDF representation of reviews is only human readable and relevant raw data is mostly not published along with the review.\n \\item {\\it Delay} -- reviews are produced and published with significant delay (often years) after original research work was done.\n \\item {\\it Coverage} -- due to the amount of work required, reviews are often only performed for relatively popular research topics and are stale or missing for less popular topics.\n \\item {\\it Lacking collaboration} -- collaboration on reviews is not possible and reviews currently represent only the viewpoint of the few authors not the community.\n \\item {\\it Missing overarching systematic semantic representation} -- the overlap between different reviews and related work sections in individual original research papers is not explicit and cannot be exploited.\n\\end{itemize}\nWe deem that these weaknesses of the current approach to scholarly literature review and synthesis significantly hinder scientific progress.","id":"34cdcc07-1e15-458e-bb25-8a7ebb70bcd8","level":"paragraph","origin_cites_number":0,"parent_id":"b8ac6e76-2ccc-4355-afcd-c2adb9788bf8","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Use cases"],["paragraph","Familiarize with the state-of-the-art"],["paragraph","Publishing of literature reviews"],["paragraph","Summary of weaknesses of the current approach to literature review"]],"subsections":[],"title":"Summary of weaknesses of the current approach to literature review"},{"cite_extract_rate":0.04,"cites":[8369],"content":"\\label{section:related-literature}\nThe task of comparing research contributions can be reviewed in light of the more general task of comparing resources (or entities) in a knowledge graph. While this is a well-known task in multiple domains (for instance in e-commerce systems~), not much work has focused on comparison in knowledge graphs, specifically. One of the few works with this focus is by \\citeauthor{Petrova2017}~ who created a framework for comparing entities in RDF graphs using SPARQL queries. In order to compare contributions, they first have to be found. Finding is an information retrieval problem. As a well-known technique, TF-IDF~ can be used for this task. More sophisticated techniques can be used to determine the structural similarity between graphs (e.g., ) and matching semantically similar predicates. This relates to dataset interlinking~ or more generally ontology alignment~. For property alignment, techniques of interest include edit distance (e.g., Jaro-Winkler~ or Levenshtein~) and vector distance. \\citeauthor{Gromann2019}~ found that fastText~ performs best for ontology alignment. \nIn light of the FAIR Data Principles~, scholarly data should be Findable, Accessible, Interoperable and Reusable both for humans and machines. Due to the publication format, literature survey tables published in scholarly articles weakly adhere to the FAIR guidelines, particularly so for machines. Scholarly data should be considered first-class objects~, including data used to create literature surveys. \\citeauthor{Rodriguez-Iglesias2016}~ describe the difficulties of making data FAIR within the plant sciences. They argue that it is more complicated than reformatting data. On the other hand they suggest that most FAIR principles can be implemented relatively easily by using off-the-shelf technologies. \\citeauthor{Boeckhout2018}~ argue that the FAIR principles alone are not sufficient to lead to responsible data sharing. More applied principles are needed to ensure better scholarly data. This claim is supported by the findings of \\citeauthor{Mons2017}~ who suggest that there are very diverse interpretations of the guidelines. In their work, they try to clarify what is FAIR and what is not. \nAn efficient literature comparison relies on scholarly knowledge being represented in a structured way. There is substantial related work on representing scholarly knowledge in structured form~. Building on the work of numerous philosophers of science, \\citeauthor{Hars2001}~ proposed a comprehensive scientific knowledge model that includes concepts such as theory, methodology and statement. More recently, ontologies were engineered to describe different aspects of the scholarly communication process. Semantic Publishing and Referencing (SPAR)\\footnote{http://purl.org/spar/\\{cito,c4o,fabio,biro,pro,pso,pwo,doco,deo\\}} is a collection of ontologies that can be used to describe scholarly publishing and referencing of documents~. \\citeauthor{IniestaCorcho:SePublica2014}~ reviewed the state-of-the-art ontologies to describe scholarly articles. \\citeauthor{Sateli2015SemanticRO}~ use some of these scholarly ontologies to add semantic representations of scholarly articles to the Linked Open Data cloud. A literature survey comparing scholarly ontologies is available via the ORKG.\\footnote{\\url{https://www.orkg.org/orkg/comparison/R8342}} Most of these ontologies are designed to capture metadata about and structure of scholarly articles, not the content communicated in articles. \nAnother literature survey is created to compare approaches for semantically representing scholarly communication.\\footnote{\\url{https://www.orkg.org/orkg/comparison/R8364}}\nAn initial attempt for semantifying review articles was done in~. The work comprises a relatively rigid ontology for describing contributions (mainly centered around research problems, approaches, implementations and evaluations) and a prototypical implementation using Semantic MediaWiki. We relax this constraint, since we are not limited by a rigid ontology schema but rather allow arbitrary domain-specific semantic structures for research contributions. The work by \\citeauthor{DBLP:conf/ercimdl/VahdatiFALV19}~ focuses on semantic article representations for generating literature overviews. Their method is to use crowdsourcing to generate the overviews. \\citeauthor{Kohl2018}~ present CADIMA, a system that supports systematic literature reviews. The tool supports the formal process of performing a literature review but does, for example, not publish data in machine actionable form for reuse.\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=\\columnwidth]{img/methodology.pdf}\n \\caption{Research contribution comparison methodology.}\n \\label{fig:workflow-comparison}\n\\end{figure}","id":"32d3f690-7c93-4310-b607-6553d1915112","level":"section","origin_cites_number":25,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Related work"]],"subsections":[],"title":"Related work"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:system-design}\nWe now present the system design of the literature comparison service. It consists of a methodology that describes how to perform a comparison of research contributions. An early version of this methodology has been presented at the 3rd SciKnow workshop~. The methodology consists of five steps: 1) finding comparison candidates, 2) selecting related statements, 3) aligning contribution descriptions, 4) visualizing comparisons and 5) publishing FAIR comparisons. The methodology is depicted in Figure~\\ref{fig:workflow-comparison}. First, we discuss the data structure of the ORKG, which forms the foundation of the comparison. Then, each step of the methodology is described in more detail. Finally, we discuss the implementation.","id":"a1917066-53fd-432d-9c65-ea4825cda6c7","level":"section","origin_cites_number":1,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"]],"subsections":["c64100b1-b57e-46ef-a8f6-a6c7131b671e","d4eda852-27fa-4aa8-880a-559597b2dc69","5d1f0f97-df75-4699-9f80-7b7af6dea650","fbebdd62-b2b1-408f-9c69-5d4f7f13cfce","1b062162-78e7-427e-8efd-91ae002ca800","da78b76d-a336-4f78-9f93-2fb8a21e6591","252651d1-6cd2-4bc9-abba-721f972b6d1e"],"title":"System design"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:ontology}\nIn ORKG, each paper is typed as \\textit{paper} class. A paper consists of at least one \\textit{research contribution}, which addresses at least one \\textit{research problem}. Research contributions consist of \\textit{contribution data} that describe the contribution. For instance, a paper in Computer Science might have descriptions for materials, methods, implementation and results as contribution data. These predefined core concepts can be easily extended with domain specific research problems, methods, etc. in ORKG curation using crowdsourcing or other curation approaches. The underlying data structure uses the notion of statements. Statements are triples that consist of a subject, a predicate (also called a property) and an object. The granularity of a comparison is at the research contribution, meaning that contributions are compared rather than papers. For simplicity, we use the terms ``paper comparison'' and ``contribution comparison'' interchangeably. Because a comparison happens on contribution level, it is possible to compare specific elements of a paper instead of the complete paper. The benefit of this is that a comparison does not contain data from irrelevant contributions. The ORKG OWL ontology is available online.\\footnote{\\url{https://gitlab.com/TIBHannover/orkg/orkg-ontology}}","id":"c64100b1-b57e-46ef-a8f6-a6c7131b671e","level":"subsection","origin_cites_number":0,"parent_id":"a1917066-53fd-432d-9c65-ea4825cda6c7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","ORKG ontology"]],"subsections":[],"title":"ORKG ontology"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:workflow-comparison-candidates}\nTo perform a comparison, a starting contribution is needed. This contribution is called \\textit{main contribution} and is always manually selected by a user. The main contribution is compared against other \\textit{comparison contributions}. There are two different approaches for selecting the comparison contributions. The first approach automatically selects comparison contributions based on similarity. The second approach lets users manually select contributions.","id":"d4eda852-27fa-4aa8-880a-559597b2dc69","level":"subsection","origin_cites_number":0,"parent_id":"a1917066-53fd-432d-9c65-ea4825cda6c7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Select comparison candidates"]],"subsections":["dceac5ad-70aa-4f9b-94ef-abad8d3758f3","2079aa06-4aa7-4551-8f0b-5dcb0826279d"],"title":"Select comparison candidates"},{"cite_extract_rate":0,"cites":[],"content":"Comparing contributions makes only sense when contributions can sensibly be compared. For example, it does not make (much) sense to compare a biology paper to a history paper. \nWe thus argue that it makes only sense to compare contributions that are similar. More specifically, contributions that share the same (or a similar set of) properties are good comparison candidates. For instance, a paper about question answering has the property \\textit{orkg:disambiguationTask}\\footnote{\\textit{orkg:} denotes the ontology of the ORKG system described in Section~\\ref{section:ontology}} and another paper is using the same property to describe what disambiguation tasks are performed. Since they share the same property it makes them likely candidates for comparison. Finding similar contributions is therefore based on finding contributions that share the same or similar informative description properties. To achieve this, each comparison contribution is converted into a string by concatenating all properties of the contribution. \nTF-IDF~ is used to query these strings with the string of the main contribution as query. The search returns the most similar contributions by weighting the most informative properties higher due to TF-IDF. The top-k contributions are selected and form a set of contributions that are used in the next step.\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=0.9\\columnwidth]{img/similar-contributions.png}\n \\caption{Implementation of the first step of the methodology: the selection of comparison candidates. Showing both the similarity-based and the manual selection approaches.}\n \\label{fig:comparison-candidates}\n\\end{figure}\nFigure~\\ref{fig:comparison-candidates} displays how the similar contribution selection is implemented. As depicted, three similar contributions are suggested to the user (with the corresponding similarity percentage being displayed next to paper title). These suggested contributions can be directly compared.","id":"dceac5ad-70aa-4f9b-94ef-abad8d3758f3","level":"subsubsection","origin_cites_number":1,"parent_id":"d4eda852-27fa-4aa8-880a-559597b2dc69","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Select comparison candidates"],["subsubsection","Find similar contributions"]],"subsections":[],"title":"Find similar contributions"},{"cite_extract_rate":0,"cites":[],"content":"There are scenarios where comparison based on similarity computation is not suitable or desired. For example, a researcher wants to compare a specific set of implementations to see which performs best. \nTherefore, the manual selection method is implemented in a similar fashion to an e-commerce shopping cart. When the ``Add to comparison'' checkbox is checked, a box appears listing the selected contributions (Figure~\\ref{fig:comparison-box}). \n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=.45\\columnwidth]{img/comparison-box.png}\n \\caption{Box showing the manually selected contributions.}\n \\label{fig:comparison-box}\n\\end{figure}","id":"2079aa06-4aa7-4551-8f0b-5dcb0826279d","level":"subsubsection","origin_cites_number":0,"parent_id":"d4eda852-27fa-4aa8-880a-559597b2dc69","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Select comparison candidates"],["subsubsection","Manual selection"]],"subsections":[],"title":"Manual selection"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:workflow-related-statements}\nThis step selects the statements from the graph related to the set of contributions selected in the previous step. Statements are selected transitively to match contributions in subject or object position. This search is performed until a predefined maximum transitive depth $\\delta$ has been reached. The intuition is that the deeper a property is nested the less likely is its relevance for the comparison. The process of selecting statements is repeated until depth $\\delta=5$ is reached. This number is chosen empirically to include statements that are not directly related to the contribution, but to exclude statements that are less relevant because they are nested too deep.","id":"5d1f0f97-df75-4699-9f80-7b7af6dea650","level":"subsection","origin_cites_number":0,"parent_id":"a1917066-53fd-432d-9c65-ea4825cda6c7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Select related statements"]],"subsections":[],"title":"Select related statements"},{"cite_extract_rate":1,"cites":[8369,6548],"content":"As described in the first step, comparisons are built using shared or similar properties of contributions. In case the same property has been used between contributions, these properties are grouped and form one \\textit{comparison row}. However, often different properties are used to describe the same concept. This occurs for various reasons. The most obvious reason is when two different ontologies are used to describe the same property. For example, for describing the population of a city, DBpedia uses \\textit{dbo:populationTotal} while WikiData uses \\textit{WikiData:population} (actually the property identifier is P1082; for the purpose here we use the label). When comparing contributions, these properties should be considered as equivalent. Especially for community-created knowledge graphs, differently identified properties likely exist that are, in fact, equivalent.\nTo overcome this problem, we use pre-trained fastText~ word embeddings to determine the similarity of properties. If the similarity is higher than a predetermined threshold $\\tau$, the properties are considered equivalent and are grouped. This happens when the similarity threshold $\\tau \\geq 0.9$ (also empirically determined). In the end, each group of properties will be visualized as one row in the comparison table. The result of this step is a list of statements for each contribution, where similar properties are grouped. Based on this similarity matrix $\\gamma$ is generated \n\\begin{equation}\n \\gamma_{p_{i}} = \\left [ cos(\\overrightarrow{p_i}, \\overrightarrow{p_j}) \\right ]\n \\label{eq:similarity_matrix}\n\\end{equation}\nwith $cos(.)$ as the cosine similarity of vector embeddings for property pairs $(p_i, p_j) \\in \\mathcal{P}$, whereby $\\mathcal{P}$ is the set of all contributions.\nFurthermore, we create a mask matrix $\\Phi$ that selects properties of contributions $c_i \\in \\mathcal {C}$, whereby $\\mathcal{C}$ is the set of contributions to be compared. Formally,\n\\begin{equation}\n \\Phi_{i,j} = \\begin{cases}\n1 & \\text{ if } p_{j} \\in c_i \\\\ \n0 & \\text{ otherwise } \n\\end{cases}\n \\label{eq:mask_matrix}\n\\end{equation}\nNext, for each selected property $p$ we create the matrix $\\varphi$ that slices $\\Phi$ to include only similar properties. Formally,\n\\begin{equation}\n\\varphi_{i,j} =(\\Phi_{i,j})_{\\substack{c_i \\in \\mathcal{C}\\\\ p_j \\in sim(p) }}\n\\label{eq:slice_mask}\n\\end{equation}\nwhere $sim(p)$ is the set of properties with similarity values $\\gamma[p] \\geq \\tau$ with property $p$. Finally, $\\varphi$ is used to efficiently compute the common set of properties~. This process is displayed in Algorithm~\\ref{alg:alining-properties}.\n\\begin{algorithm}\n\\caption{Align contribution descriptions}\\label{alg:alining-properties}\n\\begin{algorithmic}[1]\n\\Procedure{AlignProperties}{properties, threshold}\n\\ForEach {property $p_1 \\in properties$}\n \\ForEach {property $p_2 \\in properties$}\n \\State $similarity \\gets$ cos(Embb($p_1$), Embb($p_2$))\n \\If {$similarity > threshold$}\n \\State $similarProps \\gets similarProps \\cup \\{ p_1, p_2 \\}$\n \\EndIf\n \\EndFor\n\\EndFor\n\\Return $similarProps$\n\\EndProcedure\n\\end{algorithmic}\n\\end{algorithm}","id":"fbebdd62-b2b1-408f-9c69-5d4f7f13cfce","level":"subsection","origin_cites_number":2,"parent_id":"a1917066-53fd-432d-9c65-ea4825cda6c7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Align contribution descriptions"]],"subsections":[],"title":"Align contribution descriptions"},{"cite_extract_rate":0,"cites":[],"content":"The next step of the workflow is to visualize the comparison and present the data in a human understandable format. Tabular format is often appropriate for visualizing comparisons since tables provide a good overview of data. Another aspect of the visualization is determining which properties should be displayed and which ones should be hidden. A property is displayed when it is shared among a predetermined amount $\\alpha$ of contributions, where $\\alpha$ mainly depends on comparison use and can be determined based on the total amount of contributions in the comparison. By default, only properties that are common to at least two contributions ($\\alpha \\geq 2$) are displayed. \nAnother aspect of comparison visualization is the possibility to customize the resulting table. This is needed because of the similarity-based matching of properties and the use of predetermined thresholds. For example, users should be able to enable or disable properties. They should also get feedback on property provenance (i.e., the property's path in the graph). Ultimately, this contributes to a better user experience, with the possibility to manually correct mistakes made by the system. \nFigure~\\ref{fig:comparison-table} displays a comparison for research contributions related to visualization tools published in the literature. In this example, four properties are displayed. Literals are displayed as plain text while resources are displayed as links. When a resource link is selected, a popup is displayed showing the statements related to this resource. The UI implements some additional features that are particularly useful to compare research contributions.","id":"1b062162-78e7-427e-8efd-91ae002ca800","level":"subsection","origin_cites_number":0,"parent_id":"a1917066-53fd-432d-9c65-ea4825cda6c7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Visualize comparison"]],"subsections":["04b99168-d893-480d-972b-bd9514984b9f"],"title":"Visualize comparison"},{"cite_extract_rate":0,"cites":[],"content":"Users can customize comparisons including transposing the table as well as hiding and rearranging the properties. Especially the option to hide properties is helpful when contributions with many statements are compared. Only properties considered relevant to the user can be selected to display. Customizing the comparison table can be useful before exporting or sharing the comparison.","id":"04b99168-d893-480d-972b-bd9514984b9f","level":"paragraph","origin_cites_number":0,"parent_id":"1b062162-78e7-427e-8efd-91ae002ca800","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Visualize comparison"],["paragraph","Customization"]],"subsections":["6d0b1e3c-a8a3-4dd7-a2ef-5cb37393a9a0","0851ca07-6e0f-4b3d-8617-8e8005a1d81b"],"title":"Customization"},{"cite_extract_rate":0,"cites":[],"content":"Comparisons can be shared using a persistent link. Especially when sharing the comparison for research purposes, it is important to refer to the original comparison. Since contribution descriptions may change over time comparisons may also change. To support persistency, the whole state of the comparison is stored in a document-oriented database and retrieved when the permalink is invoked.","id":"6d0b1e3c-a8a3-4dd7-a2ef-5cb37393a9a0","level":"paragraph","origin_cites_number":0,"parent_id":"04b99168-d893-480d-972b-bd9514984b9f","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Visualize comparison"],["paragraph","Customization"],["paragraph","Sharing and persistence"]],"subsections":[],"title":"Sharing and persistence"},{"cite_extract_rate":0,"cites":[],"content":"It is possible to export comparisons in different output formats such as PDF, CSV, RDF and \\LaTeX. The \\LaTeX~export is useful for direct integration in research papers. Together with the \\LaTeX~table, a BibTeX file containing the bibliographic information of the papers used in the comparison is also generated. Also, a persistent link referring back to the comparison in ORKG is showed as table footnote.\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=0.95\\columnwidth]{img/comparison-table.png}\n \\caption{Comparison of research contributions related to visualization tools.}\n \\label{fig:comparison-table}\n\\end{figure}","id":"0851ca07-6e0f-4b3d-8617-8e8005a1d81b","level":"paragraph","origin_cites_number":0,"parent_id":"04b99168-d893-480d-972b-bd9514984b9f","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Visualize comparison"],["paragraph","Customization"],["paragraph","Export"]],"subsections":[],"title":"Export"},{"cite_extract_rate":0,"cites":[],"content":"Visualized and customized comparison tables can be stored. Storing tables is part of the publishing process and therefore only needed when a generated table is going to be used in a paper. In order to regenerate the table the whole state of the comparison should be saved. The knowledge graph from which the comparison was generated changes over time and thus storing just the URIs of the respective papers would not suffice. While saving a comparison, the user can provide additional metadata to ensure findability, an aspect of the FAIR principles. Metadata include a comparison title, which would normally consist of a one sentence description of the comparison. Additionally, a longer textual description can be provided. This metadata is extended with machine generated data, such as the creation date and the creator of the comparison. The metadata is stored in the knowledge graph to support easy access and interoperability. In Figure~\\ref{fig:RDF-comparison}, the structure of the metadata is displayed using the Dublin Core Metadata Terms\\footnote{\\url{https://dublincore.org/specifications/dublin-core/dcmi-terms}}. The comparison data itself is stored in a document-oriented database. An RDF export of both the metadata and the comparison data can be generated. The comparison data is modeled with the RDF Data Cube Vocabulary\\footnote{\\url{https://www.w3.org/TR/vocab-data-cube}}. A unique identifier is attached when the comparison is saved. This ID is used when the comparison is shared or when it is referenced in a paper. The literature comparison can also be performed without publishing. Although the workflow and the steps to create a comparison stay the same, the goal is different. Instead of creating a comparison that will be published and referenced in a paper, the comparison will be used by the researcher herself. \n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=.85\\columnwidth]{img/RDF-comparison.pdf}\n \\caption{The graph structure of the metadata for a published comparison. The \\textit{dcterms:} prefix denotes the Dublin Core Metadata Terms ontology.}\n \\label{fig:RDF-comparison}\n\\end{figure}","id":"da78b76d-a336-4f78-9f93-2fb8a21e6591","level":"subsection","origin_cites_number":0,"parent_id":"a1917066-53fd-432d-9c65-ea4825cda6c7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Publish comparison"]],"subsections":[],"title":"Publish comparison"},{"cite_extract_rate":0.25,"cites":[8253],"content":"The user interface of the comparison feature is seamlessly integrated with the ORKG front end, which is written in JavaScript and is publicly available\\footnote{\\url{https://gitlab.com/TIBHannover/orkg/orkg-frontend}}. The back end of the comparison feature is a service separate from the ORKG back end written in Python and also available Open Source\\footnote{\\url{https://gitlab.com/TIBHannover/orkg/orkg-similarity}}. The comparison back end is responsible for step two and three of the comparison methodology. The input in step two is the set of contribution IDs. The API selects the related statements and aligns the properties and returns the data needed to visualize the comparison. This data includes the list of papers, list of all properties and the values per property. \n\\begin{table*}[t]\n\\caption{List of imported survey tables in the ORKG. The paper and table reference can be used to identify the original table. }\n\\label{table:imported-review-tables}\n\\begin{adjustbox}{scale=.90}\n\\begin{tabular}{l|l|l|r|l|l}\n\\toprule\n\\textbf{Paper reference} & \\textbf{Table reference} & \\textbf{Research problem} & \\textbf{Papers} & \\textbf{ORKG representation} & \\textbf{Information loss} \n\\\\ \\midrule\n\\citeauthor{Bikakis2016}~ & \nTable 1 & Generic visualizations & \n11 &\n\\url{https://orkg.org/orkg/c/pdLJDk} & \nNo \n\\\\ \n\\citeauthor{Bikakis2016}~ & \nTable 2 & Graph visualizations & \n21 &\n\\url{https://orkg.org/orkg/c/Rx476Z} & \nNo\n\\\\ \n\\citeauthor{Diefenbach2018a}~ & \nTable 2 & \nQuestion answering evaluations & \n33 &\n\\url{https://orkg.org/orkg/c/gaVisD} & \nNo\n\\\\ \n\\citeauthor{Diefenbach2018a}~ & \nTable 3,4,5,6 & \nQuestion answering systems & \n26 &\n\\url{https://orkg.org/orkg/c/IuEWl2} & \nNo\n\\\\ \n\\citeauthor{Hussain2017a}~ & \nTable 4 &\nAuthor name disambiguation &\n5 &\n\\url{https://orkg.org/orkg/c/vDxKdr} & \nNo\n\\\\\n\\citeauthor{Hussain2017a}~ & \nTable 5 &\nAuthor name disambiguation &\n6 &\n\\url{https://orkg.org/orkg/c/XXg8Wg} & \nNo\n\\\\ \n\\citeauthor{Hussain2017a}~ & \nTable 6 &\nAuthor name disambiguation &\n9 &\n\\url{https://orkg.org/orkg/c/9rOwPV} & \nNo\n\\\\ \n\\citeauthor{Hussain2017a}~ & \nTable 7 &\nAuthor name disambiguation &\n6 &\n\\url{https://orkg.org/orkg/c/mB7kIK} & \nNo\n\\\\ \n\\citeauthor{Naidu2018}~ & \nTable 4 &\nText summarization &\n52 &\n\\url{https://orkg.org/orkg/c/OUqYB9} & \nNo\n\\\\ \n\\bottomrule\n\\end{tabular}\n\\end{adjustbox}\n\\end{table*}","id":"252651d1-6cd2-4bc9-abba-721f972b6d1e","level":"subsection","origin_cites_number":4,"parent_id":"a1917066-53fd-432d-9c65-ea4825cda6c7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","System design"],["subsection","Technical details"]],"subsections":[],"title":"Technical details"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:data-collection}\nIn order to generate useful literature reviews it is crucial for the knowledge graph to contain sufficient and relevant papers. Populating the knowledge graph with high quality paper descriptions it not straightforward. Structured descriptions of papers should be created in such a way that it is possible to compare papers based on shared properties. Both published papers and papers that will be published in the future should be added to the ORKG, retrospectively or prospectively. Although a comprehensive description on how to populate the ORKG is out-of-scope here, we now briefly describe how we envision populating the ORKG in a manner that would facilitate comparing contributions.\nProspectively, authors can become part of generating structured descriptions of their papers. This should be done in a crowdsourced manner and can become part of the paper submission process. Input templates that collect relevant properties can be used to ensure structured and comparable paper descriptions. Retrospectively, automated (machine learning) methods can be helpful ensure scalability of the process of adding a paper.","id":"def65367-1d9c-4420-92cd-cd2f5cd6b277","level":"section","origin_cites_number":0,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Data collection"]],"subsections":["a72860f0-4d1e-4b2f-aa43-73a96ef0f3e7"],"title":"Data collection"},{"cite_extract_rate":0,"cites":[],"content":"To populate the ORKG with comparable paper descriptions, we leverage the data published in review papers. Review papers consist of high quality, curated and often structured data that is collected from a set of papers that address the same (or a similar) research problem. Hence, using reviews to populate a scholarly knowledge graph is a relatively straightforward approach to obtain high quality structured paper descriptions. We now present a methodology to convert survey paper data into a knowledge graph structure. The steps are as follows:\n\\begin{enumerate}\n \\item \\textbf{Survey paper selection.} The first step is the selection of survey papers that are suitable for building a knowledge graph. Firstly, the survey should compare peer-reviewed scientific articles. For instance, a comparison of different systems without a reference to peer-reviewed work is not suitable for the scholarly knowledge graph. Secondly, the review should compare the papers' content in a structured way and should not merely list work in a field. Especially reviews that present their results and literature comparisons in tabular format are suitable. The result of this step is a list of papers that will be added to the ORKG. \n \\item \\textbf{Table selection.} Given the selected survey papers, tables have to be selected. Some surveys contain only one table while in others multiple tables are presented. In some cases a collection of tables can be joined into one larger table. \n \\item \\textbf{Data modeling.} Given the selected tables, a suitable graph structure has to be determined. The data structure has to be modeled. For instance, when implemented systems are compared, a suitable structure could be: \\texttt{[has implementation] -> System name}. The referenced system can be described with a list of properties to be compared. Additionally, a research problem has to be defined, which is typically the same for all papers that are part of the table. \n \\item \\textbf{Metadata collection.} Next, the metadata for the papers that are referenced in the survey table is collected. In case a referenced paper has a DOI\\footnote{Digital Object Identifier}, the metadata can be automatically retrieved via a lookup service (e.g. Crossref\\footnote{\\url{https://www.crossref.org}}). Otherwise, at least the title, authors and publication date have to be collected.\n \\item \\textbf{Data ingestion.} Finally, the paper data is ingestion into the knowledge graph. The paper data consists of both the paper's metadata and the extracted data from the comparison table. \n This does not result in a single description of the survey paper. Each paper referenced in the survey table is ingested individually. In order to speed up the process of adding papers, we developed a Python package\\footnote{\\url{https://gitlab.com/TIBHannover/orkg/orkg-pypi}} that has a function to add a paper to the knowledge graph. \n\\end{enumerate}\nThis methodology has been used to populate the ORKG with comparable paper data. The data is used to evaluate the presented literature review tool. The imported paper data is not only useful for the evaluation, but does also provide significant value to the ORKG itself.\nIn total, four review papers were selected for importing into the ORKG. The Python script for importing the table data is available online.\\footnote{\\url{https://gitlab.com/TIBHannover/orkg/orkg-papers}} From those papers, 12 different tables were imported. Together, 169 papers were reviewed in those four survey papers. This resulted in a total amount of $3\\,750$ statements being added to the knowledge graph. Table~\\ref{table:imported-review-tables} lists the imported review papers and tables. The survey papers address different research problems. Figure~\\ref{fig:graph-example-import} depicts an excerpt of the resulting graph for one particular paper. A set of comparison tables made with the imported data is available online.\\footnote{\\url{https://orkg.org/orkg/featured-comparisons}} This list includes some alternative comparison tables that were generated with the same data. \n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=0.8\\columnwidth]{img/graph-example-import.pdf}\n \\caption{Partial graph structure of an imported paper. Orange colored resources indicate potentially interesting values for a paper comparison.}\n \\label{fig:graph-example-import}\n\\end{figure}","id":"a72860f0-4d1e-4b2f-aa43-73a96ef0f3e7","level":"subsection","origin_cites_number":0,"parent_id":"def65367-1d9c-4420-92cd-cd2f5cd6b277","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Data collection"],["subsection","Leverage legacy review paper tables"]],"subsections":[],"title":"Leverage legacy review paper tables"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:evaluation}\nIn this section, we present an evaluation of multiple aspects of the presented comparison methodology and implementation. Firstly, we evaluate information representation. Then, we evaluate the FAIRness of published reviews. Finally, we present a performance evaluation that tests the scalability.","id":"9c67205a-5ea7-42a6-af06-049ac9b5d00b","level":"section","origin_cites_number":0,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"]],"subsections":["52143d93-69a3-4b3d-a716-7d4ad8e48679","c31bac6f-f294-4381-8adf-fd5c75c400a7","34ac5be8-a18c-4ef0-bc7d-dc420b39ab8d"],"title":"Evaluation"},{"cite_extract_rate":0,"cites":[],"content":"This part of the evaluation focuses on the aspect of information representation. We use the data from the imported review papers, as described in Section~\\ref{section:data-collection}. In order to build and publish useful and correct literature reviews, at a minimum our service should display the same information that was originally presented in the review tables. This means that there should not be information loss when review tables are published using our service. If there is no information loss, it means our service can be used as an alternative to the current way of publishing review tables. Apart from generating the same table, the added value comes from the ability to aggregate new (tabular) views using the same data as well as the increased FAIRness of the data published via our service. For each of the imported review tables, listed in Table~\\ref{table:imported-review-tables}, we can evaluate whether the same table can be generated with our service. For this, we have compared the table from the review paper to the table generated by the ORKG comparison service. A collection of 169 paper with 9 distinct literature views/tables are part of this evaluation. These tables can be viewed online, the links are listed in the ``ORKG representation'' column. The results of this evaluation are displayed in the same table, in column ``Information loss''. As the results show, using our service it is possible to recreate the same tabular views as originally published in the review papers.","id":"52143d93-69a3-4b3d-a716-7d4ad8e48679","level":"subsection","origin_cites_number":0,"parent_id":"9c67205a-5ea7-42a6-af06-049ac9b5d00b","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"],["subsection","Information representation"]],"subsections":[],"title":"Information representation"},{"cite_extract_rate":0,"cites":[],"content":"As described before, with the presented service it is possible to publish a generated comparison that adheres with the FAIR principles. Because the service leverages a knowledge graph to generate and save comparisons, complying with the FAIR principles is more obvious for the ORKG comparison service than for tables in published PDF articles. In order to evaluate the FAIRness of a published comparison, we evaluate each of the four FAIR principles in detail. \\citeauthor{Wilkinson2016}~ described each principle by assigning sub-principles.\\footnote{For a more detailed definition of the FAIR principles, see: \\url{https://go-fair.org/fair-principles}} We discuss the relevant sub-principles and explain how they are met. We use the term (meta)data to refer to both the actual comparison data (i.e., the data that is used to create the comparison table) and the associated metadata (e.g., the title, description and creator of a comparison). Table \\ref{table:fair-principles-compliance} presents an overview for the evaluation of the FAIR principles.","id":"c31bac6f-f294-4381-8adf-fd5c75c400a7","level":"subsection","origin_cites_number":1,"parent_id":"9c67205a-5ea7-42a6-af06-049ac9b5d00b","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"],["subsection","FAIR data evaluation"]],"subsections":["ac11071f-8ae2-423d-ab33-85fce076119e"],"title":"FAIR data evaluation"},{"cite_extract_rate":0,"cites":[],"content":"To make data findable for both humans and machines (i.e., agents), a unique and persistent identifier should be attached to the data (F1). Additionally, metadata should describe the data (F2). In the metadata, the unique identifier of the data should be mentioned (F3). Also a search interface should be available to find the data (F4). To ensure the findability of comparisons, users can title and describe them. Furthermore, machine generated metadata is attached to a comparison (e.g., the number of papers and the creation date). A unique identifier is generated and attached to the data and included in the metadata. Finally, the ORKG search interface allows users to search the whole graph and has a dedicated filter to specifically find comparisons. Additionally, comparisons can be indexed and found by third-party search engines (such as Google or Bing).","id":"ac11071f-8ae2-423d-ab33-85fce076119e","level":"paragraph","origin_cites_number":0,"parent_id":"c31bac6f-f294-4381-8adf-fd5c75c400a7","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"],["subsection","FAIR data evaluation"],["paragraph","Findable."]],"subsections":["154af8e0-ed85-46d2-9570-9da5491ee971","68397955-d4af-4601-8b25-5a6aef856a6b","a1388af8-78d9-4c21-9ff4-2ccd2d4a9694"],"title":"Findable."},{"cite_extract_rate":0,"cites":[],"content":"Having found data, agents need to know how to access it. This principle is primarily about using accessible standardised communication protocols (A1). Additionally, metadata should be available even when the data is not (A2). The metadata is part of the knowledge graph, which can be accessed via the HTTP protocol. The data can be accessed without authentication. To support A2, the metadata and the actual comparison data are stored separately. Therefore, it is possible to access only metadata when the original data is not available anymore (for example when data is retracted by the author).","id":"154af8e0-ed85-46d2-9570-9da5491ee971","level":"paragraph","origin_cites_number":0,"parent_id":"ac11071f-8ae2-423d-ab33-85fce076119e","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"],["subsection","FAIR data evaluation"],["paragraph","Findable."],["paragraph","Accessible."]],"subsections":[],"title":"Accessible."},{"cite_extract_rate":0,"cites":[],"content":"To ensure the interoperability of data, it should use a formal language for knowledge representation (I1) and should use vocabularies that are FAIR (I2). Finally, references or links to other (meta)data should be made (I3). As argued before, thanks to highly structured data and the integration of shared vocabularies, interoperability is an inherent feature of knowledge graphs. Data is (partially) described using the ORKG core ontology and other ontologies we use to canonicalize the representation of relevant information content types. Links to other data are present in the knowledge graph. For example, if a comparison uses the ``Web'' resource to specify the domain of an application, this resource is generic, can be shared among paper descriptions and comparisons, and can be described in more detail, independently of a particular comparison.","id":"68397955-d4af-4601-8b25-5a6aef856a6b","level":"paragraph","origin_cites_number":0,"parent_id":"ac11071f-8ae2-423d-ab33-85fce076119e","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"],["subsection","FAIR data evaluation"],["paragraph","Findable."],["paragraph","Interoperable."]],"subsections":[],"title":"Interoperable."},{"cite_extract_rate":0,"cites":[],"content":"Finally, data should be reuseable. This can be accomplished by adding relevant (meta)data (R1). Required are an accessible data license (R1.1) and detailed provenance (R1.2) data. Finally, (meta)data should use community standards to describe data (R1.3). It is possible to add additional metadata to a comparison, e.g. metadata about the scope of the comparison, which could be a reference to the paper in which the comparison is being used. The metadata is complemented with the metadata that is already part of the Findability principle, e.g. provenance data about the creator of the comparison. The data license of the graph data is CC BY-SA\\footnote{\\url{https://creativecommons.org/licenses/by-sa/2.0}} (Attribution-ShareAlike), which allows reuse of the data. There is currently no community standard to describe the comparison data. However, standard ontologies are used to describe metadata (e.g., Dublin Core).\n\\begin{table}[t]\n\\centering\n\\caption{Overview of FAIR principles compliance.}\n\\label{table:fair-principles-compliance}\n\\begin{adjustbox}{scale=.85}\n\\begin{tabular}{l|c|p{67mm}}\n\\toprule\n\\textbf{Principle} & \\textbf{Level} & \\textbf{Explanation} \\\\ \\midrule\n\\multicolumn{3}{l}{\\textit{Findable}} \\\\ \\midrule\nF1 & 3 & Unique IDs exist, DOI assignment for future work \\\\ \nF2 & 2 & Machine and user generated metadata is attached \\\\ \nF3 & 1 & Properties used to link data to metadata \\\\ \nF4 & 1 & Comparisons are findable via a search interface \\\\ \\midrule\n\\multicolumn{3}{l}{\\textit{Accessible}} \\\\ \\midrule\nA1 & 2 & Data is accessed over HTTP (via REST or a user interface), requires user effort to integrate the ORKG API specification \\\\ \nA1.1 & 1 & The protocol is free and widely used \\\\ \nA1.2 & 1 & No authentication is required to access the data \\\\ \nA2 & 1 & Metadata is stored in a persistent way and available without the data itself \\\\ \\midrule\n\\multicolumn{3}{l}{\\textit{Interoperable}} \\\\ \\midrule\nI1 & 1 & RDF (with type assertions) and CSV export of comparisons\\\\\nI2 & 2 & Reuse of ontologies where possible (ORKG core, Dublin core, RDF Data Cube Vocabulary). User responsible for other ontology reuse. \\\\ \nI3 & 3 & For comparisons, the compared paper metadata is linked. More references are needed and can be created by users. \\\\ \\midrule\n\\multicolumn{3}{l}{\\textit{Reusable}} \\\\ \\midrule\nR1 & 1 & Machine and user generated metadata is created while publishing \\\\\nR1.1 & 1 & CC-BY SA license \\\\ \nR1.2 & 1 & If a registered user publishes a comparison, the user is associated with the published data \\\\ \nR1.3 & 2 & Users can describe contributions using domain-relevant ontologies \\\\ \n\\bottomrule\n\\end{tabular}\n\\end{adjustbox}\n1=Yes; 2=Yes, requires user effort; 3=Partially/future work\n\\end{table}\n\\noindent\nThe evaluation of the FAIR principles shows that comparisons published with our service rank high in FAIRness, which can be even further increased with some effort from users. Users are mainly responsible for adding the correct information to the comparison and reuse vocabularies. Otherwise, findability, accessibility and to some extent also interoperability are largely handled by the service.","id":"a1388af8-78d9-4c21-9ff4-2ccd2d4a9694","level":"paragraph","origin_cites_number":0,"parent_id":"ac11071f-8ae2-423d-ab33-85fce076119e","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"],["subsection","FAIR data evaluation"],["paragraph","Findable."],["paragraph","Reusable."]],"subsections":[],"title":"Reusable."},{"cite_extract_rate":0,"cites":[],"content":"In order to evaluate the performance of the overall comparison, we compared the implemented ORKG approach to a naive approach for comparing multiple resources. The naive approach compares each property against all other properties to perform the property alignment. Table~\\ref{tab:sota-eval} shows the time needed to generate comparisons, for both the naive and the ORKG approach. In total, eight papers are compared with on average ten properties per paper. In the naive approach, the ``Align contribution descriptions'' step is not scaling well, since each property is compared against all others. If multiple contributions are selected, the number of property similarity checks grows exponentially. Table~\\ref{tab:sota-eval} shows that the ORKG approach outperforms the naive approach. The total number of papers used for the evaluation is limited to eight because the naive approach does not scale to larger sets. \n\\begin{table}[t]\n\\caption{Time (in seconds) to perform comparisons with 2-8 contributions using the naive and ORKG approaches.}\n\\begin{adjustbox}{scale=.9}\n\\begin{tabular}{l|lllllll}\n& \\multicolumn{7}{c}{\\textbf{Number of compared research contributions}} \\\\\n\\hline\n & \\textbf{2} & \\textbf{3} & \\textbf{4} & \\textbf{5} & \\textbf{6} & \\textbf{7} & \\textbf{8} \\\\ \\hline\n\\textbf{Naive} & 0.00026 & 0.1714 & 0.763 & 4.99 & 112.74 & 1772.8 & 14421 \\\\\n\\textbf{ORKG} & 0.0035 & 0.0013 & 0.01158 & 0.02 & 0.0206 & 0.0189 & 0.0204 \n\\end{tabular}\n\\end{adjustbox}\n\\label{tab:sota-eval}\n\\end{table}","id":"34ac5be8-a18c-4ef0-bc7d-dc420b39ab8d","level":"subsection","origin_cites_number":0,"parent_id":"9c67205a-5ea7-42a6-af06-049ac9b5d00b","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Evaluation"],["subsection","Performance evaluation"]],"subsections":[],"title":"Performance evaluation"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section:discussion}\nOne of the aims of the contribution comparison functionality is to support literature reviews and make this activity less cumbersome and time consuming for researchers. To live up to this aim, more structured contribution descriptions are needed. Existing scholarly knowledge graph initiatives focus primarily on scholarly metadata, while with ORKG we focus on making the actual research contributions machine readable. Currently, the ORKG does not yet contain sufficient contribution descriptions in order for the comparison functionally to be practically useful for researchers. Furthermore, for an evaluation of the effectiveness of certain components of the methodology (such as finding related papers or aligning similar properties), more contribution data is needed. Publishing surveys does not rely on data quantity and is therefore evaluated more extensively in this work. The performance evaluation results indicate that the comparison feature performs well. This means the technical infrastructure is in place for the literature survey service.\nIn the evaluation, we focused on the aspects of the system that are necessary for researchers to use the system in practice. The information representation evaluation is a straightforward evaluation to see if existing survey tables can be regenerated with the ORKG. This is a minimal requirement for researchers when using the system, since they should at least be able to recreate tables. This evaluation does not give insight to the usefulness and usability of the system, but still provides an indication that the service can be successfully used to publish literature surveys. One of the reasons for using the service is that also ``dark data'' in comparisons is published (as discussed in Section~\\ref{section:motivating-example:publish-reviews}).\nAnother interesting aspect of the service is that published literature surveys rank high in FAIRness. Therefore, the second part of the evaluation focuses on how the FAIR principles are met. Merely publishing data as RDF is not sufficient to fully meet the FAIR principles. Hence, we conducted a more detailed evaluation that describes how the service complies with each sub-principle. Since FAIR is not a \\textit{standard}, the principles are permissive and not prescriptive~. No technical requirements are specified. Both the implementation and evaluation of the guidelines are therefore subject to interpretation. With respect to data interoperability and reusability, certain aspects of the service can be improved. For example, to improve interoperability, the contribution data should be reusing existing vocabularies where possible. Additionally, although most of FAIRification is done by the system, the researcher is responsible for adding correct and relevant metadata while publishing a survey.\nAs indicated earlier, the usefulness of the presented tool depends on the number of papers present in the knowledge graph. Therefore, future work will focus on data collection, both in a crowdsourced and automated manner. We plan on extending the methodology presented in Section~\\ref{section:data-collection} with automated extraction of data and tables from literature review papers. With the extracted review data, the knowledge graph can be extended more quickly than the previously presented manual method. It could form the basis of a high quality scholarly knowledge graph that contains relevant and FAIR survey table data. Furthermore, in the future we will assign (DataCite) DOIs to published surveys. They will serve as a persistent identifier for the survey data~.","id":"f87ef12c-02fb-41f5-9f16-69dee5190666","level":"section","origin_cites_number":2,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Discussion \\& future work"]],"subsections":[],"title":"Discussion \\& future work"},{"cite_extract_rate":0,"cites":[],"content":"Reviewing existing literature is an important but cumbersome and time consuming activity. To address this problem, we presented a methodology and service that can be used to generate literature surveys from a scholarly knowledge graph. This service can be used by researchers in order to get familiar with existing literature. Additionally, the tool can be used to publish literature surveys in a way that they largely adhere to the FAIR data principles. The presented methodology addresses multiple aspects, including finding suitable contributions, aligning contribution descriptions, visualization and publishing. The methodology is implemented within the Open Research Knowledge Graph (ORKG). Since the comparison relies on structured scholarly knowledge, we discussed how to populate the ORKG with relevant data. This is done by extracting tabular survey data from existing literature reviews. In order to evaluate whether the proposed service can be used to publish literature surveys, the original survey table representations were compared with the ones generated by our service. As the results indicate, it is possible to use the service as an addition or potentially even replacement of the current publishing approach, since the same tables can be generated. The evaluation also showed how the published literature surveys largely adhere to the FAIR data principles. This is crucial for data reusability and machine actionability. To conclude, the proposed literature comparison service addresses multiple weaknesses of the current survey publishing approach and can be used by researchers to generate, publish and reuse literature surveys. \n\\begin{acks}\nThis work was co-funded by the European Research Council for the project ScienceGRAPH (Grant agreement ID: 819536) and the TIB Leibniz Information Centre for Science and Technology. We want to thank Kheir Eddine Farfar for his contributions to this work.\n\\end{acks}\n\\bibliographystyle{ACM-Reference-Format}\n\\bibliography{refs,mendeley}\n\\end{document}\n\\endinput","id":"4c8c53d6-d473-464f-a034-d85b2b90268e","level":"section","origin_cites_number":0,"parent_id":"3313ad50-7887-4970-bb3f-a8d5311e5f61","prefix_titles":[["title","Generate FAIR Literature Surveys with Scholarly Knowledge Graphs"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ]\n ],\n \"subsections\": [\n \"b5fdf9c3-044e-45b3-8bb1-0b7d691c74d6\",\n \"b95d3fb3-5cd7-4c69-8c96-6cadc3ef7b13\",\n \"32d3f690-7c93-4310-b607-6553d1915112\",\n \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"def65367-1d9c-4420-92cd-cd2f5cd6b277\",\n \"9c67205a-5ea7-42a6-af06-049ac9b5d00b\",\n \"f87ef12c-02fb-41f5-9f16-69dee5190666\",\n \"4c8c53d6-d473-464f-a034-d85b2b90268e\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.18181818181818102,\n \"cites\": [\n 6548,\n 6549\n ],\n \"content\": \"When conducting scientific research, reviewing the existing literature is an essential activity~. \\nFamiliarity with the state-of-the-art is required to effectively contribute to advancing it and do relevant research. Mainly because published scholarly knowledge is unstructured~, it is currently very tedious to review existing literature. Relevant literature has to be found among hundreds and increasingly thousands of PDF articles. This activity is supported by library catalogs and online search engines, such as Scopus or Google Scholar~. Because the search is keyword based, typically large numbers of articles are returned by search engines. Researchers have to manually identify the relevant papers. Having identified the relevant papers, the relevant pieces of information need to be extracted in order to obtain an overview of the literature. Overall, these are manual and time consuming steps. We argue that a key issue is that the scholarly knowledge communicated in the literature does not meet the FAIR Data Principles~. While PDF articles can be found and accessed (assuming Open Access or an institutional subscription), the scholarly literature is insufficiently interoperable and reusable, especially for machines. For units more granular than the PDF article, such as a specific result, findability and accessibility score low even for humans.\\nWe present a methodology and its implementation integrated into the Open Research Knowledge Graph (ORKG)~ that can be used to generate and publish literature surveys in form of machine actionable, comparable descriptions of research contributions. Machine actionability of research contributions relates to the ability of machines to access and interpret the contribution data. The benefits for researchers of such an infrastructure are (at least) two-fold. Firstly, it supports researchers in creating state-of-the-art overviews for specific research problems efficiently. Secondly, it supports researchers in publishing literature surveys that adhere to the FAIR principles, thus contributing substantially to reuse of state-of-the-art overviews and therein contained information, for both humans and machines. \\nLiterature reviews are articles that focus on analysing existing literature. Among other things, reviews can be used to gain understanding about a research problem or to identify further research directions~. Reviews can be used by authors to quickly obtain an overview of either emerging or mature research topics~. Review papers are important for research fields to develop. When review papers are lacking, the development of a research field is weakened~. Compiling literature review papers is a complicated task~ and is often more time consuming than performing original research~. The structure of such articles often consists of tables that compare published research contributions. Although in the literature the terms ``literature review'' and ``literature survey'' are sometimes used interchangeably, we make the following distinction. We refer to the tables in review articles as \\\\textit{literature surveys}. Together with a (textual) analysis and explanation, they form the \\\\textit{literature review}. The state-of-the-art (SoTA) analysis is a special kind of literature review with the objective of comparing the latest and most relevant papers in a specific domain. \\nWe implement the presented methodology in the ORKG. The ORKG is a scholarly infrastructure designed to acquire, publish and process \\\\emph{structured} scholarly knowledge published in the literature~. ORKG is part of a larger research agenda aiming at machine actionable scholarly knowledge that understands the ability to more efficiently compare literature as a key feature. \\nWe tackle the following research questions: \\n\\\\begin{enumerate}\\n \\\\item How to generate literature surveys using scholarly knowledge graphs?\\n \\\\item How to ensure that published literature surveys comply with the FAIR principles?\\n \\\\item How to effectively specify and visualize literature surveys in a user interface?\\n\\\\end{enumerate}\\nIn support of the first research question, we present a methodology that describes the steps required to generate literature surveys. In support of the second research question, we describe how the FAIRness of the published literature review is ensured. Finally, in support of the third research question, we demonstrate how the methodology is implemented within the ORKG. \\nThe paper is structured as follows. Section~\\\\ref{section:motivating-example} motivates the work. Section \\\\ref{section:related-literature} reviews related work. Section~\\\\ref{section:system-design} presents the system design, the underlying methodology and its implementation. Section~\\\\ref{section:data-collection} explains how the knowledge graph is populated with data. Section~\\\\ref{section:evaluation} presents the evaluation of the system, specifically system FAIRness and performance. Finally, Section~\\\\ref{section:discussion} discusses the presented and future work.\",\n \"id\": \"b5fdf9c3-044e-45b3-8bb1-0b7d691c74d6\",\n \"level\": \"section\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:motivating-example}\\nWe motivate our work by means of two use cases that underscore the usefulness of a literature survey generation system. In the first use case, a researcher wants to obtain an overview on state-of-the-art research addressing a specific problem. The second use case describes how a researcher can publish a FAIR compliant literature review with the ORKG.\",\n \"id\": \"b95d3fb3-5cd7-4c69-8c96-6cadc3ef7b13\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Use cases\"\n ]\n ],\n \"subsections\": [\n \"86a170e1-f5ee-40dd-9553-058683d000c4\"\n ],\n \"title\": \"Use cases\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"A state-of-the-art (SoTA) analysis reviews new and emerging research. They are useful for multiple reasons. Firstly, they provide a broad overview of a research problem and support understanding. Secondly, they juxtapose different approaches for a problem. Thirdly, they can support claims on why certain research is relevant by giving an overview of the breadth of research addressing a problem. \\nThe proposed approach enables automated generation of surveys to quickly obtain an overview of state-of-the-art research as well as sharing of surveys for others to reuse.\",\n \"id\": \"86a170e1-f5ee-40dd-9553-058683d000c4\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"b95d3fb3-5cd7-4c69-8c96-6cadc3ef7b13\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Use cases\"\n ],\n [\n \"paragraph\",\n \"Familiarize with the state-of-the-art\"\n ]\n ],\n \"subsections\": [\n \"b8ac6e76-2ccc-4355-afcd-c2adb9788bf8\"\n ],\n \"title\": \"Familiarize with the state-of-the-art\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:motivating-example:publish-reviews}\\nLiterature reviews typically consist of multiple (survey) tables in which different approaches from original papers are compared based on a set of properties. These tables can be seen as the main contribution and most informative part of the review paper, since the tables juxtapose and compare existing work. Comparison tables are published in review papers as static content in PDF documents. This presentational format is generated from datasets that typically contain more (structured) information than what is presented in the published table. However, the additional information is not published. It is ``dark data'' which is not stored or indexed and likely lost over time~. Furthermore, published tables are not machine actionable. Their overall low FAIRness hinders reusability of the published content. With the presented service, it is possible to publish a literature survey with high FAIRness, i.e. that is compliant with the FAIR principles to a high degree. Section~\\\\ref{section:related-literature} discusses this aspect in more details.\",\n \"id\": \"b8ac6e76-2ccc-4355-afcd-c2adb9788bf8\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"86a170e1-f5ee-40dd-9553-058683d000c4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Use cases\"\n ],\n [\n \"paragraph\",\n \"Familiarize with the state-of-the-art\"\n ],\n [\n \"paragraph\",\n \"Publishing of literature reviews\"\n ]\n ],\n \"subsections\": [\n \"34cdcc07-1e15-458e-bb25-8a7ebb70bcd8\"\n ],\n \"title\": \"Publishing of literature reviews\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The weaknesses of the current approach to literature review can be summarized as follows:\\n\\\\begin{itemize}\\n \\\\item {\\\\it Static} -- reviews are static, since once published as PDF they are rarely updated and there are no possibilities or incentives for creating new or updated reviews for considerable time.\\n \\\\item {\\\\it Lack of machine assistance} -- machine assistance is hardly possible, since the PDF representation of reviews is only human readable and relevant raw data is mostly not published along with the review.\\n \\\\item {\\\\it Delay} -- reviews are produced and published with significant delay (often years) after original research work was done.\\n \\\\item {\\\\it Coverage} -- due to the amount of work required, reviews are often only performed for relatively popular research topics and are stale or missing for less popular topics.\\n \\\\item {\\\\it Lacking collaboration} -- collaboration on reviews is not possible and reviews currently represent only the viewpoint of the few authors not the community.\\n \\\\item {\\\\it Missing overarching systematic semantic representation} -- the overlap between different reviews and related work sections in individual original research papers is not explicit and cannot be exploited.\\n\\\\end{itemize}\\nWe deem that these weaknesses of the current approach to scholarly literature review and synthesis significantly hinder scientific progress.\",\n \"id\": \"34cdcc07-1e15-458e-bb25-8a7ebb70bcd8\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"b8ac6e76-2ccc-4355-afcd-c2adb9788bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Use cases\"\n ],\n [\n \"paragraph\",\n \"Familiarize with the state-of-the-art\"\n ],\n [\n \"paragraph\",\n \"Publishing of literature reviews\"\n ],\n [\n \"paragraph\",\n \"Summary of weaknesses of the current approach to literature review\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary of weaknesses of the current approach to literature review\"\n },\n {\n \"cite_extract_rate\": 0.04,\n \"cites\": [\n 8369\n ],\n \"content\": \"\\\\label{section:related-literature}\\nThe task of comparing research contributions can be reviewed in light of the more general task of comparing resources (or entities) in a knowledge graph. While this is a well-known task in multiple domains (for instance in e-commerce systems~), not much work has focused on comparison in knowledge graphs, specifically. One of the few works with this focus is by \\\\citeauthor{Petrova2017}~ who created a framework for comparing entities in RDF graphs using SPARQL queries. In order to compare contributions, they first have to be found. Finding is an information retrieval problem. As a well-known technique, TF-IDF~ can be used for this task. More sophisticated techniques can be used to determine the structural similarity between graphs (e.g., ) and matching semantically similar predicates. This relates to dataset interlinking~ or more generally ontology alignment~. For property alignment, techniques of interest include edit distance (e.g., Jaro-Winkler~ or Levenshtein~) and vector distance. \\\\citeauthor{Gromann2019}~ found that fastText~ performs best for ontology alignment. \\nIn light of the FAIR Data Principles~, scholarly data should be Findable, Accessible, Interoperable and Reusable both for humans and machines. Due to the publication format, literature survey tables published in scholarly articles weakly adhere to the FAIR guidelines, particularly so for machines. Scholarly data should be considered first-class objects~, including data used to create literature surveys. \\\\citeauthor{Rodriguez-Iglesias2016}~ describe the difficulties of making data FAIR within the plant sciences. They argue that it is more complicated than reformatting data. On the other hand they suggest that most FAIR principles can be implemented relatively easily by using off-the-shelf technologies. \\\\citeauthor{Boeckhout2018}~ argue that the FAIR principles alone are not sufficient to lead to responsible data sharing. More applied principles are needed to ensure better scholarly data. This claim is supported by the findings of \\\\citeauthor{Mons2017}~ who suggest that there are very diverse interpretations of the guidelines. In their work, they try to clarify what is FAIR and what is not. \\nAn efficient literature comparison relies on scholarly knowledge being represented in a structured way. There is substantial related work on representing scholarly knowledge in structured form~. Building on the work of numerous philosophers of science, \\\\citeauthor{Hars2001}~ proposed a comprehensive scientific knowledge model that includes concepts such as theory, methodology and statement. More recently, ontologies were engineered to describe different aspects of the scholarly communication process. Semantic Publishing and Referencing (SPAR)\\\\footnote{http://purl.org/spar/\\\\{cito,c4o,fabio,biro,pro,pso,pwo,doco,deo\\\\}} is a collection of ontologies that can be used to describe scholarly publishing and referencing of documents~. \\\\citeauthor{IniestaCorcho:SePublica2014}~ reviewed the state-of-the-art ontologies to describe scholarly articles. \\\\citeauthor{Sateli2015SemanticRO}~ use some of these scholarly ontologies to add semantic representations of scholarly articles to the Linked Open Data cloud. A literature survey comparing scholarly ontologies is available via the ORKG.\\\\footnote{\\\\url{https://www.orkg.org/orkg/comparison/R8342}} Most of these ontologies are designed to capture metadata about and structure of scholarly articles, not the content communicated in articles. \\nAnother literature survey is created to compare approaches for semantically representing scholarly communication.\\\\footnote{\\\\url{https://www.orkg.org/orkg/comparison/R8364}}\\nAn initial attempt for semantifying review articles was done in~. The work comprises a relatively rigid ontology for describing contributions (mainly centered around research problems, approaches, implementations and evaluations) and a prototypical implementation using Semantic MediaWiki. We relax this constraint, since we are not limited by a rigid ontology schema but rather allow arbitrary domain-specific semantic structures for research contributions. The work by \\\\citeauthor{DBLP:conf/ercimdl/VahdatiFALV19}~ focuses on semantic article representations for generating literature overviews. Their method is to use crowdsourcing to generate the overviews. \\\\citeauthor{Kohl2018}~ present CADIMA, a system that supports systematic literature reviews. The tool supports the formal process of performing a literature review but does, for example, not publish data in machine actionable form for reuse.\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=\\\\columnwidth]{img/methodology.pdf}\\n \\\\caption{Research contribution comparison methodology.}\\n \\\\label{fig:workflow-comparison}\\n\\\\end{figure}\",\n \"id\": \"32d3f690-7c93-4310-b607-6553d1915112\",\n \"level\": \"section\",\n \"origin_cites_number\": 25,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Related work\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Related work\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:system-design}\\nWe now present the system design of the literature comparison service. It consists of a methodology that describes how to perform a comparison of research contributions. An early version of this methodology has been presented at the 3rd SciKnow workshop~. The methodology consists of five steps: 1) finding comparison candidates, 2) selecting related statements, 3) aligning contribution descriptions, 4) visualizing comparisons and 5) publishing FAIR comparisons. The methodology is depicted in Figure~\\\\ref{fig:workflow-comparison}. First, we discuss the data structure of the ORKG, which forms the foundation of the comparison. Then, each step of the methodology is described in more detail. Finally, we discuss the implementation.\",\n \"id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ]\n ],\n \"subsections\": [\n \"c64100b1-b57e-46ef-a8f6-a6c7131b671e\",\n \"d4eda852-27fa-4aa8-880a-559597b2dc69\",\n \"5d1f0f97-df75-4699-9f80-7b7af6dea650\",\n \"fbebdd62-b2b1-408f-9c69-5d4f7f13cfce\",\n \"1b062162-78e7-427e-8efd-91ae002ca800\",\n \"da78b76d-a336-4f78-9f93-2fb8a21e6591\",\n \"252651d1-6cd2-4bc9-abba-721f972b6d1e\"\n ],\n \"title\": \"System design\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:ontology}\\nIn ORKG, each paper is typed as \\\\textit{paper} class. A paper consists of at least one \\\\textit{research contribution}, which addresses at least one \\\\textit{research problem}. Research contributions consist of \\\\textit{contribution data} that describe the contribution. For instance, a paper in Computer Science might have descriptions for materials, methods, implementation and results as contribution data. These predefined core concepts can be easily extended with domain specific research problems, methods, etc. in ORKG curation using crowdsourcing or other curation approaches. The underlying data structure uses the notion of statements. Statements are triples that consist of a subject, a predicate (also called a property) and an object. The granularity of a comparison is at the research contribution, meaning that contributions are compared rather than papers. For simplicity, we use the terms ``paper comparison'' and ``contribution comparison'' interchangeably. Because a comparison happens on contribution level, it is possible to compare specific elements of a paper instead of the complete paper. The benefit of this is that a comparison does not contain data from irrelevant contributions. The ORKG OWL ontology is available online.\\\\footnote{\\\\url{https://gitlab.com/TIBHannover/orkg/orkg-ontology}}\",\n \"id\": \"c64100b1-b57e-46ef-a8f6-a6c7131b671e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"ORKG ontology\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ORKG ontology\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:workflow-comparison-candidates}\\nTo perform a comparison, a starting contribution is needed. This contribution is called \\\\textit{main contribution} and is always manually selected by a user. The main contribution is compared against other \\\\textit{comparison contributions}. There are two different approaches for selecting the comparison contributions. The first approach automatically selects comparison contributions based on similarity. The second approach lets users manually select contributions.\",\n \"id\": \"d4eda852-27fa-4aa8-880a-559597b2dc69\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Select comparison candidates\"\n ]\n ],\n \"subsections\": [\n \"dceac5ad-70aa-4f9b-94ef-abad8d3758f3\",\n \"2079aa06-4aa7-4551-8f0b-5dcb0826279d\"\n ],\n \"title\": \"Select comparison candidates\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Comparing contributions makes only sense when contributions can sensibly be compared. For example, it does not make (much) sense to compare a biology paper to a history paper. \\nWe thus argue that it makes only sense to compare contributions that are similar. More specifically, contributions that share the same (or a similar set of) properties are good comparison candidates. For instance, a paper about question answering has the property \\\\textit{orkg:disambiguationTask}\\\\footnote{\\\\textit{orkg:} denotes the ontology of the ORKG system described in Section~\\\\ref{section:ontology}} and another paper is using the same property to describe what disambiguation tasks are performed. Since they share the same property it makes them likely candidates for comparison. Finding similar contributions is therefore based on finding contributions that share the same or similar informative description properties. To achieve this, each comparison contribution is converted into a string by concatenating all properties of the contribution. \\nTF-IDF~ is used to query these strings with the string of the main contribution as query. The search returns the most similar contributions by weighting the most informative properties higher due to TF-IDF. The top-k contributions are selected and form a set of contributions that are used in the next step.\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=0.9\\\\columnwidth]{img/similar-contributions.png}\\n \\\\caption{Implementation of the first step of the methodology: the selection of comparison candidates. Showing both the similarity-based and the manual selection approaches.}\\n \\\\label{fig:comparison-candidates}\\n\\\\end{figure}\\nFigure~\\\\ref{fig:comparison-candidates} displays how the similar contribution selection is implemented. As depicted, three similar contributions are suggested to the user (with the corresponding similarity percentage being displayed next to paper title). These suggested contributions can be directly compared.\",\n \"id\": \"dceac5ad-70aa-4f9b-94ef-abad8d3758f3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"d4eda852-27fa-4aa8-880a-559597b2dc69\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Select comparison candidates\"\n ],\n [\n \"subsubsection\",\n \"Find similar contributions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Find similar contributions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"There are scenarios where comparison based on similarity computation is not suitable or desired. For example, a researcher wants to compare a specific set of implementations to see which performs best. \\nTherefore, the manual selection method is implemented in a similar fashion to an e-commerce shopping cart. When the ``Add to comparison'' checkbox is checked, a box appears listing the selected contributions (Figure~\\\\ref{fig:comparison-box}). \\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=.45\\\\columnwidth]{img/comparison-box.png}\\n \\\\caption{Box showing the manually selected contributions.}\\n \\\\label{fig:comparison-box}\\n\\\\end{figure}\",\n \"id\": \"2079aa06-4aa7-4551-8f0b-5dcb0826279d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"d4eda852-27fa-4aa8-880a-559597b2dc69\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Select comparison candidates\"\n ],\n [\n \"subsubsection\",\n \"Manual selection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Manual selection\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:workflow-related-statements}\\nThis step selects the statements from the graph related to the set of contributions selected in the previous step. Statements are selected transitively to match contributions in subject or object position. This search is performed until a predefined maximum transitive depth $\\\\delta$ has been reached. The intuition is that the deeper a property is nested the less likely is its relevance for the comparison. The process of selecting statements is repeated until depth $\\\\delta=5$ is reached. This number is chosen empirically to include statements that are not directly related to the contribution, but to exclude statements that are less relevant because they are nested too deep.\",\n \"id\": \"5d1f0f97-df75-4699-9f80-7b7af6dea650\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Select related statements\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Select related statements\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 8369,\n 6548\n ],\n \"content\": \"As described in the first step, comparisons are built using shared or similar properties of contributions. In case the same property has been used between contributions, these properties are grouped and form one \\\\textit{comparison row}. However, often different properties are used to describe the same concept. This occurs for various reasons. The most obvious reason is when two different ontologies are used to describe the same property. For example, for describing the population of a city, DBpedia uses \\\\textit{dbo:populationTotal} while WikiData uses \\\\textit{WikiData:population} (actually the property identifier is P1082; for the purpose here we use the label). When comparing contributions, these properties should be considered as equivalent. Especially for community-created knowledge graphs, differently identified properties likely exist that are, in fact, equivalent.\\nTo overcome this problem, we use pre-trained fastText~ word embeddings to determine the similarity of properties. If the similarity is higher than a predetermined threshold $\\\\tau$, the properties are considered equivalent and are grouped. This happens when the similarity threshold $\\\\tau \\\\geq 0.9$ (also empirically determined). In the end, each group of properties will be visualized as one row in the comparison table. The result of this step is a list of statements for each contribution, where similar properties are grouped. Based on this similarity matrix $\\\\gamma$ is generated \\n\\\\begin{equation}\\n \\\\gamma_{p_{i}} = \\\\left [ cos(\\\\overrightarrow{p_i}, \\\\overrightarrow{p_j}) \\\\right ]\\n \\\\label{eq:similarity_matrix}\\n\\\\end{equation}\\nwith $cos(.)$ as the cosine similarity of vector embeddings for property pairs $(p_i, p_j) \\\\in \\\\mathcal{P}$, whereby $\\\\mathcal{P}$ is the set of all contributions.\\nFurthermore, we create a mask matrix $\\\\Phi$ that selects properties of contributions $c_i \\\\in \\\\mathcal {C}$, whereby $\\\\mathcal{C}$ is the set of contributions to be compared. Formally,\\n\\\\begin{equation}\\n \\\\Phi_{i,j} = \\\\begin{cases}\\n1 & \\\\text{ if } p_{j} \\\\in c_i \\\\\\\\ \\n0 & \\\\text{ otherwise } \\n\\\\end{cases}\\n \\\\label{eq:mask_matrix}\\n\\\\end{equation}\\nNext, for each selected property $p$ we create the matrix $\\\\varphi$ that slices $\\\\Phi$ to include only similar properties. Formally,\\n\\\\begin{equation}\\n\\\\varphi_{i,j} =(\\\\Phi_{i,j})_{\\\\substack{c_i \\\\in \\\\mathcal{C}\\\\\\\\ p_j \\\\in sim(p) }}\\n\\\\label{eq:slice_mask}\\n\\\\end{equation}\\nwhere $sim(p)$ is the set of properties with similarity values $\\\\gamma[p] \\\\geq \\\\tau$ with property $p$. Finally, $\\\\varphi$ is used to efficiently compute the common set of properties~. This process is displayed in Algorithm~\\\\ref{alg:alining-properties}.\\n\\\\begin{algorithm}\\n\\\\caption{Align contribution descriptions}\\\\label{alg:alining-properties}\\n\\\\begin{algorithmic}[1]\\n\\\\Procedure{AlignProperties}{properties, threshold}\\n\\\\ForEach {property $p_1 \\\\in properties$}\\n \\\\ForEach {property $p_2 \\\\in properties$}\\n \\\\State $similarity \\\\gets$ cos(Embb($p_1$), Embb($p_2$))\\n \\\\If {$similarity > threshold$}\\n \\\\State $similarProps \\\\gets similarProps \\\\cup \\\\{ p_1, p_2 \\\\}$\\n \\\\EndIf\\n \\\\EndFor\\n\\\\EndFor\\n\\\\Return $similarProps$\\n\\\\EndProcedure\\n\\\\end{algorithmic}\\n\\\\end{algorithm}\",\n \"id\": \"fbebdd62-b2b1-408f-9c69-5d4f7f13cfce\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Align contribution descriptions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Align contribution descriptions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The next step of the workflow is to visualize the comparison and present the data in a human understandable format. Tabular format is often appropriate for visualizing comparisons since tables provide a good overview of data. Another aspect of the visualization is determining which properties should be displayed and which ones should be hidden. A property is displayed when it is shared among a predetermined amount $\\\\alpha$ of contributions, where $\\\\alpha$ mainly depends on comparison use and can be determined based on the total amount of contributions in the comparison. By default, only properties that are common to at least two contributions ($\\\\alpha \\\\geq 2$) are displayed. \\nAnother aspect of comparison visualization is the possibility to customize the resulting table. This is needed because of the similarity-based matching of properties and the use of predetermined thresholds. For example, users should be able to enable or disable properties. They should also get feedback on property provenance (i.e., the property's path in the graph). Ultimately, this contributes to a better user experience, with the possibility to manually correct mistakes made by the system. \\nFigure~\\\\ref{fig:comparison-table} displays a comparison for research contributions related to visualization tools published in the literature. In this example, four properties are displayed. Literals are displayed as plain text while resources are displayed as links. When a resource link is selected, a popup is displayed showing the statements related to this resource. The UI implements some additional features that are particularly useful to compare research contributions.\",\n \"id\": \"1b062162-78e7-427e-8efd-91ae002ca800\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Visualize comparison\"\n ]\n ],\n \"subsections\": [\n \"04b99168-d893-480d-972b-bd9514984b9f\"\n ],\n \"title\": \"Visualize comparison\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Users can customize comparisons including transposing the table as well as hiding and rearranging the properties. Especially the option to hide properties is helpful when contributions with many statements are compared. Only properties considered relevant to the user can be selected to display. Customizing the comparison table can be useful before exporting or sharing the comparison.\",\n \"id\": \"04b99168-d893-480d-972b-bd9514984b9f\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1b062162-78e7-427e-8efd-91ae002ca800\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Visualize comparison\"\n ],\n [\n \"paragraph\",\n \"Customization\"\n ]\n ],\n \"subsections\": [\n \"6d0b1e3c-a8a3-4dd7-a2ef-5cb37393a9a0\",\n \"0851ca07-6e0f-4b3d-8617-8e8005a1d81b\"\n ],\n \"title\": \"Customization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Comparisons can be shared using a persistent link. Especially when sharing the comparison for research purposes, it is important to refer to the original comparison. Since contribution descriptions may change over time comparisons may also change. To support persistency, the whole state of the comparison is stored in a document-oriented database and retrieved when the permalink is invoked.\",\n \"id\": \"6d0b1e3c-a8a3-4dd7-a2ef-5cb37393a9a0\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"04b99168-d893-480d-972b-bd9514984b9f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Visualize comparison\"\n ],\n [\n \"paragraph\",\n \"Customization\"\n ],\n [\n \"paragraph\",\n \"Sharing and persistence\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Sharing and persistence\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"It is possible to export comparisons in different output formats such as PDF, CSV, RDF and \\\\LaTeX. The \\\\LaTeX~export is useful for direct integration in research papers. Together with the \\\\LaTeX~table, a BibTeX file containing the bibliographic information of the papers used in the comparison is also generated. Also, a persistent link referring back to the comparison in ORKG is showed as table footnote.\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=0.95\\\\columnwidth]{img/comparison-table.png}\\n \\\\caption{Comparison of research contributions related to visualization tools.}\\n \\\\label{fig:comparison-table}\\n\\\\end{figure}\",\n \"id\": \"0851ca07-6e0f-4b3d-8617-8e8005a1d81b\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"04b99168-d893-480d-972b-bd9514984b9f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Visualize comparison\"\n ],\n [\n \"paragraph\",\n \"Customization\"\n ],\n [\n \"paragraph\",\n \"Export\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Export\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Visualized and customized comparison tables can be stored. Storing tables is part of the publishing process and therefore only needed when a generated table is going to be used in a paper. In order to regenerate the table the whole state of the comparison should be saved. The knowledge graph from which the comparison was generated changes over time and thus storing just the URIs of the respective papers would not suffice. While saving a comparison, the user can provide additional metadata to ensure findability, an aspect of the FAIR principles. Metadata include a comparison title, which would normally consist of a one sentence description of the comparison. Additionally, a longer textual description can be provided. This metadata is extended with machine generated data, such as the creation date and the creator of the comparison. The metadata is stored in the knowledge graph to support easy access and interoperability. In Figure~\\\\ref{fig:RDF-comparison}, the structure of the metadata is displayed using the Dublin Core Metadata Terms\\\\footnote{\\\\url{https://dublincore.org/specifications/dublin-core/dcmi-terms}}. The comparison data itself is stored in a document-oriented database. An RDF export of both the metadata and the comparison data can be generated. The comparison data is modeled with the RDF Data Cube Vocabulary\\\\footnote{\\\\url{https://www.w3.org/TR/vocab-data-cube}}. A unique identifier is attached when the comparison is saved. This ID is used when the comparison is shared or when it is referenced in a paper. The literature comparison can also be performed without publishing. Although the workflow and the steps to create a comparison stay the same, the goal is different. Instead of creating a comparison that will be published and referenced in a paper, the comparison will be used by the researcher herself. \\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=.85\\\\columnwidth]{img/RDF-comparison.pdf}\\n \\\\caption{The graph structure of the metadata for a published comparison. The \\\\textit{dcterms:} prefix denotes the Dublin Core Metadata Terms ontology.}\\n \\\\label{fig:RDF-comparison}\\n\\\\end{figure}\",\n \"id\": \"da78b76d-a336-4f78-9f93-2fb8a21e6591\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Publish comparison\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Publish comparison\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 8253\n ],\n \"content\": \"The user interface of the comparison feature is seamlessly integrated with the ORKG front end, which is written in JavaScript and is publicly available\\\\footnote{\\\\url{https://gitlab.com/TIBHannover/orkg/orkg-frontend}}. The back end of the comparison feature is a service separate from the ORKG back end written in Python and also available Open Source\\\\footnote{\\\\url{https://gitlab.com/TIBHannover/orkg/orkg-similarity}}. The comparison back end is responsible for step two and three of the comparison methodology. The input in step two is the set of contribution IDs. The API selects the related statements and aligns the properties and returns the data needed to visualize the comparison. This data includes the list of papers, list of all properties and the values per property. \\n\\\\begin{table*}[t]\\n\\\\caption{List of imported survey tables in the ORKG. The paper and table reference can be used to identify the original table. }\\n\\\\label{table:imported-review-tables}\\n\\\\begin{adjustbox}{scale=.90}\\n\\\\begin{tabular}{l|l|l|r|l|l}\\n\\\\toprule\\n\\\\textbf{Paper reference} & \\\\textbf{Table reference} & \\\\textbf{Research problem} & \\\\textbf{Papers} & \\\\textbf{ORKG representation} & \\\\textbf{Information loss} \\n\\\\\\\\ \\\\midrule\\n\\\\citeauthor{Bikakis2016}~ & \\nTable 1 & Generic visualizations & \\n11 &\\n\\\\url{https://orkg.org/orkg/c/pdLJDk} & \\nNo \\n\\\\\\\\ \\n\\\\citeauthor{Bikakis2016}~ & \\nTable 2 & Graph visualizations & \\n21 &\\n\\\\url{https://orkg.org/orkg/c/Rx476Z} & \\nNo\\n\\\\\\\\ \\n\\\\citeauthor{Diefenbach2018a}~ & \\nTable 2 & \\nQuestion answering evaluations & \\n33 &\\n\\\\url{https://orkg.org/orkg/c/gaVisD} & \\nNo\\n\\\\\\\\ \\n\\\\citeauthor{Diefenbach2018a}~ & \\nTable 3,4,5,6 & \\nQuestion answering systems & \\n26 &\\n\\\\url{https://orkg.org/orkg/c/IuEWl2} & \\nNo\\n\\\\\\\\ \\n\\\\citeauthor{Hussain2017a}~ & \\nTable 4 &\\nAuthor name disambiguation &\\n5 &\\n\\\\url{https://orkg.org/orkg/c/vDxKdr} & \\nNo\\n\\\\\\\\\\n\\\\citeauthor{Hussain2017a}~ & \\nTable 5 &\\nAuthor name disambiguation &\\n6 &\\n\\\\url{https://orkg.org/orkg/c/XXg8Wg} & \\nNo\\n\\\\\\\\ \\n\\\\citeauthor{Hussain2017a}~ & \\nTable 6 &\\nAuthor name disambiguation &\\n9 &\\n\\\\url{https://orkg.org/orkg/c/9rOwPV} & \\nNo\\n\\\\\\\\ \\n\\\\citeauthor{Hussain2017a}~ & \\nTable 7 &\\nAuthor name disambiguation &\\n6 &\\n\\\\url{https://orkg.org/orkg/c/mB7kIK} & \\nNo\\n\\\\\\\\ \\n\\\\citeauthor{Naidu2018}~ & \\nTable 4 &\\nText summarization &\\n52 &\\n\\\\url{https://orkg.org/orkg/c/OUqYB9} & \\nNo\\n\\\\\\\\ \\n\\\\bottomrule\\n\\\\end{tabular}\\n\\\\end{adjustbox}\\n\\\\end{table*}\",\n \"id\": \"252651d1-6cd2-4bc9-abba-721f972b6d1e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"a1917066-53fd-432d-9c65-ea4825cda6c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"System design\"\n ],\n [\n \"subsection\",\n \"Technical details\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Technical details\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:data-collection}\\nIn order to generate useful literature reviews it is crucial for the knowledge graph to contain sufficient and relevant papers. Populating the knowledge graph with high quality paper descriptions it not straightforward. Structured descriptions of papers should be created in such a way that it is possible to compare papers based on shared properties. Both published papers and papers that will be published in the future should be added to the ORKG, retrospectively or prospectively. Although a comprehensive description on how to populate the ORKG is out-of-scope here, we now briefly describe how we envision populating the ORKG in a manner that would facilitate comparing contributions.\\nProspectively, authors can become part of generating structured descriptions of their papers. This should be done in a crowdsourced manner and can become part of the paper submission process. Input templates that collect relevant properties can be used to ensure structured and comparable paper descriptions. Retrospectively, automated (machine learning) methods can be helpful ensure scalability of the process of adding a paper.\",\n \"id\": \"def65367-1d9c-4420-92cd-cd2f5cd6b277\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Data collection\"\n ]\n ],\n \"subsections\": [\n \"a72860f0-4d1e-4b2f-aa43-73a96ef0f3e7\"\n ],\n \"title\": \"Data collection\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To populate the ORKG with comparable paper descriptions, we leverage the data published in review papers. Review papers consist of high quality, curated and often structured data that is collected from a set of papers that address the same (or a similar) research problem. Hence, using reviews to populate a scholarly knowledge graph is a relatively straightforward approach to obtain high quality structured paper descriptions. We now present a methodology to convert survey paper data into a knowledge graph structure. The steps are as follows:\\n\\\\begin{enumerate}\\n \\\\item \\\\textbf{Survey paper selection.} The first step is the selection of survey papers that are suitable for building a knowledge graph. Firstly, the survey should compare peer-reviewed scientific articles. For instance, a comparison of different systems without a reference to peer-reviewed work is not suitable for the scholarly knowledge graph. Secondly, the review should compare the papers' content in a structured way and should not merely list work in a field. Especially reviews that present their results and literature comparisons in tabular format are suitable. The result of this step is a list of papers that will be added to the ORKG. \\n \\\\item \\\\textbf{Table selection.} Given the selected survey papers, tables have to be selected. Some surveys contain only one table while in others multiple tables are presented. In some cases a collection of tables can be joined into one larger table. \\n \\\\item \\\\textbf{Data modeling.} Given the selected tables, a suitable graph structure has to be determined. The data structure has to be modeled. For instance, when implemented systems are compared, a suitable structure could be: \\\\texttt{[has implementation] -> System name}. The referenced system can be described with a list of properties to be compared. Additionally, a research problem has to be defined, which is typically the same for all papers that are part of the table. \\n \\\\item \\\\textbf{Metadata collection.} Next, the metadata for the papers that are referenced in the survey table is collected. In case a referenced paper has a DOI\\\\footnote{Digital Object Identifier}, the metadata can be automatically retrieved via a lookup service (e.g. Crossref\\\\footnote{\\\\url{https://www.crossref.org}}). Otherwise, at least the title, authors and publication date have to be collected.\\n \\\\item \\\\textbf{Data ingestion.} Finally, the paper data is ingestion into the knowledge graph. The paper data consists of both the paper's metadata and the extracted data from the comparison table. \\n This does not result in a single description of the survey paper. Each paper referenced in the survey table is ingested individually. In order to speed up the process of adding papers, we developed a Python package\\\\footnote{\\\\url{https://gitlab.com/TIBHannover/orkg/orkg-pypi}} that has a function to add a paper to the knowledge graph. \\n\\\\end{enumerate}\\nThis methodology has been used to populate the ORKG with comparable paper data. The data is used to evaluate the presented literature review tool. The imported paper data is not only useful for the evaluation, but does also provide significant value to the ORKG itself.\\nIn total, four review papers were selected for importing into the ORKG. The Python script for importing the table data is available online.\\\\footnote{\\\\url{https://gitlab.com/TIBHannover/orkg/orkg-papers}} From those papers, 12 different tables were imported. Together, 169 papers were reviewed in those four survey papers. This resulted in a total amount of $3\\\\,750$ statements being added to the knowledge graph. Table~\\\\ref{table:imported-review-tables} lists the imported review papers and tables. The survey papers address different research problems. Figure~\\\\ref{fig:graph-example-import} depicts an excerpt of the resulting graph for one particular paper. A set of comparison tables made with the imported data is available online.\\\\footnote{\\\\url{https://orkg.org/orkg/featured-comparisons}} This list includes some alternative comparison tables that were generated with the same data. \\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=0.8\\\\columnwidth]{img/graph-example-import.pdf}\\n \\\\caption{Partial graph structure of an imported paper. Orange colored resources indicate potentially interesting values for a paper comparison.}\\n \\\\label{fig:graph-example-import}\\n\\\\end{figure}\",\n \"id\": \"a72860f0-4d1e-4b2f-aa43-73a96ef0f3e7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"def65367-1d9c-4420-92cd-cd2f5cd6b277\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Data collection\"\n ],\n [\n \"subsection\",\n \"Leverage legacy review paper tables\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Leverage legacy review paper tables\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:evaluation}\\nIn this section, we present an evaluation of multiple aspects of the presented comparison methodology and implementation. Firstly, we evaluate information representation. Then, we evaluate the FAIRness of published reviews. Finally, we present a performance evaluation that tests the scalability.\",\n \"id\": \"9c67205a-5ea7-42a6-af06-049ac9b5d00b\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ]\n ],\n \"subsections\": [\n \"52143d93-69a3-4b3d-a716-7d4ad8e48679\",\n \"c31bac6f-f294-4381-8adf-fd5c75c400a7\",\n \"34ac5be8-a18c-4ef0-bc7d-dc420b39ab8d\"\n ],\n \"title\": \"Evaluation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"This part of the evaluation focuses on the aspect of information representation. We use the data from the imported review papers, as described in Section~\\\\ref{section:data-collection}. In order to build and publish useful and correct literature reviews, at a minimum our service should display the same information that was originally presented in the review tables. This means that there should not be information loss when review tables are published using our service. If there is no information loss, it means our service can be used as an alternative to the current way of publishing review tables. Apart from generating the same table, the added value comes from the ability to aggregate new (tabular) views using the same data as well as the increased FAIRness of the data published via our service. For each of the imported review tables, listed in Table~\\\\ref{table:imported-review-tables}, we can evaluate whether the same table can be generated with our service. For this, we have compared the table from the review paper to the table generated by the ORKG comparison service. A collection of 169 paper with 9 distinct literature views/tables are part of this evaluation. These tables can be viewed online, the links are listed in the ``ORKG representation'' column. The results of this evaluation are displayed in the same table, in column ``Information loss''. As the results show, using our service it is possible to recreate the same tabular views as originally published in the review papers.\",\n \"id\": \"52143d93-69a3-4b3d-a716-7d4ad8e48679\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"9c67205a-5ea7-42a6-af06-049ac9b5d00b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ],\n [\n \"subsection\",\n \"Information representation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Information representation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As described before, with the presented service it is possible to publish a generated comparison that adheres with the FAIR principles. Because the service leverages a knowledge graph to generate and save comparisons, complying with the FAIR principles is more obvious for the ORKG comparison service than for tables in published PDF articles. In order to evaluate the FAIRness of a published comparison, we evaluate each of the four FAIR principles in detail. \\\\citeauthor{Wilkinson2016}~ described each principle by assigning sub-principles.\\\\footnote{For a more detailed definition of the FAIR principles, see: \\\\url{https://go-fair.org/fair-principles}} We discuss the relevant sub-principles and explain how they are met. We use the term (meta)data to refer to both the actual comparison data (i.e., the data that is used to create the comparison table) and the associated metadata (e.g., the title, description and creator of a comparison). Table \\\\ref{table:fair-principles-compliance} presents an overview for the evaluation of the FAIR principles.\",\n \"id\": \"c31bac6f-f294-4381-8adf-fd5c75c400a7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"9c67205a-5ea7-42a6-af06-049ac9b5d00b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ],\n [\n \"subsection\",\n \"FAIR data evaluation\"\n ]\n ],\n \"subsections\": [\n \"ac11071f-8ae2-423d-ab33-85fce076119e\"\n ],\n \"title\": \"FAIR data evaluation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To make data findable for both humans and machines (i.e., agents), a unique and persistent identifier should be attached to the data (F1). Additionally, metadata should describe the data (F2). In the metadata, the unique identifier of the data should be mentioned (F3). Also a search interface should be available to find the data (F4). To ensure the findability of comparisons, users can title and describe them. Furthermore, machine generated metadata is attached to a comparison (e.g., the number of papers and the creation date). A unique identifier is generated and attached to the data and included in the metadata. Finally, the ORKG search interface allows users to search the whole graph and has a dedicated filter to specifically find comparisons. Additionally, comparisons can be indexed and found by third-party search engines (such as Google or Bing).\",\n \"id\": \"ac11071f-8ae2-423d-ab33-85fce076119e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"c31bac6f-f294-4381-8adf-fd5c75c400a7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ],\n [\n \"subsection\",\n \"FAIR data evaluation\"\n ],\n [\n \"paragraph\",\n \"Findable.\"\n ]\n ],\n \"subsections\": [\n \"154af8e0-ed85-46d2-9570-9da5491ee971\",\n \"68397955-d4af-4601-8b25-5a6aef856a6b\",\n \"a1388af8-78d9-4c21-9ff4-2ccd2d4a9694\"\n ],\n \"title\": \"Findable.\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Having found data, agents need to know how to access it. This principle is primarily about using accessible standardised communication protocols (A1). Additionally, metadata should be available even when the data is not (A2). The metadata is part of the knowledge graph, which can be accessed via the HTTP protocol. The data can be accessed without authentication. To support A2, the metadata and the actual comparison data are stored separately. Therefore, it is possible to access only metadata when the original data is not available anymore (for example when data is retracted by the author).\",\n \"id\": \"154af8e0-ed85-46d2-9570-9da5491ee971\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"ac11071f-8ae2-423d-ab33-85fce076119e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ],\n [\n \"subsection\",\n \"FAIR data evaluation\"\n ],\n [\n \"paragraph\",\n \"Findable.\"\n ],\n [\n \"paragraph\",\n \"Accessible.\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Accessible.\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To ensure the interoperability of data, it should use a formal language for knowledge representation (I1) and should use vocabularies that are FAIR (I2). Finally, references or links to other (meta)data should be made (I3). As argued before, thanks to highly structured data and the integration of shared vocabularies, interoperability is an inherent feature of knowledge graphs. Data is (partially) described using the ORKG core ontology and other ontologies we use to canonicalize the representation of relevant information content types. Links to other data are present in the knowledge graph. For example, if a comparison uses the ``Web'' resource to specify the domain of an application, this resource is generic, can be shared among paper descriptions and comparisons, and can be described in more detail, independently of a particular comparison.\",\n \"id\": \"68397955-d4af-4601-8b25-5a6aef856a6b\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"ac11071f-8ae2-423d-ab33-85fce076119e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ],\n [\n \"subsection\",\n \"FAIR data evaluation\"\n ],\n [\n \"paragraph\",\n \"Findable.\"\n ],\n [\n \"paragraph\",\n \"Interoperable.\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Interoperable.\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Finally, data should be reuseable. This can be accomplished by adding relevant (meta)data (R1). Required are an accessible data license (R1.1) and detailed provenance (R1.2) data. Finally, (meta)data should use community standards to describe data (R1.3). It is possible to add additional metadata to a comparison, e.g. metadata about the scope of the comparison, which could be a reference to the paper in which the comparison is being used. The metadata is complemented with the metadata that is already part of the Findability principle, e.g. provenance data about the creator of the comparison. The data license of the graph data is CC BY-SA\\\\footnote{\\\\url{https://creativecommons.org/licenses/by-sa/2.0}} (Attribution-ShareAlike), which allows reuse of the data. There is currently no community standard to describe the comparison data. However, standard ontologies are used to describe metadata (e.g., Dublin Core).\\n\\\\begin{table}[t]\\n\\\\centering\\n\\\\caption{Overview of FAIR principles compliance.}\\n\\\\label{table:fair-principles-compliance}\\n\\\\begin{adjustbox}{scale=.85}\\n\\\\begin{tabular}{l|c|p{67mm}}\\n\\\\toprule\\n\\\\textbf{Principle} & \\\\textbf{Level} & \\\\textbf{Explanation} \\\\\\\\ \\\\midrule\\n\\\\multicolumn{3}{l}{\\\\textit{Findable}} \\\\\\\\ \\\\midrule\\nF1 & 3 & Unique IDs exist, DOI assignment for future work \\\\\\\\ \\nF2 & 2 & Machine and user generated metadata is attached \\\\\\\\ \\nF3 & 1 & Properties used to link data to metadata \\\\\\\\ \\nF4 & 1 & Comparisons are findable via a search interface \\\\\\\\ \\\\midrule\\n\\\\multicolumn{3}{l}{\\\\textit{Accessible}} \\\\\\\\ \\\\midrule\\nA1 & 2 & Data is accessed over HTTP (via REST or a user interface), requires user effort to integrate the ORKG API specification \\\\\\\\ \\nA1.1 & 1 & The protocol is free and widely used \\\\\\\\ \\nA1.2 & 1 & No authentication is required to access the data \\\\\\\\ \\nA2 & 1 & Metadata is stored in a persistent way and available without the data itself \\\\\\\\ \\\\midrule\\n\\\\multicolumn{3}{l}{\\\\textit{Interoperable}} \\\\\\\\ \\\\midrule\\nI1 & 1 & RDF (with type assertions) and CSV export of comparisons\\\\\\\\\\nI2 & 2 & Reuse of ontologies where possible (ORKG core, Dublin core, RDF Data Cube Vocabulary). User responsible for other ontology reuse. \\\\\\\\ \\nI3 & 3 & For comparisons, the compared paper metadata is linked. More references are needed and can be created by users. \\\\\\\\ \\\\midrule\\n\\\\multicolumn{3}{l}{\\\\textit{Reusable}} \\\\\\\\ \\\\midrule\\nR1 & 1 & Machine and user generated metadata is created while publishing \\\\\\\\\\nR1.1 & 1 & CC-BY SA license \\\\\\\\ \\nR1.2 & 1 & If a registered user publishes a comparison, the user is associated with the published data \\\\\\\\ \\nR1.3 & 2 & Users can describe contributions using domain-relevant ontologies \\\\\\\\ \\n\\\\bottomrule\\n\\\\end{tabular}\\n\\\\end{adjustbox}\\n1=Yes; 2=Yes, requires user effort; 3=Partially/future work\\n\\\\end{table}\\n\\\\noindent\\nThe evaluation of the FAIR principles shows that comparisons published with our service rank high in FAIRness, which can be even further increased with some effort from users. Users are mainly responsible for adding the correct information to the comparison and reuse vocabularies. Otherwise, findability, accessibility and to some extent also interoperability are largely handled by the service.\",\n \"id\": \"a1388af8-78d9-4c21-9ff4-2ccd2d4a9694\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"ac11071f-8ae2-423d-ab33-85fce076119e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ],\n [\n \"subsection\",\n \"FAIR data evaluation\"\n ],\n [\n \"paragraph\",\n \"Findable.\"\n ],\n [\n \"paragraph\",\n \"Reusable.\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Reusable.\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In order to evaluate the performance of the overall comparison, we compared the implemented ORKG approach to a naive approach for comparing multiple resources. The naive approach compares each property against all other properties to perform the property alignment. Table~\\\\ref{tab:sota-eval} shows the time needed to generate comparisons, for both the naive and the ORKG approach. In total, eight papers are compared with on average ten properties per paper. In the naive approach, the ``Align contribution descriptions'' step is not scaling well, since each property is compared against all others. If multiple contributions are selected, the number of property similarity checks grows exponentially. Table~\\\\ref{tab:sota-eval} shows that the ORKG approach outperforms the naive approach. The total number of papers used for the evaluation is limited to eight because the naive approach does not scale to larger sets. \\n\\\\begin{table}[t]\\n\\\\caption{Time (in seconds) to perform comparisons with 2-8 contributions using the naive and ORKG approaches.}\\n\\\\begin{adjustbox}{scale=.9}\\n\\\\begin{tabular}{l|lllllll}\\n& \\\\multicolumn{7}{c}{\\\\textbf{Number of compared research contributions}} \\\\\\\\\\n\\\\hline\\n & \\\\textbf{2} & \\\\textbf{3} & \\\\textbf{4} & \\\\textbf{5} & \\\\textbf{6} & \\\\textbf{7} & \\\\textbf{8} \\\\\\\\ \\\\hline\\n\\\\textbf{Naive} & 0.00026 & 0.1714 & 0.763 & 4.99 & 112.74 & 1772.8 & 14421 \\\\\\\\\\n\\\\textbf{ORKG} & 0.0035 & 0.0013 & 0.01158 & 0.02 & 0.0206 & 0.0189 & 0.0204 \\n\\\\end{tabular}\\n\\\\end{adjustbox}\\n\\\\label{tab:sota-eval}\\n\\\\end{table}\",\n \"id\": \"34ac5be8-a18c-4ef0-bc7d-dc420b39ab8d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"9c67205a-5ea7-42a6-af06-049ac9b5d00b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Evaluation\"\n ],\n [\n \"subsection\",\n \"Performance evaluation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Performance evaluation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section:discussion}\\nOne of the aims of the contribution comparison functionality is to support literature reviews and make this activity less cumbersome and time consuming for researchers. To live up to this aim, more structured contribution descriptions are needed. Existing scholarly knowledge graph initiatives focus primarily on scholarly metadata, while with ORKG we focus on making the actual research contributions machine readable. Currently, the ORKG does not yet contain sufficient contribution descriptions in order for the comparison functionally to be practically useful for researchers. Furthermore, for an evaluation of the effectiveness of certain components of the methodology (such as finding related papers or aligning similar properties), more contribution data is needed. Publishing surveys does not rely on data quantity and is therefore evaluated more extensively in this work. The performance evaluation results indicate that the comparison feature performs well. This means the technical infrastructure is in place for the literature survey service.\\nIn the evaluation, we focused on the aspects of the system that are necessary for researchers to use the system in practice. The information representation evaluation is a straightforward evaluation to see if existing survey tables can be regenerated with the ORKG. This is a minimal requirement for researchers when using the system, since they should at least be able to recreate tables. This evaluation does not give insight to the usefulness and usability of the system, but still provides an indication that the service can be successfully used to publish literature surveys. One of the reasons for using the service is that also ``dark data'' in comparisons is published (as discussed in Section~\\\\ref{section:motivating-example:publish-reviews}).\\nAnother interesting aspect of the service is that published literature surveys rank high in FAIRness. Therefore, the second part of the evaluation focuses on how the FAIR principles are met. Merely publishing data as RDF is not sufficient to fully meet the FAIR principles. Hence, we conducted a more detailed evaluation that describes how the service complies with each sub-principle. Since FAIR is not a \\\\textit{standard}, the principles are permissive and not prescriptive~. No technical requirements are specified. Both the implementation and evaluation of the guidelines are therefore subject to interpretation. With respect to data interoperability and reusability, certain aspects of the service can be improved. For example, to improve interoperability, the contribution data should be reusing existing vocabularies where possible. Additionally, although most of FAIRification is done by the system, the researcher is responsible for adding correct and relevant metadata while publishing a survey.\\nAs indicated earlier, the usefulness of the presented tool depends on the number of papers present in the knowledge graph. Therefore, future work will focus on data collection, both in a crowdsourced and automated manner. We plan on extending the methodology presented in Section~\\\\ref{section:data-collection} with automated extraction of data and tables from literature review papers. With the extracted review data, the knowledge graph can be extended more quickly than the previously presented manual method. It could form the basis of a high quality scholarly knowledge graph that contains relevant and FAIR survey table data. Furthermore, in the future we will assign (DataCite) DOIs to published surveys. They will serve as a persistent identifier for the survey data~.\",\n \"id\": \"f87ef12c-02fb-41f5-9f16-69dee5190666\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Discussion \\\\& future work\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion \\\\& future work\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Reviewing existing literature is an important but cumbersome and time consuming activity. To address this problem, we presented a methodology and service that can be used to generate literature surveys from a scholarly knowledge graph. This service can be used by researchers in order to get familiar with existing literature. Additionally, the tool can be used to publish literature surveys in a way that they largely adhere to the FAIR data principles. The presented methodology addresses multiple aspects, including finding suitable contributions, aligning contribution descriptions, visualization and publishing. The methodology is implemented within the Open Research Knowledge Graph (ORKG). Since the comparison relies on structured scholarly knowledge, we discussed how to populate the ORKG with relevant data. This is done by extracting tabular survey data from existing literature reviews. In order to evaluate whether the proposed service can be used to publish literature surveys, the original survey table representations were compared with the ones generated by our service. As the results indicate, it is possible to use the service as an addition or potentially even replacement of the current publishing approach, since the same tables can be generated. The evaluation also showed how the published literature surveys largely adhere to the FAIR data principles. This is crucial for data reusability and machine actionability. To conclude, the proposed literature comparison service addresses multiple weaknesses of the current survey publishing approach and can be used by researchers to generate, publish and reuse literature surveys. \\n\\\\begin{acks}\\nThis work was co-funded by the European Research Council for the project ScienceGRAPH (Grant agreement ID: 819536) and the TIB Leibniz Information Centre for Science and Technology. We want to thank Kheir Eddine Farfar for his contributions to this work.\\n\\\\end{acks}\\n\\\\bibliographystyle{ACM-Reference-Format}\\n\\\\bibliography{refs,mendeley}\\n\\\\end{document}\\n\\\\endinput\",\n \"id\": \"4c8c53d6-d473-464f-a034-d85b2b90268e\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3313ad50-7887-4970-bb3f-a8d5311e5f61\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Generate FAIR Literature Surveys with Scholarly Knowledge Graphs\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":72,"string":"72"},"all_cites":{"kind":"list like","value":[6548,6549,8369,8253],"string":"[\n 6548,\n 6549,\n 8369,\n 8253\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":2.126841101429957,"string":"2.126841"}}},{"rowIdx":38,"cells":{"authors":{"kind":"list like","value":["Yantong Wang","Vasilis Friderikos"],"string":"[\n \"Yantong Wang\",\n \"Vasilis Friderikos\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey of Deep Learning for Data Caching in Edge Network"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-08-17T12:02:32Z","string":"2020-08-17T12:02:32Z"},"category":{"kind":"string","value":"cs.NI"},"abstract":{"kind":"string","value":" The concept of edge caching provision in emerging 5G and beyond mobile networks is a promising method to deal both with the traffic congestion problem in the core network as well as reducing latency to access popular content. In that respect end user demand for popular content can be satisfied by proactively caching it at the network edge, i.e, at close proximity to the users. In addition to model based caching schemes learning-based edge caching optimizations has recently attracted significant attention and the aim hereafter is to capture these recent advances for both model based and data driven techniques in the area of proactive caching. This paper summarizes the utilization of deep learning for data caching in edge network. We first outline the typical research topics in content caching and formulate a taxonomy based on network hierarchical structure. Then, a number of key types of deep learning algorithms are presented, ranging from supervised learning to unsupervised learning as well as reinforcement learning. Furthermore, a comparison of state-of-the-art literature is provided from the aspects of caching topics and deep learning methods. Finally, we discuss research challenges and future directions of applying deep learning for caching. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"347fe6bc-3506-4394-85f3-39521077022f","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"]],"subsections":["c1e3d948-68a4-41aa-9453-ed02e0504669","77d13750-5344-4e36-b9ce-8573ab07d97a","c0dca0bc-798e-46e6-903c-5acb8a815d31","3f275de1-b1c2-48e0-ae07-49947668a760","bcb206fa-772b-4e9f-9da9-3ca19215514a","9f55687b-d2fc-4fa2-a01d-e44198b97b4c"],"title":"root"},{"cite_extract_rate":0.777777777777777,"cites":[3705,3582,3704,3706,3420,7633,3703],"content":"\\label{sec:introduction}\nUndoubtedly, future 5G and beyond mobile communication networks will have to address stringent requirements of delivering popular content at ultra high speeds and low latency due to the proliferation of advanced mobile devices and data rich applications. In that ecosystem, edge-caching has received significant research attention over the last decade as an efficient technique to reduce delivery latency and network congestion especially during peak-traffic times or during unexpected network congestion episodes by bringing popular data closer to the end users. One of the main reasons of enabling edge caching in the network is to reduce the number of requests that traverse the access and core mobile network as well as reducing the load at the origin servers that would have to, otherwise, respond to all requests directly in absence of edge caching. In that case popular content and objects can be stored and served from edge locations, which are closer to the end users. This operation is also beneficial from the end user perspective since edge caching can dramatically reduce the overall latency to access the content and increase in the sense overall user experience. It is also important to note that the notion of popular content means that the requests of top 10\\% of video content on the Internet account for almost 80\\% of all traffic; which relates to multiple requests from different end users of the same content\n.\nRecently, deep learning (DL) has attracted significant attention from both academia and industry and has been applied to diverse domains like self-driving, medical diagnosis, playing complex games such as Go . DL has also made their way into communication areas . In this paper, we pay attention to the application of DL in caching policy. Though there are some earlier surveys related to machine learning applications, they either focus on general machine learning techniques for caching , or concentrate on overall wireless applications . The work provides a big picture of applying machine learning in wireless communications. In , the authors consider the machine learning on both caching and routing strategy. A comprehensive survey on machine learning applications for caching content in edge networks is provided in . The researchers provide a survey about machine learning on mobile edge caching and communication resources. On the other hand, overviews how artificial neural networks can be employed for various wireless network problems. The authors in detail a survey on deep reinforcement learning (DRL) for issues in communications and networking. presents a comprehensive on deep learning applications and edge computing paradigm. Our work can be distinguished from the aforementioned papers based on the fact that we focus on the deep learning techniques on content caching and both wired and wireless caching are taken into account. Our main contributions are listed as follows:\n\\begin{itemize}\n\t\\item We classify the content caching problem into Layer 1 Caching and Layer 2 Caching. Each layer caching consists of four tightly coupled subproblems: where to cache, what to cache, cache dimensioning and content delivery. Related researches are provided accordingly. \n\t\\item We present the fundamentals of DL techniques which are widely used in content caching, such as convolutional neural network, recurrent neural network, actor-critic model based deep reinforcement learning, etc.\n\t\\item We analyze a broad range of state-of-the-art literature which use DL to content caching. These papers are compared based on the DL structure, layer caching coupled subproblems and the objective of DL in each scenarios. Then we discuss research challenges and potential directions for the utilization of DL in caching. \n\\end{itemize}\n\\begin{figure}[htb]\n\t\\centering\n\t\\includegraphics[trim=0mm 0mm 0mm 0mm, clip, width=\\textwidth]{figure/paper.eps}\n\t\\caption{Survey Architecture}\n\t\\label{fig:organization}\n\\end{figure}\nThe rest of this survey is organized as follows (as illustrated in Figure \\ref{fig:organization} ). Section \\ref{sec:caching} presents the categories of content caching problem. Section \\ref{sec:dl} reviews typical deep neural network structures. In Section \\ref{sec:dl_caching}, we list state-of-the-art DL-based caching strategies and their comparison. Section \\ref{sec:challenges} debates challenges as well as potential research directions. In the end, Section \\ref{sec:conclusions} concludes this paper. For better readability, the abbreviations in this paper is listed as Table \\ref{tab:abbr} shows.\n\\begin{table}[htb]\n\t\\centering\n\t\\caption{\\label{tab:abbr} List of Abbreviations.}\n\t\\begin{tabular}{l|p{.38\\textwidth}|l|p{.38\\textwidth}}\n\t\t\\hline\n\t\t\\textbf{Abbr.} & \\textbf{Description} & \\textbf{Abbr.} & \\textbf{Description} \\\\\n\t\t\\hline\n\t\t3C & Computing, Caching and Communication & A3C & Asynchronous Advantage Actor-Critic \\\\\n\t\t\\hline\n\t\tBBU & Baseband Unit & CCN & Content-Centric Network \\\\\n\t\t\\hline\n\t\tCNN & Convolutional Neural Network & CoMP-JT & Coordinated Multi Point Joint Transmission \\\\\n\t\t\\hline\n\t\tCR & Content Router & C-RAN & Cloud-Radio Access Network \\\\\n\t\t\\hline\n\t\tCSI & Channel State Information & D2D & Device to Device \\\\\n\t\t\\hline\n\t\tDDPG & Deep Deterministic Policy Gradient & DL & Deep Learning \\\\\n\t\t\\hline\n\t\tDNN & Deep Neural Network & DQN & Deep Q Network \\\\\n\t\t\\hline\n\t\tDRL & Deep Reinforcement Learning & DT & Digital Twin \\\\\n\t\t\\hline\n\t\tED & End Device & ES & Edge Server \\\\\n\t\t\\hline\n\t\tETSI & \\multicolumn{3}{l}{European Telecommunication Standardization Institute} \\\\\n\t\t\\hline\n\t\tESN & Echo-State Network & FIFO & First In First Out\\\\\n\t\t\\hline\n\t\tFNN & Feedforward Neural Network & FBS & Femto Base Station \\\\\n\t\t\\hline\n\t\tICN & Information-Centric Network & LFU & Least Frequently Used \\\\\n\t\t\\hline\n\t\tLP & Linear Programming & LRU & Least Recently Used \\\\\n\t\t\\hline\n\t\tLSTM & Long Short-Term Memory & MAR & Mobile Augmented Reality \\\\\n\t\t\\hline\n\t\tMD & Mobile Device & MILP & Mixed Integer Linear Programming \\\\\n\t\t\\hline\n\t\tMBS & Macro Base Station & NFV & Network Function Virtualization\\\\\n\t\t\\hline\n\t\tPNF & Physical Network Function & PPO & Proximal Policy Optimization \\\\\n\t\t\\hline\n\t\tQoE & Quality of Experience & RL & Reinforcement Learning \\\\\n\t\t\\hline\n\t\tRNN & Recurrent Neural Network & RRH & Remote Radio Head \\\\\n\t\t\\hline\n\t\tSAE & Sparse Auto Encoder & SDN & Software Defined Network \\\\\n\t\t\\hline\n\t\tseq2seq & Sequence to Sequence & SNM & Shot Noise Model \\\\\n\t\t\\hline\n\t\tTRPO & Trust Region Policy Optimization & TTL & Time to Live \\\\\n\t\t\\hline\n\t\tVNF & Virtual Network Function & WSN & Wireless Sensor Network \\\\\n\t\t\\hline\n\t\\end{tabular}\n\\end{table}","id":"c1e3d948-68a4-41aa-9453-ed02e0504669","level":"section","origin_cites_number":9,"parent_id":"347fe6bc-3506-4394-85f3-39521077022f","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:caching}\nThe paradigm of data caching in edge networks is illustrated in Figure \\ref{fig:caching}. Similarity to , the scope of edge in this paper is along the path between end user and data server, which contains Content Router (CR), Macro Base Station (MSB), Femto Base Station (FSB) and End Device (ED). In the context of Cloud-Radio Access Network (C-RAN), both baseband unit (BBU) and remote radio head (RRH) are considered as potential caching candidates to hosting content, where the BBUs are clustered as a BBU pool centrally and RRHs are deployed near BS's antenna distributively. According to the hierarchical structure of edge network, the data caching is classified into two categories: Layer 1 Caching and Layer 2 Caching. In this section we illustrate the typical research topics in these two areas.\n\\begin{figure}[htb]\n \\centering\n \\includegraphics[trim=1mm 1mm 5mm 1mm, clip, width=\\textwidth]{figure/Cache.eps}\n \\caption{Data Caching in Edge Network}\n \\label{fig:caching}\n\\end{figure}","id":"77d13750-5344-4e36-b9ce-8573ab07d97a","level":"section","origin_cites_number":2,"parent_id":"347fe6bc-3506-4394-85f3-39521077022f","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Data Caching Review"]],"subsections":["93519284-cc11-4b04-b82f-f74d3c6a42fc","a2d45d41-6ccf-4b15-99cb-2ea399aa1ef2"],"title":"Data Caching Review"},{"cite_extract_rate":0.25,"cites":[3709,3707,3710,3708],"content":"In Layer 1 Caching, the popular content is considered to be hosted in CRs. In the context of Information-Centric Network (ICN), CR plays dual roles both as a typical router (i.e. data flow forwarding) and content store (i.e. local area data caching facility). Generally, the CR is connected via wired networks. Layer 1 Caching consists of four tightly coupled problems: where to cache, what to cache, cache dimensioning and content delivery.\nWhere to cache focuses on selecting the proper CRs to host the content. For instance, in Figure \\ref{fig:caching}, contents replicas can be placed in lower hierarchical level CRs, such as router B and C, as the mean of reducing transmission cost but with extra cost pay for hosting contents; reversely, consolidating caching in CR A can be adopted to saving caching cost at the expense of more transmission cost and has the risk of expiring end users' delay requirement. Here the caching/hosting cost is the cost to deploy the content, which could be measured by space utilization, energy consumption or other metrics. The transmission cost represents the price for delivering the content from cached CR (or data server) to end user and is basically estimated via the number of hops. Where to cache problem usually has been modelled as a Mixed Integer Linear Programming (MILP):\n\\begin{subequations} \n\\label{fml:MILP}\n\\begin{align}\n\\label{MILP:obj}\n\\mathop{\\min}_{\\substack{x}}\\; & c^T x\\\\\n\\textrm{s.t.}\\quad \\label{MILP:con1}\n& Ax\\leq b \\\\\n\\label{MILP:con2}\n& x\\in\\{0,1\\} \\\\\n\\label{MILP:con3}\n\\text{or}\\quad & x\\geq 0\n\\end{align}\n\\end{subequations}\nwhere $x$ is the decision variable. Normally it is a binary variable indicating the CR assignment. In special cases, with the aim of modelling or linearization, some non-binary auxiliary variables are introduced as constraint \\eqref{MILP:con3} shows. If taking caching a part of a file not the complete into consideration, the decision variable $x$ is a continuous variable representing the segments host in the CR, then constraint \\eqref{MILP:con2} becomes $x\\in[0,1]$ and MILP model turns to linear programming (LP). There are many work allocating contents via MILP with different objectives and limitations. The authors in propose a model to minimize the user delay and load balancing level of CRs with the satisfaction of cache space. The work in considers a trade-off between caching and transmission cost with cache space, link bandwidth and user latency constraints. In , an energy efficient optimization model is constructed consisting of caching energy and transport energy. provides more details of mathematical model and related heuristic algorithms in caching deployment of wired networks. \nWhat to cache concentrates on selecting the proper contents in CRs for the purpose of maximizing the cache hit ratio. Via exploiting the statistical patterns of user requests, the popularity of requested information and user preference can be forecasted and play a very significant role in determining caching content. On the one hand, from the view of aggregated request contents, researchers propose many different models and algorithms for popularity estimation. One widely used model in web caching is the Zipf model based the assumption that the content popularity is static and each users' request is independent. However, this method fails to reflect the temporal and spatial correlations of the content, where the temporal correlation reflects the popularity varies over time and the spatial correlation represents the content preference is different on the geographical area and social cultural media. A temporal model named the shot noise model (SNM) is built in which enables users to estimate the content popularity dynamically. Inspired by SNM, the work in considers both spatial and temporal characteristics during caching decision. On the other hand, from the view of a specific end user during a certain period, caching his/her preference content (may not be the popular in network) can also help to reduce the traffic flow. Many approaches in recommendation systems can be applied in this case. Another aspect of what to cache problem is the designing of cache eviction strategies when storage space faces the risk of overflow. Depending on the life of caching contents, these policies can be divided into two categories roughly: one is like first in first out (FIFO), least frequently used (LFU), least recently used (LRU) and randomized replacement, the contents would not be removed until no more memory is available; the other one is called time to live (TTL) strategy, where the eviction happens once the related timer expires. presents analytic model for hit ratio in TTL-based cache requested by independent and identically distributed flows. It worth noting that in , the TTL-based cache policy is used for the consistency of dynamic contents instead of contents replacement. In , the authors introduce a TTL model for cache eviction and the timer is reset once related content cache hit happens.\nCache dimensioning highlights how much storage space to be allocated. Benefit from the softwarization and virtualization technologies, the cache size in each CR or edge cloud can be managed in a more flexible and dynamical way, which makes the cache dimensioning decisions become an important feature in data caching. Technically, the cache hit ratio rises with the increasing of cache memory, and consequently eases the traffic congestion in the core network. However, excessive space allocation would waste the resource like energy to support the caching function. Hence there is a trade-off between cache size cost and network congestion. Economically, taking such scenario into consideration: a small content provider wants to rent service from a CDN provider such as Akamai or Huawei Cloud, and there is also a balance between investment saving and network performance. In , the proper cache size of individual CR in Content-Centric Network (CCN) is investigated via exploiting the network topology. In , the authors consider the effect of network traffic distribution and user behaviours when designing cache size.\nContent delivery considers how to transform the caching content to the requested user. The delivery traffic embraces single cache file downloading and video content steaming and the metrics for these two scenarios vary. Regarding file downloading, the content cannot be consumed until the delivery is completed. Therefore the downloading time of the entire file is viewed as a metric to reflect the quality of experience (QoE). For video steaming, especially for those large video splitted into several chunks, the delay limitation only works on the first chunk. In that case, delivering the first chunk in time and keep the smooth transmission of the rest chunks are the key aims . Apart from those measuring metrics, another problem in content delivery is the routing policy. In CCN , one implementation of ICN architecture, employs a flooding-based name routing protocol to publish the request among cached CRs. On one hand, flooding strategy simplifies the designing complexity and reduce the maintaining cost particularly in an unstable scenario; on the other hand, it costly wastes bandwidth resources. In , the authors discuss the optimal radius in scoped flooding. The deliver route is often considered jointly with where to cache problem, in which the objective function \\eqref{MILP:obj} includes both deployment and routing cost.","id":"93519284-cc11-4b04-b82f-f74d3c6a42fc","level":"subsection","origin_cites_number":16,"parent_id":"77d13750-5344-4e36-b9ce-8573ab07d97a","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Data Caching Review"],["subsection","Layer 1 Caching"]],"subsections":[],"title":"Layer 1 Caching"},{"cite_extract_rate":0.6000000000000001,"cites":[3712,3716,8682,3714,3717,3713,8681,3715,3711],"content":"Contrast to Layer 1 caching in wired connection, Layer 2 caching considers implementing caching techniques in wireless network. Though both of them need solve where to cache, what to cache, cache dimensioning and content delivery problems, wireless caching is more challenging and some mature strategies in wired caching cannot be migrated directly to wireless case. Some reasons come from the listed aspects: the resources in wireless environment, such as caching storage and spectrum, are limited compared with CRs in Layer 1 Caching; the mobility of end users and dynamic network typologies are also required to be considered during the design of caching strategies; moreover, the wireless channels are uncertain since they can be effected by fading and interference.\nIn wireless caching, where to cache focus on finding the proper candidates among MBS, FBS, ED, even BBU pool and RRU in C-RAN to host the content. Caching at MBS and FBS can alleviate backhaul congestion since end users obtain the requested content from BS directly instead of from CR via backhaul links. Compared with FBS, MBS has wider coverage and typically, there is no overlap among different MBSs . As mentioned above, the caching space in BSs is limited and it is impractical to cache all popular content. With the aim of improving cache-hit ratio, a MILP-modelled collaborative caching strategy among MBSs is proposed in . If the accessed MBS does not host the content, the request will be served by a neighbour MBS which cache the file rather than by the data server. For FBS caching, a distributed caching method is presented in and the main idea is that the ED locating in the FBS coverage overlap is able to obtain contents from multiple hosters. Caching at ED can not only ease backhaul congestion but also improve the area spectral efficiency . When the end user requests a content, he/she would be severed by the local storage if the content is precached in his/her ED or by adjacent ED via D2D communication if the content is host accordingly. In , the authors model the cache-enabled D2D network as a Poisson cluster process, where end users are grouped into several clusters and the collective performance is improved. Individually, caching the interested contents for other users affects personal benefit. In , a Stackelberg game model is applied to formulate the conflict among end users and a related incentive mechanism is designed to encourage content sharing. For the case of cache-enabled C-RAN, caching at BBU can ease the traffic congestion in the backhaul while caching at RRH can reduce the fronthaul communication cost. On the other hand, caching all at BBU raises the signaling overhead of BBU pool while at RRH weakens the processing capability. Therefore, where to cache the content in C-RAN makes a substantial contribution to balancing the signal processing capability at the BBU pool and the backhaul/fronthaul costs . The work in investigates caching at RRHs with jointly considering cell outage probability and fronthaul utilization. Due to the end users' mobility, the prediction/awareness of user moving behaviour also influence the proper hoster selection. There are some researches exploiting user mobility in cache strategy designing like and .\nSimilar with Layer 1, what to cache decision as well as eviction policy of layer 2 depends on the accurate prediction on content popularity or user preference in proactive caching method. The content popularity contains the feature of temporal and spatial correlations, which has already been described in Layer 1 Caching. In Layer 2 caching, the proper spatial granularity in popular contents estimation needs to take special attentions . For example, the coverage of MBS and FBS are different, which makes the popularity in MBS and FBS are different as well. Because the former based on a large number of users' behaviors but the individual may prefer specific content categories. For small cells, the preference estimation requires more accurate information like historical data . In order to capture the temporal and spatial dynamics of user preference, many different deep learning based algorithms are proposed, which will be illustrated in Section \\ref{sec:dl_caching}.\nCache dimensioning in Layer 2 Caching has more complicated factors need to be considered, not only including the network topology and content popularity as Layer 1 Caching, but also containing backhaul transmission status and wireless channel features. The proper cache size assignment is studied in the scenario of backhaul limited cellular network . It also provides the closed-form boundary of minimum cache size in one cell case. In the case of dense wireless network, the work in quantifies the minimum required cache to achieve the linear capacity scaling of network throughput. The authors of also consider the scenario of dense networks. They derive the closed-form of the optimal memory size which can reduce the consumption of backhaul capacity as well as guarantee wireless QoS.\nAccording to the number of transmitters and receivers, we divide the content delivery in Layer 2 caching into three categories: one candidate serves one end user, such as unicast and D2D transmission; one candidate serves multiple users like multicast; and coordinated delivery including multiple transmitters serve one or more receivers like coordinated multi-point joint transmission (CoMP-JT). Once the requested content is cached locally, BS can serve the end user via unicast or the adjacent device shares the contents by implementing D2D transmission. Concurrent transmission has the risk of co-channel interference in dense deployed networks. In D2D network, link scheduling is introduced to select subsets of links to transmit simultaneously . With the aim of improving the spectral efficiency, multicast is applied in content delivery when serving multiple requests simultaneously with the same content. Therefore there is a trade off between spectral efficiency and service delay. For the aim of serving more users in one transmission as well as higher spectral efficiency, the BS will wait to collect enough requirement for the same content which makes the first request a long waiting time. An optimal dynamic multicast scheduling is proposed in to balance these two factors. Multicast can also serve multiple requests with different contents. In , the authors provides a coded caching scheme which requires the communication link is error free and each user caches a part of its own content and partial of other users. Then BS multicasts the coded data to all users. Each user can decode his own requested content by XOR operation between the received data and the precached other users' file. However, the coding complexity increases exponentially as the quantity of end users grows. The CoMP-JT can improve the spectral efficiency as well via sharing channel state information (CSI) and contents among BSs but it also needs high-capacity backhaul consumption for exchanging data. In C-RAN, the BBUs are centralized in the BBU pool, which makes communication among BSs very efficiency. designs CoMP-JT in C-RAN for the purpose of minimizing power consumption with limitations of transmission energy, link capacity and requested QoS.","id":"a2d45d41-6ccf-4b15-99cb-2ea399aa1ef2","level":"subsection","origin_cites_number":15,"parent_id":"77d13750-5344-4e36-b9ce-8573ab07d97a","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Data Caching Review"],["subsection","Layer 2 Caching"]],"subsections":[],"title":"Layer 2 Caching"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:dl}\nAs Figure \\ref{fig:DNN} shows, some typical deep neural network (DNN) methods are stated. These models are classified into three categories depending on the training methods: supervised learning, unsupervised learning and reinforcement learning. \n\\begin{figure}[htb]\n \\centering\n \\includegraphics[trim=0mm 0mm 0mm 0mm, clip, width=\\textwidth]{figure/DNN.eps}\n \\caption{Typical DNN Structures}\n \\label{fig:DNN}\n\\end{figure}","id":"c0dca0bc-798e-46e6-903c-5acb8a815d31","level":"section","origin_cites_number":0,"parent_id":"347fe6bc-3506-4394-85f3-39521077022f","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"]],"subsections":["17a86b07-b7c3-41f4-9b7f-05a0e6ea6ae0","f6c49080-7dcd-4236-b8c7-a294adda5983","12cbfae2-b31a-489d-8357-cb9dedb5c4fa","66cec1fd-1f07-4527-b02b-b98ccb6507c3","cbd77acf-a82e-43b7-9bdd-bef9313a0f73"],"title":"Deep Learning Outline"},{"cite_extract_rate":1,"cites":[166],"content":"FNN is a kind of DNNs whose information propagation direction is forward and there is no cycle in neurons. In this paper, the term FNN is used to represent fully connected neural network, which indicates the connection between two adjacent layers is filled. According to the Universal Approximation Theorem, FNN has the ability to approximate any closed and bounded function with enough neurons in hidden layer . The hidden layer is applied to extract features of input vector, and then feed the output layer, which works as a classifier. Though FNN is very powerful, it gets into trouble when dealing with real-world task such as image recognition due to enormous weight parameters (because of fully connected) and lack of data augmentation.","id":"17a86b07-b7c3-41f4-9b7f-05a0e6ea6ae0","level":"subsection","origin_cites_number":1,"parent_id":"c0dca0bc-798e-46e6-903c-5acb8a815d31","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Feedforward Neural Network (FNN)"]],"subsections":[],"title":"Feedforward Neural Network (FNN)"},{"cite_extract_rate":1,"cites":[810],"content":"For the aim of overcoming the aforementioned drawback of FNN, CNN employs convolution and pooling operations, where the former applies sliding convolutional filters to the input vector and the later does down sampling, usually via maximum or mean pooling. Generally, CNN tends to contain deeper layers and smaller convolutional filters, and the structure becomes fully convolutional network , reducing the ratio of pooling layers as well as fully connected layers. Taxonomically, CNN belongs to FNN and has been broadly employed in image recognition, video analysis, natural language processing, etc. Including CNN, one of the limitations of FNN is that the output only depends on current input vectors. So it is hard to deal with sequential tasks.","id":"f6c49080-7dcd-4236-b8c7-a294adda5983","level":"subsection","origin_cites_number":1,"parent_id":"c0dca0bc-798e-46e6-903c-5acb8a815d31","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Convolutional Neural Network (CNN)"]],"subsections":[],"title":"Convolutional Neural Network (CNN)"},{"cite_extract_rate":0.5,"cites":[166],"content":"In order to deal with sequential tasks and using historical information, RNN employs neurons with self feedback in hidden layers. Unlike the hidden neuron in FNN, the output of recurrent neuron depends on both current output of precious layer and last hidden state. Compared with FNN approximates any continues functions, RNN with Sigmoid activation function can simulate a universal Turing Machine and has the ability to solve all computational problems . It is worth noting that RNN has the risk to suffer from long-term dependencies problem including gradient exploding and vanishing. Additionally, RNN has more parameters waiting to be trained due to adding recurrent weights. In the following, we introduce some RNN variants as Figure \\ref{fig:RNN} shows.\n\\begin{figure}[htb]\n \\centering\n \\includegraphics[trim=0mm 0mm 35mm 0mm, clip, width=\\textwidth]{figure/RNN.eps}\n \\caption{RNN Variants}\n \\label{fig:RNN}\n\\end{figure}","id":"12cbfae2-b31a-489d-8357-cb9dedb5c4fa","level":"subsection","origin_cites_number":2,"parent_id":"c0dca0bc-798e-46e6-903c-5acb8a815d31","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Recurrent Neural Network (RNN)"]],"subsections":["15615e34-4c33-4644-b6a2-80e552ca205d","f028c925-baae-474d-ae9c-0ae54c3029f3","bf6576fa-f9cd-4916-b753-6a9371a05fee"],"title":"Recurrent Neural Network (RNN)"},{"cite_extract_rate":1,"cites":[166],"content":"As aforementioned, simple RNN contains more parameters in training step, where the recurrent weights and input weights are difficult to learn . The basic idea of ESN is fixing these two kinds of weights and only learn the output weights (as links highlighted in Figure \\ref{fig:RNN}). The hidden layer is renamed as reservoir in ESN, where the neurons are sparsely connected and the weights are randomly assigned. The recurrent weights keep constant so the information of previous moments is stored in the reservoir with constant weight like voice echoing.","id":"15615e34-4c33-4644-b6a2-80e552ca205d","level":"subsubsection","origin_cites_number":1,"parent_id":"12cbfae2-b31a-489d-8357-cb9dedb5c4fa","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Recurrent Neural Network (RNN)"],["subsubsection","Echo-State Network (ESN)"]],"subsections":[],"title":"Echo-State Network (ESN)"},{"cite_extract_rate":1,"cites":[166],"content":"Recently, an efficient way to cope with long-term dependencies in practical is employing gated RNN, including LSTM . Then we compare with the recurrent neuron in simple RNN: Internally LSTM introduces three gates to control signal propagation, where input gate $I$ decides the partition of input signal to be stored, forget gate $F$ controls ratio of last moment memory to be kept until next period (the name \"forget gate\" may be a little misleading because it actually represents the ratio to be remembered) and output gate $O$ influences the proportion of current state to be delivered; Externally LSTM has four inputs embracing one input signal and three control signals for three gates. All these four signals are derived via the calculation of current network input and last moment delivered state.","id":"f028c925-baae-474d-ae9c-0ae54c3029f3","level":"subsubsection","origin_cites_number":1,"parent_id":"12cbfae2-b31a-489d-8357-cb9dedb5c4fa","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Recurrent Neural Network (RNN)"],["subsubsection","Long Short-Term Memory (LSTM)"]],"subsections":[],"title":"Long Short-Term Memory (LSTM)"},{"cite_extract_rate":1,"cites":[167],"content":"A typical application of RNN is converting one sequence to another sequence (seq2seq) such as machine translation. Conventionally, the output of seq2seq architecture is a probability distribution of output dictionary. However, it cannot deal with the problem that the size of output relies on the length of input due to fixed output dictionary. In , the authors modify the output to be the distribution of input sequence, which is analogous to pointers in C/C++. Pointer network has been widely used in text condensation.","id":"bf6576fa-f9cd-4916-b753-6a9371a05fee","level":"subsubsection","origin_cites_number":1,"parent_id":"12cbfae2-b31a-489d-8357-cb9dedb5c4fa","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Recurrent Neural Network (RNN)"],["subsubsection","Pointer Network"]],"subsections":[],"title":"Pointer Network"},{"cite_extract_rate":1,"cites":[166],"content":"Auto Encoder is a stack of two NNs named encoder and decoder respectively, where the former tries to learn the representative characteristics of input and generate a related code, and the later reads the code and reconstructs the original input. In order to avoid the auto encoder simply copying the input, some restrictions are considered like the dimension of code is smaller than input vector . The quality of auto encoder can be measured via reconstruction error, which estimates the similarity between input and output. In most cases, the auto encoder is used for the proper representation of input vector so the decoder part is removed after unsupervised training. The code can be employed as input for further deep learning models.","id":"66cec1fd-1f07-4527-b02b-b98ccb6507c3","level":"subsection","origin_cites_number":1,"parent_id":"c0dca0bc-798e-46e6-903c-5acb8a815d31","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Auto Encoder"]],"subsections":[],"title":"Auto Encoder"},{"cite_extract_rate":0,"cites":[],"content":"Reinforcement Learning (RL) is a Markov Decision Process represented by a quintuple $\\{\\mathcal{S,A,P,R},\\gamma\\}$, where $\\mathcal{S}$ is the state space controlled by environment; $\\mathcal{A}$ is the action space determined by agent; $\\mathcal{P}$ is the state transition function measuring the probability of moving to a new state $s_{t+1}$ given previous state $s_t$ and action $a_t$; $R$ is reward function calculated by environment considering state and action; $\\gamma$ is a discount factor for estimating total reward. During the interaction between agent and environment, agent observes current state $s_t$ from environment, and then takes action $a_t$ following its policy $\\pi$. The environment moves to a new state $s_{t+1}$ stochastically based on $\\mathcal{P}(s_t,a_t)$ and returns a reward $r_t$ to agent. The RL's aim is finding the policy $\\pi$ to maximum accumulated reward $\\sum_t\\gamma^tr_t$. In the early stage, RL focuses on scenarios whose $\\mathcal{S}$ and $\\mathcal{A}$ are discrete and limited. So the agent can use a table to record these information. Recently, some tasks have enormous discrete states and actions such as playing go and even continuous value such as self-driving, which makes table recording impractical. In order to solve this, DRL combines RL and DL, where RL defines the problem and optimization object; DL models the policy and the reward expectation. Depending on the roles of DNN in DRL, we classify the DRL into 3 categories as Figure shows.","id":"cbd77acf-a82e-43b7-9bdd-bef9313a0f73","level":"subsection","origin_cites_number":0,"parent_id":"c0dca0bc-798e-46e6-903c-5acb8a815d31","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Deep Reinforcement Learning (DRL)"]],"subsections":["d2c418b7-118c-42b8-b9d8-37f74087f743","4e5ada48-bb01-4d53-88dc-d7516f8cb54c","92185b95-3c92-401c-8a45-161c99f6f228"],"title":"Deep Reinforcement Learning (DRL)"},{"cite_extract_rate":0.6666666666666661,"cites":[1409,1408],"content":"In value-based method, DNN does not get involved with policy decision but estimates the policy performance. Two functions are introduced for the measurement: $V^\\pi(s)$ represents the reward expectation of policy $\\pi$ starting from state $s$; $Q^\\pi(s,a)$ illustrates the reward expectation of policy $\\pi$ starting from state $s$ and taking action $a$. In addition, $V^\\pi(s)$ is the expected value of $Q^\\pi(s,a)$. If we can estimate $Q^\\pi(s,a)$, the policy $\\pi$ can also be improved by choosing the action $a^*$ hold $Q^\\pi(s,a^*)\\geq V^\\pi(s)$. So the DNN employed in agent is approximating function $Q^\\pi(s,a)$, where the inputs are state $s$ and action $a$ and output is the estimated value $Q^\\pi(s,a)$. There are some representative critic methods like Deep Q Networks (DQN) and its variants Double DQN , Dueling DQN , etc.","id":"d2c418b7-118c-42b8-b9d8-37f74087f743","level":"subsubsection","origin_cites_number":3,"parent_id":"cbd77acf-a82e-43b7-9bdd-bef9313a0f73","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Deep Reinforcement Learning (DRL)"],["subsubsection","DNN as Critic (Value-Based)"]],"subsections":[],"title":"DNN as Critic (Value-Based)"},{"cite_extract_rate":1,"cites":[2219,1391],"content":"In policy-based method, DNN gets involved in the action selection directly instead of via $Q^\\pi(s,a)$. The policy can be viewed as an optimization problem, where the objective function is maximizing reward expectation and the search space is policy space. The input of DNN is current state and output is the probability distribution of potential actions. By employing gradient ascent, we can update the DNN to provide better action then maximize total reward. Some popular algorithms include Trust Region Policy Optimization (TRPO) , Proximal Policy Optimization (PPO) .","id":"4e5ada48-bb01-4d53-88dc-d7516f8cb54c","level":"subsubsection","origin_cites_number":2,"parent_id":"cbd77acf-a82e-43b7-9bdd-bef9313a0f73","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Deep Reinforcement Learning (DRL)"],["subsubsection","DNN as Actor (Policy-Based)"]],"subsections":[],"title":"DNN as Actor (Policy-Based)"},{"cite_extract_rate":0.5,"cites":[1390],"content":"Generally, compared with policy-based approach, the value-based method is less stable and suffer from poor convergence since the policy is derived based on $Q^\\pi(s,a)$ approximation. But value-based method is more sample efficient, while policy-based method is easier to fall into local optimal solution because the search space is vast. The actor-critic model combines these two approaches, i.e. the agent contains two DNNs named actor and critic respectively. In each training iteration, the actor considers current state $s$ and policy $\\pi$ for deciding action $a$. Then the environment changes to state $s'$ and returns reward $r$. The critic updates its own parameters based on the feedback from environment and output a mark for the actor's action. The actor updates the policy $\\pi$ depending on critic's mark. Some typical algorithms are proposed recent years like Deep Deterministic Policy Gradient (DDPG) and Asynchronous Advantage Actor-Critic (A3C) .","id":"92185b95-3c92-401c-8a45-161c99f6f228","level":"subsubsection","origin_cites_number":2,"parent_id":"cbd77acf-a82e-43b7-9bdd-bef9313a0f73","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning Outline"],["subsection","Deep Reinforcement Learning (DRL)"],["subsubsection","Actor-Critic Model"]],"subsections":[],"title":"Actor-Critic Model"},{"cite_extract_rate":0.162790697674418,"cites":[3721,7746,3718,3719,3723,3722,3720],"content":"\\label{sec:dl_caching}\nWe divide the studies regarding deep learning for data caching in edge networks into four categories depending on the DL tools employed: FNN and CNN; RNN; Auto Encoder; DRL. Recently many works utilize more than one DL techniques for jointly considered caching problems. For instance, at the beginning we applies a RNN to predict content popularity, and then a DRL to find suboptimal solutions of content placement for the purpose of reducing time complexity. In such case, we classify the related work into DRL since it represents the caching allocation policy. Unless mention the caching location (such as CRs, MBSs, FBSs, EDs and BBUs) otherwise, the approaches in this section can be utilized for both Layer 1 and Layer 2 caching. Table \\ref{tab:summary} summarize some studies of DL for caching.\n\\begin{comment}\n\\begin{longtable}[c]{p{.12\\textwidth}|l|p{.2\\textwidth}|p{.23\\textwidth}|p{.26\\textwidth}}\n \\caption{Summary of Deep Learning for Data Caching in Edge Network}\n \\label{tab:summary1}\\\\\n \\hline\n \\textbf{DL method} & \\textbf{Study} & \\textbf{Caching Problem} & \\textbf{Caching Objective} & \\textbf{DL Objective} \\\\\n \\endfirsthead\n \\hline\n \\textbf{DL method} & \\textbf{Study} & \\textbf{Caching Problem} & \\textbf{Caching Objective} & \\textbf{DL Objective} \\\\\n \\endhead\n \\hline\n CNN & & L1 where to cache \\& Content Delivery & minimize caching \\& transmission cost & nominate proper CRs for caching \\\\\n \\hline\n CNN & & L1 where to cache \\& Content Delivery & minimize caching \\& transmission cost & reduce feasible region of CRs for caching \\\\\n \\hline\n CNN\\& FNN & & L2 content delivery & minimize transmission time/energy & reduce feasible region of time slot allocation for delivery \\\\\n \\hline\n FNN & & L2 where to cache \\& content delivery & minimize energy cost & determine proper MBSs for caching \\& time duration for delivery \\\\\n \\hline\n Auto-encoder & & L1 \\& L2 what to cache & improve cache hit ratio & predict content popularity \\\\\n \\hline\n ESN & & L2 what to cache \\& content delivery & minimize transmit power & predict content popularity \\& user mobility \\\\\n \\hline\n DRL & & L2 what to cache \\& content delivery & minimize service latency & decide content caching, computing offloading and radio resource allocation \\\\\n \\hline\n LSTM & & L2 what to cache & improve cache hit ratio & predict content popularity \\\\\n \\hline\n FNN & & L2 what to cache & save traffic data & predict visiting content \\\\\n \\hline\n RNN \\& DRL & & L2 what to cache \\& where to cache & minimize service latency & predict popularity \\& decide content caching and offloading \\\\\n \\hline\n FNN & & L1 what to cache & improve cache hit ratio & predict content popularity \\\\\n \\hline\n Auto-encoder \\& RNN & & L2 content delivery & minimize transmission delay & reduce traffic load \\& select optimal BS subset\\\\\n \\hline\n CNN & & L2 what to cache & minimize service latency & predict visiting content \\\\\n \\hline\n ESN \\& LSTM \\& DRL & & L2 where to cache \\& what to cache \\& content delivery & minimize delay \\& power consumption & predict user mobility, content popularity \\& determine D2D link \\\\\n \\hline\n DRL & & L2 where to cache \\& content delivery & minimize content access cost & decide content caching \\& bandwidth allocation \\\\\n \\hline \n DRL & & L2 where to cache \\& what to cache & maximize cache hit ratio \\& minimize transmission delay & decide cache replacement \\\\\n \\hline\n auto-encoder \\& DRL & & L2 content delivery & minimize average delay and power consumption & decide multicast scheduling and caching jointly\\\\\n \\hline\n DRL & & L2 what to cache & maximize traffic load served by FBSs & decide which and how many contents be cached\\\\\n \\hline\n DRL & & L2 what to cache & minimize transmission power & predict users' preference and content popularity \\\\\n \\hline\n DRL & & L2 what to cache & maximize network operator's utility & decide content replacement \\\\\n \\hline\n DRL & & L2 what to cache & minimize service latency & decide cache replacement \\& power allocation \\\\\n \\hline\n RNN & & L2 what to cache & improve cache hit ratio & predict content popularity \\\\\n \\hline\n DRL & & L2 what to cache & maximize overall content quality & decide cache replacement \\\\\n \\hline \n DRL & & L2 what to cache & maximize cache hit ratio & decide caching strategy \\\\\n \\hline\n DRL & & L2 what to cache & maximize energy efficient & decide which content to be cached \\\\\n \\hline\n CNN \\& RNN \\& DRL & & L2 what to cache & maximize cost reduction of operators & predict popularity \\& searching best NN model \\\\\n \\hline\n LSTM & & L2 what to cache & maximize cache hit ratio & decide content replacement \\\\\n \\hline\n DRL & & L2 where to cache & maximize QoE & decide content placement \\\\\n \\hline\n DRL & & L2 what to cache & improve cache hit ratio & decide content update \\\\\n \\hline\n DRL & & L2 what to cache & minimize transmitting cost & decide cache replacement \\\\\n \\hline\n auto-encoder & & L2 what to cache & improve backhaul offloading \\& user satisfaction & predict content popularity \\\\\n \\hline\n DRL & & L2 what to cache & minimize cumulative cost & determine caching content \\\\\n \\hline \n DRL & & L2 what to cache & maximize chunk hit ratio \\& minimize service time & determine caching chunks \\\\\n \\hline\n\\end{longtable}\n\\end{comment}\n\\begin{table}[htb]\n\\caption{Summary of Deep Learning for Data Caching}\n\\label{tab:summary}\n\\begin{tabular}{p{.07\\textwidth}|p{.1\\textwidth}|p{.28\\textwidth}|p{.43\\textwidth}}\n \\hline\n \\textbf{method} & \\textbf{Study} & \\textbf{Caching Problem} & \\textbf{DL Objective}\\\\\n \\hline\n \\multirow{8}{.07\\textwidth}{FNN and CNN} & & content delivery & reduce feasible region of time slot allocation\\\\\n \\cline{2-4}\n & & where to cache, content delivery & determine MBSs for caching \\& delivery duration \\\\\n \\cline{2-4}\n & & where to cache, content delivery & nominate proper CRs for caching \\\\\n \\cline{2-4}\n & & where to cache, content delivery & reduce feasible region for caching \\\\\n \\cline{2-4}\n & & what to cache & extract video features\\\\\n \\cline{2-4}\n & & what to cache & predict requested content \\& frequency \\\\\n \\cline{2-4}\n & & what to cache & predict requested content \\\\\n \\cline{2-4}\n & & what to cache & predict content popularity\\\\\n \\hline\n \\multirow{4}{.07\\textwidth}{RNN} & & what to cache & predict requested content \\& user mobility \\\\\n \\cline{2-4}\n & & what to cache & predict content popularity \\\\\n \\cline{2-4}\n & & content delivery & reduce traffic load, select optimal BS subset\\\\\n \\hline\n \\multirow{3}{.07\\textwidth}{Auto Encoder} & & what to cache & predict content popularity \\\\\n \\cline{2-4}\n & & what to cache & predict top popular contents \\\\\n \\hline\n \\multirow{17}{.07\\textwidth}{DRL} & & what to cache & decide cache placement \\\\\n \\cline{2-4}\n & & what to cache & decide cache replacement \\& power allocation \\\\\n \\cline{2-4}\n & & what to cache & predict popularity \\& searching best NN model \\\\\n \\cline{2-4}\n & & where to cache & decide cache location \\\\\n \\cline{2-4}\n & & content delivery & users grouping\\\\\n \\cline{2-4}\n & & where to cache, content delivery & decide BS connection, computation offloading \\& caching location\\\\\n \\cline{2-4}\n & & what to cache, content delivery & decide caching \\& bandwidth allocation \\\\\n \\cline{2-4}\n & & what to cache, content delivery & decide caching, computing offloading \\& radio resource allocation \\\\\n \\cline{2-4}\n & & what to cache, content delivery & decide multicast scheduling \\& caching replacement\\\\\n \\cline{2-4}\n & & where \\& what to cache & predict popularity, decide caching \\& task offloading \\\\\n \\cline{2-4}\n & & where \\& what to cache, content delivery & predict user mobility \\& content popularity, determine D2D link\\\\\n \\hline\n\\end{tabular}\n\\end{table}","id":"3f275de1-b1c2-48e0-ae07-49947668a760","level":"section","origin_cites_number":43,"parent_id":"347fe6bc-3506-4394-85f3-39521077022f","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning for Data Caching"]],"subsections":["64ba4f69-8c92-44ec-9c4d-6de443abc8ba","8a40aabe-271f-491f-92cb-1da70950927c","67ee9973-0fe6-40d1-b44e-d9d39d452c5f","c63a6eb6-9774-4886-b68b-d0543cf7f42c"],"title":"Deep Learning for Data Caching"},{"cite_extract_rate":0.2,"cites":[3723,3721],"content":"In , the content delivery problem in wireless network is formulated as two MILP optimization models with the aims of minimum delivery time slot and energy consumption respectively. Both models consider the data rate for content delivery. Considering the computational complexity of solving MILP, a CNN is introduced to reduce the feasible region of decision variables, where the input is channel coefficients matrix. The FNN in paper plays a similar role as to simplify the searching space of the content delivery optimization model. \nFor resource allocation problem, the authors of model it as linear sum assignment problems then utilize CNN and FNN to solve the model. The idea is extended in and , where the authors consider where to cache problem among potential CRs and content delivery jointly, which is modeled as MILP with the aim of balancing caching and transmission cost by considering the user mobility, space utilization and bandwidth limitations. The cache allocation is viewed as multi-label classification problem and is decomposed into several independent sub-problems, where each one correlates with a CNN to predict assignment. The input of CNN is a grey-scale image which combines the information of user mobility, space and link utilization level. In , a hill climbing local search algorithm is provided to improve the performance of CNN while in , the prediction of CNN is used to feed a smaller MILP model. \nFor these above works , the FNN or CNN input is extracted from the optimization model. The work in trains a CNN via original graph instead of parameters matrix/image, which makes the process human recognizable and interpretable. Though the authors take traveling salesman problem not data caching as an example, the method can be viewed as a potential research direction. \nIn , an ILP model is proposed to minimize the backhaul video-data type load by determining the portion of cached content in BSs. Considering the fact that the mobile users covered by a BS change frequently, therefore predicting user preference is unnecessary. Instead, the authors concentrate on the popular content in general. At the beginning, a 3D CNN is introduced to extract spatio-temporal features of videos. The popularity of new contents without historical information is determined via comparing similar video features. The authors of also considers the spatio-temporal features among visiting contents in a mobile bus WiFI environment. By exploiting the previous 9 days collecting data, the content that the user may visit on the last day and corresponding visiting frequency can be forecast. The social property is taken into account in . By observing users interests on tweets during 2016 U.S. election, a CNN based predicted model can foresee the content category that is most likely to be requested. Such kind of content would be cached in MBSs and FBSs.\nThe work of examines the role of DNN in caching from another aspect. The authors propose a FNN to predict content popularity as a regression problem. The results show that FNN outperforms RNN, though the later is believed to be effective to solve sequential predictions. Moreover, replacing the FNN by a linear estimator does not devalue the performance significantly. The author provides explanation that FNN would work better than linear predictor in the case of incomplete information, and RNN has more advantages to model the popularity prediction as a classification rather than a regression problem.","id":"64ba4f69-8c92-44ec-9c4d-6de443abc8ba","level":"subsection","origin_cites_number":10,"parent_id":"3f275de1-b1c2-48e0-ae07-49947668a760","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning for Data Caching"],["subsection","FNN and CNN"]],"subsections":[],"title":"FNN and CNN"},{"cite_extract_rate":0.11111111111111101,"cites":[3720],"content":"Considering RNN is superior in dealing with sequential tasks, the work applies a bidirectional RNN for online content popularity prediction in mobile edge network. Simple RNN's output depends on previous and current storage, but the bidirectional can also take future information into account. The forecast model consists three blocks cascadingly: a CNN reads user requests and extract features; bidirectional LTSM learns association of requests over time step; FNN is added in the end to improve the prediction performance. Then content eviction is based on the popularity prediction. \nThe authors in utilize ESN to predict both content request distribution and end user mobility pattern. The user's preference is viewed as context which links with personal information combining gender, age, job, location, etc.. For the request prediction, the input of ESN is user's information vector and the output represent the probability distribution of content. For mobility prediction, the input includes historical and present user's location and the output is the expected position for next time duration. Eventually, the prediction influences the caching content decisions in BBUs and RRHs for the purpose of minimizing traffic load and delay in CRAN. The authors extend their work in by introducing conceptor-based ESN which can split users' context into different patterns and learn them independently. Therefore a more accurate prediction is achieved. \nIn , a caching decision policy named PA-Cache is proposed to predict time-variant video popularity for cache eviction when the space is full. The temporal content popularity is exploited by attaching every hidden layer representation of RNN to an output regression. In order to improve the accuracy, hedge backpropagation is introduced during training process which decides when and how to adapt the depth of the DNN in an evolving manner. Similarly, the work in also considers caching replacement of video content. A deep LSTM network is utilized for popularity prediction consisting of stacking multiple LSTM layers and one softmax layer, where the input of the network is request sequence data (device, timestamp, location, title of video) without any prepossessing and the output is estimated content popularity. Another work concentrates on prediction and interactions between user mobility and content popularity can be found .\nThe work of recognizes the popularity prediction as a seq2seq modeling problem and proposes LSTM Encoder-Decoder model. The input vector consists of past probabilities where each vectors are calculated during a predefined time window. In , the authors focus on caching content delivery with the aim of minimizing BSs to cover all requested users, i.e. set cover problem, via coded caching. Unlike , an auto encoder is introduced in coded caching stage for file conversion to reduce transmission load. In addition, a RNN model is employed to select BSs for broadcasting. \nThe paper shows the potential of RNN in solving where to cache problem. In , a task allocation model is formulated as a knapsack problem and the decision variables represent the task is processed locally in mobile devices (MDs) or remotely in edge servers (ESs). The authors design a multi-pointer network structure of 3 RNNs, where 2 encoders encode MDs and ESs respectively, 1 decoder demonstrates ES and MD pairing. Considering the similarity of where to cache optimization model and knapsack problem, the multi-pointer network can be transferred for caching location decision after according parameter modifications.","id":"8a40aabe-271f-491f-92cb-1da70950927c","level":"subsection","origin_cites_number":9,"parent_id":"3f275de1-b1c2-48e0-ae07-49947668a760","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning for Data Caching"],["subsection","RNN"]],"subsections":[],"title":"RNN"},{"cite_extract_rate":0,"cites":[],"content":"Generally, auto encoder is utilized to learn efficient representation or extract features of raw data in an unsupervised manner. The work in considers the cache replacement in wireless sensor network (WSN) based on content popularity. Considering sparse auto encoder (SAE) can extract representative expression of input data, the authors employ a SAE followed by a classifier where the input contains collecting user content requests and the output represents the contents popularity level. The authors also think about the implementation in a distributed way by SDN/NFV technical, i.e. the input layer is deployed on sink node, while the rest layers are implemented on the main controller. A related work applying auto encoder in 5G network proactive caching can be found in . In , two auto encoders are utilized for extracting the features of users and content respectively. Then the extracted information is explored to estimate popularity at the core network. Similarly, the auto encoder in is for spatio-temporal popularity features extraction and auto encoders work collaboratively in to predict top K popular videos.","id":"67ee9973-0fe6-40d1-b44e-d9d39d452c5f","level":"subsection","origin_cites_number":5,"parent_id":"3f275de1-b1c2-48e0-ae07-49947668a760","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning for Data Caching"],["subsection","Auto Encoder"]],"subsections":[],"title":"Auto Encoder"},{"cite_extract_rate":0.23809523809523803,"cites":[7746,3718,3719,3724,3722],"content":"The work in focuses on the cooperative caching policy at FBSs with maximum distance separable coding in ultra dense networks. A value-based model is utilized to determine caching categories and the content quantity at FBSs during off peak duration. The authors in study the problem of caching 360\\degree videos and virtual viewports in FBSs with unknown content popularity. The virtual viewport represents the most popular tiles of a 360\\degree video over users' population. A DQN is introduced to decide which tiles of a video to be hosted and in which quality. Additionally, employs DQN for content eviction decision offering a satisfactory quality of experience and is for the purpose of minimizing energy consumption. In , the authors also apply DQN to decide cache eviction in a single BS. Moreover, the critic is generated with stacking LSTM and FNN to evaluate Q value and an external memory is added for recording learned knowledge. For the purpose of improving the prediction accuracy, the Q value update is determined by the similarity of estimated value of critic and recording information in the external memory, instead of critic domination. The paper puts forth DQN a two-level network caching, where a parent node links with multiple leaf nodes to cache content instead of a single BS. In , a DRL framework with Wolpertinger architecture is presented for content caching at BSs. The Wolpertinger architecture is based on actor-critic model and performs efficiently in large discrete action space. employs two FNNs working as actor and critic respectively, where the former determines requested content is cached or not and the later estimates the reward. The whole framework consists two phases: in offline phase, these two FNNs are trained in supervised learning; in online phase, the critic and actor update via the interaction with environment. The authors extend their work to a multi agent actor-critic model for decentralized cooperative caching at multiple BSs . In , an actor-critic model is used for solving cache replacement problem, which balance the data freshness and communication cost. The aforementioned papers put attention on the network performance while ignore the influence of caching on information processing and resource consumption. Therefore, authors of design cache policy considering both network performance during content transmission and processing efficiency during data consumption. A DQN is employed to determine the number of chunks of the requested file to be updated. The paper investigates a joint cache replacement and power allocation optimization problem to minimize latency in a downlink F-RAN. A DQN is proposed for finding a suboptimal solution. Though is regarded as solving what to cache problem like , the reinforcement learning approach plays a different role. In , a DNN is utilized for content popularity prediction and then a RL is used for DNN hyperparameters tuning instead of determining caching content. Therefore the action space consists of choosing model architectures (i.e. CNN, LSTM, etc.), number of layers and layer configurations.\nIn , the authors generate an optimization model with the aim of maximizing network operator's utility in mobile social networks under the framework of mobile edge computing, in network caching and D2D communications (3C). The trust value which if estimated through social relationships among users are also considered. Then a DQN model is utilized for solving optimization problem, including determine video provider and subscriber association, video transcoding offloading and the video cache allocation for video providers.. The DQN employs two CNNs for training process, where one generates target Q value and the other is for estimated Q value. Unlike the conventional DQN, the authors in introduces a dueling structure, i.e. the Q value is not computed in the final fully connected layer, but is decomposed into two components and use the summary as estimated Q value, which helps achieve a more robust result. The authors also consider utilizing dueling DQN model in different scenarios like cache-enabled opportunistic interference alignment and orchestrating 3C in vehicular network . The work provides a DDPG model to cope with continuous valued control decision for 3C in vehicular edge networks, which is combined with the idea of DQN and actor-critic model. The DDPG structure can be divided into two parts as DQN, one is for estimated Q value and the other for target Q value. Each part consists of two DNNs, which play the role of actor and critic respectively. The critic updates its parameters like DQN while the actor learns policy via deterministic policy gradient approach. The proposed DRL is used for deciding content caching/replacement, vehicle organization and bandwidth resource assignment on different duration. \nThe paper provides an optimization model which takes what to cache and content delivery into consideration in the fog-enabled IoT network in order to minimize service latency. Since the wireless signals and user requests are stochastic, a actor-critic model is engaged where the actor makes decision for requesting contents while critic estimates the reward. Specially, the action space $S$ consists of decision variables and reward function is a variant of the objective function. A caching replacement strategy and dynamic multicast scheduling strategy are studied in . In order to get a suboptimal result, an auto encoder is used to approximate the state. Further, a weighted double DQN scheme is utilized for avoiding overestimation of Q value. applies a RNN to predict content popularity by collecting historical requests and the output represents the popularity in the near future. Then the prediction is employed for cooperative caching and computation offloading among MEC servers, which is modelled as a ILP problem. For the purpose of solving it efficiently, a multi-agent DQN is applied where each user is viewed as an agent. The action space consists of task local computing and offloading decision as well as local caching and cooperative caching determination. The reward is measured by accumulated latency. The agent choose its own action based on current state without cooperation. The where to cache, what to cache and content delivery decision of D2D network are jointly modelled in . Two RNNs, ESN and LSTM, are considered to predict mobile users' location and requested content popularity. Then the prediction result is used for determining content categories and cache locations. The content delivery is formulated as the actor-critic based DRL framework. The state spaces include CSI, transmission distances and communication power between requested user and other available candidates. The function of DRL is determining the communication link among users with the aim of minimizing power consumption and content delay.\nWe notice that most papers prefer to use value-based model (critic) and value-policy-based (actor-critic) model in DRL framework, but rare paper considers only policy-based model to solve data caching problem. One proper reason is that the search space of caching problem is enormous so policy-based model is easier to fall into local optimal solution, resulting in poor performance. Though the value-based model is less stable, some variant structures are utilized like Double DQN in to avoid value overestimation and dueling DQN in to improve robust.","id":"c63a6eb6-9774-4886-b68b-d0543cf7f42c","level":"subsection","origin_cites_number":21,"parent_id":"3f275de1-b1c2-48e0-ae07-49947668a760","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Deep Learning for Data Caching"],["subsection","DRL"]],"subsections":[],"title":"DRL"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:challenges}\nA serious of open issues on content caching and potential research directions are discussed in this section. We first extend the idea of content caching to virtual network function chain since caching can be viewed as a specific network function. Then we consider the caching for augmented reality applications. Moreover, we notice that the cache dimensioning has not been covered yet by DL methods. Finally, we debate the addition cost introduced by DL.","id":"bcb206fa-772b-4e9f-9da9-3ca19215514a","level":"section","origin_cites_number":0,"parent_id":"347fe6bc-3506-4394-85f3-39521077022f","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Research Challenges and Future Directions"]],"subsections":["e584c09a-75d8-4201-90c6-a0fd658b4342","f03f2b46-f5ed-47c7-b95e-abf7079d9dd7","76a7e5ab-c389-46f2-914c-27d36e617810","bf719ec3-fb89-4a44-9c42-480d45208596"],"title":"Research Challenges and Future Directions"},{"cite_extract_rate":0,"cites":[],"content":"The concept of Network Function Virtualization (NFV) has been firstly discussed and proposed within the realms of the European Telecommunication Standardization Institute (ETSI)\\footnote{Network Functions Virtualisation, An Introduction, Benefits, Enablers, Challenges and Call for Action, ETSI, 2012 https://portal.etsi.org/NFV/NFV\\_White\\_Paper.pdf}. The rational is to facilitate the dynamic provisioning of network services through virtualization technologies to decouple the service creation process form the underlying hardware. The framework allows network services to be implemented by a specific chaining and ordering of a set of functions which can be implemented either on a more traditional dedicated hardware which in this this case are called Physical Network Functions (PNFs), or alternatively as Virtual Network Functions (VNFs) which is a software running on top of virtualized general-purpose hardware. The decoupling between the hardware and the software is one of the important considerations the other – equally important – is that a virtualized service lend itself naturally to a dynamic programmable service creation where VNF resources can be deployed as required. Hence, edge cloud and network resource usage can be adapted to the instantaneous user demand whilst avoiding a more static over-provisioned configurations.\nWithin that framework, the incoming network service requests include the specification of the service function chain that need to be created in the form of an ordered sequence of VNFs. For example different type of VNFs such as a firewall or a NAT mechanism need to be visited in a specific order. In such constructed service chain each independent VNF requires specific underlying resources in terms for example of CPU cycles and/or memory. \nUnder this framework, caching of popular content can be considered as a specialized VNF chain function since inevitably delivery of the cached popular content to users will require a set of other functions to be supported related to security, optimization of the content etc. However, the issue of data caching and VNF chaining have evolved rather independently in the literature and the issue on how to optimize data caching when seeing it as part of VNF chain is still an interesting open ended issue.","id":"e584c09a-75d8-4201-90c6-a0fd658b4342","level":"subsection","origin_cites_number":0,"parent_id":"bcb206fa-772b-4e9f-9da9-3ca19215514a","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Research Challenges and Future Directions"],["subsection","Caching as a Virtual Network Function Chain"]],"subsections":[],"title":"Caching as a Virtual Network Function Chain"},{"cite_extract_rate":0,"cites":[],"content":"Mobile augmented reality (MAR) applications can be considered as a way to augment the physical real-world environment\nwith artificial computer-based generated information and is an area that has received significant research attention recently. In order to successfully superimpose different digital object in the physical world MAR applications include several computationally and storage complex concepts such as image recognition, mobile camera calibration, and also the use of advanced 2D and 3D graphics rendering. These functionalities are highly computationally intensive and as such require support from an edge cloud, in addition the virtual objects to be embedded in the physical world are expected to be proactively cached closer to the end user so that latency is minimized. Ultra low latency in these type of applications is of paramount importance so that to provide a photorealistic embedding of virtual objects in the video view of the end user. However, since computational and augmented reality objects need to be readily available, the caching of those objects should be considered in conjunction with the computational capabilities of he edge cloud. In addition to the above when MAR is considered under the lenses of an NFV environment the application might inherently require access to some VNFs and therefore the above discussion on VNF chaining for MAR applications is also valid in this case. \nRecently the concept of Digital Twin (DT) , has received significant research attention due to the plethora of applications ranging from industrial manufacturing and health to smart cities. In a nutshell, a DT can be defined as an accurate digital replica of a real world object across multiple granularity levels; and this real world object could be a machine, a robot or an industrial process or (sub) system. By reflecting the physical status of the system under consideration in a virtual space open up a plethora of optimization, prediction, fault tolerance and automation process that cannot be done using solely the physical object. At the core of DT applications is the requirement of stringent two–way real time communication between the digital replica and the physical object. This requirement inevitably require support from edge clouds to minimize latency and efficient storage and computational resources including caching. In that setting, the use of the aforementioned deep learning technologies will have a key role to play in order to provide high quality real time decision making to avoid misalignment between the digital replica of the physical object under consideration. Efficient machine-to-DT connectivity would require capabilities similar to the above mentioned augmented reality application but due to the continuous real-time control-loop operation DTs will require a complete new set of network optimization capabilities and in that frontier efficient caching and data-driven techniques will have a central role to play. Hence, as the research regarding the inter-play between low latency communications and DTs is still in embryonic stage there is significant scope in the investigation of suitable data driven deep learning techniques to be utilized for distributed allocation of caching and computing resources.","id":"f03f2b46-f5ed-47c7-b95e-abf7079d9dd7","level":"subsection","origin_cites_number":2,"parent_id":"bcb206fa-772b-4e9f-9da9-3ca19215514a","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Research Challenges and Future Directions"],["subsection","Caching for Mobile Augmented Reality (MAR) applications and Digital Twins (DTs)"]],"subsections":[],"title":"Caching for Mobile Augmented Reality (MAR) applications and Digital Twins (DTs)"},{"cite_extract_rate":0,"cites":[],"content":"As introduced in Section \\ref{sec:caching}, cache dimensioning explores the appropriate cache size allocation for content host such as CRs and BSs. Disappointingly, there is rare paper applies DL on cache dimensioning decisions. One proper reason is lack of training data set in contrast to content popular prediction, where we have historical user request log to train a DNN. In addition, the caching size allocation affects the network performance and economic investment. Recently, network slicing is identified as an important tool to enable 5G to provide multi-services with diverse characteristics. The slice is established on physical infrastructure including network storage. Therefore it is a very interesting topic to consider the allocation of the memory space to support content caching and other storage services, which guarantees QoE and satisfies task requirements. Furthermore, for the case lack of training data set, DRL can be viewed as a promising technology to configure slicing settings as well as cache dimensioning. For the action space designing, it can be either discrete by setting storage levels, or continuous which is allocate the memory space directly. However, there are requirements to design caching-enabled network slicing model especially for dynamic allocation as well as associated DRL framework including state space, detailed action space, reward function and agent structure.","id":"76a7e5ab-c389-46f2-914c-27d36e617810","level":"subsection","origin_cites_number":0,"parent_id":"bcb206fa-772b-4e9f-9da9-3ca19215514a","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Research Challenges and Future Directions"],["subsection","Deep Learning for Cache Dimensioning"]],"subsections":[],"title":"Deep Learning for Cache Dimensioning"},{"cite_extract_rate":0,"cites":[],"content":"Though the application of DL brings performance efficiency for caching policy, additional cost introduced by DL is unneglected, since training and deploying DL model require not only network resources but also time duration. Naturally there is a trade-off between the cost which DL-assisted caching policy saved and the consumption which supports DL itself running, which indicates the trading with DL results in either profit, loss, or break even. Therefore, where and when to apply DL should be carefully investigated. In addition, for the purpose of reducing resource consumption and accelerating training process, some knowledge transfer methods like transfer learning can be utilized, which can transform the knowledge already learnt from the source domain to a relevant target domain.","id":"bf719ec3-fb89-4a44-9c42-480d45208596","level":"subsection","origin_cites_number":1,"parent_id":"bcb206fa-772b-4e9f-9da9-3ca19215514a","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Research Challenges and Future Directions"],["subsection","The Cost of Deep Learning"]],"subsections":[],"title":"The Cost of Deep Learning"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclusions}\nThis article presents a comprehensive study for the application of deep learning methods in the area of content caching. Particularly, the data caching is divided into two classifications according to the caching location of edge network. Each category contains where to cache, what to cache, cache dimensioning and content delivery. Then we introduce typical DNN methods which are categorised via training process into supervised learning, unsupervised learning and RL. Further, this paper critically compares and analyzes state-of-the-art papers on parameters, such as DL methods employed, the caching problems solved and the objective of applying DL. The challenges and research directions of DL on caching is also examined on the topic of extending caching to VNF chains, the application of caching for MAR as well as DTs, DL for cache size allocation and the additional cost of employing DL. Undoubtedly, DL is playing a significant role in 5G and beyond. We hope this paper will increase discussions and interests on DL for caching policy design and relevant applications, which will advance future network communications. \n\\vspace{6pt} \n\\bibliographystyle{unsrt}\n\\bibliography{reference}\n\\end{document}","id":"9f55687b-d2fc-4fa2-a01d-e44198b97b4c","level":"section","origin_cites_number":0,"parent_id":"347fe6bc-3506-4394-85f3-39521077022f","prefix_titles":[["title","A Survey of Deep Learning for Data Caching in Edge Network"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"347fe6bc-3506-4394-85f3-39521077022f\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ]\n ],\n \"subsections\": [\n \"c1e3d948-68a4-41aa-9453-ed02e0504669\",\n \"77d13750-5344-4e36-b9ce-8573ab07d97a\",\n \"c0dca0bc-798e-46e6-903c-5acb8a815d31\",\n \"3f275de1-b1c2-48e0-ae07-49947668a760\",\n \"bcb206fa-772b-4e9f-9da9-3ca19215514a\",\n \"9f55687b-d2fc-4fa2-a01d-e44198b97b4c\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.777777777777777,\n \"cites\": [\n 3705,\n 3582,\n 3704,\n 3706,\n 3420,\n 7633,\n 3703\n ],\n \"content\": \"\\\\label{sec:introduction}\\nUndoubtedly, future 5G and beyond mobile communication networks will have to address stringent requirements of delivering popular content at ultra high speeds and low latency due to the proliferation of advanced mobile devices and data rich applications. In that ecosystem, edge-caching has received significant research attention over the last decade as an efficient technique to reduce delivery latency and network congestion especially during peak-traffic times or during unexpected network congestion episodes by bringing popular data closer to the end users. One of the main reasons of enabling edge caching in the network is to reduce the number of requests that traverse the access and core mobile network as well as reducing the load at the origin servers that would have to, otherwise, respond to all requests directly in absence of edge caching. In that case popular content and objects can be stored and served from edge locations, which are closer to the end users. This operation is also beneficial from the end user perspective since edge caching can dramatically reduce the overall latency to access the content and increase in the sense overall user experience. It is also important to note that the notion of popular content means that the requests of top 10\\\\% of video content on the Internet account for almost 80\\\\% of all traffic; which relates to multiple requests from different end users of the same content\\n.\\nRecently, deep learning (DL) has attracted significant attention from both academia and industry and has been applied to diverse domains like self-driving, medical diagnosis, playing complex games such as Go . DL has also made their way into communication areas . In this paper, we pay attention to the application of DL in caching policy. Though there are some earlier surveys related to machine learning applications, they either focus on general machine learning techniques for caching , or concentrate on overall wireless applications . The work provides a big picture of applying machine learning in wireless communications. In , the authors consider the machine learning on both caching and routing strategy. A comprehensive survey on machine learning applications for caching content in edge networks is provided in . The researchers provide a survey about machine learning on mobile edge caching and communication resources. On the other hand, overviews how artificial neural networks can be employed for various wireless network problems. The authors in detail a survey on deep reinforcement learning (DRL) for issues in communications and networking. presents a comprehensive on deep learning applications and edge computing paradigm. Our work can be distinguished from the aforementioned papers based on the fact that we focus on the deep learning techniques on content caching and both wired and wireless caching are taken into account. Our main contributions are listed as follows:\\n\\\\begin{itemize}\\n\\t\\\\item We classify the content caching problem into Layer 1 Caching and Layer 2 Caching. Each layer caching consists of four tightly coupled subproblems: where to cache, what to cache, cache dimensioning and content delivery. Related researches are provided accordingly. \\n\\t\\\\item We present the fundamentals of DL techniques which are widely used in content caching, such as convolutional neural network, recurrent neural network, actor-critic model based deep reinforcement learning, etc.\\n\\t\\\\item We analyze a broad range of state-of-the-art literature which use DL to content caching. These papers are compared based on the DL structure, layer caching coupled subproblems and the objective of DL in each scenarios. Then we discuss research challenges and potential directions for the utilization of DL in caching. \\n\\\\end{itemize}\\n\\\\begin{figure}[htb]\\n\\t\\\\centering\\n\\t\\\\includegraphics[trim=0mm 0mm 0mm 0mm, clip, width=\\\\textwidth]{figure/paper.eps}\\n\\t\\\\caption{Survey Architecture}\\n\\t\\\\label{fig:organization}\\n\\\\end{figure}\\nThe rest of this survey is organized as follows (as illustrated in Figure \\\\ref{fig:organization} ). Section \\\\ref{sec:caching} presents the categories of content caching problem. Section \\\\ref{sec:dl} reviews typical deep neural network structures. In Section \\\\ref{sec:dl_caching}, we list state-of-the-art DL-based caching strategies and their comparison. Section \\\\ref{sec:challenges} debates challenges as well as potential research directions. In the end, Section \\\\ref{sec:conclusions} concludes this paper. For better readability, the abbreviations in this paper is listed as Table \\\\ref{tab:abbr} shows.\\n\\\\begin{table}[htb]\\n\\t\\\\centering\\n\\t\\\\caption{\\\\label{tab:abbr} List of Abbreviations.}\\n\\t\\\\begin{tabular}{l|p{.38\\\\textwidth}|l|p{.38\\\\textwidth}}\\n\\t\\t\\\\hline\\n\\t\\t\\\\textbf{Abbr.} & \\\\textbf{Description} & \\\\textbf{Abbr.} & \\\\textbf{Description} \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\t3C & Computing, Caching and Communication & A3C & Asynchronous Advantage Actor-Critic \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tBBU & Baseband Unit & CCN & Content-Centric Network \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tCNN & Convolutional Neural Network & CoMP-JT & Coordinated Multi Point Joint Transmission \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tCR & Content Router & C-RAN & Cloud-Radio Access Network \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tCSI & Channel State Information & D2D & Device to Device \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tDDPG & Deep Deterministic Policy Gradient & DL & Deep Learning \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tDNN & Deep Neural Network & DQN & Deep Q Network \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tDRL & Deep Reinforcement Learning & DT & Digital Twin \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tED & End Device & ES & Edge Server \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tETSI & \\\\multicolumn{3}{l}{European Telecommunication Standardization Institute} \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tESN & Echo-State Network & FIFO & First In First Out\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tFNN & Feedforward Neural Network & FBS & Femto Base Station \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tICN & Information-Centric Network & LFU & Least Frequently Used \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tLP & Linear Programming & LRU & Least Recently Used \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tLSTM & Long Short-Term Memory & MAR & Mobile Augmented Reality \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tMD & Mobile Device & MILP & Mixed Integer Linear Programming \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tMBS & Macro Base Station & NFV & Network Function Virtualization\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tPNF & Physical Network Function & PPO & Proximal Policy Optimization \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tQoE & Quality of Experience & RL & Reinforcement Learning \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tRNN & Recurrent Neural Network & RRH & Remote Radio Head \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tSAE & Sparse Auto Encoder & SDN & Software Defined Network \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tseq2seq & Sequence to Sequence & SNM & Shot Noise Model \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tTRPO & Trust Region Policy Optimization & TTL & Time to Live \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tVNF & Virtual Network Function & WSN & Wireless Sensor Network \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"c1e3d948-68a4-41aa-9453-ed02e0504669\",\n \"level\": \"section\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"347fe6bc-3506-4394-85f3-39521077022f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:caching}\\nThe paradigm of data caching in edge networks is illustrated in Figure \\\\ref{fig:caching}. Similarity to , the scope of edge in this paper is along the path between end user and data server, which contains Content Router (CR), Macro Base Station (MSB), Femto Base Station (FSB) and End Device (ED). In the context of Cloud-Radio Access Network (C-RAN), both baseband unit (BBU) and remote radio head (RRH) are considered as potential caching candidates to hosting content, where the BBUs are clustered as a BBU pool centrally and RRHs are deployed near BS's antenna distributively. According to the hierarchical structure of edge network, the data caching is classified into two categories: Layer 1 Caching and Layer 2 Caching. In this section we illustrate the typical research topics in these two areas.\\n\\\\begin{figure}[htb]\\n \\\\centering\\n \\\\includegraphics[trim=1mm 1mm 5mm 1mm, clip, width=\\\\textwidth]{figure/Cache.eps}\\n \\\\caption{Data Caching in Edge Network}\\n \\\\label{fig:caching}\\n\\\\end{figure}\",\n \"id\": \"77d13750-5344-4e36-b9ce-8573ab07d97a\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"347fe6bc-3506-4394-85f3-39521077022f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Data Caching Review\"\n ]\n ],\n \"subsections\": [\n \"93519284-cc11-4b04-b82f-f74d3c6a42fc\",\n \"a2d45d41-6ccf-4b15-99cb-2ea399aa1ef2\"\n ],\n \"title\": \"Data Caching Review\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 3709,\n 3707,\n 3710,\n 3708\n ],\n \"content\": \"In Layer 1 Caching, the popular content is considered to be hosted in CRs. In the context of Information-Centric Network (ICN), CR plays dual roles both as a typical router (i.e. data flow forwarding) and content store (i.e. local area data caching facility). Generally, the CR is connected via wired networks. Layer 1 Caching consists of four tightly coupled problems: where to cache, what to cache, cache dimensioning and content delivery.\\nWhere to cache focuses on selecting the proper CRs to host the content. For instance, in Figure \\\\ref{fig:caching}, contents replicas can be placed in lower hierarchical level CRs, such as router B and C, as the mean of reducing transmission cost but with extra cost pay for hosting contents; reversely, consolidating caching in CR A can be adopted to saving caching cost at the expense of more transmission cost and has the risk of expiring end users' delay requirement. Here the caching/hosting cost is the cost to deploy the content, which could be measured by space utilization, energy consumption or other metrics. The transmission cost represents the price for delivering the content from cached CR (or data server) to end user and is basically estimated via the number of hops. Where to cache problem usually has been modelled as a Mixed Integer Linear Programming (MILP):\\n\\\\begin{subequations} \\n\\\\label{fml:MILP}\\n\\\\begin{align}\\n\\\\label{MILP:obj}\\n\\\\mathop{\\\\min}_{\\\\substack{x}}\\\\; & c^T x\\\\\\\\\\n\\\\textrm{s.t.}\\\\quad \\\\label{MILP:con1}\\n& Ax\\\\leq b \\\\\\\\\\n\\\\label{MILP:con2}\\n& x\\\\in\\\\{0,1\\\\} \\\\\\\\\\n\\\\label{MILP:con3}\\n\\\\text{or}\\\\quad & x\\\\geq 0\\n\\\\end{align}\\n\\\\end{subequations}\\nwhere $x$ is the decision variable. Normally it is a binary variable indicating the CR assignment. In special cases, with the aim of modelling or linearization, some non-binary auxiliary variables are introduced as constraint \\\\eqref{MILP:con3} shows. If taking caching a part of a file not the complete into consideration, the decision variable $x$ is a continuous variable representing the segments host in the CR, then constraint \\\\eqref{MILP:con2} becomes $x\\\\in[0,1]$ and MILP model turns to linear programming (LP). There are many work allocating contents via MILP with different objectives and limitations. The authors in propose a model to minimize the user delay and load balancing level of CRs with the satisfaction of cache space. The work in considers a trade-off between caching and transmission cost with cache space, link bandwidth and user latency constraints. In , an energy efficient optimization model is constructed consisting of caching energy and transport energy. provides more details of mathematical model and related heuristic algorithms in caching deployment of wired networks. \\nWhat to cache concentrates on selecting the proper contents in CRs for the purpose of maximizing the cache hit ratio. Via exploiting the statistical patterns of user requests, the popularity of requested information and user preference can be forecasted and play a very significant role in determining caching content. On the one hand, from the view of aggregated request contents, researchers propose many different models and algorithms for popularity estimation. One widely used model in web caching is the Zipf model based the assumption that the content popularity is static and each users' request is independent. However, this method fails to reflect the temporal and spatial correlations of the content, where the temporal correlation reflects the popularity varies over time and the spatial correlation represents the content preference is different on the geographical area and social cultural media. A temporal model named the shot noise model (SNM) is built in which enables users to estimate the content popularity dynamically. Inspired by SNM, the work in considers both spatial and temporal characteristics during caching decision. On the other hand, from the view of a specific end user during a certain period, caching his/her preference content (may not be the popular in network) can also help to reduce the traffic flow. Many approaches in recommendation systems can be applied in this case. Another aspect of what to cache problem is the designing of cache eviction strategies when storage space faces the risk of overflow. Depending on the life of caching contents, these policies can be divided into two categories roughly: one is like first in first out (FIFO), least frequently used (LFU), least recently used (LRU) and randomized replacement, the contents would not be removed until no more memory is available; the other one is called time to live (TTL) strategy, where the eviction happens once the related timer expires. presents analytic model for hit ratio in TTL-based cache requested by independent and identically distributed flows. It worth noting that in , the TTL-based cache policy is used for the consistency of dynamic contents instead of contents replacement. In , the authors introduce a TTL model for cache eviction and the timer is reset once related content cache hit happens.\\nCache dimensioning highlights how much storage space to be allocated. Benefit from the softwarization and virtualization technologies, the cache size in each CR or edge cloud can be managed in a more flexible and dynamical way, which makes the cache dimensioning decisions become an important feature in data caching. Technically, the cache hit ratio rises with the increasing of cache memory, and consequently eases the traffic congestion in the core network. However, excessive space allocation would waste the resource like energy to support the caching function. Hence there is a trade-off between cache size cost and network congestion. Economically, taking such scenario into consideration: a small content provider wants to rent service from a CDN provider such as Akamai or Huawei Cloud, and there is also a balance between investment saving and network performance. In , the proper cache size of individual CR in Content-Centric Network (CCN) is investigated via exploiting the network topology. In , the authors consider the effect of network traffic distribution and user behaviours when designing cache size.\\nContent delivery considers how to transform the caching content to the requested user. The delivery traffic embraces single cache file downloading and video content steaming and the metrics for these two scenarios vary. Regarding file downloading, the content cannot be consumed until the delivery is completed. Therefore the downloading time of the entire file is viewed as a metric to reflect the quality of experience (QoE). For video steaming, especially for those large video splitted into several chunks, the delay limitation only works on the first chunk. In that case, delivering the first chunk in time and keep the smooth transmission of the rest chunks are the key aims . Apart from those measuring metrics, another problem in content delivery is the routing policy. In CCN , one implementation of ICN architecture, employs a flooding-based name routing protocol to publish the request among cached CRs. On one hand, flooding strategy simplifies the designing complexity and reduce the maintaining cost particularly in an unstable scenario; on the other hand, it costly wastes bandwidth resources. In , the authors discuss the optimal radius in scoped flooding. The deliver route is often considered jointly with where to cache problem, in which the objective function \\\\eqref{MILP:obj} includes both deployment and routing cost.\",\n \"id\": \"93519284-cc11-4b04-b82f-f74d3c6a42fc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"77d13750-5344-4e36-b9ce-8573ab07d97a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Data Caching Review\"\n ],\n [\n \"subsection\",\n \"Layer 1 Caching\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Layer 1 Caching\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 3712,\n 3716,\n 8682,\n 3714,\n 3717,\n 3713,\n 8681,\n 3715,\n 3711\n ],\n \"content\": \"Contrast to Layer 1 caching in wired connection, Layer 2 caching considers implementing caching techniques in wireless network. Though both of them need solve where to cache, what to cache, cache dimensioning and content delivery problems, wireless caching is more challenging and some mature strategies in wired caching cannot be migrated directly to wireless case. Some reasons come from the listed aspects: the resources in wireless environment, such as caching storage and spectrum, are limited compared with CRs in Layer 1 Caching; the mobility of end users and dynamic network typologies are also required to be considered during the design of caching strategies; moreover, the wireless channels are uncertain since they can be effected by fading and interference.\\nIn wireless caching, where to cache focus on finding the proper candidates among MBS, FBS, ED, even BBU pool and RRU in C-RAN to host the content. Caching at MBS and FBS can alleviate backhaul congestion since end users obtain the requested content from BS directly instead of from CR via backhaul links. Compared with FBS, MBS has wider coverage and typically, there is no overlap among different MBSs . As mentioned above, the caching space in BSs is limited and it is impractical to cache all popular content. With the aim of improving cache-hit ratio, a MILP-modelled collaborative caching strategy among MBSs is proposed in . If the accessed MBS does not host the content, the request will be served by a neighbour MBS which cache the file rather than by the data server. For FBS caching, a distributed caching method is presented in and the main idea is that the ED locating in the FBS coverage overlap is able to obtain contents from multiple hosters. Caching at ED can not only ease backhaul congestion but also improve the area spectral efficiency . When the end user requests a content, he/she would be severed by the local storage if the content is precached in his/her ED or by adjacent ED via D2D communication if the content is host accordingly. In , the authors model the cache-enabled D2D network as a Poisson cluster process, where end users are grouped into several clusters and the collective performance is improved. Individually, caching the interested contents for other users affects personal benefit. In , a Stackelberg game model is applied to formulate the conflict among end users and a related incentive mechanism is designed to encourage content sharing. For the case of cache-enabled C-RAN, caching at BBU can ease the traffic congestion in the backhaul while caching at RRH can reduce the fronthaul communication cost. On the other hand, caching all at BBU raises the signaling overhead of BBU pool while at RRH weakens the processing capability. Therefore, where to cache the content in C-RAN makes a substantial contribution to balancing the signal processing capability at the BBU pool and the backhaul/fronthaul costs . The work in investigates caching at RRHs with jointly considering cell outage probability and fronthaul utilization. Due to the end users' mobility, the prediction/awareness of user moving behaviour also influence the proper hoster selection. There are some researches exploiting user mobility in cache strategy designing like and .\\nSimilar with Layer 1, what to cache decision as well as eviction policy of layer 2 depends on the accurate prediction on content popularity or user preference in proactive caching method. The content popularity contains the feature of temporal and spatial correlations, which has already been described in Layer 1 Caching. In Layer 2 caching, the proper spatial granularity in popular contents estimation needs to take special attentions . For example, the coverage of MBS and FBS are different, which makes the popularity in MBS and FBS are different as well. Because the former based on a large number of users' behaviors but the individual may prefer specific content categories. For small cells, the preference estimation requires more accurate information like historical data . In order to capture the temporal and spatial dynamics of user preference, many different deep learning based algorithms are proposed, which will be illustrated in Section \\\\ref{sec:dl_caching}.\\nCache dimensioning in Layer 2 Caching has more complicated factors need to be considered, not only including the network topology and content popularity as Layer 1 Caching, but also containing backhaul transmission status and wireless channel features. The proper cache size assignment is studied in the scenario of backhaul limited cellular network . It also provides the closed-form boundary of minimum cache size in one cell case. In the case of dense wireless network, the work in quantifies the minimum required cache to achieve the linear capacity scaling of network throughput. The authors of also consider the scenario of dense networks. They derive the closed-form of the optimal memory size which can reduce the consumption of backhaul capacity as well as guarantee wireless QoS.\\nAccording to the number of transmitters and receivers, we divide the content delivery in Layer 2 caching into three categories: one candidate serves one end user, such as unicast and D2D transmission; one candidate serves multiple users like multicast; and coordinated delivery including multiple transmitters serve one or more receivers like coordinated multi-point joint transmission (CoMP-JT). Once the requested content is cached locally, BS can serve the end user via unicast or the adjacent device shares the contents by implementing D2D transmission. Concurrent transmission has the risk of co-channel interference in dense deployed networks. In D2D network, link scheduling is introduced to select subsets of links to transmit simultaneously . With the aim of improving the spectral efficiency, multicast is applied in content delivery when serving multiple requests simultaneously with the same content. Therefore there is a trade off between spectral efficiency and service delay. For the aim of serving more users in one transmission as well as higher spectral efficiency, the BS will wait to collect enough requirement for the same content which makes the first request a long waiting time. An optimal dynamic multicast scheduling is proposed in to balance these two factors. Multicast can also serve multiple requests with different contents. In , the authors provides a coded caching scheme which requires the communication link is error free and each user caches a part of its own content and partial of other users. Then BS multicasts the coded data to all users. Each user can decode his own requested content by XOR operation between the received data and the precached other users' file. However, the coding complexity increases exponentially as the quantity of end users grows. The CoMP-JT can improve the spectral efficiency as well via sharing channel state information (CSI) and contents among BSs but it also needs high-capacity backhaul consumption for exchanging data. In C-RAN, the BBUs are centralized in the BBU pool, which makes communication among BSs very efficiency. designs CoMP-JT in C-RAN for the purpose of minimizing power consumption with limitations of transmission energy, link capacity and requested QoS.\",\n \"id\": \"a2d45d41-6ccf-4b15-99cb-2ea399aa1ef2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"77d13750-5344-4e36-b9ce-8573ab07d97a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Data Caching Review\"\n ],\n [\n \"subsection\",\n \"Layer 2 Caching\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Layer 2 Caching\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:dl}\\nAs Figure \\\\ref{fig:DNN} shows, some typical deep neural network (DNN) methods are stated. These models are classified into three categories depending on the training methods: supervised learning, unsupervised learning and reinforcement learning. \\n\\\\begin{figure}[htb]\\n \\\\centering\\n \\\\includegraphics[trim=0mm 0mm 0mm 0mm, clip, width=\\\\textwidth]{figure/DNN.eps}\\n \\\\caption{Typical DNN Structures}\\n \\\\label{fig:DNN}\\n\\\\end{figure}\",\n \"id\": \"c0dca0bc-798e-46e6-903c-5acb8a815d31\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"347fe6bc-3506-4394-85f3-39521077022f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ]\n ],\n \"subsections\": [\n \"17a86b07-b7c3-41f4-9b7f-05a0e6ea6ae0\",\n \"f6c49080-7dcd-4236-b8c7-a294adda5983\",\n \"12cbfae2-b31a-489d-8357-cb9dedb5c4fa\",\n \"66cec1fd-1f07-4527-b02b-b98ccb6507c3\",\n \"cbd77acf-a82e-43b7-9bdd-bef9313a0f73\"\n ],\n \"title\": \"Deep Learning Outline\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 166\n ],\n \"content\": \"FNN is a kind of DNNs whose information propagation direction is forward and there is no cycle in neurons. In this paper, the term FNN is used to represent fully connected neural network, which indicates the connection between two adjacent layers is filled. According to the Universal Approximation Theorem, FNN has the ability to approximate any closed and bounded function with enough neurons in hidden layer . The hidden layer is applied to extract features of input vector, and then feed the output layer, which works as a classifier. Though FNN is very powerful, it gets into trouble when dealing with real-world task such as image recognition due to enormous weight parameters (because of fully connected) and lack of data augmentation.\",\n \"id\": \"17a86b07-b7c3-41f4-9b7f-05a0e6ea6ae0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"c0dca0bc-798e-46e6-903c-5acb8a815d31\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Feedforward Neural Network (FNN)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Feedforward Neural Network (FNN)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 810\n ],\n \"content\": \"For the aim of overcoming the aforementioned drawback of FNN, CNN employs convolution and pooling operations, where the former applies sliding convolutional filters to the input vector and the later does down sampling, usually via maximum or mean pooling. Generally, CNN tends to contain deeper layers and smaller convolutional filters, and the structure becomes fully convolutional network , reducing the ratio of pooling layers as well as fully connected layers. Taxonomically, CNN belongs to FNN and has been broadly employed in image recognition, video analysis, natural language processing, etc. Including CNN, one of the limitations of FNN is that the output only depends on current input vectors. So it is hard to deal with sequential tasks.\",\n \"id\": \"f6c49080-7dcd-4236-b8c7-a294adda5983\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"c0dca0bc-798e-46e6-903c-5acb8a815d31\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Convolutional Neural Network (CNN)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Convolutional Neural Network (CNN)\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 166\n ],\n \"content\": \"In order to deal with sequential tasks and using historical information, RNN employs neurons with self feedback in hidden layers. Unlike the hidden neuron in FNN, the output of recurrent neuron depends on both current output of precious layer and last hidden state. Compared with FNN approximates any continues functions, RNN with Sigmoid activation function can simulate a universal Turing Machine and has the ability to solve all computational problems . It is worth noting that RNN has the risk to suffer from long-term dependencies problem including gradient exploding and vanishing. Additionally, RNN has more parameters waiting to be trained due to adding recurrent weights. In the following, we introduce some RNN variants as Figure \\\\ref{fig:RNN} shows.\\n\\\\begin{figure}[htb]\\n \\\\centering\\n \\\\includegraphics[trim=0mm 0mm 35mm 0mm, clip, width=\\\\textwidth]{figure/RNN.eps}\\n \\\\caption{RNN Variants}\\n \\\\label{fig:RNN}\\n\\\\end{figure}\",\n \"id\": \"12cbfae2-b31a-489d-8357-cb9dedb5c4fa\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"c0dca0bc-798e-46e6-903c-5acb8a815d31\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Recurrent Neural Network (RNN)\"\n ]\n ],\n \"subsections\": [\n \"15615e34-4c33-4644-b6a2-80e552ca205d\",\n \"f028c925-baae-474d-ae9c-0ae54c3029f3\",\n \"bf6576fa-f9cd-4916-b753-6a9371a05fee\"\n ],\n \"title\": \"Recurrent Neural Network (RNN)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 166\n ],\n \"content\": \"As aforementioned, simple RNN contains more parameters in training step, where the recurrent weights and input weights are difficult to learn . The basic idea of ESN is fixing these two kinds of weights and only learn the output weights (as links highlighted in Figure \\\\ref{fig:RNN}). The hidden layer is renamed as reservoir in ESN, where the neurons are sparsely connected and the weights are randomly assigned. The recurrent weights keep constant so the information of previous moments is stored in the reservoir with constant weight like voice echoing.\",\n \"id\": \"15615e34-4c33-4644-b6a2-80e552ca205d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"12cbfae2-b31a-489d-8357-cb9dedb5c4fa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Recurrent Neural Network (RNN)\"\n ],\n [\n \"subsubsection\",\n \"Echo-State Network (ESN)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Echo-State Network (ESN)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 166\n ],\n \"content\": \"Recently, an efficient way to cope with long-term dependencies in practical is employing gated RNN, including LSTM . Then we compare with the recurrent neuron in simple RNN: Internally LSTM introduces three gates to control signal propagation, where input gate $I$ decides the partition of input signal to be stored, forget gate $F$ controls ratio of last moment memory to be kept until next period (the name \\\"forget gate\\\" may be a little misleading because it actually represents the ratio to be remembered) and output gate $O$ influences the proportion of current state to be delivered; Externally LSTM has four inputs embracing one input signal and three control signals for three gates. All these four signals are derived via the calculation of current network input and last moment delivered state.\",\n \"id\": \"f028c925-baae-474d-ae9c-0ae54c3029f3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"12cbfae2-b31a-489d-8357-cb9dedb5c4fa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Recurrent Neural Network (RNN)\"\n ],\n [\n \"subsubsection\",\n \"Long Short-Term Memory (LSTM)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Long Short-Term Memory (LSTM)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 167\n ],\n \"content\": \"A typical application of RNN is converting one sequence to another sequence (seq2seq) such as machine translation. Conventionally, the output of seq2seq architecture is a probability distribution of output dictionary. However, it cannot deal with the problem that the size of output relies on the length of input due to fixed output dictionary. In , the authors modify the output to be the distribution of input sequence, which is analogous to pointers in C/C++. Pointer network has been widely used in text condensation.\",\n \"id\": \"bf6576fa-f9cd-4916-b753-6a9371a05fee\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"12cbfae2-b31a-489d-8357-cb9dedb5c4fa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Recurrent Neural Network (RNN)\"\n ],\n [\n \"subsubsection\",\n \"Pointer Network\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pointer Network\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 166\n ],\n \"content\": \"Auto Encoder is a stack of two NNs named encoder and decoder respectively, where the former tries to learn the representative characteristics of input and generate a related code, and the later reads the code and reconstructs the original input. In order to avoid the auto encoder simply copying the input, some restrictions are considered like the dimension of code is smaller than input vector . The quality of auto encoder can be measured via reconstruction error, which estimates the similarity between input and output. In most cases, the auto encoder is used for the proper representation of input vector so the decoder part is removed after unsupervised training. The code can be employed as input for further deep learning models.\",\n \"id\": \"66cec1fd-1f07-4527-b02b-b98ccb6507c3\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"c0dca0bc-798e-46e6-903c-5acb8a815d31\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Auto Encoder\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Auto Encoder\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Reinforcement Learning (RL) is a Markov Decision Process represented by a quintuple $\\\\{\\\\mathcal{S,A,P,R},\\\\gamma\\\\}$, where $\\\\mathcal{S}$ is the state space controlled by environment; $\\\\mathcal{A}$ is the action space determined by agent; $\\\\mathcal{P}$ is the state transition function measuring the probability of moving to a new state $s_{t+1}$ given previous state $s_t$ and action $a_t$; $R$ is reward function calculated by environment considering state and action; $\\\\gamma$ is a discount factor for estimating total reward. During the interaction between agent and environment, agent observes current state $s_t$ from environment, and then takes action $a_t$ following its policy $\\\\pi$. The environment moves to a new state $s_{t+1}$ stochastically based on $\\\\mathcal{P}(s_t,a_t)$ and returns a reward $r_t$ to agent. The RL's aim is finding the policy $\\\\pi$ to maximum accumulated reward $\\\\sum_t\\\\gamma^tr_t$. In the early stage, RL focuses on scenarios whose $\\\\mathcal{S}$ and $\\\\mathcal{A}$ are discrete and limited. So the agent can use a table to record these information. Recently, some tasks have enormous discrete states and actions such as playing go and even continuous value such as self-driving, which makes table recording impractical. In order to solve this, DRL combines RL and DL, where RL defines the problem and optimization object; DL models the policy and the reward expectation. Depending on the roles of DNN in DRL, we classify the DRL into 3 categories as Figure shows.\",\n \"id\": \"cbd77acf-a82e-43b7-9bdd-bef9313a0f73\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"c0dca0bc-798e-46e6-903c-5acb8a815d31\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Deep Reinforcement Learning (DRL)\"\n ]\n ],\n \"subsections\": [\n \"d2c418b7-118c-42b8-b9d8-37f74087f743\",\n \"4e5ada48-bb01-4d53-88dc-d7516f8cb54c\",\n \"92185b95-3c92-401c-8a45-161c99f6f228\"\n ],\n \"title\": \"Deep Reinforcement Learning (DRL)\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 1409,\n 1408\n ],\n \"content\": \"In value-based method, DNN does not get involved with policy decision but estimates the policy performance. Two functions are introduced for the measurement: $V^\\\\pi(s)$ represents the reward expectation of policy $\\\\pi$ starting from state $s$; $Q^\\\\pi(s,a)$ illustrates the reward expectation of policy $\\\\pi$ starting from state $s$ and taking action $a$. In addition, $V^\\\\pi(s)$ is the expected value of $Q^\\\\pi(s,a)$. If we can estimate $Q^\\\\pi(s,a)$, the policy $\\\\pi$ can also be improved by choosing the action $a^*$ hold $Q^\\\\pi(s,a^*)\\\\geq V^\\\\pi(s)$. So the DNN employed in agent is approximating function $Q^\\\\pi(s,a)$, where the inputs are state $s$ and action $a$ and output is the estimated value $Q^\\\\pi(s,a)$. There are some representative critic methods like Deep Q Networks (DQN) and its variants Double DQN , Dueling DQN , etc.\",\n \"id\": \"d2c418b7-118c-42b8-b9d8-37f74087f743\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"cbd77acf-a82e-43b7-9bdd-bef9313a0f73\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Deep Reinforcement Learning (DRL)\"\n ],\n [\n \"subsubsection\",\n \"DNN as Critic (Value-Based)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"DNN as Critic (Value-Based)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 2219,\n 1391\n ],\n \"content\": \"In policy-based method, DNN gets involved in the action selection directly instead of via $Q^\\\\pi(s,a)$. The policy can be viewed as an optimization problem, where the objective function is maximizing reward expectation and the search space is policy space. The input of DNN is current state and output is the probability distribution of potential actions. By employing gradient ascent, we can update the DNN to provide better action then maximize total reward. Some popular algorithms include Trust Region Policy Optimization (TRPO) , Proximal Policy Optimization (PPO) .\",\n \"id\": \"4e5ada48-bb01-4d53-88dc-d7516f8cb54c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"cbd77acf-a82e-43b7-9bdd-bef9313a0f73\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Deep Reinforcement Learning (DRL)\"\n ],\n [\n \"subsubsection\",\n \"DNN as Actor (Policy-Based)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"DNN as Actor (Policy-Based)\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1390\n ],\n \"content\": \"Generally, compared with policy-based approach, the value-based method is less stable and suffer from poor convergence since the policy is derived based on $Q^\\\\pi(s,a)$ approximation. But value-based method is more sample efficient, while policy-based method is easier to fall into local optimal solution because the search space is vast. The actor-critic model combines these two approaches, i.e. the agent contains two DNNs named actor and critic respectively. In each training iteration, the actor considers current state $s$ and policy $\\\\pi$ for deciding action $a$. Then the environment changes to state $s'$ and returns reward $r$. The critic updates its own parameters based on the feedback from environment and output a mark for the actor's action. The actor updates the policy $\\\\pi$ depending on critic's mark. Some typical algorithms are proposed recent years like Deep Deterministic Policy Gradient (DDPG) and Asynchronous Advantage Actor-Critic (A3C) .\",\n \"id\": \"92185b95-3c92-401c-8a45-161c99f6f228\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"cbd77acf-a82e-43b7-9bdd-bef9313a0f73\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning Outline\"\n ],\n [\n \"subsection\",\n \"Deep Reinforcement Learning (DRL)\"\n ],\n [\n \"subsubsection\",\n \"Actor-Critic Model\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Actor-Critic Model\"\n },\n {\n \"cite_extract_rate\": 0.162790697674418,\n \"cites\": [\n 3721,\n 7746,\n 3718,\n 3719,\n 3723,\n 3722,\n 3720\n ],\n \"content\": \"\\\\label{sec:dl_caching}\\nWe divide the studies regarding deep learning for data caching in edge networks into four categories depending on the DL tools employed: FNN and CNN; RNN; Auto Encoder; DRL. Recently many works utilize more than one DL techniques for jointly considered caching problems. For instance, at the beginning we applies a RNN to predict content popularity, and then a DRL to find suboptimal solutions of content placement for the purpose of reducing time complexity. In such case, we classify the related work into DRL since it represents the caching allocation policy. Unless mention the caching location (such as CRs, MBSs, FBSs, EDs and BBUs) otherwise, the approaches in this section can be utilized for both Layer 1 and Layer 2 caching. Table \\\\ref{tab:summary} summarize some studies of DL for caching.\\n\\\\begin{comment}\\n\\\\begin{longtable}[c]{p{.12\\\\textwidth}|l|p{.2\\\\textwidth}|p{.23\\\\textwidth}|p{.26\\\\textwidth}}\\n \\\\caption{Summary of Deep Learning for Data Caching in Edge Network}\\n \\\\label{tab:summary1}\\\\\\\\\\n \\\\hline\\n \\\\textbf{DL method} & \\\\textbf{Study} & \\\\textbf{Caching Problem} & \\\\textbf{Caching Objective} & \\\\textbf{DL Objective} \\\\\\\\\\n \\\\endfirsthead\\n \\\\hline\\n \\\\textbf{DL method} & \\\\textbf{Study} & \\\\textbf{Caching Problem} & \\\\textbf{Caching Objective} & \\\\textbf{DL Objective} \\\\\\\\\\n \\\\endhead\\n \\\\hline\\n CNN & & L1 where to cache \\\\& Content Delivery & minimize caching \\\\& transmission cost & nominate proper CRs for caching \\\\\\\\\\n \\\\hline\\n CNN & & L1 where to cache \\\\& Content Delivery & minimize caching \\\\& transmission cost & reduce feasible region of CRs for caching \\\\\\\\\\n \\\\hline\\n CNN\\\\& FNN & & L2 content delivery & minimize transmission time/energy & reduce feasible region of time slot allocation for delivery \\\\\\\\\\n \\\\hline\\n FNN & & L2 where to cache \\\\& content delivery & minimize energy cost & determine proper MBSs for caching \\\\& time duration for delivery \\\\\\\\\\n \\\\hline\\n Auto-encoder & & L1 \\\\& L2 what to cache & improve cache hit ratio & predict content popularity \\\\\\\\\\n \\\\hline\\n ESN & & L2 what to cache \\\\& content delivery & minimize transmit power & predict content popularity \\\\& user mobility \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache \\\\& content delivery & minimize service latency & decide content caching, computing offloading and radio resource allocation \\\\\\\\\\n \\\\hline\\n LSTM & & L2 what to cache & improve cache hit ratio & predict content popularity \\\\\\\\\\n \\\\hline\\n FNN & & L2 what to cache & save traffic data & predict visiting content \\\\\\\\\\n \\\\hline\\n RNN \\\\& DRL & & L2 what to cache \\\\& where to cache & minimize service latency & predict popularity \\\\& decide content caching and offloading \\\\\\\\\\n \\\\hline\\n FNN & & L1 what to cache & improve cache hit ratio & predict content popularity \\\\\\\\\\n \\\\hline\\n Auto-encoder \\\\& RNN & & L2 content delivery & minimize transmission delay & reduce traffic load \\\\& select optimal BS subset\\\\\\\\\\n \\\\hline\\n CNN & & L2 what to cache & minimize service latency & predict visiting content \\\\\\\\\\n \\\\hline\\n ESN \\\\& LSTM \\\\& DRL & & L2 where to cache \\\\& what to cache \\\\& content delivery & minimize delay \\\\& power consumption & predict user mobility, content popularity \\\\& determine D2D link \\\\\\\\\\n \\\\hline\\n DRL & & L2 where to cache \\\\& content delivery & minimize content access cost & decide content caching \\\\& bandwidth allocation \\\\\\\\\\n \\\\hline \\n DRL & & L2 where to cache \\\\& what to cache & maximize cache hit ratio \\\\& minimize transmission delay & decide cache replacement \\\\\\\\\\n \\\\hline\\n auto-encoder \\\\& DRL & & L2 content delivery & minimize average delay and power consumption & decide multicast scheduling and caching jointly\\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & maximize traffic load served by FBSs & decide which and how many contents be cached\\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & minimize transmission power & predict users' preference and content popularity \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & maximize network operator's utility & decide content replacement \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & minimize service latency & decide cache replacement \\\\& power allocation \\\\\\\\\\n \\\\hline\\n RNN & & L2 what to cache & improve cache hit ratio & predict content popularity \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & maximize overall content quality & decide cache replacement \\\\\\\\\\n \\\\hline \\n DRL & & L2 what to cache & maximize cache hit ratio & decide caching strategy \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & maximize energy efficient & decide which content to be cached \\\\\\\\\\n \\\\hline\\n CNN \\\\& RNN \\\\& DRL & & L2 what to cache & maximize cost reduction of operators & predict popularity \\\\& searching best NN model \\\\\\\\\\n \\\\hline\\n LSTM & & L2 what to cache & maximize cache hit ratio & decide content replacement \\\\\\\\\\n \\\\hline\\n DRL & & L2 where to cache & maximize QoE & decide content placement \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & improve cache hit ratio & decide content update \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & minimize transmitting cost & decide cache replacement \\\\\\\\\\n \\\\hline\\n auto-encoder & & L2 what to cache & improve backhaul offloading \\\\& user satisfaction & predict content popularity \\\\\\\\\\n \\\\hline\\n DRL & & L2 what to cache & minimize cumulative cost & determine caching content \\\\\\\\\\n \\\\hline \\n DRL & & L2 what to cache & maximize chunk hit ratio \\\\& minimize service time & determine caching chunks \\\\\\\\\\n \\\\hline\\n\\\\end{longtable}\\n\\\\end{comment}\\n\\\\begin{table}[htb]\\n\\\\caption{Summary of Deep Learning for Data Caching}\\n\\\\label{tab:summary}\\n\\\\begin{tabular}{p{.07\\\\textwidth}|p{.1\\\\textwidth}|p{.28\\\\textwidth}|p{.43\\\\textwidth}}\\n \\\\hline\\n \\\\textbf{method} & \\\\textbf{Study} & \\\\textbf{Caching Problem} & \\\\textbf{DL Objective}\\\\\\\\\\n \\\\hline\\n \\\\multirow{8}{.07\\\\textwidth}{FNN and CNN} & & content delivery & reduce feasible region of time slot allocation\\\\\\\\\\n \\\\cline{2-4}\\n & & where to cache, content delivery & determine MBSs for caching \\\\& delivery duration \\\\\\\\\\n \\\\cline{2-4}\\n & & where to cache, content delivery & nominate proper CRs for caching \\\\\\\\\\n \\\\cline{2-4}\\n & & where to cache, content delivery & reduce feasible region for caching \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & extract video features\\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & predict requested content \\\\& frequency \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & predict requested content \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & predict content popularity\\\\\\\\\\n \\\\hline\\n \\\\multirow{4}{.07\\\\textwidth}{RNN} & & what to cache & predict requested content \\\\& user mobility \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & predict content popularity \\\\\\\\\\n \\\\cline{2-4}\\n & & content delivery & reduce traffic load, select optimal BS subset\\\\\\\\\\n \\\\hline\\n \\\\multirow{3}{.07\\\\textwidth}{Auto Encoder} & & what to cache & predict content popularity \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & predict top popular contents \\\\\\\\\\n \\\\hline\\n \\\\multirow{17}{.07\\\\textwidth}{DRL} & & what to cache & decide cache placement \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & decide cache replacement \\\\& power allocation \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache & predict popularity \\\\& searching best NN model \\\\\\\\\\n \\\\cline{2-4}\\n & & where to cache & decide cache location \\\\\\\\\\n \\\\cline{2-4}\\n & & content delivery & users grouping\\\\\\\\\\n \\\\cline{2-4}\\n & & where to cache, content delivery & decide BS connection, computation offloading \\\\& caching location\\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache, content delivery & decide caching \\\\& bandwidth allocation \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache, content delivery & decide caching, computing offloading \\\\& radio resource allocation \\\\\\\\\\n \\\\cline{2-4}\\n & & what to cache, content delivery & decide multicast scheduling \\\\& caching replacement\\\\\\\\\\n \\\\cline{2-4}\\n & & where \\\\& what to cache & predict popularity, decide caching \\\\& task offloading \\\\\\\\\\n \\\\cline{2-4}\\n & & where \\\\& what to cache, content delivery & predict user mobility \\\\& content popularity, determine D2D link\\\\\\\\\\n \\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"3f275de1-b1c2-48e0-ae07-49947668a760\",\n \"level\": \"section\",\n \"origin_cites_number\": 43,\n \"parent_id\": \"347fe6bc-3506-4394-85f3-39521077022f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning for Data Caching\"\n ]\n ],\n \"subsections\": [\n \"64ba4f69-8c92-44ec-9c4d-6de443abc8ba\",\n \"8a40aabe-271f-491f-92cb-1da70950927c\",\n \"67ee9973-0fe6-40d1-b44e-d9d39d452c5f\",\n \"c63a6eb6-9774-4886-b68b-d0543cf7f42c\"\n ],\n \"title\": \"Deep Learning for Data Caching\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 3723,\n 3721\n ],\n \"content\": \"In , the content delivery problem in wireless network is formulated as two MILP optimization models with the aims of minimum delivery time slot and energy consumption respectively. Both models consider the data rate for content delivery. Considering the computational complexity of solving MILP, a CNN is introduced to reduce the feasible region of decision variables, where the input is channel coefficients matrix. The FNN in paper plays a similar role as to simplify the searching space of the content delivery optimization model. \\nFor resource allocation problem, the authors of model it as linear sum assignment problems then utilize CNN and FNN to solve the model. The idea is extended in and , where the authors consider where to cache problem among potential CRs and content delivery jointly, which is modeled as MILP with the aim of balancing caching and transmission cost by considering the user mobility, space utilization and bandwidth limitations. The cache allocation is viewed as multi-label classification problem and is decomposed into several independent sub-problems, where each one correlates with a CNN to predict assignment. The input of CNN is a grey-scale image which combines the information of user mobility, space and link utilization level. In , a hill climbing local search algorithm is provided to improve the performance of CNN while in , the prediction of CNN is used to feed a smaller MILP model. \\nFor these above works , the FNN or CNN input is extracted from the optimization model. The work in trains a CNN via original graph instead of parameters matrix/image, which makes the process human recognizable and interpretable. Though the authors take traveling salesman problem not data caching as an example, the method can be viewed as a potential research direction. \\nIn , an ILP model is proposed to minimize the backhaul video-data type load by determining the portion of cached content in BSs. Considering the fact that the mobile users covered by a BS change frequently, therefore predicting user preference is unnecessary. Instead, the authors concentrate on the popular content in general. At the beginning, a 3D CNN is introduced to extract spatio-temporal features of videos. The popularity of new contents without historical information is determined via comparing similar video features. The authors of also considers the spatio-temporal features among visiting contents in a mobile bus WiFI environment. By exploiting the previous 9 days collecting data, the content that the user may visit on the last day and corresponding visiting frequency can be forecast. The social property is taken into account in . By observing users interests on tweets during 2016 U.S. election, a CNN based predicted model can foresee the content category that is most likely to be requested. Such kind of content would be cached in MBSs and FBSs.\\nThe work of examines the role of DNN in caching from another aspect. The authors propose a FNN to predict content popularity as a regression problem. The results show that FNN outperforms RNN, though the later is believed to be effective to solve sequential predictions. Moreover, replacing the FNN by a linear estimator does not devalue the performance significantly. The author provides explanation that FNN would work better than linear predictor in the case of incomplete information, and RNN has more advantages to model the popularity prediction as a classification rather than a regression problem.\",\n \"id\": \"64ba4f69-8c92-44ec-9c4d-6de443abc8ba\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"3f275de1-b1c2-48e0-ae07-49947668a760\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning for Data Caching\"\n ],\n [\n \"subsection\",\n \"FNN and CNN\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"FNN and CNN\"\n },\n {\n \"cite_extract_rate\": 0.11111111111111101,\n \"cites\": [\n 3720\n ],\n \"content\": \"Considering RNN is superior in dealing with sequential tasks, the work applies a bidirectional RNN for online content popularity prediction in mobile edge network. Simple RNN's output depends on previous and current storage, but the bidirectional can also take future information into account. The forecast model consists three blocks cascadingly: a CNN reads user requests and extract features; bidirectional LTSM learns association of requests over time step; FNN is added in the end to improve the prediction performance. Then content eviction is based on the popularity prediction. \\nThe authors in utilize ESN to predict both content request distribution and end user mobility pattern. The user's preference is viewed as context which links with personal information combining gender, age, job, location, etc.. For the request prediction, the input of ESN is user's information vector and the output represent the probability distribution of content. For mobility prediction, the input includes historical and present user's location and the output is the expected position for next time duration. Eventually, the prediction influences the caching content decisions in BBUs and RRHs for the purpose of minimizing traffic load and delay in CRAN. The authors extend their work in by introducing conceptor-based ESN which can split users' context into different patterns and learn them independently. Therefore a more accurate prediction is achieved. \\nIn , a caching decision policy named PA-Cache is proposed to predict time-variant video popularity for cache eviction when the space is full. The temporal content popularity is exploited by attaching every hidden layer representation of RNN to an output regression. In order to improve the accuracy, hedge backpropagation is introduced during training process which decides when and how to adapt the depth of the DNN in an evolving manner. Similarly, the work in also considers caching replacement of video content. A deep LSTM network is utilized for popularity prediction consisting of stacking multiple LSTM layers and one softmax layer, where the input of the network is request sequence data (device, timestamp, location, title of video) without any prepossessing and the output is estimated content popularity. Another work concentrates on prediction and interactions between user mobility and content popularity can be found .\\nThe work of recognizes the popularity prediction as a seq2seq modeling problem and proposes LSTM Encoder-Decoder model. The input vector consists of past probabilities where each vectors are calculated during a predefined time window. In , the authors focus on caching content delivery with the aim of minimizing BSs to cover all requested users, i.e. set cover problem, via coded caching. Unlike , an auto encoder is introduced in coded caching stage for file conversion to reduce transmission load. In addition, a RNN model is employed to select BSs for broadcasting. \\nThe paper shows the potential of RNN in solving where to cache problem. In , a task allocation model is formulated as a knapsack problem and the decision variables represent the task is processed locally in mobile devices (MDs) or remotely in edge servers (ESs). The authors design a multi-pointer network structure of 3 RNNs, where 2 encoders encode MDs and ESs respectively, 1 decoder demonstrates ES and MD pairing. Considering the similarity of where to cache optimization model and knapsack problem, the multi-pointer network can be transferred for caching location decision after according parameter modifications.\",\n \"id\": \"8a40aabe-271f-491f-92cb-1da70950927c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"3f275de1-b1c2-48e0-ae07-49947668a760\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning for Data Caching\"\n ],\n [\n \"subsection\",\n \"RNN\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"RNN\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Generally, auto encoder is utilized to learn efficient representation or extract features of raw data in an unsupervised manner. The work in considers the cache replacement in wireless sensor network (WSN) based on content popularity. Considering sparse auto encoder (SAE) can extract representative expression of input data, the authors employ a SAE followed by a classifier where the input contains collecting user content requests and the output represents the contents popularity level. The authors also think about the implementation in a distributed way by SDN/NFV technical, i.e. the input layer is deployed on sink node, while the rest layers are implemented on the main controller. A related work applying auto encoder in 5G network proactive caching can be found in . In , two auto encoders are utilized for extracting the features of users and content respectively. Then the extracted information is explored to estimate popularity at the core network. Similarly, the auto encoder in is for spatio-temporal popularity features extraction and auto encoders work collaboratively in to predict top K popular videos.\",\n \"id\": \"67ee9973-0fe6-40d1-b44e-d9d39d452c5f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"3f275de1-b1c2-48e0-ae07-49947668a760\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning for Data Caching\"\n ],\n [\n \"subsection\",\n \"Auto Encoder\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Auto Encoder\"\n },\n {\n \"cite_extract_rate\": 0.23809523809523803,\n \"cites\": [\n 7746,\n 3718,\n 3719,\n 3724,\n 3722\n ],\n \"content\": \"The work in focuses on the cooperative caching policy at FBSs with maximum distance separable coding in ultra dense networks. A value-based model is utilized to determine caching categories and the content quantity at FBSs during off peak duration. The authors in study the problem of caching 360\\\\degree videos and virtual viewports in FBSs with unknown content popularity. The virtual viewport represents the most popular tiles of a 360\\\\degree video over users' population. A DQN is introduced to decide which tiles of a video to be hosted and in which quality. Additionally, employs DQN for content eviction decision offering a satisfactory quality of experience and is for the purpose of minimizing energy consumption. In , the authors also apply DQN to decide cache eviction in a single BS. Moreover, the critic is generated with stacking LSTM and FNN to evaluate Q value and an external memory is added for recording learned knowledge. For the purpose of improving the prediction accuracy, the Q value update is determined by the similarity of estimated value of critic and recording information in the external memory, instead of critic domination. The paper puts forth DQN a two-level network caching, where a parent node links with multiple leaf nodes to cache content instead of a single BS. In , a DRL framework with Wolpertinger architecture is presented for content caching at BSs. The Wolpertinger architecture is based on actor-critic model and performs efficiently in large discrete action space. employs two FNNs working as actor and critic respectively, where the former determines requested content is cached or not and the later estimates the reward. The whole framework consists two phases: in offline phase, these two FNNs are trained in supervised learning; in online phase, the critic and actor update via the interaction with environment. The authors extend their work to a multi agent actor-critic model for decentralized cooperative caching at multiple BSs . In , an actor-critic model is used for solving cache replacement problem, which balance the data freshness and communication cost. The aforementioned papers put attention on the network performance while ignore the influence of caching on information processing and resource consumption. Therefore, authors of design cache policy considering both network performance during content transmission and processing efficiency during data consumption. A DQN is employed to determine the number of chunks of the requested file to be updated. The paper investigates a joint cache replacement and power allocation optimization problem to minimize latency in a downlink F-RAN. A DQN is proposed for finding a suboptimal solution. Though is regarded as solving what to cache problem like , the reinforcement learning approach plays a different role. In , a DNN is utilized for content popularity prediction and then a RL is used for DNN hyperparameters tuning instead of determining caching content. Therefore the action space consists of choosing model architectures (i.e. CNN, LSTM, etc.), number of layers and layer configurations.\\nIn , the authors generate an optimization model with the aim of maximizing network operator's utility in mobile social networks under the framework of mobile edge computing, in network caching and D2D communications (3C). The trust value which if estimated through social relationships among users are also considered. Then a DQN model is utilized for solving optimization problem, including determine video provider and subscriber association, video transcoding offloading and the video cache allocation for video providers.. The DQN employs two CNNs for training process, where one generates target Q value and the other is for estimated Q value. Unlike the conventional DQN, the authors in introduces a dueling structure, i.e. the Q value is not computed in the final fully connected layer, but is decomposed into two components and use the summary as estimated Q value, which helps achieve a more robust result. The authors also consider utilizing dueling DQN model in different scenarios like cache-enabled opportunistic interference alignment and orchestrating 3C in vehicular network . The work provides a DDPG model to cope with continuous valued control decision for 3C in vehicular edge networks, which is combined with the idea of DQN and actor-critic model. The DDPG structure can be divided into two parts as DQN, one is for estimated Q value and the other for target Q value. Each part consists of two DNNs, which play the role of actor and critic respectively. The critic updates its parameters like DQN while the actor learns policy via deterministic policy gradient approach. The proposed DRL is used for deciding content caching/replacement, vehicle organization and bandwidth resource assignment on different duration. \\nThe paper provides an optimization model which takes what to cache and content delivery into consideration in the fog-enabled IoT network in order to minimize service latency. Since the wireless signals and user requests are stochastic, a actor-critic model is engaged where the actor makes decision for requesting contents while critic estimates the reward. Specially, the action space $S$ consists of decision variables and reward function is a variant of the objective function. A caching replacement strategy and dynamic multicast scheduling strategy are studied in . In order to get a suboptimal result, an auto encoder is used to approximate the state. Further, a weighted double DQN scheme is utilized for avoiding overestimation of Q value. applies a RNN to predict content popularity by collecting historical requests and the output represents the popularity in the near future. Then the prediction is employed for cooperative caching and computation offloading among MEC servers, which is modelled as a ILP problem. For the purpose of solving it efficiently, a multi-agent DQN is applied where each user is viewed as an agent. The action space consists of task local computing and offloading decision as well as local caching and cooperative caching determination. The reward is measured by accumulated latency. The agent choose its own action based on current state without cooperation. The where to cache, what to cache and content delivery decision of D2D network are jointly modelled in . Two RNNs, ESN and LSTM, are considered to predict mobile users' location and requested content popularity. Then the prediction result is used for determining content categories and cache locations. The content delivery is formulated as the actor-critic based DRL framework. The state spaces include CSI, transmission distances and communication power between requested user and other available candidates. The function of DRL is determining the communication link among users with the aim of minimizing power consumption and content delay.\\nWe notice that most papers prefer to use value-based model (critic) and value-policy-based (actor-critic) model in DRL framework, but rare paper considers only policy-based model to solve data caching problem. One proper reason is that the search space of caching problem is enormous so policy-based model is easier to fall into local optimal solution, resulting in poor performance. Though the value-based model is less stable, some variant structures are utilized like Double DQN in to avoid value overestimation and dueling DQN in to improve robust.\",\n \"id\": \"c63a6eb6-9774-4886-b68b-d0543cf7f42c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"3f275de1-b1c2-48e0-ae07-49947668a760\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Deep Learning for Data Caching\"\n ],\n [\n \"subsection\",\n \"DRL\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"DRL\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:challenges}\\nA serious of open issues on content caching and potential research directions are discussed in this section. We first extend the idea of content caching to virtual network function chain since caching can be viewed as a specific network function. Then we consider the caching for augmented reality applications. Moreover, we notice that the cache dimensioning has not been covered yet by DL methods. Finally, we debate the addition cost introduced by DL.\",\n \"id\": \"bcb206fa-772b-4e9f-9da9-3ca19215514a\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"347fe6bc-3506-4394-85f3-39521077022f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Research Challenges and Future Directions\"\n ]\n ],\n \"subsections\": [\n \"e584c09a-75d8-4201-90c6-a0fd658b4342\",\n \"f03f2b46-f5ed-47c7-b95e-abf7079d9dd7\",\n \"76a7e5ab-c389-46f2-914c-27d36e617810\",\n \"bf719ec3-fb89-4a44-9c42-480d45208596\"\n ],\n \"title\": \"Research Challenges and Future Directions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The concept of Network Function Virtualization (NFV) has been firstly discussed and proposed within the realms of the European Telecommunication Standardization Institute (ETSI)\\\\footnote{Network Functions Virtualisation, An Introduction, Benefits, Enablers, Challenges and Call for Action, ETSI, 2012 https://portal.etsi.org/NFV/NFV\\\\_White\\\\_Paper.pdf}. The rational is to facilitate the dynamic provisioning of network services through virtualization technologies to decouple the service creation process form the underlying hardware. The framework allows network services to be implemented by a specific chaining and ordering of a set of functions which can be implemented either on a more traditional dedicated hardware which in this this case are called Physical Network Functions (PNFs), or alternatively as Virtual Network Functions (VNFs) which is a software running on top of virtualized general-purpose hardware. The decoupling between the hardware and the software is one of the important considerations the other – equally important – is that a virtualized service lend itself naturally to a dynamic programmable service creation where VNF resources can be deployed as required. Hence, edge cloud and network resource usage can be adapted to the instantaneous user demand whilst avoiding a more static over-provisioned configurations.\\nWithin that framework, the incoming network service requests include the specification of the service function chain that need to be created in the form of an ordered sequence of VNFs. For example different type of VNFs such as a firewall or a NAT mechanism need to be visited in a specific order. In such constructed service chain each independent VNF requires specific underlying resources in terms for example of CPU cycles and/or memory. \\nUnder this framework, caching of popular content can be considered as a specialized VNF chain function since inevitably delivery of the cached popular content to users will require a set of other functions to be supported related to security, optimization of the content etc. However, the issue of data caching and VNF chaining have evolved rather independently in the literature and the issue on how to optimize data caching when seeing it as part of VNF chain is still an interesting open ended issue.\",\n \"id\": \"e584c09a-75d8-4201-90c6-a0fd658b4342\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"bcb206fa-772b-4e9f-9da9-3ca19215514a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Research Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Caching as a Virtual Network Function Chain\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Caching as a Virtual Network Function Chain\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Mobile augmented reality (MAR) applications can be considered as a way to augment the physical real-world environment\\nwith artificial computer-based generated information and is an area that has received significant research attention recently. In order to successfully superimpose different digital object in the physical world MAR applications include several computationally and storage complex concepts such as image recognition, mobile camera calibration, and also the use of advanced 2D and 3D graphics rendering. These functionalities are highly computationally intensive and as such require support from an edge cloud, in addition the virtual objects to be embedded in the physical world are expected to be proactively cached closer to the end user so that latency is minimized. Ultra low latency in these type of applications is of paramount importance so that to provide a photorealistic embedding of virtual objects in the video view of the end user. However, since computational and augmented reality objects need to be readily available, the caching of those objects should be considered in conjunction with the computational capabilities of he edge cloud. In addition to the above when MAR is considered under the lenses of an NFV environment the application might inherently require access to some VNFs and therefore the above discussion on VNF chaining for MAR applications is also valid in this case. \\nRecently the concept of Digital Twin (DT) , has received significant research attention due to the plethora of applications ranging from industrial manufacturing and health to smart cities. In a nutshell, a DT can be defined as an accurate digital replica of a real world object across multiple granularity levels; and this real world object could be a machine, a robot or an industrial process or (sub) system. By reflecting the physical status of the system under consideration in a virtual space open up a plethora of optimization, prediction, fault tolerance and automation process that cannot be done using solely the physical object. At the core of DT applications is the requirement of stringent two–way real time communication between the digital replica and the physical object. This requirement inevitably require support from edge clouds to minimize latency and efficient storage and computational resources including caching. In that setting, the use of the aforementioned deep learning technologies will have a key role to play in order to provide high quality real time decision making to avoid misalignment between the digital replica of the physical object under consideration. Efficient machine-to-DT connectivity would require capabilities similar to the above mentioned augmented reality application but due to the continuous real-time control-loop operation DTs will require a complete new set of network optimization capabilities and in that frontier efficient caching and data-driven techniques will have a central role to play. Hence, as the research regarding the inter-play between low latency communications and DTs is still in embryonic stage there is significant scope in the investigation of suitable data driven deep learning techniques to be utilized for distributed allocation of caching and computing resources.\",\n \"id\": \"f03f2b46-f5ed-47c7-b95e-abf7079d9dd7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"bcb206fa-772b-4e9f-9da9-3ca19215514a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Research Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Caching for Mobile Augmented Reality (MAR) applications and Digital Twins (DTs)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Caching for Mobile Augmented Reality (MAR) applications and Digital Twins (DTs)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As introduced in Section \\\\ref{sec:caching}, cache dimensioning explores the appropriate cache size allocation for content host such as CRs and BSs. Disappointingly, there is rare paper applies DL on cache dimensioning decisions. One proper reason is lack of training data set in contrast to content popular prediction, where we have historical user request log to train a DNN. In addition, the caching size allocation affects the network performance and economic investment. Recently, network slicing is identified as an important tool to enable 5G to provide multi-services with diverse characteristics. The slice is established on physical infrastructure including network storage. Therefore it is a very interesting topic to consider the allocation of the memory space to support content caching and other storage services, which guarantees QoE and satisfies task requirements. Furthermore, for the case lack of training data set, DRL can be viewed as a promising technology to configure slicing settings as well as cache dimensioning. For the action space designing, it can be either discrete by setting storage levels, or continuous which is allocate the memory space directly. However, there are requirements to design caching-enabled network slicing model especially for dynamic allocation as well as associated DRL framework including state space, detailed action space, reward function and agent structure.\",\n \"id\": \"76a7e5ab-c389-46f2-914c-27d36e617810\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"bcb206fa-772b-4e9f-9da9-3ca19215514a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Research Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Deep Learning for Cache Dimensioning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Learning for Cache Dimensioning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Though the application of DL brings performance efficiency for caching policy, additional cost introduced by DL is unneglected, since training and deploying DL model require not only network resources but also time duration. Naturally there is a trade-off between the cost which DL-assisted caching policy saved and the consumption which supports DL itself running, which indicates the trading with DL results in either profit, loss, or break even. Therefore, where and when to apply DL should be carefully investigated. In addition, for the purpose of reducing resource consumption and accelerating training process, some knowledge transfer methods like transfer learning can be utilized, which can transform the knowledge already learnt from the source domain to a relevant target domain.\",\n \"id\": \"bf719ec3-fb89-4a44-9c42-480d45208596\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"bcb206fa-772b-4e9f-9da9-3ca19215514a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Research Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"The Cost of Deep Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"The Cost of Deep Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclusions}\\nThis article presents a comprehensive study for the application of deep learning methods in the area of content caching. Particularly, the data caching is divided into two classifications according to the caching location of edge network. Each category contains where to cache, what to cache, cache dimensioning and content delivery. Then we introduce typical DNN methods which are categorised via training process into supervised learning, unsupervised learning and RL. Further, this paper critically compares and analyzes state-of-the-art papers on parameters, such as DL methods employed, the caching problems solved and the objective of applying DL. The challenges and research directions of DL on caching is also examined on the topic of extending caching to VNF chains, the application of caching for MAR as well as DTs, DL for cache size allocation and the additional cost of employing DL. Undoubtedly, DL is playing a significant role in 5G and beyond. We hope this paper will increase discussions and interests on DL for caching policy design and relevant applications, which will advance future network communications. \\n\\\\vspace{6pt} \\n\\\\bibliographystyle{unsrt}\\n\\\\bibliography{reference}\\n\\\\end{document}\",\n \"id\": \"9f55687b-d2fc-4fa2-a01d-e44198b97b4c\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"347fe6bc-3506-4394-85f3-39521077022f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Learning for Data Caching in Edge Network\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":73,"string":"73"},"all_cites":{"kind":"list like","value":[3705,3582,3704,3706,3420,7633,3703,3709,3707,3710,3708,3712,3716,8682,3714,3717,3713,8681,3715,3711,166,810,167,1409,1408,2219,1391,1390,3721,7746,3718,3719,3723,3722,3720,3724],"string":"[\n 3705,\n 3582,\n 3704,\n 3706,\n 3420,\n 7633,\n 3703,\n 3709,\n 3707,\n 3710,\n 3708,\n 3712,\n 3716,\n 8682,\n 3714,\n 3717,\n 3713,\n 8681,\n 3715,\n 3711,\n 166,\n 810,\n 167,\n 1409,\n 1408,\n 2219,\n 1391,\n 1390,\n 3721,\n 7746,\n 3718,\n 3719,\n 3723,\n 3722,\n 3720,\n 3724\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.553620417173992,"string":"1.55362"}}},{"rowIdx":39,"cells":{"authors":{"kind":"list like","value":["Dominik Sisejkovic","Lennart M. Reimann","Elmira Moussavi","Farhad Merchant","Rainer Leupers"],"string":"[\n \"Dominik Sisejkovic\",\n \"Lennart M. Reimann\",\n \"Elmira Moussavi\",\n \"Farhad Merchant\",\n \"Rainer Leupers\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-07-05T10:18:26Z","string":"2021-07-05T10:18:26Z"},"category":{"kind":"string","value":"cs.CR"},"abstract":{"kind":"string","value":" \t\tIn the past decade, a lot of progress has been made in the design and evaluation of \t\tlogic locking; a premier technique to safeguard the integrity of integrated circuits throughout the electronics supply chain. \t\tHowever, the widespread proliferation of machine learning has recently introduced a new pathway to evaluating logic locking schemes. This paper summarizes the recent developments in logic locking attacks and countermeasures at the frontiers of contemporary machine learning models. Based on the presented work, the key takeaways, opportunities, and challenges are highlighted to offer recommendations for the design of next-generation logic locking. \t"},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"]],"subsections":["4af06642-e7a5-4612-b225-f1cae2b32e2a","75631a2a-145f-4b09-90ff-408cabadfd55","cb342b15-09e2-41e4-9c1f-0d7558736604","4e814a9b-2e2f-4194-94e3-5e0b2adbe038"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"The involvement of untrusted parties in the modern Integrated Circuit (IC) supply has given rise to a plethora of security concerns, including Intellectual Property (IP) piracy, reverse engineering, counterfeiting, and hardware Trojans~. Consequently, a variety of countermeasures have been introduced, including IC metering~, split manufacturing~, camouflaging~, and logic locking~. Among these, only logic locking can protect a design against all untrusted parties in the supply chain~.","id":"4af06642-e7a5-4612-b225-f1cae2b32e2a","level":"section","origin_cites_number":7,"parent_id":"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Introduction"]],"subsections":["a08af25e-5d31-4208-aa5a-4800589c0307","54ad821b-9b66-4cd9-b802-b507cf6fc674"],"title":"Introduction"},{"cite_extract_rate":0.1,"cites":[8606],"content":"Logic locking performs design manipulations by binding the correct functionality of a hardware design to a secret key that is only known to the legitimate IP owner. Hereby, both the original functionality and the structure of the design remain concealed while passing through the hands of external design houses and the foundry. In the past decade, various security aspects of logic locking have been thoroughly evaluated through the introduction of key-recovery attacks~, among which the Boolean satisfiability (SAT) attack has gained a lot of attention~. This has led to a division of logic locking into \\textit{pre- and post-SAT schemes}. Pre-SAT schemes were focusing on specific security features, such as random XOR/XNOR key-gate insertion~, thwarting the path-sensitization attack~ or maximizing output corruption for incorrect keys~. With the introduction of SAT-based attacks, the design objective has shifted towards achieving SAT-resilience, resulting in a new generation of schemes, including SARLock~, Anti-SAT~, CASLock~, SFLL~, and others~.","id":"a08af25e-5d31-4208-aa5a-4800589c0307","level":"subsection","origin_cites_number":10,"parent_id":"4af06642-e7a5-4612-b225-f1cae2b32e2a","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Introduction"],["subsection","Logic Locking: A Brief Overview"]],"subsections":[],"title":"Logic Locking: A Brief Overview"},{"cite_extract_rate":0.125,"cites":[8606,7636],"content":"With the advent of efficient and easy-to-use Machine Learning (ML) models, ML-based techniques have been gradually introduced into various hardware-security domains~. The latest efforts in the logic locking community have been invested in challenging the security properties of locking schemes using ML. Recent works were able to efficiently attack pre- and post-SAT schemes~. {The introduction of ML-based tools for the security analysis of logic locking \\textit{has opened up a new chapter} in the design of locking schemes and attacks, thereby initiating the start of the \\textit{post-ML} locking-scheme era. Herewith, the ML ecosystem offers a novel path to uncover hidden vulnerabilities and provide new directions in the development of future ML-resilient locking schemes.\n\t\t\\textbf{Contributions}\n\t\tThe ML era has undoubtedly initiated a new stage in logic locking design and evaluation. In this paper, we review all major developments in the domain of \\textit{ML-based attacks and countermeasures in logic locking}, and analyze major challenges and research opportunities. Note that a comprehensive overview of the state of pre-ML schemes and attacks can be found in~.\n\t\tThe rest of this paper is organized as follows. Section~\\ref{sec:background} introduces the relevant background on logic locking. Section~\\ref{sec:ll-in-ml-era} reviews the major developments in ML-based logic locking attacks and compiles a summary of the open challenges and opportunities. Finally, Section~\\ref{sec:conclusion} concludes the paper.","id":"54ad821b-9b66-4cd9-b802-b507cf6fc674","level":"subsection","origin_cites_number":16,"parent_id":"4af06642-e7a5-4612-b225-f1cae2b32e2a","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Introduction"],["subsection","The Advent of Machine Learning"]],"subsections":[],"title":"The Advent of Machine Learning"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:background}\n\t\tThis section introduces the preliminaries on classification, working principles and attack models of logic locking.","id":"75631a2a-145f-4b09-90ff-408cabadfd55","level":"section","origin_cites_number":0,"parent_id":"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Background"]],"subsections":["5c43bb1f-f9b2-45b3-995b-7a7bcddebadf","12a9727d-dd75-4b28-bfdf-0fa62287f301","0272a9b9-bb42-4a7b-b2d3-dbebdd61d188","156b835f-9d2c-44be-96fe-edb4453da9fb"],"title":"Background"},{"cite_extract_rate":0,"cites":[],"content":"Logic locking can be generally classified into two orthogonal classes: \\textit{combinational} and \\textit{sequential}~. Combinational logic locking performs key-dependent manipulations in the combinational path of a design. On the other hand, sequential logic locking focuses on transforming and obfuscating the state space of a circuit. As the reviewed work operates in the domain of combinational locking, in the rest of this work, the term \\textit{logic locking} refers to combinational locking schemes.","id":"5c43bb1f-f9b2-45b3-995b-7a7bcddebadf","level":"subsection","origin_cites_number":1,"parent_id":"75631a2a-145f-4b09-90ff-408cabadfd55","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Background"],["subsection","Classification"]],"subsections":[],"title":"Classification"},{"cite_extract_rate":0,"cites":[],"content":"The idea of logic locking lies in the functional and structural manipulation of a hardware design that creates a dependency to an activation key, hereby trading area, power, and delay for security. If the correct key is provided, the locked design will perform as originally intended \\textit{for all input patterns}. Otherwise, an incorrect key will yield an incorrect output for at least \\textit{some input patterns}. Logic locking can be performed on different design levels. However, typically logic locking is deployed on a gate-level netlist through the insertion of additional gates (known as \\textit{key gates}) or more complex structures. A visual example of a locked design is shown in Fig.~\\ref{fig:enc-example}~(b). Here, the original netlist in Fig.~\\ref{fig:enc-example}~(a) is locked through the insertion of two key-controlled gates in the form of an XOR and XNOR (XOR + INV) gate, marked as $KG_{1}$ and $KG_{2}$, respectively. To understand the functional implications of the key gates, let us consider the gate $KG_{1}$. This key gate takes two inputs: the original wire $x$ (output of gate $G_{1}$) and the input key bit $k_{1}$. If a correct key value is set, i.e., $k_{1}=0$, the value of $x$ is preserved and forwarded to $x'$. However, if an incorrect key value is set, i.e., $k_{1}=1$, the value of $x$ is inverted, leading to incorrect output values. Based on this concept, throughout the past decade, a variety of locking schemes have been introduced, based on XOR, XNOR, AND, OR, and MUX gates as well as more elaborate structures~. \n\t\t\\begin{figure}[t]\n\t\t\t\\centering\n\t\t\t\\subfloat[Original Circuit]{\n\t\t\t\t\\includegraphics[width=0.35\\columnwidth]{epic_example.eps}\n\t\t\t}\n\t\t\t\\subfloat[Locked Circuit]{\n\t\t\t\t\\includegraphics[width=0.4\\columnwidth]{epic_example_locked.eps}\n\t\t\t}\n\t\t\t\\caption{Example: logic locking using XOR/XNOR gates.}\n\t\t\t\\label{fig:enc-example}\n\t\t\t\\vspace{-0.1in}\n\t\t\\end{figure}","id":"12a9727d-dd75-4b28-bfdf-0fa62287f301","level":"subsection","origin_cites_number":1,"parent_id":"75631a2a-145f-4b09-90ff-408cabadfd55","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Background"],["subsection","Working Principles"]],"subsections":[],"title":"Working Principles"},{"cite_extract_rate":0,"cites":[],"content":"The role of logic locking in the IC supply chain is demonstrated in Fig.~\\ref{fig:supplychain}. Based on a trusted Register-Transfer Level (RTL) design, the legitimate IP owner performs logic synthesis to generate a gate-level netlist. At this point, logic locking is deployed, resulting in a locked netlist and an activation key. Typically, after the netlist is locked, another synthesis round is performed to facilitate the structural integration of scheme-induced netlist changes. Therefore, we can differentiate between the pre-resynthesis and post-resynthesis netlist. The former is locked but not resynthesized, while the latter is locked and resynthesized. In the next step, the locked netlist proceeds in the untrusted part of the supply chain. This often includes an untrusted external design house (for layout synthesis) and the foundry. After fabrication, the produced IC is returned to the IP owner for activation. Herewith, logic locking protects a design by concealing its functional and structural secrets in the activation key, thereby bridging the untrusted regime gap. \n\t\tIn terms of hardware Trojans, it is assumed that a sound understanding of the design's functionality and structure is required to insert an intelligible, controllable and design-specific Trojan (e.g, a targeted denial-of-service attack). However, functionality-independent Trojans remain viable. These include, e.g., the manipulation of the circuit's physical characteristics, leading to performance or reliability degradation. In the former case, finding the activation key is a prerequisite for successfully performing the reverse-engineering process. \n\t\t\\begin{figure}[!t]\n\t\t\t\\centering\n\t\t\t\\includegraphics[width=\\columnwidth]{flow4.eps}\n\t\t\t\\caption{Logic locking in the IC design and fabrication flow.}\n\t\t\t\\label{fig:supplychain}\n\t\t\t\\vspace{-0.15in}\n\t\t\\end{figure}\n\t\t\\begin{figure*}[t]\n\t\t\t\\centering\n\t\t\t\\includegraphics[width=\\textwidth]{all_attacks.eps}\n\t\t\t\\caption{ML-based attacks on logic locking (deployment phase).}\n\t\t\t\\label{fig:attack-visualization}\n\t\t\t\\vspace{-0.1in}\n\t\t\\end{figure*}","id":"0272a9b9-bb42-4a7b-b2d3-dbebdd61d188","level":"subsection","origin_cites_number":0,"parent_id":"75631a2a-145f-4b09-90ff-408cabadfd55","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Background"],["subsection","Logic Locking in the IC Supply Chain"]],"subsections":[],"title":"Logic Locking in the IC Supply Chain"},{"cite_extract_rate":0,"cites":[],"content":"The attack model includes the following: ($i$) the attacker has access to the locked netlist, either as an untrusted design house or by reverse engineering the layout, ($ii$) the location of the key inputs (pins) is known, ($iii$) the deployed locking scheme is known, and ($iv$) the attacker has access to an activated IC to use as oracle for retrieving golden Input/Output (I/O) patterns.\n\t\tThe fourth assumption is the differentiating factor that classifies all key-retrieval attacks into \\textit{oracle-less} and \\textit{oracle-guided} attacks. Oracle-less attacks assume that an activated design is not available. This is often the case in low-volume production for security-critical applications~. On the contrary, the oracle-guided scenario assumes the availability of an activated IC, thereby representing a high-volume production setting~.\n\t\tMoreover, sometimes, an attack assumes the availability of only a few golden I/O patterns, e.g., in the form of test vectors. Due to the available knowledge that is provided by these patterns, these attacks fall into the oracle-guided class. As discussed in the next section, ML-based attacks have been explored in both classes. Henceforth, we refer to the gate-level netlist under attack as the \\textit{target} netlist.","id":"156b835f-9d2c-44be-96fe-edb4453da9fb","level":"subsection","origin_cites_number":2,"parent_id":"75631a2a-145f-4b09-90ff-408cabadfd55","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Background"],["subsection","Attack Model"]],"subsections":[],"title":"Attack Model"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:ll-in-ml-era}\n\t\tThis section describes the recent developments of ML-based applications in the logic locking domain.","id":"cb342b15-09e2-41e4-9c1f-0d7558736604","level":"section","origin_cites_number":0,"parent_id":"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"]],"subsections":["8a41fbf8-901a-45b8-9e50-9a86985de29d","b5e83d76-9451-4150-b043-b622bb4e368c","ca9a962e-8ebf-4b88-8027-88e803e66e7d","1d12de22-82f0-4f45-8eba-64bb646b3352"],"title":"Logic Locking in the Machine Learning Era"},{"cite_extract_rate":0,"cites":[],"content":"Previous work in this domain has mostly been focusing on the development of novel ML-based attacks on logic locking, both for the oracle-less and oracle-guided model. A simplified visualization of all reviewed attack flows is presented in Fig.~\\ref{fig:attack-visualization}. Hereby, the attacks are presented in the \\textit{deployment} stage (after training). An exhaustive comparison of the attacks is summarized in Table~\\ref{table:attacks}. The comparison lists the attacks in order of appearance per attack class. For convenience, a glossary of the used acronyms is given in Table~\\ref{table:glossary}. \n\t\tThe following review reflects the descriptions in Fig.~\\ref{fig:attack-visualization} and Table~\\ref{table:attacks}, thereby only focusing on the \\textit{basic attack mechanisms}. All other details can be found in the mentioned \\textit{comparison table} and the provided references.","id":"8a41fbf8-901a-45b8-9e50-9a86985de29d","level":"subsection","origin_cites_number":0,"parent_id":"cb342b15-09e2-41e4-9c1f-0d7558736604","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","ML-Based Attacks"]],"subsections":["50b37d47-0116-40d6-afbd-43bf9dacb32b","eda58d36-635e-44a7-9a87-c2f7c7682303"],"title":"ML-Based Attacks"},{"cite_extract_rate":0.103448275862068,"cites":[2932,7636,2931],"content":"ML-based attacks in this class have exploited \\textit{scheme-related structural residue} to identify a correct key-bit value or the locking circuitry itself. This category includes SAIL~, SnapShot~, and GNNUnlock~.\n\t\t\\textbf{SAIL:} This attack deploys ML algorithms to retrieve local logic structures from the post-resynthesis target netlist to predict the correct key values (Fig.~\\ref{fig:attack-visualization}~(a)). The attack targets XOR/XNOR-based logic locking exclusively. Hereby, the attack exploits two leakage points of the locking flow: ($i$) the deterministic structural changes induced by logic synthesis around the key-affected netlist gates and ($ii$) the nature of XOR/XNOR-based locking (XOR for key bit 0 and XNOR for key bit 1). Therefore, the attack encloses two components: the Change Prediction Model (CPM) and the Reconstruction Model (RM). For each netlist subgraph around a key input (known as \\textit{locality}), CPM is trained to predict whether a synthesis-induced change has occurred. If a change is predicted, RM is deployed to reconstruct the pre-resynthesis netlist structures, i.e., it reverses the effect of the synthesis process. Finally, based on the intrinsic nature of XOR/XNOR-based locking, SAIL extracts the correct key value.\n\t\t\\textbf{SnapShot:} The attack utilizes a neuroevolutionary approach to automatically design suitable neural networks to \\textit{directly} predict correct key values from a locked post-resynthesis target netlist (Fig.~\\ref{fig:attack-visualization}~(b)). The attack exploits the structural alterations induced by locking schemes to learn a correlation between the key-induced structural patterns and the key value. Compared to SAIL, SnapShot implements an end-to-end ML approach, thereby having the advantage of being applicable to any locking scheme as well as not relying on learning specific transformation rules of the logic synthesis process.\n\t\t\\textbf{GNNUnlock:} The attack leverages a graph neural network to identify all the gates in a post-resynthesis target netlist that belong to the locking circuitry (Fig.~\\ref{fig:attack-visualization}~(c)). Therefore, compared to SAIL and SnapShot, GNNUnlock learns to differentiate locking gates from regular gates instead of learning a correlation between the netlist sub-graphs and the correct key. To enhance the removal accuracy, after identification, a deterministic post-processing mechanism is deployed to remove the gates depending on the intrinsic features of the underlying schemes. Hereby, GNNUnlock has specifically been trained and deployed to target SAT-attack-resilient schemes, i.e., Provably Secure Logic Locking (PSLL). Herein lies the success of the attack: PSLLs often induce isolable structural netlist changes to produce a specific SAT-resilient circuit behavior. In the pre-ML era, it has been long assumed that PSLLs can be protected through additional locking and resynthesis.\n\t\t\\begin{table}[b]\\footnotesize\n\t\t\t\\centering\n\t\t\t\\vspace{-0.2in}\n\t\t\t\\caption{Glossary}\n\t\t\t\\label{table:glossary}\n\t\t\t\\resizebox{\\columnwidth}{!}{\n\t\t\t\t\\begin{tabular}\n\t\t\t\t\t{l|l||l|l}\n\t\t\t\t\t\\toprule\n\t\t\t\t\t\\textbf{Acronym} & \\textbf{Definition}&\\textbf{Acronym} & \\textbf{Definition}\\\\\n\t\t\t\t\t\\midrule\n\t\t\t\t\t\\midrule\n\t\t\t\t\tAND-OR & AND/OR-based LL~ & LUT & Lookup table \\\\\n\t\t\t\t\tAnti-SAT & Anti-Boolean satisfiability~ & MLP & Multi-layer perceptron\\\\\n\t\t\t\t\tC & Combinational circuits & MPNN & Message-passing neural network\\\\\n\t\t\t\t\tCNN & Convolutional neural network & OG & Oracle-guided attack\\\\\n\t\t\t\t\tCS & Logic-cone-size-based LL~ & OL & Oracle-less attack\\\\\n\t\t\t\t\tD-RNN & Deep recurrent neural network & OLL & Optimal LL~\\\\\n\t\t\t\t\tFLL & Fault analysis-based LL~ & PSO & Particle swarm optimization\\\\\n\t\t\t\t\tFU & Functional & RLL & Random LL~\\\\\n\t\t\t\t\tGA & Genetic algorithm & S & Sequential circuits\\\\\n\t\t\t\t\tGL & Gate level & SFLL-HD & Stripped functionality LL~\\\\\n\t\t\t\t\tGNN & Graph neural network & SLL & Strong (secure) LL~\\\\\n\t\t\t\t\tLL & Logic locking & ST & Structural\\\\\n\t\t\t\t\tLSTM & Long short-term memory & TTLock & Tenacious and traceless LL~\\\\\n\t\t\t\t\t\\bottomrule\n\t\t\t\t\t\\bottomrule\n\t\t\t\t\\end{tabular}\n\t\t\t}\n\t\t\\end{table}\n\t\t\\begin{table*}[hbtp]\\footnotesize\n\t\t\t\\setlength{\\tabcolsep}{2pt}\n\t\t\t\\caption{Overview of ML-Based Attacks on Logic Locking}\n\t\t\t\\label{table:attacks}\n\t\t\t\\resizebox{\\textwidth}{!}{\n\t\t\t\t\\begin{threeparttable}\n\t\t\t\t\t\\centering\n\t\t\t\t\t\\begin{tabular}{l||>{\\centering\\arraybackslash}m{1.3cm}>{\\centering\\arraybackslash}m{0.5cm}>{\\centering\\arraybackslash}m{1.7cm}>{\\centering\\arraybackslash}m{1.2cm}>{\\centering}m{1.7cm}>{\\centering\\arraybackslash}m{1.2cm}>{\\centering}m{1cm}>{\\centering}m{1.2cm}>{\\centering\\arraybackslash}m{2.1cm}>{\\centering\\arraybackslash}m{1.7cm}>{\\centering\\arraybackslash}m{1.5cm}>{\\centering\\arraybackslash}m{1.7cm}}\n\t\t\t\t\t\t\\toprule\n\t\t\t\t\t\t\\textbf{Attack}& \\textbf{Objective}&\\textbf{Class}& \\textbf{Level/IC Type/Attack Basis}&\\textbf{ML Model}&\\textbf{Benchmarks}\\textsuperscript{16}&\\textbf{Evaluated Schemes}&\\textbf{Scheme Independent}&\\textbf{Exact Output}\\textsuperscript{15}&\\textbf{Evaluate Key Length in Bits}&\\textbf{\\% Accuracy [min, max]}&\\textbf{Time Complexity\\textsuperscript{17}}&\\textbf{Known Protection}\\\\\n\t\t\t\t\t\t\\midrule\n\t\t\t\t\t\tSAIL~& key retrieval & OL&GL/C,S\\textsuperscript{1}/ST&Random Forest& ISCAS'85&RLL, SLL, CS& \\xmark&\\xmark&$\\{4,\\cdots,192\\}$\\textsuperscript{4}&$[66.89, 73.88]$&$\\bigO(l)$\\textsuperscript{13}& UNSAIL~, SARO~\\textsuperscript{22}\\\\\\hline\n\t\t\t\t\t\tSnapShot~& ($i$) GSS and ($ii$) SRS key retrieval\\textsuperscript{14}& OL&GL/C,S/ST&MLP, CNN, GA& ISCAS'85, ITC'99, Ariane RV&RLL& \\cmark&\\xmark&$\\{64\\}$&$[57.71, 61.56]^{(i)}$, $[71.57, 81.67]^{(ii)}$&$\\bigO(l)$\\textsuperscript{13}& D-MUX~\\\\\\hline\n\t\t\t\t\t\tGNNUnlock~&key gates removal&OL&GL/C,S/ST&GNN&ISCAS'85, ITC'99&Anti-SAT, TTLock, SFLL-HD&\\xmark\\textsuperscript{12}&\\xmark&$\\{8,\\cdots,128\\}$&$[100.00]$\\textsuperscript{11}&$\\bigO(n)$\\textsuperscript{13}&\\omark\\\\\n\t\t\t\t\t\t\\midrule\n\t\t\t\t\t\t\\midrule\n\t\t\t\t\t\tBOCANet~& key retrieval & OG\\textsuperscript{2}&GL/C/FU&D-RNN \\& LSTM& Trust-Hub, ISCAS'85&RLL&\\cmark&\\xmark & $\\{32,64,128,256\\}$&$[89.00, 100.00]$&$\\bigO(\\alpha)$\\textsuperscript{19}&\\omark\\\\\\hline\n\t\t\t\t\t\tSURF~&key retrieval& OG &GL/C,S\\textsuperscript{1}/ST,FU&SAIL \\& heuristic optimization&ISCAS'85&RLL, SLL, CS&\\xmark\\textsuperscript{3}&\\xmark&$\\{4,\\cdots,192\\}$\\textsuperscript{4} &$[90.58,98.83]$&$\\bigO(t)$\\textsuperscript{13}&UNSAIL~\\\\\\hline\n\t\t\t\t\t\tGenUnlock~&key retrieval& OG &GL/C,S/FU&GA&ISCAS'85, MCNC&RLL,SLL, AND-OR, FLL&\\cmark&\\xmark&$\\{8,\\cdots, 1618\\}$\\textsuperscript{7}&$[100.00]$\\textsuperscript{8}&$\\bigO(\\beta)$\\textsuperscript{20}&\\omark\\\\\\hline\n\t\t\t\t\t\tNNgSAT~&key retrieval& OG &GL/C,S/FU&MPNN&ISCAS'85, ITC'99&SAT-hard\\textsuperscript{5}&\\cmark&\\cmark&n/a\\textsuperscript{5}&$[93.50]$\\textsuperscript{6}&$\\bigO(\\lambda)$\\textsuperscript{18}&\\omark\\\\\t\\hline\n\t\t\t\t\t\tPSO Attack~&key retrieval& OG\\textsuperscript{9} &GL/C,S/FU&PSO&ISCAS'85, ITC'99&RLL, OLL&\\cmark&\\xmark&$\\{64,128\\}$&$[82.07,99.80]$\\textsuperscript{10}&$\\bigO(\\gamma)$\\textsuperscript{21}&point-function locking~\\\\\n\t\t\t\t\t\t\\bottomrule\n\t\t\t\t\t\\end{tabular}\n\t\t\t\t\t\\begin{tablenotes}\n\t\t\t\t\t\t\\scriptsize\n\t\t\t\t\t\t\\item [1]~~The attack is in theory applicable to sequential circuits, however no evaluation has been performed yet.\n\t\t\t\t\t\t\\item [2]~~The attack relies on having access to at least some golden input/output patterns ($<0.5\\%$ of total I/O pairs).\n\t\t\t\t\t\t\\item [3]~~In theory, the key-refinement search algorithm could be utilized based on any seed key. However, this has not been addresses thus far.\n\t\t\t\t\t\t\\item [4]~~The following key lengths have been evaluated: $\\{4,8,16,24,32,64,96,128,192\\}$.\n\t\t\t\t\t\t\\item [5]~~$n\\times m$ bitwise multipliers $(8<m,n<32)$, $n\\times m$ crossbar network of 2-to-1 MUXes $(16<m, n<36)$, $n$-input LUTs built by 2-to-1 MUXes $(n<16)$, and $n$-to-1 AND-trees.\n\t\t\t\t\t\t\\item [6]~~Indicates the percentage of successfully de-obfuscated circuits compared to the baseline~.\n\t\t\t\t\t\t\\item [7]~~The key length is selected based on an area overhead of 5\\%, 10\\% or 25\\%.\n\t\t\t\t\t\t\\item [8]~~The quality of the retrieved approximate keys is quantified by a user-defined output-fidelity measure.\n\t\t\t\t\t\t\\item [9]~~The attack relies on an oracle without access to the scan chain.\n\t\t\t\t\t\t\\item [10]~Accuracy refers to the average number of cases where the retrieved key results in 0\\% erroneous outputs for $10^{6}$ random patterns.\n\t\t\t\t\t\t\\item [11]~The accuracy refers to the successful removal of the locking circuitry (not the key retrieval).\n\t\t\t\t\t\t\\item [12]~The attack has not been evaluated for other locking schemes so far and the post-processing steps are scheme-specific.\n\t\t\t\t\t\t\\item [13]~Notation: the key length $l$, the number of netlist nodes $n$, the total number of iterations $t=p\\cdot{n} + p\\cdot{w}+r\\cdot{l}\\cdot{i}\\cdot{n}+r\\cdot{l}\\cdot{i}\\cdot{p}$, where $p$ is the number of output pins, $i$ is the number of IO pairs, $r$ is the number of runs, and $w$ is the number of wires.\n\t\t\t\t\t\t\\item [14]~GSS refers to the generalized set scenario which trains the ML model based on a set of locked benchmarks that are different from the target. SRS captures the self-referencing scenario where the training data is generated by re-locking the target benchmark.\n\t\t\t\t\t\t\\item [15]~If the output of the attack is an exact result, the attack can guarantee a 100\\% correct deobfuscation for the complete I/O space.\n\t\t\t\t\t\t\\item [16]~ISCAS'85~, MCNC~, ITC'99~, RISC-V Ariane core~, and Trust-Hub~.\n\t\t\t\t\t\t\\item [17]~If the time complexity can be clearly determined, it refers to the time complexity after the training process.\n\t\t\t\t\t\t\\item [18]~The execution time of the SAT attack is $\\sum_{i=1}^{\\lambda}{t_{i}}$, where $\\lambda$ is the number of iterations and $t_{i}$ the time required for one SAT-solver call. $\\lambda$ depends on the characteristics of the search space and the branching preferences of the SAT solver. Therefore, the time complexity of the attack is typically measured in terms of $\\lambda$. More details can be found in~.\n\t\t\t\t\t\t\\item [19]~The complexity is linear to the number of training samples $\\alpha$, as the final key is determined based on the MSE of the trained outputs and the key-induced generated outputs.\n\t\t\t\t\t\t\\item [20]~$\\beta=g\\cdot{p}\\cdot{l}$, where $g$ is the number of generations, $p$ is the population size, and $l$ is the key length. Note that the complexity changes with any adaptations of the GA.\n\t\t\t\t\t\t\\item [21]~$\\gamma=g\\cdot{p}\\cdot{\\delta}$, where $g$ is the number of generations, $p$ is the population size, and $\\delta$ the complexity of performing circuit simulation for the fitness evaluation.\n\t\t\t\t\t\t\\item [22]~An empirical evaluation of the resilience against SAIL has not been presented in the paper; only a discussion based on a proposed metric system has been provided.\n\t\t\t\t\t\\end{tablenotes}\n\t\t\t\t\\end{threeparttable}\n\t\t\t}\n\t\t\t\\vspace{-0.1in}\n\t\t\\end{table*}","id":"50b37d47-0116-40d6-afbd-43bf9dacb32b","level":"subsubsection","origin_cites_number":29,"parent_id":"8a41fbf8-901a-45b8-9e50-9a86985de29d","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","ML-Based Attacks"],["subsubsection","Oracle-Less Attacks"]],"subsections":[],"title":"Oracle-Less Attacks"},{"cite_extract_rate":0.16666666666666602,"cites":[7636],"content":"Due to the availability of an oracle, existing ML-based attacks in this class have mostly exploited functional features of the target IC. This class includes the following attacks: BOCANet~, SURF~, GenUnlock~, NNgSAT~, and the PSO-guided attack~.\n\t\t\\textbf{BOCANet:}~This attack leverages Recurrent Neural Networks (RNN) based on long short-term memory to construct a correct activation key (Fig.~\\ref{fig:attack-visualization}~(d)). The ML model is trained on a sequence of I/O observations taken from an activated IC, thereby learning the functional I/O mapping of the circuit, i.e., its Boolean function. Once trained, the key retrieval consists of two steps. First, a random key is applied to the model as input. Second, the initial key value is subsequently updated based on the Mean-Squared-Error (MSE) of the trained outputs and the newly generated outputs that are affected by the introduced key. Note that the ML model can be utilized to predict correct inputs or outputs as well. BOCANet exploits the functional effect a correct key has on generating a correct I/O mapping.\n\t\t\\textbf{SURF:}~This attack is based on a joint structural and functional analysis of the circuit to retrieve the activation key (Fig.~\\ref{fig:attack-visualization}~(e)). The ML-aspect of the attack lies in \\textit{utilizing the SAIL attack} to generate the pre-resynthesis netlist structures and a seed key. Afterwards, based on the outputs of SAIL, SURF iteratively refines the key by means of a structure-aware greedy optimization algorithm guided by a functional simulation of the obfuscated netlist and a set of golden I/O pairs. The optimization is guided by the observation that specific key gates only affect a specific set of outputs. Thus, the key bits can be partitioned based on which output they affect. Performing a systematic perturbation of the key bits can lead to a more refined key. The success of the heuristic is grounded in the limited local effects that traditional locking schemes have on the value of the output for incorrect keys.\n\t\t\\textbf{GenUnlock:}~The attack flow of GenUnlock leverages a Genetic Algorithm (GA)-based exploration of suitable activation keys for a locked circuit (Fig.~\\ref{fig:attack-visualization}~(f)). The heuristic search is steered by the key fitness that is computed based on the matching ratio of the key on the golden I/O training set. Through multiple generations, the fitness of the key population is subsequently improved through the application of genetic operators (selection, crossover, and mutation). Once the accepted tolerance for the correctness of the key is reached, the algorithm returns the set of the fittest keys. Similarly to SURF, GenUnlock exploits the fact that a heuristic key-refinement procedure eventually leads to more accurate activation keys. \n\t\t\\textbf{NNgSAT:}~The main objective of NNgSAT is the deployment of a Message-Passing Neural Network (MPNN) to facilitate the resolution of SAT-hard circuit structures during the application of a SAT-based attack (Fig.~\\ref{fig:attack-visualization}~(g)). The motivation is driven by the fact that common SAT solvers run into scalability issues when tackling hard-to-be-solved locked circuit structures, e.g, multipliers and AND trees. Therefore, in NNgSAT, a neural network is trained to predict the satisfying assignment on a set of SAT-hard cases. In deployment, the SAT-attack flow offloads the SAT-hard problems to the trained model to speed up the attack procedure. The effectiveness of NNgSAT lies in the fact that it is possible to transfer prediction knowledge from learned clauses to unseen problems. \n\t\t\\textbf{PSO-guided Attack:}~This attack is based on a Particle Swarm Optimization (PSO) heuristic that searches through the key space directed by a selected cost function (Fig.~\\ref{fig:attack-visualization}~(f)). The cost function is modeled as the Hamming distance between the golden and the obtained output responses (for a selected key). Therefore, the search algorithm relies on having access to an activated IC to compare against. A major motivator for this attack is its applicability without having access to an open scan chain, as this is often a limiting factor for SAT-based attacks. In essence, the PSO-guided attack and GenUnlock are similar in nature, as both rely on black-box evolutionary procedures guided by a functionality-driven objective function.","id":"eda58d36-635e-44a7-9a87-c2f7c7682303","level":"subsubsection","origin_cites_number":6,"parent_id":"8a41fbf8-901a-45b8-9e50-9a86985de29d","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","ML-Based Attacks"],["subsubsection","Oracle-Guided Attacks"]],"subsections":[],"title":"Oracle-Guided Attacks"},{"cite_extract_rate":0.6666666666666661,"cites":[2932,2931],"content":"\\textbf{UNSAIL:}~This logic locking scheme has been developed to thwart attacks that target the resolution of structural transformations of logic synthesis~. The core idea of UNSAIL is to generate confusing training data that leads to false predictions in the CPM and RM modules of SAIL. This is realized through the additional manipulation of the netlist after synthesis to force the existence of \\textit{equivalent} netlist sub-graph observations that are linked to \\textit{different} key values.\n\t\t\\textbf{SARO:}~The Scalable Attack-Resistant Obfuscation (SARO) operates in two steps~. First, SARO splits the design into smaller partitions to maximize the structural alterations in the netlist. Second, a systematic truth table transformation is deployed to lock the partitions. In order to increase the complexity of pattern recognition attacks (such as SAIL), the transformations aim to maximize randomness in the netlist.\n\t\t\\textbf{Point-Functions and PSO:} As mentioned in Table~\\ref{table:attacks}, the PSO-guided attack is not applicable to point-function-based locking schemes. The reason is that this type of locking yields SAT-resilient behavior in which any incorrect key corrupts only \\textit{a very limited amount of outputs}. Consequently, this behavior offers no advantage in the guidance of the heuristic search, as it does not yield differentiating fitness values. Note that the applicability of point functions depends on the design of the fitness function that is used to guide the heuristic.\n\t\t\\textbf{D-MUX:} The recently introduced Deceptive Multiplexer (D-MUX) LL scheme builds on the concept of multiple MUX-based insertion strategies that create structural paths that are equally likely to be driven by 0 or 1 key values~. Hence, D-MUX offers efficient protection against data-driven attacks.","id":"b5e83d76-9451-4150-b043-b622bb4e368c","level":"subsection","origin_cites_number":3,"parent_id":"cb342b15-09e2-41e4-9c1f-0d7558736604","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","ML-Resilient Schemes"]],"subsections":[],"title":"ML-Resilient Schemes"},{"cite_extract_rate":0,"cites":[],"content":"\\textbf{Deobfuscation Runtime Prediction:}~Apart from ML-based attacks, machine learning has also found its way into other aspects of logic locking. A recent work has designed a framework named ICNet for the prediction of the key-retrieval runtime for SAT-based attacks using graph neural networks~. The framework obtains the predicted deobfuscation runtime based on the characteristics of the circuit topology. ICNet offers an end-to-end approach to evaluate the hardness of logic locking with respect to the SAT attack, thereby increasing the development efficiency of novel locking schemes. \n\t\t\\textbf{ML-Attack on Analog IC Locking:}~ML has started to have an impact on locking mechanisms even beyond digital circuits. The authors in~ have developed an oracle-guided attack on locked analog ICs using genetic algorithms. The approach has successfully broken all known analog logic locking techniques.","id":"ca9a962e-8ebf-4b88-8027-88e803e66e7d","level":"subsection","origin_cites_number":2,"parent_id":"cb342b15-09e2-41e4-9c1f-0d7558736604","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Other Applications of ML"]],"subsections":[],"title":"Other Applications of ML"},{"cite_extract_rate":0,"cites":[],"content":"The presented efforts gather around two focal points: oracle-less and oracle-guided attacks. The intrinsic mechanisms of these attacks shed light on major vulnerabilities in existing logic locking that are exploitable by ML. We summarize the observations as follows:","id":"1d12de22-82f0-4f45-8eba-64bb646b3352","level":"subsection","origin_cites_number":0,"parent_id":"cb342b15-09e2-41e4-9c1f-0d7558736604","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"]],"subsections":["f1a22310-9774-4923-88e9-39031f0c57c0","45fa94fe-3444-4d74-801c-6eb313e3eb65","0f226960-767d-435e-ba81-0709f1f4a0d8","058cfcba-944a-40ba-a6a6-3673887f8227","7bee07b9-543e-4474-ab4e-8d349ee21c61","4f146a54-7ec3-42e7-9ae0-dbd895f2b463","71768a97-ef1a-4c84-ad10-da8f3f475652","6c43d9c7-1dfb-4c58-8706-caf10292d0cd"],"title":"Lessons Learned - Challenges and Opportunities"},{"cite_extract_rate":0,"cites":[],"content":"Oracle-guided attacks focus on functional aspects of schemes, whereas structural leakage is exploited in the oracle-less model due to the absence of an activated IC. This indicates two pitfalls. First, the evaluated schemes have a predictable effect on the functionality of the circuit for incorrect keys, enabling the possibility to perform a guided heuristic search of the key space. Second, the existing schemes induce structural changes which strongly correlate with the value of the key. A mitigation depends on what an attempted attack tries to exploit. For example, to overcome SAIL or SnapShot, a scheme must not reveal anything about the correctness of the key through the induced structural change. This is achieved if the inserted change does not differ depending on the key value. Similarly, to protect against GNNUnlock, the deployed schemes have to ensure resilience against isolation, i.e., the locking circuitry must not be structurally independent from the original design. In the case of GenUnlock and the PSO attack, the behavior of the underlying scheme must be constant or fully random for all incorrect key values; disabling any chance for a guided heuristic convergence towards a correct key. Similar observations can be made for the other attacks as well. Nevertheless, conveying all necessary security objectives into a uniform locking scheme remains an open challenge.","id":"f1a22310-9774-4923-88e9-39031f0c57c0","level":"subsubsection","origin_cites_number":0,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","Structural vs. Functional Analysis"]],"subsections":[],"title":"Structural vs. Functional Analysis"},{"cite_extract_rate":0,"cites":[],"content":"The reliance on logic synthesis transformations to enable the security of logic locking schemes has to be revised. As shown in SAIL, the synthesis rules are predictable and reversible, thereby having little impact on deducing a correct key for traditional schemes. Nevertheless, resynthesis can increase the difficulty to reverse engineering the functionality of a design. However, this needs further evaluation in the context of novel locking policies.","id":"45fa94fe-3444-4d74-801c-6eb313e3eb65","level":"subsubsection","origin_cites_number":0,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","Logic Synthesis for Security"]],"subsections":[],"title":"Logic Synthesis for Security"},{"cite_extract_rate":0,"cites":[],"content":"PSLL schemes induce isolable changes which create a clear structural distinction between the locking-related gates and the original (regular) gates. The main reason for this vulnerability is the requirement of achieving SAT-resistant behavior for incorrect keys. This specific functional pattern requires significant structural netlist changes consolidated in isolable components (e.g., the tree of complementary logic in Anti-SAT or the restore/perturb units in TTLock and SFLL-HD).","id":"0f226960-767d-435e-ba81-0709f1f4a0d8","level":"subsubsection","origin_cites_number":0,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","Structural Isolation of Post-SAT Schemes"]],"subsections":[],"title":"Structural Isolation of Post-SAT Schemes"},{"cite_extract_rate":0,"cites":[],"content":"Regardless of the attack type, ML-based attacks tend to generate an approximate output (as marked in Table~\\ref{table:attacks}), meaning that the retrieved key or deobfuscated circuit cannot be evaluated as correct with an exact certainty. Note that this is not the case for NNgSAT. Nevertheless, the approximate output can be utilized as seed for other attacks to speed up the reverse engineering procedure.","id":"058cfcba-944a-40ba-a6a6-3673887f8227","level":"subsubsection","origin_cites_number":0,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","Output Uncertainty"]],"subsections":[],"title":"Output Uncertainty"},{"cite_extract_rate":0,"cites":[],"content":"As shown in Table~\\ref{table:attacks}, the existing attacks focus on extrapolating information leakage from gate-level netlists. So far, it is unclear if the same leakage is present in RTL-based locking schemes as well.","id":"7bee07b9-543e-4474-ab4e-8d349ee21c61","level":"subsubsection","origin_cites_number":0,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","RTL vs Gate-Level"]],"subsections":[],"title":"RTL vs Gate-Level"},{"cite_extract_rate":0,"cites":[],"content":"Available mitigation schemes suffer from overspecialization for thwarting specific attack vectors. So far, it has not been evaluated which form of logic locking offers a comprehensive protection against any form of ML-based guessing attack. Hence, a potentially fruitful opportunity lies within the automatic design of resilient schemes~.","id":"4f146a54-7ec3-42e7-9ae0-dbd895f2b463","level":"subsubsection","origin_cites_number":1,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","Overspecialization"]],"subsections":[],"title":"Overspecialization"},{"cite_extract_rate":0,"cites":[],"content":"To the best of our knowledge, ML-based techniques have not been deployed yet for the evaluation of sequential logic locking techniques.","id":"71768a97-ef1a-4c84-ad10-da8f3f475652","level":"subsubsection","origin_cites_number":0,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","ML for Sequential Locking"]],"subsections":[],"title":"ML for Sequential Locking"},{"cite_extract_rate":0,"cites":[],"content":"There is still a need to have access to a rich benchmark set for data-driven analysis~.","id":"6c43d9c7-1dfb-4c58-8706-caf10292d0cd","level":"subsubsection","origin_cites_number":1,"parent_id":"1d12de22-82f0-4f45-8eba-64bb646b3352","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Logic Locking in the Machine Learning Era"],["subsection","Lessons Learned - Challenges and Opportunities"],["subsubsection","The Need for Benchmarks"]],"subsections":[],"title":"The Need for Benchmarks"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclusion}\n\t\tThis work summarizes the recent developments of logic locking at the frontiers of machine learning. The presented ML-based attacks indicate the presence of structural and functional leakage in contemporary locking schemes which has been overlooked by the traditional interpretation of security. We show that an ML-based analysis is able to pinpoint novel vulnerabilities and challenge existing security assumptions. The offered discussion consolidates the major ML-induced challenges in logic locking design, offering a fruitful ground for future research directions in the post-ML era.\n\t\t\\bibliographystyle{IEEEtran}\n\t\t\\bibliography{bibliography_short}\n\t\\end{document}","id":"4e814a9b-2e2f-4194-94e3-5e0b2adbe038","level":"section","origin_cites_number":0,"parent_id":"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee","prefix_titles":[["title","Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ]\n ],\n \"subsections\": [\n \"4af06642-e7a5-4612-b225-f1cae2b32e2a\",\n \"75631a2a-145f-4b09-90ff-408cabadfd55\",\n \"cb342b15-09e2-41e4-9c1f-0d7558736604\",\n \"4e814a9b-2e2f-4194-94e3-5e0b2adbe038\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The involvement of untrusted parties in the modern Integrated Circuit (IC) supply has given rise to a plethora of security concerns, including Intellectual Property (IP) piracy, reverse engineering, counterfeiting, and hardware Trojans~. Consequently, a variety of countermeasures have been introduced, including IC metering~, split manufacturing~, camouflaging~, and logic locking~. Among these, only logic locking can protect a design against all untrusted parties in the supply chain~.\",\n \"id\": \"4af06642-e7a5-4612-b225-f1cae2b32e2a\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [\n \"a08af25e-5d31-4208-aa5a-4800589c0307\",\n \"54ad821b-9b66-4cd9-b802-b507cf6fc674\"\n ],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.1,\n \"cites\": [\n 8606\n ],\n \"content\": \"Logic locking performs design manipulations by binding the correct functionality of a hardware design to a secret key that is only known to the legitimate IP owner. Hereby, both the original functionality and the structure of the design remain concealed while passing through the hands of external design houses and the foundry. In the past decade, various security aspects of logic locking have been thoroughly evaluated through the introduction of key-recovery attacks~, among which the Boolean satisfiability (SAT) attack has gained a lot of attention~. This has led to a division of logic locking into \\\\textit{pre- and post-SAT schemes}. Pre-SAT schemes were focusing on specific security features, such as random XOR/XNOR key-gate insertion~, thwarting the path-sensitization attack~ or maximizing output corruption for incorrect keys~. With the introduction of SAT-based attacks, the design objective has shifted towards achieving SAT-resilience, resulting in a new generation of schemes, including SARLock~, Anti-SAT~, CASLock~, SFLL~, and others~.\",\n \"id\": \"a08af25e-5d31-4208-aa5a-4800589c0307\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"4af06642-e7a5-4612-b225-f1cae2b32e2a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Logic Locking: A Brief Overview\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Logic Locking: A Brief Overview\"\n },\n {\n \"cite_extract_rate\": 0.125,\n \"cites\": [\n 8606,\n 7636\n ],\n \"content\": \"With the advent of efficient and easy-to-use Machine Learning (ML) models, ML-based techniques have been gradually introduced into various hardware-security domains~. The latest efforts in the logic locking community have been invested in challenging the security properties of locking schemes using ML. Recent works were able to efficiently attack pre- and post-SAT schemes~. {The introduction of ML-based tools for the security analysis of logic locking \\\\textit{has opened up a new chapter} in the design of locking schemes and attacks, thereby initiating the start of the \\\\textit{post-ML} locking-scheme era. Herewith, the ML ecosystem offers a novel path to uncover hidden vulnerabilities and provide new directions in the development of future ML-resilient locking schemes.\\n\\t\\t\\\\textbf{Contributions}\\n\\t\\tThe ML era has undoubtedly initiated a new stage in logic locking design and evaluation. In this paper, we review all major developments in the domain of \\\\textit{ML-based attacks and countermeasures in logic locking}, and analyze major challenges and research opportunities. Note that a comprehensive overview of the state of pre-ML schemes and attacks can be found in~.\\n\\t\\tThe rest of this paper is organized as follows. Section~\\\\ref{sec:background} introduces the relevant background on logic locking. Section~\\\\ref{sec:ll-in-ml-era} reviews the major developments in ML-based logic locking attacks and compiles a summary of the open challenges and opportunities. Finally, Section~\\\\ref{sec:conclusion} concludes the paper.\",\n \"id\": \"54ad821b-9b66-4cd9-b802-b507cf6fc674\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"4af06642-e7a5-4612-b225-f1cae2b32e2a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"The Advent of Machine Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"The Advent of Machine Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:background}\\n\\t\\tThis section introduces the preliminaries on classification, working principles and attack models of logic locking.\",\n \"id\": \"75631a2a-145f-4b09-90ff-408cabadfd55\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Background\"\n ]\n ],\n \"subsections\": [\n \"5c43bb1f-f9b2-45b3-995b-7a7bcddebadf\",\n \"12a9727d-dd75-4b28-bfdf-0fa62287f301\",\n \"0272a9b9-bb42-4a7b-b2d3-dbebdd61d188\",\n \"156b835f-9d2c-44be-96fe-edb4453da9fb\"\n ],\n \"title\": \"Background\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Logic locking can be generally classified into two orthogonal classes: \\\\textit{combinational} and \\\\textit{sequential}~. Combinational logic locking performs key-dependent manipulations in the combinational path of a design. On the other hand, sequential logic locking focuses on transforming and obfuscating the state space of a circuit. As the reviewed work operates in the domain of combinational locking, in the rest of this work, the term \\\\textit{logic locking} refers to combinational locking schemes.\",\n \"id\": \"5c43bb1f-f9b2-45b3-995b-7a7bcddebadf\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"75631a2a-145f-4b09-90ff-408cabadfd55\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Classification\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Classification\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The idea of logic locking lies in the functional and structural manipulation of a hardware design that creates a dependency to an activation key, hereby trading area, power, and delay for security. If the correct key is provided, the locked design will perform as originally intended \\\\textit{for all input patterns}. Otherwise, an incorrect key will yield an incorrect output for at least \\\\textit{some input patterns}. Logic locking can be performed on different design levels. However, typically logic locking is deployed on a gate-level netlist through the insertion of additional gates (known as \\\\textit{key gates}) or more complex structures. A visual example of a locked design is shown in Fig.~\\\\ref{fig:enc-example}~(b). Here, the original netlist in Fig.~\\\\ref{fig:enc-example}~(a) is locked through the insertion of two key-controlled gates in the form of an XOR and XNOR (XOR + INV) gate, marked as $KG_{1}$ and $KG_{2}$, respectively. To understand the functional implications of the key gates, let us consider the gate $KG_{1}$. This key gate takes two inputs: the original wire $x$ (output of gate $G_{1}$) and the input key bit $k_{1}$. If a correct key value is set, i.e., $k_{1}=0$, the value of $x$ is preserved and forwarded to $x'$. However, if an incorrect key value is set, i.e., $k_{1}=1$, the value of $x$ is inverted, leading to incorrect output values. Based on this concept, throughout the past decade, a variety of locking schemes have been introduced, based on XOR, XNOR, AND, OR, and MUX gates as well as more elaborate structures~. \\n\\t\\t\\\\begin{figure}[t]\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\subfloat[Original Circuit]{\\n\\t\\t\\t\\t\\\\includegraphics[width=0.35\\\\columnwidth]{epic_example.eps}\\n\\t\\t\\t}\\n\\t\\t\\t\\\\subfloat[Locked Circuit]{\\n\\t\\t\\t\\t\\\\includegraphics[width=0.4\\\\columnwidth]{epic_example_locked.eps}\\n\\t\\t\\t}\\n\\t\\t\\t\\\\caption{Example: logic locking using XOR/XNOR gates.}\\n\\t\\t\\t\\\\label{fig:enc-example}\\n\\t\\t\\t\\\\vspace{-0.1in}\\n\\t\\t\\\\end{figure}\",\n \"id\": \"12a9727d-dd75-4b28-bfdf-0fa62287f301\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"75631a2a-145f-4b09-90ff-408cabadfd55\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Working Principles\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Working Principles\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The role of logic locking in the IC supply chain is demonstrated in Fig.~\\\\ref{fig:supplychain}. Based on a trusted Register-Transfer Level (RTL) design, the legitimate IP owner performs logic synthesis to generate a gate-level netlist. At this point, logic locking is deployed, resulting in a locked netlist and an activation key. Typically, after the netlist is locked, another synthesis round is performed to facilitate the structural integration of scheme-induced netlist changes. Therefore, we can differentiate between the pre-resynthesis and post-resynthesis netlist. The former is locked but not resynthesized, while the latter is locked and resynthesized. In the next step, the locked netlist proceeds in the untrusted part of the supply chain. This often includes an untrusted external design house (for layout synthesis) and the foundry. After fabrication, the produced IC is returned to the IP owner for activation. Herewith, logic locking protects a design by concealing its functional and structural secrets in the activation key, thereby bridging the untrusted regime gap. \\n\\t\\tIn terms of hardware Trojans, it is assumed that a sound understanding of the design's functionality and structure is required to insert an intelligible, controllable and design-specific Trojan (e.g, a targeted denial-of-service attack). However, functionality-independent Trojans remain viable. These include, e.g., the manipulation of the circuit's physical characteristics, leading to performance or reliability degradation. In the former case, finding the activation key is a prerequisite for successfully performing the reverse-engineering process. \\n\\t\\t\\\\begin{figure}[!t]\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\includegraphics[width=\\\\columnwidth]{flow4.eps}\\n\\t\\t\\t\\\\caption{Logic locking in the IC design and fabrication flow.}\\n\\t\\t\\t\\\\label{fig:supplychain}\\n\\t\\t\\t\\\\vspace{-0.15in}\\n\\t\\t\\\\end{figure}\\n\\t\\t\\\\begin{figure*}[t]\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\includegraphics[width=\\\\textwidth]{all_attacks.eps}\\n\\t\\t\\t\\\\caption{ML-based attacks on logic locking (deployment phase).}\\n\\t\\t\\t\\\\label{fig:attack-visualization}\\n\\t\\t\\t\\\\vspace{-0.1in}\\n\\t\\t\\\\end{figure*}\",\n \"id\": \"0272a9b9-bb42-4a7b-b2d3-dbebdd61d188\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"75631a2a-145f-4b09-90ff-408cabadfd55\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Logic Locking in the IC Supply Chain\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Logic Locking in the IC Supply Chain\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The attack model includes the following: ($i$) the attacker has access to the locked netlist, either as an untrusted design house or by reverse engineering the layout, ($ii$) the location of the key inputs (pins) is known, ($iii$) the deployed locking scheme is known, and ($iv$) the attacker has access to an activated IC to use as oracle for retrieving golden Input/Output (I/O) patterns.\\n\\t\\tThe fourth assumption is the differentiating factor that classifies all key-retrieval attacks into \\\\textit{oracle-less} and \\\\textit{oracle-guided} attacks. Oracle-less attacks assume that an activated design is not available. This is often the case in low-volume production for security-critical applications~. On the contrary, the oracle-guided scenario assumes the availability of an activated IC, thereby representing a high-volume production setting~.\\n\\t\\tMoreover, sometimes, an attack assumes the availability of only a few golden I/O patterns, e.g., in the form of test vectors. Due to the available knowledge that is provided by these patterns, these attacks fall into the oracle-guided class. As discussed in the next section, ML-based attacks have been explored in both classes. Henceforth, we refer to the gate-level netlist under attack as the \\\\textit{target} netlist.\",\n \"id\": \"156b835f-9d2c-44be-96fe-edb4453da9fb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"75631a2a-145f-4b09-90ff-408cabadfd55\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Attack Model\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attack Model\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:ll-in-ml-era}\\n\\t\\tThis section describes the recent developments of ML-based applications in the logic locking domain.\",\n \"id\": \"cb342b15-09e2-41e4-9c1f-0d7558736604\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ]\n ],\n \"subsections\": [\n \"8a41fbf8-901a-45b8-9e50-9a86985de29d\",\n \"b5e83d76-9451-4150-b043-b622bb4e368c\",\n \"ca9a962e-8ebf-4b88-8027-88e803e66e7d\",\n \"1d12de22-82f0-4f45-8eba-64bb646b3352\"\n ],\n \"title\": \"Logic Locking in the Machine Learning Era\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Previous work in this domain has mostly been focusing on the development of novel ML-based attacks on logic locking, both for the oracle-less and oracle-guided model. A simplified visualization of all reviewed attack flows is presented in Fig.~\\\\ref{fig:attack-visualization}. Hereby, the attacks are presented in the \\\\textit{deployment} stage (after training). An exhaustive comparison of the attacks is summarized in Table~\\\\ref{table:attacks}. The comparison lists the attacks in order of appearance per attack class. For convenience, a glossary of the used acronyms is given in Table~\\\\ref{table:glossary}. \\n\\t\\tThe following review reflects the descriptions in Fig.~\\\\ref{fig:attack-visualization} and Table~\\\\ref{table:attacks}, thereby only focusing on the \\\\textit{basic attack mechanisms}. All other details can be found in the mentioned \\\\textit{comparison table} and the provided references.\",\n \"id\": \"8a41fbf8-901a-45b8-9e50-9a86985de29d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"cb342b15-09e2-41e4-9c1f-0d7558736604\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"ML-Based Attacks\"\n ]\n ],\n \"subsections\": [\n \"50b37d47-0116-40d6-afbd-43bf9dacb32b\",\n \"eda58d36-635e-44a7-9a87-c2f7c7682303\"\n ],\n \"title\": \"ML-Based Attacks\"\n },\n {\n \"cite_extract_rate\": 0.103448275862068,\n \"cites\": [\n 2932,\n 7636,\n 2931\n ],\n \"content\": \"ML-based attacks in this class have exploited \\\\textit{scheme-related structural residue} to identify a correct key-bit value or the locking circuitry itself. This category includes SAIL~, SnapShot~, and GNNUnlock~.\\n\\t\\t\\\\textbf{SAIL:} This attack deploys ML algorithms to retrieve local logic structures from the post-resynthesis target netlist to predict the correct key values (Fig.~\\\\ref{fig:attack-visualization}~(a)). The attack targets XOR/XNOR-based logic locking exclusively. Hereby, the attack exploits two leakage points of the locking flow: ($i$) the deterministic structural changes induced by logic synthesis around the key-affected netlist gates and ($ii$) the nature of XOR/XNOR-based locking (XOR for key bit 0 and XNOR for key bit 1). Therefore, the attack encloses two components: the Change Prediction Model (CPM) and the Reconstruction Model (RM). For each netlist subgraph around a key input (known as \\\\textit{locality}), CPM is trained to predict whether a synthesis-induced change has occurred. If a change is predicted, RM is deployed to reconstruct the pre-resynthesis netlist structures, i.e., it reverses the effect of the synthesis process. Finally, based on the intrinsic nature of XOR/XNOR-based locking, SAIL extracts the correct key value.\\n\\t\\t\\\\textbf{SnapShot:} The attack utilizes a neuroevolutionary approach to automatically design suitable neural networks to \\\\textit{directly} predict correct key values from a locked post-resynthesis target netlist (Fig.~\\\\ref{fig:attack-visualization}~(b)). The attack exploits the structural alterations induced by locking schemes to learn a correlation between the key-induced structural patterns and the key value. Compared to SAIL, SnapShot implements an end-to-end ML approach, thereby having the advantage of being applicable to any locking scheme as well as not relying on learning specific transformation rules of the logic synthesis process.\\n\\t\\t\\\\textbf{GNNUnlock:} The attack leverages a graph neural network to identify all the gates in a post-resynthesis target netlist that belong to the locking circuitry (Fig.~\\\\ref{fig:attack-visualization}~(c)). Therefore, compared to SAIL and SnapShot, GNNUnlock learns to differentiate locking gates from regular gates instead of learning a correlation between the netlist sub-graphs and the correct key. To enhance the removal accuracy, after identification, a deterministic post-processing mechanism is deployed to remove the gates depending on the intrinsic features of the underlying schemes. Hereby, GNNUnlock has specifically been trained and deployed to target SAT-attack-resilient schemes, i.e., Provably Secure Logic Locking (PSLL). Herein lies the success of the attack: PSLLs often induce isolable structural netlist changes to produce a specific SAT-resilient circuit behavior. In the pre-ML era, it has been long assumed that PSLLs can be protected through additional locking and resynthesis.\\n\\t\\t\\\\begin{table}[b]\\\\footnotesize\\n\\t\\t\\t\\\\centering\\n\\t\\t\\t\\\\vspace{-0.2in}\\n\\t\\t\\t\\\\caption{Glossary}\\n\\t\\t\\t\\\\label{table:glossary}\\n\\t\\t\\t\\\\resizebox{\\\\columnwidth}{!}{\\n\\t\\t\\t\\t\\\\begin{tabular}\\n\\t\\t\\t\\t\\t{l|l||l|l}\\n\\t\\t\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\t\\t\\\\textbf{Acronym} & \\\\textbf{Definition}&\\\\textbf{Acronym} & \\\\textbf{Definition}\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\midrule\\n\\t\\t\\t\\t\\t\\\\midrule\\n\\t\\t\\t\\t\\tAND-OR & AND/OR-based LL~ & LUT & Lookup table \\\\\\\\\\n\\t\\t\\t\\t\\tAnti-SAT & Anti-Boolean satisfiability~ & MLP & Multi-layer perceptron\\\\\\\\\\n\\t\\t\\t\\t\\tC & Combinational circuits & MPNN & Message-passing neural network\\\\\\\\\\n\\t\\t\\t\\t\\tCNN & Convolutional neural network & OG & Oracle-guided attack\\\\\\\\\\n\\t\\t\\t\\t\\tCS & Logic-cone-size-based LL~ & OL & Oracle-less attack\\\\\\\\\\n\\t\\t\\t\\t\\tD-RNN & Deep recurrent neural network & OLL & Optimal LL~\\\\\\\\\\n\\t\\t\\t\\t\\tFLL & Fault analysis-based LL~ & PSO & Particle swarm optimization\\\\\\\\\\n\\t\\t\\t\\t\\tFU & Functional & RLL & Random LL~\\\\\\\\\\n\\t\\t\\t\\t\\tGA & Genetic algorithm & S & Sequential circuits\\\\\\\\\\n\\t\\t\\t\\t\\tGL & Gate level & SFLL-HD & Stripped functionality LL~\\\\\\\\\\n\\t\\t\\t\\t\\tGNN & Graph neural network & SLL & Strong (secure) LL~\\\\\\\\\\n\\t\\t\\t\\t\\tLL & Logic locking & ST & Structural\\\\\\\\\\n\\t\\t\\t\\t\\tLSTM & Long short-term memory & TTLock & Tenacious and traceless LL~\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\bottomrule\\n\\t\\t\\t\\t\\t\\\\bottomrule\\n\\t\\t\\t\\t\\\\end{tabular}\\n\\t\\t\\t}\\n\\t\\t\\\\end{table}\\n\\t\\t\\\\begin{table*}[hbtp]\\\\footnotesize\\n\\t\\t\\t\\\\setlength{\\\\tabcolsep}{2pt}\\n\\t\\t\\t\\\\caption{Overview of ML-Based Attacks on Logic Locking}\\n\\t\\t\\t\\\\label{table:attacks}\\n\\t\\t\\t\\\\resizebox{\\\\textwidth}{!}{\\n\\t\\t\\t\\t\\\\begin{threeparttable}\\n\\t\\t\\t\\t\\t\\\\centering\\n\\t\\t\\t\\t\\t\\\\begin{tabular}{l||>{\\\\centering\\\\arraybackslash}m{1.3cm}>{\\\\centering\\\\arraybackslash}m{0.5cm}>{\\\\centering\\\\arraybackslash}m{1.7cm}>{\\\\centering\\\\arraybackslash}m{1.2cm}>{\\\\centering}m{1.7cm}>{\\\\centering\\\\arraybackslash}m{1.2cm}>{\\\\centering}m{1cm}>{\\\\centering}m{1.2cm}>{\\\\centering\\\\arraybackslash}m{2.1cm}>{\\\\centering\\\\arraybackslash}m{1.7cm}>{\\\\centering\\\\arraybackslash}m{1.5cm}>{\\\\centering\\\\arraybackslash}m{1.7cm}}\\n\\t\\t\\t\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\t\\t\\t\\\\textbf{Attack}& \\\\textbf{Objective}&\\\\textbf{Class}& \\\\textbf{Level/IC Type/Attack Basis}&\\\\textbf{ML Model}&\\\\textbf{Benchmarks}\\\\textsuperscript{16}&\\\\textbf{Evaluated Schemes}&\\\\textbf{Scheme Independent}&\\\\textbf{Exact Output}\\\\textsuperscript{15}&\\\\textbf{Evaluate Key Length in Bits}&\\\\textbf{\\\\% Accuracy [min, max]}&\\\\textbf{Time Complexity\\\\textsuperscript{17}}&\\\\textbf{Known Protection}\\\\\\\\\\n\\t\\t\\t\\t\\t\\t\\\\midrule\\n\\t\\t\\t\\t\\t\\tSAIL~& key retrieval & OL&GL/C,S\\\\textsuperscript{1}/ST&Random Forest& ISCAS'85&RLL, SLL, CS& \\\\xmark&\\\\xmark&$\\\\{4,\\\\cdots,192\\\\}$\\\\textsuperscript{4}&$[66.89, 73.88]$&$\\\\bigO(l)$\\\\textsuperscript{13}& UNSAIL~, SARO~\\\\textsuperscript{22}\\\\\\\\\\\\hline\\n\\t\\t\\t\\t\\t\\tSnapShot~& ($i$) GSS and ($ii$) SRS key retrieval\\\\textsuperscript{14}& OL&GL/C,S/ST&MLP, CNN, GA& ISCAS'85, ITC'99, Ariane RV&RLL& \\\\cmark&\\\\xmark&$\\\\{64\\\\}$&$[57.71, 61.56]^{(i)}$, $[71.57, 81.67]^{(ii)}$&$\\\\bigO(l)$\\\\textsuperscript{13}& D-MUX~\\\\\\\\\\\\hline\\n\\t\\t\\t\\t\\t\\tGNNUnlock~&key gates removal&OL&GL/C,S/ST&GNN&ISCAS'85, ITC'99&Anti-SAT, TTLock, SFLL-HD&\\\\xmark\\\\textsuperscript{12}&\\\\xmark&$\\\\{8,\\\\cdots,128\\\\}$&$[100.00]$\\\\textsuperscript{11}&$\\\\bigO(n)$\\\\textsuperscript{13}&\\\\omark\\\\\\\\\\n\\t\\t\\t\\t\\t\\t\\\\midrule\\n\\t\\t\\t\\t\\t\\t\\\\midrule\\n\\t\\t\\t\\t\\t\\tBOCANet~& key retrieval & OG\\\\textsuperscript{2}&GL/C/FU&D-RNN \\\\& LSTM& Trust-Hub, ISCAS'85&RLL&\\\\cmark&\\\\xmark & $\\\\{32,64,128,256\\\\}$&$[89.00, 100.00]$&$\\\\bigO(\\\\alpha)$\\\\textsuperscript{19}&\\\\omark\\\\\\\\\\\\hline\\n\\t\\t\\t\\t\\t\\tSURF~&key retrieval& OG &GL/C,S\\\\textsuperscript{1}/ST,FU&SAIL \\\\& heuristic optimization&ISCAS'85&RLL, SLL, CS&\\\\xmark\\\\textsuperscript{3}&\\\\xmark&$\\\\{4,\\\\cdots,192\\\\}$\\\\textsuperscript{4} &$[90.58,98.83]$&$\\\\bigO(t)$\\\\textsuperscript{13}&UNSAIL~\\\\\\\\\\\\hline\\n\\t\\t\\t\\t\\t\\tGenUnlock~&key retrieval& OG &GL/C,S/FU&GA&ISCAS'85, MCNC&RLL,SLL, AND-OR, FLL&\\\\cmark&\\\\xmark&$\\\\{8,\\\\cdots, 1618\\\\}$\\\\textsuperscript{7}&$[100.00]$\\\\textsuperscript{8}&$\\\\bigO(\\\\beta)$\\\\textsuperscript{20}&\\\\omark\\\\\\\\\\\\hline\\n\\t\\t\\t\\t\\t\\tNNgSAT~&key retrieval& OG &GL/C,S/FU&MPNN&ISCAS'85, ITC'99&SAT-hard\\\\textsuperscript{5}&\\\\cmark&\\\\cmark&n/a\\\\textsuperscript{5}&$[93.50]$\\\\textsuperscript{6}&$\\\\bigO(\\\\lambda)$\\\\textsuperscript{18}&\\\\omark\\\\\\\\\\t\\\\hline\\n\\t\\t\\t\\t\\t\\tPSO Attack~&key retrieval& OG\\\\textsuperscript{9} &GL/C,S/FU&PSO&ISCAS'85, ITC'99&RLL, OLL&\\\\cmark&\\\\xmark&$\\\\{64,128\\\\}$&$[82.07,99.80]$\\\\textsuperscript{10}&$\\\\bigO(\\\\gamma)$\\\\textsuperscript{21}&point-function locking~\\\\\\\\\\n\\t\\t\\t\\t\\t\\t\\\\bottomrule\\n\\t\\t\\t\\t\\t\\\\end{tabular}\\n\\t\\t\\t\\t\\t\\\\begin{tablenotes}\\n\\t\\t\\t\\t\\t\\t\\\\scriptsize\\n\\t\\t\\t\\t\\t\\t\\\\item [1]~~The attack is in theory applicable to sequential circuits, however no evaluation has been performed yet.\\n\\t\\t\\t\\t\\t\\t\\\\item [2]~~The attack relies on having access to at least some golden input/output patterns ($<0.5\\\\%$ of total I/O pairs).\\n\\t\\t\\t\\t\\t\\t\\\\item [3]~~In theory, the key-refinement search algorithm could be utilized based on any seed key. However, this has not been addresses thus far.\\n\\t\\t\\t\\t\\t\\t\\\\item [4]~~The following key lengths have been evaluated: $\\\\{4,8,16,24,32,64,96,128,192\\\\}$.\\n\\t\\t\\t\\t\\t\\t\\\\item [5]~~$n\\\\times m$ bitwise multipliers $(8<m,n<32)$, $n\\\\times m$ crossbar network of 2-to-1 MUXes $(16<m, n<36)$, $n$-input LUTs built by 2-to-1 MUXes $(n<16)$, and $n$-to-1 AND-trees.\\n\\t\\t\\t\\t\\t\\t\\\\item [6]~~Indicates the percentage of successfully de-obfuscated circuits compared to the baseline~.\\n\\t\\t\\t\\t\\t\\t\\\\item [7]~~The key length is selected based on an area overhead of 5\\\\%, 10\\\\% or 25\\\\%.\\n\\t\\t\\t\\t\\t\\t\\\\item [8]~~The quality of the retrieved approximate keys is quantified by a user-defined output-fidelity measure.\\n\\t\\t\\t\\t\\t\\t\\\\item [9]~~The attack relies on an oracle without access to the scan chain.\\n\\t\\t\\t\\t\\t\\t\\\\item [10]~Accuracy refers to the average number of cases where the retrieved key results in 0\\\\% erroneous outputs for $10^{6}$ random patterns.\\n\\t\\t\\t\\t\\t\\t\\\\item [11]~The accuracy refers to the successful removal of the locking circuitry (not the key retrieval).\\n\\t\\t\\t\\t\\t\\t\\\\item [12]~The attack has not been evaluated for other locking schemes so far and the post-processing steps are scheme-specific.\\n\\t\\t\\t\\t\\t\\t\\\\item [13]~Notation: the key length $l$, the number of netlist nodes $n$, the total number of iterations $t=p\\\\cdot{n} + p\\\\cdot{w}+r\\\\cdot{l}\\\\cdot{i}\\\\cdot{n}+r\\\\cdot{l}\\\\cdot{i}\\\\cdot{p}$, where $p$ is the number of output pins, $i$ is the number of IO pairs, $r$ is the number of runs, and $w$ is the number of wires.\\n\\t\\t\\t\\t\\t\\t\\\\item [14]~GSS refers to the generalized set scenario which trains the ML model based on a set of locked benchmarks that are different from the target. SRS captures the self-referencing scenario where the training data is generated by re-locking the target benchmark.\\n\\t\\t\\t\\t\\t\\t\\\\item [15]~If the output of the attack is an exact result, the attack can guarantee a 100\\\\% correct deobfuscation for the complete I/O space.\\n\\t\\t\\t\\t\\t\\t\\\\item [16]~ISCAS'85~, MCNC~, ITC'99~, RISC-V Ariane core~, and Trust-Hub~.\\n\\t\\t\\t\\t\\t\\t\\\\item [17]~If the time complexity can be clearly determined, it refers to the time complexity after the training process.\\n\\t\\t\\t\\t\\t\\t\\\\item [18]~The execution time of the SAT attack is $\\\\sum_{i=1}^{\\\\lambda}{t_{i}}$, where $\\\\lambda$ is the number of iterations and $t_{i}$ the time required for one SAT-solver call. $\\\\lambda$ depends on the characteristics of the search space and the branching preferences of the SAT solver. Therefore, the time complexity of the attack is typically measured in terms of $\\\\lambda$. More details can be found in~.\\n\\t\\t\\t\\t\\t\\t\\\\item [19]~The complexity is linear to the number of training samples $\\\\alpha$, as the final key is determined based on the MSE of the trained outputs and the key-induced generated outputs.\\n\\t\\t\\t\\t\\t\\t\\\\item [20]~$\\\\beta=g\\\\cdot{p}\\\\cdot{l}$, where $g$ is the number of generations, $p$ is the population size, and $l$ is the key length. Note that the complexity changes with any adaptations of the GA.\\n\\t\\t\\t\\t\\t\\t\\\\item [21]~$\\\\gamma=g\\\\cdot{p}\\\\cdot{\\\\delta}$, where $g$ is the number of generations, $p$ is the population size, and $\\\\delta$ the complexity of performing circuit simulation for the fitness evaluation.\\n\\t\\t\\t\\t\\t\\t\\\\item [22]~An empirical evaluation of the resilience against SAIL has not been presented in the paper; only a discussion based on a proposed metric system has been provided.\\n\\t\\t\\t\\t\\t\\\\end{tablenotes}\\n\\t\\t\\t\\t\\\\end{threeparttable}\\n\\t\\t\\t}\\n\\t\\t\\t\\\\vspace{-0.1in}\\n\\t\\t\\\\end{table*}\",\n \"id\": \"50b37d47-0116-40d6-afbd-43bf9dacb32b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 29,\n \"parent_id\": \"8a41fbf8-901a-45b8-9e50-9a86985de29d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"ML-Based Attacks\"\n ],\n [\n \"subsubsection\",\n \"Oracle-Less Attacks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Oracle-Less Attacks\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 7636\n ],\n \"content\": \"Due to the availability of an oracle, existing ML-based attacks in this class have mostly exploited functional features of the target IC. This class includes the following attacks: BOCANet~, SURF~, GenUnlock~, NNgSAT~, and the PSO-guided attack~.\\n\\t\\t\\\\textbf{BOCANet:}~This attack leverages Recurrent Neural Networks (RNN) based on long short-term memory to construct a correct activation key (Fig.~\\\\ref{fig:attack-visualization}~(d)). The ML model is trained on a sequence of I/O observations taken from an activated IC, thereby learning the functional I/O mapping of the circuit, i.e., its Boolean function. Once trained, the key retrieval consists of two steps. First, a random key is applied to the model as input. Second, the initial key value is subsequently updated based on the Mean-Squared-Error (MSE) of the trained outputs and the newly generated outputs that are affected by the introduced key. Note that the ML model can be utilized to predict correct inputs or outputs as well. BOCANet exploits the functional effect a correct key has on generating a correct I/O mapping.\\n\\t\\t\\\\textbf{SURF:}~This attack is based on a joint structural and functional analysis of the circuit to retrieve the activation key (Fig.~\\\\ref{fig:attack-visualization}~(e)). The ML-aspect of the attack lies in \\\\textit{utilizing the SAIL attack} to generate the pre-resynthesis netlist structures and a seed key. Afterwards, based on the outputs of SAIL, SURF iteratively refines the key by means of a structure-aware greedy optimization algorithm guided by a functional simulation of the obfuscated netlist and a set of golden I/O pairs. The optimization is guided by the observation that specific key gates only affect a specific set of outputs. Thus, the key bits can be partitioned based on which output they affect. Performing a systematic perturbation of the key bits can lead to a more refined key. The success of the heuristic is grounded in the limited local effects that traditional locking schemes have on the value of the output for incorrect keys.\\n\\t\\t\\\\textbf{GenUnlock:}~The attack flow of GenUnlock leverages a Genetic Algorithm (GA)-based exploration of suitable activation keys for a locked circuit (Fig.~\\\\ref{fig:attack-visualization}~(f)). The heuristic search is steered by the key fitness that is computed based on the matching ratio of the key on the golden I/O training set. Through multiple generations, the fitness of the key population is subsequently improved through the application of genetic operators (selection, crossover, and mutation). Once the accepted tolerance for the correctness of the key is reached, the algorithm returns the set of the fittest keys. Similarly to SURF, GenUnlock exploits the fact that a heuristic key-refinement procedure eventually leads to more accurate activation keys. \\n\\t\\t\\\\textbf{NNgSAT:}~The main objective of NNgSAT is the deployment of a Message-Passing Neural Network (MPNN) to facilitate the resolution of SAT-hard circuit structures during the application of a SAT-based attack (Fig.~\\\\ref{fig:attack-visualization}~(g)). The motivation is driven by the fact that common SAT solvers run into scalability issues when tackling hard-to-be-solved locked circuit structures, e.g, multipliers and AND trees. Therefore, in NNgSAT, a neural network is trained to predict the satisfying assignment on a set of SAT-hard cases. In deployment, the SAT-attack flow offloads the SAT-hard problems to the trained model to speed up the attack procedure. The effectiveness of NNgSAT lies in the fact that it is possible to transfer prediction knowledge from learned clauses to unseen problems. \\n\\t\\t\\\\textbf{PSO-guided Attack:}~This attack is based on a Particle Swarm Optimization (PSO) heuristic that searches through the key space directed by a selected cost function (Fig.~\\\\ref{fig:attack-visualization}~(f)). The cost function is modeled as the Hamming distance between the golden and the obtained output responses (for a selected key). Therefore, the search algorithm relies on having access to an activated IC to compare against. A major motivator for this attack is its applicability without having access to an open scan chain, as this is often a limiting factor for SAT-based attacks. In essence, the PSO-guided attack and GenUnlock are similar in nature, as both rely on black-box evolutionary procedures guided by a functionality-driven objective function.\",\n \"id\": \"eda58d36-635e-44a7-9a87-c2f7c7682303\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"8a41fbf8-901a-45b8-9e50-9a86985de29d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"ML-Based Attacks\"\n ],\n [\n \"subsubsection\",\n \"Oracle-Guided Attacks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Oracle-Guided Attacks\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 2932,\n 2931\n ],\n \"content\": \"\\\\textbf{UNSAIL:}~This logic locking scheme has been developed to thwart attacks that target the resolution of structural transformations of logic synthesis~. The core idea of UNSAIL is to generate confusing training data that leads to false predictions in the CPM and RM modules of SAIL. This is realized through the additional manipulation of the netlist after synthesis to force the existence of \\\\textit{equivalent} netlist sub-graph observations that are linked to \\\\textit{different} key values.\\n\\t\\t\\\\textbf{SARO:}~The Scalable Attack-Resistant Obfuscation (SARO) operates in two steps~. First, SARO splits the design into smaller partitions to maximize the structural alterations in the netlist. Second, a systematic truth table transformation is deployed to lock the partitions. In order to increase the complexity of pattern recognition attacks (such as SAIL), the transformations aim to maximize randomness in the netlist.\\n\\t\\t\\\\textbf{Point-Functions and PSO:} As mentioned in Table~\\\\ref{table:attacks}, the PSO-guided attack is not applicable to point-function-based locking schemes. The reason is that this type of locking yields SAT-resilient behavior in which any incorrect key corrupts only \\\\textit{a very limited amount of outputs}. Consequently, this behavior offers no advantage in the guidance of the heuristic search, as it does not yield differentiating fitness values. Note that the applicability of point functions depends on the design of the fitness function that is used to guide the heuristic.\\n\\t\\t\\\\textbf{D-MUX:} The recently introduced Deceptive Multiplexer (D-MUX) LL scheme builds on the concept of multiple MUX-based insertion strategies that create structural paths that are equally likely to be driven by 0 or 1 key values~. Hence, D-MUX offers efficient protection against data-driven attacks.\",\n \"id\": \"b5e83d76-9451-4150-b043-b622bb4e368c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"cb342b15-09e2-41e4-9c1f-0d7558736604\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"ML-Resilient Schemes\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ML-Resilient Schemes\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\textbf{Deobfuscation Runtime Prediction:}~Apart from ML-based attacks, machine learning has also found its way into other aspects of logic locking. A recent work has designed a framework named ICNet for the prediction of the key-retrieval runtime for SAT-based attacks using graph neural networks~. The framework obtains the predicted deobfuscation runtime based on the characteristics of the circuit topology. ICNet offers an end-to-end approach to evaluate the hardness of logic locking with respect to the SAT attack, thereby increasing the development efficiency of novel locking schemes. \\n\\t\\t\\\\textbf{ML-Attack on Analog IC Locking:}~ML has started to have an impact on locking mechanisms even beyond digital circuits. The authors in~ have developed an oracle-guided attack on locked analog ICs using genetic algorithms. The approach has successfully broken all known analog logic locking techniques.\",\n \"id\": \"ca9a962e-8ebf-4b88-8027-88e803e66e7d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"cb342b15-09e2-41e4-9c1f-0d7558736604\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Other Applications of ML\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Other Applications of ML\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The presented efforts gather around two focal points: oracle-less and oracle-guided attacks. The intrinsic mechanisms of these attacks shed light on major vulnerabilities in existing logic locking that are exploitable by ML. We summarize the observations as follows:\",\n \"id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"cb342b15-09e2-41e4-9c1f-0d7558736604\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ]\n ],\n \"subsections\": [\n \"f1a22310-9774-4923-88e9-39031f0c57c0\",\n \"45fa94fe-3444-4d74-801c-6eb313e3eb65\",\n \"0f226960-767d-435e-ba81-0709f1f4a0d8\",\n \"058cfcba-944a-40ba-a6a6-3673887f8227\",\n \"7bee07b9-543e-4474-ab4e-8d349ee21c61\",\n \"4f146a54-7ec3-42e7-9ae0-dbd895f2b463\",\n \"71768a97-ef1a-4c84-ad10-da8f3f475652\",\n \"6c43d9c7-1dfb-4c58-8706-caf10292d0cd\"\n ],\n \"title\": \"Lessons Learned - Challenges and Opportunities\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Oracle-guided attacks focus on functional aspects of schemes, whereas structural leakage is exploited in the oracle-less model due to the absence of an activated IC. This indicates two pitfalls. First, the evaluated schemes have a predictable effect on the functionality of the circuit for incorrect keys, enabling the possibility to perform a guided heuristic search of the key space. Second, the existing schemes induce structural changes which strongly correlate with the value of the key. A mitigation depends on what an attempted attack tries to exploit. For example, to overcome SAIL or SnapShot, a scheme must not reveal anything about the correctness of the key through the induced structural change. This is achieved if the inserted change does not differ depending on the key value. Similarly, to protect against GNNUnlock, the deployed schemes have to ensure resilience against isolation, i.e., the locking circuitry must not be structurally independent from the original design. In the case of GenUnlock and the PSO attack, the behavior of the underlying scheme must be constant or fully random for all incorrect key values; disabling any chance for a guided heuristic convergence towards a correct key. Similar observations can be made for the other attacks as well. Nevertheless, conveying all necessary security objectives into a uniform locking scheme remains an open challenge.\",\n \"id\": \"f1a22310-9774-4923-88e9-39031f0c57c0\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"Structural vs. Functional Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Structural vs. Functional Analysis\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The reliance on logic synthesis transformations to enable the security of logic locking schemes has to be revised. As shown in SAIL, the synthesis rules are predictable and reversible, thereby having little impact on deducing a correct key for traditional schemes. Nevertheless, resynthesis can increase the difficulty to reverse engineering the functionality of a design. However, this needs further evaluation in the context of novel locking policies.\",\n \"id\": \"45fa94fe-3444-4d74-801c-6eb313e3eb65\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"Logic Synthesis for Security\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Logic Synthesis for Security\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"PSLL schemes induce isolable changes which create a clear structural distinction between the locking-related gates and the original (regular) gates. The main reason for this vulnerability is the requirement of achieving SAT-resistant behavior for incorrect keys. This specific functional pattern requires significant structural netlist changes consolidated in isolable components (e.g., the tree of complementary logic in Anti-SAT or the restore/perturb units in TTLock and SFLL-HD).\",\n \"id\": \"0f226960-767d-435e-ba81-0709f1f4a0d8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"Structural Isolation of Post-SAT Schemes\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Structural Isolation of Post-SAT Schemes\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Regardless of the attack type, ML-based attacks tend to generate an approximate output (as marked in Table~\\\\ref{table:attacks}), meaning that the retrieved key or deobfuscated circuit cannot be evaluated as correct with an exact certainty. Note that this is not the case for NNgSAT. Nevertheless, the approximate output can be utilized as seed for other attacks to speed up the reverse engineering procedure.\",\n \"id\": \"058cfcba-944a-40ba-a6a6-3673887f8227\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"Output Uncertainty\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Output Uncertainty\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As shown in Table~\\\\ref{table:attacks}, the existing attacks focus on extrapolating information leakage from gate-level netlists. So far, it is unclear if the same leakage is present in RTL-based locking schemes as well.\",\n \"id\": \"7bee07b9-543e-4474-ab4e-8d349ee21c61\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"RTL vs Gate-Level\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"RTL vs Gate-Level\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Available mitigation schemes suffer from overspecialization for thwarting specific attack vectors. So far, it has not been evaluated which form of logic locking offers a comprehensive protection against any form of ML-based guessing attack. Hence, a potentially fruitful opportunity lies within the automatic design of resilient schemes~.\",\n \"id\": \"4f146a54-7ec3-42e7-9ae0-dbd895f2b463\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"Overspecialization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Overspecialization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To the best of our knowledge, ML-based techniques have not been deployed yet for the evaluation of sequential logic locking techniques.\",\n \"id\": \"71768a97-ef1a-4c84-ad10-da8f3f475652\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"ML for Sequential Locking\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ML for Sequential Locking\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"There is still a need to have access to a rich benchmark set for data-driven analysis~.\",\n \"id\": \"6c43d9c7-1dfb-4c58-8706-caf10292d0cd\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"1d12de22-82f0-4f45-8eba-64bb646b3352\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Logic Locking in the Machine Learning Era\"\n ],\n [\n \"subsection\",\n \"Lessons Learned - Challenges and Opportunities\"\n ],\n [\n \"subsubsection\",\n \"The Need for Benchmarks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"The Need for Benchmarks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclusion}\\n\\t\\tThis work summarizes the recent developments of logic locking at the frontiers of machine learning. The presented ML-based attacks indicate the presence of structural and functional leakage in contemporary locking schemes which has been overlooked by the traditional interpretation of security. We show that an ML-based analysis is able to pinpoint novel vulnerabilities and challenge existing security assumptions. The offered discussion consolidates the major ML-induced challenges in logic locking design, offering a fruitful ground for future research directions in the post-ML era.\\n\\t\\t\\\\bibliographystyle{IEEEtran}\\n\\t\\t\\\\bibliography{bibliography_short}\\n\\t\\\\end{document}\",\n \"id\": \"4e814a9b-2e2f-4194-94e3-5e0b2adbe038\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"34e673cd-0e21-4b9c-bcb0-f8f8f8c633ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Logic Locking at the Frontiers of Machine Learning: A Survey on Developments and Opportunities\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":74,"string":"74"},"all_cites":{"kind":"list like","value":[8606,7636,2932,2931],"string":"[\n 8606,\n 7636,\n 2932,\n 2931\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.070896343956243,"string":"1.070896"}}},{"rowIdx":40,"cells":{"authors":{"kind":"list like","value":["Sicheng Zhao","Shangfei Wang","Mohammad Soleymani","Dhiraj Joshi","Qiang Ji"],"string":"[\n \"Sicheng Zhao\",\n \"Shangfei Wang\",\n \"Mohammad Soleymani\",\n \"Dhiraj Joshi\",\n \"Qiang Ji\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"},"year":{"kind":"string","value":"2019"},"date":{"kind":"timestamp","value":"2019-10-03T21:22:47Z","string":"2019-10-03T21:22:47Z"},"category":{"kind":"string","value":"cs.MM"},"abstract":{"kind":"string","value":" The wide popularity of digital photography and social networks has generated a rapidly growing volume of multimedia data (\\textit{i.e.}, image, music, and video), resulting in a great demand for managing, retrieving, and understanding these data. Affective computing (AC) of these data can help to understand human behaviors and enable wide applications. In this article, we survey the state-of-the-art AC technologies comprehensively for large-scale heterogeneous multimedia data. We begin this survey by introducing the typical emotion representation models from psychology that are widely employed in AC. We briefly describe the available datasets for evaluating AC algorithms. We then summarize and compare the representative methods on AC of different multimedia types, \\textit{i.e.}, images, music, videos, and multimodal data, with the focus on both handcrafted features-based methods and deep learning methods. Finally, we discuss some challenges and future directions for multimedia affective computing. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"]],"subsections":["6716f171-5279-4be6-829f-92b9ec4bb350","62f1d80f-e95a-45c5-ae8a-d6a71fc0fefa","21f5b1a5-8481-4ed1-83d5-a0396cc43cfa","50a2810e-2777-44d0-85b6-12b160fc2d4d","6a4a2789-5122-4ca2-8207-b0cf83ceda85","14b7a32f-91ab-4f67-baaf-fabb6149902f","20c98c17-ea71-4ba4-a9aa-815a42eced5b","5da3fa17-09f5-4d0c-9b55-5610d97ddda5","fc0b3d97-cde9-409b-91c4-fe25e8140878"],"title":"root"},{"cite_extract_rate":0.08333333333333301,"cites":[1016,1015],"content":"\\label{sec:Introduction}\nUsers are increasingly recording their daily activities, sharing interesting experiences, and expressing personal viewpoints using mobile devices on social networks, such as Twitter\\footnote{\\url{https://twitter.com}}, Facebook\\footnote{\\url{https://www.facebook.com}}, and Weibo\\footnote{\\url{https://www.weibo.com}}, \\textit{etc}. As a result, a rapidly growing volume of multimedia data (\\textit{i.e.}, image, music, and video) has been generated, as shown in Figure~\\ref{fig:multimedia}, which results in a great demand for the management, retrieval, and understanding of these data. Most existing work on multimedia analysis focus on the cognitive aspects, \\textit{i.e.}, understanding the objective content, such as object detection in images~, speaker recognition in speech~, and action recognition in videos~. Since what people feel have a direct influence on their decision making, affective computing (AC) of these multimedia data is of significant importance and has attracted increasing attention . For example, companies would like to know how customers evaluate their products and can thus improve their services~; depression and anxiety detection from social media can help understand psychological distress and thus potentially prevent suicidal actions~.\nWhile the sentiment analysis in text~ has long been a standard task, AC from other modalities, such as image and video, has just begun to be considered recently. In this article, we aim to review the existing AC technologies comprehensively for large-scale heterogeneous multimedia data, including image, music, video, and multimodal data.\nAffective computing of multimedia (ACM) aims to recognize the emotions that are expected to be evoked in viewers by a given stimuli. Similar to other supervised learning tasks, ACM is typically composed of three steps: data collection and annotation, feature extraction, and mapping learning between features and emotions~. One main challenge for ACM is the affective gap, \\textit{i.e.}, ``the lack of coincidence between the features and the expected affective state in which the user is brought by perceiving the signal''~. In the early stage, various hand-crafted features were designed to bridge this gap with traditional machine learning algorithms, while more recently researchers have focused on end-to-end deep learning from raw multimedia data to recognize emotions. Existing ACM methods mainly assign the dominant (average) emotion category (DEC) to an input stimuli, based on the assumption that different viewers have similar reactions to the same stimuli. We can usually formulate this task as a single-label learning problem.\nHowever, emotions are influenced by subjective and contextual factors, such as the educational background, cultural diversity, and social interaction~. As a result, different viewers may react differently to the same stimuli, which creates the subjective perception challenge. Therefore, the perception inconsistency makes it insufficient to simply predict the DEC for the highly subjective variable. As stated in~, we can perform two kinds of ACM tasks to deal with the subjectivity challenge: predicting personalized emotion perception for each viewer and assigning multiple emotion labels for each stimuli. For the latter one, we can either assign multiple labels to each stimuli with equal importance using multi-label learning methods, or predict the emotion distributions which tries to learn the degrees of each emotion~.\nIn this article, we concentrate on surveying the existing methods on ACM and analyzing potential research trends. Section~\\ref{sec:EmotionModels} introduces the widely-used emotion representation models from psychology. Section~\\ref{sec:Datasets} summarizes the existing available datasets for evaluating ACM tasks. Section~\\ref{sec:Image}, Section~\\ref{sec:Music}, Section~\\ref{sec:Video}, and Section~\\ref{sec:multimodal} survey the representative methods on AC of images, music, videos, and multimodal data, respectively, including both handcrafted features-based methods and deep learning methods. Section~\\ref{sec:FutureDirections} provides some suggestions for future research, followed by conclusion in Section~\\ref{sec:Conclusion}.\nTo the best of our knowledge, this article is among the first that provide a comprehensive survey of affective computing of multimedia data from different modalities. Previous surveys mainly focus on a single modality, such as images~, speech~, music~, video~, and multimodal data~. From this survey, readers can more easily compare the correlations and differences among different AC settings. We believe that this will be instrumental in generating novel research ideas.","id":"6716f171-5279-4be6-829f-92b9ec4bb350","level":"section","origin_cites_number":24,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:EmotionModels}\nThere are two dominant emotion representation models deployed by psychologists: categorical emotions (CE), and dimensional emotion space (DES). CE models classify emotions into a few basic categories, such as \\emph{happiness} and \\emph{anger}, \\emph{etc}. Some commonly used models include Ekman's six basic emotions and Mikels's eight emotions~. When classifying emotions into \\emph{positive} and \\emph{negative} (polarity)~, sometimes including \\emph{neutral}, ``emotion'' is called ``sentiment''. However, sentiment is usually defined as an atitude held toward an object . Emotions are usually represented by DES models as continuous coordinate points in a 3D or 2D Cartesian space, such as valence-arousal-dominance (VAD) and activity-temperature-weight~. VAD is the most widely used DES model, where valence represents the pleasantness ranging from positive to negative, arousal represents the intensity of emotion ranging from excited to calm, and dominance represents the degree of control ranging from controlled to in control. Dominance is difficult to measure and is often omitted, leading to the commonly used two dimensional VA space~.\nThe relationship between CE and DES and the transformation from one to the other are studied in~. For example, positive valence relates to a happy state, while negative valence relates to a sad or angry state. CE models are easier for users to understand and label, but the limited set of categories may not well reflect the subtlety and complexity of emotions. DES can better describe detailed emotions with subtle differences flexibly, but it is difficult for uses to distinguish the absolute continuous values, which may also be problematic. \nCE and DES are mainly employed in classification and regression tasks, respectively, with discrete and continuous emotion labels. If we discretize DES into several constant values, we can also use it for classification~. Ranking based labeling can be applied to ease DEC comprehension difficulties in raters.\nAlthough less explored in this context, one of the most well-known theories that explains the development of emotional experience is appraisal theory. According to this theory, cognitive evaluation or appraisal of a situation or content in case of multimedia results in emergence of emotions~. According to Ortony, Clore and Collins (OCC)~, emotions are experienced following a scenario comprising a series of phases. First, there is a perception of an event, object or an action. Then, there is an evaluation of events, objects or action according to personal wishes or norms. Finally, perception and evaluation result in a specific emotion or emotions arising. Certain appraisal dimensions such as novelty and complexity can be labeled and detected from content. For example, \\citeauthor{Soleymani2015} automatically recognized image novelty and complexity that are related to interestingness. There are also domain specific emotion taxonomy and scales. Geneva Emotional Music Scale is a music specific emotion model for describing emotions induced by music. It consists of a hierarchical structure with 45 emotions, nine emotional categories and three superfactors that can describe emotion in music.\n\\begin{table}[!t]\n\\centering\\scriptsize\n\\caption{Representative emotion models employed in ACM.}\n\\begin{tabular}\n{cccc}\n\\hline\n\\textbf{Model} & \\textbf{Ref} & \\textbf{Type} & \\multicolumn{1}{c}{\\textbf{Emotion states/dimensions}} \\\\\n\\hline\nEkman & & CE & happiness, sadness, anger, disgust, fear, surprise\\\\\nMikels & & CE & amusement, anger, awe, contentment, disgust, excitement, fear, sadness\\\\\nPlutchik & & CE & ($\\times$ 3 scales) anger, anticipation, disgust, joy, sadness, surprise, fear, trust\\\\\nClusters & & CE & 29 discrete labels are consistently grouped into 5 clusters at a similar distance level \\\\\nSentiment & & CE & positive, negative, (or neutral)\\\\\nVA(D) & & DES & valence-arousal(-dominance)\\\\\nATW & &DES & activity-temperature-weight\\\\\n\\hline\n\\end{tabular}\n\\label{tab:EmotionModels}\n\\end{table}\nAnother relevant concept worth mentioning is that emotion in response to multimedia can be expected, induced or perceived emotion. \nExpected emotion is the emotion that the multimedia creator intends to make people feel, perceived emotion is what people perceive as being expressed, while induced/felt emotion is the actual emotion that is felt by a viewer. Discussing the difference or correlation of various emotion models is out of the scope of this article. The typical emotion models that have been widely used in ACM are listed in in Table~\\ref{tab:EmotionModels}.","id":"62f1d80f-e95a-45c5-ae8a-d6a71fc0fefa","level":"section","origin_cites_number":14,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Emotion Models from Psychology"]],"subsections":[],"title":"Emotion Models from Psychology"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Datasets}","id":"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa","level":"section","origin_cites_number":0,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Datasets"]],"subsections":["31b031e5-4c1a-41d6-91e4-966f262d426b","6be4124c-0549-4c40-a1cf-f2efaaa9747b","d278bc85-2fef-47c9-b4d6-d1b6bb157aad","72c21e7b-1086-4048-bb25-1bc86a5c6b31"],"title":"Datasets"},{"cite_extract_rate":0.08333333333333301,"cites":[1017],"content":"\\label{sec:ImageDatasets}\nThe early datasets for AC of images mainly come from the psychology community with small-scale images. The International Affective Picture System (\\textbf{IAPS}) is an image set that is widely used in psychology to evoke emotions~. Each image that depicts complex scenes is associated with the mean and standard deviation (STD) of VAD ratings in a 9-point scale by about 100 college students. The \\textbf{IAPSa} dataset is selected from IAPS with 246 images~, which are labeled by 20 undergraduate students. The \\textbf{Abstract} dataset consists of 279 abstract paintings without contextual content. Approximately 230 people peer rated these paintings. The Artistic dataset (\\textbf{ArtPhoto}) includes 806 artistic photographs from a photo sharing site~ with emotions determined by the artist uploading the photos. The Geneva affective picture database (\\textbf{GAPED}) is composed of 520 negative, 121 positive, and 89 neutral images~. Besides, these images are also rated with valence and arousal values, ranging from 0 to 100 points. There are 500 abstract paintings in both \\textbf{MART} and \\textbf{devArt} datasets, which are collected from the Museum of Modern and Contemporary Art of Trento and Rovereto~, and the ``DeviantArt'' online social network~, respectively.\n\\begin{table*}[!t]\n\\centering\\scriptsize\n\\caption{Released and freely available datasets for AC of images, where `Ref' is short for Reference, `\\# Images' and `\\# Annotators' respectively represent the total number of images and annotators (f: female, m: male), `Labeling' represents the method to obtain labels, such as human annotation (annotation) and keyword searching (keyword), and `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).}\n\\begin{tabular}\n{cccccccc}\n\\hline\n\\textbf{Dataset} & \\textbf{Ref} & \\textbf{\\# Images} & \\textbf{Type} & \\textbf{\\# Annotators}& \\textbf{Emotion model} & \\textbf{Labeling} & \\textbf{Labels}\\\\\n\\hline\nIAPS & & 1,182 & natural & $\\approx$100 (half f) & VAD & annotation & average \\\\\nIAPSa & & 246 & natural & 20 (10f,10m) & Mikels & annotation & dominant \\\\\nAbstract & & 279 & abstract & $\\approx$230 & Mikels & annotation & dominant \\\\\nArtPhoto & & 806 & artistic & -- & Mikels & keyword & dominant \\\\\nGAPED & & 730 & natural & 60 & Sentiment, VA & annotation & dominant, average \\\\\nMART & & 500 & abstract & 25 (11f,14m) & Sentiment & annotation & dominant\\\\\ndevArt & & 500 & abstract & 60 (27f,33m) & Sentiment & annotation & dominant\\\\\nTweet & & 603 & social & 9 & Sentiment & annotation & dominant \\\\\nFlickrCC & & $\\approx$500,000 & social & -- & Plutchik & keyword & dominant \\\\\nFlickr & & 301,903 & social & 6,735 & Ekman & keyword & dominant \\\\\nEmotion6 & & 1,980 & social & 432 & Ekman+neutral & annotation & distribution \\\\\nFI & & 23,308 & social & 225 & Mikels & annotation & dominant \\\\\nIESN & & 1,012,901 & social & 118,035 & Mikels, VAD & keyword & personalized \\\\\nFlickrLDL & & 10,700 & social & 11 & Mikels & annotation & distribution \\\\\nTwitterLDL & & 10,045 & social & 8 & Mikels & annotation & distribution \\\\\n\\hline\n\\end{tabular}\n\\label{tab:ImageDataset}\n\\end{table*}\nRecent datasets, especially the large-scale ones, are constructed using images from social networks. The Tweet dataset (\\textbf{Tweet}) consists of 470 and 113 tweets for positive and negative sentiments, respectively~. The \\textbf{FlickrCC} dataset includes about 500k Flickr creative common (CC) images which are generated based on 1,553 adjective noun pairs (ANPs)~. The images are mapped to the Plutchnik's Wheel of Emotions with 8 basic emotions, each with 3 scales. The \\textbf{Flickr} dataset contains about 300k images~ with the emotion category defined by the synonym word list which has the most same words as the adjective words of an image's tags and comments. The \\textbf{FI} dataset consists of 23,308 images which are collected from Flicker and Instagram by searching the emotion keywords~ and labeled by 225 Amazon Mechanical Turk (MTurk) workers. The number of images in each Mikels emotion category is larger than 1,000. The \\textbf{Emotion6} dataset~ consists of 1,980 images collected from Flickr with 330 images for each Ekman's emotion category. Each image was scored by 15 MTurk workers to obtain the discrete emotion distribution information. The \\textbf{IESN} dataset that is constructed for personalized emotion prediction contains about 1M images from Flickr. Lexicon-based methods and VAD averaging~ are used to segment the text of metadata from uploaders for expected emotions and comments from viewers for personalized emotions. There are 7,723 active users with more than 50 involved images. We can also easily obtain the DEC and emotion distribution for each image. \\textbf{FlickrLDL} and \\textbf{TwitterLDL} datasets~ are constructed for discrete emotion distribution learning. The former one is a subset of FlickrCC, which are labeled by 11 viewers. The latter one consists of 10,045 images which are collected by searching various sentiment key words from Twitter and labeled by 8 viewers. These datasets are summarized in Table~\\ref{tab:ImageDataset}.","id":"31b031e5-4c1a-41d6-91e4-966f262d426b","level":"subsection","origin_cites_number":12,"parent_id":"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Datasets"],["subsection","Datasets for AC of Images"]],"subsections":[],"title":"Datasets for AC of Images"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:MusicDatasets}\nA notable benchmark for music recognition is music mood classification (AMC) task, organized by annual Music Information Retrieval Evaluation eXchange\\footnote{\\url{http://www.music-ir.org/mirex/wiki/}} (MIREX)~. In MIREX mood classification task, initially 600 songs were shared with the participants. Starting from 2013, 1,438 30 seconds excerpts from Korean pop songs have been added to MIREX. MIREX benchmark aims to automatically classify songs into five emotion clusters derived from cluster analysis of online tags. MIREX mood challenge emotional representation has been debated in the literature due to its data-driven origin rather than psychology of emotion. For example, in~, semantic and acoustic overlaps have been found between clusters. MIREX mood challenge considers only one label for the whole song and disregards the dynamic time evolving nature of music.\nComputer Audition Lab 500 (CAL500) is a dataset of 500 popular songs which is labeled by multiple tags including emotions~. The dataset is labeled in the lab by 66 labelers. Soundtracks datasets~ for music and emotion is developed by~\\citeauthor{Eerola2011} and contains instrumental music from soundtrack of 60 movies. The expert annotators selected songs based on five basic discrete categories (\\textit{anger}, \\textit{fear}, \\textit{sadness}, \\textit{happiness}, and \\textit{tenderness}) and dimensional VA representation of emotions. Although not developed with music content analysis in mind, the Database for Emotion Analysis using Physiological Signals or DEAP dataset~ also includes valence, arousal and dominance ratings for 120 one-minute music video clips of western pop music. Each video clip is annotated by by 14--16 participants who were asked to report their felt valence, arousal, and dominance on a 9-point scale. AMG1608~ is another music dataset that contains arousal and valence ratings for 1,608 Western songs in different genres and is annotated through MTurk.\nMusic datasets with emotion labels usually consider one emotion label per song (static). MoodSwings dataset~ was the first to annotate music dynamically over time. MoodSwings was developed by Schmidt \\etal and includes 240 15s excerpts of western pop songs with per-second valence and arousal labels, collected on MTurk. The MediaEval ``Emotion in Music'' challenge was organized in years 2013--2015 in MediaEval Multimedia Evaluation initiative\\footnote{http://www.multimediaeval.org}. MediaEval is a community-driven benchmarking campaign dedicated to evaluating algorithms for social and human-centered multimedia access and retrieval~. Unlike MIREX, ``Emotion in Music'' task focused on dynamic emotion recognition in music tracking arousal and valence over time~. The data from MediaEval tasks were compiled in MediaEval Database for Emotional Analysis in Music (DEAM) which is the largest available dataset with dynamic annotations, at 2Hz, with valence and arousal annotations for 1,802 songs and song excerpts licensed under Creative Commons license. PMEmo is a dataset of 794 songs with dynamic and static arousal and valence annotations in addition to electrodermal responses from ten participants~.\nThese datasets are summarized in Table~\\ref{tab:MusicDataset}. For a more detailed review of available music datasets with emotional labels, we refer the readers to~.\n\\begin{table*}[!t]\n\\centering\\scriptsize\n\\caption{Released and freely available datasets for music emotion recognition, where `\\# Songs' and `\\# Annotators' respectively represent the total number of songs and annotators per song, `Labeling' represents the method to obtain labels, such as human annotation (annotation), and `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).}\n\\begin{tabular}\n{cccccccc}\n\\hline\n\\textbf{Dataset} & \\textbf{Ref} & \\textbf{\\# Songs} & \\textbf{Type} & \\textbf{\\# Annotators}& \\textbf{Emotion model} & \\textbf{Labeling} & \\textbf{Labels}\\\\\n\\hline\nMIREX mood& & 2,038 & western and kpop & 2--3 & Clusters & annotation & dominant, distribution\\\\\nCAL500 & & 500 & western & >2 & -- & annotation & dominant\\\\\nSoundtracks& & 110 & instrumental &110 & self-defined, VA & annotation & distribution \\\\\nMoodSwings& & 240 & western & Unknown & VA & annotation & distribution\\\\\nDEAP & & 120 & western & 14--16 & VAD & annotation & average \\\\\nAMG1608& & 1,608 & western & 15 & VA & annotation & distribution\\\\\nDEAM & & 1,802 & diverse & 5--10 & VA & annotation & average, distribution \\\\\nPMEmo & & 794 & western & 10 & VA & annotation & distribution \\\\\n\\hline\n\\end{tabular}\n\\label{tab:MusicDataset}\n\\end{table*}","id":"6be4124c-0549-4c40-a1cf-f2efaaa9747b","level":"subsection","origin_cites_number":12,"parent_id":"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Datasets"],["subsection","Datasets for AC of Music"]],"subsections":[],"title":"Datasets for AC of Music"},{"cite_extract_rate":0,"cites":[],"content":"The target of video affective content computing is to recognize the emotions evoked by videos. In this field, it is necessary to construct a large benchmark dataset with precise emotional tags. However, the majority of existing research evaluate their proposed methods on their own collected datasets. The scarce video resources in those self-collected datasets, combined with the copyright restrictions result in limited accessibility for other researchers to reproduce existing work. Therefore, it is beneficial to summarize some publicly available datasets in this field. In general, publicly available datasets can be classified into two types: datasets consisting of videos only, such as movie clips or user generated videos, and datasets including both videos and audience's information.","id":"d278bc85-2fef-47c9-b4d6-d1b6bb157aad","level":"subsection","origin_cites_number":0,"parent_id":"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Datasets"],["subsection","Datasets for AC of Videos"]],"subsections":["d6fc556f-5306-4560-b61f-84603258d71a","7eba031f-c063-4798-bf20-621f78f3118a"],"title":"Datasets for AC of Videos"},{"cite_extract_rate":0.1,"cites":[1018],"content":"The LIRIS-ACCEDE dataset~ is one of the largest datasets in this area. Because it is collected under Creative Commons licenses, there are no copyright issues. The LIRIS-ACCEDE dataset is a living database in development. In order to fulfill requirements of different tasks, new data, features and tags are included. The LIRIS-ACCEDE dataset includes the Discrete LIRIS-ACCEDE collection and the Continuous LIRIS-ACCEDE collection in 2015 and was used for the MediaEval Emotional Impact of Movies tasks from 2015 to 2018.\nThe Discrete LIRIS-ACCEDE collection~ includes 9,800 clips, which is derived from 40 feature films and 120 short films. Specifically, the majority of the 160 films are collected from the video platform VODO. The duration of all 160 films is about 73.6 hours in total. All of the 9,800 video clips last about 27 hour in total and the duration of each clip is between 8 to 12 seconds, which is long enough for viewers to feel emotions.\nIn this collection, all the 9,800 video clips are labeled by values of valence and arousal.\nThe Continuous LIRIS-ACCEDE collection~ differs from the Discrete LIRIS-ACCEDE collection in annotation type. Roughly speaking, the annotations for movie clips in the Discrete LIRIS-ACCEDE collection are global. It means that a whole 8 to 12 second video clip is represented by a single value of valence and arousal. This annotation type limits the possiblity for tracking emotions. To address this issue, 30 longer films are selected from the 160 films mentioned above. The total duration of all the selected films is about 7.4 hours. There are emotional annotations according to valence and arousal of each second of the films in the collection.\nThe MediaEval Affective Impact of Movies collections between 2015 and 2018 are used for the MediaEval affective Impact of Movies tasks in each corresponding year. Specifically, the MediaEval 2015 Affective Impact of Movies~ includes two sub-tasks: affect detection and violence detection. The Discrete LIRIS-ACCEDE collection was used as the development set. And 1,100 additional video clips were extracted from 39 new movies and included. Indeed, all the new collected data were shared under Creative Commons licenses. In addition, three values were used to label the 10,900 video clips: a binary signal representing the presence of violence, a class tag of the excerpt for felt arousal and an annotation for felt valence.\nThe MediaEval 2016 Affective Impact of Movies Task~ also includes two sub-tasks: Global emotion prediction and Continuous emotion prediction. The Discrete LIRIS-ACCEDE collection and the Continuous LIRS-ACCEDE collection were used as the development sets for the first and second sub-tasks, respectively. In addition, 49 new movies were chosen as the test sets. 1,200 short video clips from the new movies were extracted for the first task, and 10 long movies were selected for the second task. For the first sub-task, the tags include scores of valence and arousal for each whole movie clip. And for the second sub-task, scores of valence and arousal for each second of the movies are evaluated.\n\\label{sec:VideoDatasets}\nThe MediaEval 2017 Affective Impact of Movies Task~ is focused on long movies for two sub-tasks: valence/arousal prediction and Fear prediction. The Continuous LIRIS-ACCEDE collection was selected as the development set, and an additional 14 new movies were collected as the test set. The annotations contain a valence value and an arousal value. In addition, there are a binary value to represent whether the segment is supposed to induce fear or not for each 10-second segment.\nThe MediaEval 2018 Affective Impact of Movies task~ is also dedicated to valence/arousal prediction and fear prediction. The Continuous LIRIS-ACCEDE collection and the test set of the MediaEval 2017 Emotional Impact of Movies task were used as the development set. In addition, 12 other movies selected from the set of the 160 movies mentioned in the Discrete LIRIS-ACCEDE part were used as test set. Specifically, for the first sub-task, there are annotations containing valence and arousal values for each second of the movies. And the beginning and ending times of each sequence in movies that induce fear are recorded for the second sub-task. \nThe VideoEmotion dataset~ is a well-designed user-generated video collection. It contains 1,101 videos downloaded from web platforms, such as YouTube and Flickr. The annotations of the videos in this dataset are based on Plutchik's wheel of emotions~.\nBoth the YF-E6 Dataset and the VideoStory-P14 Dataset are introduced in~. In order to collect the YF-E6 emotion dataset, six basic emotion types are used as keywords to search videos on YouTube and Flickr. There are 3,000 videos collected in the YF-E6 dataset totally. Then there were 10 annotators performing the labeling tasks. Only when all tags for a video clip were more than 50 percent consistent, the video clip was added to the dataset. Finally, the dataset includes 1,637 videos labeled with six basic emotion types. The VideoStory-P14 Dataset is based on the VideoStory dataset. Similar to the VideoEmotion Dataset, the keywords in Plutchik's Wheel of Emotions were used for the search process of the construction of the VideoStory dataset. Finally, there are 626 videos in the videoStory-P14 dataset with each having a unique emotion tag.","id":"d6fc556f-5306-4560-b61f-84603258d71a","level":"subsubsection","origin_cites_number":10,"parent_id":"d278bc85-2fef-47c9-b4d6-d1b6bb157aad","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Datasets"],["subsection","Datasets for AC of Videos"],["subsubsection","Datasets consisting of videos only"]],"subsections":[],"title":"Datasets consisting of videos only"},{"cite_extract_rate":0.07692307692307601,"cites":[1018],"content":"The DEAP dataset~ includes the EEG and peripheral physiological signals that are collected from 32 participants during watching 40 one-minute long excerpts of music videos. In addition, frontal face videos collected from 22 of the 32 participants are gathered. Annotators labeled each video according to the level of like/dislike, familiarity, arousal, valence, and dominance. Though the DEAP dataset is publicly available, it should be noted that it does not include the actual videos because of the licensing issues, but the links of videos are provided.\nThe MAHNOB-HCI~ is a multimodal dataset including multi-class information recorded in response to video affective stimuli. Particularly, speeches, face videos, and eye gazes are recorded. In addition, two experiments were conducted to record both peripheral and central nervous system physiological signals from 27 subjects. In the first experiment, subjects were assigned to report their emotional responses to 20 affective induced videos, including the level of arousal, valence and dominance, and predictability as well as emotion categories. In the second experiment, the participants evaluated whether they agreed with the displayed labels after watching short videos and images. The dataset is available for academic use through a web-interface.\n\\begin{table*}[!t]\n\\centering\\scriptsize\n \\caption{Released and freely available datasets for video emotion recognition, where `\\#Clips' and `Hours' respectively represent the total number and hours of video clips, `Type' means the genre of the videos in the dataset, `Emotion model' represents the labeling type, `Labeling' represents the method to obtain labels, such as human annotation (annotation) and keyword searching (keyword), and `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).}\n\\resizebox{\\textwidth}{!}{\n\\begin{tabular}\n{C{1.8 cm} cccccccc}\n\\hline\n\\textbf{Dataset} & \\textbf{Ref} & \\textbf{\\#Clips}& \\textbf{Hours} & \\textbf{Type} & \\textbf{\\# Annotators} & \\textbf{Emotion model} & \\textbf{Labeling} & \\textbf{Labels} \\\\\n\\hline\nDiscrete LIRIS-ACCEDE & & 9,800 & 26.9 & film & - & VA & annotation & dominant \\\\\nContinuous LIRIS-ACCEDE & & 30 & 7.4 & film & 10 (7f,3m) & VA & annotation & average \\\\\nMediaEval 2015 & & 1,100 & - & film & - & 3 discrete VA values & annotation & dominant \\\\\nMediaEval 2016 & & 1,210 & - & film & - & VA & annotation & distribution, average \\\\\nMediaEval 2017 & & 14 & 8 & film & - & VA, fear & annotation & average \\\\\nMediaEval 2018 & & 12 & 9 & film & - & VA, fear & annotation & average \\\\\nVideoEmotion & & 1,101 & 32.7 & user-generated & 10 (5f,5m) & Plutchik & annotation & dominant \\\\\nYF-E6 & & 1,637 & 50.9 & user-generated & 10(5f,5m) & Emkan & annotation & dominant \\\\\nVideoStory-P14 & & 626 & - & user-generated & - & Plutchik & keyword & dominant \\\\\nDEAP & & 120 & 2 & music video & - & VAD & annotation & personalized \\\\\nMAHNOB-HCI & & - & - & multiple types & - & VAD, Ekman+neutral & annotation & personalized \\\\\nDECAF & & 76 & - & music video/movies & - & VAD & annotation & personalized \\\\\nAMIGOS & & 20 & - & movies collection & - & VAD, Ekman & annotation & personalized \\\\\nASCERTAIN & & 36 & - & movies collection & 58 (21f,37m) & VA & annotation & personalized \\\\\n\\hline\n\\end{tabular}\n}\n\\label{tab:VideoDataset}\n\\end{table*}\nThe DECAF dataset~ consists of Infra-red facial video signals, Electrocardiogram (ECG), Magnetoencephalogram (MEG), horizontal Electrooculogram (hEOG) and Trapezius Electromyogram (tEMG), recorded from 30 participants watching 36 movie clips and 40 one-minute music videos, which are derived from the DEAP dataset~. The subjective feedback is based on valence, arousal, and dominance space. In addition, time-continuous emotion annotations for movie clips are also included in the dataset.\nThe AMIGOS dataset~ includes multi-class affective data, individual and groups of viewers' responses to both short and long videos. The EEG, ECG, GSR, frontal, and full body video were recorded in two experimental settings, \\textit{i.e.}, 40 participants watching 16 short emotional clips and 4 long clips. The duration of each selected short videos is between 51 and 150 seconds, and the duration of each long excerpt is about 20 minutes. Finally, participants annotated the affective level of valence, arousal, control, familiarity, liking, and basic emotions.\nBig-five personality scales and affective self-ratings of 58 users together with their EEG, ECG, GSR, and facial activity data were included in the ASCERTAIN dataset~ . The number of videos used as the stimulus is 36 and the length of each video clip is between 51 and 128 seconds. It is the first physiological dataset that is useful for both affective and personality recognition.\nThe publicly available datasets for video affective content analysis are summarized in Table~\\ref{tab:VideoDataset}.","id":"7eba031f-c063-4798-bf20-621f78f3118a","level":"subsubsection","origin_cites_number":13,"parent_id":"d278bc85-2fef-47c9-b4d6-d1b6bb157aad","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Datasets"],["subsection","Datasets for AC of Videos"],["subsubsection","Datasets including both videos and audience's reactions"]],"subsections":[],"title":"Datasets including both videos and audience's reactions"},{"cite_extract_rate":0.44444444444444403,"cites":[1019,1020,1021,7322],"content":"\\label{sec:multimodalDatasets}\nIn addition to audiovisual content and viewers' reactions, other modalities, such as language, also contain significant information for affective understanding of multimedia content. \nVisual sentiment is the sentiment associated with the concepts depicted in images. Two datasets were developed through mining images associated with the adjective-noun pair (ANP) representations that have affective significance~. ANPs in were generated by first using seed terms from Plutchik's Wheel of Emotion~ to query Flickr\\footnote{\\url{https://www.flickr.com}} and YouTube\\footnote{\\url{https://www.youtube.com/}}. After mining the tags associated with visual content on YouTube and Flickr, adjective and noun candidates were identified through part-of-speech tagging. Then adjective and nouns were paired to create ANP candidates which were filtered by sentiment strength, named entities, and popularity. The Visual Sentiment Ontology (VSO), \\footnote{\\url{https://visual-sentiment-ontology.appspot.com}}, is the results of this process. Sentibank resulted in the creation of a set of photo-tweet sentiment dataset, with both visual and textual data with polarity labels, collected on Amazon Mechanical Turk\\footnote{\\url{http://www.ee.columbia.edu/ln/dvmm/vso/download/twitter_dataset.html}}. This work was later extended to form a multilingual ANP set and its dataset, in~\\footnote{\\url{http://mvso.cs.columbia.edu}}, containing 15,630 ANPs from 12 major languages and 7.37M images~. My Reaction When (MRW) dataset contains 50,107 video-sentence pairs crawled from social media, depicting physical or emotional reactions to the situations described in sentences~. The GIFs are sourced from Giphy\\footnote{\\url{https://giphy.com/}}. Even though there is no emotional labels, the language and visual associations are mainly based on sentiment which makes this dataset an interesting resource for affective content analysis.\nCMU Multimodal Opinion Sentiment and Emotion Intensity (CMU-MOSEI) is a collection of multiple datasets for multimodal sentiment analysis and emotion recognition. This collection includes more than 23,500 sentence utterance videos from more than 1,000 people from YouTube~\\footnote{\\url{https://github.com/A2Zadeh/CMU-MultimodalSDK}}. All the videos are transcribed and aligned with audiovisual modalities. \nA multimodal multi-party dataset for emotion recognition in conversation (MELD) was primarily developed for emotion recognition in multiparty interaction purposes~. MELD contains visual, audio, and textual modalities and includes 13,000 utterances from 1,433 dialogues from the TV-series Friends, with each utterance labeled with emotion and sentiment. \n\\begin{table*}[!t]\n\\centering\\scriptsize\n \\caption{Released and freely available datasets for multimodal multimodal emotion recognition. Disc. for MELD corresponds to six Ekman emotions in addition to neutral. `Labeling' represents the method to obtain labels, such as human annotation (annotation), self-reported felt emotion and keyword searching (keyword), `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).} \n\\resizebox{\\textwidth}{!}{\n\\begin{tabular}\n{c c c c c c c c}\n\\hline\n\\textbf{Dataset} & \\textbf{Ref} & \\textbf{\\#Samples}& \\textbf{Modalities} & \\textbf{Type} & \\textbf{Emotion model} & \\textbf{Labeling}&\\textbf{Labels}\\\\\n\\hline\nSentiBank tweet & & 603 & images, text & images & Sentiment&annotation& dominant\\\\\nMVSO & & 7.36M & image, metadata & photos & Sentiment &automatic & average\\\\\nCMU-MOSEI & & 23,500 & video, audio, text & YouTube videos & Sentiment &annotation& average\\\\\nMELD & & 13,000 & video, audio, text & TV series & Sentiment, Disc. &annotation& dominant\\\\\nCOGNIMUSE & & 3.5h & video, audio, text & movies & VA&annotation, self-report& dominant\\\\\nVR & & 73 & video, audio & VR videos & VA&self-report&average \\\\\n\\hline\n\\end{tabular}\n}\n\\label{tab:MMDataset}\n\\end{table*}\nCOGNINMUSE is a collection of videos annotated with sensory and semantic saliency, events, cross-media semantics, and emotions~. A subset of 3.5h extracted from movies, including textual modality, are annotated on arousal and valence. \n\\citeauthor{Li_VR} collected a dataset of 360 degrees virtual reality videos that can elicit different emotions~. Even though the dataset consists of 73 short videos, on average 183s long, it is one of the first datasets of its kind whose content understanding stays limited. These multimodal datasets are summarized in Table~\\ref{tab:MMDataset}.","id":"72c21e7b-1086-4048-bb25-1bc86a5c6b31","level":"subsection","origin_cites_number":9,"parent_id":"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Datasets"],["subsection","Datasets for AC of Multimodal Data"]],"subsections":[],"title":"Datasets for AC of Multimodal Data"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Image}\nIn the early stages, AC researchers mainly worked on designing handcrafted features to bridge the affective gap. Recently, with the advent of deep learning especially convolutional neural networks (CNNs), current methods have shifted to an end-to-end deep representation learning. Motivated by the fact that the perception of image emotions may be dependent on different types of features~, some methods employ fusion strategies to jointly consider multiple features. In this section, we summarize and compare these methods. Please note that here we classify the directly extracted CNN features based on pre-trained deep models into handcrafted features category.","id":"50a2810e-2777-44d0-85b6-12b160fc2d4d","level":"section","origin_cites_number":1,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Images"]],"subsections":["6469fc7b-8c4d-423d-88c3-f522313e7cbb","c7360062-0ae0-4488-a262-c17a829a8a11"],"title":"Affective Computing of Images"},{"cite_extract_rate":0.043478260869565,"cites":[1022],"content":"\\label{sec:ImageHandcrafted}\n\\textbf{Low-level Features} are difficult to be understood by viewers. These features are often directly derived from other computer vision tasks. Some widely extracted features include GIST, HOG2x2, self-similarity and geometric context color histogram features as in~, because of their individual power and distinct description of visual phenomena in a scene perspective.\nCompared with the above generic features, some specific features derived from art theory and psychology have been designed. For example, \\citeauthor{machajdik2010affective}~\\shortcite{machajdik2010affective} extracted elements-of-art features, including \\emph{color} and \\emph{texture}. The MPEG-7 visual descriptors are employed in~, which include four color-related ideas and two texture-related ideas. How shape features in natural images influence emotions is investigated in~ by modeling the concepts of roundness-angularity and simplicity-complexity. \\citeauthor{sartori2015s}~\\shortcite{sartori2015s} designed two kinds of visual features to represent different color combinations based on Itten's color wheel.\n\\textbf{Mid-level Features} contain more semantics, are more easily interpreted by viewers than low-level features, and thus are more relevant to emotions. \\citeauthor{patterson2012sun}~\\shortcite{patterson2012sun} proposed to detect 102 attributes in 5 different categories, including materials, surface properties, functions or affordances, spatial envelop attributes, and object presence. Besides these attributes, eigenfaces that may contribute to facial images are also incorporated in~. More recently, in~, SIFT features are first extracted as basic features, which are fed into bag-of-visual-words (BoVW) to represent the multi-scale blocks. Another mid-level representation is the latent topic distribution estimated by probabilistic latent semantic analysis.\nHarmonious composition is essential in an artwork. Several compositional features, such as low depth of field, are designed to analyze such characteristics of an image~. Based on the fact that figure-ground relationships, color patterns, shapes and their diverse combinations are often jointly employed by artists to express emotions in their artworks, \\citeauthor{wang2013interpretable}~\\shortcite{wang2013interpretable} proposed to extract interpretable aesthetic features.\nInspired by princiles-of-art, \\citeauthor{zhao2014exploring}~\\shortcite{zhao2014exploring} designed corresponding mid-level features, including \\emph{balance}, \\emph{emphasis}, \\emph{harmony}, \\emph{variety}, \\emph{gradation}, and \\emph{movement}. For example, Itten's color contrasts and the rate of focused attention are employed to measure \\textit{emphasis}.\n\\begin{table*}\n\\centering\\scriptsize\n\\caption{Summary of the hand-crafted features at different levels for AC of images. `\\# Feat' indicates the dimension of each feature.}\n\\resizebox{\\textwidth}{!}{\n\\begin{tabular}{ccc p{8cm} c}\n\\hline\n\\textbf{Feature} & \\textbf{Ref} & \\textbf{Level} & \\multicolumn{1}{c}{\\textbf{Short description}} & \\textbf{\\# Feat} \\\\\n\\hline\nLOW\\_C & & low & GIST, HOG2x2, self-similarity and geometric context color histogram features & 17,032\\\\\nElements & & low & color: mean saturation, brightness and hue, emotional coordinates, colorfulness, color names, Itten contrast, Wang's semantic descriptions of colors, area statistics; texture: Tamura, Wavelet and gray-level co-occurrence matrix & 97\\\\\nMPEG-7 & & low & color: layout, structure, scalable color, dominant color; texture: edge histogram, texture browsing & $\\approx$200 \\\\\nShape & & low & line segments, continuous lines, angles, curves & 219\\\\\nIttenColor & & low & color co-occurrence features and patch-based color-combination features & 16,485\\\\\n\\hline\nAttributes & & mid & scene attributes & 102 \\\\\nSentributes & & mid & scene attributes, eigenfaces & 109 \\\\\nComposition & & mid & level of detail, low depth of field, dynamics, rule of thirds & 45 \\\\\nAesthetics & & mid & figure-ground relationship, color pattern, shape, composition & 13 \\\\\nPrinciples & & mid & principles-of-art: balance, contrast, harmony, variety, gradation, movement & 165\\\\\nBoVW & & mid & bag-of-visual-words on SIFT, latent topics & 330 \\\\\n\\hline\nFS & & high & number of faces and skin pixels, size of the biggest face, amount of skin w.r.t. the size of faces & 4 \\\\\nANP & & high & semantic concepts based on adjective noun pairs & 1,200 \\\\\nExpressions & & high & automatically assessed facial expressions (anger, contempt, disgust, fear, happiness, sadness, surprise, neutral) & 8 \\\\\n\\hline\n\\end{tabular}\n}\n\\label{tab:ImageHandCraftedFeatures}\n\\end{table*}\n\\textbf{High-level Features} that represent the semantic content contained in images can be easily understood by viewers. We can also well recognize the conveyed emotions in images through these semantics. In the early years, simple semantic content including faces and skins contained in images are extracted in~. For the images that contain faces, facial expressions may directly determine the emotions. \\citeauthor{yang2010exploring}~\\shortcite{yang2010exploring} extracted 8 kinds of facial expressions as high-level features. They built compositional features of local Haar appearances by a minimum error based optimization strategy, which are embedded into an improved AdaBoost algorithm. For the images detected without faces, the experessions are simply set as \\emph{neutral}. Finally, they generated a 8 dimensional vector with each element representing the number of corresponding facial expressions.\n\\begin{table*}[!t]\n\\centering\\scriptsize\n\\caption{Representative work on AC of images using hand-crafted features, where `Fusion' indicates the fusion strategy of different features, `cla, reg, ret, cla\\_p, dis\\_d, dis\\_c' in the Task column are short for classification, regression, retrieval, personalized classification, discrete distribution learning, continuous distribution learning (the same below), respectively, `Result' is the reported best accuracy for classification, mean squared error for regression, discounted cumulative gain for retrieval, F1 for personalized classification, and KL divergence for distribution learning (the first line~ is the result on sum of squared difference) on the corresponding datasets.}\n\\begin{tabular}\n{l p{2.5cm} l p{1.5cm} p{2.5cm} l R{2.5cm}} \n\\hline\n\\textbf{Ref} & \\textbf{Feature} & \\textbf{Fusion} & \\textbf{Learning} & \\textbf{Dataset} & \\textbf{Task} & \\textbf{Result} \\\\\n\\hline\n & Elements, Composition, FS & early & NB & IAPSa, Abstract, ArtPhoto & cla & 0.471, 0.357, 0.495\\\\\n & MPEG-7 & -- & KNN & unreleased & cla & 0.827\\\\\n & Shape, Elements & early & SVM, SVR & IAPSa; IAPS & cla; reg & 0.314; V-1.350, A-0.912\\\\%shape:0.299, V: 1.708, A: 0.943\\\\\n & Segmented objects & -- & SL & IAPS, ArtPhoto & cla & 0.612, 0.610\\\\\n & Sentributes & -- & SVM, LR & Tweet & cla & 0.824\\\\\n & Aesthetics & -- & NB & Abstract, ArtPhoto & cla & 0.726, 0.631\\\\\n & Principles & -- & SVM, SVR & IAPSa, Abstract, ArtPhoto; IAPS & cla; reg & 0.635, 0.605, 0.669; V-1.270, A-0.820\\\\\n & LOW\\_C, Elements, Attributes, Principles, ANP, Expressions & graph & MGL & IAPSa, Abstract, ArtPhoto, GAPED, Tweet & ret & 0.773, 0.735, 0.658, 0.811, 0.701\\\\\n & IttenColor & -- & SL & MART, devArt & cla & 0.751, 0.745\\\\\n & BoVW & -- & MIL & IAPSa, Abstract, ArtPhoto & cla & 0.699, 0.636, 0.707\\\\\n & IttenColor & -- & MC & MART, devArt & cla & 0.728, 0.761\\\\\n\\hline\n & GIST, Elements, Attributes, Principles, ANP, Expressions & graph & RMTHG & IESN & cla\\_p & 0.582\\\\\n\\hline\n & GIST, Elements, Principles & - & SSL & Abstract & dis\\_d & 0.134\\\\\n & GIST, Elements, Attributes, Principles, ANP, deep features from AlexNet & weighted & WMMSSL & Abstract, Emotion6, IESN & dis\\_d & 0.482, 0.479, 0.478 \\\\\n & ANP, VGG16 & - & ACPNN & Abstract, Emotion6, FlickrLDL, TwitterLDL & dis\\_d & 0.480, 0,506, 0,469, 0.555\\\\\n & GIST, Elements, Attributes, Principles, ANP, AlexNet & weighted & WMMCPNN & Abstract, Emotion6, IESN & dis\\_d & 0.461, 0.464, 0.470\\\\\n & GIST, Elements, Attributes, Principles, ANP, AlexNet & -- & MTSSR & IESN & dis\\_c & 0.436\\\\\n\\hline\n\\end{tabular}\n\\label{tab:ImageHandCraftedMethods}\n\\end{table*}\nMore recently, the semantic concepts are described by adjective noun pairs (ANPs)~, which are detected by SentiBank~ or DeepSentiBank~. The advantages of ANP are that it turns a neutral noun into an ANP with strong emotions and makes the concepts more detectable, compared to nouns and adjectives, individually. A 1,200 dimensional vector representing the probability of the ANPs can form a feature vector.\nTable~\\ref{tab:ImageHandCraftedFeatures} summarizes the above-mentioned hand-crafted features at different levels for AC of images. Some recent methods also extracted CNN features from pre-trained deep models, such as AlexNet~ and VGGNet~.\nTo map the extracted handcrafted features to emotions,\n\\textbf{Machine Learning Methods} are commonly employed. Some typical learning models include Naive Bayes (NB)~, support vector machine (SVM)~, $K$ nearest neighbor (KNN)~, sparse learning (SL)~, logistic regression (LR)~, multiple instance learning (MIL)~, and matrix completion (MC)~ for emotion classification , support vector regression (SVR)~ for emotion regression, and multi-graph learning (MGL)~ for emotion retrieval.\nInstead of assigning the DEC to an image, some recent methods began to focus on the perception subjectivity challenge, \\textit{i.e.}, predicting personalized emotions for each viewer or learning emotion distributions for each image. The personalized emotion perceptions of a specified user after viewing an image is predicted in~, associated with online social networks. They considered different types of factors that may contribute to emotion recognition, including the images' visual content, the social context related to the corresponding users, the emotions' temporal evolution, and the images' location information. To jointly model these factors, they proposed rolling multi-task hypergraph learning (RMTHG), which can also easily hanlde the data incompleteness issue.\nGenerally, the distribution learning task can be formulated as a regression problem, which slightly differs for different distribution categories (\\textit{i.e.}, discrete or continuous). For example, if emotion is represented by CE, the regression problem targets predicting the discrete probability of each emotion category with the sum equal to 1; if we represent emotion based on DES, the regression problem is typically transformed to the prediction of the parameters of specified continuous probability distributions. For the latter one, we usually need to firstly determine the form of continuous distributions, such as exponential distribution and Gaussian distribution. Some representative learning methods for emotion distribution learning of discrete emotions include shared sparse learning (SSL)~, weighted multimodal SSL (WMMSSL) , augmented conditional probability neural network (ACPNN)~, and weighted multi-model CPNN (WMMCPNN)~. Both SSL and WMMSSL can only model one test image each time, which is computationally inefficient. After the parameters are learned, ACPNN and WMMCPNN can easily predict the emotion distributions of a test image. Based on the assumption that the VA emotion labels can be well modeled by a mixture of 2 bidimensional Gaussian mixture models (GMMs), \\citeauthor{zhao2017continuous}~\\shortcite{zhao2017continuous} proposed to learn continuous emotion distributions in VA space by multi-task shared sparse regression (MTSSR). Specifically, the parameters of GMMs are regressed, including the mean vector and covariance matrix of the 2 Gaussian components as well as the mixing coefficients.\nTable~\\ref{tab:ImageHandCraftedMethods} summarizes some representative work based on hand-crafted features. Generally, high-level features (such as ANP) can achieve better recognition performance for images with rich semantics, mid-level features (such as Principles) are more effective for artistic photos, while low-level features (such as Elements) perform better for abstract paintings.","id":"6469fc7b-8c4d-423d-88c3-f522313e7cbb","level":"subsection","origin_cites_number":23,"parent_id":"50a2810e-2777-44d0-85b6-12b160fc2d4d","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Images"],["subsection","Handcrafted Features-Based Methods for AC of Images"]],"subsections":[],"title":"Handcrafted Features-Based Methods for AC of Images"},{"cite_extract_rate":0.27272727272727204,"cites":[1023,1017,1024],"content":"\\label{sec:ImageDeep}\nTo deal with the situation where images are weakly labeled, a potentially cleaner subset of the training instances are selected progressively~. First, they trained an initial CNN model based on the training data. Second, they selected the training samples with distinct sentiment scores between the two classes with a high probability based on the prediction score of the trained model on the training data itself. Finally, the pre-trained AlexNet on ImageNet is fine-tuned to classify emotions into 8 categories by changing the last layer of the CNN from 1000 to 8~. Besides using the fully connected layer as classifier, they also trained an SVM classifier based on the extracted features from the second to the last layer of the pre-trained AlexNet model.\nMulti-level deep representations (MldrNet) are learned in~ for image emotion classification. They segmented the input image into 3 levels of patches, which are input to 3 different CNN models, including Alexnet, aesthetics CNN (ACNN), and texture CNN (TCNN). The fused features are fed into multiple instance learning (MIL) to obtain the emotion labels. \\citeauthor{zhu2017dependency}~\\shortcite{zhu2017dependency} proposed to integrate the different levels of features by a Bidirectional GRU model (BiGRU) to exploit their dependencies based on MldrNet. They generated two features from the Bi-GRU model and concatenated them as the final feature representations. To enforce the feature vectors extracted from each pair of images from the same category to be close enough, and those from different categories to be far away, they proposed to jointly optimize a contrastive loss together with the traditional cross-entropy loss.\nMore recently, \\citeauthor{yang2018retrieving}~\\shortcite{yang2018retrieving} employed deep metric learning to explore the correlation of emotional labels with the same polarity, and proposed a multi-task deep framework to optimize both retrieval and classification tasks. By considering the relations among emotional categories in the Mikels' wheel, they jointly optimized a novel sentiment constraint with the cross-entropy loss. Extending triplet constraints to a hierarchical structure, the sentiment constraint employs a sentiment vector based on the texture information from the convolutional layer to measure the difference between affective images. In~, \\citeauthor{yang2018retrieving} proposed a weakley supervised coupled convolutional neural network to exploit the discriminability of localized regions for emotion classification. Based on the image-level labels, a sentiment map is firstly detected in one branch with the cross spatial pooling strategy. And then the holistic and localized information are jointly combined in the other branch to conduct a classification task. The detected sentiment map can easily explain which regions of an image determine the emotions.\nThe above deep methods mainly focused on the dominant emotion prediction. There are also some work on emotion distribution learning based on deep models. The very first work is a mixed bag of emotions, which trains a deep CNN regressor (CNNR) for each emotion category in Emotion6~ based on the AlexNet architecture. They changed the number of output nodes to 1 to predict a real value for each emotion category and replaced the Softmax loss with Euclidean loss. To ensure the sum of different probabilities to be 1, they normalized the predicted probabilities of all emotion categories. However, CNNR has some limitations. First, the predicted probability cannot be guaranteed to be non-negative. Second, the probability correlations among different emotions are ignored, since the regressor for each emotion category is trained independently. In~, \\citeauthor{yang2017joint} designed a multi-task deep framework based on VGG16 by jointly optimizing the cross-entropy loss for emotion classification and Kullback-Leibler (KL) divergence loss for emotion distribution learning. To match the single emotion dataset to emotion distribution learning settings, they transformed each single label to emotion distribution with emotion distances computed on Mikels' wheel~. By extending the size of training samples, this method achieves the state-of-the-art performance for discrete emotion distribution learning.\n\\begin{table*}[!t]\n\\centering\\scriptsize\n\\caption{Representative work on deep learning based AC of images, where `Pre' indicates whether the network is pre-trained using ImageNet, `\\# Feat' indicates the dimension of last feature mapping layer before the emotion output layer, `Cla' indicates the classifier used after the last feature mapping with default Softmax, `Loss' indicates the loss objectives (besides the common cross-entropy loss for classification), and `Result' is the reported best accuracy for classification, discounted cumulative gain for retrieval, and KL divergence for distribution learning on the corresponding datasets.}\n\\resizebox{\\textwidth}{!}{\n\\begin{tabular}\n{c C{1.5cm} c C{1cm} cc C{3cm} C{0.4cm} C{2cm}} \n\\hline\n\\textbf{Ref} & \\textbf{Base net} & \\textbf{Pre} & \\textbf{\\#Feat} & \\textbf{Cla} & \\textbf{Loss} & \\textbf{Dataset} & \\textbf{Task} & \\textbf{Result} \\\\\n\\hline\n & self-defined & no & 24 & -- & -- & FlickrCC & cla & 0.781\\\\\n & AlexNet & yes & 4,096 & SVM & -- & FI, IAPSa, Abstract, ArtPhoto & cla & 0.583, 0.872, 0.776, 0.737\\\\\n & AlexNet, ACNN, TCNN & yes & 4,096, 256, 4,096 & MIL & -- & \\tiny{FI, IAPSa, Abstract, ArtPhoto, MART} & cla & \\tiny{0.652, 0.889, 0.825, 0.834, 0.764}\\\\\n & self-defined & no & 512 & -- & contrastive & FI, IAPSa, ArtPhoto & cla & 0.730, 0.902, 0.855\\\\\n & GoogleNet-Inception & yes & 1,024 & -- & sentiment & FI, IAPSa, Abstract, ArtPhoto & cla; ret & 0.676, 0.442, 0.382, 0.400; 0.780, 0.819, 0.788, 0.704\\\\\n & ResNet-101 & yes & 2,048 & -- & -- & FI, Tweet & cla & 0.701, 0.814\\\\\n\\hline\n & AlexNet & yes & 4,096 & -- & Euclidean & Emotion6 & dis\\_d & 0.480\\\\\n & VGG16 & yes & 4,096 & -- & KL & Emotion6, FlickrLDL, TwitterLDL & dis\\_d & 0.420, 0,530, 0,530\\\\\n\\hline\n\\end{tabular}\n}\n\\label{tab:ImageDeepMethods}\n\\end{table*}\nThe representative deep learning based methods are summarized in Table~\\ref{tab:ImageDeepMethods}. The deep representation features generally perform better than the hand-crafted ones, which are intuitively designed for specific domains based on several small-scale datasets. However, how the deep features correlate to specific emotions is unclear.","id":"c7360062-0ae0-4488-a262-c17a829a8a11","level":"subsection","origin_cites_number":11,"parent_id":"50a2810e-2777-44d0-85b6-12b160fc2d4d","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Images"],["subsection","Deep Learning-Based Methods for AC of Images"]],"subsections":[],"title":"Deep Learning-Based Methods for AC of Images"},{"cite_extract_rate":0.0625,"cites":[1025],"content":"\\label{sec:Music}\nMusic emotion recognition (MER) strives to identify emotion expressed by music and subsequently predict listener's felt emotion from acoustic content and music metadata, \\textit{e.g.}, lyrics, genre, \\textit{etc.} Emotional understanding of music have applications in music recommendation and is particularly useful for producing music retrieval. An analysis of search queries from creative professionals showed that 80\\% contain emotional terms, showing emotions prominence in that field~. A growing number of work have tried to address emotional understanding of music from acoustic content and metadata (see~ for earlier reviews on this topic). \nEarlier work on emotion recognition from music relied on extracting acoustic features similar to the ones used in speech analysis, such as audio energy and formants. Acoustic features describe attributes related to musical dimensions. Musical dimensions include melody, harmony, rhythm, dynamics, timbre (tone color), expressive techniques, musical texture, and musical form~, as shown in Table~\\ref{tab:MusicFeatures}. Some also add energy as a musical feature which is important for MER~. \nMelody is a linear succession of tones and can be captured by features representing key, pitch and tonality. Among others, chroma is often used to represent melodic features~. \nHarmony is how the combination of various pitches are processed during hearing. Understanding harmony involves chords or multiple notes played together. Examples of acoustic features capturing harmony include chromagram, key, mode, and chords~.\nRhythm consists of repeated patterns of musical sounds, \\textit{i.e.}, notes and pulses that can be describes in terms of tempo and meter. Higher tempo songs often induce higher arousal and fluent rhythm is associated with higher valence and firm rhythm is associated with sad songs~. Mid-level acoustic features, such as onset rate, tempo and beat histogram, can represent rhythmic characteristics of music. \nDynamics of music involve the variation in softness or loudness of notes which include change of loudness (contrast) and emphasis on individual sounds (accent)~. Dynamics of music can be captured by changes in acoustic features related to energy such as root mean square (RMS) energy.\nTimbre is the perceived sound quality of musical notes. Timbre is what differentiates different voices and instruments playing the same sound. Acoustic features capturing timbre, such as MFCC and spectrum shape, describe sound quality~. Acoustic features describing timbre include MFCC, spectral features (centroid, contract, flatness), and zero crossing rate~.\nExpressive techniques are the way a musical piece is played including tempo and articulation~. Acoustic features, such as tempo, attack slope, and time, can be used to describe this dimension.\nMusical texture is how rhythmic, melodic, and harmonic features are combined in music production~. It is related to the range of tones played at the same time. Musical form describes how a song is structured, such as introduction verse and chorus~. \nEnergy whose dynamics are described in music dynamic features is strongly associated with arousal perception. \n\\begin{table*}[!t]\n\\centering\\scriptsize\n \\caption{Musical dimensions and acoustic features describing them.}\n\\begin{tabular}\n{ l p{10cm} }\n\\hline\n\\textbf{Musical dimension} & \\textbf{Acoustic features} \\\\\n\\hline\nMelody & Pitch \\\\%\\hline\nHarmony & chromagram, chromagram peak, key, mode, key clarity, harmonic, change, chords \\\\%\\hline\nRhythm & tempo, beat histograms, rhythm regularity, rhythm strength, onset rate\\\\%\\hline\nDynamics and loudness & RMS energy, loudness, timpral width\\\\%\\hline\nTimbre & MFCC, spectral shapres (centroid, shape, spread, skewness, kurtosis, contrast and flatness), brightness, rolloff frequency, zero crossing rate, spectral contrast, auditory modulation features, inharmonicity, roughness, dissonance, odd to even harmonic ratio \\\\%\\hline\nMusical form & Similarity Matrix (similarity between all possible frames) \\\\%\\hline\nTexture & attack slope, attack time\\\\\n\\hline\n\\end{tabular}\n\\label{tab:MusicFeatures}\n\\end{table*}\nThere are a number of toolboxes available for extracting acoustic features from music that can be used for music emotion recognition. \nMusic Analysis, Retrieval and Synthesis for Audio Signals (Marsyas)~ is an open source framework developed in C++ that supports extracting a large range of acoustic features with music information retrieval applications in mind, including time-domain zero-crossings, spectral centroid, rolloff, flux, and Mel-Frequency Cepstral Coefficients (MFCC) \\textit{etc.}\nMIRToolbox~ is an open source toolbox implemented in MATLAB for music information retrieval applications. MIRToolbox offers the ability to extract a comprehensive set of acoustic features at different levels including features related to tonality, rhythm, and structures.\nSpeech and music interpretation by large-space extraction or OpenSMILE~ is an open source software developed in C++ with the ability to extract a large number of acoustic features for speech and music analysis in real-time. \nLibROSA~ is a Python package for music and audio analysis. It is mainly developed with music information retrieval application in mind and supports importing from different audio sources and extracting musical features such as onsets chroma and tempo in addition to the low-level acoustic features. \nESSENTIA~ is an open source library developed in C++ with Python interface that is developed for audio analysis. ESSENTIA contains an extensive collection of algorithms supporting audio input/output functionality, standard digital signal processing blocks, statistical characterization of data, and a large set of spectral, temporal, tonal and high-level music descriptors. \nMusic emotion recognition either attempts to classify songs or excerpt into categories (classification) or estimate their expressed emotions on continuous dimensions (regression). The choice of machine learning model in music emotion recognition depends on the emotional representation used. Mood clusters~, dimensional representations such as arousal, tension and valence as well as music specific emotion representation can be used. An analysis of the methods proposed for MediaEval ``Music in Emotion'' task submissions revealed that using deep learning accounted for the superior performance for emotion recognition much more than the choice features~. Recent methods for emotion recognition in music rely on deep learning and often use spectrogram features that are converted to images~. \\citeauthor{aljanaki2018data} proposed learning musically meaningful mid-level perceptual features that can describe emotions in music~. They demonstrated that perceptual features such as melodiousness, modality, rhythmic complexity and dissonance can describe a large portion of emotional variance in music both in dimensional representation and MIREX clusters. They also trained a deep convolutional neural network to recognize these mid-level attributes. There have been also work attempting to use lyrics in addition to acoustic content for recognizing emotion in music~. However, lyrics are copyrighted and not easily available which hinders further work in this direction.","id":"6a4a2789-5122-4ca2-8207-b0cf83ceda85","level":"section","origin_cites_number":16,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Music"]],"subsections":[],"title":"Affective Computing of Music"},{"cite_extract_rate":0.5,"cites":[1026],"content":"\\label{sec:Video}\nCurrently, the features used in affective video content analysis are mainly from two categories~. One is considering the stimulus of video content and extracting the features reflecting the emotions conveyed by the video content itself. And the other is extracting features from the viewers. Features extracted from the video content are content-based features, and features formed from the signals of the viewers' responses are viewer-related features.","id":"14b7a32f-91ab-4f67-baaf-fabb6149902f","level":"section","origin_cites_number":2,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"]],"subsections":["0091d8de-1e85-4000-afb6-6cc1c3dcc3de","849377a8-415a-42d7-8a8d-eb0aab34165c","930f941a-bd9c-4939-bb1e-c81c0b2dac96","cefa92e9-5af7-4000-8441-381e9420e9cf","dc8e2b0f-60af-45ac-bc34-2b55fa2b0f2b"],"title":"Affective Computing of Videos"},{"cite_extract_rate":0,"cites":[],"content":"Generally speaking, the video content comprise of a series of ordered frames as well as corresponding audio signals. Therefore, it is natural to extract features from these two modalities. The audiovisual features can further be divided into low-level and mid-level according to their ability to describe the semantics of video content.","id":"0091d8de-1e85-4000-afb6-6cc1c3dcc3de","level":"subsection","origin_cites_number":0,"parent_id":"14b7a32f-91ab-4f67-baaf-fabb6149902f","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"],["subsection","Content-related Features"]],"subsections":["688b2f4a-8d63-4f40-b57e-7f57ac458bc1","1257a6cb-e4fc-4f9d-88bf-5637943576d0"],"title":"Content-related Features"},{"cite_extract_rate":0.05,"cites":[1018],"content":"Commonly, the low-level features are directly computed from the raw visual and audio content, and usually carry no semantic information. As for visual content, color, lighting, and tempo are important elements that can endow the video with strong emotional rendering and further give viewers direct visual stimuli. In many cases, computations are conducted over each frame of the video, and the average values of the computational results of the overall video are considered as visual features. Specifically, the color-related features often contain the histogram and variance of color~, the proportions of color~, the number of white frame and fades~, the grayness~, darkness ratio, color energy~, brightness ratio and saturation~, \\textit{etc}. In addition, the differences of dark and light can be reflected by the lighting key, which is used to evoke emotions in video and draw the attention of viewers by creating an emotional atmosphere~. As for the tempo-related features, properties of shot can reinforce the expression of video, such as shot change rate and shot length variance~ according to movie grammar. To better take advantage of the temporal information of the video, the motion vectors have been computed as features in~. Since the optical flow can characterize the influence of camera motions, the histogram of optimal flow matrix (HOF) has been computed as features in~. Additionally, \\citeauthor{yi2018multi}~\\shortcite{yi2018multi} traced motion key points at multiple spatial scales and computed the mean motion magnitude of each frame as features.\nTo represent audio content, pitch, zero crossing rates (ZCR), Mel frequency cepstrum coefficients (MFCC), and energy are the most popular features . In particular, the MFCC~ and its $\\Delta{MFCC}$ are used to characterize emotions in video clips frequently; while the derivatives and statistics (min, max,mean) of MFCC or $\\Delta{MFCC}$ are also explored widely. As for pitch, ~ shows that pitch of sound is associated closely with some emotions, such as anger with higher pitch and sadness with lower standard deviation of pitch. Similar situation can also occur in the energy~. For example, the total energy of anger or happiness is higher than the counterpart of unexciting emotions. ZCR~ is used to separate different types of audio signals, such as music, environmental sound and speech of human. Besides these frequent related features, audio flatness~, spectral flux~, delta spectrum magnitude, harmony~, band energy ratio, spectral centroid~, and spectral contrast~ are also utilized.\nEvidently, the aforementioned features are mostly handcrafted. With the emergence of deep learning, features can be automatically learned through deep neural networks. Some pre-trained convolutional neural networks (CNNs) are used to learn static representations from every frame or some selected key frames, while a Long-short term memory (LSTM) is exploited to capture dynamic representations existing in videos.\nFor instance, in~, an AlexNet with seven fully-connected layers trained on 2600 ImageNet classes is used to learn features. A Convolutional Auto-Encoder (CAE) is designed to ensure the CNNs can extract the visual features effectively in~. \\citeauthor{ben2018deep}~ first used the pre-trained ResNet-152 to extract feature vectors. And then, these vectors are fed into an LSTM according to their temporal order to extract high-order representations. Pre-trained model, SoundNet, is utilized to learn audio features. Because the expressive emotions of video are induced and communicated by the protagonist in video in many cases, the features of protagonist are extracted from the key frame by a pre-trained CNN and used in video affective analysis in~. In addition to the protagonist, other objects in each frame of video also give insights into emotional expression of video. For example, in~, \\citeauthor{shukla2018looking} removed the non-gaze regions from video frames (Eye ROI) and built the coarse grained scene structure remaining gist information by Guassian filter with variance. After the operations above in~, the next video affective analysis may pay more attention to important information and reduce unnecessary noise.","id":"688b2f4a-8d63-4f40-b57e-7f57ac458bc1","level":"subsubsection","origin_cites_number":20,"parent_id":"0091d8de-1e85-4000-afb6-6cc1c3dcc3de","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"],["subsection","Content-related Features"],["subsubsection","Low-level features"]],"subsections":[],"title":"Low-level features"},{"cite_extract_rate":0,"cites":[],"content":"Unlike low-level features, mid-level features often contain semantic information. For example, EmoBase10 feature depicting audio cues is computed in~. \\citeauthor{hu2016multi}~ proposed a method of combining the audio and visual features to model contextual structures of the key frames selected from video. This can produce a kind of so-called multi-instance sparse coding (MI-SC) for next analysis. In addition, the lexical features~ are extracted from the dialogues of speakers by using a natural language toolkit. These features can reflect the emotional changes in videos and can also represent a certain emotional expression in overall videos. \\citeauthor{muszynski2019recognizing}~ used aesthetic movie highlights related to occurrences of meaningful movie scenes to define some experts. These features produced by experts are more knowledgeable and abstract for video affective analysis, especially movies. HHTC features, which are computed on the basis of combination of Huang Transform in visual-audio and cross-correlation features, are proposed in~.","id":"1257a6cb-e4fc-4f9d-88bf-5637943576d0","level":"subsubsection","origin_cites_number":4,"parent_id":"0091d8de-1e85-4000-afb6-6cc1c3dcc3de","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"],["subsection","Content-related Features"],["subsubsection","Mid-level features"]],"subsections":[],"title":"Mid-level features"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Viewer-relatedFeatures}\nBesides the content-related features, viewers' facial expressions and changes of physiological signals evoked by content of videos are the most common sources for extracting viewer-related features. \\citeauthor{mcduff2017large}~ coded the facial actions of viewers for further affective video analysis. Among various physiological signals, electrocardiography (ECG), galvanic skin response (GSR), electroencephalography (EEG) are the mostly frequently ones and their statistical measures, such as mean, median, spectral power bands, \\textit{etc}., are often recommended as features. \\citeauthor{Wang2015Emotion}~ used EEG signals to construct a new EEG feature with the assistance of the relationship among video content by exploiting canonical correlation analysis (CCA).\nIn~, some viewer-related features are extracted from the whole pupil dilation ratio time-series without the differences among pupil diameter in human eyes, such as its average and derivation for global features as well as the four spectral power bands for local features.\nIn addition to viewers' responses mentioned above, the comments or other textual information produced by viewers can also reflect their attitudes or emotional reactions toward the videos. In the light of this, it is reasonable to consider users' textual comments or other textual information to extract features. In~, the ``sentiment analysis'' module using Unigrams and Bigrams is built to learn comment-related features of the collected data according to the YouTube link provided by the DEAP dataset.","id":"849377a8-415a-42d7-8a8d-eb0aab34165c","level":"subsection","origin_cites_number":3,"parent_id":"14b7a32f-91ab-4f67-baaf-fabb6149902f","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"],["subsection","Viewer-related Features"]],"subsections":[],"title":"Viewer-related Features"},{"cite_extract_rate":0.125,"cites":[1028,1026,1027],"content":"After feature extracting, a classifier or a regressor is used to obtain emotional analysis results. For classification, there are several frequently used classifiers, including support vector machines (SVM)~, Naive Bayes (NB)~, Linear Discriminant Analysis~, logistic regression (LR)~, and ensemble learning~, \\textit{etc}.\nRecent work show that the SVM-based methods are very popular for affective video content analysis due to its simplicity, max-margin training property, and use of kernels ~. For example, \\citeauthor{yi2018multi}' work ~ demonstrated that linear SVM is more suitable for classification than RBM, MLP, and LR . In~, LDA, linear SVM (LSVM), and Radial Basis SVM (RSVM) classifiers are employed in emotion recognition experiments, and the RSVM obtained the best F1 scores. In~, both Navie Bayes and SVM are used as classifiers in unimodal and multimodal conditions. In the unimodal experimental condition, NB is not better than SVM. And the fusion results showed that SVM is much better than NB in multimodal situations. However, SVM also has its shortages, such as the difficulty of selecting suitable kernel functions. Indeed, SVM is not always the best choice. In~, the results demonstrated that ensemble learning outperforms SVM in terms of classification accuracy. Ensemble learning has acquired a lot of attention in many fields because of its accuracy, simplicity, and robustness. In addition, in~, LR is adopted as the classifier for its effectiveness and simplicity. In fact, LR is used frequently in many transfer learning tasks.\nHowever, all the classifiers mentioned above are not able to capture the temporal information. Some other methods try to use temporal information. For example, \\citeauthor{gui2018implicit}~ combined SVM and LSTM to predict emotion labels. Specifically, global features and sequence features are proposed to represent the pupillary response signals. Then a SVM classifier is trained with the global features and a LSTM classifier is trained with the sequence features. Finally, a decision fusion strategy is proposed to combine these two classifiers.\n\\begin{table*}[!t]\n\\centering\\scriptsize\n\\caption{Representative work on AC of videos using kinds of features, where $P_{a}$, $P_{v}$, $MSE_{a}$, $MSE_{v}$, $Acc_{a}$, $Acc_{v}$, $Acc$ and $MAP$ indicates the Pearson correlation coefficients of arousal and valence, the mean sum error of arousal and valence, the accuracy of arousal and valence, the average accuracy and mean average precision respectively. `statistics' means (min, max, mean).}\n\\resizebox{\\textwidth}{!}{\n\\begin{tabular}\n{p{0.4 cm} p{4 cm} p{0.7 cm} p{2 cm} p{2 cm} p{0.3 cm} R{2cm}} \n\\hline\n\\textbf{Ref} & \\textbf{Feature} & \\textbf{Fusion} & \\textbf{Learning} & \\textbf{Dataset} & \\textbf{Task} & \\textbf{Result} \\\\\n\\hline\n & Mel frequency spectral; MFCC, Chroma and their derivatives and statistics; Audio compressibility; Harmonicity; Shot frequency; HOF and statistics; Histogram of 3D HSV and statistics; Video compressibility; Histogram of facial area & decision & LSTM & Dataset described by Malandrakis & reg & $P_{a}$:$0.84 \\pm 0.06$ $MSE_{a}$:$0.08 \\pm 0.04$ $P_{v}$:$0.50 \\pm 0.14$ $MSE_{v}$:$0.21 \\pm 0.06$\\\\\n & Average, standard deviation and four spectral power bands of pupil dilation ratio time-series & decision & SVM, LSTM & MAHNOB-HCI & cla & $Acc_{a}$:0.730, $Acc_{v}$:0.780\\\\\n & Audio and visual deep features from pretrained model & feature & CNN, LSTM, SVM & PMIT & cla & $MAP$:0.0.2122\\\\\n & MFCC; Color values; HoG; Dense trajectory descriptor; CNN-learned features & decision & CNN, SVM, Ensemble & DEAP & cla & $Acc$: 0.81, 0.49 \\\\\n & HHTC features & feature & SVR & Discrete LIRIS-ACCEDE & reg & $MSE_{a}$:0.294, $MSE_{v}$: 0.290 \\\\\n & Statistical measures (such as mean, median, skewness kurtosis) for EEG data, power spectral features, ECG, GSR, Face/Head-pose & decision & SVM/NB & music excerpts~ & cla & F1 (v: 0.59, 0.58, a: 0.60, 0.57) \\\\\n & Time-span visual and visual features & feature & CNN Opensmile toolbox & music excerpts~ & cla & $MSE_{a}$:0.082 $MSE_{v}:0.071$\\\\\n & tempo; pitch; zero cross; roll off; MFCCs; Saturation; Color heat; Shot length feature; General preferences; Visual excitement; Motion feature; fMRI feature & feature & DBM SVM & TRECVID & cla & - \\\\\n & Colorfulness; MFCC; CNN-learned features from the keyframes containing protagonist & decision & CNN, SVM, SVR & LIRIS-ACCEDE, PMSZU & cla/reg & - \\\\\n & Multi-frame motion vectors & decision & CNN & SumMe, TVsum, Continuous LIRIS-ACCEDE & reg & - \\\\\n & The median of the L values in Luv space; means and variances of components in HSV space; texture feature; mean and standard deviation of motions between frames in a short; MFCC; Spectral power; mean and variance of the spectral centroids; Time domain zero crossings rate; Multi-instance sparse coding & feature & SVM & Musk1, Musk2, Elephant, Fox, Tiger & cla & $Acc$: 0.911, 0.906, 0.885, 0.627, 0.868\\\\\n & Lighting key; Color; Motion vectors; ZCR; energy; MFCC; pitch; Textual features & decision & SMO, Navie Bayes & DEAP & cla & F1:0.849 0.811 $Acc$: 0.911 0.883 \\\\\n & Key lighting; Grayness; Fast motion; Shot chanage rate; Shot length variation; MFCC; CNN-learned features; power spectral density; EEG; ECG; respiation; galvanic skin resistance & feature & SVM & DEAP & cla & $Acc_v$:0.7 0.7 0.7125, $Acc_a$:0.6876 0.7 0.8 F1 (A:0.664 0.687 0.789) \\\\\n & MFCC; ZCR; energy; pitch, color histograms; lighting key; motion vector & decision & SVM, Navie Bayes & DEAP & cla & F1: 0.869, 0.846 $Acc$: 0.925, 0.897 \\\\\n & CNN feature, low-level audio visual features, EEG & decision & LDA, LSVM, RSVM & Dataset introduced by the authors & cla & - \\\\\n & MKT; ConvNets feature; EmoLarge; IS13; MFCC; EmoBase10; DSIFT; HSH & decision & SVM, LR, RBM, MLP & MediaEval 2015, 2016 Affective Impact of Movies & cla & $Acc_a$:0.574, $Acc_v$:0.462\\\\\n & CNN feature & - & SVM, LDA & dataset in~ & cla & - \\\\\n & CNN feature & - & SVR & LIRIS-ACCEDE & reg & $MSE_{a}$: 0.021, $MSE_{v}$: 0.027\\\\\n\\hline\n\\end{tabular}\n}\n\\label{tab:VideoHandCraftedMethods}\n\\end{table*}\nA regressor is needed when mapping the extracted features to the continuous dimensional emotion space. Recently, one of the most popular regression method is support vector regression (SVR)~. For example, in~, video features like audio, color, aesthetic are fed into SVR in the SVR-Standard experiment. And in the SVR-Transfer learning experiment, the pre-trained CNN is treated as a feature extractor. The CNN's outputs are used as the input to the SVR. The experimental results showed that the SVR-Transfer learning outperforms other methods. Indeed, the various kernel functions in SVR provide a stronger adaptability.","id":"930f941a-bd9c-4939-bb1e-c81c0b2dac96","level":"subsection","origin_cites_number":24,"parent_id":"14b7a32f-91ab-4f67-baaf-fabb6149902f","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"],["subsection","Machine Learning Methods"]],"subsections":[],"title":"Machine Learning Methods"},{"cite_extract_rate":0,"cites":[],"content":"In total, there are two fusion strategies for multimodal information: feature-level fusion and decision-level fusion. Feature-level fusion means that the multimodal features are combined and then used as the input of a classifier or a regressor. Decision-level fusion fuses several results of different classifiers, and the final results are computed according to the fusion methods.\nOne way of feature-level fusion is implemented by feature accumulation or concatenation~. In~, two feature vectors for visual and audio data are averaged as the global genre representations. In~, multi-class features are concatenated to generate a high dimensional joint representation. Some machine learning methods are also employed to learn joint features~. In~, a two-branch network is used to combine the visual and audio features. The outputs of the two-branch network are then fed into a classifier, and the experiment results showed that the joint features outperform other methods. In~, the low-level audio-visual features and fMRI-derived features are fed into multimodal DBM to learn joint representations. The target of this method is to learn the relation between audio-visual features and fMRI-derived features. In~, PCA is used to learn the multimodal joint features. In~, canonical correlation analysis (CCA) is used to construct a new video feature space with the help of EEG features and a new EEG feature space with the assistance of video content, so only one modality is needed to predict emotion during the testing process.\nBy combining the results of different classifiers, decision-level fusion strategy is able to achieve better results~. In~, linear fusion and SVM-based fusion techniques are explored to combine outputs of several classifiers. Specifically, the output of each classifier has its own weight in linear fusion. The final result is the weighted sum of all outputs. In SVM-based fusion, the outputs of unimodal classifiers are concatenated together. And then the higher level representations for each video clip are fed into a fusion SVM to predict the emotion. Based on these results, linear fusion is better than SVM-based fusion. In~, linear fusion is also used to fuse the outputs of multiple classifiers. The differences among these linear fusion methods depend on the distribution of weights.","id":"cefa92e9-5af7-4000-8441-381e9420e9cf","level":"subsection","origin_cites_number":13,"parent_id":"14b7a32f-91ab-4f67-baaf-fabb6149902f","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"],["subsection","Data Fusion"]],"subsections":[],"title":"Data Fusion"},{"cite_extract_rate":0.5,"cites":[1029],"content":"\\label{sec:Deep learning method}\nIn tradition, video emotional recognition includes two steps, \\textit{i.e.}, feature extraction step and regression or classification step. Because of the lack of consensus on the most relevant emotional features, we may not be able to extract the best features for the problem at hand. As a result, this two-step mode has hampered the development of affective video content analysis. In order to solve this problem, some methods based on end-to-end training frameworks are proposed. \\citeauthor{khorrami2016deep} combined CNN and RNN to recognize the emotional information of videos. According to their method, a CNN is trained using frame facial images sampled from videos to extract features. Then the features are fed into a RNN to perform continuous emotion recognition. In~, a single network using ConvLSTM is proposed, where videos are input to the network and the predicted emotional information is output directly. In fact, due to the complexity of CNNs and RNNs, the training of these frameworks needs large amounts of data. However, in video affective content analysis, the samples in existing datasets are usually limited. This is the reason why end-to-end methods are still less common compared to the traditional two step methods, despite of their influential potentials.","id":"dc8e2b0f-60af-45ac-bc34-2b55fa2b0f2b","level":"subsection","origin_cites_number":2,"parent_id":"14b7a32f-91ab-4f67-baaf-fabb6149902f","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Videos"],["subsection","Deep Learning Methods"]],"subsections":[],"title":"Deep Learning Methods"},{"cite_extract_rate":0.185185185185185,"cites":[7,1030,1032,1684,1031],"content":"\\label{sec:multimodal}\nIn this section, we survey the work that analyze multimodal data beyond audiovisual content. Most of the existing work on affective understanding of multimedia rely on one modality, even when additional modalities are available, for example in videos~. Earlier work on emotional understanding of multimedia used hand crafted features from different modalities that are fused at feature or decision levels~. The more recent work mainly use deep learning models~. \nLanguage is a commonly used modality in addition to vision and audio. There is a large body of work on text-based sentiment analysis~. Sentiment analysis from text is well-established and is deployed at scale in industry at a broad set of applications involving opinion mining~. With the shift toward an increasingly multimodal social web, multimodal sentiment analysis is becoming more relevant. For example, vloggers post their opinions on YouTube, and photos commonly accompany user posts on Instagram and Twitter. \nAnalyzing text for emotion recognition requires representing terms by features. Lexically-based approaches are one of the most popular methods for text-based emotion recognition. They involve using knowledge of words' affect for estimating document or content's affect. Linguistic Inquiry and Word Count (LIWC) is a well-known lexical tool that matches the terms in a document with its dictionary and generates scores along different dimensions including affective and cognitive constructs such as ``present focus'' and ``positive emotion''~. The terms in each category or selected by experts is extensively validated on different content. AffectNet is another notable lexical resource which includes a semantic netowrk of 10,000 items with representations for ``pleasantness'', ``attention'', ``sensitivity'', and ``aptitude''~. The continuous representations can be mapped to 24 distinct emotions. DepecheMood is a lexicon created through a data-driven method mining a news website annotated with its particular set of discrete emotions, namely, ``afraid'', ``amusemed'', ``anger'', ``annoyed'', ``don't care'', ``happy'', and ``inspired''~. DepecheMood is extended to DepecheMood++ by including Italian~.\nThe more recent development in text-based affective analysis is models powered by deep learning. Leveraging large scale data, deep neural networks are able to learn representations that are relevant for affective analysis in language. Word embeddings are one of the most common representations used to represent language. Word embeddings, such as Word2Vec~ or GloVe~, learn language context of the word by learning a representation (a vector), that can capture semantic and syntactic similarities. More recently, representation learning models that can encode the whole sequence of terms (sentences, documents) showed impressive performance in different language understanding tasks, including sentiment and emotional analysis. \nBidirectional Encoder Representations from Transformers (BERT)~ is a method for learning a language model that can be trained on large amount of data in an unsupervised manner. This pre-trained model is very effective in representing a sequence of terms as a fixed-length representation (vector). BERT architecture is a multi-layer bidirectional Transformer network that encodes the whole sequence at once. BERT representation achieves state-of-the-art results in multiple natural language understanding tasks. \nThe audiovisual features that are used for multimodal understanding of affect are similar to the ones discussed in previous sections. The main technique between miltimodal models lies in methods for multimodal fusion. Multimodal methods involve extracting features from multiple modalities, \\textit{e.g.}, audiovisual, and training joint or separate machine learning models for fusion~. Multimodal fusion can be done in model-based and model-agnostic ways. The model-agnostic fusion methods do not rely on a specific classification or regression method and include feature-level, decision-level, or hybrid fusion techniques. Model-based methods address multimodal fusion in model construction. Examples of model-based fusion methods include Multiple Kernel Learning (MKL)~, graphical models, such as Conditional Random Fields~ and neural networks~.\n\\citeauthor{Pang2015}~ used Deep Boltzmann Machine (DBM) to learn a joint representation across text, vision, and audio to recognize expected emotions from social media videos. Each modality is separately encoded with stacking multiple Restricted Boltzmann Machines (RBM) and pathways are merged to a joint representation layer. The model was evaluated for recognizing eight emotion categories for 1,101 videos from~. \\citeauthor{muszynski2019recognizing}~ studied perceived vs induced emotion in movies. To this end, they collected additional labels on a subset of LIRIS-ACCEDE dataset~. They found that perceived and induced emotions do not always agree. Using multimodal Deep Belief Networks (DBN), they could demonstrate that fusion of electrodermal responses with audiovisual content features improves the overall accuracy for emotion recognition~.\nIn~, authors performed regression to estimate intended arousal and valence levels (as judged by experts in~). LSTM recurrent neural networks are used for unimodal regressions and fused via early and late fusion for audiovisual estimation with late fusion achieving the best results. \\citeauthor{Tarvainen2018}~ performed an in-depth analysis on how emotions are constructed in movies. They identified scene type as a major factor in emotions in movies. They then used content features to recognize emotions along three dimensions of hedonic tone (valence), energetic arousal (awake--tired) and tense arousal (tense--calm).\nBilinear fusion is a method that is proposed to model inter- and intra- modality interaction among modalities by performing outer product between unimodal embeddings~. \\citeauthor{zadeh2017tensor}~ extended this to a Tensor Fusion Network to model intra-modality and inter-modality dynamics in multimodal sentiment analysis. The tensor fusion network includes modality embedding sub-networks, a tensor fusion layer modeling the unimodal, bimodal and trimodal interactions using a three-fold Cartesian product from modality embeddings along with a final sentiment inference sub-network conditioned on the tensor fusion layer. The main drawback of such methods is the increase in the dimensionality of the resulting multimodal representation.","id":"20c98c17-ea71-4ba4-a9aa-815a42eced5b","level":"section","origin_cites_number":27,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Affective Computing of Multimodal Data"]],"subsections":[],"title":"Affective Computing of Multimodal Data"},{"cite_extract_rate":0.25,"cites":[7022,1033],"content":"\\label{sec:FutureDirections}\nAlthough remarkable progress has been made on affective computing of multimedia (ACM) data, there are still several open issues and directions that can boost the performance of ACM.\n\\textbf{Multimedia Content Understanding.} As emotions may be directly evoked by the multimedia content in viewers, accurately understanding what is contained in multimedia data can significantly improve the performance of ACM. Sometimes it is even necessary to analyze the subtle details. For example, we may feel ``amused'' on a video with a laughing baby; but if the laugh is from a negative character, it is more possible for us to feel ``angry''. In such cases, besides the common property, such as ``laugh'', we may need to further recognize the identity, such as ``a lovely baby'' and ``an evil antagonist''.\n\\textbf{Multimedia Summarization.} Emotions can play a vital role in selection of multimedia for creation of summaries or highlights. This is an important application in entertainment and sports industries (\\textit{e.g.} movie trailers, sports highlights). There has been some recent work in this direction where affect information from audio visual cues has led to the successful creation of video summaries~. In particular, work reported in~ used audiovisual emotions in part to create an AI trailer for a $20^{th}$ Century Fox film in 2016. Similarly, AI Highlights described in~ hinged on audiovisual emotional cues and have successfully been employed to create the official highlights at Wimbledon and US Open since 2017. This is a very promising direction for affective multimedia computing which can have a direct impact on real world media applications.\n\\textbf{Contextual Knowledge Modeling.} The contextual information of a viewer watching some multimedia is very important. Similar multimedia data under different contexts may evoke totally different emotions. For example, we may feel ``happy\" when listening a song about love in a wedding; but if the same song is played when two lovers are departing, it is more likely that we feel ``sad\". The prior knowledge of viewers or multimedia data may also influence the emotion perceptions. An optimistic viewer and a pessimistic viewer may have totally different emotions about the same multimedia data.\n\\textbf{Group Emotion Clustering.} It is too generic to simply recognize the dominant emotion, while it is too specific to predict personalized emotion. It would make more sense to model emotions for groups or cliques of viewers with similar interests and backgrounds. Clustering different viewers into corresponding groups possibly based on the user profiles may provide a feasible solution to this problem.\n\\textbf{New AC Setting Adaptation.} Because of the domain shift~, the deep learning models trained on one labeled source domain may not work well on the other unlabeled or sparsely labeled target domain, which results in the models' low transferability to new domains. Exploring domain adaptation techniques that fit well on the AC tasks is worth investigating. One possible solution is to translate the source data to an intermediate domain that are indistinguishable from the target data while preserving the source labels~ using Generative Adversarial Networks~. How do deal with some practical settings, such as multiple labeled source domains and emotion models' homogeneity, is more challenging.\n\\textbf{Regions-of-Interest Selection.} The contributions of different regions of given multimedia may vary to the emotion recognition. For example, the regions that contain the most important semantic information in images are more discriminative than background; some video frames are of no use to emotion recognition. Detecting and selecting the regions-of-interest may significantly improve the recognition performance as well as the computation efficiency.\n\\textbf{Viewer-Multimedia Interaction.}\nInstead of direct analysis of the multimedia content or implicit consideration of viewers' physiological signals (such as facial expressions, Electroencephalogram signals, \\textit{etc}.), joint modeling of both multimedia content and viewers' responses may better bridge the affective gap and result in superior performances. We should also study how to deal with missing or corrupted data. For example, some physiological signals are unavailable during the data collection stage.\n\\textbf{Affective Computing Applications.} Although AC is claimed to be important in real-world applications, few practical systems have been developed due to the relatively low performance. With the availability of larger datasets and improvements in self-supervised and semi-supervised learning, we foresee the deployment of ACM in real-world applications. For example, in media analytics, the content understanding methods will identify the emotional preferences of users and emotional nuances of social media content to better target advertising effort; in fashion recommendation, intelligent costumer service, such as customer-multimedia interaction, can provide better experience to customers; in advertisement, generating or curating multimedia that evokes strong emotions can attract more attention. We believe that an emotional artificial intelligence will become a significant component of mainstream multimedia applications.\n\\textbf{Benchmark Dataset Construction.}\nExisting studies on ACM mainly adopt small-scale datasets or construct relatively larger-scale ones using keyword searching strategy without annotation quality guaranteed. To advance the development of ACM, creating a large-scale and high-quality dataset is in urgent need. It has shown that there are three critical factors for dataset construction of ACM, \\textit{i.e.}, the context of viewer response, personal variation among viewers, and the effectiveness and efficiency of corpus creation~. In order to include a large number of samples, we may exploit online systems and crowdsourcing platforms to recruit large numbers of viewers with a representative spread of backgrounds to annotate multimedia and provide contextual information on their emotional responses. Since emotion is a subjective variable, personalized emotion annotation would make more sense, from which we can obtain the dominant emotion and emotion distribution. Further, accurate understanding of multimedia content can boost the affective computing performance. Inferring emotional labels from social media users' interaction with data, \\textit{e.g.}, likes, comments, in addition to their spontaneous responses, \\textit{e.g.}, facial expression, where possible, will provide new avenues for enriching affective datasets.","id":"5da3fa17-09f5-4d0c-9b55-5610d97ddda5","level":"section","origin_cites_number":8,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Future Directions"]],"subsections":[],"title":"Future Directions"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Conclusion}\nIn this article, we have surveyed affective computing (AC) methods for heterogeneous multimedia data. For each multimedia type, \\textit{i.e.}, image, music, video, and multimodal data, we summarized and compared available datasets, handcrafted features, machine learning methods, deep learning models, and experimental results. We also briefly introduced the commonly employed emotion modelds and outlined potential research directions in this area. Although deep learning-based AC methods have achieved remarkable progress in recent years, an efficient and robust AC method that is able to obtain high accuracy under unconstrained conditions is yet to be achieved. With the advent of deep understanding of emotion evocation in brain science, accurate emotion measurement in psychology, and novel deep learning network architectures in machine learning, affective computing of multimedia data will remain an\nactive research topic for a long time.\n\\begin{acks}\nThis work was supported by Berkeley DeepDrive, the National Natural Science Foundation of China (Nos. 61701273, 91748129), and the National Key R\\&D Program of China (Grant No. 2017YFC011300). The work of MS is supported in part by the U.S. Army. Any opinion, content or information presented does not necessarily reflect the position or the policy of the United States Government, and no official endorsement should be inferred.\n\\end{acks}\n\\bibliographystyle{ACM-Reference-Format}\n\\bibliography{references}\n\\end{document}","id":"fc0b3d97-cde9-409b-91c4-fe25e8140878","level":"section","origin_cites_number":0,"parent_id":"36832e43-191f-4e88-aa11-6c7b91fe1bf8","prefix_titles":[["title","Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ]\n ],\n \"subsections\": [\n \"6716f171-5279-4be6-829f-92b9ec4bb350\",\n \"62f1d80f-e95a-45c5-ae8a-d6a71fc0fefa\",\n \"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa\",\n \"50a2810e-2777-44d0-85b6-12b160fc2d4d\",\n \"6a4a2789-5122-4ca2-8207-b0cf83ceda85\",\n \"14b7a32f-91ab-4f67-baaf-fabb6149902f\",\n \"20c98c17-ea71-4ba4-a9aa-815a42eced5b\",\n \"5da3fa17-09f5-4d0c-9b55-5610d97ddda5\",\n \"fc0b3d97-cde9-409b-91c4-fe25e8140878\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.08333333333333301,\n \"cites\": [\n 1016,\n 1015\n ],\n \"content\": \"\\\\label{sec:Introduction}\\nUsers are increasingly recording their daily activities, sharing interesting experiences, and expressing personal viewpoints using mobile devices on social networks, such as Twitter\\\\footnote{\\\\url{https://twitter.com}}, Facebook\\\\footnote{\\\\url{https://www.facebook.com}}, and Weibo\\\\footnote{\\\\url{https://www.weibo.com}}, \\\\textit{etc}. As a result, a rapidly growing volume of multimedia data (\\\\textit{i.e.}, image, music, and video) has been generated, as shown in Figure~\\\\ref{fig:multimedia}, which results in a great demand for the management, retrieval, and understanding of these data. Most existing work on multimedia analysis focus on the cognitive aspects, \\\\textit{i.e.}, understanding the objective content, such as object detection in images~, speaker recognition in speech~, and action recognition in videos~. Since what people feel have a direct influence on their decision making, affective computing (AC) of these multimedia data is of significant importance and has attracted increasing attention . For example, companies would like to know how customers evaluate their products and can thus improve their services~; depression and anxiety detection from social media can help understand psychological distress and thus potentially prevent suicidal actions~.\\nWhile the sentiment analysis in text~ has long been a standard task, AC from other modalities, such as image and video, has just begun to be considered recently. In this article, we aim to review the existing AC technologies comprehensively for large-scale heterogeneous multimedia data, including image, music, video, and multimodal data.\\nAffective computing of multimedia (ACM) aims to recognize the emotions that are expected to be evoked in viewers by a given stimuli. Similar to other supervised learning tasks, ACM is typically composed of three steps: data collection and annotation, feature extraction, and mapping learning between features and emotions~. One main challenge for ACM is the affective gap, \\\\textit{i.e.}, ``the lack of coincidence between the features and the expected affective state in which the user is brought by perceiving the signal''~. In the early stage, various hand-crafted features were designed to bridge this gap with traditional machine learning algorithms, while more recently researchers have focused on end-to-end deep learning from raw multimedia data to recognize emotions. Existing ACM methods mainly assign the dominant (average) emotion category (DEC) to an input stimuli, based on the assumption that different viewers have similar reactions to the same stimuli. We can usually formulate this task as a single-label learning problem.\\nHowever, emotions are influenced by subjective and contextual factors, such as the educational background, cultural diversity, and social interaction~. As a result, different viewers may react differently to the same stimuli, which creates the subjective perception challenge. Therefore, the perception inconsistency makes it insufficient to simply predict the DEC for the highly subjective variable. As stated in~, we can perform two kinds of ACM tasks to deal with the subjectivity challenge: predicting personalized emotion perception for each viewer and assigning multiple emotion labels for each stimuli. For the latter one, we can either assign multiple labels to each stimuli with equal importance using multi-label learning methods, or predict the emotion distributions which tries to learn the degrees of each emotion~.\\nIn this article, we concentrate on surveying the existing methods on ACM and analyzing potential research trends. Section~\\\\ref{sec:EmotionModels} introduces the widely-used emotion representation models from psychology. Section~\\\\ref{sec:Datasets} summarizes the existing available datasets for evaluating ACM tasks. Section~\\\\ref{sec:Image}, Section~\\\\ref{sec:Music}, Section~\\\\ref{sec:Video}, and Section~\\\\ref{sec:multimodal} survey the representative methods on AC of images, music, videos, and multimodal data, respectively, including both handcrafted features-based methods and deep learning methods. Section~\\\\ref{sec:FutureDirections} provides some suggestions for future research, followed by conclusion in Section~\\\\ref{sec:Conclusion}.\\nTo the best of our knowledge, this article is among the first that provide a comprehensive survey of affective computing of multimedia data from different modalities. Previous surveys mainly focus on a single modality, such as images~, speech~, music~, video~, and multimodal data~. From this survey, readers can more easily compare the correlations and differences among different AC settings. We believe that this will be instrumental in generating novel research ideas.\",\n \"id\": \"6716f171-5279-4be6-829f-92b9ec4bb350\",\n \"level\": \"section\",\n \"origin_cites_number\": 24,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:EmotionModels}\\nThere are two dominant emotion representation models deployed by psychologists: categorical emotions (CE), and dimensional emotion space (DES). CE models classify emotions into a few basic categories, such as \\\\emph{happiness} and \\\\emph{anger}, \\\\emph{etc}. Some commonly used models include Ekman's six basic emotions and Mikels's eight emotions~. When classifying emotions into \\\\emph{positive} and \\\\emph{negative} (polarity)~, sometimes including \\\\emph{neutral}, ``emotion'' is called ``sentiment''. However, sentiment is usually defined as an atitude held toward an object . Emotions are usually represented by DES models as continuous coordinate points in a 3D or 2D Cartesian space, such as valence-arousal-dominance (VAD) and activity-temperature-weight~. VAD is the most widely used DES model, where valence represents the pleasantness ranging from positive to negative, arousal represents the intensity of emotion ranging from excited to calm, and dominance represents the degree of control ranging from controlled to in control. Dominance is difficult to measure and is often omitted, leading to the commonly used two dimensional VA space~.\\nThe relationship between CE and DES and the transformation from one to the other are studied in~. For example, positive valence relates to a happy state, while negative valence relates to a sad or angry state. CE models are easier for users to understand and label, but the limited set of categories may not well reflect the subtlety and complexity of emotions. DES can better describe detailed emotions with subtle differences flexibly, but it is difficult for uses to distinguish the absolute continuous values, which may also be problematic. \\nCE and DES are mainly employed in classification and regression tasks, respectively, with discrete and continuous emotion labels. If we discretize DES into several constant values, we can also use it for classification~. Ranking based labeling can be applied to ease DEC comprehension difficulties in raters.\\nAlthough less explored in this context, one of the most well-known theories that explains the development of emotional experience is appraisal theory. According to this theory, cognitive evaluation or appraisal of a situation or content in case of multimedia results in emergence of emotions~. According to Ortony, Clore and Collins (OCC)~, emotions are experienced following a scenario comprising a series of phases. First, there is a perception of an event, object or an action. Then, there is an evaluation of events, objects or action according to personal wishes or norms. Finally, perception and evaluation result in a specific emotion or emotions arising. Certain appraisal dimensions such as novelty and complexity can be labeled and detected from content. For example, \\\\citeauthor{Soleymani2015} automatically recognized image novelty and complexity that are related to interestingness. There are also domain specific emotion taxonomy and scales. Geneva Emotional Music Scale is a music specific emotion model for describing emotions induced by music. It consists of a hierarchical structure with 45 emotions, nine emotional categories and three superfactors that can describe emotion in music.\\n\\\\begin{table}[!t]\\n\\\\centering\\\\scriptsize\\n\\\\caption{Representative emotion models employed in ACM.}\\n\\\\begin{tabular}\\n{cccc}\\n\\\\hline\\n\\\\textbf{Model} & \\\\textbf{Ref} & \\\\textbf{Type} & \\\\multicolumn{1}{c}{\\\\textbf{Emotion states/dimensions}} \\\\\\\\\\n\\\\hline\\nEkman & & CE & happiness, sadness, anger, disgust, fear, surprise\\\\\\\\\\nMikels & & CE & amusement, anger, awe, contentment, disgust, excitement, fear, sadness\\\\\\\\\\nPlutchik & & CE & ($\\\\times$ 3 scales) anger, anticipation, disgust, joy, sadness, surprise, fear, trust\\\\\\\\\\nClusters & & CE & 29 discrete labels are consistently grouped into 5 clusters at a similar distance level \\\\\\\\\\nSentiment & & CE & positive, negative, (or neutral)\\\\\\\\\\nVA(D) & & DES & valence-arousal(-dominance)\\\\\\\\\\nATW & &DES & activity-temperature-weight\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\label{tab:EmotionModels}\\n\\\\end{table}\\nAnother relevant concept worth mentioning is that emotion in response to multimedia can be expected, induced or perceived emotion. \\nExpected emotion is the emotion that the multimedia creator intends to make people feel, perceived emotion is what people perceive as being expressed, while induced/felt emotion is the actual emotion that is felt by a viewer. Discussing the difference or correlation of various emotion models is out of the scope of this article. The typical emotion models that have been widely used in ACM are listed in in Table~\\\\ref{tab:EmotionModels}.\",\n \"id\": \"62f1d80f-e95a-45c5-ae8a-d6a71fc0fefa\",\n \"level\": \"section\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Emotion Models from Psychology\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Emotion Models from Psychology\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Datasets}\",\n \"id\": \"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [\n \"31b031e5-4c1a-41d6-91e4-966f262d426b\",\n \"6be4124c-0549-4c40-a1cf-f2efaaa9747b\",\n \"d278bc85-2fef-47c9-b4d6-d1b6bb157aad\",\n \"72c21e7b-1086-4048-bb25-1bc86a5c6b31\"\n ],\n \"title\": \"Datasets\"\n },\n {\n \"cite_extract_rate\": 0.08333333333333301,\n \"cites\": [\n 1017\n ],\n \"content\": \"\\\\label{sec:ImageDatasets}\\nThe early datasets for AC of images mainly come from the psychology community with small-scale images. The International Affective Picture System (\\\\textbf{IAPS}) is an image set that is widely used in psychology to evoke emotions~. Each image that depicts complex scenes is associated with the mean and standard deviation (STD) of VAD ratings in a 9-point scale by about 100 college students. The \\\\textbf{IAPSa} dataset is selected from IAPS with 246 images~, which are labeled by 20 undergraduate students. The \\\\textbf{Abstract} dataset consists of 279 abstract paintings without contextual content. Approximately 230 people peer rated these paintings. The Artistic dataset (\\\\textbf{ArtPhoto}) includes 806 artistic photographs from a photo sharing site~ with emotions determined by the artist uploading the photos. The Geneva affective picture database (\\\\textbf{GAPED}) is composed of 520 negative, 121 positive, and 89 neutral images~. Besides, these images are also rated with valence and arousal values, ranging from 0 to 100 points. There are 500 abstract paintings in both \\\\textbf{MART} and \\\\textbf{devArt} datasets, which are collected from the Museum of Modern and Contemporary Art of Trento and Rovereto~, and the ``DeviantArt'' online social network~, respectively.\\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n\\\\caption{Released and freely available datasets for AC of images, where `Ref' is short for Reference, `\\\\# Images' and `\\\\# Annotators' respectively represent the total number of images and annotators (f: female, m: male), `Labeling' represents the method to obtain labels, such as human annotation (annotation) and keyword searching (keyword), and `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).}\\n\\\\begin{tabular}\\n{cccccccc}\\n\\\\hline\\n\\\\textbf{Dataset} & \\\\textbf{Ref} & \\\\textbf{\\\\# Images} & \\\\textbf{Type} & \\\\textbf{\\\\# Annotators}& \\\\textbf{Emotion model} & \\\\textbf{Labeling} & \\\\textbf{Labels}\\\\\\\\\\n\\\\hline\\nIAPS & & 1,182 & natural & $\\\\approx$100 (half f) & VAD & annotation & average \\\\\\\\\\nIAPSa & & 246 & natural & 20 (10f,10m) & Mikels & annotation & dominant \\\\\\\\\\nAbstract & & 279 & abstract & $\\\\approx$230 & Mikels & annotation & dominant \\\\\\\\\\nArtPhoto & & 806 & artistic & -- & Mikels & keyword & dominant \\\\\\\\\\nGAPED & & 730 & natural & 60 & Sentiment, VA & annotation & dominant, average \\\\\\\\\\nMART & & 500 & abstract & 25 (11f,14m) & Sentiment & annotation & dominant\\\\\\\\\\ndevArt & & 500 & abstract & 60 (27f,33m) & Sentiment & annotation & dominant\\\\\\\\\\nTweet & & 603 & social & 9 & Sentiment & annotation & dominant \\\\\\\\\\nFlickrCC & & $\\\\approx$500,000 & social & -- & Plutchik & keyword & dominant \\\\\\\\\\nFlickr & & 301,903 & social & 6,735 & Ekman & keyword & dominant \\\\\\\\\\nEmotion6 & & 1,980 & social & 432 & Ekman+neutral & annotation & distribution \\\\\\\\\\nFI & & 23,308 & social & 225 & Mikels & annotation & dominant \\\\\\\\\\nIESN & & 1,012,901 & social & 118,035 & Mikels, VAD & keyword & personalized \\\\\\\\\\nFlickrLDL & & 10,700 & social & 11 & Mikels & annotation & distribution \\\\\\\\\\nTwitterLDL & & 10,045 & social & 8 & Mikels & annotation & distribution \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\label{tab:ImageDataset}\\n\\\\end{table*}\\nRecent datasets, especially the large-scale ones, are constructed using images from social networks. The Tweet dataset (\\\\textbf{Tweet}) consists of 470 and 113 tweets for positive and negative sentiments, respectively~. The \\\\textbf{FlickrCC} dataset includes about 500k Flickr creative common (CC) images which are generated based on 1,553 adjective noun pairs (ANPs)~. The images are mapped to the Plutchnik's Wheel of Emotions with 8 basic emotions, each with 3 scales. The \\\\textbf{Flickr} dataset contains about 300k images~ with the emotion category defined by the synonym word list which has the most same words as the adjective words of an image's tags and comments. The \\\\textbf{FI} dataset consists of 23,308 images which are collected from Flicker and Instagram by searching the emotion keywords~ and labeled by 225 Amazon Mechanical Turk (MTurk) workers. The number of images in each Mikels emotion category is larger than 1,000. The \\\\textbf{Emotion6} dataset~ consists of 1,980 images collected from Flickr with 330 images for each Ekman's emotion category. Each image was scored by 15 MTurk workers to obtain the discrete emotion distribution information. The \\\\textbf{IESN} dataset that is constructed for personalized emotion prediction contains about 1M images from Flickr. Lexicon-based methods and VAD averaging~ are used to segment the text of metadata from uploaders for expected emotions and comments from viewers for personalized emotions. There are 7,723 active users with more than 50 involved images. We can also easily obtain the DEC and emotion distribution for each image. \\\\textbf{FlickrLDL} and \\\\textbf{TwitterLDL} datasets~ are constructed for discrete emotion distribution learning. The former one is a subset of FlickrCC, which are labeled by 11 viewers. The latter one consists of 10,045 images which are collected by searching various sentiment key words from Twitter and labeled by 8 viewers. These datasets are summarized in Table~\\\\ref{tab:ImageDataset}.\",\n \"id\": \"31b031e5-4c1a-41d6-91e4-966f262d426b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ],\n [\n \"subsection\",\n \"Datasets for AC of Images\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets for AC of Images\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:MusicDatasets}\\nA notable benchmark for music recognition is music mood classification (AMC) task, organized by annual Music Information Retrieval Evaluation eXchange\\\\footnote{\\\\url{http://www.music-ir.org/mirex/wiki/}} (MIREX)~. In MIREX mood classification task, initially 600 songs were shared with the participants. Starting from 2013, 1,438 30 seconds excerpts from Korean pop songs have been added to MIREX. MIREX benchmark aims to automatically classify songs into five emotion clusters derived from cluster analysis of online tags. MIREX mood challenge emotional representation has been debated in the literature due to its data-driven origin rather than psychology of emotion. For example, in~, semantic and acoustic overlaps have been found between clusters. MIREX mood challenge considers only one label for the whole song and disregards the dynamic time evolving nature of music.\\nComputer Audition Lab 500 (CAL500) is a dataset of 500 popular songs which is labeled by multiple tags including emotions~. The dataset is labeled in the lab by 66 labelers. Soundtracks datasets~ for music and emotion is developed by~\\\\citeauthor{Eerola2011} and contains instrumental music from soundtrack of 60 movies. The expert annotators selected songs based on five basic discrete categories (\\\\textit{anger}, \\\\textit{fear}, \\\\textit{sadness}, \\\\textit{happiness}, and \\\\textit{tenderness}) and dimensional VA representation of emotions. Although not developed with music content analysis in mind, the Database for Emotion Analysis using Physiological Signals or DEAP dataset~ also includes valence, arousal and dominance ratings for 120 one-minute music video clips of western pop music. Each video clip is annotated by by 14--16 participants who were asked to report their felt valence, arousal, and dominance on a 9-point scale. AMG1608~ is another music dataset that contains arousal and valence ratings for 1,608 Western songs in different genres and is annotated through MTurk.\\nMusic datasets with emotion labels usually consider one emotion label per song (static). MoodSwings dataset~ was the first to annotate music dynamically over time. MoodSwings was developed by Schmidt \\\\etal and includes 240 15s excerpts of western pop songs with per-second valence and arousal labels, collected on MTurk. The MediaEval ``Emotion in Music'' challenge was organized in years 2013--2015 in MediaEval Multimedia Evaluation initiative\\\\footnote{http://www.multimediaeval.org}. MediaEval is a community-driven benchmarking campaign dedicated to evaluating algorithms for social and human-centered multimedia access and retrieval~. Unlike MIREX, ``Emotion in Music'' task focused on dynamic emotion recognition in music tracking arousal and valence over time~. The data from MediaEval tasks were compiled in MediaEval Database for Emotional Analysis in Music (DEAM) which is the largest available dataset with dynamic annotations, at 2Hz, with valence and arousal annotations for 1,802 songs and song excerpts licensed under Creative Commons license. PMEmo is a dataset of 794 songs with dynamic and static arousal and valence annotations in addition to electrodermal responses from ten participants~.\\nThese datasets are summarized in Table~\\\\ref{tab:MusicDataset}. For a more detailed review of available music datasets with emotional labels, we refer the readers to~.\\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n\\\\caption{Released and freely available datasets for music emotion recognition, where `\\\\# Songs' and `\\\\# Annotators' respectively represent the total number of songs and annotators per song, `Labeling' represents the method to obtain labels, such as human annotation (annotation), and `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).}\\n\\\\begin{tabular}\\n{cccccccc}\\n\\\\hline\\n\\\\textbf{Dataset} & \\\\textbf{Ref} & \\\\textbf{\\\\# Songs} & \\\\textbf{Type} & \\\\textbf{\\\\# Annotators}& \\\\textbf{Emotion model} & \\\\textbf{Labeling} & \\\\textbf{Labels}\\\\\\\\\\n\\\\hline\\nMIREX mood& & 2,038 & western and kpop & 2--3 & Clusters & annotation & dominant, distribution\\\\\\\\\\nCAL500 & & 500 & western & >2 & -- & annotation & dominant\\\\\\\\\\nSoundtracks& & 110 & instrumental &110 & self-defined, VA & annotation & distribution \\\\\\\\\\nMoodSwings& & 240 & western & Unknown & VA & annotation & distribution\\\\\\\\\\nDEAP & & 120 & western & 14--16 & VAD & annotation & average \\\\\\\\\\nAMG1608& & 1,608 & western & 15 & VA & annotation & distribution\\\\\\\\\\nDEAM & & 1,802 & diverse & 5--10 & VA & annotation & average, distribution \\\\\\\\\\nPMEmo & & 794 & western & 10 & VA & annotation & distribution \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\label{tab:MusicDataset}\\n\\\\end{table*}\",\n \"id\": \"6be4124c-0549-4c40-a1cf-f2efaaa9747b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ],\n [\n \"subsection\",\n \"Datasets for AC of Music\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets for AC of Music\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The target of video affective content computing is to recognize the emotions evoked by videos. In this field, it is necessary to construct a large benchmark dataset with precise emotional tags. However, the majority of existing research evaluate their proposed methods on their own collected datasets. The scarce video resources in those self-collected datasets, combined with the copyright restrictions result in limited accessibility for other researchers to reproduce existing work. Therefore, it is beneficial to summarize some publicly available datasets in this field. In general, publicly available datasets can be classified into two types: datasets consisting of videos only, such as movie clips or user generated videos, and datasets including both videos and audience's information.\",\n \"id\": \"d278bc85-2fef-47c9-b4d6-d1b6bb157aad\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ],\n [\n \"subsection\",\n \"Datasets for AC of Videos\"\n ]\n ],\n \"subsections\": [\n \"d6fc556f-5306-4560-b61f-84603258d71a\",\n \"7eba031f-c063-4798-bf20-621f78f3118a\"\n ],\n \"title\": \"Datasets for AC of Videos\"\n },\n {\n \"cite_extract_rate\": 0.1,\n \"cites\": [\n 1018\n ],\n \"content\": \"The LIRIS-ACCEDE dataset~ is one of the largest datasets in this area. Because it is collected under Creative Commons licenses, there are no copyright issues. The LIRIS-ACCEDE dataset is a living database in development. In order to fulfill requirements of different tasks, new data, features and tags are included. The LIRIS-ACCEDE dataset includes the Discrete LIRIS-ACCEDE collection and the Continuous LIRIS-ACCEDE collection in 2015 and was used for the MediaEval Emotional Impact of Movies tasks from 2015 to 2018.\\nThe Discrete LIRIS-ACCEDE collection~ includes 9,800 clips, which is derived from 40 feature films and 120 short films. Specifically, the majority of the 160 films are collected from the video platform VODO. The duration of all 160 films is about 73.6 hours in total. All of the 9,800 video clips last about 27 hour in total and the duration of each clip is between 8 to 12 seconds, which is long enough for viewers to feel emotions.\\nIn this collection, all the 9,800 video clips are labeled by values of valence and arousal.\\nThe Continuous LIRIS-ACCEDE collection~ differs from the Discrete LIRIS-ACCEDE collection in annotation type. Roughly speaking, the annotations for movie clips in the Discrete LIRIS-ACCEDE collection are global. It means that a whole 8 to 12 second video clip is represented by a single value of valence and arousal. This annotation type limits the possiblity for tracking emotions. To address this issue, 30 longer films are selected from the 160 films mentioned above. The total duration of all the selected films is about 7.4 hours. There are emotional annotations according to valence and arousal of each second of the films in the collection.\\nThe MediaEval Affective Impact of Movies collections between 2015 and 2018 are used for the MediaEval affective Impact of Movies tasks in each corresponding year. Specifically, the MediaEval 2015 Affective Impact of Movies~ includes two sub-tasks: affect detection and violence detection. The Discrete LIRIS-ACCEDE collection was used as the development set. And 1,100 additional video clips were extracted from 39 new movies and included. Indeed, all the new collected data were shared under Creative Commons licenses. In addition, three values were used to label the 10,900 video clips: a binary signal representing the presence of violence, a class tag of the excerpt for felt arousal and an annotation for felt valence.\\nThe MediaEval 2016 Affective Impact of Movies Task~ also includes two sub-tasks: Global emotion prediction and Continuous emotion prediction. The Discrete LIRIS-ACCEDE collection and the Continuous LIRS-ACCEDE collection were used as the development sets for the first and second sub-tasks, respectively. In addition, 49 new movies were chosen as the test sets. 1,200 short video clips from the new movies were extracted for the first task, and 10 long movies were selected for the second task. For the first sub-task, the tags include scores of valence and arousal for each whole movie clip. And for the second sub-task, scores of valence and arousal for each second of the movies are evaluated.\\n\\\\label{sec:VideoDatasets}\\nThe MediaEval 2017 Affective Impact of Movies Task~ is focused on long movies for two sub-tasks: valence/arousal prediction and Fear prediction. The Continuous LIRIS-ACCEDE collection was selected as the development set, and an additional 14 new movies were collected as the test set. The annotations contain a valence value and an arousal value. In addition, there are a binary value to represent whether the segment is supposed to induce fear or not for each 10-second segment.\\nThe MediaEval 2018 Affective Impact of Movies task~ is also dedicated to valence/arousal prediction and fear prediction. The Continuous LIRIS-ACCEDE collection and the test set of the MediaEval 2017 Emotional Impact of Movies task were used as the development set. In addition, 12 other movies selected from the set of the 160 movies mentioned in the Discrete LIRIS-ACCEDE part were used as test set. Specifically, for the first sub-task, there are annotations containing valence and arousal values for each second of the movies. And the beginning and ending times of each sequence in movies that induce fear are recorded for the second sub-task. \\nThe VideoEmotion dataset~ is a well-designed user-generated video collection. It contains 1,101 videos downloaded from web platforms, such as YouTube and Flickr. The annotations of the videos in this dataset are based on Plutchik's wheel of emotions~.\\nBoth the YF-E6 Dataset and the VideoStory-P14 Dataset are introduced in~. In order to collect the YF-E6 emotion dataset, six basic emotion types are used as keywords to search videos on YouTube and Flickr. There are 3,000 videos collected in the YF-E6 dataset totally. Then there were 10 annotators performing the labeling tasks. Only when all tags for a video clip were more than 50 percent consistent, the video clip was added to the dataset. Finally, the dataset includes 1,637 videos labeled with six basic emotion types. The VideoStory-P14 Dataset is based on the VideoStory dataset. Similar to the VideoEmotion Dataset, the keywords in Plutchik's Wheel of Emotions were used for the search process of the construction of the VideoStory dataset. Finally, there are 626 videos in the videoStory-P14 dataset with each having a unique emotion tag.\",\n \"id\": \"d6fc556f-5306-4560-b61f-84603258d71a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"d278bc85-2fef-47c9-b4d6-d1b6bb157aad\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ],\n [\n \"subsection\",\n \"Datasets for AC of Videos\"\n ],\n [\n \"subsubsection\",\n \"Datasets consisting of videos only\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets consisting of videos only\"\n },\n {\n \"cite_extract_rate\": 0.07692307692307601,\n \"cites\": [\n 1018\n ],\n \"content\": \"The DEAP dataset~ includes the EEG and peripheral physiological signals that are collected from 32 participants during watching 40 one-minute long excerpts of music videos. In addition, frontal face videos collected from 22 of the 32 participants are gathered. Annotators labeled each video according to the level of like/dislike, familiarity, arousal, valence, and dominance. Though the DEAP dataset is publicly available, it should be noted that it does not include the actual videos because of the licensing issues, but the links of videos are provided.\\nThe MAHNOB-HCI~ is a multimodal dataset including multi-class information recorded in response to video affective stimuli. Particularly, speeches, face videos, and eye gazes are recorded. In addition, two experiments were conducted to record both peripheral and central nervous system physiological signals from 27 subjects. In the first experiment, subjects were assigned to report their emotional responses to 20 affective induced videos, including the level of arousal, valence and dominance, and predictability as well as emotion categories. In the second experiment, the participants evaluated whether they agreed with the displayed labels after watching short videos and images. The dataset is available for academic use through a web-interface.\\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n \\\\caption{Released and freely available datasets for video emotion recognition, where `\\\\#Clips' and `Hours' respectively represent the total number and hours of video clips, `Type' means the genre of the videos in the dataset, `Emotion model' represents the labeling type, `Labeling' represents the method to obtain labels, such as human annotation (annotation) and keyword searching (keyword), and `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).}\\n\\\\resizebox{\\\\textwidth}{!}{\\n\\\\begin{tabular}\\n{C{1.8 cm} cccccccc}\\n\\\\hline\\n\\\\textbf{Dataset} & \\\\textbf{Ref} & \\\\textbf{\\\\#Clips}& \\\\textbf{Hours} & \\\\textbf{Type} & \\\\textbf{\\\\# Annotators} & \\\\textbf{Emotion model} & \\\\textbf{Labeling} & \\\\textbf{Labels} \\\\\\\\\\n\\\\hline\\nDiscrete LIRIS-ACCEDE & & 9,800 & 26.9 & film & - & VA & annotation & dominant \\\\\\\\\\nContinuous LIRIS-ACCEDE & & 30 & 7.4 & film & 10 (7f,3m) & VA & annotation & average \\\\\\\\\\nMediaEval 2015 & & 1,100 & - & film & - & 3 discrete VA values & annotation & dominant \\\\\\\\\\nMediaEval 2016 & & 1,210 & - & film & - & VA & annotation & distribution, average \\\\\\\\\\nMediaEval 2017 & & 14 & 8 & film & - & VA, fear & annotation & average \\\\\\\\\\nMediaEval 2018 & & 12 & 9 & film & - & VA, fear & annotation & average \\\\\\\\\\nVideoEmotion & & 1,101 & 32.7 & user-generated & 10 (5f,5m) & Plutchik & annotation & dominant \\\\\\\\\\nYF-E6 & & 1,637 & 50.9 & user-generated & 10(5f,5m) & Emkan & annotation & dominant \\\\\\\\\\nVideoStory-P14 & & 626 & - & user-generated & - & Plutchik & keyword & dominant \\\\\\\\\\nDEAP & & 120 & 2 & music video & - & VAD & annotation & personalized \\\\\\\\\\nMAHNOB-HCI & & - & - & multiple types & - & VAD, Ekman+neutral & annotation & personalized \\\\\\\\\\nDECAF & & 76 & - & music video/movies & - & VAD & annotation & personalized \\\\\\\\\\nAMIGOS & & 20 & - & movies collection & - & VAD, Ekman & annotation & personalized \\\\\\\\\\nASCERTAIN & & 36 & - & movies collection & 58 (21f,37m) & VA & annotation & personalized \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n}\\n\\\\label{tab:VideoDataset}\\n\\\\end{table*}\\nThe DECAF dataset~ consists of Infra-red facial video signals, Electrocardiogram (ECG), Magnetoencephalogram (MEG), horizontal Electrooculogram (hEOG) and Trapezius Electromyogram (tEMG), recorded from 30 participants watching 36 movie clips and 40 one-minute music videos, which are derived from the DEAP dataset~. The subjective feedback is based on valence, arousal, and dominance space. In addition, time-continuous emotion annotations for movie clips are also included in the dataset.\\nThe AMIGOS dataset~ includes multi-class affective data, individual and groups of viewers' responses to both short and long videos. The EEG, ECG, GSR, frontal, and full body video were recorded in two experimental settings, \\\\textit{i.e.}, 40 participants watching 16 short emotional clips and 4 long clips. The duration of each selected short videos is between 51 and 150 seconds, and the duration of each long excerpt is about 20 minutes. Finally, participants annotated the affective level of valence, arousal, control, familiarity, liking, and basic emotions.\\nBig-five personality scales and affective self-ratings of 58 users together with their EEG, ECG, GSR, and facial activity data were included in the ASCERTAIN dataset~ . The number of videos used as the stimulus is 36 and the length of each video clip is between 51 and 128 seconds. It is the first physiological dataset that is useful for both affective and personality recognition.\\nThe publicly available datasets for video affective content analysis are summarized in Table~\\\\ref{tab:VideoDataset}.\",\n \"id\": \"7eba031f-c063-4798-bf20-621f78f3118a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"d278bc85-2fef-47c9-b4d6-d1b6bb157aad\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ],\n [\n \"subsection\",\n \"Datasets for AC of Videos\"\n ],\n [\n \"subsubsection\",\n \"Datasets including both videos and audience's reactions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets including both videos and audience's reactions\"\n },\n {\n \"cite_extract_rate\": 0.44444444444444403,\n \"cites\": [\n 1019,\n 1020,\n 1021,\n 7322\n ],\n \"content\": \"\\\\label{sec:multimodalDatasets}\\nIn addition to audiovisual content and viewers' reactions, other modalities, such as language, also contain significant information for affective understanding of multimedia content. \\nVisual sentiment is the sentiment associated with the concepts depicted in images. Two datasets were developed through mining images associated with the adjective-noun pair (ANP) representations that have affective significance~. ANPs in were generated by first using seed terms from Plutchik's Wheel of Emotion~ to query Flickr\\\\footnote{\\\\url{https://www.flickr.com}} and YouTube\\\\footnote{\\\\url{https://www.youtube.com/}}. After mining the tags associated with visual content on YouTube and Flickr, adjective and noun candidates were identified through part-of-speech tagging. Then adjective and nouns were paired to create ANP candidates which were filtered by sentiment strength, named entities, and popularity. The Visual Sentiment Ontology (VSO), \\\\footnote{\\\\url{https://visual-sentiment-ontology.appspot.com}}, is the results of this process. Sentibank resulted in the creation of a set of photo-tweet sentiment dataset, with both visual and textual data with polarity labels, collected on Amazon Mechanical Turk\\\\footnote{\\\\url{http://www.ee.columbia.edu/ln/dvmm/vso/download/twitter_dataset.html}}. This work was later extended to form a multilingual ANP set and its dataset, in~\\\\footnote{\\\\url{http://mvso.cs.columbia.edu}}, containing 15,630 ANPs from 12 major languages and 7.37M images~. My Reaction When (MRW) dataset contains 50,107 video-sentence pairs crawled from social media, depicting physical or emotional reactions to the situations described in sentences~. The GIFs are sourced from Giphy\\\\footnote{\\\\url{https://giphy.com/}}. Even though there is no emotional labels, the language and visual associations are mainly based on sentiment which makes this dataset an interesting resource for affective content analysis.\\nCMU Multimodal Opinion Sentiment and Emotion Intensity (CMU-MOSEI) is a collection of multiple datasets for multimodal sentiment analysis and emotion recognition. This collection includes more than 23,500 sentence utterance videos from more than 1,000 people from YouTube~\\\\footnote{\\\\url{https://github.com/A2Zadeh/CMU-MultimodalSDK}}. All the videos are transcribed and aligned with audiovisual modalities. \\nA multimodal multi-party dataset for emotion recognition in conversation (MELD) was primarily developed for emotion recognition in multiparty interaction purposes~. MELD contains visual, audio, and textual modalities and includes 13,000 utterances from 1,433 dialogues from the TV-series Friends, with each utterance labeled with emotion and sentiment. \\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n \\\\caption{Released and freely available datasets for multimodal multimodal emotion recognition. Disc. for MELD corresponds to six Ekman emotions in addition to neutral. `Labeling' represents the method to obtain labels, such as human annotation (annotation), self-reported felt emotion and keyword searching (keyword), `Labels' means the detailed labels in the dataset, such as dominant emotion category (dominant), average dimension values (average), personalized emotion (personalized), and emotion distribution (distribution).} \\n\\\\resizebox{\\\\textwidth}{!}{\\n\\\\begin{tabular}\\n{c c c c c c c c}\\n\\\\hline\\n\\\\textbf{Dataset} & \\\\textbf{Ref} & \\\\textbf{\\\\#Samples}& \\\\textbf{Modalities} & \\\\textbf{Type} & \\\\textbf{Emotion model} & \\\\textbf{Labeling}&\\\\textbf{Labels}\\\\\\\\\\n\\\\hline\\nSentiBank tweet & & 603 & images, text & images & Sentiment&annotation& dominant\\\\\\\\\\nMVSO & & 7.36M & image, metadata & photos & Sentiment &automatic & average\\\\\\\\\\nCMU-MOSEI & & 23,500 & video, audio, text & YouTube videos & Sentiment &annotation& average\\\\\\\\\\nMELD & & 13,000 & video, audio, text & TV series & Sentiment, Disc. &annotation& dominant\\\\\\\\\\nCOGNIMUSE & & 3.5h & video, audio, text & movies & VA&annotation, self-report& dominant\\\\\\\\\\nVR & & 73 & video, audio & VR videos & VA&self-report&average \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n}\\n\\\\label{tab:MMDataset}\\n\\\\end{table*}\\nCOGNINMUSE is a collection of videos annotated with sensory and semantic saliency, events, cross-media semantics, and emotions~. A subset of 3.5h extracted from movies, including textual modality, are annotated on arousal and valence. \\n\\\\citeauthor{Li_VR} collected a dataset of 360 degrees virtual reality videos that can elicit different emotions~. Even though the dataset consists of 73 short videos, on average 183s long, it is one of the first datasets of its kind whose content understanding stays limited. These multimodal datasets are summarized in Table~\\\\ref{tab:MMDataset}.\",\n \"id\": \"72c21e7b-1086-4048-bb25-1bc86a5c6b31\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"21f5b1a5-8481-4ed1-83d5-a0396cc43cfa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ],\n [\n \"subsection\",\n \"Datasets for AC of Multimodal Data\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets for AC of Multimodal Data\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Image}\\nIn the early stages, AC researchers mainly worked on designing handcrafted features to bridge the affective gap. Recently, with the advent of deep learning especially convolutional neural networks (CNNs), current methods have shifted to an end-to-end deep representation learning. Motivated by the fact that the perception of image emotions may be dependent on different types of features~, some methods employ fusion strategies to jointly consider multiple features. In this section, we summarize and compare these methods. Please note that here we classify the directly extracted CNN features based on pre-trained deep models into handcrafted features category.\",\n \"id\": \"50a2810e-2777-44d0-85b6-12b160fc2d4d\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Images\"\n ]\n ],\n \"subsections\": [\n \"6469fc7b-8c4d-423d-88c3-f522313e7cbb\",\n \"c7360062-0ae0-4488-a262-c17a829a8a11\"\n ],\n \"title\": \"Affective Computing of Images\"\n },\n {\n \"cite_extract_rate\": 0.043478260869565,\n \"cites\": [\n 1022\n ],\n \"content\": \"\\\\label{sec:ImageHandcrafted}\\n\\\\textbf{Low-level Features} are difficult to be understood by viewers. These features are often directly derived from other computer vision tasks. Some widely extracted features include GIST, HOG2x2, self-similarity and geometric context color histogram features as in~, because of their individual power and distinct description of visual phenomena in a scene perspective.\\nCompared with the above generic features, some specific features derived from art theory and psychology have been designed. For example, \\\\citeauthor{machajdik2010affective}~\\\\shortcite{machajdik2010affective} extracted elements-of-art features, including \\\\emph{color} and \\\\emph{texture}. The MPEG-7 visual descriptors are employed in~, which include four color-related ideas and two texture-related ideas. How shape features in natural images influence emotions is investigated in~ by modeling the concepts of roundness-angularity and simplicity-complexity. \\\\citeauthor{sartori2015s}~\\\\shortcite{sartori2015s} designed two kinds of visual features to represent different color combinations based on Itten's color wheel.\\n\\\\textbf{Mid-level Features} contain more semantics, are more easily interpreted by viewers than low-level features, and thus are more relevant to emotions. \\\\citeauthor{patterson2012sun}~\\\\shortcite{patterson2012sun} proposed to detect 102 attributes in 5 different categories, including materials, surface properties, functions or affordances, spatial envelop attributes, and object presence. Besides these attributes, eigenfaces that may contribute to facial images are also incorporated in~. More recently, in~, SIFT features are first extracted as basic features, which are fed into bag-of-visual-words (BoVW) to represent the multi-scale blocks. Another mid-level representation is the latent topic distribution estimated by probabilistic latent semantic analysis.\\nHarmonious composition is essential in an artwork. Several compositional features, such as low depth of field, are designed to analyze such characteristics of an image~. Based on the fact that figure-ground relationships, color patterns, shapes and their diverse combinations are often jointly employed by artists to express emotions in their artworks, \\\\citeauthor{wang2013interpretable}~\\\\shortcite{wang2013interpretable} proposed to extract interpretable aesthetic features.\\nInspired by princiles-of-art, \\\\citeauthor{zhao2014exploring}~\\\\shortcite{zhao2014exploring} designed corresponding mid-level features, including \\\\emph{balance}, \\\\emph{emphasis}, \\\\emph{harmony}, \\\\emph{variety}, \\\\emph{gradation}, and \\\\emph{movement}. For example, Itten's color contrasts and the rate of focused attention are employed to measure \\\\textit{emphasis}.\\n\\\\begin{table*}\\n\\\\centering\\\\scriptsize\\n\\\\caption{Summary of the hand-crafted features at different levels for AC of images. `\\\\# Feat' indicates the dimension of each feature.}\\n\\\\resizebox{\\\\textwidth}{!}{\\n\\\\begin{tabular}{ccc p{8cm} c}\\n\\\\hline\\n\\\\textbf{Feature} & \\\\textbf{Ref} & \\\\textbf{Level} & \\\\multicolumn{1}{c}{\\\\textbf{Short description}} & \\\\textbf{\\\\# Feat} \\\\\\\\\\n\\\\hline\\nLOW\\\\_C & & low & GIST, HOG2x2, self-similarity and geometric context color histogram features & 17,032\\\\\\\\\\nElements & & low & color: mean saturation, brightness and hue, emotional coordinates, colorfulness, color names, Itten contrast, Wang's semantic descriptions of colors, area statistics; texture: Tamura, Wavelet and gray-level co-occurrence matrix & 97\\\\\\\\\\nMPEG-7 & & low & color: layout, structure, scalable color, dominant color; texture: edge histogram, texture browsing & $\\\\approx$200 \\\\\\\\\\nShape & & low & line segments, continuous lines, angles, curves & 219\\\\\\\\\\nIttenColor & & low & color co-occurrence features and patch-based color-combination features & 16,485\\\\\\\\\\n\\\\hline\\nAttributes & & mid & scene attributes & 102 \\\\\\\\\\nSentributes & & mid & scene attributes, eigenfaces & 109 \\\\\\\\\\nComposition & & mid & level of detail, low depth of field, dynamics, rule of thirds & 45 \\\\\\\\\\nAesthetics & & mid & figure-ground relationship, color pattern, shape, composition & 13 \\\\\\\\\\nPrinciples & & mid & principles-of-art: balance, contrast, harmony, variety, gradation, movement & 165\\\\\\\\\\nBoVW & & mid & bag-of-visual-words on SIFT, latent topics & 330 \\\\\\\\\\n\\\\hline\\nFS & & high & number of faces and skin pixels, size of the biggest face, amount of skin w.r.t. the size of faces & 4 \\\\\\\\\\nANP & & high & semantic concepts based on adjective noun pairs & 1,200 \\\\\\\\\\nExpressions & & high & automatically assessed facial expressions (anger, contempt, disgust, fear, happiness, sadness, surprise, neutral) & 8 \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n}\\n\\\\label{tab:ImageHandCraftedFeatures}\\n\\\\end{table*}\\n\\\\textbf{High-level Features} that represent the semantic content contained in images can be easily understood by viewers. We can also well recognize the conveyed emotions in images through these semantics. In the early years, simple semantic content including faces and skins contained in images are extracted in~. For the images that contain faces, facial expressions may directly determine the emotions. \\\\citeauthor{yang2010exploring}~\\\\shortcite{yang2010exploring} extracted 8 kinds of facial expressions as high-level features. They built compositional features of local Haar appearances by a minimum error based optimization strategy, which are embedded into an improved AdaBoost algorithm. For the images detected without faces, the experessions are simply set as \\\\emph{neutral}. Finally, they generated a 8 dimensional vector with each element representing the number of corresponding facial expressions.\\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n\\\\caption{Representative work on AC of images using hand-crafted features, where `Fusion' indicates the fusion strategy of different features, `cla, reg, ret, cla\\\\_p, dis\\\\_d, dis\\\\_c' in the Task column are short for classification, regression, retrieval, personalized classification, discrete distribution learning, continuous distribution learning (the same below), respectively, `Result' is the reported best accuracy for classification, mean squared error for regression, discounted cumulative gain for retrieval, F1 for personalized classification, and KL divergence for distribution learning (the first line~ is the result on sum of squared difference) on the corresponding datasets.}\\n\\\\begin{tabular}\\n{l p{2.5cm} l p{1.5cm} p{2.5cm} l R{2.5cm}} \\n\\\\hline\\n\\\\textbf{Ref} & \\\\textbf{Feature} & \\\\textbf{Fusion} & \\\\textbf{Learning} & \\\\textbf{Dataset} & \\\\textbf{Task} & \\\\textbf{Result} \\\\\\\\\\n\\\\hline\\n & Elements, Composition, FS & early & NB & IAPSa, Abstract, ArtPhoto & cla & 0.471, 0.357, 0.495\\\\\\\\\\n & MPEG-7 & -- & KNN & unreleased & cla & 0.827\\\\\\\\\\n & Shape, Elements & early & SVM, SVR & IAPSa; IAPS & cla; reg & 0.314; V-1.350, A-0.912\\\\\\\\%shape:0.299, V: 1.708, A: 0.943\\\\\\\\\\n & Segmented objects & -- & SL & IAPS, ArtPhoto & cla & 0.612, 0.610\\\\\\\\\\n & Sentributes & -- & SVM, LR & Tweet & cla & 0.824\\\\\\\\\\n & Aesthetics & -- & NB & Abstract, ArtPhoto & cla & 0.726, 0.631\\\\\\\\\\n & Principles & -- & SVM, SVR & IAPSa, Abstract, ArtPhoto; IAPS & cla; reg & 0.635, 0.605, 0.669; V-1.270, A-0.820\\\\\\\\\\n & LOW\\\\_C, Elements, Attributes, Principles, ANP, Expressions & graph & MGL & IAPSa, Abstract, ArtPhoto, GAPED, Tweet & ret & 0.773, 0.735, 0.658, 0.811, 0.701\\\\\\\\\\n & IttenColor & -- & SL & MART, devArt & cla & 0.751, 0.745\\\\\\\\\\n & BoVW & -- & MIL & IAPSa, Abstract, ArtPhoto & cla & 0.699, 0.636, 0.707\\\\\\\\\\n & IttenColor & -- & MC & MART, devArt & cla & 0.728, 0.761\\\\\\\\\\n\\\\hline\\n & GIST, Elements, Attributes, Principles, ANP, Expressions & graph & RMTHG & IESN & cla\\\\_p & 0.582\\\\\\\\\\n\\\\hline\\n & GIST, Elements, Principles & - & SSL & Abstract & dis\\\\_d & 0.134\\\\\\\\\\n & GIST, Elements, Attributes, Principles, ANP, deep features from AlexNet & weighted & WMMSSL & Abstract, Emotion6, IESN & dis\\\\_d & 0.482, 0.479, 0.478 \\\\\\\\\\n & ANP, VGG16 & - & ACPNN & Abstract, Emotion6, FlickrLDL, TwitterLDL & dis\\\\_d & 0.480, 0,506, 0,469, 0.555\\\\\\\\\\n & GIST, Elements, Attributes, Principles, ANP, AlexNet & weighted & WMMCPNN & Abstract, Emotion6, IESN & dis\\\\_d & 0.461, 0.464, 0.470\\\\\\\\\\n & GIST, Elements, Attributes, Principles, ANP, AlexNet & -- & MTSSR & IESN & dis\\\\_c & 0.436\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\label{tab:ImageHandCraftedMethods}\\n\\\\end{table*}\\nMore recently, the semantic concepts are described by adjective noun pairs (ANPs)~, which are detected by SentiBank~ or DeepSentiBank~. The advantages of ANP are that it turns a neutral noun into an ANP with strong emotions and makes the concepts more detectable, compared to nouns and adjectives, individually. A 1,200 dimensional vector representing the probability of the ANPs can form a feature vector.\\nTable~\\\\ref{tab:ImageHandCraftedFeatures} summarizes the above-mentioned hand-crafted features at different levels for AC of images. Some recent methods also extracted CNN features from pre-trained deep models, such as AlexNet~ and VGGNet~.\\nTo map the extracted handcrafted features to emotions,\\n\\\\textbf{Machine Learning Methods} are commonly employed. Some typical learning models include Naive Bayes (NB)~, support vector machine (SVM)~, $K$ nearest neighbor (KNN)~, sparse learning (SL)~, logistic regression (LR)~, multiple instance learning (MIL)~, and matrix completion (MC)~ for emotion classification , support vector regression (SVR)~ for emotion regression, and multi-graph learning (MGL)~ for emotion retrieval.\\nInstead of assigning the DEC to an image, some recent methods began to focus on the perception subjectivity challenge, \\\\textit{i.e.}, predicting personalized emotions for each viewer or learning emotion distributions for each image. The personalized emotion perceptions of a specified user after viewing an image is predicted in~, associated with online social networks. They considered different types of factors that may contribute to emotion recognition, including the images' visual content, the social context related to the corresponding users, the emotions' temporal evolution, and the images' location information. To jointly model these factors, they proposed rolling multi-task hypergraph learning (RMTHG), which can also easily hanlde the data incompleteness issue.\\nGenerally, the distribution learning task can be formulated as a regression problem, which slightly differs for different distribution categories (\\\\textit{i.e.}, discrete or continuous). For example, if emotion is represented by CE, the regression problem targets predicting the discrete probability of each emotion category with the sum equal to 1; if we represent emotion based on DES, the regression problem is typically transformed to the prediction of the parameters of specified continuous probability distributions. For the latter one, we usually need to firstly determine the form of continuous distributions, such as exponential distribution and Gaussian distribution. Some representative learning methods for emotion distribution learning of discrete emotions include shared sparse learning (SSL)~, weighted multimodal SSL (WMMSSL) , augmented conditional probability neural network (ACPNN)~, and weighted multi-model CPNN (WMMCPNN)~. Both SSL and WMMSSL can only model one test image each time, which is computationally inefficient. After the parameters are learned, ACPNN and WMMCPNN can easily predict the emotion distributions of a test image. Based on the assumption that the VA emotion labels can be well modeled by a mixture of 2 bidimensional Gaussian mixture models (GMMs), \\\\citeauthor{zhao2017continuous}~\\\\shortcite{zhao2017continuous} proposed to learn continuous emotion distributions in VA space by multi-task shared sparse regression (MTSSR). Specifically, the parameters of GMMs are regressed, including the mean vector and covariance matrix of the 2 Gaussian components as well as the mixing coefficients.\\nTable~\\\\ref{tab:ImageHandCraftedMethods} summarizes some representative work based on hand-crafted features. Generally, high-level features (such as ANP) can achieve better recognition performance for images with rich semantics, mid-level features (such as Principles) are more effective for artistic photos, while low-level features (such as Elements) perform better for abstract paintings.\",\n \"id\": \"6469fc7b-8c4d-423d-88c3-f522313e7cbb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"50a2810e-2777-44d0-85b6-12b160fc2d4d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Images\"\n ],\n [\n \"subsection\",\n \"Handcrafted Features-Based Methods for AC of Images\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Handcrafted Features-Based Methods for AC of Images\"\n },\n {\n \"cite_extract_rate\": 0.27272727272727204,\n \"cites\": [\n 1023,\n 1017,\n 1024\n ],\n \"content\": \"\\\\label{sec:ImageDeep}\\nTo deal with the situation where images are weakly labeled, a potentially cleaner subset of the training instances are selected progressively~. First, they trained an initial CNN model based on the training data. Second, they selected the training samples with distinct sentiment scores between the two classes with a high probability based on the prediction score of the trained model on the training data itself. Finally, the pre-trained AlexNet on ImageNet is fine-tuned to classify emotions into 8 categories by changing the last layer of the CNN from 1000 to 8~. Besides using the fully connected layer as classifier, they also trained an SVM classifier based on the extracted features from the second to the last layer of the pre-trained AlexNet model.\\nMulti-level deep representations (MldrNet) are learned in~ for image emotion classification. They segmented the input image into 3 levels of patches, which are input to 3 different CNN models, including Alexnet, aesthetics CNN (ACNN), and texture CNN (TCNN). The fused features are fed into multiple instance learning (MIL) to obtain the emotion labels. \\\\citeauthor{zhu2017dependency}~\\\\shortcite{zhu2017dependency} proposed to integrate the different levels of features by a Bidirectional GRU model (BiGRU) to exploit their dependencies based on MldrNet. They generated two features from the Bi-GRU model and concatenated them as the final feature representations. To enforce the feature vectors extracted from each pair of images from the same category to be close enough, and those from different categories to be far away, they proposed to jointly optimize a contrastive loss together with the traditional cross-entropy loss.\\nMore recently, \\\\citeauthor{yang2018retrieving}~\\\\shortcite{yang2018retrieving} employed deep metric learning to explore the correlation of emotional labels with the same polarity, and proposed a multi-task deep framework to optimize both retrieval and classification tasks. By considering the relations among emotional categories in the Mikels' wheel, they jointly optimized a novel sentiment constraint with the cross-entropy loss. Extending triplet constraints to a hierarchical structure, the sentiment constraint employs a sentiment vector based on the texture information from the convolutional layer to measure the difference between affective images. In~, \\\\citeauthor{yang2018retrieving} proposed a weakley supervised coupled convolutional neural network to exploit the discriminability of localized regions for emotion classification. Based on the image-level labels, a sentiment map is firstly detected in one branch with the cross spatial pooling strategy. And then the holistic and localized information are jointly combined in the other branch to conduct a classification task. The detected sentiment map can easily explain which regions of an image determine the emotions.\\nThe above deep methods mainly focused on the dominant emotion prediction. There are also some work on emotion distribution learning based on deep models. The very first work is a mixed bag of emotions, which trains a deep CNN regressor (CNNR) for each emotion category in Emotion6~ based on the AlexNet architecture. They changed the number of output nodes to 1 to predict a real value for each emotion category and replaced the Softmax loss with Euclidean loss. To ensure the sum of different probabilities to be 1, they normalized the predicted probabilities of all emotion categories. However, CNNR has some limitations. First, the predicted probability cannot be guaranteed to be non-negative. Second, the probability correlations among different emotions are ignored, since the regressor for each emotion category is trained independently. In~, \\\\citeauthor{yang2017joint} designed a multi-task deep framework based on VGG16 by jointly optimizing the cross-entropy loss for emotion classification and Kullback-Leibler (KL) divergence loss for emotion distribution learning. To match the single emotion dataset to emotion distribution learning settings, they transformed each single label to emotion distribution with emotion distances computed on Mikels' wheel~. By extending the size of training samples, this method achieves the state-of-the-art performance for discrete emotion distribution learning.\\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n\\\\caption{Representative work on deep learning based AC of images, where `Pre' indicates whether the network is pre-trained using ImageNet, `\\\\# Feat' indicates the dimension of last feature mapping layer before the emotion output layer, `Cla' indicates the classifier used after the last feature mapping with default Softmax, `Loss' indicates the loss objectives (besides the common cross-entropy loss for classification), and `Result' is the reported best accuracy for classification, discounted cumulative gain for retrieval, and KL divergence for distribution learning on the corresponding datasets.}\\n\\\\resizebox{\\\\textwidth}{!}{\\n\\\\begin{tabular}\\n{c C{1.5cm} c C{1cm} cc C{3cm} C{0.4cm} C{2cm}} \\n\\\\hline\\n\\\\textbf{Ref} & \\\\textbf{Base net} & \\\\textbf{Pre} & \\\\textbf{\\\\#Feat} & \\\\textbf{Cla} & \\\\textbf{Loss} & \\\\textbf{Dataset} & \\\\textbf{Task} & \\\\textbf{Result} \\\\\\\\\\n\\\\hline\\n & self-defined & no & 24 & -- & -- & FlickrCC & cla & 0.781\\\\\\\\\\n & AlexNet & yes & 4,096 & SVM & -- & FI, IAPSa, Abstract, ArtPhoto & cla & 0.583, 0.872, 0.776, 0.737\\\\\\\\\\n & AlexNet, ACNN, TCNN & yes & 4,096, 256, 4,096 & MIL & -- & \\\\tiny{FI, IAPSa, Abstract, ArtPhoto, MART} & cla & \\\\tiny{0.652, 0.889, 0.825, 0.834, 0.764}\\\\\\\\\\n & self-defined & no & 512 & -- & contrastive & FI, IAPSa, ArtPhoto & cla & 0.730, 0.902, 0.855\\\\\\\\\\n & GoogleNet-Inception & yes & 1,024 & -- & sentiment & FI, IAPSa, Abstract, ArtPhoto & cla; ret & 0.676, 0.442, 0.382, 0.400; 0.780, 0.819, 0.788, 0.704\\\\\\\\\\n & ResNet-101 & yes & 2,048 & -- & -- & FI, Tweet & cla & 0.701, 0.814\\\\\\\\\\n\\\\hline\\n & AlexNet & yes & 4,096 & -- & Euclidean & Emotion6 & dis\\\\_d & 0.480\\\\\\\\\\n & VGG16 & yes & 4,096 & -- & KL & Emotion6, FlickrLDL, TwitterLDL & dis\\\\_d & 0.420, 0,530, 0,530\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n}\\n\\\\label{tab:ImageDeepMethods}\\n\\\\end{table*}\\nThe representative deep learning based methods are summarized in Table~\\\\ref{tab:ImageDeepMethods}. The deep representation features generally perform better than the hand-crafted ones, which are intuitively designed for specific domains based on several small-scale datasets. However, how the deep features correlate to specific emotions is unclear.\",\n \"id\": \"c7360062-0ae0-4488-a262-c17a829a8a11\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"50a2810e-2777-44d0-85b6-12b160fc2d4d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Images\"\n ],\n [\n \"subsection\",\n \"Deep Learning-Based Methods for AC of Images\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Learning-Based Methods for AC of Images\"\n },\n {\n \"cite_extract_rate\": 0.0625,\n \"cites\": [\n 1025\n ],\n \"content\": \"\\\\label{sec:Music}\\nMusic emotion recognition (MER) strives to identify emotion expressed by music and subsequently predict listener's felt emotion from acoustic content and music metadata, \\\\textit{e.g.}, lyrics, genre, \\\\textit{etc.} Emotional understanding of music have applications in music recommendation and is particularly useful for producing music retrieval. An analysis of search queries from creative professionals showed that 80\\\\% contain emotional terms, showing emotions prominence in that field~. A growing number of work have tried to address emotional understanding of music from acoustic content and metadata (see~ for earlier reviews on this topic). \\nEarlier work on emotion recognition from music relied on extracting acoustic features similar to the ones used in speech analysis, such as audio energy and formants. Acoustic features describe attributes related to musical dimensions. Musical dimensions include melody, harmony, rhythm, dynamics, timbre (tone color), expressive techniques, musical texture, and musical form~, as shown in Table~\\\\ref{tab:MusicFeatures}. Some also add energy as a musical feature which is important for MER~. \\nMelody is a linear succession of tones and can be captured by features representing key, pitch and tonality. Among others, chroma is often used to represent melodic features~. \\nHarmony is how the combination of various pitches are processed during hearing. Understanding harmony involves chords or multiple notes played together. Examples of acoustic features capturing harmony include chromagram, key, mode, and chords~.\\nRhythm consists of repeated patterns of musical sounds, \\\\textit{i.e.}, notes and pulses that can be describes in terms of tempo and meter. Higher tempo songs often induce higher arousal and fluent rhythm is associated with higher valence and firm rhythm is associated with sad songs~. Mid-level acoustic features, such as onset rate, tempo and beat histogram, can represent rhythmic characteristics of music. \\nDynamics of music involve the variation in softness or loudness of notes which include change of loudness (contrast) and emphasis on individual sounds (accent)~. Dynamics of music can be captured by changes in acoustic features related to energy such as root mean square (RMS) energy.\\nTimbre is the perceived sound quality of musical notes. Timbre is what differentiates different voices and instruments playing the same sound. Acoustic features capturing timbre, such as MFCC and spectrum shape, describe sound quality~. Acoustic features describing timbre include MFCC, spectral features (centroid, contract, flatness), and zero crossing rate~.\\nExpressive techniques are the way a musical piece is played including tempo and articulation~. Acoustic features, such as tempo, attack slope, and time, can be used to describe this dimension.\\nMusical texture is how rhythmic, melodic, and harmonic features are combined in music production~. It is related to the range of tones played at the same time. Musical form describes how a song is structured, such as introduction verse and chorus~. \\nEnergy whose dynamics are described in music dynamic features is strongly associated with arousal perception. \\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n \\\\caption{Musical dimensions and acoustic features describing them.}\\n\\\\begin{tabular}\\n{ l p{10cm} }\\n\\\\hline\\n\\\\textbf{Musical dimension} & \\\\textbf{Acoustic features} \\\\\\\\\\n\\\\hline\\nMelody & Pitch \\\\\\\\%\\\\hline\\nHarmony & chromagram, chromagram peak, key, mode, key clarity, harmonic, change, chords \\\\\\\\%\\\\hline\\nRhythm & tempo, beat histograms, rhythm regularity, rhythm strength, onset rate\\\\\\\\%\\\\hline\\nDynamics and loudness & RMS energy, loudness, timpral width\\\\\\\\%\\\\hline\\nTimbre & MFCC, spectral shapres (centroid, shape, spread, skewness, kurtosis, contrast and flatness), brightness, rolloff frequency, zero crossing rate, spectral contrast, auditory modulation features, inharmonicity, roughness, dissonance, odd to even harmonic ratio \\\\\\\\%\\\\hline\\nMusical form & Similarity Matrix (similarity between all possible frames) \\\\\\\\%\\\\hline\\nTexture & attack slope, attack time\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\label{tab:MusicFeatures}\\n\\\\end{table*}\\nThere are a number of toolboxes available for extracting acoustic features from music that can be used for music emotion recognition. \\nMusic Analysis, Retrieval and Synthesis for Audio Signals (Marsyas)~ is an open source framework developed in C++ that supports extracting a large range of acoustic features with music information retrieval applications in mind, including time-domain zero-crossings, spectral centroid, rolloff, flux, and Mel-Frequency Cepstral Coefficients (MFCC) \\\\textit{etc.}\\nMIRToolbox~ is an open source toolbox implemented in MATLAB for music information retrieval applications. MIRToolbox offers the ability to extract a comprehensive set of acoustic features at different levels including features related to tonality, rhythm, and structures.\\nSpeech and music interpretation by large-space extraction or OpenSMILE~ is an open source software developed in C++ with the ability to extract a large number of acoustic features for speech and music analysis in real-time. \\nLibROSA~ is a Python package for music and audio analysis. It is mainly developed with music information retrieval application in mind and supports importing from different audio sources and extracting musical features such as onsets chroma and tempo in addition to the low-level acoustic features. \\nESSENTIA~ is an open source library developed in C++ with Python interface that is developed for audio analysis. ESSENTIA contains an extensive collection of algorithms supporting audio input/output functionality, standard digital signal processing blocks, statistical characterization of data, and a large set of spectral, temporal, tonal and high-level music descriptors. \\nMusic emotion recognition either attempts to classify songs or excerpt into categories (classification) or estimate their expressed emotions on continuous dimensions (regression). The choice of machine learning model in music emotion recognition depends on the emotional representation used. Mood clusters~, dimensional representations such as arousal, tension and valence as well as music specific emotion representation can be used. An analysis of the methods proposed for MediaEval ``Music in Emotion'' task submissions revealed that using deep learning accounted for the superior performance for emotion recognition much more than the choice features~. Recent methods for emotion recognition in music rely on deep learning and often use spectrogram features that are converted to images~. \\\\citeauthor{aljanaki2018data} proposed learning musically meaningful mid-level perceptual features that can describe emotions in music~. They demonstrated that perceptual features such as melodiousness, modality, rhythmic complexity and dissonance can describe a large portion of emotional variance in music both in dimensional representation and MIREX clusters. They also trained a deep convolutional neural network to recognize these mid-level attributes. There have been also work attempting to use lyrics in addition to acoustic content for recognizing emotion in music~. However, lyrics are copyrighted and not easily available which hinders further work in this direction.\",\n \"id\": \"6a4a2789-5122-4ca2-8207-b0cf83ceda85\",\n \"level\": \"section\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Music\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Affective Computing of Music\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1026\n ],\n \"content\": \"\\\\label{sec:Video}\\nCurrently, the features used in affective video content analysis are mainly from two categories~. One is considering the stimulus of video content and extracting the features reflecting the emotions conveyed by the video content itself. And the other is extracting features from the viewers. Features extracted from the video content are content-based features, and features formed from the signals of the viewers' responses are viewer-related features.\",\n \"id\": \"14b7a32f-91ab-4f67-baaf-fabb6149902f\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ]\n ],\n \"subsections\": [\n \"0091d8de-1e85-4000-afb6-6cc1c3dcc3de\",\n \"849377a8-415a-42d7-8a8d-eb0aab34165c\",\n \"930f941a-bd9c-4939-bb1e-c81c0b2dac96\",\n \"cefa92e9-5af7-4000-8441-381e9420e9cf\",\n \"dc8e2b0f-60af-45ac-bc34-2b55fa2b0f2b\"\n ],\n \"title\": \"Affective Computing of Videos\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Generally speaking, the video content comprise of a series of ordered frames as well as corresponding audio signals. Therefore, it is natural to extract features from these two modalities. The audiovisual features can further be divided into low-level and mid-level according to their ability to describe the semantics of video content.\",\n \"id\": \"0091d8de-1e85-4000-afb6-6cc1c3dcc3de\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"14b7a32f-91ab-4f67-baaf-fabb6149902f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ],\n [\n \"subsection\",\n \"Content-related Features\"\n ]\n ],\n \"subsections\": [\n \"688b2f4a-8d63-4f40-b57e-7f57ac458bc1\",\n \"1257a6cb-e4fc-4f9d-88bf-5637943576d0\"\n ],\n \"title\": \"Content-related Features\"\n },\n {\n \"cite_extract_rate\": 0.05,\n \"cites\": [\n 1018\n ],\n \"content\": \"Commonly, the low-level features are directly computed from the raw visual and audio content, and usually carry no semantic information. As for visual content, color, lighting, and tempo are important elements that can endow the video with strong emotional rendering and further give viewers direct visual stimuli. In many cases, computations are conducted over each frame of the video, and the average values of the computational results of the overall video are considered as visual features. Specifically, the color-related features often contain the histogram and variance of color~, the proportions of color~, the number of white frame and fades~, the grayness~, darkness ratio, color energy~, brightness ratio and saturation~, \\\\textit{etc}. In addition, the differences of dark and light can be reflected by the lighting key, which is used to evoke emotions in video and draw the attention of viewers by creating an emotional atmosphere~. As for the tempo-related features, properties of shot can reinforce the expression of video, such as shot change rate and shot length variance~ according to movie grammar. To better take advantage of the temporal information of the video, the motion vectors have been computed as features in~. Since the optical flow can characterize the influence of camera motions, the histogram of optimal flow matrix (HOF) has been computed as features in~. Additionally, \\\\citeauthor{yi2018multi}~\\\\shortcite{yi2018multi} traced motion key points at multiple spatial scales and computed the mean motion magnitude of each frame as features.\\nTo represent audio content, pitch, zero crossing rates (ZCR), Mel frequency cepstrum coefficients (MFCC), and energy are the most popular features . In particular, the MFCC~ and its $\\\\Delta{MFCC}$ are used to characterize emotions in video clips frequently; while the derivatives and statistics (min, max,mean) of MFCC or $\\\\Delta{MFCC}$ are also explored widely. As for pitch, ~ shows that pitch of sound is associated closely with some emotions, such as anger with higher pitch and sadness with lower standard deviation of pitch. Similar situation can also occur in the energy~. For example, the total energy of anger or happiness is higher than the counterpart of unexciting emotions. ZCR~ is used to separate different types of audio signals, such as music, environmental sound and speech of human. Besides these frequent related features, audio flatness~, spectral flux~, delta spectrum magnitude, harmony~, band energy ratio, spectral centroid~, and spectral contrast~ are also utilized.\\nEvidently, the aforementioned features are mostly handcrafted. With the emergence of deep learning, features can be automatically learned through deep neural networks. Some pre-trained convolutional neural networks (CNNs) are used to learn static representations from every frame or some selected key frames, while a Long-short term memory (LSTM) is exploited to capture dynamic representations existing in videos.\\nFor instance, in~, an AlexNet with seven fully-connected layers trained on 2600 ImageNet classes is used to learn features. A Convolutional Auto-Encoder (CAE) is designed to ensure the CNNs can extract the visual features effectively in~. \\\\citeauthor{ben2018deep}~ first used the pre-trained ResNet-152 to extract feature vectors. And then, these vectors are fed into an LSTM according to their temporal order to extract high-order representations. Pre-trained model, SoundNet, is utilized to learn audio features. Because the expressive emotions of video are induced and communicated by the protagonist in video in many cases, the features of protagonist are extracted from the key frame by a pre-trained CNN and used in video affective analysis in~. In addition to the protagonist, other objects in each frame of video also give insights into emotional expression of video. For example, in~, \\\\citeauthor{shukla2018looking} removed the non-gaze regions from video frames (Eye ROI) and built the coarse grained scene structure remaining gist information by Guassian filter with variance. After the operations above in~, the next video affective analysis may pay more attention to important information and reduce unnecessary noise.\",\n \"id\": \"688b2f4a-8d63-4f40-b57e-7f57ac458bc1\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"0091d8de-1e85-4000-afb6-6cc1c3dcc3de\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ],\n [\n \"subsection\",\n \"Content-related Features\"\n ],\n [\n \"subsubsection\",\n \"Low-level features\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Low-level features\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Unlike low-level features, mid-level features often contain semantic information. For example, EmoBase10 feature depicting audio cues is computed in~. \\\\citeauthor{hu2016multi}~ proposed a method of combining the audio and visual features to model contextual structures of the key frames selected from video. This can produce a kind of so-called multi-instance sparse coding (MI-SC) for next analysis. In addition, the lexical features~ are extracted from the dialogues of speakers by using a natural language toolkit. These features can reflect the emotional changes in videos and can also represent a certain emotional expression in overall videos. \\\\citeauthor{muszynski2019recognizing}~ used aesthetic movie highlights related to occurrences of meaningful movie scenes to define some experts. These features produced by experts are more knowledgeable and abstract for video affective analysis, especially movies. HHTC features, which are computed on the basis of combination of Huang Transform in visual-audio and cross-correlation features, are proposed in~.\",\n \"id\": \"1257a6cb-e4fc-4f9d-88bf-5637943576d0\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"0091d8de-1e85-4000-afb6-6cc1c3dcc3de\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ],\n [\n \"subsection\",\n \"Content-related Features\"\n ],\n [\n \"subsubsection\",\n \"Mid-level features\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Mid-level features\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Viewer-relatedFeatures}\\nBesides the content-related features, viewers' facial expressions and changes of physiological signals evoked by content of videos are the most common sources for extracting viewer-related features. \\\\citeauthor{mcduff2017large}~ coded the facial actions of viewers for further affective video analysis. Among various physiological signals, electrocardiography (ECG), galvanic skin response (GSR), electroencephalography (EEG) are the mostly frequently ones and their statistical measures, such as mean, median, spectral power bands, \\\\textit{etc}., are often recommended as features. \\\\citeauthor{Wang2015Emotion}~ used EEG signals to construct a new EEG feature with the assistance of the relationship among video content by exploiting canonical correlation analysis (CCA).\\nIn~, some viewer-related features are extracted from the whole pupil dilation ratio time-series without the differences among pupil diameter in human eyes, such as its average and derivation for global features as well as the four spectral power bands for local features.\\nIn addition to viewers' responses mentioned above, the comments or other textual information produced by viewers can also reflect their attitudes or emotional reactions toward the videos. In the light of this, it is reasonable to consider users' textual comments or other textual information to extract features. In~, the ``sentiment analysis'' module using Unigrams and Bigrams is built to learn comment-related features of the collected data according to the YouTube link provided by the DEAP dataset.\",\n \"id\": \"849377a8-415a-42d7-8a8d-eb0aab34165c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"14b7a32f-91ab-4f67-baaf-fabb6149902f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ],\n [\n \"subsection\",\n \"Viewer-related Features\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Viewer-related Features\"\n },\n {\n \"cite_extract_rate\": 0.125,\n \"cites\": [\n 1028,\n 1026,\n 1027\n ],\n \"content\": \"After feature extracting, a classifier or a regressor is used to obtain emotional analysis results. For classification, there are several frequently used classifiers, including support vector machines (SVM)~, Naive Bayes (NB)~, Linear Discriminant Analysis~, logistic regression (LR)~, and ensemble learning~, \\\\textit{etc}.\\nRecent work show that the SVM-based methods are very popular for affective video content analysis due to its simplicity, max-margin training property, and use of kernels ~. For example, \\\\citeauthor{yi2018multi}' work ~ demonstrated that linear SVM is more suitable for classification than RBM, MLP, and LR . In~, LDA, linear SVM (LSVM), and Radial Basis SVM (RSVM) classifiers are employed in emotion recognition experiments, and the RSVM obtained the best F1 scores. In~, both Navie Bayes and SVM are used as classifiers in unimodal and multimodal conditions. In the unimodal experimental condition, NB is not better than SVM. And the fusion results showed that SVM is much better than NB in multimodal situations. However, SVM also has its shortages, such as the difficulty of selecting suitable kernel functions. Indeed, SVM is not always the best choice. In~, the results demonstrated that ensemble learning outperforms SVM in terms of classification accuracy. Ensemble learning has acquired a lot of attention in many fields because of its accuracy, simplicity, and robustness. In addition, in~, LR is adopted as the classifier for its effectiveness and simplicity. In fact, LR is used frequently in many transfer learning tasks.\\nHowever, all the classifiers mentioned above are not able to capture the temporal information. Some other methods try to use temporal information. For example, \\\\citeauthor{gui2018implicit}~ combined SVM and LSTM to predict emotion labels. Specifically, global features and sequence features are proposed to represent the pupillary response signals. Then a SVM classifier is trained with the global features and a LSTM classifier is trained with the sequence features. Finally, a decision fusion strategy is proposed to combine these two classifiers.\\n\\\\begin{table*}[!t]\\n\\\\centering\\\\scriptsize\\n\\\\caption{Representative work on AC of videos using kinds of features, where $P_{a}$, $P_{v}$, $MSE_{a}$, $MSE_{v}$, $Acc_{a}$, $Acc_{v}$, $Acc$ and $MAP$ indicates the Pearson correlation coefficients of arousal and valence, the mean sum error of arousal and valence, the accuracy of arousal and valence, the average accuracy and mean average precision respectively. `statistics' means (min, max, mean).}\\n\\\\resizebox{\\\\textwidth}{!}{\\n\\\\begin{tabular}\\n{p{0.4 cm} p{4 cm} p{0.7 cm} p{2 cm} p{2 cm} p{0.3 cm} R{2cm}} \\n\\\\hline\\n\\\\textbf{Ref} & \\\\textbf{Feature} & \\\\textbf{Fusion} & \\\\textbf{Learning} & \\\\textbf{Dataset} & \\\\textbf{Task} & \\\\textbf{Result} \\\\\\\\\\n\\\\hline\\n & Mel frequency spectral; MFCC, Chroma and their derivatives and statistics; Audio compressibility; Harmonicity; Shot frequency; HOF and statistics; Histogram of 3D HSV and statistics; Video compressibility; Histogram of facial area & decision & LSTM & Dataset described by Malandrakis & reg & $P_{a}$:$0.84 \\\\pm 0.06$ $MSE_{a}$:$0.08 \\\\pm 0.04$ $P_{v}$:$0.50 \\\\pm 0.14$ $MSE_{v}$:$0.21 \\\\pm 0.06$\\\\\\\\\\n & Average, standard deviation and four spectral power bands of pupil dilation ratio time-series & decision & SVM, LSTM & MAHNOB-HCI & cla & $Acc_{a}$:0.730, $Acc_{v}$:0.780\\\\\\\\\\n & Audio and visual deep features from pretrained model & feature & CNN, LSTM, SVM & PMIT & cla & $MAP$:0.0.2122\\\\\\\\\\n & MFCC; Color values; HoG; Dense trajectory descriptor; CNN-learned features & decision & CNN, SVM, Ensemble & DEAP & cla & $Acc$: 0.81, 0.49 \\\\\\\\\\n & HHTC features & feature & SVR & Discrete LIRIS-ACCEDE & reg & $MSE_{a}$:0.294, $MSE_{v}$: 0.290 \\\\\\\\\\n & Statistical measures (such as mean, median, skewness kurtosis) for EEG data, power spectral features, ECG, GSR, Face/Head-pose & decision & SVM/NB & music excerpts~ & cla & F1 (v: 0.59, 0.58, a: 0.60, 0.57) \\\\\\\\\\n & Time-span visual and visual features & feature & CNN Opensmile toolbox & music excerpts~ & cla & $MSE_{a}$:0.082 $MSE_{v}:0.071$\\\\\\\\\\n & tempo; pitch; zero cross; roll off; MFCCs; Saturation; Color heat; Shot length feature; General preferences; Visual excitement; Motion feature; fMRI feature & feature & DBM SVM & TRECVID & cla & - \\\\\\\\\\n & Colorfulness; MFCC; CNN-learned features from the keyframes containing protagonist & decision & CNN, SVM, SVR & LIRIS-ACCEDE, PMSZU & cla/reg & - \\\\\\\\\\n & Multi-frame motion vectors & decision & CNN & SumMe, TVsum, Continuous LIRIS-ACCEDE & reg & - \\\\\\\\\\n & The median of the L values in Luv space; means and variances of components in HSV space; texture feature; mean and standard deviation of motions between frames in a short; MFCC; Spectral power; mean and variance of the spectral centroids; Time domain zero crossings rate; Multi-instance sparse coding & feature & SVM & Musk1, Musk2, Elephant, Fox, Tiger & cla & $Acc$: 0.911, 0.906, 0.885, 0.627, 0.868\\\\\\\\\\n & Lighting key; Color; Motion vectors; ZCR; energy; MFCC; pitch; Textual features & decision & SMO, Navie Bayes & DEAP & cla & F1:0.849 0.811 $Acc$: 0.911 0.883 \\\\\\\\\\n & Key lighting; Grayness; Fast motion; Shot chanage rate; Shot length variation; MFCC; CNN-learned features; power spectral density; EEG; ECG; respiation; galvanic skin resistance & feature & SVM & DEAP & cla & $Acc_v$:0.7 0.7 0.7125, $Acc_a$:0.6876 0.7 0.8 F1 (A:0.664 0.687 0.789) \\\\\\\\\\n & MFCC; ZCR; energy; pitch, color histograms; lighting key; motion vector & decision & SVM, Navie Bayes & DEAP & cla & F1: 0.869, 0.846 $Acc$: 0.925, 0.897 \\\\\\\\\\n & CNN feature, low-level audio visual features, EEG & decision & LDA, LSVM, RSVM & Dataset introduced by the authors & cla & - \\\\\\\\\\n & MKT; ConvNets feature; EmoLarge; IS13; MFCC; EmoBase10; DSIFT; HSH & decision & SVM, LR, RBM, MLP & MediaEval 2015, 2016 Affective Impact of Movies & cla & $Acc_a$:0.574, $Acc_v$:0.462\\\\\\\\\\n & CNN feature & - & SVM, LDA & dataset in~ & cla & - \\\\\\\\\\n & CNN feature & - & SVR & LIRIS-ACCEDE & reg & $MSE_{a}$: 0.021, $MSE_{v}$: 0.027\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n}\\n\\\\label{tab:VideoHandCraftedMethods}\\n\\\\end{table*}\\nA regressor is needed when mapping the extracted features to the continuous dimensional emotion space. Recently, one of the most popular regression method is support vector regression (SVR)~. For example, in~, video features like audio, color, aesthetic are fed into SVR in the SVR-Standard experiment. And in the SVR-Transfer learning experiment, the pre-trained CNN is treated as a feature extractor. The CNN's outputs are used as the input to the SVR. The experimental results showed that the SVR-Transfer learning outperforms other methods. Indeed, the various kernel functions in SVR provide a stronger adaptability.\",\n \"id\": \"930f941a-bd9c-4939-bb1e-c81c0b2dac96\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 24,\n \"parent_id\": \"14b7a32f-91ab-4f67-baaf-fabb6149902f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ],\n [\n \"subsection\",\n \"Machine Learning Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Machine Learning Methods\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In total, there are two fusion strategies for multimodal information: feature-level fusion and decision-level fusion. Feature-level fusion means that the multimodal features are combined and then used as the input of a classifier or a regressor. Decision-level fusion fuses several results of different classifiers, and the final results are computed according to the fusion methods.\\nOne way of feature-level fusion is implemented by feature accumulation or concatenation~. In~, two feature vectors for visual and audio data are averaged as the global genre representations. In~, multi-class features are concatenated to generate a high dimensional joint representation. Some machine learning methods are also employed to learn joint features~. In~, a two-branch network is used to combine the visual and audio features. The outputs of the two-branch network are then fed into a classifier, and the experiment results showed that the joint features outperform other methods. In~, the low-level audio-visual features and fMRI-derived features are fed into multimodal DBM to learn joint representations. The target of this method is to learn the relation between audio-visual features and fMRI-derived features. In~, PCA is used to learn the multimodal joint features. In~, canonical correlation analysis (CCA) is used to construct a new video feature space with the help of EEG features and a new EEG feature space with the assistance of video content, so only one modality is needed to predict emotion during the testing process.\\nBy combining the results of different classifiers, decision-level fusion strategy is able to achieve better results~. In~, linear fusion and SVM-based fusion techniques are explored to combine outputs of several classifiers. Specifically, the output of each classifier has its own weight in linear fusion. The final result is the weighted sum of all outputs. In SVM-based fusion, the outputs of unimodal classifiers are concatenated together. And then the higher level representations for each video clip are fed into a fusion SVM to predict the emotion. Based on these results, linear fusion is better than SVM-based fusion. In~, linear fusion is also used to fuse the outputs of multiple classifiers. The differences among these linear fusion methods depend on the distribution of weights.\",\n \"id\": \"cefa92e9-5af7-4000-8441-381e9420e9cf\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"14b7a32f-91ab-4f67-baaf-fabb6149902f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ],\n [\n \"subsection\",\n \"Data Fusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Fusion\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1029\n ],\n \"content\": \"\\\\label{sec:Deep learning method}\\nIn tradition, video emotional recognition includes two steps, \\\\textit{i.e.}, feature extraction step and regression or classification step. Because of the lack of consensus on the most relevant emotional features, we may not be able to extract the best features for the problem at hand. As a result, this two-step mode has hampered the development of affective video content analysis. In order to solve this problem, some methods based on end-to-end training frameworks are proposed. \\\\citeauthor{khorrami2016deep} combined CNN and RNN to recognize the emotional information of videos. According to their method, a CNN is trained using frame facial images sampled from videos to extract features. Then the features are fed into a RNN to perform continuous emotion recognition. In~, a single network using ConvLSTM is proposed, where videos are input to the network and the predicted emotional information is output directly. In fact, due to the complexity of CNNs and RNNs, the training of these frameworks needs large amounts of data. However, in video affective content analysis, the samples in existing datasets are usually limited. This is the reason why end-to-end methods are still less common compared to the traditional two step methods, despite of their influential potentials.\",\n \"id\": \"dc8e2b0f-60af-45ac-bc34-2b55fa2b0f2b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"14b7a32f-91ab-4f67-baaf-fabb6149902f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Videos\"\n ],\n [\n \"subsection\",\n \"Deep Learning Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Learning Methods\"\n },\n {\n \"cite_extract_rate\": 0.185185185185185,\n \"cites\": [\n 7,\n 1030,\n 1032,\n 1684,\n 1031\n ],\n \"content\": \"\\\\label{sec:multimodal}\\nIn this section, we survey the work that analyze multimodal data beyond audiovisual content. Most of the existing work on affective understanding of multimedia rely on one modality, even when additional modalities are available, for example in videos~. Earlier work on emotional understanding of multimedia used hand crafted features from different modalities that are fused at feature or decision levels~. The more recent work mainly use deep learning models~. \\nLanguage is a commonly used modality in addition to vision and audio. There is a large body of work on text-based sentiment analysis~. Sentiment analysis from text is well-established and is deployed at scale in industry at a broad set of applications involving opinion mining~. With the shift toward an increasingly multimodal social web, multimodal sentiment analysis is becoming more relevant. For example, vloggers post their opinions on YouTube, and photos commonly accompany user posts on Instagram and Twitter. \\nAnalyzing text for emotion recognition requires representing terms by features. Lexically-based approaches are one of the most popular methods for text-based emotion recognition. They involve using knowledge of words' affect for estimating document or content's affect. Linguistic Inquiry and Word Count (LIWC) is a well-known lexical tool that matches the terms in a document with its dictionary and generates scores along different dimensions including affective and cognitive constructs such as ``present focus'' and ``positive emotion''~. The terms in each category or selected by experts is extensively validated on different content. AffectNet is another notable lexical resource which includes a semantic netowrk of 10,000 items with representations for ``pleasantness'', ``attention'', ``sensitivity'', and ``aptitude''~. The continuous representations can be mapped to 24 distinct emotions. DepecheMood is a lexicon created through a data-driven method mining a news website annotated with its particular set of discrete emotions, namely, ``afraid'', ``amusemed'', ``anger'', ``annoyed'', ``don't care'', ``happy'', and ``inspired''~. DepecheMood is extended to DepecheMood++ by including Italian~.\\nThe more recent development in text-based affective analysis is models powered by deep learning. Leveraging large scale data, deep neural networks are able to learn representations that are relevant for affective analysis in language. Word embeddings are one of the most common representations used to represent language. Word embeddings, such as Word2Vec~ or GloVe~, learn language context of the word by learning a representation (a vector), that can capture semantic and syntactic similarities. More recently, representation learning models that can encode the whole sequence of terms (sentences, documents) showed impressive performance in different language understanding tasks, including sentiment and emotional analysis. \\nBidirectional Encoder Representations from Transformers (BERT)~ is a method for learning a language model that can be trained on large amount of data in an unsupervised manner. This pre-trained model is very effective in representing a sequence of terms as a fixed-length representation (vector). BERT architecture is a multi-layer bidirectional Transformer network that encodes the whole sequence at once. BERT representation achieves state-of-the-art results in multiple natural language understanding tasks. \\nThe audiovisual features that are used for multimodal understanding of affect are similar to the ones discussed in previous sections. The main technique between miltimodal models lies in methods for multimodal fusion. Multimodal methods involve extracting features from multiple modalities, \\\\textit{e.g.}, audiovisual, and training joint or separate machine learning models for fusion~. Multimodal fusion can be done in model-based and model-agnostic ways. The model-agnostic fusion methods do not rely on a specific classification or regression method and include feature-level, decision-level, or hybrid fusion techniques. Model-based methods address multimodal fusion in model construction. Examples of model-based fusion methods include Multiple Kernel Learning (MKL)~, graphical models, such as Conditional Random Fields~ and neural networks~.\\n\\\\citeauthor{Pang2015}~ used Deep Boltzmann Machine (DBM) to learn a joint representation across text, vision, and audio to recognize expected emotions from social media videos. Each modality is separately encoded with stacking multiple Restricted Boltzmann Machines (RBM) and pathways are merged to a joint representation layer. The model was evaluated for recognizing eight emotion categories for 1,101 videos from~. \\\\citeauthor{muszynski2019recognizing}~ studied perceived vs induced emotion in movies. To this end, they collected additional labels on a subset of LIRIS-ACCEDE dataset~. They found that perceived and induced emotions do not always agree. Using multimodal Deep Belief Networks (DBN), they could demonstrate that fusion of electrodermal responses with audiovisual content features improves the overall accuracy for emotion recognition~.\\nIn~, authors performed regression to estimate intended arousal and valence levels (as judged by experts in~). LSTM recurrent neural networks are used for unimodal regressions and fused via early and late fusion for audiovisual estimation with late fusion achieving the best results. \\\\citeauthor{Tarvainen2018}~ performed an in-depth analysis on how emotions are constructed in movies. They identified scene type as a major factor in emotions in movies. They then used content features to recognize emotions along three dimensions of hedonic tone (valence), energetic arousal (awake--tired) and tense arousal (tense--calm).\\nBilinear fusion is a method that is proposed to model inter- and intra- modality interaction among modalities by performing outer product between unimodal embeddings~. \\\\citeauthor{zadeh2017tensor}~ extended this to a Tensor Fusion Network to model intra-modality and inter-modality dynamics in multimodal sentiment analysis. The tensor fusion network includes modality embedding sub-networks, a tensor fusion layer modeling the unimodal, bimodal and trimodal interactions using a three-fold Cartesian product from modality embeddings along with a final sentiment inference sub-network conditioned on the tensor fusion layer. The main drawback of such methods is the increase in the dimensionality of the resulting multimodal representation.\",\n \"id\": \"20c98c17-ea71-4ba4-a9aa-815a42eced5b\",\n \"level\": \"section\",\n \"origin_cites_number\": 27,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Affective Computing of Multimodal Data\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Affective Computing of Multimodal Data\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 7022,\n 1033\n ],\n \"content\": \"\\\\label{sec:FutureDirections}\\nAlthough remarkable progress has been made on affective computing of multimedia (ACM) data, there are still several open issues and directions that can boost the performance of ACM.\\n\\\\textbf{Multimedia Content Understanding.} As emotions may be directly evoked by the multimedia content in viewers, accurately understanding what is contained in multimedia data can significantly improve the performance of ACM. Sometimes it is even necessary to analyze the subtle details. For example, we may feel ``amused'' on a video with a laughing baby; but if the laugh is from a negative character, it is more possible for us to feel ``angry''. In such cases, besides the common property, such as ``laugh'', we may need to further recognize the identity, such as ``a lovely baby'' and ``an evil antagonist''.\\n\\\\textbf{Multimedia Summarization.} Emotions can play a vital role in selection of multimedia for creation of summaries or highlights. This is an important application in entertainment and sports industries (\\\\textit{e.g.} movie trailers, sports highlights). There has been some recent work in this direction where affect information from audio visual cues has led to the successful creation of video summaries~. In particular, work reported in~ used audiovisual emotions in part to create an AI trailer for a $20^{th}$ Century Fox film in 2016. Similarly, AI Highlights described in~ hinged on audiovisual emotional cues and have successfully been employed to create the official highlights at Wimbledon and US Open since 2017. This is a very promising direction for affective multimedia computing which can have a direct impact on real world media applications.\\n\\\\textbf{Contextual Knowledge Modeling.} The contextual information of a viewer watching some multimedia is very important. Similar multimedia data under different contexts may evoke totally different emotions. For example, we may feel ``happy\\\" when listening a song about love in a wedding; but if the same song is played when two lovers are departing, it is more likely that we feel ``sad\\\". The prior knowledge of viewers or multimedia data may also influence the emotion perceptions. An optimistic viewer and a pessimistic viewer may have totally different emotions about the same multimedia data.\\n\\\\textbf{Group Emotion Clustering.} It is too generic to simply recognize the dominant emotion, while it is too specific to predict personalized emotion. It would make more sense to model emotions for groups or cliques of viewers with similar interests and backgrounds. Clustering different viewers into corresponding groups possibly based on the user profiles may provide a feasible solution to this problem.\\n\\\\textbf{New AC Setting Adaptation.} Because of the domain shift~, the deep learning models trained on one labeled source domain may not work well on the other unlabeled or sparsely labeled target domain, which results in the models' low transferability to new domains. Exploring domain adaptation techniques that fit well on the AC tasks is worth investigating. One possible solution is to translate the source data to an intermediate domain that are indistinguishable from the target data while preserving the source labels~ using Generative Adversarial Networks~. How do deal with some practical settings, such as multiple labeled source domains and emotion models' homogeneity, is more challenging.\\n\\\\textbf{Regions-of-Interest Selection.} The contributions of different regions of given multimedia may vary to the emotion recognition. For example, the regions that contain the most important semantic information in images are more discriminative than background; some video frames are of no use to emotion recognition. Detecting and selecting the regions-of-interest may significantly improve the recognition performance as well as the computation efficiency.\\n\\\\textbf{Viewer-Multimedia Interaction.}\\nInstead of direct analysis of the multimedia content or implicit consideration of viewers' physiological signals (such as facial expressions, Electroencephalogram signals, \\\\textit{etc}.), joint modeling of both multimedia content and viewers' responses may better bridge the affective gap and result in superior performances. We should also study how to deal with missing or corrupted data. For example, some physiological signals are unavailable during the data collection stage.\\n\\\\textbf{Affective Computing Applications.} Although AC is claimed to be important in real-world applications, few practical systems have been developed due to the relatively low performance. With the availability of larger datasets and improvements in self-supervised and semi-supervised learning, we foresee the deployment of ACM in real-world applications. For example, in media analytics, the content understanding methods will identify the emotional preferences of users and emotional nuances of social media content to better target advertising effort; in fashion recommendation, intelligent costumer service, such as customer-multimedia interaction, can provide better experience to customers; in advertisement, generating or curating multimedia that evokes strong emotions can attract more attention. We believe that an emotional artificial intelligence will become a significant component of mainstream multimedia applications.\\n\\\\textbf{Benchmark Dataset Construction.}\\nExisting studies on ACM mainly adopt small-scale datasets or construct relatively larger-scale ones using keyword searching strategy without annotation quality guaranteed. To advance the development of ACM, creating a large-scale and high-quality dataset is in urgent need. It has shown that there are three critical factors for dataset construction of ACM, \\\\textit{i.e.}, the context of viewer response, personal variation among viewers, and the effectiveness and efficiency of corpus creation~. In order to include a large number of samples, we may exploit online systems and crowdsourcing platforms to recruit large numbers of viewers with a representative spread of backgrounds to annotate multimedia and provide contextual information on their emotional responses. Since emotion is a subjective variable, personalized emotion annotation would make more sense, from which we can obtain the dominant emotion and emotion distribution. Further, accurate understanding of multimedia content can boost the affective computing performance. Inferring emotional labels from social media users' interaction with data, \\\\textit{e.g.}, likes, comments, in addition to their spontaneous responses, \\\\textit{e.g.}, facial expression, where possible, will provide new avenues for enriching affective datasets.\",\n \"id\": \"5da3fa17-09f5-4d0c-9b55-5610d97ddda5\",\n \"level\": \"section\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Future Directions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Conclusion}\\nIn this article, we have surveyed affective computing (AC) methods for heterogeneous multimedia data. For each multimedia type, \\\\textit{i.e.}, image, music, video, and multimodal data, we summarized and compared available datasets, handcrafted features, machine learning methods, deep learning models, and experimental results. We also briefly introduced the commonly employed emotion modelds and outlined potential research directions in this area. Although deep learning-based AC methods have achieved remarkable progress in recent years, an efficient and robust AC method that is able to obtain high accuracy under unconstrained conditions is yet to be achieved. With the advent of deep understanding of emotion evocation in brain science, accurate emotion measurement in psychology, and novel deep learning network architectures in machine learning, affective computing of multimedia data will remain an\\nactive research topic for a long time.\\n\\\\begin{acks}\\nThis work was supported by Berkeley DeepDrive, the National Natural Science Foundation of China (Nos. 61701273, 91748129), and the National Key R\\\\&D Program of China (Grant No. 2017YFC011300). The work of MS is supported in part by the U.S. Army. Any opinion, content or information presented does not necessarily reflect the position or the policy of the United States Government, and no official endorsement should be inferred.\\n\\\\end{acks}\\n\\\\bibliographystyle{ACM-Reference-Format}\\n\\\\bibliography{references}\\n\\\\end{document}\",\n \"id\": \"fc0b3d97-cde9-409b-91c4-fe25e8140878\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"36832e43-191f-4e88-aa11-6c7b91fe1bf8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Affective Computing for Large-Scale Heterogeneous Multimedia Data: A Survey\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":75,"string":"75"},"all_cites":{"kind":"list like","value":[1016,1015,1017,1018,1019,1020,1021,7322,1022,1023,1024,1025,1026,1028,1027,1029,7,1030,1032,1684,1031,7022,1033],"string":"[\n 1016,\n 1015,\n 1017,\n 1018,\n 1019,\n 1020,\n 1021,\n 7322,\n 1022,\n 1023,\n 1024,\n 1025,\n 1026,\n 1028,\n 1027,\n 1029,\n 7,\n 1030,\n 1032,\n 1684,\n 1031,\n 7022,\n 1033\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.336731744741586,"string":"1.336732"}}},{"rowIdx":41,"cells":{"authors":{"kind":"list like","value":["Dingwen Zhang","Junwei Han","Gong Cheng","Ming-Hsuan Yang"],"string":"[\n \"Dingwen Zhang\",\n \"Junwei Han\",\n \"Gong Cheng\",\n \"Ming-Hsuan Yang\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Weakly Supervised Object Localization and Detection: A Survey"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-04-16T06:44:50Z","string":"2021-04-16T06:44:50Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" As an emerging and challenging problem in the computer vision community, weakly supervised object localization and detection plays an important role for developing new generation computer vision systems and has received significant attention in the past decade. As methods have been proposed, a comprehensive survey of these topics is of great importance. In this work, we review (1) classic models, (2) approaches with feature representations from off-the-shelf deep networks, (3) approaches solely based on deep learning, and (4) publicly available datasets and standard evaluation metrics that are widely used in this field. We also discuss the key challenges in this field, development history of this field, advantages/disadvantages of the methods in each category, the relationships between methods in different categories, applications of the weakly supervised object localization and detection methods, and potential future directions to further promote the development of this research field. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"3a2d911b-3708-4f3a-880a-851136519e79","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"]],"subsections":["6019d392-fb6b-4290-8c58-a72268a80ff9","a3e94559-bb78-4175-a692-379986435e13","bba63d48-815b-4407-9de8-7df72823a004","ce302070-cf4f-4d53-a26c-8a51743b26f9","930ae35a-10a8-4a65-83e8-a12f89f6fe36","79d342a2-5acc-4290-aaff-2761ce99e494","5cef946e-16aa-4550-aeb5-c885a8308ec2","d8feaee7-5d65-4327-96a9-38b968bfe1c4","92209b0e-e020-4aa4-8af9-77a6bcf0b9af"],"title":"root"},{"cite_extract_rate":0.2,"cites":[2259,209,2260],"content":"\\label{sec:introduction}}\n\\IEEEPARstart{W}{eakly} supervised learning (WSL) has recently received much attention in computer vision community.\nA plethora of methods on this topic have been proposed in the past decade to address the challenging computer vision tasks including semantic segmentation~, object detection~, and 3D reconstruction~, to name a few.\nAs shown in Fig.~\\ref{illustration}, a WSL problem is defined as the learning process when some partial information regarding the task (e.g., class label or object location) on a small subset of the data points is at our disposal.\nCompared to the conventional learning framework, e.g., fully supervised learning approaches, the WSL framework needs to operate on the small amount of weakly-labelled training data to learn the target model, which alleviates a huge amount of human labor to annotate training samples.\nIt can also facilitate the learning process when the fine-grained annotation\nis extremely labor intensive and time consuming to even obtain the whole labeled data required by the fully-supervised approaches.\nWhile a plethora of WSL-based vision methods have been developed, this survey mainly focuses on the task of weakly supervised object localization and detection, which is shown as the red dot in Fig.~\\ref{illustration}.\nIt is well-known that object localization and detection is a fundamental research problem in computer vision.\nLearning object localization and detection models under weak supervision has attracted much attention in the past decades.\nWhile existing methods treat weakly supervised object localization (WSOL) and weakly supervised object detection (WSOD) as two different tasks\\footnote{ {The difference between WSOL and WSOD mainly lies in that WSOL mainly aims at localizing a single known (or unknown) object from each given image scene. The goal of WSOD is to instead detect every possible object instance from the given image scene. This makes WSOD a little more difficult than WSOL.}}, we consider these as a common task due to several reasons: 1) these tasks learn with the same image-level human annotation; 2) these two tasks need certain supervision as input and usually aim to localize objects on the bounding-box level as output; 3) WSOD task can be accomplished by directly training off-the-shelf fully supervised object detectors on the object locations obtained from WSOL.\nDuring the last decade, considerable efforts have been made to develop various approaches for learning object detectors with weak supervision.\nSome of the existing algorithms only learn weakly supervised object detectors for one or several certain object categories, such as vehicles~, traffic signs~, pedestrians~, faces~, tuberculosis bacilli~, aircrafts~, and human actions~.\nWhile other approaches, e.g., , focus on developing weakly supervised learning frameworks for unconstrained object categories, i.e., learning frameworks can be extended to learn object detector for the given category-specific weakly-labelled training images.\nAs enormous methods have been developed for these important tasks, a comprehensive review of the literature concerning weakly supervised object localization and detection is of great importance.\nAs weakly supervised object localization and detection methods mainly exploit the image-level manual annotation, the learning frameworks not only need to address the typical issues, such as the intra-class variations in appearance, transformation, scale and aspect ratio, encountered in conventional fully supervised object localization and detection task, but also the \\textbf{learning under uncertainty} challenges caused by the inconsistency between human annotations and real supervisory signals.\nIn weakly supervised object localization and detection, the accuracy of object locations and learning processes are closely related.\nThe key is to propagate the image-level supervisory signals to the instance-level (bounding-box-level) training data for the learning processes.\nAs each training image can be labeled by numerous bounding boxes of different accuracy, propagating such weak supervision inevitably involves a large amount of ambiguous and noisy information as each training instance.\nMore specifically, the \\textbf{learning under uncertainty} issue would cause the following challenges that make the weakly supervised learning process challenging:\n\\begin{itemize}\n \\item \\textbf{Learning with inaccurate instance locations:} This issue is mainly caused by the definition ambiguity in object parts and context.\n Without precise annotation or definition, it is difficult for a learner to decide whether an object category label associates with a discriminative object part, the whole object region, or the object with a certain context region.\n As a result, the bounding-box instance locations inferred by the learner may contain many inaccurate samples including the ones with local object parts or undesired contextual regions.\n These samples would negatively affect the performance of WSL-based detectors.\n \\item \\textbf{Learning with noisy samples:} Even when the bounding-box locations can be precisely labeled, the training examples enclosed by bounding-boxes may still be noisy as background pixels are usually included.\n As there is no additional information to separate foreground objects from the background, the learner may tag a ``background'' label to an object region when it fails to recognize the object category.\n In addition, the learner may mistakenly label a bounding-box that contains a bicycle as a motorcycle, as these two object categories share many similar features.\n \\item \\textbf{Learning with domain shifts:} For a certain object category, the image regions localized during the learning process may only contain samples with limited diversity in object shape, appearance, scale, and view angle.\n This makes the subsequent learning process biased to limited knowledge of the object category and does not generalize well for test samples.\n For instance, a learner can hardly localize or detect a flying swan when all the training samples contain the swimming ones on lakes.\n This issue happens frequently among the weakly supervised learning process when there is a large gap between the training and testing domains.\n \\item \\textbf{Learning with insufficient instance samples:}\n Similar to the issues in conventional learning methods, it is difficult to train effective object detectors under the weakly-supervised setting when the amount of training samples is limited.\n In addition, the number of positive samples is usually much smaller than that of negative samples for binary classes.\n Furthermore, the data distributions for a large number of categories is usually long-tailed.\n This issues are significantly exaggerated for the WSL-based methods using deep learning.\n\\end{itemize}\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=1\\linewidth]{illustration.jpg}\n \\caption{Illustration of the weakly supervised learning tasks in computer vision community. The blue area in the top block indicates the conventional fully-supervised learning tasks, while the red area in the top block indicates the weakly supervised learning tasks. The coordinate axes shows different levels of human annotation or supervision requirement, from low cost to high cost. Notice that the high cost annotation can be transformed to low cost annotation easily, e.g., from bounding-box level to image level, whereas the low cost annotation is hard to be transformed to high cost annotation. In the bottom block, we also show the label cost, in terms of annotation time, and the examples of different type of annotations. In this survey, we mainly focus on reviewing the research progress in weakly supervised object localization and detection, i.e., the red dot in the top block.}\n \\label{illustration}\n\\end{figure}\nTo address the above-mentioned issues in learning weakly supervised object detectors, existing methods are usually constructed based on two steps: initialization and refinement.\nThe initialization stage is used to leverage certain prior knowledge to propagate image-level annotation into instance level, and thus can generate instance-level annotation (but with label noise, sample bias, and limited quality in location accuracy) for the learning process.\nThe refinement stage is used to leverage new instance samples obtained from the first stage to mine truthful knowledge about the objects of interest gradually and finally obtain the desired object models for localization and detection.\nThese two learning stages need to collaborate to address the aforementioned five-fold challenges.\nIn initialization stage, efforts should be made to improve the annotation quality as much as possible to generate training instances with proper locations, accurate labels, high diversity, and high recall rate.\nAs the annotation quality obtained in the learning stage cannot be perfect, in the refinement stage, further efforts should be made to improve the learner's robustness to cope with the inaccurate instance location, noisy examples, biased instance sample, insufficient instance sample issues as well as the capacity to take advantage of the unlabelled instance samples.\nWhen properly addressing the problems in each learning stage,\ngood weakly supervised object detectors can be learned.\nIn this work, we review the existing weakly supervised object localization and detection approaches\\footnote{Some early methods, such as~, learn to localize category-wise key points under the weak supervision, while this survey mainly focuses on the methods for localizing instances with bounding-boxes.}, which are divided into three main categories and eight subcategories.\nThese three main categories are based on classic approaches, feature representations from off-the-shelf deep models, and deep learning frameworks.\nThe eight subcategories include approaches for initialization, refinement, initialization and refinement, pre-trained deep features, inherent cues in deep models, fine-tuned deep models, single-network training, and multi-network training.\nWe further discuss the relationship between the approaches in different categories.\nIn addition, we also discuss open problems and challenges of current studies and propose several promising research directions in the future for constructing more effective weakly supervised object localization and detection frameworks.\n\\begin{figure*}[t]\n \\centering\n \\includegraphics[width=1\\linewidth]{category2.jpg}\n \\caption{In the left block, taxonomy of the existing approaches for weakly supervised object localization and detection, which includes three main\n categories and eight subcategories.\n In the right block, the relationships between the approaches in different categories are shown.\n}\n \\label{taxonomy}\n\\end{figure*}\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=1\\linewidth]{history.JPG}\n \\caption{Developments of weakly supervised localization and detection methods. The yellow histogram shows the number of publications in this research field in each year, and the curves show the number of proposed methods each year for a particular category of approach.}\n \\label{history}\n\\end{figure}","id":"6019d392-fb6b-4290-8c58-a72268a80ff9","level":"section","origin_cites_number":15,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0.1,"cites":[2261,2262,2263],"content":"\\label{Taxonomy}\nIn the last decade, a plethora of methods have been developed for weakly supervised object localization and detection.\nWe can generally categorize existing methods based on classic formulations, feature representations from off-the-shelf deep models, and deep weakly supervised learning algorithms.\nWhile inside each main category, we further divide the approaches into two or three subcategories.\nFig.~\\ref{taxonomy} shows our taxonomy of the studies in the research field of weakly supervised object localization and detection.\nIn addition, Fig. \\ref{history} reviews the development history of each of the main category as well as that of the whole research field.\nA few approaches based on classic formulations appeared around 2002.\nFrom 2002 to 2009, the research in this field went through a very slow pace.\nSince 2014, numerous approaches based on both classic formulations and learned feature representations from deep models have been developed and received much attention.\nWhile in the last few years, more approaches solely based on deep learning have become the main stream to address the problems of weakly supervised object localization and detection.\nWhile a plethora of methods have been developed to address different aspects of these problems in the past decades, this field is gaining increasing attention.\n\\renewcommand{\\multirowsetup}{\\centering}\n \\begin{table*} [t]\n \\tiny\n \\caption{Summary of the approaches for initialization, which is a subcategory in the weakly supervised object localization and detection approaches that learn by classic models. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\n \\label{table initialization}\n \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{2.5cm}<{\\centering}m{1.2cm}<{\\centering}m{1.5cm}<{\\centering}m{3.1cm}<{\\centering}m{1.8cm}<{\\centering}m{1.5cm}<{\\centering}m{3.5cm}<{\\centering}m{1.8cm}<{\\centering}}\n \\hline\n \\multirow{2}{*}{\\textbf{Methods}} & \\multirow{2}{*}{\\textbf{Detector}} & \\multirow{2}{*}{\\textbf{Descriptor}} & \\multirow{2}{*}{\\textbf{Prior knowledge}} & \\multirow{2}{*}{\\textbf{Extra training data}} & \\multirow{2}{*}{\\textbf{Learning model}} & \\multirow{2}{*}{\\textbf{Learning strategy}} & \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline \\hline\nCao-PR-2017 & SVM & HOG+PCA & Road map prior + density prior & None & SVM & MIL with density estimation & Vehicle in satellite imagery \\\\\nShi-TPAMI-2015 & DPM & SIFT+Lab+LBP, BOW & Appearance prior + geometry prior & None & Topic model & Bayesian inference & General objects \\\\\nTang-ICIP-2014 & DPM & HOG & Saliency (objectness score) & None & DPM & Select initial boxes + DPM training & General objects \\\\\n Xie-VCIP-2013 & None & SIFT & Intra-class consistency & None & None & Low-rank and sparse coding & General objects \\\\\nSiva-CVPR-2013 & DPM & Lab+SIFT & Saliency & Labelme + PASCAL07, 12 (unlabelled) &None & None & General objects \\\\\n Shi-ICCV-2013 & DPM & SIFT & Appearance prior (spatial distribution, size, saliency) & None & Topic model & Bayesian inference & General objects \\\\\n Sikka-FG-2013 & None & SIFT, LLC, BOW & None & None & MilBoost & Generating multiple segments for initialization and use Milboost for learning & Pain (on face) \\\\\n Siva-ECCV-2012& None & SIFT+BOW & Iter-class variance + saliency & None & None & Negative mining & General objects \\\\\n Shi-BMVC-2012 & None & BOW & Mapping relationship between the box overlap and appearance similarity & Part of PASCAL 07 (box annotation) & RankSVM & Transfer learning by ranking & General objects \\\\\n Khan-AAPRW-2011 & DPM & Phog/phow & None & Internet image (weakly annotated) & MIL & Learning from internet image & Pascal@8 \\\\\nZhang-BMVC-2010 & SVM & IHOF & Co-occurrence & None & SVM & High order feature learning by exploring co-occurrence & General objects \\\\\n\\hline\n \\end{tabular}}\n\\end{table*}\n\\renewcommand{\\multirowsetup}{\\centering}\n\\begin{table*} [t]\n \\tiny\n\t\\caption{Summary of the approaches for refinement, which is a subcategory in the weakly supervised object localization and detection approaches that learn by classic models.\n\tHere, * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\n\t \\label{table refinement} \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{2.5cm}<{\\centering}m{1.5cm}<{\\centering}m{3cm}<{\\centering}m{2cm}<{\\centering}m{2cm}<{\\centering}m{2cm}<{\\centering}m{3.1cm}<{\\centering}m{2.5cm}<{\\centering}}\n\\hline\n \\multirow{2}{*}{\\textbf{Methods}} &\n \\multirow{2}{*}{\\textbf{Detector}} &\n \\multirow{2}{*}{\\textbf{Descriptor}} &\n \\multirow{2}{*}{\\textbf{Prior knowledge} } &\n \\multirow{2}{*}{\\textbf{Extra training data}} &\n \\multirow{2}{*}{\\textbf{Learning model}} &\n \\multirow{2}{*}{\\textbf{Learning strategy}} &\n \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline\n\\hline\n Wang-TMI-2018 & None & Color & None & None & Low-rank model & Low-rank Factorization & General lesion \\\\\n Wang-TC-2017 & None & SIFT+LAB & None & None & Probability model & BOW learning+instance labeling & General objects \\\\\n Cholakkal-CVPR-2016 & None & SIFT & Saliency & None & SVM* & ScSPM-based top down saliency & Salient objects \\\\\n Zadrija-GCPR-2015 & None & SIFT+FV & None & None & GMM + linear classifier & Patch-level spatial layout learning & Traffic sign \\\\\n Krapac-ICCVTA-2015 & None & SIFT+FV & None & None & Sparse logistic regression & Sparse classification & Traffic sign \\\\\n Cinbis-CVPR-2014 & SVM & FV & Center prior & None & SVM & Multi-fold MIL & General objects \\\\\n Wang-ICIP-2014 & SVM + graph model & SIFT & None & None & SVM + graph model & Maximal entropy random walk & Car, dog \\\\\n Wang-ICIP-2014 & SVM & SIFT & None & None & SVM & Clustering for window mining & General objects \\\\\n Tang-CVPR-2014 & None & SIFT & Saliency & None & Boolean constrained quadratic program & Mine similarity and discriminativeness both for image and box & General objects \\\\\n Hoai-PR-2014 & SVM & SIFT,BOW & None & None & SVM* & Localization-classification SVM & Face,car, human motion \\\\\n Wang-WACV-2013 & Task-specific detectors & HOG/SC & Background\\ saliency & None & MIL+AdaBoost* & Soft-label Boosting after MIL & Vehicle, pedestrain\\\\\n Kanezaki-MM-2013 & Linear classifiers & 3D voxel feature (color+C3HLAC +Intensity, texture, GRSD) & None & None & Linear classifiers & Multi-class MIL & Balls, tools \\\\\n Pandey-ICCV-2011 & DPM* & HOG & None & None & DPM* & Learning DPM with fully latent variable & General objects \\\\\nBlaschko-NIPS-2010 & None & BOW/HOG & None & None & SVM* & Learning SVM with structured output ranking objective & Cat, pedestrian \\\\\nHoai-ICCV-2009 & SVM & SIFT,BOW & None & None & SVM* & Localization-classification SVM & Face,car, human motion \\\\\n Galleguillos-ECCV-2008 & None & SIFT+BOW & None & None & MilBoost & Train MIL classifier for localization & Landmarks, faces, airplanes, leopard, motorbike, car \\\\\n Rosenberg-BMVC-2002 & GMM & Orientation deriviate filters & Exampler prior & Training exampler (box annotation) & GMM & Learning from exampler training data to weakly labelled training data & Telephone \\\\\n\\hline\n \\end{tabular}}\n\\end{table*}\n\\renewcommand{\\multirowsetup}{\\centering}\n\\begin{table*} [t]\n \\tiny\n\t\\caption{Approaches for both initialization and refinement, which is a subcategory in the weakly supervised object localization and detection approaches that learn by classic models. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\n\t \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{1.8cm}<{\\centering}m{1.2cm}<{\\centering}m{1.5cm}<{\\centering}m{2cm}<{\\centering}m{1.8cm}<{\\centering}m{1.5cm}<{\\centering}m{3cm}<{\\centering}m{1.5cm}<{\\centering}}\n\\hline\n \\multirow{2}{*}{\\textbf{Methods}} &\n \\multirow{2}{*}{\\textbf{Detector}} &\n \\multirow{2}{*}{\\textbf{Descriptor}} &\n \\multirow{2}{*}{\\textbf{Prior knowledge} } &\n \\multirow{2}{*}{\\textbf{Extra training data}} &\n \\multirow{2}{*}{\\textbf{Learning model}} &\n \\multirow{2}{*}{\\textbf{Learning strategy}} &\n \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline\n\\hline\n Wang-cvpr-2013 & None & Color+SIFT & None & None & HST+SVM & Joint parsing and attribute localization & Scene attributes \\\\\n Deselaers-IJCV-2012 & DPM & GIST+CH+BOW+HOG & Generic knowledge & Meta-training data with box annotation & CRF+DPM & Learning appearance model by trainsferring genearic knowledge & General objects \\\\\n Siva-ICCV-2011 & DPM & SIFT+BOW+HOG & Inter-class prior + intra-class prior & None & DPM & Model drift learning & General objects \\\\\nDeselaers-ECCV-2010 & DPM & GIST+CH+BOW+HOG & Generic knowledge & Meta-training data with box annotation & CRF+DPM & Learning appearance model by transferring generic knowledge & General objects \\\\ \\hline\n \\end{tabular}}\n\\end{table*}\nAs shown in the right block of Fig.~\\ref{taxonomy}, methods in\nmain categories are related in several aspects.\nNumerous methods are developed based on the classic formulations with the advances of feature representations from deep models.\nSimilarly, a number of methods based solely on deep models are end-to-end trainable by considering classic formulations and feature extraction schemes.\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=7 cm]{subcategory1.jpg}\n \\caption{Flowchart of the weakly supervised object localization and detection approaches using classic models.}\n \\label{nondeep flowchart}\n\\end{figure}","id":"a3e94559-bb78-4175-a692-379986435e13","level":"section","origin_cites_number":30,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Taxonomy"]],"subsections":[],"title":"Taxonomy"},{"cite_extract_rate":0.16666666666666602,"cites":[2261,2262,2263],"content":"\\label{Classic Models}\nIn this section, we review the classic approaches that learn weakly supervised object localizer or detector without using deep features.\nThese methods typically consist of one initialization module followed by one refinement process as shown in\nFig.~\\ref{nondeep flowchart}.\nIn , the detector is based on the deformable part model (DPM)~.\nIn other approaches , the detector is based on the support vector machine (SVM) classifier.\nThe features used by these approaches are hand-crafted feature descriptors, such as HOG in , SIFT in , and Lab color in , which are sometimes used to build higher level representation such as bag-of-words (BOW) in , ,,,, Fisher vector representation in , and subspace-based representation in .\nIn the following, we divide these approaches for initialization and refinement process.","id":"bba63d48-815b-4407-9de8-7df72823a004","level":"section","origin_cites_number":18,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Classic Models"]],"subsections":["bfae7607-5833-4424-b0e8-9e0a00e92605","8f1d3bcb-1c87-4161-b6dc-8cbe17a169ff","31104e17-5d59-42f1-88ae-903542c63cd2","19588196-93c9-4245-8a54-e9fb29d5fd1e"],"title":"Classic Models"},{"cite_extract_rate":0.40740740740740705,"cites":[2261,2262,2267,2263,2269,2270,2266,2264,2265,8537,2268],"content":"Numerous methods have been developed to mine reliable instance samples, using prior knowledge, as weak supervision for the following processes.\nA brief summary of these approaches are shown in Table \\ref{table initialization}.\nZhang et al.~ leverage the prior-knowledge of object co-occurrence to identify translation and scale invariant high order features for weakly supervised object localization.\nIn~ Shi et al. propose a transfer learning paradigm to first use a RankSVM to learn the mapping relationship between the box overlap and appearance similarity from an auxiliary training data (with bounding-box level annotation) and then transfer the learned prior-knowledge for localizing objects of interest in the given weakly labeled images.\nA simple yet effective approach, named as negative mining, is developed by Siva et al. to explore the inter-class variance among the object regions in weakly labeled training images.\nThe final object locations are obtained by\nusing a linear combination of the inter-class variance and saliency prior.\nSimilarly, Tang et al.~ and Xie et al. use the saliency prior and intra-class consistency to mine the initial object locations, respectively.\nIn , Shi et al. explore the appearance prior and geometry prior in their topic model to build a Bayesian joint modeling framework for weakly supervised object localization.\nOn the other hand, Cao et al.~ exploit the road map prior and density prior to mine the initial vehicle locations from the weakly labeled satellite images and then trained the vehicle detector under a modified multiple-instance learning (MIL) model.\n\\renewcommand{\\multirowsetup}{\\centering}\n\\begin{table*} [t]\n \\tiny\n\t\\caption{Summary of the approaches using pre-trained feature representations, which is a subcategory in the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\n \\label{table off-the-shelf feature}\n \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{2cm}<{\\centering}m{1cm}<{\\centering}m{1.6cm}<{\\centering}m{2.8cm}<{\\centering}m{3.8cm}<{\\centering}m{1.2cm}<{\\centering}m{3.5cm}<{\\centering}m{2.2cm}<{\\centering}}\n\\hline\n \\multirow{2}{*}{\\textbf{Methods}} &\n \\multirow{2}{*}{\\textbf{Detector}} &\n \\multirow{2}{*}{\\textbf{Descriptor}} &\n \\multirow{2}{*}{\\textbf{Prior knowledge} } &\n \\multirow{2}{*}{\\textbf{Extra training data}} &\n \\multirow{2}{*}{\\textbf{Learning model}} &\n \\multirow{2}{*}{\\textbf{Learning strategy}} &\n \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline\n\\hline\n Gonthier-arxiv-2018 & None & CNN & Supervised objectness score (Fast R-CNN(Resnet)) & ImageNet(tag label) & SVM & MIL & Objects in art (watercolor2K, people-art) \\\\\nZadrija-CVIU-2018 & None & VGG19 Conv 5\\_4. SIFT, Fisher vector & None & ImageNet(tag label) & Sparse model & None & Zebra crossings, traffic signs \\\\\n Cinbis-TPAMI-2017 & SVM* & FV+CNN & Center prior & ImageNet(tag label) & SVM* & Multi-fold MIL & General objects \\\\\nWei-IJCAI-2017 & None & CNN & None & ImageNet(tag label) & None & Deep descriptor transforming & General objects \\\\\n Zhang-IJCAI-2016 & SVM & CNN & Saliency prior & ImageNet(tag label) & SVM & Easy-to-hard(SPL+CL) & General objects \\\\\n Li-ECCV-2016 & None & FC6 & Strong detector prior (sparsity) & ImageNet(tag label) & SVM & Regularizing score distribution & General objects \\\\\n Ren-TPAMI-2016 & SVM* & FC6 & None & ImageNet(tag label) & SVM* & MIL+bag splitting & General objects \\\\\n Wan-ICIP-2016 & SVM* & FC7 & None & ImageNet(tag label) & SVM* & Correlation suppression+part suppression & General objects \\\\\n Rochan-IVC-2016 & None & Color histogram +CNN & Saliency & PASCAL (edge box), ImageNet & SVM & None & General objects \\\\\n Shi-ECCV-2016 & SVM & FC7(Alexnet) & Size prior & ImageNet(tag label), PASCAL2012 (object size) & SVM & Easy-to-hard(curriculum) & General objects \\\\\n Wang-TIP-2015 & SVM & FC6 & None & ImageNet(tag label) & pLSA, SVM & Online latent category learning & General objects \\\\\nRochan-CVPR-2015 & SVM & CNN & Objectness score, word embedding prior & YouTube-Objects (for parameter validation), Familiar object categories(detector) & SVM, Sparse reconstruction & Appearance transfer from text representation & General objects \\\\\nBilen-CVPR-2015 & LSVM & FC7+spatial features & None & ImageNet(tag label) & LSVM & Convex clustering & General objects \\\\\nWang-ICCV-2015 & None & FC6 & None & PASCAL(edge box), ImageNet & SVM* & Relaxed multiple-Instance SVM & General objects \\\\\nZhou-ICMBD-2015 & SVM & FC7 & Saliency prior & ImageNet(tag label) & SVM & Negative Bootstrapping & Airplanes in remote sensing \\\\\nHan-TGRS-2015 & SVM & DBM & Saliency, intra-class compactness, inter-class separability & None & DBM + SVM & Bayesian framework for initialization + refinement detector training & Objects in remote sensing \\\\\nMathe-Arxiv-2014 & Sequential detector & FC6 & Human fixation & ImageNet(tag label) & MIL*+RL & Constrained multiple instance SVM learning + reinforcement learning of detector & Human actions \\\\\n Wang-ECCV-2014 & SVM & FC6 & None & ImageNet(tag label) & pLSA, SVM & Online latent category learning & General objects \\\\\n Bilen-BMVC-2014 & LSVM & DeCAF & None & ImageNet(tag label) & LSVM* & LSVM with posterior regularization on symmetry and mutual exclusion & General objects \\\\\nSong-NIPS-2014 & DPM & FC7 & Objectness score & ImageNet(tag label) & LSVM & Frequent configuration mining+detector training & General objects \\\\\n Song-ICML-2014 & LSVM & DeCAF & None & ImageNet(tag label) & Graph model+LSVM* & Initialization via discriminative submodular cover+smoothed LSVM learning & General objects \\\\\n\\hline\n \\end{tabular}}\n\\end{table*}","id":"bfae7607-5833-4424-b0e8-9e0a00e92605","level":"subsection","origin_cites_number":27,"parent_id":"bba63d48-815b-4407-9de8-7df72823a004","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Classic Models"],["subsection","Initialization"]],"subsections":[],"title":"Initialization"},{"cite_extract_rate":0.28571428571428503,"cites":[2272,2271,2274,2273],"content":"After potential object instances are obtained, these hypotheses are verified in the following refinement processes.\nThe goals of these approaches are to design learning objective functions, optimization strategies, or learning mechanisms to gradually determine objects of interest from the extracted initial instance training samples.\nA brief summary of these approaches is shown in Table \\ref{table refinement}.\n Hoai et al. propose an approach which localizes the instances of the positive class and learns a sub-window classifier to recognize the corresponding object class.\nBlaschko et al. use a structured output SVM to learn a regressor from the weakly labeled training images to object locations that are parameterized by the coordinates of the bounding boxes.\nThe object locations were treated as latent variables, while the image-level annotation was used to constrain the set of values the latent variable can take.\nSimilarly, Pandey et al. learn weakly supervised object detectors by using DPMs with latent SVM training.\nIn , a soft-label boosting approach is developed to exploit the soft labels that are estimated during the MIL process to train object detectors based on Boosting algorithm.\nIn , Tang et al. treat the weakly supervised object localization problem as an object co-localization task, and present a joint image-box formulation to mine reliable object locations via a Boolean constrained quadratic program.\nThis approach can handle noisy labels in the image-level annotations.\nTo address the property that the MIL process may converge to poor local optima after the initialization, Cinbis et al. design a multi-fold MIL training paradigm.\nThis method divides the whole weakly labelled training images into multiple folds and implements the detector training process and object re-localization process in different folds, thereby alleviating the issue with convergence of poor local optima.\n\\renewcommand{\\multirowsetup}{\\centering}\n\\begin{table*} [t]\n \\tiny\n\t\\caption{Summary of the approaches using visual cues, which is a subcategory in the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\n\t \\label{table pre-learning} \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{2.5cm}<{\\centering}m{1.5cm}<{\\centering}m{1.5cm}<{\\centering}m{2cm}<{\\centering}m{3cm}<{\\centering}m{1.5cm}<{\\centering}m{3.1cm}<{\\centering}m{1.8cm}<{\\centering}}\n\\hline\n \\multirow{2}{*}{\\textbf{Methods}} &\n \\multirow{2}{*}{\\textbf{Detector}} &\n \\multirow{2}{*}{\\textbf{Descriptor}} &\n \\multirow{2}{*}{\\textbf{Prior knowledge} } &\n \\multirow{2}{*}{\\textbf{Extra training data}} &\n \\multirow{2}{*}{\\textbf{Learning model}} &\n \\multirow{2}{*}{\\textbf{Learning strategy}} &\n \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline\n\\hline\n Li-ISPRS-2018 & CAM* (VGG-F) & CNN & None & ImageNet(tag label) & VGG-F* & Learning learning + CAM learning (patch level) & Remote sensing objects \\\\\n Wilhelem-DICTA-2017 & None & CNN & None & ImageNEt(tage label) & CAM & CAM+KDE refine & General objects \\\\\n Tang-TMM-2017 & DPM & CNN & Saliency + objectness score & ImageNet(tag label) & DPM+ CNN & Region initialization+DPM and feature learning+bounding box modification & General objects \\\\\n Kolesnikov -BMVC-2016 & None & CNN & Human feedback annotation & ImageNet(tag label) & CAM & Active learning for identifying object cluster & General objects \\\\\n Bency-ECCV-2016 & None & CNN & None & ImageNet(tag label) & VGG16 & Beam-search based on CNN classifier & General objects \\\\\nZhou-MSSP-2016 & SVM & FC7 & Saliency prior & ImageNet(tag label), remote sensing data(unlabelled) & CNN (AlexNet), SVM & Deep feature transfer +MIL & Remote sensing objects (airplane, car, airport) \\\\\nBergamo-WACV-2016 & SVM & CNN & None & ImageNEt(tage label) & CNN, SVM & Mask out initialization + SVM detector training & General objects \\\\\nHoffman-CVPR-2015 & SVM & FC7 & Detector prior+ representation prior & ImageNet(tag label), ILSVRC13 validation subset(box annotation) & CNN, Latent SVM & Transferring detectors and representation from auxiliary data & General objects \\\\\n\\hline\n \\end{tabular}}\n\\end{table*}","id":"8f1d3bcb-1c87-4161-b6dc-8cbe17a169ff","level":"subsection","origin_cites_number":14,"parent_id":"bba63d48-815b-4407-9de8-7df72823a004","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Classic Models"],["subsection","Refinement"]],"subsections":[],"title":"Refinement"},{"cite_extract_rate":0,"cites":[],"content":"A number of iterative approaches have been developed that take both initialization and refinement into account.\nIn , Siva et al. propose an intra-class metric and an inter-class metric to initialize the potential object locations.\nAfter obtaining the initial object locations, this method iteratively trains a DPM object detector and uses a model drift detection approach to identify the termination refinement dynamically.\nDeselaers et al. present a conditional random field (CRF) model, which is used to learn generic prior knowledge of the objects from meta-training data firstly to localize the potential objects of interest in the weakly labelled training images.\nThis algorithm updates the CRF model to learn the appearance and shape models for the target object category and localizes the objects of interest in the refinement stage. \nThe alternation of localization and learning processes progressively transforms the CRF model from class-generic prior knowledge into the specialized knowledge for a certain target class.\nFor learning weakly supervised attribute localizer, Wang et al. initialize the learning process by building a Hierarchical Space Tiling (HST) scene configuration model and the corresponding appearance models are trained based on HST.\nA joint inference and learning process is designed to update the scene attributes and the correlations between the scene parts and attributes gradually.","id":"31104e17-5d59-42f1-88ae-903542c63cd2","level":"subsection","origin_cites_number":4,"parent_id":"bba63d48-815b-4407-9de8-7df72823a004","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Classic Models"],["subsection","Initialization and Refinement"]],"subsections":[],"title":"Initialization and Refinement"},{"cite_extract_rate":0,"cites":[],"content":"Although the classic weakly supervised learning models are studied in early age, the two-stage learning frameworks, i.e., the learning initialization stage and refinement stage, built by these methods have been widely applied in future works. In the learning initialization stage, these methods provide two kinds of information cues to infer the candidate object regions. The first one is the bottom-up cues, including the region saliency, objectness, intra-class consistency, inter-class discriminability, et al. The other one is the top-down cues, which usually provide the appearance prior for the learning process. Notice that as such top-down cues are hard to obtain from the weakly labeled data, auxiliary training data (with instance-level manual annotation) are usually leveraged to explore the top-down cues which are then transferred to the weakly labeled target data. In the refinement stage, classic machine learning models, such as SVM and CRF, are adopted to gradually refine both the appearance model and locations of the objects of interest.\nThe advantage of the classic weakly supervised learning methods is that the learning processes can be implemented on small-scale training data and the whole frameworks are quick both in the training phase and the testing phase. The disadvantage is that their performance is not satisfactory, which is due to the limitation in feature representation and model complexity.","id":"19588196-93c9-4245-8a54-e9fb29d5fd1e","level":"subsection","origin_cites_number":0,"parent_id":"bba63d48-815b-4407-9de8-7df72823a004","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Classic Models"],["subsection","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":0.5714285714285711,"cites":[1223,514,2272,8429,2275,895,2278,2264,2273,2276,2279,2277],"content":"In this section, we review the approaches that learn weakly supervised object localizer or detector based on classic formulations and feature representations based on the deep neural networks, either pre-trained from the ImageNet dataset (with image tag annotation) or further fine-tuned on the weakly supervised training images in the target domain.\nThe feature representations are based on the widely used deep models for image classification, such as AlexNet and VGG .\nThe detectors are constructed based on classic formulations such as DPM and SVM~, or recent models such as RCNN and fast RCNN .\nWe further divide these approaches into three subcategories using pre-trained deep features, inherent cues in deep models, and fine-tuned deep models as shown in Fig. \\ref{off-the-shelf-flowchart}.\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=8.5 cm]{subcategory.jpg}\n \\caption{Illustration of different usages of the off-the-shelf deep neural networks by the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. }\n \\label{off-the-shelf-flowchart}\n\\end{figure}\n\\renewcommand{\\multirowsetup}{\\centering}\n\\begin{table*} [t]\n \\tiny\n\t\\caption{Summary of the approaches with fine-tuned deep models, which is a subcategory in the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. }\n\\label{table prepost-learning}.\n \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{2cm}<{\\centering}m{2cm}<{\\centering}m{1.6cm}<{\\centering}m{2.8cm}<{\\centering}m{2.5cm}<{\\centering}m{1.5cm}<{\\centering}m{3.7cm}<{\\centering}m{1.5cm}<{\\centering}}\n\\hline\n \\multirow{2}{*}{\\textbf{Methods}} &\n \\multirow{2}{*}{\\textbf{Detector}} &\n \\multirow{2}{*}{\\textbf{Descriptor}} &\n \\multirow{2}{*}{\\textbf{Prior knowledge} } &\n \\multirow{2}{*}{\\textbf{Extra training data}} &\n \\multirow{2}{*}{\\textbf{Learning model}} &\n \\multirow{2}{*}{\\textbf{Learning strategy}} &\n \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline\n\\hline\n Zhang-IJCV-2019 & Fast RCNN (VGG16) & Pre-trained FC7 & Tag number + mask out prior(AlexNet) & ImageNet(tage label) & SVM & Easy-to-hard & General objects \\\\\n Uijlings-CVPR-2018 & None & CNN & Semantic objectness(SSD*) & ImageNet(tage label), ILSVRC(full annotation) & SSD*+SVM+ Fast RCNN & MIL+knowledge transfer & General objects \\\\\n Jie-CVPR-2017 & Fast RCNN (VGG16) & CNN & Image-to-object transfer prior & ImageNet(tage label) & Fast RCNN (VGG16) & Initialization based on classification network and subgraph discovery + iterative Fast RCNN learning & General objects \\\\\nShi-ICCV-2017 & Fast RCNN & CNN & Things and stuff prior & ImageNet(tage label), PASCAL Context (full annotation) & FCN,Fast RCNN & Localizing objects based on things and stuff prior and training Fast TCNN iteratively & General objects \\\\\nSingh-CVPR-2016 & Fast RCNN & CNN & Tracking prior & ImageNEt(tage label), Youtube-objects (unlabelled) & Fast RCNN & Discriminative region minning+transferring tracking object pattern + learn object detector & General objects \\\\\nLi-CVPR-2016 & VGG* & CNN & Mask out prior(AlexNet) & ImageNEt(tage label) & VGG*, SVM & Progressive Domain Adaptation & General objects \\\\\nLiang-ICCV-2015 & CNN & CNN & Instance example, motion prior & ImageNEt(tage label) & CNN,R-CNN & Seed selection based on instance example and instance tracking & General objects \\\\\nChen-ICCV-2015 & RCNN & CNN & Online data type & Web data (weak label) & BLVC net +E-LDA + RCNN & Simple image initialization + graph-based representation adaptation on hard image & General objects \\\\\nZhou-CVPR-2015 & RCNN & FC7 & None & ImageNet(tag label) & SVM, R-CNN & .Max-margin visual concept discovery + Domain-specific detector selection & General objects\\\\\n\\hline\n \\end{tabular}}\n\\end{table*}","id":"ce302070-cf4f-4d53-a26c-8a51743b26f9","level":"section","origin_cites_number":21,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Off-the-shelf Deep Models"]],"subsections":["72d7b986-53df-40a6-a211-40fcbb5e65ab","a62482ff-4ed2-4ba2-b72d-45432f53b8f2","eab65442-451f-481c-9954-9b19dd3d9a08","0873c3d1-8648-437a-a297-c566cfb8420e"],"title":"Off-the-shelf Deep Models"},{"cite_extract_rate":0.6136363636363631,"cites":[2282,2297,2288,2289,2291,2286,2294,7089,2285,2293,2287,2283,2269,2299,2259,2296,2284,2292,2264,2265,737,8537,2295,2298,2281,2290,2280],"content":"The methods of this category replace the hand-crated feature representations with the pre-trained deep features typically from as AlexNet and VGG.\nA brief summary of these approaches is shown in Table \\ref{table off-the-shelf feature}.\nSong et al. determine discriminative feature configurations of an object class via graph modeling, and train object detectors within the multiple-instance learning paradigm.\nThe deep features of this work are extracted based on the DeCAF scheme~ and AlexNet~.\nBy using the deep features and spatial features to represent each proposal region, Bilen et al. propose a convex clustering process for learning the object models under the weak supervision.\nThe learning objective is able to enforce the similarity among the selected proposal windows.\nIn , Shi and Ferrari develop a curriculum learning strategy to feed training images into the MIL loop in a pre-defined order, where images containing larger objects are learned at the early stages while images containing smaller objects are learned at later stages.\nRen et al. present a bag-splitting-based MIL mechanism that iteratively generated new negative bags from the positive ones.\nThis algorithm can gradually reduce the ambiguity in positive images and thus facilitate the learning of more reliable training instance samples.\nIn , Wei et al. leverage the pre-trained CNN model to implement a Deep Descriptor Transforming process, which can obtain the category-consistent image regions via evaluating the correlations of the descriptors in the convolutional activations of the CNN model.\n\\renewcommand{\\multirowsetup}{\\centering}\n\\begin{table*} [t]\n \\tiny\n \t\\caption{A brief summary of the approaches using single-network training scheme, which is a subcategory in the weakly supervised object localization and detection approaches with deep weakly supervised learning algorithms. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\n \\label{table endtoend}\n \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{2.5cm}<{\\centering}m{1.5cm}<{\\centering}m{1cm}<{\\centering}m{1cm}<{\\centering}m{2.5cm}<{\\centering}m{1.5cm}<{\\centering}m{2.8cm}<{\\centering}m{1.8cm}<{\\centering}}\n\\hline\n \\multirow{2}{*}{\\textbf{Methods}} &\n \\multirow{2}{*}{\\textbf{Detector}} &\n \\multirow{2}{*}{\\textbf{Descriptor}} &\n \\multirow{2}{*}{\\textbf{Prior knowledge} } &\n \\multirow{2}{*}{\\textbf{Extra training data}} &\n \\multirow{2}{*}{\\textbf{Learning model}} &\n \\multirow{2}{*}{\\textbf{Learning strategy}} &\n \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline\n\\hline\nHuang-NIPS-2020&Faster RCNN &CNN &None &ImageNet(tag label) &Faster RCNN*(VGG16/ResNet50) &Proposal attention aggregation and distillation & General objects \\\\\nShen-CVPR-2020&Faster RCNN* &CNN &None &ImageNet(tag label) + Flickr &Faster RCNN*(VGG16) &bagging-mixup + background noise decomposition + clean dat modelling& General objects \\\\\nMai-CVPR-2020&None &CNN &None &ImageNet(tag label) &VGG/InceptionV3 &Integrating discriminative region mining and adversarial erasing & General objects \\\\\nZhang-ECCV-2020 &None &CNN &Cross-image consistency &ImageNet(tag label) &VGG/InceptionV3/ ResNet50 &Inter-image stochastic consistency and global consistency & General objects \\\\\nYang-WACV-2020&None &CNN &None &ImageNet(tag label) &VGG &Weighted classification activation map combination & General objects \\\\\nYang-ICCV-2019 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Faster RCNN* (VGG16) + CAM & Online classifier learning with bounding box regression & General objects \\\\\n Wan-CVPR-2019 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Faster RCNN* (VGG16) & Continuation MIL & General objects \\\\\nShen-CVPR-2019 & WSDDN* (VGG16) & CNN & None & ImageNet(tag label) & Two-stream CNN (WSDDN+DeepLab) & Joint detection and segmentation with cyclic guidance & General objects \\\\\nWan-CVPR-2019 & Fast RCNN & CNN & None & ImageNet(tag label) & Two-stream CNN & continuation instance selection and detector estimation & General objects \\\\\nChoe-CVPR-2019 & None & CNN & None & ImageNet(tag label) & CNN & Feature learning by attention-based dropout & General objects \\\\\nJiang-ICCV-2019 & None &CNN & None & ImageNet(tag label) & VGG16/Resnet101 & Online attention accumulation on CAM & General objects\\\\\nSangineto-TPAMI-2018 & Fast RCNN (VGG16) & CNN & None & ImageNet(tag label) & Fast RCNN (VGG16) & Easy-to-hard & General objects \\\\\nInoue-CVPR-2018 & SSD & CNN & None & PASCAL (full label as source domain),ImageNet & SSD & Domain transfer + pseudo-labeling & Cartoon objects \\\\\n Wan-CVPR-2018 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Faster RCNN* (VGG16) & Min-entropy latent modeling & General objects \\\\\nShen-TNNLS-2018 & VGG16* & CNN & None & ImageNet (tag label) & vgg16* & Object-specific pixel gradient mapping+Iterative component mining & General objects \\\\\nTang-TPAMI-2018 & Fast RCNN* (model ensemble) & CNN & None & ImageNet(tag label) & Fast RCNN* (model ensemble) & MIL+oicr+multi-scale+proposal cluster learning & General objects \\\\\n Zhang-CVPR-2018 & None & CNN & None & ImageNet(tag label) & VGG16* & Adversarial complementary erasing & General objects (for ILSVRC) \\\\\n Choe-BMVC-2018 & ResNet & CNN & None & Tiny ImageNet(tag label) & ResNet & GoogLeNet Resize (GR) augmentation & General objects (for Tiny ImageNet) \\\\\n Zhang-ECCV-2018 & Inception-v3+CAM & CNN & None & ImageNet(tag label) & Inception-v3+CAM & Self-produced guidance learning & General objects (for ILSVRC and CUB) \\\\\n Gao-ECCV-2018 & Fast RCNN* & CNN & Count (human label) & ImageNet(tag label) & Fast RCNN*+Fast RCNN & WSL with count-based region selection & General objects \\\\\n Singh-ICCV-2017 & CAM* (GoogLeNet)& CNN & None & ImageNet (tag label) & CAM* (GoogLeNet) & Random hidding patches & General objects (for ILSVRC) \\\\\n Zhu-ICCV-2017 & None & CNN & None & ImageNet(tag label) & GoogLeNet* & Soft proposal layer+CAM & General objects \\\\\nWan-ICIP-2017 & None & CNN & None & ImageNet(tag label) & CAM*(VGG) & CAM with spatial pyramid pooling layer & General objects \\\\\n Durand-CVPR-2017 & None & CNN & None & ImageNet(tag label) & CAM* (ResNet101) & CAM with multi-map transfer layer & General objects \\\\\nTang-CVPR-2017 & Fast RCNN* (model ensemble) & CNN & None & ImageNet(tag label) & Fast RCNN*+Fast RCNN & MIL+oicr+multi-scale & General objects \\\\\n Jiang-CVPR-2017 & Fast RCNN* (AlexNEt) & CNN & None & PASCAL (edge box),ImageNet & AlexNet+ ROIpool & Region calssification+region selection+multi-scale & General objects \\\\\nDiba-CVPR-2017 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Multi-stream CNN & Cascading LocNet (CAM), SegNet, and MILNet+multi-scale & General objects \\\\\nSelvaraju-ICCV-2017 & None & CNN & None & ImageNet(tag label) & VGG* & Gradient-based class activation mapping & General objects \\\\\nTang-PR-2017 & None & CNN & None & ImageNet(tag label) & Fast RCNN* (VGG-16) & SPP with discovery block and classification block & General objects \\\\\nGudi-BMVC-2017 & None & CNN & None & ImageNet(tag label) & CAM* (VGG-16) & CAM with Spatial Pyramid Averaged Max (SPAM) Pooling & General objects\\\\\n Bilen-CVPR-2016 & Fast RCNN* (model ensemble) & CNN & None & PASCAL (edge box),ImageNet & Fast RCNN* & MIL+multi scale & General objects \\\\\n Kantorov-ECCV-2016 & Fast RCNN* (VGG-F) & CNN & Context & ImageNet(tag label) & Fast RCNN* & MIL+multi-scale & General objects \\\\\n Teh-BMVC-2016 & CNN & CNN & None & PASCAL (edge box),ImageNet & CNN & Proposal attention learning & General objects\\\\\n Durand-CVPR-2016 & None & CNN & None & ImageNet(tag label) & CNN & Feature extraction network+weakly supervised prediction module & General objects \\\\\n Zhou-CVPR-2016 & None & CNN & None & ImageNet(tag label) & GoogLeNet* & Class activation mapping & General objects\\\\\n Oquab-CVPR-2015 & None & CNN & None & ImageNet(tag label) & CNN & Fully convolutional CNN with global max pooling & General objects \\\\\n Wu-CVPR-2015 & None & CNN & None & PASCAL (BING),ImageNet & CNN & Deep multiple instance learning network & General objects\\\\\n\\hline\n \\end{tabular}}\n\\end{table*}","id":"72d7b986-53df-40a6-a211-40fcbb5e65ab","level":"subsection","origin_cites_number":44,"parent_id":"ce302070-cf4f-4d53-a26c-8a51743b26f9","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Off-the-shelf Deep Models"],["subsection","Pre-trained Deep Features"]],"subsections":[],"title":"Pre-trained Deep Features"},{"cite_extract_rate":1,"cites":[2271,2273,2274,2272],"content":"Instead of using the pre-trained deep models as feature extractor, the methods of this category obtain useful information cues (such as the activations in the intermediate network layers and the semantic scores in the output network layer) from the pre-trained deep neural networks to facilitate the weakly supervised learning process.\nThe focus of these approaches mainly lies in the initialization stage of the weakly supervised learning process.\nA brief summary of these approaches is shown in Table \\ref{table pre-learning}.\nBergamo et al. propose a self-taught deep learning approach for localizing objects of interest under weak supervision.\nIn the initialization stage, they design a mask-out strategy based on the deep semantic cues from a pre-trained classification network.\nSpecifically, this method first calculates the degeneration of the image-level classification scores when masking out a certain object proposal region and then selects those with large differences as the interested object regions.\nAfter the initialization stage, this method trains an SVM-based object detector in the subsequent refinement stage for final object localization.\nSimilar to , Bency et al. propose a beam search algorithm to leverage the activation maps of a pre-trained classification network to localize the objects of interest.\nThis method is based on the observation that when image regions centered around objects of interest are classified by a pre-trained DNN, they obtain higher semantic scores than other image regions.\nHoffman et al. develop a transfer learning-based algorithm, where the deep neural network is first trained on both the weakly labeled auxiliary training data and the strongly labeled training data to obtain the background detector.\nThen, an SVM-MIL model is adopted to learn the object detectors based on the potential foreground regions that are obtained by using the pre-trained DNN to screen the image background regions.\nTo overcome the problem that the objects of interest would sometimes co-occur with the distracting image background, Kolesnikov et al. {propose a user-guided weakly supervised learning framework to improve the localization capacity.}\nThis approach first trains a classification network under the image level annotation.\nFor each image, the intermediate feature maps of the pre-trained network are clustered, and the object clutter is identified by a user.","id":"a62482ff-4ed2-4ba2-b72d-45432f53b8f2","level":"subsection","origin_cites_number":4,"parent_id":"ce302070-cf4f-4d53-a26c-8a51743b26f9","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Off-the-shelf Deep Models"],["subsection","Inherent Cues in Deep Models"]],"subsections":[],"title":"Inherent Cues in Deep Models"},{"cite_extract_rate":0.5714285714285711,"cites":[8538,2306,2303,2301,802,2302,2309,2300,2304,2308,2276,2307,2279,2310,2305,2277],"content":"The methods of this category fine-tune the off-the-shelf DNN models during the weakly supervised learning process to obtain strong object detectors .\nA brief summary of these approaches is shown in Table \\ref{table prepost-learning}.\nChen et al. first train CNNs from the web image data via an easy-to-hard learning scheme.\nThe learned deep features are used to mine object locations by using the exemplar-LDA detector .\nThe off-the-shelf RCNN detector is then adopted to learn object models based on their localization results.\nIn , Li et al. first use the mask-out strategy based on the pre-trained classification network to obtain the class-specific object proposals.\nAn SVM-based MIL process is used to localize object instances and\nthe classification network is further fine-tuned on the localized object instances for better performance.\nShi et al. propose to transfer the prior knowledge of \\textit{things} and \\textit{stuff} to help the weakly supervised learning process.\nA semantic segmentation network is trained from the source set (with available bounding-box annotations) to generate the stuff map and thing map for the weakly labeled images in the target set.\nThese maps are used to obtain potential object locations, and the fast RCNN model is adopted in a Deep MIL scheme to train the object detectors.\nRecently, revisits knowledge transfer for training the weakly supervised object detector.\nIn this method, a DNN-based proposal extractor is learned from the source data firstly.\nThe DNN is designed based on the SSD architecture and trained with a semantic hierarchy.\nThe network is then used to provide proposals and other prior knowledge for the weakly labeled images in the target set.\nAn MIL process is used to determine the proposals that cover the objects of interest\nbased on which the fast RCNN model is adopted to learn the final object detectors.\nIn , Zhang et al. first learn to localize the objects of interest via a collaborative self-paced curriculum learning mechanism based on pre-trained deep features.\nThe fast RCNN model is applied to learn object detectors.\n\\renewcommand{\\multirowsetup}{\\centering}\n\\begin{table*} [t]\n \\tiny\n\t\\caption{A brief summary of the approaches with multi-network training, which is a subcategory in the weakly supervised object localization and detection approaches with deep weakly supervised learning algorithms. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\n\t\\label{table multinet} \\resizebox{\\linewidth}{!}{\\begin{tabular}{m{2cm}<{\\centering}m{1.5cm}<{\\centering}m{1.5cm}<{\\centering}m{1.8cm}<{\\centering}m{3cm}<{\\centering}m{2.6cm}<{\\centering}m{3cm}<{\\centering}m{1.5cm}<{\\centering}}\n\\hline\n \\multirow{2}{*}{\\textbf{Methods}} &\n \\multirow{2}{*}{\\textbf{Detector}} &\n \\multirow{2}{*}{\\textbf{Descriptor}} &\n \\multirow{2}{*}{\\textbf{Prior knowledge} } &\n \\multirow{2}{*}{\\textbf{Extra training data}} &\n \\multirow{2}{*}{\\textbf{Learning model}} &\n \\multirow{2}{*}{\\textbf{Learning strategy}} &\n \\multirow{2}{*}{\\textbf{Object category}} \\\\\n & & & & & & &\\\\\n\\hline\n\\hline\nZhang-CVPR-2020 &None &CNN &Common object co-localization &ImageNet(tag label) &VGG/InceptionV3/ResNet50/ DenseNet161 &Classification + pseudo supervised object localization & General objects \\\\\nZhong-ECCV-2020 &Faster RCNN &CNN &Location prior &ImageNet(tag label) + COCO (box label) &Oneclass universal detector + MIL classifier (on ResNet50) &Progressive knowledge transfer & General objects \\\\\nKosugi-ICCV-2019 & Fast RCNN* &CNN & Mask-out prior & ImageNet(tag label) & Mask-out net + OICR* & Mask-our prior-guided label refinement & General objects \\\\\nSingh-CVPR-2019 & Fast RCNN* & CNN & Motion prior & ImageNet(tag label), videos & RPN+WSDDN/OICR (VGG16) & Training RPN using weakly-labeled videos for WSOD & General objects \\\\\nArun-CVPR-2019 & Fast RCNN & CNN & None & ImageNet(tag label) & Fast RCNN (VGG16) + Fast RCNN* (VGG16) & Employ dissimilarity coefficient for modeling uncertainty & General objects \\\\\nLi-TPAMI-2019 & Faster RCNN* (VGG16) & CNN & Objectness (classifier) prior & ImageNet(tag label), ILSVRC2013(box label for unseen categories) & Faster RCNN* (VGG16) *2 & Objectness transfer+MIL +multi-scale & General objects \\\\\n Zhang-ECCV-2018 & fast RCNN* (VGG16) & CNN & None & PASCAL (edge box), ImageNet(tag label) & Multi-view WSDDN+multi view Fast RCNN & Two phase multi-view learning & General objects \\\\\nShen-CVPR-2018 & SSD & CNN & None & ImageNet(tag label) & SSD+Fast RCNN* & MIL+GAN & General objects \\\\\n Zhang-CVPR-2018 & Faster RCNN (VGG16) & CNN & None & ImageNet(tag label) & MIDN+Faster RCNN & WSOD+Pseudo Ground-truth Mining+FOD & General objects \\\\\nZhang-CVPR-2018 & Fast RCNN* (VGG16) & CNN & None & PASCAL (edge box), ImageNet & WSDDN + Fast RCNN* & WSDDN+easy-to-hard FOD & General objects \\\\\nTang-ECCV-2018 & Fast RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Fast RCNN* (VGG16) & Alternating training of WSRPN and WSOD+multi-scale & General objects \\\\\nTao-TMM-2018 & Fast RCNN* (VGG16) & CNN & Web image & Web dataset(weak label),imageNet(tag label) & Midn & Easy-to-hard & General objects \\\\\n Wang-IJCAI-2018 & Faster RCNN (VGG16) & CNN & Model consistency & ImageNet(tag label) & Faster RCNN+Fast RCNN* & MIL+collaborative learning & General objects \\\\\nWei-ECCV-2018 & Faster RCNN* (VGG16) & CNN & Shape prior+ context prior & ImageNet(tag label) & MIDN+CAM+ Deeplab & Tight Box Mining+MIL +OICR+multi-scale & General objects \\\\\n Ge-CVPR-2018 & Faster RCNN (VGG16) & CNN & Local objectness and global attention & ImageNet(tag label) & MIDN,TripNet, GoogleNet, FCN, Fast RCNN & Multi evidence fusion+ outlier filtering +pixel label preidction +box generation+multi-scale & General objects \\\\\nDong-MM-2017 & Fast RCNN* & CNN & None & ImageNet(tag label) & Fast RCNN*+R-FCN & Easy-to-hard & General objects \\\\\n Li-BMVC-2017 & Fast RCNN* & CNN & Shape prior & ImageNet(tag label) & Fast RCNN* +CAM+DeepLab & Easy-to-hard & General objects \\\\\nWang-CVPR-2017 & CNN & CNN & None & ImageNet(tag label) & CAM* & Image-level training+pixel-level fine tuning & Salient objects \\\\\n Sun-CVPR-2016 & None & CNN & None & ImageNet(tag label) & Multi-scale FCN+ CNN(vgg16) & Cascade localization and recognition & General objects \\\\\n Zhang-TGRS-2016 & CNN & CNN & None & ImageNet(tag label), auxiliary data(image label) & CPRNet+LocNet & Alternative training CPRNet and LocNet & Aircraft \\\\\n\\hline\n \\end{tabular}}\n\\end{table*}","id":"eab65442-451f-481c-9954-9b19dd3d9a08","level":"subsection","origin_cites_number":28,"parent_id":"ce302070-cf4f-4d53-a26c-8a51743b26f9","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Off-the-shelf Deep Models"],["subsection","Fine-tuned Deep Models"]],"subsections":[],"title":"Fine-tuned Deep Models"},{"cite_extract_rate":0,"cites":[],"content":"Introducing the off-the-shelf deep models into the weakly supervised object localization or detection framework is the most straightforward approach to integrate deep learning and weakly supervised learning.\nThe methods in this category show that: 1) feature learning is an important factor to improve the weakly supervised learning process; 2) DCNN models can infer discriminative spatial locations when learned under the image-level supervision; 3) pre-training DNN models on large-scale auxiliary training data is a simple but effective way to encode useful cues for the weakly supervised learning process.\nCompared with the classic models,\nthe methods of this category exploit large-scale auxiliary training data to learn powerful feature representations and top-down cues.\nBy using DNN models as the object detector or localizer, a significant performance gain can be obtained.\nHowever, more effective feature learning models can be exploited in the weakly supervised learning process.","id":"0873c3d1-8648-437a-a297-c566cfb8420e","level":"subsection","origin_cites_number":0,"parent_id":"ce302070-cf4f-4d53-a26c-8a51743b26f9","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Off-the-shelf Deep Models"],["subsection","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":0.529411764705882,"cites":[2282,2291,7089,2300,2312,2284,2304,2311,2303],"content":"In this section, we review the methods that learn weakly supervised object localizers or detectors by designing novel deep weakly supervised learning frameworks\\footnote{Notice that here we mainly indicate that the core learning blocks used in the learning frameworks are based on deep learning, while some minor computational components in pre-processing or post-processing, such as proposal extraction and bounding-box modification, et al., are not necessarily be implemented by deep learning.}.\nDifferent from the approaches discussed in previous sections, both the feature representations and the object detectors of the approaches in this category are learned by newly-designed deep neural networks.\nThe whole weakly supervised learning framework may be designed in a compact network model, such as in , or contain several function-distinct DNN components, such as in .\nWe categorize these approaches into two groups using single-network training and multi-network training, respectively.","id":"930ae35a-10a8-4a65-83e8-a12f89f6fe36","level":"section","origin_cites_number":17,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Deep Weakly Supervised Learning"]],"subsections":["352d0dfb-6a12-4bf2-b2bf-5544ab013008","65da26f9-ce63-494b-9e38-4a72e74d3aa0","c9fa8b67-133f-4688-92de-bc8e20c858b9"],"title":"Deep Weakly Supervised Learning"},{"cite_extract_rate":0.777777777777777,"cites":[2259,2282,2298,2290,2314,2280,2315,737,107,7090,2286,687,2313,2295],"content":"The methods of this category are designed with a single deep neural network using the training images (or together with the extracted object proposals) as inputs and the image-level classification labels as the outputs.\nThese approaches do not usually rely on meticulously designed initialization processes to obtain the potential object regions.\nInstead, these methods discover the interested object regions solely based on the end-to-end learning process of the designed DNN models.\nThe DNN models used in these approaches usually have similar feature learning layers as the conventional image classification network, e.g., AlexNet, VGG, GoogleNet, and ResNet, followed by the instance label inferring and image label propagation layers to inherently predict the labels of each proposal region and generate the final image labels from the predicted proposal labels, respectively. Some of the methods contain multiple network streams for online inferring multiple informative cues.\nA brief summary of these approaches are\nshown in Table \\ref{table endtoend}.\nIn the DNN models proposed by Wu et al. and Bilen et al. , the network inputs are the training images and extracted object proposal regions while the outputs are the image-level semantic scores.\nThe first parts of these networks extract the features and infer the labels for each proposal region, and the second parts propagate the proposal scores to the image-level via the max pooling scheme or the two-stream score regularization method .\nZhou et al. present an end-to-end weakly supervised deep learning based on the class activation mapping (CAM).\nThe weights of the feature maps in the intermediate layers are inferred based on the correspondence between the feature map and a certain object category.\nThe feature maps are then combined to form the class activation maps based on the inferred weights, which highlight the locations of the objects of interest. {Notably, this method is determined to be highly efficient in recent work . }\nBuilt on CAM , Durand et al. introduce the multi-map transfer layer and the WILDCAT pooling layer to facilitate the more accurate deep MIL process.\nRecently, a number of two-branch MIL models have been developed in which one is based on a typical deep network and the other one is introduced for weakly supervised learning.\nBased on the WSDDN , Tang et al. integrate MIL branch and the instance classifier refinement branch into a unified deep learning framework such that more accurate online instance classifier learning is realized under the weak supervision.\nIn , Diba et al. propose a weakly supervised cascaded convolutional network, which contains three branches.\nThe first branch adopts the CAM module to generate the class activation maps. \nThe second branch uses the generated class activation maps as the supervision signal to learn a segmentation module to generate the segmentation masks of the objects of interest.\nUsing the candidate object proposals selected based on the obtained segmentation masks as supervision, the third network branch uses a MIL process to mine accurate object locations from the candidate object regions.\nIn , Zhang et al. present a CAM-based network architecture which contains a classification branch and a counterpart classifier branch for object localization.\nSpecifically, the classification branch is used to localize the discriminative object regions, which drives the counterpart classifier branch to discover new and complementary object regions by erasing its discovered regions from the feature maps.\nWan et al. propose a min-entropy latent model (MELM) for weakly supervised object detection based on the assumption that minimizing entropy results in minimum randomness of a system.\nThe network architecture is similar to , but global min-entropy and local min-entropy losses are introduced to train a DNN model to select the proposal cliques of largest object probability and mine truthful object locations from the selected proposal cliques.\nZhang et al. develop a self-generated guidance method for weakly supervised object localization.\nIn this work, a self-generated guidance map is derived from a CAM {layer to help learning features and object location masks from the previous network layers.}\n {More recently, Gao et al. propose a token semantic coupled attention mapping for WSOL, which models the long-range visual dependency of the image regions and thus avoid partial activation. Ren et al. introduce the instance-associative spatial diversification constraints and build the parametric spatial dropout block to address the instance ambiguity and incomplete localization problems. Besides, they additionally adopt a sequential batch back-propagation algorithm, which enables their model to use a large ResNet as the backbone\\footnote{{As discussed in , it is non-trivial to introduce the deeper network backbones, e.g, the deep residual network, into the weakly supervised object detection frameworks as it would encounter dramatic deterioration of detection accuracy and training non-convergence. }}. Although there are other methods that use ResNet as the backbone , there is very limited exploration of using more recent backbone architectures, e.g., DesNet and Res2net , in both WSOL and WSOD frameworks. }","id":"352d0dfb-6a12-4bf2-b2bf-5544ab013008","level":"subsection","origin_cites_number":18,"parent_id":"930ae35a-10a8-4a65-83e8-a12f89f6fe36","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Deep Weakly Supervised Learning"],["subsection","Single-Network Training"]],"subsections":[],"title":"Single-Network Training"},{"cite_extract_rate":0.6666666666666661,"cites":[2259,2298,2306,737,2301,2304,2310,209,8429,2309],"content":"The methods in this school collaborate multiple networks, either in one training stage or in multiple training stages, to accomplish the weakly supervised object localization or detection task.\nThe approaches of this category usually train a network to mine the initial object regions and another network for the detection task under the MIL framework .\nAn additional object detection network, e.g., Fast RCNN, may also be used to train the final object detectors .\nBy integrating multiple networks, these methods tend to achieve better performance both in object localization and detection.\nA brief summary of these approaches is shown in Table \\ref{table multinet}.\nLi et al. propose a multiple instance curriculum learning method, where a network based on the WSDDN model is used to mine candidate object proposals and another one based on the CAM algorithm to generate saliency maps from the selected proposals.\nA curriculum is designed to select confident training examples based on the consistency between the regions outputted by the two networks.\nThe object detectors are then trained by using the confident training examples iteratively.\nDong et al. present a dual-network progressive approach for weakly supervised object detection, where a positive instance selection network and a region refinement network are adopted to minimize the classification error and modify object localization, respectively.\nThese two networks are worked under a co-training paradigm.\nIn , Ge et al. first obtain intermediate object localization and pixel labeling results using a classification network.\nA triplet-loss network and an instance classification network are then constructed to detect outlier and filter object instances.\nFinally, the filtered object instances are used as the supervision to train another Fast RCNN-based detection network.\n {In order to overcome the limitations brought by the imprecise of the extracted object proposals,} Wei et al. propose to mine object proposals with tight boxes to learn weakly supervised object detector. {The assumption is that the proposals with tight boxes are more likely to contain the objects of interest thus mining such kind of proposals would help screen the cluttered background regions. In their approach, }\na semantic segmentation network is first learned using the object localization map generated by CAM as the pseudo ground-truth.\nThe predicted segmentation masks are used to mine object proposals with tight boxes, and fed into the online instance classifier refinement (OICR) network to learn weakly supervised object detector.\n {With the same motivation as , Tang et al. propose to combine a two-stage region proposal network with an OICR network to learn the weakly supervised object detectors. Instead of mining proposals based on the semantic segmentation cue, Tang et al. use both the low-level cues (feature maps in shallow layers) and the high-level cues (semantic scores in deep layers) to mine reliable proposals in the two stages, respectively. The parameters of the region proposal network are obtained based on the network trained by .}\n {Zhang et al. explore the reliability of each training image by evaluating the image difficulty and then feed the images into the learning procedure in an easy-to-hard order. Specifically, the image difficulty is evaluated by diagnosing the localization outputs of the pre-trained WSDDN model based on the concentrateness of the high-scored proposal locations. }\nIn , Zhang et al. use three networks to learn weakly supervised object detectors.\nFirst, an OICR network is trained to generate the initial object regions.\nThen, they train an RPN based on {the pseudo ground-truth boxes obtained after implementing a post-process on the initial object regions, and use the learned RPN} to generate more accurate object locations.\nFinally, fully supervised object detectors are trained based on the obtained object locations.","id":"65da26f9-ce63-494b-9e38-4a72e74d3aa0","level":"subsection","origin_cites_number":15,"parent_id":"930ae35a-10a8-4a65-83e8-a12f89f6fe36","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Deep Weakly Supervised Learning"],["subsection","Multi-Network Training "]],"subsections":[],"title":"Multi-Network Training "},{"cite_extract_rate":0,"cites":[],"content":"Compared with the off-the-shelf deep model-based weakly supervised object detection and localization methods, the deep weakly supervised learning methods exploits the merits of deep learning and weakly supervised learning approaches.\nWithout complex design in the learning initialization stage, the end-to-end deep weakly supervised learning methods have been shown to perform well by introducing the MIL mechanism into the network design of the DCNN models.\nWhile the multi-network training methods can further improve the learning performance by combining multiple function-specific networks.\nOn the other hand, the performance of these methods is limited by whether the information extracted from the weakly-supervised module is effective or not.\nAs such, prior knowledge may be useful to guide the deep learning process without solely relying on the weakly-supervised network.","id":"c9fa8b67-133f-4688-92de-bc8e20c858b9","level":"subsection","origin_cites_number":0,"parent_id":"930ae35a-10a8-4a65-83e8-a12f89f6fe36","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Deep Weakly Supervised Learning"],["subsection","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":0.14285714285714202,"cites":[2286],"content":"During the last decades, significant efforts have been made to develop various methods for learning weakly supervised object localizer or detector.\nFor fair performance evaluation, it is of great importance to introduce some publicly available benchmark datasets and evaluation metrics.\n\\begin{table}[t]\n \\centering\n \\caption{ {Brief summarization of the characteristics of the datasets. The top three are the datasets usually used for the WSOD task, while the bottom two are the common datasets for the WSOL task. \\#Categories indicates the number of object categories. \\#Images indicates the number of images. ``GO'' is short for Generic Objects. }}\n \\begin{tabular}{ccccc}\n \\hline\n Dataset & \\multicolumn{1}{c}{Content} & \\multicolumn{1}{c}{\\#Categories} & \\multicolumn{1}{c}{\\#Images} & \\multicolumn{1}{c}{Metrics} \\\\\n \\hline\\hline\n PASCAL VOC 07 & \\multirow{3}[2]{*}{GO} & \\multirow{3}[2]{*}{20} & 9,962 & \\multirow{3}[2]{*}{mAP, CorLoc} \\\\\n PASCAL VOC 10 & & & 21,738 & \\\\\n PASCAL VOC 12 & & & 22,531 & \\\\\n \\hline\n CUB-200-2011 & Birds & 200 & 11,788 & Top-1/5 Loc \\\\\n ILSVRC 2016 & GO & 1000 & 1.2 M & GT Loc \\\\\n \\hline\n \\end{tabular}\n \\label{tab:dataset}\n\\end{table}\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=9.5 cm]{dataset.jpg}\n \\caption{ {Illustration of examples from the PASCAL VOC (top block), CUB-200-2011 (bottom-left block), and ILSVRC 2016 (bottom-right block) datasets.}}\n \\label{dataset}\n\\end{figure}\nExisting weakly supervised object detection methods are usually evaluated on the PASCAL VOC datasets, including the PASCAL VOC 2007, 2010, and 2012 sets.\nThe PASCAL VOC 2007 , PASCAL VOC 2010 and PASCAL VOC 2012 contain 9,962, 21,738, and 22,531 images of 20 object classes.\nThese three datasets are divided into train, val, and test sets, where the trainval set (5,011 images for PASCAL VOC 2007, 10,869 images for PASCAL VOC 2010, and 11,540 images for PASCAL VOC 2012) are used to train the weakly supervised object detector, and the rest for evaluation.\nThe mean of AP (mAP) metric is used to measure the performance where one object is successfully detected if the intersection over union (IoU) between the ground-truth and predicted boxes is more than 50 percentage.\nThe weakly supervised object localization performance is usually evaluated on the PASCAL VOC, ILSVRC, and CUB datasets.\nOn the PASCAL VOC datasets , weakly supervised object localization methods only use the trainval sets, which are different from the setting in weakly supervised object detection.\nThat is, both the weakly supervised learning process and localization process are implemented on the same image data.\nTo evaluate the localization performance on the PASCAL VOC datasets, the correct localization metric (CorLoc) is adopted, where the bounding-box with the highest class-specific score from each image is examined to be whether correct (with more than 50\\% overlap with the ground-truth box) or not.\nIn addition to the PASCAL VOC datasets, the ILSVRC 2016 dataset \\cite {russakovsky2015imagenet} (i.e., the ImageNet) and CUB-200-2011 dataset are also widely used for performance evaluation.\nThe ILSVRC 2016 dataset contains more than 1.2 million images of 1,000 classes for training, while the validation set, which contains 50,000 images, is used for testing.\nThe CUB-200-2011 dataset contains 11,788 images of 200 categories with 5,994 images for training and 5,794 for testing.\nFor these two datasets, {The commonly-used evaluation metrics are GT-known localization accuracy (GT Loc), Top-1 localization accuracy (Top-1 Loc), and Top-5 localization accuracy (Top-5 Loc). Specifically, GT Loc judges the localization results as correct when the intersection over union (IoU) between the ground-truth bounding box and the estimated box is no less than 50\\%, while Top-1 Loc considers the localization results as correct when the class predicted with the highest score is equal to the ground-truth class and the estimated bounding box has no less than 50\\% IoU with the ground-truth bounding box . Top-5 Loc differs from Top-1 Loc in that it checks if the target label is one of the top 5 predictions. As can be seen, the Top-1 Loc is a harder metric than the GT Loc as it needs to additionally predict the class label correctly. This will dramatically increase the task difficulty when performing under very large or fine-grained semantic spaces. The difficulty of Top-5 Loc is between Top-1 Loc and GT Loc as it requires the model to predict the class label but does not restrict the prediction to be perfectly correct.}\n {We provide a brief summarization of the characteristics of the aforementioned datasets in Table \\ref{tab:dataset}. We additionally show some examples from each dataset to illustrate the bias of image content in different datasets. As displayed in Fig. \\ref{dataset}, the PASCAL datasets contain relatively more complex image content, where multiple object instances and categories may appear in a single image and different images would contain objects with significant scale variations. Although it only contains 20 object categories, its category diversity is higher than the CUB-200-2011 dataset as all the 200 categories in the CUB-200-2011 dataset are related to birds. ILSVRC 2016 contains far more images and categories than the PASCAL datasets. However, the content of the images from the ILSVRC 2016 dataset tends to be simpler than that of the PASCAL datasets---each of the images from the ILSVRC 2016 dataset typically contains only one single object and the objects have more consistent sizes and are placed in clearer background context relative to those from the PASCAL datasets. }\\\\\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=9.5 cm]{videoapp.jpg}\n \\caption{Weakly supervised object localization or detection methods for video understanding. The examples are from . }\n \\label{videoap}\n\\end{figure}","id":"79d342a2-5acc-4290-aaff-2761ce99e494","level":"section","origin_cites_number":7,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Datasets and Evaluation Metrics"]],"subsections":[],"title":"Datasets and Evaluation Metrics"},{"cite_extract_rate":0,"cites":[],"content":"In recent years, weakly supervised object localization and detection techniques have been used in numerous vision problems, especially when it is difficult to collect ground-truth labels.","id":"5cef946e-16aa-4550-aeb5-c885a8308ec2","level":"section","origin_cites_number":0,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Applications"]],"subsections":["2093fbc7-1f16-499f-9309-a24b9da6de5b","5bcd51f6-c4c7-4e80-b0c3-9cec6a4e37d4","6c796257-03d1-4c11-b23e-711e3eea5646","ec87da5b-525f-4f14-a7b7-8c787fb56bb5"],"title":"Applications"},{"cite_extract_rate":0.5,"cites":[2319,2317,2316,2318,8539],"content":"As it is time-consuming to obtain object-level annotations for each video frame, weakly supervised object localization and detection methods have also been applied in the field of video understanding (see Fig. \\ref{videoap}).\nFor example, Chanda et al. build a two-stream learning framework, which adapts the information from the labeled images (source domain) to the weakly labeled videos (target domain).\nIn Zhang et al. propose a self-paced fine-tuning network for learning two network heads to localize and segment the object of interest from the weakly labeled training videos.\nThe network is equipped with the multi-task self-paced learning function which can integrate confident knowledge from each single task (localization or segmentation) and use it to build stronger deep feature representation for both tasks.\nOn the other hand, develop methods to localize temporal actions in the given untrimmed videos, where the main goal is to predict the temporal boundary of each action instance contained in the weakly labeled training videos.\n {Essentially, such a weakly supervised action localization (WSAL) task is an emerging, yet rapidly developing topic in recent years, and the methods for solving this task are highly related to the weakly supervised object detection and localization methods. The additional challenges are: (i) the duration of the interesting action has very large variation, i.e., from a few seconds to thousands of seconds; and (ii) the features extracted to represent the interesting action would be entangled with those of the complex scenes of the video frame. }\n {Notice that when applying to video understanding, there are strong correlations among adjacent video frames. So, additional informative constraints can be introduced to facilitate the weakly supervised object detection or localization under this scenario. }","id":"2093fbc7-1f16-499f-9309-a24b9da6de5b","level":"subsection","origin_cites_number":10,"parent_id":"5cef946e-16aa-4550-aeb5-c885a8308ec2","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Applications"],["subsection","Video Understanding"]],"subsections":[],"title":"Video Understanding"},{"cite_extract_rate":0.6666666666666661,"cites":[2267,2299],"content":"One interesting application of the weakly supervised object localization and detection techniques is the analysis of art images (see Fig. \\ref{art}).\nInoue et al. propose a cross-domain weakly-supervised object detection framework for learning the object detectors from weakly labeled watercolor images.\nA progressive domain adaptation method to transfer the style of the fully-labeled data from the source domain (the normal RGB domain) to the target domain (the watercolor domain) is developed.\nIn Gonthier et al. propose a weakly supervised learning algorithm for detecting objects in paintings.\nThe IconArt database which contains object classes that are absent from the photographs in daily life is developed for performance evaluation.\nIn addition, Crowley\nand Zisserman adopt a weakly supervised object localization scheme for automatically annotating images of gods and animals in decorations on classical Greek vases.\n {When applying to art image analysis, a key challenge arises due to the distinctiveness of the content domain---even the same semantics and image scenes would be presented differently to those in the natural environment. Under this scenario, models with stronger self-domain adaptation capacity would be required for the task.}\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=9.5 cm]{artapp.jpg}\n \\caption{Examples of the application of weakly supervised object localization or detection approaches for the analysis of art images.\n The examples are from , where detection results in different colors in the painting images indicate different types of objects. }\n \\label{art}\n\\end{figure}","id":"5bcd51f6-c4c7-4e80-b0c3-9cec6a4e37d4","level":"subsection","origin_cites_number":3,"parent_id":"5cef946e-16aa-4550-aeb5-c885a8308ec2","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Applications"],["subsection","Art Image Analysis"]],"subsections":[],"title":"Art Image Analysis"},{"cite_extract_rate":0.33333333333333304,"cites":[2321,2322,2320],"content":"As shown in Fig. \\ref{bio}, medical image analysis is one area where the weakly supervised object localization and detection methods are of critical importance as only few annotations of target objects by trained experts in bio-image (e.g., organ or tissues).\nTo alleviate this problem, Hwang and Kim develop a two-stream DNN model to localize the tuberculosis regions from the chest X-ray images.\n {Considering that the medical image-based applications usually do not have the pre-trained networks, this work proposes a weakly supervised learning scheme without requiring any pre-trained network parameters. The proposed network contains a fully connected layer-based classification branch and a CAM-based localization branch with shared convolutional layers for feature extraction. Both of the two branches are supervised by image label annotation, where a weighting parameter is introduced to dynamically control the relative importance between them to gradually switch the focus of the learning process from the classification branch to the localization branch. The authors demonstrate that the features learned from the classification layer at the early stage can provide informative cues to learn the localization branch at the late stage.}\n {For detecting a general type of lesions, Wang et al. model the normal image as the combination of background and noise, while modeling the abnormal images as the combination of background, blood vessels, and noise. With the assumption that the noise for the normal image and abnormal image is the unified distribution, the image data can then be decomposed by the low-rank subspace learning technique to obtain the vessel areas. }\nIn , Gondal et al. apply the weakly supervised object detection network on the retina images and achieve few false positives with high sensitivity on the lesion-level prediction.\nLi et al. apply a sparse coding-based weakly supervised learning method for localizing actinophrys in microscopic images.\nDubost et al. propose weakly supervised regression neural networks for detecting brain lesions. Besides, some recent works also show great research interests in weakly supervised learning-based brain image analysis, such as brain disease prognosis , brain tumor or lesion segmentation , brain structure estimation , etc.\n {Notice that compared to common images, medical imaging data usually suffers from issues of low contrast and limited texture. Fortunately, some spatial priors could be obtained for different organs or lesions. These priors can be used to guide the weakly supervised learning process on medical imaging data.}\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=9.5 cm]{bioimage.jpg}\n \\caption{Examples of the application of weakly supervised object localization or detection approaches in medical image analysis. The examples are from . }\n \\label{bio}\n\\end{figure}","id":"6c796257-03d1-4c11-b23e-711e3eea5646","level":"subsection","origin_cites_number":9,"parent_id":"5cef946e-16aa-4550-aeb5-c885a8308ec2","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Applications"],["subsection","Medical Imaging"]],"subsections":[],"title":"Medical Imaging"},{"cite_extract_rate":0,"cites":[],"content":"Remote sensing imagery analysis is one of the most widely studied applications based on weakly supervised object localization and detection, where the input images are usually of large scale and the annotation process tends to be very time-consuming (see Fig. \\ref{remote}).\nZhang et al. propose a saliency-based weakly supervised detector learning method to learn the detectors of the airplane, vehicle, and airport from the remote sensing images collected from different sensors.\nIn this work, a weakly supervised object detection benchmark dataset for remote sensing imagery analysis is developed.\nHan et al. and Zhou et al. introduce the Bayesian inference and negative bootstrapping methods, respectively, for effective weakly supervised detectors for remote sensing images.\nIn , Zhang et al. propose a coupled CNN method which combines a candidate region proposal network and a localization network to detect aircrafts in images.\n {When applying to remote sensing imagery analysis, the target objects are usually very small in size, which would dramatically increase the localization and detection difficulty given only weak supervision. }\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=9.5 cm]{remote.jpg}\n \\caption{Examples of the application of weakly supervised object localization or detection approaches in remote sensing imagery analysis. The examples are from . }\n \\label{remote}\n\\end{figure}","id":"ec87da5b-525f-4f14-a7b7-8c787fb56bb5","level":"subsection","origin_cites_number":8,"parent_id":"5cef946e-16aa-4550-aeb5-c885a8308ec2","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Applications"],["subsection","Remote Sensing Imagery Analysis"]],"subsections":[],"title":"Remote Sensing Imagery Analysis"},{"cite_extract_rate":0,"cites":[],"content":"We discuss the issues to be addressed in this field for future research.","id":"d8feaee7-5d65-4327-96a9-38b968bfe1c4","level":"section","origin_cites_number":0,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Future Directions"]],"subsections":["e95954d7-ff46-4ac2-807f-3d34b6e59526","abffd2fa-2c1a-4f41-8253-4df7d27f9686","56c3e88b-fc74-49e6-aa73-cff57609f8ef","47c758a2-7983-451c-ad2b-fa45ee0b2b84","8b5b77db-9af6-4869-b487-c1496c6cc1b1"],"title":"Future Directions"},{"cite_extract_rate":0.33333333333333304,"cites":[2269,2323,737],"content":"Weakly supervised object localization or detection methods can be easily formulated within the MIL framework.\nEarly methods in this filed usually add prior knowledge or post regularization on the classic MIL models, such as LSVM , while the current research obtains the breakthrough by building deep MIL models . To further improve the weakly supervised learning performance, efforts should be made to introduce the most advanced ideas and techniques in the research filed of MIL, such as the set-level problem the key instance shift issue and the scalable issue in MIL.\nFurther research towards more advanced MIL techniques would also bring helpful insights for the WSOL and WSOD in the future.","id":"e95954d7-ff46-4ac2-807f-3d34b6e59526","level":"subsection","origin_cites_number":9,"parent_id":"d8feaee7-5d65-4327-96a9-38b968bfe1c4","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Future Directions"],["subsection","Multiple Instance Learning"]],"subsections":[],"title":"Multiple Instance Learning"},{"cite_extract_rate":0.5,"cites":[7091,2260,2324,1751,8429,8540],"content":"Another future direction is to combine multiple weakly supervised learning tasks into a unified learning framework. These tasks may include object detection , semantic segmentation , instance segmentation , 3D shape reconstruction , and depth estimation .\nEssentially, efforts for simultaneously accomplishing multiple aforementioned tasks have been made in the conventional fully supervised learning scenario , which have demonstrated that such learning mechanism can bring helpful information from one task to the other ones.\nThe methods proposed by Zhang et al. is an early attempt to implement such a weakly supervised multi-task learning mechanism and the experimental results show that training object segmentation and 3D shape reconsecration models jointly indeed benefits the both weakly supervised learning tasks. With similar spirits to , Zhang et al. and Shen et al. establish a self-paced fine-tuning network and a cyclic guidance network to jointly learn object localization and segmentation models under the weak supervision, respectively. {Under the multi-task weakly supervised learning scenario, one key problem is that the learning ambiguity of each individual task might be aggregated and the imprecise prediction on one task might affect the learning on other tasks. To deal with this problem, one needs to disentangle the complex multi-task learning, separately learning each individual task first, and then leveraging the confidence knowledge from each task to provide informative priors to guide the learning processes of the other tasks. }","id":"abffd2fa-2c1a-4f41-8253-4df7d27f9686","level":"subsection","origin_cites_number":12,"parent_id":"d8feaee7-5d65-4327-96a9-38b968bfe1c4","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Future Directions"],["subsection","Multi-Task Learning"]],"subsections":[],"title":"Multi-Task Learning"},{"cite_extract_rate":0.4,"cites":[2306,8537],"content":"To address the learning under uncertainty issue that is inherently existed in the weakly supervised learning process, robust learning strategy will become one of the key techniques in the future. The goal is to alleviate the influence of the noisy samples during the learning process. In implementation, such learning strategy is usually achieved by selecting easy and confident training samples in the early learning stages while using hard and more ambiguous training samples in the late learning stages. Essentially, a number of recent methods have already introduced the robust learning strategies into their learning frameworks. For example, Shi and Ferrari propose a curriculum learning strategy to feed training images into the WSOL learning loop in order from images containing bigger objects down to smaller ones. The training order is determined by the size of the object, which is estimated based on a regression model. Similarly, Zhang et al. design a zigzag learning strategy, where they first develop a criterion to automatically rank the localization difficulty of an image, and then learn the detector progressively by feeding examples with increasing difficulty. As can be seen, these methods are just intuitive ways to introduce robust learning strategy into the weakly supervised object localization and detection frameworks, while they have already achieved obvious performance gains when compared with the conventional learning strategy. Along this line, Zhang et al. propose a self-paced curriculum learning framework for weakly supervised object detection. By integrating the curriculum learning with the self-paced learning , the established learning framework provides a more theoretical-sounded way to improve the learning robustness. However, the solid robust learning theory is still lack in this research field.","id":"56c3e88b-fc74-49e6-aa73-cff57609f8ef","level":"subsection","origin_cites_number":5,"parent_id":"d8feaee7-5d65-4327-96a9-38b968bfe1c4","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Future Directions"],["subsection","Robust Learning Theory"]],"subsections":[],"title":"Robust Learning Theory"},{"cite_extract_rate":0.5714285714285711,"cites":[2329,2328,2326,2280,2325,2330,2327,8541],"content":"Besides the conventional CNN models, it is also worth trying to apply some more advanced learning models into the learning process of the weakly supervised object detector. Here we give two examples. The first one is the deep reinforcement learning.\nAccording to~, biological vision systems are believed to have a sequential process with changing retinal fixations that gradually accumulate evidence of certainty when searching or localizing objects. Several existing methods~ have also demonstrated that designing deep reinforcement learning frameworks to model such a sequential searching process can indeed help to address the object localization, detection, and tracking problems in the computer vision community. Thus, it is highly desirable, both biologically and computationally, to explore deep reinforcement learning models that facilitate the weakly supervised object localization and detection systems in such a sequential searching process .\nThe second one is the generative adversary learning. As we know, generative adversary learning has been demonstrated to have advantages in unsupervised and semi-supervised learning scenarios . It can generate the desired data distribution based on very weak supervision, i.e, ``real'' or ``fake''. Such capacity endows generative adversary learning very large potential in solving the weakly supervised object localization and detection problems. Although existing methods, such as , have already made efforts to introduce such a learning mechanism into the weakly supervised object localization and detection, there is still much room for improvement along this research direction.","id":"47c758a2-7983-451c-ad2b-fa45ee0b2b84","level":"subsection","origin_cites_number":14,"parent_id":"d8feaee7-5d65-4327-96a9-38b968bfe1c4","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Future Directions"],["subsection","Reinforcement and Adversarial Learning"]],"subsections":[],"title":"Reinforcement and Adversarial Learning"},{"cite_extract_rate":0.533333333333333,"cites":[2332,2289,8538,2334,2307,2331,2333,2309],"content":"From Table \\ref{table endtoend} and Table \\ref{table multinet}, we can observe that most of the current deep weakly supervised object detection methods have not introduced any prior knowledge into their learning frameworks. However, from our review on classic models (see Sec. \\ref{Classic Models}), prior knowledges actually play important roles in avoiding the weakly supervised learning process from drifting to trivial solutions. Considering this issue, some recent works utilize prior knowledges of saliency , objectness , shape , count , human action , human object interaction , mask-out scoring in their frameworks. However, research towards building effective deep MIL frameworks (such as the one with prior knowledge distillation or cross domain adaptation ) to embed helpful prior knowledge into the weakly supervised learning process needs to be further explored in the future. In addition, the co-occurring patterns mined in co-saliency detection and object co-localization approaches can also be used as informative priors to guide the deep multiple instance learning process in weakly supervised object localization and detection.","id":"8b5b77db-9af6-4869-b487-c1496c6cc1b1","level":"subsection","origin_cites_number":15,"parent_id":"d8feaee7-5d65-4327-96a9-38b968bfe1c4","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Future Directions"],["subsection","Prior-guided Deep MIL"]],"subsections":[],"title":"Prior-guided Deep MIL"},{"cite_extract_rate":0,"cites":[],"content":"In this paper, we provide a comprehensive survey of existing literatures in the research field of weakly supervised object localization and detection. We start with the introduction of the definition of the task and the key challenges that make the weakly supervised learning process hard to implement. Then, we introduce the development history of this field, the taxonomy of methods for weakly supervised object localization and detection, and the relationship between different categories. After reviewing existing literatures in each category of methodology, we introduce the benchmark datasets and evaluation metrics that are widely used in this filed, which are followed by the reviewing of the applications of the existing weakly supervised object localization and detection algorithms. Finally, we point out several future directions that may further promote the development of this research field.\n{\\small\n\\bibliographystyle{ieee}\n\\bibliography{WSsurvey}\n}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{dingwen_zhang}}]{Dingwen Zhang}\nreceived his Ph.D. degree from the Northwestern Polytechnical University, Xi'an, China, in 2018. He is currently a full professor in School of Automation, Northwestern Polytechnical University. From 2015 to 2017, he was a visiting scholar at the Robotic Institute, Carnegie Mellon University. His research interests include computer vision and multimedia processing, especially on saliency detection, video object segmentation, and weakly supervised learning.\n\\vspace*{-30pt}\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{HJW}}]{Junwei Han}\n(M'12-SM'15) is a Professor with Northwestern Polytechnical University, Xi'an, China. He received Ph.D. degree in Northwestern Polytechnical University in 2003. He was a Research Fellow in Nanyang Technological University, The Chinese University of Hong Kong, and University of Dundee. His research interests include computer vision and brain imaging analysis. He has published over 100 papers in IEEE TRANSACTIONS and top tier conferences. He is currently an Associate Editor of IEEE Trans. on Neural Networks and Learning Systems, IEEE Trans. on Circuits and Systems for Video Technology, IEEE Trans. Cognitive and Developmental Systems, IEEE Trans. on Human-Machine Systems, Neurocomputing, and Machine Vision and Applications.\n\\vspace*{-30pt}\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{GC}}]{Gong Cheng}\nreceived the B.S. degree from Xidian University, Xi'an, China, in 2007, and the M.S. and Ph.D. degrees from Northwestern Polytechnical University, Xi'an, in 2010 and 2013, respectively. He is currently a Professor with Northwestern Polytechnical University. His main research interests are computer vision and pattern recognition.\n\\vspace*{-30pt}\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{mhyang}}]{Ming-Hsuan Yang}\nreceived the PhD degree in computer science from the University of Illinois at Urbana-Champaign, in 2000. He is a professor in Electrical Engineering and Computer Science from the University of California, Merced. He served as an associate editor of the IEEE Transactions on Pattern Analysis and Machine Intelligence from 2007 to 2011, and is an associate editor of the International Journal of Computer Vision, the Computer Vision and Image Understanding, the Image and Vision Computing, and the Journal of Artificial Intelligence Research. He received the NSF CAREER award in 2012, and the Google Faculty Award in 2009. He is a Fellow of the IEEE and Senior Member of the ACM.\n\\vspace*{-30pt}\n\\end{IEEEbiography}\n\\end{document}","id":"92209b0e-e020-4aa4-8af9-77a6bcf0b9af","level":"section","origin_cites_number":0,"parent_id":"3a2d911b-3708-4f3a-880a-851136519e79","prefix_titles":[["title","Weakly Supervised Object Localization and Detection: A Survey"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ]\n ],\n \"subsections\": [\n \"6019d392-fb6b-4290-8c58-a72268a80ff9\",\n \"a3e94559-bb78-4175-a692-379986435e13\",\n \"bba63d48-815b-4407-9de8-7df72823a004\",\n \"ce302070-cf4f-4d53-a26c-8a51743b26f9\",\n \"930ae35a-10a8-4a65-83e8-a12f89f6fe36\",\n \"79d342a2-5acc-4290-aaff-2761ce99e494\",\n \"5cef946e-16aa-4550-aeb5-c885a8308ec2\",\n \"d8feaee7-5d65-4327-96a9-38b968bfe1c4\",\n \"92209b0e-e020-4aa4-8af9-77a6bcf0b9af\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 2259,\n 209,\n 2260\n ],\n \"content\": \"\\\\label{sec:introduction}}\\n\\\\IEEEPARstart{W}{eakly} supervised learning (WSL) has recently received much attention in computer vision community.\\nA plethora of methods on this topic have been proposed in the past decade to address the challenging computer vision tasks including semantic segmentation~, object detection~, and 3D reconstruction~, to name a few.\\nAs shown in Fig.~\\\\ref{illustration}, a WSL problem is defined as the learning process when some partial information regarding the task (e.g., class label or object location) on a small subset of the data points is at our disposal.\\nCompared to the conventional learning framework, e.g., fully supervised learning approaches, the WSL framework needs to operate on the small amount of weakly-labelled training data to learn the target model, which alleviates a huge amount of human labor to annotate training samples.\\nIt can also facilitate the learning process when the fine-grained annotation\\nis extremely labor intensive and time consuming to even obtain the whole labeled data required by the fully-supervised approaches.\\nWhile a plethora of WSL-based vision methods have been developed, this survey mainly focuses on the task of weakly supervised object localization and detection, which is shown as the red dot in Fig.~\\\\ref{illustration}.\\nIt is well-known that object localization and detection is a fundamental research problem in computer vision.\\nLearning object localization and detection models under weak supervision has attracted much attention in the past decades.\\nWhile existing methods treat weakly supervised object localization (WSOL) and weakly supervised object detection (WSOD) as two different tasks\\\\footnote{ {The difference between WSOL and WSOD mainly lies in that WSOL mainly aims at localizing a single known (or unknown) object from each given image scene. The goal of WSOD is to instead detect every possible object instance from the given image scene. This makes WSOD a little more difficult than WSOL.}}, we consider these as a common task due to several reasons: 1) these tasks learn with the same image-level human annotation; 2) these two tasks need certain supervision as input and usually aim to localize objects on the bounding-box level as output; 3) WSOD task can be accomplished by directly training off-the-shelf fully supervised object detectors on the object locations obtained from WSOL.\\nDuring the last decade, considerable efforts have been made to develop various approaches for learning object detectors with weak supervision.\\nSome of the existing algorithms only learn weakly supervised object detectors for one or several certain object categories, such as vehicles~, traffic signs~, pedestrians~, faces~, tuberculosis bacilli~, aircrafts~, and human actions~.\\nWhile other approaches, e.g., , focus on developing weakly supervised learning frameworks for unconstrained object categories, i.e., learning frameworks can be extended to learn object detector for the given category-specific weakly-labelled training images.\\nAs enormous methods have been developed for these important tasks, a comprehensive review of the literature concerning weakly supervised object localization and detection is of great importance.\\nAs weakly supervised object localization and detection methods mainly exploit the image-level manual annotation, the learning frameworks not only need to address the typical issues, such as the intra-class variations in appearance, transformation, scale and aspect ratio, encountered in conventional fully supervised object localization and detection task, but also the \\\\textbf{learning under uncertainty} challenges caused by the inconsistency between human annotations and real supervisory signals.\\nIn weakly supervised object localization and detection, the accuracy of object locations and learning processes are closely related.\\nThe key is to propagate the image-level supervisory signals to the instance-level (bounding-box-level) training data for the learning processes.\\nAs each training image can be labeled by numerous bounding boxes of different accuracy, propagating such weak supervision inevitably involves a large amount of ambiguous and noisy information as each training instance.\\nMore specifically, the \\\\textbf{learning under uncertainty} issue would cause the following challenges that make the weakly supervised learning process challenging:\\n\\\\begin{itemize}\\n \\\\item \\\\textbf{Learning with inaccurate instance locations:} This issue is mainly caused by the definition ambiguity in object parts and context.\\n Without precise annotation or definition, it is difficult for a learner to decide whether an object category label associates with a discriminative object part, the whole object region, or the object with a certain context region.\\n As a result, the bounding-box instance locations inferred by the learner may contain many inaccurate samples including the ones with local object parts or undesired contextual regions.\\n These samples would negatively affect the performance of WSL-based detectors.\\n \\\\item \\\\textbf{Learning with noisy samples:} Even when the bounding-box locations can be precisely labeled, the training examples enclosed by bounding-boxes may still be noisy as background pixels are usually included.\\n As there is no additional information to separate foreground objects from the background, the learner may tag a ``background'' label to an object region when it fails to recognize the object category.\\n In addition, the learner may mistakenly label a bounding-box that contains a bicycle as a motorcycle, as these two object categories share many similar features.\\n \\\\item \\\\textbf{Learning with domain shifts:} For a certain object category, the image regions localized during the learning process may only contain samples with limited diversity in object shape, appearance, scale, and view angle.\\n This makes the subsequent learning process biased to limited knowledge of the object category and does not generalize well for test samples.\\n For instance, a learner can hardly localize or detect a flying swan when all the training samples contain the swimming ones on lakes.\\n This issue happens frequently among the weakly supervised learning process when there is a large gap between the training and testing domains.\\n \\\\item \\\\textbf{Learning with insufficient instance samples:}\\n Similar to the issues in conventional learning methods, it is difficult to train effective object detectors under the weakly-supervised setting when the amount of training samples is limited.\\n In addition, the number of positive samples is usually much smaller than that of negative samples for binary classes.\\n Furthermore, the data distributions for a large number of categories is usually long-tailed.\\n This issues are significantly exaggerated for the WSL-based methods using deep learning.\\n\\\\end{itemize}\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=1\\\\linewidth]{illustration.jpg}\\n \\\\caption{Illustration of the weakly supervised learning tasks in computer vision community. The blue area in the top block indicates the conventional fully-supervised learning tasks, while the red area in the top block indicates the weakly supervised learning tasks. The coordinate axes shows different levels of human annotation or supervision requirement, from low cost to high cost. Notice that the high cost annotation can be transformed to low cost annotation easily, e.g., from bounding-box level to image level, whereas the low cost annotation is hard to be transformed to high cost annotation. In the bottom block, we also show the label cost, in terms of annotation time, and the examples of different type of annotations. In this survey, we mainly focus on reviewing the research progress in weakly supervised object localization and detection, i.e., the red dot in the top block.}\\n \\\\label{illustration}\\n\\\\end{figure}\\nTo address the above-mentioned issues in learning weakly supervised object detectors, existing methods are usually constructed based on two steps: initialization and refinement.\\nThe initialization stage is used to leverage certain prior knowledge to propagate image-level annotation into instance level, and thus can generate instance-level annotation (but with label noise, sample bias, and limited quality in location accuracy) for the learning process.\\nThe refinement stage is used to leverage new instance samples obtained from the first stage to mine truthful knowledge about the objects of interest gradually and finally obtain the desired object models for localization and detection.\\nThese two learning stages need to collaborate to address the aforementioned five-fold challenges.\\nIn initialization stage, efforts should be made to improve the annotation quality as much as possible to generate training instances with proper locations, accurate labels, high diversity, and high recall rate.\\nAs the annotation quality obtained in the learning stage cannot be perfect, in the refinement stage, further efforts should be made to improve the learner's robustness to cope with the inaccurate instance location, noisy examples, biased instance sample, insufficient instance sample issues as well as the capacity to take advantage of the unlabelled instance samples.\\nWhen properly addressing the problems in each learning stage,\\ngood weakly supervised object detectors can be learned.\\nIn this work, we review the existing weakly supervised object localization and detection approaches\\\\footnote{Some early methods, such as~, learn to localize category-wise key points under the weak supervision, while this survey mainly focuses on the methods for localizing instances with bounding-boxes.}, which are divided into three main categories and eight subcategories.\\nThese three main categories are based on classic approaches, feature representations from off-the-shelf deep models, and deep learning frameworks.\\nThe eight subcategories include approaches for initialization, refinement, initialization and refinement, pre-trained deep features, inherent cues in deep models, fine-tuned deep models, single-network training, and multi-network training.\\nWe further discuss the relationship between the approaches in different categories.\\nIn addition, we also discuss open problems and challenges of current studies and propose several promising research directions in the future for constructing more effective weakly supervised object localization and detection frameworks.\\n\\\\begin{figure*}[t]\\n \\\\centering\\n \\\\includegraphics[width=1\\\\linewidth]{category2.jpg}\\n \\\\caption{In the left block, taxonomy of the existing approaches for weakly supervised object localization and detection, which includes three main\\n categories and eight subcategories.\\n In the right block, the relationships between the approaches in different categories are shown.\\n}\\n \\\\label{taxonomy}\\n\\\\end{figure*}\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=1\\\\linewidth]{history.JPG}\\n \\\\caption{Developments of weakly supervised localization and detection methods. The yellow histogram shows the number of publications in this research field in each year, and the curves show the number of proposed methods each year for a particular category of approach.}\\n \\\\label{history}\\n\\\\end{figure}\",\n \"id\": \"6019d392-fb6b-4290-8c58-a72268a80ff9\",\n \"level\": \"section\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.1,\n \"cites\": [\n 2261,\n 2262,\n 2263\n ],\n \"content\": \"\\\\label{Taxonomy}\\nIn the last decade, a plethora of methods have been developed for weakly supervised object localization and detection.\\nWe can generally categorize existing methods based on classic formulations, feature representations from off-the-shelf deep models, and deep weakly supervised learning algorithms.\\nWhile inside each main category, we further divide the approaches into two or three subcategories.\\nFig.~\\\\ref{taxonomy} shows our taxonomy of the studies in the research field of weakly supervised object localization and detection.\\nIn addition, Fig. \\\\ref{history} reviews the development history of each of the main category as well as that of the whole research field.\\nA few approaches based on classic formulations appeared around 2002.\\nFrom 2002 to 2009, the research in this field went through a very slow pace.\\nSince 2014, numerous approaches based on both classic formulations and learned feature representations from deep models have been developed and received much attention.\\nWhile in the last few years, more approaches solely based on deep learning have become the main stream to address the problems of weakly supervised object localization and detection.\\nWhile a plethora of methods have been developed to address different aspects of these problems in the past decades, this field is gaining increasing attention.\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n \\\\begin{table*} [t]\\n \\\\tiny\\n \\\\caption{Summary of the approaches for initialization, which is a subcategory in the weakly supervised object localization and detection approaches that learn by classic models. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\\n \\\\label{table initialization}\\n \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{2.5cm}<{\\\\centering}m{1.2cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{3.1cm}<{\\\\centering}m{1.8cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{3.5cm}<{\\\\centering}m{1.8cm}<{\\\\centering}}\\n \\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} & \\\\multirow{2}{*}{\\\\textbf{Detector}} & \\\\multirow{2}{*}{\\\\textbf{Descriptor}} & \\\\multirow{2}{*}{\\\\textbf{Prior knowledge}} & \\\\multirow{2}{*}{\\\\textbf{Extra training data}} & \\\\multirow{2}{*}{\\\\textbf{Learning model}} & \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} & \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline \\\\hline\\nCao-PR-2017 & SVM & HOG+PCA & Road map prior + density prior & None & SVM & MIL with density estimation & Vehicle in satellite imagery \\\\\\\\\\nShi-TPAMI-2015 & DPM & SIFT+Lab+LBP, BOW & Appearance prior + geometry prior & None & Topic model & Bayesian inference & General objects \\\\\\\\\\nTang-ICIP-2014 & DPM & HOG & Saliency (objectness score) & None & DPM & Select initial boxes + DPM training & General objects \\\\\\\\\\n Xie-VCIP-2013 & None & SIFT & Intra-class consistency & None & None & Low-rank and sparse coding & General objects \\\\\\\\\\nSiva-CVPR-2013 & DPM & Lab+SIFT & Saliency & Labelme + PASCAL07, 12 (unlabelled) &None & None & General objects \\\\\\\\\\n Shi-ICCV-2013 & DPM & SIFT & Appearance prior (spatial distribution, size, saliency) & None & Topic model & Bayesian inference & General objects \\\\\\\\\\n Sikka-FG-2013 & None & SIFT, LLC, BOW & None & None & MilBoost & Generating multiple segments for initialization and use Milboost for learning & Pain (on face) \\\\\\\\\\n Siva-ECCV-2012& None & SIFT+BOW & Iter-class variance + saliency & None & None & Negative mining & General objects \\\\\\\\\\n Shi-BMVC-2012 & None & BOW & Mapping relationship between the box overlap and appearance similarity & Part of PASCAL 07 (box annotation) & RankSVM & Transfer learning by ranking & General objects \\\\\\\\\\n Khan-AAPRW-2011 & DPM & Phog/phow & None & Internet image (weakly annotated) & MIL & Learning from internet image & Pascal@8 \\\\\\\\\\nZhang-BMVC-2010 & SVM & IHOF & Co-occurrence & None & SVM & High order feature learning by exploring co-occurrence & General objects \\\\\\\\\\n\\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n\\\\begin{table*} [t]\\n \\\\tiny\\n\\t\\\\caption{Summary of the approaches for refinement, which is a subcategory in the weakly supervised object localization and detection approaches that learn by classic models.\\n\\tHere, * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\\n\\t \\\\label{table refinement} \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{2.5cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{3cm}<{\\\\centering}m{2cm}<{\\\\centering}m{2cm}<{\\\\centering}m{2cm}<{\\\\centering}m{3.1cm}<{\\\\centering}m{2.5cm}<{\\\\centering}}\\n\\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} &\\n \\\\multirow{2}{*}{\\\\textbf{Detector}} &\\n \\\\multirow{2}{*}{\\\\textbf{Descriptor}} &\\n \\\\multirow{2}{*}{\\\\textbf{Prior knowledge} } &\\n \\\\multirow{2}{*}{\\\\textbf{Extra training data}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning model}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} &\\n \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline\\n\\\\hline\\n Wang-TMI-2018 & None & Color & None & None & Low-rank model & Low-rank Factorization & General lesion \\\\\\\\\\n Wang-TC-2017 & None & SIFT+LAB & None & None & Probability model & BOW learning+instance labeling & General objects \\\\\\\\\\n Cholakkal-CVPR-2016 & None & SIFT & Saliency & None & SVM* & ScSPM-based top down saliency & Salient objects \\\\\\\\\\n Zadrija-GCPR-2015 & None & SIFT+FV & None & None & GMM + linear classifier & Patch-level spatial layout learning & Traffic sign \\\\\\\\\\n Krapac-ICCVTA-2015 & None & SIFT+FV & None & None & Sparse logistic regression & Sparse classification & Traffic sign \\\\\\\\\\n Cinbis-CVPR-2014 & SVM & FV & Center prior & None & SVM & Multi-fold MIL & General objects \\\\\\\\\\n Wang-ICIP-2014 & SVM + graph model & SIFT & None & None & SVM + graph model & Maximal entropy random walk & Car, dog \\\\\\\\\\n Wang-ICIP-2014 & SVM & SIFT & None & None & SVM & Clustering for window mining & General objects \\\\\\\\\\n Tang-CVPR-2014 & None & SIFT & Saliency & None & Boolean constrained quadratic program & Mine similarity and discriminativeness both for image and box & General objects \\\\\\\\\\n Hoai-PR-2014 & SVM & SIFT,BOW & None & None & SVM* & Localization-classification SVM & Face,car, human motion \\\\\\\\\\n Wang-WACV-2013 & Task-specific detectors & HOG/SC & Background\\\\ saliency & None & MIL+AdaBoost* & Soft-label Boosting after MIL & Vehicle, pedestrain\\\\\\\\\\n Kanezaki-MM-2013 & Linear classifiers & 3D voxel feature (color+C3HLAC +Intensity, texture, GRSD) & None & None & Linear classifiers & Multi-class MIL & Balls, tools \\\\\\\\\\n Pandey-ICCV-2011 & DPM* & HOG & None & None & DPM* & Learning DPM with fully latent variable & General objects \\\\\\\\\\nBlaschko-NIPS-2010 & None & BOW/HOG & None & None & SVM* & Learning SVM with structured output ranking objective & Cat, pedestrian \\\\\\\\\\nHoai-ICCV-2009 & SVM & SIFT,BOW & None & None & SVM* & Localization-classification SVM & Face,car, human motion \\\\\\\\\\n Galleguillos-ECCV-2008 & None & SIFT+BOW & None & None & MilBoost & Train MIL classifier for localization & Landmarks, faces, airplanes, leopard, motorbike, car \\\\\\\\\\n Rosenberg-BMVC-2002 & GMM & Orientation deriviate filters & Exampler prior & Training exampler (box annotation) & GMM & Learning from exampler training data to weakly labelled training data & Telephone \\\\\\\\\\n\\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n\\\\begin{table*} [t]\\n \\\\tiny\\n\\t\\\\caption{Approaches for both initialization and refinement, which is a subcategory in the weakly supervised object localization and detection approaches that learn by classic models. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\\n\\t \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{1.8cm}<{\\\\centering}m{1.2cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{2cm}<{\\\\centering}m{1.8cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{3cm}<{\\\\centering}m{1.5cm}<{\\\\centering}}\\n\\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} &\\n \\\\multirow{2}{*}{\\\\textbf{Detector}} &\\n \\\\multirow{2}{*}{\\\\textbf{Descriptor}} &\\n \\\\multirow{2}{*}{\\\\textbf{Prior knowledge} } &\\n \\\\multirow{2}{*}{\\\\textbf{Extra training data}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning model}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} &\\n \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline\\n\\\\hline\\n Wang-cvpr-2013 & None & Color+SIFT & None & None & HST+SVM & Joint parsing and attribute localization & Scene attributes \\\\\\\\\\n Deselaers-IJCV-2012 & DPM & GIST+CH+BOW+HOG & Generic knowledge & Meta-training data with box annotation & CRF+DPM & Learning appearance model by trainsferring genearic knowledge & General objects \\\\\\\\\\n Siva-ICCV-2011 & DPM & SIFT+BOW+HOG & Inter-class prior + intra-class prior & None & DPM & Model drift learning & General objects \\\\\\\\\\nDeselaers-ECCV-2010 & DPM & GIST+CH+BOW+HOG & Generic knowledge & Meta-training data with box annotation & CRF+DPM & Learning appearance model by transferring generic knowledge & General objects \\\\\\\\ \\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\\nAs shown in the right block of Fig.~\\\\ref{taxonomy}, methods in\\nmain categories are related in several aspects.\\nNumerous methods are developed based on the classic formulations with the advances of feature representations from deep models.\\nSimilarly, a number of methods based solely on deep models are end-to-end trainable by considering classic formulations and feature extraction schemes.\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=7 cm]{subcategory1.jpg}\\n \\\\caption{Flowchart of the weakly supervised object localization and detection approaches using classic models.}\\n \\\\label{nondeep flowchart}\\n\\\\end{figure}\",\n \"id\": \"a3e94559-bb78-4175-a692-379986435e13\",\n \"level\": \"section\",\n \"origin_cites_number\": 30,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Taxonomy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Taxonomy\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 2261,\n 2262,\n 2263\n ],\n \"content\": \"\\\\label{Classic Models}\\nIn this section, we review the classic approaches that learn weakly supervised object localizer or detector without using deep features.\\nThese methods typically consist of one initialization module followed by one refinement process as shown in\\nFig.~\\\\ref{nondeep flowchart}.\\nIn , the detector is based on the deformable part model (DPM)~.\\nIn other approaches , the detector is based on the support vector machine (SVM) classifier.\\nThe features used by these approaches are hand-crafted feature descriptors, such as HOG in , SIFT in , and Lab color in , which are sometimes used to build higher level representation such as bag-of-words (BOW) in , ,,,, Fisher vector representation in , and subspace-based representation in .\\nIn the following, we divide these approaches for initialization and refinement process.\",\n \"id\": \"bba63d48-815b-4407-9de8-7df72823a004\",\n \"level\": \"section\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Classic Models\"\n ]\n ],\n \"subsections\": [\n \"bfae7607-5833-4424-b0e8-9e0a00e92605\",\n \"8f1d3bcb-1c87-4161-b6dc-8cbe17a169ff\",\n \"31104e17-5d59-42f1-88ae-903542c63cd2\",\n \"19588196-93c9-4245-8a54-e9fb29d5fd1e\"\n ],\n \"title\": \"Classic Models\"\n },\n {\n \"cite_extract_rate\": 0.40740740740740705,\n \"cites\": [\n 2261,\n 2262,\n 2267,\n 2263,\n 2269,\n 2270,\n 2266,\n 2264,\n 2265,\n 8537,\n 2268\n ],\n \"content\": \"Numerous methods have been developed to mine reliable instance samples, using prior knowledge, as weak supervision for the following processes.\\nA brief summary of these approaches are shown in Table \\\\ref{table initialization}.\\nZhang et al.~ leverage the prior-knowledge of object co-occurrence to identify translation and scale invariant high order features for weakly supervised object localization.\\nIn~ Shi et al. propose a transfer learning paradigm to first use a RankSVM to learn the mapping relationship between the box overlap and appearance similarity from an auxiliary training data (with bounding-box level annotation) and then transfer the learned prior-knowledge for localizing objects of interest in the given weakly labeled images.\\nA simple yet effective approach, named as negative mining, is developed by Siva et al. to explore the inter-class variance among the object regions in weakly labeled training images.\\nThe final object locations are obtained by\\nusing a linear combination of the inter-class variance and saliency prior.\\nSimilarly, Tang et al.~ and Xie et al. use the saliency prior and intra-class consistency to mine the initial object locations, respectively.\\nIn , Shi et al. explore the appearance prior and geometry prior in their topic model to build a Bayesian joint modeling framework for weakly supervised object localization.\\nOn the other hand, Cao et al.~ exploit the road map prior and density prior to mine the initial vehicle locations from the weakly labeled satellite images and then trained the vehicle detector under a modified multiple-instance learning (MIL) model.\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n\\\\begin{table*} [t]\\n \\\\tiny\\n\\t\\\\caption{Summary of the approaches using pre-trained feature representations, which is a subcategory in the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\\n \\\\label{table off-the-shelf feature}\\n \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{2cm}<{\\\\centering}m{1cm}<{\\\\centering}m{1.6cm}<{\\\\centering}m{2.8cm}<{\\\\centering}m{3.8cm}<{\\\\centering}m{1.2cm}<{\\\\centering}m{3.5cm}<{\\\\centering}m{2.2cm}<{\\\\centering}}\\n\\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} &\\n \\\\multirow{2}{*}{\\\\textbf{Detector}} &\\n \\\\multirow{2}{*}{\\\\textbf{Descriptor}} &\\n \\\\multirow{2}{*}{\\\\textbf{Prior knowledge} } &\\n \\\\multirow{2}{*}{\\\\textbf{Extra training data}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning model}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} &\\n \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline\\n\\\\hline\\n Gonthier-arxiv-2018 & None & CNN & Supervised objectness score (Fast R-CNN(Resnet)) & ImageNet(tag label) & SVM & MIL & Objects in art (watercolor2K, people-art) \\\\\\\\\\nZadrija-CVIU-2018 & None & VGG19 Conv 5\\\\_4. SIFT, Fisher vector & None & ImageNet(tag label) & Sparse model & None & Zebra crossings, traffic signs \\\\\\\\\\n Cinbis-TPAMI-2017 & SVM* & FV+CNN & Center prior & ImageNet(tag label) & SVM* & Multi-fold MIL & General objects \\\\\\\\\\nWei-IJCAI-2017 & None & CNN & None & ImageNet(tag label) & None & Deep descriptor transforming & General objects \\\\\\\\\\n Zhang-IJCAI-2016 & SVM & CNN & Saliency prior & ImageNet(tag label) & SVM & Easy-to-hard(SPL+CL) & General objects \\\\\\\\\\n Li-ECCV-2016 & None & FC6 & Strong detector prior (sparsity) & ImageNet(tag label) & SVM & Regularizing score distribution & General objects \\\\\\\\\\n Ren-TPAMI-2016 & SVM* & FC6 & None & ImageNet(tag label) & SVM* & MIL+bag splitting & General objects \\\\\\\\\\n Wan-ICIP-2016 & SVM* & FC7 & None & ImageNet(tag label) & SVM* & Correlation suppression+part suppression & General objects \\\\\\\\\\n Rochan-IVC-2016 & None & Color histogram +CNN & Saliency & PASCAL (edge box), ImageNet & SVM & None & General objects \\\\\\\\\\n Shi-ECCV-2016 & SVM & FC7(Alexnet) & Size prior & ImageNet(tag label), PASCAL2012 (object size) & SVM & Easy-to-hard(curriculum) & General objects \\\\\\\\\\n Wang-TIP-2015 & SVM & FC6 & None & ImageNet(tag label) & pLSA, SVM & Online latent category learning & General objects \\\\\\\\\\nRochan-CVPR-2015 & SVM & CNN & Objectness score, word embedding prior & YouTube-Objects (for parameter validation), Familiar object categories(detector) & SVM, Sparse reconstruction & Appearance transfer from text representation & General objects \\\\\\\\\\nBilen-CVPR-2015 & LSVM & FC7+spatial features & None & ImageNet(tag label) & LSVM & Convex clustering & General objects \\\\\\\\\\nWang-ICCV-2015 & None & FC6 & None & PASCAL(edge box), ImageNet & SVM* & Relaxed multiple-Instance SVM & General objects \\\\\\\\\\nZhou-ICMBD-2015 & SVM & FC7 & Saliency prior & ImageNet(tag label) & SVM & Negative Bootstrapping & Airplanes in remote sensing \\\\\\\\\\nHan-TGRS-2015 & SVM & DBM & Saliency, intra-class compactness, inter-class separability & None & DBM + SVM & Bayesian framework for initialization + refinement detector training & Objects in remote sensing \\\\\\\\\\nMathe-Arxiv-2014 & Sequential detector & FC6 & Human fixation & ImageNet(tag label) & MIL*+RL & Constrained multiple instance SVM learning + reinforcement learning of detector & Human actions \\\\\\\\\\n Wang-ECCV-2014 & SVM & FC6 & None & ImageNet(tag label) & pLSA, SVM & Online latent category learning & General objects \\\\\\\\\\n Bilen-BMVC-2014 & LSVM & DeCAF & None & ImageNet(tag label) & LSVM* & LSVM with posterior regularization on symmetry and mutual exclusion & General objects \\\\\\\\\\nSong-NIPS-2014 & DPM & FC7 & Objectness score & ImageNet(tag label) & LSVM & Frequent configuration mining+detector training & General objects \\\\\\\\\\n Song-ICML-2014 & LSVM & DeCAF & None & ImageNet(tag label) & Graph model+LSVM* & Initialization via discriminative submodular cover+smoothed LSVM learning & General objects \\\\\\\\\\n\\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\",\n \"id\": \"bfae7607-5833-4424-b0e8-9e0a00e92605\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 27,\n \"parent_id\": \"bba63d48-815b-4407-9de8-7df72823a004\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Classic Models\"\n ],\n [\n \"subsection\",\n \"Initialization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Initialization\"\n },\n {\n \"cite_extract_rate\": 0.28571428571428503,\n \"cites\": [\n 2272,\n 2271,\n 2274,\n 2273\n ],\n \"content\": \"After potential object instances are obtained, these hypotheses are verified in the following refinement processes.\\nThe goals of these approaches are to design learning objective functions, optimization strategies, or learning mechanisms to gradually determine objects of interest from the extracted initial instance training samples.\\nA brief summary of these approaches is shown in Table \\\\ref{table refinement}.\\n Hoai et al. propose an approach which localizes the instances of the positive class and learns a sub-window classifier to recognize the corresponding object class.\\nBlaschko et al. use a structured output SVM to learn a regressor from the weakly labeled training images to object locations that are parameterized by the coordinates of the bounding boxes.\\nThe object locations were treated as latent variables, while the image-level annotation was used to constrain the set of values the latent variable can take.\\nSimilarly, Pandey et al. learn weakly supervised object detectors by using DPMs with latent SVM training.\\nIn , a soft-label boosting approach is developed to exploit the soft labels that are estimated during the MIL process to train object detectors based on Boosting algorithm.\\nIn , Tang et al. treat the weakly supervised object localization problem as an object co-localization task, and present a joint image-box formulation to mine reliable object locations via a Boolean constrained quadratic program.\\nThis approach can handle noisy labels in the image-level annotations.\\nTo address the property that the MIL process may converge to poor local optima after the initialization, Cinbis et al. design a multi-fold MIL training paradigm.\\nThis method divides the whole weakly labelled training images into multiple folds and implements the detector training process and object re-localization process in different folds, thereby alleviating the issue with convergence of poor local optima.\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n\\\\begin{table*} [t]\\n \\\\tiny\\n\\t\\\\caption{Summary of the approaches using visual cues, which is a subcategory in the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\\n\\t \\\\label{table pre-learning} \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{2.5cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{2cm}<{\\\\centering}m{3cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{3.1cm}<{\\\\centering}m{1.8cm}<{\\\\centering}}\\n\\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} &\\n \\\\multirow{2}{*}{\\\\textbf{Detector}} &\\n \\\\multirow{2}{*}{\\\\textbf{Descriptor}} &\\n \\\\multirow{2}{*}{\\\\textbf{Prior knowledge} } &\\n \\\\multirow{2}{*}{\\\\textbf{Extra training data}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning model}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} &\\n \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline\\n\\\\hline\\n Li-ISPRS-2018 & CAM* (VGG-F) & CNN & None & ImageNet(tag label) & VGG-F* & Learning learning + CAM learning (patch level) & Remote sensing objects \\\\\\\\\\n Wilhelem-DICTA-2017 & None & CNN & None & ImageNEt(tage label) & CAM & CAM+KDE refine & General objects \\\\\\\\\\n Tang-TMM-2017 & DPM & CNN & Saliency + objectness score & ImageNet(tag label) & DPM+ CNN & Region initialization+DPM and feature learning+bounding box modification & General objects \\\\\\\\\\n Kolesnikov -BMVC-2016 & None & CNN & Human feedback annotation & ImageNet(tag label) & CAM & Active learning for identifying object cluster & General objects \\\\\\\\\\n Bency-ECCV-2016 & None & CNN & None & ImageNet(tag label) & VGG16 & Beam-search based on CNN classifier & General objects \\\\\\\\\\nZhou-MSSP-2016 & SVM & FC7 & Saliency prior & ImageNet(tag label), remote sensing data(unlabelled) & CNN (AlexNet), SVM & Deep feature transfer +MIL & Remote sensing objects (airplane, car, airport) \\\\\\\\\\nBergamo-WACV-2016 & SVM & CNN & None & ImageNEt(tage label) & CNN, SVM & Mask out initialization + SVM detector training & General objects \\\\\\\\\\nHoffman-CVPR-2015 & SVM & FC7 & Detector prior+ representation prior & ImageNet(tag label), ILSVRC13 validation subset(box annotation) & CNN, Latent SVM & Transferring detectors and representation from auxiliary data & General objects \\\\\\\\\\n\\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\",\n \"id\": \"8f1d3bcb-1c87-4161-b6dc-8cbe17a169ff\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"bba63d48-815b-4407-9de8-7df72823a004\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Classic Models\"\n ],\n [\n \"subsection\",\n \"Refinement\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Refinement\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"A number of iterative approaches have been developed that take both initialization and refinement into account.\\nIn , Siva et al. propose an intra-class metric and an inter-class metric to initialize the potential object locations.\\nAfter obtaining the initial object locations, this method iteratively trains a DPM object detector and uses a model drift detection approach to identify the termination refinement dynamically.\\nDeselaers et al. present a conditional random field (CRF) model, which is used to learn generic prior knowledge of the objects from meta-training data firstly to localize the potential objects of interest in the weakly labelled training images.\\nThis algorithm updates the CRF model to learn the appearance and shape models for the target object category and localizes the objects of interest in the refinement stage. \\nThe alternation of localization and learning processes progressively transforms the CRF model from class-generic prior knowledge into the specialized knowledge for a certain target class.\\nFor learning weakly supervised attribute localizer, Wang et al. initialize the learning process by building a Hierarchical Space Tiling (HST) scene configuration model and the corresponding appearance models are trained based on HST.\\nA joint inference and learning process is designed to update the scene attributes and the correlations between the scene parts and attributes gradually.\",\n \"id\": \"31104e17-5d59-42f1-88ae-903542c63cd2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"bba63d48-815b-4407-9de8-7df72823a004\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Classic Models\"\n ],\n [\n \"subsection\",\n \"Initialization and Refinement\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Initialization and Refinement\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Although the classic weakly supervised learning models are studied in early age, the two-stage learning frameworks, i.e., the learning initialization stage and refinement stage, built by these methods have been widely applied in future works. In the learning initialization stage, these methods provide two kinds of information cues to infer the candidate object regions. The first one is the bottom-up cues, including the region saliency, objectness, intra-class consistency, inter-class discriminability, et al. The other one is the top-down cues, which usually provide the appearance prior for the learning process. Notice that as such top-down cues are hard to obtain from the weakly labeled data, auxiliary training data (with instance-level manual annotation) are usually leveraged to explore the top-down cues which are then transferred to the weakly labeled target data. In the refinement stage, classic machine learning models, such as SVM and CRF, are adopted to gradually refine both the appearance model and locations of the objects of interest.\\nThe advantage of the classic weakly supervised learning methods is that the learning processes can be implemented on small-scale training data and the whole frameworks are quick both in the training phase and the testing phase. The disadvantage is that their performance is not satisfactory, which is due to the limitation in feature representation and model complexity.\",\n \"id\": \"19588196-93c9-4245-8a54-e9fb29d5fd1e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"bba63d48-815b-4407-9de8-7df72823a004\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Classic Models\"\n ],\n [\n \"subsection\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 1223,\n 514,\n 2272,\n 8429,\n 2275,\n 895,\n 2278,\n 2264,\n 2273,\n 2276,\n 2279,\n 2277\n ],\n \"content\": \"In this section, we review the approaches that learn weakly supervised object localizer or detector based on classic formulations and feature representations based on the deep neural networks, either pre-trained from the ImageNet dataset (with image tag annotation) or further fine-tuned on the weakly supervised training images in the target domain.\\nThe feature representations are based on the widely used deep models for image classification, such as AlexNet and VGG .\\nThe detectors are constructed based on classic formulations such as DPM and SVM~, or recent models such as RCNN and fast RCNN .\\nWe further divide these approaches into three subcategories using pre-trained deep features, inherent cues in deep models, and fine-tuned deep models as shown in Fig. \\\\ref{off-the-shelf-flowchart}.\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=8.5 cm]{subcategory.jpg}\\n \\\\caption{Illustration of different usages of the off-the-shelf deep neural networks by the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. }\\n \\\\label{off-the-shelf-flowchart}\\n\\\\end{figure}\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n\\\\begin{table*} [t]\\n \\\\tiny\\n\\t\\\\caption{Summary of the approaches with fine-tuned deep models, which is a subcategory in the weakly supervised object localization and detection approaches based on the off-the-shelf deep models. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. }\\n\\\\label{table prepost-learning}.\\n \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{2cm}<{\\\\centering}m{2cm}<{\\\\centering}m{1.6cm}<{\\\\centering}m{2.8cm}<{\\\\centering}m{2.5cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{3.7cm}<{\\\\centering}m{1.5cm}<{\\\\centering}}\\n\\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} &\\n \\\\multirow{2}{*}{\\\\textbf{Detector}} &\\n \\\\multirow{2}{*}{\\\\textbf{Descriptor}} &\\n \\\\multirow{2}{*}{\\\\textbf{Prior knowledge} } &\\n \\\\multirow{2}{*}{\\\\textbf{Extra training data}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning model}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} &\\n \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline\\n\\\\hline\\n Zhang-IJCV-2019 & Fast RCNN (VGG16) & Pre-trained FC7 & Tag number + mask out prior(AlexNet) & ImageNet(tage label) & SVM & Easy-to-hard & General objects \\\\\\\\\\n Uijlings-CVPR-2018 & None & CNN & Semantic objectness(SSD*) & ImageNet(tage label), ILSVRC(full annotation) & SSD*+SVM+ Fast RCNN & MIL+knowledge transfer & General objects \\\\\\\\\\n Jie-CVPR-2017 & Fast RCNN (VGG16) & CNN & Image-to-object transfer prior & ImageNet(tage label) & Fast RCNN (VGG16) & Initialization based on classification network and subgraph discovery + iterative Fast RCNN learning & General objects \\\\\\\\\\nShi-ICCV-2017 & Fast RCNN & CNN & Things and stuff prior & ImageNet(tage label), PASCAL Context (full annotation) & FCN,Fast RCNN & Localizing objects based on things and stuff prior and training Fast TCNN iteratively & General objects \\\\\\\\\\nSingh-CVPR-2016 & Fast RCNN & CNN & Tracking prior & ImageNEt(tage label), Youtube-objects (unlabelled) & Fast RCNN & Discriminative region minning+transferring tracking object pattern + learn object detector & General objects \\\\\\\\\\nLi-CVPR-2016 & VGG* & CNN & Mask out prior(AlexNet) & ImageNEt(tage label) & VGG*, SVM & Progressive Domain Adaptation & General objects \\\\\\\\\\nLiang-ICCV-2015 & CNN & CNN & Instance example, motion prior & ImageNEt(tage label) & CNN,R-CNN & Seed selection based on instance example and instance tracking & General objects \\\\\\\\\\nChen-ICCV-2015 & RCNN & CNN & Online data type & Web data (weak label) & BLVC net +E-LDA + RCNN & Simple image initialization + graph-based representation adaptation on hard image & General objects \\\\\\\\\\nZhou-CVPR-2015 & RCNN & FC7 & None & ImageNet(tag label) & SVM, R-CNN & .Max-margin visual concept discovery + Domain-specific detector selection & General objects\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\",\n \"id\": \"ce302070-cf4f-4d53-a26c-8a51743b26f9\",\n \"level\": \"section\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Off-the-shelf Deep Models\"\n ]\n ],\n \"subsections\": [\n \"72d7b986-53df-40a6-a211-40fcbb5e65ab\",\n \"a62482ff-4ed2-4ba2-b72d-45432f53b8f2\",\n \"eab65442-451f-481c-9954-9b19dd3d9a08\",\n \"0873c3d1-8648-437a-a297-c566cfb8420e\"\n ],\n \"title\": \"Off-the-shelf Deep Models\"\n },\n {\n \"cite_extract_rate\": 0.6136363636363631,\n \"cites\": [\n 2282,\n 2297,\n 2288,\n 2289,\n 2291,\n 2286,\n 2294,\n 7089,\n 2285,\n 2293,\n 2287,\n 2283,\n 2269,\n 2299,\n 2259,\n 2296,\n 2284,\n 2292,\n 2264,\n 2265,\n 737,\n 8537,\n 2295,\n 2298,\n 2281,\n 2290,\n 2280\n ],\n \"content\": \"The methods of this category replace the hand-crated feature representations with the pre-trained deep features typically from as AlexNet and VGG.\\nA brief summary of these approaches is shown in Table \\\\ref{table off-the-shelf feature}.\\nSong et al. determine discriminative feature configurations of an object class via graph modeling, and train object detectors within the multiple-instance learning paradigm.\\nThe deep features of this work are extracted based on the DeCAF scheme~ and AlexNet~.\\nBy using the deep features and spatial features to represent each proposal region, Bilen et al. propose a convex clustering process for learning the object models under the weak supervision.\\nThe learning objective is able to enforce the similarity among the selected proposal windows.\\nIn , Shi and Ferrari develop a curriculum learning strategy to feed training images into the MIL loop in a pre-defined order, where images containing larger objects are learned at the early stages while images containing smaller objects are learned at later stages.\\nRen et al. present a bag-splitting-based MIL mechanism that iteratively generated new negative bags from the positive ones.\\nThis algorithm can gradually reduce the ambiguity in positive images and thus facilitate the learning of more reliable training instance samples.\\nIn , Wei et al. leverage the pre-trained CNN model to implement a Deep Descriptor Transforming process, which can obtain the category-consistent image regions via evaluating the correlations of the descriptors in the convolutional activations of the CNN model.\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n\\\\begin{table*} [t]\\n \\\\tiny\\n \\t\\\\caption{A brief summary of the approaches using single-network training scheme, which is a subcategory in the weakly supervised object localization and detection approaches with deep weakly supervised learning algorithms. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\\n \\\\label{table endtoend}\\n \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{2.5cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{1cm}<{\\\\centering}m{1cm}<{\\\\centering}m{2.5cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{2.8cm}<{\\\\centering}m{1.8cm}<{\\\\centering}}\\n\\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} &\\n \\\\multirow{2}{*}{\\\\textbf{Detector}} &\\n \\\\multirow{2}{*}{\\\\textbf{Descriptor}} &\\n \\\\multirow{2}{*}{\\\\textbf{Prior knowledge} } &\\n \\\\multirow{2}{*}{\\\\textbf{Extra training data}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning model}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} &\\n \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline\\n\\\\hline\\nHuang-NIPS-2020&Faster RCNN &CNN &None &ImageNet(tag label) &Faster RCNN*(VGG16/ResNet50) &Proposal attention aggregation and distillation & General objects \\\\\\\\\\nShen-CVPR-2020&Faster RCNN* &CNN &None &ImageNet(tag label) + Flickr &Faster RCNN*(VGG16) &bagging-mixup + background noise decomposition + clean dat modelling& General objects \\\\\\\\\\nMai-CVPR-2020&None &CNN &None &ImageNet(tag label) &VGG/InceptionV3 &Integrating discriminative region mining and adversarial erasing & General objects \\\\\\\\\\nZhang-ECCV-2020 &None &CNN &Cross-image consistency &ImageNet(tag label) &VGG/InceptionV3/ ResNet50 &Inter-image stochastic consistency and global consistency & General objects \\\\\\\\\\nYang-WACV-2020&None &CNN &None &ImageNet(tag label) &VGG &Weighted classification activation map combination & General objects \\\\\\\\\\nYang-ICCV-2019 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Faster RCNN* (VGG16) + CAM & Online classifier learning with bounding box regression & General objects \\\\\\\\\\n Wan-CVPR-2019 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Faster RCNN* (VGG16) & Continuation MIL & General objects \\\\\\\\\\nShen-CVPR-2019 & WSDDN* (VGG16) & CNN & None & ImageNet(tag label) & Two-stream CNN (WSDDN+DeepLab) & Joint detection and segmentation with cyclic guidance & General objects \\\\\\\\\\nWan-CVPR-2019 & Fast RCNN & CNN & None & ImageNet(tag label) & Two-stream CNN & continuation instance selection and detector estimation & General objects \\\\\\\\\\nChoe-CVPR-2019 & None & CNN & None & ImageNet(tag label) & CNN & Feature learning by attention-based dropout & General objects \\\\\\\\\\nJiang-ICCV-2019 & None &CNN & None & ImageNet(tag label) & VGG16/Resnet101 & Online attention accumulation on CAM & General objects\\\\\\\\\\nSangineto-TPAMI-2018 & Fast RCNN (VGG16) & CNN & None & ImageNet(tag label) & Fast RCNN (VGG16) & Easy-to-hard & General objects \\\\\\\\\\nInoue-CVPR-2018 & SSD & CNN & None & PASCAL (full label as source domain),ImageNet & SSD & Domain transfer + pseudo-labeling & Cartoon objects \\\\\\\\\\n Wan-CVPR-2018 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Faster RCNN* (VGG16) & Min-entropy latent modeling & General objects \\\\\\\\\\nShen-TNNLS-2018 & VGG16* & CNN & None & ImageNet (tag label) & vgg16* & Object-specific pixel gradient mapping+Iterative component mining & General objects \\\\\\\\\\nTang-TPAMI-2018 & Fast RCNN* (model ensemble) & CNN & None & ImageNet(tag label) & Fast RCNN* (model ensemble) & MIL+oicr+multi-scale+proposal cluster learning & General objects \\\\\\\\\\n Zhang-CVPR-2018 & None & CNN & None & ImageNet(tag label) & VGG16* & Adversarial complementary erasing & General objects (for ILSVRC) \\\\\\\\\\n Choe-BMVC-2018 & ResNet & CNN & None & Tiny ImageNet(tag label) & ResNet & GoogLeNet Resize (GR) augmentation & General objects (for Tiny ImageNet) \\\\\\\\\\n Zhang-ECCV-2018 & Inception-v3+CAM & CNN & None & ImageNet(tag label) & Inception-v3+CAM & Self-produced guidance learning & General objects (for ILSVRC and CUB) \\\\\\\\\\n Gao-ECCV-2018 & Fast RCNN* & CNN & Count (human label) & ImageNet(tag label) & Fast RCNN*+Fast RCNN & WSL with count-based region selection & General objects \\\\\\\\\\n Singh-ICCV-2017 & CAM* (GoogLeNet)& CNN & None & ImageNet (tag label) & CAM* (GoogLeNet) & Random hidding patches & General objects (for ILSVRC) \\\\\\\\\\n Zhu-ICCV-2017 & None & CNN & None & ImageNet(tag label) & GoogLeNet* & Soft proposal layer+CAM & General objects \\\\\\\\\\nWan-ICIP-2017 & None & CNN & None & ImageNet(tag label) & CAM*(VGG) & CAM with spatial pyramid pooling layer & General objects \\\\\\\\\\n Durand-CVPR-2017 & None & CNN & None & ImageNet(tag label) & CAM* (ResNet101) & CAM with multi-map transfer layer & General objects \\\\\\\\\\nTang-CVPR-2017 & Fast RCNN* (model ensemble) & CNN & None & ImageNet(tag label) & Fast RCNN*+Fast RCNN & MIL+oicr+multi-scale & General objects \\\\\\\\\\n Jiang-CVPR-2017 & Fast RCNN* (AlexNEt) & CNN & None & PASCAL (edge box),ImageNet & AlexNet+ ROIpool & Region calssification+region selection+multi-scale & General objects \\\\\\\\\\nDiba-CVPR-2017 & Faster RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Multi-stream CNN & Cascading LocNet (CAM), SegNet, and MILNet+multi-scale & General objects \\\\\\\\\\nSelvaraju-ICCV-2017 & None & CNN & None & ImageNet(tag label) & VGG* & Gradient-based class activation mapping & General objects \\\\\\\\\\nTang-PR-2017 & None & CNN & None & ImageNet(tag label) & Fast RCNN* (VGG-16) & SPP with discovery block and classification block & General objects \\\\\\\\\\nGudi-BMVC-2017 & None & CNN & None & ImageNet(tag label) & CAM* (VGG-16) & CAM with Spatial Pyramid Averaged Max (SPAM) Pooling & General objects\\\\\\\\\\n Bilen-CVPR-2016 & Fast RCNN* (model ensemble) & CNN & None & PASCAL (edge box),ImageNet & Fast RCNN* & MIL+multi scale & General objects \\\\\\\\\\n Kantorov-ECCV-2016 & Fast RCNN* (VGG-F) & CNN & Context & ImageNet(tag label) & Fast RCNN* & MIL+multi-scale & General objects \\\\\\\\\\n Teh-BMVC-2016 & CNN & CNN & None & PASCAL (edge box),ImageNet & CNN & Proposal attention learning & General objects\\\\\\\\\\n Durand-CVPR-2016 & None & CNN & None & ImageNet(tag label) & CNN & Feature extraction network+weakly supervised prediction module & General objects \\\\\\\\\\n Zhou-CVPR-2016 & None & CNN & None & ImageNet(tag label) & GoogLeNet* & Class activation mapping & General objects\\\\\\\\\\n Oquab-CVPR-2015 & None & CNN & None & ImageNet(tag label) & CNN & Fully convolutional CNN with global max pooling & General objects \\\\\\\\\\n Wu-CVPR-2015 & None & CNN & None & PASCAL (BING),ImageNet & CNN & Deep multiple instance learning network & General objects\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\",\n \"id\": \"72d7b986-53df-40a6-a211-40fcbb5e65ab\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 44,\n \"parent_id\": \"ce302070-cf4f-4d53-a26c-8a51743b26f9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Off-the-shelf Deep Models\"\n ],\n [\n \"subsection\",\n \"Pre-trained Deep Features\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pre-trained Deep Features\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 2271,\n 2273,\n 2274,\n 2272\n ],\n \"content\": \"Instead of using the pre-trained deep models as feature extractor, the methods of this category obtain useful information cues (such as the activations in the intermediate network layers and the semantic scores in the output network layer) from the pre-trained deep neural networks to facilitate the weakly supervised learning process.\\nThe focus of these approaches mainly lies in the initialization stage of the weakly supervised learning process.\\nA brief summary of these approaches is shown in Table \\\\ref{table pre-learning}.\\nBergamo et al. propose a self-taught deep learning approach for localizing objects of interest under weak supervision.\\nIn the initialization stage, they design a mask-out strategy based on the deep semantic cues from a pre-trained classification network.\\nSpecifically, this method first calculates the degeneration of the image-level classification scores when masking out a certain object proposal region and then selects those with large differences as the interested object regions.\\nAfter the initialization stage, this method trains an SVM-based object detector in the subsequent refinement stage for final object localization.\\nSimilar to , Bency et al. propose a beam search algorithm to leverage the activation maps of a pre-trained classification network to localize the objects of interest.\\nThis method is based on the observation that when image regions centered around objects of interest are classified by a pre-trained DNN, they obtain higher semantic scores than other image regions.\\nHoffman et al. develop a transfer learning-based algorithm, where the deep neural network is first trained on both the weakly labeled auxiliary training data and the strongly labeled training data to obtain the background detector.\\nThen, an SVM-MIL model is adopted to learn the object detectors based on the potential foreground regions that are obtained by using the pre-trained DNN to screen the image background regions.\\nTo overcome the problem that the objects of interest would sometimes co-occur with the distracting image background, Kolesnikov et al. {propose a user-guided weakly supervised learning framework to improve the localization capacity.}\\nThis approach first trains a classification network under the image level annotation.\\nFor each image, the intermediate feature maps of the pre-trained network are clustered, and the object clutter is identified by a user.\",\n \"id\": \"a62482ff-4ed2-4ba2-b72d-45432f53b8f2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"ce302070-cf4f-4d53-a26c-8a51743b26f9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Off-the-shelf Deep Models\"\n ],\n [\n \"subsection\",\n \"Inherent Cues in Deep Models\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Inherent Cues in Deep Models\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 8538,\n 2306,\n 2303,\n 2301,\n 802,\n 2302,\n 2309,\n 2300,\n 2304,\n 2308,\n 2276,\n 2307,\n 2279,\n 2310,\n 2305,\n 2277\n ],\n \"content\": \"The methods of this category fine-tune the off-the-shelf DNN models during the weakly supervised learning process to obtain strong object detectors .\\nA brief summary of these approaches is shown in Table \\\\ref{table prepost-learning}.\\nChen et al. first train CNNs from the web image data via an easy-to-hard learning scheme.\\nThe learned deep features are used to mine object locations by using the exemplar-LDA detector .\\nThe off-the-shelf RCNN detector is then adopted to learn object models based on their localization results.\\nIn , Li et al. first use the mask-out strategy based on the pre-trained classification network to obtain the class-specific object proposals.\\nAn SVM-based MIL process is used to localize object instances and\\nthe classification network is further fine-tuned on the localized object instances for better performance.\\nShi et al. propose to transfer the prior knowledge of \\\\textit{things} and \\\\textit{stuff} to help the weakly supervised learning process.\\nA semantic segmentation network is trained from the source set (with available bounding-box annotations) to generate the stuff map and thing map for the weakly labeled images in the target set.\\nThese maps are used to obtain potential object locations, and the fast RCNN model is adopted in a Deep MIL scheme to train the object detectors.\\nRecently, revisits knowledge transfer for training the weakly supervised object detector.\\nIn this method, a DNN-based proposal extractor is learned from the source data firstly.\\nThe DNN is designed based on the SSD architecture and trained with a semantic hierarchy.\\nThe network is then used to provide proposals and other prior knowledge for the weakly labeled images in the target set.\\nAn MIL process is used to determine the proposals that cover the objects of interest\\nbased on which the fast RCNN model is adopted to learn the final object detectors.\\nIn , Zhang et al. first learn to localize the objects of interest via a collaborative self-paced curriculum learning mechanism based on pre-trained deep features.\\nThe fast RCNN model is applied to learn object detectors.\\n\\\\renewcommand{\\\\multirowsetup}{\\\\centering}\\n\\\\begin{table*} [t]\\n \\\\tiny\\n\\t\\\\caption{A brief summary of the approaches with multi-network training, which is a subcategory in the weakly supervised object localization and detection approaches with deep weakly supervised learning algorithms. * indicates a certain variation of the corresponding model. An approach is considered for general object category when it is tested for detecting more than five object categories in the corresponding literature. The approaches with None detector indicate the weakly supervised object localization approaches.}\\n\\t\\\\label{table multinet} \\\\resizebox{\\\\linewidth}{!}{\\\\begin{tabular}{m{2cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{1.5cm}<{\\\\centering}m{1.8cm}<{\\\\centering}m{3cm}<{\\\\centering}m{2.6cm}<{\\\\centering}m{3cm}<{\\\\centering}m{1.5cm}<{\\\\centering}}\\n\\\\hline\\n \\\\multirow{2}{*}{\\\\textbf{Methods}} &\\n \\\\multirow{2}{*}{\\\\textbf{Detector}} &\\n \\\\multirow{2}{*}{\\\\textbf{Descriptor}} &\\n \\\\multirow{2}{*}{\\\\textbf{Prior knowledge} } &\\n \\\\multirow{2}{*}{\\\\textbf{Extra training data}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning model}} &\\n \\\\multirow{2}{*}{\\\\textbf{Learning strategy}} &\\n \\\\multirow{2}{*}{\\\\textbf{Object category}} \\\\\\\\\\n & & & & & & &\\\\\\\\\\n\\\\hline\\n\\\\hline\\nZhang-CVPR-2020 &None &CNN &Common object co-localization &ImageNet(tag label) &VGG/InceptionV3/ResNet50/ DenseNet161 &Classification + pseudo supervised object localization & General objects \\\\\\\\\\nZhong-ECCV-2020 &Faster RCNN &CNN &Location prior &ImageNet(tag label) + COCO (box label) &Oneclass universal detector + MIL classifier (on ResNet50) &Progressive knowledge transfer & General objects \\\\\\\\\\nKosugi-ICCV-2019 & Fast RCNN* &CNN & Mask-out prior & ImageNet(tag label) & Mask-out net + OICR* & Mask-our prior-guided label refinement & General objects \\\\\\\\\\nSingh-CVPR-2019 & Fast RCNN* & CNN & Motion prior & ImageNet(tag label), videos & RPN+WSDDN/OICR (VGG16) & Training RPN using weakly-labeled videos for WSOD & General objects \\\\\\\\\\nArun-CVPR-2019 & Fast RCNN & CNN & None & ImageNet(tag label) & Fast RCNN (VGG16) + Fast RCNN* (VGG16) & Employ dissimilarity coefficient for modeling uncertainty & General objects \\\\\\\\\\nLi-TPAMI-2019 & Faster RCNN* (VGG16) & CNN & Objectness (classifier) prior & ImageNet(tag label), ILSVRC2013(box label for unseen categories) & Faster RCNN* (VGG16) *2 & Objectness transfer+MIL +multi-scale & General objects \\\\\\\\\\n Zhang-ECCV-2018 & fast RCNN* (VGG16) & CNN & None & PASCAL (edge box), ImageNet(tag label) & Multi-view WSDDN+multi view Fast RCNN & Two phase multi-view learning & General objects \\\\\\\\\\nShen-CVPR-2018 & SSD & CNN & None & ImageNet(tag label) & SSD+Fast RCNN* & MIL+GAN & General objects \\\\\\\\\\n Zhang-CVPR-2018 & Faster RCNN (VGG16) & CNN & None & ImageNet(tag label) & MIDN+Faster RCNN & WSOD+Pseudo Ground-truth Mining+FOD & General objects \\\\\\\\\\nZhang-CVPR-2018 & Fast RCNN* (VGG16) & CNN & None & PASCAL (edge box), ImageNet & WSDDN + Fast RCNN* & WSDDN+easy-to-hard FOD & General objects \\\\\\\\\\nTang-ECCV-2018 & Fast RCNN* (VGG16) & CNN & None & ImageNet(tag label) & Fast RCNN* (VGG16) & Alternating training of WSRPN and WSOD+multi-scale & General objects \\\\\\\\\\nTao-TMM-2018 & Fast RCNN* (VGG16) & CNN & Web image & Web dataset(weak label),imageNet(tag label) & Midn & Easy-to-hard & General objects \\\\\\\\\\n Wang-IJCAI-2018 & Faster RCNN (VGG16) & CNN & Model consistency & ImageNet(tag label) & Faster RCNN+Fast RCNN* & MIL+collaborative learning & General objects \\\\\\\\\\nWei-ECCV-2018 & Faster RCNN* (VGG16) & CNN & Shape prior+ context prior & ImageNet(tag label) & MIDN+CAM+ Deeplab & Tight Box Mining+MIL +OICR+multi-scale & General objects \\\\\\\\\\n Ge-CVPR-2018 & Faster RCNN (VGG16) & CNN & Local objectness and global attention & ImageNet(tag label) & MIDN,TripNet, GoogleNet, FCN, Fast RCNN & Multi evidence fusion+ outlier filtering +pixel label preidction +box generation+multi-scale & General objects \\\\\\\\\\nDong-MM-2017 & Fast RCNN* & CNN & None & ImageNet(tag label) & Fast RCNN*+R-FCN & Easy-to-hard & General objects \\\\\\\\\\n Li-BMVC-2017 & Fast RCNN* & CNN & Shape prior & ImageNet(tag label) & Fast RCNN* +CAM+DeepLab & Easy-to-hard & General objects \\\\\\\\\\nWang-CVPR-2017 & CNN & CNN & None & ImageNet(tag label) & CAM* & Image-level training+pixel-level fine tuning & Salient objects \\\\\\\\\\n Sun-CVPR-2016 & None & CNN & None & ImageNet(tag label) & Multi-scale FCN+ CNN(vgg16) & Cascade localization and recognition & General objects \\\\\\\\\\n Zhang-TGRS-2016 & CNN & CNN & None & ImageNet(tag label), auxiliary data(image label) & CPRNet+LocNet & Alternative training CPRNet and LocNet & Aircraft \\\\\\\\\\n\\\\hline\\n \\\\end{tabular}}\\n\\\\end{table*}\",\n \"id\": \"eab65442-451f-481c-9954-9b19dd3d9a08\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 28,\n \"parent_id\": \"ce302070-cf4f-4d53-a26c-8a51743b26f9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Off-the-shelf Deep Models\"\n ],\n [\n \"subsection\",\n \"Fine-tuned Deep Models\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Fine-tuned Deep Models\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Introducing the off-the-shelf deep models into the weakly supervised object localization or detection framework is the most straightforward approach to integrate deep learning and weakly supervised learning.\\nThe methods in this category show that: 1) feature learning is an important factor to improve the weakly supervised learning process; 2) DCNN models can infer discriminative spatial locations when learned under the image-level supervision; 3) pre-training DNN models on large-scale auxiliary training data is a simple but effective way to encode useful cues for the weakly supervised learning process.\\nCompared with the classic models,\\nthe methods of this category exploit large-scale auxiliary training data to learn powerful feature representations and top-down cues.\\nBy using DNN models as the object detector or localizer, a significant performance gain can be obtained.\\nHowever, more effective feature learning models can be exploited in the weakly supervised learning process.\",\n \"id\": \"0873c3d1-8648-437a-a297-c566cfb8420e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"ce302070-cf4f-4d53-a26c-8a51743b26f9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Off-the-shelf Deep Models\"\n ],\n [\n \"subsection\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 0.529411764705882,\n \"cites\": [\n 2282,\n 2291,\n 7089,\n 2300,\n 2312,\n 2284,\n 2304,\n 2311,\n 2303\n ],\n \"content\": \"In this section, we review the methods that learn weakly supervised object localizers or detectors by designing novel deep weakly supervised learning frameworks\\\\footnote{Notice that here we mainly indicate that the core learning blocks used in the learning frameworks are based on deep learning, while some minor computational components in pre-processing or post-processing, such as proposal extraction and bounding-box modification, et al., are not necessarily be implemented by deep learning.}.\\nDifferent from the approaches discussed in previous sections, both the feature representations and the object detectors of the approaches in this category are learned by newly-designed deep neural networks.\\nThe whole weakly supervised learning framework may be designed in a compact network model, such as in , or contain several function-distinct DNN components, such as in .\\nWe categorize these approaches into two groups using single-network training and multi-network training, respectively.\",\n \"id\": \"930ae35a-10a8-4a65-83e8-a12f89f6fe36\",\n \"level\": \"section\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deep Weakly Supervised Learning\"\n ]\n ],\n \"subsections\": [\n \"352d0dfb-6a12-4bf2-b2bf-5544ab013008\",\n \"65da26f9-ce63-494b-9e38-4a72e74d3aa0\",\n \"c9fa8b67-133f-4688-92de-bc8e20c858b9\"\n ],\n \"title\": \"Deep Weakly Supervised Learning\"\n },\n {\n \"cite_extract_rate\": 0.777777777777777,\n \"cites\": [\n 2259,\n 2282,\n 2298,\n 2290,\n 2314,\n 2280,\n 2315,\n 737,\n 107,\n 7090,\n 2286,\n 687,\n 2313,\n 2295\n ],\n \"content\": \"The methods of this category are designed with a single deep neural network using the training images (or together with the extracted object proposals) as inputs and the image-level classification labels as the outputs.\\nThese approaches do not usually rely on meticulously designed initialization processes to obtain the potential object regions.\\nInstead, these methods discover the interested object regions solely based on the end-to-end learning process of the designed DNN models.\\nThe DNN models used in these approaches usually have similar feature learning layers as the conventional image classification network, e.g., AlexNet, VGG, GoogleNet, and ResNet, followed by the instance label inferring and image label propagation layers to inherently predict the labels of each proposal region and generate the final image labels from the predicted proposal labels, respectively. Some of the methods contain multiple network streams for online inferring multiple informative cues.\\nA brief summary of these approaches are\\nshown in Table \\\\ref{table endtoend}.\\nIn the DNN models proposed by Wu et al. and Bilen et al. , the network inputs are the training images and extracted object proposal regions while the outputs are the image-level semantic scores.\\nThe first parts of these networks extract the features and infer the labels for each proposal region, and the second parts propagate the proposal scores to the image-level via the max pooling scheme or the two-stream score regularization method .\\nZhou et al. present an end-to-end weakly supervised deep learning based on the class activation mapping (CAM).\\nThe weights of the feature maps in the intermediate layers are inferred based on the correspondence between the feature map and a certain object category.\\nThe feature maps are then combined to form the class activation maps based on the inferred weights, which highlight the locations of the objects of interest. {Notably, this method is determined to be highly efficient in recent work . }\\nBuilt on CAM , Durand et al. introduce the multi-map transfer layer and the WILDCAT pooling layer to facilitate the more accurate deep MIL process.\\nRecently, a number of two-branch MIL models have been developed in which one is based on a typical deep network and the other one is introduced for weakly supervised learning.\\nBased on the WSDDN , Tang et al. integrate MIL branch and the instance classifier refinement branch into a unified deep learning framework such that more accurate online instance classifier learning is realized under the weak supervision.\\nIn , Diba et al. propose a weakly supervised cascaded convolutional network, which contains three branches.\\nThe first branch adopts the CAM module to generate the class activation maps. \\nThe second branch uses the generated class activation maps as the supervision signal to learn a segmentation module to generate the segmentation masks of the objects of interest.\\nUsing the candidate object proposals selected based on the obtained segmentation masks as supervision, the third network branch uses a MIL process to mine accurate object locations from the candidate object regions.\\nIn , Zhang et al. present a CAM-based network architecture which contains a classification branch and a counterpart classifier branch for object localization.\\nSpecifically, the classification branch is used to localize the discriminative object regions, which drives the counterpart classifier branch to discover new and complementary object regions by erasing its discovered regions from the feature maps.\\nWan et al. propose a min-entropy latent model (MELM) for weakly supervised object detection based on the assumption that minimizing entropy results in minimum randomness of a system.\\nThe network architecture is similar to , but global min-entropy and local min-entropy losses are introduced to train a DNN model to select the proposal cliques of largest object probability and mine truthful object locations from the selected proposal cliques.\\nZhang et al. develop a self-generated guidance method for weakly supervised object localization.\\nIn this work, a self-generated guidance map is derived from a CAM {layer to help learning features and object location masks from the previous network layers.}\\n {More recently, Gao et al. propose a token semantic coupled attention mapping for WSOL, which models the long-range visual dependency of the image regions and thus avoid partial activation. Ren et al. introduce the instance-associative spatial diversification constraints and build the parametric spatial dropout block to address the instance ambiguity and incomplete localization problems. Besides, they additionally adopt a sequential batch back-propagation algorithm, which enables their model to use a large ResNet as the backbone\\\\footnote{{As discussed in , it is non-trivial to introduce the deeper network backbones, e.g, the deep residual network, into the weakly supervised object detection frameworks as it would encounter dramatic deterioration of detection accuracy and training non-convergence. }}. Although there are other methods that use ResNet as the backbone , there is very limited exploration of using more recent backbone architectures, e.g., DesNet and Res2net , in both WSOL and WSOD frameworks. }\",\n \"id\": \"352d0dfb-6a12-4bf2-b2bf-5544ab013008\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"930ae35a-10a8-4a65-83e8-a12f89f6fe36\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deep Weakly Supervised Learning\"\n ],\n [\n \"subsection\",\n \"Single-Network Training\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Single-Network Training\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 2259,\n 2298,\n 2306,\n 737,\n 2301,\n 2304,\n 2310,\n 209,\n 8429,\n 2309\n ],\n \"content\": \"The methods in this school collaborate multiple networks, either in one training stage or in multiple training stages, to accomplish the weakly supervised object localization or detection task.\\nThe approaches of this category usually train a network to mine the initial object regions and another network for the detection task under the MIL framework .\\nAn additional object detection network, e.g., Fast RCNN, may also be used to train the final object detectors .\\nBy integrating multiple networks, these methods tend to achieve better performance both in object localization and detection.\\nA brief summary of these approaches is shown in Table \\\\ref{table multinet}.\\nLi et al. propose a multiple instance curriculum learning method, where a network based on the WSDDN model is used to mine candidate object proposals and another one based on the CAM algorithm to generate saliency maps from the selected proposals.\\nA curriculum is designed to select confident training examples based on the consistency between the regions outputted by the two networks.\\nThe object detectors are then trained by using the confident training examples iteratively.\\nDong et al. present a dual-network progressive approach for weakly supervised object detection, where a positive instance selection network and a region refinement network are adopted to minimize the classification error and modify object localization, respectively.\\nThese two networks are worked under a co-training paradigm.\\nIn , Ge et al. first obtain intermediate object localization and pixel labeling results using a classification network.\\nA triplet-loss network and an instance classification network are then constructed to detect outlier and filter object instances.\\nFinally, the filtered object instances are used as the supervision to train another Fast RCNN-based detection network.\\n {In order to overcome the limitations brought by the imprecise of the extracted object proposals,} Wei et al. propose to mine object proposals with tight boxes to learn weakly supervised object detector. {The assumption is that the proposals with tight boxes are more likely to contain the objects of interest thus mining such kind of proposals would help screen the cluttered background regions. In their approach, }\\na semantic segmentation network is first learned using the object localization map generated by CAM as the pseudo ground-truth.\\nThe predicted segmentation masks are used to mine object proposals with tight boxes, and fed into the online instance classifier refinement (OICR) network to learn weakly supervised object detector.\\n {With the same motivation as , Tang et al. propose to combine a two-stage region proposal network with an OICR network to learn the weakly supervised object detectors. Instead of mining proposals based on the semantic segmentation cue, Tang et al. use both the low-level cues (feature maps in shallow layers) and the high-level cues (semantic scores in deep layers) to mine reliable proposals in the two stages, respectively. The parameters of the region proposal network are obtained based on the network trained by .}\\n {Zhang et al. explore the reliability of each training image by evaluating the image difficulty and then feed the images into the learning procedure in an easy-to-hard order. Specifically, the image difficulty is evaluated by diagnosing the localization outputs of the pre-trained WSDDN model based on the concentrateness of the high-scored proposal locations. }\\nIn , Zhang et al. use three networks to learn weakly supervised object detectors.\\nFirst, an OICR network is trained to generate the initial object regions.\\nThen, they train an RPN based on {the pseudo ground-truth boxes obtained after implementing a post-process on the initial object regions, and use the learned RPN} to generate more accurate object locations.\\nFinally, fully supervised object detectors are trained based on the obtained object locations.\",\n \"id\": \"65da26f9-ce63-494b-9e38-4a72e74d3aa0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"930ae35a-10a8-4a65-83e8-a12f89f6fe36\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deep Weakly Supervised Learning\"\n ],\n [\n \"subsection\",\n \"Multi-Network Training \"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-Network Training \"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Compared with the off-the-shelf deep model-based weakly supervised object detection and localization methods, the deep weakly supervised learning methods exploits the merits of deep learning and weakly supervised learning approaches.\\nWithout complex design in the learning initialization stage, the end-to-end deep weakly supervised learning methods have been shown to perform well by introducing the MIL mechanism into the network design of the DCNN models.\\nWhile the multi-network training methods can further improve the learning performance by combining multiple function-specific networks.\\nOn the other hand, the performance of these methods is limited by whether the information extracted from the weakly-supervised module is effective or not.\\nAs such, prior knowledge may be useful to guide the deep learning process without solely relying on the weakly-supervised network.\",\n \"id\": \"c9fa8b67-133f-4688-92de-bc8e20c858b9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"930ae35a-10a8-4a65-83e8-a12f89f6fe36\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deep Weakly Supervised Learning\"\n ],\n [\n \"subsection\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 0.14285714285714202,\n \"cites\": [\n 2286\n ],\n \"content\": \"During the last decades, significant efforts have been made to develop various methods for learning weakly supervised object localizer or detector.\\nFor fair performance evaluation, it is of great importance to introduce some publicly available benchmark datasets and evaluation metrics.\\n\\\\begin{table}[t]\\n \\\\centering\\n \\\\caption{ {Brief summarization of the characteristics of the datasets. The top three are the datasets usually used for the WSOD task, while the bottom two are the common datasets for the WSOL task. \\\\#Categories indicates the number of object categories. \\\\#Images indicates the number of images. ``GO'' is short for Generic Objects. }}\\n \\\\begin{tabular}{ccccc}\\n \\\\hline\\n Dataset & \\\\multicolumn{1}{c}{Content} & \\\\multicolumn{1}{c}{\\\\#Categories} & \\\\multicolumn{1}{c}{\\\\#Images} & \\\\multicolumn{1}{c}{Metrics} \\\\\\\\\\n \\\\hline\\\\hline\\n PASCAL VOC 07 & \\\\multirow{3}[2]{*}{GO} & \\\\multirow{3}[2]{*}{20} & 9,962 & \\\\multirow{3}[2]{*}{mAP, CorLoc} \\\\\\\\\\n PASCAL VOC 10 & & & 21,738 & \\\\\\\\\\n PASCAL VOC 12 & & & 22,531 & \\\\\\\\\\n \\\\hline\\n CUB-200-2011 & Birds & 200 & 11,788 & Top-1/5 Loc \\\\\\\\\\n ILSVRC 2016 & GO & 1000 & 1.2 M & GT Loc \\\\\\\\\\n \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:dataset}\\n\\\\end{table}\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=9.5 cm]{dataset.jpg}\\n \\\\caption{ {Illustration of examples from the PASCAL VOC (top block), CUB-200-2011 (bottom-left block), and ILSVRC 2016 (bottom-right block) datasets.}}\\n \\\\label{dataset}\\n\\\\end{figure}\\nExisting weakly supervised object detection methods are usually evaluated on the PASCAL VOC datasets, including the PASCAL VOC 2007, 2010, and 2012 sets.\\nThe PASCAL VOC 2007 , PASCAL VOC 2010 and PASCAL VOC 2012 contain 9,962, 21,738, and 22,531 images of 20 object classes.\\nThese three datasets are divided into train, val, and test sets, where the trainval set (5,011 images for PASCAL VOC 2007, 10,869 images for PASCAL VOC 2010, and 11,540 images for PASCAL VOC 2012) are used to train the weakly supervised object detector, and the rest for evaluation.\\nThe mean of AP (mAP) metric is used to measure the performance where one object is successfully detected if the intersection over union (IoU) between the ground-truth and predicted boxes is more than 50 percentage.\\nThe weakly supervised object localization performance is usually evaluated on the PASCAL VOC, ILSVRC, and CUB datasets.\\nOn the PASCAL VOC datasets , weakly supervised object localization methods only use the trainval sets, which are different from the setting in weakly supervised object detection.\\nThat is, both the weakly supervised learning process and localization process are implemented on the same image data.\\nTo evaluate the localization performance on the PASCAL VOC datasets, the correct localization metric (CorLoc) is adopted, where the bounding-box with the highest class-specific score from each image is examined to be whether correct (with more than 50\\\\% overlap with the ground-truth box) or not.\\nIn addition to the PASCAL VOC datasets, the ILSVRC 2016 dataset \\\\cite {russakovsky2015imagenet} (i.e., the ImageNet) and CUB-200-2011 dataset are also widely used for performance evaluation.\\nThe ILSVRC 2016 dataset contains more than 1.2 million images of 1,000 classes for training, while the validation set, which contains 50,000 images, is used for testing.\\nThe CUB-200-2011 dataset contains 11,788 images of 200 categories with 5,994 images for training and 5,794 for testing.\\nFor these two datasets, {The commonly-used evaluation metrics are GT-known localization accuracy (GT Loc), Top-1 localization accuracy (Top-1 Loc), and Top-5 localization accuracy (Top-5 Loc). Specifically, GT Loc judges the localization results as correct when the intersection over union (IoU) between the ground-truth bounding box and the estimated box is no less than 50\\\\%, while Top-1 Loc considers the localization results as correct when the class predicted with the highest score is equal to the ground-truth class and the estimated bounding box has no less than 50\\\\% IoU with the ground-truth bounding box . Top-5 Loc differs from Top-1 Loc in that it checks if the target label is one of the top 5 predictions. As can be seen, the Top-1 Loc is a harder metric than the GT Loc as it needs to additionally predict the class label correctly. This will dramatically increase the task difficulty when performing under very large or fine-grained semantic spaces. The difficulty of Top-5 Loc is between Top-1 Loc and GT Loc as it requires the model to predict the class label but does not restrict the prediction to be perfectly correct.}\\n {We provide a brief summarization of the characteristics of the aforementioned datasets in Table \\\\ref{tab:dataset}. We additionally show some examples from each dataset to illustrate the bias of image content in different datasets. As displayed in Fig. \\\\ref{dataset}, the PASCAL datasets contain relatively more complex image content, where multiple object instances and categories may appear in a single image and different images would contain objects with significant scale variations. Although it only contains 20 object categories, its category diversity is higher than the CUB-200-2011 dataset as all the 200 categories in the CUB-200-2011 dataset are related to birds. ILSVRC 2016 contains far more images and categories than the PASCAL datasets. However, the content of the images from the ILSVRC 2016 dataset tends to be simpler than that of the PASCAL datasets---each of the images from the ILSVRC 2016 dataset typically contains only one single object and the objects have more consistent sizes and are placed in clearer background context relative to those from the PASCAL datasets. }\\\\\\\\\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=9.5 cm]{videoapp.jpg}\\n \\\\caption{Weakly supervised object localization or detection methods for video understanding. The examples are from . }\\n \\\\label{videoap}\\n\\\\end{figure}\",\n \"id\": \"79d342a2-5acc-4290-aaff-2761ce99e494\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Datasets and Evaluation Metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets and Evaluation Metrics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In recent years, weakly supervised object localization and detection techniques have been used in numerous vision problems, especially when it is difficult to collect ground-truth labels.\",\n \"id\": \"5cef946e-16aa-4550-aeb5-c885a8308ec2\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Applications\"\n ]\n ],\n \"subsections\": [\n \"2093fbc7-1f16-499f-9309-a24b9da6de5b\",\n \"5bcd51f6-c4c7-4e80-b0c3-9cec6a4e37d4\",\n \"6c796257-03d1-4c11-b23e-711e3eea5646\",\n \"ec87da5b-525f-4f14-a7b7-8c787fb56bb5\"\n ],\n \"title\": \"Applications\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 2319,\n 2317,\n 2316,\n 2318,\n 8539\n ],\n \"content\": \"As it is time-consuming to obtain object-level annotations for each video frame, weakly supervised object localization and detection methods have also been applied in the field of video understanding (see Fig. \\\\ref{videoap}).\\nFor example, Chanda et al. build a two-stream learning framework, which adapts the information from the labeled images (source domain) to the weakly labeled videos (target domain).\\nIn Zhang et al. propose a self-paced fine-tuning network for learning two network heads to localize and segment the object of interest from the weakly labeled training videos.\\nThe network is equipped with the multi-task self-paced learning function which can integrate confident knowledge from each single task (localization or segmentation) and use it to build stronger deep feature representation for both tasks.\\nOn the other hand, develop methods to localize temporal actions in the given untrimmed videos, where the main goal is to predict the temporal boundary of each action instance contained in the weakly labeled training videos.\\n {Essentially, such a weakly supervised action localization (WSAL) task is an emerging, yet rapidly developing topic in recent years, and the methods for solving this task are highly related to the weakly supervised object detection and localization methods. The additional challenges are: (i) the duration of the interesting action has very large variation, i.e., from a few seconds to thousands of seconds; and (ii) the features extracted to represent the interesting action would be entangled with those of the complex scenes of the video frame. }\\n {Notice that when applying to video understanding, there are strong correlations among adjacent video frames. So, additional informative constraints can be introduced to facilitate the weakly supervised object detection or localization under this scenario. }\",\n \"id\": \"2093fbc7-1f16-499f-9309-a24b9da6de5b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"5cef946e-16aa-4550-aeb5-c885a8308ec2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Video Understanding\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Video Understanding\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 2267,\n 2299\n ],\n \"content\": \"One interesting application of the weakly supervised object localization and detection techniques is the analysis of art images (see Fig. \\\\ref{art}).\\nInoue et al. propose a cross-domain weakly-supervised object detection framework for learning the object detectors from weakly labeled watercolor images.\\nA progressive domain adaptation method to transfer the style of the fully-labeled data from the source domain (the normal RGB domain) to the target domain (the watercolor domain) is developed.\\nIn Gonthier et al. propose a weakly supervised learning algorithm for detecting objects in paintings.\\nThe IconArt database which contains object classes that are absent from the photographs in daily life is developed for performance evaluation.\\nIn addition, Crowley\\nand Zisserman adopt a weakly supervised object localization scheme for automatically annotating images of gods and animals in decorations on classical Greek vases.\\n {When applying to art image analysis, a key challenge arises due to the distinctiveness of the content domain---even the same semantics and image scenes would be presented differently to those in the natural environment. Under this scenario, models with stronger self-domain adaptation capacity would be required for the task.}\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=9.5 cm]{artapp.jpg}\\n \\\\caption{Examples of the application of weakly supervised object localization or detection approaches for the analysis of art images.\\n The examples are from , where detection results in different colors in the painting images indicate different types of objects. }\\n \\\\label{art}\\n\\\\end{figure}\",\n \"id\": \"5bcd51f6-c4c7-4e80-b0c3-9cec6a4e37d4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"5cef946e-16aa-4550-aeb5-c885a8308ec2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Art Image Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Art Image Analysis\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 2321,\n 2322,\n 2320\n ],\n \"content\": \"As shown in Fig. \\\\ref{bio}, medical image analysis is one area where the weakly supervised object localization and detection methods are of critical importance as only few annotations of target objects by trained experts in bio-image (e.g., organ or tissues).\\nTo alleviate this problem, Hwang and Kim develop a two-stream DNN model to localize the tuberculosis regions from the chest X-ray images.\\n {Considering that the medical image-based applications usually do not have the pre-trained networks, this work proposes a weakly supervised learning scheme without requiring any pre-trained network parameters. The proposed network contains a fully connected layer-based classification branch and a CAM-based localization branch with shared convolutional layers for feature extraction. Both of the two branches are supervised by image label annotation, where a weighting parameter is introduced to dynamically control the relative importance between them to gradually switch the focus of the learning process from the classification branch to the localization branch. The authors demonstrate that the features learned from the classification layer at the early stage can provide informative cues to learn the localization branch at the late stage.}\\n {For detecting a general type of lesions, Wang et al. model the normal image as the combination of background and noise, while modeling the abnormal images as the combination of background, blood vessels, and noise. With the assumption that the noise for the normal image and abnormal image is the unified distribution, the image data can then be decomposed by the low-rank subspace learning technique to obtain the vessel areas. }\\nIn , Gondal et al. apply the weakly supervised object detection network on the retina images and achieve few false positives with high sensitivity on the lesion-level prediction.\\nLi et al. apply a sparse coding-based weakly supervised learning method for localizing actinophrys in microscopic images.\\nDubost et al. propose weakly supervised regression neural networks for detecting brain lesions. Besides, some recent works also show great research interests in weakly supervised learning-based brain image analysis, such as brain disease prognosis , brain tumor or lesion segmentation , brain structure estimation , etc.\\n {Notice that compared to common images, medical imaging data usually suffers from issues of low contrast and limited texture. Fortunately, some spatial priors could be obtained for different organs or lesions. These priors can be used to guide the weakly supervised learning process on medical imaging data.}\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=9.5 cm]{bioimage.jpg}\\n \\\\caption{Examples of the application of weakly supervised object localization or detection approaches in medical image analysis. The examples are from . }\\n \\\\label{bio}\\n\\\\end{figure}\",\n \"id\": \"6c796257-03d1-4c11-b23e-711e3eea5646\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"5cef946e-16aa-4550-aeb5-c885a8308ec2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Medical Imaging\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Medical Imaging\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Remote sensing imagery analysis is one of the most widely studied applications based on weakly supervised object localization and detection, where the input images are usually of large scale and the annotation process tends to be very time-consuming (see Fig. \\\\ref{remote}).\\nZhang et al. propose a saliency-based weakly supervised detector learning method to learn the detectors of the airplane, vehicle, and airport from the remote sensing images collected from different sensors.\\nIn this work, a weakly supervised object detection benchmark dataset for remote sensing imagery analysis is developed.\\nHan et al. and Zhou et al. introduce the Bayesian inference and negative bootstrapping methods, respectively, for effective weakly supervised detectors for remote sensing images.\\nIn , Zhang et al. propose a coupled CNN method which combines a candidate region proposal network and a localization network to detect aircrafts in images.\\n {When applying to remote sensing imagery analysis, the target objects are usually very small in size, which would dramatically increase the localization and detection difficulty given only weak supervision. }\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=9.5 cm]{remote.jpg}\\n \\\\caption{Examples of the application of weakly supervised object localization or detection approaches in remote sensing imagery analysis. The examples are from . }\\n \\\\label{remote}\\n\\\\end{figure}\",\n \"id\": \"ec87da5b-525f-4f14-a7b7-8c787fb56bb5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"5cef946e-16aa-4550-aeb5-c885a8308ec2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Remote Sensing Imagery Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Remote Sensing Imagery Analysis\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"We discuss the issues to be addressed in this field for future research.\",\n \"id\": \"d8feaee7-5d65-4327-96a9-38b968bfe1c4\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ]\n ],\n \"subsections\": [\n \"e95954d7-ff46-4ac2-807f-3d34b6e59526\",\n \"abffd2fa-2c1a-4f41-8253-4df7d27f9686\",\n \"56c3e88b-fc74-49e6-aa73-cff57609f8ef\",\n \"47c758a2-7983-451c-ad2b-fa45ee0b2b84\",\n \"8b5b77db-9af6-4869-b487-c1496c6cc1b1\"\n ],\n \"title\": \"Future Directions\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 2269,\n 2323,\n 737\n ],\n \"content\": \"Weakly supervised object localization or detection methods can be easily formulated within the MIL framework.\\nEarly methods in this filed usually add prior knowledge or post regularization on the classic MIL models, such as LSVM , while the current research obtains the breakthrough by building deep MIL models . To further improve the weakly supervised learning performance, efforts should be made to introduce the most advanced ideas and techniques in the research filed of MIL, such as the set-level problem the key instance shift issue and the scalable issue in MIL.\\nFurther research towards more advanced MIL techniques would also bring helpful insights for the WSOL and WSOD in the future.\",\n \"id\": \"e95954d7-ff46-4ac2-807f-3d34b6e59526\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"d8feaee7-5d65-4327-96a9-38b968bfe1c4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Multiple Instance Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multiple Instance Learning\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7091,\n 2260,\n 2324,\n 1751,\n 8429,\n 8540\n ],\n \"content\": \"Another future direction is to combine multiple weakly supervised learning tasks into a unified learning framework. These tasks may include object detection , semantic segmentation , instance segmentation , 3D shape reconstruction , and depth estimation .\\nEssentially, efforts for simultaneously accomplishing multiple aforementioned tasks have been made in the conventional fully supervised learning scenario , which have demonstrated that such learning mechanism can bring helpful information from one task to the other ones.\\nThe methods proposed by Zhang et al. is an early attempt to implement such a weakly supervised multi-task learning mechanism and the experimental results show that training object segmentation and 3D shape reconsecration models jointly indeed benefits the both weakly supervised learning tasks. With similar spirits to , Zhang et al. and Shen et al. establish a self-paced fine-tuning network and a cyclic guidance network to jointly learn object localization and segmentation models under the weak supervision, respectively. {Under the multi-task weakly supervised learning scenario, one key problem is that the learning ambiguity of each individual task might be aggregated and the imprecise prediction on one task might affect the learning on other tasks. To deal with this problem, one needs to disentangle the complex multi-task learning, separately learning each individual task first, and then leveraging the confidence knowledge from each task to provide informative priors to guide the learning processes of the other tasks. }\",\n \"id\": \"abffd2fa-2c1a-4f41-8253-4df7d27f9686\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"d8feaee7-5d65-4327-96a9-38b968bfe1c4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Multi-Task Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-Task Learning\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 2306,\n 8537\n ],\n \"content\": \"To address the learning under uncertainty issue that is inherently existed in the weakly supervised learning process, robust learning strategy will become one of the key techniques in the future. The goal is to alleviate the influence of the noisy samples during the learning process. In implementation, such learning strategy is usually achieved by selecting easy and confident training samples in the early learning stages while using hard and more ambiguous training samples in the late learning stages. Essentially, a number of recent methods have already introduced the robust learning strategies into their learning frameworks. For example, Shi and Ferrari propose a curriculum learning strategy to feed training images into the WSOL learning loop in order from images containing bigger objects down to smaller ones. The training order is determined by the size of the object, which is estimated based on a regression model. Similarly, Zhang et al. design a zigzag learning strategy, where they first develop a criterion to automatically rank the localization difficulty of an image, and then learn the detector progressively by feeding examples with increasing difficulty. As can be seen, these methods are just intuitive ways to introduce robust learning strategy into the weakly supervised object localization and detection frameworks, while they have already achieved obvious performance gains when compared with the conventional learning strategy. Along this line, Zhang et al. propose a self-paced curriculum learning framework for weakly supervised object detection. By integrating the curriculum learning with the self-paced learning , the established learning framework provides a more theoretical-sounded way to improve the learning robustness. However, the solid robust learning theory is still lack in this research field.\",\n \"id\": \"56c3e88b-fc74-49e6-aa73-cff57609f8ef\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"d8feaee7-5d65-4327-96a9-38b968bfe1c4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Robust Learning Theory\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Robust Learning Theory\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 2329,\n 2328,\n 2326,\n 2280,\n 2325,\n 2330,\n 2327,\n 8541\n ],\n \"content\": \"Besides the conventional CNN models, it is also worth trying to apply some more advanced learning models into the learning process of the weakly supervised object detector. Here we give two examples. The first one is the deep reinforcement learning.\\nAccording to~, biological vision systems are believed to have a sequential process with changing retinal fixations that gradually accumulate evidence of certainty when searching or localizing objects. Several existing methods~ have also demonstrated that designing deep reinforcement learning frameworks to model such a sequential searching process can indeed help to address the object localization, detection, and tracking problems in the computer vision community. Thus, it is highly desirable, both biologically and computationally, to explore deep reinforcement learning models that facilitate the weakly supervised object localization and detection systems in such a sequential searching process .\\nThe second one is the generative adversary learning. As we know, generative adversary learning has been demonstrated to have advantages in unsupervised and semi-supervised learning scenarios . It can generate the desired data distribution based on very weak supervision, i.e, ``real'' or ``fake''. Such capacity endows generative adversary learning very large potential in solving the weakly supervised object localization and detection problems. Although existing methods, such as , have already made efforts to introduce such a learning mechanism into the weakly supervised object localization and detection, there is still much room for improvement along this research direction.\",\n \"id\": \"47c758a2-7983-451c-ad2b-fa45ee0b2b84\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"d8feaee7-5d65-4327-96a9-38b968bfe1c4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Reinforcement and Adversarial Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Reinforcement and Adversarial Learning\"\n },\n {\n \"cite_extract_rate\": 0.533333333333333,\n \"cites\": [\n 2332,\n 2289,\n 8538,\n 2334,\n 2307,\n 2331,\n 2333,\n 2309\n ],\n \"content\": \"From Table \\\\ref{table endtoend} and Table \\\\ref{table multinet}, we can observe that most of the current deep weakly supervised object detection methods have not introduced any prior knowledge into their learning frameworks. However, from our review on classic models (see Sec. \\\\ref{Classic Models}), prior knowledges actually play important roles in avoiding the weakly supervised learning process from drifting to trivial solutions. Considering this issue, some recent works utilize prior knowledges of saliency , objectness , shape , count , human action , human object interaction , mask-out scoring in their frameworks. However, research towards building effective deep MIL frameworks (such as the one with prior knowledge distillation or cross domain adaptation ) to embed helpful prior knowledge into the weakly supervised learning process needs to be further explored in the future. In addition, the co-occurring patterns mined in co-saliency detection and object co-localization approaches can also be used as informative priors to guide the deep multiple instance learning process in weakly supervised object localization and detection.\",\n \"id\": \"8b5b77db-9af6-4869-b487-c1496c6cc1b1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"d8feaee7-5d65-4327-96a9-38b968bfe1c4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Prior-guided Deep MIL\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Prior-guided Deep MIL\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this paper, we provide a comprehensive survey of existing literatures in the research field of weakly supervised object localization and detection. We start with the introduction of the definition of the task and the key challenges that make the weakly supervised learning process hard to implement. Then, we introduce the development history of this field, the taxonomy of methods for weakly supervised object localization and detection, and the relationship between different categories. After reviewing existing literatures in each category of methodology, we introduce the benchmark datasets and evaluation metrics that are widely used in this filed, which are followed by the reviewing of the applications of the existing weakly supervised object localization and detection algorithms. Finally, we point out several future directions that may further promote the development of this research field.\\n{\\\\small\\n\\\\bibliographystyle{ieee}\\n\\\\bibliography{WSsurvey}\\n}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{dingwen_zhang}}]{Dingwen Zhang}\\nreceived his Ph.D. degree from the Northwestern Polytechnical University, Xi'an, China, in 2018. He is currently a full professor in School of Automation, Northwestern Polytechnical University. From 2015 to 2017, he was a visiting scholar at the Robotic Institute, Carnegie Mellon University. His research interests include computer vision and multimedia processing, especially on saliency detection, video object segmentation, and weakly supervised learning.\\n\\\\vspace*{-30pt}\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{HJW}}]{Junwei Han}\\n(M'12-SM'15) is a Professor with Northwestern Polytechnical University, Xi'an, China. He received Ph.D. degree in Northwestern Polytechnical University in 2003. He was a Research Fellow in Nanyang Technological University, The Chinese University of Hong Kong, and University of Dundee. His research interests include computer vision and brain imaging analysis. He has published over 100 papers in IEEE TRANSACTIONS and top tier conferences. He is currently an Associate Editor of IEEE Trans. on Neural Networks and Learning Systems, IEEE Trans. on Circuits and Systems for Video Technology, IEEE Trans. Cognitive and Developmental Systems, IEEE Trans. on Human-Machine Systems, Neurocomputing, and Machine Vision and Applications.\\n\\\\vspace*{-30pt}\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{GC}}]{Gong Cheng}\\nreceived the B.S. degree from Xidian University, Xi'an, China, in 2007, and the M.S. and Ph.D. degrees from Northwestern Polytechnical University, Xi'an, in 2010 and 2013, respectively. He is currently a Professor with Northwestern Polytechnical University. His main research interests are computer vision and pattern recognition.\\n\\\\vspace*{-30pt}\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{mhyang}}]{Ming-Hsuan Yang}\\nreceived the PhD degree in computer science from the University of Illinois at Urbana-Champaign, in 2000. He is a professor in Electrical Engineering and Computer Science from the University of California, Merced. He served as an associate editor of the IEEE Transactions on Pattern Analysis and Machine Intelligence from 2007 to 2011, and is an associate editor of the International Journal of Computer Vision, the Computer Vision and Image Understanding, the Image and Vision Computing, and the Journal of Artificial Intelligence Research. He received the NSF CAREER award in 2012, and the Google Faculty Award in 2009. He is a Fellow of the IEEE and Senior Member of the ACM.\\n\\\\vspace*{-30pt}\\n\\\\end{IEEEbiography}\\n\\\\end{document}\",\n \"id\": \"92209b0e-e020-4aa4-8af9-77a6bcf0b9af\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3a2d911b-3708-4f3a-880a-851136519e79\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Weakly Supervised Object Localization and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":76,"string":"76"},"all_cites":{"kind":"list like","value":[2259,209,2260,2261,2262,2263,2267,2269,2270,2266,2264,2265,8537,2268,2272,2271,2274,2273,1223,514,8429,2275,895,2278,2276,2279,2277,2282,2297,2288,2289,2291,2286,2294,7089,2285,2293,2287,2283,2299,2296,2284,2292,737,2295,2298,2281,2290,2280,8538,2306,2303,2301,802,2302,2309,2300,2304,2308,2307,2310,2305,2312,2311,2314,2315,107,7090,687,2313,2319,2317,2316,2318,8539,2321,2322,2320,2323,7091,2324,1751,8540,2329,2328,2326,2325,2330,2327,8541,2332,2334,2331,2333],"string":"[\n 2259,\n 209,\n 2260,\n 2261,\n 2262,\n 2263,\n 2267,\n 2269,\n 2270,\n 2266,\n 2264,\n 2265,\n 8537,\n 2268,\n 2272,\n 2271,\n 2274,\n 2273,\n 1223,\n 514,\n 8429,\n 2275,\n 895,\n 2278,\n 2276,\n 2279,\n 2277,\n 2282,\n 2297,\n 2288,\n 2289,\n 2291,\n 2286,\n 2294,\n 7089,\n 2285,\n 2293,\n 2287,\n 2283,\n 2299,\n 2296,\n 2284,\n 2292,\n 737,\n 2295,\n 2298,\n 2281,\n 2290,\n 2280,\n 8538,\n 2306,\n 2303,\n 2301,\n 802,\n 2302,\n 2309,\n 2300,\n 2304,\n 2308,\n 2307,\n 2310,\n 2305,\n 2312,\n 2311,\n 2314,\n 2315,\n 107,\n 7090,\n 687,\n 2313,\n 2319,\n 2317,\n 2316,\n 2318,\n 8539,\n 2321,\n 2322,\n 2320,\n 2323,\n 7091,\n 2324,\n 1751,\n 8540,\n 2329,\n 2328,\n 2326,\n 2325,\n 2330,\n 2327,\n 8541,\n 2332,\n 2334,\n 2331,\n 2333\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.6972244941615191,"string":"0.697224"}}},{"rowIdx":42,"cells":{"authors":{"kind":"list like","value":["Manish Gupta","Puneet Agrawal"],"string":"[\n \"Manish Gupta\",\n \"Puneet Agrawal\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Compression of Deep Learning Models for Text: A Survey"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-08-12T10:42:14Z","string":"2020-08-12T10:42:14Z"},"category":{"kind":"string","value":"cs.CL"},"abstract":{"kind":"string","value":" In recent years, the fields of natural language processing (NLP) and information retrieval (IR) have made tremendous progress thanks to deep learning models like Recurrent Neural Networks (RNNs), Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTMs) networks, and Transformer~\\cite{vaswani2017attention} based models like Bidirectional Encoder Representations from Transformers (BERT)~\\cite{devlin2018bert}, Generative Pre-training Transformer (GPT-2)~\\cite{radford2019language}, Multi-task Deep Neural Network (MT-DNN)~\\cite{liu2019multi}, Extra-Long Network (XLNet)~\\cite{yang2019xlnet}, Text-to-text transfer transformer (T5)~\\cite{raffel2019exploring}, T-NLG~\\cite{rosset2019turing} and GShard~\\cite{lepikhin2020gshard}. But these models are humongous in size. On the other hand, real world applications demand small model size, low response times and low computational power wattage. In this survey, we discuss six different types of methods (Pruning, Quantization, Knowledge Distillation, Parameter Sharing, Tensor Decomposition, and Sub-quadratic Transformer based methods) for compression of such models to enable their deployment in real industry NLP projects. Given the critical need of building applications with efficient and small models, and the large amount of recently published work in this area, we believe that this survey organizes the plethora of work done by the `deep learning for NLP' community in the past few years and presents it as a coherent story. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"]],"subsections":["57df7b91-17f0-42ac-b043-8a5e492f9515","92ac6101-b828-4042-8708-2e0386e701bb","c0c0e7e6-55dc-426b-8669-a8adfb8fe243","2c5b1b2b-3f74-442d-8ee6-503425cf4207","57903397-1a0f-4365-9fd9-a6bf95e76421","a059247f-141a-43b7-a6a9-f3440e7622cd","65404f84-cbf0-4721-8f05-178f8230aa55","d06ca9cc-bcd7-4f53-9be8-ee86f8754664"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:overview}\nIn this survey, we discuss compression methods using pruning, quantization, knowledge distillation, parameter sharing, tensor decomposition and sub-quadratic Transformers. \nThe most obvious way to reduce model size is to sparsify weight matrices. Pruning methods differ based on what is pruned and the actual logic used to prune. Given a matrix, one can prune \n\\begin{itemize}\n \\item Some weight entries\n \\item Rows or columns (i.e., neurons)\n \\item Weight blocks\n \\item Attention heads (in case of Transformer based methods)\n \\item Layers or a combination of the structures.\n\\end{itemize}\nOther interesting aspects of pruning methods include the following.\n\\begin{itemize}\n \\item How to decide which weights/neurons/blocks/heads to prune?\n \\item Should you prune large networks or build small networks?\n \\item Should you do one-shot pruning versus iterative/gradual pruning? \n \\item How does regularization help when pruning?\n\\end{itemize}\n We discuss these aspects of pruning based methods in Section~\\ref{sec:pruning}.\nBesides removing the weights themselves, another way to reduce the size of weight matrices is to reduce the number of bits needed to represent each weight. Typically weights are stored as 32-bit double values. In an extreme case, weights can be quantized to two values (binary 1 bit). But other popular ways include quantization to three values (ternary) or multiple bits. Quantization can be uniform or non-uniform. Quantization methods can be deterministic or stochastic. Quantization can be performed in a loss-aware or unaware manner. Quantization bin ranges can be trained versus tuned. Finally, the level of quantization needs to be different across layers of RNNs, LSTMs or Transformers to attain a favorable model size versus accuracy tradeoff. We discuss these aspects of quantization based methods in Section~\\ref{sec:quantization}.\nA third way of doing model compression is using knowledge distillation (also broadly known as student teacher networks). In such methods, the idea is to first train a deep teacher model using labeled data, and then transfer ``dark knowledge'' from teacher to train a shallow student network. Thus, the student model is trained to mimic a pre-trained, larger teacher. After the student is trained, it is deployed while the teacher network can be thrown. Distillation methods vary based on the following design choices.\n\\begin{itemize}\n \\item Different types of teacher model\n \\item Different types of loss function like squared error between the logits of the models, KL divergence between the predictive distributions, or some other measure of agreement between the model predictions\n \\item Different choices for what dataset the student model trains on (a large unlabeled dataset, a held-out data set, or the original training set)\n \\item Different choices for what to mimic from the teacher -- teacher's class probabilities, teacher's feature representation.\n \\item Learn from whom -- teacher, teacher assistant, or other fellow students. \n\\end{itemize}\nWe discuss these aspects of knowledge distillation based methods in Section~\\ref{sec:knowledgeDistillation}.\nAnother way that reduces overall weights is to have parameters shared across multiple weight structures. Broadly, the method is called parameter sharing. Methods differ depending on the following aspects.\n\\begin{itemize}\n \\item Which parameters are shared\n \\item Technique used to share parameters\n \\item The level at which sharing is performed\n\\end{itemize}\nWe discuss these aspects of parameter sharing based methods in Section~\\ref{sec:parameterSharing}.\nYet another way to avoid large matrices is to approximate them using a combination of smaller matrices. Such tensor decomposition methods for model compression factorize large tensors into multiple smaller components. Methods differ based on the following aspects.\n\\begin{itemize}\n \\item Type of factorization technique\n \\item Matrices being factorized\n \\item The property of weight matrix being exploited\n\\end{itemize}\nWe discuss these aspects of tensor decomposition methods in Section~\\ref{sec:matrixDecomposition}.\nIn Transformer based models, besides the model size, latency is a concern because of quadratic complexity in terms of the input sequence size. Also, RAM needed for Transformer models is quadratic in nature. Hence, very recently, there have been several efforts on designing Transformers with sub-quadratic complexity. Some of these methods are super-linear while many are linear. Linear complexity methods use various techniques for enforcing linearity -- the broad idea is to compute a transformed representation for every token using attention over a fixed small number of other tokens. Methods differ in terms of defining the set of other tokens to be used to compute a transformed representation for the current token. We discuss such methods in Section~\\ref{sec:linearTransformers}.","id":"57df7b91-17f0-42ac-b043-8a5e492f9515","level":"section","origin_cites_number":0,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Model Compression Methods: Overview"]],"subsections":[],"title":"Model Compression Methods: Overview"},{"cite_extract_rate":0.5,"cites":[844],"content":"\\label{sec:pruning} \nThe first proposed methods for model compression were based on pruning. Pruning can be done in two ways: structured versus unstructured. In unstructured pruning, individual weight connections are removed from a network by setting them to 0. One can prune away weights from a weight matrix depending on various criteria (e.g., prune away low magnitude weights). Such unstructured pruning methods lead to sparse matrices and need special sparse matrix manipulation libraries at inference time. Hence, various structured pruning methods have also been proposed which aim to prune away structures like neurons, weight matrix blocks, attention heads or layers. Fig.~\\ref{fig:pruning} provides a broad overview of various pruning styles. Fig~\\ref{fig:pruning}(B) depicts unstructured pruning. Fig~\\ref{fig:pruning}(C)-(G) shows various structured pruning methods. In this section, we discuss unstructured weight pruning methods in Section~\\ref{subsec:PruningWeights} and structured pruning methods in Sections~\\ref{subsec:PruningNeurons}-\\ref{subsec:PruningHeadsLayers}.\n\\begin{figure*}\n \\centering\n \\includegraphics[width=\\textwidth]{pruning}\n \\caption{Different Types of Pruning: (A) represents no pruning. Filled cells represent pruned entries. (B), (H) and (I) are unstructured pruning methods discussed in Section~\\ref{subsec:PruningWeights}. Remaining are structured pruning methods: (C) is discussed in Section~\\ref{subsec:PruningNeurons}, (D) in Section~\\ref{subsec:PruningBlock}, (E), (F) and (G) in Section~\\ref{subsec:PruningHeadsLayers}.}\n \\label{fig:pruning}\n\\end{figure*}\nIn pruning, the main idea is to grow a large model and then prune away weights to end up with a much smaller but effective model. This is inspired by the following biological observation. Trillions of synapses are generated in the human brain during the first few months of birth. At one year old, synapse count peaks at 1000 trillion. And then natural pruning begins to occur. A ten year old child has nearly 500 trillion synapses. This `pruning' mechanism removes redundant connections in the brain~. \nOne natural question is should you prune large networks or build small dense networks? Pruning involves extra processing plus sparse matrices need special handling. Can we avoid it by training smaller models? Zhu et al.~ experimented with models of various sizes with/ without pruning of stacked LSTMs models for language modeling, and seq2seq models for Neural Machine Translation (NMT). They found that large-sparse models consistently outperform small-dense models and achieve up to 10x reduction in number of non-zero parameters with minimal loss in accuracy.","id":"92ac6101-b828-4042-8708-2e0386e701bb","level":"section","origin_cites_number":2,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"]],"subsections":["2362b1c7-732c-4073-87eb-1e3f045414b0","56b21370-ac94-4d19-9ed7-985faad42cc8","49a2e5f4-1aec-4e81-ac60-4671c95948a6","10a33a73-e952-454b-b5ce-09086a71de83","80f88ea7-37a5-4100-9fe4-121a616869c9"],"title":"Pruning"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:PruningWeights}\nThe weights to be pruned can be chosen using two heuristics: (1) Using Hessian matrix computation, or (2) using magnitude of the weights. Further, magnitude based pruning methods could do pruning in one shot (typically statically after training is done), or iteratively (also called dynamic or gradual pruning), or iteratively with pruning and densification. Accordingly, there are four main ways of performing unstructured weight pruning: (1) Hessian based methods, (2) magnitude pruning, (3) iterative magnitude pruning, and (4) iterative magnitude pruning and densification. In this subsection, we discuss these methods in detail.","id":"2362b1c7-732c-4073-87eb-1e3f045414b0","level":"subsection","origin_cites_number":0,"parent_id":"92ac6101-b828-4042-8708-2e0386e701bb","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Weights"]],"subsections":["17e91dfe-059c-4ae9-8bb2-0fdcc9c148d7","db3c2850-1885-440e-a725-4d53c91cb9a3","bdf39c62-744f-439a-a336-cc2914c01b9e","8ead8d95-de3b-4322-8928-4d2d5dd4066f"],"title":"Pruning Weights"},{"cite_extract_rate":0,"cites":[],"content":"In their seminal work (Optimal Brain Damage or OBD) on proposing weight pruning as a method for model compression, LeCun et al.~ defined saliency of a weight as change in the objective function $E$ caused by deleting that parameter. Using Taylor series and making multiple assumptions, they propose to use the following as a measure of saliency of weight $u_i$.\n\\begin{eqnarray}\n s(u_i)=\\frac{1}{2}\\sum_i h_{ii} \\delta u_i^2\n\\end{eqnarray}\n\\noindent where $h_{ii}=\\frac{\\partial^2 E}{\\partial u_i \\partial u_j}$ is the $i^{th}$ element on the diagonal of the Hessian matrix. Weights with low saliency can be pruned and the pruned network can be retrained to adjust the remaining weights. \nThe procedure for computation of the diagonal of the Hessian is very similar to the back-propagation algorithm used for computing the first derivatives. Hence, computing the diagonal of the Hessian is of the same order of complexity as computing the gradient. \nOBD ignores cross terms in the Hessian matrix. But on most real datasets, Hessian is strongly non-diagonal. Hence, to avoid pruning of important weights, Hassibi et al.~ proposed a method called Optimal Brain Surgeon (OBS) which considers cross terms as well. Using a similar derivative of $E$ wrt weight $w_i$, saliency of the weight is computed as follows.\n\\begin{eqnarray}\nL_i=\\frac{1}{2}\\frac{w_i^2}{[H^{-1}]_{ii}} \n\\end{eqnarray}\nComputing $H^{-1}$ is difficult. Hence, they provide a faster recursion relation for calculating $H^{-1}$ from training data and structural information of the network. Also, unlike other methods (like OBD or magnitude pruning), OBS does not demand (typically slow) retraining after the pruning of a weight. In every step, we delete $w_i$ with min $L_i$ and update all weights as follows.\n\\begin{eqnarray}\n \\delta w=-\\frac{w_i}{[H^{-1}]_{ii}}H^{-1}e_i\n\\end{eqnarray}\n\\noindent where $e_i$ is the unit vector in weight space corresponding to (scalar) weight $w_i$. Unfortunately, these methods (OBD and OBS) are computationally prohibitive as second derivative (i.e. Hessian) computations are expensive.","id":"17e91dfe-059c-4ae9-8bb2-0fdcc9c148d7","level":"subsubsection","origin_cites_number":2,"parent_id":"2362b1c7-732c-4073-87eb-1e3f045414b0","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Weights"],["subsubsection","Hessian based Methods"]],"subsections":[],"title":"Hessian based Methods"},{"cite_extract_rate":0.5,"cites":[4478],"content":"A more computationally feasible method for pruning connections and relearning weights based solely on the magnitude of the original weights is to simply prune away weights with small magnitudes. The idea was first proposed by Han et al.~. Pruning away low magnitude weights makes the matrix sparse. Sparse matrices can be stored in Compressed Sparse Row/Column (CSR/CSC) formats. Further space can be saved by storing the index difference instead of the absolute position, and encoding this difference using small fixed number of bits. See et al.~ experimented with these pruning methods on encoder-decoder LSTM NMT (neural machine translation) models. They perform magnitude pruning on all weight matrices of a 4-layer LSTM. They find that higher layers, attention and softmax weights are the most important, while lower layers and the embedding weights hold a lot of redundancy. At the lower layers the parameters for the input are most crucial, but at higher layers the parameters for the gates also become important. These methods typically have a target pruning percent as a hyper-parameter and pruning is either performed statically (after training the full model) or dynamically (while training itself). Retraining the sparse pruned network helps in improving accuracy.\nIn a typical encoder-decoder LSTM model, there are these weight classes: source embedding weights, target embedding weights, source layer weights, target layer weights, attention weights and softmax weights. An important consideration related to magnitude pruning is how do we distribute the pruning over these different weight classes of a model, given a target $x$\\% pruning? Three ways suggested by See et al.~ include class-blind, class-uniform and class-distribution. In the class-blind way, we take all parameters, sort them by magnitude and prune the $x$\\% with smallest magnitude, regardless of the weight class. So some classes are pruned proportionally more than others. In the class-uniform way, Within each class, we sort the weights by magnitude and prune the x\\% with smallest magnitude. So all classes have exactly x\\% of their parameters pruned. In the class-distribution scheme, for each class $c$, weights with magnitude less than $\\lambda\\sigma_c$ are pruned. Here, $\\sigma_c$ is the standard deviation of that class and $\\lambda$ is a universal parameter chosen such that in total, $x$\\% of all parameters are pruned. Class-blind pruning is the simplest and adheres to the principle that pruning weights (or equivalently, setting them to zero) is least damaging when those weights are small, regardless of their locations in the architecture. Class-uniform pruning and class-distribution pruning both seek to prune proportionally within each weight class, either absolutely, or relative to the standard deviation of that class. They observe that class-blind pruning outperforms both other schemes.","id":"db3c2850-1885-440e-a725-4d53c91cb9a3","level":"subsubsection","origin_cites_number":2,"parent_id":"2362b1c7-732c-4073-87eb-1e3f045414b0","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Weights"],["subsubsection","Magnitude Pruning Methods"]],"subsections":[],"title":"Magnitude Pruning Methods"},{"cite_extract_rate":0.6000000000000001,"cites":[4331,4479,844],"content":"Typically, it has been seen that rather than pruning in one-shot, it is a good idea to prune gradually over epochs. This way of pruning is called iterative or gradual pruning (see Fig.~\\ref{fig:pruning}(H)). For starting proportion $x$\\% and ending proportion $y$\\%, iterative magnitude pruning procedure prunes $x$\\% of each of the weights, does re-training, and then prunes $(y-x)/T$\\% of the weights every $K$ iterations. $T$ is the number of times, pruning is done. Sometimes, pruning is started after few warmup iterations have already been performed. Magnitude pruning has been seen to be very effective with regularization ($L_1/L_2$) while training. Dropouts also help in effective pruning. \nIn some pruning methods, a weight once pruned cannot be a part of the network in future iterations. On the other hand, other methods do not modify the gradients of a pruned weight in the back-propagation step. In that case, it is possible for the updates of a pruned weight to be larger than the threshold of that layer, and then the weight will be involved in the forward pass again. Also, in every pruning iteration, we could either use a fixed threshold~ or monotonically increase it~. \nIn case of gradual pruning~, where pruning threshold $\\epsilon$ is monotonically increased, $\\epsilon$ is determined as follows in every iteration $i$. Let $f$ be the number of iterations after which $\\epsilon$ is updated. Also, after a few warmup iterations, weights are sorted using absolute values and we pick the weight corresponding to the $90^{th}$ percentile as $q$. Pruning threshold $\\epsilon$ is increased in two stages. In the first stage (which starts at start iteration $s$ and continues until ramp iteration $r$, we use $\\theta$ as the initial slope to prune weights. In the second stage (which starts at ramp iteration $r$ and continues until end iteration $e$), we use $\\phi$ as the ramp slope to change the rate of pruning. Typically, $\\phi$ is set to 1.5$\\theta$ where $\\theta$ is calculated as follows.\n\\begin{eqnarray}\n \\theta=\\frac{2qf}{2(r-s)+3(e-r)}\n\\end{eqnarray}\nThus, from iteration $s$ to $r$, we set the pruning threshold as follows.\n\\begin{eqnarray}\n\\epsilon=\\frac{\\theta(i-s+1)}{f} \n\\end{eqnarray}\nFrom iterations $r+1$ to $e$, we set the pruning threshold as follows.\n\\begin{eqnarray}\n\\epsilon=\\frac{\\theta(r-s+1)+\\phi(i-r+1)}{f}\n\\end{eqnarray}\nTypically when pruning, biases are not pruned since they are much fewer in number. Overall, RNN/LSTM model size can be reduced by 90\\% and speed-up is around 2x to 7x using gradual pruning with no deterioration in accuracy. Also, layers closer to input are pruned more aggressively compared to the final layers.\nAnother way of performing iterative pruning is to set a pruning target per iteration~. In this scheme, we start with an initial sparsity value $s_0$. To achieve a final sparsity value of $s_f$ after $n$ pruning steps with pruning frequency $f$, pruning target in iteration $i$ can be computed as follows.\n\\begin{eqnarray}\ns_i=s_f+(s_0-s_f)\\left(1-\\frac{i}{nf}\\right)^3\n\\end{eqnarray}\nThus, the sparsity of the network is gradually increased while allowing the network training steps to recover from any pruning-induced loss in accuracy. We prune the network rapidly in the initial phase when the redundant connections are abundant and gradually reduce the number of weights being pruned each time as there are fewer and fewer weights remaining in the network.\nCheong et al.~ found that iterative pruning leads to poor results when pruning Transformer models like BERT. Guo et al.~ found that there are two problems with pruning especially when done with regularization. \n\\begin{itemize}\n \\item The larger weights $w_j$ are penalized more heavily than smaller weights $w_i$ in $L_1$ regularization, which violates the original intention of weight pruning, ``removing the unimportant connections''. \n \\item Direct optimization of a regularization penalty term causes divergence from the original loss function and has negative effect on the effectiveness of gradient-based update.\n\\end{itemize}\nThey propose to perform reweighted $L_1$ minimization where $\\alpha_i>0$ are inversely proportional to magnitude of corresponding weights $|w_i|$. Thus, they solve the following optimization problem\n\\begin{eqnarray}\n \\min_w f(w)+\\gamma \\sum_i \\alpha_i |w_i|\n\\end{eqnarray}\n\\noindent where $f(w)$ is the original loss function for the network. This optimization is solved using a reweighted proximal pruning (RPP) method (which depends on proximal operators). RPP decouples the goals of high sparsity from minimizing loss, and hence leads to improved accuracy even with high levels of pruning for BERT.","id":"bdf39c62-744f-439a-a336-cc2914c01b9e","level":"subsubsection","origin_cites_number":5,"parent_id":"2362b1c7-732c-4073-87eb-1e3f045414b0","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Weights"],["subsubsection","Iterative Magnitude Pruning Methods"]],"subsections":[],"title":"Iterative Magnitude Pruning Methods"},{"cite_extract_rate":0.5,"cites":[4480],"content":"Further, the effectiveness of pruning can be improved by performing pruning and densification~ alternately across multiple iterations (see Fig.~\\ref{fig:pruning}(I)). There are two ways of doing this. In the first method~, in each iteration, either pruning is performed or densification. The sparse training regularizes the model, and the dense training restores the pruned weights, increasing the model capacity without overfitting. Sparsification helps the optimizer escape saddle points, and leads to regularized training which converges to a significantly better minima. In the second method~, in every iteration some dormant weights can reappear in the network while other active ones can get pruned out. A dormant $w\\in W$ is activated iff $|w.grad|$ is larger than the $(100\\alpha)^{th}$ percentile of all elements in $|W.grad|$. A $w\\in W$ is removed iff $|w|$ is smaller than the $(100\\beta)^{th}$ percentile of all elements in $|W|$. $\\alpha$ and $\\beta$ refer to growth ratio, and pruning ratio, respectively.","id":"8ead8d95-de3b-4322-8928-4d2d5dd4066f","level":"subsubsection","origin_cites_number":2,"parent_id":"2362b1c7-732c-4073-87eb-1e3f045414b0","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Weights"],["subsubsection","Iterative Magnitude Pruning and Densification"]],"subsections":[],"title":"Iterative Magnitude Pruning and Densification"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:PruningNeurons}\nIt is difficult to implement unstructured pruning practically since, at inference time, special support is needed for matrix multiplication in the sparse space. Pruning away neurons leads to removal of a row or a column from a weight matrix, thereby avoiding sparse matrix handling (see Fig.~\\ref{fig:pruning}(C)). However, compared to pruning weights, pruning neurons is less flexible since we need to find entire rows/columns for deletion. In this section, we discuss ways of determining neurons that can be pruned.","id":"56b21370-ac94-4d19-9ed7-985faad42cc8","level":"subsection","origin_cites_number":0,"parent_id":"92ac6101-b828-4042-8708-2e0386e701bb","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Neurons"]],"subsections":["60bed51b-08a1-4c61-9777-2265a58630d5","e7c0d686-3915-4133-8281-2cebccd6dd52"],"title":"Pruning Neurons"},{"cite_extract_rate":0.6666666666666661,"cites":[4481,4482],"content":"He et al.~ proposed three node importance functions to determine importance score for neurons. \n\\begin{itemize}\n \\item Entropy: For a neuron $i$, let $a_i$ ($d_i$) be the \\#instances with node output $> (or\\leq)$ 0.5 for binary classification with a sigmoid activation. Then entropy for node $i$ can be written as follows.\n \\begin{eqnarray}\n \\text{Entropy}(i)=\\frac{d_i}{a_i+d_i}\\log_2 \\frac{d_i}{a_i+d_i}+\\frac{a_i}{a_i+d_i}\\log_2 \\frac{a_i}{a_i+d_i} \n \\end{eqnarray}\n The intuition is that if one node's outputs are almost identical on all training data, these outputs do not generate variations to later layers and consequently the node may not be useful. \n \\item Output-weights Norm (onorm): average $L_1$-norm of the weights of its outgoing links.\n \\item Input-weights norm (inorm): average $L_1$-norm of the weights of its incoming links.\n\\end{itemize}\nAll the neurons are sorted by their scores and nodes with less importance values are removed. In most cases, onorm has been found to be the best among these importance functions. \nSpecial regularizers have also been proposed to force neurons to push either all incoming or outgoing connection weights towards zero~. Specifically, for handling incoming connections, the following two regularizers are popular.\n\\begin{itemize}\n \\item $L_2$ norm on weight matrix $W$ defined as follows.\n \\begin{eqnarray}\n \\sum_i ||W_{i:}||_2=\\sum_i \\left(\\sum_j W^2_{ij}\\right)^{1/2}\n \\end{eqnarray}\n This puts equal pressure on each row, but within each row, the larger values contribute more, and therefore there is more pressure on larger values towards zero. \n \\item $L_\\infty$ norm on weight matrix $W$ defined as follows.\n \\begin{eqnarray}\n \\sum_i ||W_{i:}||_\\infty=\\sum_i \\max_j |W_{ij}|\n \\end{eqnarray}\n This puts equal pressure on each row, but within each row, only the maximum value (or values) matter, and therefore the pressure towards zero is entirely on the maximum value(s).\n\\end{itemize}\nSimilar regularizers can easily be defined for outgoing connections as well.","id":"60bed51b-08a1-4c61-9777-2265a58630d5","level":"subsubsection","origin_cites_number":3,"parent_id":"56b21370-ac94-4d19-9ed7-985faad42cc8","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Neurons"],["subsubsection","Removing Low Importance Neurons"]],"subsections":[],"title":"Removing Low Importance Neurons"},{"cite_extract_rate":1,"cites":[4483],"content":"Consider a simple network with one hidden layer with $n$ neurons. Thus, the output can be computed as follows.\n\\begin{eqnarray}\nz=a_1 h(W_1^TX)+a_2 h(W_2^TX)+...+a_n h(W_n^TX)\n\\end{eqnarray}\n\\noindent where $a_i$ and $W_i$ indicate weights. In case $W_1==W_2$, $h(w_1^TX)=h(w_2^TX)$. Thus, we can compute output as follows.\n\\begin{eqnarray}\nz=(a_1+a_2) h(W_1^TX)+0. h(W_2^TX)+...+a_n h(W_n^TX)\n\\end{eqnarray} \nIn general, whenever two weight sets ($W_1$, $W_2$) are equal, one of them can effectively be removed. This should be done with a surgery step, i.e., we need to alter the co-efficient $a_1$ to $a_1+a_2$. Of course, for many pairs of weight sets (i.e., neurons), $W_1$ and $W_2$ are not exactly the same. Hence, Srinivas et al.~ proposed this 3 step method for redundant neuron identification and removal. \n\\begin{itemize}\n \\item Compute saliency $s_{ij}$ for all possible neuron pairs (i, j) as follows.\n \\begin{eqnarray}\n s_{ij}=\\langle a_j^2 \\rangle||\\epsilon_{ij}||_2^2\n \\end{eqnarray} \n \\noindent where $\\langle a_j^2 \\rangle$ denotes the average of the quantity over all output neurons. Let $S$ be the matrix with all $s_{ij}$ values. \\item Pick the indices $(i',j')$ corresponding to the minimum $s_{ij}$. Delete the $j'$ neuron, and update $a_i'$ as follows.\n \\begin{eqnarray}\n a_i'\\leftarrow a_i'+a_j' \n \\end{eqnarray}\n \\item Update $S$ by removing the $j'^{th}$ column and row, and updating the $i'^{th}$ column (to account for the updated $a_i'$).\n\\end{itemize}","id":"e7c0d686-3915-4133-8281-2cebccd6dd52","level":"subsubsection","origin_cites_number":1,"parent_id":"56b21370-ac94-4d19-9ed7-985faad42cc8","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Neurons"],["subsubsection","Removing Redundant Neurons"]],"subsections":[],"title":"Removing Redundant Neurons"},{"cite_extract_rate":0.5,"cites":[7826],"content":"\\label{subsec:PruningBlock}\nIn weight pruning, irregularity of sparse matrices limits the maximum performance and energy efficiency achievable on hardware accelerators. Pruning neurons avoids sparse matrix issues but is limited in term of overall pruning possible. Hence, block based pruning methods were introduced (see Fig.~\\ref{fig:pruning}(D)).\nBlock-sparse formats store blocks contiguously in memory reducing irregular memory accesses. If the maximum magnitude weight of a block is below the current threshold, we set all the weights in that block to zeros. Similar to iterative weight pruning, block pruning~ can also be done iteratively using a monotonically growing threshold. Any blocks that had been zeroed out are held at zero even after pruning has ended resulting in a sparse model at the end of training. Just like weight pruning (as discussed in Section~\\ref{subsec:PruningWeights}), the start slope $\\theta$ and ramp slope $\\phi$ determine the rate at which the threshold increases. For block pruning, we need to modify the start slope to account for the number of elements in a block ($N_b$). Thus, the start slope for block pruning is typically set as follows.\n\\begin{eqnarray}\n\\theta_b=\\theta\\times\\sqrt[\\leftroot{-2}\\uproot{2}4]{N_b} \n\\end{eqnarray}\nFurther, to enable effective removal of blocks, Narang et al.~ propose group Lasso regularization method. Group lasso is a type of weight regularization that works on groups of weights and can zero out all the weights in a group. For each block, we add a loss term proportional to the $L_2$ norm of the block. Thus, we optimize for the following.\n\\begin{eqnarray}\n\\min_w f(w)+\\lambda_g \\sum_{g=1}^{G} ||w^{(g)}||_2 \n\\end{eqnarray}\nWhen we combine group lasso with block pruning, group lasso guides the selection of blocks to prune. Group lasso regularization is applied to coincide with the pruning schedule, i.e., we turn off regularization when the pruning schedule ends. Typically, inducing block sparsity with 4x4 blocks in vanilla RNNs and GRUs works well, compared to larger block sizes. Larger blocks require lower sparsity to maintain similar accuracy. \nUnfortunately, it becomes challenging to maintain the same model accuracy when block sparsity is applied. Also, block sizes (i.e., pruning granularity) are application-sensitive, making it another hyper-parameter to tune. To avoid these problems, Cao et al.~ proposed a new method called Bank-Balanced Sparsity (BBS). BBS splits each weight matrix row into multiple equal-sized banks, and adopts fine-grained pruning to each bank independently to obtain identical sparsity among banks. Each bank has the same number of non-zero values. For example, retaining top two weights in each bank of size 4 implies a sparsity of 50\\%. We apply the BBS pruning method iteratively to a pre-trained network, and fine-tune the network after each pruning iteration to restore the model accuracy. BBS achieves almost the same model accuracy as unstructured sparsity and significantly outperforms block sparsity when pruning weights at the same sparsity level. BBS is also amenable to FPGA (Field Programmable Gate Arrays) acceleration because it inherently provides a balanced matrix partitioning for parallel computing.","id":"49a2e5f4-1aec-4e81-ac60-4671c95948a6","level":"subsection","origin_cites_number":2,"parent_id":"92ac6101-b828-4042-8708-2e0386e701bb","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Blocks"]],"subsections":[],"title":"Pruning Blocks"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:PruningHeadsLayers}\nBesides neurons and blocks, for Transformer based models, structured pruning can also be applied to attention heads and entire layers. In this section, we discuss such methods.","id":"10a33a73-e952-454b-b5ce-09086a71de83","level":"subsection","origin_cites_number":0,"parent_id":"92ac6101-b828-4042-8708-2e0386e701bb","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Heads and Layers"]],"subsections":["bc2b6e76-0a87-4b66-8b3c-2bc732d0a821","c4355812-9f8a-4da4-86ff-8739c39b93e8","7dc1a143-838c-41ca-90a0-ebe43039cc74"],"title":"Pruning Heads and Layers"},{"cite_extract_rate":0.6666666666666661,"cites":[4485,4484],"content":"BERT-base model consists of 12 layers each with 12 attention heads. Similarly, a typical NMT encoder-decoder Transformer with 6 layers each for encoder as well as decoder contains 16 attention heads per layer. Michel et al.~ found that majority of attention heads can be removed without deviating too much from the original score. Surprisingly, in some cases removing an attention head (see Fig.~\\ref{fig:pruning}(E)) results in an increase in accuracy. When these heads are removed individually, only 8 (out of 96) heads in 6-layer WMT NMT Transformer (16 heads/layer) cause a statistically significant change in performance when they are removed from the model, half of which actually result in a higher BLEU score. For most layers, one head is indeed sufficient at test time, even though the network was trained with 12 (BERT) or 16 (WMT Transformer) attention heads. One can also do iterative pruning of multiple heads (rather than just one at a time) across layers. For iterative pruning, head importance score is defined using the expected sensitivity of the model to the mask variables $\\xi_h$ as follows.\n\\begin{equation}\n I_h=E_{x\\sim X}\\left|\\frac{\\partial L(x)}{\\partial \\xi_h}\\right|=E_{x\\sim X}\\left|\\text{Att}_h(x)^T\\frac{\\partial L(x)}{\\partial \\text{Att}_h(x)}\\right|\n\\end{equation}\n\\noindent where $X$ is the data distribution, $L(x)$ is the loss on sample $x$, and $Att_h(x)$ is the output of the attention head $h$ for instance $x$, . Intuitively, if $I_h$ has a high value then changing $\\xi_h$ is liable to have a large effect on the model. Hence, in every iteration heads with low $I_h$ values are pruned out. Michel et al.~ observed that pruning up to 20\\% and 40\\% of heads from NMT and BERT models respectively, did not lead to any noticeable negative impact on accuracy.\nVoita et al.~ used two other head importance scores to prune attention heads from the NMT model. The two scoring methods were: (1) Layer-wise relevance propagation (LRP)~. LRP is a method for computing the relative contribution of neurons at one point in a network to neurons at another. (2) ``confidence'' of a head which is computed as the average of its maximum attention weight excluding the end of sentence symbol, where the average is taken over tokens in a set of sentences used for evaluation. For pruning the heads, they propose a method based on stochastic gates and a differentiable relaxation of the $L_0$ penalty. $L_0$ norm equals the number of non-zero components and pushes the model to switch off less important heads. They find that only a small subset of heads are important for translation. On the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU.","id":"bc2b6e76-0a87-4b66-8b3c-2bc732d0a821","level":"subsubsection","origin_cites_number":3,"parent_id":"10a33a73-e952-454b-b5ce-09086a71de83","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Heads and Layers"],["subsubsection","Pruning Attention Heads"]],"subsections":[],"title":"Pruning Attention Heads"},{"cite_extract_rate":1,"cites":[4486],"content":"Note that dropping attention heads does not reduce runtime as they are usually computed in parallel. While one can prune weights, neurons or attention heads, how can we design a scheme to prune away layers (see Fig.~\\ref{fig:pruning}(G))? The LayerDrop idea proposed in~ is inspired by DropConnect. DropConnect randomly drops weights while training on a batch. LayerDrop does structured dropout: it drops groups of weights, heads, feed forward network (FFN) matrices, or layers. The layers to be pruned can be decided using one of these ways: \n\\begin{itemize}\n \\item Every Other: Prune every other layer (with rate $p$), e.g., every $3^{rd}$ layer in a 12-layer BERT model. \n \\item Search on Validation: Search for a set of layers to be pruned by checking their impact on a validation set. This entails trying various combinations. \n \\item Data Driven Pruning: Learn the drop rate $p_d$ of each layer in a data driven manner.\n\\end{itemize}\nGiven a target drop rate $p$, we learn an individual drop rate $p_d$ for the layer at depth $d$ such that the average rate over layers is equal to $p$. At inference time, we forward only the fixed top-$k$ highest scoring layers based on the softmax output. Across the three methods, ``Every Other'' strategy works surprisingly well across many tasks and configurations. ``Search on Validation'' and ``Data Driven Pruning'' only offer marginal gains.","id":"c4355812-9f8a-4da4-86ff-8739c39b93e8","level":"subsubsection","origin_cites_number":1,"parent_id":"10a33a73-e952-454b-b5ce-09086a71de83","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Heads and Layers"],["subsubsection","Pruning Layers"]],"subsections":[],"title":"Pruning Layers"},{"cite_extract_rate":0.7142857142857141,"cites":[4331,4479,4487,4478,4486,844,4482,4480,4485,7826],"content":"Lastly, Prasanna et al.~ experiment with pruning both the FFN layers (see Fig.~\\ref{fig:pruning}(F)) as well as attention heads (see Fig.~\\ref{fig:pruning}(E)) in a BERT network. Just like~, they assign a mask variable to each of these structures. To decide which structures to prune, we look at the expected sensitivity of the model to the mask variables. High sensitivity implies large impact on the model output and hence corresponding structures should be retained. They find that it is possible to find a subnetwork of elements that achieves performance comparable with that of the full model, and similarly-sized subnetworks sampled from the rest of the model perform worse. \n\\begin{table}\n \\centering\n \\scriptsize\n \\begin{tabular}{|l|l|l|l|l|l|l|l|l|}\n \\hline\nTask&Dataset&Model&Method&Size (Pruned; Orig)&Eval. (Pruned; Orig)&Metric\\\\\n\\hline\n\\hline\nLanguage modeling&Europarl v7 English &2-layer MLP&Prune Neurons~&5.06M; 5.1M&57; 100&Perplexity (L)\\\\\n\\hline\nLanguage modeling&PTB&2-layer LSTM&Iter. Mag.~&6.6M; 66M&80.24; 78.45&Perplexity (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Iter. Mag.~&20M; 66M&78.5; 78.8&Perplexity (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Block Sparsity~&20M; 66M&83; 78.8&Perplexity (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&BBS~&20M; 66M&78.5; 78.8&Perplexity (L)\\\\\n\\hline\nLanguage modeling&Wikitext-103&Transformer&LayerDrop~&22M; 44M&19.5; 18.2&Perplexity (L)\\\\\n\\hline\nLanguage modeling&AFP from English Gigaword&2-layer MLP&Prune Neurons~&5.07M; 5.1M&107; 100&Perplexity (L)\\\\\n\\hline\nLinguistic acceptability&CoLA&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&82.8/76.3; 81.5&Matthews (H)\\\\\n\\hline\nMachine reading comp.&MRPC&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&88.1/83.5; 89.3&Acc (H)\\\\\n\\hline\nMachine reading comp.&MRPC&BERT-base&LayerDrop~&66M; 110M&85.3; 88.9&Acc (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de) &WMT14&Multi-layer LSTM&Mag.~&43M; 216M&20.91; 20.5&BLEU (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de) &WMT16&4-layer LSTM&Iter. Mag.~&23M; 211M&26.19; 26.77&BLEU (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de) &WMT16 &Transformer&LayerDrop~&22M; 44M&29; 29&BLEU (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de) &WMT17&Transformer&Iter. Mag.~&22M; 44M&26.4; 28.09&BLEU (H)\\\\\n\\hline\nParaphrasing&QQP&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&91.2/85.1; 91.2&Acc (H)\\\\\n\\hline\nQuestion answering&SQuAD 1.1&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&90.23/85.3; 90.9&Acc (H)\\\\\n\\hline\nQuestion answering&SQuAD 2.0&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&81.3/75.3; 81.9&Acc (H)\\\\\n\\hline\nSentiment analysis&SST-2&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&92.4/91.3; 93.2&Acc (H)\\\\\n\\hline\nSentiment analysis&SST-2&BERT-base&LayerDrop~&66M; 110M&92.5; 93.5&Acc (H)\\\\\n\\hline\nSpeech recognition&2100 hours English Speech&2 CONV+7 BiRNNs+CTC&Iter. Mag.~&11.1M; 67M&10.59; 10.67&CER (L) on dev\\\\\n\\hline\nSpeech recognition&2100 hours English Speech&2 CONV+7 BiGRUs+CTC&Iter. Mag.~&17.8M; 115M&9.76; 9.55&CER (L) on dev\\\\\n\\hline\nSpeech recognition&2100 hours English speech&2 CONV+7 BiRNNs+CTC&Block Sparsity~&25.8M; 67M&15.66; 15.36&CER (L) on test\\\\\n\\hline\nSpeech recognition&2100 hours English speech&2 CONV+7 BiRNNs+CTC&Block Sparsity+Group Lasso~&12.9M; 67M&15.89; 15.36&CER (L) on test\\\\\n\\hline\nSpeech recognition&2100 hours English speech&2 CONV+3 BiGRUs+CTC&Block Sparsity~&25.6M; 115M&16.23; 15.42&CER (L) on test\\\\\n\\hline\nSpeech recognition&2100 hours English speech&2 CONV+3 BiGRUs+CTC&Block Sparsity+Group Lasso~&10.8M; 115M&16.78; 15.42&CER (L) on test\\\\\n\\hline\nSpeech recognition&AN4&2 CONV+3 HLSTMs+CTC&Grow and Prune~&2.6M; 44.72M&10.37; 8.92&WER (L)\\\\\n\\hline\nSpeech recognition&Switchboard (swb/fsh)&7-layer MLP&Prune Neurons~&12.2M; 32.19M&25.5/28.8; 25.7/28.8&WER (L)\\\\\n\\hline\nSpeech recognition&TIMIT&5-layer MLP&Prune Neurons~&3.5M; 5.76M&20.7; 20.79&PER (L)\\\\\n\\hline\nSpeech recognition&TIMIT&LSTM&Iter. Mag.~&0.32M; 3.2M&23.5; 23.5&PER (L)\\\\\n\\hline\nSpeech recognition&TIMIT&LSTM&Block Sparsity~&0.32M; 3.2M&26.5; 23.5&PER (L)\\\\\n\\hline\nSpeech recognition&TIMIT&LSTM&BBS~&0.32M; 3.2M&23.5; 23.5&PER (L)\\\\\n\\hline\nSpeech recognition&WSJ 92&1 CONV+3 FC+1 BiRNN&DSD~&4.07M; 8.14M&27.9; 27.45&WER (L)\\\\\n\\hline\nSpeech recognition&WSJ 92&2 CONV+7 BiRNNs+CTC&DSD~&33.5M; 67M&10.65; 9.02&WER (L)\\\\\n\\hline\nSpeech recognition&WSJ 93&1 CONV+3 FC+1 BiRNN&DSD~&4.07M; 8.14M&32.99; 31.6&WER (L)\\\\\n\\hline\nSpeech recognition&WSJ 93&2 CONV+7 BiRNNs+CTC&DSD~&33.5M; 67M&14.84; 13.44&WER (L)\\\\\n\\hline\nSummarization&CNN-Dailymail&Transformer&LayerDrop~&22M; 44M&39; 40&ROUGE (H)\\\\\n\\hline\nTextual entailment&MNLI&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&86.1/77; 86.1&Acc (H)\\\\\n\\hline\nTextual entailment&MNLI-m&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&85.7/82.5; 85.9&Acc (H)\\\\\n\\hline\nTextual entailment&MNLI-m&BERT-base&LayerDrop~&66M; 110M&82.9; 84.6&Acc (H)\\\\\n\\hline\nTextual entailment&QNLI&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&92.3/90.2; 91.3&Acc (H)\\\\\n\\hline\nTextual entailment&QNLI&BERT-base&LayerDrop~&66M; 110M&89.4; 90.5&Acc (H)\\\\\n\\hline\nTextual entailment&RTE&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&70.1/68.6; 70.1&Acc (H)\\\\\n\\hline\n \\end{tabular}\n \\caption{Comparison of various pruning methods (sorted by Task and then Dataset). CONV=Convolution. CTC=Connectionist temporal classification. FC=Fully connected. HLSTM=hidden-layer LSTM~. In the metric column, H means high is better while L means low is better. PER/CER/WER=Phone/Character/Word error rate. For~, embedding weights have not been considered when computing model size in the table. Block sparsity methods use block size of 4x4. BBS uses 64 banks.}\n \\label{tab:pruningSummary}\n\\end{table}","id":"7dc1a143-838c-41ca-90a0-ebe43039cc74","level":"subsubsection","origin_cites_number":14,"parent_id":"10a33a73-e952-454b-b5ce-09086a71de83","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Pruning Heads and Layers"],["subsubsection","Pruning General Structures"]],"subsections":[],"title":"Pruning General Structures"},{"cite_extract_rate":0.75,"cites":[4331,4484,4479,4478,4486,844,1568,4485,7826],"content":"Table~\\ref{tab:pruningSummary} compares various pruning methods across different tasks and datasets. Size and accuracy of both the original and the pruned model are shown. The papers report multiple (model size, model accuracy) pairs but we carefully chose the pair such that accuracy is typically within 10\\% of the original or the best pruned model accuracy reported. For language modeling, most popular datasets include PTB, Europarl v7 English, Wikitext-103 and AFP from English Gigaword. On PTB using LSTMs, we observe that Bank-Balanced Sparsity method~ leads to lowest perplexity. For the Linguistic Acceptability CoLA task, RPP~ resulted into a smaller and a more accurate model compared to iterative magnitude pruning. As expected in some senses, the pruned model provides better accuracy than the unpruned one since pruning causes regularization. For the machine reading comprehension, question answering and paraphrasing tasks also, RPP seems to work better than iterative magnitude pruning. For the machine translation (NMT) task, on English-German WMT datasets, pruned Transformer models provide better accuracy than pruned LSTM with comparable number of parameters. For the sentiment analysis task, on SST-2, although RPP leads to a better pruned model compared to iterative pruning, LayerDrop~ improves further on it with a model less than half of the RPP-pruned model. For the speech recognition task, experiments have been reported on 2100 hours English speech data, TIMIT, WSJ, Switchboard and AN4 datasets. On 2100 hours English speech data, Block Sparsity+Group Lasso is better than Block sparsity without regularization. Also, it is better than plain iterative magnitude pruning. On TIMIT, block sparsity~ leads to a more accurate 90\\% pruned LSTM model compared to the unpruned one. For the summarization task, LayerDrop~ can compress the model to half without any noticeable accuracy change. Finally, for the textual entailment task, experiments have been done on GLUE~ datasets: MNLI, MNLI-m, QNLI and RTE. Models pruned from BERT-large perform better than models pruned from BERT-base; RPP performs better than iterative magnitude pruning.\nWhile older methods~ claimed that Hessian based methods were more effective than magnitude based pruning, almost all recent methods have been based on magnitude based pruning. See et al.~ proposed three pruning schemes. They make the following observations: (1) Class-blind pruning outperforms both other schemes. Further, the overall performance loss is caused disproportionately by a few classes: softmax weights, source and target embedding weights. (2) It seems that higher layers are more important than lower layers, and that attention and softmax weights are crucial in LSTMs. (3) After retraining the pruned NMT models, baseline performance (20.48 BLEU) is both recovered and improved upon, up to 80\\% pruning (20.91 BLEU), with only a small performance loss at 90\\% pruning (20.13 BLEU). (4) In LSTMs, the parameters corresponding to the less common words are more dispensable. Weights connecting to the input are most crucial, followed by the input gate, then the output gate, then the forget gate. This is particularly true of the lower layers, which focus primarily on the input. However for higher layers, especially on the target side, weights connecting to the gates are as important as those connecting to the input. \nNarang et al.~ observe that for approximately same number of parameters, gradual/iterative pruning is 7\\% to 9\\% better than hard pruning. They also conclude that the initial layers are pruned more aggressively compared to the final layers. Zhu et al.~ advise that in order to get the best-performing sparse model of a certain size, we should train a dense model that is 5x-10x larger and then prune to the desired number of parameters rather than taking the largest and best-performing dense model and pruning this model by 20x or more to the desired number of parameters. Guo et al.~ find that RPP is much better than typical iterative pruning. In their experiments with BERT they find that for both original BERT and BERT pruned with RPP, the low-dimensional manifolds of the language representations are similar, showing the similar projection. This implies that the BERT applied with RPP keeps most of the language representation information similar to that from the original BERT.\nFor block pruning, Narang et al.~ make the following observations: (1) We can create block-sparse RNNs with sparsity ranging from 80\\% to 90\\% with small loss in accuracy. This allows us to reduce the model size by roughly 10×. Block sparsity works with a variety of block sizes up to 32×32. (2) For block size 4×4, models with sparsity greater 90\\% yield a relative accuracy loss of 30\\% or higher. Similarly, for blocks of 16×16, models with sparsity greater than 86\\% have 30\\% or more accuracy loss. A similar trend is observed for block size 32×32. This indicates that there is a tradeoff between sparsity, block size and accuracy of the model. (3) For both block pruning and weight pruning, we see that the initial layers are pruned more aggressively compared to the final layers. Increasing sparsity in the layers closer to the output results in poor accuracy. Additionally, the variance in sparsity across the layers increases with the block size. Further, Cao et al.~ make the following observations comparing block sparsity with BBS: (1) BBS achieves almost the same model accuracy regardless of the change of bank size. For block sparsity, however, increasing the block size adversely affects model accuracy. \nFor pruning of attention heads, Michel et al.~ observe that one can prune up to 20\\% and 40\\% of heads from 6-layer NMT Transformer and BERT resp., without incurring any noticeable negative impact. When trying to remove a head at a time, only 8 (out of 96) heads in 6-layer NMT Transformer (16 heads/layer) cause a statistically significant change in performance when they are removed from the model, half of which actually result in a higher BLEU score. Further Voita et al.~ find that on the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU. \nOverall, to summarize, pruning has been the most popular method for model compression. Pruning methods can be unstructured (prune weights) or structured (prune neurons, blocks, attention heads, layers). While weight pruning theoretically leads to pruning to a large extent, practical implementation of sparse data structures is difficult. Pruning and regularization need to be done together carefully. Also, it is critical to define the importance functions for various structures carefully. Among weight pruning methods, while iterative magnitude pruning with regularization works well for RNNs and LSTMs, RPP performs better for Transformer based models. Pruning blocks using BBS is better than pruning neurons. For Transformer models, pruning just the heads do not provide much model compression, but dropping a combination of attention heads and layers is better.","id":"80f88ea7-37a5-4100-9fe4-121a616869c9","level":"subsection","origin_cites_number":12,"parent_id":"92ac6101-b828-4042-8708-2e0386e701bb","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Pruning"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:quantization}\nWhile pruning saves on the model size by removing weights, quantization aims to reduce the number of bits needed to store weights. Most computer architectures use 32 bits to represent weights. However, estimated precision of the brain (hippocampal spine) synapses is around 4.6 bits~. Empirical evidence suggests that most quantities in the nervous system (for instance, firing of the neurons) have variability of a few percent due to biological noise, or a precision of 1 in 100 at best~. Thus, each decision could depend on $\\log_2 (100)$=6.64 bits. Thus, we should be able to store weights in our artificial neural networks on average in a space of 4--7 bits. Given this motivation, various methods have been proposed which perform 1-bit (or binary quantization), ternary quantization, and general quantization exploring the spectrum between 3 and 32 bits. We discuss such methods in this section. Fig.~\\ref{fig:quantization} provides a broad overview of various quantization styles.\n\\begin{figure}\n \\centering\n \\includegraphics[width=\\columnwidth]{quantization}\n \\caption{Different Types of Quantization: Binary (A), Ternary (B) and General Quantized (C and D). Note that X axis denotes the weight value while the Y axis denotes frequency.}\n \\label{fig:quantization}\n\\end{figure}","id":"c0c0e7e6-55dc-426b-8669-a8adfb8fe243","level":"section","origin_cites_number":2,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"]],"subsections":["11228467-1126-4614-9475-560634f98a9b","60a24ca4-58df-4063-9d44-f538e8a26b98","be929a80-0b3b-4005-bf19-e84fee8ab111","97e94011-bad6-4d07-98cb-26de46e32fe4"],"title":"Quantization"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:binQuant}\nBinarized networks use binary quantization (see Fig.~\\ref{fig:quantization}(A)), which quantizes weights using 1 bit. Quantizing weights to 1 bit provides a compression of 32x but leads to a significant drop in accuracy across many tasks. However, in a hybrid quantization scheme, such binary quantization can be very helpful for some layers in a network. Binarization can be done using deterministic methods or could be stochastic in nature. Also, while na\\\"ive binarization has a very simple way of fixing the binary boundary threshold, one could perform a complex loss aware binarization as well. We discuss these variants of binarization in this section.","id":"11228467-1126-4614-9475-560634f98a9b","level":"subsection","origin_cites_number":0,"parent_id":"c0c0e7e6-55dc-426b-8669-a8adfb8fe243","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Binarized Networks"]],"subsections":["f998a362-cba7-42b7-aa49-83d293cf4b58","3c274eec-1522-4b1c-b25f-1f3d927c3167","bb389108-4900-443c-bee1-d687f6265cf4"],"title":"Binarized Networks"},{"cite_extract_rate":0.33333333333333304,"cites":[4351],"content":"Simplest way of binary quantization is to set the weight as 1 for non-negative weights, and to -1 for negative weights. This leads to 32x compression. Also, the matrix multiplication for binary matrices is $\\sim$7x faster~ leading to faster model inference. In the forward pass, binary networks drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which leads to great increases of power efficiency. Also, in the simplest version, binarization can be performed in a static manner, i.e., after the training is done. However, this method leads to large loss in accuracy. \nA variant of this simple method is to set the weight to a constant $c_1$ for non-negative weights, and to another constant $c_2$ for negative weights. Binary Scheme (BS)-Fixed method~ stores the original weights and during the forward pass replaces the values with a masked value of $c_1$ or $c_2$, where $c_1$ and $c_2$ are fixed and chosen with hyperparameter tuning. Full precision weights are used during training. At the end of training, the weights are replaced with the index of its masked value. Choosing the values of $c_1$ and $c_2$ can be difficult and time-consuming in BS-Fixed. Thus, in the BS-flexible method~, we initialize $c_1$ and $c_2$ using KMeans with two centroids over the weights, and then update $c_1$ and $c_2$ using back-propagation. Also, in the BS-Flexible method, weights are quantized as follows.\n\\begin{equation}\n w_b= \\begin{cases}\nc_1 &\\text{if } w \\geq (c_1+c_2)/2\\\\\nc_2 &\\text{if } w < (c_1+c_2)/2\n\\end{cases}\n\\end{equation}\nNote that $w$ is the original weight value while $w_b$ is the binarized weight value. These changes eliminate the need for hyper-parameter tuning.","id":"f998a362-cba7-42b7-aa49-83d293cf4b58","level":"subsubsection","origin_cites_number":3,"parent_id":"11228467-1126-4614-9475-560634f98a9b","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Binarized Networks"],["subsubsection","Deterministic Binarization"]],"subsections":[],"title":"Deterministic Binarization"},{"cite_extract_rate":1,"cites":[685,4488,8375,2670],"content":"Stochastic~ binarization is performed as follows. \n\\begin{equation}\n w_b= \\begin{cases}\n+1 &\\text{with probability } p=\\sigma(w)\\\\\n-1 &\\text{with probability } 1-p\n\\end{cases}\n\\end{equation}\n\\noindent where\n\\begin{eqnarray}\n \\sigma(w)=\\text{clip}\\left(\\frac{w+1}{2},0,1\\right)=\\max\\left(0, \\min\\left(1, \\frac{w+1}{2}\\right)\\right)\n\\end{eqnarray}\nWe only binarize the weights during the forward and backward propagations but not during the parameter update. Keeping good precision weights during the updates is necessary for Stochastic Gradient Descent (SGD). This is possible using something called as ``Straight Through Estimator (STE) trick''~. As per STE, as the quantized value is an approximation of the original value, we can substitute the gradient with respect to the quantized value for the gradient of original value. The trick allows the inclusion of quantization into the computation graph of back-propagation and allows QNNs to represent parameters, activations and gradients with low bitwidth numbers. For test-time inference, there are three options using such a quantization method: \n\\begin{itemize}\n \\item Use the resulting binary weights $w_b$ (this makes most sense with the deterministic binarization). \n \\item In the stochastic case, many different networks can be sampled by sampling a $w_b$ for each weight. The ensemble output of these networks can then be obtained by averaging the outputs from individual networks. \n \\item Use original weights. But this does not reduce model size. \n\\end{itemize}\nBesides this, there have been further efforts that make train/test faster but do not reduce model size. For example, Lin et al.~ convert multiplications in the backward pass into bit-shifts by restricting the activations to be power-of-two integers. Hubara et al.~ binarize weights and activations, at the inference phase and the entire training phase of a deep network.","id":"3c274eec-1522-4b1c-b25f-1f3d927c3167","level":"subsubsection","origin_cites_number":4,"parent_id":"11228467-1126-4614-9475-560634f98a9b","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Binarized Networks"],["subsubsection","Stochastic Binarization"]],"subsections":[],"title":"Stochastic Binarization"},{"cite_extract_rate":1,"cites":[4489,4490,851],"content":"\\label{subsubsec:lossAwareBin}\nThe na\\\"ive binary quantization methods divide the real number line into two parts and each part was mapped to a quantized weight value. Can we decide per weight value which of the two weights it should be quantized to? Thus the idea behind Binary Weight Networks (BWN)~ is to approximate the weight vector $W\\in R^n$ using a binary vector $B\\in\\{+1,-1\\}^n$ and a scaling factor $\\alpha\\in R^+$ such that $W\\approx\\alpha B$. This can be expressed as an optimization problem as follows.\n\\begin{eqnarray}\n \\alpha^*, B^*=\\argmin_{\\alpha, B}||W-\\alpha B||^2\n\\end{eqnarray}\nWe can expand and write this as follows.\n\\begin{eqnarray}\n ||W-\\alpha B||^2=\\alpha^2B^TB-2\\alpha W^TB+W^TW\n\\end{eqnarray}\nSince $B\\in{+1,-1}^n$, $B^TB=n$. Also $W^TW$ is a constant. Thus $B^*=\\argmax_B W^B$ such that $B\\in{+1,-1}^n$. This optimization can be solved by simply assigning $B_i=+1$ when $W_i\\geq 0$, and $B_i=-1$ otherwise. To compute $\\alpha^*$, we set the derivative of $||W-\\alpha B||^2$ wrt $\\alpha$ to 0 and get the solution as follows.\n\\begin{eqnarray}\n \\alpha^*=\\frac{\\sum|W_i|}{n}\n\\end{eqnarray}\nThus, besides the binarized weight matrix, a scaling parameter is also learned in BWN.\nTo take this idea further, can we learn $\\alpha$ and $B$ to minimize the overall network's loss function? Thus, now, the Weight binarization can be formulated as the following optimization problem.\n\\begin{eqnarray}\n &&\\min_{\\hat{w}} \\text{loss}(\\hat{w})\\\\\n &&\\text{such that } \\hat{w}_l=\\alpha_l b_l; \\alpha_l>0; b_l\\in\\{+1,-1\\}^{n_l}; l=1,..., L\n\\end{eqnarray}\n\\noindent where $L$ is the number of layers, $n_l$ is the number of weights in layer $l$. This loss aware binarization~ problem can be solved using proximal Newton algorithm~ to find the best $\\alpha_l$ and $B_l$.","id":"bb389108-4900-443c-bee1-d687f6265cf4","level":"subsubsection","origin_cites_number":3,"parent_id":"11228467-1126-4614-9475-560634f98a9b","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Binarized Networks"],["subsubsection","Loss Aware Binarization (LAB)"]],"subsections":[],"title":"Loss Aware Binarization (LAB)"},{"cite_extract_rate":1,"cites":[4491],"content":"\\label{subsec:ternaryQuant}\nUnfortunately, binary quantization of the recurrent weights in RNNs/LSTMs never worked~. When the true value of a weight is near zero, its quantized value is either set to -1 or 1. This results into an artificial increase in the magnitude of the weights and the vanishing/exploding gradients problem becomes more severe. Hence, another popular form of quantization is ternary quantization (see Fig.~\\ref{fig:quantization}(B)). Ternary quantization can help achieve a min of 16x compression (up to 32x compression if hardware allows to avoid storing zeros). In this section, we discuss different variants of ternary quantization from the simplest ternary connect networks to hybrid ternary networks like HitNets.","id":"60a24ca4-58df-4063-9d44-f538e8a26b98","level":"subsection","origin_cites_number":1,"parent_id":"c0c0e7e6-55dc-426b-8669-a8adfb8fe243","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Ternarized Networks"]],"subsections":["8b4d9559-79ce-43e9-8312-ccc718a576a7","05681eea-3f45-4e85-b114-d5a86ce4f547","3a480858-aec9-47ce-85a6-135a6bfa320a"],"title":"Ternarized Networks"},{"cite_extract_rate":1,"cites":[4492,4488],"content":"The simplest method for ternary quantization is ternary connect~ whose deterministic form is as follows.\n\\begin{equation}\n w_t= \\begin{cases}\n+1 &\\text{if } w > 0.5\\\\\n0 &\\text{if } -0.5 < w \\leq 0.5\\\\\n-1 &\\text{if } w \\leq -0.5\\\\\n\\end{cases}\n\\end{equation}\nNote that $w$ is the original weight value while $w_t$ is the ternarized weight value. Like binary connect, ternary connect also eliminates all multiplications in the forward pass. In the stochastic form, assuming original weights have been normalized to be in the range [-1,1], ternary quantization is done as follows.\n\\begin{equation}\n w_t= \\begin{cases}\n+1 &\\text{with prob } w \\text{ if } w\\in (0,1]\\\\\n0 &\\text{with prob } 1-w \\text{ if } w\\in (0,1]\\\\\n0 &\\text{with prob } 1+w \\text{ if } w\\in[-1,0]\\\\\n-1 &\\text{with prob } -w \\text{ if } w \\in [-1,0]\\\\\n\\end{cases}\n\\end{equation}\nA slightly related way called as Bernoulli Ternary Quantization where $w_t$ is set to +1 (or -1) with prob $p$ if $w>0$ (or) $<0$, and set to 0 with prob 1-p where $p\\sim$Bernoulli($|x|$). Yet another way to set the boundaries for the three ranges is to use Gaussian based ternary weights~ as follows.\n\\begin{equation}\n w_t= \\begin{cases}\n+1 &\\text{if } w > -(\\mu+\\sigma/2)\\\\\n0 &\\text{if } -(\\mu+\\sigma/2)<w\\leq (\\mu+\\sigma/2)\\\\\n-1 &\\text{if } w \\leq -(\\mu+\\sigma/2)\\\\\n\\end{cases}\n\\end{equation}\n\\noindent where $\\mu$ and $\\sigma$ are the mean and standard deviation of the weight matrix being quantized.","id":"8b4d9559-79ce-43e9-8312-ccc718a576a7","level":"subsubsection","origin_cites_number":2,"parent_id":"60a24ca4-58df-4063-9d44-f538e8a26b98","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Ternarized Networks"],["subsubsection","Ternary Weight Networks"]],"subsections":[],"title":"Ternary Weight Networks"},{"cite_extract_rate":0.6000000000000001,"cites":[852,4493,851],"content":"Rather than using the rules for ternary quantization as mentioned above, one can learn the boundary ranges or the quantized values for individual weights. One way of learning the right ternary representation per weight value is to minimize the Euclidean distance between full precision weights $W$ and the ternary weights $T$ along with a scaling factor~. This can be expressed as the following optimization problem.\n\\begin{eqnarray}\n\\alpha^*, T^*=\\argmin_{\\alpha,T}||W-\\alpha T||_2^2\\\\\n\\text{such that } \\alpha\\geq 0; T_i\\in \\{-1,0,1\\}; i=1,2,..., n\n\\end{eqnarray}\nNote that this is equivalent to the BWN method~. This does not lead to a closed form solution. Hence, we approximate the solution with threshold-based ternary function. \n\\begin{equation}\n w_t= \\begin{cases}\n+1 &\\text{if } w > \\Delta\\\\\n0 &\\text{if } -\\Delta < w \\leq \\Delta\\\\\n-1 &\\text{if } w \\leq -\\Delta\\\\\n\\end{cases}\n\\label{eq:eq2}\n\\end{equation}\nThe approximation works when we set $\\Delta$ as follows.\n\\begin{eqnarray}\n\\Delta^*=\\argmax_{\\Delta>0}\\frac{1}{|I_\\Delta|}\\left(\\sum_{i\\in I_\\Delta}|W_i|\\right)^2\n\\end{eqnarray}\n\\noindent where $I_\\Delta$ is the number of weights with magnitude$>\\Delta$. Again, this has no straightforward solution, unless we assume that original weights $W_i$'s are generated from uniform or normal distribution. When $W_i$'s are uniformly distributed in $[-a, a]$ and $\\Delta$ lies in $(0, a]$, the approximated $\\Delta^*$ is $a/3$, which equals to $\\frac{2}{3}E(W)$. When $W_i$'s are generated from normal distributions $N(0,\\sigma^2)$, the approximated $\\Delta^*$ is 0.6$\\sigma$ which equals to 0.75$E(|W|)$. Thus, we can use the following rule of thumb for fast and easy computation.\n\\begin{eqnarray}\n\\Delta^*\\approx 0.7 E(W)=\\frac{0.7}{n}\\sum_{i=1}^n |W_i|\n\\end{eqnarray}\nAnother way to learn the quantization step size $\\Delta$ in Eq.~\\ref{eq:eq2} is to learn in a loss-aware manner~, i.e., tuning it to minimize the overall network loss. Given a multi-layered network, we need to perform such quantization layer by layer in a greedy manner. We first train the network with full precision weights. We quantize all input data and signals of hidden layers. Next, we start with the weight quantizer between the input layer and the first hidden layer, try several step sizes around the initial step size and measure the output error of the network with the training set. The initial step size is determined using Lloyd-Max algorithm~. Choose the step size that minimizes the output error and quantize the weights. Further, we perform these steps for the next few layers until the output layer. Finally, the quantized neural network is retrained.\nYet another way of training ternary quantization~ is to quantize weights to one of {$-W_l^n$, 0, $W_l^p$} for each layer $l$, where $W_l^n$ and $W_l^p$ are trainable parameters, learned using back-propagation. First, we normalize the full-precision weights to the range [-1, +1] by dividing each weight by the maximum weight. During SGD, we back propagate the gradient to both $W_l^n$ and $W_l^p$ and to the latent full-precision weights. This makes it possible to adjust the ternary assignment (i.e. which of the three values a weight is assigned). To decide the quantization step size $\\Delta_l$ for a layer $l$, two heuristics can be used: (1) set $\\Delta_l=t\\times \\max(|w_l|)$ where $t$ is a constant and $w_l$ are the full precision weights in layer $l$. (2) maintain a constant sparsity $r$ for all layers throughout training. By adjusting the hyper-parameter $r$ we can obtain ternary weight networks with various sparsities.","id":"05681eea-3f45-4e85-b114-d5a86ce4f547","level":"subsubsection","origin_cites_number":5,"parent_id":"60a24ca4-58df-4063-9d44-f538e8a26b98","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Ternarized Networks"],["subsubsection","Trained Ternary Quantization"]],"subsections":[],"title":"Trained Ternary Quantization"},{"cite_extract_rate":0,"cites":[],"content":"Given various ternary quantization methods proposed so far, one can combine them and use different methods for different layers. Wang et al.~ found that threshold ternary quantization (TTQ) (Eq.~\\ref{eq:eq2}) is preferable for weights in an RNN while Bernoulli Ternary Quantization (BTQ) is preferable for activations. This is based on the observation that in an RNN, the distribution of weights follows normal distribution (with different ranges across different weight matrices), while for activations, the range is [0,1] and most of the values are located near to the two poles instead of the middle of the range. In the training phase (where we need to store the full precision weights), ternary quantization of weights only saves 1.4x memory consumption but quantizing both weights and activations can achieve up to 16x memory savings.\nThe HitNet architecture~ with this hybrid ternary quantization can be defined using these equations, where $i_t, f_t, o_t$ are the input, forget and output gates; $x_t$ is input at time $t$; $c_t$ is the cell output; and $h_t$ is the hidden layer output; $W_x$, $W_h$, $b_x$, $b_h$ are weights. \n\\begin{eqnarray}\n \\nonumber i_t,f_t,g_t,o_t&=&\\sigma(\\text{TTQ}(W_x)x_t+\\text{TTQ}(b_x)\\\\\n \\nonumber &+&\\text{TTQ}(W_h)h_{t-1}+\\text{TTQ}(b_h))\\\\\n \\nonumber c_t&=&f_t\\times c_{t-1}+i_t\\times g_t\\\\\n h_t&=&\\text{BTQ}(o_t\\times \\sigma(c_t))\n\\end{eqnarray}","id":"3a480858-aec9-47ce-85a6-135a6bfa320a","level":"subsubsection","origin_cites_number":1,"parent_id":"60a24ca4-58df-4063-9d44-f538e8a26b98","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Ternarized Networks"],["subsubsection","Hybrid Ternary Quantization"]],"subsections":[],"title":"Hybrid Ternary Quantization"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:generalQuant}\nSo far we discussed methods designed specifically for binary and ternary quantization. Now, we discuss general $k$-bit quantization methods. We will discuss (1) uniform quantization methods which perform equal width binning, (2) non-uniform methods which are closer to equal frequency binning, (3) loss-aware quantization methods, and (4) methods specifically designed for Transformer models.","id":"be929a80-0b3b-4005-bf19-e84fee8ab111","level":"subsection","origin_cites_number":0,"parent_id":"c0c0e7e6-55dc-426b-8669-a8adfb8fe243","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","General Quantized Networks"]],"subsections":["f96f7375-a19b-4b18-9a52-99c14e0b21ea","6c78c9b7-5b2e-4d63-90b9-bd800235109c","4b1349e1-295c-4789-9acd-ec92a1cf9f02","96bae440-819d-4a68-8df5-216ce70bd5a8","c55e72b6-a137-4167-a584-9f7c420c0df0"],"title":"General Quantized Networks"},{"cite_extract_rate":1,"cites":[4491,8797,4351,851],"content":"Uniform $k$-bit Quantization simply splits the range of original weights into $2^k-1$ equal size intervals~. Refer Fig.~\\ref{fig:quantization}(C).\nIf original weights are in range [-1,1], they can be quantized as follows.\n\\begin{equation}\n q_k(x)=2\\left(\\frac{\\text{round}[(2^k-1)\\left(\\frac{x+1}{2}\\right)]}{2^k-1}-\\frac{1}{2}\\right)\n \\label{eq:uq}\n\\end{equation}\nSimilarly, if entries are in range [0,1], we could use the following formula.\n\\begin{eqnarray}\n q_k(x)=\\frac{1}{2^k-1}\\lfloor(2^k-1)x+\\frac{1}{2}\\rfloor\n\\end{eqnarray}\nWhen the weights in matrix $X$ are not in the range [0,1], we can first scale weights as follows.\n\\begin{eqnarray}\n \\tilde{X}=\\frac{X-\\beta}{\\alpha}\n\\end{eqnarray}\n\\noindent where $\\alpha=\\max(X)-\\min(X)$ and $\\beta=\\min(X)$. After quantization, we can apply a reverse transform to approximate the original values. Overall, the quantized result can be written as follows.\n\\begin{eqnarray}\n Q(X)=\\alpha q_k(\\tilde{X})+\\beta\n\\end{eqnarray}\nGiven any quantization function $q_k(x)$, one can use it for quantizing weight matrices of various recurrent models like RNNs, GRUs and LSTMs~. Typical inference equations for a GRU can be written as follows.\n\\begin{eqnarray}\n z_t=\\sigma(W_z.[h_{t-1},x_t]); r_t=\\sigma(W_r.[h_{t-1},x_t])\\\\\n \\tilde{h_t}=\\text{tanh}(W.[r_t\\times h_{t-1},x_t]); h_t=(1-z_t)h_{t-1}+z_t \\tilde{h_t}\n\\end{eqnarray}\nBesides the matrix multiplications needed to compute $z_t$, $r_t$ and $\\tilde{h_t}$, the gate structure of $\\tilde{h_t}$ and $h_t$ brings in the need for element-wise multiplication. \nAs $\\tilde{h_t}$ and $h_t$ are also the inputs to computations at the next timestamp, and noting that a quantized value multiplied by a quantized value will have a larger bit-width, we need to insert additional quantization steps after element-wise multiplications.\nAnother problem with quantization of GRU structure lies in the different value range of gates. The range of tanh is [-1, 1], which is different from the value range [0, 1] of $z_t$ and $r_t$. Keeping in mind these observations, the equations for a quantized GRU can be written as follows, after the weights $W_z$, $W_r$ and $W$ and input $x_t$ have already been quantized to [-1,1].\n\\begin{eqnarray}\n z_t&=&\\sigma(W_z.[h_{t-1},x_t])\\\\\n r_t&=&\\sigma(W_r.[h_{t-1},x_t])\\\\\n \\tilde{h_t}&=&tanh\\left(W.\\left[2q_k\\left(\\frac{1}{2}(r_t h_{t-1})+\\frac{1}{2}\\right)-1,x_t\\right]\\right)\\\\\n h_t&=&2q_k\\left(\\frac{1}{2}((1-z_t) h_{t-1}+z_t \\tilde{h_t})+\\frac{1}{2}\\right)-1\n\\end{eqnarray}\nFollowing a similar method, we can also quantize LSTM networks.","id":"f96f7375-a19b-4b18-9a52-99c14e0b21ea","level":"subsubsection","origin_cites_number":4,"parent_id":"be929a80-0b3b-4005-bf19-e84fee8ab111","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","General Quantized Networks"],["subsubsection","Uniform Quantization"]],"subsections":[],"title":"Uniform Quantization"},{"cite_extract_rate":1,"cites":[4494,4491],"content":"Uniform quantization is easy to implement but far from optimum when quantizing non-uniform data, which is believed to be the trained weights and activations of deep neural network. One way of performing non-uniform quantization is exponential quantization~. It quantizes the weight values to an integer power of 2. If we let \n\\begin{eqnarray}\n p=\\frac{|W|}{2^{\\lfloor \\log_2|W|\\rfloor}}-1\n\\end{eqnarray}\ndeterministic exponential quantization can be written as follows.\n\\begin{equation}\n \\log_2 W_q= \\begin{cases}\n\\lceil \\log_2|W| \\rceil &\\text{if } p>0.5\\\\\n\\lfloor \\log_2|W| \\rfloor &\\text{otherwise }\n\\end{cases}\n\\end{equation}\nSimilarly, stochastic exponential quantization can be written as follows.\n\\begin{equation}\n \\log_2 W_q= \\begin{cases}\n\\lceil \\log_2|W| \\rceil &\\text{with prob } p\\\\\n\\lfloor \\log_2|W| \\rfloor &\\text{with prob } 1-p\n\\end{cases}\n\\end{equation}\nExponential quantization enables storing weights in low precision and eliminating multiplications. However, it still does not perform quantization in a way which is sensitive to the distribution of the weights. Distributions of parameters in neural networks are often imbalanced, such that the uniform quantization determined from extremal values may under utilize available bitwidth. When we quantize values, it may be desirable to make the quantized values have balanced distributions, to take full advantage of the available parameter space. Balanced quantization method~ starts by partitioning numbers into $2^{k}$ bins containing roughly the same number of entries (percentiles). Refer Fig.~\\ref{fig:quantization}(D). Each partition is then mapped to an evenly-divided interval in the closed interval [0, 1]. Finally, the quantization step maps intervals into discrete values using Eq.~\\ref{eq:uq} and transforms the value range to be approximately the same as input.\nA na\\\"ive implementation using percentiles as thresholds would require sorting of weight values during each forward operation in back-propagation, which may slow down the training process. The $2^{k}$ evenly spaced percentiles required in histogram equalization can be computed from the recursive application of partitioning of numbers by medians. Further, the mean $\\mu$ can be used to approximate the median $m$. Thus, we can perform approximate histogram equalization without doing sorting.","id":"6c78c9b7-5b2e-4d63-90b9-bd800235109c","level":"subsubsection","origin_cites_number":2,"parent_id":"be929a80-0b3b-4005-bf19-e84fee8ab111","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","General Quantized Networks"],["subsubsection","Balanced Quantization"]],"subsections":[],"title":"Balanced Quantization"},{"cite_extract_rate":0.6666666666666661,"cites":[4496,4495],"content":"Yet another way of performing non-uniform quantization is to decide bin boundaries using clustering in a static manner. In this static-KMeans method~, We first train the neural network with full-precision parameters. Then apply KMeans to the weights. After clustering, the value of each pixel is set to the value of the center of the cluster it belongs to. We also need to store mapping from integers to cluster centers. Given $k$ clusters, we only need $\\log (k)$ bits to code the clusters.\nA better approach is to perform KMeans clustering during training. In this method~, multiple connections (belonging to the same cluster) share the same weight, and we fine-tune those shared weights. For the forward pass, the cluster index stored for each connection is mapped to a centroid which is then used as the weight. For back-propagation, during update, all the gradients are grouped by the cluster index and summed together, multiplied by the learning rate and subtracted from the shared centroids from last iteration. We use KMeans clustering to identify the shared weights for each layer of a trained network, so that all the weights that fall into the same cluster will share the same weight. Weights are not shared across layers. To calculate the compression rate, given $k$ clusters, we only need $\\log_2 k$ bits to encode the index. In general, for a network with $n$ connections and each connection is represented with $b$ bits, constraining the connections to have only $k$ shared weights will result in a compression rate of\n\\begin{eqnarray}\n r=\\frac{nb}{n\\log_2 k+kb} \n\\end{eqnarray}\nThere are two other ways of using KMeans for non-uniform quantization: Product Quantization (PQ) and Residual Quantization (RQ)~. In product quantization (PQ), we partition the vector space into many disjoint subspaces, and perform quantization (KMeans) in each subspace. Weight matrix $W$ is partitioned columnwise: $W=[W^1, W^2, ..., W^s]$ where $W^i\\in R^{m\\times n/s}$ assuming $n$ is divisible by $s$. Then we perform KMeans on each submatrix $W^i$ to obtain clusters $c^i_1, ..., c^i_k$. Thus, we get $s$ codebooks. The reconstructed matrix is $\\hat{W}=[\\hat{W}^1, \\hat{W}^2, ..., \\hat{W}^s]$ where $\\hat{W}^i_j$ is the closest centroid $c^i_j$. PQ can be applied to either the x-axis or the y-axis of the matrix. We need to store the cluster indexes and codebooks for each subvector. The compression rate for this method can be written as follows.\n\\begin{eqnarray}\n r=\\frac{32mn}{32kn + log_2(k)ms} \n\\end{eqnarray}\nResidual quantization (RQ) is similar. In RQ, we first quantize the vectors into k-centers. Next we find out the residuals for each data point ($w-c$) and perform KMeans on the residuals. Do it recursively $t$ times.\nThen the resultant weight vectors are calculated as $\\hat{W}_z=c^1_j+c^2_j+...+c^t_j$ given we have recursively performed $t$ iterations. We need to store all the codebooks for each iteration, which potentially needs large amount of memory. The compression rate for this method can be written as follows.\n\\begin{eqnarray}\n r=\\frac{m}{tk + log_2(k)tn} \n\\end{eqnarray}","id":"4b1349e1-295c-4789-9acd-ec92a1cf9f02","level":"subsubsection","origin_cites_number":3,"parent_id":"be929a80-0b3b-4005-bf19-e84fee8ab111","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","General Quantized Networks"],["subsubsection","KMeans based Quantization Schemes"]],"subsections":[],"title":"KMeans based Quantization Schemes"},{"cite_extract_rate":1,"cites":[4498,4497,851],"content":"Generalizing the loss aware binarization approach (Sec.~\\ref{subsubsec:lossAwareBin})~, we can perform $k$-bit quantization~ by attempting to solve the following problem. \n\\begin{eqnarray}\n \\min_{\\{\\alpha_i,b_i\\}_{i=1}^k} \\left|\\left|w-\\sum_{i=1}^k \\alpha_i b_i\\right|\\right|^2 \n\\end{eqnarray}\nwhere $w\\in R^n$ is the original weight vector, $\\alpha_i\\in R$ and $b_i\\in\\{-1,+1\\}^n$ are variables to be learned. This NP-hard problem can be solved using an iterative greedy approximation which sequentially minimizes the residue. In each iteration, first the residue is computed as \n\\begin{eqnarray}\nr_{i-1}=w-\\sum_{j=1}^{i-1}\\alpha_j b_j,\n\\end{eqnarray}\nand then $\\alpha_i$ and $b_i$ are computed as $\\alpha_i=\\frac{1}{n}||r_{i-1}||_1$ and $b_i=\\text{sign}(r_{i-1})$. Further, refined greedy approximation~ extends this to further decrease the quantization error. In the $j^{th}$ iteration after $\\alpha_j$ and $b_j$ have been updated, the method adds one extra step to refine all computed $\\{\\alpha_i\\}_{i=1}^j$ with the least squares solution as follows.\n\\begin{eqnarray}\n[\\alpha_1, ..., \\alpha_j]=((B_j^TB_j)^{-1}B_j^Tw)^T\n\\end{eqnarray}\nwhere $B_j=[b_1,...,b_j]$. Typically refined greedy is more accurate than the greedy approach. In refined greedy approximation, after modification on the computed $\\alpha$'s, $b$'s are no longer optimal while the method keeps all of them fixed. To improve the refined greedy approximation, alternating minimizing $\\alpha$'s and $b$'s becomes a natural choice. Xu et al.~ find that only two alternating cycles is good enough to find high precision quantization. Further, similar to~, for an LSTM, we can combine overall network loss minimization with the multi-bit quantization loss minimization using this bi-level optimization.\n\\begin{eqnarray}\n\\min_{w,\\{\\alpha_i, b_i\\}_{i=1}^k} \\text{LSTM}\\left(\\sum_{i=1}^k\\alpha_i b_i\\right)\\\\\n\\text{such that }\\{\\alpha_i, b_i\\}_{i=1}^k=\\argmin_{\\{\\alpha_i',b_i'\\}_{i=1}^k}\\left|\\left|w-\\sum_{i=1}^k \\alpha_i' b_i'\\right|\\right|^2\n\\end{eqnarray}","id":"96bae440-819d-4a68-8df5-216ce70bd5a8","level":"subsubsection","origin_cites_number":3,"parent_id":"be929a80-0b3b-4005-bf19-e84fee8ab111","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","General Quantized Networks"],["subsubsection","Loss Aware Quantization (LAQ)"]],"subsections":[],"title":"Loss Aware Quantization (LAQ)"},{"cite_extract_rate":0.25,"cites":[2488],"content":"Each word vector is typically represented as a 300--500 dimensional vector, with each parameter being 32 bits. As there are millions of words, word vectors may take up to 3--6 GB of memory/storage. Can we quantize word vectors? We can clearly quantize them after training. But, we could also quantize when learning word embeddings. For example, Lam et al.~ perform 1-bit and 2-bit quantization while performing word2vec~ training using the Continuous Bag of Words (CBOW) method. They observe that quantization while training leads to better results compared to quantization after training. \nCheong et al.~ applied BS-Fixed and BS-Flexible binary quantization to Transformer models. They observed that the Transformer architecture is highly resistant to quantization, and is able to match the original model up to a 4-bit representation. Simple iterative pruning is much worse compared to quantization. Lastly, Shen et al.~ propose mixed-precision quantization for BERT based on the observation that different encoder layers should use different number of bits for quantization. Layers that exhibit flatter curvature of the loss gradient surface can be quantized to lower bit precision. Thus, they use different number of bits at different levels of granularity: layers, attention heads and groups of neurons. They observe that quantizing embedding layers with 8 bits and other weight matrices with 2--4 bits leads to results comparable with full-precision BERT.","id":"c55e72b6-a137-4167-a584-9f7c420c0df0","level":"subsubsection","origin_cites_number":4,"parent_id":"be929a80-0b3b-4005-bf19-e84fee8ab111","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","General Quantized Networks"],["subsubsection","Quantization for Word Embeddings and Transformers"]],"subsections":[],"title":"Quantization for Word Embeddings and Transformers"},{"cite_extract_rate":0.7333333333333331,"cites":[4498,8797,4494,4497,2670,4490,4351,4491,4492,2488,851],"content":"\\begin{table}\n \\centering\n \\scriptsize\n \\begin{tabular}{|l|l|l|l|l|l|l|l|}\n \\hline\nTask&Dataset&Model&Method&Eval&Bits (weights)&Bits (activation)&Metric\\\\\n\\hline\n\\hline\nLanguage modeling&IMDB&GRU&Uniform Q.~&0.882; 0.905&4&4&Acc (H)\\\\\n\\hline\nLanguage modeling&Linux Kernel&LSTM&BinaryConnect~&3.532; 1.329&1&FP&CE (L)\\\\\n\\hline\nLanguage modeling&Linux Kernel&LSTM&Loss Aware B.~&1.305/1.409; 1.329&1&FP/1&CE (L)\\\\\n\\hline\nLanguage modeling&Linux Kernel&LSTM&BNN~&3.624; 1.329&1&1&CE (L)\\\\\n\\hline\nLanguage modeling&Linux Kernel&LSTM&BWN~&1.307; 1.329&1&FP&CE (L)\\\\\n\\hline\nLanguage modeling&PTB&GRU&Uniform Q.~&102; 100&4&4&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&GRU&Balanced Q.~&116; 100&4&4&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Greedy~&118.9; 89.8&2&FP&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Refined Loss Aware~&95.6; 89.8&2&3&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Uniform Q.~&152/114; 109&2/4&2/4&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&BNN~&100; 97&4&4&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&HitNet~&110.3; 97.2&2&2&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Alternating LAQ~&103.1/91.4; 89.8&2/4&FP&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Alternating LAQ~&91.9; 89.8&2&3&PPW (L)\\\\\n\\hline\nLanguage modeling&PTB&LSTM&Balanced Q.~&126/123; 106&2&2/3&PPW (L)\\\\\n\\hline\nLanguage modeling&Text-8&LSTM&Refined Loss Aware~&122.3; 101.1&2&3&PPW (L)\\\\\n\\hline\nLanguage modeling&Text-8&LSTM&HitNet~&169.1; 151.4&2&2&PPW (L)\\\\\n\\hline\nLanguage modeling&Text-8&LSTM&Alternating LAQ~&105.1; 101.1&2&3&PPW (L)\\\\\n\\hline\nLanguage modeling&Text-8&RNN&Exponential Q.~&1.639; 1.588&2&FP&BPC (L)\\\\\n\\hline\nLanguage modeling&War and Peace&LSTM&BinaryConnect~&2.942; 1.268&1&FP&CE (L)\\\\\n\\hline\nLanguage modeling&War and Peace&LSTM&Loss Aware B.~&1.291/1.376; 1.268&1&FP/1&CE (L)\\\\\n\\hline\nLanguage modeling&War and Peace&LSTM&BNN~&3.05; 1.268&1&1&CE (L)\\\\\n\\hline\nLanguage modeling&War and Peace&LSTM&BWN~&1.313; 1.268&1&FP&CE (L)\\\\\n\\hline\nLanguage modeling&Wikidata-2&LSTM&HitNet~&126.72; 114.37&2&2&PPW (L)\\\\\n\\hline\nLanguage modeling&WikiText-2&LSTM&Refined Loss Aware~&105.8; 114.37&2&3&PPW (L)\\\\\n\\hline\nLanguage modeling&WikiText-2&LSTM&Alternating LAQ~&102.7; 100.1&2&3&PPW (L)\\\\\n\\hline\nNamed Entity Recognition&CoNLL-03&BERT-base&QBERT~&91.06; 95&2w8e&8&F1 (H)\\\\\n\\hline\nNamed Entity Recognition&CoNLL-03&BERT-base&Mixed-precision Q.~&94.37; 95&2-3w8e&8&F1 (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT17&Transformer&BS-Fixed/BS-Flexible~&11.61/12.11; 28.09&1&FP&BLEU (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT17&Transformer&K-Means 1/4-bit Q.~&12.07/27.65; 28.09&1&FP&BLEU (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT17&Transformer&K-Means 1-bit att-Q.~&24.96; 28.09&1&FP&BLEU (H)\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT17&Transformer&BS-Flexible 1-bit att-Q.~&25.54; 28.09&1&FP&BLEU (H)\\\\\n\\hline\nQuestion answering&SQuAD&BERT-base&QBERT~&79.6; 88.69&2w8e&8&F1 (H)\\\\\n\\hline\nQuestion answering&SQuAD&BERT-base&Mixed-precision Q.~&86.95; 88.69&2-3w8e&8&F1 (H)\\\\\n\\hline\nQuestion answering&SQuAD&Facebook’s DrQA&BS-Fixed~&77.04; 75.28&2&FP&F1 (H)\\\\\n\\hline\nSentiment analysis&IMDB&LSTM&Gaussian Q.~&79.64; 82.87&1&FP&Acc (H)\\\\\n\\hline\nSentiment analysis&IMDB&LSTM&Gaussian T./B.~&76.86/76.25; 82.87&2&FP&Acc (H)\\\\\n\\hline\nSentiment analysis&SST-2&BERT-base&QBERT~&84.63; 93&2w8e&8&Acc (H)\\\\\n\\hline\nSentiment analysis&SST-2&BERT-base&Mixed-precision Q.~&92.08; 93&2-3w8e&8&Acc (H)\\\\\n\\hline\nSpeech recognition&TIDIGITS&GRU&Pow2 T.~&99.18; 99.1&2&FP&Acc (H)\\\\\n\\hline\nSpeech recognition&TIMIT&4-layer MLP&Loss Aware T.~&29.97/28.35; 26.24&1/2&1&FER (L)\\\\\n\\hline\nSpeech recognition&WSJ&4-layer BiLSTM&Pow2 T.~&10.49; 11.16&2&FP&WER (L)\\\\\n\\hline\nTextual entailment&MNLI&BERT-base&QBERT~&77.02; 84.4&2w8e&8&Acc (H)\\\\\n\\hline\nTextual entailment&MNLI&BERT-base&Mixed-precision Q.~&82.29; 84.4&2-3w8e&8&Acc (H)\\\\\n\\hline\nTextual entailment&MNLI-m&BERT-base&QBERT~&76.56; 84&2w8e&8&Acc (H)\\\\\n\\hline\nTextual entailment&MNLI-m&BERT-base&Mixed-precision Q.~&81.75; 84&2-3w8e&8&Acc (H)\\\\\n\\hline\nWord similarity&M. Turk &Word embeddings&BS-Fixed~&0.602; 0.617&2&FP&CHR (H)\\\\\n\\hline\nWord similarity&MEN &Word embeddings&BS-Fixed~&0.764; 0.745&2&FP&CHR (H)\\\\\n\\hline\nWord similarity&Rare Words &Word embeddings&BS-Fixed~&0.362; 0.4&2&FP&CHR (H)\\\\\n\\hline\nWord similarity&SimLex&Word embeddings&BS-Fixed~&0.387; 0.358&2&FP&CHR (H)\\\\\n\\hline\nWord similarity&WordSim Relatedness &Word embeddings&BS-Fixed~&0.594; 0.529&2&FP&CHR (H)\\\\\n\\hline\nWord similarity&WordSim Similarity&Word embeddings&BS-Fixed~&0.752; 0.741&2&FP&CHR (H)\\\\\n\\hline\n \\end{tabular}\n\\caption{Comparison of various quantization methods (sorted by Task and then Dataset). Q.=Quantization, B.=Binarization, T.=Ternarization, PPW=Perplexity per word, BPC=Bits per character, CE=cross-entropy, FER=frame error rate, CHR=correlation with human rankings. FP=full precision (32 bits). For~, we report results with word embedding dimensions set to 1000. In the metric column, H means high is better while L means low is better. For quantization of the BERT-base model~, we report number of bits used for encoders as well as for embeddings. `2-3w8e' means 2 or 3 bits were used for encoder weights while 8 bits were used for embeddings. For NMT results by Cheong et al.~, ``att-Q'' means only attention layers were quantized.}\n \\label{tab:quantizationSummary}\n\\end{table}\nOtt et al.~ observed that the weight binarization methods do not work with RNNs. Hubara et al.~ were the first to attempt to quantize both weights and activations by trying to evaluate the accuracy of quantized recurrent models trained on the Penn Treebank dataset. Similar to~, Hubara et al.~ found that binarization of weight matrices lead to large accuracy degradation. Later techniques like the one by Xu et al.~ with 2 bits for weights and 3 bits for activations showed better results.\nTable~\\ref{tab:quantizationSummary} compares various quantization methods across different tasks and datasets. Accuracy of both the original and the quantized model are shown. Also, we report number of bits used for weights (which indicate the model size) as well as activations. For the same task, dataset and model combination, different papers report different accuracy of the full precision model because of slight changes in training hyper-parameters; hence we report accuracy of full precision model for each row. \nFor language modeling, PTB, Text-8, WikiText-2, Linux Kernel, IMDB and ``War and Peace'' are the popular datasets. Across all the datasets, loss aware binarization outperforms other weight binarization schemes. On the Linux Kernel dataset, it is even better than the full-precision network. BinaryConnect does not work well here because of the problem of exploding gradients. On PTB, Xu et al.'s Alternating LAQ~ with 2 bits for weights and 3 bits for activations leads to an LSTM which is just 2.1 points worse in terms of perplexity per word. By 3-bit quantization, Alternating LAQ can achieve $\\sim$10.5x memory saving and $\\sim$3× real inference acceleration. Uniform and Balanced quantization are rule-based and not aim at minimizing the error. Balanced quantization proposed by Zhou et al.~ performs better than HitNet~ and uniform quantization~. Balanced quantization leads to better results compared to unbalanced counterparts, especially when quantizing to 2-bit weights. However, for 4-bit weights, there is no clear gap between scaling by mean and scaling by max (i.e. balanced and unbalanced quantization).\nAcross multiple tasks like named entity recognition with CoNLL-03, question answering with SQuAD, sentiment analysis with SST-2, textual entailment using MNLI, Shen et al.~ show that mixed precision quantization (where different number of bits are used for different groups of neurons -- 128 groups in each layer) of BERT is better than QBERT. The reason behind this is the discrepancy (in QBERT) that not all the layers have the same sensitivity to quantization. For more sensitive layers, higher bit precision needs to be set, while for layers that are less sensitive, 2-bits setting is already sufficient. With only additional 5MB memory storage, 2/3-bits mixed-precision Q-BERT is able to retain the performance drop within 2.3\\% for MNLI, SQuAD and 1.1\\% for SST-2, CoNLL-03, with up to 13x compression ratio in weights. For machine translation, Cheong et al.~ observe that Transformer architecture is highly resistant to quantization (unlike to pruning), and are, in essence, able to match the original model up to a 4-bit representation. Binary Scheme Flexible outperforms 1-bit k-means in both pure 1-bit compression and quantizing only the attention layers, suggesting not tying the weights to particular centroids improves performance, and outperform Binary Scheme Fixed, indicating learning the values to be a superior method. After binarizing only the attention layers, we are still able to recover 90\\% of the model's performance. Lam et al.~ experimented with quantization of word embeddings and showed that 2-bit quantized word vectors outperform full precision vectors on word similarity tasks, but do worse on word analogy tasks. Intuitively, they reason that full precision Word2Vec is prone to overfitting with increased epochs of training; quantized training does not seem to suffer as much from this. For sentiment analysis on IMDB, Alom et al.~ show that quantization to 4 bits is better than to 3 or 2 bits, which is expected. They also show that the normal distribution shows better performance against uniform distribution with quantized weights. For speech recognition, Ott et al.~ show that pow-2 ternarization is the best. \nCheong et al~ were the first to quantize Transformers. They observed that in the last attention layer of the decoder over the encoder hidden states, the attention distribution of the original and 4-bit model are highly similar, indicating 4 bit weights, i.e weights that take on one of 16 values, is enough to get the full effects of attention. Attention distributions in the encoder layers of the Transformer for the original and 4-bit models are almost indistinguishable from one another. This again highlights the idea that self-attention is highly resistant to quantization and could be heavily compressed. Later Shen et al.~ showed that comparable performance to full precision BERT can be achieved with at most 2.3\\% performance degradation across many tasks, even with ultra-low precision quantization down to 2 bits, corresponding up to 13x compression of the model parameters, and up to 4x compression of the embedding table as well as activations. \nOverall, quantization performs model compression by reducing the number of bits per weight value. Binary quantization does not work well by itself for text based neural models. But ternary and higher-bit quantization lead to significant model size reduction without loss in accuracy across tasks. One consideration for quantization is that 3-bit quantized execution is typically not supported in hardware. It is however possible to load 3-bit quantized values and cast them to higher bit precision such as 4 or 8 bits in the execution units. This would still have the benefit of reduced memory volume to/from DRAM. Non-uniform quantization methods like balanced quantization or KMeans based quantization methods are better than uniform quantization methods. Loss aware quantization done while training is better than static loss-unaware quantization. Mixed-precision quantization combined with pruning is highly effective for Transformer based models.","id":"97e94011-bad6-4d07-98cb-26de46e32fe4","level":"subsection","origin_cites_number":15,"parent_id":"c0c0e7e6-55dc-426b-8669-a8adfb8fe243","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Quantization"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:knowledgeDistillation}\nKD methods are the most popular model compression methods for Transformer networks. Also called student-teacher networks, the main idea is to first train a deep teacher network, and then learn a shallow student network that mimics the teacher. After training, the student model is deployed. What information (``dark knowledge'') from the teacher can be used to train the student? What loss functions can be used to ensure right flow of information from teacher to student? Can we have an ensemble of teachers, or teacher assistants or rather fellow students who can train the student? \nWe discuss these aspects in this section.","id":"2c5b1b2b-3f74-442d-8ee6-503425cf4207","level":"section","origin_cites_number":0,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Knowledge Distillation (KD)"]],"subsections":["5addca6a-ae25-4e99-8ffc-a608a63231c7","d2899d20-743a-43e0-84c7-0fc90c1a975b","e1a31918-0804-4d5f-9f78-0ea43f83fc23","8fabba12-aca9-4230-b6a7-cc461dfdceaa","4cd2b9eb-147c-41ba-996b-a30e824f168f"],"title":"Knowledge Distillation (KD)"},{"cite_extract_rate":0.9166666666666661,"cites":[8798,4499,4500,4503,4504,681,8375,8799,8390,4501,4502],"content":"Ba and Caruna~ proposed Student Teacher networks (or mimic models) where the student uses the logits before softmax from the teacher network for training (see Fig.~\\ref{fig:kd}(A)). The student model is not trained on the original labels; it is trained to learn the function that was learned by the teacher model. Thus, the student model is optimized to minimize the L2 loss between the teacher logits and the student logits across all training instances. Such distilled student models are more accurate than the same shallow student trained directly on the original labeled training data mainly because: (1) Teacher removes noisy labels, if any. (2) The uncertainty from the teacher is more informative to the student than the original 0/1 labels. (3) The original targets may depend in part on features not available as inputs for learning, but the student sees targets that depend only on the input features. The dependence on unavailable features has been eliminated by filtering targets through the teacher.\nYet another way of utilizing logits is to have the student learn from noisy teacher logits~. After obtaining logits from the teacher, Gaussian noise with mean 0 and standard deviation $\\sigma$ is added to teacher’s logits. This perturbation can be applied to samples selected with probability $\\alpha$. The perturbed outputs produce the effect of a regularizer. \n\\begin{figure}\n \\centering\n \\includegraphics[width=\\textwidth]{kd}\n \\caption{Different Types of Knowledge Distillation methods.}\n \\label{fig:kd}\n\\end{figure}\nWhile Ba and Caruna~ suggested using only the logits, Hinton et al.~ suggested training the student by minimizing the cross entropy loss between the teacher softmax output and the student softmax output, besides minimizing the cross entropy between student prediction and actual label. The first part is called the soft loss and the second one is called the hard loss. Typically hard loss is given much lower weight compared to the soft loss term. To make the softmax output non-peaked and thereby transfer more useful information from teacher to student, softmax with temperature $>$1 should be used. The same temperature should be used for training both the teacher and the student, but after the student has been trained the temperature can be set to 1 at test time. Besides logits and softmax output, Sobolev training for neural networks is a method for incorporating target derivatives in addition to the target values while training student network (see Fig.~\\ref{fig:kd}(C)). Czarnecki et al.~ experiment with first two derivatives of the targets.\nKD has also been used along with quantization for better model compression~. We start with a trained full-precision large teacher network and an apprentice (student) network that has been initialised with full-precision weights. The apprentice network's precision is lowered and is fine-tuned using KD. \nWhy just use the output from the last layer of the teacher for training the student? In FitNets~, the student performs hint-based training, i.e., the student is trained using not only the outputs but also the intermediate representations learned by the teacher as hints to improve the training process and final performance of the student (see Fig.~\\ref{fig:kd}(B)). we choose a hidden layer of the FitNet, the guided layer, to learn from the teacher's hint layer. Because the student intermediate hidden layer will generally be smaller than the teacher's intermediate hidden layer, additional parameters are introduced to map the student hidden layer to the prediction of the teacher hidden layer. \nWhile the methods discussed so far use logits, softmax output or their derivatives to transfer knowledge, Yim et al.~ proposed a ``flow of solution procedure (FSP)'' method where the distilled knowledge is transferred in terms of flow between layers, which is calculated by computing the inner product between features from two layers (see Fig.~\\ref{fig:kd}(D)). What does this ``flow'' capture intuitively? If we view the input of a deep network as the question and the output as the answer, we can think of the generated features at the middle of the network as the intermediate result in the solution process. There are many ways to solve the problem of generating the output from the input. Hence, mimicking the generated features of the teacher can be a hard constraint for the student. Learning the solution process from teacher is important. More concretely, the student is trained to minimize the L2 difference between the teacher and student FSP matrices computed across various pairs of layers and across multiple training instances. A similar method called Representational distance learning (RDL) has also been proposed in~.\nLastly, multiple KD variants have been proposed for sequence-level predictions~, e.g., for neural machine translation (NMT). In word-level KD, cross-entropy is minimized between the student/teacher distributions for each word in the actual target sequence, as well as between the student distribution and the degenerate data distribution, which has all of its probability mass on one word. In sequence-level KD (Seq-KD) the student network is trained on the output from beam search of the teacher network that had the highest score. In sequence-level interpolation (Seq-Inter) the student is trained on the output from beam search of the teacher network that had the highest similarity (say using BLEU score) with the target sequence.","id":"5addca6a-ae25-4e99-8ffc-a608a63231c7","level":"subsection","origin_cites_number":12,"parent_id":"2c5b1b2b-3f74-442d-8ee6-503425cf4207","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Knowledge Distillation (KD)"],["subsection","Various Distillation Architectures"]],"subsections":[],"title":"Various Distillation Architectures"},{"cite_extract_rate":1,"cites":[4506,4505],"content":"Can multiple students learn from each other? Is a powerful teacher really required? In the deep mutual learning (DML) method~, different from the one-way transfer between a static pre-defined teacher and a student in model distillation, with DML, an ensemble of students learn collaboratively and teach each other throughout the training process. Surprisingly, no prior powerful teacher network is necessary -- mutual learning of a collection of simple student networks works, and moreover outperforms distillation from a more powerful yet static teacher. Specifically, each student is trained with two losses: a conventional supervised learning loss, and a mimicry loss that aligns each student's class posterior with the class probabilities of other students. \nAnil et al.~ propose a similar method but suggest letting the students learn independently just using the conventional supervised learning (hard) loss at least for a few burn in iterations (see Fig.~\\ref{fig:kd}(E)). After this, the mutual learning can be done as in DML. They also propose a variant of their Co-Distillation method to perform this training in a distributed scenario where communication efficiency is also important. To update the parameters of one network using co-distillation one only needs the predictions of the other networks, which can be computed locally from copies of the other networks' weights. Empirically, using stale predictions instead of up-to-date predictions for the other neural networks has little to no adverse effect on the quality of the final trained model produced by co-distillation.","id":"d2899d20-743a-43e0-84c7-0fc90c1a975b","level":"subsection","origin_cites_number":2,"parent_id":"2c5b1b2b-3f74-442d-8ee6-503425cf4207","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Knowledge Distillation (KD)"],["subsection","Collaborative Learning"]],"subsections":[],"title":"Collaborative Learning"},{"cite_extract_rate":0.75,"cites":[4508,4507,681],"content":"So far we have talked about a student mimicing a single teacher. However, it is interesting to explore if the student can learn better in presence of multiple teachers or from a teacher assistant.\nIntuitively and also observed empirically, student network performance degrades when the gap between student and teacher is large. Given a fixed student network, one cannot employ an arbitrarily large teacher, or in other words, a teacher can effectively transfer its knowledge to students up to a certain size, not smaller. To alleviate this shortcoming, Mirzadeh et al.~ introduced multi-step KD, which employs an intermediate-sized network (teacher assistant) to bridge the gap between the student and the teacher (see Fig.~\\ref{fig:kd}(F)).\nThe teacher assistant (TA) models are distilled from the teacher, and the student is then only distilled from the TAs. One could also perform multi-step TA distillation, for example, distillation path could be $10 \\rightarrow 6 \\rightarrow 4 \\rightarrow 2$.\nA simple way to do KD with multiple teachers is to train student with cross entropy loss between student predictions and average prediction from multiple teachers (see Fig.~\\ref{fig:kd}(G)). A more effective method is to augment this with a relative dissimilarity (RD) loss~ defined over intermediate layer outputs generated for a triplet of instances between the student and an ensemble of teachers. For the student, the middle layer is selected. For each teacher, we select the layer such that most teachers are consistent with the resulting order relationships under the voting strategy. We discuss the RD loss given a student and a teacher. Consider a triplet of instances ($x_i$, $x_i^+$, $x_i^-$) such that at an intermediate layer of the teacher network, distance between activations for $x_i^+$ and $x_i$ is smaller than the distance between activations for $x_i^-$ and $x_i$. Let $p_i$ be the intermediate output from student for example $x_i$. Then the RD loss for the triplet ($x_i$, $x_i^+$, $x_i^-$) can be written as follows. \n\\begin{eqnarray}\n\\text{Loss}=\\max(0,d(p_i, p_i^+)-d(p_i, p_i^-)+\\delta) \n\\end{eqnarray}\nwhere $d$ is the distance function, and $\\delta$ is a small constant to prevent the trivial solution. To extend this loss function definition to multiple teachers, the order between the instances $x_i^+$ and $x_i^-$ given $x_i$ is decided based on majority voting between the teachers. \nThere are also specific settings when distilling from multiple teachers becomes natural, e.g., when the number of classes is large~ or in multi-lingual settings~. When the number of classes is very large, the teacher model could be an ensemble that contains one generalist model trained on all the data and many ``specialist'' models, each of which is trained on data that is highly enriched in examples from a very confusable subset of the classes (like different types of mushroom). Softmax distribution vector of this type of specialist can be made much smaller by combining all of the classes it does not care about into a single dustbin class. Each specialist model is initialized with the weights of the generalist model. These weights are then slightly modified by training the specialist with half its examples coming from its special subset and half sampled at random from the remainder of the training set. To derive groupings of object categories for the specialists, we focus on categories that the full generalist network often confuses. When training the student, for each instance, we first find the $set $k$ of n$ most probable classes according to the generalist model. Then, we take all the specialist models, $m$, whose special subset of confusable classes has a non-empty intersection with $k$ and call this the active set of specialists $A_k$. Given student's full probability distribution $q$ over all the classes, we minimize the following.\n\\begin{eqnarray}\n\\text{Loss}=KL(p^g,q)+\\sum_{m\\in A_k} KL(p^m, q) \n\\end{eqnarray}\nwhere $p^g$ is output distribution from the generalist model, and $p^m$ is the output distribution from the $m^{th}$ specialist model.\nAn ensemble of teachers is also very useful in a multi-lingual NMT setting~. Individual models for each language pair are first trained and regarded as teachers, and then the multilingual model is trained to fit the training data and match the outputs of individual models simultaneously through KD. When the accuracy of multilingual model surpasses the individual model for the accuracy threshold $\\tau$ on a certain language pair, we remove the distillation loss and just train the model with original negative log-likelihood loss for this pair. Lastly, when learning from a teacher ensemble, it is burdensome to load all the teacher models in the GPU memory for distillation. Alternatively, we first generate the output probability distribution of each teacher model for each instance offline, and then just load the top-$K$ probabilities of the distribution into memory and normalize them so that they sum to 1 for distillation. This reduces the memory cost from the scale of $|V|$ (the vocabulary size) to $K$.","id":"e1a31918-0804-4d5f-9f78-0ea43f83fc23","level":"subsection","origin_cites_number":4,"parent_id":"2c5b1b2b-3f74-442d-8ee6-503425cf4207","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Knowledge Distillation (KD)"],["subsection","Multiple Teachers"]],"subsections":[],"title":"Multiple Teachers"},{"cite_extract_rate":0.6666666666666661,"cites":[4513,4509,4511,2481,4512,4514,4510,854],"content":"Recently, there has been a lot of work around distilling Transformers to smaller Transformers with less number of layers or to Bidirectional LSTMs. Some of these methods aim at improving the accuracy versus model size tradeoff, while others focus on complex settings like mismatching student-teacher vocabulary~ or mismatch number of attention heads. \nZhao et al.~ learn a student with small vocabulary compared to the teacher using a dual training method. During distillation, for a given training sequence input to the teacher model, they mix the teacher and student vocabularies by randomly selecting tokens from the sequence to segment using the student vocabulary, with the other tokens segmented using the teacher vocabulary. As part of the masked language model (MLM) task, the model now needs to learn to predict words from the student vocabulary using context words segmented using the teacher vocabulary, and vice versa. The expectation is that the student embeddings can be learned effectively this way from the teacher embeddings as well as teacher model parameters. We perform dual training only for the teacher model inputs. The student model receives words segmented exclusively using the student vocabulary. Also, during MLM, the model uses different softmax layers for the teacher and the student vocabularies depending on which one was used to segment the word in question. Instead of distilling solely on the teacher model's final-layer outputs, layer-wise teacher model parameters can also be leveraged to directly optimize parameters of corresponding layers in the student model.\nIn Patient KD (PKD)~, the student learns from the teacher's output after every $k$ layers (see Fig.~\\ref{fig:kd}(H)) or the output from the last few layers of the teacher (see Fig.~\\ref{fig:kd}(I)). The student BERT is initialized using some layers of the pre-trained teacher BERT. TinyBERT~ further extends this idea by using extensive knowledge from embedding layer, and attention and hidden sub-layers of multiple teacher layers, and also the overall teacher output (see Fig.~\\ref{fig:kd}(J)). Each student layer is first mapped to a teacher layer before the student training. Liu et al.~ distill a multi-task student from a multi-task teacher, given the soft targets of the training data across multiple tasks. If task $t$ has a teacher, the task-specific loss is the average of two objective functions, one for the correct targets and the other for the soft targets assigned by the teacher. In MiniLM~, the student is trained by deeply mimicking the self-attention behavior of the last Transformer layer of the teacher (see Fig.~\\ref{fig:kd}(K)). Besides self-attention distributions, MiniLM introduces the self-attention value-relation transfer to help the student achieve a deeper mimicry. The value-relation is computed as pairwise correlation between different components of the value matrix across various attention heads of the final layer. Pretrained Distillation~ pretrains the student model with a self-supervised masked LM objective on a\nlarge corpus first, then performs a standard KD on supervised tasks. \nMost of these models learn one-to-one layer mapping, where each student layer is guided by only one specific teacher layer. Li et al.~ propose a method where each student intermediate layer learns from every teacher intermediate layer with learnable attention weights. Both the embedding-layer distillation and the prediction-layer distillation employ the one-to-one layer mapping as in TinyBERT and BERT-PKD.\nTang et al.~ propose distillation of a BERT model to a single layer BiLSTM using KL divergence between student and teacher logits (see Fig.~\\ref{fig:kd}(L)). Mukherjee et al.~ also distill a multi-lingual BERT (mBERT) model to a BiLSTM. Representation transfer is done from Transformer-based teacher model to BiLSTM-based student model with different embedding dimensions and disparate output spaces. Distillation features include teacher logits and internal teacher representations for one teacher layer. To make all output spaces compatible, a non-linear projection of the parameters in student representation is done to have same shape as teacher representation for each token. The projection parameters are learned by minimizing the KL-divergence (KLD) between the representations of the student and the chosen layer from the teacher. Overall there are multiple loss functions for the student training: supervised hard loss, soft loss wrt output logits, and soft loss wrt internal teacher layer. Rather than optimizing for all loss functions jointly, stage-wise training is performed where each loss function is sequentially used for optimization. \nLastly, there have been recent efforts~ to distill Transformer models to slightly modified Transformer architectures. MobileBERT~ is a thin version of BERT-large, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks (FFN). To train MobileBERT, we first train a specially designed teacher model, an inverted-bottleneck incorporated BERT-large model (IB-BERT). The IB-BERT uses the inverted-bottleneck structures to adjust its feature map size to 512. Thus, in the bottleneck structure, the inputs to the multi-head attention (MHA) module are from wider feature maps (of inter-block size), while the inputs to the FFN are from narrower bottlenecks (of intra-block size). To fix this issue, MobileBERT uses stacked feed-forward networks to re-balance the relative size between MHA and FFN. Each MobileBERT layer contains one MHA but 4 stacked FFN after each MHA. Then, we conduct knowledge transfer from this teacher to MobileBERT using feature map transfer and attention transfer across all layers. Also, while distilling they perform progressive knowledge transfer, i.e., they progressively train each layer in $L$ stages, where $L$ is the number of layers. When the $l$-th layer is trained, all the trainable parameters in the layers below are frozen. Another difference in the MobileBERT student architecture is usage of high information flow residual connections between the high-channel-count layers. MobileBERT is 4.3x smaller and 5.5x faster than BERT-base while achieving competitive results on well-known benchmarks. Iandola et al.~ propose a new Transformer model architecture called SqueezeBERT which is much like BERT-base, but with the position-wise fully connected layers implemented as convolutions, and grouped convolutions for many of the layers. Distillation for SqueezeBERT is rather straightforward. Distillation is applied only to the final layer, and only during finetuning using soft cross entropy loss with respect to a weighted sum of the teacher's logits and a one-hot encoding of the ground-truth. Teacher model is a BERT-base model finetuned independently on each GLUE task, and these task-specific teacher weights are used for distillation. Xu et al.~ propose a method for progressive module replacement for compressing BERT. Their approach first divides the original BERT into several modules and builds their compact substitutes. Then, the original modules are randomly replaced with their substitutes to train the compact modules to mimic the behavior of\nthe original modules. We progressively increase the probability of replacement through\nthe training. In this way, their approach brings\na deeper level of interaction between the original and compact models.","id":"8fabba12-aca9-4230-b6a7-cc461dfdceaa","level":"subsection","origin_cites_number":12,"parent_id":"2c5b1b2b-3f74-442d-8ee6-503425cf4207","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Knowledge Distillation (KD)"],["subsection","Distilling Transformers"]],"subsections":[],"title":"Distilling Transformers"},{"cite_extract_rate":0.8518518518518511,"cites":[4479,4515,1150,4505,4514,9120,4513,856,681,854,7333,4486,4500,1568,7,8390,4509,4511,2481,4512,4507,2488,4510],"content":"\\begin{table}\n \\centering\n \\scriptsize\n \\begin{tabular}{|l|l|l|l|l|l|p{0.5in}|}\n \\hline\nTask&Dataset&Teacher/Student models&Method&Metric&Size (Distilled; Orig)&Eval. (Distilled; Orig)\\\\\n\\hline\n\\hline\nAbs. summarization&CNN/DailyMail&UniLM-large/12-layer BERT&MiniLM~&ROUGE-L (H)&33M; 340M&39.73; 40.34\\\\\n\\hline\nAd CTR prediction&Criteo&No teacher/DNN&CoDistillation~&MAE (L)&-&0.019; 0.022\\\\\n\\hline\nCross-lingual NLI&XNLI (15 langs.)&XLM-R-base/12-layer BERT&MiniLM~&Acc (H)&21M; 85M&71.1; 74.5\\\\\n\\hline\nCross-lingual QA&MLQA (7 langs.)&XLM-R-base/12-layer BERT&MiniLM~&F1 (H)&21M; 85M&63.2; 64.9\\\\\n\\hline\nIntent detection&SNIPS (7-class)&BERT-base/6-layer BERT&Mixed-vocabulary training~&Acc (H)&6.2M; 109M&98.7; 98.6\\\\\n\\hline\nLanguage modeling&Common Crawl&No teacher/2-layer LSTM&CoDistillation~&CE (L)&-&3.91; 3.92\\\\\n\\hline\nMachine reading comp.&RACE&BERT-base/6-layer BERT&Patient KD~&Acc (H)&67M; 109M&60.34; 65.30\\\\\n\\hline\nMachine reading comp.&RACE&BERT-base/6-layer BERT&Vanilla KD~&Acc (H)&67M; 109M&58.74; 65.30\\\\\n\\hline\nNER (41 langs.)&Wikiann-41&Multiple mBERT/BiLSTM&XtremeDistill~&F1 (H)&31.8M; 179M*41&88.64; 90.76\\\\\n\\hline\nNMT (de$\\rightarrow$en)&OpenNMT&4-layer LSTM/2-layer LSTM&Vanilla KD~&BLEU (H)&64.8M; 84.8M&15.48; 15.88\\\\\n\\hline\nNMT (de$\\rightarrow$en)&OpenNMT&4-layer LSTM/2-layer LSTM&Quantized Distillation (4 bits)~&BLEU (H)&64.8M; 84.8M&15.26; 15.88\\\\\n\\hline\nNMT (de$\\rightarrow$en)&WMT13&4-layer LSTM/2-layer LSTM&Quantized Distillation (4 bits)~&BLEU (H)&81.6M; 84.8M&35.32; 34.7\\\\\n\\hline\nNMT (de$\\rightarrow$en)&WMT13&4-layer LSTM/2-layer LSTM&Vanilla KD~&BLEU (H)&81.6M; 84.8M&30.21; 34.7\\\\\n\\hline\nNMT (en$\\rightarrow$Others)&IWSLT (13 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*12&22.96; 22.72\\\\\n\\hline\nNMT (en$\\rightarrow$Others)&IWSLT (13 langs.)&Multiple Transformers/Transformer&Seq-KD with Multiple teachers~&BLEU (H)&44M; 44M*12&21.98; 22.72\\\\\n\\hline\nNMT (en$\\rightarrow$Others)&WMT (7 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*6&24.47; 24.50\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Word-KD~&BLEU (H)&49M; 221M&14.9; 17.7\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Seq-KD~&BLEU (H)&49M; 221M&18.1; 17.7\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Seq-KD + Seq-Inter + Word-KD~&BLEU (H)&49M; 221M&18.5; 17.7\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Pruned Seq-KD + Seq-Inter~&BLEU@5 (H)&8M/17M; 221M&18.5/19.1; 19.5\\\\\n\\hline\nNMT (Others$\\rightarrow$en)&IWSLT (13 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*12&30.34; 29.7\\\\\n\\hline\nNMT (Others$\\rightarrow$en)&Ted Talk (45 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*43&28.95; 25.17\\\\\n\\hline\nNMT (Others$\\rightarrow$en)&WMT (7 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*6&28.61; 27.07\\\\\n\\hline\nNMT (th$\\rightarrow$en)&IWSLT15&4-layer LSTM/2-layer LSTM&Word-KD~&BLEU (H)&8M; 47M&11.8; 14.3\\\\\n\\hline\nNMT (th$\\rightarrow$en)&IWSLT15&4-layer LSTM/2-layer LSTM&Seq-KD~&BLEU (H)&8M; 47M&12.8; 14.3\\\\\n\\hline\nNMT (th$\\rightarrow$en)&IWSLT15&4-layer LSTM/2-layer LSTM&Seq-KD + Seq-Inter + Word-KD~&BLEU (H)&8M; 47M&14.2; 14.3\\\\\n\\hline\nQuestion generation&SQuAD 1.1&UniLM-large/12-layer BERT&MiniLM~&BLEU@4 (H)&33M; 340M&23.27; 24.32\\\\\n\\hline\nSlot filling&SNIPS (39 slots)&BERT-base/6-layer BERT&Mixed-vocabulary training~&F1 (H)&6.2M; 109M&95.0; 97.0\\\\\n\\hline\n \\end{tabular}\n\\caption{Comparison of various knowledge distillation methods (sorted by Task and then Dataset). CE=cross entropy, MAE=mean absolute error. en=English, th=Thai. MRC=Machine Reading Comprehension. NLI=Natural Language Inference. QA=Question Answering. NER=Named Entity Recognition. In the metric column, H means high is better while L means low is better.}\n \\label{tab:kdSummary1}\n\\end{table}\nTable~\\ref{tab:kdSummary1} compares various knowledge distillation methods across different tasks and datasets. Accuracy of both the original and the distilled model are shown in the Eval column. Also, we report model size for both the student as well as the teacher models. Note that sometimes smaller student models perform better than the teacher models. This could be because (1) for some (task, dataset) pairs, the smaller models are a better regularized fit compared to potentially overfitted teacher models, and (2) student models can be rigorously trained using additional semi-supervision while teacher models depend on limited labeled training data.\nWhile knowledge distillation has been used for distilling NLP models across many applications, NMT is the most popular one. For abstractive summarization, MiniLM~ leads to a student models which is less than one tenth of the teacher without much loss in ROUGE-L. Similarly, MiniLM shows good results for cross-lingual NLI and multi-lingual QA as well. For Ad click through rate (CTR) prediction and language modeling, Anil et al.~ show that co-distillation leads to lower MAE and cross entropy respectively compared to the individually trained models. Zhao et al.~'s mixed-vocab training leads to 6-layer model that retains over 95\\% of the BERT-base model’s slot filling F1 score while being 30x smaller ($<$10 MB without quantization) and 57x faster on a mobile device, yet task-agnostic. For Named Entity Recognition (NER), Mukherjee et al.~ show that XtremeDistil leads to massive compression of teacher models like mBERT by upto 35x in terms of parameters and 51x in terms of latency for batch inference while retaining 95\\% of its F1-score for NER over 41 languages. \nFor NMT, experiments have been done on OpenNMT, WMT, IWSLT and TedTalk datasets. Kim et al.~ make the following observations: (1) Sequence-level knowledge distillation (Seq-KD) does better than word-level knowledge distillation (Word-KD) on English $\\rightarrow$ German and performs similarly on Thai $\\rightarrow$ English. (2) Combining them (Seq-KD + Word-KD) results in further gains, indicating that these methods provide orthogonal means of transferring knowledge from the teacher to the student: Word-KD is transferring knowledge at the the local (i.e. word) level while Seq-KD is transferring knowledge at the global (i.e. sequence) level. (3) Applying weight pruning on top of knowledge distillation results in a student model that has 13x fewer parameters than the original teacher model, with a decrease of 0.4 BLEU. Tan et al.~ show that one model is enough to handle multiple languages (up to 44 languages), with comparable or even better accuracy than individual models. Their method achieves larger improvements on some languages, such as Da, Et, Fi, Hi and Hy, than others. This is correlated with the data size of the languages. When a language is of smaller data size, it may get more improvement due to the benefit of multilingual training.\n\\begin{table}\n \\centering\n \\scriptsize\n \\begin{tabular}{|p{0.3in}|l|l|l|l|l|l|l|l|l|l|l|l|l|l|}\n\\hline\nMethod&Method&Teacher/Student&Size&MRPC&MNLI&MNLI-m&SST-2&QQP&QNLI&RTE&CoLA&STS-B&SQuAD&SQuAD\\\\ \nCategory&& models&& F1& Acc& Acc& Acc&F1& Acc& Acc& MCC&Spearman& 1.1 F1& 2.0 F1\\\\\n\\hline\n\\hline\n\\multirow{3}{0.3in}{Original Models}&BERT-B~&-& 109M&88.9&83.4&84.6&93.5&71.2&90.5&66.4&52.1&85.8&88.4&77.7\\\\\n\\cline{2-15}\n&BERT-L~&-& 340M&89.3&85.9&86.7&94.9&72.1&92.7&70.1&60.5&86.5&90.9&81.9\\\\\n\\cline{2-15}\n&MTDNN (ensemble)~&-& 340*4M&90.0&87.2&86.7&95.6&72.4&93.9&80.9&61.5&88.3&-&-\\\\\n\\hline\n\\hline\n\\multirow{19}{*}{\\rotatebox{90}{Knowledge Distillation Methods}}&Distilled-BiLSTM~&BERT-L/BiLSTM&0.96M&-&72.6&73.0&90.7&68.2&-&-&-&-&-&-\\\\\n\\cline{2-15}\n&Mixed-vocab. training~&BERT-L/BERT-12&10.9M&87.2&80.5&80.7&90.6&-&-&-&-&-&-&-\\\\\n\\cline{2-15}\n&TinyBERT~&BERT-B/BERT-4&14.5M&86.4&81.8&82.5&92.6&71.3&87.7&66.6&44.1&80.4&82.1&71.8\\\\\n\\cline{2-15}\n&BERT-EMD~&BERT-B/BERT-4&14.5M&87.6&80.6&82.1&91.0&69.3&87.2&66.2&25.6&82.3&-&-\\\\\n\\cline{2-15}\n&MobileBERT~&BERT-L/BERT-6&25.3M&88.8&82.6&83.3&92.8&70.2&90.6&66.2&50.5&84.4&90.0&79.2\\\\\n\\cline{2-15}\n&MobileBERT+Quantization~&BERT-L/BERT-6&25.3M&87.0&-&83.9&91.9&-&90.8&-&-&-&90.0&-\\\\\n\\cline{2-15}\n&MiniLM~&BERT-B/BERT-12&33M&89.5&-&85.7&93.0&91.3&91.5&73.3&58.5&-&-&81.7\\\\\n\\cline{2-15}\n&SqueezeBERT~&BERT-B/SqueezeBERT&51.1M&87.8&81.1&82.0&91.4&80.3&90.1&73.2&46.5&86.7&-&-\\\\\n\\cline{2-15}\n&DistilBERT~&BERT-B/BERT-4&52.2M&82.4&78.0&78.9&91.4&68.5&85.2&54.1&32.8&76.1&81.2&64.1\\\\\n\\cline{2-15}\n&Patient KD~&BERT-B/BERT-4&52.2M&82.6&79.3&79.9&89.4&70.2&85.1&62.3&24.8&79.8&79.5&64.6\\\\\n\\cline{2-15}\n&BERT-of-Theseus~&BERT-B/BERT-6&66M&87.6&82.1&82.4&92.2&71.6&89.6&66.2&47.8&84.1&-&-\\\\\n\\cline{2-15}\n&MiniLM~&BERT-B/BERT-6&66M&88.4&-&84.0&92.0&91.0&91.0&71.5&49.2&-&-&76.4\\\\\n\\cline{2-15}\n&BERT-EMD~&BERT-B/BERT-6&66M&89.8&83.5&84.7&93.3&72.0&90.7&71.7&47.5&86.8&-&-\\\\\n\\cline{2-15}\n&Patient KD~&BERT-B/BERT-6&67M&85.0&81.0&81.5&92.0&70.7&89.0&65.5&43.5&81.6&85.3&69.8\\\\\n\\cline{2-15}\n&Vanilla KD~&BERT-B/BERT-6&67M&86.2&79.8&80.2&91.5&70.1&88.3&64.7&-&-&-&-\\\\\n\\cline{2-15}\n&DistilBERT~&BERT-B/BERT-6&67M&86.9&81.3&82.6&92.5&70.1&88.9&58.4&49.0&81.3&86.2&69.5\\\\\n\\cline{2-15}\n&TinyBERT~&BERT-B/BERT-6&67M&87.3&83.2&84.6&93.1&71.6&90.4&70.0&51.1&83.7&87.5&77.7\\\\\n\\cline{2-15}\n&Pretrained Distillation~&BERT-B/BERT-6&67M&86.8&82.2&82.8&91.8&70.4&88.9&65.3&-&-&-&-\\\\\n\\cline{2-15}\n&MTDNN-KD~&MTDNN/MTDNN-KD&340M&91.1&86.7&87.5&95.6&72.7&96.0&85.1&65.4&89.6&-&-\\\\\n\\hline\n\\hline\nParam.&ALBERT-B~ (dev)&-&12M&-&-&81.6&90.3&-&-&-&-&-&89.3&80\\\\\n\\cline{2-15}\nSharing&ALBERT-L~ (dev)&-&18M&-&-&83.5&91.7&-&-&-&-&-&90.6&82.3\\\\\n\\hline\nTensor Decomp.&FLOP~&-&80M&88.61&-&-&92.09&-&89.05&-&-&88.18&-&-\\\\\n\\hline\n\\multirow{3}{*}{Pruning}&RPP Iterative Magnitude Pruning~&-&138M&88.1&86.1&85.7&92.4&91.2&92.3&70.1&82.8&-&90.23&75.3\\\\\n\\cline{2-15}\n&Iterative Magnitude Pruning~&-&170M&83.5&77&82.5&91.3&85.1&90.2&68.6&76.3&-&85.3&-\\\\\n\\cline{2-15}\n&LayerDrop~&-&66M&85.3&-&82.9&92.5&-&89.4&-&-&-&-&-\\\\\n\\hline\nQuant.&Mixed-precision quant. (QBERTMP)~&-&-&-&82.29&81.75&92.08&-&-&-&-&-&86.95&-\\\\\n\\hline\nSubQuad Trans.&Linformer~&-&-&-&-&-&93.1&90.8&91.2&-&-&-&-&-\\\\\n\\hline\n \\end{tabular}\n\\caption{Comparison of various methods across various GLUE~ and SQuAD tasks. Please refer~ for detailed description of tasks. Top part (first three rows) shows results for basic Transformer methods (teacher models). Middle part shows results for knowledge distillation methods. Bottom part shows results for a mix of other methods across categories. BERT-L=BERT-large, BERT-B=BERT-base, BERT-$i$=$i$-layer BERT. MCC refers to Matthews correlation. Results for SQuAD are on dev set. Empty entries indicate that the papers do not report those results, or NA entries.}\n \\label{tab:kdSummary2}\n\\end{table}\nTable~\\ref{tab:kdSummary2} compares various knowledge distillation methods (besides other model compression methods) across various GLUE~ and SQuAD tasks. Also, we report model size for both the student as well as the teacher models. Different distillation methods use one of these as the teacher: BERT-large, BERT-base or MTDNN. PKD~ outperforms the Vanilla KD~ on almost all the datasets except for MRPC. TinyBERT-4 significantly outperforms the 4-layer BERT-PKD and DistilBERT by a margin of at least 4.4\\%, with ~28\\% parameters and 3.1x inference speedup. Compared with the teacher BERT-base, 4-layer TinyBERT is 7.5x smaller and 9.4x faster in the model efficiency, while maintaining competitive performances. The 6-layer TinyBERT achieves comparable results with the teacher. Overall, TinyBERT consistently outperforms both the 4-layer and 6-layer baselines like PKD, DistilBERT and MiniLM.\nTurc et al.~ show how appropriate pretraining can improve the quality of distillation. MiniLM outperforms DistilBERT and TinyBERT across most tasks. The 6-layer MiniLM is 2.0x faster than original BERTBASE, while retaining more than 99\\% performance on a variety of tasks, such as SQuAD 2.0 and MNLI. Distilled MT-DNN significantly outperforms the original MT-DNN on 7 out of 9 GLUE tasks. 4-layer BERT-EMD outperforms 4-layer DistilBERT and BERT-PKD by a substantial margin, even with only 30\\% parameters and inference time. Furthermore, it exceeds the TinyBERT model by 2.3\\% accuracy on RTE, 2.2\\% F1 on MRPC, and 1.9\\% Spearman correlation on STS-B. 6-layer BERT-EMD performs better than the 12-layer BERT-base model on 7 out of 9 tasks, with only about 50\\% parameters and inference time of the original BERT-base model. Tang et al.~ distill BERT to BiLSTMs. They observe that the distilled BiLSTM model uses 349 times fewer parameters than BERT-large and is 434 times faster. Also, mixed vocab training by Zhao et al.~ produces a small 12-layer model that performs competitively with 6-layer PKD and 4-layer DistilBERT while being $\\sim$5-6x smaller.\nMobileBERT is 4.3x smaller and 5.5x faster than BERT-base. On the SQuAD v1.1/v2.0 question answering task, MobileBERT achieves a dev F1 score of 90.0/79.2 (1.5/2.1 higher than BERT-base). On the natural language inference tasks of GLUE, MobileBERT can achieve a GLUE score of 77.7, which is only 0.6 lower than BERT-base, with a latency of 62 ms on a Pixel 4 phone. While quantization can further compress MobileBERT by 4x, there is nearly no performance degradation from it. SqueezeBERT is approximately half the size of BERT-base. MobileBERT is half the size of SqueezeBERT. SqueezeBERT is 4.3x faster than BERT-base, while MobileBERT is 3.0x faster than BERT-base. On 4 of GLUE tasks SqueezeBERT outperforms the accuracy of MobileBERT, while on the other 4 of the GLUE tasks MobileBERT outperforms SqueezeBERT. MobileBERT and SqueezeBERT outperform BERT-base significantly across all tasks.\nTo summarize, KD is a popular method for text based model compression. Various methods have proposed information copying using logits, softmax output, attention sub-layer output, value relation, relative dissimilarity information from both the last layer as well as intermediate layers of the teacher. Many methods have been proposed to handle complex teacher-student configuration mismatches in terms of vocabulary, number of attention heads, and hidden layer sizes. Also, KD has been found to be very effective in complex problem settings like multi-lingual tasks and tasks with large number of classes. Learning from noisy teachers, teacher assistants, an ensemble of teachers has been found to be effective as well. KD is the best model compression method especially in settings where a large amount of unlabeled data exists; distillation with data pseudo-labeled by teacher leads to very effective students.","id":"4cd2b9eb-147c-41ba-996b-a30e824f168f","level":"subsection","origin_cites_number":27,"parent_id":"2c5b1b2b-3f74-442d-8ee6-503425cf4207","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Knowledge Distillation (KD)"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:parameterSharing}\nRather than removing weights or reducing \\#bits to store them, parameter sharing methods reduce model size by finding weight blocks that can share the same weight. Character-based language models learn embeddings for characters and use them to compose word embeddings. In some senses, we can think of various words sharing these character embedding parameters. Further, various parameter sharing methods have been proposed to reduce the large word embedding matrix size. Finally, there are multiple Transformer architectures which benefit from the parameter sharing philosophy. We discuss these methods in this section.","id":"57903397-1a0f-4365-9fd9-a6bf95e76421","level":"section","origin_cites_number":0,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Parameter sharing"]],"subsections":["1d632f6e-824d-41b6-8395-ec40e6c541ab","e8e447ef-9838-46fc-a0ca-4f8617f32377","82aba295-2827-4559-b383-3667b9bd0577","736e641f-1998-44af-a324-58eb82e0693d"],"title":"Parameter sharing"},{"cite_extract_rate":1,"cites":[4516,4517,4036],"content":"Fig.~\\ref{fig:charCNN} illustrates various character-aware language model architectures. Ling et al.~ proposed their character to word (C2W) model which constructs vector representations of words by composing characters using BiLSTMs. Relative to traditional word representation models that have independent vectors for each word type, C2W requires only a single vector per character type and a fixed set of parameters for the compositional model. As input, we define an alphabet of characters $C$. For English, this vocabulary would contain an entry for each uppercase and lowercase letter as well as numbers and punctuation. Thus compared to the word embedding matrix, this model is much smaller. Despite the compactness of this model, this ``composed'' word representations yield comparable results across multiple text classification tasks. \nJozefowicz et al.~ propose two variants for composing word embeddings using character embeddings. In the first CNN-Softmax variant, they use character CNNs (Convolutional Neural Networks) to compose word embeddings from character embeddings both at the input side as well as at the output softmax layer. The character-CNN sub-networks at the input or the output do not share weights. The composed word embeddings are fed to an LSTM to generate the output. In the second Char-LSTM variant, character CNN is used to compose word embeddings on the input side. The composed word embeddings are fed to an LSTM to generate an output which is further fed to a small LSTM that predicts the target word one character at a time. Thus, the word and character-level models are combined, and predictions are made one character at a time, thus allowing to compute probabilities over a much smaller vocabulary. Kim et al.~ propose another variant where at the output side they continue to use word embeddings, but at the input side they compose word embeddings using a highway network on top of a character CNN. The highway network's output is used as the input to a multi-layer LSTM, whose last hidden state output is fed to the output softmax layer.\n\\begin{figure}\n \\centering\n \\includegraphics[width=0.6\\columnwidth]{charCNN}\n \\caption{Character-aware Language Models}\n \\label{fig:charCNN}\n\\end{figure}","id":"1d632f6e-824d-41b6-8395-ec40e6c541ab","level":"subsection","origin_cites_number":3,"parent_id":"57903397-1a0f-4365-9fd9-a6bf95e76421","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Parameter sharing"],["subsection","Character-aware Language Models"]],"subsections":[],"title":"Character-aware Language Models"},{"cite_extract_rate":0.6666666666666661,"cites":[4522,4520,4036,4518,4521,4519],"content":"Given a weight matrix $W$ and a budget $K$, we want to share weights within $W$ to have a max of $K$ unique values. A na\\\"ive implementation of random weight sharing can be trivially achieved by maintaining a secondary matrix consisting of each connection's group assignment. But this needs memory space itself. Hence, Chen et al.~ propose to use hashing. HashedNets use a low-cost hash function (like xxhash\\footnote{\\url{https://code.google.com/p/xxhash/\n}}) to randomly group connection weights into hash buckets, and all connections within the same hash bucket share a single parameter value.\nUnlike HashedNets where weights are randomly grouped, parameter sharing mechanisms in Toeplitz-like structured matrices~ are highly specific and deterministic. Toeplitz matrices have parameters tied along diagonals. The displacement rank of all Toeplitz matrices is up to 2. Toeplitz-like matrices allow the displacement rank $r$ to be higher. They include products and inverses of Toeplitz matrices, and their linear combinations. The displacement rank $r$ serves as a knob on modeling capacity. High displacement rank matrices are increasingly unstructured. With displacement rank $r$, there are $2nr$ free parameters in the Toeplitz-like structured matrix. Toeplitz transforms can be applied not just to embedding matrix but to all weight matrices in an RNN model. Tay et al.~ use a similar Toeplitz-like structured matrix method with Hamilton Products in Quaternion Algebra to propose Quaternion Transformers which lead to 75\\% parameter reduction in the Transformer architecture.\nAnother method for parameter sharing is to share low-rank factors across layers in a recurrent model. In this method, we first represent a weight matrix $W$ using matrix factorization as $W=W_a W_b$. Thus, hidden layer output for layer $l$ at time $t$ can be written as follows.\n\\begin{eqnarray}\nh_t^l=\\sigma\\left[W_a^l W_b^l h_t^{l-1}+U_a^l U_b^l h_{t-1}^l + b^l\\right] \n\\end{eqnarray}\nBut we can share some low-rank factors by setting $W_b^l=U_b^{l-1}$. The combination of matrix factorization and parameter sharing leads to large model compression.\nAnother way of compressing the embedding matrix is to divide the vocabulary $V$ into frequent and infrequent word sets $B$ and $C$ respectively. Infrequent words' embeddings are represented with frequent words' by sparse linear combinations~. This is inspired by the observation that, in a dictionary, an unfamiliar word is typically defined by common words. A dense embedding is assigned to each common word; an infrequent word, on the other hand, computes its vector representation by a sparse combination of common words' embeddings. This compression is useful for both word embedding matrix as well as output layer of RNNs/LSTMs. Let $U\\in R^{E\\times |B|}$ be the learned embedding matrix of common words where $E$ is the embedding dimension. For a word $w\\in C$, we shall learn a sparse vector $x\\in R^{|B|}$ as the sparse code of the word. Once we know $x$, embedding for a word $w\\in C$ can be written as follows.\n\\begin{eqnarray}\n \\text{embedding}(w)=\\sum_{j=1}^B x_jU_j\n\\end{eqnarray}\nwhere $U_j$ is the $j^{th}$ column of $U$. To learn the sparse representation of word $w\\in C$, the following problem needs to be solved.\n\\begin{eqnarray}\n\\min_x ||Ux-A||_2^2+\\alpha ||x||_1+\\beta |1^Tx-1|+\\gamma 1^T \\max(0,-x)\n\\end{eqnarray}\nwhere $A$ is embedding for the rare word $w$. The last two regularization terms favor a solution that sums to 1 and that is non-negative (for psychological interpretation concerns), respectively. \nLightRNN~ compresses word embedding matrix from $O(|V|)$ to $O(\\sqrt{|V|})$. It uses a 2-Component shared embedding for word representations. We allocate every word in the vocabulary into a word-allocation table, each row of which is associated with a learned vector, and each column associated with another learned vector. Table~\\ref{tab:wordAllocation} shows an example of a word allocation table. Depending on its position in the table, a word is jointly represented by two components: a row vector and a column vector. Thus, we only need $2\\sqrt{|V|}$ vectors to represent a vocabulary of $|V|$ unique words, which are far less than the $|V|$ vectors. \nThe input and output use different embedding row/column vectors but they share the same word-allocation table. Word Allocation table creation uses a bootstrap procedure to iteratively refine word allocation based on the learned word embedding. Embeddings (i.e. row and column vectors) are learned using language modeling loss using an RNN on top of the embedding layer.\n\\begin{table}\n \\centering\n \\begin{tabular}{|c|c|c|c|c|}\n \\hline\nEmbedding&$x_1^c$&$x_2^c$&$x_3^c$&$x_4^c$\\\\\n\\hline\n$x_1^r$&january&february&...&...\\\\\n\\hline\n$x_2^r$&one&two&...&...\\\\\n\\hline\n$x_3^r$&...&...&...&...\\\\\n\\hline\n$x_4^r$&...&...&...&...\\\\\n\\hline\n\\end{tabular}\n \\caption{An Example of a Word Allocation Table}\n \\label{tab:wordAllocation}\n\\end{table}\nFinally, Suzuki et al.~ propose a Skipgram~ training method with parameter sharing as follows. Split every embedding vector of size $D$ into $B$ equal sub-vectors of size $C$. Thus $D=B\\times C$. We assign a limited number of reference vectors to each block of block-splitting vectors. E.g., the number of reference vectors becomes $K\\times B$ if we assign $K$ reference vectors to each block. Each reference vector is of size $C$. Skipgram training optimization remains the same except for these extra parameter sharing constraints (applied to both the input and output embedding vectors). Liu et al.~ propose a very similar method, Slim Embeddings, where the embeddings are learned using a RNN language model rather than the Skipgram method. Slim Embeddings is very related to HashedNets~, LightRNN~ and\nCharacter Aware Language model~. In HashedNets, all elements in a parameter matrix are mapped into a vector through a hash function. However in Slim Embeddings approach, we randomly share subvectors instead of single elements. Slim Embeddings differs from LightRNN in\nthat it is able to control the compression ratio\nto any arbitrary value, while LightRNN can only compress\nat the rate of square or cube root of vocabulary size, which\ncould be too harsh in practical applications. In character aware language model, if we treat the sequence of subvector ids (virtual characters) as each word's representation,\nthe word embedding then can be treated as concatenated unigram character feature vectors. The drawback of such an approach is that the model is\nmore complicated and to speed up inference, it needs to pre-compute the word embeddings for the words, so it couldn't\nstay in its compact form during inference. The Slim embeddings model is much simpler, and easier to tune. And during inference, it uses much less space and can even decrease the complexity of inference.","id":"e8e447ef-9838-46fc-a0ca-4f8617f32377","level":"subsection","origin_cites_number":9,"parent_id":"57903397-1a0f-4365-9fd9-a6bf95e76421","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Parameter sharing"],["subsection","Parameter Sharing in the Embedding Matrix"]],"subsections":[],"title":"Parameter Sharing in the Embedding Matrix"},{"cite_extract_rate":0.6666666666666661,"cites":[1150,4523],"content":"A standard Transformer does not share parameters across layers and also has a fixed number of encoder layers. ALBERT~ incorporates two parameter reduction techniques: \n\\begin{itemize}\n \\item Factorized embedding parameterization. That is, it decomposes large vocabulary embedding matrix into two small matrices. Thus, it reduces the embedding parameters from $O(V \\times H)$ to $O(V \\times E + E \\times H)$ where $H >> E$.\n \\item Cross-layer parameter sharing: There are multiple ways to share parameters, e.g., only sharing feed-forward network (FFN) parameters across layers, or only sharing attention parameters. The default decision for ALBERT is to share all parameters across layers.\n\\end{itemize}\nAn ALBERT configuration similar to BERT-large has 18x fewer parameters and can be trained about 1.7x faster. Dehghani et al.~ propose Universal Transformers where the number of encoder layers are not pre-decided, and all the encoder layers share the parameters. Certain symbols (e.g. some words or phonemes) are usually more ambiguous than others. It is therefore reasonable to allocate more processing resources to these more ambiguous symbols. Thus, ambiguous symbols undergo more self-attention transformations compared to non-ambiguous ones. Thus, they provide a dynamic per-position halting mechanism for dynamically modulating the number of computational steps needed to process each input symbol (called the ``ponder time'') before the representation is passed on as input to the decoder. The idea of sharing weights across layers in Transformers has also been explored in~.","id":"82aba295-2827-4559-b383-3667b9bd0577","level":"subsection","origin_cites_number":3,"parent_id":"57903397-1a0f-4365-9fd9-a6bf95e76421","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Parameter sharing"],["subsection","Parameter Sharing in Transformers"]],"subsections":[],"title":"Parameter Sharing in Transformers"},{"cite_extract_rate":0.7000000000000001,"cites":[4522,4516,4036,1150,4517,4521,4519],"content":"\\begin{table}\n \\centering\n \\scriptsize\n \\begin{tabular}{|l|p{0.8in}|l|l|l|l|l|l|}\n \\hline\nTask&Dataset&Method&Base Model&Metric&Size (Comp; Orig)&Eval. (Comp; Orig)\n\\\\\n\\hline\n\\hline\nLanguage modeling&1B Word Benchmark&Char-CNN (input embeddings)~&2-layer word-BiLSTM&Perplexity (L)&1.04B; 1.8B&30.0; 30.6\\\\\n\\hline\nLanguage modeling&1B Word Benchmark&Char-CNN (in/output embeddings)~&2-layer word-BiLSTM&Perplexity (L)&0.39B; 1.8B&35.8; 30.6\\\\\n\\hline\nLanguage modeling&1B Word Benchmark&LightRNN~&word-LSTM&Perplexity (L)&41M; 1.6G&66; 85\\\\\n\\hline\nLanguage modeling&1B Word Benchmark&Slim Embeddings~&2-layer word-BiLSTM&Perplexity (L)&0.25B; 1.8B&38.3; 30.6\\\\\n\\hline\nLanguage modeling&ACLW-Czech&CNN+Highway network~&word-LSTM&Perplexity (L)&64M; 83M&578; 701\\\\\n\\hline\nLanguage modeling&ACLW-Czech&LightRNN~&word-LSTM&Perplexity (L)&18M; 83M&558; 701\\\\\n\\hline\nLanguage modeling&ACLW-Czech&Slim Embeddings~&word-LSTM&Perplexity (L)&17M; 83M&528; 701\\\\\n\\hline\nLanguage modeling&ACLW-English&CNN+Highway network~&word-LSTM&Perplexity (L)&20M; 25M&216; 236\\\\\n\\hline\nLanguage modeling&ACLW-English&LightRNN~&word-LSTM&Perplexity (L)&17M; 25M&191; 236\\\\\n\\hline\nLanguage modeling&ACLW-English&Slim Embeddings~&word-LSTM&Perplexity (L)&7M; 25M&187; 236\\\\\n\\hline\nLanguage modeling&ACLW-French &CNN+Highway network~&word-LSTM&Perplexity (L)&44M; 56M&190; 202\\\\\n\\hline\nLanguage modeling&ACLW-French &LightRNN~&word-LSTM&Perplexity (L)&17M; 56M&176; 202\\\\\n\\hline\nLanguage modeling&ACLW-French &Slim Embeddings~&word-LSTM&Perplexity (L)&12M; 56M&162; 202\\\\\n\\hline\nLanguage modeling&ACLW-German &CNN+Highway network~&word-LSTM&Perplexity (L)&104M; 137M&305; 347\\\\\n\\hline\nLanguage modeling&ACLW-German &LightRNN~&word-LSTM&Perplexity (L)&18M; 137M&281; 347\\\\\n\\hline\nLanguage modeling&ACLW-German &Slim Embeddings~&word-LSTM&Perplexity (L)&17M; 137M&261; 347\\\\\n\\hline\nLanguage modeling&ACLW-Russian&CNN+Highway network~&word-LSTM&Perplexity (L)&152M; 200M&313; 353\\\\\n\\hline\nLanguage modeling&ACLW-Russian&LightRNN~&word-LSTM&Perplexity (L)&19M; 200M&288; 353\\\\\n\\hline\nLanguage modeling&ACLW-Russian&Slim Embeddings~&word-LSTM&Perplexity (L)&19M; 200M&274; 353\\\\\n\\hline\nLanguage modeling&ACLW-Spanish&CNN+Highway network~&word-LSTM&Perplexity (L)&48M; 61M&169; 186\\\\\n\\hline\nLanguage modeling&ACLW-Spanish&LightRNN~&word-LSTM&Perplexity (L)&18M; 61M&157; 186\\\\\n\\hline\nLanguage modeling&ACLW-Spanish&Slim Embeddings~&word-LSTM&Perplexity (L)&8M; 61M&149; 186\\\\\n\\hline\nLanguage modeling&PTB&CNN+Highway network~&word-LSTM&Perplexity (L)&5M; 20M&92.3; 85.4\\\\\n\\hline\nLanguage modeling&Wikipedia articles (ca)&C2W~&word emb.&Perplexity (L)&182K; 4.3M&34.92; 35.34\\\\\n\\hline\nLanguage modeling&Wikipedia articles (de)&C2W~&word emb.&Perplexity (L)&183K; 6.3M&41.94; 43.02\\\\\n\\hline\nLanguage modeling&Wikipedia articles (en)&C2W~&word emb.&Perplexity (L)&180K; 4.3M&57.39; 59.38\\\\\n\\hline\nLanguage modeling&Wikipedia articles (pt)&C2W~&word emb.&Perplexity (L)&178K; 4.2M&40.92; 46.17\\\\\n\\hline\nLanguage modeling&Wikipedia articles (tr)&C2W~&word emb.&Perplexity (L)&174K; 5.7M&32.88; 44.01\\\\\n\\hline\nMachine reading comp.&ReAding Comprehension from Examinations&ALBERT~&BERT-B&Acc (H)&18M; 108M&68.5; 68.2\\\\\n\\hline\nNLI&MNLI&Quaternion Attention~&2-layer att-GloVe&Acc (H)&200K; 700K&72.3; 73.6\\\\\n\\hline\nNLI&SciTail&Quaternion Attention~&2-layer att-GloVe&Acc (H)&200K; 700K&79.6; 79.0\\\\\n\\hline\nNLI&SNLI&Quaternion Attention~&2-layer att-GloVe&Acc (H)&200K; 700K&85.4; 86.2\\\\\n\\hline\nNMT (en$\\rightarrow$et)&IWSLT15&Quaternion Transformer~&Transformer&BLEU (H)&11M; 44M&13.1; 14.1\\\\\n\\hline\nNMT (en$\\rightarrow$ro)&IWSLT15&Quaternion Transformer~&Transformer&BLEU (H)&11M; 44M&18.5; 22.8\\\\\n\\hline\nNMT (en$\\rightarrow$vi)&IWSLT15&Quaternion Transformer~&Transformer&BLEU (H)&11M; 44M&28.0; 28.4\\\\\n\\hline\nPOS Tagging&PTB (ca)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&98.92; 98.09\\\\\n\\hline\nPOS Tagging&PTB (de)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&98.08; 97.51\\\\\n\\hline\nPOS Tagging&PTB (en)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&97.36; 96.97\\\\\n\\hline\nPOS Tagging&PTB (pt)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&97.47; 95.67\\\\\n\\hline\nPOS Tagging&PTB (tr)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&91.59; 83.43\\\\\n\\hline\nQuestion answering&BABI&Universal Transformer~&Transformer&Avg error (L)&7.3M; 44M&0.21; 15.2\\\\\n\\hline\nQuestion answering&WikiQA&Quaternion Attention~&2-layer att-GloVe&MAP (H)&200K; 700K&66.2; 67.2\\\\\n\\hline\nSentiment analysis&IMDB&Quaternion Transformer~&2-layer Transformer&Acc (H)&100K; 400K&83.9; 82.6\\\\\n\\hline\nSentiment analysis&SST&Quaternion Transformer~&2-layer Transformer&Acc (H)&100K; 400K&80.5; 78.9\\\\\n\\hline\nSpeech recognition&2000 hour En. Speech&Toeplitz-like~&3-layer RNN&WER (L)&790K; 1.85M&48.4; 43.5\\\\\n\\hline\nSpeech recognition&2000 hour En. Speech&Toeplitz-like~&5-layer LSTM&WER (L)&2.00M; 9.12M&33.5; 33.1\\\\\n\\hline\nSubject verb agreement&SVA dataset&Quaternion Transformer~&2-layer Transformer&Acc (H)&100K; 400K&94.7; 94.8\\\\\n\\hline\nSubject verb agreement&SVA dataset&Universal Transformer~&Transformer&Acc (H)&7.3M; 44M&0.992; 0.962\\\\\n\\hline\nWord analogy &GSEm, GSYN, MSYN&Shared ref. vec. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&64.6/64.5; 64.7\\\\\n\\hline\nWord analogy &GSEm, GSYN, MSYN&k-means quant. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&64.0/64.5; 64.7\\\\\n\\hline\nWord similarity&7 datasets&Shared ref. vec. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&65.5/67.1; 67.2\\\\\n\\hline\nWord similarity&7 datasets&k-means quant. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&64.4/67.0; 67.2\\\\\n\\hline\n \\end{tabular}\n\\caption{Comparison of various parameter sharing methods (sorted by Task and then Dataset). 7 datasets for word similarity are MEN, MTurk, RARE, SLex, SCWS, WSR, WSS. SGNS=SkipGram with negative sampling. In the metric column, H means high is better while L means low is better.}\n \\label{tab:paramSharingSummary}\n\\end{table}\nTable~\\ref{tab:paramSharingSummary} compares various parameter sharing methods across different tasks and datasets. Accuracy of both the original and the compressed (comp.) model are shown. Also, we report model size (in terms of number of parameters or memory size) for both the original as well as the compressed models. For the same task, dataset and model combination, different papers report different accuracy of the original model because of slight changes in training hyper-parameters; hence we report accuracy of the original model for each row. Since many parameter sharing methods have been used to compress word embeddings, the most common application is language modeling. \nFor language modeling, experiments have been done on the One Billion Word Benchmark, ACLW, PTB and Wikipedia articles datasets across multiple languages using parameter sharing methods like Char-CNN~, LightRNN~, Slim embeddings~, CNN+Highway networks~ and C2W~. C2W~ always outperforms word lookup tables and improvements are especially pronounced in Turkish, which is a highly morphological language, where word meanings differ radically depending on the suffixes used. Jozefowicz et al.~ observe that Char-CNNs especially with character embeddings being used both at input as well as output can lead to 4.6x model size reduction with a slight increase in perplexity. On the One-Billion-Word benchmark, LightRNN~ achieves much lower perplexity compared to word-LSTMs, whilst reducing the model size by a factor of 40-100, and speeding up the training process by a factor of 2. LightRNN~ significantly reduces the model size, while at the same time outperforms CNN+Highway network~ method. While the model sizes of the CNN+Highway network method increase linearly with respect to the vocabulary size, the model size of LightRNN almost keeps constant on the ACLW datasets. Slim embeddings~ is way better than LightRNN (low perplexity with smaller models). On PTB and 44M giga word corpus datasets, Slim embeddings applied at input layer can maintain same perplexity for a word-LSTM using just 1\\% (and 0.2\\%) of trainable parameters respectively. \nk-means quantization and shared reference vectors are also methods for compression of word embeddings using parameter sharing. Suzuki et al.~ showed significant gains over typical skipgram (SGNS) embeddings in terms of model size reduction while retaining similar accuracies for word analogy and similarity tasks. The model size of their method with shared reference vectors with `K = 16, B = 64' was just 24MB, approximately 65 times smaller than that of original SGNS method. They also showed that SGNS with shared reference vectors was better than SGNS with block-wise k-means post-processing method. Unfortunately, there exists no good comparison between the slim embeddings and the shared reference vectors methods.\nFor the MRC task, ALBERT~ pushed the accuracy from 68.5 to 68.2 using 6x fewer parameters compared to BERT-base. On the NLI task, a tiny Quaternion-Att model (50 dimensions) achieves comparable (or occasionally marginally better or worse) performance compared to typical attention over GloVe (200 dimensions), gaining a 68\\% parameter savings across three datasets. For sentiment analysis, Quaternion Transformers~ leads by +1.3\\%/1.6\\% gains on IMDb and SST datasets respectively while maintaining a 75\\% reduction in parameter cost. Quaternion Transformers~ have been shown to outperform the vanilla Transformer for the mathematical language understanding task as well, with 25\\% parameters. At the same compression rate, Quaternion Transformers lose only very minor BLEU on IWSLT 2015 NMT datasets.\nFor POS tagging, we can observe that the model using word lookup tables performs consistently worse than the C2W model~. Universal Transformers~ (while being 1/6th the size) outperform standard Transformers on a wide range of algorithmic and language understanding tasks, including the challenging LAMBADA language modeling task. On speech recognition, Lu et al.~ study mechanisms for learning compact RNNs and LSTMs via low-rank factorizations and parameter sharing schemes. A hybrid strategy of using structured matrices in the bottom layers and shared low-rank factors on the top layers is found to be particularly effective, reducing the parameters of a standard LSTM by 75\\%, at a small cost of 0.3\\% increase in WER, on a 2000-hr English Voice Search task. For LSTM parameter reduction, architecting upper layers with projection nodes to moderate rank, and bottom layers with Toeplitz-like transforms was found to be a particularly effective strategy. \nOverall, besides model compression, parameter sharing methods also act as a good regularizer. Parameter sharing in Transformers has been very successful. ALBERT was at the top of the GLUE leaderboard when it was proposed. Parameter sharing methods have also been widely used for compressing embedding matrix. Slim embeddings has the best method for compressing embedding matrices.","id":"736e641f-1998-44af-a324-58eb82e0693d","level":"subsection","origin_cites_number":10,"parent_id":"57903397-1a0f-4365-9fd9-a6bf95e76421","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Parameter sharing"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:matrixDecomposition}\nSparse Matrix decomposition has been traditionally used for applications like feature selection, collaborative filtering, topic mining from text, etc. In this section, we discuss how various popular tensor decomposition methods like Singular Value Decomposition (SVD), Tensor-Train~, CP (CANDECOMP/PARAFAC)~ and Tucker~ can be used for model compression.","id":"a059247f-141a-43b7-a6a9-f3440e7622cd","level":"section","origin_cites_number":3,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Tensor decomposition"]],"subsections":["da1505f4-583c-4b3b-a84f-338a8c53f73d","1b3ed6c2-9088-4bfb-8866-160b3884224a","4c1e5567-aabc-47e9-8344-624bb81f275f","2c213350-b4bf-49d2-a046-a4fe16413654"],"title":"Tensor decomposition"},{"cite_extract_rate":0.6000000000000001,"cites":[4515,4524,4525],"content":"In this part, we will discuss methods where a matrix is factorized into two low-rank factors. Specifically, we replace a weight matrix $W$ with $W_1\\times W_2$ such that the total number of parameters are significantly lesser.\nA multi-layer RNN can be represented as follows.\n\\begin{eqnarray}\n h_t^l=\\sigma(W_x^{l-1}h_t^{l-1}+W_h^l h_{t-1}^l +b^l)\\\\\n h_t^{l+1}=\\sigma(W_x^l h_t^l + W_h^{l+1}h_{t-1}^{l+1}+b^{l+1})\n\\end{eqnarray}\nThus, there are two important weight matrices: the recurrent $W_h^l$ and inter-layer matrices $W_l^x$. Prabhavalkar et al.~ propose a method to jointly compress the recurrent and inter-layer matrices corresponding to a specific layer $l$ by determining a suitable recurrent projection matrix, denoted by $P^l\\in R^{r_l\\times N_l}$ of rank $r^l<N^l$ such that $W_h^l=Z_h^l P^l$ and $W_l^x=Z_x^l P^l$. First, $P^l$ is determined by computing a truncated SVD of the recurrent weight matrix, which we then truncate, retaining only the top $r^l$ singular values. Thus, $W_h^l$ can be written as follows.\n\\begin{eqnarray}\n W_h^l=(U_h^l \\Sigma_h^l)(V_h^l)^T=Z_h^lP^l\n\\end{eqnarray}\nThus, $P^l$ is set to $(V_h^l)^T$. Further, we determine $Z_x^l$ as the solution to the following least-squares problem.\n\\begin{eqnarray}\n Z_x^l=\\argmin_Y ||YP^l-W_x^l||_2^2\n\\end{eqnarray}\nThis solution can also be easily extended to LSTMs. Sak et al.~ also proposed a similar solution based on a combination of parameter sharing and matrix decomposition but without SVD initialization. However, typically SVD initialization has been found to perform better.\nBesides SVD, another way of matrix decomposition is sparse coding. Faruqui et al.~ propose using sparse coding to decompose word embedding matrices. Thus, given vocabulary of size $V$, word embedding matrix $X\\in R^{L\\times V}$, sparse coding aims at representing each input vector $x_i$ as a sparse linear combination of basis vectors $a_i$ by solving the following problem. \n\\begin{eqnarray}\n \\argmin_{D,A} \\sum_{i=1}^V ||x_i-Da_i||_2^2+\\lambda||a_i||_1+\\tau ||D||_2^2\n\\end{eqnarray}\nwhere $D\\in R^{L\\times K}$ and $A\\in R^{K\\times V}$. Further, for interpretability, one can enforce all elements of $A$ and $D$ to be non-negative. For further compression, one can also enforce $A$ to be binary or ensure that each column of $A$ is a $K$ sized one hot vector~. \nLastly, Wang et al.~ combine pruning with matrix factorization for model compression and propose the FLOP (Factorized Low-rank Pruning) method. Let $W$ be a weight matrix. Structured pruning (removing a neuron, i.e., removing a column from weight matrix) can be achieved by replacing the computation $Wx$ by $WGx$ where diagonal sparsity-inducing matrix $G$ is learned using $L_0$ regularization over $WG$ along with the supervised loss. This effectively removes a subset of columns of $W$ for column indices $k$ with $z_k=0$. One limitation is that this structured pruning method tends to produce lower performance than its unstructured counterpart. Hence, in the FL0P (Factorized L0 Pruning) model, we first factorize $W=PQ$. Let $r$ be \\#columns of $P$ (or equivalently \\#rows of $Q$), $p_k$ and $q_k$ be the $k$-th column of $P$ and $k$-th row of $Q$ respectively. We achieve structured pruning by introducing a pruning variable $z_k$ for each component. Thus, now, we can write $W$ as follows.\n\\begin{eqnarray}\n W=PGQ=\\sum_{k=1}^r z_k\\times (p_kq_k)\n\\end{eqnarray}\nwhere $G$ is again the diagonal matrix of pruning variables. After training, only columns and rows corresponding to non-zero diagonal values need to be stored, resulting in much smaller (but still dense) matrices $P$ and $Q$. The nonzero values of $G$ can be absorbed into either $P$ or $Q$. This structured pruning with factorization is much more effective compared to the vanilla structured pruning.","id":"da1505f4-583c-4b3b-a84f-338a8c53f73d","level":"subsection","origin_cites_number":5,"parent_id":"a059247f-141a-43b7-a6a9-f3440e7622cd","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Tensor decomposition"],["subsection","Two Low-Rank Factors"]],"subsections":[],"title":"Two Low-Rank Factors"},{"cite_extract_rate":1,"cites":[4526,8800],"content":"The last layer of a language model is very large of the size $HV$ where $H$ is the size of the hidden layer and $V$ is vocabulary size. Each word by an output embedding of the same size $H$. Chen et al.~ propose a differentiated softmax method which varies the dimension of the output embeddings across words depending on how much model capacity is deemed suitable for a given word. In particular, it is meaningful to assign more parameters to frequent words than to rare words. By definition, frequent words occur more often in the training data than rare words and therefore allow to fit more parameters. They define partitions of the output vocabulary based on word frequency and the words in each partition share the same embedding size. Partitioning results in a sparse final weight matrix which arranges the embeddings of the output words in blocks, each one corresponding to a separate partition. The size of the final hidden layer $H$ is the sum of the embedding sizes of the partitions. While this method does not involve creation of multiple factors, it factorizes the original matrix into multiple blocks while setting the remaining part of the matrix to 0.\nVariani et al.~ propose a method called Word Encoded Sequence Transducers (WEST) which factorizes a matrix $E=C\\times D$ where $D$ is constrained to be a block diagonal matrix. The block diagonal nature of the second factor leads to large compression rates.","id":"1b3ed6c2-9088-4bfb-8866-160b3884224a","level":"subsection","origin_cites_number":2,"parent_id":"a059247f-141a-43b7-a6a9-f3440e7622cd","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Tensor decomposition"],["subsection","Factorizing into Block Diagonal Matrices"]],"subsections":[],"title":"Factorizing into Block Diagonal Matrices"},{"cite_extract_rate":0.8,"cites":[4529,4527,4530,4528],"content":"Tensor train decomposition (TTD)~ is a standard tensor decomposition technique which decomposes a high dimensional tensor into multiple 2D and 3D tensors which can be multiplied together to reconstruct the original tensor. These factors are called TT-cores and the other dimensions are referred to as TT-ranks. TTD can be leveraged to compress various weight matrices in RNNs and LSTMs~. The first step is to represent a matrix as a multi-dimensional tensor by simple reshaping transformation and then use TTD on it. The values of TT–ranks directly define the compression ratio, so choosing them to be too small or too large will result into either significant performance drop or little reduction of the number of parameters. Typically TT-ranks around 16 for small matrices and 64-192 for larger matrices result in a good trade-off between compression ratio and the accuracy metric of interest. Also, when we use TTD for weight matrices, we also need change the inputs appropriately to be compatible in terms of dimensions. \nCompared with TT-RNN, Block-Term RNN (BTRNN)~ is not only more concise (when using the same rank), but also able to attain a better approximation to the original RNNs with much fewer parameters. BTD decomposes a high order tensor into a sum of multiple Tucker decomposition models. The redundant dense connections between input and hidden state is first tensorized to a $d$-dimensional tensor and then decomposed using low-rank BTD into a sum of $N$ different Tucker decompositions where $N$ is the CP-rank. Each Tucker decomposition in turn consists of a core $d$-dimensional tensor and $d$ 3-dimensional factor tensors. While Ye et al.~ used BTD to compress RNNs, Ma et al.~ used BTD to compress the self-attention matrix in Transformers. They first build a single-block attention based on the Tucker decomposition where the query, key and value are mapped into three factor matrices and the core tensor is trainable and randomly initialized. It is then straightforward to represent the multi-head attention using BTD.","id":"4c1e5567-aabc-47e9-8344-624bb81f275f","level":"subsection","origin_cites_number":5,"parent_id":"a059247f-141a-43b7-a6a9-f3440e7622cd","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Tensor decomposition"],["subsection","Tensor Train and Block Term Decomposition"]],"subsections":[],"title":"Tensor Train and Block Term Decomposition"},{"cite_extract_rate":0.777777777777777,"cites":[4527,4530,1568,4515,4524,4525,4526],"content":"\\begin{table}\n \\centering\n \\scriptsize\n \\begin{tabular}{|l|l|l|l|l|l|l|l|}\n \\hline\nTask&Dataset&Method&Base Model&Metric&Size (Comp; Orig)&Eval. (Comp; Orig)\n\\\\\n\\hline\n\\hline\nCTR prediction&Criteo CTR Challenge&TT-embedding~&MLP&LogLoss (L)&4.7M; 41.2M&0.4433; 0.4440\\\\\n\\hline\nLanguage modeling&1B Word Benchmark&BTD~&Transformer-XL Large&Perplexity (L)&0.16B; 0.8B&19.5; 21.8\\\\\n\\hline\nLanguage modeling&Enwiki-8&FLOP~&Transformer&BPC (L)&8M; 41M&1.13; 1.08\\\\\n\\hline\nLanguage modeling&PTB&WEST~&LSTM&Perplexity (L)&3.5M; 4.51M&116.84; 115.91\\\\\n\\hline\nLanguage modeling&PTB&BTD~&Transformer-XL-Base&Perplexity (L)&12M; 24M&49.8; 54.52\\\\\n\\hline\nLanguage modeling&PTB&TT-embedding~&Transformer-XL-Base&Perplexity (L)&18M; 24M&55.4; 54.52\\\\\n\\hline\nLanguage modeling&WikiText-103&FLOP~&Transformer&Perplexity (L)&50M; 151M&25.3; 24.1\\\\\n\\hline\nLanguage modeling&WikiText-103&TT-embedding~&Transformer-XL&Perplexity (L)&91M; 192M&25.67; 24.37\\\\\n\\hline\nLanguage modeling&WikiText-103&BTD~&Transformer-XL-Base&Perplexity (L)&85.3M; 151M&20.9; 24.0\\\\\n\\hline\nLanguage modeling&WikiText-103&TT-embedding~&Transformer-XL-Base&Perplexity (L)&130M; 151M&25.7; 24.0\\\\\n\\hline\nNMT (en$\\rightarrow$ja)&ASPEC&Compositional codes~&LSTM&BLEU (H)&2.97MB; 274MB&38.89; 37.93\\\\\n\\hline\nNMT (de$\\rightarrow$en)&IWSLT14&Compositional codes~&LSTM&BLEU (H)&2.11MB; 35MB&29.56; 29.45\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT14&TT-embedding~&Transformer-Big&BLEU (H)&179M; 210M&28.53; 28.84\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT16&BTD~&Transformer&BLEU (H)&21.2M; 52M&34.91; 34.5\\\\\n\\hline\nNP bracketing&Subset of PTB&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&82.3; 77.9\\\\\n\\hline\nQuestion classification&TREC Questions&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&81.5; 76.2\\\\\n\\hline\nSentiment analysis&IMDB&Compositional codes~&LSTM&Acc (H)&1.23MB; 78MB&87.37; 87.18\\\\\n\\hline\nSentiment analysis&IMDB&TT-embedding~&LSTM&Acc (H)&0.81M; 7.19M&89.7; 88.6\\\\\n\\hline\nSentiment analysis&SST&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&81.4; 77.7\\\\\n\\hline\nSpeech recognition&3M Google voice utterances&Joint-SVD~&5-layer RNN&WER (L)&3.1M; 9.7M&12.9; 12.4\\\\\n\\hline\nSpeech recognition&3M Google voice utterances&Projections~&LSTM&WER (L)&2M; 2.2M&14.8; 17.5\\\\\n\\hline\nSpeech recognition&Live traffic utterances&WEST~&3-layer LSTM&WER (L)&4.75MB; 15MB&13.6; 13.7\\\\\n\\hline\nSpeech recognition&Live traffic utterances&WEST+Quantization~&3-layer LSTM&WER (L)&1.35MB; 15MB&13.7; 13.7\\\\\n\\hline\nText classification&20 Newsgroup (Computers)&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&87.0; 79.7\\\\\n\\hline\nText classification&20 Newsgroup (Religion)&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&88.8; 86.7\\\\\n\\hline\nText classification&20 Newsgroup (Sports)&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&96.3; 95.9\\\\\n\\hline\nWord similarity&Simlex-999&Sparse coding~&Logistic regression&Correlation (H)&120M; 120M&38.9; 36.9\\\\\n\\hline\n \\end{tabular}\n\\caption{Comparison of various tensor decomposition methods (sorted by Task and then Dataset). In the metric column, H means high is better while L means low is better. For compositional codes, 16x32 coding was used. For BTD, two block term tensors were used. In~, logistic regression uses GloVe embeddings.}\n \\label{tab:tensorDecompSummary}\n\\end{table}\nTable~\\ref{tab:tensorDecompSummary} compares various tensor decomposition methods across different tasks and datasets. Accuracy of both the original and the compressed (comp.) model are shown. Also, we report model size (in terms of number of parameters or memory size) for both the student as well as the teacher models. For the same task, dataset and model combination, different papers report different accuracy of the original model because of slight changes in training hyper-parameters; hence we report accuracy of the original model for each row. \nFor CTR prediction, the TT-embedding method~ in Table~\\ref{tab:tensorDecompSummary} uses 3 factors with TT-rank of 16. It actually leads to an embedding compression of 16. With 4 factors and TT-rank=2, test loss increases to 0.4530 with a massive embedding compression of 4193 and overall model size of 0.53M. Thus, substitution of large embedding layers with TT–embeddings leads to significant compression ratios (up to 2011 times) with a slight improvement in the test loss, and up to 4200 with a small drop in the test loss. \nFor language modeling, BTD~ leads to an improved model with 20\\% of the Transformer-XL large model. For character level language modeling using FLOP~, an 8M sized FLOP model achieves 1.13 on Enwiki8 while gradual pruning achieves 1.14. Thus, FLOP based pruning which combines pruning with matrix factorization is better than both structured neuron pruning as well as gradual unstructured pruning. On PTB, BTD is clearly much better than TT-embedding. With half the model size compared to Transformer-XL-Base, BTD leads to a model with $\\sim$10\\% lower perplexity. On WikiText-103, while FLOP pruned model achieves 25.3 perplexity with model size of 50M, gradual unstructured pruning and neuron pruning achieve 25.7 and 26.7 respectively with the same model size for language modeling on Wiki-103 dataset. Thus, FLOP is better than other pruning methods. Again on Wiki-103, BTD is superior to TT-embedding. \nFor NMT, the loss-free compression rate reaches 92\\% on ASPEC dataset by pruning 90\\% of the connections. However, with the same pruning ratio, a modest performance loss is observed in IWSLT14 dataset. For the models using compositional coding~, the loss-free compression rate is 94\\% for the IWSLT14 dataset and 99\\% for the ASPEC dataset. Thus, compositional codes are much better compared to pruning. TT-embedding with TT-rank=64 leads to embedding compression of 15.3 on WMT14 dataset with marginal loss in the BLEU score. Lastly, BTD has been used to reduce Transformer size by more than half with improved BLEU on WMT16 (en$\\rightarrow$de) data.\nSparse coding for GloVe vectors~ has led to improved accuracies for multiple tasks like Noun Phrase (NP) bracketing, question classification, text classification and word similarity, while retaining the same model size. For sentiment analysis on the IMDB dataset, compositional codes method achieves a compression rate of 98\\% without performance loss. Further, TT-embedding method leads to a much smaller model compared to compositional codes with better accuracy. In the TT-embedding case, embedding compression rate is 441 with TT-rank=16.\nFor Speech recognition, Prabhavalkar et al.~ experimented with a 3M Google voice utterances dataset and found that a joint SVD with explained variance retained after the SVD as 0.6 leads to a 3x smaller RNN model without much performance loss. They improved upon Sak et al.'s projections method which led to a much higher WER on the same dataset. On live traffic utterances dataset, WEST~ leads to a 3x smaller model with a slightly reduced word error rate when using 3-layer LSTMs. Variani et al.~ further compress this model using quantization to obtain a 11x compressed 3-layer LSTM model without any performance loss. Thus, using quantization along with matrix decomposition seems to work well.\nOverall, matrix decomposition techniques are usually used in combination with parameter sharing and sometimes with quantization. They have been very effective in dealing with large input/output embedding matrices in RNNs and LSTMs. SVD, Tensor Train, CP, Tucker, BTD have been the most popular decomposition techniques found to be useful for model compression.","id":"2c213350-b4bf-49d2-a046-a4fe16413654","level":"subsection","origin_cites_number":9,"parent_id":"a059247f-141a-43b7-a6a9-f3440e7622cd","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Tensor decomposition"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":1,"cites":[4531],"content":"\\label{sec:linearTransformers}\nMemory usage throughout neural network training can be categorized into three main types: (1) Model memory is used to store model parameters; (2) Optimizer memory is the additional memory used by the specific learning algorithm during the process; (3) Activation memory consists of the outputs of each layer, which are cached for reuse in backpropagation to compute gradients. For a BERT-base model, model memory is around 0.2 GB, optimizer memory is around 1GB, while activation memory is around 8.5GB~. Time and activation memory in Transformers grows quadratically with the sequence length. This is because in every layer, every attention head attempts to come up with a transformed representation for every position by ``paying attention'' to tokens at every other position. Quadratic complexity implies that practically the maximum input size is rather limited. Thus, we cannot extract semantic representation for long documents by passing them as input to Transformers.","id":"65404f84-cbf0-4721-8f05-178f8230aa55","level":"section","origin_cites_number":1,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Transformers with Sub-Quadratic Complexity"]],"subsections":["4db79252-def3-4c73-83ab-c7863b1703fc","f00ad6f8-f45a-4b14-ab1c-135e291338eb","2202feab-4da3-4fb5-8d59-2d1d4fbac3f9"],"title":"Transformers with Sub-Quadratic Complexity"},{"cite_extract_rate":1,"cites":[4531,793,794],"content":"A few efforts like BlockBERT~ try to reduce this quadratic complexity by a constant factor by introducing sparse block structures into the attention matrix. If we split the length-$N$ input sequence into $n$ blocks, $N\\times N$ attention matrix gets partitioned into $n\\times n$ blocks, where each block matrix is of the size $\\frac{N}{n}\\times\\frac{N}{n}$. Thus, BlockBERT reduces $O(N^2)$ memory consumption by a factor of $n$. \nChild et al.~ propose sparse transformers where sparse factorizations of the attention matrix reduce the quadratic complexity to $O(n\\sqrt{n})$. The key idea is to reduce the dense attention matrix to a sparse version by only computing attention on a sparse number of (query, key) pairs. They propose two kinds of sparse factorizations: strided and fixed. Strided attention implies having one head attend to the previous $l$ locations, and the other head attend to every $l^{th}$ location, where $l$ is the stride and chosen to be close to $\\sqrt{n}$. More heads could be used with a different stride value. Fixed attention assumes that specific positions summarize previous locations and propagate that information to all future positions. \nThe Reformer architecture~ replaces the dot-product attention in a typical Transformer by one that uses locality-sensitive hashing (LSH), changing its complexity from $O(n^2)$ to $O(n\\log n)$, where $n$ is the length of the sequence. In a standard Transformer, we compute scaled dot-product attention as follows.\n\\begin{eqnarray}\n \\text{Attention}(Q,K,V)=\\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d}}\\right)V\n\\end{eqnarray}\nwhere $Q$, $K$ and $V$ are the standard query, key and value components and $d$ is a scaling factor. Reformer uses a Shared QK Transformer, i.e., $Q=K$ enabled by sharing the matrix that projects words/hidden layer to $Q$ or $K$. Further, note that we are actually only interested in $\\text{softmax}(QK^T)$. Since softmax is dominated by the largest elements, for each query $q_i$ we only need to focus on the keys in $K$ that are closest to $q_i$. How can we find the nearest neighbors among the keys? Reformer uses LSH. LSH is used to cluster (hash-bucket) the positions into various groups, and then every position needs to focus only on others within the same bucket.","id":"4db79252-def3-4c73-83ab-c7863b1703fc","level":"subsection","origin_cites_number":3,"parent_id":"65404f84-cbf0-4721-8f05-178f8230aa55","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Transformers with Sub-Quadratic Complexity"],["subsection","Transformers with Super-Linear Complexity"]],"subsections":[],"title":"Transformers with Super-Linear Complexity"},{"cite_extract_rate":1,"cites":[7333,7298,798,7371,1473,7827],"content":"Even better, there have been several efforts recently to reduce this quadratic complexity to linear. Most of these efforts choose a constant number of other positions to ``pay attention'' to so as to compute a transformed representation for any given position. They can model sequences tens of thousands of timesteps long using hundreds of layers. The methods differ in their approach towards selecting this constant number of other positions. We discuss a few of such recently proposed methods in this section.\nIn Star-Transformers~, to reduce model complexity from $O(n^2)$ to linear, we replace the fully-connected attention matrix structure with a star-shaped topology, in which every two non-adjacent nodes are connected through a shared relay node. While ring connections connect a satellite node with two other satellite nodes, a radical connection connects a satellite node with the relay node. The idea is to update the star-center relay node based on satellite nodes and then update satellite nodes using information from the star node, and adjacent satellite nodes. \nLinformer architecture~ exploits low-rank factorization of the self-attention matrix to reduce overall self-attention complexity from $O(n^2)$ to $O(n)$ in both time and space. The main idea is to add two linear projection matrices $E_i, F_i\\in R^{n\\times k}$ when computing key and value. We first project the original $(n\\times d)$-dimensional key and value layers into $(k\\times d)$-dimensional projected key and value layers. We then compute an $(n \\times k)$-dimensional context mapping matrix using scaled dot-product attention. If we can choose a very small projected dimension $k$, such that $k<<n$, then we can significantly reduce the memory and space consumption. Overall, it is $O(nk)$. Further, we can do three other forms of parameter sharing: (1) Headwise sharing: $E_i=E$ and $F_i=F$ across all heads $i$ in a layer. (2) Key-value sharing: $E_i=F_i=E$ across all heads $i$ in a layer. (3) Layerwise sharing: Single projection matrix $E$ is used across all layers, all heads for both key and value.\nSparse Sinkhorn Attention based Transformer~ is based on differentiable sorting of internal representations. First, they divide the input sequence into $B$ equal sized blocks each of size $n/B$. A meta sorting network learns to generate latent permutations over these block sequences. Given sorted sequences, we are then able to compute quasi-global attention with only local windows, improving the memory efficiency of the attention module. They also propose Causal Sinkhorn Balancing and SortCut algorithms for causal scenarios for tailoring Sinkhorn Attention for encoding and/or decoding purposes. Their method reduces the memory complexity from $O(n^2)$ to $O(B^2+(n/B)^2)$. The SortCut variant further reduces complexity to linear-time, i.e., $O(nk)$ where $k$ is a user defined budget hyper-parameter much smaller than $n$.\nShen et al.~ propose a very simple mathematical trick to reduce quadratic complexity to linear. A typical dot-product attention can be written as $\\text{softmax}(QK^T)V$ ignoring the scale factor. This is quadratic because $QK^T$ is $n^2$ in size. This can be rewritten as follows.\n\\begin{eqnarray}\n \\text{Attention}(Q,K,V)=\\text{softmax}_r(Q)(\\text{softmax}_c(K^TV))\n\\end{eqnarray}\nwhere $\\text{softmax}_r$ and $\\text{softmax}_c$ are softmax applied to rows and columns respectively. This revised formulation has terms which are only linear in $n$. Finally, Katharopoulos et al.~ express the self-attention as a linear dot-product of kernel feature maps and make use of the associativity property of matrix products to reduce the complexity from $O(n^2)$ to $O(n)$, where $n$ is the sequence length.\nFinally, Longformer~ propose to reduce Transformer complexity to linear by sparsifying the full self-attention matrix using multiple kinds of attention patterns. The simplest attention pattern is the sliding window pattern which employs a fixed-size window attention surrounding each token. Given a fixed window size $w$, each token attends to $0.5w$ tokens on each side. The computation complexity of this pattern is $O(nw)$ which scales linearly with input sequence length $n$. To further increase the receptive field without increasing computation, the sliding window can be ``dilated''. This is analogous to dilated CNNs where the window has gaps of size dilation $d$. The windowed and dilated attention are not flexible enough to learn task-specific representations. Accordingly, Longformer also has ``global attention'' on few pre-selected input locations. Moreover, this attention operation is symmetric: that is, a token with a global attention attends to all tokens across the sequence, and all tokens in the sequence attend to it. Further, two different sets of projections are used: $Q_s$, $K_s$, $V_s$ to compute attention scores of sliding window attention, and $Q_g$, $K_g$, $V_g$ to compute attention scores for the global attention. The pretrained Longformer consistently outperforms RoBERTa on multiple downstream long document tasks.","id":"f00ad6f8-f45a-4b14-ab1c-135e291338eb","level":"subsection","origin_cites_number":6,"parent_id":"65404f84-cbf0-4721-8f05-178f8230aa55","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Transformers with Sub-Quadratic Complexity"],["subsection","Transformers with Linear Complexity"]],"subsections":[],"title":"Transformers with Linear Complexity"},{"cite_extract_rate":1,"cites":[7333,4531,7298,793,798,7371,794,1473],"content":"\\begin{table}\n \\centering\n \\scriptsize\n \\begin{tabular}{|l|l|l|l|l|l|}\n \\hline\nTask&Dataset&Model&Base Model&Metric&Eval. (Opt.; Orig)\\\\\n\\hline\n\\hline\nChar-level language modeling&1B Word Benchmark&Sinkhorn Mixture Big~&Transformer Big&BPC (L)&1.134; 1.825\\\\\n\\hline\nChar-level language modeling&1B Word Benchmark&Sparse Transformer~&Transformer Big&BPC (L)&1.119; 1.825\\\\\n\\hline\nChar-level language modeling&Enwik8&Longformer~&Transformer&BPC (L)&1.00; 1.11\\\\\n\\hline\nChar-level language modeling&Enwik8&Reformer~&Transformer&BPC (L)&1.05; 1.11\\\\\n\\hline\nChar-level language modeling&text8&Longformer~&Transformer&BPC (L)&1.10; 1.18\\\\\n\\hline\nCoreference resolution&OntoNotes&Longformer-base~&RoBERTa-base&F1 (H)&78.6; 78.4\\\\\n\\hline\nLanguage modeling&1B Word Benchmark&Sinkhorn Mixture Big~&Transformer Big&Perplexity (L)&27.34; 27.59\\\\\n\\hline\nLanguage modeling&1B Word Benchmark&Sparse Transformer~&Transformer Big&Perplexity (L)&28.77; 27.59\\\\\n\\hline\nNamed Entity Recognition&CoNLL2003&Star Transformer~&Transformer&Acc (H)&90.93; 86.48\\\\\n\\hline\nNamed Entity Recognition&CoNLL2012&Star Transformer~&Transformer&Acc (H)&86.30; 83.57\\\\\n\\hline\nNLI&MNLI&SortCut Sinkhorn~&Transformer&Acc (H)&55.80; 53.69\\\\\n\\hline\nNLI&QNLI&Linformer~&BERT-base&Acc (H)&91.2; 91.8\\\\\n\\hline\nNLI&SNLI&Star Transformer~&Transformer&Acc (H)&86.0; 82.2\\\\\n\\hline\nNLI&SNLI&SortCut Sinkhorn~&Transformer&Acc (H)&80.30; 78.87\\\\\n\\hline\nNMT (en$\\rightarrow$de)&WMT14&Reformer big~&Transformer Big&BLEU (H)&29.1; 27.3\\\\\n\\hline\nPOS Tagging&PTB&Star Transformer~&Transformer&Acc (H)&97.14; 96.31\\\\\n\\hline\nQuestion answering&HotpotQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&78.94; 77.08\\\\\n\\hline\nQuestion answering&HotpotQA&SparseBERT~&BERT&F1 (H)&76.02; 77.08\\\\\n\\hline\nQuestion answering&NaturalQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&79.39; 78.29\\\\\n\\hline\nQuestion answering&NaturalQA&SparseBERT~&BERT&F1 (H)&77.31; 78.29\\\\\n\\hline\nQuestion answering&NewsQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&70.08; 66.25\\\\\n\\hline\nQuestion answering&NewsQA&SparseBERT~&BERT&F1 (H)&67.16; 66.25\\\\\n\\hline\nQuestion answering&SearchQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&83.51; 80.37\\\\\n\\hline\nQuestion answering&SearchQA&SparseBERT~&BERT&F1 (H)&80.54; 80.37\\\\\n\\hline\nQuestion answering&SQuAD 1.1&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&90.74; 88.45\\\\\n\\hline\nQuestion answering&SQuAD 1.1&SparseBERT~&BERT&F1 (H)&88.37; 88.45\\\\\n\\hline\nQuestion answering&SQuAD 2.0&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&81.45; 77.16\\\\\n\\hline\nQuestion answering&SQuAD 2.0&SparseBERT~&BERT&F1 (H)&77.57; 77.16\\\\\n\\hline\nQuestion answering&TriviaQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&79.41; 75.35\\\\\n\\hline\nQuestion answering&TriviaQA&SparseBERT~&BERT&F1 (H)&75.34; 75.35\\\\\n\\hline\nQuestion answering&TriviaQA&Longformer-base~&RoBERTa-base&F1 (H)&75.2; 74.3\\\\\n\\hline\nQuestion answering&WikiHop&Longformer-base~&RoBERTa-base&Acc (H)&75.0; 72.4\\\\\n\\hline\nSentiment analysis&IMDB&Linformer~&BERT-base&Acc (H)&94.1; 93.5\\\\\n\\hline\nSentiment analysis&IMDB&Longformer-base~&RoBERTa-base&Acc (H)&95.7; 95.3\\\\\n\\hline\nSentiment analysis&SST&Star Transformer~&Transformer&Acc (H)&52.9; 50.4\\\\\n\\hline\nSentiment analysis&SST&Sinkhorn ~&Transformer&Acc (H)&77.52; 76.83\\\\\n\\hline\nSentiment analysis&SST-2&Linformer~&BERT-base&Acc (H)&93.1; 92.7\\\\\n\\hline\nSpeech recognition&WSJ&Linear Transformer~&Reformer&Phoneme Error Rate (L)&8.08; 9.33\\\\\n\\hline\nText classification&Hyperpartisan&Longformer-base~&RoBERTa-base&F1 (H)&94.8; 87.4\\\\\n\\hline\nText classification&MTL-16&Star Transformer~&Transformer&Acc (H)&86.98; 82.78\\\\\n\\hline\nTextual similarity&QQP&Linformer~&BERT-base&Acc (H)&90.8; 89.6\\\\\n\\hline\n \\end{tabular}\n\\caption{Comparison of various sub-quadratic complexity Transformer methods (sorted by Task and then Dataset). In the metric column, H means high is better while L means low is better. Note that in this case, model sizes do not reduce much; activation memory reduces as described in Section~\\ref{sec:linearTransformers} (with comparable or better accuracy) compared to the standard Transformer.}\n \\label{tab:linearTransSummary}\n\\end{table}\nTable~\\ref{tab:linearTransSummary} compares various sub-quadratic Transformer methods across different tasks and datasets. Accuracy of both the original Transformer and the optimized (opt.) model are shown. These models have been applied for various applications like language modeling (both word level as well as character level), coreference resolution, NER, NLI, NMT, POS, question answering, sentiment analysis, speech recognition, text classification and text similarity analysis. For the same task, dataset and model combination, different papers report different accuracy of the original model because of slight changes in training hyper-parameters; hence we report accuracy of the original model for each row. \nStar Transformers were shown to outperform the vanilla Transformer model across various tasks like Sentiment analysis, Text classification, NLI, POS and NER. On the SST, the Star Transformer achieves 2.5 points improvement against the standard Transformer. On MTL-16, the Star-Transformer outperform the standard Transformer in all 16 datasets, the improvement of the average accuracy is 4.2. Average test times are 10.94 and 49.31ms per batch with batch-size=128 for Star Transformer and standard Transformer respectively.\nLongformer achieves a new state-of-the-art on both text8 and enwik8 using the small models with BPC of 1.10 and 1.00 on text8 and enwik8 respectively. Longformer-large model of the same size as Sparse Transformer achieves a BPC of 0.99 on Enwik8, which is the same as that obtained using Sparse Transformer. Also, for both character-level as well as word-level language modeling, on 1B word benchmark, we observe that Sinkhorn mixture (which is a combination of the Sinkhorn attention by mixing it with the vanilla standard dot product attention) performs better than Sparse Transformer. \nOn question answering, results have been reported across multiple datasets like HotpotQA, NaturalQA, NewsQA, SearchQA, TriviaQA, WikiHop, SQuAD 1.1 and SQuAD 2.0. We observe that BlockBERT performs better than SparseBERT. Also, it is not surprising that BlockBERT with 2 blocks (n = 2) performs better than that with 3 blocks (n = 3), because it keeps more attention matrix entries. Also, not shown in the table, Longformer-large achieves scores of 81.9 and 73.2 for WikiHop and HotpotQA beating state-of-the-art results by 3.6 and 4 points respectively. \nLongformer consistently outperforms the RoBERTa baseline across many tasks like coreference resolution, question answering, sentiment analysis and text classification. Its performance gain is especially obvious for tasks that require long context such as WikiHop (question answering) and Hyperpartisan (text classification). For TriviaQA (question answering), the improvement is more modest as the local context is often sufficient to answer the question. In the case of HotpotQA (question answering), the supporting fact auxiliary supervision allows models to easily find relevant contexts and then focus on local context, leading to smaller gains. On the IMDB (sentiment analysis) and OntoNotes (coreference resolution) datasets the performance gains are smaller. For IMDB, the majority of the dataset consists of short documents and thus it is expected to see smaller improvements. For OntoNotes, the distance between any two mentions is typically quite small so that a baseline that processes smaller chunks separately is able to stitch together mentions into coreference chains without considering cross chunk interactions.\nOn speech recognition task, Linear transformers achieve similar performance to vanilla transformers and they are up to 4000x faster on autoregressive prediction of very long sequences. The linear Transformer model outperforms the LSTM and Reformer while being faster to train and evaluate. Reformer takes 2250 seconds per epoch, while Linear Transformers take just 824s/epoch.\nTo summarize, multiple methods have been proposed to reduce the quadratic complexity of the standard Transformer model. While Sparse Transformers reduce it to $O(n\\sqrt{n})$, Reformers reduce it to $O(n\\log n)$. Other methods like Star Transformer, Linformer, Sparse Sinkhorn Transformer, Efficient Attention, Linear Transformers and Longformer promise linear complexity. In particular Sparse Sinkhorn Transformer and Longformer have been shown to result into very good accuracy, latency and RAM tradeoff across many tasks.","id":"2202feab-4da3-4fb5-8d59-2d1d4fbac3f9","level":"subsection","origin_cites_number":8,"parent_id":"65404f84-cbf0-4721-8f05-178f8230aa55","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Transformers with Sub-Quadratic Complexity"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:summary}\nWe discussed various methods for compression of deep learning models for text. Broadly, we discussed pruning, quantization, knowledge distillation, parameter sharing, tensor decomposition, and sub-quadratic Transformer based methods. These methods not just help reduce the model size, but also lead to lower prediction latencies and low power consumption due to reduced computations. Pruning is the oldest method, but not commonly applied for Transformers. Quantization is effective however it is important to use mixed precision balanced quantization with GPU architectures that support efficient low-bit computations. Knowledge Distillation is the most popular method for compression of Transformer models. Parameter sharing is a very useful method but often needs to be combined with other techniques. Matrix decomposition is not very common but has lots of potential, especially the BTD method. Sub-quadratic Transformers are very important to enable processing long documents in applications like query-document similarity, long document summarization, etc.","id":"d06ca9cc-bcd7-4f53-9be8-ee86f8754664","level":"section","origin_cites_number":0,"parent_id":"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Summary and Future Directions"]],"subsections":["ab02f8bb-9d54-4f84-82fc-3f831c1bc39a","fe1a6e49-c5b1-4394-b0be-1a452c8b0f80","7ea3f834-f595-4c1b-afed-044d53e835ea"],"title":"Summary and Future Directions"},{"cite_extract_rate":0.76,"cites":[4479,4478,7298,4515,1150,4517,4480,4521,4514,4485,4492,1473,4491,8375,9120,4522,4536,4497,4533,8799,4525,4519,4526,854,4484,7829,7333,8797,4486,4516,4527,4500,793,4530,1568,4036,4523,7371,4532,4524,8390,4499,4331,4509,4534,4511,2481,844,4535,4482,7828,4518,794,4351,4507,2488,4510],"content":"In the previous six sections, we have compared across multiple methods within each of the six broad types of model compression. In this subsection, we attempt to compare across multiple model compression method types.\nTable~\\ref{tab:kdSummary2} provides a comparison of model size versus accuracy for various methods across various GLUE~ and SQuAD tasks. We observe that most of the methods tried on GLUE datasets have been based on knowledge distillation. However, some pruning methods (iterative magnitude pruning, RPP and LayerDrop), quantization (mixed precision QBERT), parameter sharing (ALBERT), tensor decomposition (FLOP) and linear Transformer (Linformer) have also been tried. Quantization methods do not reduce the number of parameters but reduce the number of bits per parameter. Particularly, mixed precision QBERT uses 8 bits for embeddings and 2/3/4 bits for encoder layer weights. Similarly, Linformer does not reduce number of weights but reduces activation memory as well as latency of the overall Transformer model. \nFor the remaining models, for each model, we first computed an average GLUE score based on any of the 9 tasks for which scores have been reported. Next, we computed the ratio GLUE score/model size (in Millions). We find the following as the top three \n\\begin{itemize}\n \\item Distilled-BiLSTM~ (ratio=79.3)\n \\item Mixed-vocab. training~ (ratio=7.8)\n \\item ALBERT-B~ (ratio=7.2)\n\\end{itemize}\nThis clearly tells us that Distilled BiLSTMs provide us the best accuracy versus size tradeoff on GLUE. However, each of these three models actually report results on only 4 out of 9 GLUE tasks. Hence, further, we considered only those methods for which results on at least 5 tasks have been reported and computed the GLUE score/model size ratio. We find the following as the top three \n\\begin{itemize}\n\\item TinyBERT~ (ratio=5.31)\n\\item BERT-EMD~ (ratio=5.15)\n\\item MobileBERT+Quantization~ (ratio=3.49)\n\\end{itemize}\nThus, on GLUE tasks, it is clear that distillation based methods (combined with quantization) are better than other types of methods. \nTables~\\ref{tab:pruningSummary},~\\ref{tab:quantizationSummary},~\\ref{tab:kdSummary1},~\\ref{tab:paramSharingSummary},~\\ref{tab:tensorDecompSummary} and~\\ref{tab:linearTransSummary} compare various pruning, quantization, knowledge distillation, parameter sharing, tensor decomposition and sub-quadratic Transformer methods across different tasks and datasets. Overall, there are 123 (task, dataset) combinations across these six tables which implies that unlike GLUE, not many methods have been applied on the same set of (task, dataset) combinations\\footnote{We make the entire statistics available as an excel file at \\url{https://bit.ly/3vmaxZ9}.}. The most popular tasks are language modeling (on PTB and 1B word benchmark), sentiment analysis (on SST) and NMT (WMT14 en$\\rightarrow$de). \nFor language modeling on PTB, on LSTM models, bank balanced sparsity~ based pruning method worked best. With Transformer models, Block Term Decomposition (BTD)~ method seems to work best. Among various methods like parameter sharing, tensor decomposition and sub-quadratic complexity Transformer which have been tried for language modeling on 1B Word Benchmark, again, BTD~ method seems to work best leading to a model with 0.16B parameters and a perplexity as low as 19.5. Multiple datasets have been used for Neural machine translation (NMT). Datasets from WMT and IWSLT are the most popular. Among these, the en$\\rightarrow$de from WMT14 is the most popular dataset used for testing various NMT models. For en$\\rightarrow$de NMT with WMT14, using 2-layer LSTMs, the best accuracy versus size tradeoff is using Pruned Seq-KD + Seq-Inter~ which gives a 8M size model leading to 18.5 BLEU. Among Transformer based models, BTD~ leads to a Transformer model which provides 34.91 BLEU with 21.2M model size.\n\\begin{table*}\n \\centering\n \\scriptsize\n \\begin{tabular}{|p{1in}|p{2in}|p{2.7in}|}\n \\hline\nTask&Popular Datasets&References\\\\\n\\hline\n\\hline\nLanguage modeling&Penn TreeBank Corpus, One billion word benchmark, Europarl, WikiText-103, text8, source code of Linux kernel, 2013 ACL Workshop Morphological Language Datasets (ACLW), Arabic news commentary corpus, 2013 ACL workshop on MT, enwik8 (from Wikipedia), Lambada&Neuron Pruning~, Iterative magnitude pruning~, Block sparsity~,Loss -Aware Quantization~, Uniform Quantization~, Binary Quantization~, HitNet~, Sparse Word Representations~, LightRNN~, Slim Embeddings~, C2W~, LayerDrop~, Reformer~, Linformer~, Char-CNN~, CNN+Highway Network~, SparseBERT~, FLOP~, Deep Equilibrium Models~, WEST~, Sparse Sinkhorn Attention~, BTD~, Universal Transformers~, TT-embedding~, multiple methods~\\\\\n\\hline\nNeural Machine translation (NMT)&IWSLT German-English, IWSLT Thai-English, ASPEC English-Japanese, WMT English-German, WMT German-English, WMT English-Russian, IWSLT English Vietnamese, WMT English-Romanian, WMT English-Estonian, Ted Talk&Compositional codes~, LayerDrop~, Pruning attention heads~, Neuron Pruning~, Magnitude Pruning~, Iterative magnitude pruning~, Pruned Seq-KD + Seq-Inter~, Quantized Distillation~, Teacher ensembles KD~, Multiple teachers KD~, BTD~, Quaternion Attention~, Universal Transformers~, TT-embedding~\\\\\n\\hline\nSentiment Analysis&IMDB movie review, SST, SST-2, Elec (electronic product reviews)&Compositional codes~, Star Transformer~, TinyBERT~, MiniLM~, Linformer~, XtremeDistil~, Gaussian Quantization~, Uniform Quantization~, Sparse coding~, Quaternion Attention~, Sparse Sinkhorn Attention~, RPP~, ALBERT~, Patient KD~, Mixed-vocabulary KD training~, Distilled-BiLSTM~, MTDNN~, TT-embedding~\\\\\n\\hline\nQuestion Answering&SQuAD1.1, SQuAD2.0, ELI5, SemEval, BABI&LayerDrop~, MiniLM~, RPP~, BS-Fixed Quantization~, ALBERT~, KD~, Universal Transformers~\\\\\n\\hline\nNatural Language Inference&SNLI, MNLI-m, MNLI-mm, QNLI, RTE, WNLI, XNLI&Star Transformer~, LayerDrop~, TinyBERT~, MiniLM~, Linformer~, Sparse Sinkhorn Attention~, RPP~, ALBERT~, Patient KD~, Mixed-vocabulary KD training~, Distilled-BiLSTM~, MTDNN~\\\\\n\\hline\nParaphrasing&QQP, STS-B&TinyBERT~, MiniLM~, Linformer~, RPP~, ALBERT~, Patient KD~, Distilled-BiLSTM~, MTDNN~\\\\\n\\hline\nImage captioning&MSCOCO&Grow and Prune~, Magnitude Pruning~, Iterative Magnitude Pruning and Densification~\\\\\n\\hline\nHandwritten character recognition&ICDAR&SVD and Pruning\\\\\n\\hline\nPart-of-speech (POS) tagging&Wall Street Journal of the Penn Treebank dataset, WikiAnn NER corpus&C2W~, XtremeDistil~\\\\\n\\hline\nSummarization&CNN-DailyMail, XSum&LayerDrop~, MiniLM~\\\\\n\\hline\nMachine Reading Comprehension& Microsoft Research Paraphrase Corpus (MRPC), ReAding Comprehension from Examinations (RACE)&LayerDrop~, TinyBERT~, MiniLM~, RPP~, ALBERT~, Patient KD~, Mixed-vocabulary KD training~, MTDNN~\\\\\n\\hline\nLinguistic Acceptability&CoLA&TinyBERT~, MiniLM~, RPP~, ALBERT~, MTDNN~\\\\\n\\hline\nTopic Classification&DbPedia, Ag News,20 Newsgroup&XtremeDistil~, Sparse coding~\\\\\n\\hline\nQuestion Type Classification&TREC&Sparse coding~\\\\\n\\hline\nNoun Phrase Bracketing&Lazaridou~&Sparse coding~\\\\\n\\hline\nWord Similarity&SimLex-999, MEN, MTurk, RARE, SCWS, WSR, WSS&Sparse coding~, Shared reference vectors~\\\\\n\\hline\nMathematical Language Understanding&Wangperawong's MLU~&Quaternion Attention~\\\\\n\\hline\nSubject Verb Agreement&Linzen~&Quaternion Attention~, Universal Transformers~\\\\\n\\hline\nWord Analogy&GSEM, GSYN, MSYN&Shared reference vectors~\\\\\n\\hline\nSentence Completion&MSC&Shared reference vectors~\\\\\n\\hline\nLearning to execute&Zaremba and Sutskever~&Universal Transformers~\\\\\n\\hline\nAd Click Through Rate Prediction&Criteo Kaggle&TT-embedding~\\\\\n\\hline\nSpeech Recognition&2100 hours English Speech, AN4, Switchboard, TIMIT, WSJ 92, WSJ 93, TIDIGITS, 3M Google voice utterances, Live traffic utterances&Iterative Magnitude Pruning~, Grow and Prune~, Neuron Pruning~, Block Sparsity~, BBS~, DSD~, Pow2 Ternarization~, Loss Aware Quantization~, Toeplitz-like~, Joint-SVD~, Projections~, WEST~\\\\\n\\hline\nNamed entity recognition (NER)&CoNLL2003, Wikiann-41, CoNLL2012&QBERT~, XtremeDistill~, Star Transformer~\\\\\n\\hline\nIntent Detection&SNIPS&Mixed-vocabulary KD training~\\\\\n\\hline\nQuestion Generation&SQuAD 1.1&MiniLM~\\\\\n\\hline\nSlot Filling&SNIPS&Mixed-vocabulary KD training~\\\\\n\\hline\nText classification&20 Newsgroup, Hyperpartisan, MTL-16&Sparse coding~, Longformer~, Star Transformer~\\\\\n\\hline\nCoreference Resolution&OntoNotes&Longformer~\\\\\n\\hline\n \\end{tabular}\n \\caption{Applications of Model Compression Methods for Text}\n \\label{tab:applications}\n\\end{table*}\nCombinations of multiple model compression method types has also been experimented with and found to be effective. Some examples of such combinations include the following:\n\\begin{itemize}\n \\item Pruning + Tensor Decomposition ~\n \\item Pruning + Quantization~\n \\item Knowledge distillation + Quantization~\n \\item Knowledge distillation + Pruning~\n \\item Tensor decomposition + Parameter sharing~\n \\item Tensor decomposition + Quantization~\n\\end{itemize}\nRecently, Kim et al.~ combined knowledge distillation, structured pruning and quantization leading to drastic improvements on inference efficiency. First, they investigate the efficacy of various Knowledge Distillation techniques to significantly reduce the size of the models with respect to the depth and hidden state sizes while preserving the\naccuracy. Second, they explore Structured Pruning that further reduces the size of the models by reducing the number of self-attention heads and the number of intermediate hidden states in the feedforward layers to achieve more efficiency while\ntrying to preserve the accuracy as well. Finally, they \nexplore model quantization which enables faster model executions by optimally utilizing hardware acceleration capabilities. Such a combined method leads to heavily reduced model size, 12.4x GPU speed-up and 6.9x-125.8x reduction in energy consumption.","id":"ab02f8bb-9d54-4f84-82fc-3f831c1bc39a","level":"subsection","origin_cites_number":75,"parent_id":"d06ca9cc-bcd7-4f53-9be8-ee86f8754664","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Summary and Future Directions"],["subsection","Comparison across model compression method types."]],"subsections":[],"title":"Comparison across model compression method types."},{"cite_extract_rate":0,"cites":[],"content":"The model compression methods mentioned in this survey have been used across a wide variety of text processing tasks. In Table~\\ref{tab:applications}, we list down the tasks, popular datasets and references where the readers can find more discussion around model size versus accuracy tradeoff.","id":"fe1a6e49-c5b1-4394-b0be-1a452c8b0f80","level":"subsection","origin_cites_number":0,"parent_id":"d06ca9cc-bcd7-4f53-9be8-ee86f8754664","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Summary and Future Directions"],["subsection","Summary of Applications"]],"subsections":[],"title":"Summary of Applications"},{"cite_extract_rate":0.6666666666666661,"cites":[849,4536],"content":"Although there has been so much of work already in this field, there is a lot more work to be done. \n\\begin{itemize}\n \\item With linear Transformer models, one can afford to have input with tens of thousands of tokens. Hence, many tasks need to be redesigned where large context can now be included as input to improve accuracy. \n \\item Combinations of several methods have not been tested well. Recently, Fastformers method~ showed that combining multiple methods like knowledge distillation, 16-bit quantization, structured pruning and numerical optimizations can lead to drastic improvements. However, lot of experiments are needed to further check how models respond to combination of model compression methods.\n \\item Latency results vary based on GPU architectures. With new GPU architectures (Nvidia RTX 3080, Nvidia T4), some methods like quantization may become more impactful.\n \\item Real world settings are often complex: multi-modal, multi-task, multi-label, small-data, noisy labels, multi-teachers, mismatching teacher-student architectures. Efficient ways of recommending the most promising method is necessary.\n \\item Different components/structures of a model may respond to different kinds of compression methods with specific hyper-parameters. A generic method to choose the right method for various structures is needed.\n \\item How does compression of models impact their interpretability? Can we design model compression mechanisms aimed at looking at a tradeoff between model accuracy, size, latency and interpretability.\n \\item None of the model compression methods performs any application specific compression. Can we obtain further compression by exploiting some task-specific patterns?\n \\item Recently, deep reinforcement learning based methods have been proposed in the computer vision community~. It will be nice to check the effectiveness of such methods for NLP tasks.\n\\end{itemize}\nWe hope that this survey acts as a good guide to folks across academia and industry. Also, we hope that a significantly large chunk of research gets done in the area of model compression to enable good accuracy across many NLP tasks while keeping model sizes and latencies in check. \n\\begin{thebibliography}{100}\n\\scriptsize\n\\providecommand{\\url}[1]{#1}\n\\csname url@samestyle\\endcsname\n\\providecommand{\\newblock}{\\relax}\n\\providecommand{\\bibinfo}[2]{#2}\n\\providecommand{\\BIBentrySTDinterwordspacing}{\\spaceskip=0pt\\relax}\n\\providecommand{\\BIBentryALTinterwordstretchfactor}{4}\n\\providecommand{\\BIBentryALTinterwordspacing}{\\spaceskip=\\fontdimen2\\font plus\n\\BIBentryALTinterwordstretchfactor\\fontdimen3\\font minus\n \\fontdimen4\\font\\relax}\n\\providecommand{\\BIBforeignlanguage}[2]{{\n\\expandafter\\ifx\\csname l@#1\\endcsname\\relax\n\\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}\n\\typeout{** loaded for the language `#1'. Using the pattern for}\n\\typeout{** the default language instead.}\n\\else\n\\language=\\csname l@#1\\endcsname\n\\fi\n#2}}\n\\providecommand{\\BIBdecl}{\\relax}\n\\BIBdecl\n\\bibitem{vaswani2017attention}\nA.~Vaswani, N.~Shazeer, N.~Parmar, J.~Uszkoreit, L.~Jones, A.~N. Gomez,\n {\\L}.~Kaiser, and I.~Polosukhin, ``Attention is all you need,'' in\n \\emph{NIPS}, 2017, pp. 5998--6008.\n\\bibitem{devlin2018bert}\nJ.~Devlin, M.-W. Chang, K.~Lee, and K.~Toutanova, ``Bert: Pre-training of deep\n bidirectional transformers for language understanding,''\n \\emph{arXiv:1810.04805}, 2018.\n\\bibitem{wang2019glue}\nA.~Wang, A.~Singh, J.~Michael, F.~Hill, O.~Levy, and S.~R. Bowman, ``{GLUE}: A\n multi-task benchmark and analysis platform for natural language\n understanding,'' in \\emph{ICLR}, 2019.\n\\bibitem{wang2019superglue}\nA.~Wang, Y.~Pruksachatkun, N.~Nangia, A.~Singh, J.~Michael, F.~Hill, O.~Levy,\n and S.~R. Bowman, ``Super{GLUE}: A stickier benchmark for general-purpose\n language understanding systems,'' \\emph{1905.00537}, 2019.\n\\bibitem{radford2019language}\nA.~Radford, J.~Wu, R.~Child, D.~Luan, D.~Amodei, and I.~Sutskever, ``Language\n models are unsupervised multitask learners,'' \\emph{OpenAI Blog}, vol.~1,\n no.~8, 2019.\n\\bibitem{liu2019multi}\nX.~Liu, P.~He, W.~Chen, and J.~Gao, ``Multi-task deep neural networks for\n natural language understanding,'' \\emph{arXiv:1901.11504}, 2019.\n\\bibitem{yang2019xlnet}\nZ.~Yang, Z.~Dai, Y.~Yang, J.~Carbonell, R.~Salakhutdinov, and Q.~V. Le,\n ``Xlnet: Generalized autoregressive pretraining for language understanding,''\n \\emph{arXiv:1906.08237}, 2019.\n\\bibitem{shoeybi2019megatron}\nM.~Shoeybi, M.~Patwary, R.~Puri, P.~LeGresley, J.~Casper, and B.~Catanzaro,\n ``Megatron-lm: Training multi-billion parameter language models using gpu\n model parallelism,'' \\emph{arXiv:1909.08053}, 2019.\n\\bibitem{raffel2019exploring}\nC.~Raffel, N.~Shazeer, A.~Roberts, K.~Lee, S.~Narang, M.~Matena, Y.~Zhou,\n W.~Li, and P.~J. Liu, ``Exploring the limits of transfer learning with a\n unified text-to-text transformer,'' \\emph{arXiv:1910.10683}, 2019.\n\\bibitem{rosset2019turing}\nC.~Rosset, ``Turing-nlg: A 17-billion-parameter language model by microsoft,''\n \\emph{Microsoft Blog}, 2019.\n\\bibitem{lepikhin2020gshard}\nD.~Lepikhin, H.~Lee, Y.~Xu, D.~Chen, O.~Firat, Y.~Huang, M.~Krikun, N.~Shazeer,\n and Z.~Chen, ``Gshard: Scaling giant models with conditional computation and\n automatic sharding,'' \\emph{arXiv:2006.16668}, 2020.\n\\bibitem{bianco2018benchmark}\nS.~Bianco, R.~Cadene, L.~Celona, and P.~Napoletano, ``Benchmark analysis of\n representative deep neural network architectures,'' \\emph{IEEE Access},\n vol.~6, pp. 64\\,270--64\\,277, 2018.\n\\bibitem{sanh2019distilbert}\nV.~Sanh, L.~Debut, J.~Chaumond, and T.~Wolf, ``Distilbert, a distilled version\n of bert: smaller, faster, cheaper and lighter,'' \\emph{arXiv:1910.01108},\n 2019.\n\\bibitem{han2016eie}\nS.~Han, X.~Liu, H.~Mao, J.~Pu, A.~Pedram, M.~A. Horowitz, and W.~J. Dally,\n ``Eie: efficient inference engine on compressed deep neural network,''\n \\emph{ACM SIGARCH Computer Architecture News}, vol.~44, no.~3, pp. 243--254,\n 2016.\n\\bibitem{diamos2016persistent}\nG.~Diamos, S.~Sengupta, B.~Catanzaro, M.~Chrzanowski, A.~Coates, E.~Elsen,\n J.~Engel, A.~Hannun, and S.~Satheesh, ``Persistent rnns: Stashing recurrent\n weights on-chip,'' in \\emph{ICML}, 2016, pp. 2024--2033.\n\\bibitem{denil2013predicting}\nM.~Denil, B.~Shakibi, L.~Dinh, M.~Ranzato, and N.~De~Freitas, ``Predicting\n parameters in deep learning,'' in \\emph{NIPS}, 2013, pp. 2148--2156.\n\\bibitem{cheng2017survey}\nY.~Cheng, D.~Wang, P.~Zhou, and T.~Zhang, ``A survey of model compression and\n acceleration for deep neural networks,'' \\emph{arXiv:1710.09282}, 2017.\n\\bibitem{deng2020model}\nL.~Deng, G.~Li, S.~Han, L.~Shi, and Y.~Xie, ``Model compression and hardware\n acceleration for neural networks: A comprehensive survey,'' \\emph{IEEE}, vol.\n 108, no.~4, pp. 485--532, 2020.\n\\bibitem{walsh2013peter}\nC.~A. Walsh, ``Peter huttenlocher (1931--2013),'' \\emph{Nature}, vol. 502, no.\n 7470, pp. 172--172, 2013.\n\\bibitem{zhu2017prune}\nM.~Zhu and S.~Gupta, ``To prune, or not to prune: exploring the efficacy of\n pruning for model compression,'' \\emph{arXiv:1710.01878}, 2017.\n\\bibitem{lecun1990optimal}\nY.~LeCun, J.~S. Denker, and S.~A. Solla, ``Optimal brain damage,'' in\n \\emph{NIPS}, 1990, pp. 598--605.\n\\bibitem{hassibi1993second}\nB.~Hassibi and D.~G. Stork, ``Second order derivatives for network pruning:\n Optimal brain surgeon,'' in \\emph{NIPS}, 1993, pp. 164--171.\n\\bibitem{han2015deep}\nS.~Han, H.~Mao, and W.~J. Dally, ``Deep compression: Compressing deep neural\n networks with pruning, trained quantization and huffman coding,''\n \\emph{arXiv:1510.00149}, 2015.\n\\bibitem{see2016compression}\nA.~See, M.-T. Luong, and C.~D. Manning, ``Compression of neural machine\n translation models via pruning,'' \\emph{arXiv:1606.09274}, 2016.\n\\bibitem{han2015learning}\nS.~Han, J.~Pool, J.~Tran, and W.~Dally, ``Learning both weights and connections\n for efficient neural network,'' in \\emph{NIPS}, 2015, pp. 1135--1143.\n\\bibitem{narang2017exploring}\nS.~Narang, E.~Elsen, G.~Diamos, and S.~Sengupta, ``Exploring sparsity in\n recurrent neural networks,'' \\emph{arXiv:1704.05119}, 2017.\n\\bibitem{cheong2019transformers}\nR.~Cheong and R.~Daniel, ``transformers. zip: Compressing transformers with\n pruning and quantization,'' Technical report, Stanford University, Stanford,\n California, 2019., Tech. Rep., 2019.\n\\bibitem{guo2019reweighted}\nF.-M. Guo, S.~Liu, F.~S. Mungall, X.~Lin, and Y.~Wang, ``Reweighted proximal\n pruning for large-scale language representation,'' \\emph{arXiv:1909.12486},\n 2019.\n\\bibitem{han2016dsd}\nS.~Han, J.~Pool, S.~Narang, H.~Mao, E.~Gong, S.~Tang, E.~Elsen, P.~Vajda,\n M.~Paluri, J.~Tran \\emph{et~al.}, ``Dsd: Dense-sparse-dense training for deep\n neural networks,'' \\emph{arXiv:1607.04381}, 2016.\n\\bibitem{dai2018grow}\nX.~Dai, H.~Yin, and N.~K. Jha, ``Grow and prune compact, fast, and accurate\n lstms,'' \\emph{arXiv:1805.11797}, 2018.\n\\bibitem{he2014reshaping}\nT.~He, Y.~Fan, Y.~Qian, T.~Tan, and K.~Yu, ``Reshaping deep neural network for\n fast decoding by node-pruning,'' in \\emph{ICASSP}.\\hskip 1em plus 0.5em minus\n 0.4em\\relax IEEE, 2014, pp. 245--249.\n\\bibitem{murray2015auto}\nK.~Murray and D.~Chiang, ``Auto-sizing neural networks: With applications to\n n-gram language models,'' \\emph{arXiv:1508.05051}, 2015.\n\\bibitem{pan2016dropneuron}\nW.~Pan, H.~Dong, and Y.~Guo, ``Dropneuron: Simplifying the structure of deep\n neural networks,'' \\emph{arXiv:1606.07326}, 2016.\n\\bibitem{srinivas2015data}\nS.~Srinivas and R.~V. Babu, ``Data-free parameter pruning for deep neural\n networks,'' \\emph{arXiv:1507.06149}, 2015.\n\\bibitem{narang2017block}\nS.~Narang, E.~Undersander, and G.~Diamos, ``Block-sparse recurrent neural\n networks,'' \\emph{arXiv:1711.02782}, 2017.\n\\bibitem{cao2019efficient}\nS.~Cao, C.~Zhang, Z.~Yao, W.~Xiao, L.~Nie, D.~Zhan, Y.~Liu, M.~Wu, and\n L.~Zhang, ``Efficient and effective sparse lstm on fpga with bank-balanced\n sparsity,'' in \\emph{SIGDA Intl. Symp. on FPGA}.\\hskip 1em plus 0.5em minus\n 0.4em\\relax ACM, 2019, pp. 63--72.\n\\bibitem{michel2019sixteen}\nP.~Michel, O.~Levy, and G.~Neubig, ``Are sixteen heads really better than\n one?'' \\emph{arXiv:1905.10650}, 2019.\n\\bibitem{voita2019analyzing}\nE.~Voita, D.~Talbot, F.~Moiseev, R.~Sennrich, and I.~Titov, ``Analyzing\n multi-head self-attention: Specialized heads do the heavy lifting, the rest\n can be pruned,'' \\emph{arXiv:1905.09418}, 2019.\n\\bibitem{ding2017visualizing}\nY.~Ding, Y.~Liu, H.~Luan, and M.~Sun, ``Visualizing and understanding neural\n machine translation,'' in \\emph{ACL}, 2017, pp. 1150--1159.\n\\bibitem{fan2019reducing}\nA.~Fan, E.~Grave, and A.~Joulin, ``Reducing transformer depth on demand with\n structured dropout,'' \\emph{arXiv:1909.11556}, 2019.\n\\bibitem{prasanna2020bert}\nS.~Prasanna, A.~Rogers, and A.~Rumshisky, ``When bert plays the lottery, all\n tickets are winning,'' \\emph{arXiv:2005.00561}, 2020.\n\\bibitem{bartol2015hippocampal}\nT.~M. Bartol, C.~Bromer, J.~Kinney, M.~A. Chirillo, J.~N. Bourne, K.~M. Harris,\n and T.~J. Sejnowski, ``Hippocampal spine head sizes are highly precise,''\n \\emph{bioRxiv}, p. 016329, 2015.\n\\bibitem{linden2018think}\nD.~J. Linden, \\emph{Think Tank: Forty Neuroscientists Explore the Biological\n Roots of Human Experience}.\\hskip 1em plus 0.5em minus 0.4em\\relax Yale\n University Press, 2018.\n\\bibitem{hubara2017quantized}\nI.~Hubara, M.~Courbariaux, D.~Soudry, R.~El-Yaniv, and Y.~Bengio, ``Quantized\n neural networks: Training neural networks with low precision weights and\n activations,'' \\emph{JMLR}, vol.~18, no.~1, pp. 6869--6898, 2017.\n\\bibitem{lam2018word2bits}\nM.~Lam, ``Word2bits-quantized word vectors,'' \\emph{arXiv:1803.05651}, 2018.\n\\bibitem{courbariaux2015binaryconnect}\nM.~Courbariaux, Y.~Bengio, and J.-P. David, ``Binaryconnect: Training deep\n neural networks with binary weights during propagations,'' in \\emph{NIPS},\n 2015, pp. 3123--3131.\n\\bibitem{bengio2013estimating}\nY.~Bengio, N.~L{\\'e}onard, and A.~Courville, ``Estimating or propagating\n gradients through stochastic neurons for conditional computation,''\n \\emph{arXiv:1308.3432}, 2013.\n\\bibitem{lin2015neural}\nZ.~Lin, M.~Courbariaux, R.~Memisevic, and Y.~Bengio, ``Neural networks with few\n multiplications,'' \\emph{arXiv:1510.03009}, 2015.\n\\bibitem{hubara2016binarized}\nI.~Hubara, M.~Courbariaux, D.~Soudry, R.~El-Yaniv, and Y.~Bengio, ``Binarized\n neural networks,'' in \\emph{NIPS}, 2016, pp. 4107--4115.\n\\bibitem{rastegari2016xnor}\nM.~Rastegari, V.~Ordonez, J.~Redmon, and A.~Farhadi, ``Xnor-net: Imagenet\n classification using binary convolutional neural networks,'' in\n \\emph{ECCV}.\\hskip 1em plus 0.5em minus 0.4em\\relax Springer, 2016, pp.\n 525--542.\n\\bibitem{hou2016loss}\nL.~Hou, Q.~Yao, and J.~T. Kwok, ``Loss-aware binarization of deep networks,''\n \\emph{arXiv:1611.01600}, 2016.\n\\bibitem{lee2014proximal}\nJ.~D. Lee, Y.~Sun, and M.~A. Saunders, ``Proximal newton-type methods for\n minimizing composite functions,'' \\emph{J. Optimization}, vol.~24, no.~3, pp.\n 1420--1443, 2014.\n\\bibitem{ott2016recurrent}\nJ.~Ott, Z.~Lin, Y.~Zhang, S.-C. Liu, and Y.~Bengio, ``Recurrent neural networks\n with limited numerical precision,'' \\emph{arXiv:1608.06902}, 2016.\n\\bibitem{alom2018effective}\nM.~Z. Alom, A.~T. Moody, N.~Maruyama, B.~C. Van~Essen, and T.~M. Taha,\n ``Effective quantization approaches for recurrent neural networks,'' in\n \\emph{IJCNN}.\\hskip 1em plus 0.5em minus 0.4em\\relax IEEE, 2018, pp. 1--8.\n\\bibitem{li2016ternary}\nF.~Li, B.~Zhang, and B.~Liu, ``Ternary weight networks,''\n \\emph{arXiv:1605.04711}, 2016.\n\\bibitem{hwang2014fixed}\nK.~Hwang and W.~Sung, ``Fixed-point feedforward deep neural network design\n using weights+ 1, 0, and- 1,'' in \\emph{SiPS}.\\hskip 1em plus 0.5em minus\n 0.4em\\relax IEEE, 2014, pp. 1--6.\n\\bibitem{lloyd1982least}\nS.~Lloyd, ``Least squares quantization in pcm,'' \\emph{Tran. on information\n theory}, vol.~28, no.~2, pp. 129--137, 1982.\n\\bibitem{zhu2016trained}\nC.~Zhu, S.~Han, H.~Mao, and W.~J. Dally, ``Trained ternary quantization,''\n \\emph{arXiv:1612.01064}, 2016.\n\\bibitem{wang2018hitnet}\nP.~Wang, X.~Xie, L.~Deng, G.~Li, D.~Wang, and Y.~Xie, ``Hitnet: hybrid ternary\n recurrent neural network,'' in \\emph{NIPS}, 2018, pp. 604--614.\n\\bibitem{he2016effective}\nQ.~He, H.~Wen, S.~Zhou, Y.~Wu, C.~Yao, X.~Zhou, and Y.~Zou, ``Effective\n quantization methods for recurrent neural networks,''\n \\emph{arXiv:1611.10176}, 2016.\n\\bibitem{zhou2017balanced}\nS.-C. Zhou, Y.-Z. Wang, H.~Wen, Q.-Y. He, and Y.-H. Zou, ``Balanced\n quantization: An effective and efficient approach to quantized neural\n networks,'' \\emph{J. of Computer Science and Technology}, vol.~32, no.~4, pp.\n 667--682, 2017.\n\\bibitem{muller2015rounding}\nL.~K. Muller and G.~Indiveri, ``Rounding methods for neural networks with low\n resolution synaptic weights,'' \\emph{arXiv:1504.05767}, 2015.\n\\bibitem{gong2014compressing}\nY.~Gong, L.~Liu, M.~Yang, and L.~Bourdev, ``Compressing deep convolutional\n networks using vector quantization,'' \\emph{arXiv:1412.6115}, 2014.\n\\bibitem{guo2017network}\nY.~Guo, A.~Yao, H.~Zhao, and Y.~Chen, ``Network sketching: Exploiting binary\n structure in deep cnns,'' in \\emph{CVPR}, 2017, pp. 5955--5963.\n\\bibitem{xu2018alternating}\nC.~Xu, J.~Yao, Z.~Lin, W.~Ou, Y.~Cao, Z.~Wang, and H.~Zha, ``Alternating\n multi-bit quantization for recurrent neural networks,''\n \\emph{arXiv:1802.00150}, 2018.\n\\bibitem{mikolov2013word2vec}\nT.~Mikolov, K.~Chen, G.~Corrado, J.~Dean, L.~Sutskever, and G.~Zweig,\n ``word2vec,'' \\emph{URL https://code. google. com/p/word2vec}, vol.~22, 2013.\n\\bibitem{shen2019q}\nS.~Shen, Z.~Dong, J.~Ye, L.~Ma, Z.~Yao, A.~Gholami, M.~W. Mahoney, and\n K.~Keutzer, ``Q-bert: Hessian based ultra low precision quantization of\n bert,'' \\emph{arXiv:1909.05840}, 2019.\n\\bibitem{ba2014deep}\nJ.~Ba and R.~Caruana, ``Do deep nets really need to be deep?'' in \\emph{NIPS},\n 2014, pp. 2654--2662.\n\\bibitem{sau2016deep}\nB.~B. Sau and V.~N. Balasubramanian, ``Deep model compression: Distilling\n knowledge from noisy teachers,'' \\emph{arXiv:1610.09650}, 2016.\n\\bibitem{hinton2015distilling}\nG.~Hinton, O.~Vinyals, and J.~Dean, ``Distilling the knowledge in a neural\n network,'' \\emph{arXiv:1503.02531}, 2015.\n\\bibitem{czarnecki2017sobolev}\nW.~M. Czarnecki, S.~Osindero, M.~Jaderberg, G.~Swirszcz, and R.~Pascanu,\n ``Sobolev training for neural networks,'' in \\emph{NIPS}, 2017, pp.\n 4278--4287.\n\\bibitem{mishra2017apprentice}\nA.~Mishra and D.~Marr, ``Apprentice: Using knowledge distillation techniques to\n improve low-precision network accuracy,'' \\emph{arXiv:1711.05852}, 2017.\n\\bibitem{polino2018model}\nA.~Polino, R.~Pascanu, and D.~Alistarh, ``Model compression via distillation\n and quantization,'' \\emph{arXiv:1802.05668}, 2018.\n\\bibitem{romero2014fitnets}\nA.~Romero, N.~Ballas, S.~E. Kahou, A.~Chassang, C.~Gatta, and Y.~Bengio,\n ``Fitnets: Hints for thin deep nets,'' \\emph{arXiv:1412.6550}, 2014.\n\\bibitem{yim2017gift}\nJ.~Yim, D.~Joo, J.~Bae, and J.~Kim, ``A gift from knowledge distillation: Fast\n optimization, network minimization and transfer learning,'' in \\emph{CVPR},\n 2017, pp. 4133--4141.\n\\bibitem{mcclure2016representational}\nP.~McClure and N.~Kriegeskorte, ``Representational distance learning for deep\n neural networks,'' \\emph{Frontiers in computational neuroscience}, vol.~10,\n p. 131, 2016.\n\\bibitem{kim2016sequence}\nY.~Kim and A.~M. Rush, ``Sequence-level knowledge distillation,''\n \\emph{arXiv:1606.07947}, 2016.\n\\bibitem{freitag2017ensemble}\nM.~Freitag, Y.~Al-Onaizan, and B.~Sankaran, ``Ensemble distillation for neural\n machine translation,'' \\emph{arXiv:1702.01802}, 2017.\n\\bibitem{zhang2018deep}\nY.~Zhang, T.~Xiang, T.~M. Hospedales, and H.~Lu, ``Deep mutual learning,'' in\n \\emph{CVPR}, 2018, pp. 4320--4328.\n\\bibitem{anil2018large}\nR.~Anil, G.~Pereyra, A.~Passos, R.~Ormandi, G.~E. Dahl, and G.~E. Hinton,\n ``Large scale distributed neural network training through online\n distillation,'' \\emph{arXiv:1804.03235}, 2018.\n\\bibitem{mirzadeh2019improved}\nS.-I. Mirzadeh, M.~Farajtabar, A.~Li, N.~Levine, A.~Matsukawa, and\n H.~Ghasemzadeh, ``Improved knowledge distillation via teacher assistant,''\n \\emph{arXiv:1902.03393}, 2019.\n\\bibitem{you2017learning}\nS.~You, C.~Xu, C.~Xu, and D.~Tao, ``Learning from multiple teacher networks,''\n in \\emph{KDD}, 2017, pp. 1285--1294.\n\\bibitem{tan2019multilingual}\nX.~Tan, Y.~Ren, D.~He, T.~Qin, Z.~Zhao, and T.-Y. Liu, ``Multilingual neural\n machine translation with knowledge distillation,'' \\emph{arXiv:1902.10461},\n 2019.\n\\bibitem{heo2019knowledge}\nB.~Heo, M.~Lee, S.~Yun, and J.~Y. Choi, ``Knowledge distillation with\n adversarial samples supporting decision boundary,'' in \\emph{AAAI}, vol.~33,\n 2019, pp. 3771--3778.\n\\bibitem{xu2017training}\nZ.~Xu, Y.-C. Hsu, and J.~Huang, ``Training shallow and thin networks for\n acceleration via knowledge distillation with conditional adversarial\n networks,'' \\emph{arXiv:1709.00513}, 2017.\n\\bibitem{wang2018kdgan}\nX.~Wang, R.~Zhang, Y.~Sun, and J.~Qi, ``Kdgan: Knowledge distillation with\n generative adversarial networks,'' in \\emph{NIPS}, 2018, pp. 775--786.\n\\bibitem{zhao2019extreme}\nS.~Zhao, R.~Gupta, Y.~Song, and D.~Zhou, ``Extreme language model compression\n with optimal subwords and shared projections,'' \\emph{arXiv:1909.11687},\n 2019.\n\\bibitem{sun2019patient}\nS.~Sun, Y.~Cheng, Z.~Gan, and J.~Liu, ``Patient knowledge distillation for bert\n model compression,'' \\emph{arXiv:1908.09355}, 2019.\n\\bibitem{jiao2019tinybert}\nX.~Jiao, Y.~Yin, L.~Shang, X.~Jiang, X.~Chen, L.~Li, F.~Wang, and Q.~Liu,\n ``Tinybert: Distilling bert for natural language understanding,''\n \\emph{arXiv:1909.10351}, 2019.\n\\bibitem{liu2019improving}\nX.~Liu, P.~He, W.~Chen, and J.~Gao, ``Improving multi-task deep neural networks\n via knowledge distillation for natural language understanding,''\n \\emph{arXiv:1904.09482}, 2019.\n\\bibitem{wang2020minilm}\nW.~Wang, F.~Wei, L.~Dong, H.~Bao, N.~Yang, and M.~Zhou, ``Minilm: Deep\n self-attention distillation for task-agnostic compression of pre-trained\n transformers,'' \\emph{arXiv:2002.10957}, 2020.\n\\bibitem{tang2019distilling}\nR.~Tang, Y.~Lu, L.~Liu, L.~Mou, O.~Vechtomova, and J.~Lin, ``Distilling\n task-specific knowledge from bert into simple neural networks,''\n \\emph{arXiv:1903.12136}, 2019.\n\\bibitem{mukherjee2020xtremedistil}\nS.~Mukherjee and A.~H. Awadallah, ``Xtremedistil: Multi-stage distillation for\n massive multilingual models,'' in \\emph{ACL}, 2020, pp. 2221--2234.\n\\bibitem{ling2015finding}\nW.~Ling, T.~Lu{\\'\\i}s, L.~Marujo, R.~F. Astudillo, S.~Amir, C.~Dyer, A.~W.\n Black, and I.~Trancoso, ``Finding function in form: Compositional character\n models for open vocabulary word representation,'' \\emph{arXiv:1508.02096},\n 2015.\n\\bibitem{jozefowicz2016exploring}\nR.~Jozefowicz, O.~Vinyals, M.~Schuster, N.~Shazeer, and Y.~Wu, ``Exploring the\n limits of language modeling,'' \\emph{arXiv:1602.02410}, 2016.\n\\bibitem{kim2016character}\nY.~Kim, Y.~Jernite, D.~Sontag, and A.~M. Rush, ``Character-aware neural\n language models,'' in \\emph{AAAI}, 2016.\n\\bibitem{chen2015compressing}\nW.~Chen, J.~Wilson, S.~Tyree, K.~Weinberger, and Y.~Chen, ``Compressing neural\n networks with the hashing trick,'' in \\emph{ICML}, 2015, pp. 2285--2294.\n\\bibitem{lu2016learning}\nZ.~Lu, V.~Sindhwani, and T.~N. Sainath, ``Learning compact recurrent neural\n networks,'' in \\emph{ICASSP}.\\hskip 1em plus 0.5em minus 0.4em\\relax IEEE,\n 2016, pp. 5960--5964.\n\\bibitem{tay2019lightweight}\nY.~Tay, A.~Zhang, L.~A. Tuan, J.~Rao, S.~Zhang, S.~Wang, J.~Fu, and S.~C. Hui,\n ``Lightweight and efficient neural natural language processing with\n quaternion networks,'' \\emph{arXiv:1906.04393}, 2019.\n\\bibitem{chen2016compressing}\nY.~Chen, L.~Mou, Y.~Xu, G.~Li, and Z.~Jin, ``Compressing neural language models\n by sparse word representations,'' \\emph{arXiv:1610.03950}, 2016.\n\\bibitem{li2016lightrnn}\nX.~Li, T.~Qin, J.~Yang, and T.-Y. Liu, ``Lightrnn: Memory and\n computation-efficient recurrent neural networks,'' in \\emph{NIPS}, 2016, pp.\n 4385--4393.\n\\bibitem{suzuki2016learning}\nJ.~Suzuki and M.~Nagata, ``Learning compact neural word embeddings by parameter\n space sharing.'' in \\emph{IJCAI}, 2016, pp. 2046--2052.\n\\bibitem{li2018slim}\nZ.~Li, R.~Kulhanek, S.~Wang, Y.~Zhao, and S.~Wu, ``Slim embedding layers for\n recurrent neural language models,'' in \\emph{AAAI}, 2018.\n\\bibitem{lan2019albert}\nZ.~Lan, M.~Chen, S.~Goodman, K.~Gimpel, P.~Sharma, and R.~Soricut, ``Albert: A\n lite bert for self-supervised learning of language representations,''\n \\emph{arXiv:1909.11942}, 2019.\n\\bibitem{dehghani2018universal}\nM.~Dehghani, S.~Gouws, O.~Vinyals, J.~Uszkoreit, and {\\L}.~Kaiser, ``Universal\n transformers,'' \\emph{arXiv:1807.03819}, 2018.\n\\bibitem{bai2019deep}\nS.~Bai, J.~Z. Kolter, and V.~Koltun, ``Deep equilibrium models,''\n \\emph{arXiv:1909.01377}, 2019.\n\\bibitem{oseledets2011tensor}\nI.~V. Oseledets, ``Tensor-train decomposition,'' \\emph{SIAM J. on Scientific\n Computing}, vol.~33, no.~5, pp. 2295--2317, 2011.\n\\bibitem{carroll1970analysis}\nJ.~D. Carroll and J.-J. Chang, ``Analysis of individual differences in\n multidimensional scaling via an n-way generalization of “eckart-young”\n decomposition,'' \\emph{Psychometrika}, vol.~35, no.~3, pp. 283--319, 1970.\n\\bibitem{tucker1966some}\nL.~R. Tucker, ``Some mathematical notes on three-mode factor analysis,''\n \\emph{Psychometrika}, vol.~31, no.~3, pp. 279--311, 1966.\n\\bibitem{prabhavalkar2016compression}\nR.~Prabhavalkar, O.~Alsharif, A.~Bruguier, and L.~McGraw, ``On the compression\n of recurrent neural networks with an application to lvcsr acoustic modeling\n for embedded speech recognition,'' in \\emph{ICASSP}.\\hskip 1em plus 0.5em\n minus 0.4em\\relax IEEE, 2016, pp. 5970--5974.\n\\bibitem{sak2014long}\nH.~Sak, A.~Senior, and F.~Beaufays, ``Long short-term memory based recurrent\n neural network architectures for large vocabulary speech recognition,''\n \\emph{arXiv:1402.1128}, 2014.\n\\bibitem{faruqui2015sparse}\nM.~Faruqui, Y.~Tsvetkov, D.~Yogatama, C.~Dyer, and N.~Smith, ``Sparse\n overcomplete word vector representations,'' \\emph{arXiv:1506.02004}, 2015.\n\\bibitem{shu2017compressing}\nR.~Shu and H.~Nakayama, ``Compressing word embeddings via deep compositional\n code learning,'' \\emph{arXiv:1711.01068}, 2017.\n\\bibitem{wang2019structured}\nZ.~Wang, J.~Wohlwend, and T.~Lei, ``Structured pruning of large language\n models,'' \\emph{arXiv:1910.04732}, 2019.\n\\bibitem{chen2015strategies}\nW.~Chen, D.~Grangier, and M.~Auli, ``Strategies for training large vocabulary\n neural language models,'' \\emph{arXiv:1512.04906}, 2015.\n\\bibitem{variani2019west}\nE.~Variani, A.~T. Suresh, and M.~Weintraub, ``West: Word encoded sequence\n transducers,'' in \\emph{ICASSP}.\\hskip 1em plus 0.5em minus 0.4em\\relax IEEE,\n 2019, pp. 7340--7344.\n\\bibitem{tjandra2017compressing}\nA.~Tjandra, S.~Sakti, and S.~Nakamura, ``Compressing recurrent neural network\n with tensor train,'' in \\emph{IJCNN}.\\hskip 1em plus 0.5em minus 0.4em\\relax\n IEEE, 2017, pp. 4451--4458.\n\\bibitem{khrulkov2019tensorized}\nV.~Khrulkov, O.~Hrinchuk, L.~Mirvakhabova, and I.~Oseledets, ``Tensorized\n embedding layers for efficient model compression,'' \\emph{arXiv:1901.10787},\n 2019.\n\\bibitem{ye2018learning}\nJ.~Ye, L.~Wang, G.~Li, D.~Chen, S.~Zhe, X.~Chu, and Z.~Xu, ``Learning compact\n recurrent neural networks with block-term tensor decomposition,'' in\n \\emph{CVPR}, 2018, pp. 9378--9387.\n\\bibitem{ma2019tensorized}\nX.~Ma, P.~Zhang, S.~Zhang, N.~Duan, Y.~Hou, D.~Song, and M.~Zhou, ``A\n tensorized transformer for language modeling,'' \\emph{arXiv:1906.09777},\n 2019.\n\\bibitem{child2019generating}\nR.~Child, S.~Gray, A.~Radford, and I.~Sutskever, ``Generating long sequences\n with sparse transformers,'' \\emph{arXiv:1904.10509}, 2019.\n\\bibitem{guo2019star}\nQ.~Guo, X.~Qiu, P.~Liu, Y.~Shao, X.~Xue, and Z.~Zhang, ``Star-transformer,'' in\n \\emph{NAACL-HLT}, 2019, pp. 1315--1325.\n\\bibitem{kitaev2020reformer}\nN.~Kitaev, {\\L}.~Kaiser, and A.~Levskaya, ``Reformer: The efficient\n transformer,'' \\emph{arXiv:2001.04451}, 2020.\n\\bibitem{wang2020linformer}\nS.~Wang, B.~Li, M.~Khabsa, H.~Fang, and H.~Ma, ``Linformer: Self-attention with\n linear complexity,'' \\emph{arXiv:2006.04768}, 2020.\n\\bibitem{tay2020sparse}\nY.~Tay, D.~Bahri, L.~Yang, D.~Metzler, and D.-C. Juan, ``Sparse sinkhorn\n attention,'' \\emph{arXiv:2002.11296}, 2020.\n\\bibitem{shen2018efficient}\nZ.~Shen, M.~Zhang, H.~Zhao, S.~Yi, and H.~Li, ``Efficient attention: Attention\n with linear complexities,'' \\emph{arXiv:1812.01243}, 2018.\n\\bibitem{katharopoulos2020transformers}\nA.~Katharopoulos, A.~Vyas, N.~Pappas, and F.~Fleuret, ``Transformers are rnns:\n Fast autoregressive transformers with linear attention,''\n \\emph{arXiv:2006.16236}, 2020.\n\\bibitem{kapur2017low}\nS.~Kapur, A.~Mishra, and D.~Marr, ``Low precision rnns: Quantizing rnns without\n losing accuracy,'' \\emph{arXiv:1710.07706}, 2017.\n\\bibitem{hou2018loss}\nL.~Hou and J.~T. Kwok, ``Loss-aware weight quantization of deep networks,''\n \\emph{arXiv:1802.08635}, 2018.\n\\bibitem{grachev2019compression}\nA.~M. Grachev, D.~I. Ignatov, and A.~V. Savchenko, ``Compression of recurrent\n neural networks for efficient language modeling,'' \\emph{Applied Soft\n Computing}, vol.~79, pp. 354--362, 2019.\n\\bibitem{damani20_pakdd}\nS.~Damani, K.~N. Narahari, A.~Chatterjee, M.~Gupta, and P.~Agrawal, ``Optimized\n transformer models for faq answering,'' in \\emph{PAKDD}, 2020, p. To appear.\n\\bibitem{yang2018accelerating}\nY.~Yang, K.~Liang, X.~Xiao, Z.~Xie, L.~Jin, J.~Sun, and W.~Zhou, ``Accelerating\n and compressing lstm based model for online handwritten chinese character\n recognition,'' in \\emph{ICFHR}.\\hskip 1em plus 0.5em minus 0.4em\\relax IEEE,\n 2018, pp. 110--115.\n\\bibitem{lazaridou2013fish}\nA.~Lazaridou, E.~M. Vecchi, and M.~Baroni, ``Fish transporters and miracle\n homes: How compositional distributional semantics can help np parsing,'' in\n \\emph{EMNLP}, 2013, pp. 1908--1913.\n\\bibitem{wangperawong2018attending}\nA.~Wangperawong, ``Attending to mathematical language with transformers,''\n \\emph{arXiv:1812.02825}, 2018.\n\\bibitem{linzen2016assessing}\nT.~Linzen, E.~Dupoux, and Y.~Goldberg, ``Assessing the ability of lstms to\n learn syntax-sensitive dependencies,'' \\emph{TACL}, vol.~4, pp. 521--535,\n 2016.\n\\bibitem{zaremba2014learning}\nW.~Zaremba and I.~Sutskever, ``Learning to execute,'' \\emph{arXiv:1410.4615},\n 2014.\n\\end{thebibliography}\n\\bibliographystyle{ACM-Reference-Format}\n\\bibliography{referencesShort}\n\\end{document}\n\\endinput","id":"7ea3f834-f595-4c1b-afed-044d53e835ea","level":"subsection","origin_cites_number":3,"parent_id":"d06ca9cc-bcd7-4f53-9be8-ee86f8754664","prefix_titles":[["title","Compression of Deep Learning Models for Text: A Survey"],["section","Summary and Future Directions"],["subsection","Future Trends"]],"subsections":[],"title":"Future Trends"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ]\n ],\n \"subsections\": [\n \"57df7b91-17f0-42ac-b043-8a5e492f9515\",\n \"92ac6101-b828-4042-8708-2e0386e701bb\",\n \"c0c0e7e6-55dc-426b-8669-a8adfb8fe243\",\n \"2c5b1b2b-3f74-442d-8ee6-503425cf4207\",\n \"57903397-1a0f-4365-9fd9-a6bf95e76421\",\n \"a059247f-141a-43b7-a6a9-f3440e7622cd\",\n \"65404f84-cbf0-4721-8f05-178f8230aa55\",\n \"d06ca9cc-bcd7-4f53-9be8-ee86f8754664\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:overview}\\nIn this survey, we discuss compression methods using pruning, quantization, knowledge distillation, parameter sharing, tensor decomposition and sub-quadratic Transformers. \\nThe most obvious way to reduce model size is to sparsify weight matrices. Pruning methods differ based on what is pruned and the actual logic used to prune. Given a matrix, one can prune \\n\\\\begin{itemize}\\n \\\\item Some weight entries\\n \\\\item Rows or columns (i.e., neurons)\\n \\\\item Weight blocks\\n \\\\item Attention heads (in case of Transformer based methods)\\n \\\\item Layers or a combination of the structures.\\n\\\\end{itemize}\\nOther interesting aspects of pruning methods include the following.\\n\\\\begin{itemize}\\n \\\\item How to decide which weights/neurons/blocks/heads to prune?\\n \\\\item Should you prune large networks or build small networks?\\n \\\\item Should you do one-shot pruning versus iterative/gradual pruning? \\n \\\\item How does regularization help when pruning?\\n\\\\end{itemize}\\n We discuss these aspects of pruning based methods in Section~\\\\ref{sec:pruning}.\\nBesides removing the weights themselves, another way to reduce the size of weight matrices is to reduce the number of bits needed to represent each weight. Typically weights are stored as 32-bit double values. In an extreme case, weights can be quantized to two values (binary 1 bit). But other popular ways include quantization to three values (ternary) or multiple bits. Quantization can be uniform or non-uniform. Quantization methods can be deterministic or stochastic. Quantization can be performed in a loss-aware or unaware manner. Quantization bin ranges can be trained versus tuned. Finally, the level of quantization needs to be different across layers of RNNs, LSTMs or Transformers to attain a favorable model size versus accuracy tradeoff. We discuss these aspects of quantization based methods in Section~\\\\ref{sec:quantization}.\\nA third way of doing model compression is using knowledge distillation (also broadly known as student teacher networks). In such methods, the idea is to first train a deep teacher model using labeled data, and then transfer ``dark knowledge'' from teacher to train a shallow student network. Thus, the student model is trained to mimic a pre-trained, larger teacher. After the student is trained, it is deployed while the teacher network can be thrown. Distillation methods vary based on the following design choices.\\n\\\\begin{itemize}\\n \\\\item Different types of teacher model\\n \\\\item Different types of loss function like squared error between the logits of the models, KL divergence between the predictive distributions, or some other measure of agreement between the model predictions\\n \\\\item Different choices for what dataset the student model trains on (a large unlabeled dataset, a held-out data set, or the original training set)\\n \\\\item Different choices for what to mimic from the teacher -- teacher's class probabilities, teacher's feature representation.\\n \\\\item Learn from whom -- teacher, teacher assistant, or other fellow students. \\n\\\\end{itemize}\\nWe discuss these aspects of knowledge distillation based methods in Section~\\\\ref{sec:knowledgeDistillation}.\\nAnother way that reduces overall weights is to have parameters shared across multiple weight structures. Broadly, the method is called parameter sharing. Methods differ depending on the following aspects.\\n\\\\begin{itemize}\\n \\\\item Which parameters are shared\\n \\\\item Technique used to share parameters\\n \\\\item The level at which sharing is performed\\n\\\\end{itemize}\\nWe discuss these aspects of parameter sharing based methods in Section~\\\\ref{sec:parameterSharing}.\\nYet another way to avoid large matrices is to approximate them using a combination of smaller matrices. Such tensor decomposition methods for model compression factorize large tensors into multiple smaller components. Methods differ based on the following aspects.\\n\\\\begin{itemize}\\n \\\\item Type of factorization technique\\n \\\\item Matrices being factorized\\n \\\\item The property of weight matrix being exploited\\n\\\\end{itemize}\\nWe discuss these aspects of tensor decomposition methods in Section~\\\\ref{sec:matrixDecomposition}.\\nIn Transformer based models, besides the model size, latency is a concern because of quadratic complexity in terms of the input sequence size. Also, RAM needed for Transformer models is quadratic in nature. Hence, very recently, there have been several efforts on designing Transformers with sub-quadratic complexity. Some of these methods are super-linear while many are linear. Linear complexity methods use various techniques for enforcing linearity -- the broad idea is to compute a transformed representation for every token using attention over a fixed small number of other tokens. Methods differ in terms of defining the set of other tokens to be used to compute a transformed representation for the current token. We discuss such methods in Section~\\\\ref{sec:linearTransformers}.\",\n \"id\": \"57df7b91-17f0-42ac-b043-8a5e492f9515\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Model Compression Methods: Overview\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model Compression Methods: Overview\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 844\n ],\n \"content\": \"\\\\label{sec:pruning} \\nThe first proposed methods for model compression were based on pruning. Pruning can be done in two ways: structured versus unstructured. In unstructured pruning, individual weight connections are removed from a network by setting them to 0. One can prune away weights from a weight matrix depending on various criteria (e.g., prune away low magnitude weights). Such unstructured pruning methods lead to sparse matrices and need special sparse matrix manipulation libraries at inference time. Hence, various structured pruning methods have also been proposed which aim to prune away structures like neurons, weight matrix blocks, attention heads or layers. Fig.~\\\\ref{fig:pruning} provides a broad overview of various pruning styles. Fig~\\\\ref{fig:pruning}(B) depicts unstructured pruning. Fig~\\\\ref{fig:pruning}(C)-(G) shows various structured pruning methods. In this section, we discuss unstructured weight pruning methods in Section~\\\\ref{subsec:PruningWeights} and structured pruning methods in Sections~\\\\ref{subsec:PruningNeurons}-\\\\ref{subsec:PruningHeadsLayers}.\\n\\\\begin{figure*}\\n \\\\centering\\n \\\\includegraphics[width=\\\\textwidth]{pruning}\\n \\\\caption{Different Types of Pruning: (A) represents no pruning. Filled cells represent pruned entries. (B), (H) and (I) are unstructured pruning methods discussed in Section~\\\\ref{subsec:PruningWeights}. Remaining are structured pruning methods: (C) is discussed in Section~\\\\ref{subsec:PruningNeurons}, (D) in Section~\\\\ref{subsec:PruningBlock}, (E), (F) and (G) in Section~\\\\ref{subsec:PruningHeadsLayers}.}\\n \\\\label{fig:pruning}\\n\\\\end{figure*}\\nIn pruning, the main idea is to grow a large model and then prune away weights to end up with a much smaller but effective model. This is inspired by the following biological observation. Trillions of synapses are generated in the human brain during the first few months of birth. At one year old, synapse count peaks at 1000 trillion. And then natural pruning begins to occur. A ten year old child has nearly 500 trillion synapses. This `pruning' mechanism removes redundant connections in the brain~. \\nOne natural question is should you prune large networks or build small dense networks? Pruning involves extra processing plus sparse matrices need special handling. Can we avoid it by training smaller models? Zhu et al.~ experimented with models of various sizes with/ without pruning of stacked LSTMs models for language modeling, and seq2seq models for Neural Machine Translation (NMT). They found that large-sparse models consistently outperform small-dense models and achieve up to 10x reduction in number of non-zero parameters with minimal loss in accuracy.\",\n \"id\": \"92ac6101-b828-4042-8708-2e0386e701bb\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ]\n ],\n \"subsections\": [\n \"2362b1c7-732c-4073-87eb-1e3f045414b0\",\n \"56b21370-ac94-4d19-9ed7-985faad42cc8\",\n \"49a2e5f4-1aec-4e81-ac60-4671c95948a6\",\n \"10a33a73-e952-454b-b5ce-09086a71de83\",\n \"80f88ea7-37a5-4100-9fe4-121a616869c9\"\n ],\n \"title\": \"Pruning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:PruningWeights}\\nThe weights to be pruned can be chosen using two heuristics: (1) Using Hessian matrix computation, or (2) using magnitude of the weights. Further, magnitude based pruning methods could do pruning in one shot (typically statically after training is done), or iteratively (also called dynamic or gradual pruning), or iteratively with pruning and densification. Accordingly, there are four main ways of performing unstructured weight pruning: (1) Hessian based methods, (2) magnitude pruning, (3) iterative magnitude pruning, and (4) iterative magnitude pruning and densification. In this subsection, we discuss these methods in detail.\",\n \"id\": \"2362b1c7-732c-4073-87eb-1e3f045414b0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"92ac6101-b828-4042-8708-2e0386e701bb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Weights\"\n ]\n ],\n \"subsections\": [\n \"17e91dfe-059c-4ae9-8bb2-0fdcc9c148d7\",\n \"db3c2850-1885-440e-a725-4d53c91cb9a3\",\n \"bdf39c62-744f-439a-a336-cc2914c01b9e\",\n \"8ead8d95-de3b-4322-8928-4d2d5dd4066f\"\n ],\n \"title\": \"Pruning Weights\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In their seminal work (Optimal Brain Damage or OBD) on proposing weight pruning as a method for model compression, LeCun et al.~ defined saliency of a weight as change in the objective function $E$ caused by deleting that parameter. Using Taylor series and making multiple assumptions, they propose to use the following as a measure of saliency of weight $u_i$.\\n\\\\begin{eqnarray}\\n s(u_i)=\\\\frac{1}{2}\\\\sum_i h_{ii} \\\\delta u_i^2\\n\\\\end{eqnarray}\\n\\\\noindent where $h_{ii}=\\\\frac{\\\\partial^2 E}{\\\\partial u_i \\\\partial u_j}$ is the $i^{th}$ element on the diagonal of the Hessian matrix. Weights with low saliency can be pruned and the pruned network can be retrained to adjust the remaining weights. \\nThe procedure for computation of the diagonal of the Hessian is very similar to the back-propagation algorithm used for computing the first derivatives. Hence, computing the diagonal of the Hessian is of the same order of complexity as computing the gradient. \\nOBD ignores cross terms in the Hessian matrix. But on most real datasets, Hessian is strongly non-diagonal. Hence, to avoid pruning of important weights, Hassibi et al.~ proposed a method called Optimal Brain Surgeon (OBS) which considers cross terms as well. Using a similar derivative of $E$ wrt weight $w_i$, saliency of the weight is computed as follows.\\n\\\\begin{eqnarray}\\nL_i=\\\\frac{1}{2}\\\\frac{w_i^2}{[H^{-1}]_{ii}} \\n\\\\end{eqnarray}\\nComputing $H^{-1}$ is difficult. Hence, they provide a faster recursion relation for calculating $H^{-1}$ from training data and structural information of the network. Also, unlike other methods (like OBD or magnitude pruning), OBS does not demand (typically slow) retraining after the pruning of a weight. In every step, we delete $w_i$ with min $L_i$ and update all weights as follows.\\n\\\\begin{eqnarray}\\n \\\\delta w=-\\\\frac{w_i}{[H^{-1}]_{ii}}H^{-1}e_i\\n\\\\end{eqnarray}\\n\\\\noindent where $e_i$ is the unit vector in weight space corresponding to (scalar) weight $w_i$. Unfortunately, these methods (OBD and OBS) are computationally prohibitive as second derivative (i.e. Hessian) computations are expensive.\",\n \"id\": \"17e91dfe-059c-4ae9-8bb2-0fdcc9c148d7\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"2362b1c7-732c-4073-87eb-1e3f045414b0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Weights\"\n ],\n [\n \"subsubsection\",\n \"Hessian based Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hessian based Methods\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 4478\n ],\n \"content\": \"A more computationally feasible method for pruning connections and relearning weights based solely on the magnitude of the original weights is to simply prune away weights with small magnitudes. The idea was first proposed by Han et al.~. Pruning away low magnitude weights makes the matrix sparse. Sparse matrices can be stored in Compressed Sparse Row/Column (CSR/CSC) formats. Further space can be saved by storing the index difference instead of the absolute position, and encoding this difference using small fixed number of bits. See et al.~ experimented with these pruning methods on encoder-decoder LSTM NMT (neural machine translation) models. They perform magnitude pruning on all weight matrices of a 4-layer LSTM. They find that higher layers, attention and softmax weights are the most important, while lower layers and the embedding weights hold a lot of redundancy. At the lower layers the parameters for the input are most crucial, but at higher layers the parameters for the gates also become important. These methods typically have a target pruning percent as a hyper-parameter and pruning is either performed statically (after training the full model) or dynamically (while training itself). Retraining the sparse pruned network helps in improving accuracy.\\nIn a typical encoder-decoder LSTM model, there are these weight classes: source embedding weights, target embedding weights, source layer weights, target layer weights, attention weights and softmax weights. An important consideration related to magnitude pruning is how do we distribute the pruning over these different weight classes of a model, given a target $x$\\\\% pruning? Three ways suggested by See et al.~ include class-blind, class-uniform and class-distribution. In the class-blind way, we take all parameters, sort them by magnitude and prune the $x$\\\\% with smallest magnitude, regardless of the weight class. So some classes are pruned proportionally more than others. In the class-uniform way, Within each class, we sort the weights by magnitude and prune the x\\\\% with smallest magnitude. So all classes have exactly x\\\\% of their parameters pruned. In the class-distribution scheme, for each class $c$, weights with magnitude less than $\\\\lambda\\\\sigma_c$ are pruned. Here, $\\\\sigma_c$ is the standard deviation of that class and $\\\\lambda$ is a universal parameter chosen such that in total, $x$\\\\% of all parameters are pruned. Class-blind pruning is the simplest and adheres to the principle that pruning weights (or equivalently, setting them to zero) is least damaging when those weights are small, regardless of their locations in the architecture. Class-uniform pruning and class-distribution pruning both seek to prune proportionally within each weight class, either absolutely, or relative to the standard deviation of that class. They observe that class-blind pruning outperforms both other schemes.\",\n \"id\": \"db3c2850-1885-440e-a725-4d53c91cb9a3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"2362b1c7-732c-4073-87eb-1e3f045414b0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Weights\"\n ],\n [\n \"subsubsection\",\n \"Magnitude Pruning Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Magnitude Pruning Methods\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 4331,\n 4479,\n 844\n ],\n \"content\": \"Typically, it has been seen that rather than pruning in one-shot, it is a good idea to prune gradually over epochs. This way of pruning is called iterative or gradual pruning (see Fig.~\\\\ref{fig:pruning}(H)). For starting proportion $x$\\\\% and ending proportion $y$\\\\%, iterative magnitude pruning procedure prunes $x$\\\\% of each of the weights, does re-training, and then prunes $(y-x)/T$\\\\% of the weights every $K$ iterations. $T$ is the number of times, pruning is done. Sometimes, pruning is started after few warmup iterations have already been performed. Magnitude pruning has been seen to be very effective with regularization ($L_1/L_2$) while training. Dropouts also help in effective pruning. \\nIn some pruning methods, a weight once pruned cannot be a part of the network in future iterations. On the other hand, other methods do not modify the gradients of a pruned weight in the back-propagation step. In that case, it is possible for the updates of a pruned weight to be larger than the threshold of that layer, and then the weight will be involved in the forward pass again. Also, in every pruning iteration, we could either use a fixed threshold~ or monotonically increase it~. \\nIn case of gradual pruning~, where pruning threshold $\\\\epsilon$ is monotonically increased, $\\\\epsilon$ is determined as follows in every iteration $i$. Let $f$ be the number of iterations after which $\\\\epsilon$ is updated. Also, after a few warmup iterations, weights are sorted using absolute values and we pick the weight corresponding to the $90^{th}$ percentile as $q$. Pruning threshold $\\\\epsilon$ is increased in two stages. In the first stage (which starts at start iteration $s$ and continues until ramp iteration $r$, we use $\\\\theta$ as the initial slope to prune weights. In the second stage (which starts at ramp iteration $r$ and continues until end iteration $e$), we use $\\\\phi$ as the ramp slope to change the rate of pruning. Typically, $\\\\phi$ is set to 1.5$\\\\theta$ where $\\\\theta$ is calculated as follows.\\n\\\\begin{eqnarray}\\n \\\\theta=\\\\frac{2qf}{2(r-s)+3(e-r)}\\n\\\\end{eqnarray}\\nThus, from iteration $s$ to $r$, we set the pruning threshold as follows.\\n\\\\begin{eqnarray}\\n\\\\epsilon=\\\\frac{\\\\theta(i-s+1)}{f} \\n\\\\end{eqnarray}\\nFrom iterations $r+1$ to $e$, we set the pruning threshold as follows.\\n\\\\begin{eqnarray}\\n\\\\epsilon=\\\\frac{\\\\theta(r-s+1)+\\\\phi(i-r+1)}{f}\\n\\\\end{eqnarray}\\nTypically when pruning, biases are not pruned since they are much fewer in number. Overall, RNN/LSTM model size can be reduced by 90\\\\% and speed-up is around 2x to 7x using gradual pruning with no deterioration in accuracy. Also, layers closer to input are pruned more aggressively compared to the final layers.\\nAnother way of performing iterative pruning is to set a pruning target per iteration~. In this scheme, we start with an initial sparsity value $s_0$. To achieve a final sparsity value of $s_f$ after $n$ pruning steps with pruning frequency $f$, pruning target in iteration $i$ can be computed as follows.\\n\\\\begin{eqnarray}\\ns_i=s_f+(s_0-s_f)\\\\left(1-\\\\frac{i}{nf}\\\\right)^3\\n\\\\end{eqnarray}\\nThus, the sparsity of the network is gradually increased while allowing the network training steps to recover from any pruning-induced loss in accuracy. We prune the network rapidly in the initial phase when the redundant connections are abundant and gradually reduce the number of weights being pruned each time as there are fewer and fewer weights remaining in the network.\\nCheong et al.~ found that iterative pruning leads to poor results when pruning Transformer models like BERT. Guo et al.~ found that there are two problems with pruning especially when done with regularization. \\n\\\\begin{itemize}\\n \\\\item The larger weights $w_j$ are penalized more heavily than smaller weights $w_i$ in $L_1$ regularization, which violates the original intention of weight pruning, ``removing the unimportant connections''. \\n \\\\item Direct optimization of a regularization penalty term causes divergence from the original loss function and has negative effect on the effectiveness of gradient-based update.\\n\\\\end{itemize}\\nThey propose to perform reweighted $L_1$ minimization where $\\\\alpha_i>0$ are inversely proportional to magnitude of corresponding weights $|w_i|$. Thus, they solve the following optimization problem\\n\\\\begin{eqnarray}\\n \\\\min_w f(w)+\\\\gamma \\\\sum_i \\\\alpha_i |w_i|\\n\\\\end{eqnarray}\\n\\\\noindent where $f(w)$ is the original loss function for the network. This optimization is solved using a reweighted proximal pruning (RPP) method (which depends on proximal operators). RPP decouples the goals of high sparsity from minimizing loss, and hence leads to improved accuracy even with high levels of pruning for BERT.\",\n \"id\": \"bdf39c62-744f-439a-a336-cc2914c01b9e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"2362b1c7-732c-4073-87eb-1e3f045414b0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Weights\"\n ],\n [\n \"subsubsection\",\n \"Iterative Magnitude Pruning Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Iterative Magnitude Pruning Methods\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 4480\n ],\n \"content\": \"Further, the effectiveness of pruning can be improved by performing pruning and densification~ alternately across multiple iterations (see Fig.~\\\\ref{fig:pruning}(I)). There are two ways of doing this. In the first method~, in each iteration, either pruning is performed or densification. The sparse training regularizes the model, and the dense training restores the pruned weights, increasing the model capacity without overfitting. Sparsification helps the optimizer escape saddle points, and leads to regularized training which converges to a significantly better minima. In the second method~, in every iteration some dormant weights can reappear in the network while other active ones can get pruned out. A dormant $w\\\\in W$ is activated iff $|w.grad|$ is larger than the $(100\\\\alpha)^{th}$ percentile of all elements in $|W.grad|$. A $w\\\\in W$ is removed iff $|w|$ is smaller than the $(100\\\\beta)^{th}$ percentile of all elements in $|W|$. $\\\\alpha$ and $\\\\beta$ refer to growth ratio, and pruning ratio, respectively.\",\n \"id\": \"8ead8d95-de3b-4322-8928-4d2d5dd4066f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"2362b1c7-732c-4073-87eb-1e3f045414b0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Weights\"\n ],\n [\n \"subsubsection\",\n \"Iterative Magnitude Pruning and Densification\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Iterative Magnitude Pruning and Densification\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:PruningNeurons}\\nIt is difficult to implement unstructured pruning practically since, at inference time, special support is needed for matrix multiplication in the sparse space. Pruning away neurons leads to removal of a row or a column from a weight matrix, thereby avoiding sparse matrix handling (see Fig.~\\\\ref{fig:pruning}(C)). However, compared to pruning weights, pruning neurons is less flexible since we need to find entire rows/columns for deletion. In this section, we discuss ways of determining neurons that can be pruned.\",\n \"id\": \"56b21370-ac94-4d19-9ed7-985faad42cc8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"92ac6101-b828-4042-8708-2e0386e701bb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Neurons\"\n ]\n ],\n \"subsections\": [\n \"60bed51b-08a1-4c61-9777-2265a58630d5\",\n \"e7c0d686-3915-4133-8281-2cebccd6dd52\"\n ],\n \"title\": \"Pruning Neurons\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4481,\n 4482\n ],\n \"content\": \"He et al.~ proposed three node importance functions to determine importance score for neurons. \\n\\\\begin{itemize}\\n \\\\item Entropy: For a neuron $i$, let $a_i$ ($d_i$) be the \\\\#instances with node output $> (or\\\\leq)$ 0.5 for binary classification with a sigmoid activation. Then entropy for node $i$ can be written as follows.\\n \\\\begin{eqnarray}\\n \\\\text{Entropy}(i)=\\\\frac{d_i}{a_i+d_i}\\\\log_2 \\\\frac{d_i}{a_i+d_i}+\\\\frac{a_i}{a_i+d_i}\\\\log_2 \\\\frac{a_i}{a_i+d_i} \\n \\\\end{eqnarray}\\n The intuition is that if one node's outputs are almost identical on all training data, these outputs do not generate variations to later layers and consequently the node may not be useful. \\n \\\\item Output-weights Norm (onorm): average $L_1$-norm of the weights of its outgoing links.\\n \\\\item Input-weights norm (inorm): average $L_1$-norm of the weights of its incoming links.\\n\\\\end{itemize}\\nAll the neurons are sorted by their scores and nodes with less importance values are removed. In most cases, onorm has been found to be the best among these importance functions. \\nSpecial regularizers have also been proposed to force neurons to push either all incoming or outgoing connection weights towards zero~. Specifically, for handling incoming connections, the following two regularizers are popular.\\n\\\\begin{itemize}\\n \\\\item $L_2$ norm on weight matrix $W$ defined as follows.\\n \\\\begin{eqnarray}\\n \\\\sum_i ||W_{i:}||_2=\\\\sum_i \\\\left(\\\\sum_j W^2_{ij}\\\\right)^{1/2}\\n \\\\end{eqnarray}\\n This puts equal pressure on each row, but within each row, the larger values contribute more, and therefore there is more pressure on larger values towards zero. \\n \\\\item $L_\\\\infty$ norm on weight matrix $W$ defined as follows.\\n \\\\begin{eqnarray}\\n \\\\sum_i ||W_{i:}||_\\\\infty=\\\\sum_i \\\\max_j |W_{ij}|\\n \\\\end{eqnarray}\\n This puts equal pressure on each row, but within each row, only the maximum value (or values) matter, and therefore the pressure towards zero is entirely on the maximum value(s).\\n\\\\end{itemize}\\nSimilar regularizers can easily be defined for outgoing connections as well.\",\n \"id\": \"60bed51b-08a1-4c61-9777-2265a58630d5\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"56b21370-ac94-4d19-9ed7-985faad42cc8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Neurons\"\n ],\n [\n \"subsubsection\",\n \"Removing Low Importance Neurons\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Removing Low Importance Neurons\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4483\n ],\n \"content\": \"Consider a simple network with one hidden layer with $n$ neurons. Thus, the output can be computed as follows.\\n\\\\begin{eqnarray}\\nz=a_1 h(W_1^TX)+a_2 h(W_2^TX)+...+a_n h(W_n^TX)\\n\\\\end{eqnarray}\\n\\\\noindent where $a_i$ and $W_i$ indicate weights. In case $W_1==W_2$, $h(w_1^TX)=h(w_2^TX)$. Thus, we can compute output as follows.\\n\\\\begin{eqnarray}\\nz=(a_1+a_2) h(W_1^TX)+0. h(W_2^TX)+...+a_n h(W_n^TX)\\n\\\\end{eqnarray} \\nIn general, whenever two weight sets ($W_1$, $W_2$) are equal, one of them can effectively be removed. This should be done with a surgery step, i.e., we need to alter the co-efficient $a_1$ to $a_1+a_2$. Of course, for many pairs of weight sets (i.e., neurons), $W_1$ and $W_2$ are not exactly the same. Hence, Srinivas et al.~ proposed this 3 step method for redundant neuron identification and removal. \\n\\\\begin{itemize}\\n \\\\item Compute saliency $s_{ij}$ for all possible neuron pairs (i, j) as follows.\\n \\\\begin{eqnarray}\\n s_{ij}=\\\\langle a_j^2 \\\\rangle||\\\\epsilon_{ij}||_2^2\\n \\\\end{eqnarray} \\n \\\\noindent where $\\\\langle a_j^2 \\\\rangle$ denotes the average of the quantity over all output neurons. Let $S$ be the matrix with all $s_{ij}$ values. \\\\item Pick the indices $(i',j')$ corresponding to the minimum $s_{ij}$. Delete the $j'$ neuron, and update $a_i'$ as follows.\\n \\\\begin{eqnarray}\\n a_i'\\\\leftarrow a_i'+a_j' \\n \\\\end{eqnarray}\\n \\\\item Update $S$ by removing the $j'^{th}$ column and row, and updating the $i'^{th}$ column (to account for the updated $a_i'$).\\n\\\\end{itemize}\",\n \"id\": \"e7c0d686-3915-4133-8281-2cebccd6dd52\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"56b21370-ac94-4d19-9ed7-985faad42cc8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Neurons\"\n ],\n [\n \"subsubsection\",\n \"Removing Redundant Neurons\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Removing Redundant Neurons\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7826\n ],\n \"content\": \"\\\\label{subsec:PruningBlock}\\nIn weight pruning, irregularity of sparse matrices limits the maximum performance and energy efficiency achievable on hardware accelerators. Pruning neurons avoids sparse matrix issues but is limited in term of overall pruning possible. Hence, block based pruning methods were introduced (see Fig.~\\\\ref{fig:pruning}(D)).\\nBlock-sparse formats store blocks contiguously in memory reducing irregular memory accesses. If the maximum magnitude weight of a block is below the current threshold, we set all the weights in that block to zeros. Similar to iterative weight pruning, block pruning~ can also be done iteratively using a monotonically growing threshold. Any blocks that had been zeroed out are held at zero even after pruning has ended resulting in a sparse model at the end of training. Just like weight pruning (as discussed in Section~\\\\ref{subsec:PruningWeights}), the start slope $\\\\theta$ and ramp slope $\\\\phi$ determine the rate at which the threshold increases. For block pruning, we need to modify the start slope to account for the number of elements in a block ($N_b$). Thus, the start slope for block pruning is typically set as follows.\\n\\\\begin{eqnarray}\\n\\\\theta_b=\\\\theta\\\\times\\\\sqrt[\\\\leftroot{-2}\\\\uproot{2}4]{N_b} \\n\\\\end{eqnarray}\\nFurther, to enable effective removal of blocks, Narang et al.~ propose group Lasso regularization method. Group lasso is a type of weight regularization that works on groups of weights and can zero out all the weights in a group. For each block, we add a loss term proportional to the $L_2$ norm of the block. Thus, we optimize for the following.\\n\\\\begin{eqnarray}\\n\\\\min_w f(w)+\\\\lambda_g \\\\sum_{g=1}^{G} ||w^{(g)}||_2 \\n\\\\end{eqnarray}\\nWhen we combine group lasso with block pruning, group lasso guides the selection of blocks to prune. Group lasso regularization is applied to coincide with the pruning schedule, i.e., we turn off regularization when the pruning schedule ends. Typically, inducing block sparsity with 4x4 blocks in vanilla RNNs and GRUs works well, compared to larger block sizes. Larger blocks require lower sparsity to maintain similar accuracy. \\nUnfortunately, it becomes challenging to maintain the same model accuracy when block sparsity is applied. Also, block sizes (i.e., pruning granularity) are application-sensitive, making it another hyper-parameter to tune. To avoid these problems, Cao et al.~ proposed a new method called Bank-Balanced Sparsity (BBS). BBS splits each weight matrix row into multiple equal-sized banks, and adopts fine-grained pruning to each bank independently to obtain identical sparsity among banks. Each bank has the same number of non-zero values. For example, retaining top two weights in each bank of size 4 implies a sparsity of 50\\\\%. We apply the BBS pruning method iteratively to a pre-trained network, and fine-tune the network after each pruning iteration to restore the model accuracy. BBS achieves almost the same model accuracy as unstructured sparsity and significantly outperforms block sparsity when pruning weights at the same sparsity level. BBS is also amenable to FPGA (Field Programmable Gate Arrays) acceleration because it inherently provides a balanced matrix partitioning for parallel computing.\",\n \"id\": \"49a2e5f4-1aec-4e81-ac60-4671c95948a6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"92ac6101-b828-4042-8708-2e0386e701bb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Blocks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pruning Blocks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:PruningHeadsLayers}\\nBesides neurons and blocks, for Transformer based models, structured pruning can also be applied to attention heads and entire layers. In this section, we discuss such methods.\",\n \"id\": \"10a33a73-e952-454b-b5ce-09086a71de83\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"92ac6101-b828-4042-8708-2e0386e701bb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Heads and Layers\"\n ]\n ],\n \"subsections\": [\n \"bc2b6e76-0a87-4b66-8b3c-2bc732d0a821\",\n \"c4355812-9f8a-4da4-86ff-8739c39b93e8\",\n \"7dc1a143-838c-41ca-90a0-ebe43039cc74\"\n ],\n \"title\": \"Pruning Heads and Layers\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4485,\n 4484\n ],\n \"content\": \"BERT-base model consists of 12 layers each with 12 attention heads. Similarly, a typical NMT encoder-decoder Transformer with 6 layers each for encoder as well as decoder contains 16 attention heads per layer. Michel et al.~ found that majority of attention heads can be removed without deviating too much from the original score. Surprisingly, in some cases removing an attention head (see Fig.~\\\\ref{fig:pruning}(E)) results in an increase in accuracy. When these heads are removed individually, only 8 (out of 96) heads in 6-layer WMT NMT Transformer (16 heads/layer) cause a statistically significant change in performance when they are removed from the model, half of which actually result in a higher BLEU score. For most layers, one head is indeed sufficient at test time, even though the network was trained with 12 (BERT) or 16 (WMT Transformer) attention heads. One can also do iterative pruning of multiple heads (rather than just one at a time) across layers. For iterative pruning, head importance score is defined using the expected sensitivity of the model to the mask variables $\\\\xi_h$ as follows.\\n\\\\begin{equation}\\n I_h=E_{x\\\\sim X}\\\\left|\\\\frac{\\\\partial L(x)}{\\\\partial \\\\xi_h}\\\\right|=E_{x\\\\sim X}\\\\left|\\\\text{Att}_h(x)^T\\\\frac{\\\\partial L(x)}{\\\\partial \\\\text{Att}_h(x)}\\\\right|\\n\\\\end{equation}\\n\\\\noindent where $X$ is the data distribution, $L(x)$ is the loss on sample $x$, and $Att_h(x)$ is the output of the attention head $h$ for instance $x$, . Intuitively, if $I_h$ has a high value then changing $\\\\xi_h$ is liable to have a large effect on the model. Hence, in every iteration heads with low $I_h$ values are pruned out. Michel et al.~ observed that pruning up to 20\\\\% and 40\\\\% of heads from NMT and BERT models respectively, did not lead to any noticeable negative impact on accuracy.\\nVoita et al.~ used two other head importance scores to prune attention heads from the NMT model. The two scoring methods were: (1) Layer-wise relevance propagation (LRP)~. LRP is a method for computing the relative contribution of neurons at one point in a network to neurons at another. (2) ``confidence'' of a head which is computed as the average of its maximum attention weight excluding the end of sentence symbol, where the average is taken over tokens in a set of sentences used for evaluation. For pruning the heads, they propose a method based on stochastic gates and a differentiable relaxation of the $L_0$ penalty. $L_0$ norm equals the number of non-zero components and pushes the model to switch off less important heads. They find that only a small subset of heads are important for translation. On the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU.\",\n \"id\": \"bc2b6e76-0a87-4b66-8b3c-2bc732d0a821\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"10a33a73-e952-454b-b5ce-09086a71de83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Heads and Layers\"\n ],\n [\n \"subsubsection\",\n \"Pruning Attention Heads\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pruning Attention Heads\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4486\n ],\n \"content\": \"Note that dropping attention heads does not reduce runtime as they are usually computed in parallel. While one can prune weights, neurons or attention heads, how can we design a scheme to prune away layers (see Fig.~\\\\ref{fig:pruning}(G))? The LayerDrop idea proposed in~ is inspired by DropConnect. DropConnect randomly drops weights while training on a batch. LayerDrop does structured dropout: it drops groups of weights, heads, feed forward network (FFN) matrices, or layers. The layers to be pruned can be decided using one of these ways: \\n\\\\begin{itemize}\\n \\\\item Every Other: Prune every other layer (with rate $p$), e.g., every $3^{rd}$ layer in a 12-layer BERT model. \\n \\\\item Search on Validation: Search for a set of layers to be pruned by checking their impact on a validation set. This entails trying various combinations. \\n \\\\item Data Driven Pruning: Learn the drop rate $p_d$ of each layer in a data driven manner.\\n\\\\end{itemize}\\nGiven a target drop rate $p$, we learn an individual drop rate $p_d$ for the layer at depth $d$ such that the average rate over layers is equal to $p$. At inference time, we forward only the fixed top-$k$ highest scoring layers based on the softmax output. Across the three methods, ``Every Other'' strategy works surprisingly well across many tasks and configurations. ``Search on Validation'' and ``Data Driven Pruning'' only offer marginal gains.\",\n \"id\": \"c4355812-9f8a-4da4-86ff-8739c39b93e8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"10a33a73-e952-454b-b5ce-09086a71de83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Heads and Layers\"\n ],\n [\n \"subsubsection\",\n \"Pruning Layers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pruning Layers\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 4331,\n 4479,\n 4487,\n 4478,\n 4486,\n 844,\n 4482,\n 4480,\n 4485,\n 7826\n ],\n \"content\": \"Lastly, Prasanna et al.~ experiment with pruning both the FFN layers (see Fig.~\\\\ref{fig:pruning}(F)) as well as attention heads (see Fig.~\\\\ref{fig:pruning}(E)) in a BERT network. Just like~, they assign a mask variable to each of these structures. To decide which structures to prune, we look at the expected sensitivity of the model to the mask variables. High sensitivity implies large impact on the model output and hence corresponding structures should be retained. They find that it is possible to find a subnetwork of elements that achieves performance comparable with that of the full model, and similarly-sized subnetworks sampled from the rest of the model perform worse. \\n\\\\begin{table}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|l|l|l|l|l|l|l|l|l|}\\n \\\\hline\\nTask&Dataset&Model&Method&Size (Pruned; Orig)&Eval. (Pruned; Orig)&Metric\\\\\\\\\\n\\\\hline\\n\\\\hline\\nLanguage modeling&Europarl v7 English &2-layer MLP&Prune Neurons~&5.06M; 5.1M&57; 100&Perplexity (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&2-layer LSTM&Iter. Mag.~&6.6M; 66M&80.24; 78.45&Perplexity (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Iter. Mag.~&20M; 66M&78.5; 78.8&Perplexity (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Block Sparsity~&20M; 66M&83; 78.8&Perplexity (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&BBS~&20M; 66M&78.5; 78.8&Perplexity (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Wikitext-103&Transformer&LayerDrop~&22M; 44M&19.5; 18.2&Perplexity (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&AFP from English Gigaword&2-layer MLP&Prune Neurons~&5.07M; 5.1M&107; 100&Perplexity (L)\\\\\\\\\\n\\\\hline\\nLinguistic acceptability&CoLA&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&82.8/76.3; 81.5&Matthews (H)\\\\\\\\\\n\\\\hline\\nMachine reading comp.&MRPC&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&88.1/83.5; 89.3&Acc (H)\\\\\\\\\\n\\\\hline\\nMachine reading comp.&MRPC&BERT-base&LayerDrop~&66M; 110M&85.3; 88.9&Acc (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de) &WMT14&Multi-layer LSTM&Mag.~&43M; 216M&20.91; 20.5&BLEU (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de) &WMT16&4-layer LSTM&Iter. Mag.~&23M; 211M&26.19; 26.77&BLEU (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de) &WMT16 &Transformer&LayerDrop~&22M; 44M&29; 29&BLEU (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de) &WMT17&Transformer&Iter. Mag.~&22M; 44M&26.4; 28.09&BLEU (H)\\\\\\\\\\n\\\\hline\\nParaphrasing&QQP&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&91.2/85.1; 91.2&Acc (H)\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD 1.1&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&90.23/85.3; 90.9&Acc (H)\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD 2.0&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&81.3/75.3; 81.9&Acc (H)\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST-2&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&92.4/91.3; 93.2&Acc (H)\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST-2&BERT-base&LayerDrop~&66M; 110M&92.5; 93.5&Acc (H)\\\\\\\\\\n\\\\hline\\nSpeech recognition&2100 hours English Speech&2 CONV+7 BiRNNs+CTC&Iter. Mag.~&11.1M; 67M&10.59; 10.67&CER (L) on dev\\\\\\\\\\n\\\\hline\\nSpeech recognition&2100 hours English Speech&2 CONV+7 BiGRUs+CTC&Iter. Mag.~&17.8M; 115M&9.76; 9.55&CER (L) on dev\\\\\\\\\\n\\\\hline\\nSpeech recognition&2100 hours English speech&2 CONV+7 BiRNNs+CTC&Block Sparsity~&25.8M; 67M&15.66; 15.36&CER (L) on test\\\\\\\\\\n\\\\hline\\nSpeech recognition&2100 hours English speech&2 CONV+7 BiRNNs+CTC&Block Sparsity+Group Lasso~&12.9M; 67M&15.89; 15.36&CER (L) on test\\\\\\\\\\n\\\\hline\\nSpeech recognition&2100 hours English speech&2 CONV+3 BiGRUs+CTC&Block Sparsity~&25.6M; 115M&16.23; 15.42&CER (L) on test\\\\\\\\\\n\\\\hline\\nSpeech recognition&2100 hours English speech&2 CONV+3 BiGRUs+CTC&Block Sparsity+Group Lasso~&10.8M; 115M&16.78; 15.42&CER (L) on test\\\\\\\\\\n\\\\hline\\nSpeech recognition&AN4&2 CONV+3 HLSTMs+CTC&Grow and Prune~&2.6M; 44.72M&10.37; 8.92&WER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&Switchboard (swb/fsh)&7-layer MLP&Prune Neurons~&12.2M; 32.19M&25.5/28.8; 25.7/28.8&WER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&TIMIT&5-layer MLP&Prune Neurons~&3.5M; 5.76M&20.7; 20.79&PER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&TIMIT&LSTM&Iter. Mag.~&0.32M; 3.2M&23.5; 23.5&PER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&TIMIT&LSTM&Block Sparsity~&0.32M; 3.2M&26.5; 23.5&PER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&TIMIT&LSTM&BBS~&0.32M; 3.2M&23.5; 23.5&PER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&WSJ 92&1 CONV+3 FC+1 BiRNN&DSD~&4.07M; 8.14M&27.9; 27.45&WER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&WSJ 92&2 CONV+7 BiRNNs+CTC&DSD~&33.5M; 67M&10.65; 9.02&WER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&WSJ 93&1 CONV+3 FC+1 BiRNN&DSD~&4.07M; 8.14M&32.99; 31.6&WER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&WSJ 93&2 CONV+7 BiRNNs+CTC&DSD~&33.5M; 67M&14.84; 13.44&WER (L)\\\\\\\\\\n\\\\hline\\nSummarization&CNN-Dailymail&Transformer&LayerDrop~&22M; 44M&39; 40&ROUGE (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&MNLI&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&86.1/77; 86.1&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&MNLI-m&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&85.7/82.5; 85.9&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&MNLI-m&BERT-base&LayerDrop~&66M; 110M&82.9; 84.6&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&QNLI&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&92.3/90.2; 91.3&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&QNLI&BERT-base&LayerDrop~&66M; 110M&89.4; 90.5&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&RTE&BERT-large&RPP/Iter. Mag.~&138M/170M; 340M&70.1/68.6; 70.1&Acc (H)\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n \\\\caption{Comparison of various pruning methods (sorted by Task and then Dataset). CONV=Convolution. CTC=Connectionist temporal classification. FC=Fully connected. HLSTM=hidden-layer LSTM~. In the metric column, H means high is better while L means low is better. PER/CER/WER=Phone/Character/Word error rate. For~, embedding weights have not been considered when computing model size in the table. Block sparsity methods use block size of 4x4. BBS uses 64 banks.}\\n \\\\label{tab:pruningSummary}\\n\\\\end{table}\",\n \"id\": \"7dc1a143-838c-41ca-90a0-ebe43039cc74\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"10a33a73-e952-454b-b5ce-09086a71de83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Pruning Heads and Layers\"\n ],\n [\n \"subsubsection\",\n \"Pruning General Structures\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pruning General Structures\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 4331,\n 4484,\n 4479,\n 4478,\n 4486,\n 844,\n 1568,\n 4485,\n 7826\n ],\n \"content\": \"Table~\\\\ref{tab:pruningSummary} compares various pruning methods across different tasks and datasets. Size and accuracy of both the original and the pruned model are shown. The papers report multiple (model size, model accuracy) pairs but we carefully chose the pair such that accuracy is typically within 10\\\\% of the original or the best pruned model accuracy reported. For language modeling, most popular datasets include PTB, Europarl v7 English, Wikitext-103 and AFP from English Gigaword. On PTB using LSTMs, we observe that Bank-Balanced Sparsity method~ leads to lowest perplexity. For the Linguistic Acceptability CoLA task, RPP~ resulted into a smaller and a more accurate model compared to iterative magnitude pruning. As expected in some senses, the pruned model provides better accuracy than the unpruned one since pruning causes regularization. For the machine reading comprehension, question answering and paraphrasing tasks also, RPP seems to work better than iterative magnitude pruning. For the machine translation (NMT) task, on English-German WMT datasets, pruned Transformer models provide better accuracy than pruned LSTM with comparable number of parameters. For the sentiment analysis task, on SST-2, although RPP leads to a better pruned model compared to iterative pruning, LayerDrop~ improves further on it with a model less than half of the RPP-pruned model. For the speech recognition task, experiments have been reported on 2100 hours English speech data, TIMIT, WSJ, Switchboard and AN4 datasets. On 2100 hours English speech data, Block Sparsity+Group Lasso is better than Block sparsity without regularization. Also, it is better than plain iterative magnitude pruning. On TIMIT, block sparsity~ leads to a more accurate 90\\\\% pruned LSTM model compared to the unpruned one. For the summarization task, LayerDrop~ can compress the model to half without any noticeable accuracy change. Finally, for the textual entailment task, experiments have been done on GLUE~ datasets: MNLI, MNLI-m, QNLI and RTE. Models pruned from BERT-large perform better than models pruned from BERT-base; RPP performs better than iterative magnitude pruning.\\nWhile older methods~ claimed that Hessian based methods were more effective than magnitude based pruning, almost all recent methods have been based on magnitude based pruning. See et al.~ proposed three pruning schemes. They make the following observations: (1) Class-blind pruning outperforms both other schemes. Further, the overall performance loss is caused disproportionately by a few classes: softmax weights, source and target embedding weights. (2) It seems that higher layers are more important than lower layers, and that attention and softmax weights are crucial in LSTMs. (3) After retraining the pruned NMT models, baseline performance (20.48 BLEU) is both recovered and improved upon, up to 80\\\\% pruning (20.91 BLEU), with only a small performance loss at 90\\\\% pruning (20.13 BLEU). (4) In LSTMs, the parameters corresponding to the less common words are more dispensable. Weights connecting to the input are most crucial, followed by the input gate, then the output gate, then the forget gate. This is particularly true of the lower layers, which focus primarily on the input. However for higher layers, especially on the target side, weights connecting to the gates are as important as those connecting to the input. \\nNarang et al.~ observe that for approximately same number of parameters, gradual/iterative pruning is 7\\\\% to 9\\\\% better than hard pruning. They also conclude that the initial layers are pruned more aggressively compared to the final layers. Zhu et al.~ advise that in order to get the best-performing sparse model of a certain size, we should train a dense model that is 5x-10x larger and then prune to the desired number of parameters rather than taking the largest and best-performing dense model and pruning this model by 20x or more to the desired number of parameters. Guo et al.~ find that RPP is much better than typical iterative pruning. In their experiments with BERT they find that for both original BERT and BERT pruned with RPP, the low-dimensional manifolds of the language representations are similar, showing the similar projection. This implies that the BERT applied with RPP keeps most of the language representation information similar to that from the original BERT.\\nFor block pruning, Narang et al.~ make the following observations: (1) We can create block-sparse RNNs with sparsity ranging from 80\\\\% to 90\\\\% with small loss in accuracy. This allows us to reduce the model size by roughly 10×. Block sparsity works with a variety of block sizes up to 32×32. (2) For block size 4×4, models with sparsity greater 90\\\\% yield a relative accuracy loss of 30\\\\% or higher. Similarly, for blocks of 16×16, models with sparsity greater than 86\\\\% have 30\\\\% or more accuracy loss. A similar trend is observed for block size 32×32. This indicates that there is a tradeoff between sparsity, block size and accuracy of the model. (3) For both block pruning and weight pruning, we see that the initial layers are pruned more aggressively compared to the final layers. Increasing sparsity in the layers closer to the output results in poor accuracy. Additionally, the variance in sparsity across the layers increases with the block size. Further, Cao et al.~ make the following observations comparing block sparsity with BBS: (1) BBS achieves almost the same model accuracy regardless of the change of bank size. For block sparsity, however, increasing the block size adversely affects model accuracy. \\nFor pruning of attention heads, Michel et al.~ observe that one can prune up to 20\\\\% and 40\\\\% of heads from 6-layer NMT Transformer and BERT resp., without incurring any noticeable negative impact. When trying to remove a head at a time, only 8 (out of 96) heads in 6-layer NMT Transformer (16 heads/layer) cause a statistically significant change in performance when they are removed from the model, half of which actually result in a higher BLEU score. Further Voita et al.~ find that on the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU. \\nOverall, to summarize, pruning has been the most popular method for model compression. Pruning methods can be unstructured (prune weights) or structured (prune neurons, blocks, attention heads, layers). While weight pruning theoretically leads to pruning to a large extent, practical implementation of sparse data structures is difficult. Pruning and regularization need to be done together carefully. Also, it is critical to define the importance functions for various structures carefully. Among weight pruning methods, while iterative magnitude pruning with regularization works well for RNNs and LSTMs, RPP performs better for Transformer based models. Pruning blocks using BBS is better than pruning neurons. For Transformer models, pruning just the heads do not provide much model compression, but dropping a combination of attention heads and layers is better.\",\n \"id\": \"80f88ea7-37a5-4100-9fe4-121a616869c9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"92ac6101-b828-4042-8708-2e0386e701bb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Pruning\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:quantization}\\nWhile pruning saves on the model size by removing weights, quantization aims to reduce the number of bits needed to store weights. Most computer architectures use 32 bits to represent weights. However, estimated precision of the brain (hippocampal spine) synapses is around 4.6 bits~. Empirical evidence suggests that most quantities in the nervous system (for instance, firing of the neurons) have variability of a few percent due to biological noise, or a precision of 1 in 100 at best~. Thus, each decision could depend on $\\\\log_2 (100)$=6.64 bits. Thus, we should be able to store weights in our artificial neural networks on average in a space of 4--7 bits. Given this motivation, various methods have been proposed which perform 1-bit (or binary quantization), ternary quantization, and general quantization exploring the spectrum between 3 and 32 bits. We discuss such methods in this section. Fig.~\\\\ref{fig:quantization} provides a broad overview of various quantization styles.\\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=\\\\columnwidth]{quantization}\\n \\\\caption{Different Types of Quantization: Binary (A), Ternary (B) and General Quantized (C and D). Note that X axis denotes the weight value while the Y axis denotes frequency.}\\n \\\\label{fig:quantization}\\n\\\\end{figure}\",\n \"id\": \"c0c0e7e6-55dc-426b-8669-a8adfb8fe243\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ]\n ],\n \"subsections\": [\n \"11228467-1126-4614-9475-560634f98a9b\",\n \"60a24ca4-58df-4063-9d44-f538e8a26b98\",\n \"be929a80-0b3b-4005-bf19-e84fee8ab111\",\n \"97e94011-bad6-4d07-98cb-26de46e32fe4\"\n ],\n \"title\": \"Quantization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:binQuant}\\nBinarized networks use binary quantization (see Fig.~\\\\ref{fig:quantization}(A)), which quantizes weights using 1 bit. Quantizing weights to 1 bit provides a compression of 32x but leads to a significant drop in accuracy across many tasks. However, in a hybrid quantization scheme, such binary quantization can be very helpful for some layers in a network. Binarization can be done using deterministic methods or could be stochastic in nature. Also, while na\\\\\\\"ive binarization has a very simple way of fixing the binary boundary threshold, one could perform a complex loss aware binarization as well. We discuss these variants of binarization in this section.\",\n \"id\": \"11228467-1126-4614-9475-560634f98a9b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"c0c0e7e6-55dc-426b-8669-a8adfb8fe243\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Binarized Networks\"\n ]\n ],\n \"subsections\": [\n \"f998a362-cba7-42b7-aa49-83d293cf4b58\",\n \"3c274eec-1522-4b1c-b25f-1f3d927c3167\",\n \"bb389108-4900-443c-bee1-d687f6265cf4\"\n ],\n \"title\": \"Binarized Networks\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 4351\n ],\n \"content\": \"Simplest way of binary quantization is to set the weight as 1 for non-negative weights, and to -1 for negative weights. This leads to 32x compression. Also, the matrix multiplication for binary matrices is $\\\\sim$7x faster~ leading to faster model inference. In the forward pass, binary networks drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which leads to great increases of power efficiency. Also, in the simplest version, binarization can be performed in a static manner, i.e., after the training is done. However, this method leads to large loss in accuracy. \\nA variant of this simple method is to set the weight to a constant $c_1$ for non-negative weights, and to another constant $c_2$ for negative weights. Binary Scheme (BS)-Fixed method~ stores the original weights and during the forward pass replaces the values with a masked value of $c_1$ or $c_2$, where $c_1$ and $c_2$ are fixed and chosen with hyperparameter tuning. Full precision weights are used during training. At the end of training, the weights are replaced with the index of its masked value. Choosing the values of $c_1$ and $c_2$ can be difficult and time-consuming in BS-Fixed. Thus, in the BS-flexible method~, we initialize $c_1$ and $c_2$ using KMeans with two centroids over the weights, and then update $c_1$ and $c_2$ using back-propagation. Also, in the BS-Flexible method, weights are quantized as follows.\\n\\\\begin{equation}\\n w_b= \\\\begin{cases}\\nc_1 &\\\\text{if } w \\\\geq (c_1+c_2)/2\\\\\\\\\\nc_2 &\\\\text{if } w < (c_1+c_2)/2\\n\\\\end{cases}\\n\\\\end{equation}\\nNote that $w$ is the original weight value while $w_b$ is the binarized weight value. These changes eliminate the need for hyper-parameter tuning.\",\n \"id\": \"f998a362-cba7-42b7-aa49-83d293cf4b58\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"11228467-1126-4614-9475-560634f98a9b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Binarized Networks\"\n ],\n [\n \"subsubsection\",\n \"Deterministic Binarization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deterministic Binarization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 685,\n 4488,\n 8375,\n 2670\n ],\n \"content\": \"Stochastic~ binarization is performed as follows. \\n\\\\begin{equation}\\n w_b= \\\\begin{cases}\\n+1 &\\\\text{with probability } p=\\\\sigma(w)\\\\\\\\\\n-1 &\\\\text{with probability } 1-p\\n\\\\end{cases}\\n\\\\end{equation}\\n\\\\noindent where\\n\\\\begin{eqnarray}\\n \\\\sigma(w)=\\\\text{clip}\\\\left(\\\\frac{w+1}{2},0,1\\\\right)=\\\\max\\\\left(0, \\\\min\\\\left(1, \\\\frac{w+1}{2}\\\\right)\\\\right)\\n\\\\end{eqnarray}\\nWe only binarize the weights during the forward and backward propagations but not during the parameter update. Keeping good precision weights during the updates is necessary for Stochastic Gradient Descent (SGD). This is possible using something called as ``Straight Through Estimator (STE) trick''~. As per STE, as the quantized value is an approximation of the original value, we can substitute the gradient with respect to the quantized value for the gradient of original value. The trick allows the inclusion of quantization into the computation graph of back-propagation and allows QNNs to represent parameters, activations and gradients with low bitwidth numbers. For test-time inference, there are three options using such a quantization method: \\n\\\\begin{itemize}\\n \\\\item Use the resulting binary weights $w_b$ (this makes most sense with the deterministic binarization). \\n \\\\item In the stochastic case, many different networks can be sampled by sampling a $w_b$ for each weight. The ensemble output of these networks can then be obtained by averaging the outputs from individual networks. \\n \\\\item Use original weights. But this does not reduce model size. \\n\\\\end{itemize}\\nBesides this, there have been further efforts that make train/test faster but do not reduce model size. For example, Lin et al.~ convert multiplications in the backward pass into bit-shifts by restricting the activations to be power-of-two integers. Hubara et al.~ binarize weights and activations, at the inference phase and the entire training phase of a deep network.\",\n \"id\": \"3c274eec-1522-4b1c-b25f-1f3d927c3167\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"11228467-1126-4614-9475-560634f98a9b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Binarized Networks\"\n ],\n [\n \"subsubsection\",\n \"Stochastic Binarization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Stochastic Binarization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4489,\n 4490,\n 851\n ],\n \"content\": \"\\\\label{subsubsec:lossAwareBin}\\nThe na\\\\\\\"ive binary quantization methods divide the real number line into two parts and each part was mapped to a quantized weight value. Can we decide per weight value which of the two weights it should be quantized to? Thus the idea behind Binary Weight Networks (BWN)~ is to approximate the weight vector $W\\\\in R^n$ using a binary vector $B\\\\in\\\\{+1,-1\\\\}^n$ and a scaling factor $\\\\alpha\\\\in R^+$ such that $W\\\\approx\\\\alpha B$. This can be expressed as an optimization problem as follows.\\n\\\\begin{eqnarray}\\n \\\\alpha^*, B^*=\\\\argmin_{\\\\alpha, B}||W-\\\\alpha B||^2\\n\\\\end{eqnarray}\\nWe can expand and write this as follows.\\n\\\\begin{eqnarray}\\n ||W-\\\\alpha B||^2=\\\\alpha^2B^TB-2\\\\alpha W^TB+W^TW\\n\\\\end{eqnarray}\\nSince $B\\\\in{+1,-1}^n$, $B^TB=n$. Also $W^TW$ is a constant. Thus $B^*=\\\\argmax_B W^B$ such that $B\\\\in{+1,-1}^n$. This optimization can be solved by simply assigning $B_i=+1$ when $W_i\\\\geq 0$, and $B_i=-1$ otherwise. To compute $\\\\alpha^*$, we set the derivative of $||W-\\\\alpha B||^2$ wrt $\\\\alpha$ to 0 and get the solution as follows.\\n\\\\begin{eqnarray}\\n \\\\alpha^*=\\\\frac{\\\\sum|W_i|}{n}\\n\\\\end{eqnarray}\\nThus, besides the binarized weight matrix, a scaling parameter is also learned in BWN.\\nTo take this idea further, can we learn $\\\\alpha$ and $B$ to minimize the overall network's loss function? Thus, now, the Weight binarization can be formulated as the following optimization problem.\\n\\\\begin{eqnarray}\\n &&\\\\min_{\\\\hat{w}} \\\\text{loss}(\\\\hat{w})\\\\\\\\\\n &&\\\\text{such that } \\\\hat{w}_l=\\\\alpha_l b_l; \\\\alpha_l>0; b_l\\\\in\\\\{+1,-1\\\\}^{n_l}; l=1,..., L\\n\\\\end{eqnarray}\\n\\\\noindent where $L$ is the number of layers, $n_l$ is the number of weights in layer $l$. This loss aware binarization~ problem can be solved using proximal Newton algorithm~ to find the best $\\\\alpha_l$ and $B_l$.\",\n \"id\": \"bb389108-4900-443c-bee1-d687f6265cf4\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"11228467-1126-4614-9475-560634f98a9b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Binarized Networks\"\n ],\n [\n \"subsubsection\",\n \"Loss Aware Binarization (LAB)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Loss Aware Binarization (LAB)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4491\n ],\n \"content\": \"\\\\label{subsec:ternaryQuant}\\nUnfortunately, binary quantization of the recurrent weights in RNNs/LSTMs never worked~. When the true value of a weight is near zero, its quantized value is either set to -1 or 1. This results into an artificial increase in the magnitude of the weights and the vanishing/exploding gradients problem becomes more severe. Hence, another popular form of quantization is ternary quantization (see Fig.~\\\\ref{fig:quantization}(B)). Ternary quantization can help achieve a min of 16x compression (up to 32x compression if hardware allows to avoid storing zeros). In this section, we discuss different variants of ternary quantization from the simplest ternary connect networks to hybrid ternary networks like HitNets.\",\n \"id\": \"60a24ca4-58df-4063-9d44-f538e8a26b98\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"c0c0e7e6-55dc-426b-8669-a8adfb8fe243\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Ternarized Networks\"\n ]\n ],\n \"subsections\": [\n \"8b4d9559-79ce-43e9-8312-ccc718a576a7\",\n \"05681eea-3f45-4e85-b114-d5a86ce4f547\",\n \"3a480858-aec9-47ce-85a6-135a6bfa320a\"\n ],\n \"title\": \"Ternarized Networks\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4492,\n 4488\n ],\n \"content\": \"The simplest method for ternary quantization is ternary connect~ whose deterministic form is as follows.\\n\\\\begin{equation}\\n w_t= \\\\begin{cases}\\n+1 &\\\\text{if } w > 0.5\\\\\\\\\\n0 &\\\\text{if } -0.5 < w \\\\leq 0.5\\\\\\\\\\n-1 &\\\\text{if } w \\\\leq -0.5\\\\\\\\\\n\\\\end{cases}\\n\\\\end{equation}\\nNote that $w$ is the original weight value while $w_t$ is the ternarized weight value. Like binary connect, ternary connect also eliminates all multiplications in the forward pass. In the stochastic form, assuming original weights have been normalized to be in the range [-1,1], ternary quantization is done as follows.\\n\\\\begin{equation}\\n w_t= \\\\begin{cases}\\n+1 &\\\\text{with prob } w \\\\text{ if } w\\\\in (0,1]\\\\\\\\\\n0 &\\\\text{with prob } 1-w \\\\text{ if } w\\\\in (0,1]\\\\\\\\\\n0 &\\\\text{with prob } 1+w \\\\text{ if } w\\\\in[-1,0]\\\\\\\\\\n-1 &\\\\text{with prob } -w \\\\text{ if } w \\\\in [-1,0]\\\\\\\\\\n\\\\end{cases}\\n\\\\end{equation}\\nA slightly related way called as Bernoulli Ternary Quantization where $w_t$ is set to +1 (or -1) with prob $p$ if $w>0$ (or) $<0$, and set to 0 with prob 1-p where $p\\\\sim$Bernoulli($|x|$). Yet another way to set the boundaries for the three ranges is to use Gaussian based ternary weights~ as follows.\\n\\\\begin{equation}\\n w_t= \\\\begin{cases}\\n+1 &\\\\text{if } w > -(\\\\mu+\\\\sigma/2)\\\\\\\\\\n0 &\\\\text{if } -(\\\\mu+\\\\sigma/2)<w\\\\leq (\\\\mu+\\\\sigma/2)\\\\\\\\\\n-1 &\\\\text{if } w \\\\leq -(\\\\mu+\\\\sigma/2)\\\\\\\\\\n\\\\end{cases}\\n\\\\end{equation}\\n\\\\noindent where $\\\\mu$ and $\\\\sigma$ are the mean and standard deviation of the weight matrix being quantized.\",\n \"id\": \"8b4d9559-79ce-43e9-8312-ccc718a576a7\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"60a24ca4-58df-4063-9d44-f538e8a26b98\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Ternarized Networks\"\n ],\n [\n \"subsubsection\",\n \"Ternary Weight Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Ternary Weight Networks\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 852,\n 4493,\n 851\n ],\n \"content\": \"Rather than using the rules for ternary quantization as mentioned above, one can learn the boundary ranges or the quantized values for individual weights. One way of learning the right ternary representation per weight value is to minimize the Euclidean distance between full precision weights $W$ and the ternary weights $T$ along with a scaling factor~. This can be expressed as the following optimization problem.\\n\\\\begin{eqnarray}\\n\\\\alpha^*, T^*=\\\\argmin_{\\\\alpha,T}||W-\\\\alpha T||_2^2\\\\\\\\\\n\\\\text{such that } \\\\alpha\\\\geq 0; T_i\\\\in \\\\{-1,0,1\\\\}; i=1,2,..., n\\n\\\\end{eqnarray}\\nNote that this is equivalent to the BWN method~. This does not lead to a closed form solution. Hence, we approximate the solution with threshold-based ternary function. \\n\\\\begin{equation}\\n w_t= \\\\begin{cases}\\n+1 &\\\\text{if } w > \\\\Delta\\\\\\\\\\n0 &\\\\text{if } -\\\\Delta < w \\\\leq \\\\Delta\\\\\\\\\\n-1 &\\\\text{if } w \\\\leq -\\\\Delta\\\\\\\\\\n\\\\end{cases}\\n\\\\label{eq:eq2}\\n\\\\end{equation}\\nThe approximation works when we set $\\\\Delta$ as follows.\\n\\\\begin{eqnarray}\\n\\\\Delta^*=\\\\argmax_{\\\\Delta>0}\\\\frac{1}{|I_\\\\Delta|}\\\\left(\\\\sum_{i\\\\in I_\\\\Delta}|W_i|\\\\right)^2\\n\\\\end{eqnarray}\\n\\\\noindent where $I_\\\\Delta$ is the number of weights with magnitude$>\\\\Delta$. Again, this has no straightforward solution, unless we assume that original weights $W_i$'s are generated from uniform or normal distribution. When $W_i$'s are uniformly distributed in $[-a, a]$ and $\\\\Delta$ lies in $(0, a]$, the approximated $\\\\Delta^*$ is $a/3$, which equals to $\\\\frac{2}{3}E(W)$. When $W_i$'s are generated from normal distributions $N(0,\\\\sigma^2)$, the approximated $\\\\Delta^*$ is 0.6$\\\\sigma$ which equals to 0.75$E(|W|)$. Thus, we can use the following rule of thumb for fast and easy computation.\\n\\\\begin{eqnarray}\\n\\\\Delta^*\\\\approx 0.7 E(W)=\\\\frac{0.7}{n}\\\\sum_{i=1}^n |W_i|\\n\\\\end{eqnarray}\\nAnother way to learn the quantization step size $\\\\Delta$ in Eq.~\\\\ref{eq:eq2} is to learn in a loss-aware manner~, i.e., tuning it to minimize the overall network loss. Given a multi-layered network, we need to perform such quantization layer by layer in a greedy manner. We first train the network with full precision weights. We quantize all input data and signals of hidden layers. Next, we start with the weight quantizer between the input layer and the first hidden layer, try several step sizes around the initial step size and measure the output error of the network with the training set. The initial step size is determined using Lloyd-Max algorithm~. Choose the step size that minimizes the output error and quantize the weights. Further, we perform these steps for the next few layers until the output layer. Finally, the quantized neural network is retrained.\\nYet another way of training ternary quantization~ is to quantize weights to one of {$-W_l^n$, 0, $W_l^p$} for each layer $l$, where $W_l^n$ and $W_l^p$ are trainable parameters, learned using back-propagation. First, we normalize the full-precision weights to the range [-1, +1] by dividing each weight by the maximum weight. During SGD, we back propagate the gradient to both $W_l^n$ and $W_l^p$ and to the latent full-precision weights. This makes it possible to adjust the ternary assignment (i.e. which of the three values a weight is assigned). To decide the quantization step size $\\\\Delta_l$ for a layer $l$, two heuristics can be used: (1) set $\\\\Delta_l=t\\\\times \\\\max(|w_l|)$ where $t$ is a constant and $w_l$ are the full precision weights in layer $l$. (2) maintain a constant sparsity $r$ for all layers throughout training. By adjusting the hyper-parameter $r$ we can obtain ternary weight networks with various sparsities.\",\n \"id\": \"05681eea-3f45-4e85-b114-d5a86ce4f547\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"60a24ca4-58df-4063-9d44-f538e8a26b98\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Ternarized Networks\"\n ],\n [\n \"subsubsection\",\n \"Trained Ternary Quantization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Trained Ternary Quantization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Given various ternary quantization methods proposed so far, one can combine them and use different methods for different layers. Wang et al.~ found that threshold ternary quantization (TTQ) (Eq.~\\\\ref{eq:eq2}) is preferable for weights in an RNN while Bernoulli Ternary Quantization (BTQ) is preferable for activations. This is based on the observation that in an RNN, the distribution of weights follows normal distribution (with different ranges across different weight matrices), while for activations, the range is [0,1] and most of the values are located near to the two poles instead of the middle of the range. In the training phase (where we need to store the full precision weights), ternary quantization of weights only saves 1.4x memory consumption but quantizing both weights and activations can achieve up to 16x memory savings.\\nThe HitNet architecture~ with this hybrid ternary quantization can be defined using these equations, where $i_t, f_t, o_t$ are the input, forget and output gates; $x_t$ is input at time $t$; $c_t$ is the cell output; and $h_t$ is the hidden layer output; $W_x$, $W_h$, $b_x$, $b_h$ are weights. \\n\\\\begin{eqnarray}\\n \\\\nonumber i_t,f_t,g_t,o_t&=&\\\\sigma(\\\\text{TTQ}(W_x)x_t+\\\\text{TTQ}(b_x)\\\\\\\\\\n \\\\nonumber &+&\\\\text{TTQ}(W_h)h_{t-1}+\\\\text{TTQ}(b_h))\\\\\\\\\\n \\\\nonumber c_t&=&f_t\\\\times c_{t-1}+i_t\\\\times g_t\\\\\\\\\\n h_t&=&\\\\text{BTQ}(o_t\\\\times \\\\sigma(c_t))\\n\\\\end{eqnarray}\",\n \"id\": \"3a480858-aec9-47ce-85a6-135a6bfa320a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"60a24ca4-58df-4063-9d44-f538e8a26b98\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Ternarized Networks\"\n ],\n [\n \"subsubsection\",\n \"Hybrid Ternary Quantization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hybrid Ternary Quantization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:generalQuant}\\nSo far we discussed methods designed specifically for binary and ternary quantization. Now, we discuss general $k$-bit quantization methods. We will discuss (1) uniform quantization methods which perform equal width binning, (2) non-uniform methods which are closer to equal frequency binning, (3) loss-aware quantization methods, and (4) methods specifically designed for Transformer models.\",\n \"id\": \"be929a80-0b3b-4005-bf19-e84fee8ab111\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"c0c0e7e6-55dc-426b-8669-a8adfb8fe243\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"General Quantized Networks\"\n ]\n ],\n \"subsections\": [\n \"f96f7375-a19b-4b18-9a52-99c14e0b21ea\",\n \"6c78c9b7-5b2e-4d63-90b9-bd800235109c\",\n \"4b1349e1-295c-4789-9acd-ec92a1cf9f02\",\n \"96bae440-819d-4a68-8df5-216ce70bd5a8\",\n \"c55e72b6-a137-4167-a584-9f7c420c0df0\"\n ],\n \"title\": \"General Quantized Networks\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4491,\n 8797,\n 4351,\n 851\n ],\n \"content\": \"Uniform $k$-bit Quantization simply splits the range of original weights into $2^k-1$ equal size intervals~. Refer Fig.~\\\\ref{fig:quantization}(C).\\nIf original weights are in range [-1,1], they can be quantized as follows.\\n\\\\begin{equation}\\n q_k(x)=2\\\\left(\\\\frac{\\\\text{round}[(2^k-1)\\\\left(\\\\frac{x+1}{2}\\\\right)]}{2^k-1}-\\\\frac{1}{2}\\\\right)\\n \\\\label{eq:uq}\\n\\\\end{equation}\\nSimilarly, if entries are in range [0,1], we could use the following formula.\\n\\\\begin{eqnarray}\\n q_k(x)=\\\\frac{1}{2^k-1}\\\\lfloor(2^k-1)x+\\\\frac{1}{2}\\\\rfloor\\n\\\\end{eqnarray}\\nWhen the weights in matrix $X$ are not in the range [0,1], we can first scale weights as follows.\\n\\\\begin{eqnarray}\\n \\\\tilde{X}=\\\\frac{X-\\\\beta}{\\\\alpha}\\n\\\\end{eqnarray}\\n\\\\noindent where $\\\\alpha=\\\\max(X)-\\\\min(X)$ and $\\\\beta=\\\\min(X)$. After quantization, we can apply a reverse transform to approximate the original values. Overall, the quantized result can be written as follows.\\n\\\\begin{eqnarray}\\n Q(X)=\\\\alpha q_k(\\\\tilde{X})+\\\\beta\\n\\\\end{eqnarray}\\nGiven any quantization function $q_k(x)$, one can use it for quantizing weight matrices of various recurrent models like RNNs, GRUs and LSTMs~. Typical inference equations for a GRU can be written as follows.\\n\\\\begin{eqnarray}\\n z_t=\\\\sigma(W_z.[h_{t-1},x_t]); r_t=\\\\sigma(W_r.[h_{t-1},x_t])\\\\\\\\\\n \\\\tilde{h_t}=\\\\text{tanh}(W.[r_t\\\\times h_{t-1},x_t]); h_t=(1-z_t)h_{t-1}+z_t \\\\tilde{h_t}\\n\\\\end{eqnarray}\\nBesides the matrix multiplications needed to compute $z_t$, $r_t$ and $\\\\tilde{h_t}$, the gate structure of $\\\\tilde{h_t}$ and $h_t$ brings in the need for element-wise multiplication. \\nAs $\\\\tilde{h_t}$ and $h_t$ are also the inputs to computations at the next timestamp, and noting that a quantized value multiplied by a quantized value will have a larger bit-width, we need to insert additional quantization steps after element-wise multiplications.\\nAnother problem with quantization of GRU structure lies in the different value range of gates. The range of tanh is [-1, 1], which is different from the value range [0, 1] of $z_t$ and $r_t$. Keeping in mind these observations, the equations for a quantized GRU can be written as follows, after the weights $W_z$, $W_r$ and $W$ and input $x_t$ have already been quantized to [-1,1].\\n\\\\begin{eqnarray}\\n z_t&=&\\\\sigma(W_z.[h_{t-1},x_t])\\\\\\\\\\n r_t&=&\\\\sigma(W_r.[h_{t-1},x_t])\\\\\\\\\\n \\\\tilde{h_t}&=&tanh\\\\left(W.\\\\left[2q_k\\\\left(\\\\frac{1}{2}(r_t h_{t-1})+\\\\frac{1}{2}\\\\right)-1,x_t\\\\right]\\\\right)\\\\\\\\\\n h_t&=&2q_k\\\\left(\\\\frac{1}{2}((1-z_t) h_{t-1}+z_t \\\\tilde{h_t})+\\\\frac{1}{2}\\\\right)-1\\n\\\\end{eqnarray}\\nFollowing a similar method, we can also quantize LSTM networks.\",\n \"id\": \"f96f7375-a19b-4b18-9a52-99c14e0b21ea\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"be929a80-0b3b-4005-bf19-e84fee8ab111\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"General Quantized Networks\"\n ],\n [\n \"subsubsection\",\n \"Uniform Quantization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Uniform Quantization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4494,\n 4491\n ],\n \"content\": \"Uniform quantization is easy to implement but far from optimum when quantizing non-uniform data, which is believed to be the trained weights and activations of deep neural network. One way of performing non-uniform quantization is exponential quantization~. It quantizes the weight values to an integer power of 2. If we let \\n\\\\begin{eqnarray}\\n p=\\\\frac{|W|}{2^{\\\\lfloor \\\\log_2|W|\\\\rfloor}}-1\\n\\\\end{eqnarray}\\ndeterministic exponential quantization can be written as follows.\\n\\\\begin{equation}\\n \\\\log_2 W_q= \\\\begin{cases}\\n\\\\lceil \\\\log_2|W| \\\\rceil &\\\\text{if } p>0.5\\\\\\\\\\n\\\\lfloor \\\\log_2|W| \\\\rfloor &\\\\text{otherwise }\\n\\\\end{cases}\\n\\\\end{equation}\\nSimilarly, stochastic exponential quantization can be written as follows.\\n\\\\begin{equation}\\n \\\\log_2 W_q= \\\\begin{cases}\\n\\\\lceil \\\\log_2|W| \\\\rceil &\\\\text{with prob } p\\\\\\\\\\n\\\\lfloor \\\\log_2|W| \\\\rfloor &\\\\text{with prob } 1-p\\n\\\\end{cases}\\n\\\\end{equation}\\nExponential quantization enables storing weights in low precision and eliminating multiplications. However, it still does not perform quantization in a way which is sensitive to the distribution of the weights. Distributions of parameters in neural networks are often imbalanced, such that the uniform quantization determined from extremal values may under utilize available bitwidth. When we quantize values, it may be desirable to make the quantized values have balanced distributions, to take full advantage of the available parameter space. Balanced quantization method~ starts by partitioning numbers into $2^{k}$ bins containing roughly the same number of entries (percentiles). Refer Fig.~\\\\ref{fig:quantization}(D). Each partition is then mapped to an evenly-divided interval in the closed interval [0, 1]. Finally, the quantization step maps intervals into discrete values using Eq.~\\\\ref{eq:uq} and transforms the value range to be approximately the same as input.\\nA na\\\\\\\"ive implementation using percentiles as thresholds would require sorting of weight values during each forward operation in back-propagation, which may slow down the training process. The $2^{k}$ evenly spaced percentiles required in histogram equalization can be computed from the recursive application of partitioning of numbers by medians. Further, the mean $\\\\mu$ can be used to approximate the median $m$. Thus, we can perform approximate histogram equalization without doing sorting.\",\n \"id\": \"6c78c9b7-5b2e-4d63-90b9-bd800235109c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"be929a80-0b3b-4005-bf19-e84fee8ab111\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"General Quantized Networks\"\n ],\n [\n \"subsubsection\",\n \"Balanced Quantization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Balanced Quantization\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4496,\n 4495\n ],\n \"content\": \"Yet another way of performing non-uniform quantization is to decide bin boundaries using clustering in a static manner. In this static-KMeans method~, We first train the neural network with full-precision parameters. Then apply KMeans to the weights. After clustering, the value of each pixel is set to the value of the center of the cluster it belongs to. We also need to store mapping from integers to cluster centers. Given $k$ clusters, we only need $\\\\log (k)$ bits to code the clusters.\\nA better approach is to perform KMeans clustering during training. In this method~, multiple connections (belonging to the same cluster) share the same weight, and we fine-tune those shared weights. For the forward pass, the cluster index stored for each connection is mapped to a centroid which is then used as the weight. For back-propagation, during update, all the gradients are grouped by the cluster index and summed together, multiplied by the learning rate and subtracted from the shared centroids from last iteration. We use KMeans clustering to identify the shared weights for each layer of a trained network, so that all the weights that fall into the same cluster will share the same weight. Weights are not shared across layers. To calculate the compression rate, given $k$ clusters, we only need $\\\\log_2 k$ bits to encode the index. In general, for a network with $n$ connections and each connection is represented with $b$ bits, constraining the connections to have only $k$ shared weights will result in a compression rate of\\n\\\\begin{eqnarray}\\n r=\\\\frac{nb}{n\\\\log_2 k+kb} \\n\\\\end{eqnarray}\\nThere are two other ways of using KMeans for non-uniform quantization: Product Quantization (PQ) and Residual Quantization (RQ)~. In product quantization (PQ), we partition the vector space into many disjoint subspaces, and perform quantization (KMeans) in each subspace. Weight matrix $W$ is partitioned columnwise: $W=[W^1, W^2, ..., W^s]$ where $W^i\\\\in R^{m\\\\times n/s}$ assuming $n$ is divisible by $s$. Then we perform KMeans on each submatrix $W^i$ to obtain clusters $c^i_1, ..., c^i_k$. Thus, we get $s$ codebooks. The reconstructed matrix is $\\\\hat{W}=[\\\\hat{W}^1, \\\\hat{W}^2, ..., \\\\hat{W}^s]$ where $\\\\hat{W}^i_j$ is the closest centroid $c^i_j$. PQ can be applied to either the x-axis or the y-axis of the matrix. We need to store the cluster indexes and codebooks for each subvector. The compression rate for this method can be written as follows.\\n\\\\begin{eqnarray}\\n r=\\\\frac{32mn}{32kn + log_2(k)ms} \\n\\\\end{eqnarray}\\nResidual quantization (RQ) is similar. In RQ, we first quantize the vectors into k-centers. Next we find out the residuals for each data point ($w-c$) and perform KMeans on the residuals. Do it recursively $t$ times.\\nThen the resultant weight vectors are calculated as $\\\\hat{W}_z=c^1_j+c^2_j+...+c^t_j$ given we have recursively performed $t$ iterations. We need to store all the codebooks for each iteration, which potentially needs large amount of memory. The compression rate for this method can be written as follows.\\n\\\\begin{eqnarray}\\n r=\\\\frac{m}{tk + log_2(k)tn} \\n\\\\end{eqnarray}\",\n \"id\": \"4b1349e1-295c-4789-9acd-ec92a1cf9f02\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"be929a80-0b3b-4005-bf19-e84fee8ab111\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"General Quantized Networks\"\n ],\n [\n \"subsubsection\",\n \"KMeans based Quantization Schemes\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"KMeans based Quantization Schemes\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4498,\n 4497,\n 851\n ],\n \"content\": \"Generalizing the loss aware binarization approach (Sec.~\\\\ref{subsubsec:lossAwareBin})~, we can perform $k$-bit quantization~ by attempting to solve the following problem. \\n\\\\begin{eqnarray}\\n \\\\min_{\\\\{\\\\alpha_i,b_i\\\\}_{i=1}^k} \\\\left|\\\\left|w-\\\\sum_{i=1}^k \\\\alpha_i b_i\\\\right|\\\\right|^2 \\n\\\\end{eqnarray}\\nwhere $w\\\\in R^n$ is the original weight vector, $\\\\alpha_i\\\\in R$ and $b_i\\\\in\\\\{-1,+1\\\\}^n$ are variables to be learned. This NP-hard problem can be solved using an iterative greedy approximation which sequentially minimizes the residue. In each iteration, first the residue is computed as \\n\\\\begin{eqnarray}\\nr_{i-1}=w-\\\\sum_{j=1}^{i-1}\\\\alpha_j b_j,\\n\\\\end{eqnarray}\\nand then $\\\\alpha_i$ and $b_i$ are computed as $\\\\alpha_i=\\\\frac{1}{n}||r_{i-1}||_1$ and $b_i=\\\\text{sign}(r_{i-1})$. Further, refined greedy approximation~ extends this to further decrease the quantization error. In the $j^{th}$ iteration after $\\\\alpha_j$ and $b_j$ have been updated, the method adds one extra step to refine all computed $\\\\{\\\\alpha_i\\\\}_{i=1}^j$ with the least squares solution as follows.\\n\\\\begin{eqnarray}\\n[\\\\alpha_1, ..., \\\\alpha_j]=((B_j^TB_j)^{-1}B_j^Tw)^T\\n\\\\end{eqnarray}\\nwhere $B_j=[b_1,...,b_j]$. Typically refined greedy is more accurate than the greedy approach. In refined greedy approximation, after modification on the computed $\\\\alpha$'s, $b$'s are no longer optimal while the method keeps all of them fixed. To improve the refined greedy approximation, alternating minimizing $\\\\alpha$'s and $b$'s becomes a natural choice. Xu et al.~ find that only two alternating cycles is good enough to find high precision quantization. Further, similar to~, for an LSTM, we can combine overall network loss minimization with the multi-bit quantization loss minimization using this bi-level optimization.\\n\\\\begin{eqnarray}\\n\\\\min_{w,\\\\{\\\\alpha_i, b_i\\\\}_{i=1}^k} \\\\text{LSTM}\\\\left(\\\\sum_{i=1}^k\\\\alpha_i b_i\\\\right)\\\\\\\\\\n\\\\text{such that }\\\\{\\\\alpha_i, b_i\\\\}_{i=1}^k=\\\\argmin_{\\\\{\\\\alpha_i',b_i'\\\\}_{i=1}^k}\\\\left|\\\\left|w-\\\\sum_{i=1}^k \\\\alpha_i' b_i'\\\\right|\\\\right|^2\\n\\\\end{eqnarray}\",\n \"id\": \"96bae440-819d-4a68-8df5-216ce70bd5a8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"be929a80-0b3b-4005-bf19-e84fee8ab111\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"General Quantized Networks\"\n ],\n [\n \"subsubsection\",\n \"Loss Aware Quantization (LAQ)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Loss Aware Quantization (LAQ)\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 2488\n ],\n \"content\": \"Each word vector is typically represented as a 300--500 dimensional vector, with each parameter being 32 bits. As there are millions of words, word vectors may take up to 3--6 GB of memory/storage. Can we quantize word vectors? We can clearly quantize them after training. But, we could also quantize when learning word embeddings. For example, Lam et al.~ perform 1-bit and 2-bit quantization while performing word2vec~ training using the Continuous Bag of Words (CBOW) method. They observe that quantization while training leads to better results compared to quantization after training. \\nCheong et al.~ applied BS-Fixed and BS-Flexible binary quantization to Transformer models. They observed that the Transformer architecture is highly resistant to quantization, and is able to match the original model up to a 4-bit representation. Simple iterative pruning is much worse compared to quantization. Lastly, Shen et al.~ propose mixed-precision quantization for BERT based on the observation that different encoder layers should use different number of bits for quantization. Layers that exhibit flatter curvature of the loss gradient surface can be quantized to lower bit precision. Thus, they use different number of bits at different levels of granularity: layers, attention heads and groups of neurons. They observe that quantizing embedding layers with 8 bits and other weight matrices with 2--4 bits leads to results comparable with full-precision BERT.\",\n \"id\": \"c55e72b6-a137-4167-a584-9f7c420c0df0\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"be929a80-0b3b-4005-bf19-e84fee8ab111\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"General Quantized Networks\"\n ],\n [\n \"subsubsection\",\n \"Quantization for Word Embeddings and Transformers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Quantization for Word Embeddings and Transformers\"\n },\n {\n \"cite_extract_rate\": 0.7333333333333331,\n \"cites\": [\n 4498,\n 8797,\n 4494,\n 4497,\n 2670,\n 4490,\n 4351,\n 4491,\n 4492,\n 2488,\n 851\n ],\n \"content\": \"\\\\begin{table}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|l|l|l|l|l|l|l|l|}\\n \\\\hline\\nTask&Dataset&Model&Method&Eval&Bits (weights)&Bits (activation)&Metric\\\\\\\\\\n\\\\hline\\n\\\\hline\\nLanguage modeling&IMDB&GRU&Uniform Q.~&0.882; 0.905&4&4&Acc (H)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Linux Kernel&LSTM&BinaryConnect~&3.532; 1.329&1&FP&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Linux Kernel&LSTM&Loss Aware B.~&1.305/1.409; 1.329&1&FP/1&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Linux Kernel&LSTM&BNN~&3.624; 1.329&1&1&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Linux Kernel&LSTM&BWN~&1.307; 1.329&1&FP&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&GRU&Uniform Q.~&102; 100&4&4&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&GRU&Balanced Q.~&116; 100&4&4&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Greedy~&118.9; 89.8&2&FP&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Refined Loss Aware~&95.6; 89.8&2&3&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Uniform Q.~&152/114; 109&2/4&2/4&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&BNN~&100; 97&4&4&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&HitNet~&110.3; 97.2&2&2&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Alternating LAQ~&103.1/91.4; 89.8&2/4&FP&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Alternating LAQ~&91.9; 89.8&2&3&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&LSTM&Balanced Q.~&126/123; 106&2&2/3&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Text-8&LSTM&Refined Loss Aware~&122.3; 101.1&2&3&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Text-8&LSTM&HitNet~&169.1; 151.4&2&2&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Text-8&LSTM&Alternating LAQ~&105.1; 101.1&2&3&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Text-8&RNN&Exponential Q.~&1.639; 1.588&2&FP&BPC (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&War and Peace&LSTM&BinaryConnect~&2.942; 1.268&1&FP&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&War and Peace&LSTM&Loss Aware B.~&1.291/1.376; 1.268&1&FP/1&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&War and Peace&LSTM&BNN~&3.05; 1.268&1&1&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&War and Peace&LSTM&BWN~&1.313; 1.268&1&FP&CE (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&Wikidata-2&LSTM&HitNet~&126.72; 114.37&2&2&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&WikiText-2&LSTM&Refined Loss Aware~&105.8; 114.37&2&3&PPW (L)\\\\\\\\\\n\\\\hline\\nLanguage modeling&WikiText-2&LSTM&Alternating LAQ~&102.7; 100.1&2&3&PPW (L)\\\\\\\\\\n\\\\hline\\nNamed Entity Recognition&CoNLL-03&BERT-base&QBERT~&91.06; 95&2w8e&8&F1 (H)\\\\\\\\\\n\\\\hline\\nNamed Entity Recognition&CoNLL-03&BERT-base&Mixed-precision Q.~&94.37; 95&2-3w8e&8&F1 (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT17&Transformer&BS-Fixed/BS-Flexible~&11.61/12.11; 28.09&1&FP&BLEU (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT17&Transformer&K-Means 1/4-bit Q.~&12.07/27.65; 28.09&1&FP&BLEU (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT17&Transformer&K-Means 1-bit att-Q.~&24.96; 28.09&1&FP&BLEU (H)\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT17&Transformer&BS-Flexible 1-bit att-Q.~&25.54; 28.09&1&FP&BLEU (H)\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD&BERT-base&QBERT~&79.6; 88.69&2w8e&8&F1 (H)\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD&BERT-base&Mixed-precision Q.~&86.95; 88.69&2-3w8e&8&F1 (H)\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD&Facebook’s DrQA&BS-Fixed~&77.04; 75.28&2&FP&F1 (H)\\\\\\\\\\n\\\\hline\\nSentiment analysis&IMDB&LSTM&Gaussian Q.~&79.64; 82.87&1&FP&Acc (H)\\\\\\\\\\n\\\\hline\\nSentiment analysis&IMDB&LSTM&Gaussian T./B.~&76.86/76.25; 82.87&2&FP&Acc (H)\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST-2&BERT-base&QBERT~&84.63; 93&2w8e&8&Acc (H)\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST-2&BERT-base&Mixed-precision Q.~&92.08; 93&2-3w8e&8&Acc (H)\\\\\\\\\\n\\\\hline\\nSpeech recognition&TIDIGITS&GRU&Pow2 T.~&99.18; 99.1&2&FP&Acc (H)\\\\\\\\\\n\\\\hline\\nSpeech recognition&TIMIT&4-layer MLP&Loss Aware T.~&29.97/28.35; 26.24&1/2&1&FER (L)\\\\\\\\\\n\\\\hline\\nSpeech recognition&WSJ&4-layer BiLSTM&Pow2 T.~&10.49; 11.16&2&FP&WER (L)\\\\\\\\\\n\\\\hline\\nTextual entailment&MNLI&BERT-base&QBERT~&77.02; 84.4&2w8e&8&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&MNLI&BERT-base&Mixed-precision Q.~&82.29; 84.4&2-3w8e&8&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&MNLI-m&BERT-base&QBERT~&76.56; 84&2w8e&8&Acc (H)\\\\\\\\\\n\\\\hline\\nTextual entailment&MNLI-m&BERT-base&Mixed-precision Q.~&81.75; 84&2-3w8e&8&Acc (H)\\\\\\\\\\n\\\\hline\\nWord similarity&M. Turk &Word embeddings&BS-Fixed~&0.602; 0.617&2&FP&CHR (H)\\\\\\\\\\n\\\\hline\\nWord similarity&MEN &Word embeddings&BS-Fixed~&0.764; 0.745&2&FP&CHR (H)\\\\\\\\\\n\\\\hline\\nWord similarity&Rare Words &Word embeddings&BS-Fixed~&0.362; 0.4&2&FP&CHR (H)\\\\\\\\\\n\\\\hline\\nWord similarity&SimLex&Word embeddings&BS-Fixed~&0.387; 0.358&2&FP&CHR (H)\\\\\\\\\\n\\\\hline\\nWord similarity&WordSim Relatedness &Word embeddings&BS-Fixed~&0.594; 0.529&2&FP&CHR (H)\\\\\\\\\\n\\\\hline\\nWord similarity&WordSim Similarity&Word embeddings&BS-Fixed~&0.752; 0.741&2&FP&CHR (H)\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n\\\\caption{Comparison of various quantization methods (sorted by Task and then Dataset). Q.=Quantization, B.=Binarization, T.=Ternarization, PPW=Perplexity per word, BPC=Bits per character, CE=cross-entropy, FER=frame error rate, CHR=correlation with human rankings. FP=full precision (32 bits). For~, we report results with word embedding dimensions set to 1000. In the metric column, H means high is better while L means low is better. For quantization of the BERT-base model~, we report number of bits used for encoders as well as for embeddings. `2-3w8e' means 2 or 3 bits were used for encoder weights while 8 bits were used for embeddings. For NMT results by Cheong et al.~, ``att-Q'' means only attention layers were quantized.}\\n \\\\label{tab:quantizationSummary}\\n\\\\end{table}\\nOtt et al.~ observed that the weight binarization methods do not work with RNNs. Hubara et al.~ were the first to attempt to quantize both weights and activations by trying to evaluate the accuracy of quantized recurrent models trained on the Penn Treebank dataset. Similar to~, Hubara et al.~ found that binarization of weight matrices lead to large accuracy degradation. Later techniques like the one by Xu et al.~ with 2 bits for weights and 3 bits for activations showed better results.\\nTable~\\\\ref{tab:quantizationSummary} compares various quantization methods across different tasks and datasets. Accuracy of both the original and the quantized model are shown. Also, we report number of bits used for weights (which indicate the model size) as well as activations. For the same task, dataset and model combination, different papers report different accuracy of the full precision model because of slight changes in training hyper-parameters; hence we report accuracy of full precision model for each row. \\nFor language modeling, PTB, Text-8, WikiText-2, Linux Kernel, IMDB and ``War and Peace'' are the popular datasets. Across all the datasets, loss aware binarization outperforms other weight binarization schemes. On the Linux Kernel dataset, it is even better than the full-precision network. BinaryConnect does not work well here because of the problem of exploding gradients. On PTB, Xu et al.'s Alternating LAQ~ with 2 bits for weights and 3 bits for activations leads to an LSTM which is just 2.1 points worse in terms of perplexity per word. By 3-bit quantization, Alternating LAQ can achieve $\\\\sim$10.5x memory saving and $\\\\sim$3× real inference acceleration. Uniform and Balanced quantization are rule-based and not aim at minimizing the error. Balanced quantization proposed by Zhou et al.~ performs better than HitNet~ and uniform quantization~. Balanced quantization leads to better results compared to unbalanced counterparts, especially when quantizing to 2-bit weights. However, for 4-bit weights, there is no clear gap between scaling by mean and scaling by max (i.e. balanced and unbalanced quantization).\\nAcross multiple tasks like named entity recognition with CoNLL-03, question answering with SQuAD, sentiment analysis with SST-2, textual entailment using MNLI, Shen et al.~ show that mixed precision quantization (where different number of bits are used for different groups of neurons -- 128 groups in each layer) of BERT is better than QBERT. The reason behind this is the discrepancy (in QBERT) that not all the layers have the same sensitivity to quantization. For more sensitive layers, higher bit precision needs to be set, while for layers that are less sensitive, 2-bits setting is already sufficient. With only additional 5MB memory storage, 2/3-bits mixed-precision Q-BERT is able to retain the performance drop within 2.3\\\\% for MNLI, SQuAD and 1.1\\\\% for SST-2, CoNLL-03, with up to 13x compression ratio in weights. For machine translation, Cheong et al.~ observe that Transformer architecture is highly resistant to quantization (unlike to pruning), and are, in essence, able to match the original model up to a 4-bit representation. Binary Scheme Flexible outperforms 1-bit k-means in both pure 1-bit compression and quantizing only the attention layers, suggesting not tying the weights to particular centroids improves performance, and outperform Binary Scheme Fixed, indicating learning the values to be a superior method. After binarizing only the attention layers, we are still able to recover 90\\\\% of the model's performance. Lam et al.~ experimented with quantization of word embeddings and showed that 2-bit quantized word vectors outperform full precision vectors on word similarity tasks, but do worse on word analogy tasks. Intuitively, they reason that full precision Word2Vec is prone to overfitting with increased epochs of training; quantized training does not seem to suffer as much from this. For sentiment analysis on IMDB, Alom et al.~ show that quantization to 4 bits is better than to 3 or 2 bits, which is expected. They also show that the normal distribution shows better performance against uniform distribution with quantized weights. For speech recognition, Ott et al.~ show that pow-2 ternarization is the best. \\nCheong et al~ were the first to quantize Transformers. They observed that in the last attention layer of the decoder over the encoder hidden states, the attention distribution of the original and 4-bit model are highly similar, indicating 4 bit weights, i.e weights that take on one of 16 values, is enough to get the full effects of attention. Attention distributions in the encoder layers of the Transformer for the original and 4-bit models are almost indistinguishable from one another. This again highlights the idea that self-attention is highly resistant to quantization and could be heavily compressed. Later Shen et al.~ showed that comparable performance to full precision BERT can be achieved with at most 2.3\\\\% performance degradation across many tasks, even with ultra-low precision quantization down to 2 bits, corresponding up to 13x compression of the model parameters, and up to 4x compression of the embedding table as well as activations. \\nOverall, quantization performs model compression by reducing the number of bits per weight value. Binary quantization does not work well by itself for text based neural models. But ternary and higher-bit quantization lead to significant model size reduction without loss in accuracy across tasks. One consideration for quantization is that 3-bit quantized execution is typically not supported in hardware. It is however possible to load 3-bit quantized values and cast them to higher bit precision such as 4 or 8 bits in the execution units. This would still have the benefit of reduced memory volume to/from DRAM. Non-uniform quantization methods like balanced quantization or KMeans based quantization methods are better than uniform quantization methods. Loss aware quantization done while training is better than static loss-unaware quantization. Mixed-precision quantization combined with pruning is highly effective for Transformer based models.\",\n \"id\": \"97e94011-bad6-4d07-98cb-26de46e32fe4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"c0c0e7e6-55dc-426b-8669-a8adfb8fe243\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Quantization\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:knowledgeDistillation}\\nKD methods are the most popular model compression methods for Transformer networks. Also called student-teacher networks, the main idea is to first train a deep teacher network, and then learn a shallow student network that mimics the teacher. After training, the student model is deployed. What information (``dark knowledge'') from the teacher can be used to train the student? What loss functions can be used to ensure right flow of information from teacher to student? Can we have an ensemble of teachers, or teacher assistants or rather fellow students who can train the student? \\nWe discuss these aspects in this section.\",\n \"id\": \"2c5b1b2b-3f74-442d-8ee6-503425cf4207\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Knowledge Distillation (KD)\"\n ]\n ],\n \"subsections\": [\n \"5addca6a-ae25-4e99-8ffc-a608a63231c7\",\n \"d2899d20-743a-43e0-84c7-0fc90c1a975b\",\n \"e1a31918-0804-4d5f-9f78-0ea43f83fc23\",\n \"8fabba12-aca9-4230-b6a7-cc461dfdceaa\",\n \"4cd2b9eb-147c-41ba-996b-a30e824f168f\"\n ],\n \"title\": \"Knowledge Distillation (KD)\"\n },\n {\n \"cite_extract_rate\": 0.9166666666666661,\n \"cites\": [\n 8798,\n 4499,\n 4500,\n 4503,\n 4504,\n 681,\n 8375,\n 8799,\n 8390,\n 4501,\n 4502\n ],\n \"content\": \"Ba and Caruna~ proposed Student Teacher networks (or mimic models) where the student uses the logits before softmax from the teacher network for training (see Fig.~\\\\ref{fig:kd}(A)). The student model is not trained on the original labels; it is trained to learn the function that was learned by the teacher model. Thus, the student model is optimized to minimize the L2 loss between the teacher logits and the student logits across all training instances. Such distilled student models are more accurate than the same shallow student trained directly on the original labeled training data mainly because: (1) Teacher removes noisy labels, if any. (2) The uncertainty from the teacher is more informative to the student than the original 0/1 labels. (3) The original targets may depend in part on features not available as inputs for learning, but the student sees targets that depend only on the input features. The dependence on unavailable features has been eliminated by filtering targets through the teacher.\\nYet another way of utilizing logits is to have the student learn from noisy teacher logits~. After obtaining logits from the teacher, Gaussian noise with mean 0 and standard deviation $\\\\sigma$ is added to teacher’s logits. This perturbation can be applied to samples selected with probability $\\\\alpha$. The perturbed outputs produce the effect of a regularizer. \\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=\\\\textwidth]{kd}\\n \\\\caption{Different Types of Knowledge Distillation methods.}\\n \\\\label{fig:kd}\\n\\\\end{figure}\\nWhile Ba and Caruna~ suggested using only the logits, Hinton et al.~ suggested training the student by minimizing the cross entropy loss between the teacher softmax output and the student softmax output, besides minimizing the cross entropy between student prediction and actual label. The first part is called the soft loss and the second one is called the hard loss. Typically hard loss is given much lower weight compared to the soft loss term. To make the softmax output non-peaked and thereby transfer more useful information from teacher to student, softmax with temperature $>$1 should be used. The same temperature should be used for training both the teacher and the student, but after the student has been trained the temperature can be set to 1 at test time. Besides logits and softmax output, Sobolev training for neural networks is a method for incorporating target derivatives in addition to the target values while training student network (see Fig.~\\\\ref{fig:kd}(C)). Czarnecki et al.~ experiment with first two derivatives of the targets.\\nKD has also been used along with quantization for better model compression~. We start with a trained full-precision large teacher network and an apprentice (student) network that has been initialised with full-precision weights. The apprentice network's precision is lowered and is fine-tuned using KD. \\nWhy just use the output from the last layer of the teacher for training the student? In FitNets~, the student performs hint-based training, i.e., the student is trained using not only the outputs but also the intermediate representations learned by the teacher as hints to improve the training process and final performance of the student (see Fig.~\\\\ref{fig:kd}(B)). we choose a hidden layer of the FitNet, the guided layer, to learn from the teacher's hint layer. Because the student intermediate hidden layer will generally be smaller than the teacher's intermediate hidden layer, additional parameters are introduced to map the student hidden layer to the prediction of the teacher hidden layer. \\nWhile the methods discussed so far use logits, softmax output or their derivatives to transfer knowledge, Yim et al.~ proposed a ``flow of solution procedure (FSP)'' method where the distilled knowledge is transferred in terms of flow between layers, which is calculated by computing the inner product between features from two layers (see Fig.~\\\\ref{fig:kd}(D)). What does this ``flow'' capture intuitively? If we view the input of a deep network as the question and the output as the answer, we can think of the generated features at the middle of the network as the intermediate result in the solution process. There are many ways to solve the problem of generating the output from the input. Hence, mimicking the generated features of the teacher can be a hard constraint for the student. Learning the solution process from teacher is important. More concretely, the student is trained to minimize the L2 difference between the teacher and student FSP matrices computed across various pairs of layers and across multiple training instances. A similar method called Representational distance learning (RDL) has also been proposed in~.\\nLastly, multiple KD variants have been proposed for sequence-level predictions~, e.g., for neural machine translation (NMT). In word-level KD, cross-entropy is minimized between the student/teacher distributions for each word in the actual target sequence, as well as between the student distribution and the degenerate data distribution, which has all of its probability mass on one word. In sequence-level KD (Seq-KD) the student network is trained on the output from beam search of the teacher network that had the highest score. In sequence-level interpolation (Seq-Inter) the student is trained on the output from beam search of the teacher network that had the highest similarity (say using BLEU score) with the target sequence.\",\n \"id\": \"5addca6a-ae25-4e99-8ffc-a608a63231c7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"2c5b1b2b-3f74-442d-8ee6-503425cf4207\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Knowledge Distillation (KD)\"\n ],\n [\n \"subsection\",\n \"Various Distillation Architectures\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Various Distillation Architectures\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4506,\n 4505\n ],\n \"content\": \"Can multiple students learn from each other? Is a powerful teacher really required? In the deep mutual learning (DML) method~, different from the one-way transfer between a static pre-defined teacher and a student in model distillation, with DML, an ensemble of students learn collaboratively and teach each other throughout the training process. Surprisingly, no prior powerful teacher network is necessary -- mutual learning of a collection of simple student networks works, and moreover outperforms distillation from a more powerful yet static teacher. Specifically, each student is trained with two losses: a conventional supervised learning loss, and a mimicry loss that aligns each student's class posterior with the class probabilities of other students. \\nAnil et al.~ propose a similar method but suggest letting the students learn independently just using the conventional supervised learning (hard) loss at least for a few burn in iterations (see Fig.~\\\\ref{fig:kd}(E)). After this, the mutual learning can be done as in DML. They also propose a variant of their Co-Distillation method to perform this training in a distributed scenario where communication efficiency is also important. To update the parameters of one network using co-distillation one only needs the predictions of the other networks, which can be computed locally from copies of the other networks' weights. Empirically, using stale predictions instead of up-to-date predictions for the other neural networks has little to no adverse effect on the quality of the final trained model produced by co-distillation.\",\n \"id\": \"d2899d20-743a-43e0-84c7-0fc90c1a975b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"2c5b1b2b-3f74-442d-8ee6-503425cf4207\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Knowledge Distillation (KD)\"\n ],\n [\n \"subsection\",\n \"Collaborative Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Collaborative Learning\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 4508,\n 4507,\n 681\n ],\n \"content\": \"So far we have talked about a student mimicing a single teacher. However, it is interesting to explore if the student can learn better in presence of multiple teachers or from a teacher assistant.\\nIntuitively and also observed empirically, student network performance degrades when the gap between student and teacher is large. Given a fixed student network, one cannot employ an arbitrarily large teacher, or in other words, a teacher can effectively transfer its knowledge to students up to a certain size, not smaller. To alleviate this shortcoming, Mirzadeh et al.~ introduced multi-step KD, which employs an intermediate-sized network (teacher assistant) to bridge the gap between the student and the teacher (see Fig.~\\\\ref{fig:kd}(F)).\\nThe teacher assistant (TA) models are distilled from the teacher, and the student is then only distilled from the TAs. One could also perform multi-step TA distillation, for example, distillation path could be $10 \\\\rightarrow 6 \\\\rightarrow 4 \\\\rightarrow 2$.\\nA simple way to do KD with multiple teachers is to train student with cross entropy loss between student predictions and average prediction from multiple teachers (see Fig.~\\\\ref{fig:kd}(G)). A more effective method is to augment this with a relative dissimilarity (RD) loss~ defined over intermediate layer outputs generated for a triplet of instances between the student and an ensemble of teachers. For the student, the middle layer is selected. For each teacher, we select the layer such that most teachers are consistent with the resulting order relationships under the voting strategy. We discuss the RD loss given a student and a teacher. Consider a triplet of instances ($x_i$, $x_i^+$, $x_i^-$) such that at an intermediate layer of the teacher network, distance between activations for $x_i^+$ and $x_i$ is smaller than the distance between activations for $x_i^-$ and $x_i$. Let $p_i$ be the intermediate output from student for example $x_i$. Then the RD loss for the triplet ($x_i$, $x_i^+$, $x_i^-$) can be written as follows. \\n\\\\begin{eqnarray}\\n\\\\text{Loss}=\\\\max(0,d(p_i, p_i^+)-d(p_i, p_i^-)+\\\\delta) \\n\\\\end{eqnarray}\\nwhere $d$ is the distance function, and $\\\\delta$ is a small constant to prevent the trivial solution. To extend this loss function definition to multiple teachers, the order between the instances $x_i^+$ and $x_i^-$ given $x_i$ is decided based on majority voting between the teachers. \\nThere are also specific settings when distilling from multiple teachers becomes natural, e.g., when the number of classes is large~ or in multi-lingual settings~. When the number of classes is very large, the teacher model could be an ensemble that contains one generalist model trained on all the data and many ``specialist'' models, each of which is trained on data that is highly enriched in examples from a very confusable subset of the classes (like different types of mushroom). Softmax distribution vector of this type of specialist can be made much smaller by combining all of the classes it does not care about into a single dustbin class. Each specialist model is initialized with the weights of the generalist model. These weights are then slightly modified by training the specialist with half its examples coming from its special subset and half sampled at random from the remainder of the training set. To derive groupings of object categories for the specialists, we focus on categories that the full generalist network often confuses. When training the student, for each instance, we first find the $set $k$ of n$ most probable classes according to the generalist model. Then, we take all the specialist models, $m$, whose special subset of confusable classes has a non-empty intersection with $k$ and call this the active set of specialists $A_k$. Given student's full probability distribution $q$ over all the classes, we minimize the following.\\n\\\\begin{eqnarray}\\n\\\\text{Loss}=KL(p^g,q)+\\\\sum_{m\\\\in A_k} KL(p^m, q) \\n\\\\end{eqnarray}\\nwhere $p^g$ is output distribution from the generalist model, and $p^m$ is the output distribution from the $m^{th}$ specialist model.\\nAn ensemble of teachers is also very useful in a multi-lingual NMT setting~. Individual models for each language pair are first trained and regarded as teachers, and then the multilingual model is trained to fit the training data and match the outputs of individual models simultaneously through KD. When the accuracy of multilingual model surpasses the individual model for the accuracy threshold $\\\\tau$ on a certain language pair, we remove the distillation loss and just train the model with original negative log-likelihood loss for this pair. Lastly, when learning from a teacher ensemble, it is burdensome to load all the teacher models in the GPU memory for distillation. Alternatively, we first generate the output probability distribution of each teacher model for each instance offline, and then just load the top-$K$ probabilities of the distribution into memory and normalize them so that they sum to 1 for distillation. This reduces the memory cost from the scale of $|V|$ (the vocabulary size) to $K$.\",\n \"id\": \"e1a31918-0804-4d5f-9f78-0ea43f83fc23\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"2c5b1b2b-3f74-442d-8ee6-503425cf4207\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Knowledge Distillation (KD)\"\n ],\n [\n \"subsection\",\n \"Multiple Teachers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multiple Teachers\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4513,\n 4509,\n 4511,\n 2481,\n 4512,\n 4514,\n 4510,\n 854\n ],\n \"content\": \"Recently, there has been a lot of work around distilling Transformers to smaller Transformers with less number of layers or to Bidirectional LSTMs. Some of these methods aim at improving the accuracy versus model size tradeoff, while others focus on complex settings like mismatching student-teacher vocabulary~ or mismatch number of attention heads. \\nZhao et al.~ learn a student with small vocabulary compared to the teacher using a dual training method. During distillation, for a given training sequence input to the teacher model, they mix the teacher and student vocabularies by randomly selecting tokens from the sequence to segment using the student vocabulary, with the other tokens segmented using the teacher vocabulary. As part of the masked language model (MLM) task, the model now needs to learn to predict words from the student vocabulary using context words segmented using the teacher vocabulary, and vice versa. The expectation is that the student embeddings can be learned effectively this way from the teacher embeddings as well as teacher model parameters. We perform dual training only for the teacher model inputs. The student model receives words segmented exclusively using the student vocabulary. Also, during MLM, the model uses different softmax layers for the teacher and the student vocabularies depending on which one was used to segment the word in question. Instead of distilling solely on the teacher model's final-layer outputs, layer-wise teacher model parameters can also be leveraged to directly optimize parameters of corresponding layers in the student model.\\nIn Patient KD (PKD)~, the student learns from the teacher's output after every $k$ layers (see Fig.~\\\\ref{fig:kd}(H)) or the output from the last few layers of the teacher (see Fig.~\\\\ref{fig:kd}(I)). The student BERT is initialized using some layers of the pre-trained teacher BERT. TinyBERT~ further extends this idea by using extensive knowledge from embedding layer, and attention and hidden sub-layers of multiple teacher layers, and also the overall teacher output (see Fig.~\\\\ref{fig:kd}(J)). Each student layer is first mapped to a teacher layer before the student training. Liu et al.~ distill a multi-task student from a multi-task teacher, given the soft targets of the training data across multiple tasks. If task $t$ has a teacher, the task-specific loss is the average of two objective functions, one for the correct targets and the other for the soft targets assigned by the teacher. In MiniLM~, the student is trained by deeply mimicking the self-attention behavior of the last Transformer layer of the teacher (see Fig.~\\\\ref{fig:kd}(K)). Besides self-attention distributions, MiniLM introduces the self-attention value-relation transfer to help the student achieve a deeper mimicry. The value-relation is computed as pairwise correlation between different components of the value matrix across various attention heads of the final layer. Pretrained Distillation~ pretrains the student model with a self-supervised masked LM objective on a\\nlarge corpus first, then performs a standard KD on supervised tasks. \\nMost of these models learn one-to-one layer mapping, where each student layer is guided by only one specific teacher layer. Li et al.~ propose a method where each student intermediate layer learns from every teacher intermediate layer with learnable attention weights. Both the embedding-layer distillation and the prediction-layer distillation employ the one-to-one layer mapping as in TinyBERT and BERT-PKD.\\nTang et al.~ propose distillation of a BERT model to a single layer BiLSTM using KL divergence between student and teacher logits (see Fig.~\\\\ref{fig:kd}(L)). Mukherjee et al.~ also distill a multi-lingual BERT (mBERT) model to a BiLSTM. Representation transfer is done from Transformer-based teacher model to BiLSTM-based student model with different embedding dimensions and disparate output spaces. Distillation features include teacher logits and internal teacher representations for one teacher layer. To make all output spaces compatible, a non-linear projection of the parameters in student representation is done to have same shape as teacher representation for each token. The projection parameters are learned by minimizing the KL-divergence (KLD) between the representations of the student and the chosen layer from the teacher. Overall there are multiple loss functions for the student training: supervised hard loss, soft loss wrt output logits, and soft loss wrt internal teacher layer. Rather than optimizing for all loss functions jointly, stage-wise training is performed where each loss function is sequentially used for optimization. \\nLastly, there have been recent efforts~ to distill Transformer models to slightly modified Transformer architectures. MobileBERT~ is a thin version of BERT-large, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks (FFN). To train MobileBERT, we first train a specially designed teacher model, an inverted-bottleneck incorporated BERT-large model (IB-BERT). The IB-BERT uses the inverted-bottleneck structures to adjust its feature map size to 512. Thus, in the bottleneck structure, the inputs to the multi-head attention (MHA) module are from wider feature maps (of inter-block size), while the inputs to the FFN are from narrower bottlenecks (of intra-block size). To fix this issue, MobileBERT uses stacked feed-forward networks to re-balance the relative size between MHA and FFN. Each MobileBERT layer contains one MHA but 4 stacked FFN after each MHA. Then, we conduct knowledge transfer from this teacher to MobileBERT using feature map transfer and attention transfer across all layers. Also, while distilling they perform progressive knowledge transfer, i.e., they progressively train each layer in $L$ stages, where $L$ is the number of layers. When the $l$-th layer is trained, all the trainable parameters in the layers below are frozen. Another difference in the MobileBERT student architecture is usage of high information flow residual connections between the high-channel-count layers. MobileBERT is 4.3x smaller and 5.5x faster than BERT-base while achieving competitive results on well-known benchmarks. Iandola et al.~ propose a new Transformer model architecture called SqueezeBERT which is much like BERT-base, but with the position-wise fully connected layers implemented as convolutions, and grouped convolutions for many of the layers. Distillation for SqueezeBERT is rather straightforward. Distillation is applied only to the final layer, and only during finetuning using soft cross entropy loss with respect to a weighted sum of the teacher's logits and a one-hot encoding of the ground-truth. Teacher model is a BERT-base model finetuned independently on each GLUE task, and these task-specific teacher weights are used for distillation. Xu et al.~ propose a method for progressive module replacement for compressing BERT. Their approach first divides the original BERT into several modules and builds their compact substitutes. Then, the original modules are randomly replaced with their substitutes to train the compact modules to mimic the behavior of\\nthe original modules. We progressively increase the probability of replacement through\\nthe training. In this way, their approach brings\\na deeper level of interaction between the original and compact models.\",\n \"id\": \"8fabba12-aca9-4230-b6a7-cc461dfdceaa\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"2c5b1b2b-3f74-442d-8ee6-503425cf4207\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Knowledge Distillation (KD)\"\n ],\n [\n \"subsection\",\n \"Distilling Transformers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Distilling Transformers\"\n },\n {\n \"cite_extract_rate\": 0.8518518518518511,\n \"cites\": [\n 4479,\n 4515,\n 1150,\n 4505,\n 4514,\n 9120,\n 4513,\n 856,\n 681,\n 854,\n 7333,\n 4486,\n 4500,\n 1568,\n 7,\n 8390,\n 4509,\n 4511,\n 2481,\n 4512,\n 4507,\n 2488,\n 4510\n ],\n \"content\": \"\\\\begin{table}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|l|l|l|l|l|l|p{0.5in}|}\\n \\\\hline\\nTask&Dataset&Teacher/Student models&Method&Metric&Size (Distilled; Orig)&Eval. (Distilled; Orig)\\\\\\\\\\n\\\\hline\\n\\\\hline\\nAbs. summarization&CNN/DailyMail&UniLM-large/12-layer BERT&MiniLM~&ROUGE-L (H)&33M; 340M&39.73; 40.34\\\\\\\\\\n\\\\hline\\nAd CTR prediction&Criteo&No teacher/DNN&CoDistillation~&MAE (L)&-&0.019; 0.022\\\\\\\\\\n\\\\hline\\nCross-lingual NLI&XNLI (15 langs.)&XLM-R-base/12-layer BERT&MiniLM~&Acc (H)&21M; 85M&71.1; 74.5\\\\\\\\\\n\\\\hline\\nCross-lingual QA&MLQA (7 langs.)&XLM-R-base/12-layer BERT&MiniLM~&F1 (H)&21M; 85M&63.2; 64.9\\\\\\\\\\n\\\\hline\\nIntent detection&SNIPS (7-class)&BERT-base/6-layer BERT&Mixed-vocabulary training~&Acc (H)&6.2M; 109M&98.7; 98.6\\\\\\\\\\n\\\\hline\\nLanguage modeling&Common Crawl&No teacher/2-layer LSTM&CoDistillation~&CE (L)&-&3.91; 3.92\\\\\\\\\\n\\\\hline\\nMachine reading comp.&RACE&BERT-base/6-layer BERT&Patient KD~&Acc (H)&67M; 109M&60.34; 65.30\\\\\\\\\\n\\\\hline\\nMachine reading comp.&RACE&BERT-base/6-layer BERT&Vanilla KD~&Acc (H)&67M; 109M&58.74; 65.30\\\\\\\\\\n\\\\hline\\nNER (41 langs.)&Wikiann-41&Multiple mBERT/BiLSTM&XtremeDistill~&F1 (H)&31.8M; 179M*41&88.64; 90.76\\\\\\\\\\n\\\\hline\\nNMT (de$\\\\rightarrow$en)&OpenNMT&4-layer LSTM/2-layer LSTM&Vanilla KD~&BLEU (H)&64.8M; 84.8M&15.48; 15.88\\\\\\\\\\n\\\\hline\\nNMT (de$\\\\rightarrow$en)&OpenNMT&4-layer LSTM/2-layer LSTM&Quantized Distillation (4 bits)~&BLEU (H)&64.8M; 84.8M&15.26; 15.88\\\\\\\\\\n\\\\hline\\nNMT (de$\\\\rightarrow$en)&WMT13&4-layer LSTM/2-layer LSTM&Quantized Distillation (4 bits)~&BLEU (H)&81.6M; 84.8M&35.32; 34.7\\\\\\\\\\n\\\\hline\\nNMT (de$\\\\rightarrow$en)&WMT13&4-layer LSTM/2-layer LSTM&Vanilla KD~&BLEU (H)&81.6M; 84.8M&30.21; 34.7\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$Others)&IWSLT (13 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*12&22.96; 22.72\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$Others)&IWSLT (13 langs.)&Multiple Transformers/Transformer&Seq-KD with Multiple teachers~&BLEU (H)&44M; 44M*12&21.98; 22.72\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$Others)&WMT (7 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*6&24.47; 24.50\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Word-KD~&BLEU (H)&49M; 221M&14.9; 17.7\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Seq-KD~&BLEU (H)&49M; 221M&18.1; 17.7\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Seq-KD + Seq-Inter + Word-KD~&BLEU (H)&49M; 221M&18.5; 17.7\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT14&4-layer LSTM/2-layer LSTM&Pruned Seq-KD + Seq-Inter~&BLEU@5 (H)&8M/17M; 221M&18.5/19.1; 19.5\\\\\\\\\\n\\\\hline\\nNMT (Others$\\\\rightarrow$en)&IWSLT (13 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*12&30.34; 29.7\\\\\\\\\\n\\\\hline\\nNMT (Others$\\\\rightarrow$en)&Ted Talk (45 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*43&28.95; 25.17\\\\\\\\\\n\\\\hline\\nNMT (Others$\\\\rightarrow$en)&WMT (7 langs.)&Multiple Transformers/Transformer&Multiple teachers~&BLEU (H)&44M; 44M*6&28.61; 27.07\\\\\\\\\\n\\\\hline\\nNMT (th$\\\\rightarrow$en)&IWSLT15&4-layer LSTM/2-layer LSTM&Word-KD~&BLEU (H)&8M; 47M&11.8; 14.3\\\\\\\\\\n\\\\hline\\nNMT (th$\\\\rightarrow$en)&IWSLT15&4-layer LSTM/2-layer LSTM&Seq-KD~&BLEU (H)&8M; 47M&12.8; 14.3\\\\\\\\\\n\\\\hline\\nNMT (th$\\\\rightarrow$en)&IWSLT15&4-layer LSTM/2-layer LSTM&Seq-KD + Seq-Inter + Word-KD~&BLEU (H)&8M; 47M&14.2; 14.3\\\\\\\\\\n\\\\hline\\nQuestion generation&SQuAD 1.1&UniLM-large/12-layer BERT&MiniLM~&BLEU@4 (H)&33M; 340M&23.27; 24.32\\\\\\\\\\n\\\\hline\\nSlot filling&SNIPS (39 slots)&BERT-base/6-layer BERT&Mixed-vocabulary training~&F1 (H)&6.2M; 109M&95.0; 97.0\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n\\\\caption{Comparison of various knowledge distillation methods (sorted by Task and then Dataset). CE=cross entropy, MAE=mean absolute error. en=English, th=Thai. MRC=Machine Reading Comprehension. NLI=Natural Language Inference. QA=Question Answering. NER=Named Entity Recognition. In the metric column, H means high is better while L means low is better.}\\n \\\\label{tab:kdSummary1}\\n\\\\end{table}\\nTable~\\\\ref{tab:kdSummary1} compares various knowledge distillation methods across different tasks and datasets. Accuracy of both the original and the distilled model are shown in the Eval column. Also, we report model size for both the student as well as the teacher models. Note that sometimes smaller student models perform better than the teacher models. This could be because (1) for some (task, dataset) pairs, the smaller models are a better regularized fit compared to potentially overfitted teacher models, and (2) student models can be rigorously trained using additional semi-supervision while teacher models depend on limited labeled training data.\\nWhile knowledge distillation has been used for distilling NLP models across many applications, NMT is the most popular one. For abstractive summarization, MiniLM~ leads to a student models which is less than one tenth of the teacher without much loss in ROUGE-L. Similarly, MiniLM shows good results for cross-lingual NLI and multi-lingual QA as well. For Ad click through rate (CTR) prediction and language modeling, Anil et al.~ show that co-distillation leads to lower MAE and cross entropy respectively compared to the individually trained models. Zhao et al.~'s mixed-vocab training leads to 6-layer model that retains over 95\\\\% of the BERT-base model’s slot filling F1 score while being 30x smaller ($<$10 MB without quantization) and 57x faster on a mobile device, yet task-agnostic. For Named Entity Recognition (NER), Mukherjee et al.~ show that XtremeDistil leads to massive compression of teacher models like mBERT by upto 35x in terms of parameters and 51x in terms of latency for batch inference while retaining 95\\\\% of its F1-score for NER over 41 languages. \\nFor NMT, experiments have been done on OpenNMT, WMT, IWSLT and TedTalk datasets. Kim et al.~ make the following observations: (1) Sequence-level knowledge distillation (Seq-KD) does better than word-level knowledge distillation (Word-KD) on English $\\\\rightarrow$ German and performs similarly on Thai $\\\\rightarrow$ English. (2) Combining them (Seq-KD + Word-KD) results in further gains, indicating that these methods provide orthogonal means of transferring knowledge from the teacher to the student: Word-KD is transferring knowledge at the the local (i.e. word) level while Seq-KD is transferring knowledge at the global (i.e. sequence) level. (3) Applying weight pruning on top of knowledge distillation results in a student model that has 13x fewer parameters than the original teacher model, with a decrease of 0.4 BLEU. Tan et al.~ show that one model is enough to handle multiple languages (up to 44 languages), with comparable or even better accuracy than individual models. Their method achieves larger improvements on some languages, such as Da, Et, Fi, Hi and Hy, than others. This is correlated with the data size of the languages. When a language is of smaller data size, it may get more improvement due to the benefit of multilingual training.\\n\\\\begin{table}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|p{0.3in}|l|l|l|l|l|l|l|l|l|l|l|l|l|l|}\\n\\\\hline\\nMethod&Method&Teacher/Student&Size&MRPC&MNLI&MNLI-m&SST-2&QQP&QNLI&RTE&CoLA&STS-B&SQuAD&SQuAD\\\\\\\\ \\nCategory&& models&& F1& Acc& Acc& Acc&F1& Acc& Acc& MCC&Spearman& 1.1 F1& 2.0 F1\\\\\\\\\\n\\\\hline\\n\\\\hline\\n\\\\multirow{3}{0.3in}{Original Models}&BERT-B~&-& 109M&88.9&83.4&84.6&93.5&71.2&90.5&66.4&52.1&85.8&88.4&77.7\\\\\\\\\\n\\\\cline{2-15}\\n&BERT-L~&-& 340M&89.3&85.9&86.7&94.9&72.1&92.7&70.1&60.5&86.5&90.9&81.9\\\\\\\\\\n\\\\cline{2-15}\\n&MTDNN (ensemble)~&-& 340*4M&90.0&87.2&86.7&95.6&72.4&93.9&80.9&61.5&88.3&-&-\\\\\\\\\\n\\\\hline\\n\\\\hline\\n\\\\multirow{19}{*}{\\\\rotatebox{90}{Knowledge Distillation Methods}}&Distilled-BiLSTM~&BERT-L/BiLSTM&0.96M&-&72.6&73.0&90.7&68.2&-&-&-&-&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&Mixed-vocab. training~&BERT-L/BERT-12&10.9M&87.2&80.5&80.7&90.6&-&-&-&-&-&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&TinyBERT~&BERT-B/BERT-4&14.5M&86.4&81.8&82.5&92.6&71.3&87.7&66.6&44.1&80.4&82.1&71.8\\\\\\\\\\n\\\\cline{2-15}\\n&BERT-EMD~&BERT-B/BERT-4&14.5M&87.6&80.6&82.1&91.0&69.3&87.2&66.2&25.6&82.3&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&MobileBERT~&BERT-L/BERT-6&25.3M&88.8&82.6&83.3&92.8&70.2&90.6&66.2&50.5&84.4&90.0&79.2\\\\\\\\\\n\\\\cline{2-15}\\n&MobileBERT+Quantization~&BERT-L/BERT-6&25.3M&87.0&-&83.9&91.9&-&90.8&-&-&-&90.0&-\\\\\\\\\\n\\\\cline{2-15}\\n&MiniLM~&BERT-B/BERT-12&33M&89.5&-&85.7&93.0&91.3&91.5&73.3&58.5&-&-&81.7\\\\\\\\\\n\\\\cline{2-15}\\n&SqueezeBERT~&BERT-B/SqueezeBERT&51.1M&87.8&81.1&82.0&91.4&80.3&90.1&73.2&46.5&86.7&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&DistilBERT~&BERT-B/BERT-4&52.2M&82.4&78.0&78.9&91.4&68.5&85.2&54.1&32.8&76.1&81.2&64.1\\\\\\\\\\n\\\\cline{2-15}\\n&Patient KD~&BERT-B/BERT-4&52.2M&82.6&79.3&79.9&89.4&70.2&85.1&62.3&24.8&79.8&79.5&64.6\\\\\\\\\\n\\\\cline{2-15}\\n&BERT-of-Theseus~&BERT-B/BERT-6&66M&87.6&82.1&82.4&92.2&71.6&89.6&66.2&47.8&84.1&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&MiniLM~&BERT-B/BERT-6&66M&88.4&-&84.0&92.0&91.0&91.0&71.5&49.2&-&-&76.4\\\\\\\\\\n\\\\cline{2-15}\\n&BERT-EMD~&BERT-B/BERT-6&66M&89.8&83.5&84.7&93.3&72.0&90.7&71.7&47.5&86.8&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&Patient KD~&BERT-B/BERT-6&67M&85.0&81.0&81.5&92.0&70.7&89.0&65.5&43.5&81.6&85.3&69.8\\\\\\\\\\n\\\\cline{2-15}\\n&Vanilla KD~&BERT-B/BERT-6&67M&86.2&79.8&80.2&91.5&70.1&88.3&64.7&-&-&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&DistilBERT~&BERT-B/BERT-6&67M&86.9&81.3&82.6&92.5&70.1&88.9&58.4&49.0&81.3&86.2&69.5\\\\\\\\\\n\\\\cline{2-15}\\n&TinyBERT~&BERT-B/BERT-6&67M&87.3&83.2&84.6&93.1&71.6&90.4&70.0&51.1&83.7&87.5&77.7\\\\\\\\\\n\\\\cline{2-15}\\n&Pretrained Distillation~&BERT-B/BERT-6&67M&86.8&82.2&82.8&91.8&70.4&88.9&65.3&-&-&-&-\\\\\\\\\\n\\\\cline{2-15}\\n&MTDNN-KD~&MTDNN/MTDNN-KD&340M&91.1&86.7&87.5&95.6&72.7&96.0&85.1&65.4&89.6&-&-\\\\\\\\\\n\\\\hline\\n\\\\hline\\nParam.&ALBERT-B~ (dev)&-&12M&-&-&81.6&90.3&-&-&-&-&-&89.3&80\\\\\\\\\\n\\\\cline{2-15}\\nSharing&ALBERT-L~ (dev)&-&18M&-&-&83.5&91.7&-&-&-&-&-&90.6&82.3\\\\\\\\\\n\\\\hline\\nTensor Decomp.&FLOP~&-&80M&88.61&-&-&92.09&-&89.05&-&-&88.18&-&-\\\\\\\\\\n\\\\hline\\n\\\\multirow{3}{*}{Pruning}&RPP Iterative Magnitude Pruning~&-&138M&88.1&86.1&85.7&92.4&91.2&92.3&70.1&82.8&-&90.23&75.3\\\\\\\\\\n\\\\cline{2-15}\\n&Iterative Magnitude Pruning~&-&170M&83.5&77&82.5&91.3&85.1&90.2&68.6&76.3&-&85.3&-\\\\\\\\\\n\\\\cline{2-15}\\n&LayerDrop~&-&66M&85.3&-&82.9&92.5&-&89.4&-&-&-&-&-\\\\\\\\\\n\\\\hline\\nQuant.&Mixed-precision quant. (QBERTMP)~&-&-&-&82.29&81.75&92.08&-&-&-&-&-&86.95&-\\\\\\\\\\n\\\\hline\\nSubQuad Trans.&Linformer~&-&-&-&-&-&93.1&90.8&91.2&-&-&-&-&-\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n\\\\caption{Comparison of various methods across various GLUE~ and SQuAD tasks. Please refer~ for detailed description of tasks. Top part (first three rows) shows results for basic Transformer methods (teacher models). Middle part shows results for knowledge distillation methods. Bottom part shows results for a mix of other methods across categories. BERT-L=BERT-large, BERT-B=BERT-base, BERT-$i$=$i$-layer BERT. MCC refers to Matthews correlation. Results for SQuAD are on dev set. Empty entries indicate that the papers do not report those results, or NA entries.}\\n \\\\label{tab:kdSummary2}\\n\\\\end{table}\\nTable~\\\\ref{tab:kdSummary2} compares various knowledge distillation methods (besides other model compression methods) across various GLUE~ and SQuAD tasks. Also, we report model size for both the student as well as the teacher models. Different distillation methods use one of these as the teacher: BERT-large, BERT-base or MTDNN. PKD~ outperforms the Vanilla KD~ on almost all the datasets except for MRPC. TinyBERT-4 significantly outperforms the 4-layer BERT-PKD and DistilBERT by a margin of at least 4.4\\\\%, with ~28\\\\% parameters and 3.1x inference speedup. Compared with the teacher BERT-base, 4-layer TinyBERT is 7.5x smaller and 9.4x faster in the model efficiency, while maintaining competitive performances. The 6-layer TinyBERT achieves comparable results with the teacher. Overall, TinyBERT consistently outperforms both the 4-layer and 6-layer baselines like PKD, DistilBERT and MiniLM.\\nTurc et al.~ show how appropriate pretraining can improve the quality of distillation. MiniLM outperforms DistilBERT and TinyBERT across most tasks. The 6-layer MiniLM is 2.0x faster than original BERTBASE, while retaining more than 99\\\\% performance on a variety of tasks, such as SQuAD 2.0 and MNLI. Distilled MT-DNN significantly outperforms the original MT-DNN on 7 out of 9 GLUE tasks. 4-layer BERT-EMD outperforms 4-layer DistilBERT and BERT-PKD by a substantial margin, even with only 30\\\\% parameters and inference time. Furthermore, it exceeds the TinyBERT model by 2.3\\\\% accuracy on RTE, 2.2\\\\% F1 on MRPC, and 1.9\\\\% Spearman correlation on STS-B. 6-layer BERT-EMD performs better than the 12-layer BERT-base model on 7 out of 9 tasks, with only about 50\\\\% parameters and inference time of the original BERT-base model. Tang et al.~ distill BERT to BiLSTMs. They observe that the distilled BiLSTM model uses 349 times fewer parameters than BERT-large and is 434 times faster. Also, mixed vocab training by Zhao et al.~ produces a small 12-layer model that performs competitively with 6-layer PKD and 4-layer DistilBERT while being $\\\\sim$5-6x smaller.\\nMobileBERT is 4.3x smaller and 5.5x faster than BERT-base. On the SQuAD v1.1/v2.0 question answering task, MobileBERT achieves a dev F1 score of 90.0/79.2 (1.5/2.1 higher than BERT-base). On the natural language inference tasks of GLUE, MobileBERT can achieve a GLUE score of 77.7, which is only 0.6 lower than BERT-base, with a latency of 62 ms on a Pixel 4 phone. While quantization can further compress MobileBERT by 4x, there is nearly no performance degradation from it. SqueezeBERT is approximately half the size of BERT-base. MobileBERT is half the size of SqueezeBERT. SqueezeBERT is 4.3x faster than BERT-base, while MobileBERT is 3.0x faster than BERT-base. On 4 of GLUE tasks SqueezeBERT outperforms the accuracy of MobileBERT, while on the other 4 of the GLUE tasks MobileBERT outperforms SqueezeBERT. MobileBERT and SqueezeBERT outperform BERT-base significantly across all tasks.\\nTo summarize, KD is a popular method for text based model compression. Various methods have proposed information copying using logits, softmax output, attention sub-layer output, value relation, relative dissimilarity information from both the last layer as well as intermediate layers of the teacher. Many methods have been proposed to handle complex teacher-student configuration mismatches in terms of vocabulary, number of attention heads, and hidden layer sizes. Also, KD has been found to be very effective in complex problem settings like multi-lingual tasks and tasks with large number of classes. Learning from noisy teachers, teacher assistants, an ensemble of teachers has been found to be effective as well. KD is the best model compression method especially in settings where a large amount of unlabeled data exists; distillation with data pseudo-labeled by teacher leads to very effective students.\",\n \"id\": \"4cd2b9eb-147c-41ba-996b-a30e824f168f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 27,\n \"parent_id\": \"2c5b1b2b-3f74-442d-8ee6-503425cf4207\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Knowledge Distillation (KD)\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:parameterSharing}\\nRather than removing weights or reducing \\\\#bits to store them, parameter sharing methods reduce model size by finding weight blocks that can share the same weight. Character-based language models learn embeddings for characters and use them to compose word embeddings. In some senses, we can think of various words sharing these character embedding parameters. Further, various parameter sharing methods have been proposed to reduce the large word embedding matrix size. Finally, there are multiple Transformer architectures which benefit from the parameter sharing philosophy. We discuss these methods in this section.\",\n \"id\": \"57903397-1a0f-4365-9fd9-a6bf95e76421\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Parameter sharing\"\n ]\n ],\n \"subsections\": [\n \"1d632f6e-824d-41b6-8395-ec40e6c541ab\",\n \"e8e447ef-9838-46fc-a0ca-4f8617f32377\",\n \"82aba295-2827-4559-b383-3667b9bd0577\",\n \"736e641f-1998-44af-a324-58eb82e0693d\"\n ],\n \"title\": \"Parameter sharing\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4516,\n 4517,\n 4036\n ],\n \"content\": \"Fig.~\\\\ref{fig:charCNN} illustrates various character-aware language model architectures. Ling et al.~ proposed their character to word (C2W) model which constructs vector representations of words by composing characters using BiLSTMs. Relative to traditional word representation models that have independent vectors for each word type, C2W requires only a single vector per character type and a fixed set of parameters for the compositional model. As input, we define an alphabet of characters $C$. For English, this vocabulary would contain an entry for each uppercase and lowercase letter as well as numbers and punctuation. Thus compared to the word embedding matrix, this model is much smaller. Despite the compactness of this model, this ``composed'' word representations yield comparable results across multiple text classification tasks. \\nJozefowicz et al.~ propose two variants for composing word embeddings using character embeddings. In the first CNN-Softmax variant, they use character CNNs (Convolutional Neural Networks) to compose word embeddings from character embeddings both at the input side as well as at the output softmax layer. The character-CNN sub-networks at the input or the output do not share weights. The composed word embeddings are fed to an LSTM to generate the output. In the second Char-LSTM variant, character CNN is used to compose word embeddings on the input side. The composed word embeddings are fed to an LSTM to generate an output which is further fed to a small LSTM that predicts the target word one character at a time. Thus, the word and character-level models are combined, and predictions are made one character at a time, thus allowing to compute probabilities over a much smaller vocabulary. Kim et al.~ propose another variant where at the output side they continue to use word embeddings, but at the input side they compose word embeddings using a highway network on top of a character CNN. The highway network's output is used as the input to a multi-layer LSTM, whose last hidden state output is fed to the output softmax layer.\\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=0.6\\\\columnwidth]{charCNN}\\n \\\\caption{Character-aware Language Models}\\n \\\\label{fig:charCNN}\\n\\\\end{figure}\",\n \"id\": \"1d632f6e-824d-41b6-8395-ec40e6c541ab\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"57903397-1a0f-4365-9fd9-a6bf95e76421\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Parameter sharing\"\n ],\n [\n \"subsection\",\n \"Character-aware Language Models\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Character-aware Language Models\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4522,\n 4520,\n 4036,\n 4518,\n 4521,\n 4519\n ],\n \"content\": \"Given a weight matrix $W$ and a budget $K$, we want to share weights within $W$ to have a max of $K$ unique values. A na\\\\\\\"ive implementation of random weight sharing can be trivially achieved by maintaining a secondary matrix consisting of each connection's group assignment. But this needs memory space itself. Hence, Chen et al.~ propose to use hashing. HashedNets use a low-cost hash function (like xxhash\\\\footnote{\\\\url{https://code.google.com/p/xxhash/\\n}}) to randomly group connection weights into hash buckets, and all connections within the same hash bucket share a single parameter value.\\nUnlike HashedNets where weights are randomly grouped, parameter sharing mechanisms in Toeplitz-like structured matrices~ are highly specific and deterministic. Toeplitz matrices have parameters tied along diagonals. The displacement rank of all Toeplitz matrices is up to 2. Toeplitz-like matrices allow the displacement rank $r$ to be higher. They include products and inverses of Toeplitz matrices, and their linear combinations. The displacement rank $r$ serves as a knob on modeling capacity. High displacement rank matrices are increasingly unstructured. With displacement rank $r$, there are $2nr$ free parameters in the Toeplitz-like structured matrix. Toeplitz transforms can be applied not just to embedding matrix but to all weight matrices in an RNN model. Tay et al.~ use a similar Toeplitz-like structured matrix method with Hamilton Products in Quaternion Algebra to propose Quaternion Transformers which lead to 75\\\\% parameter reduction in the Transformer architecture.\\nAnother method for parameter sharing is to share low-rank factors across layers in a recurrent model. In this method, we first represent a weight matrix $W$ using matrix factorization as $W=W_a W_b$. Thus, hidden layer output for layer $l$ at time $t$ can be written as follows.\\n\\\\begin{eqnarray}\\nh_t^l=\\\\sigma\\\\left[W_a^l W_b^l h_t^{l-1}+U_a^l U_b^l h_{t-1}^l + b^l\\\\right] \\n\\\\end{eqnarray}\\nBut we can share some low-rank factors by setting $W_b^l=U_b^{l-1}$. The combination of matrix factorization and parameter sharing leads to large model compression.\\nAnother way of compressing the embedding matrix is to divide the vocabulary $V$ into frequent and infrequent word sets $B$ and $C$ respectively. Infrequent words' embeddings are represented with frequent words' by sparse linear combinations~. This is inspired by the observation that, in a dictionary, an unfamiliar word is typically defined by common words. A dense embedding is assigned to each common word; an infrequent word, on the other hand, computes its vector representation by a sparse combination of common words' embeddings. This compression is useful for both word embedding matrix as well as output layer of RNNs/LSTMs. Let $U\\\\in R^{E\\\\times |B|}$ be the learned embedding matrix of common words where $E$ is the embedding dimension. For a word $w\\\\in C$, we shall learn a sparse vector $x\\\\in R^{|B|}$ as the sparse code of the word. Once we know $x$, embedding for a word $w\\\\in C$ can be written as follows.\\n\\\\begin{eqnarray}\\n \\\\text{embedding}(w)=\\\\sum_{j=1}^B x_jU_j\\n\\\\end{eqnarray}\\nwhere $U_j$ is the $j^{th}$ column of $U$. To learn the sparse representation of word $w\\\\in C$, the following problem needs to be solved.\\n\\\\begin{eqnarray}\\n\\\\min_x ||Ux-A||_2^2+\\\\alpha ||x||_1+\\\\beta |1^Tx-1|+\\\\gamma 1^T \\\\max(0,-x)\\n\\\\end{eqnarray}\\nwhere $A$ is embedding for the rare word $w$. The last two regularization terms favor a solution that sums to 1 and that is non-negative (for psychological interpretation concerns), respectively. \\nLightRNN~ compresses word embedding matrix from $O(|V|)$ to $O(\\\\sqrt{|V|})$. It uses a 2-Component shared embedding for word representations. We allocate every word in the vocabulary into a word-allocation table, each row of which is associated with a learned vector, and each column associated with another learned vector. Table~\\\\ref{tab:wordAllocation} shows an example of a word allocation table. Depending on its position in the table, a word is jointly represented by two components: a row vector and a column vector. Thus, we only need $2\\\\sqrt{|V|}$ vectors to represent a vocabulary of $|V|$ unique words, which are far less than the $|V|$ vectors. \\nThe input and output use different embedding row/column vectors but they share the same word-allocation table. Word Allocation table creation uses a bootstrap procedure to iteratively refine word allocation based on the learned word embedding. Embeddings (i.e. row and column vectors) are learned using language modeling loss using an RNN on top of the embedding layer.\\n\\\\begin{table}\\n \\\\centering\\n \\\\begin{tabular}{|c|c|c|c|c|}\\n \\\\hline\\nEmbedding&$x_1^c$&$x_2^c$&$x_3^c$&$x_4^c$\\\\\\\\\\n\\\\hline\\n$x_1^r$&january&february&...&...\\\\\\\\\\n\\\\hline\\n$x_2^r$&one&two&...&...\\\\\\\\\\n\\\\hline\\n$x_3^r$&...&...&...&...\\\\\\\\\\n\\\\hline\\n$x_4^r$&...&...&...&...\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n \\\\caption{An Example of a Word Allocation Table}\\n \\\\label{tab:wordAllocation}\\n\\\\end{table}\\nFinally, Suzuki et al.~ propose a Skipgram~ training method with parameter sharing as follows. Split every embedding vector of size $D$ into $B$ equal sub-vectors of size $C$. Thus $D=B\\\\times C$. We assign a limited number of reference vectors to each block of block-splitting vectors. E.g., the number of reference vectors becomes $K\\\\times B$ if we assign $K$ reference vectors to each block. Each reference vector is of size $C$. Skipgram training optimization remains the same except for these extra parameter sharing constraints (applied to both the input and output embedding vectors). Liu et al.~ propose a very similar method, Slim Embeddings, where the embeddings are learned using a RNN language model rather than the Skipgram method. Slim Embeddings is very related to HashedNets~, LightRNN~ and\\nCharacter Aware Language model~. In HashedNets, all elements in a parameter matrix are mapped into a vector through a hash function. However in Slim Embeddings approach, we randomly share subvectors instead of single elements. Slim Embeddings differs from LightRNN in\\nthat it is able to control the compression ratio\\nto any arbitrary value, while LightRNN can only compress\\nat the rate of square or cube root of vocabulary size, which\\ncould be too harsh in practical applications. In character aware language model, if we treat the sequence of subvector ids (virtual characters) as each word's representation,\\nthe word embedding then can be treated as concatenated unigram character feature vectors. The drawback of such an approach is that the model is\\nmore complicated and to speed up inference, it needs to pre-compute the word embeddings for the words, so it couldn't\\nstay in its compact form during inference. The Slim embeddings model is much simpler, and easier to tune. And during inference, it uses much less space and can even decrease the complexity of inference.\",\n \"id\": \"e8e447ef-9838-46fc-a0ca-4f8617f32377\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"57903397-1a0f-4365-9fd9-a6bf95e76421\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Parameter sharing\"\n ],\n [\n \"subsection\",\n \"Parameter Sharing in the Embedding Matrix\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Parameter Sharing in the Embedding Matrix\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 1150,\n 4523\n ],\n \"content\": \"A standard Transformer does not share parameters across layers and also has a fixed number of encoder layers. ALBERT~ incorporates two parameter reduction techniques: \\n\\\\begin{itemize}\\n \\\\item Factorized embedding parameterization. That is, it decomposes large vocabulary embedding matrix into two small matrices. Thus, it reduces the embedding parameters from $O(V \\\\times H)$ to $O(V \\\\times E + E \\\\times H)$ where $H >> E$.\\n \\\\item Cross-layer parameter sharing: There are multiple ways to share parameters, e.g., only sharing feed-forward network (FFN) parameters across layers, or only sharing attention parameters. The default decision for ALBERT is to share all parameters across layers.\\n\\\\end{itemize}\\nAn ALBERT configuration similar to BERT-large has 18x fewer parameters and can be trained about 1.7x faster. Dehghani et al.~ propose Universal Transformers where the number of encoder layers are not pre-decided, and all the encoder layers share the parameters. Certain symbols (e.g. some words or phonemes) are usually more ambiguous than others. It is therefore reasonable to allocate more processing resources to these more ambiguous symbols. Thus, ambiguous symbols undergo more self-attention transformations compared to non-ambiguous ones. Thus, they provide a dynamic per-position halting mechanism for dynamically modulating the number of computational steps needed to process each input symbol (called the ``ponder time'') before the representation is passed on as input to the decoder. The idea of sharing weights across layers in Transformers has also been explored in~.\",\n \"id\": \"82aba295-2827-4559-b383-3667b9bd0577\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"57903397-1a0f-4365-9fd9-a6bf95e76421\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Parameter sharing\"\n ],\n [\n \"subsection\",\n \"Parameter Sharing in Transformers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Parameter Sharing in Transformers\"\n },\n {\n \"cite_extract_rate\": 0.7000000000000001,\n \"cites\": [\n 4522,\n 4516,\n 4036,\n 1150,\n 4517,\n 4521,\n 4519\n ],\n \"content\": \"\\\\begin{table}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|l|p{0.8in}|l|l|l|l|l|l|}\\n \\\\hline\\nTask&Dataset&Method&Base Model&Metric&Size (Comp; Orig)&Eval. (Comp; Orig)\\n\\\\\\\\\\n\\\\hline\\n\\\\hline\\nLanguage modeling&1B Word Benchmark&Char-CNN (input embeddings)~&2-layer word-BiLSTM&Perplexity (L)&1.04B; 1.8B&30.0; 30.6\\\\\\\\\\n\\\\hline\\nLanguage modeling&1B Word Benchmark&Char-CNN (in/output embeddings)~&2-layer word-BiLSTM&Perplexity (L)&0.39B; 1.8B&35.8; 30.6\\\\\\\\\\n\\\\hline\\nLanguage modeling&1B Word Benchmark&LightRNN~&word-LSTM&Perplexity (L)&41M; 1.6G&66; 85\\\\\\\\\\n\\\\hline\\nLanguage modeling&1B Word Benchmark&Slim Embeddings~&2-layer word-BiLSTM&Perplexity (L)&0.25B; 1.8B&38.3; 30.6\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Czech&CNN+Highway network~&word-LSTM&Perplexity (L)&64M; 83M&578; 701\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Czech&LightRNN~&word-LSTM&Perplexity (L)&18M; 83M&558; 701\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Czech&Slim Embeddings~&word-LSTM&Perplexity (L)&17M; 83M&528; 701\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-English&CNN+Highway network~&word-LSTM&Perplexity (L)&20M; 25M&216; 236\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-English&LightRNN~&word-LSTM&Perplexity (L)&17M; 25M&191; 236\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-English&Slim Embeddings~&word-LSTM&Perplexity (L)&7M; 25M&187; 236\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-French &CNN+Highway network~&word-LSTM&Perplexity (L)&44M; 56M&190; 202\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-French &LightRNN~&word-LSTM&Perplexity (L)&17M; 56M&176; 202\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-French &Slim Embeddings~&word-LSTM&Perplexity (L)&12M; 56M&162; 202\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-German &CNN+Highway network~&word-LSTM&Perplexity (L)&104M; 137M&305; 347\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-German &LightRNN~&word-LSTM&Perplexity (L)&18M; 137M&281; 347\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-German &Slim Embeddings~&word-LSTM&Perplexity (L)&17M; 137M&261; 347\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Russian&CNN+Highway network~&word-LSTM&Perplexity (L)&152M; 200M&313; 353\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Russian&LightRNN~&word-LSTM&Perplexity (L)&19M; 200M&288; 353\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Russian&Slim Embeddings~&word-LSTM&Perplexity (L)&19M; 200M&274; 353\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Spanish&CNN+Highway network~&word-LSTM&Perplexity (L)&48M; 61M&169; 186\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Spanish&LightRNN~&word-LSTM&Perplexity (L)&18M; 61M&157; 186\\\\\\\\\\n\\\\hline\\nLanguage modeling&ACLW-Spanish&Slim Embeddings~&word-LSTM&Perplexity (L)&8M; 61M&149; 186\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&CNN+Highway network~&word-LSTM&Perplexity (L)&5M; 20M&92.3; 85.4\\\\\\\\\\n\\\\hline\\nLanguage modeling&Wikipedia articles (ca)&C2W~&word emb.&Perplexity (L)&182K; 4.3M&34.92; 35.34\\\\\\\\\\n\\\\hline\\nLanguage modeling&Wikipedia articles (de)&C2W~&word emb.&Perplexity (L)&183K; 6.3M&41.94; 43.02\\\\\\\\\\n\\\\hline\\nLanguage modeling&Wikipedia articles (en)&C2W~&word emb.&Perplexity (L)&180K; 4.3M&57.39; 59.38\\\\\\\\\\n\\\\hline\\nLanguage modeling&Wikipedia articles (pt)&C2W~&word emb.&Perplexity (L)&178K; 4.2M&40.92; 46.17\\\\\\\\\\n\\\\hline\\nLanguage modeling&Wikipedia articles (tr)&C2W~&word emb.&Perplexity (L)&174K; 5.7M&32.88; 44.01\\\\\\\\\\n\\\\hline\\nMachine reading comp.&ReAding Comprehension from Examinations&ALBERT~&BERT-B&Acc (H)&18M; 108M&68.5; 68.2\\\\\\\\\\n\\\\hline\\nNLI&MNLI&Quaternion Attention~&2-layer att-GloVe&Acc (H)&200K; 700K&72.3; 73.6\\\\\\\\\\n\\\\hline\\nNLI&SciTail&Quaternion Attention~&2-layer att-GloVe&Acc (H)&200K; 700K&79.6; 79.0\\\\\\\\\\n\\\\hline\\nNLI&SNLI&Quaternion Attention~&2-layer att-GloVe&Acc (H)&200K; 700K&85.4; 86.2\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$et)&IWSLT15&Quaternion Transformer~&Transformer&BLEU (H)&11M; 44M&13.1; 14.1\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$ro)&IWSLT15&Quaternion Transformer~&Transformer&BLEU (H)&11M; 44M&18.5; 22.8\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$vi)&IWSLT15&Quaternion Transformer~&Transformer&BLEU (H)&11M; 44M&28.0; 28.4\\\\\\\\\\n\\\\hline\\nPOS Tagging&PTB (ca)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&98.92; 98.09\\\\\\\\\\n\\\\hline\\nPOS Tagging&PTB (de)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&98.08; 97.51\\\\\\\\\\n\\\\hline\\nPOS Tagging&PTB (en)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&97.36; 96.97\\\\\\\\\\n\\\\hline\\nPOS Tagging&PTB (pt)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&97.47; 95.67\\\\\\\\\\n\\\\hline\\nPOS Tagging&PTB (tr)&C2W~&word-BiLSTM&Acc (H)&150K; 2M&91.59; 83.43\\\\\\\\\\n\\\\hline\\nQuestion answering&BABI&Universal Transformer~&Transformer&Avg error (L)&7.3M; 44M&0.21; 15.2\\\\\\\\\\n\\\\hline\\nQuestion answering&WikiQA&Quaternion Attention~&2-layer att-GloVe&MAP (H)&200K; 700K&66.2; 67.2\\\\\\\\\\n\\\\hline\\nSentiment analysis&IMDB&Quaternion Transformer~&2-layer Transformer&Acc (H)&100K; 400K&83.9; 82.6\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST&Quaternion Transformer~&2-layer Transformer&Acc (H)&100K; 400K&80.5; 78.9\\\\\\\\\\n\\\\hline\\nSpeech recognition&2000 hour En. Speech&Toeplitz-like~&3-layer RNN&WER (L)&790K; 1.85M&48.4; 43.5\\\\\\\\\\n\\\\hline\\nSpeech recognition&2000 hour En. Speech&Toeplitz-like~&5-layer LSTM&WER (L)&2.00M; 9.12M&33.5; 33.1\\\\\\\\\\n\\\\hline\\nSubject verb agreement&SVA dataset&Quaternion Transformer~&2-layer Transformer&Acc (H)&100K; 400K&94.7; 94.8\\\\\\\\\\n\\\\hline\\nSubject verb agreement&SVA dataset&Universal Transformer~&Transformer&Acc (H)&7.3M; 44M&0.992; 0.962\\\\\\\\\\n\\\\hline\\nWord analogy &GSEm, GSYN, MSYN&Shared ref. vec. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&64.6/64.5; 64.7\\\\\\\\\\n\\\\hline\\nWord analogy &GSEm, GSYN, MSYN&k-means quant. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&64.0/64.5; 64.7\\\\\\\\\\n\\\\hline\\nWord similarity&7 datasets&Shared ref. vec. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&65.5/67.1; 67.2\\\\\\\\\\n\\\\hline\\nWord similarity&7 datasets&k-means quant. ($K$=16/256, $B$=256)~&SGNS&Acc (H)&98/196MB; 1565MB&64.4/67.0; 67.2\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n\\\\caption{Comparison of various parameter sharing methods (sorted by Task and then Dataset). 7 datasets for word similarity are MEN, MTurk, RARE, SLex, SCWS, WSR, WSS. SGNS=SkipGram with negative sampling. In the metric column, H means high is better while L means low is better.}\\n \\\\label{tab:paramSharingSummary}\\n\\\\end{table}\\nTable~\\\\ref{tab:paramSharingSummary} compares various parameter sharing methods across different tasks and datasets. Accuracy of both the original and the compressed (comp.) model are shown. Also, we report model size (in terms of number of parameters or memory size) for both the original as well as the compressed models. For the same task, dataset and model combination, different papers report different accuracy of the original model because of slight changes in training hyper-parameters; hence we report accuracy of the original model for each row. Since many parameter sharing methods have been used to compress word embeddings, the most common application is language modeling. \\nFor language modeling, experiments have been done on the One Billion Word Benchmark, ACLW, PTB and Wikipedia articles datasets across multiple languages using parameter sharing methods like Char-CNN~, LightRNN~, Slim embeddings~, CNN+Highway networks~ and C2W~. C2W~ always outperforms word lookup tables and improvements are especially pronounced in Turkish, which is a highly morphological language, where word meanings differ radically depending on the suffixes used. Jozefowicz et al.~ observe that Char-CNNs especially with character embeddings being used both at input as well as output can lead to 4.6x model size reduction with a slight increase in perplexity. On the One-Billion-Word benchmark, LightRNN~ achieves much lower perplexity compared to word-LSTMs, whilst reducing the model size by a factor of 40-100, and speeding up the training process by a factor of 2. LightRNN~ significantly reduces the model size, while at the same time outperforms CNN+Highway network~ method. While the model sizes of the CNN+Highway network method increase linearly with respect to the vocabulary size, the model size of LightRNN almost keeps constant on the ACLW datasets. Slim embeddings~ is way better than LightRNN (low perplexity with smaller models). On PTB and 44M giga word corpus datasets, Slim embeddings applied at input layer can maintain same perplexity for a word-LSTM using just 1\\\\% (and 0.2\\\\%) of trainable parameters respectively. \\nk-means quantization and shared reference vectors are also methods for compression of word embeddings using parameter sharing. Suzuki et al.~ showed significant gains over typical skipgram (SGNS) embeddings in terms of model size reduction while retaining similar accuracies for word analogy and similarity tasks. The model size of their method with shared reference vectors with `K = 16, B = 64' was just 24MB, approximately 65 times smaller than that of original SGNS method. They also showed that SGNS with shared reference vectors was better than SGNS with block-wise k-means post-processing method. Unfortunately, there exists no good comparison between the slim embeddings and the shared reference vectors methods.\\nFor the MRC task, ALBERT~ pushed the accuracy from 68.5 to 68.2 using 6x fewer parameters compared to BERT-base. On the NLI task, a tiny Quaternion-Att model (50 dimensions) achieves comparable (or occasionally marginally better or worse) performance compared to typical attention over GloVe (200 dimensions), gaining a 68\\\\% parameter savings across three datasets. For sentiment analysis, Quaternion Transformers~ leads by +1.3\\\\%/1.6\\\\% gains on IMDb and SST datasets respectively while maintaining a 75\\\\% reduction in parameter cost. Quaternion Transformers~ have been shown to outperform the vanilla Transformer for the mathematical language understanding task as well, with 25\\\\% parameters. At the same compression rate, Quaternion Transformers lose only very minor BLEU on IWSLT 2015 NMT datasets.\\nFor POS tagging, we can observe that the model using word lookup tables performs consistently worse than the C2W model~. Universal Transformers~ (while being 1/6th the size) outperform standard Transformers on a wide range of algorithmic and language understanding tasks, including the challenging LAMBADA language modeling task. On speech recognition, Lu et al.~ study mechanisms for learning compact RNNs and LSTMs via low-rank factorizations and parameter sharing schemes. A hybrid strategy of using structured matrices in the bottom layers and shared low-rank factors on the top layers is found to be particularly effective, reducing the parameters of a standard LSTM by 75\\\\%, at a small cost of 0.3\\\\% increase in WER, on a 2000-hr English Voice Search task. For LSTM parameter reduction, architecting upper layers with projection nodes to moderate rank, and bottom layers with Toeplitz-like transforms was found to be a particularly effective strategy. \\nOverall, besides model compression, parameter sharing methods also act as a good regularizer. Parameter sharing in Transformers has been very successful. ALBERT was at the top of the GLUE leaderboard when it was proposed. Parameter sharing methods have also been widely used for compressing embedding matrix. Slim embeddings has the best method for compressing embedding matrices.\",\n \"id\": \"736e641f-1998-44af-a324-58eb82e0693d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"57903397-1a0f-4365-9fd9-a6bf95e76421\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Parameter sharing\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:matrixDecomposition}\\nSparse Matrix decomposition has been traditionally used for applications like feature selection, collaborative filtering, topic mining from text, etc. In this section, we discuss how various popular tensor decomposition methods like Singular Value Decomposition (SVD), Tensor-Train~, CP (CANDECOMP/PARAFAC)~ and Tucker~ can be used for model compression.\",\n \"id\": \"a059247f-141a-43b7-a6a9-f3440e7622cd\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Tensor decomposition\"\n ]\n ],\n \"subsections\": [\n \"da1505f4-583c-4b3b-a84f-338a8c53f73d\",\n \"1b3ed6c2-9088-4bfb-8866-160b3884224a\",\n \"4c1e5567-aabc-47e9-8344-624bb81f275f\",\n \"2c213350-b4bf-49d2-a046-a4fe16413654\"\n ],\n \"title\": \"Tensor decomposition\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 4515,\n 4524,\n 4525\n ],\n \"content\": \"In this part, we will discuss methods where a matrix is factorized into two low-rank factors. Specifically, we replace a weight matrix $W$ with $W_1\\\\times W_2$ such that the total number of parameters are significantly lesser.\\nA multi-layer RNN can be represented as follows.\\n\\\\begin{eqnarray}\\n h_t^l=\\\\sigma(W_x^{l-1}h_t^{l-1}+W_h^l h_{t-1}^l +b^l)\\\\\\\\\\n h_t^{l+1}=\\\\sigma(W_x^l h_t^l + W_h^{l+1}h_{t-1}^{l+1}+b^{l+1})\\n\\\\end{eqnarray}\\nThus, there are two important weight matrices: the recurrent $W_h^l$ and inter-layer matrices $W_l^x$. Prabhavalkar et al.~ propose a method to jointly compress the recurrent and inter-layer matrices corresponding to a specific layer $l$ by determining a suitable recurrent projection matrix, denoted by $P^l\\\\in R^{r_l\\\\times N_l}$ of rank $r^l<N^l$ such that $W_h^l=Z_h^l P^l$ and $W_l^x=Z_x^l P^l$. First, $P^l$ is determined by computing a truncated SVD of the recurrent weight matrix, which we then truncate, retaining only the top $r^l$ singular values. Thus, $W_h^l$ can be written as follows.\\n\\\\begin{eqnarray}\\n W_h^l=(U_h^l \\\\Sigma_h^l)(V_h^l)^T=Z_h^lP^l\\n\\\\end{eqnarray}\\nThus, $P^l$ is set to $(V_h^l)^T$. Further, we determine $Z_x^l$ as the solution to the following least-squares problem.\\n\\\\begin{eqnarray}\\n Z_x^l=\\\\argmin_Y ||YP^l-W_x^l||_2^2\\n\\\\end{eqnarray}\\nThis solution can also be easily extended to LSTMs. Sak et al.~ also proposed a similar solution based on a combination of parameter sharing and matrix decomposition but without SVD initialization. However, typically SVD initialization has been found to perform better.\\nBesides SVD, another way of matrix decomposition is sparse coding. Faruqui et al.~ propose using sparse coding to decompose word embedding matrices. Thus, given vocabulary of size $V$, word embedding matrix $X\\\\in R^{L\\\\times V}$, sparse coding aims at representing each input vector $x_i$ as a sparse linear combination of basis vectors $a_i$ by solving the following problem. \\n\\\\begin{eqnarray}\\n \\\\argmin_{D,A} \\\\sum_{i=1}^V ||x_i-Da_i||_2^2+\\\\lambda||a_i||_1+\\\\tau ||D||_2^2\\n\\\\end{eqnarray}\\nwhere $D\\\\in R^{L\\\\times K}$ and $A\\\\in R^{K\\\\times V}$. Further, for interpretability, one can enforce all elements of $A$ and $D$ to be non-negative. For further compression, one can also enforce $A$ to be binary or ensure that each column of $A$ is a $K$ sized one hot vector~. \\nLastly, Wang et al.~ combine pruning with matrix factorization for model compression and propose the FLOP (Factorized Low-rank Pruning) method. Let $W$ be a weight matrix. Structured pruning (removing a neuron, i.e., removing a column from weight matrix) can be achieved by replacing the computation $Wx$ by $WGx$ where diagonal sparsity-inducing matrix $G$ is learned using $L_0$ regularization over $WG$ along with the supervised loss. This effectively removes a subset of columns of $W$ for column indices $k$ with $z_k=0$. One limitation is that this structured pruning method tends to produce lower performance than its unstructured counterpart. Hence, in the FL0P (Factorized L0 Pruning) model, we first factorize $W=PQ$. Let $r$ be \\\\#columns of $P$ (or equivalently \\\\#rows of $Q$), $p_k$ and $q_k$ be the $k$-th column of $P$ and $k$-th row of $Q$ respectively. We achieve structured pruning by introducing a pruning variable $z_k$ for each component. Thus, now, we can write $W$ as follows.\\n\\\\begin{eqnarray}\\n W=PGQ=\\\\sum_{k=1}^r z_k\\\\times (p_kq_k)\\n\\\\end{eqnarray}\\nwhere $G$ is again the diagonal matrix of pruning variables. After training, only columns and rows corresponding to non-zero diagonal values need to be stored, resulting in much smaller (but still dense) matrices $P$ and $Q$. The nonzero values of $G$ can be absorbed into either $P$ or $Q$. This structured pruning with factorization is much more effective compared to the vanilla structured pruning.\",\n \"id\": \"da1505f4-583c-4b3b-a84f-338a8c53f73d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"a059247f-141a-43b7-a6a9-f3440e7622cd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Tensor decomposition\"\n ],\n [\n \"subsection\",\n \"Two Low-Rank Factors\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Two Low-Rank Factors\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4526,\n 8800\n ],\n \"content\": \"The last layer of a language model is very large of the size $HV$ where $H$ is the size of the hidden layer and $V$ is vocabulary size. Each word by an output embedding of the same size $H$. Chen et al.~ propose a differentiated softmax method which varies the dimension of the output embeddings across words depending on how much model capacity is deemed suitable for a given word. In particular, it is meaningful to assign more parameters to frequent words than to rare words. By definition, frequent words occur more often in the training data than rare words and therefore allow to fit more parameters. They define partitions of the output vocabulary based on word frequency and the words in each partition share the same embedding size. Partitioning results in a sparse final weight matrix which arranges the embeddings of the output words in blocks, each one corresponding to a separate partition. The size of the final hidden layer $H$ is the sum of the embedding sizes of the partitions. While this method does not involve creation of multiple factors, it factorizes the original matrix into multiple blocks while setting the remaining part of the matrix to 0.\\nVariani et al.~ propose a method called Word Encoded Sequence Transducers (WEST) which factorizes a matrix $E=C\\\\times D$ where $D$ is constrained to be a block diagonal matrix. The block diagonal nature of the second factor leads to large compression rates.\",\n \"id\": \"1b3ed6c2-9088-4bfb-8866-160b3884224a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"a059247f-141a-43b7-a6a9-f3440e7622cd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Tensor decomposition\"\n ],\n [\n \"subsection\",\n \"Factorizing into Block Diagonal Matrices\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Factorizing into Block Diagonal Matrices\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 4529,\n 4527,\n 4530,\n 4528\n ],\n \"content\": \"Tensor train decomposition (TTD)~ is a standard tensor decomposition technique which decomposes a high dimensional tensor into multiple 2D and 3D tensors which can be multiplied together to reconstruct the original tensor. These factors are called TT-cores and the other dimensions are referred to as TT-ranks. TTD can be leveraged to compress various weight matrices in RNNs and LSTMs~. The first step is to represent a matrix as a multi-dimensional tensor by simple reshaping transformation and then use TTD on it. The values of TT–ranks directly define the compression ratio, so choosing them to be too small or too large will result into either significant performance drop or little reduction of the number of parameters. Typically TT-ranks around 16 for small matrices and 64-192 for larger matrices result in a good trade-off between compression ratio and the accuracy metric of interest. Also, when we use TTD for weight matrices, we also need change the inputs appropriately to be compatible in terms of dimensions. \\nCompared with TT-RNN, Block-Term RNN (BTRNN)~ is not only more concise (when using the same rank), but also able to attain a better approximation to the original RNNs with much fewer parameters. BTD decomposes a high order tensor into a sum of multiple Tucker decomposition models. The redundant dense connections between input and hidden state is first tensorized to a $d$-dimensional tensor and then decomposed using low-rank BTD into a sum of $N$ different Tucker decompositions where $N$ is the CP-rank. Each Tucker decomposition in turn consists of a core $d$-dimensional tensor and $d$ 3-dimensional factor tensors. While Ye et al.~ used BTD to compress RNNs, Ma et al.~ used BTD to compress the self-attention matrix in Transformers. They first build a single-block attention based on the Tucker decomposition where the query, key and value are mapped into three factor matrices and the core tensor is trainable and randomly initialized. It is then straightforward to represent the multi-head attention using BTD.\",\n \"id\": \"4c1e5567-aabc-47e9-8344-624bb81f275f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"a059247f-141a-43b7-a6a9-f3440e7622cd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Tensor decomposition\"\n ],\n [\n \"subsection\",\n \"Tensor Train and Block Term Decomposition\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Tensor Train and Block Term Decomposition\"\n },\n {\n \"cite_extract_rate\": 0.777777777777777,\n \"cites\": [\n 4527,\n 4530,\n 1568,\n 4515,\n 4524,\n 4525,\n 4526\n ],\n \"content\": \"\\\\begin{table}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|l|l|l|l|l|l|l|l|}\\n \\\\hline\\nTask&Dataset&Method&Base Model&Metric&Size (Comp; Orig)&Eval. (Comp; Orig)\\n\\\\\\\\\\n\\\\hline\\n\\\\hline\\nCTR prediction&Criteo CTR Challenge&TT-embedding~&MLP&LogLoss (L)&4.7M; 41.2M&0.4433; 0.4440\\\\\\\\\\n\\\\hline\\nLanguage modeling&1B Word Benchmark&BTD~&Transformer-XL Large&Perplexity (L)&0.16B; 0.8B&19.5; 21.8\\\\\\\\\\n\\\\hline\\nLanguage modeling&Enwiki-8&FLOP~&Transformer&BPC (L)&8M; 41M&1.13; 1.08\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&WEST~&LSTM&Perplexity (L)&3.5M; 4.51M&116.84; 115.91\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&BTD~&Transformer-XL-Base&Perplexity (L)&12M; 24M&49.8; 54.52\\\\\\\\\\n\\\\hline\\nLanguage modeling&PTB&TT-embedding~&Transformer-XL-Base&Perplexity (L)&18M; 24M&55.4; 54.52\\\\\\\\\\n\\\\hline\\nLanguage modeling&WikiText-103&FLOP~&Transformer&Perplexity (L)&50M; 151M&25.3; 24.1\\\\\\\\\\n\\\\hline\\nLanguage modeling&WikiText-103&TT-embedding~&Transformer-XL&Perplexity (L)&91M; 192M&25.67; 24.37\\\\\\\\\\n\\\\hline\\nLanguage modeling&WikiText-103&BTD~&Transformer-XL-Base&Perplexity (L)&85.3M; 151M&20.9; 24.0\\\\\\\\\\n\\\\hline\\nLanguage modeling&WikiText-103&TT-embedding~&Transformer-XL-Base&Perplexity (L)&130M; 151M&25.7; 24.0\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$ja)&ASPEC&Compositional codes~&LSTM&BLEU (H)&2.97MB; 274MB&38.89; 37.93\\\\\\\\\\n\\\\hline\\nNMT (de$\\\\rightarrow$en)&IWSLT14&Compositional codes~&LSTM&BLEU (H)&2.11MB; 35MB&29.56; 29.45\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT14&TT-embedding~&Transformer-Big&BLEU (H)&179M; 210M&28.53; 28.84\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT16&BTD~&Transformer&BLEU (H)&21.2M; 52M&34.91; 34.5\\\\\\\\\\n\\\\hline\\nNP bracketing&Subset of PTB&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&82.3; 77.9\\\\\\\\\\n\\\\hline\\nQuestion classification&TREC Questions&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&81.5; 76.2\\\\\\\\\\n\\\\hline\\nSentiment analysis&IMDB&Compositional codes~&LSTM&Acc (H)&1.23MB; 78MB&87.37; 87.18\\\\\\\\\\n\\\\hline\\nSentiment analysis&IMDB&TT-embedding~&LSTM&Acc (H)&0.81M; 7.19M&89.7; 88.6\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&81.4; 77.7\\\\\\\\\\n\\\\hline\\nSpeech recognition&3M Google voice utterances&Joint-SVD~&5-layer RNN&WER (L)&3.1M; 9.7M&12.9; 12.4\\\\\\\\\\n\\\\hline\\nSpeech recognition&3M Google voice utterances&Projections~&LSTM&WER (L)&2M; 2.2M&14.8; 17.5\\\\\\\\\\n\\\\hline\\nSpeech recognition&Live traffic utterances&WEST~&3-layer LSTM&WER (L)&4.75MB; 15MB&13.6; 13.7\\\\\\\\\\n\\\\hline\\nSpeech recognition&Live traffic utterances&WEST+Quantization~&3-layer LSTM&WER (L)&1.35MB; 15MB&13.7; 13.7\\\\\\\\\\n\\\\hline\\nText classification&20 Newsgroup (Computers)&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&87.0; 79.7\\\\\\\\\\n\\\\hline\\nText classification&20 Newsgroup (Religion)&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&88.8; 86.7\\\\\\\\\\n\\\\hline\\nText classification&20 Newsgroup (Sports)&Sparse coding~&Logistic regression&Acc (H)&120M; 120M&96.3; 95.9\\\\\\\\\\n\\\\hline\\nWord similarity&Simlex-999&Sparse coding~&Logistic regression&Correlation (H)&120M; 120M&38.9; 36.9\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n\\\\caption{Comparison of various tensor decomposition methods (sorted by Task and then Dataset). In the metric column, H means high is better while L means low is better. For compositional codes, 16x32 coding was used. For BTD, two block term tensors were used. In~, logistic regression uses GloVe embeddings.}\\n \\\\label{tab:tensorDecompSummary}\\n\\\\end{table}\\nTable~\\\\ref{tab:tensorDecompSummary} compares various tensor decomposition methods across different tasks and datasets. Accuracy of both the original and the compressed (comp.) model are shown. Also, we report model size (in terms of number of parameters or memory size) for both the student as well as the teacher models. For the same task, dataset and model combination, different papers report different accuracy of the original model because of slight changes in training hyper-parameters; hence we report accuracy of the original model for each row. \\nFor CTR prediction, the TT-embedding method~ in Table~\\\\ref{tab:tensorDecompSummary} uses 3 factors with TT-rank of 16. It actually leads to an embedding compression of 16. With 4 factors and TT-rank=2, test loss increases to 0.4530 with a massive embedding compression of 4193 and overall model size of 0.53M. Thus, substitution of large embedding layers with TT–embeddings leads to significant compression ratios (up to 2011 times) with a slight improvement in the test loss, and up to 4200 with a small drop in the test loss. \\nFor language modeling, BTD~ leads to an improved model with 20\\\\% of the Transformer-XL large model. For character level language modeling using FLOP~, an 8M sized FLOP model achieves 1.13 on Enwiki8 while gradual pruning achieves 1.14. Thus, FLOP based pruning which combines pruning with matrix factorization is better than both structured neuron pruning as well as gradual unstructured pruning. On PTB, BTD is clearly much better than TT-embedding. With half the model size compared to Transformer-XL-Base, BTD leads to a model with $\\\\sim$10\\\\% lower perplexity. On WikiText-103, while FLOP pruned model achieves 25.3 perplexity with model size of 50M, gradual unstructured pruning and neuron pruning achieve 25.7 and 26.7 respectively with the same model size for language modeling on Wiki-103 dataset. Thus, FLOP is better than other pruning methods. Again on Wiki-103, BTD is superior to TT-embedding. \\nFor NMT, the loss-free compression rate reaches 92\\\\% on ASPEC dataset by pruning 90\\\\% of the connections. However, with the same pruning ratio, a modest performance loss is observed in IWSLT14 dataset. For the models using compositional coding~, the loss-free compression rate is 94\\\\% for the IWSLT14 dataset and 99\\\\% for the ASPEC dataset. Thus, compositional codes are much better compared to pruning. TT-embedding with TT-rank=64 leads to embedding compression of 15.3 on WMT14 dataset with marginal loss in the BLEU score. Lastly, BTD has been used to reduce Transformer size by more than half with improved BLEU on WMT16 (en$\\\\rightarrow$de) data.\\nSparse coding for GloVe vectors~ has led to improved accuracies for multiple tasks like Noun Phrase (NP) bracketing, question classification, text classification and word similarity, while retaining the same model size. For sentiment analysis on the IMDB dataset, compositional codes method achieves a compression rate of 98\\\\% without performance loss. Further, TT-embedding method leads to a much smaller model compared to compositional codes with better accuracy. In the TT-embedding case, embedding compression rate is 441 with TT-rank=16.\\nFor Speech recognition, Prabhavalkar et al.~ experimented with a 3M Google voice utterances dataset and found that a joint SVD with explained variance retained after the SVD as 0.6 leads to a 3x smaller RNN model without much performance loss. They improved upon Sak et al.'s projections method which led to a much higher WER on the same dataset. On live traffic utterances dataset, WEST~ leads to a 3x smaller model with a slightly reduced word error rate when using 3-layer LSTMs. Variani et al.~ further compress this model using quantization to obtain a 11x compressed 3-layer LSTM model without any performance loss. Thus, using quantization along with matrix decomposition seems to work well.\\nOverall, matrix decomposition techniques are usually used in combination with parameter sharing and sometimes with quantization. They have been very effective in dealing with large input/output embedding matrices in RNNs and LSTMs. SVD, Tensor Train, CP, Tucker, BTD have been the most popular decomposition techniques found to be useful for model compression.\",\n \"id\": \"2c213350-b4bf-49d2-a046-a4fe16413654\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"a059247f-141a-43b7-a6a9-f3440e7622cd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Tensor decomposition\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4531\n ],\n \"content\": \"\\\\label{sec:linearTransformers}\\nMemory usage throughout neural network training can be categorized into three main types: (1) Model memory is used to store model parameters; (2) Optimizer memory is the additional memory used by the specific learning algorithm during the process; (3) Activation memory consists of the outputs of each layer, which are cached for reuse in backpropagation to compute gradients. For a BERT-base model, model memory is around 0.2 GB, optimizer memory is around 1GB, while activation memory is around 8.5GB~. Time and activation memory in Transformers grows quadratically with the sequence length. This is because in every layer, every attention head attempts to come up with a transformed representation for every position by ``paying attention'' to tokens at every other position. Quadratic complexity implies that practically the maximum input size is rather limited. Thus, we cannot extract semantic representation for long documents by passing them as input to Transformers.\",\n \"id\": \"65404f84-cbf0-4721-8f05-178f8230aa55\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Transformers with Sub-Quadratic Complexity\"\n ]\n ],\n \"subsections\": [\n \"4db79252-def3-4c73-83ab-c7863b1703fc\",\n \"f00ad6f8-f45a-4b14-ab1c-135e291338eb\",\n \"2202feab-4da3-4fb5-8d59-2d1d4fbac3f9\"\n ],\n \"title\": \"Transformers with Sub-Quadratic Complexity\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4531,\n 793,\n 794\n ],\n \"content\": \"A few efforts like BlockBERT~ try to reduce this quadratic complexity by a constant factor by introducing sparse block structures into the attention matrix. If we split the length-$N$ input sequence into $n$ blocks, $N\\\\times N$ attention matrix gets partitioned into $n\\\\times n$ blocks, where each block matrix is of the size $\\\\frac{N}{n}\\\\times\\\\frac{N}{n}$. Thus, BlockBERT reduces $O(N^2)$ memory consumption by a factor of $n$. \\nChild et al.~ propose sparse transformers where sparse factorizations of the attention matrix reduce the quadratic complexity to $O(n\\\\sqrt{n})$. The key idea is to reduce the dense attention matrix to a sparse version by only computing attention on a sparse number of (query, key) pairs. They propose two kinds of sparse factorizations: strided and fixed. Strided attention implies having one head attend to the previous $l$ locations, and the other head attend to every $l^{th}$ location, where $l$ is the stride and chosen to be close to $\\\\sqrt{n}$. More heads could be used with a different stride value. Fixed attention assumes that specific positions summarize previous locations and propagate that information to all future positions. \\nThe Reformer architecture~ replaces the dot-product attention in a typical Transformer by one that uses locality-sensitive hashing (LSH), changing its complexity from $O(n^2)$ to $O(n\\\\log n)$, where $n$ is the length of the sequence. In a standard Transformer, we compute scaled dot-product attention as follows.\\n\\\\begin{eqnarray}\\n \\\\text{Attention}(Q,K,V)=\\\\text{softmax}\\\\left(\\\\frac{QK^T}{\\\\sqrt{d}}\\\\right)V\\n\\\\end{eqnarray}\\nwhere $Q$, $K$ and $V$ are the standard query, key and value components and $d$ is a scaling factor. Reformer uses a Shared QK Transformer, i.e., $Q=K$ enabled by sharing the matrix that projects words/hidden layer to $Q$ or $K$. Further, note that we are actually only interested in $\\\\text{softmax}(QK^T)$. Since softmax is dominated by the largest elements, for each query $q_i$ we only need to focus on the keys in $K$ that are closest to $q_i$. How can we find the nearest neighbors among the keys? Reformer uses LSH. LSH is used to cluster (hash-bucket) the positions into various groups, and then every position needs to focus only on others within the same bucket.\",\n \"id\": \"4db79252-def3-4c73-83ab-c7863b1703fc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"65404f84-cbf0-4721-8f05-178f8230aa55\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Transformers with Sub-Quadratic Complexity\"\n ],\n [\n \"subsection\",\n \"Transformers with Super-Linear Complexity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Transformers with Super-Linear Complexity\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7333,\n 7298,\n 798,\n 7371,\n 1473,\n 7827\n ],\n \"content\": \"Even better, there have been several efforts recently to reduce this quadratic complexity to linear. Most of these efforts choose a constant number of other positions to ``pay attention'' to so as to compute a transformed representation for any given position. They can model sequences tens of thousands of timesteps long using hundreds of layers. The methods differ in their approach towards selecting this constant number of other positions. We discuss a few of such recently proposed methods in this section.\\nIn Star-Transformers~, to reduce model complexity from $O(n^2)$ to linear, we replace the fully-connected attention matrix structure with a star-shaped topology, in which every two non-adjacent nodes are connected through a shared relay node. While ring connections connect a satellite node with two other satellite nodes, a radical connection connects a satellite node with the relay node. The idea is to update the star-center relay node based on satellite nodes and then update satellite nodes using information from the star node, and adjacent satellite nodes. \\nLinformer architecture~ exploits low-rank factorization of the self-attention matrix to reduce overall self-attention complexity from $O(n^2)$ to $O(n)$ in both time and space. The main idea is to add two linear projection matrices $E_i, F_i\\\\in R^{n\\\\times k}$ when computing key and value. We first project the original $(n\\\\times d)$-dimensional key and value layers into $(k\\\\times d)$-dimensional projected key and value layers. We then compute an $(n \\\\times k)$-dimensional context mapping matrix using scaled dot-product attention. If we can choose a very small projected dimension $k$, such that $k<<n$, then we can significantly reduce the memory and space consumption. Overall, it is $O(nk)$. Further, we can do three other forms of parameter sharing: (1) Headwise sharing: $E_i=E$ and $F_i=F$ across all heads $i$ in a layer. (2) Key-value sharing: $E_i=F_i=E$ across all heads $i$ in a layer. (3) Layerwise sharing: Single projection matrix $E$ is used across all layers, all heads for both key and value.\\nSparse Sinkhorn Attention based Transformer~ is based on differentiable sorting of internal representations. First, they divide the input sequence into $B$ equal sized blocks each of size $n/B$. A meta sorting network learns to generate latent permutations over these block sequences. Given sorted sequences, we are then able to compute quasi-global attention with only local windows, improving the memory efficiency of the attention module. They also propose Causal Sinkhorn Balancing and SortCut algorithms for causal scenarios for tailoring Sinkhorn Attention for encoding and/or decoding purposes. Their method reduces the memory complexity from $O(n^2)$ to $O(B^2+(n/B)^2)$. The SortCut variant further reduces complexity to linear-time, i.e., $O(nk)$ where $k$ is a user defined budget hyper-parameter much smaller than $n$.\\nShen et al.~ propose a very simple mathematical trick to reduce quadratic complexity to linear. A typical dot-product attention can be written as $\\\\text{softmax}(QK^T)V$ ignoring the scale factor. This is quadratic because $QK^T$ is $n^2$ in size. This can be rewritten as follows.\\n\\\\begin{eqnarray}\\n \\\\text{Attention}(Q,K,V)=\\\\text{softmax}_r(Q)(\\\\text{softmax}_c(K^TV))\\n\\\\end{eqnarray}\\nwhere $\\\\text{softmax}_r$ and $\\\\text{softmax}_c$ are softmax applied to rows and columns respectively. This revised formulation has terms which are only linear in $n$. Finally, Katharopoulos et al.~ express the self-attention as a linear dot-product of kernel feature maps and make use of the associativity property of matrix products to reduce the complexity from $O(n^2)$ to $O(n)$, where $n$ is the sequence length.\\nFinally, Longformer~ propose to reduce Transformer complexity to linear by sparsifying the full self-attention matrix using multiple kinds of attention patterns. The simplest attention pattern is the sliding window pattern which employs a fixed-size window attention surrounding each token. Given a fixed window size $w$, each token attends to $0.5w$ tokens on each side. The computation complexity of this pattern is $O(nw)$ which scales linearly with input sequence length $n$. To further increase the receptive field without increasing computation, the sliding window can be ``dilated''. This is analogous to dilated CNNs where the window has gaps of size dilation $d$. The windowed and dilated attention are not flexible enough to learn task-specific representations. Accordingly, Longformer also has ``global attention'' on few pre-selected input locations. Moreover, this attention operation is symmetric: that is, a token with a global attention attends to all tokens across the sequence, and all tokens in the sequence attend to it. Further, two different sets of projections are used: $Q_s$, $K_s$, $V_s$ to compute attention scores of sliding window attention, and $Q_g$, $K_g$, $V_g$ to compute attention scores for the global attention. The pretrained Longformer consistently outperforms RoBERTa on multiple downstream long document tasks.\",\n \"id\": \"f00ad6f8-f45a-4b14-ab1c-135e291338eb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"65404f84-cbf0-4721-8f05-178f8230aa55\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Transformers with Sub-Quadratic Complexity\"\n ],\n [\n \"subsection\",\n \"Transformers with Linear Complexity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Transformers with Linear Complexity\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7333,\n 4531,\n 7298,\n 793,\n 798,\n 7371,\n 794,\n 1473\n ],\n \"content\": \"\\\\begin{table}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|l|l|l|l|l|l|}\\n \\\\hline\\nTask&Dataset&Model&Base Model&Metric&Eval. (Opt.; Orig)\\\\\\\\\\n\\\\hline\\n\\\\hline\\nChar-level language modeling&1B Word Benchmark&Sinkhorn Mixture Big~&Transformer Big&BPC (L)&1.134; 1.825\\\\\\\\\\n\\\\hline\\nChar-level language modeling&1B Word Benchmark&Sparse Transformer~&Transformer Big&BPC (L)&1.119; 1.825\\\\\\\\\\n\\\\hline\\nChar-level language modeling&Enwik8&Longformer~&Transformer&BPC (L)&1.00; 1.11\\\\\\\\\\n\\\\hline\\nChar-level language modeling&Enwik8&Reformer~&Transformer&BPC (L)&1.05; 1.11\\\\\\\\\\n\\\\hline\\nChar-level language modeling&text8&Longformer~&Transformer&BPC (L)&1.10; 1.18\\\\\\\\\\n\\\\hline\\nCoreference resolution&OntoNotes&Longformer-base~&RoBERTa-base&F1 (H)&78.6; 78.4\\\\\\\\\\n\\\\hline\\nLanguage modeling&1B Word Benchmark&Sinkhorn Mixture Big~&Transformer Big&Perplexity (L)&27.34; 27.59\\\\\\\\\\n\\\\hline\\nLanguage modeling&1B Word Benchmark&Sparse Transformer~&Transformer Big&Perplexity (L)&28.77; 27.59\\\\\\\\\\n\\\\hline\\nNamed Entity Recognition&CoNLL2003&Star Transformer~&Transformer&Acc (H)&90.93; 86.48\\\\\\\\\\n\\\\hline\\nNamed Entity Recognition&CoNLL2012&Star Transformer~&Transformer&Acc (H)&86.30; 83.57\\\\\\\\\\n\\\\hline\\nNLI&MNLI&SortCut Sinkhorn~&Transformer&Acc (H)&55.80; 53.69\\\\\\\\\\n\\\\hline\\nNLI&QNLI&Linformer~&BERT-base&Acc (H)&91.2; 91.8\\\\\\\\\\n\\\\hline\\nNLI&SNLI&Star Transformer~&Transformer&Acc (H)&86.0; 82.2\\\\\\\\\\n\\\\hline\\nNLI&SNLI&SortCut Sinkhorn~&Transformer&Acc (H)&80.30; 78.87\\\\\\\\\\n\\\\hline\\nNMT (en$\\\\rightarrow$de)&WMT14&Reformer big~&Transformer Big&BLEU (H)&29.1; 27.3\\\\\\\\\\n\\\\hline\\nPOS Tagging&PTB&Star Transformer~&Transformer&Acc (H)&97.14; 96.31\\\\\\\\\\n\\\\hline\\nQuestion answering&HotpotQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&78.94; 77.08\\\\\\\\\\n\\\\hline\\nQuestion answering&HotpotQA&SparseBERT~&BERT&F1 (H)&76.02; 77.08\\\\\\\\\\n\\\\hline\\nQuestion answering&NaturalQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&79.39; 78.29\\\\\\\\\\n\\\\hline\\nQuestion answering&NaturalQA&SparseBERT~&BERT&F1 (H)&77.31; 78.29\\\\\\\\\\n\\\\hline\\nQuestion answering&NewsQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&70.08; 66.25\\\\\\\\\\n\\\\hline\\nQuestion answering&NewsQA&SparseBERT~&BERT&F1 (H)&67.16; 66.25\\\\\\\\\\n\\\\hline\\nQuestion answering&SearchQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&83.51; 80.37\\\\\\\\\\n\\\\hline\\nQuestion answering&SearchQA&SparseBERT~&BERT&F1 (H)&80.54; 80.37\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD 1.1&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&90.74; 88.45\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD 1.1&SparseBERT~&BERT&F1 (H)&88.37; 88.45\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD 2.0&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&81.45; 77.16\\\\\\\\\\n\\\\hline\\nQuestion answering&SQuAD 2.0&SparseBERT~&BERT&F1 (H)&77.57; 77.16\\\\\\\\\\n\\\\hline\\nQuestion answering&TriviaQA&BlockBERT (n=2, N=1024)~&BERT&F1 (H)&79.41; 75.35\\\\\\\\\\n\\\\hline\\nQuestion answering&TriviaQA&SparseBERT~&BERT&F1 (H)&75.34; 75.35\\\\\\\\\\n\\\\hline\\nQuestion answering&TriviaQA&Longformer-base~&RoBERTa-base&F1 (H)&75.2; 74.3\\\\\\\\\\n\\\\hline\\nQuestion answering&WikiHop&Longformer-base~&RoBERTa-base&Acc (H)&75.0; 72.4\\\\\\\\\\n\\\\hline\\nSentiment analysis&IMDB&Linformer~&BERT-base&Acc (H)&94.1; 93.5\\\\\\\\\\n\\\\hline\\nSentiment analysis&IMDB&Longformer-base~&RoBERTa-base&Acc (H)&95.7; 95.3\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST&Star Transformer~&Transformer&Acc (H)&52.9; 50.4\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST&Sinkhorn ~&Transformer&Acc (H)&77.52; 76.83\\\\\\\\\\n\\\\hline\\nSentiment analysis&SST-2&Linformer~&BERT-base&Acc (H)&93.1; 92.7\\\\\\\\\\n\\\\hline\\nSpeech recognition&WSJ&Linear Transformer~&Reformer&Phoneme Error Rate (L)&8.08; 9.33\\\\\\\\\\n\\\\hline\\nText classification&Hyperpartisan&Longformer-base~&RoBERTa-base&F1 (H)&94.8; 87.4\\\\\\\\\\n\\\\hline\\nText classification&MTL-16&Star Transformer~&Transformer&Acc (H)&86.98; 82.78\\\\\\\\\\n\\\\hline\\nTextual similarity&QQP&Linformer~&BERT-base&Acc (H)&90.8; 89.6\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n\\\\caption{Comparison of various sub-quadratic complexity Transformer methods (sorted by Task and then Dataset). In the metric column, H means high is better while L means low is better. Note that in this case, model sizes do not reduce much; activation memory reduces as described in Section~\\\\ref{sec:linearTransformers} (with comparable or better accuracy) compared to the standard Transformer.}\\n \\\\label{tab:linearTransSummary}\\n\\\\end{table}\\nTable~\\\\ref{tab:linearTransSummary} compares various sub-quadratic Transformer methods across different tasks and datasets. Accuracy of both the original Transformer and the optimized (opt.) model are shown. These models have been applied for various applications like language modeling (both word level as well as character level), coreference resolution, NER, NLI, NMT, POS, question answering, sentiment analysis, speech recognition, text classification and text similarity analysis. For the same task, dataset and model combination, different papers report different accuracy of the original model because of slight changes in training hyper-parameters; hence we report accuracy of the original model for each row. \\nStar Transformers were shown to outperform the vanilla Transformer model across various tasks like Sentiment analysis, Text classification, NLI, POS and NER. On the SST, the Star Transformer achieves 2.5 points improvement against the standard Transformer. On MTL-16, the Star-Transformer outperform the standard Transformer in all 16 datasets, the improvement of the average accuracy is 4.2. Average test times are 10.94 and 49.31ms per batch with batch-size=128 for Star Transformer and standard Transformer respectively.\\nLongformer achieves a new state-of-the-art on both text8 and enwik8 using the small models with BPC of 1.10 and 1.00 on text8 and enwik8 respectively. Longformer-large model of the same size as Sparse Transformer achieves a BPC of 0.99 on Enwik8, which is the same as that obtained using Sparse Transformer. Also, for both character-level as well as word-level language modeling, on 1B word benchmark, we observe that Sinkhorn mixture (which is a combination of the Sinkhorn attention by mixing it with the vanilla standard dot product attention) performs better than Sparse Transformer. \\nOn question answering, results have been reported across multiple datasets like HotpotQA, NaturalQA, NewsQA, SearchQA, TriviaQA, WikiHop, SQuAD 1.1 and SQuAD 2.0. We observe that BlockBERT performs better than SparseBERT. Also, it is not surprising that BlockBERT with 2 blocks (n = 2) performs better than that with 3 blocks (n = 3), because it keeps more attention matrix entries. Also, not shown in the table, Longformer-large achieves scores of 81.9 and 73.2 for WikiHop and HotpotQA beating state-of-the-art results by 3.6 and 4 points respectively. \\nLongformer consistently outperforms the RoBERTa baseline across many tasks like coreference resolution, question answering, sentiment analysis and text classification. Its performance gain is especially obvious for tasks that require long context such as WikiHop (question answering) and Hyperpartisan (text classification). For TriviaQA (question answering), the improvement is more modest as the local context is often sufficient to answer the question. In the case of HotpotQA (question answering), the supporting fact auxiliary supervision allows models to easily find relevant contexts and then focus on local context, leading to smaller gains. On the IMDB (sentiment analysis) and OntoNotes (coreference resolution) datasets the performance gains are smaller. For IMDB, the majority of the dataset consists of short documents and thus it is expected to see smaller improvements. For OntoNotes, the distance between any two mentions is typically quite small so that a baseline that processes smaller chunks separately is able to stitch together mentions into coreference chains without considering cross chunk interactions.\\nOn speech recognition task, Linear transformers achieve similar performance to vanilla transformers and they are up to 4000x faster on autoregressive prediction of very long sequences. The linear Transformer model outperforms the LSTM and Reformer while being faster to train and evaluate. Reformer takes 2250 seconds per epoch, while Linear Transformers take just 824s/epoch.\\nTo summarize, multiple methods have been proposed to reduce the quadratic complexity of the standard Transformer model. While Sparse Transformers reduce it to $O(n\\\\sqrt{n})$, Reformers reduce it to $O(n\\\\log n)$. Other methods like Star Transformer, Linformer, Sparse Sinkhorn Transformer, Efficient Attention, Linear Transformers and Longformer promise linear complexity. In particular Sparse Sinkhorn Transformer and Longformer have been shown to result into very good accuracy, latency and RAM tradeoff across many tasks.\",\n \"id\": \"2202feab-4da3-4fb5-8d59-2d1d4fbac3f9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"65404f84-cbf0-4721-8f05-178f8230aa55\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Transformers with Sub-Quadratic Complexity\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:summary}\\nWe discussed various methods for compression of deep learning models for text. Broadly, we discussed pruning, quantization, knowledge distillation, parameter sharing, tensor decomposition, and sub-quadratic Transformer based methods. These methods not just help reduce the model size, but also lead to lower prediction latencies and low power consumption due to reduced computations. Pruning is the oldest method, but not commonly applied for Transformers. Quantization is effective however it is important to use mixed precision balanced quantization with GPU architectures that support efficient low-bit computations. Knowledge Distillation is the most popular method for compression of Transformer models. Parameter sharing is a very useful method but often needs to be combined with other techniques. Matrix decomposition is not very common but has lots of potential, especially the BTD method. Sub-quadratic Transformers are very important to enable processing long documents in applications like query-document similarity, long document summarization, etc.\",\n \"id\": \"d06ca9cc-bcd7-4f53-9be8-ee86f8754664\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3efda6e0-cf0a-4f26-8cb1-3dde0497c32c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Future Directions\"\n ]\n ],\n \"subsections\": [\n \"ab02f8bb-9d54-4f84-82fc-3f831c1bc39a\",\n \"fe1a6e49-c5b1-4394-b0be-1a452c8b0f80\",\n \"7ea3f834-f595-4c1b-afed-044d53e835ea\"\n ],\n \"title\": \"Summary and Future Directions\"\n },\n {\n \"cite_extract_rate\": 0.76,\n \"cites\": [\n 4479,\n 4478,\n 7298,\n 4515,\n 1150,\n 4517,\n 4480,\n 4521,\n 4514,\n 4485,\n 4492,\n 1473,\n 4491,\n 8375,\n 9120,\n 4522,\n 4536,\n 4497,\n 4533,\n 8799,\n 4525,\n 4519,\n 4526,\n 854,\n 4484,\n 7829,\n 7333,\n 8797,\n 4486,\n 4516,\n 4527,\n 4500,\n 793,\n 4530,\n 1568,\n 4036,\n 4523,\n 7371,\n 4532,\n 4524,\n 8390,\n 4499,\n 4331,\n 4509,\n 4534,\n 4511,\n 2481,\n 844,\n 4535,\n 4482,\n 7828,\n 4518,\n 794,\n 4351,\n 4507,\n 2488,\n 4510\n ],\n \"content\": \"In the previous six sections, we have compared across multiple methods within each of the six broad types of model compression. In this subsection, we attempt to compare across multiple model compression method types.\\nTable~\\\\ref{tab:kdSummary2} provides a comparison of model size versus accuracy for various methods across various GLUE~ and SQuAD tasks. We observe that most of the methods tried on GLUE datasets have been based on knowledge distillation. However, some pruning methods (iterative magnitude pruning, RPP and LayerDrop), quantization (mixed precision QBERT), parameter sharing (ALBERT), tensor decomposition (FLOP) and linear Transformer (Linformer) have also been tried. Quantization methods do not reduce the number of parameters but reduce the number of bits per parameter. Particularly, mixed precision QBERT uses 8 bits for embeddings and 2/3/4 bits for encoder layer weights. Similarly, Linformer does not reduce number of weights but reduces activation memory as well as latency of the overall Transformer model. \\nFor the remaining models, for each model, we first computed an average GLUE score based on any of the 9 tasks for which scores have been reported. Next, we computed the ratio GLUE score/model size (in Millions). We find the following as the top three \\n\\\\begin{itemize}\\n \\\\item Distilled-BiLSTM~ (ratio=79.3)\\n \\\\item Mixed-vocab. training~ (ratio=7.8)\\n \\\\item ALBERT-B~ (ratio=7.2)\\n\\\\end{itemize}\\nThis clearly tells us that Distilled BiLSTMs provide us the best accuracy versus size tradeoff on GLUE. However, each of these three models actually report results on only 4 out of 9 GLUE tasks. Hence, further, we considered only those methods for which results on at least 5 tasks have been reported and computed the GLUE score/model size ratio. We find the following as the top three \\n\\\\begin{itemize}\\n\\\\item TinyBERT~ (ratio=5.31)\\n\\\\item BERT-EMD~ (ratio=5.15)\\n\\\\item MobileBERT+Quantization~ (ratio=3.49)\\n\\\\end{itemize}\\nThus, on GLUE tasks, it is clear that distillation based methods (combined with quantization) are better than other types of methods. \\nTables~\\\\ref{tab:pruningSummary},~\\\\ref{tab:quantizationSummary},~\\\\ref{tab:kdSummary1},~\\\\ref{tab:paramSharingSummary},~\\\\ref{tab:tensorDecompSummary} and~\\\\ref{tab:linearTransSummary} compare various pruning, quantization, knowledge distillation, parameter sharing, tensor decomposition and sub-quadratic Transformer methods across different tasks and datasets. Overall, there are 123 (task, dataset) combinations across these six tables which implies that unlike GLUE, not many methods have been applied on the same set of (task, dataset) combinations\\\\footnote{We make the entire statistics available as an excel file at \\\\url{https://bit.ly/3vmaxZ9}.}. The most popular tasks are language modeling (on PTB and 1B word benchmark), sentiment analysis (on SST) and NMT (WMT14 en$\\\\rightarrow$de). \\nFor language modeling on PTB, on LSTM models, bank balanced sparsity~ based pruning method worked best. With Transformer models, Block Term Decomposition (BTD)~ method seems to work best. Among various methods like parameter sharing, tensor decomposition and sub-quadratic complexity Transformer which have been tried for language modeling on 1B Word Benchmark, again, BTD~ method seems to work best leading to a model with 0.16B parameters and a perplexity as low as 19.5. Multiple datasets have been used for Neural machine translation (NMT). Datasets from WMT and IWSLT are the most popular. Among these, the en$\\\\rightarrow$de from WMT14 is the most popular dataset used for testing various NMT models. For en$\\\\rightarrow$de NMT with WMT14, using 2-layer LSTMs, the best accuracy versus size tradeoff is using Pruned Seq-KD + Seq-Inter~ which gives a 8M size model leading to 18.5 BLEU. Among Transformer based models, BTD~ leads to a Transformer model which provides 34.91 BLEU with 21.2M model size.\\n\\\\begin{table*}\\n \\\\centering\\n \\\\scriptsize\\n \\\\begin{tabular}{|p{1in}|p{2in}|p{2.7in}|}\\n \\\\hline\\nTask&Popular Datasets&References\\\\\\\\\\n\\\\hline\\n\\\\hline\\nLanguage modeling&Penn TreeBank Corpus, One billion word benchmark, Europarl, WikiText-103, text8, source code of Linux kernel, 2013 ACL Workshop Morphological Language Datasets (ACLW), Arabic news commentary corpus, 2013 ACL workshop on MT, enwik8 (from Wikipedia), Lambada&Neuron Pruning~, Iterative magnitude pruning~, Block sparsity~,Loss -Aware Quantization~, Uniform Quantization~, Binary Quantization~, HitNet~, Sparse Word Representations~, LightRNN~, Slim Embeddings~, C2W~, LayerDrop~, Reformer~, Linformer~, Char-CNN~, CNN+Highway Network~, SparseBERT~, FLOP~, Deep Equilibrium Models~, WEST~, Sparse Sinkhorn Attention~, BTD~, Universal Transformers~, TT-embedding~, multiple methods~\\\\\\\\\\n\\\\hline\\nNeural Machine translation (NMT)&IWSLT German-English, IWSLT Thai-English, ASPEC English-Japanese, WMT English-German, WMT German-English, WMT English-Russian, IWSLT English Vietnamese, WMT English-Romanian, WMT English-Estonian, Ted Talk&Compositional codes~, LayerDrop~, Pruning attention heads~, Neuron Pruning~, Magnitude Pruning~, Iterative magnitude pruning~, Pruned Seq-KD + Seq-Inter~, Quantized Distillation~, Teacher ensembles KD~, Multiple teachers KD~, BTD~, Quaternion Attention~, Universal Transformers~, TT-embedding~\\\\\\\\\\n\\\\hline\\nSentiment Analysis&IMDB movie review, SST, SST-2, Elec (electronic product reviews)&Compositional codes~, Star Transformer~, TinyBERT~, MiniLM~, Linformer~, XtremeDistil~, Gaussian Quantization~, Uniform Quantization~, Sparse coding~, Quaternion Attention~, Sparse Sinkhorn Attention~, RPP~, ALBERT~, Patient KD~, Mixed-vocabulary KD training~, Distilled-BiLSTM~, MTDNN~, TT-embedding~\\\\\\\\\\n\\\\hline\\nQuestion Answering&SQuAD1.1, SQuAD2.0, ELI5, SemEval, BABI&LayerDrop~, MiniLM~, RPP~, BS-Fixed Quantization~, ALBERT~, KD~, Universal Transformers~\\\\\\\\\\n\\\\hline\\nNatural Language Inference&SNLI, MNLI-m, MNLI-mm, QNLI, RTE, WNLI, XNLI&Star Transformer~, LayerDrop~, TinyBERT~, MiniLM~, Linformer~, Sparse Sinkhorn Attention~, RPP~, ALBERT~, Patient KD~, Mixed-vocabulary KD training~, Distilled-BiLSTM~, MTDNN~\\\\\\\\\\n\\\\hline\\nParaphrasing&QQP, STS-B&TinyBERT~, MiniLM~, Linformer~, RPP~, ALBERT~, Patient KD~, Distilled-BiLSTM~, MTDNN~\\\\\\\\\\n\\\\hline\\nImage captioning&MSCOCO&Grow and Prune~, Magnitude Pruning~, Iterative Magnitude Pruning and Densification~\\\\\\\\\\n\\\\hline\\nHandwritten character recognition&ICDAR&SVD and Pruning\\\\\\\\\\n\\\\hline\\nPart-of-speech (POS) tagging&Wall Street Journal of the Penn Treebank dataset, WikiAnn NER corpus&C2W~, XtremeDistil~\\\\\\\\\\n\\\\hline\\nSummarization&CNN-DailyMail, XSum&LayerDrop~, MiniLM~\\\\\\\\\\n\\\\hline\\nMachine Reading Comprehension& Microsoft Research Paraphrase Corpus (MRPC), ReAding Comprehension from Examinations (RACE)&LayerDrop~, TinyBERT~, MiniLM~, RPP~, ALBERT~, Patient KD~, Mixed-vocabulary KD training~, MTDNN~\\\\\\\\\\n\\\\hline\\nLinguistic Acceptability&CoLA&TinyBERT~, MiniLM~, RPP~, ALBERT~, MTDNN~\\\\\\\\\\n\\\\hline\\nTopic Classification&DbPedia, Ag News,20 Newsgroup&XtremeDistil~, Sparse coding~\\\\\\\\\\n\\\\hline\\nQuestion Type Classification&TREC&Sparse coding~\\\\\\\\\\n\\\\hline\\nNoun Phrase Bracketing&Lazaridou~&Sparse coding~\\\\\\\\\\n\\\\hline\\nWord Similarity&SimLex-999, MEN, MTurk, RARE, SCWS, WSR, WSS&Sparse coding~, Shared reference vectors~\\\\\\\\\\n\\\\hline\\nMathematical Language Understanding&Wangperawong's MLU~&Quaternion Attention~\\\\\\\\\\n\\\\hline\\nSubject Verb Agreement&Linzen~&Quaternion Attention~, Universal Transformers~\\\\\\\\\\n\\\\hline\\nWord Analogy&GSEM, GSYN, MSYN&Shared reference vectors~\\\\\\\\\\n\\\\hline\\nSentence Completion&MSC&Shared reference vectors~\\\\\\\\\\n\\\\hline\\nLearning to execute&Zaremba and Sutskever~&Universal Transformers~\\\\\\\\\\n\\\\hline\\nAd Click Through Rate Prediction&Criteo Kaggle&TT-embedding~\\\\\\\\\\n\\\\hline\\nSpeech Recognition&2100 hours English Speech, AN4, Switchboard, TIMIT, WSJ 92, WSJ 93, TIDIGITS, 3M Google voice utterances, Live traffic utterances&Iterative Magnitude Pruning~, Grow and Prune~, Neuron Pruning~, Block Sparsity~, BBS~, DSD~, Pow2 Ternarization~, Loss Aware Quantization~, Toeplitz-like~, Joint-SVD~, Projections~, WEST~\\\\\\\\\\n\\\\hline\\nNamed entity recognition (NER)&CoNLL2003, Wikiann-41, CoNLL2012&QBERT~, XtremeDistill~, Star Transformer~\\\\\\\\\\n\\\\hline\\nIntent Detection&SNIPS&Mixed-vocabulary KD training~\\\\\\\\\\n\\\\hline\\nQuestion Generation&SQuAD 1.1&MiniLM~\\\\\\\\\\n\\\\hline\\nSlot Filling&SNIPS&Mixed-vocabulary KD training~\\\\\\\\\\n\\\\hline\\nText classification&20 Newsgroup, Hyperpartisan, MTL-16&Sparse coding~, Longformer~, Star Transformer~\\\\\\\\\\n\\\\hline\\nCoreference Resolution&OntoNotes&Longformer~\\\\\\\\\\n\\\\hline\\n \\\\end{tabular}\\n \\\\caption{Applications of Model Compression Methods for Text}\\n \\\\label{tab:applications}\\n\\\\end{table*}\\nCombinations of multiple model compression method types has also been experimented with and found to be effective. Some examples of such combinations include the following:\\n\\\\begin{itemize}\\n \\\\item Pruning + Tensor Decomposition ~\\n \\\\item Pruning + Quantization~\\n \\\\item Knowledge distillation + Quantization~\\n \\\\item Knowledge distillation + Pruning~\\n \\\\item Tensor decomposition + Parameter sharing~\\n \\\\item Tensor decomposition + Quantization~\\n\\\\end{itemize}\\nRecently, Kim et al.~ combined knowledge distillation, structured pruning and quantization leading to drastic improvements on inference efficiency. First, they investigate the efficacy of various Knowledge Distillation techniques to significantly reduce the size of the models with respect to the depth and hidden state sizes while preserving the\\naccuracy. Second, they explore Structured Pruning that further reduces the size of the models by reducing the number of self-attention heads and the number of intermediate hidden states in the feedforward layers to achieve more efficiency while\\ntrying to preserve the accuracy as well. Finally, they \\nexplore model quantization which enables faster model executions by optimally utilizing hardware acceleration capabilities. Such a combined method leads to heavily reduced model size, 12.4x GPU speed-up and 6.9x-125.8x reduction in energy consumption.\",\n \"id\": \"ab02f8bb-9d54-4f84-82fc-3f831c1bc39a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 75,\n \"parent_id\": \"d06ca9cc-bcd7-4f53-9be8-ee86f8754664\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Future Directions\"\n ],\n [\n \"subsection\",\n \"Comparison across model compression method types.\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Comparison across model compression method types.\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The model compression methods mentioned in this survey have been used across a wide variety of text processing tasks. In Table~\\\\ref{tab:applications}, we list down the tasks, popular datasets and references where the readers can find more discussion around model size versus accuracy tradeoff.\",\n \"id\": \"fe1a6e49-c5b1-4394-b0be-1a452c8b0f80\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"d06ca9cc-bcd7-4f53-9be8-ee86f8754664\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Future Directions\"\n ],\n [\n \"subsection\",\n \"Summary of Applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary of Applications\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 849,\n 4536\n ],\n \"content\": \"Although there has been so much of work already in this field, there is a lot more work to be done. \\n\\\\begin{itemize}\\n \\\\item With linear Transformer models, one can afford to have input with tens of thousands of tokens. Hence, many tasks need to be redesigned where large context can now be included as input to improve accuracy. \\n \\\\item Combinations of several methods have not been tested well. Recently, Fastformers method~ showed that combining multiple methods like knowledge distillation, 16-bit quantization, structured pruning and numerical optimizations can lead to drastic improvements. However, lot of experiments are needed to further check how models respond to combination of model compression methods.\\n \\\\item Latency results vary based on GPU architectures. With new GPU architectures (Nvidia RTX 3080, Nvidia T4), some methods like quantization may become more impactful.\\n \\\\item Real world settings are often complex: multi-modal, multi-task, multi-label, small-data, noisy labels, multi-teachers, mismatching teacher-student architectures. Efficient ways of recommending the most promising method is necessary.\\n \\\\item Different components/structures of a model may respond to different kinds of compression methods with specific hyper-parameters. A generic method to choose the right method for various structures is needed.\\n \\\\item How does compression of models impact their interpretability? Can we design model compression mechanisms aimed at looking at a tradeoff between model accuracy, size, latency and interpretability.\\n \\\\item None of the model compression methods performs any application specific compression. Can we obtain further compression by exploiting some task-specific patterns?\\n \\\\item Recently, deep reinforcement learning based methods have been proposed in the computer vision community~. It will be nice to check the effectiveness of such methods for NLP tasks.\\n\\\\end{itemize}\\nWe hope that this survey acts as a good guide to folks across academia and industry. Also, we hope that a significantly large chunk of research gets done in the area of model compression to enable good accuracy across many NLP tasks while keeping model sizes and latencies in check. \\n\\\\begin{thebibliography}{100}\\n\\\\scriptsize\\n\\\\providecommand{\\\\url}[1]{#1}\\n\\\\csname url@samestyle\\\\endcsname\\n\\\\providecommand{\\\\newblock}{\\\\relax}\\n\\\\providecommand{\\\\bibinfo}[2]{#2}\\n\\\\providecommand{\\\\BIBentrySTDinterwordspacing}{\\\\spaceskip=0pt\\\\relax}\\n\\\\providecommand{\\\\BIBentryALTinterwordstretchfactor}{4}\\n\\\\providecommand{\\\\BIBentryALTinterwordspacing}{\\\\spaceskip=\\\\fontdimen2\\\\font plus\\n\\\\BIBentryALTinterwordstretchfactor\\\\fontdimen3\\\\font minus\\n \\\\fontdimen4\\\\font\\\\relax}\\n\\\\providecommand{\\\\BIBforeignlanguage}[2]{{\\n\\\\expandafter\\\\ifx\\\\csname l@#1\\\\endcsname\\\\relax\\n\\\\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}\\n\\\\typeout{** loaded for the language `#1'. Using the pattern for}\\n\\\\typeout{** the default language instead.}\\n\\\\else\\n\\\\language=\\\\csname l@#1\\\\endcsname\\n\\\\fi\\n#2}}\\n\\\\providecommand{\\\\BIBdecl}{\\\\relax}\\n\\\\BIBdecl\\n\\\\bibitem{vaswani2017attention}\\nA.~Vaswani, N.~Shazeer, N.~Parmar, J.~Uszkoreit, L.~Jones, A.~N. Gomez,\\n {\\\\L}.~Kaiser, and I.~Polosukhin, ``Attention is all you need,'' in\\n \\\\emph{NIPS}, 2017, pp. 5998--6008.\\n\\\\bibitem{devlin2018bert}\\nJ.~Devlin, M.-W. Chang, K.~Lee, and K.~Toutanova, ``Bert: Pre-training of deep\\n bidirectional transformers for language understanding,''\\n \\\\emph{arXiv:1810.04805}, 2018.\\n\\\\bibitem{wang2019glue}\\nA.~Wang, A.~Singh, J.~Michael, F.~Hill, O.~Levy, and S.~R. Bowman, ``{GLUE}: A\\n multi-task benchmark and analysis platform for natural language\\n understanding,'' in \\\\emph{ICLR}, 2019.\\n\\\\bibitem{wang2019superglue}\\nA.~Wang, Y.~Pruksachatkun, N.~Nangia, A.~Singh, J.~Michael, F.~Hill, O.~Levy,\\n and S.~R. Bowman, ``Super{GLUE}: A stickier benchmark for general-purpose\\n language understanding systems,'' \\\\emph{1905.00537}, 2019.\\n\\\\bibitem{radford2019language}\\nA.~Radford, J.~Wu, R.~Child, D.~Luan, D.~Amodei, and I.~Sutskever, ``Language\\n models are unsupervised multitask learners,'' \\\\emph{OpenAI Blog}, vol.~1,\\n no.~8, 2019.\\n\\\\bibitem{liu2019multi}\\nX.~Liu, P.~He, W.~Chen, and J.~Gao, ``Multi-task deep neural networks for\\n natural language understanding,'' \\\\emph{arXiv:1901.11504}, 2019.\\n\\\\bibitem{yang2019xlnet}\\nZ.~Yang, Z.~Dai, Y.~Yang, J.~Carbonell, R.~Salakhutdinov, and Q.~V. Le,\\n ``Xlnet: Generalized autoregressive pretraining for language understanding,''\\n \\\\emph{arXiv:1906.08237}, 2019.\\n\\\\bibitem{shoeybi2019megatron}\\nM.~Shoeybi, M.~Patwary, R.~Puri, P.~LeGresley, J.~Casper, and B.~Catanzaro,\\n ``Megatron-lm: Training multi-billion parameter language models using gpu\\n model parallelism,'' \\\\emph{arXiv:1909.08053}, 2019.\\n\\\\bibitem{raffel2019exploring}\\nC.~Raffel, N.~Shazeer, A.~Roberts, K.~Lee, S.~Narang, M.~Matena, Y.~Zhou,\\n W.~Li, and P.~J. Liu, ``Exploring the limits of transfer learning with a\\n unified text-to-text transformer,'' \\\\emph{arXiv:1910.10683}, 2019.\\n\\\\bibitem{rosset2019turing}\\nC.~Rosset, ``Turing-nlg: A 17-billion-parameter language model by microsoft,''\\n \\\\emph{Microsoft Blog}, 2019.\\n\\\\bibitem{lepikhin2020gshard}\\nD.~Lepikhin, H.~Lee, Y.~Xu, D.~Chen, O.~Firat, Y.~Huang, M.~Krikun, N.~Shazeer,\\n and Z.~Chen, ``Gshard: Scaling giant models with conditional computation and\\n automatic sharding,'' \\\\emph{arXiv:2006.16668}, 2020.\\n\\\\bibitem{bianco2018benchmark}\\nS.~Bianco, R.~Cadene, L.~Celona, and P.~Napoletano, ``Benchmark analysis of\\n representative deep neural network architectures,'' \\\\emph{IEEE Access},\\n vol.~6, pp. 64\\\\,270--64\\\\,277, 2018.\\n\\\\bibitem{sanh2019distilbert}\\nV.~Sanh, L.~Debut, J.~Chaumond, and T.~Wolf, ``Distilbert, a distilled version\\n of bert: smaller, faster, cheaper and lighter,'' \\\\emph{arXiv:1910.01108},\\n 2019.\\n\\\\bibitem{han2016eie}\\nS.~Han, X.~Liu, H.~Mao, J.~Pu, A.~Pedram, M.~A. Horowitz, and W.~J. Dally,\\n ``Eie: efficient inference engine on compressed deep neural network,''\\n \\\\emph{ACM SIGARCH Computer Architecture News}, vol.~44, no.~3, pp. 243--254,\\n 2016.\\n\\\\bibitem{diamos2016persistent}\\nG.~Diamos, S.~Sengupta, B.~Catanzaro, M.~Chrzanowski, A.~Coates, E.~Elsen,\\n J.~Engel, A.~Hannun, and S.~Satheesh, ``Persistent rnns: Stashing recurrent\\n weights on-chip,'' in \\\\emph{ICML}, 2016, pp. 2024--2033.\\n\\\\bibitem{denil2013predicting}\\nM.~Denil, B.~Shakibi, L.~Dinh, M.~Ranzato, and N.~De~Freitas, ``Predicting\\n parameters in deep learning,'' in \\\\emph{NIPS}, 2013, pp. 2148--2156.\\n\\\\bibitem{cheng2017survey}\\nY.~Cheng, D.~Wang, P.~Zhou, and T.~Zhang, ``A survey of model compression and\\n acceleration for deep neural networks,'' \\\\emph{arXiv:1710.09282}, 2017.\\n\\\\bibitem{deng2020model}\\nL.~Deng, G.~Li, S.~Han, L.~Shi, and Y.~Xie, ``Model compression and hardware\\n acceleration for neural networks: A comprehensive survey,'' \\\\emph{IEEE}, vol.\\n 108, no.~4, pp. 485--532, 2020.\\n\\\\bibitem{walsh2013peter}\\nC.~A. Walsh, ``Peter huttenlocher (1931--2013),'' \\\\emph{Nature}, vol. 502, no.\\n 7470, pp. 172--172, 2013.\\n\\\\bibitem{zhu2017prune}\\nM.~Zhu and S.~Gupta, ``To prune, or not to prune: exploring the efficacy of\\n pruning for model compression,'' \\\\emph{arXiv:1710.01878}, 2017.\\n\\\\bibitem{lecun1990optimal}\\nY.~LeCun, J.~S. Denker, and S.~A. Solla, ``Optimal brain damage,'' in\\n \\\\emph{NIPS}, 1990, pp. 598--605.\\n\\\\bibitem{hassibi1993second}\\nB.~Hassibi and D.~G. Stork, ``Second order derivatives for network pruning:\\n Optimal brain surgeon,'' in \\\\emph{NIPS}, 1993, pp. 164--171.\\n\\\\bibitem{han2015deep}\\nS.~Han, H.~Mao, and W.~J. Dally, ``Deep compression: Compressing deep neural\\n networks with pruning, trained quantization and huffman coding,''\\n \\\\emph{arXiv:1510.00149}, 2015.\\n\\\\bibitem{see2016compression}\\nA.~See, M.-T. Luong, and C.~D. Manning, ``Compression of neural machine\\n translation models via pruning,'' \\\\emph{arXiv:1606.09274}, 2016.\\n\\\\bibitem{han2015learning}\\nS.~Han, J.~Pool, J.~Tran, and W.~Dally, ``Learning both weights and connections\\n for efficient neural network,'' in \\\\emph{NIPS}, 2015, pp. 1135--1143.\\n\\\\bibitem{narang2017exploring}\\nS.~Narang, E.~Elsen, G.~Diamos, and S.~Sengupta, ``Exploring sparsity in\\n recurrent neural networks,'' \\\\emph{arXiv:1704.05119}, 2017.\\n\\\\bibitem{cheong2019transformers}\\nR.~Cheong and R.~Daniel, ``transformers. zip: Compressing transformers with\\n pruning and quantization,'' Technical report, Stanford University, Stanford,\\n California, 2019., Tech. Rep., 2019.\\n\\\\bibitem{guo2019reweighted}\\nF.-M. Guo, S.~Liu, F.~S. Mungall, X.~Lin, and Y.~Wang, ``Reweighted proximal\\n pruning for large-scale language representation,'' \\\\emph{arXiv:1909.12486},\\n 2019.\\n\\\\bibitem{han2016dsd}\\nS.~Han, J.~Pool, S.~Narang, H.~Mao, E.~Gong, S.~Tang, E.~Elsen, P.~Vajda,\\n M.~Paluri, J.~Tran \\\\emph{et~al.}, ``Dsd: Dense-sparse-dense training for deep\\n neural networks,'' \\\\emph{arXiv:1607.04381}, 2016.\\n\\\\bibitem{dai2018grow}\\nX.~Dai, H.~Yin, and N.~K. Jha, ``Grow and prune compact, fast, and accurate\\n lstms,'' \\\\emph{arXiv:1805.11797}, 2018.\\n\\\\bibitem{he2014reshaping}\\nT.~He, Y.~Fan, Y.~Qian, T.~Tan, and K.~Yu, ``Reshaping deep neural network for\\n fast decoding by node-pruning,'' in \\\\emph{ICASSP}.\\\\hskip 1em plus 0.5em minus\\n 0.4em\\\\relax IEEE, 2014, pp. 245--249.\\n\\\\bibitem{murray2015auto}\\nK.~Murray and D.~Chiang, ``Auto-sizing neural networks: With applications to\\n n-gram language models,'' \\\\emph{arXiv:1508.05051}, 2015.\\n\\\\bibitem{pan2016dropneuron}\\nW.~Pan, H.~Dong, and Y.~Guo, ``Dropneuron: Simplifying the structure of deep\\n neural networks,'' \\\\emph{arXiv:1606.07326}, 2016.\\n\\\\bibitem{srinivas2015data}\\nS.~Srinivas and R.~V. Babu, ``Data-free parameter pruning for deep neural\\n networks,'' \\\\emph{arXiv:1507.06149}, 2015.\\n\\\\bibitem{narang2017block}\\nS.~Narang, E.~Undersander, and G.~Diamos, ``Block-sparse recurrent neural\\n networks,'' \\\\emph{arXiv:1711.02782}, 2017.\\n\\\\bibitem{cao2019efficient}\\nS.~Cao, C.~Zhang, Z.~Yao, W.~Xiao, L.~Nie, D.~Zhan, Y.~Liu, M.~Wu, and\\n L.~Zhang, ``Efficient and effective sparse lstm on fpga with bank-balanced\\n sparsity,'' in \\\\emph{SIGDA Intl. Symp. on FPGA}.\\\\hskip 1em plus 0.5em minus\\n 0.4em\\\\relax ACM, 2019, pp. 63--72.\\n\\\\bibitem{michel2019sixteen}\\nP.~Michel, O.~Levy, and G.~Neubig, ``Are sixteen heads really better than\\n one?'' \\\\emph{arXiv:1905.10650}, 2019.\\n\\\\bibitem{voita2019analyzing}\\nE.~Voita, D.~Talbot, F.~Moiseev, R.~Sennrich, and I.~Titov, ``Analyzing\\n multi-head self-attention: Specialized heads do the heavy lifting, the rest\\n can be pruned,'' \\\\emph{arXiv:1905.09418}, 2019.\\n\\\\bibitem{ding2017visualizing}\\nY.~Ding, Y.~Liu, H.~Luan, and M.~Sun, ``Visualizing and understanding neural\\n machine translation,'' in \\\\emph{ACL}, 2017, pp. 1150--1159.\\n\\\\bibitem{fan2019reducing}\\nA.~Fan, E.~Grave, and A.~Joulin, ``Reducing transformer depth on demand with\\n structured dropout,'' \\\\emph{arXiv:1909.11556}, 2019.\\n\\\\bibitem{prasanna2020bert}\\nS.~Prasanna, A.~Rogers, and A.~Rumshisky, ``When bert plays the lottery, all\\n tickets are winning,'' \\\\emph{arXiv:2005.00561}, 2020.\\n\\\\bibitem{bartol2015hippocampal}\\nT.~M. Bartol, C.~Bromer, J.~Kinney, M.~A. Chirillo, J.~N. Bourne, K.~M. Harris,\\n and T.~J. Sejnowski, ``Hippocampal spine head sizes are highly precise,''\\n \\\\emph{bioRxiv}, p. 016329, 2015.\\n\\\\bibitem{linden2018think}\\nD.~J. Linden, \\\\emph{Think Tank: Forty Neuroscientists Explore the Biological\\n Roots of Human Experience}.\\\\hskip 1em plus 0.5em minus 0.4em\\\\relax Yale\\n University Press, 2018.\\n\\\\bibitem{hubara2017quantized}\\nI.~Hubara, M.~Courbariaux, D.~Soudry, R.~El-Yaniv, and Y.~Bengio, ``Quantized\\n neural networks: Training neural networks with low precision weights and\\n activations,'' \\\\emph{JMLR}, vol.~18, no.~1, pp. 6869--6898, 2017.\\n\\\\bibitem{lam2018word2bits}\\nM.~Lam, ``Word2bits-quantized word vectors,'' \\\\emph{arXiv:1803.05651}, 2018.\\n\\\\bibitem{courbariaux2015binaryconnect}\\nM.~Courbariaux, Y.~Bengio, and J.-P. David, ``Binaryconnect: Training deep\\n neural networks with binary weights during propagations,'' in \\\\emph{NIPS},\\n 2015, pp. 3123--3131.\\n\\\\bibitem{bengio2013estimating}\\nY.~Bengio, N.~L{\\\\'e}onard, and A.~Courville, ``Estimating or propagating\\n gradients through stochastic neurons for conditional computation,''\\n \\\\emph{arXiv:1308.3432}, 2013.\\n\\\\bibitem{lin2015neural}\\nZ.~Lin, M.~Courbariaux, R.~Memisevic, and Y.~Bengio, ``Neural networks with few\\n multiplications,'' \\\\emph{arXiv:1510.03009}, 2015.\\n\\\\bibitem{hubara2016binarized}\\nI.~Hubara, M.~Courbariaux, D.~Soudry, R.~El-Yaniv, and Y.~Bengio, ``Binarized\\n neural networks,'' in \\\\emph{NIPS}, 2016, pp. 4107--4115.\\n\\\\bibitem{rastegari2016xnor}\\nM.~Rastegari, V.~Ordonez, J.~Redmon, and A.~Farhadi, ``Xnor-net: Imagenet\\n classification using binary convolutional neural networks,'' in\\n \\\\emph{ECCV}.\\\\hskip 1em plus 0.5em minus 0.4em\\\\relax Springer, 2016, pp.\\n 525--542.\\n\\\\bibitem{hou2016loss}\\nL.~Hou, Q.~Yao, and J.~T. Kwok, ``Loss-aware binarization of deep networks,''\\n \\\\emph{arXiv:1611.01600}, 2016.\\n\\\\bibitem{lee2014proximal}\\nJ.~D. Lee, Y.~Sun, and M.~A. Saunders, ``Proximal newton-type methods for\\n minimizing composite functions,'' \\\\emph{J. Optimization}, vol.~24, no.~3, pp.\\n 1420--1443, 2014.\\n\\\\bibitem{ott2016recurrent}\\nJ.~Ott, Z.~Lin, Y.~Zhang, S.-C. Liu, and Y.~Bengio, ``Recurrent neural networks\\n with limited numerical precision,'' \\\\emph{arXiv:1608.06902}, 2016.\\n\\\\bibitem{alom2018effective}\\nM.~Z. Alom, A.~T. Moody, N.~Maruyama, B.~C. Van~Essen, and T.~M. Taha,\\n ``Effective quantization approaches for recurrent neural networks,'' in\\n \\\\emph{IJCNN}.\\\\hskip 1em plus 0.5em minus 0.4em\\\\relax IEEE, 2018, pp. 1--8.\\n\\\\bibitem{li2016ternary}\\nF.~Li, B.~Zhang, and B.~Liu, ``Ternary weight networks,''\\n \\\\emph{arXiv:1605.04711}, 2016.\\n\\\\bibitem{hwang2014fixed}\\nK.~Hwang and W.~Sung, ``Fixed-point feedforward deep neural network design\\n using weights+ 1, 0, and- 1,'' in \\\\emph{SiPS}.\\\\hskip 1em plus 0.5em minus\\n 0.4em\\\\relax IEEE, 2014, pp. 1--6.\\n\\\\bibitem{lloyd1982least}\\nS.~Lloyd, ``Least squares quantization in pcm,'' \\\\emph{Tran. on information\\n theory}, vol.~28, no.~2, pp. 129--137, 1982.\\n\\\\bibitem{zhu2016trained}\\nC.~Zhu, S.~Han, H.~Mao, and W.~J. Dally, ``Trained ternary quantization,''\\n \\\\emph{arXiv:1612.01064}, 2016.\\n\\\\bibitem{wang2018hitnet}\\nP.~Wang, X.~Xie, L.~Deng, G.~Li, D.~Wang, and Y.~Xie, ``Hitnet: hybrid ternary\\n recurrent neural network,'' in \\\\emph{NIPS}, 2018, pp. 604--614.\\n\\\\bibitem{he2016effective}\\nQ.~He, H.~Wen, S.~Zhou, Y.~Wu, C.~Yao, X.~Zhou, and Y.~Zou, ``Effective\\n quantization methods for recurrent neural networks,''\\n \\\\emph{arXiv:1611.10176}, 2016.\\n\\\\bibitem{zhou2017balanced}\\nS.-C. Zhou, Y.-Z. Wang, H.~Wen, Q.-Y. He, and Y.-H. Zou, ``Balanced\\n quantization: An effective and efficient approach to quantized neural\\n networks,'' \\\\emph{J. of Computer Science and Technology}, vol.~32, no.~4, pp.\\n 667--682, 2017.\\n\\\\bibitem{muller2015rounding}\\nL.~K. Muller and G.~Indiveri, ``Rounding methods for neural networks with low\\n resolution synaptic weights,'' \\\\emph{arXiv:1504.05767}, 2015.\\n\\\\bibitem{gong2014compressing}\\nY.~Gong, L.~Liu, M.~Yang, and L.~Bourdev, ``Compressing deep convolutional\\n networks using vector quantization,'' \\\\emph{arXiv:1412.6115}, 2014.\\n\\\\bibitem{guo2017network}\\nY.~Guo, A.~Yao, H.~Zhao, and Y.~Chen, ``Network sketching: Exploiting binary\\n structure in deep cnns,'' in \\\\emph{CVPR}, 2017, pp. 5955--5963.\\n\\\\bibitem{xu2018alternating}\\nC.~Xu, J.~Yao, Z.~Lin, W.~Ou, Y.~Cao, Z.~Wang, and H.~Zha, ``Alternating\\n multi-bit quantization for recurrent neural networks,''\\n \\\\emph{arXiv:1802.00150}, 2018.\\n\\\\bibitem{mikolov2013word2vec}\\nT.~Mikolov, K.~Chen, G.~Corrado, J.~Dean, L.~Sutskever, and G.~Zweig,\\n ``word2vec,'' \\\\emph{URL https://code. google. com/p/word2vec}, vol.~22, 2013.\\n\\\\bibitem{shen2019q}\\nS.~Shen, Z.~Dong, J.~Ye, L.~Ma, Z.~Yao, A.~Gholami, M.~W. Mahoney, and\\n K.~Keutzer, ``Q-bert: Hessian based ultra low precision quantization of\\n bert,'' \\\\emph{arXiv:1909.05840}, 2019.\\n\\\\bibitem{ba2014deep}\\nJ.~Ba and R.~Caruana, ``Do deep nets really need to be deep?'' in \\\\emph{NIPS},\\n 2014, pp. 2654--2662.\\n\\\\bibitem{sau2016deep}\\nB.~B. Sau and V.~N. Balasubramanian, ``Deep model compression: Distilling\\n knowledge from noisy teachers,'' \\\\emph{arXiv:1610.09650}, 2016.\\n\\\\bibitem{hinton2015distilling}\\nG.~Hinton, O.~Vinyals, and J.~Dean, ``Distilling the knowledge in a neural\\n network,'' \\\\emph{arXiv:1503.02531}, 2015.\\n\\\\bibitem{czarnecki2017sobolev}\\nW.~M. Czarnecki, S.~Osindero, M.~Jaderberg, G.~Swirszcz, and R.~Pascanu,\\n ``Sobolev training for neural networks,'' in \\\\emph{NIPS}, 2017, pp.\\n 4278--4287.\\n\\\\bibitem{mishra2017apprentice}\\nA.~Mishra and D.~Marr, ``Apprentice: Using knowledge distillation techniques to\\n improve low-precision network accuracy,'' \\\\emph{arXiv:1711.05852}, 2017.\\n\\\\bibitem{polino2018model}\\nA.~Polino, R.~Pascanu, and D.~Alistarh, ``Model compression via distillation\\n and quantization,'' \\\\emph{arXiv:1802.05668}, 2018.\\n\\\\bibitem{romero2014fitnets}\\nA.~Romero, N.~Ballas, S.~E. Kahou, A.~Chassang, C.~Gatta, and Y.~Bengio,\\n ``Fitnets: Hints for thin deep nets,'' \\\\emph{arXiv:1412.6550}, 2014.\\n\\\\bibitem{yim2017gift}\\nJ.~Yim, D.~Joo, J.~Bae, and J.~Kim, ``A gift from knowledge distillation: Fast\\n optimization, network minimization and transfer learning,'' in \\\\emph{CVPR},\\n 2017, pp. 4133--4141.\\n\\\\bibitem{mcclure2016representational}\\nP.~McClure and N.~Kriegeskorte, ``Representational distance learning for deep\\n neural networks,'' \\\\emph{Frontiers in computational neuroscience}, vol.~10,\\n p. 131, 2016.\\n\\\\bibitem{kim2016sequence}\\nY.~Kim and A.~M. Rush, ``Sequence-level knowledge distillation,''\\n \\\\emph{arXiv:1606.07947}, 2016.\\n\\\\bibitem{freitag2017ensemble}\\nM.~Freitag, Y.~Al-Onaizan, and B.~Sankaran, ``Ensemble distillation for neural\\n machine translation,'' \\\\emph{arXiv:1702.01802}, 2017.\\n\\\\bibitem{zhang2018deep}\\nY.~Zhang, T.~Xiang, T.~M. Hospedales, and H.~Lu, ``Deep mutual learning,'' in\\n \\\\emph{CVPR}, 2018, pp. 4320--4328.\\n\\\\bibitem{anil2018large}\\nR.~Anil, G.~Pereyra, A.~Passos, R.~Ormandi, G.~E. Dahl, and G.~E. Hinton,\\n ``Large scale distributed neural network training through online\\n distillation,'' \\\\emph{arXiv:1804.03235}, 2018.\\n\\\\bibitem{mirzadeh2019improved}\\nS.-I. Mirzadeh, M.~Farajtabar, A.~Li, N.~Levine, A.~Matsukawa, and\\n H.~Ghasemzadeh, ``Improved knowledge distillation via teacher assistant,''\\n \\\\emph{arXiv:1902.03393}, 2019.\\n\\\\bibitem{you2017learning}\\nS.~You, C.~Xu, C.~Xu, and D.~Tao, ``Learning from multiple teacher networks,''\\n in \\\\emph{KDD}, 2017, pp. 1285--1294.\\n\\\\bibitem{tan2019multilingual}\\nX.~Tan, Y.~Ren, D.~He, T.~Qin, Z.~Zhao, and T.-Y. Liu, ``Multilingual neural\\n machine translation with knowledge distillation,'' \\\\emph{arXiv:1902.10461},\\n 2019.\\n\\\\bibitem{heo2019knowledge}\\nB.~Heo, M.~Lee, S.~Yun, and J.~Y. Choi, ``Knowledge distillation with\\n adversarial samples supporting decision boundary,'' in \\\\emph{AAAI}, vol.~33,\\n 2019, pp. 3771--3778.\\n\\\\bibitem{xu2017training}\\nZ.~Xu, Y.-C. Hsu, and J.~Huang, ``Training shallow and thin networks for\\n acceleration via knowledge distillation with conditional adversarial\\n networks,'' \\\\emph{arXiv:1709.00513}, 2017.\\n\\\\bibitem{wang2018kdgan}\\nX.~Wang, R.~Zhang, Y.~Sun, and J.~Qi, ``Kdgan: Knowledge distillation with\\n generative adversarial networks,'' in \\\\emph{NIPS}, 2018, pp. 775--786.\\n\\\\bibitem{zhao2019extreme}\\nS.~Zhao, R.~Gupta, Y.~Song, and D.~Zhou, ``Extreme language model compression\\n with optimal subwords and shared projections,'' \\\\emph{arXiv:1909.11687},\\n 2019.\\n\\\\bibitem{sun2019patient}\\nS.~Sun, Y.~Cheng, Z.~Gan, and J.~Liu, ``Patient knowledge distillation for bert\\n model compression,'' \\\\emph{arXiv:1908.09355}, 2019.\\n\\\\bibitem{jiao2019tinybert}\\nX.~Jiao, Y.~Yin, L.~Shang, X.~Jiang, X.~Chen, L.~Li, F.~Wang, and Q.~Liu,\\n ``Tinybert: Distilling bert for natural language understanding,''\\n \\\\emph{arXiv:1909.10351}, 2019.\\n\\\\bibitem{liu2019improving}\\nX.~Liu, P.~He, W.~Chen, and J.~Gao, ``Improving multi-task deep neural networks\\n via knowledge distillation for natural language understanding,''\\n \\\\emph{arXiv:1904.09482}, 2019.\\n\\\\bibitem{wang2020minilm}\\nW.~Wang, F.~Wei, L.~Dong, H.~Bao, N.~Yang, and M.~Zhou, ``Minilm: Deep\\n self-attention distillation for task-agnostic compression of pre-trained\\n transformers,'' \\\\emph{arXiv:2002.10957}, 2020.\\n\\\\bibitem{tang2019distilling}\\nR.~Tang, Y.~Lu, L.~Liu, L.~Mou, O.~Vechtomova, and J.~Lin, ``Distilling\\n task-specific knowledge from bert into simple neural networks,''\\n \\\\emph{arXiv:1903.12136}, 2019.\\n\\\\bibitem{mukherjee2020xtremedistil}\\nS.~Mukherjee and A.~H. Awadallah, ``Xtremedistil: Multi-stage distillation for\\n massive multilingual models,'' in \\\\emph{ACL}, 2020, pp. 2221--2234.\\n\\\\bibitem{ling2015finding}\\nW.~Ling, T.~Lu{\\\\'\\\\i}s, L.~Marujo, R.~F. Astudillo, S.~Amir, C.~Dyer, A.~W.\\n Black, and I.~Trancoso, ``Finding function in form: Compositional character\\n models for open vocabulary word representation,'' \\\\emph{arXiv:1508.02096},\\n 2015.\\n\\\\bibitem{jozefowicz2016exploring}\\nR.~Jozefowicz, O.~Vinyals, M.~Schuster, N.~Shazeer, and Y.~Wu, ``Exploring the\\n limits of language modeling,'' \\\\emph{arXiv:1602.02410}, 2016.\\n\\\\bibitem{kim2016character}\\nY.~Kim, Y.~Jernite, D.~Sontag, and A.~M. Rush, ``Character-aware neural\\n language models,'' in \\\\emph{AAAI}, 2016.\\n\\\\bibitem{chen2015compressing}\\nW.~Chen, J.~Wilson, S.~Tyree, K.~Weinberger, and Y.~Chen, ``Compressing neural\\n networks with the hashing trick,'' in \\\\emph{ICML}, 2015, pp. 2285--2294.\\n\\\\bibitem{lu2016learning}\\nZ.~Lu, V.~Sindhwani, and T.~N. Sainath, ``Learning compact recurrent neural\\n networks,'' in \\\\emph{ICASSP}.\\\\hskip 1em plus 0.5em minus 0.4em\\\\relax IEEE,\\n 2016, pp. 5960--5964.\\n\\\\bibitem{tay2019lightweight}\\nY.~Tay, A.~Zhang, L.~A. Tuan, J.~Rao, S.~Zhang, S.~Wang, J.~Fu, and S.~C. Hui,\\n ``Lightweight and efficient neural natural language processing with\\n quaternion networks,'' \\\\emph{arXiv:1906.04393}, 2019.\\n\\\\bibitem{chen2016compressing}\\nY.~Chen, L.~Mou, Y.~Xu, G.~Li, and Z.~Jin, ``Compressing neural language models\\n by sparse word representations,'' \\\\emph{arXiv:1610.03950}, 2016.\\n\\\\bibitem{li2016lightrnn}\\nX.~Li, T.~Qin, J.~Yang, and T.-Y. Liu, ``Lightrnn: Memory and\\n computation-efficient recurrent neural networks,'' in \\\\emph{NIPS}, 2016, pp.\\n 4385--4393.\\n\\\\bibitem{suzuki2016learning}\\nJ.~Suzuki and M.~Nagata, ``Learning compact neural word embeddings by parameter\\n space sharing.'' in \\\\emph{IJCAI}, 2016, pp. 2046--2052.\\n\\\\bibitem{li2018slim}\\nZ.~Li, R.~Kulhanek, S.~Wang, Y.~Zhao, and S.~Wu, ``Slim embedding layers for\\n recurrent neural language models,'' in \\\\emph{AAAI}, 2018.\\n\\\\bibitem{lan2019albert}\\nZ.~Lan, M.~Chen, S.~Goodman, K.~Gimpel, P.~Sharma, and R.~Soricut, ``Albert: A\\n lite bert for self-supervised learning of language representations,''\\n \\\\emph{arXiv:1909.11942}, 2019.\\n\\\\bibitem{dehghani2018universal}\\nM.~Dehghani, S.~Gouws, O.~Vinyals, J.~Uszkoreit, and {\\\\L}.~Kaiser, ``Universal\\n transformers,'' \\\\emph{arXiv:1807.03819}, 2018.\\n\\\\bibitem{bai2019deep}\\nS.~Bai, J.~Z. Kolter, and V.~Koltun, ``Deep equilibrium models,''\\n \\\\emph{arXiv:1909.01377}, 2019.\\n\\\\bibitem{oseledets2011tensor}\\nI.~V. Oseledets, ``Tensor-train decomposition,'' \\\\emph{SIAM J. on Scientific\\n Computing}, vol.~33, no.~5, pp. 2295--2317, 2011.\\n\\\\bibitem{carroll1970analysis}\\nJ.~D. Carroll and J.-J. Chang, ``Analysis of individual differences in\\n multidimensional scaling via an n-way generalization of “eckart-young”\\n decomposition,'' \\\\emph{Psychometrika}, vol.~35, no.~3, pp. 283--319, 1970.\\n\\\\bibitem{tucker1966some}\\nL.~R. Tucker, ``Some mathematical notes on three-mode factor analysis,''\\n \\\\emph{Psychometrika}, vol.~31, no.~3, pp. 279--311, 1966.\\n\\\\bibitem{prabhavalkar2016compression}\\nR.~Prabhavalkar, O.~Alsharif, A.~Bruguier, and L.~McGraw, ``On the compression\\n of recurrent neural networks with an application to lvcsr acoustic modeling\\n for embedded speech recognition,'' in \\\\emph{ICASSP}.\\\\hskip 1em plus 0.5em\\n minus 0.4em\\\\relax IEEE, 2016, pp. 5970--5974.\\n\\\\bibitem{sak2014long}\\nH.~Sak, A.~Senior, and F.~Beaufays, ``Long short-term memory based recurrent\\n neural network architectures for large vocabulary speech recognition,''\\n \\\\emph{arXiv:1402.1128}, 2014.\\n\\\\bibitem{faruqui2015sparse}\\nM.~Faruqui, Y.~Tsvetkov, D.~Yogatama, C.~Dyer, and N.~Smith, ``Sparse\\n overcomplete word vector representations,'' \\\\emph{arXiv:1506.02004}, 2015.\\n\\\\bibitem{shu2017compressing}\\nR.~Shu and H.~Nakayama, ``Compressing word embeddings via deep compositional\\n code learning,'' \\\\emph{arXiv:1711.01068}, 2017.\\n\\\\bibitem{wang2019structured}\\nZ.~Wang, J.~Wohlwend, and T.~Lei, ``Structured pruning of large language\\n models,'' \\\\emph{arXiv:1910.04732}, 2019.\\n\\\\bibitem{chen2015strategies}\\nW.~Chen, D.~Grangier, and M.~Auli, ``Strategies for training large vocabulary\\n neural language models,'' \\\\emph{arXiv:1512.04906}, 2015.\\n\\\\bibitem{variani2019west}\\nE.~Variani, A.~T. Suresh, and M.~Weintraub, ``West: Word encoded sequence\\n transducers,'' in \\\\emph{ICASSP}.\\\\hskip 1em plus 0.5em minus 0.4em\\\\relax IEEE,\\n 2019, pp. 7340--7344.\\n\\\\bibitem{tjandra2017compressing}\\nA.~Tjandra, S.~Sakti, and S.~Nakamura, ``Compressing recurrent neural network\\n with tensor train,'' in \\\\emph{IJCNN}.\\\\hskip 1em plus 0.5em minus 0.4em\\\\relax\\n IEEE, 2017, pp. 4451--4458.\\n\\\\bibitem{khrulkov2019tensorized}\\nV.~Khrulkov, O.~Hrinchuk, L.~Mirvakhabova, and I.~Oseledets, ``Tensorized\\n embedding layers for efficient model compression,'' \\\\emph{arXiv:1901.10787},\\n 2019.\\n\\\\bibitem{ye2018learning}\\nJ.~Ye, L.~Wang, G.~Li, D.~Chen, S.~Zhe, X.~Chu, and Z.~Xu, ``Learning compact\\n recurrent neural networks with block-term tensor decomposition,'' in\\n \\\\emph{CVPR}, 2018, pp. 9378--9387.\\n\\\\bibitem{ma2019tensorized}\\nX.~Ma, P.~Zhang, S.~Zhang, N.~Duan, Y.~Hou, D.~Song, and M.~Zhou, ``A\\n tensorized transformer for language modeling,'' \\\\emph{arXiv:1906.09777},\\n 2019.\\n\\\\bibitem{child2019generating}\\nR.~Child, S.~Gray, A.~Radford, and I.~Sutskever, ``Generating long sequences\\n with sparse transformers,'' \\\\emph{arXiv:1904.10509}, 2019.\\n\\\\bibitem{guo2019star}\\nQ.~Guo, X.~Qiu, P.~Liu, Y.~Shao, X.~Xue, and Z.~Zhang, ``Star-transformer,'' in\\n \\\\emph{NAACL-HLT}, 2019, pp. 1315--1325.\\n\\\\bibitem{kitaev2020reformer}\\nN.~Kitaev, {\\\\L}.~Kaiser, and A.~Levskaya, ``Reformer: The efficient\\n transformer,'' \\\\emph{arXiv:2001.04451}, 2020.\\n\\\\bibitem{wang2020linformer}\\nS.~Wang, B.~Li, M.~Khabsa, H.~Fang, and H.~Ma, ``Linformer: Self-attention with\\n linear complexity,'' \\\\emph{arXiv:2006.04768}, 2020.\\n\\\\bibitem{tay2020sparse}\\nY.~Tay, D.~Bahri, L.~Yang, D.~Metzler, and D.-C. Juan, ``Sparse sinkhorn\\n attention,'' \\\\emph{arXiv:2002.11296}, 2020.\\n\\\\bibitem{shen2018efficient}\\nZ.~Shen, M.~Zhang, H.~Zhao, S.~Yi, and H.~Li, ``Efficient attention: Attention\\n with linear complexities,'' \\\\emph{arXiv:1812.01243}, 2018.\\n\\\\bibitem{katharopoulos2020transformers}\\nA.~Katharopoulos, A.~Vyas, N.~Pappas, and F.~Fleuret, ``Transformers are rnns:\\n Fast autoregressive transformers with linear attention,''\\n \\\\emph{arXiv:2006.16236}, 2020.\\n\\\\bibitem{kapur2017low}\\nS.~Kapur, A.~Mishra, and D.~Marr, ``Low precision rnns: Quantizing rnns without\\n losing accuracy,'' \\\\emph{arXiv:1710.07706}, 2017.\\n\\\\bibitem{hou2018loss}\\nL.~Hou and J.~T. Kwok, ``Loss-aware weight quantization of deep networks,''\\n \\\\emph{arXiv:1802.08635}, 2018.\\n\\\\bibitem{grachev2019compression}\\nA.~M. Grachev, D.~I. Ignatov, and A.~V. Savchenko, ``Compression of recurrent\\n neural networks for efficient language modeling,'' \\\\emph{Applied Soft\\n Computing}, vol.~79, pp. 354--362, 2019.\\n\\\\bibitem{damani20_pakdd}\\nS.~Damani, K.~N. Narahari, A.~Chatterjee, M.~Gupta, and P.~Agrawal, ``Optimized\\n transformer models for faq answering,'' in \\\\emph{PAKDD}, 2020, p. To appear.\\n\\\\bibitem{yang2018accelerating}\\nY.~Yang, K.~Liang, X.~Xiao, Z.~Xie, L.~Jin, J.~Sun, and W.~Zhou, ``Accelerating\\n and compressing lstm based model for online handwritten chinese character\\n recognition,'' in \\\\emph{ICFHR}.\\\\hskip 1em plus 0.5em minus 0.4em\\\\relax IEEE,\\n 2018, pp. 110--115.\\n\\\\bibitem{lazaridou2013fish}\\nA.~Lazaridou, E.~M. Vecchi, and M.~Baroni, ``Fish transporters and miracle\\n homes: How compositional distributional semantics can help np parsing,'' in\\n \\\\emph{EMNLP}, 2013, pp. 1908--1913.\\n\\\\bibitem{wangperawong2018attending}\\nA.~Wangperawong, ``Attending to mathematical language with transformers,''\\n \\\\emph{arXiv:1812.02825}, 2018.\\n\\\\bibitem{linzen2016assessing}\\nT.~Linzen, E.~Dupoux, and Y.~Goldberg, ``Assessing the ability of lstms to\\n learn syntax-sensitive dependencies,'' \\\\emph{TACL}, vol.~4, pp. 521--535,\\n 2016.\\n\\\\bibitem{zaremba2014learning}\\nW.~Zaremba and I.~Sutskever, ``Learning to execute,'' \\\\emph{arXiv:1410.4615},\\n 2014.\\n\\\\end{thebibliography}\\n\\\\bibliographystyle{ACM-Reference-Format}\\n\\\\bibliography{referencesShort}\\n\\\\end{document}\\n\\\\endinput\",\n \"id\": \"7ea3f834-f595-4c1b-afed-044d53e835ea\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"d06ca9cc-bcd7-4f53-9be8-ee86f8754664\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Compression of Deep Learning Models for Text: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Future Directions\"\n ],\n [\n \"subsection\",\n \"Future Trends\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Future Trends\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":77,"string":"77"},"all_cites":{"kind":"list like","value":[844,4478,4331,4479,4480,4481,4482,4483,7826,4485,4484,4486,4487,1568,4351,685,4488,8375,2670,4489,4490,851,4491,4492,852,4493,8797,4494,4496,4495,4498,4497,2488,8798,4499,4500,4503,4504,681,8799,8390,4501,4502,4506,4505,4508,4507,4513,4509,4511,2481,4512,4514,4510,854,4515,1150,9120,856,7333,7,4516,4517,4036,4522,4520,4518,4521,4519,4523,4524,4525,4526,8800,4529,4527,4530,4528,4531,793,794,7298,798,7371,1473,7827,4536,4533,7829,4532,4534,4535,7828,849],"string":"[\n 844,\n 4478,\n 4331,\n 4479,\n 4480,\n 4481,\n 4482,\n 4483,\n 7826,\n 4485,\n 4484,\n 4486,\n 4487,\n 1568,\n 4351,\n 685,\n 4488,\n 8375,\n 2670,\n 4489,\n 4490,\n 851,\n 4491,\n 4492,\n 852,\n 4493,\n 8797,\n 4494,\n 4496,\n 4495,\n 4498,\n 4497,\n 2488,\n 8798,\n 4499,\n 4500,\n 4503,\n 4504,\n 681,\n 8799,\n 8390,\n 4501,\n 4502,\n 4506,\n 4505,\n 4508,\n 4507,\n 4513,\n 4509,\n 4511,\n 2481,\n 4512,\n 4514,\n 4510,\n 854,\n 4515,\n 1150,\n 9120,\n 856,\n 7333,\n 7,\n 4516,\n 4517,\n 4036,\n 4522,\n 4520,\n 4518,\n 4521,\n 4519,\n 4523,\n 4524,\n 4525,\n 4526,\n 8800,\n 4529,\n 4527,\n 4530,\n 4528,\n 4531,\n 793,\n 794,\n 7298,\n 798,\n 7371,\n 1473,\n 7827,\n 4536,\n 4533,\n 7829,\n 4532,\n 4534,\n 4535,\n 7828,\n 849\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.403818468826191,"string":"1.403818"}}},{"rowIdx":43,"cells":{"authors":{"kind":"list like","value":["Akshay L Chandra","Sai Vikas Desai","Wei Guo","Vineeth N Balasubramanian"],"string":"[\n \"Akshay L Chandra\",\n \"Sai Vikas Desai\",\n \"Wei Guo\",\n \"Vineeth N Balasubramanian\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-06-18T14:21:19Z","string":"2020-06-18T14:21:19Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" In light of growing challenges in agriculture with ever growing food demand across the world, efficient crop management techniques are necessary to increase crop yield. Precision agriculture techniques allow the stakeholders to make effective and customized crop management decisions based on data gathered from monitoring crop environments. Plant phenotyping techniques play a major role in accurate crop monitoring. Advancements in deep learning have made previously difficult phenotyping tasks possible. This survey aims to introduce the reader to the state of the art research in deep plant phenotyping. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"40b8b988-fe57-4e55-9221-27ac302d3a6f","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"]],"subsections":["50e565bb-2ea0-4b95-baa0-7a3211a9fe18","61c27a51-9e54-4459-89d0-6fb5a235e790","be6f63d3-6b39-458c-918e-c883b38ba7a8","fcb14fe1-1d8d-475b-8ed7-cf7547edb6ff","6bf986f3-5fd9-4afa-a45b-7964c94d86eb","1cab82e7-7d3c-4e75-913b-94d66346fc31"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"Population growth, increasing incomes, and rapid urbanization in developing countries are expected to cause a drastic hike in food demand. This projected rise in food demand poses several challenges to agriculture. Owing to a continuous decline in global cultivable land , increasing the productivity of the existing agricultural land is highly necessary. This need has led to the scientific community focusing their efforts on developing efficient and sustainable ways to increase crop yield. To this end, precision agriculture techniques have attracted a lot of attention. Precision agriculture is a set of methods to monitor crops, gather data, and carry out informed crop management tasks such as applying the optimum amount of water, selecting suitable pesticides, and reducing environmental impact. These methods involve the usage of specialized devices such as sensors, UAVs, and static cameras to monitor the crops. Accurate crop monitoring goes a long way in assisting farmers in making the right choices to obtain the maximum yield. Plant phenotyping, a rapidly emerging research area, plays a significant role in understanding crop-related traits. Plant phenotyping is the science of characterizing and quantifying the physical and physiological traits of a plant. It provides a quantitative assessment of the plant's properties and its behavior in various environmental conditions. Understanding these properties is crucial in performing effective crop management. \nResearch in plant phenotyping has grown rapidly thanks to the availability of cost-effective and easy to use digital imaging devices such as RGB, multispectral, and hyperspectral cameras, which have facilitated the collection of large amounts of data. This influx of data coupled with the usage of machine learning algorithms has fueled the development of various high throughput phenotyping tools [refs] for tasks such as weed detection, fruit/organ counting, disease detection and yield estimation. A machine learning pipeline typically consists of feature extraction followed by a classification/regression module for prediction. While machine learning techniques have helped build sophisticated phenotyping tools, they are known to lack robustness. They rely heavily on handcrafted feature extraction techniques and manual hyperparameter tuning methods. As a result, if feature extraction is not carefully done under a domain expert's supervision, they tend to perform poorly in uncontrolled environments such as agricultural fields where factors such as lighting, weather, exposure, etc. often cannot be regulated. Hence, feature extraction from data has been one of the major bottlenecks in the development of efficient high throughput plant phenotyping systems.\nAdvancements in deep learning, a sub-field of machine learning which allows for automatic feature extraction and prediction on large scale data, has led to a surge in the development of visual plant phenotyping methods. Deep learning is particularly well-known for its effectiveness in handling vision-based tasks such as image classification, object detection, semantic segmentation, and scene understanding. Coincidentally, many of these tasks form the backbone for various plant phenotyping tasks such as disease detection, fruit detection, and yield estimation. Fig. \\ref{fig-ml-dl} illustrates the difference between machine learning based plant phenotyping and deep learning based plant phenotyping. We believe that the expressive power and robustness of deep learning systems can be effectively leveraged by plant researchers to identify complex patterns from raw data and devise efficient precision agriculture methodologies. The purpose of this survey is to enable the readers to get a bird's eye view of the advancements in the field of deep learning based plant phenotyping, understand the existing issues, and become familiar with some of the open problems which warrant further research.\n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.7]{figures/ml-dl.pdf}}\n\\caption{Difference between ML and DL based Plant Phenotyping.}\n\\label{fig-ml-dl}\n\\end{figure}","id":"50e565bb-2ea0-4b95-baa0-7a3211a9fe18","level":"section","origin_cites_number":5,"parent_id":"40b8b988-fe57-4e55-9221-27ac302d3a6f","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"","id":"61c27a51-9e54-4459-89d0-6fb5a235e790","level":"section","origin_cites_number":0,"parent_id":"40b8b988-fe57-4e55-9221-27ac302d3a6f","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Background"]],"subsections":["3bee11d8-8a06-40e7-a105-9445d5a61266","cd19296d-3914-418b-a3bf-11f032138dc2","ad075082-3735-469d-88a5-712030a4f4ec"],"title":"Background"},{"cite_extract_rate":0,"cites":[],"content":"Plant phenotyping is the science of quantifying the physical and physiological traits of a plant. Plant phenotyping mainly benefits two communities: farmers and plant breeders. By better understanding the traits of the crop, a farmer can optimize crop yield by making informed crop management decisions. Similarly, understanding the crop's behavior is crucial for plant breeders to select the best possible crop variety for a given location and environment. In the past, plant phenotyping was a manual endeavor. The process of manually observing a small set of crop samples and reporting observations periodically was slow, labor intensive and inefficient. The low throughput nature of these methods has impeded the progress in plant breeding research. However, the advent of modern data acquisition methods with various sensors, cameras and UAVs (Unmanned Aerial Vehicles) coupled with advances in machine learning techniques have resulted in the development of high-throughput plant phenotyping methods to be effectively used for precision agriculture. \nDepending on the method of data collection, plant phenotyping techniques can be classified into ground based, aerial and satellite based methods. In ground based phenotyping, high precision sensors are embedded in handheld devices or mounted on movable vehicles to measure useful traits such as plant height, plant biomass, crop development stage, crop yield etc. Fig. \\ref{fig-three-modes} contracts the discussed classifcations. Movable phenotyping vehicles like BoniRob have been developed where RGB cameras, hyperspectral cameras, LIDAR sensors, GPS receivers and other sensors can be mounted. Aerial based methods typically involve the usage of Unmanned Aerial Vehicles (UAVs) for crop monitoring. The recent advancements in UAVs and high resolution cameras have allowed the researchers to obtain high quality crop images. Tasks such as weed mapping, crop yield estimation, plant disease detection and pesticide spraying have been effectively carried out by UAVs. Satellite based plant phenotyping involves remote sensing of agricultural plots from satellites such as Landsat-8 and WorldView-3. Satellite based methods have been typically used for crop health monitoring over a large scale area such as a region/country. \nHowever, the cost of obtaining satellite images, the effect of clouds and the time gap between capturing and obtaining images inhibits its applicability for high throughput plant phenotyping in precision agriculture. \n\\smallskip\nWith a variety of data collection tools at our disposal, large amounts of image and sensor data have been made available for plant phenotyping research. The next section introduces deep learning, a set of methods which can effectively recognize useful patterns in huge datasets.","id":"3bee11d8-8a06-40e7-a105-9445d5a61266","level":"subsection","origin_cites_number":1,"parent_id":"61c27a51-9e54-4459-89d0-6fb5a235e790","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Background"],["subsection","Plant Phenotyping"]],"subsections":[],"title":"Plant Phenotyping"},{"cite_extract_rate":0,"cites":[],"content":"Machine Learning (ML) is a subset of Artificial Intelligence (AI), that deals with an algorithmic approach of learning from observational data without being explicitly programmed. ML has unimaginably revolutionized several fields in the last few decades. Neural Networks (NN) is a sub-field of ML and it was this sub-field that spawned Deep Learning (DL). Among the most prominent factors that contributed to the huge boost of deep learning are the appearance of large, high-quality, publicly available labelled datasets, along with the empowerment of parallel GPU computing, which enabled the transition from CPU-based to GPU-based training thus allowing for signifcant acceleration in deep models’ training. Since its redemption in 2006 , DL community has been creating ever more complex and intelligent algorithms, showing better than human performances in several intelligent tasks. The \\textit{deep} in deep learning comes from the deep architectures of learning or the hierarchical nature of its algorithms. DL algorithms stack several layers of non-linear information processing units between input and output layer, called Artificial Neurons (AN). The stacking of these ANs in a hierarchical fashion allows for exploitation of feature learning and pattern recognition through efficient learning algorithms. It is proven that NNs are universal approximator of any function, making DL task agnostic . Fig. \\ref{fig-taxonomy-of-ai} depicts the taxonomy of AI. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.35]{figures/taxonomy-of-ai.pdf}}\n\\caption{The taxonomy of AI . AI: Artificial Intelligence; ML: Machine Learning; NN: Neural Networks; DL:Deep Learning; SNN: Spiking Neural Networks.}\n\\label{fig-taxonomy-of-ai}\n\\end{figure}\nDeep learning approaches may be categorized as follows: Supervised, semi-supervised or partially supervised, and unsupervised\\footnote{Reinforcement Learning (RL) or Deep RL (DRL) is often treated as a semi-supervised or sometimes unsupervised approach.}. Supervised learning techniques use labeled data. In supervised DL, the environment includes sets of input and corresponding output pairs (often in large amounts), a criterion that evaluates model performance at all times called cost or loss function, an optimizing algorithm that minimizes the cost function with respect to the given data. Semi-supervised learning techniques use only partially labeled datasets (usually small amounts of label data, large amounts of unlabeled data). The popular Generative Adversarial Networks (GAN) are semi-supervised learning techniques. Unsupervised learning systems function without the presence of labeled data. In this case, the system learns the internal representation or important features to discover unknown relationships or structure within the input data. Often clustering, dimensionality reduction, and generative techniques are considered as unsupervised learning approaches.","id":"cd19296d-3914-418b-a3bf-11f032138dc2","level":"subsection","origin_cites_number":7,"parent_id":"61c27a51-9e54-4459-89d0-6fb5a235e790","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Background"],["subsection","Deep Learning"]],"subsections":[],"title":"Deep Learning"},{"cite_extract_rate":0.22222222222222202,"cites":[1034,486],"content":"\\subsubsection*{Convolutional Neural Networks} \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.4]{figures/cnn.pdf}}\n\\caption{The structure of a CNN , consisting of convolutional, pooling, and fully-connected layers.}\n\\label{fig-cnn}\n\\end{figure}\nConvolutional Neural Networks (CNN) is a subclass of neural networks that takes advantage of the spatial structure of the inputs. This network structure was first proposed by Fukushima in 1988 . It was not widely used then, however, due to limits of computation hardware for training the network. In the 1990s, LeCun et al. applied a gradient-based learning algorithm to CNNs and obtained successful results for the handwritten digit classification problem. CNNs have been extremely successful in computer vision applications, such as face recognition, object detection, powering vision in robotics, and self-driving cars. CNN models have a standard structure consisting of alternating convolutional layers and pooling layers (often each pooling layer is placed after a convolutional layer). The last layers are a small number of fully-connected layers, and the final layer is a softmax classifier as shown in Fig. \\ref{fig-cnn}. Every layer of a CNN transforms the input volume to an output volume of neuron activation, eventually leading to the final fully connected layers, resulting in a mapping of the input data to a 1D feature vector. In a nutshell, CNN comprises three main types of neural layers, namely, (i) convolutional layers, (ii) pooling layers, and (iii) fully connected layers. Each type of layer plays a diferent role.\n\\smallskip\n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.3]{figures/fcn-cnn.pdf}}\n\\caption{In a fully connected layer (left), each unit is connected to all units of the previous layers. In a convolutional layer (right), each unit is connected to a constant number of units in a local region of the previous layer. Furthermore, in a convolutional layer, the units all share the weights for these connections, as indicated by the shared linetypes. Figure and description are taken from .}\n\\label{fig-fcnn-cnn}\n\\end{figure}\n\\hspace{-5mm}\\textit{(i) Convolution Layers}. In the convolutional layers, a CNN\nconvolves the whole image as well as the intermediate feature maps with different kernels, generating various feature maps. Exploiting the advantages of the convolution operation, several works have proposed it as a substitute for fully connected layers with a view to attaining faster learning times. Difference between a fully connected layer and a convolutional layer is shown in Fig. \\ref{fig-fcnn-cnn}. \n\\smallskip\n\\hspace{-5mm}\\textit{(ii) Pooling Layers}. Pooling layers handle the reduction of the spatial dimensions of the input volume for the convolutional layers that immediately follow. The pooling layer does not affect the depth dimension of the volume. The operation performed by this layer is also called subsampling or downsampling, as the reduction of size leads to a simultaneous loss of information. However, such a loss is beneficial for the network because the network is forced to learn only meaningful feature representation. On top of that, the decrease in size leads to less computational overhead for the upcoming layers of the network, and also it works against overfitting. Average pooling and max pooling are the most commonly used strategies. In a detailed theoretical analysis of max pooling and average pooling performances is given, whereas in it was shown that max pooling can lead to faster convergence, select superior invariant features, and improve generalization. \n\\smallskip\n\\hspace{-5mm}\\textit{(iii) Fully Connected Layers}.\nFollowing several convolutional and pooling layers, the high-level reasoning in the neural network is performed via fully connected layers. Neurons in a fully connected layer have full connections to all activation in the previous layer, as their name implies. Their activation can hence be computed with a matrix multiplication followed by a bias offset. Fully connected layers eventually convert the 2D feature maps into a 1D feature vector. The learned vector representations either could be fed forward for classification or could be used as feature vectors for further processing.\n\\subsubsection*{Object Detection and Segmentation}\nObject detection and segmentation are two of the most important and challenging branches of computer vision, which have been widely applied in real-world applications, such as monitoring security, autonomous driving and so on, with the purpose of locating instances of semantic objects of a certain class. In a nutshell, object detection is the task of identifying locating objects (with bounding boxes) in images. While the task of segmentation is to classify each pixel of images with objects (dog, cat, airplane, etc.). We refer readers to for more information on these tasks. Fig. \\ref{fig-class-det-seg} visually contrasts the difference between these tasks.\n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.2]{figures/class_det_seg.png}}\n\\caption{Visual illustration of difference between tasks - Image Classification, Object Detection and Instance Segmentation. Example taken from MS-COCO Dataset .}\n\\label{fig-class-det-seg}\n\\end{figure}","id":"ad075082-3735-469d-88a5-712030a4f4ec","level":"subsection","origin_cites_number":9,"parent_id":"61c27a51-9e54-4459-89d0-6fb5a235e790","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Background"],["subsection","Deep Learning for Computer Vision"]],"subsections":[],"title":"Deep Learning for Computer Vision"},{"cite_extract_rate":0,"cites":[],"content":"","id":"be6f63d3-6b39-458c-918e-c883b38ba7a8","level":"section","origin_cites_number":0,"parent_id":"40b8b988-fe57-4e55-9221-27ac302d3a6f","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Application of Deep Learning in Plant Phenotyping"]],"subsections":["7b5a3160-2cf3-4478-90b2-e4ffb2d28171","a449abb1-6b4c-48cd-a598-c9b18f11bfc1","3f927eff-6ed6-4137-949b-bb7a6ee40c3a"],"title":"Application of Deep Learning in Plant Phenotyping"},{"cite_extract_rate":0.104477611940298,"cites":[1036,7043,7324,1037,1035,7323],"content":"Automation in agriculture and robotic precision agriculture activities demand a lot of information about the environment, the field, the condition and the phenotype of individual plants. An increase in availability of data allowed for successful usage of such robotic tools in real-world conditions. Taking advantage of the available data, combined with the availability of robots such as BoniRob that navigate autonomously in fields, computer vision with deep learning has played a prominent role in realizing autonomous farming. Previously laborious jobs of actively tracking certain measurements of interest such as plant growth rate, plant stem position, biomass amount, leaf count, leaf area, inter crop spacing, crop plant count and others can now be done almost seamlessly. \n\\subsubsection*{Crop Identification and Classification}\nA crucial prerequisite for selective and plant-specific treatments is that farming robots need to be equipped with an effective plant identification and classification system providing the robot with the information where and when to trigger its actuators to perform the desired action in real-time. For example, weeds generally have no useful value in terms of food, nutrition or medicine yet they have accelerated growth and parasitically compete with actual crops for nutrients and space. Inefficient processes such as hand weeding has\nled to significant losses and increasing costs due to manual labour , which is why a lot of research is being done on crop vs weed classification and weed identification and plant seedlings classification . This is extremely useful in improving the efficiency of precision farming techniques on weed control by modulating herbicide spraying appropriately to the level of weeds infestation. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.3]{figures/three-modes.pdf}}\n\\caption{Top row of \\textbf{(a)} shows BoniRob a ground-based remote sensing robot, \\textbf{(b)} shows an unmanned aerial vehicle , \\textbf{(c)} shows a satellite scanning large areas of land respectively. Bottom row across \\textbf{(a)}, \\textbf{(b)}, and \\textbf{(c)} shows corresponding example images acquired. Satellite Image Credits: NASA.}\n\\label{fig-three-modes}\n\\end{figure}\n\\subsubsection*{Crop Detection and Segmentation}\nCrop detection in the wild is arguably the most crucial step in the pipeline of several farm management tasks such as visual crop categorization , real-time plant disease and pest recognition , picking and harvesting automatic robots , healthy and quality monitoring of crop growing and yield estimation . However, existing deep learning networks achieving state-of-the-art performance in other research fields are not suitable for agricultural tasks of crop management such as irrigation , picking , pesticide spraying , and fertilization . The dominating cause is lack of diverse set of public benchmark datasets that are specifically designed for various agricultural missions. Some of the few rich datasets available are CropDeep for detection, multi-modal datasets like Rosette plant or \\textit{Arabidopsis} datasets , Sorghum-Head , Wheat-Panicle , Crop/Weed segmentation , and Crop/Tassle segmentation . Fig. \\ref{fig-cropdeep} contains some examples from the CropDeep dataset. Fig. \\ref{fig-rosette} depicts multi-modal annotations provided in the Rosette Plant Phenotyping dataset i.e., annotations for detection, segmentation, leaf center along with otherwise rarely found meta data. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.3]{figures/cropdeep.png}}\n\\caption{Some annotation examples from CropDeep dataset .}\n\\label{fig-cropdeep}\n\\end{figure}\nEfficient yield estimation from images is also one of the key tasks for farmers and plant breeders to accurately quantify the overall throughput of their ecosystem. Recent efforts in panicle or spike detection , leaf counting , fruit detection as well as pixel-wise segmentation-based tasks such as panicle segmentation show very promising results in this direction. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.7]{figures/RosettePlant.pdf}}\n\\caption{Visual illustration of all types of annotations available in dataset.}\n\\label{fig-rosette}\n\\end{figure}\n\\subsubsection*{Crop Disease and Pest Recognition}\nModern technologies have given human society the ability to produce enough food to meet the demand of more than 7 billion people. However, food security remains threatened by a number of factors including climate change , the decline in pollinators , plant diseases , and others. Plant diseases are not only a threat to food security at the global scale, but can also have disastrous consequences for smallholder farmers whose livelihoods depend on healthy crops. India loses 35\\% of the annual crop yield due to plant diseases . In the developing world, more than 80 percent of the agricultural production is generated by smallholder farmers , and reports of yield loss of more than 50\\% due to pests and diseases are frequent . Furthermore, the largest fraction of hungry people (50\\%) live in smallholder farming households , making smallholder farmers a group that’s particularly vulnerable to pathogen-derived disruptions in food supply.\nOwing to these factors, timely disease and pest recognition becomes a priority task for farmers. In addition to that, farmers do not have many options other than consulting other fellow farmers or seeking help from government funded helplines . Availability of public datasets such as PlantVillage , PlantDoc allowed for progress in the area of disease and pest detection. Recent research works in pest and insect detection , invasive species detection in marine aquaculture and disease detection in plant leafs , Rice , Tomato , Banana , Grape , Sugarcane , Eggplant , Cucumber , Soybean , Olive , Tea , Coffee and other similar works take encouraging steps towards disease-free agriculture. Fig. \\ref{fig-disease} depicts banana diseases and pest detection outputs from . This work reports solutions to extant limitations in plant disease detection. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.3]{figures/banana.png}}\n\\caption{Detected classes and expected output of the trained disease detection model. \\textbf{a} Entire plant afected by banana bunchy top virus (BBTV), \\textbf{b} leaves affected by black sigatoka (BS), \\textbf{c} cut pseudostem of Xanthomonas wilt (BXW) afected plant showing yellow bacterial ooze, \\textbf{d} fruit bunch afected by Xanthomonas wilt (BXW), e cut fruit afected by Xanthomonas wilt (BXW), \\textbf{f} corm afected by banana corm weevil (BCW). Figure and description taken from .}\n\\label{fig-disease}\n\\end{figure}","id":"7b5a3160-2cf3-4478-90b2-e4ffb2d28171","level":"subsection","origin_cites_number":67,"parent_id":"be6f63d3-6b39-458c-918e-c883b38ba7a8","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Application of Deep Learning in Plant Phenotyping"],["subsection","Ground-Based Remote Sensing for Plant Phenotyping"]],"subsections":[],"title":"Ground-Based Remote Sensing for Plant Phenotyping"},{"cite_extract_rate":0.17391304347826,"cites":[8406,1040,1039,1038],"content":"The past few decades have witnessed the great progress of unmanned aircraft vehicles (UAVs) in civilian fields, especially in photogrammetry and remote sensing. In contrast with the platforms of manned aircraft and satellite, the UAV platform holds many promising characteristics: flexibility, efficiency, high spatial/temporal resolution, low cost, easy operation, etc., which make it an effective complement to other remote-sensing platforms and a cost-effective means for remote sensing. We refer reader to literary works for the detailed reports of techniques and applications of UAVs in precision agriculture, remote sensing, search and rescue, construction and infrastructure inspection and discuss other market opportunities. UAVs can be utilized in precision agriculture (PA) for crop management and monitoring , weed detection , irrigation scheduling , agricultural pattern detection , pesticide spraying , cattle detection , disease detection , insect detection and data collection from ground sensors (moisture, soil properties, etc.,) . The deployment of UAVs in PA is a cost-effective and time saving technology which can help for improving crop yields, farms productivity and profitability in farming systems. Moreover, UAVs facilitate agricultural management, weed monitoring, and pest damage, thereby they help to meet these challenges quickly . \nUAVs can also be utilized to monitor and quantify several factors of irrigation such as availability of soil water, crop water need (which represents the amount of water needed by the various crops to grow optimally), rainfall amount, efficiency of the irrigation system . In this work , UAVs are currently being utilized to estimate the spatial distribution of surface soil moisture high-resolution multi-spectral imagery in combination with ground sampling. UAVs are also being used for thermal remote sensing to monitor the spatial and temporal patterns of crop diseases during various disease development phases which reduces crop losses for farmers. This work detects early stage development of soil-borne fungus in UAV imagery. Soil texture can be an indicative of soil quality which in turn influences crop productivity. Hence, UAV thermal images are being utilized to quantify soil texture at a regional scale by measuring the differences in land surface temperature under a relatively homogeneous climatic condition . Accurate assessment of crop residue is crucial for proper implementation of conservation tillage practices since crop residues provide a protective layer on agricultural fields that shields soil from wind and water. In , the authors demonstrated that aerial thermal images can explain more than 95\\% of the variability in crop residue cover amount compared to 77\\% using visible and near IR images.\nFarmers must monitor crop maturity to determine the harvesting time of their crops. UAVs can be a practical solution to this problem . Farmers require accurate, early estimation of crop yield for a number of reasons, including crop insurance, planning of harvest and storage requirements, and cash flow budgeting. In , UAV images were utilized to estimate yield and total biomass of rice crop in Thailand. In , UAV images were also utilized to predict corn grain yields in the early to midseason crop growth stages in Germany. \nThere have also been successful efforts that seamlessly combine aerial and ground based system for precision agriculture . With relaxed flight regulations and drastic improvement in machine learning techniques, geo-referencing, mosaicing, and other related algorithms, UAVs can provide a great potential for soil and crop monitoring . More precision agricultural researches are encouraged to design and implement special types of cameras and sensors on-board UAVs, which have the ability of remote crop monitoring and detection of soil and other agricultural characteristics in real time scenarios.","id":"a449abb1-6b4c-48cd-a598-c9b18f11bfc1","level":"subsection","origin_cites_number":23,"parent_id":"be6f63d3-6b39-458c-918e-c883b38ba7a8","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Application of Deep Learning in Plant Phenotyping"],["subsection","Unmanned Aircraft Vehicles for Plant Phenotyping"]],"subsections":[],"title":"Unmanned Aircraft Vehicles for Plant Phenotyping"},{"cite_extract_rate":0,"cites":[],"content":"The impact of climate change and its unforeseeable nature, has caused majority of the agricultural crops to be affected in terms of their production and maintenance. With more than seven billion mouths to feed greater demands are being put on agriculture than ever before, at the same time as land is being degraded by factors such as soil erosion, mineral exhaustion and drought. It becomes the utmost priority for governments to support farmers by providing crucial information about changing weather conditions, soil conditions and more. Currently, satellite imagery is making agriculture more efficient by reducing scouting efforts of farmers, by optimizing use of nitrogen based on variable rate of application, by optimizing water schedules, identifying field performance and benchmark fields, etc . India alone has 7 satellites specially designed for benefits of farmers .\nSatellites and their imagery are being applied to agriculture in several ways, initially as a means of estimating crop yields and crop types , soil salinity, soil moisture, soil pH . Optical and radar sensors can provide an accurate picture of the acreage being cultivated, while also differentiating between crop types and determining their health and maturity. Optical satellite sensors can detect visible and near-infrared wavelengths of light, reflected from agricultural land below. It is these wavelengths which combined, can be manipulated to help us understand the condition of the crops. This information helps to inform the market, and provide early warning of crop failure or famine.\nBy extension, satellites are also used as a management tool through the practice of PA, where satellite images are used to characterise a farmer's fields in detail, often used in combination with geographical information systems (GIS), to allow more intensive and efficient cultivation practices. For instance, different crops might be recommended for different fields while the farmer's use of fertiliser is optimised in a more economic and environmentally-friendly fashion. Providing access to satellite imagery also becomes very important for building trust among the involved parties (farmers and government and private bodies involved). Web-based platforms such as Google Earth Engine, Planet.com, Earth Data Search by NASA, LandViewer by Earth Observing System, Geocento and others provide access to past and present (even daily) satellite imagery of your interest. \nAgricultural monitoring is also increasingly being applied to forestry, both for forest management and as a way of characterising forests as carbon sinks to help minimise climate change – notably as part of the UN's REDD programme .","id":"3f927eff-6ed6-4137-949b-bb7a6ee40c3a","level":"subsection","origin_cites_number":10,"parent_id":"be6f63d3-6b39-458c-918e-c883b38ba7a8","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Application of Deep Learning in Plant Phenotyping"],["subsection","Satellites for Plant Phenotyping"]],"subsections":[],"title":"Satellites for Plant Phenotyping"},{"cite_extract_rate":0.5,"cites":[7217,1050,1051,1048,1045,1041,1047,1043,1049,1044,1042,1046,7323],"content":"While deep learning based plant phenotyping has shown great promise, requirement of large labeled datasets still remains to be the bottleneck. Phenotyping tasks are often specific to the environmental and genetic conditions, finding large datasets with such conditions is not always possible. This results in researchers needing to acquire their own dataset and label it, which is often a arduous and expensive affair. Moreover, small datasets often lead to models that overfit. Deep learning approaches optimized for working with limited labeled data would immensely help the plant phenotyping community, since this would encourage many more farmers, breeders, and researchers to employ reliable plant phenotyping techniques to optimize crop yield. To this end, we list out some of the recent efforts in the area of deep plant phenotyping with limited labeled data.\n\\subsection*{Data Augmentation}\nThe computer vision community has long been employing dataset augmentation techniques to grow the amount of data using artificial transformations. Artificially perturbing the original dataset with affine transformations (e.g., rotation, scale, translation) is considered a common practice now. However, this approach has some constraints: the augmented data only capture the variability of the available training set (e.g., if the dataset doesn't include a unique colored fruit, the particular unique case will never be learnt). To overcome this, several data augmentation methods proposed take advantage of recent advancements in the image generation space. In this work , the authors use Generative Adversarial Network (GAN) to generate \\textit{Arabidopsis} plant images (called ARIGAN) with unique desirable traits (over 7 leaves) that were originally less frequent in the dataset. Fig. \\ref{fig-generated} \\textbf{(a)} shows examples of images generated by ARIGAN. Other latest works use more advanced variants of GANs to generate realistic plant images with particularly favorable leaf segmentations of interest to boost leaf counting accuracy of the learning models. In , the authors proposed an unsupervised image translation technique to improve plant disease recognition performance. LeafGAN , an image-to-image translation model, generates leaf images with various plant diseases and boosts diagnostic performance by a great margin. Two sets of example images generated by LeafGAN are shown in Fig. \\ref{fig-generated} \\textbf{(b)}. Other data enhancement techniques are also being employed by researchers to train plant disease diagnosis models on generated lesions . \nThe effort to provide finely annotated data has enabled great improvement of the state of the art on segmentation performance. Some researches have started working on effectively transferring the knowledge obtained from RGB images on annotated plants either to other species or other modalities of imaging. In this work , the authors successfully transfer the knowledge gained from annotated leaves of \\textit{Arabidopsis thaliana} in RGB to images of the same plant in chlorophyll fluorescence imaging. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.55]{figures/generated.pdf}}\n\\caption{\\textbf{(a)} shows \\textit{Arabidopsis} plant images generated by ARIGAN . Bottom-right numbers refer to the leaf count. \\textbf{(b)} shows two sets of healthy leafs and their corresponding disease prone leaves generated by LeafGAN .}\n\\label{fig-generated}\n\\end{figure}\n\\subsection*{Weakly Supervised Learning}\nFruit/organ counting is a well explored task by the plant phenotyping community. However, many vision-based solutions we have currently require highly accurate instance and density labels of fruits and organs in diverse set of environments. The labeling procedures are often very burdensome and error prone and, in many agricultural scenarios, it may be impossible to acquire a sufficient number of labelled samples to achieve consistent performance that are robust to image noise or other forms of covariate shift. This is why using only weak labels can be crucial for cost-effective plant phenotyping. \nRecently, a lot of attention has been placed on engineering weakly supervised learning frameworks for plant phenotyping. In , the authors created a weakly supervised framework for the sorghum head detection task where annotators label the data only until the model reaches a desired performance level. After that, model outputs are directly passed as data labels leading to a exponential reduction in annotation costs with minimal loss in model accuracy. In other work , the authors proposed a strategy which is able to learn to count fruits without requiring task-specific supervision labels, such as manually labelled object bounding boxes or total instance count. In , the authors use a trained CNN on defect classification data and use it's activate maps to segment infected regions on potatoes. Segmentation task requires really rich labels (each pixel of the image is annotated) so this task effectively bypasses the labeling for segmentation altogether. On another note, rice heading date estimation greatly assists the breeders to understand the adaptability of the crop to various environmental and genetic conditions. Accurate estimation of heading date requires monitoring the increase in number of rice panicles in the crop. Detecting rice panicles from crop images usually requires training an object detection model such as Faster R-CNN or YOLO, which requires costly bounding box annotations. However, a recently proposed method uses a sliding window based detector which requires training an image classifier, for which annotations are much easier to obtain. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.3]{figures/bmvc.jpg}}\n\\caption{\\textbf{(a)} Standard pool-based active learning framework \\textbf{(b)} Proposed framework which interleaves weak supervision in the active learning process. This novel framework includes an adaptive supervision module which allows switching to a stronger form of supervision as required when training the model. The \\textit{oracle} is the source of labels \\textit{a.k.a} annotator.}\n\\label{fig-bmvc}\n\\end{figure}\n\\subsection*{Transfer Learning}\nTransfer learning is a type of learning that enables using knowledge gained while solving one problem and applying it to a different but related problem i.e., a model trained on one phenotyping task (say potato leaf classification) being able to assist another phenotyping (tomato leaf classification) task. Transfer learning is a very well explored area of machine learning. As part of the first steps of adopting existing transfer learning techniques for plant phenotyping, the authors of use CNNs (AlexNet, GoogleNet and VGGNet) pretrained on ImageNet dataset and fine tune on the plant dataset used in LifeCLEF 2015 challenge. With the help of transfer learning, they were able to beat then existing state-of-the-art LifeCLEF performance by 15\\% points. Similary in , the authors report better than human results in segmentation task with the help of transfer learning where they transfer learn a model trained on peanut root dataset for switchgrass root dataset (they also report results using ImageNet pretrained models). Leaf disease detection and treatment recommendation performance is also shown to be boosted with transfer learning . In , the authors interestingly combined a State-of-the-Art weakly-supervised fruit counting model with an unsupervised style transfer method for fruit counting. They used Cycle-Generative Adversarial Network (C-GAN) to perform unsupervised domain adaptation from one fruit dataset to another and train it alongside with a Presence-Absence Classifier (PAC) that discriminates images containing fruits or not and ultimately achieved better performance than fully supervised models. \n\\begin{figure}\n\\centerline{\\includegraphics[scale=0.15]{figures/bmc.jpg}}\n\\caption{Proposed point supervision framework into the pool-based active learning cycle. In this framework, strong supervision is queried for images only after deemed \\textit{informative} based on point supervision of those images.}\n\\label{fig-bmc}\n\\end{figure}\n\\subsection*{Active Learning}\nActive learning , an iterative training approach that curiously selects the best samples to train, has been shown to reduce labeled data requirement when training deep classification networks . Research in the area of active learning for object detection has been, arguably, limited. However, numerous plant phenotyping tasks such as detection and quantification of crop yield and fruit counting are directly dependent on object detection. Keeping this in mind, an active learning method has been proposed for training deep object detection models where the model can selectively query either weak labels (pointing at the object) or strong labels (drawing a box around the object). By introducing a switching module for weak labels and strong labels, the authors were able to save 24\\% of annotation time while training a wheat head detection model. Fig. \\ref{fig-bmvc} illustrates the difference between regular active learning cycle and proposed active learning cycle. This method demonstrates the applicability of active learning to plant phenotyping methods where obtaining labeled data is often difficult. Along the same lines, to alleviate the labeled data requirement for training object detection models for cereal crop detection, a weak supervision based active learning method was proposed recently. In this active learning approach, the model constantly interacts with a human annotator by iteratively querying the labels for only the most informative images, as opposed to all images in a dataset. Fig. \\ref{fig-bmc} visually illustrates the proposed framework. The active query method is specifically designed for cereal crops which usually tend to have panicles with low variance in appearance. This training method has been shown to reduce over 50\\% of annotation costs on sorghum head and wheat spike detection datasets. We expect to see more research works using active learning for limited labeled data based plant phenotyping in the near future.","id":"fcb14fe1-1d8d-475b-8ed7-cf7547edb6ff","level":"section","origin_cites_number":28,"parent_id":"40b8b988-fe57-4e55-9221-27ac302d3a6f","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Plant Phenotyping with Limited Labeled Data"]],"subsections":[],"title":"Plant Phenotyping with Limited Labeled Data"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we describe some of the challenges present in plant phenotyping methods which warrant further research. \n\\subsection*{The Training Data Bottleneck} \nModern phenotyping methods rely on deep learning which is notorious for requiring large amounts of labeled data. While some progress has been made in developing data efficient models for phenotyping, reducing the labeling efforts for training efficient phenotyping tools is still an open problem. We believe that effectively adapting techniques from deep learning such as unsupervised, self supervised, weakly supervised, active and semi-supervised learning will greatly benefit the phenotyping community in observing plant traits with small datasets. \n\\subsection*{Explainability}\nDeep neural networks are generally considered as black boxes which produce predictions without sufficient justification. This makes debugging a neural network difficult i.e., it can be tough to understand what caused a wrong prediction. Crop management decisions based on incorrect phenotyping results can cause financial losses. Hence, developing explainable models for plant phenotyping is one of the open problems in this field. Obtaining the reasons behind a given set of plant traits using explainable models has the potential to achieve breakthroughs in our understanding of plant behavior in various genetic and environmental conditions. \n\\subsection*{Data collection}\nVision based plant phenotyping suffers from challenges such as occlusion, inaccuracies in 3D reconstruction of crops and bad lighting conditions caused by the changing weather. It is therefore necessary to develop phenotyping tools which are robust to visual variations.","id":"6bf986f3-5fd9-4afa-a45b-7964c94d86eb","level":"section","origin_cites_number":0,"parent_id":"40b8b988-fe57-4e55-9221-27ac302d3a6f","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Challenges and Open Problems"]],"subsections":[],"title":"Challenges and Open Problems"},{"cite_extract_rate":0,"cites":[],"content":"High throughput plant phenotyping methods have shown great promise in efficiently monitoring crops for plant breeding and agricultural crop management. Research in deep learning has accelerated the progress in plant phenotyping research which resulted in the development of various image analysis tools to observe plant traits. However, wide applicability of high throughput phenotyping tools is limited by some issues such as 1) dependence of deep networks on large datasets, which are difficult to curate, 2) large variations of field environment which cannot always be captured, and 3) capital and maintenance which can be prohibitively expensive to be widely used in developing countries. With many open problems in plant phenotyping warranting further studies, it is indeed a great time to study plant phenotyping and achieve rapid progress by utilizing the advances in deep learning. \n\\bibliographystyle{splncs}\n\\bibliography{egbib}\n\\end{document}","id":"1cab82e7-7d3c-4e75-913b-94d66346fc31","level":"section","origin_cites_number":0,"parent_id":"40b8b988-fe57-4e55-9221-27ac302d3a6f","prefix_titles":[["title","Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"40b8b988-fe57-4e55-9221-27ac302d3a6f\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ]\n ],\n \"subsections\": [\n \"50e565bb-2ea0-4b95-baa0-7a3211a9fe18\",\n \"61c27a51-9e54-4459-89d0-6fb5a235e790\",\n \"be6f63d3-6b39-458c-918e-c883b38ba7a8\",\n \"fcb14fe1-1d8d-475b-8ed7-cf7547edb6ff\",\n \"6bf986f3-5fd9-4afa-a45b-7964c94d86eb\",\n \"1cab82e7-7d3c-4e75-913b-94d66346fc31\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Population growth, increasing incomes, and rapid urbanization in developing countries are expected to cause a drastic hike in food demand. This projected rise in food demand poses several challenges to agriculture. Owing to a continuous decline in global cultivable land , increasing the productivity of the existing agricultural land is highly necessary. This need has led to the scientific community focusing their efforts on developing efficient and sustainable ways to increase crop yield. To this end, precision agriculture techniques have attracted a lot of attention. Precision agriculture is a set of methods to monitor crops, gather data, and carry out informed crop management tasks such as applying the optimum amount of water, selecting suitable pesticides, and reducing environmental impact. These methods involve the usage of specialized devices such as sensors, UAVs, and static cameras to monitor the crops. Accurate crop monitoring goes a long way in assisting farmers in making the right choices to obtain the maximum yield. Plant phenotyping, a rapidly emerging research area, plays a significant role in understanding crop-related traits. Plant phenotyping is the science of characterizing and quantifying the physical and physiological traits of a plant. It provides a quantitative assessment of the plant's properties and its behavior in various environmental conditions. Understanding these properties is crucial in performing effective crop management. \\nResearch in plant phenotyping has grown rapidly thanks to the availability of cost-effective and easy to use digital imaging devices such as RGB, multispectral, and hyperspectral cameras, which have facilitated the collection of large amounts of data. This influx of data coupled with the usage of machine learning algorithms has fueled the development of various high throughput phenotyping tools [refs] for tasks such as weed detection, fruit/organ counting, disease detection and yield estimation. A machine learning pipeline typically consists of feature extraction followed by a classification/regression module for prediction. While machine learning techniques have helped build sophisticated phenotyping tools, they are known to lack robustness. They rely heavily on handcrafted feature extraction techniques and manual hyperparameter tuning methods. As a result, if feature extraction is not carefully done under a domain expert's supervision, they tend to perform poorly in uncontrolled environments such as agricultural fields where factors such as lighting, weather, exposure, etc. often cannot be regulated. Hence, feature extraction from data has been one of the major bottlenecks in the development of efficient high throughput plant phenotyping systems.\\nAdvancements in deep learning, a sub-field of machine learning which allows for automatic feature extraction and prediction on large scale data, has led to a surge in the development of visual plant phenotyping methods. Deep learning is particularly well-known for its effectiveness in handling vision-based tasks such as image classification, object detection, semantic segmentation, and scene understanding. Coincidentally, many of these tasks form the backbone for various plant phenotyping tasks such as disease detection, fruit detection, and yield estimation. Fig. \\\\ref{fig-ml-dl} illustrates the difference between machine learning based plant phenotyping and deep learning based plant phenotyping. We believe that the expressive power and robustness of deep learning systems can be effectively leveraged by plant researchers to identify complex patterns from raw data and devise efficient precision agriculture methodologies. The purpose of this survey is to enable the readers to get a bird's eye view of the advancements in the field of deep learning based plant phenotyping, understand the existing issues, and become familiar with some of the open problems which warrant further research.\\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.7]{figures/ml-dl.pdf}}\\n\\\\caption{Difference between ML and DL based Plant Phenotyping.}\\n\\\\label{fig-ml-dl}\\n\\\\end{figure}\",\n \"id\": \"50e565bb-2ea0-4b95-baa0-7a3211a9fe18\",\n \"level\": \"section\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"40b8b988-fe57-4e55-9221-27ac302d3a6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"61c27a51-9e54-4459-89d0-6fb5a235e790\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40b8b988-fe57-4e55-9221-27ac302d3a6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ]\n ],\n \"subsections\": [\n \"3bee11d8-8a06-40e7-a105-9445d5a61266\",\n \"cd19296d-3914-418b-a3bf-11f032138dc2\",\n \"ad075082-3735-469d-88a5-712030a4f4ec\"\n ],\n \"title\": \"Background\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Plant phenotyping is the science of quantifying the physical and physiological traits of a plant. Plant phenotyping mainly benefits two communities: farmers and plant breeders. By better understanding the traits of the crop, a farmer can optimize crop yield by making informed crop management decisions. Similarly, understanding the crop's behavior is crucial for plant breeders to select the best possible crop variety for a given location and environment. In the past, plant phenotyping was a manual endeavor. The process of manually observing a small set of crop samples and reporting observations periodically was slow, labor intensive and inefficient. The low throughput nature of these methods has impeded the progress in plant breeding research. However, the advent of modern data acquisition methods with various sensors, cameras and UAVs (Unmanned Aerial Vehicles) coupled with advances in machine learning techniques have resulted in the development of high-throughput plant phenotyping methods to be effectively used for precision agriculture. \\nDepending on the method of data collection, plant phenotyping techniques can be classified into ground based, aerial and satellite based methods. In ground based phenotyping, high precision sensors are embedded in handheld devices or mounted on movable vehicles to measure useful traits such as plant height, plant biomass, crop development stage, crop yield etc. Fig. \\\\ref{fig-three-modes} contracts the discussed classifcations. Movable phenotyping vehicles like BoniRob have been developed where RGB cameras, hyperspectral cameras, LIDAR sensors, GPS receivers and other sensors can be mounted. Aerial based methods typically involve the usage of Unmanned Aerial Vehicles (UAVs) for crop monitoring. The recent advancements in UAVs and high resolution cameras have allowed the researchers to obtain high quality crop images. Tasks such as weed mapping, crop yield estimation, plant disease detection and pesticide spraying have been effectively carried out by UAVs. Satellite based plant phenotyping involves remote sensing of agricultural plots from satellites such as Landsat-8 and WorldView-3. Satellite based methods have been typically used for crop health monitoring over a large scale area such as a region/country. \\nHowever, the cost of obtaining satellite images, the effect of clouds and the time gap between capturing and obtaining images inhibits its applicability for high throughput plant phenotyping in precision agriculture. \\n\\\\smallskip\\nWith a variety of data collection tools at our disposal, large amounts of image and sensor data have been made available for plant phenotyping research. The next section introduces deep learning, a set of methods which can effectively recognize useful patterns in huge datasets.\",\n \"id\": \"3bee11d8-8a06-40e7-a105-9445d5a61266\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"61c27a51-9e54-4459-89d0-6fb5a235e790\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Plant Phenotyping\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Plant Phenotyping\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Machine Learning (ML) is a subset of Artificial Intelligence (AI), that deals with an algorithmic approach of learning from observational data without being explicitly programmed. ML has unimaginably revolutionized several fields in the last few decades. Neural Networks (NN) is a sub-field of ML and it was this sub-field that spawned Deep Learning (DL). Among the most prominent factors that contributed to the huge boost of deep learning are the appearance of large, high-quality, publicly available labelled datasets, along with the empowerment of parallel GPU computing, which enabled the transition from CPU-based to GPU-based training thus allowing for signifcant acceleration in deep models’ training. Since its redemption in 2006 , DL community has been creating ever more complex and intelligent algorithms, showing better than human performances in several intelligent tasks. The \\\\textit{deep} in deep learning comes from the deep architectures of learning or the hierarchical nature of its algorithms. DL algorithms stack several layers of non-linear information processing units between input and output layer, called Artificial Neurons (AN). The stacking of these ANs in a hierarchical fashion allows for exploitation of feature learning and pattern recognition through efficient learning algorithms. It is proven that NNs are universal approximator of any function, making DL task agnostic . Fig. \\\\ref{fig-taxonomy-of-ai} depicts the taxonomy of AI. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.35]{figures/taxonomy-of-ai.pdf}}\\n\\\\caption{The taxonomy of AI . AI: Artificial Intelligence; ML: Machine Learning; NN: Neural Networks; DL:Deep Learning; SNN: Spiking Neural Networks.}\\n\\\\label{fig-taxonomy-of-ai}\\n\\\\end{figure}\\nDeep learning approaches may be categorized as follows: Supervised, semi-supervised or partially supervised, and unsupervised\\\\footnote{Reinforcement Learning (RL) or Deep RL (DRL) is often treated as a semi-supervised or sometimes unsupervised approach.}. Supervised learning techniques use labeled data. In supervised DL, the environment includes sets of input and corresponding output pairs (often in large amounts), a criterion that evaluates model performance at all times called cost or loss function, an optimizing algorithm that minimizes the cost function with respect to the given data. Semi-supervised learning techniques use only partially labeled datasets (usually small amounts of label data, large amounts of unlabeled data). The popular Generative Adversarial Networks (GAN) are semi-supervised learning techniques. Unsupervised learning systems function without the presence of labeled data. In this case, the system learns the internal representation or important features to discover unknown relationships or structure within the input data. Often clustering, dimensionality reduction, and generative techniques are considered as unsupervised learning approaches.\",\n \"id\": \"cd19296d-3914-418b-a3bf-11f032138dc2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"61c27a51-9e54-4459-89d0-6fb5a235e790\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Deep Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Learning\"\n },\n {\n \"cite_extract_rate\": 0.22222222222222202,\n \"cites\": [\n 1034,\n 486\n ],\n \"content\": \"\\\\subsubsection*{Convolutional Neural Networks} \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.4]{figures/cnn.pdf}}\\n\\\\caption{The structure of a CNN , consisting of convolutional, pooling, and fully-connected layers.}\\n\\\\label{fig-cnn}\\n\\\\end{figure}\\nConvolutional Neural Networks (CNN) is a subclass of neural networks that takes advantage of the spatial structure of the inputs. This network structure was first proposed by Fukushima in 1988 . It was not widely used then, however, due to limits of computation hardware for training the network. In the 1990s, LeCun et al. applied a gradient-based learning algorithm to CNNs and obtained successful results for the handwritten digit classification problem. CNNs have been extremely successful in computer vision applications, such as face recognition, object detection, powering vision in robotics, and self-driving cars. CNN models have a standard structure consisting of alternating convolutional layers and pooling layers (often each pooling layer is placed after a convolutional layer). The last layers are a small number of fully-connected layers, and the final layer is a softmax classifier as shown in Fig. \\\\ref{fig-cnn}. Every layer of a CNN transforms the input volume to an output volume of neuron activation, eventually leading to the final fully connected layers, resulting in a mapping of the input data to a 1D feature vector. In a nutshell, CNN comprises three main types of neural layers, namely, (i) convolutional layers, (ii) pooling layers, and (iii) fully connected layers. Each type of layer plays a diferent role.\\n\\\\smallskip\\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.3]{figures/fcn-cnn.pdf}}\\n\\\\caption{In a fully connected layer (left), each unit is connected to all units of the previous layers. In a convolutional layer (right), each unit is connected to a constant number of units in a local region of the previous layer. Furthermore, in a convolutional layer, the units all share the weights for these connections, as indicated by the shared linetypes. Figure and description are taken from .}\\n\\\\label{fig-fcnn-cnn}\\n\\\\end{figure}\\n\\\\hspace{-5mm}\\\\textit{(i) Convolution Layers}. In the convolutional layers, a CNN\\nconvolves the whole image as well as the intermediate feature maps with different kernels, generating various feature maps. Exploiting the advantages of the convolution operation, several works have proposed it as a substitute for fully connected layers with a view to attaining faster learning times. Difference between a fully connected layer and a convolutional layer is shown in Fig. \\\\ref{fig-fcnn-cnn}. \\n\\\\smallskip\\n\\\\hspace{-5mm}\\\\textit{(ii) Pooling Layers}. Pooling layers handle the reduction of the spatial dimensions of the input volume for the convolutional layers that immediately follow. The pooling layer does not affect the depth dimension of the volume. The operation performed by this layer is also called subsampling or downsampling, as the reduction of size leads to a simultaneous loss of information. However, such a loss is beneficial for the network because the network is forced to learn only meaningful feature representation. On top of that, the decrease in size leads to less computational overhead for the upcoming layers of the network, and also it works against overfitting. Average pooling and max pooling are the most commonly used strategies. In a detailed theoretical analysis of max pooling and average pooling performances is given, whereas in it was shown that max pooling can lead to faster convergence, select superior invariant features, and improve generalization. \\n\\\\smallskip\\n\\\\hspace{-5mm}\\\\textit{(iii) Fully Connected Layers}.\\nFollowing several convolutional and pooling layers, the high-level reasoning in the neural network is performed via fully connected layers. Neurons in a fully connected layer have full connections to all activation in the previous layer, as their name implies. Their activation can hence be computed with a matrix multiplication followed by a bias offset. Fully connected layers eventually convert the 2D feature maps into a 1D feature vector. The learned vector representations either could be fed forward for classification or could be used as feature vectors for further processing.\\n\\\\subsubsection*{Object Detection and Segmentation}\\nObject detection and segmentation are two of the most important and challenging branches of computer vision, which have been widely applied in real-world applications, such as monitoring security, autonomous driving and so on, with the purpose of locating instances of semantic objects of a certain class. In a nutshell, object detection is the task of identifying locating objects (with bounding boxes) in images. While the task of segmentation is to classify each pixel of images with objects (dog, cat, airplane, etc.). We refer readers to for more information on these tasks. Fig. \\\\ref{fig-class-det-seg} visually contrasts the difference between these tasks.\\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.2]{figures/class_det_seg.png}}\\n\\\\caption{Visual illustration of difference between tasks - Image Classification, Object Detection and Instance Segmentation. Example taken from MS-COCO Dataset .}\\n\\\\label{fig-class-det-seg}\\n\\\\end{figure}\",\n \"id\": \"ad075082-3735-469d-88a5-712030a4f4ec\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"61c27a51-9e54-4459-89d0-6fb5a235e790\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Deep Learning for Computer Vision\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Learning for Computer Vision\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"be6f63d3-6b39-458c-918e-c883b38ba7a8\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40b8b988-fe57-4e55-9221-27ac302d3a6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Application of Deep Learning in Plant Phenotyping\"\n ]\n ],\n \"subsections\": [\n \"7b5a3160-2cf3-4478-90b2-e4ffb2d28171\",\n \"a449abb1-6b4c-48cd-a598-c9b18f11bfc1\",\n \"3f927eff-6ed6-4137-949b-bb7a6ee40c3a\"\n ],\n \"title\": \"Application of Deep Learning in Plant Phenotyping\"\n },\n {\n \"cite_extract_rate\": 0.104477611940298,\n \"cites\": [\n 1036,\n 7043,\n 7324,\n 1037,\n 1035,\n 7323\n ],\n \"content\": \"Automation in agriculture and robotic precision agriculture activities demand a lot of information about the environment, the field, the condition and the phenotype of individual plants. An increase in availability of data allowed for successful usage of such robotic tools in real-world conditions. Taking advantage of the available data, combined with the availability of robots such as BoniRob that navigate autonomously in fields, computer vision with deep learning has played a prominent role in realizing autonomous farming. Previously laborious jobs of actively tracking certain measurements of interest such as plant growth rate, plant stem position, biomass amount, leaf count, leaf area, inter crop spacing, crop plant count and others can now be done almost seamlessly. \\n\\\\subsubsection*{Crop Identification and Classification}\\nA crucial prerequisite for selective and plant-specific treatments is that farming robots need to be equipped with an effective plant identification and classification system providing the robot with the information where and when to trigger its actuators to perform the desired action in real-time. For example, weeds generally have no useful value in terms of food, nutrition or medicine yet they have accelerated growth and parasitically compete with actual crops for nutrients and space. Inefficient processes such as hand weeding has\\nled to significant losses and increasing costs due to manual labour , which is why a lot of research is being done on crop vs weed classification and weed identification and plant seedlings classification . This is extremely useful in improving the efficiency of precision farming techniques on weed control by modulating herbicide spraying appropriately to the level of weeds infestation. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.3]{figures/three-modes.pdf}}\\n\\\\caption{Top row of \\\\textbf{(a)} shows BoniRob a ground-based remote sensing robot, \\\\textbf{(b)} shows an unmanned aerial vehicle , \\\\textbf{(c)} shows a satellite scanning large areas of land respectively. Bottom row across \\\\textbf{(a)}, \\\\textbf{(b)}, and \\\\textbf{(c)} shows corresponding example images acquired. Satellite Image Credits: NASA.}\\n\\\\label{fig-three-modes}\\n\\\\end{figure}\\n\\\\subsubsection*{Crop Detection and Segmentation}\\nCrop detection in the wild is arguably the most crucial step in the pipeline of several farm management tasks such as visual crop categorization , real-time plant disease and pest recognition , picking and harvesting automatic robots , healthy and quality monitoring of crop growing and yield estimation . However, existing deep learning networks achieving state-of-the-art performance in other research fields are not suitable for agricultural tasks of crop management such as irrigation , picking , pesticide spraying , and fertilization . The dominating cause is lack of diverse set of public benchmark datasets that are specifically designed for various agricultural missions. Some of the few rich datasets available are CropDeep for detection, multi-modal datasets like Rosette plant or \\\\textit{Arabidopsis} datasets , Sorghum-Head , Wheat-Panicle , Crop/Weed segmentation , and Crop/Tassle segmentation . Fig. \\\\ref{fig-cropdeep} contains some examples from the CropDeep dataset. Fig. \\\\ref{fig-rosette} depicts multi-modal annotations provided in the Rosette Plant Phenotyping dataset i.e., annotations for detection, segmentation, leaf center along with otherwise rarely found meta data. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.3]{figures/cropdeep.png}}\\n\\\\caption{Some annotation examples from CropDeep dataset .}\\n\\\\label{fig-cropdeep}\\n\\\\end{figure}\\nEfficient yield estimation from images is also one of the key tasks for farmers and plant breeders to accurately quantify the overall throughput of their ecosystem. Recent efforts in panicle or spike detection , leaf counting , fruit detection as well as pixel-wise segmentation-based tasks such as panicle segmentation show very promising results in this direction. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.7]{figures/RosettePlant.pdf}}\\n\\\\caption{Visual illustration of all types of annotations available in dataset.}\\n\\\\label{fig-rosette}\\n\\\\end{figure}\\n\\\\subsubsection*{Crop Disease and Pest Recognition}\\nModern technologies have given human society the ability to produce enough food to meet the demand of more than 7 billion people. However, food security remains threatened by a number of factors including climate change , the decline in pollinators , plant diseases , and others. Plant diseases are not only a threat to food security at the global scale, but can also have disastrous consequences for smallholder farmers whose livelihoods depend on healthy crops. India loses 35\\\\% of the annual crop yield due to plant diseases . In the developing world, more than 80 percent of the agricultural production is generated by smallholder farmers , and reports of yield loss of more than 50\\\\% due to pests and diseases are frequent . Furthermore, the largest fraction of hungry people (50\\\\%) live in smallholder farming households , making smallholder farmers a group that’s particularly vulnerable to pathogen-derived disruptions in food supply.\\nOwing to these factors, timely disease and pest recognition becomes a priority task for farmers. In addition to that, farmers do not have many options other than consulting other fellow farmers or seeking help from government funded helplines . Availability of public datasets such as PlantVillage , PlantDoc allowed for progress in the area of disease and pest detection. Recent research works in pest and insect detection , invasive species detection in marine aquaculture and disease detection in plant leafs , Rice , Tomato , Banana , Grape , Sugarcane , Eggplant , Cucumber , Soybean , Olive , Tea , Coffee and other similar works take encouraging steps towards disease-free agriculture. Fig. \\\\ref{fig-disease} depicts banana diseases and pest detection outputs from . This work reports solutions to extant limitations in plant disease detection. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.3]{figures/banana.png}}\\n\\\\caption{Detected classes and expected output of the trained disease detection model. \\\\textbf{a} Entire plant afected by banana bunchy top virus (BBTV), \\\\textbf{b} leaves affected by black sigatoka (BS), \\\\textbf{c} cut pseudostem of Xanthomonas wilt (BXW) afected plant showing yellow bacterial ooze, \\\\textbf{d} fruit bunch afected by Xanthomonas wilt (BXW), e cut fruit afected by Xanthomonas wilt (BXW), \\\\textbf{f} corm afected by banana corm weevil (BCW). Figure and description taken from .}\\n\\\\label{fig-disease}\\n\\\\end{figure}\",\n \"id\": \"7b5a3160-2cf3-4478-90b2-e4ffb2d28171\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 67,\n \"parent_id\": \"be6f63d3-6b39-458c-918e-c883b38ba7a8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Application of Deep Learning in Plant Phenotyping\"\n ],\n [\n \"subsection\",\n \"Ground-Based Remote Sensing for Plant Phenotyping\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Ground-Based Remote Sensing for Plant Phenotyping\"\n },\n {\n \"cite_extract_rate\": 0.17391304347826,\n \"cites\": [\n 8406,\n 1040,\n 1039,\n 1038\n ],\n \"content\": \"The past few decades have witnessed the great progress of unmanned aircraft vehicles (UAVs) in civilian fields, especially in photogrammetry and remote sensing. In contrast with the platforms of manned aircraft and satellite, the UAV platform holds many promising characteristics: flexibility, efficiency, high spatial/temporal resolution, low cost, easy operation, etc., which make it an effective complement to other remote-sensing platforms and a cost-effective means for remote sensing. We refer reader to literary works for the detailed reports of techniques and applications of UAVs in precision agriculture, remote sensing, search and rescue, construction and infrastructure inspection and discuss other market opportunities. UAVs can be utilized in precision agriculture (PA) for crop management and monitoring , weed detection , irrigation scheduling , agricultural pattern detection , pesticide spraying , cattle detection , disease detection , insect detection and data collection from ground sensors (moisture, soil properties, etc.,) . The deployment of UAVs in PA is a cost-effective and time saving technology which can help for improving crop yields, farms productivity and profitability in farming systems. Moreover, UAVs facilitate agricultural management, weed monitoring, and pest damage, thereby they help to meet these challenges quickly . \\nUAVs can also be utilized to monitor and quantify several factors of irrigation such as availability of soil water, crop water need (which represents the amount of water needed by the various crops to grow optimally), rainfall amount, efficiency of the irrigation system . In this work , UAVs are currently being utilized to estimate the spatial distribution of surface soil moisture high-resolution multi-spectral imagery in combination with ground sampling. UAVs are also being used for thermal remote sensing to monitor the spatial and temporal patterns of crop diseases during various disease development phases which reduces crop losses for farmers. This work detects early stage development of soil-borne fungus in UAV imagery. Soil texture can be an indicative of soil quality which in turn influences crop productivity. Hence, UAV thermal images are being utilized to quantify soil texture at a regional scale by measuring the differences in land surface temperature under a relatively homogeneous climatic condition . Accurate assessment of crop residue is crucial for proper implementation of conservation tillage practices since crop residues provide a protective layer on agricultural fields that shields soil from wind and water. In , the authors demonstrated that aerial thermal images can explain more than 95\\\\% of the variability in crop residue cover amount compared to 77\\\\% using visible and near IR images.\\nFarmers must monitor crop maturity to determine the harvesting time of their crops. UAVs can be a practical solution to this problem . Farmers require accurate, early estimation of crop yield for a number of reasons, including crop insurance, planning of harvest and storage requirements, and cash flow budgeting. In , UAV images were utilized to estimate yield and total biomass of rice crop in Thailand. In , UAV images were also utilized to predict corn grain yields in the early to midseason crop growth stages in Germany. \\nThere have also been successful efforts that seamlessly combine aerial and ground based system for precision agriculture . With relaxed flight regulations and drastic improvement in machine learning techniques, geo-referencing, mosaicing, and other related algorithms, UAVs can provide a great potential for soil and crop monitoring . More precision agricultural researches are encouraged to design and implement special types of cameras and sensors on-board UAVs, which have the ability of remote crop monitoring and detection of soil and other agricultural characteristics in real time scenarios.\",\n \"id\": \"a449abb1-6b4c-48cd-a598-c9b18f11bfc1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"be6f63d3-6b39-458c-918e-c883b38ba7a8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Application of Deep Learning in Plant Phenotyping\"\n ],\n [\n \"subsection\",\n \"Unmanned Aircraft Vehicles for Plant Phenotyping\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Unmanned Aircraft Vehicles for Plant Phenotyping\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The impact of climate change and its unforeseeable nature, has caused majority of the agricultural crops to be affected in terms of their production and maintenance. With more than seven billion mouths to feed greater demands are being put on agriculture than ever before, at the same time as land is being degraded by factors such as soil erosion, mineral exhaustion and drought. It becomes the utmost priority for governments to support farmers by providing crucial information about changing weather conditions, soil conditions and more. Currently, satellite imagery is making agriculture more efficient by reducing scouting efforts of farmers, by optimizing use of nitrogen based on variable rate of application, by optimizing water schedules, identifying field performance and benchmark fields, etc . India alone has 7 satellites specially designed for benefits of farmers .\\nSatellites and their imagery are being applied to agriculture in several ways, initially as a means of estimating crop yields and crop types , soil salinity, soil moisture, soil pH . Optical and radar sensors can provide an accurate picture of the acreage being cultivated, while also differentiating between crop types and determining their health and maturity. Optical satellite sensors can detect visible and near-infrared wavelengths of light, reflected from agricultural land below. It is these wavelengths which combined, can be manipulated to help us understand the condition of the crops. This information helps to inform the market, and provide early warning of crop failure or famine.\\nBy extension, satellites are also used as a management tool through the practice of PA, where satellite images are used to characterise a farmer's fields in detail, often used in combination with geographical information systems (GIS), to allow more intensive and efficient cultivation practices. For instance, different crops might be recommended for different fields while the farmer's use of fertiliser is optimised in a more economic and environmentally-friendly fashion. Providing access to satellite imagery also becomes very important for building trust among the involved parties (farmers and government and private bodies involved). Web-based platforms such as Google Earth Engine, Planet.com, Earth Data Search by NASA, LandViewer by Earth Observing System, Geocento and others provide access to past and present (even daily) satellite imagery of your interest. \\nAgricultural monitoring is also increasingly being applied to forestry, both for forest management and as a way of characterising forests as carbon sinks to help minimise climate change – notably as part of the UN's REDD programme .\",\n \"id\": \"3f927eff-6ed6-4137-949b-bb7a6ee40c3a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"be6f63d3-6b39-458c-918e-c883b38ba7a8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Application of Deep Learning in Plant Phenotyping\"\n ],\n [\n \"subsection\",\n \"Satellites for Plant Phenotyping\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Satellites for Plant Phenotyping\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7217,\n 1050,\n 1051,\n 1048,\n 1045,\n 1041,\n 1047,\n 1043,\n 1049,\n 1044,\n 1042,\n 1046,\n 7323\n ],\n \"content\": \"While deep learning based plant phenotyping has shown great promise, requirement of large labeled datasets still remains to be the bottleneck. Phenotyping tasks are often specific to the environmental and genetic conditions, finding large datasets with such conditions is not always possible. This results in researchers needing to acquire their own dataset and label it, which is often a arduous and expensive affair. Moreover, small datasets often lead to models that overfit. Deep learning approaches optimized for working with limited labeled data would immensely help the plant phenotyping community, since this would encourage many more farmers, breeders, and researchers to employ reliable plant phenotyping techniques to optimize crop yield. To this end, we list out some of the recent efforts in the area of deep plant phenotyping with limited labeled data.\\n\\\\subsection*{Data Augmentation}\\nThe computer vision community has long been employing dataset augmentation techniques to grow the amount of data using artificial transformations. Artificially perturbing the original dataset with affine transformations (e.g., rotation, scale, translation) is considered a common practice now. However, this approach has some constraints: the augmented data only capture the variability of the available training set (e.g., if the dataset doesn't include a unique colored fruit, the particular unique case will never be learnt). To overcome this, several data augmentation methods proposed take advantage of recent advancements in the image generation space. In this work , the authors use Generative Adversarial Network (GAN) to generate \\\\textit{Arabidopsis} plant images (called ARIGAN) with unique desirable traits (over 7 leaves) that were originally less frequent in the dataset. Fig. \\\\ref{fig-generated} \\\\textbf{(a)} shows examples of images generated by ARIGAN. Other latest works use more advanced variants of GANs to generate realistic plant images with particularly favorable leaf segmentations of interest to boost leaf counting accuracy of the learning models. In , the authors proposed an unsupervised image translation technique to improve plant disease recognition performance. LeafGAN , an image-to-image translation model, generates leaf images with various plant diseases and boosts diagnostic performance by a great margin. Two sets of example images generated by LeafGAN are shown in Fig. \\\\ref{fig-generated} \\\\textbf{(b)}. Other data enhancement techniques are also being employed by researchers to train plant disease diagnosis models on generated lesions . \\nThe effort to provide finely annotated data has enabled great improvement of the state of the art on segmentation performance. Some researches have started working on effectively transferring the knowledge obtained from RGB images on annotated plants either to other species or other modalities of imaging. In this work , the authors successfully transfer the knowledge gained from annotated leaves of \\\\textit{Arabidopsis thaliana} in RGB to images of the same plant in chlorophyll fluorescence imaging. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.55]{figures/generated.pdf}}\\n\\\\caption{\\\\textbf{(a)} shows \\\\textit{Arabidopsis} plant images generated by ARIGAN . Bottom-right numbers refer to the leaf count. \\\\textbf{(b)} shows two sets of healthy leafs and their corresponding disease prone leaves generated by LeafGAN .}\\n\\\\label{fig-generated}\\n\\\\end{figure}\\n\\\\subsection*{Weakly Supervised Learning}\\nFruit/organ counting is a well explored task by the plant phenotyping community. However, many vision-based solutions we have currently require highly accurate instance and density labels of fruits and organs in diverse set of environments. The labeling procedures are often very burdensome and error prone and, in many agricultural scenarios, it may be impossible to acquire a sufficient number of labelled samples to achieve consistent performance that are robust to image noise or other forms of covariate shift. This is why using only weak labels can be crucial for cost-effective plant phenotyping. \\nRecently, a lot of attention has been placed on engineering weakly supervised learning frameworks for plant phenotyping. In , the authors created a weakly supervised framework for the sorghum head detection task where annotators label the data only until the model reaches a desired performance level. After that, model outputs are directly passed as data labels leading to a exponential reduction in annotation costs with minimal loss in model accuracy. In other work , the authors proposed a strategy which is able to learn to count fruits without requiring task-specific supervision labels, such as manually labelled object bounding boxes or total instance count. In , the authors use a trained CNN on defect classification data and use it's activate maps to segment infected regions on potatoes. Segmentation task requires really rich labels (each pixel of the image is annotated) so this task effectively bypasses the labeling for segmentation altogether. On another note, rice heading date estimation greatly assists the breeders to understand the adaptability of the crop to various environmental and genetic conditions. Accurate estimation of heading date requires monitoring the increase in number of rice panicles in the crop. Detecting rice panicles from crop images usually requires training an object detection model such as Faster R-CNN or YOLO, which requires costly bounding box annotations. However, a recently proposed method uses a sliding window based detector which requires training an image classifier, for which annotations are much easier to obtain. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.3]{figures/bmvc.jpg}}\\n\\\\caption{\\\\textbf{(a)} Standard pool-based active learning framework \\\\textbf{(b)} Proposed framework which interleaves weak supervision in the active learning process. This novel framework includes an adaptive supervision module which allows switching to a stronger form of supervision as required when training the model. The \\\\textit{oracle} is the source of labels \\\\textit{a.k.a} annotator.}\\n\\\\label{fig-bmvc}\\n\\\\end{figure}\\n\\\\subsection*{Transfer Learning}\\nTransfer learning is a type of learning that enables using knowledge gained while solving one problem and applying it to a different but related problem i.e., a model trained on one phenotyping task (say potato leaf classification) being able to assist another phenotyping (tomato leaf classification) task. Transfer learning is a very well explored area of machine learning. As part of the first steps of adopting existing transfer learning techniques for plant phenotyping, the authors of use CNNs (AlexNet, GoogleNet and VGGNet) pretrained on ImageNet dataset and fine tune on the plant dataset used in LifeCLEF 2015 challenge. With the help of transfer learning, they were able to beat then existing state-of-the-art LifeCLEF performance by 15\\\\% points. Similary in , the authors report better than human results in segmentation task with the help of transfer learning where they transfer learn a model trained on peanut root dataset for switchgrass root dataset (they also report results using ImageNet pretrained models). Leaf disease detection and treatment recommendation performance is also shown to be boosted with transfer learning . In , the authors interestingly combined a State-of-the-Art weakly-supervised fruit counting model with an unsupervised style transfer method for fruit counting. They used Cycle-Generative Adversarial Network (C-GAN) to perform unsupervised domain adaptation from one fruit dataset to another and train it alongside with a Presence-Absence Classifier (PAC) that discriminates images containing fruits or not and ultimately achieved better performance than fully supervised models. \\n\\\\begin{figure}\\n\\\\centerline{\\\\includegraphics[scale=0.15]{figures/bmc.jpg}}\\n\\\\caption{Proposed point supervision framework into the pool-based active learning cycle. In this framework, strong supervision is queried for images only after deemed \\\\textit{informative} based on point supervision of those images.}\\n\\\\label{fig-bmc}\\n\\\\end{figure}\\n\\\\subsection*{Active Learning}\\nActive learning , an iterative training approach that curiously selects the best samples to train, has been shown to reduce labeled data requirement when training deep classification networks . Research in the area of active learning for object detection has been, arguably, limited. However, numerous plant phenotyping tasks such as detection and quantification of crop yield and fruit counting are directly dependent on object detection. Keeping this in mind, an active learning method has been proposed for training deep object detection models where the model can selectively query either weak labels (pointing at the object) or strong labels (drawing a box around the object). By introducing a switching module for weak labels and strong labels, the authors were able to save 24\\\\% of annotation time while training a wheat head detection model. Fig. \\\\ref{fig-bmvc} illustrates the difference between regular active learning cycle and proposed active learning cycle. This method demonstrates the applicability of active learning to plant phenotyping methods where obtaining labeled data is often difficult. Along the same lines, to alleviate the labeled data requirement for training object detection models for cereal crop detection, a weak supervision based active learning method was proposed recently. In this active learning approach, the model constantly interacts with a human annotator by iteratively querying the labels for only the most informative images, as opposed to all images in a dataset. Fig. \\\\ref{fig-bmc} visually illustrates the proposed framework. The active query method is specifically designed for cereal crops which usually tend to have panicles with low variance in appearance. This training method has been shown to reduce over 50\\\\% of annotation costs on sorghum head and wheat spike detection datasets. We expect to see more research works using active learning for limited labeled data based plant phenotyping in the near future.\",\n \"id\": \"fcb14fe1-1d8d-475b-8ed7-cf7547edb6ff\",\n \"level\": \"section\",\n \"origin_cites_number\": 28,\n \"parent_id\": \"40b8b988-fe57-4e55-9221-27ac302d3a6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Plant Phenotyping with Limited Labeled Data\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Plant Phenotyping with Limited Labeled Data\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we describe some of the challenges present in plant phenotyping methods which warrant further research. \\n\\\\subsection*{The Training Data Bottleneck} \\nModern phenotyping methods rely on deep learning which is notorious for requiring large amounts of labeled data. While some progress has been made in developing data efficient models for phenotyping, reducing the labeling efforts for training efficient phenotyping tools is still an open problem. We believe that effectively adapting techniques from deep learning such as unsupervised, self supervised, weakly supervised, active and semi-supervised learning will greatly benefit the phenotyping community in observing plant traits with small datasets. \\n\\\\subsection*{Explainability}\\nDeep neural networks are generally considered as black boxes which produce predictions without sufficient justification. This makes debugging a neural network difficult i.e., it can be tough to understand what caused a wrong prediction. Crop management decisions based on incorrect phenotyping results can cause financial losses. Hence, developing explainable models for plant phenotyping is one of the open problems in this field. Obtaining the reasons behind a given set of plant traits using explainable models has the potential to achieve breakthroughs in our understanding of plant behavior in various genetic and environmental conditions. \\n\\\\subsection*{Data collection}\\nVision based plant phenotyping suffers from challenges such as occlusion, inaccuracies in 3D reconstruction of crops and bad lighting conditions caused by the changing weather. It is therefore necessary to develop phenotyping tools which are robust to visual variations.\",\n \"id\": \"6bf986f3-5fd9-4afa-a45b-7964c94d86eb\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40b8b988-fe57-4e55-9221-27ac302d3a6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Challenges and Open Problems\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Challenges and Open Problems\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"High throughput plant phenotyping methods have shown great promise in efficiently monitoring crops for plant breeding and agricultural crop management. Research in deep learning has accelerated the progress in plant phenotyping research which resulted in the development of various image analysis tools to observe plant traits. However, wide applicability of high throughput phenotyping tools is limited by some issues such as 1) dependence of deep networks on large datasets, which are difficult to curate, 2) large variations of field environment which cannot always be captured, and 3) capital and maintenance which can be prohibitively expensive to be widely used in developing countries. With many open problems in plant phenotyping warranting further studies, it is indeed a great time to study plant phenotyping and achieve rapid progress by utilizing the advances in deep learning. \\n\\\\bibliographystyle{splncs}\\n\\\\bibliography{egbib}\\n\\\\end{document}\",\n \"id\": \"1cab82e7-7d3c-4e75-913b-94d66346fc31\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40b8b988-fe57-4e55-9221-27ac302d3a6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Computer Vision with Deep Learning for Plant Phenotyping in Agriculture: A Survey\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":78,"string":"78"},"all_cites":{"kind":"list like","value":[1034,486,1036,7043,7324,1037,1035,7323,8406,1040,1039,1038,7217,1050,1051,1048,1045,1041,1047,1043,1049,1044,1042,1046],"string":"[\n 1034,\n 486,\n 1036,\n 7043,\n 7324,\n 1037,\n 1035,\n 7323,\n 8406,\n 1040,\n 1039,\n 1038,\n 7217,\n 1050,\n 1051,\n 1048,\n 1045,\n 1041,\n 1047,\n 1043,\n 1049,\n 1044,\n 1042,\n 1046\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.059429035838205,"string":"1.059429"}}},{"rowIdx":44,"cells":{"authors":{"kind":"list like","value":["Yuqiao Liu","Yanan Sun","Bing Xue","Mengjie Zhang","Gary G. Yen","Kay Chen Tan"],"string":"[\n \"Yuqiao Liu\",\n \"Yanan Sun\",\n \"Bing Xue\",\n \"Mengjie Zhang\",\n \"Gary G. Yen\",\n \"Kay Chen Tan\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey on Evolutionary Neural Architecture Search"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-08-25T11:00:46Z","string":"2020-08-25T11:00:46Z"},"category":{"kind":"string","value":"cs.NE"},"abstract":{"kind":"string","value":" Deep Neural Networks (DNNs) have achieved great success in many applications. The architectures of DNNs play a crucial role in their performance, which is usually manually designed with rich expertise. However, such a design process is labour intensive because of the trial-and-error process, and also not easy to realize due to the rare expertise in practice. Neural Architecture Search (NAS) is a type of technology that can design the architectures automatically. Among different methods to realize NAS, Evolutionary Computation (EC) methods have recently gained much attention and success. Unfortunately, there has not yet been a comprehensive summary of the EC-based NAS algorithms. This paper reviews over 200 papers of most recent EC-based NAS methods in light of the core components, to systematically discuss their design principles as well as justifications on the design. Furthermore, current challenges and issues are also discussed to identify future research in this emerging field. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"40bf4670-36c8-4014-b074-8f86f999eff7","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"]],"subsections":["e9b844de-0b2f-4ecf-94b9-874602da0564","2857eea5-3541-4f2f-8282-08d5126a14b3","7fb1f2d3-d962-4d8c-8b81-c32933ee534a","13bf9942-9456-4154-adf6-fca4ab8f93d7","d3650a8a-21c2-4aa1-9b61-b69c455614a3","fa11702e-7516-4849-b9c7-1422e93b4202","390da3df-188a-48a9-991a-966d654bdd4d","e3dda7bf-8a69-4f93-8f3f-07b17269af8c","d338e930-f930-4ee2-9cea-b0e47084ac48"],"title":"root"},{"cite_extract_rate":0.5714285714285711,"cites":[96,1110,166,6822,514,8441,6821,6823,8247,97,6824,7,684,872,6820,6825],"content":"\\IEEEPARstart{D}{eep} Neural Networks (DNNs), as the cornerstone of deep learning~, have demonstrated their great success in diverse real-world applications, including image classification~, natural language processing~, speech recognition~, to name a few. The promising performance of DNNs has been widely documented due to their deep architectures~, which can learn meaningful features directly from the raw data almost without any explicit feature engineering. Generally, the performance of DNNs depends on two aspects: their architectures and the associated weights. Only when both achieve the optimal status simultaneously, the performance of the DNNs could be expected promising. The optimal weights are often obtained through a learning process: using a continuous loss function to measure the discrepencies between the real output and the desired one, and then the gradient-based algorithms are often used to minimize the loss. When the termination condition is satisfied, which is commonly a maximal iteration number, the algorithm can often find a good set of weights~. This kind of process has been very popular owing to its effectiveness in practice, and has become the dominant practice for weight optimization~, although they are principally local-search~ algorithms. On the other hand, obtaining the optimal architectures cannot be directly formulated by a continuous function, and there is even no explicit function to measure the process for finding optimal architectures.\nTo this end, there has been a long time that the promising architectures of DNNs are manually designed with rich expertise. This can be evidenced from the state of the arts, such as VGG~, ResNet~ and DenseNet~. These promising Convolutional Neural Network (CNN) models are all manually designed by the researchers with rich knowledge in both neural networks and image processing. However, in practice, most end users are not with such kinds of knowledge. Moreover, DNN architectures are often problem-dependent. If the distribution of the data is changed, the architectures must be redesigned accordingly. Neural Architecture Search (NAS), which aims to automate the architecture designs of deep neural networks, is identified as a promising way to address the challenge aforementioned. \nMathematically, NAS can be modeled by an optimization problem formulated by Equation~(\\ref{equ_defination_NAS}):\n\\begin{equation}\n\\label{equ_defination_NAS}\n\\left\\{\n\\centering\n\\begin{array}{c}\n\\argmin_A = \\mathcal{L}(A, \\mathcal{D}_{train}, \\mathcal{D}_{fitness}) \\\\\ns.t.\\quad A \\in \\mathcal{A} \\\\\n\\end{array}\n\\right.\n\\end{equation}\nwhere $\\mathcal{A}$ denotes the search space of the potential neural architectures, $\\mathcal{L}(\\cdot)$ measures the performance of the architecture $A$ on the fitness evaluation dataset $\\mathcal{D}_{fitness}$ after being trained on the training dataset $\\mathcal{D}_{train}$. The $\\mathcal{L}(\\cdot)$ is usually non-convex and non-differentiable~. In principle, NAS is a complex optimization problem experiencing several challenges, e.g., complex constraints, discrete representations, bi-level structures, computationally expensive characteristics and multiple conflicting criteria. NAS algorithms refer to the optimization algorithms which are specifically designed to effectively and efficiently solve the problem represented by Equation~(\\ref{equ_defination_NAS}). The initial work of NAS algorithms is commonly viewed as the work in~, which was proposed by Google. The pre-print version of this work was firstly released in the website of arXiv\\footnote{\\url{https://arxiv.org/abs/1611.01578}} in 2016, and then was formally accepted for publication by the International Conference on Learning Representations (ICLR) in 2017. Since then, a large number of researchers have been investing tremendous efforts in developing novel NAS algorithms.\nBased on the optimizer employed, existing NAS algorithms can be broadly classified into three different categories: Reinforcement Learning (RL)~ based NAS algorithms, gradient-based NAS algorithms, and Evolutionary Computation (EC)~ based NAS algorithms (ENAS). Specifically, RL based algorithms often require thousands of Graphics Processing Cards (GPUs) performing several days even on median-scale dataset, such as the CIFAR-10 image classification benchmark dataset~. Gradient-based algorithms are more efficient than RL based algorithms. However, they often find the ill-conditioned architectures due to the improper relation for adapting to gradient-based optimization. Unfortunately, the relation has not been mathematically proven. In addition, the gradient-based algorithms require to construct a supernet in advance, which also highly requires expertise. The ENAS algorithms solve the NAS by exploiting EC techniques. Specifically, EC is a class of population-based computational paradigms, simulating the evolution of species or the behaviors of the population in nature, to solve challenging optimization problems. In particular, Genetic Algorithms (GAs)~, Genetic Programming (GP)~, and Particle Swarm Optimization (PSO)~ are widely used EC methods in practice. Owing to the promising characteristics of EC methods in insensitiveness to the local minima and no requirement to gradient information, EC has been widely applied to solve complex non-convex optimization problems~, even when the mathematical form of the objective function is not available~.\nIn fact, the EC methods had been frequently used more than twenty years ago, searching for not only the optimal neural architectures but also the weights of neural networks simultaneously, which is also termed as neuroevolution~. The major differences between ENAS and neuroevolution lie in two aspects. Firstly, neuroevolution often uses EC to search for both the neural architectures and the optimal weight values, while ENAS for now focuses mainly on searching for the architectures and the optimal weight values are obtained by using the gradient-based algorithms\\footnote{Please note that we still categorize some existing algorithms as the ENAS algorithm, such as API~, EvoCNN~ and EvoDeep~, although they also concern the weights. This is because the optimal weight values of the DNNs searched by them are still obtained by the gradient-based algorithms. They only searched for the best weight initialization values or the best weight initialization method of the DNNs.} immediately after. Secondly, neuroevolution commonly applies to small-scale and median-scale neuron networks, while ENAS generally works on DNNs, such as the deep CNNs~ and deep stacked autoencoders~, which are stacked by the building blocks of deep learning techniques~. Generally, the first work of ENAS is often viewed as the LargeEvo algorithm~ which was proposed by Google who released its early version in March 2017 in arXiv. Afterwards, this paper got accepted into the 34th International Conference on Machine Learning in June 2017. The LargeEvo algorithm employed a GA to search for the best architecture of a CNN, and the experimental results on CIFAR-10 and CIFAR-100~ have demonstrated its effectiveness. Since then, a large number of ENAS algorithms have been proposed. Fig.~\\ref{fig_histogram} shows the number of similar works\\footnote{These ``submissions\" include the ones which have been accepted for publication after the peer-review process, and also the ones which are only available on the arXiv website without the peer-review process.} published from 2017 to 2020 when this survey paper was ready for submission. As can be seen from Fig.~\\ref{fig_histogram}, from 2017 to 2020, the number of submissions grows with multiple scales.\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.8\\linewidth]{histogram}\n\t\\caption{The number of submissions refers to the works of evolutionary neural architecture search. The data is collected from Google Scholar with the keywords of ``evolutionary\" OR ``genetic algorithm'' OR ``particle swarm optimization\" OR ``PSO\" OR ``genetic programming\" AND ``architecture search\" OR ``architecture design'' OR ``CNN'' OR ``deep learning'' OR ``deep neural network'' and the literature on Neural Architecture Search collected from the AutoML.org website by the end of 2020. With these initially collected data, we then have carefully checked each manuscript to make its scope accurately within the evolutionary neural architecture search.}\n\t\\label{fig_histogram}\n\t\\vspace{-0.3cm}\n\\end{figure}\nA large number of related submissions have been made available publicly, but there is no comprehensive survey of the literature on ENAS algorithms. Although recent reviews on NAS have been made in~, they mainly focus on reviewing different methods to realize NAS, rather than concentrating on the ENAS algorithms. Specifically, Elsken \\textit{et al.}~ divided NAS into three stages: search space, search strategy, and performance estimation strategy. Similarly, Wistuba \\textit{et al.}~ followed these three stages with an additional review about the multiple objectives in NAS. Darwish \\textit{et al.}~ made a summary of Swarm Intelligence (SI) and Evolutionary Algorithms (EAs) approaches for deep learning, with the focuses on both NAS and other hyperparameter optimization. Stanley \\textit{et al.} ~ went through a review of neuroevolution, aiming at revealing the weight optimization rather than the architecture of neural networks. Besides, most of the references in these surveys are pre-2019 and do not include an update on the papers published during the past two years when most ENAS works were published. This paper presents a survey involving a large number of ENAS papers, with the expectation to inspire some new ideas for enhancing the development of ENAS. To allow readers easily concentrating on the technical part of this survey, we also follow the three stages to introduce ENAS algorithms, which has been widely adopted by existing NAS survey papers~, but with essential modifications made to specifically suit ENAS algorithms.\nThe remainder of this paper is organized as follows. The background of ENAS is discussed in Section~\\ref{sec_background}. Section~\\ref{sec_encoding_space} documents different encoding spaces, initial spaces and search spaces of ENAS algorithms. In Section~\\ref{sec_architectureEncoding}, the encoding strategy and architecture representation are introduced. Section~\\ref{sec_Population_operators} summarizes the process of population updating, including evolutionary operators and selection strategies. Section~\\ref{sec_evaluation} shows multiple ways to speed up the evolution. Section~\\ref{sec_application} presents the applications of ENAS algorithms. Section~\\ref{sec_challenges_and_issues} discusses the challenges and prospects, and finally Section~\\ref{sec_conclusion} presents conclusions.","id":"e9b844de-0b2f-4ecf-94b9-874602da0564","level":"section","origin_cites_number":28,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec_background}\nAs discussed above, ENAS is a process of searching for the optimal architectures of DNNs by using EC methods. In this section, we will first introduce the unified flow chart of ENAS in Subsection~\\ref{sec2_3}. Then, the categories of EC methods based on their search strategies and the categories of DNN architectures in Subsections~\\ref{sec2_1} and \\ref{sec2_2}, respectively.","id":"2857eea5-3541-4f2f-8282-08d5126a14b3","level":"section","origin_cites_number":0,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Background"]],"subsections":["0869d4e9-c981-472b-b025-5a8bbce117b1","fb246415-ce9a-4259-827b-2d75a5822aa3","a76618db-7274-4a81-a316-4d0b05a79b1e"],"title":"Background"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec2_3}\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{flowchart}\n\t\\caption{The flowchart of a common ENAS algorithm.}\n\t\\label{fig_flowchart}\n\t\\vspace{-0.3cm}\n\\end{figure}\nFig.~\\ref{fig_flowchart} shows an illustration of the flowchart of an ENAS algorithm. Specifically, the evolutionary process takes place in the initial space and the search space sequentially. First of all, a population is initialized within the initial space that is defined in advance. Each individual in the population represents a solution for the ENAS, i.e., a DNN architecture. Each architecture needs to be encoded as an individual before it joins the population. Second, the fitness of the generated individuals is evaluated. Note that there are two fitness evaluation steps as shown in Fig.~\\ref{fig_flowchart}, which commonly employ the same evaluation criterion. Thirdly, after the fitness evaluation of the initial population, the whole population starts the evolutionary process within the search space, which is shown by the dashed box in Fig.~\\ref{fig_flowchart}. In the evolutionary process, the population is updated by the selection and the evolutionary operators in each iteration, until the stopping criterion is met. Please note that the selection stage is not necessary for some other EC paradigms like SI. Finally, a population that has finished the evolution is obtained. In the following sections, these key steps are documented in detail.","id":"0869d4e9-c981-472b-b025-5a8bbce117b1","level":"subsection","origin_cites_number":0,"parent_id":"2857eea5-3541-4f2f-8282-08d5126a14b3","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Background"],["subsection","Flow Chart of ENAS"]],"subsections":[],"title":"Flow Chart of ENAS"},{"cite_extract_rate":0.4,"cites":[9078,6826],"content":"\\label{sec2_1}\nENAS is distinguished from other NAS methods by its employed optimization approach, i.e., the EC methods, where the optimization approaches can be further subdivided based on the search strategy that the optimization approach adopted. Fig.~\\ref{fig_EC_taxonomy} provides such an illustration from three aspects: EAs, Swarm Intelligence (SI), and others, and more detailed statistics can be observed in Table~\\ref{table_EC}. In practice, the EA-based methods account for the majority of existing ENAS algorithms, where the GA takes a large part of EA-based methods. The other categories of EC methods are also important parts for realizing ENAS algorithms, such as GP, Evolutionary Strategy (ES), PSO, Ant Colony Optimization (ACO)~, Differential Evolution (DE)~, Firefly Algorithm (FA)~. In this paper, the Hill-Climbing Algorithm (HCA) is classified into EC paradigm because it can be regarded as an EA with a simple selection mechanism and without crossover operation~. HCA has been well known as a widely used local search algorithm in memetic algorithms~.\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=1\\linewidth]{taxonomy1}\n\t\\caption{The categories of ENAS from EC methods regarding the search strategies.}\n\t\\label{fig_EC_taxonomy}\n\t\\vspace{-0.3cm}\n\\end{figure}","id":"fb246415-ce9a-4259-827b-2d75a5822aa3","level":"subsection","origin_cites_number":5,"parent_id":"2857eea5-3541-4f2f-8282-08d5126a14b3","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Background"],["subsection","Evolutionary Search Strategy"]],"subsections":[],"title":"Evolutionary Search Strategy"},{"cite_extract_rate":0.47368421052631504,"cites":[6822,96,514,6821,870,8247,97,6827,166],"content":"\\label{sec2_2}\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{taxonomy2}\n\t\\caption{The categories of ENAS from neural network perspective.}\n\t\\label{fig_NN_taxonomy}\n\t\\vspace{-0.3cm}\n\\end{figure}\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=1\\linewidth]{NN}\n\t\\caption{Examples of different neural architectures in ENAS. (a) CNN. (b) DBN. (c) AE. (d) RNN. }\n\t\\label{fig_NN}\n\t\\vspace{-0.3cm}\n\\end{figure}\nThe end product of ENAS is all about DNNs, which can be broadly divided into five different categories: CNN, Deep Belief Network (DBN), Stacked Auto-Encoder (SAE), Recurrent Neural Network (RNN) and others. The brief fact can be seen from Fig.~\\ref{fig_NN_taxonomy}, and a detailed illustration will be shown in Table~\\ref{table_EC}. \nMost ENAS methods are proposed for searching for the optimal CNN architectures. This is because many hand-crafted CNNs, such as VGG~, ResNet~ and DenseNet~, have demonstrated their superiority in handling the image classification tasks which is the most successful applications in the deep learning field. Generally, the CNN architecture is composed of the convolutional layers, the pooling layers and the fully-connected layers, and a common example of CNNs is shown in Fig.~\\ref{fig_NN} (a). The optimization of CNN architecture is mainly composed of three aspects: the hyperparameters of each layer~, the depth of the architecture~ and the connections between layers~. In practice, the majority of the ENAS methods consider the above three aspects collectively~.\nDBN~ is made up by stacking multiple Restricted Boltzmann Machines (RBMs), and an example can be seen in Fig.~\\ref{fig_NN} (b). Specifically, RBMs have connections only between layers, while without any inter-layer connection. Meanwhile, RBMs allow the DBN to be trained in an unsupervised manner to obtain a good weight initialization. There are two commonly used types of hyperparameters needed to be optimized in DBN: the number of neurons in each layer~ and the number of layers~.\nAn SAE is constructed by stacking multiple of its building blocks which are AEs. An AE aims to learn meaningful representations from the raw input data by restoring its input data as its objective~. Generally, an AE is typically composed of two symmetrical components: the encoder and the decoder, and an example including both parts is shown in Fig.~\\ref{fig_NN} (c). Generally, some ENAS methods only take the encoder into evolution~ because the decoder part is symmetric with the encoder and can be derived accordingly. Yet, some other ENAS methods optimize the hyperparameters of encoders and decoders separately~.\nThe most significant difference between RNNs and the neural networks introduced above is its recurrent connection. Fig.~\\ref{fig_NN} (d) shows the time-expanded structure of an RNN, where the value of current hidden layer $h^{t}$ is influenced by the value at its previous time slot $h^{t-1}$ and the output value of its previous layer. Because these layers are reused, all the weights (i.e., \\textit{U}, \\textit{W} and \\textit{V} in the figure) are shared. Different from focusing on the number of neurons and the number of layers in the feedforward neural network, some ENAS methods concern about how many times the RNN should be unfold~. In addition, there remain other neural networks like typical DNNs, which are made up of only fully-connected layers, where the connections are formed by all the neurons in the two adjacent layers. Because such kind of DNNs is not the major target investigated by NAS, we will not introduce it in detail.\nWhen the ENAS algorithms are applied to these DNNs, the goal is to find the best architecture-related parameters. Specifically, for CNNs, they are the number of convolutional layers, pooling layers, fully-connected layers, and parameters related to these layers (such as the kernel size of the convolutional layers, the pooling type, and the number of neurons for fully-connected layers, and so on) as well as their connection situation (such as the dense connection and the skip connection). For DBN and SAE, they are the number of their building blocks, i.e., the RBM for DBN and the AE for SAE, and the number of neurons in each layer. For RNN, in addition to the architecture-related parameters mentioned above, the number of the time slot is also an important parameter to be optimized by ENAS algorithms. For the traditional DNNs, the ENAS algorithms often concern about the neuron number of each layer. In addition, some ENAS algorithms also concern the weights, such as the weight initialization method and weigh initial values. Table~\\ref{table_parameters} summarizes the detail of the common parameters optimized by ENAS methods in different types of neural networks, including the CNNs as the most popular type. The details of these ENAS algorithms will be documented in the following sections. \n\\begin{table}[]\n\t\\renewcommand\\arraystretch{0.8}\n\t\\caption{Common Parameters Optimized in Different DNNs.} \n\t\\label{table_parameters}\n\t\\centering\n\t\\begin{tabular}{p{1.2cm}|p{1.5cm}|p{4.5cm}}\n\t\t\\hline\n\t\t& \\multicolumn{2}{c}{Parameters} \\\\ \\hline\n\t\t\\multirow{4}{*}{CNN} & global parameters & number of layers, connections between layers \\\\ \\cline{2-3} \n\t\t& convolution layer & filter size (width and height), stride size (width and height), feature map size, convolution type, standard deviation and mean value of the filter elements \\\\ \\cline{2-3} \n\t\t& pooling layer & filter size (width and height), stride size (width and height), pooling type \\\\ \\cline{2-3} \n\t\t& fully-connected layer & number of neurons, standard deviation and mean value of weights \\\\ \\hline\n\t\tDBN, AE & \\multicolumn{2}{l}{number of hidden layers, neurons per layer} \\\\ \\hline\n\t\tRNN & \\multicolumn{2}{l}{\\shortstack[l]{number of hidden layers, neurons per layer, number of\\\\ time slot}} \\\\ \\hline\n\t\\end{tabular}\n\\vspace{-0.3cm}\n\\end{table}","id":"a76618db-7274-4a81-a316-4d0b05a79b1e","level":"subsection","origin_cites_number":19,"parent_id":"2857eea5-3541-4f2f-8282-08d5126a14b3","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Background"],["subsection","Common Neural Networks in ENAS"]],"subsections":[],"title":"Common Neural Networks in ENAS"},{"cite_extract_rate":0.330275229357798,"cites":[6851,6838,6837,6844,6841,8955,6829,6832,9079,6850,6836,6835,6843,6821,6846,8247,6847,9080,6834,6845,6842,6826,6830,6822,6848,870,6853,6840,6833,6852,6831,6849,6824,6827,6839,6828],"content":"\\label{sec_encoding_space}\n\\begin{table*}[]\n\t\\renewcommand\\arraystretch{0.8}\n\t\\caption{Different Encoding Space and the Constraints} \n\t\\label{table_constrains}\n\t\\vspace{-0.3cm}\n\t\\begin{center}\n\t\t\\begin{tabular}{p{3.8cm}|p{1.8cm}|p{2cm}|p{2.5cm}|p{5.5cm}}\n\t\t\t\\hline\n\t\t\t& Fixed depth & Rich initialization & Partial fixed structure & Relatively few constraints\\\\\n\t\t\t\\hline\n\t\t\tLayer-based encoding space & ~ & ~ & ~ & ~\\\\\n\t\t\t\\hline\n\t\t\tBlock-based encoding space & ~ & ~ & ~ & ~ \\\\\n\t\t\t\\hline\n\t\t\tCell-based encoding space & & & ~ & ~\\\\\n\t\t\t\\hline\n\t\t\tTopology-based encoding space & & ~ & & ~ \\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\vspace{-0.5cm}\n\\end{table*}\nThe encoding space contains all the valid individuals encoded in the population. In terms of the basic units, the encoding space can be divided into three categories according to the basic units they adopt. They are the layer-based encoding space, the block-based encoding space and the cell-based encoding space. In addition, some ENAS methods do not take the configuration of the basic unit into consideration, but care the connections between units. Such kind of encoding space is often called the topology-based encoding space.\nIn addition, the constraints on the encoding space are important. This is because they represent the human intervention which restricts the encoding space and lightens the burden of the evolutionary process. A method with a mass of constraints can obtain a promising architecture easily but prevent to design any novel architecture that does not follow the constraints. Furthermore, the different sizes of the search space would greatly affect the efficiency of the evolution. On the other hand, the effectiveness of the ENAS methods cannot be guaranteed if there is no constraint on the search space: one extreme case is that all the individuals in the search space are mediocre. In this case, an excellent individual can be obtained even without selection. Table~\\ref{table_constrains} shows different kinds of encoding spaces and the constraints of existing ENAS algorithms, where the first row shows the constraints and the first column displays the encoding space. In the following, we will discuss them at length.","id":"7fb1f2d3-d962-4d8c-8b81-c32933ee534a","level":"section","origin_cites_number":109,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Space"]],"subsections":["c45304b4-d500-4803-95c8-a1b032fcb338","6c612948-3e12-4637-8d86-3b3c5b4c17d5","0e3df321-149a-48ba-be12-da3f243c2af6"],"title":"Encoding Space"},{"cite_extract_rate":0.72,"cites":[96,6524,514,6821,6848,870,305,871,544,6831,97,6849,6842,872,6827,6839,6854,6826],"content":"The layer-based encoding space denotes that the basic units in the encoding space are the primitive layers, such as convolution layers and fully-connected layers. This would lead to a huge search space, since it tries to encode so much information in the search space. However, it may take more time to search for a promising individual because there are more possibilities to construct a well-performed DNN from the primitive layers. For example, a promising CNN is commonly composed of hundreds of primitive layers, accordingly, the search process will consume more time to search for such a deep architecture by iteratively stacking the primitive layers. In addition, the promising DNN may not be found if only with the primitive layers. For instance, the promising performance of ResNet is widely recognized due to its skip connections which cannot be represented by the primitive layers.\nTo alleviate the above problems, the block-based encoding space is developed, where various layers of different types are combined as the blocks to serve as the basic unit of the encoding space. Traditional blocks are ResBlock~, DenseBlock~, ConvBlock (Conv2d + BatchNormalization + Activation)~ and InceptionBlock~, etc. Specifically, layers in the blocks have a specific topological relationship, such as the residual connection in ResBlock and the dense connections in DenseBlock. These blocks have promising performance and often require fewer parameters to build the architecture. So it is principally easier to find a good architecture in the block-based encoding space compared to the layer-based encoding space. Some ENAS algorithms used these blocks directly, such as~, while other methods proposed different blocks for different purposes. For example, Chen \\textit{et al.}~ proposed eight blocks including ResBlock and InceptionBlock encoded in a 3-bit string, and used Hamming distance to determine the similar blocks. Song \\textit{et al.}~ proposed three residual-dense-mixed blocks to reduce the amount of computation due to the convolution operation of image super-resolution tasks. \nThe cell-based encoding space is similar to the block-based one, and can be regarded as a special case of the block-based space where all the blocks are the same. The ENAS algorithms employing this space build the architectures by stacking repeated motifs. Chu \\textit{et al.}~ divided the cell-based space into two independent parts: the micro part containing the parameters of cells, and the macro part defining the connections between different cells. To be more specific, the cell-based encoding space concentrates more on the micro part. Different from the block-based space, the layers in the cell can be combined more freely, and the macro part is always determined by human expertise~. Some widely used encoding spaces are classified in this category. For example, NAS-Bench-101~ and NAS-Bench-201~ search for different combinations of layers and connections in cells, and then stacked the cells sequentially. In addition, NASNet~ and DARTS~ search for two kinds of cells, namely normal cell and reduction cell, and each stacked cell is connected to the two previous cells. The cell-based space greatly reduces the size of the encoding space. This is because all the basic units in the encoding space are the same, and the number of parameters in terms of constructing the promising DNN is much fewer. However, Frachon \\textit{et al.}~ claimed that there is no theoretical basis for that the cell-based space can help to obtain a good architecture. \nIn contrast, the topology-based space does not consider the parameters or the structure of each unit (layer or block), yet they are only concerned about the connections between units. One classical example is the one-shot method which treats all the architectures as different subgraphs of a supergraph~. Yang \\textit{et al.}~ proposed CARS to search for the architecture by choosing different connections in the supergraph, and then the architecture was built by the chosen subgraph. CARS can be classified into the topology-based category because it aims at deciding whether or not to keep the connections in the supergraph. But from the perspective of building the architecture, CARS can also be classified into the cell-based category. This is because the subgraph cell was stacked several times to build the architecture. Another typical case is pruning. Wu \\textit{et al.}~ employed a shallow VGGNet~ on CIFAR-10, and the aim was to prune unimportant weight connections from the VGGNet.\nAs observed from Table~\\ref{table_constrains}, the constraints on the encoding space mainly focus on three aspects: fixed depth, rich initialization and partial fixed structure. The fixed depth means all the individuals in the population have the same depth. The fixed depth is a strong constraint and largely reduces the size of the encoding space. Please note that the fixed depth is different from the \\textit{fixed-length encoding strategy} which will be introduced in Section~\\ref{sec_architectureEncoding}. In Genetic CNN~, for example, the \\textit{fixed-length encoding strategy} only limits the maximum depth. The node which is isolated (no connection) is simply ignored. By this way, the individuals can obtain different depths. The second constraint is rich initialization (i.e., the \\textit{well-designed space} to be discussed in Section~\\ref{sec_initial_space}), and it is also a limitation in practice that requires a lot of expertise. In this case, the initialized architectures are manually designed, which goes against the original intention of NAS. The partial fixed structure means the architecture is partially settled. For example, in~, a max-pooling layer is added to the network after every set of four convolution layers.\nIn Table~\\ref{table_constrains}, the relatively few constraints category means that those methods have no restrictions like the three aspects discussed above. However, it does not imply there is no constraint. For example, in the classification task, the fully-connected layer is often used as the tail of the whole DNNs in some methods~. Moreover, the maximum length is predefined in many methods including both \\textit{fixed-length encoding strategy} methods~ and \\textit{variable-length encoding strategy} methods~ resulting in preventing the method from discovering a deeper architecture. Wang \\textit{et al.}~ tried to break the limit of maximum length by using a Gaussian distribution initialization mechanism. Irwin \\textit{et al.}~ broke the limit by using the evolutionary operators, the crossover operator and the mutation operator, to extend the depth to any size.\nGenerally, the encoding space can be served as the basis of the search space and the initial space. In practice, the encoding space is often the same as the search space, while the initial space is often a subspace of the encoding space. Fig.~\\ref{fig_spaces} shows the relationship between the three spaces. The search space is larger than the initial space when some manual constraints are added to the population initialization. When there is no such manual constrains, search space and initial space are equivalent. Furthermore, the initial space determines what kind of individuals may appear in the initial population, and the search space determines what kind of individuals may appear in the evolutionary search process, as illustrated in Fig.~\\ref{fig_flowchart}. In the following subsections, we will discuss the initial space and the search space in Subsections~\\ref{sec_initial_space} and~\\ref{sec_search_space}, respectively.\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.8\\linewidth]{spaces}\n\t\\caption{The relationship between encoding space, search space and initial space.}\n\t\\label{fig_spaces}\n\t\\vspace{-0.3cm}\n\\end{figure}","id":"c45304b4-d500-4803-95c8-a1b032fcb338","level":"subsection","origin_cites_number":25,"parent_id":"7fb1f2d3-d962-4d8c-8b81-c32933ee534a","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Space"],["subsection","Encoding Space and Constraints"]],"subsections":[],"title":"Encoding Space and Constraints"},{"cite_extract_rate":0.5714285714285711,"cites":[6829,8247,6821,6827],"content":"\\label{sec_initial_space}\nIn general, there are three types of architecture initialization approaches in the initial space: starting from trivial initial conditions~, randomly initialization in the encoding space~ and staring from a well-designed architecture (also termed as rich initialization)~. These three types of initialization correspond to three different initial spaces: \\textit{trivial space}, \\textit{random space} and \\textit{well-designed space}.\nThe \\textit{trivial space} contains only a few primitive layers. For example, the LargeEvo algorithm~ initialized the population in a \\textit{trivial space} where each individual constitutes just a single-layer model with no convolutions. Xie \\textit{et al.}~ experimentally demonstrated that a \\textit{trivial space} can evolve to a competitive architecture. The reason for using as little experience as possible is to justify the advantage of EC-based methods where some novel architectures can be discovered, and most of the discovery is different from the manually designed DNN architectures. On the contrary, the \\textit{well-designed space} contains the state-of-the-art architectures. In this way, a promising architecture can be obtained at the beginning of the evolution, whereas it can hardly evolve to other novel architectures. Actually, many of ENAS methods adopting this initial space focus on improving the performance upon the well-designed architecture. For example, the architecture pruning aims at compressing DNNs by removing less important connections~. For a \\textit{random space}, all the individuals in the initial population are randomly generated in the limited space, and it has been adopted by many methods, such as~. The aim of this type of initial spaces is also to reduce the intervention of human experience in the initial population.","id":"6c612948-3e12-4637-8d86-3b3c5b4c17d5","level":"subsection","origin_cites_number":7,"parent_id":"7fb1f2d3-d962-4d8c-8b81-c32933ee534a","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Space"],["subsection","Initial Space"]],"subsections":[],"title":"Initial Space"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec_search_space}\nAfter the initialization of the population, the ENAS methods start to search the architectures in the search space. Generally speaking, the search space is the same as the initial space when the random initial space is adopted. For other two types of initial spaces, however, due to the relatively small initial space, the search space will become much larger with the expectation that the promising architecture is included. It is worth noting that many methods do not directly define the search space, but restrict the search space by using evolutionary operators. For example, Irwin \\textit{et al.}~ did not specify the maximal depth, instead, they used the evolutionary operations to extend the architecture to any depth.","id":"0e3df321-149a-48ba-be12-da3f243c2af6","level":"subsection","origin_cites_number":1,"parent_id":"7fb1f2d3-d962-4d8c-8b81-c32933ee534a","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Space"],["subsection","Search Space"]],"subsections":[],"title":"Search Space"},{"cite_extract_rate":0.5,"cites":[6821,6827],"content":"\\label{sec_architectureEncoding}\nThis section will discuss how to encode a network architecture into an individual of the EC methods. Each ENAS method needs to determine its encoding strategy before starting the first stage of the ENAS method, i.e., the population initialization. The most intuitive difference in different encoding strategies of ENAS methods is the length of the encoded individuals. \nGenerally, the encoding strategies can be divided into two different categories according to whether the length of an individual changes or not during the evolutionary process. They are the \\textit{fixed-length encoding strategy} and the \\textit{variable-length encoding strategy}. Particularly, the individuals have the same length during the evolutionary process when the \\textit{fixed-length encoding strategy} is used. In contrast, the individuals are with different lengths during the evolutionary process if the \\textit{variable-length encoding strategy} is employed. The advantage of the \\textit{fixed-length encoding strategy} is that it is easy to use standard evolutionary operations which are originally designed for the individuals with equal lengths. In Genetic CNN~, for example, the fixed-length binary string helped the evolutionary operators (especially the crossover) with easy implementation. Another example is~, where Loni \\textit{et al.} used a fixed-length string of genes to represent the architecture which led to easy implementation of the one-point crossover in the corresponding encoded information. However, a proper maximal length must be predefined for the fixed-length encoding strategy. Because the maximal length relates to the optimal depth of the DNN architecture, which is unknown in advance, the corresponding ENAS algorithms still rely on expertise and experience. \nCompared with the \\textit{fixed-length encoding strategy}, the \\textit{variable-length encoding strategy} dose not require the human expertise regarding the optimal depth in advance, which has the potential to be fully automated. In addition, the advantage of this encoding strategy is that it can define more details of the architecture with freedom. For example, when solving a new task where there is no expertise in knowing the optimal depth of the DNN, we just initialize the individuals with a random depth, and the optimal depth can be found through the variable-length encoding strategy, where the depths of the corresponding DNN can be changed during the evolutionary process. However, the variable-length encoding strategy also brings some drawbacks. Because the traditional evolutionary operators might not be suitable for this kind of encoding strategy, the corresponding evolutionary operators need to be redesigned, where an example can be seen in~. Another disadvantage is that, due to the flexibility of variable length encoding, it could generally result in over-depth architectures, which sometimes further leads to more time-consuming fitness evaluation. Please note that some works claimed their use of the variable-length encoding strategy, where each individual is designed with a maximal length, and then the placeholder is used to indicate the gene's validation of the gene~. For example, the individual is designed to have $1,000$ genes, while some genes having the values of zeros do not participate in the evolutionary process. In this paper, we also categorize these methods into the fixed-length encoding strategy.\nIn addition, most of the DNN architectures can be represented as directed graphs which are made up of different basic units and the connections between the units. Therefore, the encoding for architecture can be divided into two aspects: configurations of basic units and connections, which will be discussed in the following subsections.","id":"13bf9942-9456-4154-adf6-fca4ab8f93d7","level":"section","origin_cites_number":4,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Strategy"]],"subsections":["c0d67ff4-2346-4359-81e9-cfc247725ee8","2960167f-5dd3-47f9-b647-47b27f2be6ee"],"title":"Encoding Strategy"},{"cite_extract_rate":0.6666666666666661,"cites":[6821,870],"content":"In practice, different basic units have different configurations, such as layers, blocks and cells, which are demonstrated in Section~\\ref{sec_encoding_space}. For example, in CNNs, there are multiple parameters in the primitive layers which can be seen in Table~\\ref{table_parameters}. As for the DenseBlock implemented by Sun \\textit{et al.}~, only two parameters are needed to build the block. Because the configuration of cell is more flexible than that of blocks in CNNs, which can be regarded as a microcosm of a complete neural network. For example, the cells in~ are made up by a combination of 10 layers selected from 8 different primitive layers. But the cells do not include some of the configurations of primitive layer, such as the feature map size which is an important parameter in the primitive layer~.","id":"c0d67ff4-2346-4359-81e9-cfc247725ee8","level":"subsection","origin_cites_number":3,"parent_id":"13bf9942-9456-4154-adf6-fca4ab8f93d7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Strategy"],["subsection","Encoding for Configurations of Basic Units"]],"subsections":[],"title":"Encoding for Configurations of Basic Units"},{"cite_extract_rate":0,"cites":[],"content":"When the parameters (configurations) of the basic units in the architecture have been determined, the corresponding architecture cannot be built up immediately. Since edges are indispensable in directed graph, connections are also part of the architecture. The connections discussed in this section include not only connections between basic units, but also the connections within the basic units.\nGenerally, the architectures of neural networks can be divided into two categories: \\textit{linear architecture} and \\textit{non-linear architecture}. The former denotes the architectures containing sequential basic units. The latter indicates that there are skip-connections or loop-connections in the architecture. Please note that, the structure could be the macro structure consisting of basic units, or the micro structure within the basic units.","id":"2960167f-5dd3-47f9-b647-47b27f2be6ee","level":"subsection","origin_cites_number":0,"parent_id":"13bf9942-9456-4154-adf6-fca4ab8f93d7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Strategy"],["subsection","Encoding for Connections"]],"subsections":["6fd9ba7f-54c9-40c3-908e-224aaa013157","a7c80d6f-1308-48d5-9a45-859a5cedcd95"],"title":"Encoding for Connections"},{"cite_extract_rate":0.75,"cites":[96,6821,97],"content":"The linear architecture can be found in different kinds of architectures including the one generated from the layer-based encoding space and the block-based encoding space. Its widespread use in ENAS stems from its simplicity. No matter how complex the internal of the basic units is, many ENAS methods stack the basic units one by one to build up the skeleton of the architecture which is linear. For example, in AE-CNN~, Sun \\textit{et al.} stacked different kinds of blocks to generate the architecture.\nOne special case is a linear architecture generated from the layer-based encoding space. In this case, there is no need to solely encode the connections, and only the parameters in each basic unit are enough to build an architecture. One classical example can be seen in~ where Sun \\textit{et al.} explored a great number of parameters based on a linear CNN architecture. However, most architectures are not designed to be linear. The skip connections in ResNet~ and the dense connections in DenseNet~ show the ability to build a good architecture.","id":"6fd9ba7f-54c9-40c3-908e-224aaa013157","level":"subsubsection","origin_cites_number":4,"parent_id":"2960167f-5dd3-47f9-b647-47b27f2be6ee","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Strategy"],["subsection","Encoding for Connections"],["subsubsection","Linear Architecture"]],"subsections":[],"title":"Linear Architecture"},{"cite_extract_rate":0.4,"cites":[870,6827],"content":"Firstly, we will introduce two approaches to encoding a non-linear architecture in this subsection. Specifically, the adjacent matrix is the most popular way to represent the connections for non-linear architectures. Genetic CNN~ used a binary string to represent the connection, and the string can be transformed into a triangular matrix. In the binary string, ``1'' denotes that there is a connection between the two nodes while ``0'' denotes no connection in between. Lorenzo \\textit{et al.}~ used a matrix to represent the skip connections, and this work revolved around the adjacent matrix. Back in the 1990s, Kitano \\textit{et al.}~ began to study the use of the adjacent matrix to represent network connection and explained the process from the connectivity matrix, to the bit-string genotype, to the network architecture phenotype. Another way to represent the connections is to use an ordered pair $G = (V,E)$ with vertices $V$ and edge $E$ associated with a direction, to represent a directed acyclic graph. Irwin \\textit{et al.}~ also used this strategy to encode the connections.\nSecondly, the non-linear architecture is a more common case in both the macro architecture and the micro architecture. AmoebaNet-A~, as an example of non-linear macro architecture, stacked these two kinds of cells for several times to build up an architecture. In addition, each cell receives two inputs from the previous two cells separately, which means that the direct connection and the skip-connection are both used in this macro structure. Also in AmoebaNet, each cell is a non-linear architecture inside, which means the micro architecture is also non-linear. The non-linear architecture provides the architecture with more flexibility, which makes it more likely to build up a promising architecture than that of the linear ones.","id":"a7c80d6f-1308-48d5-9a45-859a5cedcd95","level":"subsubsection","origin_cites_number":5,"parent_id":"2960167f-5dd3-47f9-b647-47b27f2be6ee","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Encoding Strategy"],["subsection","Encoding for Connections"],["subsubsection","Non-linear Architecture"]],"subsections":[],"title":"Non-linear Architecture"},{"cite_extract_rate":0.318840579710144,"cites":[6861,6851,6838,6837,6863,6844,6841,8955,6829,6832,6856,6857,9079,6858,6836,6835,6843,6860,6821,6846,6855,8247,6847,6864,6865,9080,6845,6842,6859,6826,6822,6830,6848,870,6853,6840,6833,6862,6831,6849,6824,6827,6839,6828],"content":"\\label{sec_Population_operators}\n\\begin{table*}[ht]\n\t\\renewcommand\\arraystretch{0.8}\n\t\\caption{Categorization of EC and Different Types of Neural Network} \n\t\\label{table_EC}\n\t\\vspace{-0.3cm}\n\t\\begin{center}\n\t\t\\begin{tabular}{|p{1cm}|p{0.5cm}|p{1.8cm}|p{4.5cm}|p{1.4cm}|p{2cm}|p{1.3cm}|p{2cm}|}\n\t\t\t\\hline\n\t\t\t\\multicolumn{3}{|l|}{} & CNN & DBN & RNN & AE & Others \\\\ \\hline\n\t\t\t\\multirow{12}{*}{\\shortstack[l]{Single\\\\objective}} & \\multirow{3}{*}{EA} & GAs &~ & ~ & ~ & ~ & ~ \\\\ \\cline{3-8} \n\t\t\t& & GP & ~ & & ~ & ~ & ~ \\\\ \\cline{3-8} \n\t\t\t& & ES & ~ & & ~ & ~ & ~ \\\\ \\cline{2-8} \n\t\t\t& \\multirow{3}{*}{SI} & ACO & ~ & & ~ & & \\\\ \\cline{3-8} \n\t\t\t& & PSO & ~ & ~ & & ~ & ~ \\\\ \\cline{2-8} \n\t\t\t& \\multirow{6}{*}{Other} & Memetic & ~ & & & & \\\\ \\cline{3-8} \n\t\t\t& & DE & ~ & & ~ & ~ & ~ \\\\ \\cline{3-8} \n\t\t\t& & HCA & ~ & & & & \\\\ \\cline{3-8} \n\t\t\t& & CVOA & & & ~ & & \\\\ \\cline{3-8} \n\t\t\t& & Hyper-heuristic & & ~ & & & \\\\ \\cline{3-8} \n\t\t\t& & FA & ~ & & & & \\\\ \\cline{3-8} \n\t\t\t& & AIS & ~ & & & & \\\\ \\hline\n\t\t\t\\multirow{2}{*}{\\shortstack[l]{Multi-\\\\objective}} & \\multicolumn{2}{l|}{EA} & ~ & ~ & ~ & & ~ \\\\ \\cline{2-8} \n\t\t\t& \\multicolumn{2}{l|}{SI} & ~ & ~ & & & \\\\ \\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\\vspace{-0.5cm}\n\\end{table*}\nThis section discusses the population updating process as shown in Fig.~\\ref{fig_flowchart}. Generally, the population updating varies greatly among existing ENAS algorithms because they may employ different EC methods which are with different updating mechanisms. Table~\\ref{table_EC} shows the ENAS algorithms which are classified according to the EC methods that they employ and different types of DNNs that they target at. Obviously, the EA-based ENAS algorithms dominate the ENAS. To be more specific, the GA-based ENAS is the most popular approach which largely owes to the convenience of architecture representation in GA. As a result, we give a detailed introduction to the EA-based ENAS including the selection strategy at first. Immediately after, we present a summary of the SI-based ENAS methods and others separately. The corresponding multi-objective ENAS algorithms are also introduced at the end of each subsection.","id":"d3650a8a-21c2-4aa1-9b61-b69c455614a3","level":"section","origin_cites_number":138,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Population Updating"]],"subsections":["a152dfdd-9b9e-46a8-8504-9097a22a6736","f362ecc4-6be9-45c6-b9e4-700fb0393594","670f0571-da2a-4b7f-ab3e-86d8a80bfc64"],"title":"Population Updating"},{"cite_extract_rate":0.4,"cites":[6851,6843,6869,6821,870,9079,6868,6867,6837,8247,6847,6863,6853,9080,6844,6841,6834,6845,6829,6831,6849,6824,6866,6842,6827,6839,6832,8302,6858,6826],"content":"\\begin{table}[]\n\t\\renewcommand\\arraystretch{0.8}\n\t\\caption{Selection Strategy} \n\t\\label{table_selection_strategy}\n\t\\vspace{-0.3cm}\n\t\\begin{center}\n\t\t\\begin{tabular}{p{2cm}|p{5cm}}\n\t\t\t\\hline\n\t\t\tElitism & ~ \\\\\n\t\t\t\\hline\n\t\t\tDiscard the worst or the oldest & ~\\\\\n\t\t\t\\hline\n\t\t\tRoulette & ~ \\\\\n\t\t\t\\hline\n\t\t\tTournament selection & ~\\\\\n\t\t\t\\hline\n\t\t\tOthers & ~ \\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\\vspace{-0.5cm}\t\n\\end{table}\nThe dash box in Fig.~\\ref{fig_flowchart} shows the general flow of population updating in EA-based ENAS. In this section, we will introduce the selection strategy and the evolutionary operations which collectively update the population. Specifically, the first stage of population updating is selection. The selection strategies can be classified into several types, where Table~\\ref{table_selection_strategy} shows the main kinds of selection strategies. Note that the selection strategy can be not only used in choosing individuals as parents to generate offspring with the evolutionary operators, but also used in the environmental selection stage which chooses individuals to make up the next population. Zhang \\textit{et al.}~ termed these two selections as mate selection and environmental selection separately.\nExisting selection strategies can be divided into five categories: elitism, discarding the worst, roulette wheel selection, tournament selection and others. The simplest strategy is elitism which retains the individuals with higher fitness. However, this can cause a loss of diversity in the population, which may lead the population falling into local optima. Discarding the worst is similar to elitism, which removes the individuals with poor fitness values from the population. Real \\textit{et al.}~ used the aging evolution which discards the oldest individual in the population. Aging evolution can explore the search space more, instead of zooming in on good models too early, as non-aging evolution would. The same selection strategy was also used in~. Zhu \\textit{et al.}~ combined these two approaches to discarding the worst individual and the oldest individual at the same time. Roulette wheel selection gives every individual a probability according to its fitness among the population to survive (or be discarded), regardless it is the best or not. Tournament selection selects the best one from an equally likely sampling of individuals. Furthermore, Johner \\textit{et al.}~ used a ranking function to choose individuals by rank. A selection trick termed as niching was used in~ to avoid stacking into local optima. This trick allows offspring worse than parent to survive for several generations until evolving to a better one.\nMost of the methods focus on preserving the well-performed individuals, however, Liu \\textit{et al.}~ emphasized the genes more than the survived individuals, where a gene can represent any components in the architecture. They believed the individuals which consist of the fine-gene set are more likely to have promising performance.\nSome selection methods aim at preserving the diversity of the population. Elsken \\textit{et al.}~ selected individuals in inverse proportion to their density. Javaheripi \\textit{et al.}~ chose the parents based on the distance (difference) during the mate selection. They chose two individuals with the highest distance to promote the exploration search.\nIn terms of evolutionary operations, mutation and crossover are two of the most commonly used operations in EA-based ENAS algorithms. Particularly, the mutation is only performed on a single individual, while the crossover takes two individuals to generate offspring.\nThe mutation operator aims to search the global optimum around the individual. A simple idea is to allow the encoded information to vary from a given range. Sun \\textit{et al.}~ used the polynomial mutation~ on the parameters of layers which are expressed by real numbers.\nTo make mutation not random, Lorenzo \\textit{et al.}~ proposed a novel Gaussian mutation based on a Gaussian regression to guide the mutation, i.e., the Gaussian regression can predict which architecture may be good, and the newly generated individuals are sampled in the regions of the search space, where the fitness values are likely to be high. This makes mutation to have a ``direction\".\nMoreover, Maziarz \\textit{et al.}~ used an RNN to guide the mutation operation. In this work, the mutation operations were not sampled at random among the possible architectural choices, but were sampled from distributions inferred by an RNN. Using an RNN to control the mutation operation can also be seen in other methods such as~.\nSome researches investigated the diversity of the population after mutation. Qiang \\textit{et al.}~ used a variable mutation probability. They used a higher probability in the early stage for better exploration and a lower probability in the later stage for better exploitation. It has been effectively applied to many other methods~. To maintain the diversity of the population after the mutation operation, Tian \\textit{et al.}~ used \\textit{force mutation} and distance calculation, which ensures the individual in the population is not particularly similar to other individuals (especially the best one). Kramer \\textit{et al.}~ used the (1+1)-evolutionary strategy that generates an offspring based on a single parent with bit-flip mutation, and used a mutation rate to control and niching to overcome local optima.\nThe intensive computational cost of ENAS presents a bottleneck which will be discussed in later sections. To reduce the unaffordable computational cost and time, some kinds of the mutation have also been designed. Zhang \\textit{et al.}~ proposed an exchange mutation which exchanges the position of two genes of the individual, i.e., exchanging the order of layers. This will not bring new layers and the weights in neural networks can be completely preserved, which means that the offspring do not have to be trained from scratch. \nChen \\textit{et al.}~ introduced two function-preserving operators for DNNs, and these operators are termed as network morphisms~. The network morphisms aim to change the DNN architecture without the loss of the acquired experience. The network morphisms change the architecture from $F(\\cdot)$ to $G(\\cdot)$, which satisfies the condition formulated by Equation~(\\ref{equ_function-preserving}):\n\\begin{equation}\n\\label{equ_function-preserving}\n\\forall x,\\quad F(x)=G(x)\n\\end{equation}\nwhere $x$ denotes the input of the DNN. The network morphisms can be regarded as a function-preserving mutation operation. With this operation, the mutated individuals will not have worse performance than their parents. To be more specific, Chen \\textit{et al.}~ proposed \\textit{net2widernet} to obtain a wider net and \\textit{net2deepernet} to obtain a deeper net. Elsken \\textit{et al.}~ extended the network morphisms with two popular network operations: skip connections and batch normalization. Zhu \\textit{et al.}~ proposed five well-designed function-preserving mutation operations to guide the evolutionary process by the information which have already learned. To avoid local optimal, Chen \\textit{et al.}~ added noises in some function-preserving mutation, and in the experiment, they found that by adding noises to pure network morphism, instead of compromising the efficiency, it by contrast, improved the final classification accuracy.\nPlease note that all the network morphisms can only increase the capacity of a network because if one would decrease the network’s capacity, the function-preserving property could not be guaranteed~. As a result, the architecture generated by network morphisms is only going to get larger and deeper, which is not suitable for a device with limited computing resources, such as a mobile phone. In order for the network architecture to be reduced, Elsken \\textit{et al.}~ proposed the approximate network morphism, which satisfies Equation~(\\ref{equ_approximate_network_morphism}):\n\\begin{equation}\n\\label{equ_approximate_network_morphism}\n\\forall x,\\quad F(x)\\approx G(x)\n\\end{equation}\nFor the crossover operator, the single-point crossover~ is the most popular method in EA-based ENAS~ because of its implementation simplicity. However, single-point crossover can typically apply to two individuals with equal lengths only. Therefore, this cannot be applied to variable-length individuals. To this end, Sun \\textit{et al.}~ proposed an efficient crossover operator for individuals with variable lengths. Sapra \\textit{et al.}~ proposed a disruptive crossover swapping the whole cluster (a sequence of layers) between both individuals at the corresponding positions rather than only focusing on the parameters of layers. Sun \\textit{et al.}~ used the Simulated Binary Crossover (SBX)~ to do a combination of the encoded parameters from two matched layers. Please note that the encoded parameters after SBX are not the same as that of both parents, which are quite different from other crossover operators.\nEAs for multi-objective ENAS is gaining more and more attention from researchers. The single objective ENAS algorithms are always concerned about only one objective, e.g., the classification accuracy, and these algorithms have only one goal: searching for the architecture with the highest accuracy. In general, most of the multi-objective ENAS algorithms aim at dealing with both the performance of the neural network and the number of parameters simultaneously~.\nHowever, these objective functions are often in conflict with each other. For example, getting a higher accuracy often requires a more complicated architecture with the need of more computational resources. On the contrary, a device with limited computational resource, e.g., a mobile phone, cannot afford such sophisticated architectures. \nThe simplest way to tackle the multi-objective optimization problem is by converting it into a single objective optimization problem with weighting factors, i.e., the weighted summation method. The Equation~(\\ref{equ_multi-fitness})\n\\begin{equation}\n\\label{equ_multi-fitness}\nF = \\lambda f_{1} + (1-\\lambda) f_{2}\n\\end{equation}\nis the classical linear form to weight two objective functions $f_{1},f_{2}$ into a single objective function, where the $\\lambda \\in (0, 1)$ denotes the weighting factor. In~, the multi-objective optimization problem was solved by using the available single objective optimization methods by Equation~(\\ref{equ_multi-fitness}) of the weighted summation. Chen \\textit{et al.}~ did not adopt the linear addition as the objective function, whereas using a nonlinear penalty term. However, the weights manually defined may incur bias~.\nSome algorithms have been designed and widely used in multi-objective optimization, such as NSGA-II~, and MOEA/D~, which have been also used in ENAS methods such as~. These methods aim to find a Pareto-font set (or non-dominant set). Only these methods are in the multi-objective category of Table~\\ref{table_EC}. Some researches have made improvements on these multi-objective optimization methods for better use in ENAS.\nBaldeon \\textit{et al.}~ chose the penalty based boundary intersection approach in MOEA/D because training a neural network involves nonconvex optimization and the form of Pareto Font is unknown.\nLEMONADE~ divided the objective function into two categories: $f_{exp}$ and $f_{cheap}$. $f_{exp}$ denotes the expensive-to-evaluate objectives (e.g., the accuracy), while $f_{cheap}$ denotes the cheap-to-evaluate objectives (e.g., the model size). In every iteration, they sampled parent networks with respect to sparsely distribution based on the cheap objectives $f_{cheap}$ to generate offspring. Therefore, they evaluated $f_{cheap}$ more times than $f_{exp}$ to save time. Schoron \\textit{et al.}~ also took the use of the LEMONADE proposed by Elsken \\textit{et al.}~.\nDue to that NSGA-III~ may fall into the small model trap (this algorithm prefers small models), Yang \\textit{et al.}~ have made some improvements to the conventional NSGA-III for favoring larger models.","id":"a152dfdd-9b9e-46a8-8504-9097a22a6736","level":"subsection","origin_cites_number":75,"parent_id":"d3650a8a-21c2-4aa1-9b61-b69c455614a3","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Population Updating"],["subsection","EAs for ENAS"]],"subsections":[],"title":"EAs for ENAS"},{"cite_extract_rate":0.18181818181818102,"cites":[6859,6835],"content":"PSO is inspired by the bird flocking or fish schooling~, and is easy to implement compared with other SI algorithms.\nJunior \\textit{et al.}~ used their implementation of PSO to update the particles based on the layer instead of the parameters of the layer. Gao \\textit{et al.}~ developed a gradient-priority particle swarm optimization algorithm to handle issues including the low convergence efficiency of PSO when there are a large number of hyper-parameters to be optimized. They expected the particle to find the locally optimal solution at first, and then move to the global optimal solution.\nFor ACO, the individuals are generated in a quite different way. Several ants are in an ant colony\\footnote{The population in ACO also termed as colony.}. Each ant moves from node to node following the pheromone instructions to build an architecture. The pheromone is updated every generation. The paths of well-performed architecture will maintain more pheromone to attract the next ant for exploitation and at the same time, the pheromone is also decaying (i.e., pheromone evaporation), which encourages other ants to explore other areas. Byla \\textit{et al.}~ let the ants choose the path from the node to node in a graph whose depth increases gradually. Elsaid \\textit{et al.}~ introduced different ant agent types to act according to specific roles to serve the needs of the colony, which is inspired by the real ants species.\nSI for multi-objective ENAS started only in the last two years and the research of this field is scarce which can be seen from Table~\\ref{table_EC}.\nLi \\textit{et al.}~ used the bias-variance framework on their proposed multi-objective PSO to get a more accurate and stable architecture.\nWu \\textit{et al.}~ used the MOPSO~ for neural networks pruning. The $G_{best}$ is selected according to the crowding distance in the non-dominant solutions set. Wang \\textit{et al.}~ used the OMOPSO~, which selects the leaders using a crowding factor and the $G_{best}$ is selected from the leaders. To better control the balance between convergence and diversity, Jiang \\textit{et. al.}~ proposed a MOPSO/D algorithm based on an adaptive penalty-based boundary intersection.","id":"f362ecc4-6be9-45c6-b9e4-700fb0393594","level":"subsection","origin_cites_number":11,"parent_id":"d3650a8a-21c2-4aa1-9b61-b69c455614a3","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Population Updating"],["subsection","SI for ENAS"]],"subsections":[],"title":"SI for ENAS"},{"cite_extract_rate":0.42857142857142805,"cites":[6831,6865,6826],"content":"Different from GAs, the mutation of DE exploits the information from three individuals. \nSome ENAS methods like~ chose DE to guide the offspring generation. However, there is little difference between different DE-based ENAS algorithms.\nWang \\textit{et al.}~ proposed a hybrid PSO-GA method. They used PSO to guide the evolution of the parameters in each block encoded in decimal notation. Meanwhile, using GA to guide the evolution of the shortcut connections is encoded in binary notation. Because PSO performs well on continuous optimization and GA is suitable for optimization with binary values, this hybrid method can search architectures effectively.\nHCA can be interpreted as a very simple evolutionary algorithm. For example, in~ the evolutionary operators only contain mutation and no crossover, and the selection strategy is relatively simple. The memetic algorithm is the hybrids of EAs and local search. Evans \\textit{et al.}~ integrated the local search (as gradient descent) into GP as a fine-tuning operation. The CVOA~ was inspired by the new respiratory virus, COVID-19. The architecture was found by simulating the virus spreads and infecting healthy individuals. Hyper-heuristic contains two levels: high-level strategy and low-level heuristics, and a domain barrier is between these two levels. Hence the high-level strategy is still useful when the application domain is changed. AIS was inspired by theories related to the mammal immune system and do not require the crossover operator compared to the GA~.","id":"670f0571-da2a-4b7f-ab3e-86d8a80bfc64","level":"subsection","origin_cites_number":7,"parent_id":"d3650a8a-21c2-4aa1-9b61-b69c455614a3","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Population Updating"],["subsection","Other EC Techniques for ENAS"]],"subsections":[],"title":"Other EC Techniques for ENAS"},{"cite_extract_rate":0.33333333333333304,"cites":[6822,6851,6843,6821,6848,9079,6846,8247,6871,6841,6833,6845,8955,6862,6831,6842,6872,6870,1461,6826,6835],"content":"\\label{sec_evaluation}\nIn this section, we will discuss the strategies to improve the efficiency of evaluations, with the consideration that the evaluation is often the most time-consuming stage of ENAS algorithms~.\nReal \\textit{et al.}~ used 250 computers to finish the LargeEvo algorithm over 11 days. Such computational resources are not available for everyone interested in NAS. Almost all of the methods evaluate individuals by training them first and evaluating them on the validation/test dataset. Since the architecture is becoming more and more complex, it will take a lot of time for training each architecture to convergence. So it is needed to investigate new methods to shorten the evaluation time and reduce the dependency on large amounts of computational resources.\nTable~\\ref{table_shorten_time} lists five of the most common methods to reduce the time: weight inheritance, early stopping policy, reduced training set, reduced population, and population memory. We would like to introduce the five kinds of methods first, then other kinds of promising methods next, and finally the surrogate-assisted methods at the end of this section.\nBecause the evolutionary operators do not completely disrupt the architecture of an individual, some parts of the newly generated individuals are the same as their parents. The weights of the same parts can be easily inherited. With the weight inheritance, the neural networks no longer need to be trained completely from scratch. This method has been used in~ 20 years ago.\nMoreover, as mentioned in Section~\\ref{sec_Population_operators}, the network morphisms change the network architecture without loss of the acquired experience. This could be regarded as the ultimate weight inheritance because it solved the weight inheritance problem in the changed architecture part. The ultimate weight inheritance lets the new individual completely inherit the knowledge from their parents, which will save a lot of time.\nThe early stopping policy is another method which has been used widely in NAS. The simplest way is to set a fixed and relatively small number of training epochs. This method is used in~, where training the individuals after a small number of epochs is sufficient. Similarly, Assunccao \\textit{et al.}~ let the individuals undergo the training for the same and short time each epoch (although this time is not fixed and will increase with the epoch), to allow the promising architectures have more training time to get a more precise evaluation, So \\textit{et al.}~ set hurdles after a fixed number of epochs. The weak individuals stop training early to save time.\nHowever, the early stopping policy can lead to inaccurate estimation about individuals' performance (especially the large and complicated architecture), which can be seen in Fig.~\\ref{fig_learning_curve}. In Fig.~\\ref{fig_learning_curve}, \\emph{individual2} performs better than \\emph{individual1} before epoch \\emph{t1}, whereas \\emph{individual1} performs better in the end. Yang \\textit{et al.}~ also discussed this phenomenon. So, it is crucial to determine at which point to stop.\nNote that neural networks can converge or hardly improve its performance after several epochs, as seen in the \\emph{t1} for \\emph{individual2} and the \\emph{t2} for \\emph{individual1} in Fig.~\\ref{fig_learning_curve}. Using the performance estimated at this point can evaluate an individual's relatively accurately with less training time. Therefore, some methods such as~ stopped training when observing there is no significant performance improvement.\nSuganuma \\textit{et al.}~ used the early stopping policy based on a reference curve. If the accuracy curve of an individual is under the reference curve for successive epochs, then the training will be terminated and this individual is regarded as a poor one. After every epoch, the reference curve is updated by the accuracy curve of the best offspring. \nThe reduced training set, i.e., using a subset of data that assuming similar properties to a large dataset, can also shorten the time effectively. Liu \\textit{et al.}~ explored promising architectures by training on a subset and used transfer learning on the large original dataset. Because there are so many benchmark datasets in the image classification field, the architecture can be evaluated on a smaller dataset (e.g., CIFAR-10) first and then applied on a large dataset (such as CIFAR-100 and ImageNet~). The smaller dataset can be regarded as the proxy for the large one.\nThe reduced population is a unique method of ENAS since other NAS approaches do not have a population. Assunccao \\textit{et al.}~ reduced the population based on their previous algorithm~ to speed up the evolution. However, simply reducing the population may not explore the search space well in each epoch and may lose the global search ability. Another way is reducing the population dynamically. For instance, Fan \\textit{et al.}~ used the ($\\mu+\\lambda$) evolution strategy and divided the evolution into three stages with the population reduction, which aims to find the balance between the limited computing resources and the efficiency of the evolution. The large population in the first stage is to ensure the global search ability, while the small population in the last stage is to shorten the evolution time. Instead of reducing the population, Liu \\textit{et al.}~ evaluated the downsizing architecture with smaller size at an early stage of evolution. Similarly, Wang \\textit{et al.}~ did not evaluate the whole architecture but starting with a single block, and then the blocks were stacked to build architecture as evolution proceeds.\n\\begin{table}[]\n\t\\renewcommand\\arraystretch{0.8}\n\t\\caption{Different Methods to Shorten the Evaluation Time} \n\t\\label{table_shorten_time}\n\t\\vspace{-0.2cm}\n\t\\begin{center}\n\t\t\\begin{tabular}{p{2.5cm}|p{5cm}}\n\t\t\t\\hline\n\t\t\tWeight inheritance & ~ \\\\\n\t\t\t\\hline\n\t\t\tEarly stopping policy & ~\\\\\n\t\t\t\\hline\n\t\t\tReduced training set & ~ \\\\\n\t\t\t\\hline\n\t\t\tReduced population & ~\\\\\n\t\t\t\\hline\n\t\t\tPopulation memory & ~\\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\\vspace{-0.3cm}\n\\end{table}\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.8\\linewidth]{learning_curve}\n\t\\caption{Two learning curve of two different individuals.}\n\t\\label{fig_learning_curve}\n\t\\vspace{-0.2cm}\n\\end{figure}\nThe population memory is another category of the unique methods of ENAS. It works by reusing the corresponding architectural information that has previously appeared in the population. In the population-based methods, especially the GA based methods (e.g., in~), it is natural to maintain well-performed individuals in the population in successive generations. Sometimes, the individuals in the next population directly inherit all the architecture information of their parents without any modification, so it is not necessary to evaluate the individuals again. Fujino \\textit{et al.}~ used memory to record the fitness of individuals, and if the same architecture encoded in an individual appears, the fitness value is retrieved from memory instead of being reevaluated. Similarly, Miahi \\textit{et al.}~ and Sun \\textit{et al.}~ employed a hashing method for saving pairs of architecture and fitness of each individual and reusing them when the same architecture appears again. Johner \\textit{et al.}~ prohibited the appearance of architectures that have appeared before offspring generation. This does reduce the time, however, the best individuals are forbidden to remain in the population which may lead the population evolve towards a bad direction.\nThere are many other well-performed methods to reduce the time in ENAS. Rather than training thousands of different architectures, one-shot model~ trained only one SuperNet to save time. Different architectures, i.e., the SubNets, are sampled from the SuperNet with the shared parameters. Yang \\textit{et al.}~ believed the traditional ENAS methods without using SuperNet which were less efficient for models to be optimized separately. In contrast, the one-shot model optimizes the architecture and the weights alternatively. But the weight sharing mechanism brings difficulty in accurately evaluating the architecture. Chu \\textit{et al.}~ scrutinized the weight-sharing NAS with a fairness perspective and demonstrated the effectiveness. However, there remain some doubts that cannot explain clearly in the one-shot model. The weights in the SuperNet are coupled. It is unclear why inherited weights for a specific architecture are still effective~.\nMaking the use of hardware can reduce the time, too. Jiang \\textit{et al.}~ used a distributed asynchronous system which contained a major computing node with 20 individual workers. Each worker is responsible for training a single block and uploading its result to the major node in every generation. Wang \\textit{et al.}~ designed an infrastructure which has the ability to leverage all of the available GPU cards across multiple machines to concurrently perform the objective evaluations for a batch of individuals.\nColangelo \\textit{et al.}~ designed a reconfigurable hardware framework that fits the ENAS. As they claimed, this was the first work of conducting NAS and hardware co-optimization.\nFurthermore, Lu \\textit{et al.}~ adopted the concept of proxy models, which are small-scale versions of the intended architectures. For example, in a CNN architecture, the number of layers and the number of channels in each layer are reduced. However, the drawback of this method is obvious: the loss of prediction accuracy. Therefore, they performed an experiment to determine the smallest proxy model that can provide a reliable estimate of performance at a larger scale.\nAll the above methods obtain the fitness of individuals by directly evaluating the performance on the validation dataset. An alternative way is to use indirect methods, namely the performance predictors. As summarized in~, the performance predictors can be divided into two categories: performance predictors based on the learning curve and end-to-end performance predictors, both of which are based on the training-predicting learning paradigm. This does not mean the performance predictor does not undergo the training phase at all, while it means learning from the information obtained in the training phase, and uses the knowledge learned to make a reasonable prediction for other architectures.\nRalwal \\textit{et al.}~ took the use of the learning curve based predictor, where the fitness is not calculated in the last epoch but is predicted by the sequence of fitness from first epochs. Specifically, they used a Long Short Term Memory (LSTM)~ as a sequence to sequence model, predicting the final performance by using the learning results of the first several epochs on the validation dataset.\nSun \\textit{et al.}~ adopted a surrogate-assisted method which is called end-to-end performance predictor. The predictor does not need any extra information about the performance of individuals to be evaluated. The performance predictor is essentially a regression model mapping from the architecture to its performance. The regression model needs to be trained with sufficient training data pairs first, and each pair consists of architecture and its corresponding performance. Specifically, they chose random forest~ as the regression model to accelerate the fitness evaluations in ENAS. When the random forest receives a newly generated architecture as input, the adaptive combination of a huge number of regression trees which have been trained in advance in the forest gives the prediction.","id":"fa11702e-7516-4849-b9c7-1422e93b4202","level":"section","origin_cites_number":63,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Efficient Evaluation"]],"subsections":[],"title":"Efficient Evaluation"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec_application}\nThis section discussed different application fields of ENAS has involved. Generally, the ENAS algorithms can be applied to wherever DNNs can be applied. Table~\\ref{table_application} shows the wide range of applications and Table~\\ref{table_CIFAR} displays the performance of extraordinary ENAS methods on two popular and challenging datasets for image classification tasks, namely CIFAR-10 and CIFAR-100. Both of these two tables can show what ENAS has achieved so far.","id":"390da3df-188a-48a9-991a-966d654bdd4d","level":"section","origin_cites_number":0,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Applications"]],"subsections":["06bc3345-4a23-45a7-a0e1-7338ed3b2e6b","ce4d03d7-51ce-4271-828b-0f60a203e534"],"title":"Applications"},{"cite_extract_rate":0.3671875,"cites":[6851,6838,6837,6863,6844,6841,8955,6829,6876,6832,6857,9079,6875,6874,6850,6870,6836,6835,6869,6843,6860,6821,6846,8247,6847,6864,6873,6865,9080,6834,6845,6842,6866,6830,6822,6848,870,6853,6840,6852,6833,6862,6831,6849,6827,6839,6828],"content":"Table~\\ref{table_application} shows the applications of existing ENAS algorithms, which contains a wide range of real-world applications. \n\\begin{table}[]\n\t\\renewcommand\\arraystretch{0.8}\n\t\\caption{Applications of Existing ENAS algorithms.} \n\t\\label{table_application}\n\t\\begin{tabular}{|p{1cm}|p{3.5cm}|p{3cm}|}\n\t\t\\hline\n\t\tCategory & Applications & References \\\\ \\hline\n\t\t(1) & Image classification & ~ \\\\ \\hline\n\t\t(1) & Image to image & ~ \\\\ \\hline\n\t\t(1) & Emotion recognition & ~ \\\\ \\hline\n\t\t(1) & Speech recognition & ~ \\\\ \\hline\n\t\t(1) & Language modeling & ~ \\\\ \\hline\n\t\t(1) & Face De-identification & ~ \\\\ \\hline\n\t\t(2) & Medical image segmentation & ~ \\\\ \\hline\n\t\t(2) & Malignant melanoma detection & ~ \\\\ \\hline\n\t\t(2) & Sleep heart study & ~ \\\\ \\hline\n\t\t(2) & Assessment of human sperm & ~ \\\\ \\hline\n\t\t(3) & Wind speed prediction & ~ \\\\ \\hline\n\t\t(3) & Electricity demand time series forecasting & ~ \\\\ \\hline\n\t\t(3) & Traffic flow forecasting & ~ \\\\ \\hline\n\t\t(3) & Electricity price forecasting & ~ \\\\ \\hline\n\t\t(3) & Car park occupancy prediction & ~ \\\\ \\hline\n\t\t(3) & Energy consumption prediction & ~ \\\\ \\hline\n\t\t(3) & Time series data prediction & ~ \\\\ \\hline\n\t\t(3) & Financial prediction & ~ \\\\ \\hline\n\t\t(3) & Usable life prediction & ~ \\\\ \\hline\n\t\t(3) & Municipal waste forecasting & ~ \\\\ \\hline\n\t\t(4) & Engine vibration prediction & ~ \\\\ \\hline\n\t\t(4) & UAV & ~ \\\\ \\hline\n\t\t(4) & Bearing fault diagnosis & ~ \\\\ \\hline\n\t\t(4) & Predicting general aviation flight data & ~ \\\\ \\hline\n\t\t(5) & Crack detection of concrete & ~ \\\\ \\hline\n\t\t(5) & Gamma-ray detection & ~ \\\\ \\hline\n\t\t(5) & Multitask learning & ~ \\\\ \\hline\n\t\t(5) & Identify Galaxies & ~ \\\\ \\hline\n\t\t(5) & Video understanding & ~ \\\\ \\hline\n\t\t(5) & Comics understanding & ~ \\\\ \\hline\n\t\\end{tabular}\n\\vspace{-0.5cm}\n\\end{table}\nGenerally, these applications can be grouped into the following five different categories:\n(1) Image and signal processing, including image classification which is the most popular and competitive field, image to image processing (including image restoration, image denoising, super-resolution and image inpainting), emotion recognition, speech recognition, language modelling, and face de-identification.\n(2) Biological and biomedical tasks, including medical image segmentation, malignant melanoma detection, sleep heart study, and assessment of human sperm.\n(3) Predictions and forecasting about all sorts of data, including the prediction of wind speed, car park occupancy, time-series data, financial and usable life, the forecasting of electricity demand time series, traffic flow, electricity price, and municipal waste.\n(4) Engineering, including engine vibration prediction, Unmanned Aerial Vehicle (UAV), bearing fault diagnosis and predicting general aviation flight data.\n(5) Others, including crack detection of concrete, gamma-ray detection, multitask learning, identify galaxies, video understanding, and comics understanding.","id":"06bc3345-4a23-45a7-a0e1-7338ed3b2e6b","level":"subsection","origin_cites_number":128,"parent_id":"390da3df-188a-48a9-991a-966d654bdd4d","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Applications"],["subsection","Overview"]],"subsections":[],"title":"Overview"},{"cite_extract_rate":0.538461538461538,"cites":[6869,6851,6877,6821,870,6837,8247,6847,6863,6878,6844,6834,6845,6829,6831,97,6842,6827,8319,6836,6826],"content":"\\label{sec_CIFAR}\n\\begin{table*}[!htbp]\n\t\\renewcommand\\arraystretch{0.8}\n\t\\caption{The Comparison of the Classification Error Rate on CIFAR-10, CIFAR-100 and ImageNet}\n\t\\label{table_CIFAR}\n\t\\vspace{-0.2cm}\n\t\\begin{tabular}{|p{3.7cm}|p{1.3cm}|p{1.8cm}|p{2.4cm}|p{2.4cm}|p{3cm}|p{0.5cm}|}\n\t\t\\hline\n\t\t\\textbf{ENAS Methods} & \\textbf{GPU Days} & \\textbf{Parameters(M)} & \\textbf{CIFAR-10(\\%)} & \\textbf{CIFAR-100(\\%)} & {\\textbf{ImageNet(Top1/Top5 \\%)}} & \\textbf{Year} \\\\ \\hline\n\t\tCGP-DCNN~ & --- & 1.1 & 8.1 & --- & --- & 2018 \\\\ \\hline\n\t\tEPT~ & 2 & --- & 7.5 & --- & --- & 2020 \\\\ \\hline\n\t\t\\multirow{2}{*}{GeNet~} & \\multirow{3}{*}{17} & --- & 7.1 & --- & --- & \\multirow{3}{*}{2017} \\\\ \\cline{3-6}\n\t\t& & --- & --- & 29.03 & --- & \\\\ \\cline{3-6}\n\t\t& & 156 & --- & --- & {27.87 / 9.74} &\n\t\t\\\\ \\hline\n\t\tEANN-Net~ & --- & --- & 7.05 $\\pm$ 0.02 & --- & --- & 2019 \\\\ \\hline\n\t\t\\multirow{2}{*}{DeepMaker~} & \\multirow{2}{*}{3.125} & 1 & 6.9 & --- & --- &\\multirow{2}{*}{2020} \\\\ \\cline{3-6}\n\t\t& & 1.89 & --- & 24.87 & --- & \\\\ \\hline\n\t\t\\multirow{2}{*}{GeneCai (ResNet-50)~ } & \\multirow{2}{*}{0.024} & --- & 6.4 & --- & --- & \\multirow{2}{*}{2020} \\\\ \\cline{3-6}\n\t\t& & --- & --- & --- & {25.7 / 7.9} & \\\\ \\hline\n\t\tCGP-CNN (ConvSet)~ & --- & 1.52 & 6.75 & --- & --- &2017 \\\\ \\hline\n\t\tCGP-CNN (ResSet)~ & --- & 1.68 & 5.98 & --- & --- &2017 \\\\ \\hline\n\t\tMOPSO/D-Net~ & 0.33 & 8.1 & 5.88 & --- & --- &2019 \\\\ \\hline\n\t\tReseNet-50 (20\\% pruned)~ & --- & 6.44 & 5.85 & --- & --- &2018 \\\\ \\hline\n\t\tImmuNeCS~ & 14 & --- & 5.58 & --- & --- &2019 \\\\ \\hline\n\t\t\\multirow{2}{*}{EIGEN~} & 2 & 2.6 & 5.4 & --- & --- & \\multirow{2}{*}{2018} \\\\ \\cline{2-6}\n\t\t& 5 & 11.8 & --- & 21.9 & --- & \\\\ \\hline\n\t\t\\multirow{2}{*}{LargeEvo~} & \\multirow{2}{*}{2750} & 5.4 & 5.4 & --- & --- & \\multirow{2}{*}{2017} \\\\ \\cline{3-6}\n\t\t& & 40.4 & --- & 23 & --- & \\\\ \\hline\n\t\t\\multirow{2}{*}{CGP-CNN (ConvSet)~} & 31 & 1.5 & 5.92 (6.48 $\\pm$ 0.48) & --- & --- & \\multirow{4}{*}{2019} \\\\ \\cline{2-6}\n\t\t& --- & 2.01 & --- & 26.7 (28.1 $\\pm$ 0.83) & --- & \\\\ \\cline{1-6}\n\t\t\\multirow{2}{*}{CGP-CNN (ResSet)~} & 30 & 2.01 & 5.01 (6.10 $\\pm$ 0.89) & --- & --- & \\\\ \\cline{2-6}\n\t\t& --- & 4.6 & --- & 25.1 (26.8 $\\pm$ 1.21) & --- & \\\\ \\hline\n\t\tMOCNN~ & 24 & --- & 4.49 & --- & --- & 2019 \\\\ \\hline\n\t\t\\multirow{2}{*}{NASH~} & 4 & 88 & 4.4 & --- & --- & \\multirow{2}{*}{2017} \\\\ \\cline{2-6}\n\t\t& 5 & 111.5 & --- & 19.6 & --- & \\\\ \\hline\n\t\tHGAPSO~ & 7+ & --- & 4.37 & --- & --- & 2018 \\\\ \\hline\n\t\t\\multirow{3}{*}{DPP-Net~,~} & \\multirow{3}{*}{2} & 11.39 & 4.36 & --- & --- & \\multirow{3}{*}{2018} \\\\ \\cline{3-6}\n\t\t& & 0.45 & 5.84 & --- & --- & \\\\ \\cline{3-6}\n\t\t& & 4.8 & --- & --- & {26.0 / 8.2} & \\\\ \\hline\n\t\t\\multirow{2}{*}{AE-CNN~} & 27 & 2 & 4.3 & --- & --- &\\multirow{2}{*}{2019} \\\\ \\cline{2-6}\n\t\t& 36 & 5.4 & --- & 20.85 & --- & \\\\ \\hline\n\t\tSI-ENAS~ & 1.8 & --- & 4.07 & 18.64 & --- & 2020 \\\\ \\hline\n\t\tEPSOCNN~ & 4- & 6.77 & 3.69 & --- & --- & 2019 \\\\ \\hline\n\t\t\\multirow{2}{*}{Hierarchical Evolution~ } & \\multirow{2}{*}{300} & --- & 3.63 $\\pm$ 0.10 & --- & --- & \\multirow{2}{*}{2017} \\\\ \\cline{3-6}\n\t\t& & --- & --- & --- & {20.3 / 5.2} & \\\\ \\hline\n\t\t\\multirow{2}{*}{EA-FPNN~} & 0.5 & 5.8 & 3.57 & --- & --- & \\multirow{2}{*}{2018} \\\\ \\cline{2-6}\n\t\t& 1 & 7.2 & --- & 21.74 & --- & \\\\ \\hline\n\t\t\\multirow{2}{*}{AmoebaNet-A~} & \\multirow{2}{*}{3150} & 3.2 & 3.34 $\\pm$ 0.06 & --- & --- & \\multirow{2}{*}{2018} \\\\ \\cline{3-6}\n\t\t&& 86.7 & --- & --- & {17.2 / 3.9} & \\\\ \\hline\n\t\tFirefly-CNN~ & --- & 3.21 & 3.3 & 22.3 & --- & 2019 \\\\ \\hline\n\t\t\\multirow{3}{*}{CNN-GA~} & 35 & 2.9 & 3.22 & --- & --- & \\multirow{3}{*}{2018} \\\\ \\cline{2-6}\n\t\t& 40 & 4.1 & --- & 20.53 & --- & \\\\ \\cline{2-6}\n\t\t& 35 & --- & --- & --- & {25.2 / 7.7} & \\\\ \\hline\n\t\t\\multirow{3}{*}{JASQNet~} & \\multirow{3}{*}{3} & 3.3 & 2.9 & --- & --- & \\multirow{3}{*}{2018} \\\\ \\cline{3-6}\n\t\t& & 1.8 & 2.97 & --- & --- & \\\\ \\cline{3-6} \n\t\t& & 4.9 & --- & --- & {27.2 / ---} & \\\\ \\hline\n\t\t\\multirow{2}{*}{RENASNet~} & \\multirow{2}{*}{6} & 3.5 & 2.88 $\\pm$ 0.02 & --- & --- & \\multirow{2}{*}{2018} \\\\ \\cline{3-6}\n\t\t& & 5.36 & --- & --- & {24.3 / 7.4} & \\\\ \\hline\n\t\tNSGA-Net~ & 4 & 3.3 & 2.75 & --- & --- & 2018 \\\\ \\hline\n\t\t\\multirow{4}{*}{CARS~} & \\multirow{4}{*}{0.4} & 2.4 & 3 & --- & --- & \\multirow{4}{*}{2019} \\\\ \\cline{3-6}\n\t\t& & 3.6 & 2.62 & --- & --- & \\\\ \\cline{3-6} \n\t\t& & 3.7 & --- & --- & {27.2 / 9.2} & \\\\ \\cline{3-6}\n\t\t& & 5.1 & --- & --- & {24.8 / 7.5} & \\\\ \\hline\n\t\t\\multirow{3}{*}{LEMONADE~} & \\multirow{3}{*}{80} & 13.1 & 2.58 & --- & --- & \\multirow{3}{*}{2018} \\\\ \\cline{3-6}\n\t\t& & 0.5 & 4.57 & --- & --- & \\\\ \\cline{3-6}\n\t\t& & --- & --- & --- & {28.3 / 9.6} & \\\\ \\hline\n\t\t\\multirow{2}{*}{EENA~} & \\multirow{2}{*}{0.65} & 8.47 & 2.56 & --- & --- & \\multirow{2}{*}{2019} \\\\ \\cline{3-6}\n\t\t& & 8.49 & --- & 17.71 & --- & \\\\ \\hline\n\t\tEEDNAS-NNMM~ & 0.5 & 4.7 & 2.55 & --- & --- & 2019 \\\\ \\hline\n\t\t\\multirow{5}{*}{NSGANet~} & \\multirow{5}{*}{27} & 0.2 & 4.67 & --- & --- & \\multirow{5}{*}{2019} \\\\ \\cline{3-6}\n\t\t& & 4 & 2.02 & --- & --- & \\\\ \\cline{3-6}\n\t\t& & 0.2 & --- & 25.17 & --- & \\\\ \\cline{3-6}\n\t\t& & 4.1 & --- & 14.38 & --- & \\\\ \\cline{3-6}\n\t\t& & 5.0 & --- & --- & {23.8 / 7.0} & \\\\ \\hline\n\t\\end{tabular}\n\\vspace{-0.3cm}\n\\end{table*}\nIn Table~\\ref{table_application}, it is obvious to see that many ENAS methods are applied to the image classification tasks. The benchmark dataset, CIFAR-10 which contains a total of ten classes, and the CIFAR-100 is the advanced dataset including a hundred classes. These two datasets have been widely used in image classification tasks, and the accuracy on these two challenging datasets can represent the ability of the architecture searched. In addition, ImageNet~ is a more challenging benchmark dataset, which contains 1,000 classes and more than one million images. Because CIFAR-10 and CIFAR-100 are relatively small and easy to be over-fitting nowadays~, the results of the ENAS methods on ImageNet are also included. We have collected the well-performed ENAS methods tested on these three datasets and showed the results in Table~\\ref{table_CIFAR}, where the methods are ranked in an ascending order of their best accuracy on CIFAR-10, i.e., the methods are ranked in a descending order of their error rate based on the conventions. The data shown under column ``CIFAR-10\" and ``CIFAR-100\" denotes the error rate of each method on the corresponding datasets. Especially, as for ImageNet, we report both top1 and top5 error rates. Furthermore, the ``GPU Days\", which was initially proposed in~, denotes the total search time of each method, it can be calculated by Equation~(\\ref{equ_GPUDays})\n\\begin{equation}\n\\label{equ_GPUDays}\nGPU\\ Days = The\\ number\\ of\\ GPUs \\times t\n\\end{equation}\nwhere the $t$ denotes the number of days that each method searched for. In Table~\\ref{table_CIFAR} ``Parameters\" denotes the total number of parameters which can represent the capability of architecture and the complexity. In addition, the symbol ``---\" in Table~\\ref{table_CIFAR} implies there is no result publically reported by the corresponding paper. The year reported in this table is its earliest time made public. Furthermore, there are additional notes provided for Table~\\ref{table_CIFAR}. Firstly, if there are several results reported from literature, such as CGP-DCNN~ and CARS~, we choose to report one or two of the most representative architectures. Secondly, we name two algorithms without proper names in Table~\\ref{table_CIFAR} (i.e., Firefly-CNN~ and EEDNAS-NNMM~) based on the EC methods they used and the first letter of the title. Thirdly, we report the classification errors of CGP-CNN~ in the format of ``best (mean $\\pm$ std)\". Finally, the symbols of ``+\" and ``-\" in the column of ``GPU Days\" denote the meaning of ``more than\" and ``less than\", respectively.\nIn principle, there is no totally fair comparison. Due to the following two reasons: (1) The encoding space including the initial space and the search space is different from each other. There are two extreme cases in the initial space: trivial initialization which starts at the simplest architecture and rich initialization which starts from a well-designed architecture (e.g., ResNet-50~). Meanwhile the size of the search space is largely different, e.g., Ref~ only takes the kernel size into the search space. (2) Different tricks exploited in the methods, e.g., the ``cutout\", can make the comparisons unfair, too. The ``cutout\" refers to a regularization method~ used in the training of CNNs, which could improve the final performance appreciably.\nTable~\\ref{table_CIFAR} shows the progress of ENAS for image classification according to the accuracy on CIFAR-10: LargeEvo algorithm~ (5.4\\%, 2017), LEMONADE~ (2.58\\%, 2018), and NSGANet~ (2.02\\%, 2019). Many ENAS methods have a lower error rate than ResNet-110~ with 6.43\\% error rate on CIFAR-10, which is a manually well-designed architecture. Therefore, the architecture found by ENAS can reach the same level or exceed the architecture designed by experts. It shows that the ENAS is reliable and can be used in other application fields.","id":"ce4d03d7-51ce-4271-828b-0f60a203e534","level":"subsection","origin_cites_number":39,"parent_id":"390da3df-188a-48a9-991a-966d654bdd4d","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Applications"],["subsection","Comparisons on CIFAR-10, CIFAR-100 and ImageNet"]],"subsections":[],"title":"Comparisons on CIFAR-10, CIFAR-100 and ImageNet"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec_challenges_and_issues}\nDespite the positive results of the existing ENAS methods, there are still some challenges and issues which need to be addressed. In the following, we will briefly discuss them.","id":"e3dda7bf-8a69-4f93-8f3f-07b17269af8c","level":"section","origin_cites_number":0,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Challenges and Issues"]],"subsections":["d6c13ce5-ab8c-4d29-a77d-718dbcbb0cf9","8f512a41-ca51-4724-a511-f83f8ff395ee","0511a9dd-9340-4ba1-ba4f-5c1c82725c9e","97b2a2f7-bddb-45d3-819a-f3dcbad41cba","ed6f44db-5133-4077-be4f-e9675dc00180","14a9080c-0216-4f7c-899b-00f8965feb35"],"title":"Challenges and Issues"},{"cite_extract_rate":0.7142857142857141,"cites":[8955,6831,8247,6880,870,6854,6825,6879,6844,6830],"content":"The effectiveness of ENAS is questioned by many researchers. Wistuba \\textit{et al.}~ noticed that the random search can get a well-performed architecture and has proven to be an extremely strong baseline. Yu \\textit{et al.}~ showed the state-of-the-art NAS algorithms performed similarly to the random policy on average. Liashchynskyi \\textit{et al.}~ compared with grid search, random search, and GA for NAS, which showed that the architecture obtained by GA and random search have similar performance. There is no need to use complicated algorithms to guide the search process if random search can outperform NAS based on EC approaches.\nHowever, the evolutionary operator in~ only contains a recombination operator, which limited the performance of ENAS algorithms. Although random search can find a well-performed architecture in the experiments, it cannot guarantee that it will find a good architecture every time. Moreover, recent researches~ also showed the evolutionary search was more effective than random search. Furthermore, the experiments in Amoebanet~ and NAS-Bench-201~ also showed that ENAS can search for and find better architectures. Thus, there is an urgent need to design an elaborated experiment to reveal what components in ENAS are responsible for the effectiveness of the algorithm, especially in a large encoding space.\nIn Section~\\ref{sec_Population_operators}, two types of operators have been introduced. We note that some methods like the LargeEvo algorithm~ only use single individual based operator (mutation) to generate offspring. The main reason that they did not involve the crossover operator in their method come from two reasons: the first is for simplicity~, and the second is that simply combining a section of one individual with a section of another individual seems ``ill-suited\" to the neural network paradigm~. Secondly in~, the authors believed that there was no indication that a recombination operation applied to two individuals with high fitness would result in an offspring with similar or better fitness.\nHowever, the supplemental materials in~ demonstrated the effectiveness of the crossover operator in this method. This method can find a good architecture with the help of the crossover operation. On the contrary, when crossover is not used, the architecture found is not promising, unless it runs for a long time. In fact, the mutation operator let an individual explore the neighbouring region, and it is a gradually incremental search process like searching step by step. Crossover (recombination) can generate offspring dramatically different from the parents, which is more likely a stride. So, this operator has the ability to efficiently find a promising architecture. Chu \\textit{et al.}~ preferred that while a crossover mainly contributes to exploitation, a mutation is usually aimed to introduce exploration. These two operators play different roles in the evolutionary process. But there is not a sufficient explanation of how the crossover operator works. Maybe some additional experiments need to be done on the methods without the crossover operator. \nThe EC approach generally performs well when dealing with practical problems (e.g., the vehicle routing problem). This is mainly because EC is designed collectively by considering the domain knowledge in most cases~. Similarly, another key reason to ensure the effectiveness of the ENAS algorithm is that the EC approach can effectively consider (and embed) some domain knowledge of the neural architecture. Since the original intention of ENAS is to design neural architectures automatically without manual experience, the future development of ENAS will need to continue to reduce any artificial experience by including prior knowledge of the encoding space. In this case, how to effectively embed domain knowledge into ENAS is a big challenge.","id":"d6c13ce5-ab8c-4d29-a77d-718dbcbb0cf9","level":"subsection","origin_cites_number":14,"parent_id":"e3dda7bf-8a69-4f93-8f3f-07b17269af8c","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Challenges and Issues"],["subsection","The Effectiveness"]],"subsections":[],"title":"The Effectiveness"},{"cite_extract_rate":0.555555555555555,"cites":[6829,6851,6881,6866,870],"content":"The scale of the datasets used in most ENAS methods is relatively large. Taking the image classification task as an example, the MNIST dataset~ is one of the earliest datasets. In total, it contains $70, 000$ $28\\times28$ grayscale images. Later in 2009, CIFAR-10 and CIFAR-100~ including $60, 000$ $32\\times32$ color images are medium-scale datasets. One of the most well-known large-scale datasets is ImageNet~, which provides more than 14 million manually annotated high-resolution images. Unlike CIFAR-10 and CIFAR-100, which are commonly used in ENAS, fewer methods choose to verify their performance on ImageNet~. This can be explained by the data shown in Table~\\ref{table_CIFAR}, where the GPU Days are usually tens or even thousands. It is unaffordable for most researchers when the medium-scale dataset changes to the lager one.\nHowever, Chen \\textit{et al.}~ believed that the results on the larger datasets like ImageNet are more convincing because CIFAR-10 is easy to be over-fitting. A popular way to deal with this is using a proxy on CIFAR-10 and transfers to ImageNet~. Another alternative approach is to use the down-scaled dataset of a large-scale dataset such as ImageNet$64\\times64$~.","id":"8f512a41-ca51-4724-a511-f83f8ff395ee","level":"subsection","origin_cites_number":9,"parent_id":"e3dda7bf-8a69-4f93-8f3f-07b17269af8c","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Challenges and Issues"],["subsection","Scalability"]],"subsections":[],"title":"Scalability"},{"cite_extract_rate":0.5,"cites":[8247],"content":"Section~\\ref{sec_evaluation} has introduced the most popular and effective ways to reduce the fitness evaluation time and the computational cost. In a nutshell, it can be described as a question that strikes the balance between the time spent and the accuracy of the evaluation. Because of the unbearable time for fully training the architecture, we must compromise as little as we can on the evaluation accuracy in exchange for significant reduction in the evaluation time without sufficient computing resources.\nAlthough a lot of ENAS methods have adopted various means to shorten the evaluation time. Even though Sun \\textit{et al.}~ specifically proposed a method for acceleration, the research direction of search acceleration is just getting started. The current approaches have many limitations that need to be addressed. For example, although the LargeEvo algorithm~ used the weight inheritance to shorten the evaluation time and reduced the computational cost, it still ran for several days with the use of lots of computational resources which cannot be easily accessed by many researchers. Furthermore, there is no baseline and common assessment criteria for search acceleration methods. It is a major challenge to propose a novel method to evaluate the architecture accurately and quickly.","id":"0511a9dd-9340-4ba1-ba4f-5c1c82725c9e","level":"subsection","origin_cites_number":2,"parent_id":"e3dda7bf-8a69-4f93-8f3f-07b17269af8c","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Challenges and Issues"],["subsection","Efficient Evaluation Method and Reduce Computational Cost"]],"subsections":[],"title":"Efficient Evaluation Method and Reduce Computational Cost"},{"cite_extract_rate":0.28571428571428503,"cites":[499,6865],"content":"CNNs are known as black-box-like solutions, which are hard to interpret~. Although some works have been done to visualize the process of feature extraction~, they are uninterpretable due to a large number of learned features~. The low interpretability of the manual designed architecture becomes a big obstacle to the development of neural networks. To overcome this obstacle, some researches~ used GP to automatically design the neural network. Being well-known for its potential interpretability, GP aims at solving problems by automatically evolving computer programs~.\nAll the above researches~ gave a further analysis to demonstrate the interpretability. Specifically, Evans \\textit{et al.}~ have made a visualization on the JAFFE dataset~ to expound how the evolved convolution filter served as a form of edge detection, and the large presence of white color in the output of the convolution can help the classification. In their subsequent work~, they made a visualization of the automatically evolved model on the Hands dataset~, where the aggregation function extracts the minimum value of a specific area in the hand image to determine whether the hand is open or closed. Furthermore, Bi \\textit{et al.}~ displayed the features described by the evolved functions like convolution, max-pooling and addition, and the generated salient features are discriminative for face classification.\nDespite the interpretability the existing work made, all these GP-based ENAS methods only aim at shallow NNs and the number of the generated features is relatively small. However, all the most successful NNs have a deep architecture. This is due partly to the fact that very few GP based methods can be run on GPUs. It is necessary to use deep-GP to evolve a deeper GP tree and make a further analysis on the deeper architecture in the future.","id":"97b2a2f7-bddb-45d3-819a-f3dcbad41cba","level":"subsection","origin_cites_number":7,"parent_id":"e3dda7bf-8a69-4f93-8f3f-07b17269af8c","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Challenges and Issues"],["subsection","Interpretability"]],"subsections":[],"title":"Interpretability"},{"cite_extract_rate":0.8,"cites":[6882,39,6876,9081],"content":"Table~\\ref{table_application} shows various applications which have been explored by ENAS. But these are just a small part of all areas of neural network applications. ENAS can be applied wherever neural networks can be applied, and automate the process of architecture designed which should have been done by experts. Moreover, plenty of the image classification successes of ENAS have proven that ENAS has the ability to replace experts in many areas. The automated architecture design is a trend.\nHowever, this process is not completely automated. The encoding space (search space) still needs to be designed by experts for different applications. For example, for the image processing tasks, CNNs are more suitable, so the encoding space contains the layers including convolution layers, pooling layers and fully connected layers. For the time-series data processing, RNNs are more suitable, so the encoding space may contain the cells including $\\Delta$-RNN cell, LSTM~, Gated Recurrent Unit (GRU)~, Minimally-Gated Unit (MGU)~, and Update-Gated RNN (UGRNN)~. The two manually determined encoding spaces already contain a great deal of artificial experience and the components without guaranteed performance are excluded. The problem is: can a method search the corresponding type of neural network for multiple tasks in a large encoding space, including all the popular widely used components? Instead of searching one multitask network~ which learns several tasks at once with the same neural network, the aim is to find appropriate networks for different tasks in one large encoding space.","id":"ed6f44db-5133-4077-be4f-e9675dc00180","level":"subsection","origin_cites_number":5,"parent_id":"e3dda7bf-8a69-4f93-8f3f-07b17269af8c","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Challenges and Issues"],["subsection","Future Applications"]],"subsections":[],"title":"Future Applications"},{"cite_extract_rate":0.875,"cites":[871,6524,6881,8319,6854,870,9080],"content":"Section~\\ref{sec_CIFAR} gives a brief introduction of the unfair comparisons. The unfairness mainly comes from two aspects: (1) the tricks including cutout~, ScheduledDropPath~, etc. (2) The different encoding spaces. For aspect (1), some ENAS methods~ have reported the results with and without the tricks. For aspect (2), the well-designed search space is widely used in different ENAS methods. For instance, the NASNet search space~ is also used in~ because it is well-constructed so that even random search can perform well. The comparison under the same condition can tell the effectiveness of different search methods.\nFortunately, the first public benchmark dataset for NAS, the NAS-Bench-101~ has been proposed. The dataset contains 432K unique convolutional architectures based on the cell-based encoding space. Each architecture can query the corresponding metrics, including test accuracy, training time, etc., directly in the dataset without the large-scale computation. NAS-Bench-201~ was proposed recently and is based on another cell-based encoding space, which does not have no limits on edges. Compared with NAS-Bench-101, which was only tested on CIFAR-10, this dataset collects the test accuracy on three different image classification datasets (CIFAR-10, CIFAR-100, ImageNet-16-120~). But the encoding space is relatively small, and only contains 15.6K architectures. Experiments with different ENAS methods on these benchmark datasets can get a fair comparison and it will not take too much time. However, these datasets are only based on the cell-based encoding spaces and cannot contain all the search space of the existing methods, because the other basic units (layers and blocks) are built using more hyper-parameters, which may lead to a larger encoding space.\nIn the future, a common platform making fair comparisons needs to be built. This platform must have several benchmarks encoding space, such as the NASNet search space, NAS-Bench-101 and NAS-Bench-201. All the ENAS methods can be directly tested on the platform. Furthermore, this platform also needs to solve the problem that different kinds of GPUs have different computing power, which may cause an inaccurate GPU Days based on different standards. The GPU Days cannot be compared directly until they have a common baseline of computing power.","id":"14a9080c-0216-4f7c-899b-00f8965feb35","level":"subsection","origin_cites_number":8,"parent_id":"e3dda7bf-8a69-4f93-8f3f-07b17269af8c","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Challenges and Issues"],["subsection","Fair Comparisons"]],"subsections":[],"title":"Fair Comparisons"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec_conclusion}\nThis paper provides a comprehensive survey of ENAS. We introduced ENAS from four aspects: population representation, encoding space, population updating, and fitness evaluation following the unified flow, which can be seen in Fig.~\\ref{fig_flowchart}. The various applications and the performance of the state-of-the-art methods on image classification are also summarized in tables to demonstrate the wide applicability and the promising ability. Challenges and issues are also discussed to identify future research direction in this field.\nTo be specific, firstly, the encoding space is introduced by categories. We divide the encoding space into two parts: initial space and search space, where the former one defines the initial conditions whereas the latter one defines the architectures that can be found in the evolutionary process. Also, different encoding strategies and architecture representations are discussed. Secondly, the process of population updating including the evolutionary operators, the multi-objective search strategy, and the selection strategy are presented. A variety of EC paradigms use respective metaphors to generate new individuals. Based on standard algorithms, many improved methods have been proposed to obtain a stable and reliable search capability. Furthermore, we introduce the existing methods to reduce the need for large amounts of time and computing resources, which is a huge obstacle to efficiency.\nAlthough the state-of-the-art methods have achieved some success, ENAS still faces challenges and issues. The first important issue is whether the EC-based search strategy has advantages. If the result is at the same level as the baseline (e.g., random search), it is unnecessary to design the complex evolutionary operators. A sophisticated experiment is in urgent need to tell the effectiveness, especially in a large encoding space. Secondly, the crossover operator is a multi-individual based operator and there is no sufficient explanation to how the crossover operator works well on ENAS. Besides, ENAS is just beginning a new era, so there is a lot of uncharted territory to be explored. \nMoreover, a unified standard or platform is demanded to make a fair comparison.\n\\bibliographystyle{IEEEtran}\n\\bibliography{IEEEabrv,mybibfile}\n\\end{document}","id":"d338e930-f930-4ee2-9cea-b0e47084ac48","level":"section","origin_cites_number":0,"parent_id":"40bf4670-36c8-4014-b074-8f86f999eff7","prefix_titles":[["title","A Survey on Evolutionary Neural Architecture Search"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ]\n ],\n \"subsections\": [\n \"e9b844de-0b2f-4ecf-94b9-874602da0564\",\n \"2857eea5-3541-4f2f-8282-08d5126a14b3\",\n \"7fb1f2d3-d962-4d8c-8b81-c32933ee534a\",\n \"13bf9942-9456-4154-adf6-fca4ab8f93d7\",\n \"d3650a8a-21c2-4aa1-9b61-b69c455614a3\",\n \"fa11702e-7516-4849-b9c7-1422e93b4202\",\n \"390da3df-188a-48a9-991a-966d654bdd4d\",\n \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"d338e930-f930-4ee2-9cea-b0e47084ac48\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 96,\n 1110,\n 166,\n 6822,\n 514,\n 8441,\n 6821,\n 6823,\n 8247,\n 97,\n 6824,\n 7,\n 684,\n 872,\n 6820,\n 6825\n ],\n \"content\": \"\\\\IEEEPARstart{D}{eep} Neural Networks (DNNs), as the cornerstone of deep learning~, have demonstrated their great success in diverse real-world applications, including image classification~, natural language processing~, speech recognition~, to name a few. The promising performance of DNNs has been widely documented due to their deep architectures~, which can learn meaningful features directly from the raw data almost without any explicit feature engineering. Generally, the performance of DNNs depends on two aspects: their architectures and the associated weights. Only when both achieve the optimal status simultaneously, the performance of the DNNs could be expected promising. The optimal weights are often obtained through a learning process: using a continuous loss function to measure the discrepencies between the real output and the desired one, and then the gradient-based algorithms are often used to minimize the loss. When the termination condition is satisfied, which is commonly a maximal iteration number, the algorithm can often find a good set of weights~. This kind of process has been very popular owing to its effectiveness in practice, and has become the dominant practice for weight optimization~, although they are principally local-search~ algorithms. On the other hand, obtaining the optimal architectures cannot be directly formulated by a continuous function, and there is even no explicit function to measure the process for finding optimal architectures.\\nTo this end, there has been a long time that the promising architectures of DNNs are manually designed with rich expertise. This can be evidenced from the state of the arts, such as VGG~, ResNet~ and DenseNet~. These promising Convolutional Neural Network (CNN) models are all manually designed by the researchers with rich knowledge in both neural networks and image processing. However, in practice, most end users are not with such kinds of knowledge. Moreover, DNN architectures are often problem-dependent. If the distribution of the data is changed, the architectures must be redesigned accordingly. Neural Architecture Search (NAS), which aims to automate the architecture designs of deep neural networks, is identified as a promising way to address the challenge aforementioned. \\nMathematically, NAS can be modeled by an optimization problem formulated by Equation~(\\\\ref{equ_defination_NAS}):\\n\\\\begin{equation}\\n\\\\label{equ_defination_NAS}\\n\\\\left\\\\{\\n\\\\centering\\n\\\\begin{array}{c}\\n\\\\argmin_A = \\\\mathcal{L}(A, \\\\mathcal{D}_{train}, \\\\mathcal{D}_{fitness}) \\\\\\\\\\ns.t.\\\\quad A \\\\in \\\\mathcal{A} \\\\\\\\\\n\\\\end{array}\\n\\\\right.\\n\\\\end{equation}\\nwhere $\\\\mathcal{A}$ denotes the search space of the potential neural architectures, $\\\\mathcal{L}(\\\\cdot)$ measures the performance of the architecture $A$ on the fitness evaluation dataset $\\\\mathcal{D}_{fitness}$ after being trained on the training dataset $\\\\mathcal{D}_{train}$. The $\\\\mathcal{L}(\\\\cdot)$ is usually non-convex and non-differentiable~. In principle, NAS is a complex optimization problem experiencing several challenges, e.g., complex constraints, discrete representations, bi-level structures, computationally expensive characteristics and multiple conflicting criteria. NAS algorithms refer to the optimization algorithms which are specifically designed to effectively and efficiently solve the problem represented by Equation~(\\\\ref{equ_defination_NAS}). The initial work of NAS algorithms is commonly viewed as the work in~, which was proposed by Google. The pre-print version of this work was firstly released in the website of arXiv\\\\footnote{\\\\url{https://arxiv.org/abs/1611.01578}} in 2016, and then was formally accepted for publication by the International Conference on Learning Representations (ICLR) in 2017. Since then, a large number of researchers have been investing tremendous efforts in developing novel NAS algorithms.\\nBased on the optimizer employed, existing NAS algorithms can be broadly classified into three different categories: Reinforcement Learning (RL)~ based NAS algorithms, gradient-based NAS algorithms, and Evolutionary Computation (EC)~ based NAS algorithms (ENAS). Specifically, RL based algorithms often require thousands of Graphics Processing Cards (GPUs) performing several days even on median-scale dataset, such as the CIFAR-10 image classification benchmark dataset~. Gradient-based algorithms are more efficient than RL based algorithms. However, they often find the ill-conditioned architectures due to the improper relation for adapting to gradient-based optimization. Unfortunately, the relation has not been mathematically proven. In addition, the gradient-based algorithms require to construct a supernet in advance, which also highly requires expertise. The ENAS algorithms solve the NAS by exploiting EC techniques. Specifically, EC is a class of population-based computational paradigms, simulating the evolution of species or the behaviors of the population in nature, to solve challenging optimization problems. In particular, Genetic Algorithms (GAs)~, Genetic Programming (GP)~, and Particle Swarm Optimization (PSO)~ are widely used EC methods in practice. Owing to the promising characteristics of EC methods in insensitiveness to the local minima and no requirement to gradient information, EC has been widely applied to solve complex non-convex optimization problems~, even when the mathematical form of the objective function is not available~.\\nIn fact, the EC methods had been frequently used more than twenty years ago, searching for not only the optimal neural architectures but also the weights of neural networks simultaneously, which is also termed as neuroevolution~. The major differences between ENAS and neuroevolution lie in two aspects. Firstly, neuroevolution often uses EC to search for both the neural architectures and the optimal weight values, while ENAS for now focuses mainly on searching for the architectures and the optimal weight values are obtained by using the gradient-based algorithms\\\\footnote{Please note that we still categorize some existing algorithms as the ENAS algorithm, such as API~, EvoCNN~ and EvoDeep~, although they also concern the weights. This is because the optimal weight values of the DNNs searched by them are still obtained by the gradient-based algorithms. They only searched for the best weight initialization values or the best weight initialization method of the DNNs.} immediately after. Secondly, neuroevolution commonly applies to small-scale and median-scale neuron networks, while ENAS generally works on DNNs, such as the deep CNNs~ and deep stacked autoencoders~, which are stacked by the building blocks of deep learning techniques~. Generally, the first work of ENAS is often viewed as the LargeEvo algorithm~ which was proposed by Google who released its early version in March 2017 in arXiv. Afterwards, this paper got accepted into the 34th International Conference on Machine Learning in June 2017. The LargeEvo algorithm employed a GA to search for the best architecture of a CNN, and the experimental results on CIFAR-10 and CIFAR-100~ have demonstrated its effectiveness. Since then, a large number of ENAS algorithms have been proposed. Fig.~\\\\ref{fig_histogram} shows the number of similar works\\\\footnote{These ``submissions\\\" include the ones which have been accepted for publication after the peer-review process, and also the ones which are only available on the arXiv website without the peer-review process.} published from 2017 to 2020 when this survey paper was ready for submission. As can be seen from Fig.~\\\\ref{fig_histogram}, from 2017 to 2020, the number of submissions grows with multiple scales.\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.8\\\\linewidth]{histogram}\\n\\t\\\\caption{The number of submissions refers to the works of evolutionary neural architecture search. The data is collected from Google Scholar with the keywords of ``evolutionary\\\" OR ``genetic algorithm'' OR ``particle swarm optimization\\\" OR ``PSO\\\" OR ``genetic programming\\\" AND ``architecture search\\\" OR ``architecture design'' OR ``CNN'' OR ``deep learning'' OR ``deep neural network'' and the literature on Neural Architecture Search collected from the AutoML.org website by the end of 2020. With these initially collected data, we then have carefully checked each manuscript to make its scope accurately within the evolutionary neural architecture search.}\\n\\t\\\\label{fig_histogram}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\\nA large number of related submissions have been made available publicly, but there is no comprehensive survey of the literature on ENAS algorithms. Although recent reviews on NAS have been made in~, they mainly focus on reviewing different methods to realize NAS, rather than concentrating on the ENAS algorithms. Specifically, Elsken \\\\textit{et al.}~ divided NAS into three stages: search space, search strategy, and performance estimation strategy. Similarly, Wistuba \\\\textit{et al.}~ followed these three stages with an additional review about the multiple objectives in NAS. Darwish \\\\textit{et al.}~ made a summary of Swarm Intelligence (SI) and Evolutionary Algorithms (EAs) approaches for deep learning, with the focuses on both NAS and other hyperparameter optimization. Stanley \\\\textit{et al.} ~ went through a review of neuroevolution, aiming at revealing the weight optimization rather than the architecture of neural networks. Besides, most of the references in these surveys are pre-2019 and do not include an update on the papers published during the past two years when most ENAS works were published. This paper presents a survey involving a large number of ENAS papers, with the expectation to inspire some new ideas for enhancing the development of ENAS. To allow readers easily concentrating on the technical part of this survey, we also follow the three stages to introduce ENAS algorithms, which has been widely adopted by existing NAS survey papers~, but with essential modifications made to specifically suit ENAS algorithms.\\nThe remainder of this paper is organized as follows. The background of ENAS is discussed in Section~\\\\ref{sec_background}. Section~\\\\ref{sec_encoding_space} documents different encoding spaces, initial spaces and search spaces of ENAS algorithms. In Section~\\\\ref{sec_architectureEncoding}, the encoding strategy and architecture representation are introduced. Section~\\\\ref{sec_Population_operators} summarizes the process of population updating, including evolutionary operators and selection strategies. Section~\\\\ref{sec_evaluation} shows multiple ways to speed up the evolution. Section~\\\\ref{sec_application} presents the applications of ENAS algorithms. Section~\\\\ref{sec_challenges_and_issues} discusses the challenges and prospects, and finally Section~\\\\ref{sec_conclusion} presents conclusions.\",\n \"id\": \"e9b844de-0b2f-4ecf-94b9-874602da0564\",\n \"level\": \"section\",\n \"origin_cites_number\": 28,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec_background}\\nAs discussed above, ENAS is a process of searching for the optimal architectures of DNNs by using EC methods. In this section, we will first introduce the unified flow chart of ENAS in Subsection~\\\\ref{sec2_3}. Then, the categories of EC methods based on their search strategies and the categories of DNN architectures in Subsections~\\\\ref{sec2_1} and \\\\ref{sec2_2}, respectively.\",\n \"id\": \"2857eea5-3541-4f2f-8282-08d5126a14b3\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Background\"\n ]\n ],\n \"subsections\": [\n \"0869d4e9-c981-472b-b025-5a8bbce117b1\",\n \"fb246415-ce9a-4259-827b-2d75a5822aa3\",\n \"a76618db-7274-4a81-a316-4d0b05a79b1e\"\n ],\n \"title\": \"Background\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec2_3}\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{flowchart}\\n\\t\\\\caption{The flowchart of a common ENAS algorithm.}\\n\\t\\\\label{fig_flowchart}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\\nFig.~\\\\ref{fig_flowchart} shows an illustration of the flowchart of an ENAS algorithm. Specifically, the evolutionary process takes place in the initial space and the search space sequentially. First of all, a population is initialized within the initial space that is defined in advance. Each individual in the population represents a solution for the ENAS, i.e., a DNN architecture. Each architecture needs to be encoded as an individual before it joins the population. Second, the fitness of the generated individuals is evaluated. Note that there are two fitness evaluation steps as shown in Fig.~\\\\ref{fig_flowchart}, which commonly employ the same evaluation criterion. Thirdly, after the fitness evaluation of the initial population, the whole population starts the evolutionary process within the search space, which is shown by the dashed box in Fig.~\\\\ref{fig_flowchart}. In the evolutionary process, the population is updated by the selection and the evolutionary operators in each iteration, until the stopping criterion is met. Please note that the selection stage is not necessary for some other EC paradigms like SI. Finally, a population that has finished the evolution is obtained. In the following sections, these key steps are documented in detail.\",\n \"id\": \"0869d4e9-c981-472b-b025-5a8bbce117b1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"2857eea5-3541-4f2f-8282-08d5126a14b3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Flow Chart of ENAS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Flow Chart of ENAS\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 9078,\n 6826\n ],\n \"content\": \"\\\\label{sec2_1}\\nENAS is distinguished from other NAS methods by its employed optimization approach, i.e., the EC methods, where the optimization approaches can be further subdivided based on the search strategy that the optimization approach adopted. Fig.~\\\\ref{fig_EC_taxonomy} provides such an illustration from three aspects: EAs, Swarm Intelligence (SI), and others, and more detailed statistics can be observed in Table~\\\\ref{table_EC}. In practice, the EA-based methods account for the majority of existing ENAS algorithms, where the GA takes a large part of EA-based methods. The other categories of EC methods are also important parts for realizing ENAS algorithms, such as GP, Evolutionary Strategy (ES), PSO, Ant Colony Optimization (ACO)~, Differential Evolution (DE)~, Firefly Algorithm (FA)~. In this paper, the Hill-Climbing Algorithm (HCA) is classified into EC paradigm because it can be regarded as an EA with a simple selection mechanism and without crossover operation~. HCA has been well known as a widely used local search algorithm in memetic algorithms~.\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=1\\\\linewidth]{taxonomy1}\\n\\t\\\\caption{The categories of ENAS from EC methods regarding the search strategies.}\\n\\t\\\\label{fig_EC_taxonomy}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\",\n \"id\": \"fb246415-ce9a-4259-827b-2d75a5822aa3\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"2857eea5-3541-4f2f-8282-08d5126a14b3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Evolutionary Search Strategy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Evolutionary Search Strategy\"\n },\n {\n \"cite_extract_rate\": 0.47368421052631504,\n \"cites\": [\n 6822,\n 96,\n 514,\n 6821,\n 870,\n 8247,\n 97,\n 6827,\n 166\n ],\n \"content\": \"\\\\label{sec2_2}\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{taxonomy2}\\n\\t\\\\caption{The categories of ENAS from neural network perspective.}\\n\\t\\\\label{fig_NN_taxonomy}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=1\\\\linewidth]{NN}\\n\\t\\\\caption{Examples of different neural architectures in ENAS. (a) CNN. (b) DBN. (c) AE. (d) RNN. }\\n\\t\\\\label{fig_NN}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\\nThe end product of ENAS is all about DNNs, which can be broadly divided into five different categories: CNN, Deep Belief Network (DBN), Stacked Auto-Encoder (SAE), Recurrent Neural Network (RNN) and others. The brief fact can be seen from Fig.~\\\\ref{fig_NN_taxonomy}, and a detailed illustration will be shown in Table~\\\\ref{table_EC}. \\nMost ENAS methods are proposed for searching for the optimal CNN architectures. This is because many hand-crafted CNNs, such as VGG~, ResNet~ and DenseNet~, have demonstrated their superiority in handling the image classification tasks which is the most successful applications in the deep learning field. Generally, the CNN architecture is composed of the convolutional layers, the pooling layers and the fully-connected layers, and a common example of CNNs is shown in Fig.~\\\\ref{fig_NN} (a). The optimization of CNN architecture is mainly composed of three aspects: the hyperparameters of each layer~, the depth of the architecture~ and the connections between layers~. In practice, the majority of the ENAS methods consider the above three aspects collectively~.\\nDBN~ is made up by stacking multiple Restricted Boltzmann Machines (RBMs), and an example can be seen in Fig.~\\\\ref{fig_NN} (b). Specifically, RBMs have connections only between layers, while without any inter-layer connection. Meanwhile, RBMs allow the DBN to be trained in an unsupervised manner to obtain a good weight initialization. There are two commonly used types of hyperparameters needed to be optimized in DBN: the number of neurons in each layer~ and the number of layers~.\\nAn SAE is constructed by stacking multiple of its building blocks which are AEs. An AE aims to learn meaningful representations from the raw input data by restoring its input data as its objective~. Generally, an AE is typically composed of two symmetrical components: the encoder and the decoder, and an example including both parts is shown in Fig.~\\\\ref{fig_NN} (c). Generally, some ENAS methods only take the encoder into evolution~ because the decoder part is symmetric with the encoder and can be derived accordingly. Yet, some other ENAS methods optimize the hyperparameters of encoders and decoders separately~.\\nThe most significant difference between RNNs and the neural networks introduced above is its recurrent connection. Fig.~\\\\ref{fig_NN} (d) shows the time-expanded structure of an RNN, where the value of current hidden layer $h^{t}$ is influenced by the value at its previous time slot $h^{t-1}$ and the output value of its previous layer. Because these layers are reused, all the weights (i.e., \\\\textit{U}, \\\\textit{W} and \\\\textit{V} in the figure) are shared. Different from focusing on the number of neurons and the number of layers in the feedforward neural network, some ENAS methods concern about how many times the RNN should be unfold~. In addition, there remain other neural networks like typical DNNs, which are made up of only fully-connected layers, where the connections are formed by all the neurons in the two adjacent layers. Because such kind of DNNs is not the major target investigated by NAS, we will not introduce it in detail.\\nWhen the ENAS algorithms are applied to these DNNs, the goal is to find the best architecture-related parameters. Specifically, for CNNs, they are the number of convolutional layers, pooling layers, fully-connected layers, and parameters related to these layers (such as the kernel size of the convolutional layers, the pooling type, and the number of neurons for fully-connected layers, and so on) as well as their connection situation (such as the dense connection and the skip connection). For DBN and SAE, they are the number of their building blocks, i.e., the RBM for DBN and the AE for SAE, and the number of neurons in each layer. For RNN, in addition to the architecture-related parameters mentioned above, the number of the time slot is also an important parameter to be optimized by ENAS algorithms. For the traditional DNNs, the ENAS algorithms often concern about the neuron number of each layer. In addition, some ENAS algorithms also concern the weights, such as the weight initialization method and weigh initial values. Table~\\\\ref{table_parameters} summarizes the detail of the common parameters optimized by ENAS methods in different types of neural networks, including the CNNs as the most popular type. The details of these ENAS algorithms will be documented in the following sections. \\n\\\\begin{table}[]\\n\\t\\\\renewcommand\\\\arraystretch{0.8}\\n\\t\\\\caption{Common Parameters Optimized in Different DNNs.} \\n\\t\\\\label{table_parameters}\\n\\t\\\\centering\\n\\t\\\\begin{tabular}{p{1.2cm}|p{1.5cm}|p{4.5cm}}\\n\\t\\t\\\\hline\\n\\t\\t& \\\\multicolumn{2}{c}{Parameters} \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{4}{*}{CNN} & global parameters & number of layers, connections between layers \\\\\\\\ \\\\cline{2-3} \\n\\t\\t& convolution layer & filter size (width and height), stride size (width and height), feature map size, convolution type, standard deviation and mean value of the filter elements \\\\\\\\ \\\\cline{2-3} \\n\\t\\t& pooling layer & filter size (width and height), stride size (width and height), pooling type \\\\\\\\ \\\\cline{2-3} \\n\\t\\t& fully-connected layer & number of neurons, standard deviation and mean value of weights \\\\\\\\ \\\\hline\\n\\t\\tDBN, AE & \\\\multicolumn{2}{l}{number of hidden layers, neurons per layer} \\\\\\\\ \\\\hline\\n\\t\\tRNN & \\\\multicolumn{2}{l}{\\\\shortstack[l]{number of hidden layers, neurons per layer, number of\\\\\\\\ time slot}} \\\\\\\\ \\\\hline\\n\\t\\\\end{tabular}\\n\\\\vspace{-0.3cm}\\n\\\\end{table}\",\n \"id\": \"a76618db-7274-4a81-a316-4d0b05a79b1e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"2857eea5-3541-4f2f-8282-08d5126a14b3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Common Neural Networks in ENAS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Common Neural Networks in ENAS\"\n },\n {\n \"cite_extract_rate\": 0.330275229357798,\n \"cites\": [\n 6851,\n 6838,\n 6837,\n 6844,\n 6841,\n 8955,\n 6829,\n 6832,\n 9079,\n 6850,\n 6836,\n 6835,\n 6843,\n 6821,\n 6846,\n 8247,\n 6847,\n 9080,\n 6834,\n 6845,\n 6842,\n 6826,\n 6830,\n 6822,\n 6848,\n 870,\n 6853,\n 6840,\n 6833,\n 6852,\n 6831,\n 6849,\n 6824,\n 6827,\n 6839,\n 6828\n ],\n \"content\": \"\\\\label{sec_encoding_space}\\n\\\\begin{table*}[]\\n\\t\\\\renewcommand\\\\arraystretch{0.8}\\n\\t\\\\caption{Different Encoding Space and the Constraints} \\n\\t\\\\label{table_constrains}\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{p{3.8cm}|p{1.8cm}|p{2cm}|p{2.5cm}|p{5.5cm}}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t& Fixed depth & Rich initialization & Partial fixed structure & Relatively few constraints\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tLayer-based encoding space & ~ & ~ & ~ & ~\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tBlock-based encoding space & ~ & ~ & ~ & ~ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tCell-based encoding space & & & ~ & ~\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tTopology-based encoding space & & ~ & & ~ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\vspace{-0.5cm}\\n\\\\end{table*}\\nThe encoding space contains all the valid individuals encoded in the population. In terms of the basic units, the encoding space can be divided into three categories according to the basic units they adopt. They are the layer-based encoding space, the block-based encoding space and the cell-based encoding space. In addition, some ENAS methods do not take the configuration of the basic unit into consideration, but care the connections between units. Such kind of encoding space is often called the topology-based encoding space.\\nIn addition, the constraints on the encoding space are important. This is because they represent the human intervention which restricts the encoding space and lightens the burden of the evolutionary process. A method with a mass of constraints can obtain a promising architecture easily but prevent to design any novel architecture that does not follow the constraints. Furthermore, the different sizes of the search space would greatly affect the efficiency of the evolution. On the other hand, the effectiveness of the ENAS methods cannot be guaranteed if there is no constraint on the search space: one extreme case is that all the individuals in the search space are mediocre. In this case, an excellent individual can be obtained even without selection. Table~\\\\ref{table_constrains} shows different kinds of encoding spaces and the constraints of existing ENAS algorithms, where the first row shows the constraints and the first column displays the encoding space. In the following, we will discuss them at length.\",\n \"id\": \"7fb1f2d3-d962-4d8c-8b81-c32933ee534a\",\n \"level\": \"section\",\n \"origin_cites_number\": 109,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Space\"\n ]\n ],\n \"subsections\": [\n \"c45304b4-d500-4803-95c8-a1b032fcb338\",\n \"6c612948-3e12-4637-8d86-3b3c5b4c17d5\",\n \"0e3df321-149a-48ba-be12-da3f243c2af6\"\n ],\n \"title\": \"Encoding Space\"\n },\n {\n \"cite_extract_rate\": 0.72,\n \"cites\": [\n 96,\n 6524,\n 514,\n 6821,\n 6848,\n 870,\n 305,\n 871,\n 544,\n 6831,\n 97,\n 6849,\n 6842,\n 872,\n 6827,\n 6839,\n 6854,\n 6826\n ],\n \"content\": \"The layer-based encoding space denotes that the basic units in the encoding space are the primitive layers, such as convolution layers and fully-connected layers. This would lead to a huge search space, since it tries to encode so much information in the search space. However, it may take more time to search for a promising individual because there are more possibilities to construct a well-performed DNN from the primitive layers. For example, a promising CNN is commonly composed of hundreds of primitive layers, accordingly, the search process will consume more time to search for such a deep architecture by iteratively stacking the primitive layers. In addition, the promising DNN may not be found if only with the primitive layers. For instance, the promising performance of ResNet is widely recognized due to its skip connections which cannot be represented by the primitive layers.\\nTo alleviate the above problems, the block-based encoding space is developed, where various layers of different types are combined as the blocks to serve as the basic unit of the encoding space. Traditional blocks are ResBlock~, DenseBlock~, ConvBlock (Conv2d + BatchNormalization + Activation)~ and InceptionBlock~, etc. Specifically, layers in the blocks have a specific topological relationship, such as the residual connection in ResBlock and the dense connections in DenseBlock. These blocks have promising performance and often require fewer parameters to build the architecture. So it is principally easier to find a good architecture in the block-based encoding space compared to the layer-based encoding space. Some ENAS algorithms used these blocks directly, such as~, while other methods proposed different blocks for different purposes. For example, Chen \\\\textit{et al.}~ proposed eight blocks including ResBlock and InceptionBlock encoded in a 3-bit string, and used Hamming distance to determine the similar blocks. Song \\\\textit{et al.}~ proposed three residual-dense-mixed blocks to reduce the amount of computation due to the convolution operation of image super-resolution tasks. \\nThe cell-based encoding space is similar to the block-based one, and can be regarded as a special case of the block-based space where all the blocks are the same. The ENAS algorithms employing this space build the architectures by stacking repeated motifs. Chu \\\\textit{et al.}~ divided the cell-based space into two independent parts: the micro part containing the parameters of cells, and the macro part defining the connections between different cells. To be more specific, the cell-based encoding space concentrates more on the micro part. Different from the block-based space, the layers in the cell can be combined more freely, and the macro part is always determined by human expertise~. Some widely used encoding spaces are classified in this category. For example, NAS-Bench-101~ and NAS-Bench-201~ search for different combinations of layers and connections in cells, and then stacked the cells sequentially. In addition, NASNet~ and DARTS~ search for two kinds of cells, namely normal cell and reduction cell, and each stacked cell is connected to the two previous cells. The cell-based space greatly reduces the size of the encoding space. This is because all the basic units in the encoding space are the same, and the number of parameters in terms of constructing the promising DNN is much fewer. However, Frachon \\\\textit{et al.}~ claimed that there is no theoretical basis for that the cell-based space can help to obtain a good architecture. \\nIn contrast, the topology-based space does not consider the parameters or the structure of each unit (layer or block), yet they are only concerned about the connections between units. One classical example is the one-shot method which treats all the architectures as different subgraphs of a supergraph~. Yang \\\\textit{et al.}~ proposed CARS to search for the architecture by choosing different connections in the supergraph, and then the architecture was built by the chosen subgraph. CARS can be classified into the topology-based category because it aims at deciding whether or not to keep the connections in the supergraph. But from the perspective of building the architecture, CARS can also be classified into the cell-based category. This is because the subgraph cell was stacked several times to build the architecture. Another typical case is pruning. Wu \\\\textit{et al.}~ employed a shallow VGGNet~ on CIFAR-10, and the aim was to prune unimportant weight connections from the VGGNet.\\nAs observed from Table~\\\\ref{table_constrains}, the constraints on the encoding space mainly focus on three aspects: fixed depth, rich initialization and partial fixed structure. The fixed depth means all the individuals in the population have the same depth. The fixed depth is a strong constraint and largely reduces the size of the encoding space. Please note that the fixed depth is different from the \\\\textit{fixed-length encoding strategy} which will be introduced in Section~\\\\ref{sec_architectureEncoding}. In Genetic CNN~, for example, the \\\\textit{fixed-length encoding strategy} only limits the maximum depth. The node which is isolated (no connection) is simply ignored. By this way, the individuals can obtain different depths. The second constraint is rich initialization (i.e., the \\\\textit{well-designed space} to be discussed in Section~\\\\ref{sec_initial_space}), and it is also a limitation in practice that requires a lot of expertise. In this case, the initialized architectures are manually designed, which goes against the original intention of NAS. The partial fixed structure means the architecture is partially settled. For example, in~, a max-pooling layer is added to the network after every set of four convolution layers.\\nIn Table~\\\\ref{table_constrains}, the relatively few constraints category means that those methods have no restrictions like the three aspects discussed above. However, it does not imply there is no constraint. For example, in the classification task, the fully-connected layer is often used as the tail of the whole DNNs in some methods~. Moreover, the maximum length is predefined in many methods including both \\\\textit{fixed-length encoding strategy} methods~ and \\\\textit{variable-length encoding strategy} methods~ resulting in preventing the method from discovering a deeper architecture. Wang \\\\textit{et al.}~ tried to break the limit of maximum length by using a Gaussian distribution initialization mechanism. Irwin \\\\textit{et al.}~ broke the limit by using the evolutionary operators, the crossover operator and the mutation operator, to extend the depth to any size.\\nGenerally, the encoding space can be served as the basis of the search space and the initial space. In practice, the encoding space is often the same as the search space, while the initial space is often a subspace of the encoding space. Fig.~\\\\ref{fig_spaces} shows the relationship between the three spaces. The search space is larger than the initial space when some manual constraints are added to the population initialization. When there is no such manual constrains, search space and initial space are equivalent. Furthermore, the initial space determines what kind of individuals may appear in the initial population, and the search space determines what kind of individuals may appear in the evolutionary search process, as illustrated in Fig.~\\\\ref{fig_flowchart}. In the following subsections, we will discuss the initial space and the search space in Subsections~\\\\ref{sec_initial_space} and~\\\\ref{sec_search_space}, respectively.\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.8\\\\linewidth]{spaces}\\n\\t\\\\caption{The relationship between encoding space, search space and initial space.}\\n\\t\\\\label{fig_spaces}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\",\n \"id\": \"c45304b4-d500-4803-95c8-a1b032fcb338\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 25,\n \"parent_id\": \"7fb1f2d3-d962-4d8c-8b81-c32933ee534a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Space\"\n ],\n [\n \"subsection\",\n \"Encoding Space and Constraints\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Encoding Space and Constraints\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 6829,\n 8247,\n 6821,\n 6827\n ],\n \"content\": \"\\\\label{sec_initial_space}\\nIn general, there are three types of architecture initialization approaches in the initial space: starting from trivial initial conditions~, randomly initialization in the encoding space~ and staring from a well-designed architecture (also termed as rich initialization)~. These three types of initialization correspond to three different initial spaces: \\\\textit{trivial space}, \\\\textit{random space} and \\\\textit{well-designed space}.\\nThe \\\\textit{trivial space} contains only a few primitive layers. For example, the LargeEvo algorithm~ initialized the population in a \\\\textit{trivial space} where each individual constitutes just a single-layer model with no convolutions. Xie \\\\textit{et al.}~ experimentally demonstrated that a \\\\textit{trivial space} can evolve to a competitive architecture. The reason for using as little experience as possible is to justify the advantage of EC-based methods where some novel architectures can be discovered, and most of the discovery is different from the manually designed DNN architectures. On the contrary, the \\\\textit{well-designed space} contains the state-of-the-art architectures. In this way, a promising architecture can be obtained at the beginning of the evolution, whereas it can hardly evolve to other novel architectures. Actually, many of ENAS methods adopting this initial space focus on improving the performance upon the well-designed architecture. For example, the architecture pruning aims at compressing DNNs by removing less important connections~. For a \\\\textit{random space}, all the individuals in the initial population are randomly generated in the limited space, and it has been adopted by many methods, such as~. The aim of this type of initial spaces is also to reduce the intervention of human experience in the initial population.\",\n \"id\": \"6c612948-3e12-4637-8d86-3b3c5b4c17d5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"7fb1f2d3-d962-4d8c-8b81-c32933ee534a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Space\"\n ],\n [\n \"subsection\",\n \"Initial Space\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Initial Space\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec_search_space}\\nAfter the initialization of the population, the ENAS methods start to search the architectures in the search space. Generally speaking, the search space is the same as the initial space when the random initial space is adopted. For other two types of initial spaces, however, due to the relatively small initial space, the search space will become much larger with the expectation that the promising architecture is included. It is worth noting that many methods do not directly define the search space, but restrict the search space by using evolutionary operators. For example, Irwin \\\\textit{et al.}~ did not specify the maximal depth, instead, they used the evolutionary operations to extend the architecture to any depth.\",\n \"id\": \"0e3df321-149a-48ba-be12-da3f243c2af6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"7fb1f2d3-d962-4d8c-8b81-c32933ee534a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Space\"\n ],\n [\n \"subsection\",\n \"Search Space\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Search Space\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6821,\n 6827\n ],\n \"content\": \"\\\\label{sec_architectureEncoding}\\nThis section will discuss how to encode a network architecture into an individual of the EC methods. Each ENAS method needs to determine its encoding strategy before starting the first stage of the ENAS method, i.e., the population initialization. The most intuitive difference in different encoding strategies of ENAS methods is the length of the encoded individuals. \\nGenerally, the encoding strategies can be divided into two different categories according to whether the length of an individual changes or not during the evolutionary process. They are the \\\\textit{fixed-length encoding strategy} and the \\\\textit{variable-length encoding strategy}. Particularly, the individuals have the same length during the evolutionary process when the \\\\textit{fixed-length encoding strategy} is used. In contrast, the individuals are with different lengths during the evolutionary process if the \\\\textit{variable-length encoding strategy} is employed. The advantage of the \\\\textit{fixed-length encoding strategy} is that it is easy to use standard evolutionary operations which are originally designed for the individuals with equal lengths. In Genetic CNN~, for example, the fixed-length binary string helped the evolutionary operators (especially the crossover) with easy implementation. Another example is~, where Loni \\\\textit{et al.} used a fixed-length string of genes to represent the architecture which led to easy implementation of the one-point crossover in the corresponding encoded information. However, a proper maximal length must be predefined for the fixed-length encoding strategy. Because the maximal length relates to the optimal depth of the DNN architecture, which is unknown in advance, the corresponding ENAS algorithms still rely on expertise and experience. \\nCompared with the \\\\textit{fixed-length encoding strategy}, the \\\\textit{variable-length encoding strategy} dose not require the human expertise regarding the optimal depth in advance, which has the potential to be fully automated. In addition, the advantage of this encoding strategy is that it can define more details of the architecture with freedom. For example, when solving a new task where there is no expertise in knowing the optimal depth of the DNN, we just initialize the individuals with a random depth, and the optimal depth can be found through the variable-length encoding strategy, where the depths of the corresponding DNN can be changed during the evolutionary process. However, the variable-length encoding strategy also brings some drawbacks. Because the traditional evolutionary operators might not be suitable for this kind of encoding strategy, the corresponding evolutionary operators need to be redesigned, where an example can be seen in~. Another disadvantage is that, due to the flexibility of variable length encoding, it could generally result in over-depth architectures, which sometimes further leads to more time-consuming fitness evaluation. Please note that some works claimed their use of the variable-length encoding strategy, where each individual is designed with a maximal length, and then the placeholder is used to indicate the gene's validation of the gene~. For example, the individual is designed to have $1,000$ genes, while some genes having the values of zeros do not participate in the evolutionary process. In this paper, we also categorize these methods into the fixed-length encoding strategy.\\nIn addition, most of the DNN architectures can be represented as directed graphs which are made up of different basic units and the connections between the units. Therefore, the encoding for architecture can be divided into two aspects: configurations of basic units and connections, which will be discussed in the following subsections.\",\n \"id\": \"13bf9942-9456-4154-adf6-fca4ab8f93d7\",\n \"level\": \"section\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Strategy\"\n ]\n ],\n \"subsections\": [\n \"c0d67ff4-2346-4359-81e9-cfc247725ee8\",\n \"2960167f-5dd3-47f9-b647-47b27f2be6ee\"\n ],\n \"title\": \"Encoding Strategy\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 6821,\n 870\n ],\n \"content\": \"In practice, different basic units have different configurations, such as layers, blocks and cells, which are demonstrated in Section~\\\\ref{sec_encoding_space}. For example, in CNNs, there are multiple parameters in the primitive layers which can be seen in Table~\\\\ref{table_parameters}. As for the DenseBlock implemented by Sun \\\\textit{et al.}~, only two parameters are needed to build the block. Because the configuration of cell is more flexible than that of blocks in CNNs, which can be regarded as a microcosm of a complete neural network. For example, the cells in~ are made up by a combination of 10 layers selected from 8 different primitive layers. But the cells do not include some of the configurations of primitive layer, such as the feature map size which is an important parameter in the primitive layer~.\",\n \"id\": \"c0d67ff4-2346-4359-81e9-cfc247725ee8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"13bf9942-9456-4154-adf6-fca4ab8f93d7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Strategy\"\n ],\n [\n \"subsection\",\n \"Encoding for Configurations of Basic Units\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Encoding for Configurations of Basic Units\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"When the parameters (configurations) of the basic units in the architecture have been determined, the corresponding architecture cannot be built up immediately. Since edges are indispensable in directed graph, connections are also part of the architecture. The connections discussed in this section include not only connections between basic units, but also the connections within the basic units.\\nGenerally, the architectures of neural networks can be divided into two categories: \\\\textit{linear architecture} and \\\\textit{non-linear architecture}. The former denotes the architectures containing sequential basic units. The latter indicates that there are skip-connections or loop-connections in the architecture. Please note that, the structure could be the macro structure consisting of basic units, or the micro structure within the basic units.\",\n \"id\": \"2960167f-5dd3-47f9-b647-47b27f2be6ee\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"13bf9942-9456-4154-adf6-fca4ab8f93d7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Strategy\"\n ],\n [\n \"subsection\",\n \"Encoding for Connections\"\n ]\n ],\n \"subsections\": [\n \"6fd9ba7f-54c9-40c3-908e-224aaa013157\",\n \"a7c80d6f-1308-48d5-9a45-859a5cedcd95\"\n ],\n \"title\": \"Encoding for Connections\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 96,\n 6821,\n 97\n ],\n \"content\": \"The linear architecture can be found in different kinds of architectures including the one generated from the layer-based encoding space and the block-based encoding space. Its widespread use in ENAS stems from its simplicity. No matter how complex the internal of the basic units is, many ENAS methods stack the basic units one by one to build up the skeleton of the architecture which is linear. For example, in AE-CNN~, Sun \\\\textit{et al.} stacked different kinds of blocks to generate the architecture.\\nOne special case is a linear architecture generated from the layer-based encoding space. In this case, there is no need to solely encode the connections, and only the parameters in each basic unit are enough to build an architecture. One classical example can be seen in~ where Sun \\\\textit{et al.} explored a great number of parameters based on a linear CNN architecture. However, most architectures are not designed to be linear. The skip connections in ResNet~ and the dense connections in DenseNet~ show the ability to build a good architecture.\",\n \"id\": \"6fd9ba7f-54c9-40c3-908e-224aaa013157\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"2960167f-5dd3-47f9-b647-47b27f2be6ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Strategy\"\n ],\n [\n \"subsection\",\n \"Encoding for Connections\"\n ],\n [\n \"subsubsection\",\n \"Linear Architecture\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Linear Architecture\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 870,\n 6827\n ],\n \"content\": \"Firstly, we will introduce two approaches to encoding a non-linear architecture in this subsection. Specifically, the adjacent matrix is the most popular way to represent the connections for non-linear architectures. Genetic CNN~ used a binary string to represent the connection, and the string can be transformed into a triangular matrix. In the binary string, ``1'' denotes that there is a connection between the two nodes while ``0'' denotes no connection in between. Lorenzo \\\\textit{et al.}~ used a matrix to represent the skip connections, and this work revolved around the adjacent matrix. Back in the 1990s, Kitano \\\\textit{et al.}~ began to study the use of the adjacent matrix to represent network connection and explained the process from the connectivity matrix, to the bit-string genotype, to the network architecture phenotype. Another way to represent the connections is to use an ordered pair $G = (V,E)$ with vertices $V$ and edge $E$ associated with a direction, to represent a directed acyclic graph. Irwin \\\\textit{et al.}~ also used this strategy to encode the connections.\\nSecondly, the non-linear architecture is a more common case in both the macro architecture and the micro architecture. AmoebaNet-A~, as an example of non-linear macro architecture, stacked these two kinds of cells for several times to build up an architecture. In addition, each cell receives two inputs from the previous two cells separately, which means that the direct connection and the skip-connection are both used in this macro structure. Also in AmoebaNet, each cell is a non-linear architecture inside, which means the micro architecture is also non-linear. The non-linear architecture provides the architecture with more flexibility, which makes it more likely to build up a promising architecture than that of the linear ones.\",\n \"id\": \"a7c80d6f-1308-48d5-9a45-859a5cedcd95\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"2960167f-5dd3-47f9-b647-47b27f2be6ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Encoding Strategy\"\n ],\n [\n \"subsection\",\n \"Encoding for Connections\"\n ],\n [\n \"subsubsection\",\n \"Non-linear Architecture\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Non-linear Architecture\"\n },\n {\n \"cite_extract_rate\": 0.318840579710144,\n \"cites\": [\n 6861,\n 6851,\n 6838,\n 6837,\n 6863,\n 6844,\n 6841,\n 8955,\n 6829,\n 6832,\n 6856,\n 6857,\n 9079,\n 6858,\n 6836,\n 6835,\n 6843,\n 6860,\n 6821,\n 6846,\n 6855,\n 8247,\n 6847,\n 6864,\n 6865,\n 9080,\n 6845,\n 6842,\n 6859,\n 6826,\n 6822,\n 6830,\n 6848,\n 870,\n 6853,\n 6840,\n 6833,\n 6862,\n 6831,\n 6849,\n 6824,\n 6827,\n 6839,\n 6828\n ],\n \"content\": \"\\\\label{sec_Population_operators}\\n\\\\begin{table*}[ht]\\n\\t\\\\renewcommand\\\\arraystretch{0.8}\\n\\t\\\\caption{Categorization of EC and Different Types of Neural Network} \\n\\t\\\\label{table_EC}\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{|p{1cm}|p{0.5cm}|p{1.8cm}|p{4.5cm}|p{1.4cm}|p{2cm}|p{1.3cm}|p{2cm}|}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multicolumn{3}{|l|}{} & CNN & DBN & RNN & AE & Others \\\\\\\\ \\\\hline\\n\\t\\t\\t\\\\multirow{12}{*}{\\\\shortstack[l]{Single\\\\\\\\objective}} & \\\\multirow{3}{*}{EA} & GAs &~ & ~ & ~ & ~ & ~ \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & GP & ~ & & ~ & ~ & ~ \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & ES & ~ & & ~ & ~ & ~ \\\\\\\\ \\\\cline{2-8} \\n\\t\\t\\t& \\\\multirow{3}{*}{SI} & ACO & ~ & & ~ & & \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & PSO & ~ & ~ & & ~ & ~ \\\\\\\\ \\\\cline{2-8} \\n\\t\\t\\t& \\\\multirow{6}{*}{Other} & Memetic & ~ & & & & \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & DE & ~ & & ~ & ~ & ~ \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & HCA & ~ & & & & \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & CVOA & & & ~ & & \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & Hyper-heuristic & & ~ & & & \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & FA & ~ & & & & \\\\\\\\ \\\\cline{3-8} \\n\\t\\t\\t& & AIS & ~ & & & & \\\\\\\\ \\\\hline\\n\\t\\t\\t\\\\multirow{2}{*}{\\\\shortstack[l]{Multi-\\\\\\\\objective}} & \\\\multicolumn{2}{l|}{EA} & ~ & ~ & ~ & & ~ \\\\\\\\ \\\\cline{2-8} \\n\\t\\t\\t& \\\\multicolumn{2}{l|}{SI} & ~ & ~ & & & \\\\\\\\ \\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\\\vspace{-0.5cm}\\n\\\\end{table*}\\nThis section discusses the population updating process as shown in Fig.~\\\\ref{fig_flowchart}. Generally, the population updating varies greatly among existing ENAS algorithms because they may employ different EC methods which are with different updating mechanisms. Table~\\\\ref{table_EC} shows the ENAS algorithms which are classified according to the EC methods that they employ and different types of DNNs that they target at. Obviously, the EA-based ENAS algorithms dominate the ENAS. To be more specific, the GA-based ENAS is the most popular approach which largely owes to the convenience of architecture representation in GA. As a result, we give a detailed introduction to the EA-based ENAS including the selection strategy at first. Immediately after, we present a summary of the SI-based ENAS methods and others separately. The corresponding multi-objective ENAS algorithms are also introduced at the end of each subsection.\",\n \"id\": \"d3650a8a-21c2-4aa1-9b61-b69c455614a3\",\n \"level\": \"section\",\n \"origin_cites_number\": 138,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Population Updating\"\n ]\n ],\n \"subsections\": [\n \"a152dfdd-9b9e-46a8-8504-9097a22a6736\",\n \"f362ecc4-6be9-45c6-b9e4-700fb0393594\",\n \"670f0571-da2a-4b7f-ab3e-86d8a80bfc64\"\n ],\n \"title\": \"Population Updating\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 6851,\n 6843,\n 6869,\n 6821,\n 870,\n 9079,\n 6868,\n 6867,\n 6837,\n 8247,\n 6847,\n 6863,\n 6853,\n 9080,\n 6844,\n 6841,\n 6834,\n 6845,\n 6829,\n 6831,\n 6849,\n 6824,\n 6866,\n 6842,\n 6827,\n 6839,\n 6832,\n 8302,\n 6858,\n 6826\n ],\n \"content\": \"\\\\begin{table}[]\\n\\t\\\\renewcommand\\\\arraystretch{0.8}\\n\\t\\\\caption{Selection Strategy} \\n\\t\\\\label{table_selection_strategy}\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{p{2cm}|p{5cm}}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tElitism & ~ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tDiscard the worst or the oldest & ~\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tRoulette & ~ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tTournament selection & ~\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tOthers & ~ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\\\vspace{-0.5cm}\\t\\n\\\\end{table}\\nThe dash box in Fig.~\\\\ref{fig_flowchart} shows the general flow of population updating in EA-based ENAS. In this section, we will introduce the selection strategy and the evolutionary operations which collectively update the population. Specifically, the first stage of population updating is selection. The selection strategies can be classified into several types, where Table~\\\\ref{table_selection_strategy} shows the main kinds of selection strategies. Note that the selection strategy can be not only used in choosing individuals as parents to generate offspring with the evolutionary operators, but also used in the environmental selection stage which chooses individuals to make up the next population. Zhang \\\\textit{et al.}~ termed these two selections as mate selection and environmental selection separately.\\nExisting selection strategies can be divided into five categories: elitism, discarding the worst, roulette wheel selection, tournament selection and others. The simplest strategy is elitism which retains the individuals with higher fitness. However, this can cause a loss of diversity in the population, which may lead the population falling into local optima. Discarding the worst is similar to elitism, which removes the individuals with poor fitness values from the population. Real \\\\textit{et al.}~ used the aging evolution which discards the oldest individual in the population. Aging evolution can explore the search space more, instead of zooming in on good models too early, as non-aging evolution would. The same selection strategy was also used in~. Zhu \\\\textit{et al.}~ combined these two approaches to discarding the worst individual and the oldest individual at the same time. Roulette wheel selection gives every individual a probability according to its fitness among the population to survive (or be discarded), regardless it is the best or not. Tournament selection selects the best one from an equally likely sampling of individuals. Furthermore, Johner \\\\textit{et al.}~ used a ranking function to choose individuals by rank. A selection trick termed as niching was used in~ to avoid stacking into local optima. This trick allows offspring worse than parent to survive for several generations until evolving to a better one.\\nMost of the methods focus on preserving the well-performed individuals, however, Liu \\\\textit{et al.}~ emphasized the genes more than the survived individuals, where a gene can represent any components in the architecture. They believed the individuals which consist of the fine-gene set are more likely to have promising performance.\\nSome selection methods aim at preserving the diversity of the population. Elsken \\\\textit{et al.}~ selected individuals in inverse proportion to their density. Javaheripi \\\\textit{et al.}~ chose the parents based on the distance (difference) during the mate selection. They chose two individuals with the highest distance to promote the exploration search.\\nIn terms of evolutionary operations, mutation and crossover are two of the most commonly used operations in EA-based ENAS algorithms. Particularly, the mutation is only performed on a single individual, while the crossover takes two individuals to generate offspring.\\nThe mutation operator aims to search the global optimum around the individual. A simple idea is to allow the encoded information to vary from a given range. Sun \\\\textit{et al.}~ used the polynomial mutation~ on the parameters of layers which are expressed by real numbers.\\nTo make mutation not random, Lorenzo \\\\textit{et al.}~ proposed a novel Gaussian mutation based on a Gaussian regression to guide the mutation, i.e., the Gaussian regression can predict which architecture may be good, and the newly generated individuals are sampled in the regions of the search space, where the fitness values are likely to be high. This makes mutation to have a ``direction\\\".\\nMoreover, Maziarz \\\\textit{et al.}~ used an RNN to guide the mutation operation. In this work, the mutation operations were not sampled at random among the possible architectural choices, but were sampled from distributions inferred by an RNN. Using an RNN to control the mutation operation can also be seen in other methods such as~.\\nSome researches investigated the diversity of the population after mutation. Qiang \\\\textit{et al.}~ used a variable mutation probability. They used a higher probability in the early stage for better exploration and a lower probability in the later stage for better exploitation. It has been effectively applied to many other methods~. To maintain the diversity of the population after the mutation operation, Tian \\\\textit{et al.}~ used \\\\textit{force mutation} and distance calculation, which ensures the individual in the population is not particularly similar to other individuals (especially the best one). Kramer \\\\textit{et al.}~ used the (1+1)-evolutionary strategy that generates an offspring based on a single parent with bit-flip mutation, and used a mutation rate to control and niching to overcome local optima.\\nThe intensive computational cost of ENAS presents a bottleneck which will be discussed in later sections. To reduce the unaffordable computational cost and time, some kinds of the mutation have also been designed. Zhang \\\\textit{et al.}~ proposed an exchange mutation which exchanges the position of two genes of the individual, i.e., exchanging the order of layers. This will not bring new layers and the weights in neural networks can be completely preserved, which means that the offspring do not have to be trained from scratch. \\nChen \\\\textit{et al.}~ introduced two function-preserving operators for DNNs, and these operators are termed as network morphisms~. The network morphisms aim to change the DNN architecture without the loss of the acquired experience. The network morphisms change the architecture from $F(\\\\cdot)$ to $G(\\\\cdot)$, which satisfies the condition formulated by Equation~(\\\\ref{equ_function-preserving}):\\n\\\\begin{equation}\\n\\\\label{equ_function-preserving}\\n\\\\forall x,\\\\quad F(x)=G(x)\\n\\\\end{equation}\\nwhere $x$ denotes the input of the DNN. The network morphisms can be regarded as a function-preserving mutation operation. With this operation, the mutated individuals will not have worse performance than their parents. To be more specific, Chen \\\\textit{et al.}~ proposed \\\\textit{net2widernet} to obtain a wider net and \\\\textit{net2deepernet} to obtain a deeper net. Elsken \\\\textit{et al.}~ extended the network morphisms with two popular network operations: skip connections and batch normalization. Zhu \\\\textit{et al.}~ proposed five well-designed function-preserving mutation operations to guide the evolutionary process by the information which have already learned. To avoid local optimal, Chen \\\\textit{et al.}~ added noises in some function-preserving mutation, and in the experiment, they found that by adding noises to pure network morphism, instead of compromising the efficiency, it by contrast, improved the final classification accuracy.\\nPlease note that all the network morphisms can only increase the capacity of a network because if one would decrease the network’s capacity, the function-preserving property could not be guaranteed~. As a result, the architecture generated by network morphisms is only going to get larger and deeper, which is not suitable for a device with limited computing resources, such as a mobile phone. In order for the network architecture to be reduced, Elsken \\\\textit{et al.}~ proposed the approximate network morphism, which satisfies Equation~(\\\\ref{equ_approximate_network_morphism}):\\n\\\\begin{equation}\\n\\\\label{equ_approximate_network_morphism}\\n\\\\forall x,\\\\quad F(x)\\\\approx G(x)\\n\\\\end{equation}\\nFor the crossover operator, the single-point crossover~ is the most popular method in EA-based ENAS~ because of its implementation simplicity. However, single-point crossover can typically apply to two individuals with equal lengths only. Therefore, this cannot be applied to variable-length individuals. To this end, Sun \\\\textit{et al.}~ proposed an efficient crossover operator for individuals with variable lengths. Sapra \\\\textit{et al.}~ proposed a disruptive crossover swapping the whole cluster (a sequence of layers) between both individuals at the corresponding positions rather than only focusing on the parameters of layers. Sun \\\\textit{et al.}~ used the Simulated Binary Crossover (SBX)~ to do a combination of the encoded parameters from two matched layers. Please note that the encoded parameters after SBX are not the same as that of both parents, which are quite different from other crossover operators.\\nEAs for multi-objective ENAS is gaining more and more attention from researchers. The single objective ENAS algorithms are always concerned about only one objective, e.g., the classification accuracy, and these algorithms have only one goal: searching for the architecture with the highest accuracy. In general, most of the multi-objective ENAS algorithms aim at dealing with both the performance of the neural network and the number of parameters simultaneously~.\\nHowever, these objective functions are often in conflict with each other. For example, getting a higher accuracy often requires a more complicated architecture with the need of more computational resources. On the contrary, a device with limited computational resource, e.g., a mobile phone, cannot afford such sophisticated architectures. \\nThe simplest way to tackle the multi-objective optimization problem is by converting it into a single objective optimization problem with weighting factors, i.e., the weighted summation method. The Equation~(\\\\ref{equ_multi-fitness})\\n\\\\begin{equation}\\n\\\\label{equ_multi-fitness}\\nF = \\\\lambda f_{1} + (1-\\\\lambda) f_{2}\\n\\\\end{equation}\\nis the classical linear form to weight two objective functions $f_{1},f_{2}$ into a single objective function, where the $\\\\lambda \\\\in (0, 1)$ denotes the weighting factor. In~, the multi-objective optimization problem was solved by using the available single objective optimization methods by Equation~(\\\\ref{equ_multi-fitness}) of the weighted summation. Chen \\\\textit{et al.}~ did not adopt the linear addition as the objective function, whereas using a nonlinear penalty term. However, the weights manually defined may incur bias~.\\nSome algorithms have been designed and widely used in multi-objective optimization, such as NSGA-II~, and MOEA/D~, which have been also used in ENAS methods such as~. These methods aim to find a Pareto-font set (or non-dominant set). Only these methods are in the multi-objective category of Table~\\\\ref{table_EC}. Some researches have made improvements on these multi-objective optimization methods for better use in ENAS.\\nBaldeon \\\\textit{et al.}~ chose the penalty based boundary intersection approach in MOEA/D because training a neural network involves nonconvex optimization and the form of Pareto Font is unknown.\\nLEMONADE~ divided the objective function into two categories: $f_{exp}$ and $f_{cheap}$. $f_{exp}$ denotes the expensive-to-evaluate objectives (e.g., the accuracy), while $f_{cheap}$ denotes the cheap-to-evaluate objectives (e.g., the model size). In every iteration, they sampled parent networks with respect to sparsely distribution based on the cheap objectives $f_{cheap}$ to generate offspring. Therefore, they evaluated $f_{cheap}$ more times than $f_{exp}$ to save time. Schoron \\\\textit{et al.}~ also took the use of the LEMONADE proposed by Elsken \\\\textit{et al.}~.\\nDue to that NSGA-III~ may fall into the small model trap (this algorithm prefers small models), Yang \\\\textit{et al.}~ have made some improvements to the conventional NSGA-III for favoring larger models.\",\n \"id\": \"a152dfdd-9b9e-46a8-8504-9097a22a6736\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 75,\n \"parent_id\": \"d3650a8a-21c2-4aa1-9b61-b69c455614a3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Population Updating\"\n ],\n [\n \"subsection\",\n \"EAs for ENAS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"EAs for ENAS\"\n },\n {\n \"cite_extract_rate\": 0.18181818181818102,\n \"cites\": [\n 6859,\n 6835\n ],\n \"content\": \"PSO is inspired by the bird flocking or fish schooling~, and is easy to implement compared with other SI algorithms.\\nJunior \\\\textit{et al.}~ used their implementation of PSO to update the particles based on the layer instead of the parameters of the layer. Gao \\\\textit{et al.}~ developed a gradient-priority particle swarm optimization algorithm to handle issues including the low convergence efficiency of PSO when there are a large number of hyper-parameters to be optimized. They expected the particle to find the locally optimal solution at first, and then move to the global optimal solution.\\nFor ACO, the individuals are generated in a quite different way. Several ants are in an ant colony\\\\footnote{The population in ACO also termed as colony.}. Each ant moves from node to node following the pheromone instructions to build an architecture. The pheromone is updated every generation. The paths of well-performed architecture will maintain more pheromone to attract the next ant for exploitation and at the same time, the pheromone is also decaying (i.e., pheromone evaporation), which encourages other ants to explore other areas. Byla \\\\textit{et al.}~ let the ants choose the path from the node to node in a graph whose depth increases gradually. Elsaid \\\\textit{et al.}~ introduced different ant agent types to act according to specific roles to serve the needs of the colony, which is inspired by the real ants species.\\nSI for multi-objective ENAS started only in the last two years and the research of this field is scarce which can be seen from Table~\\\\ref{table_EC}.\\nLi \\\\textit{et al.}~ used the bias-variance framework on their proposed multi-objective PSO to get a more accurate and stable architecture.\\nWu \\\\textit{et al.}~ used the MOPSO~ for neural networks pruning. The $G_{best}$ is selected according to the crowding distance in the non-dominant solutions set. Wang \\\\textit{et al.}~ used the OMOPSO~, which selects the leaders using a crowding factor and the $G_{best}$ is selected from the leaders. To better control the balance between convergence and diversity, Jiang \\\\textit{et. al.}~ proposed a MOPSO/D algorithm based on an adaptive penalty-based boundary intersection.\",\n \"id\": \"f362ecc4-6be9-45c6-b9e4-700fb0393594\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"d3650a8a-21c2-4aa1-9b61-b69c455614a3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Population Updating\"\n ],\n [\n \"subsection\",\n \"SI for ENAS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"SI for ENAS\"\n },\n {\n \"cite_extract_rate\": 0.42857142857142805,\n \"cites\": [\n 6831,\n 6865,\n 6826\n ],\n \"content\": \"Different from GAs, the mutation of DE exploits the information from three individuals. \\nSome ENAS methods like~ chose DE to guide the offspring generation. However, there is little difference between different DE-based ENAS algorithms.\\nWang \\\\textit{et al.}~ proposed a hybrid PSO-GA method. They used PSO to guide the evolution of the parameters in each block encoded in decimal notation. Meanwhile, using GA to guide the evolution of the shortcut connections is encoded in binary notation. Because PSO performs well on continuous optimization and GA is suitable for optimization with binary values, this hybrid method can search architectures effectively.\\nHCA can be interpreted as a very simple evolutionary algorithm. For example, in~ the evolutionary operators only contain mutation and no crossover, and the selection strategy is relatively simple. The memetic algorithm is the hybrids of EAs and local search. Evans \\\\textit{et al.}~ integrated the local search (as gradient descent) into GP as a fine-tuning operation. The CVOA~ was inspired by the new respiratory virus, COVID-19. The architecture was found by simulating the virus spreads and infecting healthy individuals. Hyper-heuristic contains two levels: high-level strategy and low-level heuristics, and a domain barrier is between these two levels. Hence the high-level strategy is still useful when the application domain is changed. AIS was inspired by theories related to the mammal immune system and do not require the crossover operator compared to the GA~.\",\n \"id\": \"670f0571-da2a-4b7f-ab3e-86d8a80bfc64\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"d3650a8a-21c2-4aa1-9b61-b69c455614a3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Population Updating\"\n ],\n [\n \"subsection\",\n \"Other EC Techniques for ENAS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Other EC Techniques for ENAS\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 6822,\n 6851,\n 6843,\n 6821,\n 6848,\n 9079,\n 6846,\n 8247,\n 6871,\n 6841,\n 6833,\n 6845,\n 8955,\n 6862,\n 6831,\n 6842,\n 6872,\n 6870,\n 1461,\n 6826,\n 6835\n ],\n \"content\": \"\\\\label{sec_evaluation}\\nIn this section, we will discuss the strategies to improve the efficiency of evaluations, with the consideration that the evaluation is often the most time-consuming stage of ENAS algorithms~.\\nReal \\\\textit{et al.}~ used 250 computers to finish the LargeEvo algorithm over 11 days. Such computational resources are not available for everyone interested in NAS. Almost all of the methods evaluate individuals by training them first and evaluating them on the validation/test dataset. Since the architecture is becoming more and more complex, it will take a lot of time for training each architecture to convergence. So it is needed to investigate new methods to shorten the evaluation time and reduce the dependency on large amounts of computational resources.\\nTable~\\\\ref{table_shorten_time} lists five of the most common methods to reduce the time: weight inheritance, early stopping policy, reduced training set, reduced population, and population memory. We would like to introduce the five kinds of methods first, then other kinds of promising methods next, and finally the surrogate-assisted methods at the end of this section.\\nBecause the evolutionary operators do not completely disrupt the architecture of an individual, some parts of the newly generated individuals are the same as their parents. The weights of the same parts can be easily inherited. With the weight inheritance, the neural networks no longer need to be trained completely from scratch. This method has been used in~ 20 years ago.\\nMoreover, as mentioned in Section~\\\\ref{sec_Population_operators}, the network morphisms change the network architecture without loss of the acquired experience. This could be regarded as the ultimate weight inheritance because it solved the weight inheritance problem in the changed architecture part. The ultimate weight inheritance lets the new individual completely inherit the knowledge from their parents, which will save a lot of time.\\nThe early stopping policy is another method which has been used widely in NAS. The simplest way is to set a fixed and relatively small number of training epochs. This method is used in~, where training the individuals after a small number of epochs is sufficient. Similarly, Assunccao \\\\textit{et al.}~ let the individuals undergo the training for the same and short time each epoch (although this time is not fixed and will increase with the epoch), to allow the promising architectures have more training time to get a more precise evaluation, So \\\\textit{et al.}~ set hurdles after a fixed number of epochs. The weak individuals stop training early to save time.\\nHowever, the early stopping policy can lead to inaccurate estimation about individuals' performance (especially the large and complicated architecture), which can be seen in Fig.~\\\\ref{fig_learning_curve}. In Fig.~\\\\ref{fig_learning_curve}, \\\\emph{individual2} performs better than \\\\emph{individual1} before epoch \\\\emph{t1}, whereas \\\\emph{individual1} performs better in the end. Yang \\\\textit{et al.}~ also discussed this phenomenon. So, it is crucial to determine at which point to stop.\\nNote that neural networks can converge or hardly improve its performance after several epochs, as seen in the \\\\emph{t1} for \\\\emph{individual2} and the \\\\emph{t2} for \\\\emph{individual1} in Fig.~\\\\ref{fig_learning_curve}. Using the performance estimated at this point can evaluate an individual's relatively accurately with less training time. Therefore, some methods such as~ stopped training when observing there is no significant performance improvement.\\nSuganuma \\\\textit{et al.}~ used the early stopping policy based on a reference curve. If the accuracy curve of an individual is under the reference curve for successive epochs, then the training will be terminated and this individual is regarded as a poor one. After every epoch, the reference curve is updated by the accuracy curve of the best offspring. \\nThe reduced training set, i.e., using a subset of data that assuming similar properties to a large dataset, can also shorten the time effectively. Liu \\\\textit{et al.}~ explored promising architectures by training on a subset and used transfer learning on the large original dataset. Because there are so many benchmark datasets in the image classification field, the architecture can be evaluated on a smaller dataset (e.g., CIFAR-10) first and then applied on a large dataset (such as CIFAR-100 and ImageNet~). The smaller dataset can be regarded as the proxy for the large one.\\nThe reduced population is a unique method of ENAS since other NAS approaches do not have a population. Assunccao \\\\textit{et al.}~ reduced the population based on their previous algorithm~ to speed up the evolution. However, simply reducing the population may not explore the search space well in each epoch and may lose the global search ability. Another way is reducing the population dynamically. For instance, Fan \\\\textit{et al.}~ used the ($\\\\mu+\\\\lambda$) evolution strategy and divided the evolution into three stages with the population reduction, which aims to find the balance between the limited computing resources and the efficiency of the evolution. The large population in the first stage is to ensure the global search ability, while the small population in the last stage is to shorten the evolution time. Instead of reducing the population, Liu \\\\textit{et al.}~ evaluated the downsizing architecture with smaller size at an early stage of evolution. Similarly, Wang \\\\textit{et al.}~ did not evaluate the whole architecture but starting with a single block, and then the blocks were stacked to build architecture as evolution proceeds.\\n\\\\begin{table}[]\\n\\t\\\\renewcommand\\\\arraystretch{0.8}\\n\\t\\\\caption{Different Methods to Shorten the Evaluation Time} \\n\\t\\\\label{table_shorten_time}\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{p{2.5cm}|p{5cm}}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tWeight inheritance & ~ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tEarly stopping policy & ~\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tReduced training set & ~ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tReduced population & ~\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tPopulation memory & ~\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\\\vspace{-0.3cm}\\n\\\\end{table}\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.8\\\\linewidth]{learning_curve}\\n\\t\\\\caption{Two learning curve of two different individuals.}\\n\\t\\\\label{fig_learning_curve}\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{figure}\\nThe population memory is another category of the unique methods of ENAS. It works by reusing the corresponding architectural information that has previously appeared in the population. In the population-based methods, especially the GA based methods (e.g., in~), it is natural to maintain well-performed individuals in the population in successive generations. Sometimes, the individuals in the next population directly inherit all the architecture information of their parents without any modification, so it is not necessary to evaluate the individuals again. Fujino \\\\textit{et al.}~ used memory to record the fitness of individuals, and if the same architecture encoded in an individual appears, the fitness value is retrieved from memory instead of being reevaluated. Similarly, Miahi \\\\textit{et al.}~ and Sun \\\\textit{et al.}~ employed a hashing method for saving pairs of architecture and fitness of each individual and reusing them when the same architecture appears again. Johner \\\\textit{et al.}~ prohibited the appearance of architectures that have appeared before offspring generation. This does reduce the time, however, the best individuals are forbidden to remain in the population which may lead the population evolve towards a bad direction.\\nThere are many other well-performed methods to reduce the time in ENAS. Rather than training thousands of different architectures, one-shot model~ trained only one SuperNet to save time. Different architectures, i.e., the SubNets, are sampled from the SuperNet with the shared parameters. Yang \\\\textit{et al.}~ believed the traditional ENAS methods without using SuperNet which were less efficient for models to be optimized separately. In contrast, the one-shot model optimizes the architecture and the weights alternatively. But the weight sharing mechanism brings difficulty in accurately evaluating the architecture. Chu \\\\textit{et al.}~ scrutinized the weight-sharing NAS with a fairness perspective and demonstrated the effectiveness. However, there remain some doubts that cannot explain clearly in the one-shot model. The weights in the SuperNet are coupled. It is unclear why inherited weights for a specific architecture are still effective~.\\nMaking the use of hardware can reduce the time, too. Jiang \\\\textit{et al.}~ used a distributed asynchronous system which contained a major computing node with 20 individual workers. Each worker is responsible for training a single block and uploading its result to the major node in every generation. Wang \\\\textit{et al.}~ designed an infrastructure which has the ability to leverage all of the available GPU cards across multiple machines to concurrently perform the objective evaluations for a batch of individuals.\\nColangelo \\\\textit{et al.}~ designed a reconfigurable hardware framework that fits the ENAS. As they claimed, this was the first work of conducting NAS and hardware co-optimization.\\nFurthermore, Lu \\\\textit{et al.}~ adopted the concept of proxy models, which are small-scale versions of the intended architectures. For example, in a CNN architecture, the number of layers and the number of channels in each layer are reduced. However, the drawback of this method is obvious: the loss of prediction accuracy. Therefore, they performed an experiment to determine the smallest proxy model that can provide a reliable estimate of performance at a larger scale.\\nAll the above methods obtain the fitness of individuals by directly evaluating the performance on the validation dataset. An alternative way is to use indirect methods, namely the performance predictors. As summarized in~, the performance predictors can be divided into two categories: performance predictors based on the learning curve and end-to-end performance predictors, both of which are based on the training-predicting learning paradigm. This does not mean the performance predictor does not undergo the training phase at all, while it means learning from the information obtained in the training phase, and uses the knowledge learned to make a reasonable prediction for other architectures.\\nRalwal \\\\textit{et al.}~ took the use of the learning curve based predictor, where the fitness is not calculated in the last epoch but is predicted by the sequence of fitness from first epochs. Specifically, they used a Long Short Term Memory (LSTM)~ as a sequence to sequence model, predicting the final performance by using the learning results of the first several epochs on the validation dataset.\\nSun \\\\textit{et al.}~ adopted a surrogate-assisted method which is called end-to-end performance predictor. The predictor does not need any extra information about the performance of individuals to be evaluated. The performance predictor is essentially a regression model mapping from the architecture to its performance. The regression model needs to be trained with sufficient training data pairs first, and each pair consists of architecture and its corresponding performance. Specifically, they chose random forest~ as the regression model to accelerate the fitness evaluations in ENAS. When the random forest receives a newly generated architecture as input, the adaptive combination of a huge number of regression trees which have been trained in advance in the forest gives the prediction.\",\n \"id\": \"fa11702e-7516-4849-b9c7-1422e93b4202\",\n \"level\": \"section\",\n \"origin_cites_number\": 63,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Efficient Evaluation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Efficient Evaluation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec_application}\\nThis section discussed different application fields of ENAS has involved. Generally, the ENAS algorithms can be applied to wherever DNNs can be applied. Table~\\\\ref{table_application} shows the wide range of applications and Table~\\\\ref{table_CIFAR} displays the performance of extraordinary ENAS methods on two popular and challenging datasets for image classification tasks, namely CIFAR-10 and CIFAR-100. Both of these two tables can show what ENAS has achieved so far.\",\n \"id\": \"390da3df-188a-48a9-991a-966d654bdd4d\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Applications\"\n ]\n ],\n \"subsections\": [\n \"06bc3345-4a23-45a7-a0e1-7338ed3b2e6b\",\n \"ce4d03d7-51ce-4271-828b-0f60a203e534\"\n ],\n \"title\": \"Applications\"\n },\n {\n \"cite_extract_rate\": 0.3671875,\n \"cites\": [\n 6851,\n 6838,\n 6837,\n 6863,\n 6844,\n 6841,\n 8955,\n 6829,\n 6876,\n 6832,\n 6857,\n 9079,\n 6875,\n 6874,\n 6850,\n 6870,\n 6836,\n 6835,\n 6869,\n 6843,\n 6860,\n 6821,\n 6846,\n 8247,\n 6847,\n 6864,\n 6873,\n 6865,\n 9080,\n 6834,\n 6845,\n 6842,\n 6866,\n 6830,\n 6822,\n 6848,\n 870,\n 6853,\n 6840,\n 6852,\n 6833,\n 6862,\n 6831,\n 6849,\n 6827,\n 6839,\n 6828\n ],\n \"content\": \"Table~\\\\ref{table_application} shows the applications of existing ENAS algorithms, which contains a wide range of real-world applications. \\n\\\\begin{table}[]\\n\\t\\\\renewcommand\\\\arraystretch{0.8}\\n\\t\\\\caption{Applications of Existing ENAS algorithms.} \\n\\t\\\\label{table_application}\\n\\t\\\\begin{tabular}{|p{1cm}|p{3.5cm}|p{3cm}|}\\n\\t\\t\\\\hline\\n\\t\\tCategory & Applications & References \\\\\\\\ \\\\hline\\n\\t\\t(1) & Image classification & ~ \\\\\\\\ \\\\hline\\n\\t\\t(1) & Image to image & ~ \\\\\\\\ \\\\hline\\n\\t\\t(1) & Emotion recognition & ~ \\\\\\\\ \\\\hline\\n\\t\\t(1) & Speech recognition & ~ \\\\\\\\ \\\\hline\\n\\t\\t(1) & Language modeling & ~ \\\\\\\\ \\\\hline\\n\\t\\t(1) & Face De-identification & ~ \\\\\\\\ \\\\hline\\n\\t\\t(2) & Medical image segmentation & ~ \\\\\\\\ \\\\hline\\n\\t\\t(2) & Malignant melanoma detection & ~ \\\\\\\\ \\\\hline\\n\\t\\t(2) & Sleep heart study & ~ \\\\\\\\ \\\\hline\\n\\t\\t(2) & Assessment of human sperm & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Wind speed prediction & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Electricity demand time series forecasting & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Traffic flow forecasting & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Electricity price forecasting & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Car park occupancy prediction & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Energy consumption prediction & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Time series data prediction & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Financial prediction & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Usable life prediction & ~ \\\\\\\\ \\\\hline\\n\\t\\t(3) & Municipal waste forecasting & ~ \\\\\\\\ \\\\hline\\n\\t\\t(4) & Engine vibration prediction & ~ \\\\\\\\ \\\\hline\\n\\t\\t(4) & UAV & ~ \\\\\\\\ \\\\hline\\n\\t\\t(4) & Bearing fault diagnosis & ~ \\\\\\\\ \\\\hline\\n\\t\\t(4) & Predicting general aviation flight data & ~ \\\\\\\\ \\\\hline\\n\\t\\t(5) & Crack detection of concrete & ~ \\\\\\\\ \\\\hline\\n\\t\\t(5) & Gamma-ray detection & ~ \\\\\\\\ \\\\hline\\n\\t\\t(5) & Multitask learning & ~ \\\\\\\\ \\\\hline\\n\\t\\t(5) & Identify Galaxies & ~ \\\\\\\\ \\\\hline\\n\\t\\t(5) & Video understanding & ~ \\\\\\\\ \\\\hline\\n\\t\\t(5) & Comics understanding & ~ \\\\\\\\ \\\\hline\\n\\t\\\\end{tabular}\\n\\\\vspace{-0.5cm}\\n\\\\end{table}\\nGenerally, these applications can be grouped into the following five different categories:\\n(1) Image and signal processing, including image classification which is the most popular and competitive field, image to image processing (including image restoration, image denoising, super-resolution and image inpainting), emotion recognition, speech recognition, language modelling, and face de-identification.\\n(2) Biological and biomedical tasks, including medical image segmentation, malignant melanoma detection, sleep heart study, and assessment of human sperm.\\n(3) Predictions and forecasting about all sorts of data, including the prediction of wind speed, car park occupancy, time-series data, financial and usable life, the forecasting of electricity demand time series, traffic flow, electricity price, and municipal waste.\\n(4) Engineering, including engine vibration prediction, Unmanned Aerial Vehicle (UAV), bearing fault diagnosis and predicting general aviation flight data.\\n(5) Others, including crack detection of concrete, gamma-ray detection, multitask learning, identify galaxies, video understanding, and comics understanding.\",\n \"id\": \"06bc3345-4a23-45a7-a0e1-7338ed3b2e6b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 128,\n \"parent_id\": \"390da3df-188a-48a9-991a-966d654bdd4d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Overview\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Overview\"\n },\n {\n \"cite_extract_rate\": 0.538461538461538,\n \"cites\": [\n 6869,\n 6851,\n 6877,\n 6821,\n 870,\n 6837,\n 8247,\n 6847,\n 6863,\n 6878,\n 6844,\n 6834,\n 6845,\n 6829,\n 6831,\n 97,\n 6842,\n 6827,\n 8319,\n 6836,\n 6826\n ],\n \"content\": \"\\\\label{sec_CIFAR}\\n\\\\begin{table*}[!htbp]\\n\\t\\\\renewcommand\\\\arraystretch{0.8}\\n\\t\\\\caption{The Comparison of the Classification Error Rate on CIFAR-10, CIFAR-100 and ImageNet}\\n\\t\\\\label{table_CIFAR}\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\begin{tabular}{|p{3.7cm}|p{1.3cm}|p{1.8cm}|p{2.4cm}|p{2.4cm}|p{3cm}|p{0.5cm}|}\\n\\t\\t\\\\hline\\n\\t\\t\\\\textbf{ENAS Methods} & \\\\textbf{GPU Days} & \\\\textbf{Parameters(M)} & \\\\textbf{CIFAR-10(\\\\%)} & \\\\textbf{CIFAR-100(\\\\%)} & {\\\\textbf{ImageNet(Top1/Top5 \\\\%)}} & \\\\textbf{Year} \\\\\\\\ \\\\hline\\n\\t\\tCGP-DCNN~ & --- & 1.1 & 8.1 & --- & --- & 2018 \\\\\\\\ \\\\hline\\n\\t\\tEPT~ & 2 & --- & 7.5 & --- & --- & 2020 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{GeNet~} & \\\\multirow{3}{*}{17} & --- & 7.1 & --- & --- & \\\\multirow{3}{*}{2017} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & --- & --- & 29.03 & --- & \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 156 & --- & --- & {27.87 / 9.74} &\\n\\t\\t\\\\\\\\ \\\\hline\\n\\t\\tEANN-Net~ & --- & --- & 7.05 $\\\\pm$ 0.02 & --- & --- & 2019 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{DeepMaker~} & \\\\multirow{2}{*}{3.125} & 1 & 6.9 & --- & --- &\\\\multirow{2}{*}{2020} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 1.89 & --- & 24.87 & --- & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{GeneCai (ResNet-50)~ } & \\\\multirow{2}{*}{0.024} & --- & 6.4 & --- & --- & \\\\multirow{2}{*}{2020} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & --- & --- & --- & {25.7 / 7.9} & \\\\\\\\ \\\\hline\\n\\t\\tCGP-CNN (ConvSet)~ & --- & 1.52 & 6.75 & --- & --- &2017 \\\\\\\\ \\\\hline\\n\\t\\tCGP-CNN (ResSet)~ & --- & 1.68 & 5.98 & --- & --- &2017 \\\\\\\\ \\\\hline\\n\\t\\tMOPSO/D-Net~ & 0.33 & 8.1 & 5.88 & --- & --- &2019 \\\\\\\\ \\\\hline\\n\\t\\tReseNet-50 (20\\\\% pruned)~ & --- & 6.44 & 5.85 & --- & --- &2018 \\\\\\\\ \\\\hline\\n\\t\\tImmuNeCS~ & 14 & --- & 5.58 & --- & --- &2019 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{EIGEN~} & 2 & 2.6 & 5.4 & --- & --- & \\\\multirow{2}{*}{2018} \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& 5 & 11.8 & --- & 21.9 & --- & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{LargeEvo~} & \\\\multirow{2}{*}{2750} & 5.4 & 5.4 & --- & --- & \\\\multirow{2}{*}{2017} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 40.4 & --- & 23 & --- & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{CGP-CNN (ConvSet)~} & 31 & 1.5 & 5.92 (6.48 $\\\\pm$ 0.48) & --- & --- & \\\\multirow{4}{*}{2019} \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& --- & 2.01 & --- & 26.7 (28.1 $\\\\pm$ 0.83) & --- & \\\\\\\\ \\\\cline{1-6}\\n\\t\\t\\\\multirow{2}{*}{CGP-CNN (ResSet)~} & 30 & 2.01 & 5.01 (6.10 $\\\\pm$ 0.89) & --- & --- & \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& --- & 4.6 & --- & 25.1 (26.8 $\\\\pm$ 1.21) & --- & \\\\\\\\ \\\\hline\\n\\t\\tMOCNN~ & 24 & --- & 4.49 & --- & --- & 2019 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{NASH~} & 4 & 88 & 4.4 & --- & --- & \\\\multirow{2}{*}{2017} \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& 5 & 111.5 & --- & 19.6 & --- & \\\\\\\\ \\\\hline\\n\\t\\tHGAPSO~ & 7+ & --- & 4.37 & --- & --- & 2018 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{3}{*}{DPP-Net~,~} & \\\\multirow{3}{*}{2} & 11.39 & 4.36 & --- & --- & \\\\multirow{3}{*}{2018} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 0.45 & 5.84 & --- & --- & \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 4.8 & --- & --- & {26.0 / 8.2} & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{AE-CNN~} & 27 & 2 & 4.3 & --- & --- &\\\\multirow{2}{*}{2019} \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& 36 & 5.4 & --- & 20.85 & --- & \\\\\\\\ \\\\hline\\n\\t\\tSI-ENAS~ & 1.8 & --- & 4.07 & 18.64 & --- & 2020 \\\\\\\\ \\\\hline\\n\\t\\tEPSOCNN~ & 4- & 6.77 & 3.69 & --- & --- & 2019 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{Hierarchical Evolution~ } & \\\\multirow{2}{*}{300} & --- & 3.63 $\\\\pm$ 0.10 & --- & --- & \\\\multirow{2}{*}{2017} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & --- & --- & --- & {20.3 / 5.2} & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{EA-FPNN~} & 0.5 & 5.8 & 3.57 & --- & --- & \\\\multirow{2}{*}{2018} \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& 1 & 7.2 & --- & 21.74 & --- & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{AmoebaNet-A~} & \\\\multirow{2}{*}{3150} & 3.2 & 3.34 $\\\\pm$ 0.06 & --- & --- & \\\\multirow{2}{*}{2018} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t&& 86.7 & --- & --- & {17.2 / 3.9} & \\\\\\\\ \\\\hline\\n\\t\\tFirefly-CNN~ & --- & 3.21 & 3.3 & 22.3 & --- & 2019 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{3}{*}{CNN-GA~} & 35 & 2.9 & 3.22 & --- & --- & \\\\multirow{3}{*}{2018} \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& 40 & 4.1 & --- & 20.53 & --- & \\\\\\\\ \\\\cline{2-6}\\n\\t\\t& 35 & --- & --- & --- & {25.2 / 7.7} & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{3}{*}{JASQNet~} & \\\\multirow{3}{*}{3} & 3.3 & 2.9 & --- & --- & \\\\multirow{3}{*}{2018} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 1.8 & 2.97 & --- & --- & \\\\\\\\ \\\\cline{3-6} \\n\\t\\t& & 4.9 & --- & --- & {27.2 / ---} & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{RENASNet~} & \\\\multirow{2}{*}{6} & 3.5 & 2.88 $\\\\pm$ 0.02 & --- & --- & \\\\multirow{2}{*}{2018} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 5.36 & --- & --- & {24.3 / 7.4} & \\\\\\\\ \\\\hline\\n\\t\\tNSGA-Net~ & 4 & 3.3 & 2.75 & --- & --- & 2018 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{4}{*}{CARS~} & \\\\multirow{4}{*}{0.4} & 2.4 & 3 & --- & --- & \\\\multirow{4}{*}{2019} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 3.6 & 2.62 & --- & --- & \\\\\\\\ \\\\cline{3-6} \\n\\t\\t& & 3.7 & --- & --- & {27.2 / 9.2} & \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 5.1 & --- & --- & {24.8 / 7.5} & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{3}{*}{LEMONADE~} & \\\\multirow{3}{*}{80} & 13.1 & 2.58 & --- & --- & \\\\multirow{3}{*}{2018} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 0.5 & 4.57 & --- & --- & \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & --- & --- & --- & {28.3 / 9.6} & \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{2}{*}{EENA~} & \\\\multirow{2}{*}{0.65} & 8.47 & 2.56 & --- & --- & \\\\multirow{2}{*}{2019} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 8.49 & --- & 17.71 & --- & \\\\\\\\ \\\\hline\\n\\t\\tEEDNAS-NNMM~ & 0.5 & 4.7 & 2.55 & --- & --- & 2019 \\\\\\\\ \\\\hline\\n\\t\\t\\\\multirow{5}{*}{NSGANet~} & \\\\multirow{5}{*}{27} & 0.2 & 4.67 & --- & --- & \\\\multirow{5}{*}{2019} \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 4 & 2.02 & --- & --- & \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 0.2 & --- & 25.17 & --- & \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 4.1 & --- & 14.38 & --- & \\\\\\\\ \\\\cline{3-6}\\n\\t\\t& & 5.0 & --- & --- & {23.8 / 7.0} & \\\\\\\\ \\\\hline\\n\\t\\\\end{tabular}\\n\\\\vspace{-0.3cm}\\n\\\\end{table*}\\nIn Table~\\\\ref{table_application}, it is obvious to see that many ENAS methods are applied to the image classification tasks. The benchmark dataset, CIFAR-10 which contains a total of ten classes, and the CIFAR-100 is the advanced dataset including a hundred classes. These two datasets have been widely used in image classification tasks, and the accuracy on these two challenging datasets can represent the ability of the architecture searched. In addition, ImageNet~ is a more challenging benchmark dataset, which contains 1,000 classes and more than one million images. Because CIFAR-10 and CIFAR-100 are relatively small and easy to be over-fitting nowadays~, the results of the ENAS methods on ImageNet are also included. We have collected the well-performed ENAS methods tested on these three datasets and showed the results in Table~\\\\ref{table_CIFAR}, where the methods are ranked in an ascending order of their best accuracy on CIFAR-10, i.e., the methods are ranked in a descending order of their error rate based on the conventions. The data shown under column ``CIFAR-10\\\" and ``CIFAR-100\\\" denotes the error rate of each method on the corresponding datasets. Especially, as for ImageNet, we report both top1 and top5 error rates. Furthermore, the ``GPU Days\\\", which was initially proposed in~, denotes the total search time of each method, it can be calculated by Equation~(\\\\ref{equ_GPUDays})\\n\\\\begin{equation}\\n\\\\label{equ_GPUDays}\\nGPU\\\\ Days = The\\\\ number\\\\ of\\\\ GPUs \\\\times t\\n\\\\end{equation}\\nwhere the $t$ denotes the number of days that each method searched for. In Table~\\\\ref{table_CIFAR} ``Parameters\\\" denotes the total number of parameters which can represent the capability of architecture and the complexity. In addition, the symbol ``---\\\" in Table~\\\\ref{table_CIFAR} implies there is no result publically reported by the corresponding paper. The year reported in this table is its earliest time made public. Furthermore, there are additional notes provided for Table~\\\\ref{table_CIFAR}. Firstly, if there are several results reported from literature, such as CGP-DCNN~ and CARS~, we choose to report one or two of the most representative architectures. Secondly, we name two algorithms without proper names in Table~\\\\ref{table_CIFAR} (i.e., Firefly-CNN~ and EEDNAS-NNMM~) based on the EC methods they used and the first letter of the title. Thirdly, we report the classification errors of CGP-CNN~ in the format of ``best (mean $\\\\pm$ std)\\\". Finally, the symbols of ``+\\\" and ``-\\\" in the column of ``GPU Days\\\" denote the meaning of ``more than\\\" and ``less than\\\", respectively.\\nIn principle, there is no totally fair comparison. Due to the following two reasons: (1) The encoding space including the initial space and the search space is different from each other. There are two extreme cases in the initial space: trivial initialization which starts at the simplest architecture and rich initialization which starts from a well-designed architecture (e.g., ResNet-50~). Meanwhile the size of the search space is largely different, e.g., Ref~ only takes the kernel size into the search space. (2) Different tricks exploited in the methods, e.g., the ``cutout\\\", can make the comparisons unfair, too. The ``cutout\\\" refers to a regularization method~ used in the training of CNNs, which could improve the final performance appreciably.\\nTable~\\\\ref{table_CIFAR} shows the progress of ENAS for image classification according to the accuracy on CIFAR-10: LargeEvo algorithm~ (5.4\\\\%, 2017), LEMONADE~ (2.58\\\\%, 2018), and NSGANet~ (2.02\\\\%, 2019). Many ENAS methods have a lower error rate than ResNet-110~ with 6.43\\\\% error rate on CIFAR-10, which is a manually well-designed architecture. Therefore, the architecture found by ENAS can reach the same level or exceed the architecture designed by experts. It shows that the ENAS is reliable and can be used in other application fields.\",\n \"id\": \"ce4d03d7-51ce-4271-828b-0f60a203e534\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 39,\n \"parent_id\": \"390da3df-188a-48a9-991a-966d654bdd4d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Comparisons on CIFAR-10, CIFAR-100 and ImageNet\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Comparisons on CIFAR-10, CIFAR-100 and ImageNet\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec_challenges_and_issues}\\nDespite the positive results of the existing ENAS methods, there are still some challenges and issues which need to be addressed. In the following, we will briefly discuss them.\",\n \"id\": \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Challenges and Issues\"\n ]\n ],\n \"subsections\": [\n \"d6c13ce5-ab8c-4d29-a77d-718dbcbb0cf9\",\n \"8f512a41-ca51-4724-a511-f83f8ff395ee\",\n \"0511a9dd-9340-4ba1-ba4f-5c1c82725c9e\",\n \"97b2a2f7-bddb-45d3-819a-f3dcbad41cba\",\n \"ed6f44db-5133-4077-be4f-e9675dc00180\",\n \"14a9080c-0216-4f7c-899b-00f8965feb35\"\n ],\n \"title\": \"Challenges and Issues\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 8955,\n 6831,\n 8247,\n 6880,\n 870,\n 6854,\n 6825,\n 6879,\n 6844,\n 6830\n ],\n \"content\": \"The effectiveness of ENAS is questioned by many researchers. Wistuba \\\\textit{et al.}~ noticed that the random search can get a well-performed architecture and has proven to be an extremely strong baseline. Yu \\\\textit{et al.}~ showed the state-of-the-art NAS algorithms performed similarly to the random policy on average. Liashchynskyi \\\\textit{et al.}~ compared with grid search, random search, and GA for NAS, which showed that the architecture obtained by GA and random search have similar performance. There is no need to use complicated algorithms to guide the search process if random search can outperform NAS based on EC approaches.\\nHowever, the evolutionary operator in~ only contains a recombination operator, which limited the performance of ENAS algorithms. Although random search can find a well-performed architecture in the experiments, it cannot guarantee that it will find a good architecture every time. Moreover, recent researches~ also showed the evolutionary search was more effective than random search. Furthermore, the experiments in Amoebanet~ and NAS-Bench-201~ also showed that ENAS can search for and find better architectures. Thus, there is an urgent need to design an elaborated experiment to reveal what components in ENAS are responsible for the effectiveness of the algorithm, especially in a large encoding space.\\nIn Section~\\\\ref{sec_Population_operators}, two types of operators have been introduced. We note that some methods like the LargeEvo algorithm~ only use single individual based operator (mutation) to generate offspring. The main reason that they did not involve the crossover operator in their method come from two reasons: the first is for simplicity~, and the second is that simply combining a section of one individual with a section of another individual seems ``ill-suited\\\" to the neural network paradigm~. Secondly in~, the authors believed that there was no indication that a recombination operation applied to two individuals with high fitness would result in an offspring with similar or better fitness.\\nHowever, the supplemental materials in~ demonstrated the effectiveness of the crossover operator in this method. This method can find a good architecture with the help of the crossover operation. On the contrary, when crossover is not used, the architecture found is not promising, unless it runs for a long time. In fact, the mutation operator let an individual explore the neighbouring region, and it is a gradually incremental search process like searching step by step. Crossover (recombination) can generate offspring dramatically different from the parents, which is more likely a stride. So, this operator has the ability to efficiently find a promising architecture. Chu \\\\textit{et al.}~ preferred that while a crossover mainly contributes to exploitation, a mutation is usually aimed to introduce exploration. These two operators play different roles in the evolutionary process. But there is not a sufficient explanation of how the crossover operator works. Maybe some additional experiments need to be done on the methods without the crossover operator. \\nThe EC approach generally performs well when dealing with practical problems (e.g., the vehicle routing problem). This is mainly because EC is designed collectively by considering the domain knowledge in most cases~. Similarly, another key reason to ensure the effectiveness of the ENAS algorithm is that the EC approach can effectively consider (and embed) some domain knowledge of the neural architecture. Since the original intention of ENAS is to design neural architectures automatically without manual experience, the future development of ENAS will need to continue to reduce any artificial experience by including prior knowledge of the encoding space. In this case, how to effectively embed domain knowledge into ENAS is a big challenge.\",\n \"id\": \"d6c13ce5-ab8c-4d29-a77d-718dbcbb0cf9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Challenges and Issues\"\n ],\n [\n \"subsection\",\n \"The Effectiveness\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"The Effectiveness\"\n },\n {\n \"cite_extract_rate\": 0.555555555555555,\n \"cites\": [\n 6829,\n 6851,\n 6881,\n 6866,\n 870\n ],\n \"content\": \"The scale of the datasets used in most ENAS methods is relatively large. Taking the image classification task as an example, the MNIST dataset~ is one of the earliest datasets. In total, it contains $70, 000$ $28\\\\times28$ grayscale images. Later in 2009, CIFAR-10 and CIFAR-100~ including $60, 000$ $32\\\\times32$ color images are medium-scale datasets. One of the most well-known large-scale datasets is ImageNet~, which provides more than 14 million manually annotated high-resolution images. Unlike CIFAR-10 and CIFAR-100, which are commonly used in ENAS, fewer methods choose to verify their performance on ImageNet~. This can be explained by the data shown in Table~\\\\ref{table_CIFAR}, where the GPU Days are usually tens or even thousands. It is unaffordable for most researchers when the medium-scale dataset changes to the lager one.\\nHowever, Chen \\\\textit{et al.}~ believed that the results on the larger datasets like ImageNet are more convincing because CIFAR-10 is easy to be over-fitting. A popular way to deal with this is using a proxy on CIFAR-10 and transfers to ImageNet~. Another alternative approach is to use the down-scaled dataset of a large-scale dataset such as ImageNet$64\\\\times64$~.\",\n \"id\": \"8f512a41-ca51-4724-a511-f83f8ff395ee\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Challenges and Issues\"\n ],\n [\n \"subsection\",\n \"Scalability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Scalability\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8247\n ],\n \"content\": \"Section~\\\\ref{sec_evaluation} has introduced the most popular and effective ways to reduce the fitness evaluation time and the computational cost. In a nutshell, it can be described as a question that strikes the balance between the time spent and the accuracy of the evaluation. Because of the unbearable time for fully training the architecture, we must compromise as little as we can on the evaluation accuracy in exchange for significant reduction in the evaluation time without sufficient computing resources.\\nAlthough a lot of ENAS methods have adopted various means to shorten the evaluation time. Even though Sun \\\\textit{et al.}~ specifically proposed a method for acceleration, the research direction of search acceleration is just getting started. The current approaches have many limitations that need to be addressed. For example, although the LargeEvo algorithm~ used the weight inheritance to shorten the evaluation time and reduced the computational cost, it still ran for several days with the use of lots of computational resources which cannot be easily accessed by many researchers. Furthermore, there is no baseline and common assessment criteria for search acceleration methods. It is a major challenge to propose a novel method to evaluate the architecture accurately and quickly.\",\n \"id\": \"0511a9dd-9340-4ba1-ba4f-5c1c82725c9e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Challenges and Issues\"\n ],\n [\n \"subsection\",\n \"Efficient Evaluation Method and Reduce Computational Cost\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Efficient Evaluation Method and Reduce Computational Cost\"\n },\n {\n \"cite_extract_rate\": 0.28571428571428503,\n \"cites\": [\n 499,\n 6865\n ],\n \"content\": \"CNNs are known as black-box-like solutions, which are hard to interpret~. Although some works have been done to visualize the process of feature extraction~, they are uninterpretable due to a large number of learned features~. The low interpretability of the manual designed architecture becomes a big obstacle to the development of neural networks. To overcome this obstacle, some researches~ used GP to automatically design the neural network. Being well-known for its potential interpretability, GP aims at solving problems by automatically evolving computer programs~.\\nAll the above researches~ gave a further analysis to demonstrate the interpretability. Specifically, Evans \\\\textit{et al.}~ have made a visualization on the JAFFE dataset~ to expound how the evolved convolution filter served as a form of edge detection, and the large presence of white color in the output of the convolution can help the classification. In their subsequent work~, they made a visualization of the automatically evolved model on the Hands dataset~, where the aggregation function extracts the minimum value of a specific area in the hand image to determine whether the hand is open or closed. Furthermore, Bi \\\\textit{et al.}~ displayed the features described by the evolved functions like convolution, max-pooling and addition, and the generated salient features are discriminative for face classification.\\nDespite the interpretability the existing work made, all these GP-based ENAS methods only aim at shallow NNs and the number of the generated features is relatively small. However, all the most successful NNs have a deep architecture. This is due partly to the fact that very few GP based methods can be run on GPUs. It is necessary to use deep-GP to evolve a deeper GP tree and make a further analysis on the deeper architecture in the future.\",\n \"id\": \"97b2a2f7-bddb-45d3-819a-f3dcbad41cba\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Challenges and Issues\"\n ],\n [\n \"subsection\",\n \"Interpretability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Interpretability\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 6882,\n 39,\n 6876,\n 9081\n ],\n \"content\": \"Table~\\\\ref{table_application} shows various applications which have been explored by ENAS. But these are just a small part of all areas of neural network applications. ENAS can be applied wherever neural networks can be applied, and automate the process of architecture designed which should have been done by experts. Moreover, plenty of the image classification successes of ENAS have proven that ENAS has the ability to replace experts in many areas. The automated architecture design is a trend.\\nHowever, this process is not completely automated. The encoding space (search space) still needs to be designed by experts for different applications. For example, for the image processing tasks, CNNs are more suitable, so the encoding space contains the layers including convolution layers, pooling layers and fully connected layers. For the time-series data processing, RNNs are more suitable, so the encoding space may contain the cells including $\\\\Delta$-RNN cell, LSTM~, Gated Recurrent Unit (GRU)~, Minimally-Gated Unit (MGU)~, and Update-Gated RNN (UGRNN)~. The two manually determined encoding spaces already contain a great deal of artificial experience and the components without guaranteed performance are excluded. The problem is: can a method search the corresponding type of neural network for multiple tasks in a large encoding space, including all the popular widely used components? Instead of searching one multitask network~ which learns several tasks at once with the same neural network, the aim is to find appropriate networks for different tasks in one large encoding space.\",\n \"id\": \"ed6f44db-5133-4077-be4f-e9675dc00180\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Challenges and Issues\"\n ],\n [\n \"subsection\",\n \"Future Applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Future Applications\"\n },\n {\n \"cite_extract_rate\": 0.875,\n \"cites\": [\n 871,\n 6524,\n 6881,\n 8319,\n 6854,\n 870,\n 9080\n ],\n \"content\": \"Section~\\\\ref{sec_CIFAR} gives a brief introduction of the unfair comparisons. The unfairness mainly comes from two aspects: (1) the tricks including cutout~, ScheduledDropPath~, etc. (2) The different encoding spaces. For aspect (1), some ENAS methods~ have reported the results with and without the tricks. For aspect (2), the well-designed search space is widely used in different ENAS methods. For instance, the NASNet search space~ is also used in~ because it is well-constructed so that even random search can perform well. The comparison under the same condition can tell the effectiveness of different search methods.\\nFortunately, the first public benchmark dataset for NAS, the NAS-Bench-101~ has been proposed. The dataset contains 432K unique convolutional architectures based on the cell-based encoding space. Each architecture can query the corresponding metrics, including test accuracy, training time, etc., directly in the dataset without the large-scale computation. NAS-Bench-201~ was proposed recently and is based on another cell-based encoding space, which does not have no limits on edges. Compared with NAS-Bench-101, which was only tested on CIFAR-10, this dataset collects the test accuracy on three different image classification datasets (CIFAR-10, CIFAR-100, ImageNet-16-120~). But the encoding space is relatively small, and only contains 15.6K architectures. Experiments with different ENAS methods on these benchmark datasets can get a fair comparison and it will not take too much time. However, these datasets are only based on the cell-based encoding spaces and cannot contain all the search space of the existing methods, because the other basic units (layers and blocks) are built using more hyper-parameters, which may lead to a larger encoding space.\\nIn the future, a common platform making fair comparisons needs to be built. This platform must have several benchmarks encoding space, such as the NASNet search space, NAS-Bench-101 and NAS-Bench-201. All the ENAS methods can be directly tested on the platform. Furthermore, this platform also needs to solve the problem that different kinds of GPUs have different computing power, which may cause an inaccurate GPU Days based on different standards. The GPU Days cannot be compared directly until they have a common baseline of computing power.\",\n \"id\": \"14a9080c-0216-4f7c-899b-00f8965feb35\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"e3dda7bf-8a69-4f93-8f3f-07b17269af8c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Challenges and Issues\"\n ],\n [\n \"subsection\",\n \"Fair Comparisons\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Fair Comparisons\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec_conclusion}\\nThis paper provides a comprehensive survey of ENAS. We introduced ENAS from four aspects: population representation, encoding space, population updating, and fitness evaluation following the unified flow, which can be seen in Fig.~\\\\ref{fig_flowchart}. The various applications and the performance of the state-of-the-art methods on image classification are also summarized in tables to demonstrate the wide applicability and the promising ability. Challenges and issues are also discussed to identify future research direction in this field.\\nTo be specific, firstly, the encoding space is introduced by categories. We divide the encoding space into two parts: initial space and search space, where the former one defines the initial conditions whereas the latter one defines the architectures that can be found in the evolutionary process. Also, different encoding strategies and architecture representations are discussed. Secondly, the process of population updating including the evolutionary operators, the multi-objective search strategy, and the selection strategy are presented. A variety of EC paradigms use respective metaphors to generate new individuals. Based on standard algorithms, many improved methods have been proposed to obtain a stable and reliable search capability. Furthermore, we introduce the existing methods to reduce the need for large amounts of time and computing resources, which is a huge obstacle to efficiency.\\nAlthough the state-of-the-art methods have achieved some success, ENAS still faces challenges and issues. The first important issue is whether the EC-based search strategy has advantages. If the result is at the same level as the baseline (e.g., random search), it is unnecessary to design the complex evolutionary operators. A sophisticated experiment is in urgent need to tell the effectiveness, especially in a large encoding space. Secondly, the crossover operator is a multi-individual based operator and there is no sufficient explanation to how the crossover operator works well on ENAS. Besides, ENAS is just beginning a new era, so there is a lot of uncharted territory to be explored. \\nMoreover, a unified standard or platform is demanded to make a fair comparison.\\n\\\\bibliographystyle{IEEEtran}\\n\\\\bibliography{IEEEabrv,mybibfile}\\n\\\\end{document}\",\n \"id\": \"d338e930-f930-4ee2-9cea-b0e47084ac48\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"40bf4670-36c8-4014-b074-8f86f999eff7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Evolutionary Neural Architecture Search\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":79,"string":"79"},"all_cites":{"kind":"list like","value":[96,1110,166,6822,514,8441,6821,6823,8247,97,6824,7,684,872,6820,6825,9078,6826,870,6827,6851,6838,6837,6844,6841,8955,6829,6832,9079,6850,6836,6835,6843,6846,6847,9080,6834,6845,6842,6830,6848,6853,6840,6833,6852,6831,6849,6839,6828,6524,305,871,544,6854,6861,6863,6856,6857,6858,6860,6855,6864,6865,6859,6862,6869,6868,6867,6866,8302,6871,6872,6870,1461,6876,6875,6874,6873,6877,6878,8319,6880,6879,6881,499,6882,39,9081],"string":"[\n 96,\n 1110,\n 166,\n 6822,\n 514,\n 8441,\n 6821,\n 6823,\n 8247,\n 97,\n 6824,\n 7,\n 684,\n 872,\n 6820,\n 6825,\n 9078,\n 6826,\n 870,\n 6827,\n 6851,\n 6838,\n 6837,\n 6844,\n 6841,\n 8955,\n 6829,\n 6832,\n 9079,\n 6850,\n 6836,\n 6835,\n 6843,\n 6846,\n 6847,\n 9080,\n 6834,\n 6845,\n 6842,\n 6830,\n 6848,\n 6853,\n 6840,\n 6833,\n 6852,\n 6831,\n 6849,\n 6839,\n 6828,\n 6524,\n 305,\n 871,\n 544,\n 6854,\n 6861,\n 6863,\n 6856,\n 6857,\n 6858,\n 6860,\n 6855,\n 6864,\n 6865,\n 6859,\n 6862,\n 6869,\n 6868,\n 6867,\n 6866,\n 8302,\n 6871,\n 6872,\n 6870,\n 1461,\n 6876,\n 6875,\n 6874,\n 6873,\n 6877,\n 6878,\n 8319,\n 6880,\n 6879,\n 6881,\n 499,\n 6882,\n 39,\n 9081\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.154586619783284,"string":"1.154587"}}},{"rowIdx":45,"cells":{"authors":{"kind":"list like","value":["Zhifeng Jiang","Wei Wang","Bo Li","Qiang Yang"],"string":"[\n \"Zhifeng Jiang\",\n \"Wei Wang\",\n \"Bo Li\",\n \"Qiang Yang\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-09-09T02:31:29Z","string":"2021-09-09T02:31:29Z"},"category":{"kind":"string","value":"cs.DC"},"abstract":{"kind":"string","value":" The increasing demand for privacy-preserving collaborative learning has given rise to a new computing paradigm called federated learning (FL), in which clients collaboratively train a machine learning (ML) model without revealing their private training data. Given an acceptable level of privacy guarantee, the goal of FL is to minimize the \\emph{time-to-accuracy} of model training. Compared with distributed ML in data centers, there are four distinct challenges to achieving short time-to-accuracy in FL training, namely the lack of information for optimization, the tradeoff between statistical and system utility, client heterogeneity, and large configuration space. In this paper, we survey recent works in addressing these challenges and present them following a typical training workflow through three phases: client selection, configuration, and reporting. We also review system works including measurement studies and benchmarking tools that aim to support FL developers. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"419df863-69f4-4c35-93d0-018d167ed4ec","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"]],"subsections":["428f0e83-f11e-4df6-8b69-0c0beb42e5b5","498f24aa-02f6-4623-986c-cb8907d7dc9c","a91a3aeb-ed4e-41b8-921a-e2c112807c50","52489191-afca-4640-b251-831b75c284de","24e6f906-e64a-4296-a2e5-eb81d7da883e"],"title":"root"},{"cite_extract_rate":0.6000000000000001,"cites":[585,589,584,581,587,580,583,586,7262,582,7261,588],"content":"\\label{chap:introduction}}\n\\IEEEPARstart{B}{uilding} high-quality machine learning (ML) models demands a massive amount of training data. Yet, the communication cost and privacy concerns impinge on the process of collecting large volumes of data from diverse sources. It is not until recently that governments started to regulate the commercial use of data with privacy-preserving legislation (e.g., GDPR~, HIPAA~, and CCPA~). Compliance violations can be costly, with hefty fines up to hundreds of millions of dollars a year~. As such, the desire for multiple entities (e.g., mobile devices or large organizations) to collaboratively train a shared model efficiently and privately gives birth to a new ML paradigm called federated learning (FL)~. FL promises not to expose the clients' raw data, and has been widely adopted in many industries with applications ranging from mobile devices~ to financial management~ and medical care~.\nApart from providing strong privacy guarantees, the key to the success of a federated training system lies in its efficiency. A typical efficiency metric is the \\emph{time-to-accuracy}, which is the wall clock time taken to train a model until it reaches the target accuracy. Despite the rich body of work that explored various optimization strategies, there is still plenty of room for further improvement due to the following distinct challenges posed to FL (\\cref{chap:background}): (1) the \\textit{lack of information for optimization}: the information needed for optimally configuring the system is usually outdated or simply unavailable due to privacy constraints and scaling issues; (2) the \\textit{tradeoff between statistical and system utility}: statistical utility (the number of iterations taken to reach a plausible target accuracy) and system utility (the duration of an iteration), the two determining factor for time-to-accuracy, are usually at odds in FL; (3) \\textit{client heterogeneity}: clients cannot be treated uniformly due to the intrinsic differences in terms of resources, data, and states; and (4) a \\textit{large configuration space}: the operational dimensions for system developers are too many to explore within an acceptable time. Given these challenges, it is worth summarizing existing research efforts to give researchers a holistic view of the lessons learned and to solicit further explorations.\nTo position existing research attempts in optimizing the time-to-accuracy\nperformance in FL, we propose a layered approach that categorizes them by the\ntraining phases in which they take effect: selection, configuration, and\nreporting (\\cref{chap:prior}). For the \\textit{selection phase} where the\nserver chooses clients for participation, there are mainly two lines of\noptimization efforts: (1) prioritizing clients either with\nhigh statistical utility or system utility~\\cite\n{zhang2021client, nishio2019client, wang2020optimizing}, and (2) explicitly considering\nboth utilities and developing a more informed solution in response to client\ndynamics in practice~. \nAs for the \\textit{configuration phase} where the server sends the global\nmodel to the selected clients with auxiliary configuration information, and\nclients perform local training, we sort out four lines of work: (1) the first\ntwo lines advocate mitigating the communication cost by reducing the model\nsize~ and\ndecreasing the synchronization frequency~\\cite\n{hsieh2017gaia, kamp2018efficient, luping2019cmfl, chen2019communication,\nchen2021communication}; (2) the last two lines minimize the computational\noverhead by accelerating the training speed in each round~\\cite\n{anh2019efficient, nguyen2020resource, wang2020towards, li2018federated, diao2020heterofl, ren2020accelerating} and reducing the number of training\nrounds~. \nIn terms of the \\textit{reporting phase}, we focus on the aggregation and\noutline two related optimizations: (1) reducing the aggregation\nlatency by adopting hierarchical methods~\\cite\n{liu2020client, abad2020hierarchical, wu2020accelerating} and developing\nlightweight privacy-preserving methods~\\cite\n{so2021turbo, zhang2020batchcrypt, jiang2021flashe}, and (2)\nimproving the long-term convergence rate with adaptive\noptimizers on the server~. For each of the attempted optimizations, our discussion\nincludes necessary details for readers to understand the motivation,\nmechanisms, and major results. In addition, works such as measurement\nstudies~ and benchmarking tools~\\cite\n{caldas2018leaf, hu2020oarf, lai2021fedscale, fate, he2020fedml,\nbeutel2020flower, plato} are indispensable in system research. We also survey the\nstatus quo as a tutorial on FL practice (\\cref{chap:concerstone}).\nOur work focuses on the system-level efforts made in improving the time-to-accuracy performance for synchronous federated training. We also share some implications derived from the literature and our survey process. It thus differs from existing surveys which mainly focus on ML algorithms and privacy-preserving algorithms. We expect this work to be an initial attempt to bridge the gap of system-oriented surveys in FL literature, as well as soliciting more contributions to related research.","id":"428f0e83-f11e-4df6-8b69-0c0beb42e5b5","level":"section","origin_cites_number":20,"parent_id":"419df863-69f4-4c35-93d0-018d167ed4ec","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{chap:background}\nIn this section, we give a detailed introduction to the system optimization problem in federated training. We start with a quick primer on the execution workflow of federated training (\\cref{sec:background_federated}), followed by the problem statement and the scope of this survey (\\cref{sec:background_problem}). We next outline two challenges that make the problem difficult: optimality and practicality (\\cref{sec:background_challenge}), which also serves as a summary of criteria for evaluating existing solutions presented thereafter.","id":"498f24aa-02f6-4623-986c-cb8907d7dc9c","level":"section","origin_cites_number":0,"parent_id":"419df863-69f4-4c35-93d0-018d167ed4ec","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"]],"subsections":["74ace61b-35de-40e1-bc2b-a3118bd28783","e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4","a9660a24-bfd4-4549-8a91-101fb1f84585"],"title":"Background, Problem and Challenges"},{"cite_extract_rate":0.636363636363636,"cites":[590,587,7261,582,583,588,591],"content":"\\label{sec:background_federated}\nFederated learning (FL)~ has recently emerged\nas a new paradigm of collaborative machine learning (ML) that allows multiple\ndistributed clients (e.g., mobile devices or business organizations) to\ncollaboratively train or evaluate a model with decentralized data. Compared\nto traditional distributed learning in datacenter environments, FL mainly\ndiffers in orchestration, resource constraints, data distribution, and\nparticipation scale~. At its core, FL keeps private\ndata on-premises, while using a central server to maintain a global model and\niteratively refine it by aggregating each client's local updates. This design\nreduces not only the communication cost but also the privacy risk in\ngathering clients' raw data. Owing to its privacy guarantees, FL has found\nwide applications in various domains. On mobile devices, Google runs FL to\nimprove the user experience for Google Keyboard~\\cite\n{yang2018applied, hard2018federated, ramaswamy2019federated,\nchen2019federated} and Assistant~, while Apple deploys FL to\nevaluate and tune speech recognition models~; in\nfintech, both IBM~ and WeBank~ utilize\nFL to detect financial misconducts; in healthcare, NVIDIA applies FL\nto create medical imaging AI~ and predict patients' needs\nfor oxygen~.\nWhile both model training and evaluation play important roles in the development of an FL model, they have different criteria in system design. In this survey, we limit the scope to the \\textit{training} process, which is the most time-consuming and resource-intensive stage throughout the development of an FL model. Due to its predominance in practice, we focus on the support for the \\textit{synchronous} mode, wherein an ML model is trained across a pool of candidate clients in rounds, and in each round, the server needs to wait until a predefined deadline or receiving a sufficient number of clients' updates prior to deriving an aggregated update. In more detail, each round consists of the following three phases (Fig.~\\ref{fig:fl}).\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=1.0\\columnwidth]{standard-fl-protocol}\n \\caption{Standard synchronous federated training protocol~.}\n \\label{fig:fl}\n\\end{figure}\n\\begin{itemize}\n \\item \\textit{Client Selection}. At the beginning of each round, the server waits for a sufficient number of clients with eligible status (i.e., currently charging and connected to an unmetered network) to check in. The server then selects a subset of them based on certain strategies (e.g., randomly or selectively) for participation, and notifies the others to reconnect later.\n \\item \\textit{Configuration}. The server next sends the global model status and configuration profiles (e.g., the number of local epochs or the reporting deadline) to each of the selected clients. Based on the instructed configuration, the clients perform local model training independently with their private data.\n \\item \\textit{Reporting}. The server then waits for the participating clients to report local updates until reaching the predefined deadline. The current round is aborted if no enough clients report in time. Otherwise, the server aggregates the received local updates, uses the aggregate to update the global model status, and concludes the round.\n\\end{itemize}","id":"74ace61b-35de-40e1-bc2b-a3118bd28783","level":"subsection","origin_cites_number":11,"parent_id":"498f24aa-02f6-4623-986c-cb8907d7dc9c","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"],["subsection","Federated Training"]],"subsections":[],"title":"Federated Training"},{"cite_extract_rate":0,"cites":[],"content":"}\nThe primary goal of system optimization in federated training is to minimize the end-to-end resource usage of performing a task. The most common metric is the wall clock time, which is typically measured from the very beginning to a certain desirable checkpoint (e.g., convergence or reaching target accuracy). When a metered network is in use (e.g., when clients are on-demand virtual machines in a public cloud), the overall monetary cost becomes another relevant metric that deserves special attention. When uncharged devices are involved, the power consumption should also be considered. Because the cost and energy consumption generally grows linearly as time flies, in this survey, we are particularly interested in reducing the \\textit{time-to-accuracy}, i.e., the wall clock time for achieving a preset accuracy target.\n\\begin{figure*}[t]\n \\centering\n \\begin{subfigure}[b]{0.19\\linewidth}\n \\centering\n \\includegraphics[width=\\columnwidth]{time-to-acc.pdf}\n \\caption{Time to accuracy performance.}\n \\label{fig:toy_time_to_acc}\n \\end{subfigure}\n \\begin{subfigure}[b]{0.19\\linewidth}\n \\centering\n \\includegraphics[width=\\columnwidth]{./round-to-acc.pdf}\n \\caption{Round to accuracy performance.}\n \\label{fig:toy_round_to_acc}\n \\end{subfigure}\n \\begin{subfigure}[b]{0.19\\linewidth }\n \\centering\n \\includegraphics[width=\\columnwidth]{./round-time}\n \\caption{Average round time w/ standard deviation.}\n \\label{fig:toy_round_time}\n \\end{subfigure}\n \\begin{subfigure}[b]{0.19\\linewidth}\n \\centering\n \\includegraphics[width=\\columnwidth]{./dirichlet}\n \\caption{Data distributions of clients in NonIID.}\n \\label{fig:toy_noniid}\n \\end{subfigure}\n \\begin{subfigure}[b]{0.19\\linewidth}\n \\centering\n \\includegraphics[width=\\columnwidth]{./figure/comp_zipf.pdf}\n \\caption{Computational speed of clients in Straggler.}\n \\label{fig:toy_straggler}\n \\end{subfigure}\n \\caption{A case study for illustrating system utility and statistical utility with FL emulation.} \n \\label{fig:toy}\n \\end{figure*}","id":"e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4","level":"subsection","origin_cites_number":0,"parent_id":"498f24aa-02f6-4623-986c-cb8907d7dc9c","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"],["subsection","System Optimization: The Problem\\label{sec:background_problem"]],"subsections":["172d329c-9179-421a-a850-1b10255f24b6","1485b3df-12ae-4a59-a7c0-b7477f1b132f"],"title":"System Optimization: The Problem\\label{sec:background_problem"},{"cite_extract_rate":0.33333333333333304,"cites":[580,582],"content":"}\nIntuitively, the time-to-accuracy performance of federated training is determined by two factors: the number of rounds taken to reach the accuracy and the average duration of training rounds. As in , we regard the former as \\textit{statistical utility} and the latter as \\textit{system utility} throughout this survey.\nTo illustrate, we conduct an FL case study where 16 clients collaboratively train the LeNet5~ model to classify the images from the MNIST~ dataset. To emulate the cross-device environment, each client is run atop an AWC EC2 \\texttt{c5.xlarge} instance (4 vCPUs and 8 GB memory) and the network bandwidth is throttled to 21 Mbps to dictate the average mobile connection speed as of 2021~. We run FedAvg~ for model aggregation without loss of generality, atop which we enforce clients' privacy by implementing the SecAgg protocol~ (also see~\\cref{sec:prior_configuration_latency}) which makes the server blind to each client's local updates. Essentially, we design three schemes to compare:\n\\begin{itemize}\n \\item \\textit{Baseline}: all clients' data is independent and identically distributed (IID). They also share the same computational speed.\n \\item \\textit{NonIID}: the same as \\textit{Baseline} except that clients' data is non-IID as depicted by Fig.~\\ref{fig:toy_noniid}. This is achieved by latent Dirichlet allocation (described in~\\cref{sec:cornerstone_benchmarking_datasets}) with the concentration vectors being all 0.1's.\n \\item \\textit{Straggler}: the same as \\textit{Baseline} except that clients’ computation speeds follow the Zipf's distribution (with $\\alpha=1.2$, i.e., moderately skewed). As shown in Fig.~\\ref{fig:toy_straggler}, the straggler is 5$\\times$ slower than the fastest who shares the same speed as clients in \\textit{Baseline}.\n\\end{itemize}\nWe first study the impact of statistical utility by comparing \\textit{Baseline} and \\textit{NonIID}. As indicated in Fig.~\\ref{fig:toy_round_time}, they proceed a round at the same speed. However, as the learning curves in Fig.~\\ref{fig:toy_time_to_acc} dictate, \\textit{Baseline} reaches the target accuracy (96.1\\%) with only 8.4 min, while \\textit{NonIID} takes 43.5 min, which is 5$\\times$ slower than \\textit{Baseline} does. Referring to the round-to-accuracy performance as in~Fig.~\\ref{fig:toy_round_to_acc}, one can know that the reason for this contrast lies in the difference in the numbers of rounds taken to convergence: 10 for \\textit{Baseline} while 40 for \\textit{NonIID}. Hence, despite having the same system utility, \\textit{Baseline} achieves much better time-to-accuracy performance for having greater statistical utility due to more evenly distributed data across clients. Of course, data distribution is uncontrollable in FL practice and developers can barely end up with IID cases as in \\textit{Baseline}. This adds up to the challenges of system-level optimizations in FL training, as later discussed in~\\cref{sec:background_challenge_optimality}.\nNext, we provide a sense of how system utility affects the end-to-end performance by comparing \\textit{Baseline} with \\textit{Straggler}. As shown in Fig.~\\ref{fig:toy_round_to_acc}, both cases converge with the same amount of rounds: 10. However, it takes \\textit{Straggler} slightly longer in time (15.8 min, i.e., 1.9$\\times$) to reach the target accuracy, as demonstrated in Fig.~\\ref{fig:toy_time_to_acc}. By observing Fig.~\\ref{fig:toy_round_to_acc} and Fig.~\\ref{fig:toy_round_time}, we know that the source of the difference does not stem from the number of rounds used but the making span of each round. As all clients have to proceed at the same speed as the slowest clients in synchronous training, \\textit{Straggler} features lower system utility than \\textit{Baseline} does due to the presence of slow clients, rendering worse time-to-accuracy. Similarly, the disparity of clients' capabilities is inevitable in FL practice, which also complicates the design of system-level optimizations as mentioned in~\\cref{sec:background_challenge_optimality}.","id":"172d329c-9179-421a-a850-1b10255f24b6","level":"subsubsection","origin_cites_number":6,"parent_id":"e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"],["subsection","System Optimization: The Problem\\label{sec:background_problem"],["subsubsection","Statistical Utility and Sytem Utility: A Case Study~\\label{sec:background_problem_utility"]],"subsections":[],"title":"Statistical Utility and Sytem Utility: A Case Study~\\label{sec:background_problem_utility"},{"cite_extract_rate":0.625,"cites":[593,595,596,600,598,599,594,592,7038,597],"content":"}\nWe emphasize that the solutions discussed in this survey operate at the \\textit{system-level}. Therefore, the following directions of work are excluded from discussion, though they could effectively improve the statistical and/or system efficiency in synchronous federated training:\n\\begin{itemize}\n \\item Hardware updates, e.g., adapting programmable switches to enjoy the communication efficiency brought by in-network aggregation~.\n \\item Security mechanisms, e.g., employing robust aggregation methods to protect the statistical utility from being impaired by model poisoning attacks~.\n \\item Paradigm innovations, e.g., (1) letting clients fit different sets of model parameters by personalization~ or models of heterogeneous architectures through knowledge distillation~ to tackle data heterogeneity (a concept mentioned in~\\cref{sec:background_challenge_optimality}), (2) allowing clients to exchange data representations for realizing prototype learning~ to reduce communication burdens, or (3) permitting clients to send encoded versions of local datasets to the server to reduce the computational complexity~.\n \\item Optimizations on upstream parts of the FL pipeline, e.g., searching for neural architectures that yield better predictive accuracy~.\n\\end{itemize}","id":"1485b3df-12ae-4a59-a7c0-b7477f1b132f","level":"subsubsection","origin_cites_number":16,"parent_id":"e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"],["subsection","System Optimization: The Problem\\label{sec:background_problem"],["subsubsection","What is beyond the Scope~\\label{sec:background_problem_beyond"]],"subsections":[],"title":"What is beyond the Scope~\\label{sec:background_problem_beyond"},{"cite_extract_rate":0,"cites":[],"content":"}\nDespite the clear objective, it is non-trivial to work out a feasible solution due to the following two challenges.","id":"a9660a24-bfd4-4549-8a91-101fb1f84585","level":"subsection","origin_cites_number":0,"parent_id":"498f24aa-02f6-4623-986c-cb8907d7dc9c","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"],["subsection","What Makes It Hard: The Challenges\\label{sec:background_challenge"]],"subsections":["e9339976-85db-45bd-8a4a-f9969c5967b9","b7818ca2-24b7-43f4-a5b2-dddedbcf0e1e"],"title":"What Makes It Hard: The Challenges\\label{sec:background_challenge"},{"cite_extract_rate":0.92,"cites":[3469,618,617,608,591,605,602,603,606,613,616,612,580,607,610,614,604,583,611,615,601,4338,609],"content":"}\nFirst, the information needed in decision-making may be \\textit{outdated or even unavailable}. For example, to estimate a client's system utility, it is common to refer to its most recent response latency~. However, due to the dynamics over time, such information may not accurately reflect the client's current status. There also exists a cold-start issue, where we are unaware of a client's system capabilities until its first participation. As for estimating a client's statistical utility, the amount of available information is further limited by privacy concerns. According to the recent FL literature, exploratory attacks such as property inference~, membership inference~, and data reconstruction~ can be made possible with model updates. As such, even exposing model updates can discourage clients from participation, let alone inquiring about their data distributions or even raw data~. Note that the uncertainty in clients' statistical utility and system utility can be accumulated over time.\nEven given a holistic view of the environment, the problem remains hard due to the \\textit{coupled nature} of statistical utility and system utility. Intuitively, improving system utility is equivalent to minimizing the average resource consumption (e.g., time or bandwidth) per task unit. On the other hand, reducing the resources invested in a task unit inevitably downgrades the quality of the outcome (e.g., statistical utility) as long as no resource is redundant. To exemplify, by constantly picking the fastest clients in client selection, the average duration of each round indeed decreases, whereas the number of rounds taken to target accuracy may be increased as well when other clients' data are under-represented in the global model. Another example can be found in model compression. To improve communication efficiency, a client can send only an important subset of model updates by sparsification~, or a low-bit representation of them by quantization~. Although the per-round communication duration can be significantly reduced by adopting a higher compression ratio, the convergence has to take more rounds to occur due to the loss of arithmetic precision.\nThe problem is further complicated by \\textit{client heterogeneity.} Federated training involves tens to potentially millions of clients, each of which intrinsically differs from one another in the following three aspects: \n\\begin{itemize}\n \\item \\textit{Resource Heterogeneity}: due to the variability in hardware specifications and system-level constraints, clients in federated training typically possess different capabilities in computation (CPU/GPU/NPU, memory, and storage), communication (connectivity and bandwidth) and power (battery level and lifespan)~. These types of heterogeneity complicate the optimization of the overall system utility. For example, merely improving the communication speed does not necessarily lead to shorter end-to-end latency, especially when the straggler is bottlenecked by the computation~.\n \\item \\textit{Data Heterogeneity}: as the training datasets of clients are typically generated based on their local activities and contexts, they are not IID. More specifically, clients' datasets mainly differ in two aspects\\footnote{See a more complete categorization of non-IID scenarios in \\cref{sec:cornerstone_benchmarking_datasets}.}: (1) sample quantity (i.e., the number of data samples), and (2) label partition (i.e., the distribution of data labels)~. As a result, not all of them are representative of the population distribution. In case we do not include all the clients in the federation, optimization for statistical utility has to additionally account for such heterogeneity.\n \\item \\textit{State Heterogeneity}: as observed from real-world traces~, the available slots of mobile device clients vary significantly in temporal distribution due to different user behaviors (e.g., screen locking or battery charging). Therefore, in each round, there can be different sets of candidate clients to choose from, as well as different client drop-out outcomes. On top of the non-IID distribution of clients' data, this type of heterogeneity further complicates statistical utility optimization. Nevertheless, in the cross-silo settings, it may be less of a concern due to the stable and dedicated nature of clients' computing power~.\n\\end{itemize}\nLast but not least, it is \\textit{infeasible to search through the entire configuration space} for the global optimum. On the one hand, the space is prohibitively large, as a federated training task typically spans $10^1$--$10^6$ users and $10^2$--$10^4$ rounds~, wherein each phase of a round (\\cref{sec:background_federated}) has multiple configurable hyperparameters and alternative policies (e.g., client selection choices in the selection phase, or the number of local steps in configuration phase). On the other hand, most of the online decisions are made on the critical path of the task, meaning that the time spent on working out a solution also counts towards the end-to-end runtime performance, the very objective of the optimization. As a result, it is desirable to be guided by efficient and effective heuristic algorithms, especially balancing the exploration and exploitation efforts made in the configuration space.","id":"e9339976-85db-45bd-8a4a-f9969c5967b9","level":"subsubsection","origin_cites_number":25,"parent_id":"a9660a24-bfd4-4549-8a91-101fb1f84585","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"],["subsection","What Makes It Hard: The Challenges\\label{sec:background_challenge"],["subsubsection","On the Optimality of a Solution\\label{sec:background_challenge_optimality"]],"subsections":[],"title":"On the Optimality of a Solution\\label{sec:background_challenge_optimality"},{"cite_extract_rate":1,"cites":[580,591],"content":"}\n\\begin{figure*}[t]\n \\centering\n \\includegraphics[width=0.9\\textwidth]{skeleton}\n \\caption{Taxonomy of the approaches discussed in \\cref{chap:prior}.}\n \\label{fig:opt}\n\\end{figure*}\nApart from navigating the performance-accuracy-privacy trade-off, the design process of a practical optimization solution should mitigate the accompanying side-effects on other aspects such as the loss of \\textit{robustness} to attacks and failures~. For example, to evaluate the statistical utility of a client, the server may require it to report the loss values generated in local training~. However, a malicious or free-rider client may intentionally respond with arbitrary values in the hope of messing with the orchestration or reaping the benefits of the federation without making solid contributions. As such backdoors are introduced by the optimization solution, the developers should take charge of eliminating the undesirable exploitations of these security loopholes. Other possible concerns that may arise as a result of a system optimization solution include but are not limited to \\textit{fairness} (e.g., whether participant bias is introduced in the solution), \\textit{generality} (e.g., whether the solution applies to diverse tasks), and \\textit{ease of deployment} (e.g., whether the solution can be implemented with moderate engineering efforts). In other words, a mature system optimization solution should not only improve the time-to-accuracy performance by enhancing statistical and system utility but also minimize the adverse impacts on other aspects that federated training also values in practice.","id":"b7818ca2-24b7-43f4-a5b2-dddedbcf0e1e","level":"subsubsection","origin_cites_number":2,"parent_id":"a9660a24-bfd4-4549-8a91-101fb1f84585","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Background, Problem and Challenges"],["subsection","What Makes It Hard: The Challenges\\label{sec:background_challenge"],["subsubsection","On the Practicality of a Solution\\label{sec:background_challenge_practicality"]],"subsections":[],"title":"On the Practicality of a Solution\\label{sec:background_challenge_practicality"},{"cite_extract_rate":0,"cites":[],"content":"}\nIn the past few years, considerable research efforts have been put into tackling the above challenges for fully unleashing the performance potential of FL training. In this section, we organize them by the training phases, i.e. selection (\\cref{sec:prior_selection}), configuration (\\cref{sec:prior_configuration}), and reporting (\\cref{sec:prior_reporting}), as visualized in Fig.~\\ref{fig:opt}.","id":"a91a3aeb-ed4e-41b8-921a-e2c112807c50","level":"section","origin_cites_number":0,"parent_id":"419df863-69f4-4c35-93d0-018d167ed4ec","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"]],"subsections":["6750018e-2b8c-43be-8977-a671addc47f2","ded2658f-08bb-4199-8696-66ba4578329c","d6884228-aa74-48a6-9eb1-ee9a7ae15668"],"title":"Recent Optimization Approaches\\label{chap:prior"},{"cite_extract_rate":1,"cites":[590],"content":"}\nDue to the (potentially) large population size and the heterogeneity across clients, the effectiveness of the used participant selection algorithm plays a critical role in the time-to-accuracy performance in federated training. However, the state-of-the-practice system still relies on randomly picking participants~, which inevitably leads to waste of resources and suboptimal convergence speed. In response, there is an array of work to guide the selection, which can be roughly categorized by the target utilities that they improve.","id":"6750018e-2b8c-43be-8977-a671addc47f2","level":"subsection","origin_cites_number":1,"parent_id":"a91a3aeb-ed4e-41b8-921a-e2c112807c50","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Selection Phase\\label{sec:prior_selection"]],"subsections":["2c47120e-34e5-44cd-b940-4eb997711491","c9c46878-34a7-4fcb-8626-cb64857d9ade"],"title":"Optimizing the Selection Phase\\label{sec:prior_selection"},{"cite_extract_rate":0.5,"cites":[619,620],"content":"}\nThis line of work does not consider the interplay between statistical utility and system utility. Instead, they mainly focus on lifting either utility while leaving the other one ignored or controlled to a limited extent.\n\\PHM{Statistics-Oriented.} To approach the convergence rate in centralized settings where the data is IID, CSFedAvg~ advocates that clients with a lower degree of non-IID data should participate more often. To this end, the authors propose weight divergence to capture the non-IID degree of data owned by a client. More precisely, it measures the normalized Euclidean distance between a client's model and the reference model trained by the server with auxiliary IID data. According to the 500-client simulation over CIFAR-10 and Fashion MNIST, CSFedAvg reduces the time-to-accuracy by up to 4.0$\\times$ and 2.7$\\times$, respectively, compared to random selection.\nBesides heuristic methods, some researchers use reinforcement learning (RL) algorithms to learn which clients to select in the presence of data heterogeneity. For example, FAVOR~ seeks to reduce the number of rounds to reach a target accuracy with a deep Q-learning network (DQN)~. To capture each client's statistical characteristics, it takes the low-dimension representations of local models as the RL states. Compared to random selection, FAVOR can reduce the communication rounds by up to 49\\% in three image classification tasks. On the other hand, the training overhead for the RL agent may become an obstacle to FAVOR's real-world applications. Specifically, it is reported to take more than 100 episodes to train an agent suited for a specific learning task, which could be prohibitive as each episode corresponds to an entire FL process, i.e., training a global model from scratch for reaching a target accuracy.\n\\PHM{System-Oriented.} In synchronous training, clients with the lowest system utility bottleneck the speed of a federation round. A straightforward way to bound the time usage is setting a deadline for randomly selected clients' to report updates and ignoring any update submitted after the deadline. To avoid waste of computing resources, FedCS~ takes a step further by proactively selecting a set of clients whose participation is not likely to miss the deadline according to latency estimation results. As there can be multiple eligible sets, FedCS further favors the solution with the largest scale of participation, which reduces part of the loss in statistical utility. Technically, the whole problem is formalized as a complex combinatorial optimization, and the authors resort to a greedy algorithm for efficient approximation. As indicated in their 1000-client simulation, FedCS outperforms FedLim (modified FedAvg with per-round deadlines imposed) by up to 1.2$\\times$ and 1.8$\\times$ in the time-to-accuracy when training over the non-IID CIFAR-10 and Fashion MNIST datasets, respectively.","id":"2c47120e-34e5-44cd-b940-4eb997711491","level":"subsubsection","origin_cites_number":4,"parent_id":"6750018e-2b8c-43be-8977-a671addc47f2","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Selection Phase\\label{sec:prior_selection"],["subsubsection","Partial Optimization Attempts\\label{sec:prior_selection_either"]],"subsections":[],"title":"Partial Optimization Attempts\\label{sec:prior_selection_either"},{"cite_extract_rate":0.6000000000000001,"cites":[7262,580,585],"content":"}\nGiven the coupled nature of clients' system utility and statistical utility, it is more desirable to navigate the sweet point of jointly maximizing both of them.\n\\PHM{Coarse-Grained.} TiFL~ first considers increasing the system utility. To that end, it divides clients into different tiers based on the observed runtime performance, and at each round only selects clients from the same tier for mitigating the waste of resources due to idle waiting for stragglers. To reduce the average iteration span, it also limits the number of times a (slow) tier can be selected. On top of that, the statistical utility is respected by prioritizing tiers with lower testing accuracy whenever there is more than one electable tier. Compared with FedCS, TiFL bears some resemblance in limiting the participation of less capable clients, while being more aware of the statistical utility. As reported in a 50-client cluster with 5 client tiers, TiFL achieves an improvement over random selection by up to 3$\\times$ speedup in overall training time and by 6\\% in accuracy.\n\\PHM{Fine-Grained.} Compared to TiFL, Oort~ reconciles the demand for enhancing both system utility and statistical utility in finer granularity. Specifically, it associates each client with a continuous score and prioritizes those clients with higher scores. The score is meant to be a principled measurement of both the statistical utility (determined by the training loss) and the system utility (estimated from historical response latency). As some components of the score cannot be known in advance until the corresponding client's first participation, or cannot be guaranteed to be stable due to the client's runtime dynamics, the score estimation process is modeled as a Multi-Armed Bandit (MAB) problem and tackled by the Upper Confidence Bound (UCB) algorithm~. Apart from the scoring backbone, Oort also aims to address other practical issues like staleness and robustness. Oort was evaluated on a 1300-client GPU cluster with realistic datasets and simulation on the client heterogeneity. Compared to random selection, it reduces the training time by up to 14$\\times$ and improves model accuracy by up to 9.8\\%.\nBesides the UCB algorithm, researchers also experiment with more sophisticated RL methods to cherrypick participants. Notably, AutoFL~ learns to select participants and execution targets for each individual based on the Q-Learning algorithm~. To achieve efficient FL execution, it identifies RL states that are critical to energy efficiency, convergence time, and accuracy. It also defines RL rewards that track the energy consumption of clients and model accuracy. Compared to random selection, AutoFL is reported to improve the average FL energy efficiency by up to 4.3$\\times$, while also exhibiting better training accuracy. Similar to FAVOR (mentioned in \\cref{sec:prior_selection_either}), however, training the Q-Learning model to converge needs multiple episodes, each of which corresponds to an entire FL training process. Its practicality could thus be challenged whenever offline training is infeasible or costly.","id":"c9c46878-34a7-4fcb-8626-cb64857d9ade","level":"subsubsection","origin_cites_number":5,"parent_id":"6750018e-2b8c-43be-8977-a671addc47f2","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Selection Phase\\label{sec:prior_selection"],["subsubsection","Co-Optimizing Statistical/System Utility\\label{sec:prior_selection_both"]],"subsections":[],"title":"Co-Optimizing Statistical/System Utility\\label{sec:prior_selection_both"},{"cite_extract_rate":0,"cites":[],"content":"}\nIn the configuration phase, there are mainly two processes that are responsible for the time-to-accuracy performance. One is downlink (i.e., server-to-client) model transmission and the other is local model training. Thus, both aspects can be investigated for system optimization. As for communication overhead reduction, one can reduce the size of model updates (\\cref{sec:prior_configuration_size}) and decrease the synchronization frequency (\\cref{sec:prior_configuration_freq}). To lower computational overhead, one can shorten the training latency by balancing the workload across clients (\\cref{sec:prior_configuration_latency}), as well as reducing the number of rounds taken to converge by adopting heterogeneity-aware training algorithms (\\cref{sec:prior_configuration_round}). As the uplink model submission (i.e., client-to-server) that takes place in the reporting phase shares the same operational space as the downlink one, we combine the discussion on both of them in this section for brevity.","id":"ded2658f-08bb-4199-8696-66ba4578329c","level":"subsection","origin_cites_number":0,"parent_id":"a91a3aeb-ed4e-41b8-921a-e2c112807c50","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Configuration Phase\\label{sec:prior_configuration"]],"subsections":["0b02d7a9-809d-41bd-95d3-97624f811da6","77437726-b7b4-48cb-ac0d-7210459652b2","75004bfc-b89a-4773-8188-0896026459ce","c1657bca-2a2d-4dc4-a8d5-1e81d9a7c3f6"],"title":"Optimizing the Configuration Phase\\label{sec:prior_configuration"},{"cite_extract_rate":0.7307692307692301,"cites":[618,617,626,628,624,622,606,613,616,621,625,610,611,8366,627,615,601,623,4338],"content":"}\nPrior arts of model update size reduction mainly fall into three camps: quantization, sketching, and sparsification. Ahead of the emergence of FL, the exploration of these directions has already been initiated in the context of traditional distributed learning. While their communication merits are mostly reproducible in FL, they also face new challenges due to the privacy regulations and client heterogeneity, which we will also point out hereafter.\n\\PHM{Quantization.} Quantization converts each scalar in a model update to its low-bit representation which takes up less space. While quantization has already gained its fame in traditional distributed learning and we refer the readers to dedicated surveys like~ for more details, here we only introduce the most representative work. As the first quantization work in model training with rigorous convergence proof, QSGD~ performs unbiased quantization with standard random dithering, a technique borrowed from image processing. Since its birth, related works have been emerging with more aggressive quantization bit-widths and more appealing empirical performance. For example, TernGrad~ advocates using only ternary values (0, $\\pm$1) in the uplink direction, while signSGD~ can use only binary signs ($\\pm$) in both uplink and downlink communication. It is worth mentioning that a popular technique in tackling the precision loss brought by quantization is error feedback, whose basic idea is to accumulate the previous quantization errors and compensate for them in the current round. Leveraging this technique, ECQ-SGD~ performs consistently better than QSD in terms of both convergence speed and accuracy, while EF-SGD~ also achieves a narrower generalization gap from centralized training compared to signSGD.\nDespite their generality, there are some practical concerns about applying these general quantization strategies to FL due to privacy constraints and client heterogeneity. For example, determining the clipping threshold for quantization needs to exploit the knowledge about its inputs (i.e., local model updates) for reducing the induced error as in dACIQ~. However, an FL client does neither possess a priori knowledge of others' model updates nor should it request these sensitive values. To work out a globally applicable clipping threshold, we may need to share some less sensitive information (e.g., the maximum and minimum values in local updates) across clients for threshold estimation as in BatchCrypt~. Still, whether such a circumvention guarantees robust estimation and immunity to privacy attacks remains an open question.\n\\PHM{Sketching.} Existing quantization approaches assume the input values follow a certain distribution (e.g., uniform or bell-shaped), which may not always be the case in model updates~. To be more general, some researchers introduce sketching methods where some memory-saving data structures are used to approximate the exact distribution of model update values in a single processing pass over the values. For example, SketchML~ utilizes a quantile sketch method to generate a non-uniform mapping from gradient values to low-bit integers. SketchML achieves empirical success such as decreasing the gradient size by around 7$\\times$ and is the first effort to introduce sketching for compressing model updates in ML training. Similar to quantization, sketch algorithms can also make use of error feedback techniques to efficiently amend the errors induced by the approximation, as demonstrated in SketchedSGD~ and FetchSGD~. There is also sketching practice that compresses auxiliary variables apart from model updates, such as sketching clients' momenta and per-coordinate learning rates as in~.\n\\PHM{Sparsification.} While quantization and sketching operate at the precision level in terms of model size reduction, sparsification operates at the coordinate level. Specifically, sparsification allows each client to transmit only a sparse subset of its model updates, while the rest are accumulated and incorporated into future training. Technically, the sparsified gradient is obtained by first performing element-wise multiplication on the original gradient with a certain 0/1 mask and then discarding zero elements. The mask is typically randomly generated as in~, while another commonly used variant is the top $s$\\% scheme where 1 is given to the coordinates that rank top $s$\\% in absolute magnitude and 0 otherwise~. The top $s$\\% method is reported to reduce network traffic by up to three orders of magnitude, while still preserving model quality~.\nWhile similar cost savings are transferable to plaintext FL, it is unclear whether sparsification can be further compatible with cryptographic techniques that are widely adopted for privacy enforcement in FL. For example, apart from the uplink model updates, it is also desirable to sparsify the downlink global update for fully releasing the potential for communication improvement. However, implementing the downlink sparsification may not be feasible when the server is not aware of the plaintext values of the aggregated update as a result of the applied Secure Multi-Party Computation (SMPC)~ or Homomorphic Encryption (HE)~ techniques.\nIt is noteworthy that as quantization (or sketching) and sparsification are orthogonal to each other, they can be combined to reap the most benefits in terms of model size reduction~.","id":"0b02d7a9-809d-41bd-95d3-97624f811da6","level":"subsubsection","origin_cites_number":26,"parent_id":"ded2658f-08bb-4199-8696-66ba4578329c","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Configuration Phase\\label{sec:prior_configuration"],["subsubsection","Model Update Size Reduction\\label{sec:prior_configuration_size"]],"subsections":[],"title":"Model Update Size Reduction\\label{sec:prior_configuration_size"},{"cite_extract_rate":0.33333333333333304,"cites":[630,629],"content":"}\nAt its core, the reduction in the synchronization frequency is achieved by identifying and precluding redundant synchronization efforts. This can be operated at different granularities ranging from clients, layers to individual parameters in a model update.\n\\PHM{Client-Level.} The importance of an entire model update is usually measured by some numerical features. In the most intuitive form, a model update in Gaia~ is considered significant if its magnitude relative to the current value $\\mid \\frac{Update}{Value} \\mid$ exceeds a specific threshold such as 1\\%. While the magnitude may serve as a good indicator for how data center learning performs, it does not work in FL where determining an appropriate threshold is hard due to clients' heterogeneity. As such, some researchers propose to involve the comparison with some reference points for more robust measurement of the importance. For example,~ tracks the Euclidean distance between the local model and a reference model, while CMFL~ focuses on the number of coordinates with the same sign in the local model and the most recent global model.\n\\PHM{Layer-Level.} Apart from considering a model update as a whole, another line of work tries to reduce the synchronization frequency on a layer basis. A representative work done in this direction is TWAFL~ where the model aggregation is conducted layer-wise. As observations made in deep neural network (DNN) fine-tuning~, shallow layers in a DNN learn general features across different datasets while deep layers learn ad hoc ones. TWAFL hence proposes to update shallow layers more frequently than deep ones as they are more responsible for the overall quality of the global model.\n\\PHM{Parameter-Level.} Some industrial practitioners also consider whether to synchronize for each round at the level of individual parameters. Noticing that each parameter usually evolves in a transient-then-stable manner, i.e., it first varies drastically and then settles down around a certain value with slight oscillation, APF~ proposes to stop synchronizing those parameters whose evolution has reached a stationary phase.","id":"77437726-b7b4-48cb-ac0d-7210459652b2","level":"subsubsection","origin_cites_number":6,"parent_id":"ded2658f-08bb-4199-8696-66ba4578329c","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Configuration Phase\\label{sec:prior_configuration"],["subsubsection","Synchronization Frequency Reduction\\label{sec:prior_configuration_freq"]],"subsections":[],"title":"Synchronization Frequency Reduction\\label{sec:prior_configuration_freq"},{"cite_extract_rate":0.33333333333333304,"cites":[631,632],"content":"}\nA client's training latency is determined by both its computational workload and resource capabilities. While the latter cannot be altered, the former still leaves room for optimization innovations. We discuss one major line of such efforts.\n\\PHM{Load Balancing.} Given the variations in computing power and data volume, clients may not finish the training process at the same time. To mitigate the resulting straggler effects, suggest balancing the amount of training data across clients. Specifically, they turn to RL techniques for determining the optimal number of data units used in a communication round for each participant, intending to minimize the time and energy consumption and maximize the volume of involved data. While these solutions can reduce the round latency, the number of rounds needed for convergence does not necessarily remain unchanged because they do not account for data heterogeneity and thus the contribution of slow clients with important data could be throttled. To jointly consider heterogeneous system speed and data distribution during load balancing, carefully formulates an optimization problem where each client is associated with a weight that diminishes exponentially with the disparity between its number of labels and that of the population. Compared to even data allocation, this work manages to reduce the end-to-end time-to-accuracy.\nLoad balancing can also be achieved by varying the number of optimization steps or the complexity of local models. For example, FedProx~ balances the system load across clients by formulating an inexact learning problem that allows variable steps of local solvers. On the other hand, HeteroFL~ assigns sub-models with different widths of hidden channels to clients so that clients with fewer capabilities can train smaller sub-models. All sub-models share the same model architecture, and thus normal model aggregation is still possible. The authors empirically show that the quality of the global model trained with heterogeneous sub-models is comparable to that trained with full-sized local models.\nApart from the computational load, it is sometimes also beneficial to balance the communication load across clients, especially when the network conditions are complicated as in wireless connections. For example, targeting a mobile edge computing (MEC) scenario where Time Division Multiple Access (TDMA) is implemented, optimizes both the data batch size and uplink/downlink frame time slots for each client to achieve the maximum learning efficiency. In addition to coping with CPU computing, the authors further extend the optimization problem to the scenario where devices are equipped with GPUs for training.","id":"75004bfc-b89a-4773-8188-0896026459ce","level":"subsubsection","origin_cites_number":6,"parent_id":"ded2658f-08bb-4199-8696-66ba4578329c","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Configuration Phase\\label{sec:prior_configuration"],["subsubsection","Training Latency Reduction\\label{sec:prior_configuration_latency"]],"subsections":[],"title":"Training Latency Reduction\\label{sec:prior_configuration_latency"},{"cite_extract_rate":0.785714285714285,"cites":[634,638,635,640,636,639,641,633,582,632,637],"content":"}\nIn the settings with heterogeneous data, more local computation in a communication round does not necessarily lead to fewer numbers of rounds for reaching satisfying optima~. Thus, adopting heterogeneity-aware techniques such as adaptive optimization and bias reduction can help guarantee the convergence speed in FL practice.\n\\PHM{Optimizer State Synchronization.} It is common practice for first-order moment optimization to apply momentum to dampen oscillations~. However, in the federated settings, if the clients' momenta are separately updated, they may deviate from each other due to non-IID data distributions. Thus, there are researchers proposing optimizer state synchronization frameworks where clients' optimizer states are synchronized by the server periodically. PR-SGD-Momentum~ is aligned with this direction and also gives proof of the linear speedup of convergence w.r.t. the number of workers. FedAC~ also applies momentum at clients with periodic synchronization, while it is proven to obtain the same linear speedup property with asymptotically fewer rounds of synchronization. MFL~ is another similar idea with theoretical guarantees, but it focuses on accelerating deterministic gradient descent (DGD) instead of SGD, unlike the previous two studies.\n\\PHM{Client Bias Reduction.} Due to data heterogeneity, clients' model updates can be biased towards the minima of local objectives, known as ``client drift'' in the literature~, which hinders the convergence of the global model. To reduce the variance across clients, a natural idea is to regularize local objective functions for minimizing the drift. For example, assuming bounded dissimilarity between local functions, FedProx~ limits the Euclidean distance between local models and the global one by regularizing the square of the distance. FedDANE~ uses the same regularization term, while it formulates the other part of the objective function following the Distributed Approximate NEwton (DANE) method~ in classic distributed optimization. Despite having a similar theoretical convergence rate as FedProx, FedDANE underperforms FedProx in practice where the data heterogeneity is high and the participation rate is low, suggesting a discrepancy between theory and practice which needs further investigation. The objective function in FedDyn~ also considers the combination of a linear term and an L2 regularizer, with a different linear term which is formulated to align the empirical loss surfaces of clients. In theory, FedDyn ensures that the consensus point of model convergence across clients will be consistent with the global stationary solution as long as local models converge, regardless of the degree of data heterogeneity.\nApart from regularizing local objectives, variants in clients' local updates can also be reduced by leveraging the idea of control variates borrowed from the convex optimization literature. For example, in SCAFFOLD~, each of the clients and the server maintains a control variate, and at each local step, a client de-biases its local updates with two control variates: one of its own and the other broadcast by the server. SCAFFOLD converges provably faster than FedAvg~ without any assumption made on the client selection or data heterogeneity. MIME~ considers a similar idea but makes a different choice on the specific definition of control variates.\nWhile the use of control variates requires persistent client states, there exists another line of work that works for stateless clients: posterior averaging. Instead of approaching FL as optimization, this line of work formulates the problem as a posterior inference one. Compared to traditional federated optimization, posterior inference can benefit from an increased amount of local computation without risking stagnating at inferior optima. FedPA~ instantiates this idea with an efficient algorithm to conduct federated posterior inference with linear computation and communication costs.","id":"c1657bca-2a2d-4dc4-a8d5-1e81d9a7c3f6","level":"subsubsection","origin_cites_number":14,"parent_id":"ded2658f-08bb-4199-8696-66ba4578329c","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Configuration Phase\\label{sec:prior_configuration"],["subsubsection","Training Round Reduction\\label{sec:prior_configuration_round"]],"subsections":[],"title":"Training Round Reduction\\label{sec:prior_configuration_round"},{"cite_extract_rate":0,"cites":[],"content":"}\nIn this phase, the operation room for system optimization is limited to either model uploading or model aggregation. As the former has already been combined in the last section, we hereafter focus on optimizing the aggregation process with two main directions explored in the literature: (1) directly reducing the aggregation latency at each round (\\cref{sec:prior_reporting_round}), and (2) expediting the convergence rate in the long run through conducting adaptive aggregation (\\cref{sec:prior_reporting_end}).","id":"d6884228-aa74-48a6-9eb1-ee9a7ae15668","level":"subsection","origin_cites_number":0,"parent_id":"a91a3aeb-ed4e-41b8-921a-e2c112807c50","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Reporting Phase\\label{sec:prior_reporting"]],"subsections":["542ef1b6-edee-4a8a-bd3d-68e0c8f5c762","cc07a19d-bfd0-4b7a-9057-e912fa55b00d"],"title":"Optimizing the Reporting Phase\\label{sec:prior_reporting"},{"cite_extract_rate":0.533333333333333,"cites":[643,646,644,621,645,647,7039,642],"content":"}\nCompared to local training in the configuration phase, model aggregation involves less intensive computation. However, its latency can still be salient because (1) large-scale participation can put pressure on the communication, and (2) the deployment of security methods can complicate the computation. We hereafter introduce the respective optimization efforts in the literature.\n\\PHM{Hierarchical Aggregation.} The downsides of the traditional two-layer (server-client) system include (1) instability: the network bandwidth may be slow or unpredictable, especially in public networks and/or under geo-distributed settings; (2) risk of scalability: the cloud server may suffer from network congestion when concurrently receiving too many local updates; (3) heterogeneity: the straggler effects could be exacerbated by imbalanced network bandwidth. \nTo address these issues, some researchers resort to a hierarchical design of model aggregation by introducing an extra level of edge servers, each of which is typically responsible for a small number of clients with proximity. For instance, in HierFAVG~, after a fixed number of local updates on clients, each edge server aggregates its own clients' models. Subsequently, after another fixed interval of edge aggregation, the cloud server aggregates all the edge servers' models. It is proven that HierFAVG still guarantees convergence, and empirical studies with synthetic FL datasets show that it reduces the time-to-accuracy by up to 2.7$\\times$ in a simulated cloud-edge-client environment. A concurrent work HFL~ also considers a similar design, while it does not attach theoretical analysis on its convergence behaviors. HybridFL~ further extends this primary design with two ideas: (1) quota-triggered edge-level aggregation: edge nodes stop waiting for more local updates once receiving a sufficient number of them; and (2) immediate cloud aggregation: cloud-level aggregation is conducted right after the edge-level one is completed. This decouples each pair of interactions (i.e., cloud-edge and edge-client), thereby further mitigating the impact of client drop-out and stragglers.\nDespite changing the aggregation rules, the hierarchical designs mentioned above still focus on establishing the convergence to a single global model, which does not deviate from the learning paradigm discussed throughout this survey. On the other hand, there also exist other formulations of collaborative learning such as clustered FL where clients are assigned to different groups and aggregation takes place within a group . As they aim to build personalized models for each group of clients, they are considered paradigm innovations to standard FL and thus fall out of the scope of this survey (\\cref{sec:background_problem}).\n\\PHM{Lightweight Privacy-Preserving Aggregation.} As mentioned in \\cref{sec:background_challenge_optimality}, uploading model updates in the clear may be vulnerable to exploratory attacks which plague clients' privacy. Therefore, model aggregation is preferably safeguarded by cryptographic techniques, which inevitably induces extra computation and communication overhead. \nOne line of this work leverages secure multiparty computation (SMPC). One of the most impactful works is Google's secure aggregation (SecAgg) protocol, where pseudorandom masks are used for data confidentiality and Shamir's secret sharing scheme~ is used to accommodate client dropout~. It has provable rigorous privacy guarantees under both passive and active adversary models. Given its quadratic communication overhead during secret sharing, SecAgg has inspired many other SMPC protocols with improved performance. For example, SecAgg+~ optimizes SecAgg in both communication and computation by replacing the complete graph with a random sparse one of logarithmic degree. TurboAgg~, on the other hand, attempts to optimize SecAgg through the use of multi-group circular topology, additive secret sharing, and Lagrange coding. Compared to SecAgg+, TurboAgg incurs a slightly higher overhead in the asymptotic sense for both communication and computation, and only focuses on tackling honest-but-curious adversaries. While all of the above schemes rely on Shamir's scheme for sharing clients' secrets, there also exist other secret sharing techniques that offer different trade-offs across performance, privacy, and dropout resilience, e.g., encoding using Maximum Distance Separable matrices as in LightSecAgg~, or Fast Fourier Transform as in FastSecAgg~.\nBesides SA, Homomorphic Encryption (HE) is another commonly used privacy-preserving aggregation technique that comes with prohibitively high message inflation and runtime overhead. In response, BatchCrypt~ implements an end-to-end solution for batching multiple plaintexts into one large plaintext so that HE-related operations can be performed in a data-parallel manner. BatchCrypt is shown to speed up the training by 23$\\times$-93$\\times$ compared to plain Paillier~ (a prevalent variant of HEs), but it still leaves the message inflation suboptimal and is incompatible with top s\\% sparsification approaches~. Instead of optimizing traditional HE schemes by batching, FLASHE~ proposes a lightweight HE scheme that is tailored for cross-silo FL. It induces negligible ($\\leq6\\%$) computational overhead and no network communication overhead compared to plaintext FL for staying symmetric. Its performance is also optimized when interacting with sparsification techniques.","id":"542ef1b6-edee-4a8a-bd3d-68e0c8f5c762","level":"subsubsection","origin_cites_number":15,"parent_id":"d6884228-aa74-48a6-9eb1-ee9a7ae15668","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Reporting Phase\\label{sec:prior_reporting"],["subsubsection","Aggregation Latency Reduction\\label{sec:prior_reporting_round"]],"subsections":[],"title":"Aggregation Latency Reduction\\label{sec:prior_reporting_round"},{"cite_extract_rate":0.555555555555555,"cites":[635,8367,582,648,649],"content":"}\nIn FedAvg~, the de facto standard aggregation method, local model updates simply get weighted by the corresponding numbers of training samples and then added up. While it guarantees convergence when even dealing with non-convex empirical risk functions in IID data settings~, it is observed to yield unstable convergence behavior or even divergence when it faces models trained with arbitrarily non-IID data. There are thus rising interests on whether the aggregation can be more adaptive w.r.t. the data heterogeneity across clients.\n\\PHM{Server-Side Optimizers.} Other than accelerating convergence with local momentum (\\cref{sec:prior_configuration_round}), there are also exploration efforts on server-side momentum. As there is originally no optimizer at the server in FL, these methods first need to generalize the existing aggregation algorithm. Specifically, at each round, instead of collecting local model weights, the server instead collects their changes and treats these changes as the ``pseudo-gradient'' for the server, which the server can use to update the global model with adaptive optimizers. FedAvgM~ initiates the empirical studies with the simplest form of momentum applied at the server, while SlowMo~ independently proposes a similar scheme and also attaches the theoretical analysis for its convergence behaviors. A recent work~ makes more sophisticated use of momentum such as adopting AdaGrad~, Adam~ and YOGI~ optimizers (which correspond to FedAdaGrad, FedAdam, and FedYOGI, respectively). It is shown that FedYOGI consistently outperforms FedAvgM in terms of validation performance for both sparse- and dense-gradient FL tasks. Server-side momentum methods feature no need for persistent states or computation burdens at the client end, making them well suited to cross-device scenarios.","id":"cc07a19d-bfd0-4b7a-9057-e912fa55b00d","level":"subsubsection","origin_cites_number":9,"parent_id":"d6884228-aa74-48a6-9eb1-ee9a7ae15668","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Recent Optimization Approaches\\label{chap:prior"],["subsection","Optimizing the Reporting Phase\\label{sec:prior_reporting"],["subsubsection","Adaptive Aggregation\\label{sec:prior_reporting_end"]],"subsections":[],"title":"Adaptive Aggregation\\label{sec:prior_reporting_end"},{"cite_extract_rate":0,"cites":[],"content":"}\nAside from innovating optimization solutions, there are also researchers contributing with cornerstone works that benefit the community with informative insights from systematical measurement studies (\\cref{sec:cornerstone_measurement}) and grounding benchmarking tools (\\cref{sec:cornerstone_benchmarking}), as visualized in Fig.~\\ref{fig:cornerstone}.\n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=0.85\\columnwidth]{skeleton-4}\n \\caption{Taxonomy of the work introduced in \\cref{chap:concerstone}.}\n \\label{fig:cornerstone}\n\\end{figure}","id":"52489191-afca-4640-b251-831b75c284de","level":"section","origin_cites_number":0,"parent_id":"419df863-69f4-4c35-93d0-018d167ed4ec","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Measurement and Benchmarking Tools\\label{chap:concerstone"]],"subsections":["b59ba9be-d908-40e7-8000-731ab0a60a79","271672fe-a947-4489-af72-cdac79330349"],"title":"Measurement and Benchmarking Tools\\label{chap:concerstone"},{"cite_extract_rate":0.75,"cites":[601,583,616],"content":"}\nDue to the complicated interplay between statistical utility and system utility, there are a few measurement studies which dedicate to conducting thorough investigations and providing actionable implications for interested researchers.\n\\begin{itemize}\n \\item \\textit{FLASH}\\footnote{We name it after the corresponding Github repository handle~.}~: FLASH particularly studies the impacts that heterogeneity has on both the statistical utility and system utility. This work contributes to the community mainly with three aspects: (1) it establishes a novel large-scale dataset that reflects the system and state heterogeneity among 136,000 real-world smartphones; (2) it implements and open-sources an FL platform that provides a heterogeneity-aware environment for experiments; (3) it reports comprehensive findings that stem from the authors' extensive measurements atop the platform.\n Among the authors' numerous implications, there are two fresh findings at the time when the study is written. First, gradient compression methods (e.g., Gradient Dropping~ and SignSGD~) can hardly shorten the convergence time under heterogeneous cross-device settings. Second, advanced aggregation algorithms that overlook some aspects of heterogeneity will be less effective in realistic settings. Readers can refer to the paper for more details.\n\\end{itemize}","id":"b59ba9be-d908-40e7-8000-731ab0a60a79","level":"subsection","origin_cites_number":4,"parent_id":"52489191-afca-4640-b251-831b75c284de","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Measurement and Benchmarking Tools\\label{chap:concerstone"],["subsection","Measurement-Based Research\\label{sec:cornerstone_measurement"]],"subsections":[],"title":"Measurement-Based Research\\label{sec:cornerstone_measurement"},{"cite_extract_rate":0,"cites":[],"content":"}\nRealistic benchmark suites are necessary for enabling fair, insightful, and reproducible evaluation of the effectiveness of system optimizations. As the time-to-accuracy performance relies on both the statistical utility and system utility (\\cref{sec:background_problem}), in the following summary of existing benchmarking tools, we aim to cover diverse aspects of simulating practical FL: data characteristics, client capabilities, and availability.","id":"271672fe-a947-4489-af72-cdac79330349","level":"subsection","origin_cites_number":0,"parent_id":"52489191-afca-4640-b251-831b75c284de","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Measurement and Benchmarking Tools\\label{chap:concerstone"],["subsection","Benchmarking Suites\\label{sec:cornerstone_benchmarking"]],"subsections":["e04c90d3-4295-47a6-81b9-5f869d19382c","6b106f87-b95f-4ecf-a35e-e685df38e3b3"],"title":"Benchmarking Suites\\label{sec:cornerstone_benchmarking"},{"cite_extract_rate":0.7333333333333331,"cites":[605,591,8367,633,652,653,582,650,637,648,651],"content":"}\nThere are two prevalent categories of training datasets. One line of work is \\textit{derived from conventional ML datasets} (e.g., \\texttt{CIFAR}~, \\texttt{MNIST}~, and \\texttt{Fashion-MNIST}~). To synthesize the non-IID nature as in real FL scenarios, the data partitions in these datasets are typically formed by restricting the number of data classes each client has (e.g., partitioning by shard-based methods as in~ or latent Dirichlet allocation (LDA) processes as in~). Although the data generated in such a way are indeed non-IID, they may not perfectly represent the real-world characteristics. For instance, besides the label distribution skew, in reality, non-IID data may also involve feature distribution skew (e.g., same words with different stroke widths), same labels with different features (e.g., images of clothing vary due to regional differences) and same features with different labels (e.g., the same context mapped to different next words due to personal habits)~.\nIn contrast, the other type of datasets is \\textit{collected in real distributed scenarios} and thus naturally captures the FL features. We briefly introduce existing open-source attempts in curating such datasets as follows.\n\\begin{itemize}\n \\item \\textit{LEAF}~: LEAF is an actively maintained project, which currently consists of 6 datasets spanning multiple applications such as computer vision (CV) (\\texttt{FEMNIST} and \\texttt{CelebA}) and natural language processing (NLP) (\\texttt{Shakespeare}, \\texttt{Reddit}, and \\texttt{Sentiment140}). Each dataset is generally formed by splitting the corresponding public dataset by the original contributors of the data samples. In other words, the non-IID nature comes from the unique behavior style of each contributor.\n \\item \\textit{FedScale}~: Similar to LEAF, FedScale also collects realistic datasets and partitions them by unique client identification. FedScale currently includes 18 datasets (including \\texttt{iNature}, \\texttt{OpenImage}, \\texttt{Google Landmark} etc.) which span 10 FL tasks. Apart from the comprehensive coverage of tasks, FedScale has further made four contributions to the community: (1) it has the training, validation, and testing set well established; (2) it streamlines different datasets into a unified format; (3) it accounts for various participation scale from hundreds to millions of clients; and (4) it provides handy APIs for users to customize their datasets.\n \\item \\textit{OARF}~: for a specific FL task, OARF assembles real-world datasets from different sources to realize data heterogeneity. For example, for sentiment analysis, it combines both the \\texttt{IMDB Movie Review} and \\texttt{Amazon Moview Review} datasets. In training, datasets belonging to different sources are distributed to different parties. As such, the data are split in a dataset-wise manner instead of a sample-wise one. OARF currently covers 9 tasks in CV and NLP.\n\\end{itemize}\nBesides the above systematic collection, there are also other separately maintained realistic datasets, such as \\texttt{Stackoverflow}~ and \\texttt{PERSONA-CHAT}~.","id":"e04c90d3-4295-47a6-81b9-5f869d19382c","level":"subsubsection","origin_cites_number":15,"parent_id":"271672fe-a947-4489-af72-cdac79330349","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Measurement and Benchmarking Tools\\label{chap:concerstone"],["subsection","Benchmarking Suites\\label{sec:cornerstone_benchmarking"],["subsubsection","Training Datasets\\label{sec:cornerstone_benchmarking_datasets"]],"subsections":[],"title":"Training Datasets\\label{sec:cornerstone_benchmarking_datasets"},{"cite_extract_rate":0.529411764705882,"cites":[656,655,612,41,7199,583,653,654,651],"content":"}\nIn addition to setting up data heterogeneity, we also need to incorporate system heterogeneity in realistic benchmarking. The most straightforward way to study FL designs with system utility borne in mind is to deploy in \\textit{production-oriented systems}. Such systems not only embed the ML backbone but also address practical problems like authentication, communication, encryption, and deployment to physical distributed environments. We sketch the open-source representatives.\n\\begin{itemize}\n \\item \\textit{FATE}~: FATE is an FL framework that can be deployed in distributed environments. In addition to its flexible ML pipeline, FATE also features several aspects that further facilitate the research on various goals of practical FL: (1) it supports privacy-preserving computation by implementing cryptographic algorithms such as the Diffie-Hellman key agreement~ and homomorphic encryption~; (2) it covers different training architectures including horizontal FL, vertical FL, and federated transfer learning; and (3) it allows a certain degree of customization on the FL pipeline such as the aggregation step. Given its heaviness in resource consumption, FATE is preferable in powering cross-silo applications instead of cross-device ones.\n \\item \\textit{FedML}~: FedML is also a secure and versatile FL framework that supports distributed mode. Compared to FATE, FedML is more flexible in communication engineering due to the ease of customizing message flow and topology definitions. It is also more lightweight and can thus accommodate training on mobile or IoT devices. Moreover, it can be accelerated with GPUs, while FATE is currently not compatible with hardware accelerators.\n \\item \\textit{Flower}~: Flower is a concurrent work with FedML, and it concentrates on providing a unified approach for FL with mobile devices. Similar to FedML, Flower bears in mind the goals of being (1) lightweight, (2) extensible, (3) scalable, and (4) compatible with diverse mobile platforms (e.g., Android and iOS) and ML-frameworks (e.g., PyTorch~ and Tensorflow~). It also implements secure aggregation methods like SecAgg (c.f. \\cref{sec:prior_reporting_round}) with easily configurable APIs~.\n \\item \\textit{Plato}~: similar to Flower, Plato also aims to facilitate FL research atop multiple ML backends including PyTorch, Tensorflow, and MindSpore~. Different from its counterparts, Plato is optimized in the development process, e.g., making extensive use of plugin mechanisms to maximize extensibility and maintainability. Apart from lightweight distributed deployment, Plato also offers solutions to integrate secure communication, reverse proxy, and load balancing for best fitting in production environments.\n\\end{itemize}\nAlthough using production systems yields the most realistic insights, it may not be practical for researchers with limited resources and time budgets. To meet the growing demand for conducting agile FL research, several platforms \\textit{that enable system-aware simulation} have been developed. As opposed to system-unaware simulators (e.g., Tensorflow Federated~, PySyft~, LEAF~, OARF~, and FedEval~), these platforms respect the impact of client system heterogeneity by associating each client with her computation and communication speed, as well as availability dynamics, which are either set manually by developers or by replaying realistic traces. In addition, these platforms also excel in producing comprehensive metrics needed in performance analysis. Compared to real deployment, on the other hand, these systems allow researchers to make fast-forward progress without being blocked by real-world bottlenecks in computation and communication.\n\\begin{itemize}\n \\item \\textit{Flower}~: besides deployment on real mobile devices as just mentioned, Flower also supports simulation in the cloud with configurable system-level parameters such as bandwidth constraints and computational capabilities. With that, researchers can experiment with larger and more compute-intensive FL workloads that cannot be run on today's mobile devices.\n \\item \\textit{FedScale}~: aside from curating real-world datasets (\\cref{sec:cornerstone_benchmarking_datasets}), FedScale also builds an automated runtime to simulate FL in realistic settings. By design, FedScale integrates \\texttt{AI Benchmark} and \\texttt{MobiPerf Measurements} system traces to simulate clients' heterogeneous training speed and network throughput, respectively. It also incorporates a large-scale user behavior dataset that was formulated in~ to emulate clients' availability dynamics. Compared to Flower, it lacks support for deployment on real distributed devices. Still, it broadly simulates realistic cross-device heterogeneity and can embrace new behavior traces with its APIs.\n\\end{itemize}","id":"6b106f87-b95f-4ecf-a35e-e685df38e3b3","level":"subsubsection","origin_cites_number":17,"parent_id":"271672fe-a947-4489-af72-cdac79330349","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Measurement and Benchmarking Tools\\label{chap:concerstone"],["subsection","Benchmarking Suites\\label{sec:cornerstone_benchmarking"],["subsubsection","Production Systems and Simulation Platforms\\label{sec:cornerstone_benchmarking_platforms"]],"subsections":[],"title":"Production Systems and Simulation Platforms\\label{sec:cornerstone_benchmarking_platforms"},{"cite_extract_rate":0,"cites":[],"content":"}","id":"24e6f906-e64a-4296-a2e5-eb81d7da883e","level":"section","origin_cites_number":0,"parent_id":"419df863-69f4-4c35-93d0-018d167ed4ec","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"]],"subsections":["526fca83-6724-45b6-99ae-76110ef046e4","1404800d-4173-49f1-b574-96677b669f93","02bbb984-68d7-4c50-804c-74430a998104","a4f483b3-14e7-4959-a5a0-300b337d0549"],"title":"Related Work and Concluding Remarks\\label{chap:conclusion"},{"cite_extract_rate":0.6923076923076921,"cites":[661,636,657,658,8368,659,660,591,628],"content":"}\nThe motivation for this survey stems from three observations. First, few in-depth surveys focus on system optimization for FL. As FL features strict compliance to privacy regulations as opposed to traditional distributed learning, many survey efforts are directed to the unique challenges such as enforcing data privacy and model robustness~, while the system optimization issues receive less attention in dedicated surveys.\nAs for common system issues shared by FL and traditional distributed learning, there do exist extensive surveys with detailed discussion. In~, the authors discuss the realm of mobile distributed machine learning, where algorithms are classified into three categories: 1) machine learning optimizers, 2) distributed optimization algorithms, and 3) data aggregation methods. In , the authors provide a detailed survey of communication-efficient distributed training algorithms. They identify four key factors that affect the communication cost in distributed learning and organize the literature accordingly: 1) synchronous scheme, 2) system architecture, 3) compression techniques and 4) parallelism of communication and computation.~ also studies the communication issues in distributed deep learning, however, with a focus on the commonly used lossless methods. It also contains a quantitative analysis of these methods based on real-world experiments. Compared to our survey, the scope of these surveys is not narrowed down to FL and thus does not fully capture all the system optimization challenges that are unique in FL. Notably, FL has in its standard workflow the selection phase which needs particular investigation due to client heterogeneity, while traditional distributed learning does not even have the notion of clients.\nWe notice that there are comprehensive surveys that cover a wide range of FL techniques including system-level optimizations. In~, the authors discuss in-depth the advances and open problems in FL at the time of writing. The presented problems and solutions are presented in four categories: 1) efficiency and effectiveness, 2) privacy preservation, 3) robustness enforcement and 4) fairness establishment. In~, the authors extensively introduce the challenges and research directions of FL at and for mobile edge network. As for FL at mobile edge network, they particularly elaborate techniques on 1) communication efficiency, 2) resource allocation, and 3) privacy and security. In~, the authors provide recommendations on algorithm-level optimization of FL applications, while they also briefly sketch the system-level constraints and practices. They discuss the reduction of communication, computation, and memory costs, as well as propose a basic model to estimate the communication efficiency of cross-device FL deployment. In~, the authors present a fine-grained multi-level classification of the FL literature, spanning six major topics: 1) statistical challenges, 2) communication efficiency, 3) client selection and scheduling, 4) security concerns, 5) privacy concerns, and 6) service pricing. Compared to this survey, the above-mentioned work is neither intended to focus on system-level optimizations nor do they categorize the literature by the phases in the synchronous training process. The survey most closely related to ours is~ which is dedicated to summarizing system-level optimizations used in asynchronous FL. As their studied architecture is fundamentally different from ours, their survey could be a good complement to this survey. It is also noteworthy that their classification mechanism on existing techniques is built upon the type of client heterogeneity, which also differs from our taxonomy.\nIn short, this survey aims to provide a succinct yet complete view of an essential domain in the FL literature: system optimization in synchronous federated training. We also summarize in Tab.~\\ref{tab:survey} the main similarities and differences between our survey and the existing surveys that are closely related.\n\\begin{table*}[h]\n \\begin{center}\n \\caption{Comparative summary between our survey and the most related work.}\n \\label{tab:survey}\n \\resizebox{1.0\\linewidth}{!}{\n \\begin{tabular}{cm{9cm}m{9cm}}\n \\toprule\n Survey & \\multicolumn{1}{c}{Similarities} & \\multicolumn{1}{c}{Differences} \\\\\n \\midrule\n Gu et al.~\n &\n Similar to our survey, this survey addresses some common challenges of synchronous federated training such as communication efficiency.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item As for machine learning optimizers, it does not discuss the server-side ones which are FL-specific.\n \\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and aggregation efficiency.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n Tang et al.~\n & Similar to our survey, this survey addresses some common challenges of synchronous federated training such as communication efficiency.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item It considers distributed vanilla SGD, which is different in convergence characteristics from FL, a special case of local SGD.\n \\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and client bias reduction.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n Shi et al.~\n & Similar to our survey, this survey addresses some common challenges of synchronous federated training such as communication efficiency.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item It focuses on controlling the system architecture and scheduling algorithms in data center learning, which is not applicable to synchronuous FL.\n \\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and data heterogeneity.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n Kairouz et al.~\n & Similar to our survey, this survey addresses some common challenges of synchronous federated training such as client bias reduction and communication efficiency.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item The scope of this survey is quite board, where for example theoretical analysis of convergence and privacy-preserving mechanisms are also extensively discussed.\n \\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and load balancing.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n Lim et al.~\n & Similar to our survey, this survey addresses some common challenges of FL such as communication efficiency and participant selection. It also mentions the standard workflow of synchronous training.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item The scope of this survey is quite board, where many applications of FL in mobile edge network, e.g., cyberattack detection, are also elaborated.\n \\item They do not organize system-level optimizations by the phases in the standard training workflow.\n \\item Our work addresses additional challenges that are not mentioned in this survey such as client bias reduction and lightweight privacy-preserving aggregation.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n Wang et al.~\n & Similar to our survey, this survey addresses some common challenges of FL such as communication efficiency and lightweight privacy-preserving aggregation.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item It centers on the algorithm-level optimizations as well as their theoretical guanrantees, while the challenges and directives of system-level issues are briefly introduced.\n \\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and load balancing.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n Wahab et al.~\n & Similar to our survey, this survey addresses some common challenges of FL such as participant selection and communication efficiency.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item The scope of this survey is quite board, where even training service pricing are also included.\n \\item Our work contains additional exploratory work in the system community that is not mentioned in this survey including measurement studies and benchmarking suites.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n Xu et al.~\n & Similar to our survey, this survey focuses on optimizing FL training process in the precence of client heterogeneity.\n & \\hspace{-5mm}\n \\begin{minipage}[b]{0.52\\textwidth}\n \\vspace{1mm}\n \\begin{itemize}\n \\item Its target architecture is asynchornous and thus the intersection of its introduced techniques and ours are fairly small.\n \\item Our work sorts the literature by the phases in the standard synchronous training workflow, which is distinct from the classification mechanism used in this survey.\n \\end{itemize}\n \\vspace{1mm}\n \\end{minipage} \\\\\n \\bottomrule\n \\end{tabular}\n }\n \\end{center}\n\\end{table*}","id":"526fca83-6724-45b6-99ae-76110ef046e4","level":"subsection","origin_cites_number":13,"parent_id":"24e6f906-e64a-4296-a2e5-eb81d7da883e","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Related Surveys\\label{sec:conclusion_related"]],"subsections":[],"title":"Related Surveys\\label{sec:conclusion_related"},{"cite_extract_rate":0,"cites":[],"content":"}\nThe primary goal of this survey is to help researchers design future optimization solutions. To stimulate more directives for FL practitioners, we discuss in the following some possible future directions that we derive from the literature as well as our development practice.","id":"1404800d-4173-49f1-b574-96677b669f93","level":"subsection","origin_cites_number":0,"parent_id":"24e6f906-e64a-4296-a2e5-eb81d7da883e","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Future Research Directions\\label{sec:conclusion_future"]],"subsections":["e4f64951-7cd3-49d1-bb82-2ae75da77f76","3dbd150d-cc33-4585-ba9e-a4c4d2278414","0e591dcb-87fb-4bba-a561-d583bde20c52"],"title":"Future Research Directions\\label{sec:conclusion_future"},{"cite_extract_rate":0.4,"cites":[585,582],"content":"}\nWhile RL agents are commonly used in guiding client selection for their beneficial balance between exploitation and exploration~, their training costs are inherently high. Before an RL agent is ready for use, it has to learn from tens to hundreds of full-sized FL training processes, whose cost may not be justifiable in practice. Inspired by the experience from container management~, we are interested in whether certain techniques, e.g., transfer learning~, can be adopted so that an RL agent can be trained from prior FL tasks, instead of learning from scratch for every single task.\nMoreover, existing algorithms all assume the participation scale (i.e., the number of clients training simultaneously) to be the magnitude of a hundred, because involving more clients in a round is observed to have marginal benefits under primary aggregation methods (e.g., FedAvg~). On the other hand, the number of available clients at each minute can be as many as thousands in the cross-device practice. It is still desirable to explore the selection of a larger crowd for boosting the time-to-accuracy performance.","id":"e4f64951-7cd3-49d1-bb82-2ae75da77f76","level":"subsubsection","origin_cites_number":5,"parent_id":"1404800d-4173-49f1-b574-96677b669f93","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Future Research Directions\\label{sec:conclusion_future"],["subsubsection","On the Selection Phase\\label{sec:conclusion_future_selection"]],"subsections":[],"title":"On the Selection Phase\\label{sec:conclusion_future_selection"},{"cite_extract_rate":0.33333333333333304,"cites":[662],"content":"}\nAligned with the observations in the literature~, most prior arts consistently configure different clients. Although there exist some heterogeneity-aware efforts like load balancing where the number of batches, batch size, and number of local epochs can vary across clients (\\cref{sec:prior_configuration_latency}), we anticipate that the design space for heterogeneity-aware client configurations could be larger, e.g., using different compression ratios or synchronization frequencies.\nTo achieve efficient and secure communication, it is also worth studying how to combine sparsification techniques with privacy-preserving methods. For example, model sparsification is by nature not compatible with secure multi-party computation protocols such as SecAgg~, because the coordinates of sparsified model values often vary across clients, preventing the pairwise masks from being canceled out during model aggregation as required by SecAgg. A recent work named SparseSecAgg~ attempts to tackle this issue, however, it implies the need for sharing some sparsified locations between each pair of clients, which cannot be directly extended to conventional sparsification schemes such as the top $s\\%$ scheme (\\cref{sec:prior_configuration_size}). More general reconciliation between the two techniques is still ripe for future investigation.","id":"3dbd150d-cc33-4585-ba9e-a4c4d2278414","level":"subsubsection","origin_cites_number":3,"parent_id":"1404800d-4173-49f1-b574-96677b669f93","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Future Research Directions\\label{sec:conclusion_future"],["subsubsection","On the Configuration Phase\\label{sec:conclusion_future_configuration"]],"subsections":[],"title":"On the Configuration Phase\\label{sec:conclusion_future_configuration"},{"cite_extract_rate":0.5454545454545451,"cites":[664,663,666,7038,665,591],"content":"}\nAs the system bottleneck is usually assumed to locate in clients instead of the server, most of the existing optimization efforts focus on improving the utility (system and statistical) of clients. It is thus interesting to investigate whether such an assumption holds in practice, especially when the scalability of the server is restricted due to rigid capabilities or limited budgets, or when its aggregation latency is not negligible due to the presence of large-scale models. \nBesides, existing lightweight privacy-preserving aggregation methods are not compatible with robustness enforcement techniques~. Because the true values of local models are concealed from the server, it has no way of inspecting their numerical features~ or validating their performance for anomaly detection~. This raises the tension between privacy and robustness. Leveraging Trusted Execution Environment (TEE) to perform model inspection securely might be a helpful workaround. However, due to current limits on TEE's hardware capabilities , there could be a foreseeably large performance downgrade of doing so. It thus remains largely unexplored as to how to navigate the sweet point of jointly maximizing accuracy, performance, privacy, and robustness in synchronous federated training.","id":"0e591dcb-87fb-4bba-a561-d583bde20c52","level":"subsubsection","origin_cites_number":11,"parent_id":"1404800d-4173-49f1-b574-96677b669f93","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Future Research Directions\\label{sec:conclusion_future"],["subsubsection","On the Reporting Phase\\label{sec:conclusion_future_reporting"]],"subsections":[],"title":"On the Reporting Phase\\label{sec:conclusion_future_reporting"},{"cite_extract_rate":0,"cites":[],"content":"}\nWe further discuss the applicability of the above-introduced optimization techniques in different contexts.","id":"02bbb984-68d7-4c50-804c-74430a998104","level":"subsection","origin_cites_number":0,"parent_id":"24e6f906-e64a-4296-a2e5-eb81d7da883e","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Discussion\\label{sec:conclusion_dicussion"]],"subsections":["19eb0c2c-4b7e-4dcc-9766-a74abafd9f2f","23cebd68-fb90-4604-b70b-25f0eba0e52b"],"title":"Discussion\\label{sec:conclusion_dicussion"},{"cite_extract_rate":0.75,"cites":[590,591,583],"content":"}\nFL applications are often categorized as either cross-device FL (where the participants are a mass of less capable mobile or IoT devices) or cross-silo FL (where the participants are 2-100 organizational entities)~. While the FL workflow that we base on throughout this survey is primarily proposed for cross-device FL~, it also generalizes to cross-silo settings. Hence, the scope of this survey does not preclude cross-silo FL, and hence many practical methods mentioned here should apply to both settings. For those techniques that are suitable for merely one setting, we have emphasized their limitations and stated the practical reasons behind them.","id":"19eb0c2c-4b7e-4dcc-9766-a74abafd9f2f","level":"subsubsection","origin_cites_number":4,"parent_id":"02bbb984-68d7-4c50-804c-74430a998104","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Discussion\\label{sec:conclusion_dicussion"],["subsubsection","Cross-Device FL and Cross-Silo FL\\label{sec:conclusion_dicussion_cross"]],"subsections":[],"title":"Cross-Device FL and Cross-Silo FL\\label{sec:conclusion_dicussion_cross"},{"cite_extract_rate":0.625,"cites":[669,670,667,671,668],"content":"}\nBased on the characteristics of data distribution, FL applications can also basically be classified as horizontal FL (where clients' data have the same feature space but different samples) or vertical FL (where clients' data share the same sample space but have different features)~. Horizontal FL is typically initiated by a service provider who wants to improve the quality of an ML application to better fit the end users' data of a specific domain. For example, by combining the statistics of users' input habits, the developers of a mobile keyboard application could achieve a more intelligent prediction of the next words typed by the users, thus enhancing their experience.\nOn the other hand, vertical FL participants are usually a few organizations who hold data in different domains while being likely to achieve win-win cooperation by knowledge sharing. For instance, a regional e-commerce company and a bank may share the same set of clients. By incorporating the knowledge of the clients' revenue and expenditure recorded at the bank as well as the purchasing traces collected by the e-commerce, they can build a more accurate prediction model on the clients' purchasing behaviors, benefiting both of their businesses.\nGiven that there exist diverse training workflows of vertical FL and the community has not yet achieved a consensus on the use of any one~, much discussion in this survey is biased towards horizontal FL that has a widely acknowledged architecture (\\cref{sec:background_federated}). Still, we can provide some insights for optimizing vertical FL training in general:\n\\begin{enumerate}\n \\item As the data is feature-partitioned, vertical FL needs the participation of all clients in each model update attempt. Thus, optimizations on participant selection (\\cref{sec:prior_selection}) are not applicable to vertical FL.\n \\item In terms of communication, model-agnostic compression techniques (\\cref{sec:prior_configuration_size}) should still apply to plaintext variants of vertical FL such as~. However, synchronization frequency reduction (\\cref{sec:prior_configuration_freq}) is not possible in vertical FL, as each client does not have a complete model and cannot conduct training independently.\n \\item As for local training, as all participants need processing data of the same sample space, load balancing techniques that assign a different amount of data samples to each client (\\cref{sec:prior_configuration_latency}) are not feasible in vertical FL. Moreover, as clients inherently have different local models, adaptive optimization and bias reduction methods (\\cref{sec:prior_configuration_round}) from horizontal FL do not generalize well to vertical FL. \n \\item Unlike horizontal FL, the forms of aggregation in vertical FL differ significantly by the model architecture. Thus, the mentioned optimizations (\\cref{sec:prior_reporting}) can hardly be a generic remedy.\n\\end{enumerate}\nBesides horizontal FL and vertical FL, there recently emerges a novel collaborative learning paradigm called Federated Transfer Learning (FTL)~ which can cope with more relaxed assumptions of client data distribution using transfer learning techniques~. Specifically, it deals with the cases where two clients' data not only differ in sample space but also feature space. Toward this end, FTL learns a common feature representation between the two clients and minimizes the empirical errors in predicting one client's labels by leveraging the labels of other clients. As FTL's training workflow ensembles some variants of vertical FL~, we believe that the above-mentioned insights for vertical FL also hold for FTL.","id":"23cebd68-fb90-4604-b70b-25f0eba0e52b","level":"subsubsection","origin_cites_number":8,"parent_id":"02bbb984-68d7-4c50-804c-74430a998104","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Discussion\\label{sec:conclusion_dicussion"],["subsubsection","Horizontal FL and Vertical FL\\label{sec:conclusion_dicussion_horizontal"]],"subsections":[],"title":"Horizontal FL and Vertical FL\\label{sec:conclusion_dicussion_horizontal"},{"cite_extract_rate":0,"cites":[],"content":"}\nIn this survey, we focus on system optimization in synchronous federated training and propose a natural taxonomy that categorizes existing solutions based on both the training phase and the type of utility at which they target. Apart from problem-driven attempts, we also include related cornerstone efforts including measurement studies and benchmarking suites. We expect this manuscript to be a \\textit{useful guideline} for the design and implementation of federated learning systems.\n\\ifCLASSOPTIONcompsoc\n \\section*{Acknowledgments}\n\\else\n \\section*{Acknowledgment}\n\\fi\nThe research was supported in part by RGC RIF grant R6021-20, and RGC GRF grants under the contracts 16209120, 16200221, and 16213120. We also thank the anonymous reviewers for their valuable feedback.\n\\ifCLASSOPTIONcaptionsoff\n \\newpage\n\\fi\n\\Urlmuskip=0mu plus 1mu\\relax\n\\bibliographystyle{IEEEtran}\n\\bibliography{./ref.bib}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/zjiangaj.jpg}}]{Zhifeng Jiang}\nreceived the BEng (Hons.) degree from the Department of Computer Science and Technology at Zhejiang University (ZJU) in 2019. Since then, he has been working towards a Ph.D. degree in the Department of Computer Science and Engineering, Hong Kong University of Science and Technology (HKUST). He is mainly interested in machine learning systems, with a special focus on efficient and scalable federated learning. He was a recipient of the Best Paper Runner-up Award at IEEE ICDCS 2021. He served on the artifact evaluation committees of ACM SOSP 2021, USENIX OSDI 2022, and USENIX ATC 2022.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/weiwa.jpeg}}]{Wei Wang} (Member, IEEE) received his B.Engr. and M.Engr. degrees from the Department of Electrical Engineering, Shanghai Jiao Tong University, China, in 2007 and 2010, respectively and his Ph.D. degree from the Department of Electrical and Computer Engineering, University of Toronto, Canada, in 2015. Since 2015, he has been with the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology (HKUST), where he is currently an Associate Professor. He is also affiliated with the HKUST Big Data Institute. Dr. Wang's research interests cover the broad area of distributed systems, with focus on serverless computing, machine learning systems, and cloud resource management. He published extensively in the premier conferences and journals of his fields. His research has won the Best Paper Runner Up awards of IEEE ICDCS 2021 and USENIX ICAC 2013.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/bli.jpg}}]{Bo Li} is a Chair Professor in Department of Computer Science and Engineering, Hong Kong University of Science and Technology. He was a Cheung Kong Scholar Visiting Chair Professor in Shanghai Jiao Tong University (2010-2016), and was the Chief Technical Advisor for ChinaCache Corp. (NASDAQ:CCIH), a leading CDN provider. He made pioneering contributions in multimedia communications and the Internet video broadcast, which attracted significant investment and received the Test-of-Time Best Paper Award from IEEE INFOCOM (2015). He received 6 Best Paper Awards from IEEE including IEEE INFOCOM (2021). He was the Co-TPC Chair for IEEE INFOCOM (2004).\nHe is a Fellow of IEEE. He received his PhD in the ECE Department, University of Massachusetts at Amherst, and his B. Eng. (summa cum laude) in the Computer Science from Tsinghua University, Beijing, China.\n\\end{IEEEbiography}\n\\vfill\n\\newpage\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/qyang.jpeg}}]{Qiang Yang} received the B.Sc. degree in astrophysics from Peking University, Beijing, China, in 1982, and the Ph.D. degree in computer science and the M.Sc. degree in astrophysics from the University of Maryland at College Park, College Park, MD, USA, in 1985 and 1989, respectively.\nHe was a Faculty Member with the University of Waterloo, Waterloo, ON, Canada, from 1989 to 1995, and Simon Fraser University, Burnaby, BC, Canada, from 1995 to 2001. He was the Founding Director of the Noah’s Ark Laboratory, Huawei, Hong Kong, from 2012 to 2014, and a Co-Founder of 4Paradigm Corporation, Beijing, an AI platform company. He is currently the Head (Chief AI Officer) with the AI Department, WeBank, Shenzhen, China, and the Chair Professor with the Department of Computer Science and Engineering (CSE), The Hong Kong University of Science and Technology, Hong Kong, where he was the Former Head of the Department of CSE and the Founding Director of the Big Data Institute, Hong Kong, from 2015 to 2018. He has authored several books, including Intelligent Planning (Springer), Crafting Your Research Future (Morgan and Claypool), and Constraint-Based Design Recovery for Software Engineering (Springer). His research interests include artificial intelligence, machine learning, and data mining, with an emphasis on transfer learning, automated planning, federated learning, and case-based reasoning.\nDr. Yang is a fellow of several international societies, including the ACM, AAAI, IAPR, and AAAS. He served as an Executive Council Member for the Association for the Advancement of AI from 2016 to 2020 and the President for the International Joint Conference on AI from 2017 to 2019. He was a recipient of several awards, including the 2004/2005 ACM KDDCUP Championship, the AAAI Innovative AI Applications Award in 2016, and the ACM SIGKDD Distinguished Service Award in 2017. He was the Founding Editor-in-Chief of the ACM Transactions on Intelligent Systems and Technology and the IEEE TRANSACTIONS ON BIG DATA.\n\\end{IEEEbiography}\n\\end{document}","id":"a4f483b3-14e7-4959-a5a0-300b337d0549","level":"subsection","origin_cites_number":0,"parent_id":"24e6f906-e64a-4296-a2e5-eb81d7da883e","prefix_titles":[["title","Towards Efficient Synchronous\\\\Federated Training: A Survey on\\\\System Optimization Strategies"],["section","Related Work and Concluding Remarks\\label{chap:conclusion"],["subsection","Conclusion\\label{sec:conclusion_conclusion"]],"subsections":[],"title":"Conclusion\\label{sec:conclusion_conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"419df863-69f4-4c35-93d0-018d167ed4ec\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ]\n ],\n \"subsections\": [\n \"428f0e83-f11e-4df6-8b69-0c0beb42e5b5\",\n \"498f24aa-02f6-4623-986c-cb8907d7dc9c\",\n \"a91a3aeb-ed4e-41b8-921a-e2c112807c50\",\n \"52489191-afca-4640-b251-831b75c284de\",\n \"24e6f906-e64a-4296-a2e5-eb81d7da883e\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 585,\n 589,\n 584,\n 581,\n 587,\n 580,\n 583,\n 586,\n 7262,\n 582,\n 7261,\n 588\n ],\n \"content\": \"\\\\label{chap:introduction}}\\n\\\\IEEEPARstart{B}{uilding} high-quality machine learning (ML) models demands a massive amount of training data. Yet, the communication cost and privacy concerns impinge on the process of collecting large volumes of data from diverse sources. It is not until recently that governments started to regulate the commercial use of data with privacy-preserving legislation (e.g., GDPR~, HIPAA~, and CCPA~). Compliance violations can be costly, with hefty fines up to hundreds of millions of dollars a year~. As such, the desire for multiple entities (e.g., mobile devices or large organizations) to collaboratively train a shared model efficiently and privately gives birth to a new ML paradigm called federated learning (FL)~. FL promises not to expose the clients' raw data, and has been widely adopted in many industries with applications ranging from mobile devices~ to financial management~ and medical care~.\\nApart from providing strong privacy guarantees, the key to the success of a federated training system lies in its efficiency. A typical efficiency metric is the \\\\emph{time-to-accuracy}, which is the wall clock time taken to train a model until it reaches the target accuracy. Despite the rich body of work that explored various optimization strategies, there is still plenty of room for further improvement due to the following distinct challenges posed to FL (\\\\cref{chap:background}): (1) the \\\\textit{lack of information for optimization}: the information needed for optimally configuring the system is usually outdated or simply unavailable due to privacy constraints and scaling issues; (2) the \\\\textit{tradeoff between statistical and system utility}: statistical utility (the number of iterations taken to reach a plausible target accuracy) and system utility (the duration of an iteration), the two determining factor for time-to-accuracy, are usually at odds in FL; (3) \\\\textit{client heterogeneity}: clients cannot be treated uniformly due to the intrinsic differences in terms of resources, data, and states; and (4) a \\\\textit{large configuration space}: the operational dimensions for system developers are too many to explore within an acceptable time. Given these challenges, it is worth summarizing existing research efforts to give researchers a holistic view of the lessons learned and to solicit further explorations.\\nTo position existing research attempts in optimizing the time-to-accuracy\\nperformance in FL, we propose a layered approach that categorizes them by the\\ntraining phases in which they take effect: selection, configuration, and\\nreporting (\\\\cref{chap:prior}). For the \\\\textit{selection phase} where the\\nserver chooses clients for participation, there are mainly two lines of\\noptimization efforts: (1) prioritizing clients either with\\nhigh statistical utility or system utility~\\\\cite\\n{zhang2021client, nishio2019client, wang2020optimizing}, and (2) explicitly considering\\nboth utilities and developing a more informed solution in response to client\\ndynamics in practice~. \\nAs for the \\\\textit{configuration phase} where the server sends the global\\nmodel to the selected clients with auxiliary configuration information, and\\nclients perform local training, we sort out four lines of work: (1) the first\\ntwo lines advocate mitigating the communication cost by reducing the model\\nsize~ and\\ndecreasing the synchronization frequency~\\\\cite\\n{hsieh2017gaia, kamp2018efficient, luping2019cmfl, chen2019communication,\\nchen2021communication}; (2) the last two lines minimize the computational\\noverhead by accelerating the training speed in each round~\\\\cite\\n{anh2019efficient, nguyen2020resource, wang2020towards, li2018federated, diao2020heterofl, ren2020accelerating} and reducing the number of training\\nrounds~. \\nIn terms of the \\\\textit{reporting phase}, we focus on the aggregation and\\noutline two related optimizations: (1) reducing the aggregation\\nlatency by adopting hierarchical methods~\\\\cite\\n{liu2020client, abad2020hierarchical, wu2020accelerating} and developing\\nlightweight privacy-preserving methods~\\\\cite\\n{so2021turbo, zhang2020batchcrypt, jiang2021flashe}, and (2)\\nimproving the long-term convergence rate with adaptive\\noptimizers on the server~. For each of the attempted optimizations, our discussion\\nincludes necessary details for readers to understand the motivation,\\nmechanisms, and major results. In addition, works such as measurement\\nstudies~ and benchmarking tools~\\\\cite\\n{caldas2018leaf, hu2020oarf, lai2021fedscale, fate, he2020fedml,\\nbeutel2020flower, plato} are indispensable in system research. We also survey the\\nstatus quo as a tutorial on FL practice (\\\\cref{chap:concerstone}).\\nOur work focuses on the system-level efforts made in improving the time-to-accuracy performance for synchronous federated training. We also share some implications derived from the literature and our survey process. It thus differs from existing surveys which mainly focus on ML algorithms and privacy-preserving algorithms. We expect this work to be an initial attempt to bridge the gap of system-oriented surveys in FL literature, as well as soliciting more contributions to related research.\",\n \"id\": \"428f0e83-f11e-4df6-8b69-0c0beb42e5b5\",\n \"level\": \"section\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"419df863-69f4-4c35-93d0-018d167ed4ec\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{chap:background}\\nIn this section, we give a detailed introduction to the system optimization problem in federated training. We start with a quick primer on the execution workflow of federated training (\\\\cref{sec:background_federated}), followed by the problem statement and the scope of this survey (\\\\cref{sec:background_problem}). We next outline two challenges that make the problem difficult: optimality and practicality (\\\\cref{sec:background_challenge}), which also serves as a summary of criteria for evaluating existing solutions presented thereafter.\",\n \"id\": \"498f24aa-02f6-4623-986c-cb8907d7dc9c\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"419df863-69f4-4c35-93d0-018d167ed4ec\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ]\n ],\n \"subsections\": [\n \"74ace61b-35de-40e1-bc2b-a3118bd28783\",\n \"e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4\",\n \"a9660a24-bfd4-4549-8a91-101fb1f84585\"\n ],\n \"title\": \"Background, Problem and Challenges\"\n },\n {\n \"cite_extract_rate\": 0.636363636363636,\n \"cites\": [\n 590,\n 587,\n 7261,\n 582,\n 583,\n 588,\n 591\n ],\n \"content\": \"\\\\label{sec:background_federated}\\nFederated learning (FL)~ has recently emerged\\nas a new paradigm of collaborative machine learning (ML) that allows multiple\\ndistributed clients (e.g., mobile devices or business organizations) to\\ncollaboratively train or evaluate a model with decentralized data. Compared\\nto traditional distributed learning in datacenter environments, FL mainly\\ndiffers in orchestration, resource constraints, data distribution, and\\nparticipation scale~. At its core, FL keeps private\\ndata on-premises, while using a central server to maintain a global model and\\niteratively refine it by aggregating each client's local updates. This design\\nreduces not only the communication cost but also the privacy risk in\\ngathering clients' raw data. Owing to its privacy guarantees, FL has found\\nwide applications in various domains. On mobile devices, Google runs FL to\\nimprove the user experience for Google Keyboard~\\\\cite\\n{yang2018applied, hard2018federated, ramaswamy2019federated,\\nchen2019federated} and Assistant~, while Apple deploys FL to\\nevaluate and tune speech recognition models~; in\\nfintech, both IBM~ and WeBank~ utilize\\nFL to detect financial misconducts; in healthcare, NVIDIA applies FL\\nto create medical imaging AI~ and predict patients' needs\\nfor oxygen~.\\nWhile both model training and evaluation play important roles in the development of an FL model, they have different criteria in system design. In this survey, we limit the scope to the \\\\textit{training} process, which is the most time-consuming and resource-intensive stage throughout the development of an FL model. Due to its predominance in practice, we focus on the support for the \\\\textit{synchronous} mode, wherein an ML model is trained across a pool of candidate clients in rounds, and in each round, the server needs to wait until a predefined deadline or receiving a sufficient number of clients' updates prior to deriving an aggregated update. In more detail, each round consists of the following three phases (Fig.~\\\\ref{fig:fl}).\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=1.0\\\\columnwidth]{standard-fl-protocol}\\n \\\\caption{Standard synchronous federated training protocol~.}\\n \\\\label{fig:fl}\\n\\\\end{figure}\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Client Selection}. At the beginning of each round, the server waits for a sufficient number of clients with eligible status (i.e., currently charging and connected to an unmetered network) to check in. The server then selects a subset of them based on certain strategies (e.g., randomly or selectively) for participation, and notifies the others to reconnect later.\\n \\\\item \\\\textit{Configuration}. The server next sends the global model status and configuration profiles (e.g., the number of local epochs or the reporting deadline) to each of the selected clients. Based on the instructed configuration, the clients perform local model training independently with their private data.\\n \\\\item \\\\textit{Reporting}. The server then waits for the participating clients to report local updates until reaching the predefined deadline. The current round is aborted if no enough clients report in time. Otherwise, the server aggregates the received local updates, uses the aggregate to update the global model status, and concludes the round.\\n\\\\end{itemize}\",\n \"id\": \"74ace61b-35de-40e1-bc2b-a3118bd28783\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"498f24aa-02f6-4623-986c-cb8907d7dc9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ],\n [\n \"subsection\",\n \"Federated Training\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Federated Training\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nThe primary goal of system optimization in federated training is to minimize the end-to-end resource usage of performing a task. The most common metric is the wall clock time, which is typically measured from the very beginning to a certain desirable checkpoint (e.g., convergence or reaching target accuracy). When a metered network is in use (e.g., when clients are on-demand virtual machines in a public cloud), the overall monetary cost becomes another relevant metric that deserves special attention. When uncharged devices are involved, the power consumption should also be considered. Because the cost and energy consumption generally grows linearly as time flies, in this survey, we are particularly interested in reducing the \\\\textit{time-to-accuracy}, i.e., the wall clock time for achieving a preset accuracy target.\\n\\\\begin{figure*}[t]\\n \\\\centering\\n \\\\begin{subfigure}[b]{0.19\\\\linewidth}\\n \\\\centering\\n \\\\includegraphics[width=\\\\columnwidth]{time-to-acc.pdf}\\n \\\\caption{Time to accuracy performance.}\\n \\\\label{fig:toy_time_to_acc}\\n \\\\end{subfigure}\\n \\\\begin{subfigure}[b]{0.19\\\\linewidth}\\n \\\\centering\\n \\\\includegraphics[width=\\\\columnwidth]{./round-to-acc.pdf}\\n \\\\caption{Round to accuracy performance.}\\n \\\\label{fig:toy_round_to_acc}\\n \\\\end{subfigure}\\n \\\\begin{subfigure}[b]{0.19\\\\linewidth }\\n \\\\centering\\n \\\\includegraphics[width=\\\\columnwidth]{./round-time}\\n \\\\caption{Average round time w/ standard deviation.}\\n \\\\label{fig:toy_round_time}\\n \\\\end{subfigure}\\n \\\\begin{subfigure}[b]{0.19\\\\linewidth}\\n \\\\centering\\n \\\\includegraphics[width=\\\\columnwidth]{./dirichlet}\\n \\\\caption{Data distributions of clients in NonIID.}\\n \\\\label{fig:toy_noniid}\\n \\\\end{subfigure}\\n \\\\begin{subfigure}[b]{0.19\\\\linewidth}\\n \\\\centering\\n \\\\includegraphics[width=\\\\columnwidth]{./figure/comp_zipf.pdf}\\n \\\\caption{Computational speed of clients in Straggler.}\\n \\\\label{fig:toy_straggler}\\n \\\\end{subfigure}\\n \\\\caption{A case study for illustrating system utility and statistical utility with FL emulation.} \\n \\\\label{fig:toy}\\n \\\\end{figure*}\",\n \"id\": \"e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"498f24aa-02f6-4623-986c-cb8907d7dc9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ],\n [\n \"subsection\",\n \"System Optimization: The Problem\\\\label{sec:background_problem\"\n ]\n ],\n \"subsections\": [\n \"172d329c-9179-421a-a850-1b10255f24b6\",\n \"1485b3df-12ae-4a59-a7c0-b7477f1b132f\"\n ],\n \"title\": \"System Optimization: The Problem\\\\label{sec:background_problem\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 580,\n 582\n ],\n \"content\": \"}\\nIntuitively, the time-to-accuracy performance of federated training is determined by two factors: the number of rounds taken to reach the accuracy and the average duration of training rounds. As in , we regard the former as \\\\textit{statistical utility} and the latter as \\\\textit{system utility} throughout this survey.\\nTo illustrate, we conduct an FL case study where 16 clients collaboratively train the LeNet5~ model to classify the images from the MNIST~ dataset. To emulate the cross-device environment, each client is run atop an AWC EC2 \\\\texttt{c5.xlarge} instance (4 vCPUs and 8 GB memory) and the network bandwidth is throttled to 21 Mbps to dictate the average mobile connection speed as of 2021~. We run FedAvg~ for model aggregation without loss of generality, atop which we enforce clients' privacy by implementing the SecAgg protocol~ (also see~\\\\cref{sec:prior_configuration_latency}) which makes the server blind to each client's local updates. Essentially, we design three schemes to compare:\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Baseline}: all clients' data is independent and identically distributed (IID). They also share the same computational speed.\\n \\\\item \\\\textit{NonIID}: the same as \\\\textit{Baseline} except that clients' data is non-IID as depicted by Fig.~\\\\ref{fig:toy_noniid}. This is achieved by latent Dirichlet allocation (described in~\\\\cref{sec:cornerstone_benchmarking_datasets}) with the concentration vectors being all 0.1's.\\n \\\\item \\\\textit{Straggler}: the same as \\\\textit{Baseline} except that clients’ computation speeds follow the Zipf's distribution (with $\\\\alpha=1.2$, i.e., moderately skewed). As shown in Fig.~\\\\ref{fig:toy_straggler}, the straggler is 5$\\\\times$ slower than the fastest who shares the same speed as clients in \\\\textit{Baseline}.\\n\\\\end{itemize}\\nWe first study the impact of statistical utility by comparing \\\\textit{Baseline} and \\\\textit{NonIID}. As indicated in Fig.~\\\\ref{fig:toy_round_time}, they proceed a round at the same speed. However, as the learning curves in Fig.~\\\\ref{fig:toy_time_to_acc} dictate, \\\\textit{Baseline} reaches the target accuracy (96.1\\\\%) with only 8.4 min, while \\\\textit{NonIID} takes 43.5 min, which is 5$\\\\times$ slower than \\\\textit{Baseline} does. Referring to the round-to-accuracy performance as in~Fig.~\\\\ref{fig:toy_round_to_acc}, one can know that the reason for this contrast lies in the difference in the numbers of rounds taken to convergence: 10 for \\\\textit{Baseline} while 40 for \\\\textit{NonIID}. Hence, despite having the same system utility, \\\\textit{Baseline} achieves much better time-to-accuracy performance for having greater statistical utility due to more evenly distributed data across clients. Of course, data distribution is uncontrollable in FL practice and developers can barely end up with IID cases as in \\\\textit{Baseline}. This adds up to the challenges of system-level optimizations in FL training, as later discussed in~\\\\cref{sec:background_challenge_optimality}.\\nNext, we provide a sense of how system utility affects the end-to-end performance by comparing \\\\textit{Baseline} with \\\\textit{Straggler}. As shown in Fig.~\\\\ref{fig:toy_round_to_acc}, both cases converge with the same amount of rounds: 10. However, it takes \\\\textit{Straggler} slightly longer in time (15.8 min, i.e., 1.9$\\\\times$) to reach the target accuracy, as demonstrated in Fig.~\\\\ref{fig:toy_time_to_acc}. By observing Fig.~\\\\ref{fig:toy_round_to_acc} and Fig.~\\\\ref{fig:toy_round_time}, we know that the source of the difference does not stem from the number of rounds used but the making span of each round. As all clients have to proceed at the same speed as the slowest clients in synchronous training, \\\\textit{Straggler} features lower system utility than \\\\textit{Baseline} does due to the presence of slow clients, rendering worse time-to-accuracy. Similarly, the disparity of clients' capabilities is inevitable in FL practice, which also complicates the design of system-level optimizations as mentioned in~\\\\cref{sec:background_challenge_optimality}.\",\n \"id\": \"172d329c-9179-421a-a850-1b10255f24b6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ],\n [\n \"subsection\",\n \"System Optimization: The Problem\\\\label{sec:background_problem\"\n ],\n [\n \"subsubsection\",\n \"Statistical Utility and Sytem Utility: A Case Study~\\\\label{sec:background_problem_utility\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Statistical Utility and Sytem Utility: A Case Study~\\\\label{sec:background_problem_utility\"\n },\n {\n \"cite_extract_rate\": 0.625,\n \"cites\": [\n 593,\n 595,\n 596,\n 600,\n 598,\n 599,\n 594,\n 592,\n 7038,\n 597\n ],\n \"content\": \"}\\nWe emphasize that the solutions discussed in this survey operate at the \\\\textit{system-level}. Therefore, the following directions of work are excluded from discussion, though they could effectively improve the statistical and/or system efficiency in synchronous federated training:\\n\\\\begin{itemize}\\n \\\\item Hardware updates, e.g., adapting programmable switches to enjoy the communication efficiency brought by in-network aggregation~.\\n \\\\item Security mechanisms, e.g., employing robust aggregation methods to protect the statistical utility from being impaired by model poisoning attacks~.\\n \\\\item Paradigm innovations, e.g., (1) letting clients fit different sets of model parameters by personalization~ or models of heterogeneous architectures through knowledge distillation~ to tackle data heterogeneity (a concept mentioned in~\\\\cref{sec:background_challenge_optimality}), (2) allowing clients to exchange data representations for realizing prototype learning~ to reduce communication burdens, or (3) permitting clients to send encoded versions of local datasets to the server to reduce the computational complexity~.\\n \\\\item Optimizations on upstream parts of the FL pipeline, e.g., searching for neural architectures that yield better predictive accuracy~.\\n\\\\end{itemize}\",\n \"id\": \"1485b3df-12ae-4a59-a7c0-b7477f1b132f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"e9b69d5b-e1b1-49ab-9b47-79a64bec7fd4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ],\n [\n \"subsection\",\n \"System Optimization: The Problem\\\\label{sec:background_problem\"\n ],\n [\n \"subsubsection\",\n \"What is beyond the Scope~\\\\label{sec:background_problem_beyond\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"What is beyond the Scope~\\\\label{sec:background_problem_beyond\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nDespite the clear objective, it is non-trivial to work out a feasible solution due to the following two challenges.\",\n \"id\": \"a9660a24-bfd4-4549-8a91-101fb1f84585\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"498f24aa-02f6-4623-986c-cb8907d7dc9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ],\n [\n \"subsection\",\n \"What Makes It Hard: The Challenges\\\\label{sec:background_challenge\"\n ]\n ],\n \"subsections\": [\n \"e9339976-85db-45bd-8a4a-f9969c5967b9\",\n \"b7818ca2-24b7-43f4-a5b2-dddedbcf0e1e\"\n ],\n \"title\": \"What Makes It Hard: The Challenges\\\\label{sec:background_challenge\"\n },\n {\n \"cite_extract_rate\": 0.92,\n \"cites\": [\n 3469,\n 618,\n 617,\n 608,\n 591,\n 605,\n 602,\n 603,\n 606,\n 613,\n 616,\n 612,\n 580,\n 607,\n 610,\n 614,\n 604,\n 583,\n 611,\n 615,\n 601,\n 4338,\n 609\n ],\n \"content\": \"}\\nFirst, the information needed in decision-making may be \\\\textit{outdated or even unavailable}. For example, to estimate a client's system utility, it is common to refer to its most recent response latency~. However, due to the dynamics over time, such information may not accurately reflect the client's current status. There also exists a cold-start issue, where we are unaware of a client's system capabilities until its first participation. As for estimating a client's statistical utility, the amount of available information is further limited by privacy concerns. According to the recent FL literature, exploratory attacks such as property inference~, membership inference~, and data reconstruction~ can be made possible with model updates. As such, even exposing model updates can discourage clients from participation, let alone inquiring about their data distributions or even raw data~. Note that the uncertainty in clients' statistical utility and system utility can be accumulated over time.\\nEven given a holistic view of the environment, the problem remains hard due to the \\\\textit{coupled nature} of statistical utility and system utility. Intuitively, improving system utility is equivalent to minimizing the average resource consumption (e.g., time or bandwidth) per task unit. On the other hand, reducing the resources invested in a task unit inevitably downgrades the quality of the outcome (e.g., statistical utility) as long as no resource is redundant. To exemplify, by constantly picking the fastest clients in client selection, the average duration of each round indeed decreases, whereas the number of rounds taken to target accuracy may be increased as well when other clients' data are under-represented in the global model. Another example can be found in model compression. To improve communication efficiency, a client can send only an important subset of model updates by sparsification~, or a low-bit representation of them by quantization~. Although the per-round communication duration can be significantly reduced by adopting a higher compression ratio, the convergence has to take more rounds to occur due to the loss of arithmetic precision.\\nThe problem is further complicated by \\\\textit{client heterogeneity.} Federated training involves tens to potentially millions of clients, each of which intrinsically differs from one another in the following three aspects: \\n\\\\begin{itemize}\\n \\\\item \\\\textit{Resource Heterogeneity}: due to the variability in hardware specifications and system-level constraints, clients in federated training typically possess different capabilities in computation (CPU/GPU/NPU, memory, and storage), communication (connectivity and bandwidth) and power (battery level and lifespan)~. These types of heterogeneity complicate the optimization of the overall system utility. For example, merely improving the communication speed does not necessarily lead to shorter end-to-end latency, especially when the straggler is bottlenecked by the computation~.\\n \\\\item \\\\textit{Data Heterogeneity}: as the training datasets of clients are typically generated based on their local activities and contexts, they are not IID. More specifically, clients' datasets mainly differ in two aspects\\\\footnote{See a more complete categorization of non-IID scenarios in \\\\cref{sec:cornerstone_benchmarking_datasets}.}: (1) sample quantity (i.e., the number of data samples), and (2) label partition (i.e., the distribution of data labels)~. As a result, not all of them are representative of the population distribution. In case we do not include all the clients in the federation, optimization for statistical utility has to additionally account for such heterogeneity.\\n \\\\item \\\\textit{State Heterogeneity}: as observed from real-world traces~, the available slots of mobile device clients vary significantly in temporal distribution due to different user behaviors (e.g., screen locking or battery charging). Therefore, in each round, there can be different sets of candidate clients to choose from, as well as different client drop-out outcomes. On top of the non-IID distribution of clients' data, this type of heterogeneity further complicates statistical utility optimization. Nevertheless, in the cross-silo settings, it may be less of a concern due to the stable and dedicated nature of clients' computing power~.\\n\\\\end{itemize}\\nLast but not least, it is \\\\textit{infeasible to search through the entire configuration space} for the global optimum. On the one hand, the space is prohibitively large, as a federated training task typically spans $10^1$--$10^6$ users and $10^2$--$10^4$ rounds~, wherein each phase of a round (\\\\cref{sec:background_federated}) has multiple configurable hyperparameters and alternative policies (e.g., client selection choices in the selection phase, or the number of local steps in configuration phase). On the other hand, most of the online decisions are made on the critical path of the task, meaning that the time spent on working out a solution also counts towards the end-to-end runtime performance, the very objective of the optimization. As a result, it is desirable to be guided by efficient and effective heuristic algorithms, especially balancing the exploration and exploitation efforts made in the configuration space.\",\n \"id\": \"e9339976-85db-45bd-8a4a-f9969c5967b9\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 25,\n \"parent_id\": \"a9660a24-bfd4-4549-8a91-101fb1f84585\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ],\n [\n \"subsection\",\n \"What Makes It Hard: The Challenges\\\\label{sec:background_challenge\"\n ],\n [\n \"subsubsection\",\n \"On the Optimality of a Solution\\\\label{sec:background_challenge_optimality\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"On the Optimality of a Solution\\\\label{sec:background_challenge_optimality\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 580,\n 591\n ],\n \"content\": \"}\\n\\\\begin{figure*}[t]\\n \\\\centering\\n \\\\includegraphics[width=0.9\\\\textwidth]{skeleton}\\n \\\\caption{Taxonomy of the approaches discussed in \\\\cref{chap:prior}.}\\n \\\\label{fig:opt}\\n\\\\end{figure*}\\nApart from navigating the performance-accuracy-privacy trade-off, the design process of a practical optimization solution should mitigate the accompanying side-effects on other aspects such as the loss of \\\\textit{robustness} to attacks and failures~. For example, to evaluate the statistical utility of a client, the server may require it to report the loss values generated in local training~. However, a malicious or free-rider client may intentionally respond with arbitrary values in the hope of messing with the orchestration or reaping the benefits of the federation without making solid contributions. As such backdoors are introduced by the optimization solution, the developers should take charge of eliminating the undesirable exploitations of these security loopholes. Other possible concerns that may arise as a result of a system optimization solution include but are not limited to \\\\textit{fairness} (e.g., whether participant bias is introduced in the solution), \\\\textit{generality} (e.g., whether the solution applies to diverse tasks), and \\\\textit{ease of deployment} (e.g., whether the solution can be implemented with moderate engineering efforts). In other words, a mature system optimization solution should not only improve the time-to-accuracy performance by enhancing statistical and system utility but also minimize the adverse impacts on other aspects that federated training also values in practice.\",\n \"id\": \"b7818ca2-24b7-43f4-a5b2-dddedbcf0e1e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"a9660a24-bfd4-4549-8a91-101fb1f84585\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Background, Problem and Challenges\"\n ],\n [\n \"subsection\",\n \"What Makes It Hard: The Challenges\\\\label{sec:background_challenge\"\n ],\n [\n \"subsubsection\",\n \"On the Practicality of a Solution\\\\label{sec:background_challenge_practicality\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"On the Practicality of a Solution\\\\label{sec:background_challenge_practicality\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nIn the past few years, considerable research efforts have been put into tackling the above challenges for fully unleashing the performance potential of FL training. In this section, we organize them by the training phases, i.e. selection (\\\\cref{sec:prior_selection}), configuration (\\\\cref{sec:prior_configuration}), and reporting (\\\\cref{sec:prior_reporting}), as visualized in Fig.~\\\\ref{fig:opt}.\",\n \"id\": \"a91a3aeb-ed4e-41b8-921a-e2c112807c50\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"419df863-69f4-4c35-93d0-018d167ed4ec\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ]\n ],\n \"subsections\": [\n \"6750018e-2b8c-43be-8977-a671addc47f2\",\n \"ded2658f-08bb-4199-8696-66ba4578329c\",\n \"d6884228-aa74-48a6-9eb1-ee9a7ae15668\"\n ],\n \"title\": \"Recent Optimization Approaches\\\\label{chap:prior\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 590\n ],\n \"content\": \"}\\nDue to the (potentially) large population size and the heterogeneity across clients, the effectiveness of the used participant selection algorithm plays a critical role in the time-to-accuracy performance in federated training. However, the state-of-the-practice system still relies on randomly picking participants~, which inevitably leads to waste of resources and suboptimal convergence speed. In response, there is an array of work to guide the selection, which can be roughly categorized by the target utilities that they improve.\",\n \"id\": \"6750018e-2b8c-43be-8977-a671addc47f2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"a91a3aeb-ed4e-41b8-921a-e2c112807c50\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Selection Phase\\\\label{sec:prior_selection\"\n ]\n ],\n \"subsections\": [\n \"2c47120e-34e5-44cd-b940-4eb997711491\",\n \"c9c46878-34a7-4fcb-8626-cb64857d9ade\"\n ],\n \"title\": \"Optimizing the Selection Phase\\\\label{sec:prior_selection\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 619,\n 620\n ],\n \"content\": \"}\\nThis line of work does not consider the interplay between statistical utility and system utility. Instead, they mainly focus on lifting either utility while leaving the other one ignored or controlled to a limited extent.\\n\\\\PHM{Statistics-Oriented.} To approach the convergence rate in centralized settings where the data is IID, CSFedAvg~ advocates that clients with a lower degree of non-IID data should participate more often. To this end, the authors propose weight divergence to capture the non-IID degree of data owned by a client. More precisely, it measures the normalized Euclidean distance between a client's model and the reference model trained by the server with auxiliary IID data. According to the 500-client simulation over CIFAR-10 and Fashion MNIST, CSFedAvg reduces the time-to-accuracy by up to 4.0$\\\\times$ and 2.7$\\\\times$, respectively, compared to random selection.\\nBesides heuristic methods, some researchers use reinforcement learning (RL) algorithms to learn which clients to select in the presence of data heterogeneity. For example, FAVOR~ seeks to reduce the number of rounds to reach a target accuracy with a deep Q-learning network (DQN)~. To capture each client's statistical characteristics, it takes the low-dimension representations of local models as the RL states. Compared to random selection, FAVOR can reduce the communication rounds by up to 49\\\\% in three image classification tasks. On the other hand, the training overhead for the RL agent may become an obstacle to FAVOR's real-world applications. Specifically, it is reported to take more than 100 episodes to train an agent suited for a specific learning task, which could be prohibitive as each episode corresponds to an entire FL process, i.e., training a global model from scratch for reaching a target accuracy.\\n\\\\PHM{System-Oriented.} In synchronous training, clients with the lowest system utility bottleneck the speed of a federation round. A straightforward way to bound the time usage is setting a deadline for randomly selected clients' to report updates and ignoring any update submitted after the deadline. To avoid waste of computing resources, FedCS~ takes a step further by proactively selecting a set of clients whose participation is not likely to miss the deadline according to latency estimation results. As there can be multiple eligible sets, FedCS further favors the solution with the largest scale of participation, which reduces part of the loss in statistical utility. Technically, the whole problem is formalized as a complex combinatorial optimization, and the authors resort to a greedy algorithm for efficient approximation. As indicated in their 1000-client simulation, FedCS outperforms FedLim (modified FedAvg with per-round deadlines imposed) by up to 1.2$\\\\times$ and 1.8$\\\\times$ in the time-to-accuracy when training over the non-IID CIFAR-10 and Fashion MNIST datasets, respectively.\",\n \"id\": \"2c47120e-34e5-44cd-b940-4eb997711491\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"6750018e-2b8c-43be-8977-a671addc47f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Selection Phase\\\\label{sec:prior_selection\"\n ],\n [\n \"subsubsection\",\n \"Partial Optimization Attempts\\\\label{sec:prior_selection_either\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Partial Optimization Attempts\\\\label{sec:prior_selection_either\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 7262,\n 580,\n 585\n ],\n \"content\": \"}\\nGiven the coupled nature of clients' system utility and statistical utility, it is more desirable to navigate the sweet point of jointly maximizing both of them.\\n\\\\PHM{Coarse-Grained.} TiFL~ first considers increasing the system utility. To that end, it divides clients into different tiers based on the observed runtime performance, and at each round only selects clients from the same tier for mitigating the waste of resources due to idle waiting for stragglers. To reduce the average iteration span, it also limits the number of times a (slow) tier can be selected. On top of that, the statistical utility is respected by prioritizing tiers with lower testing accuracy whenever there is more than one electable tier. Compared with FedCS, TiFL bears some resemblance in limiting the participation of less capable clients, while being more aware of the statistical utility. As reported in a 50-client cluster with 5 client tiers, TiFL achieves an improvement over random selection by up to 3$\\\\times$ speedup in overall training time and by 6\\\\% in accuracy.\\n\\\\PHM{Fine-Grained.} Compared to TiFL, Oort~ reconciles the demand for enhancing both system utility and statistical utility in finer granularity. Specifically, it associates each client with a continuous score and prioritizes those clients with higher scores. The score is meant to be a principled measurement of both the statistical utility (determined by the training loss) and the system utility (estimated from historical response latency). As some components of the score cannot be known in advance until the corresponding client's first participation, or cannot be guaranteed to be stable due to the client's runtime dynamics, the score estimation process is modeled as a Multi-Armed Bandit (MAB) problem and tackled by the Upper Confidence Bound (UCB) algorithm~. Apart from the scoring backbone, Oort also aims to address other practical issues like staleness and robustness. Oort was evaluated on a 1300-client GPU cluster with realistic datasets and simulation on the client heterogeneity. Compared to random selection, it reduces the training time by up to 14$\\\\times$ and improves model accuracy by up to 9.8\\\\%.\\nBesides the UCB algorithm, researchers also experiment with more sophisticated RL methods to cherrypick participants. Notably, AutoFL~ learns to select participants and execution targets for each individual based on the Q-Learning algorithm~. To achieve efficient FL execution, it identifies RL states that are critical to energy efficiency, convergence time, and accuracy. It also defines RL rewards that track the energy consumption of clients and model accuracy. Compared to random selection, AutoFL is reported to improve the average FL energy efficiency by up to 4.3$\\\\times$, while also exhibiting better training accuracy. Similar to FAVOR (mentioned in \\\\cref{sec:prior_selection_either}), however, training the Q-Learning model to converge needs multiple episodes, each of which corresponds to an entire FL training process. Its practicality could thus be challenged whenever offline training is infeasible or costly.\",\n \"id\": \"c9c46878-34a7-4fcb-8626-cb64857d9ade\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"6750018e-2b8c-43be-8977-a671addc47f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Selection Phase\\\\label{sec:prior_selection\"\n ],\n [\n \"subsubsection\",\n \"Co-Optimizing Statistical/System Utility\\\\label{sec:prior_selection_both\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Co-Optimizing Statistical/System Utility\\\\label{sec:prior_selection_both\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nIn the configuration phase, there are mainly two processes that are responsible for the time-to-accuracy performance. One is downlink (i.e., server-to-client) model transmission and the other is local model training. Thus, both aspects can be investigated for system optimization. As for communication overhead reduction, one can reduce the size of model updates (\\\\cref{sec:prior_configuration_size}) and decrease the synchronization frequency (\\\\cref{sec:prior_configuration_freq}). To lower computational overhead, one can shorten the training latency by balancing the workload across clients (\\\\cref{sec:prior_configuration_latency}), as well as reducing the number of rounds taken to converge by adopting heterogeneity-aware training algorithms (\\\\cref{sec:prior_configuration_round}). As the uplink model submission (i.e., client-to-server) that takes place in the reporting phase shares the same operational space as the downlink one, we combine the discussion on both of them in this section for brevity.\",\n \"id\": \"ded2658f-08bb-4199-8696-66ba4578329c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a91a3aeb-ed4e-41b8-921a-e2c112807c50\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Configuration Phase\\\\label{sec:prior_configuration\"\n ]\n ],\n \"subsections\": [\n \"0b02d7a9-809d-41bd-95d3-97624f811da6\",\n \"77437726-b7b4-48cb-ac0d-7210459652b2\",\n \"75004bfc-b89a-4773-8188-0896026459ce\",\n \"c1657bca-2a2d-4dc4-a8d5-1e81d9a7c3f6\"\n ],\n \"title\": \"Optimizing the Configuration Phase\\\\label{sec:prior_configuration\"\n },\n {\n \"cite_extract_rate\": 0.7307692307692301,\n \"cites\": [\n 618,\n 617,\n 626,\n 628,\n 624,\n 622,\n 606,\n 613,\n 616,\n 621,\n 625,\n 610,\n 611,\n 8366,\n 627,\n 615,\n 601,\n 623,\n 4338\n ],\n \"content\": \"}\\nPrior arts of model update size reduction mainly fall into three camps: quantization, sketching, and sparsification. Ahead of the emergence of FL, the exploration of these directions has already been initiated in the context of traditional distributed learning. While their communication merits are mostly reproducible in FL, they also face new challenges due to the privacy regulations and client heterogeneity, which we will also point out hereafter.\\n\\\\PHM{Quantization.} Quantization converts each scalar in a model update to its low-bit representation which takes up less space. While quantization has already gained its fame in traditional distributed learning and we refer the readers to dedicated surveys like~ for more details, here we only introduce the most representative work. As the first quantization work in model training with rigorous convergence proof, QSGD~ performs unbiased quantization with standard random dithering, a technique borrowed from image processing. Since its birth, related works have been emerging with more aggressive quantization bit-widths and more appealing empirical performance. For example, TernGrad~ advocates using only ternary values (0, $\\\\pm$1) in the uplink direction, while signSGD~ can use only binary signs ($\\\\pm$) in both uplink and downlink communication. It is worth mentioning that a popular technique in tackling the precision loss brought by quantization is error feedback, whose basic idea is to accumulate the previous quantization errors and compensate for them in the current round. Leveraging this technique, ECQ-SGD~ performs consistently better than QSD in terms of both convergence speed and accuracy, while EF-SGD~ also achieves a narrower generalization gap from centralized training compared to signSGD.\\nDespite their generality, there are some practical concerns about applying these general quantization strategies to FL due to privacy constraints and client heterogeneity. For example, determining the clipping threshold for quantization needs to exploit the knowledge about its inputs (i.e., local model updates) for reducing the induced error as in dACIQ~. However, an FL client does neither possess a priori knowledge of others' model updates nor should it request these sensitive values. To work out a globally applicable clipping threshold, we may need to share some less sensitive information (e.g., the maximum and minimum values in local updates) across clients for threshold estimation as in BatchCrypt~. Still, whether such a circumvention guarantees robust estimation and immunity to privacy attacks remains an open question.\\n\\\\PHM{Sketching.} Existing quantization approaches assume the input values follow a certain distribution (e.g., uniform or bell-shaped), which may not always be the case in model updates~. To be more general, some researchers introduce sketching methods where some memory-saving data structures are used to approximate the exact distribution of model update values in a single processing pass over the values. For example, SketchML~ utilizes a quantile sketch method to generate a non-uniform mapping from gradient values to low-bit integers. SketchML achieves empirical success such as decreasing the gradient size by around 7$\\\\times$ and is the first effort to introduce sketching for compressing model updates in ML training. Similar to quantization, sketch algorithms can also make use of error feedback techniques to efficiently amend the errors induced by the approximation, as demonstrated in SketchedSGD~ and FetchSGD~. There is also sketching practice that compresses auxiliary variables apart from model updates, such as sketching clients' momenta and per-coordinate learning rates as in~.\\n\\\\PHM{Sparsification.} While quantization and sketching operate at the precision level in terms of model size reduction, sparsification operates at the coordinate level. Specifically, sparsification allows each client to transmit only a sparse subset of its model updates, while the rest are accumulated and incorporated into future training. Technically, the sparsified gradient is obtained by first performing element-wise multiplication on the original gradient with a certain 0/1 mask and then discarding zero elements. The mask is typically randomly generated as in~, while another commonly used variant is the top $s$\\\\% scheme where 1 is given to the coordinates that rank top $s$\\\\% in absolute magnitude and 0 otherwise~. The top $s$\\\\% method is reported to reduce network traffic by up to three orders of magnitude, while still preserving model quality~.\\nWhile similar cost savings are transferable to plaintext FL, it is unclear whether sparsification can be further compatible with cryptographic techniques that are widely adopted for privacy enforcement in FL. For example, apart from the uplink model updates, it is also desirable to sparsify the downlink global update for fully releasing the potential for communication improvement. However, implementing the downlink sparsification may not be feasible when the server is not aware of the plaintext values of the aggregated update as a result of the applied Secure Multi-Party Computation (SMPC)~ or Homomorphic Encryption (HE)~ techniques.\\nIt is noteworthy that as quantization (or sketching) and sparsification are orthogonal to each other, they can be combined to reap the most benefits in terms of model size reduction~.\",\n \"id\": \"0b02d7a9-809d-41bd-95d3-97624f811da6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 26,\n \"parent_id\": \"ded2658f-08bb-4199-8696-66ba4578329c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Configuration Phase\\\\label{sec:prior_configuration\"\n ],\n [\n \"subsubsection\",\n \"Model Update Size Reduction\\\\label{sec:prior_configuration_size\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model Update Size Reduction\\\\label{sec:prior_configuration_size\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 630,\n 629\n ],\n \"content\": \"}\\nAt its core, the reduction in the synchronization frequency is achieved by identifying and precluding redundant synchronization efforts. This can be operated at different granularities ranging from clients, layers to individual parameters in a model update.\\n\\\\PHM{Client-Level.} The importance of an entire model update is usually measured by some numerical features. In the most intuitive form, a model update in Gaia~ is considered significant if its magnitude relative to the current value $\\\\mid \\\\frac{Update}{Value} \\\\mid$ exceeds a specific threshold such as 1\\\\%. While the magnitude may serve as a good indicator for how data center learning performs, it does not work in FL where determining an appropriate threshold is hard due to clients' heterogeneity. As such, some researchers propose to involve the comparison with some reference points for more robust measurement of the importance. For example,~ tracks the Euclidean distance between the local model and a reference model, while CMFL~ focuses on the number of coordinates with the same sign in the local model and the most recent global model.\\n\\\\PHM{Layer-Level.} Apart from considering a model update as a whole, another line of work tries to reduce the synchronization frequency on a layer basis. A representative work done in this direction is TWAFL~ where the model aggregation is conducted layer-wise. As observations made in deep neural network (DNN) fine-tuning~, shallow layers in a DNN learn general features across different datasets while deep layers learn ad hoc ones. TWAFL hence proposes to update shallow layers more frequently than deep ones as they are more responsible for the overall quality of the global model.\\n\\\\PHM{Parameter-Level.} Some industrial practitioners also consider whether to synchronize for each round at the level of individual parameters. Noticing that each parameter usually evolves in a transient-then-stable manner, i.e., it first varies drastically and then settles down around a certain value with slight oscillation, APF~ proposes to stop synchronizing those parameters whose evolution has reached a stationary phase.\",\n \"id\": \"77437726-b7b4-48cb-ac0d-7210459652b2\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"ded2658f-08bb-4199-8696-66ba4578329c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Configuration Phase\\\\label{sec:prior_configuration\"\n ],\n [\n \"subsubsection\",\n \"Synchronization Frequency Reduction\\\\label{sec:prior_configuration_freq\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Synchronization Frequency Reduction\\\\label{sec:prior_configuration_freq\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 631,\n 632\n ],\n \"content\": \"}\\nA client's training latency is determined by both its computational workload and resource capabilities. While the latter cannot be altered, the former still leaves room for optimization innovations. We discuss one major line of such efforts.\\n\\\\PHM{Load Balancing.} Given the variations in computing power and data volume, clients may not finish the training process at the same time. To mitigate the resulting straggler effects, suggest balancing the amount of training data across clients. Specifically, they turn to RL techniques for determining the optimal number of data units used in a communication round for each participant, intending to minimize the time and energy consumption and maximize the volume of involved data. While these solutions can reduce the round latency, the number of rounds needed for convergence does not necessarily remain unchanged because they do not account for data heterogeneity and thus the contribution of slow clients with important data could be throttled. To jointly consider heterogeneous system speed and data distribution during load balancing, carefully formulates an optimization problem where each client is associated with a weight that diminishes exponentially with the disparity between its number of labels and that of the population. Compared to even data allocation, this work manages to reduce the end-to-end time-to-accuracy.\\nLoad balancing can also be achieved by varying the number of optimization steps or the complexity of local models. For example, FedProx~ balances the system load across clients by formulating an inexact learning problem that allows variable steps of local solvers. On the other hand, HeteroFL~ assigns sub-models with different widths of hidden channels to clients so that clients with fewer capabilities can train smaller sub-models. All sub-models share the same model architecture, and thus normal model aggregation is still possible. The authors empirically show that the quality of the global model trained with heterogeneous sub-models is comparable to that trained with full-sized local models.\\nApart from the computational load, it is sometimes also beneficial to balance the communication load across clients, especially when the network conditions are complicated as in wireless connections. For example, targeting a mobile edge computing (MEC) scenario where Time Division Multiple Access (TDMA) is implemented, optimizes both the data batch size and uplink/downlink frame time slots for each client to achieve the maximum learning efficiency. In addition to coping with CPU computing, the authors further extend the optimization problem to the scenario where devices are equipped with GPUs for training.\",\n \"id\": \"75004bfc-b89a-4773-8188-0896026459ce\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"ded2658f-08bb-4199-8696-66ba4578329c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Configuration Phase\\\\label{sec:prior_configuration\"\n ],\n [\n \"subsubsection\",\n \"Training Latency Reduction\\\\label{sec:prior_configuration_latency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Training Latency Reduction\\\\label{sec:prior_configuration_latency\"\n },\n {\n \"cite_extract_rate\": 0.785714285714285,\n \"cites\": [\n 634,\n 638,\n 635,\n 640,\n 636,\n 639,\n 641,\n 633,\n 582,\n 632,\n 637\n ],\n \"content\": \"}\\nIn the settings with heterogeneous data, more local computation in a communication round does not necessarily lead to fewer numbers of rounds for reaching satisfying optima~. Thus, adopting heterogeneity-aware techniques such as adaptive optimization and bias reduction can help guarantee the convergence speed in FL practice.\\n\\\\PHM{Optimizer State Synchronization.} It is common practice for first-order moment optimization to apply momentum to dampen oscillations~. However, in the federated settings, if the clients' momenta are separately updated, they may deviate from each other due to non-IID data distributions. Thus, there are researchers proposing optimizer state synchronization frameworks where clients' optimizer states are synchronized by the server periodically. PR-SGD-Momentum~ is aligned with this direction and also gives proof of the linear speedup of convergence w.r.t. the number of workers. FedAC~ also applies momentum at clients with periodic synchronization, while it is proven to obtain the same linear speedup property with asymptotically fewer rounds of synchronization. MFL~ is another similar idea with theoretical guarantees, but it focuses on accelerating deterministic gradient descent (DGD) instead of SGD, unlike the previous two studies.\\n\\\\PHM{Client Bias Reduction.} Due to data heterogeneity, clients' model updates can be biased towards the minima of local objectives, known as ``client drift'' in the literature~, which hinders the convergence of the global model. To reduce the variance across clients, a natural idea is to regularize local objective functions for minimizing the drift. For example, assuming bounded dissimilarity between local functions, FedProx~ limits the Euclidean distance between local models and the global one by regularizing the square of the distance. FedDANE~ uses the same regularization term, while it formulates the other part of the objective function following the Distributed Approximate NEwton (DANE) method~ in classic distributed optimization. Despite having a similar theoretical convergence rate as FedProx, FedDANE underperforms FedProx in practice where the data heterogeneity is high and the participation rate is low, suggesting a discrepancy between theory and practice which needs further investigation. The objective function in FedDyn~ also considers the combination of a linear term and an L2 regularizer, with a different linear term which is formulated to align the empirical loss surfaces of clients. In theory, FedDyn ensures that the consensus point of model convergence across clients will be consistent with the global stationary solution as long as local models converge, regardless of the degree of data heterogeneity.\\nApart from regularizing local objectives, variants in clients' local updates can also be reduced by leveraging the idea of control variates borrowed from the convex optimization literature. For example, in SCAFFOLD~, each of the clients and the server maintains a control variate, and at each local step, a client de-biases its local updates with two control variates: one of its own and the other broadcast by the server. SCAFFOLD converges provably faster than FedAvg~ without any assumption made on the client selection or data heterogeneity. MIME~ considers a similar idea but makes a different choice on the specific definition of control variates.\\nWhile the use of control variates requires persistent client states, there exists another line of work that works for stateless clients: posterior averaging. Instead of approaching FL as optimization, this line of work formulates the problem as a posterior inference one. Compared to traditional federated optimization, posterior inference can benefit from an increased amount of local computation without risking stagnating at inferior optima. FedPA~ instantiates this idea with an efficient algorithm to conduct federated posterior inference with linear computation and communication costs.\",\n \"id\": \"c1657bca-2a2d-4dc4-a8d5-1e81d9a7c3f6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"ded2658f-08bb-4199-8696-66ba4578329c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Configuration Phase\\\\label{sec:prior_configuration\"\n ],\n [\n \"subsubsection\",\n \"Training Round Reduction\\\\label{sec:prior_configuration_round\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Training Round Reduction\\\\label{sec:prior_configuration_round\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nIn this phase, the operation room for system optimization is limited to either model uploading or model aggregation. As the former has already been combined in the last section, we hereafter focus on optimizing the aggregation process with two main directions explored in the literature: (1) directly reducing the aggregation latency at each round (\\\\cref{sec:prior_reporting_round}), and (2) expediting the convergence rate in the long run through conducting adaptive aggregation (\\\\cref{sec:prior_reporting_end}).\",\n \"id\": \"d6884228-aa74-48a6-9eb1-ee9a7ae15668\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a91a3aeb-ed4e-41b8-921a-e2c112807c50\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Reporting Phase\\\\label{sec:prior_reporting\"\n ]\n ],\n \"subsections\": [\n \"542ef1b6-edee-4a8a-bd3d-68e0c8f5c762\",\n \"cc07a19d-bfd0-4b7a-9057-e912fa55b00d\"\n ],\n \"title\": \"Optimizing the Reporting Phase\\\\label{sec:prior_reporting\"\n },\n {\n \"cite_extract_rate\": 0.533333333333333,\n \"cites\": [\n 643,\n 646,\n 644,\n 621,\n 645,\n 647,\n 7039,\n 642\n ],\n \"content\": \"}\\nCompared to local training in the configuration phase, model aggregation involves less intensive computation. However, its latency can still be salient because (1) large-scale participation can put pressure on the communication, and (2) the deployment of security methods can complicate the computation. We hereafter introduce the respective optimization efforts in the literature.\\n\\\\PHM{Hierarchical Aggregation.} The downsides of the traditional two-layer (server-client) system include (1) instability: the network bandwidth may be slow or unpredictable, especially in public networks and/or under geo-distributed settings; (2) risk of scalability: the cloud server may suffer from network congestion when concurrently receiving too many local updates; (3) heterogeneity: the straggler effects could be exacerbated by imbalanced network bandwidth. \\nTo address these issues, some researchers resort to a hierarchical design of model aggregation by introducing an extra level of edge servers, each of which is typically responsible for a small number of clients with proximity. For instance, in HierFAVG~, after a fixed number of local updates on clients, each edge server aggregates its own clients' models. Subsequently, after another fixed interval of edge aggregation, the cloud server aggregates all the edge servers' models. It is proven that HierFAVG still guarantees convergence, and empirical studies with synthetic FL datasets show that it reduces the time-to-accuracy by up to 2.7$\\\\times$ in a simulated cloud-edge-client environment. A concurrent work HFL~ also considers a similar design, while it does not attach theoretical analysis on its convergence behaviors. HybridFL~ further extends this primary design with two ideas: (1) quota-triggered edge-level aggregation: edge nodes stop waiting for more local updates once receiving a sufficient number of them; and (2) immediate cloud aggregation: cloud-level aggregation is conducted right after the edge-level one is completed. This decouples each pair of interactions (i.e., cloud-edge and edge-client), thereby further mitigating the impact of client drop-out and stragglers.\\nDespite changing the aggregation rules, the hierarchical designs mentioned above still focus on establishing the convergence to a single global model, which does not deviate from the learning paradigm discussed throughout this survey. On the other hand, there also exist other formulations of collaborative learning such as clustered FL where clients are assigned to different groups and aggregation takes place within a group . As they aim to build personalized models for each group of clients, they are considered paradigm innovations to standard FL and thus fall out of the scope of this survey (\\\\cref{sec:background_problem}).\\n\\\\PHM{Lightweight Privacy-Preserving Aggregation.} As mentioned in \\\\cref{sec:background_challenge_optimality}, uploading model updates in the clear may be vulnerable to exploratory attacks which plague clients' privacy. Therefore, model aggregation is preferably safeguarded by cryptographic techniques, which inevitably induces extra computation and communication overhead. \\nOne line of this work leverages secure multiparty computation (SMPC). One of the most impactful works is Google's secure aggregation (SecAgg) protocol, where pseudorandom masks are used for data confidentiality and Shamir's secret sharing scheme~ is used to accommodate client dropout~. It has provable rigorous privacy guarantees under both passive and active adversary models. Given its quadratic communication overhead during secret sharing, SecAgg has inspired many other SMPC protocols with improved performance. For example, SecAgg+~ optimizes SecAgg in both communication and computation by replacing the complete graph with a random sparse one of logarithmic degree. TurboAgg~, on the other hand, attempts to optimize SecAgg through the use of multi-group circular topology, additive secret sharing, and Lagrange coding. Compared to SecAgg+, TurboAgg incurs a slightly higher overhead in the asymptotic sense for both communication and computation, and only focuses on tackling honest-but-curious adversaries. While all of the above schemes rely on Shamir's scheme for sharing clients' secrets, there also exist other secret sharing techniques that offer different trade-offs across performance, privacy, and dropout resilience, e.g., encoding using Maximum Distance Separable matrices as in LightSecAgg~, or Fast Fourier Transform as in FastSecAgg~.\\nBesides SA, Homomorphic Encryption (HE) is another commonly used privacy-preserving aggregation technique that comes with prohibitively high message inflation and runtime overhead. In response, BatchCrypt~ implements an end-to-end solution for batching multiple plaintexts into one large plaintext so that HE-related operations can be performed in a data-parallel manner. BatchCrypt is shown to speed up the training by 23$\\\\times$-93$\\\\times$ compared to plain Paillier~ (a prevalent variant of HEs), but it still leaves the message inflation suboptimal and is incompatible with top s\\\\% sparsification approaches~. Instead of optimizing traditional HE schemes by batching, FLASHE~ proposes a lightweight HE scheme that is tailored for cross-silo FL. It induces negligible ($\\\\leq6\\\\%$) computational overhead and no network communication overhead compared to plaintext FL for staying symmetric. Its performance is also optimized when interacting with sparsification techniques.\",\n \"id\": \"542ef1b6-edee-4a8a-bd3d-68e0c8f5c762\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"d6884228-aa74-48a6-9eb1-ee9a7ae15668\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Reporting Phase\\\\label{sec:prior_reporting\"\n ],\n [\n \"subsubsection\",\n \"Aggregation Latency Reduction\\\\label{sec:prior_reporting_round\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Aggregation Latency Reduction\\\\label{sec:prior_reporting_round\"\n },\n {\n \"cite_extract_rate\": 0.555555555555555,\n \"cites\": [\n 635,\n 8367,\n 582,\n 648,\n 649\n ],\n \"content\": \"}\\nIn FedAvg~, the de facto standard aggregation method, local model updates simply get weighted by the corresponding numbers of training samples and then added up. While it guarantees convergence when even dealing with non-convex empirical risk functions in IID data settings~, it is observed to yield unstable convergence behavior or even divergence when it faces models trained with arbitrarily non-IID data. There are thus rising interests on whether the aggregation can be more adaptive w.r.t. the data heterogeneity across clients.\\n\\\\PHM{Server-Side Optimizers.} Other than accelerating convergence with local momentum (\\\\cref{sec:prior_configuration_round}), there are also exploration efforts on server-side momentum. As there is originally no optimizer at the server in FL, these methods first need to generalize the existing aggregation algorithm. Specifically, at each round, instead of collecting local model weights, the server instead collects their changes and treats these changes as the ``pseudo-gradient'' for the server, which the server can use to update the global model with adaptive optimizers. FedAvgM~ initiates the empirical studies with the simplest form of momentum applied at the server, while SlowMo~ independently proposes a similar scheme and also attaches the theoretical analysis for its convergence behaviors. A recent work~ makes more sophisticated use of momentum such as adopting AdaGrad~, Adam~ and YOGI~ optimizers (which correspond to FedAdaGrad, FedAdam, and FedYOGI, respectively). It is shown that FedYOGI consistently outperforms FedAvgM in terms of validation performance for both sparse- and dense-gradient FL tasks. Server-side momentum methods feature no need for persistent states or computation burdens at the client end, making them well suited to cross-device scenarios.\",\n \"id\": \"cc07a19d-bfd0-4b7a-9057-e912fa55b00d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"d6884228-aa74-48a6-9eb1-ee9a7ae15668\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Recent Optimization Approaches\\\\label{chap:prior\"\n ],\n [\n \"subsection\",\n \"Optimizing the Reporting Phase\\\\label{sec:prior_reporting\"\n ],\n [\n \"subsubsection\",\n \"Adaptive Aggregation\\\\label{sec:prior_reporting_end\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adaptive Aggregation\\\\label{sec:prior_reporting_end\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nAside from innovating optimization solutions, there are also researchers contributing with cornerstone works that benefit the community with informative insights from systematical measurement studies (\\\\cref{sec:cornerstone_measurement}) and grounding benchmarking tools (\\\\cref{sec:cornerstone_benchmarking}), as visualized in Fig.~\\\\ref{fig:cornerstone}.\\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=0.85\\\\columnwidth]{skeleton-4}\\n \\\\caption{Taxonomy of the work introduced in \\\\cref{chap:concerstone}.}\\n \\\\label{fig:cornerstone}\\n\\\\end{figure}\",\n \"id\": \"52489191-afca-4640-b251-831b75c284de\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"419df863-69f4-4c35-93d0-018d167ed4ec\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Measurement and Benchmarking Tools\\\\label{chap:concerstone\"\n ]\n ],\n \"subsections\": [\n \"b59ba9be-d908-40e7-8000-731ab0a60a79\",\n \"271672fe-a947-4489-af72-cdac79330349\"\n ],\n \"title\": \"Measurement and Benchmarking Tools\\\\label{chap:concerstone\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 601,\n 583,\n 616\n ],\n \"content\": \"}\\nDue to the complicated interplay between statistical utility and system utility, there are a few measurement studies which dedicate to conducting thorough investigations and providing actionable implications for interested researchers.\\n\\\\begin{itemize}\\n \\\\item \\\\textit{FLASH}\\\\footnote{We name it after the corresponding Github repository handle~.}~: FLASH particularly studies the impacts that heterogeneity has on both the statistical utility and system utility. This work contributes to the community mainly with three aspects: (1) it establishes a novel large-scale dataset that reflects the system and state heterogeneity among 136,000 real-world smartphones; (2) it implements and open-sources an FL platform that provides a heterogeneity-aware environment for experiments; (3) it reports comprehensive findings that stem from the authors' extensive measurements atop the platform.\\n Among the authors' numerous implications, there are two fresh findings at the time when the study is written. First, gradient compression methods (e.g., Gradient Dropping~ and SignSGD~) can hardly shorten the convergence time under heterogeneous cross-device settings. Second, advanced aggregation algorithms that overlook some aspects of heterogeneity will be less effective in realistic settings. Readers can refer to the paper for more details.\\n\\\\end{itemize}\",\n \"id\": \"b59ba9be-d908-40e7-8000-731ab0a60a79\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"52489191-afca-4640-b251-831b75c284de\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Measurement and Benchmarking Tools\\\\label{chap:concerstone\"\n ],\n [\n \"subsection\",\n \"Measurement-Based Research\\\\label{sec:cornerstone_measurement\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Measurement-Based Research\\\\label{sec:cornerstone_measurement\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nRealistic benchmark suites are necessary for enabling fair, insightful, and reproducible evaluation of the effectiveness of system optimizations. As the time-to-accuracy performance relies on both the statistical utility and system utility (\\\\cref{sec:background_problem}), in the following summary of existing benchmarking tools, we aim to cover diverse aspects of simulating practical FL: data characteristics, client capabilities, and availability.\",\n \"id\": \"271672fe-a947-4489-af72-cdac79330349\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"52489191-afca-4640-b251-831b75c284de\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Measurement and Benchmarking Tools\\\\label{chap:concerstone\"\n ],\n [\n \"subsection\",\n \"Benchmarking Suites\\\\label{sec:cornerstone_benchmarking\"\n ]\n ],\n \"subsections\": [\n \"e04c90d3-4295-47a6-81b9-5f869d19382c\",\n \"6b106f87-b95f-4ecf-a35e-e685df38e3b3\"\n ],\n \"title\": \"Benchmarking Suites\\\\label{sec:cornerstone_benchmarking\"\n },\n {\n \"cite_extract_rate\": 0.7333333333333331,\n \"cites\": [\n 605,\n 591,\n 8367,\n 633,\n 652,\n 653,\n 582,\n 650,\n 637,\n 648,\n 651\n ],\n \"content\": \"}\\nThere are two prevalent categories of training datasets. One line of work is \\\\textit{derived from conventional ML datasets} (e.g., \\\\texttt{CIFAR}~, \\\\texttt{MNIST}~, and \\\\texttt{Fashion-MNIST}~). To synthesize the non-IID nature as in real FL scenarios, the data partitions in these datasets are typically formed by restricting the number of data classes each client has (e.g., partitioning by shard-based methods as in~ or latent Dirichlet allocation (LDA) processes as in~). Although the data generated in such a way are indeed non-IID, they may not perfectly represent the real-world characteristics. For instance, besides the label distribution skew, in reality, non-IID data may also involve feature distribution skew (e.g., same words with different stroke widths), same labels with different features (e.g., images of clothing vary due to regional differences) and same features with different labels (e.g., the same context mapped to different next words due to personal habits)~.\\nIn contrast, the other type of datasets is \\\\textit{collected in real distributed scenarios} and thus naturally captures the FL features. We briefly introduce existing open-source attempts in curating such datasets as follows.\\n\\\\begin{itemize}\\n \\\\item \\\\textit{LEAF}~: LEAF is an actively maintained project, which currently consists of 6 datasets spanning multiple applications such as computer vision (CV) (\\\\texttt{FEMNIST} and \\\\texttt{CelebA}) and natural language processing (NLP) (\\\\texttt{Shakespeare}, \\\\texttt{Reddit}, and \\\\texttt{Sentiment140}). Each dataset is generally formed by splitting the corresponding public dataset by the original contributors of the data samples. In other words, the non-IID nature comes from the unique behavior style of each contributor.\\n \\\\item \\\\textit{FedScale}~: Similar to LEAF, FedScale also collects realistic datasets and partitions them by unique client identification. FedScale currently includes 18 datasets (including \\\\texttt{iNature}, \\\\texttt{OpenImage}, \\\\texttt{Google Landmark} etc.) which span 10 FL tasks. Apart from the comprehensive coverage of tasks, FedScale has further made four contributions to the community: (1) it has the training, validation, and testing set well established; (2) it streamlines different datasets into a unified format; (3) it accounts for various participation scale from hundreds to millions of clients; and (4) it provides handy APIs for users to customize their datasets.\\n \\\\item \\\\textit{OARF}~: for a specific FL task, OARF assembles real-world datasets from different sources to realize data heterogeneity. For example, for sentiment analysis, it combines both the \\\\texttt{IMDB Movie Review} and \\\\texttt{Amazon Moview Review} datasets. In training, datasets belonging to different sources are distributed to different parties. As such, the data are split in a dataset-wise manner instead of a sample-wise one. OARF currently covers 9 tasks in CV and NLP.\\n\\\\end{itemize}\\nBesides the above systematic collection, there are also other separately maintained realistic datasets, such as \\\\texttt{Stackoverflow}~ and \\\\texttt{PERSONA-CHAT}~.\",\n \"id\": \"e04c90d3-4295-47a6-81b9-5f869d19382c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"271672fe-a947-4489-af72-cdac79330349\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Measurement and Benchmarking Tools\\\\label{chap:concerstone\"\n ],\n [\n \"subsection\",\n \"Benchmarking Suites\\\\label{sec:cornerstone_benchmarking\"\n ],\n [\n \"subsubsection\",\n \"Training Datasets\\\\label{sec:cornerstone_benchmarking_datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Training Datasets\\\\label{sec:cornerstone_benchmarking_datasets\"\n },\n {\n \"cite_extract_rate\": 0.529411764705882,\n \"cites\": [\n 656,\n 655,\n 612,\n 41,\n 7199,\n 583,\n 653,\n 654,\n 651\n ],\n \"content\": \"}\\nIn addition to setting up data heterogeneity, we also need to incorporate system heterogeneity in realistic benchmarking. The most straightforward way to study FL designs with system utility borne in mind is to deploy in \\\\textit{production-oriented systems}. Such systems not only embed the ML backbone but also address practical problems like authentication, communication, encryption, and deployment to physical distributed environments. We sketch the open-source representatives.\\n\\\\begin{itemize}\\n \\\\item \\\\textit{FATE}~: FATE is an FL framework that can be deployed in distributed environments. In addition to its flexible ML pipeline, FATE also features several aspects that further facilitate the research on various goals of practical FL: (1) it supports privacy-preserving computation by implementing cryptographic algorithms such as the Diffie-Hellman key agreement~ and homomorphic encryption~; (2) it covers different training architectures including horizontal FL, vertical FL, and federated transfer learning; and (3) it allows a certain degree of customization on the FL pipeline such as the aggregation step. Given its heaviness in resource consumption, FATE is preferable in powering cross-silo applications instead of cross-device ones.\\n \\\\item \\\\textit{FedML}~: FedML is also a secure and versatile FL framework that supports distributed mode. Compared to FATE, FedML is more flexible in communication engineering due to the ease of customizing message flow and topology definitions. It is also more lightweight and can thus accommodate training on mobile or IoT devices. Moreover, it can be accelerated with GPUs, while FATE is currently not compatible with hardware accelerators.\\n \\\\item \\\\textit{Flower}~: Flower is a concurrent work with FedML, and it concentrates on providing a unified approach for FL with mobile devices. Similar to FedML, Flower bears in mind the goals of being (1) lightweight, (2) extensible, (3) scalable, and (4) compatible with diverse mobile platforms (e.g., Android and iOS) and ML-frameworks (e.g., PyTorch~ and Tensorflow~). It also implements secure aggregation methods like SecAgg (c.f. \\\\cref{sec:prior_reporting_round}) with easily configurable APIs~.\\n \\\\item \\\\textit{Plato}~: similar to Flower, Plato also aims to facilitate FL research atop multiple ML backends including PyTorch, Tensorflow, and MindSpore~. Different from its counterparts, Plato is optimized in the development process, e.g., making extensive use of plugin mechanisms to maximize extensibility and maintainability. Apart from lightweight distributed deployment, Plato also offers solutions to integrate secure communication, reverse proxy, and load balancing for best fitting in production environments.\\n\\\\end{itemize}\\nAlthough using production systems yields the most realistic insights, it may not be practical for researchers with limited resources and time budgets. To meet the growing demand for conducting agile FL research, several platforms \\\\textit{that enable system-aware simulation} have been developed. As opposed to system-unaware simulators (e.g., Tensorflow Federated~, PySyft~, LEAF~, OARF~, and FedEval~), these platforms respect the impact of client system heterogeneity by associating each client with her computation and communication speed, as well as availability dynamics, which are either set manually by developers or by replaying realistic traces. In addition, these platforms also excel in producing comprehensive metrics needed in performance analysis. Compared to real deployment, on the other hand, these systems allow researchers to make fast-forward progress without being blocked by real-world bottlenecks in computation and communication.\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Flower}~: besides deployment on real mobile devices as just mentioned, Flower also supports simulation in the cloud with configurable system-level parameters such as bandwidth constraints and computational capabilities. With that, researchers can experiment with larger and more compute-intensive FL workloads that cannot be run on today's mobile devices.\\n \\\\item \\\\textit{FedScale}~: aside from curating real-world datasets (\\\\cref{sec:cornerstone_benchmarking_datasets}), FedScale also builds an automated runtime to simulate FL in realistic settings. By design, FedScale integrates \\\\texttt{AI Benchmark} and \\\\texttt{MobiPerf Measurements} system traces to simulate clients' heterogeneous training speed and network throughput, respectively. It also incorporates a large-scale user behavior dataset that was formulated in~ to emulate clients' availability dynamics. Compared to Flower, it lacks support for deployment on real distributed devices. Still, it broadly simulates realistic cross-device heterogeneity and can embrace new behavior traces with its APIs.\\n\\\\end{itemize}\",\n \"id\": \"6b106f87-b95f-4ecf-a35e-e685df38e3b3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"271672fe-a947-4489-af72-cdac79330349\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Measurement and Benchmarking Tools\\\\label{chap:concerstone\"\n ],\n [\n \"subsection\",\n \"Benchmarking Suites\\\\label{sec:cornerstone_benchmarking\"\n ],\n [\n \"subsubsection\",\n \"Production Systems and Simulation Platforms\\\\label{sec:cornerstone_benchmarking_platforms\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Production Systems and Simulation Platforms\\\\label{sec:cornerstone_benchmarking_platforms\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\",\n \"id\": \"24e6f906-e64a-4296-a2e5-eb81d7da883e\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"419df863-69f4-4c35-93d0-018d167ed4ec\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ]\n ],\n \"subsections\": [\n \"526fca83-6724-45b6-99ae-76110ef046e4\",\n \"1404800d-4173-49f1-b574-96677b669f93\",\n \"02bbb984-68d7-4c50-804c-74430a998104\",\n \"a4f483b3-14e7-4959-a5a0-300b337d0549\"\n ],\n \"title\": \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n },\n {\n \"cite_extract_rate\": 0.6923076923076921,\n \"cites\": [\n 661,\n 636,\n 657,\n 658,\n 8368,\n 659,\n 660,\n 591,\n 628\n ],\n \"content\": \"}\\nThe motivation for this survey stems from three observations. First, few in-depth surveys focus on system optimization for FL. As FL features strict compliance to privacy regulations as opposed to traditional distributed learning, many survey efforts are directed to the unique challenges such as enforcing data privacy and model robustness~, while the system optimization issues receive less attention in dedicated surveys.\\nAs for common system issues shared by FL and traditional distributed learning, there do exist extensive surveys with detailed discussion. In~, the authors discuss the realm of mobile distributed machine learning, where algorithms are classified into three categories: 1) machine learning optimizers, 2) distributed optimization algorithms, and 3) data aggregation methods. In , the authors provide a detailed survey of communication-efficient distributed training algorithms. They identify four key factors that affect the communication cost in distributed learning and organize the literature accordingly: 1) synchronous scheme, 2) system architecture, 3) compression techniques and 4) parallelism of communication and computation.~ also studies the communication issues in distributed deep learning, however, with a focus on the commonly used lossless methods. It also contains a quantitative analysis of these methods based on real-world experiments. Compared to our survey, the scope of these surveys is not narrowed down to FL and thus does not fully capture all the system optimization challenges that are unique in FL. Notably, FL has in its standard workflow the selection phase which needs particular investigation due to client heterogeneity, while traditional distributed learning does not even have the notion of clients.\\nWe notice that there are comprehensive surveys that cover a wide range of FL techniques including system-level optimizations. In~, the authors discuss in-depth the advances and open problems in FL at the time of writing. The presented problems and solutions are presented in four categories: 1) efficiency and effectiveness, 2) privacy preservation, 3) robustness enforcement and 4) fairness establishment. In~, the authors extensively introduce the challenges and research directions of FL at and for mobile edge network. As for FL at mobile edge network, they particularly elaborate techniques on 1) communication efficiency, 2) resource allocation, and 3) privacy and security. In~, the authors provide recommendations on algorithm-level optimization of FL applications, while they also briefly sketch the system-level constraints and practices. They discuss the reduction of communication, computation, and memory costs, as well as propose a basic model to estimate the communication efficiency of cross-device FL deployment. In~, the authors present a fine-grained multi-level classification of the FL literature, spanning six major topics: 1) statistical challenges, 2) communication efficiency, 3) client selection and scheduling, 4) security concerns, 5) privacy concerns, and 6) service pricing. Compared to this survey, the above-mentioned work is neither intended to focus on system-level optimizations nor do they categorize the literature by the phases in the synchronous training process. The survey most closely related to ours is~ which is dedicated to summarizing system-level optimizations used in asynchronous FL. As their studied architecture is fundamentally different from ours, their survey could be a good complement to this survey. It is also noteworthy that their classification mechanism on existing techniques is built upon the type of client heterogeneity, which also differs from our taxonomy.\\nIn short, this survey aims to provide a succinct yet complete view of an essential domain in the FL literature: system optimization in synchronous federated training. We also summarize in Tab.~\\\\ref{tab:survey} the main similarities and differences between our survey and the existing surveys that are closely related.\\n\\\\begin{table*}[h]\\n \\\\begin{center}\\n \\\\caption{Comparative summary between our survey and the most related work.}\\n \\\\label{tab:survey}\\n \\\\resizebox{1.0\\\\linewidth}{!}{\\n \\\\begin{tabular}{cm{9cm}m{9cm}}\\n \\\\toprule\\n Survey & \\\\multicolumn{1}{c}{Similarities} & \\\\multicolumn{1}{c}{Differences} \\\\\\\\\\n \\\\midrule\\n Gu et al.~\\n &\\n Similar to our survey, this survey addresses some common challenges of synchronous federated training such as communication efficiency.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item As for machine learning optimizers, it does not discuss the server-side ones which are FL-specific.\\n \\\\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and aggregation efficiency.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n Tang et al.~\\n & Similar to our survey, this survey addresses some common challenges of synchronous federated training such as communication efficiency.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item It considers distributed vanilla SGD, which is different in convergence characteristics from FL, a special case of local SGD.\\n \\\\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and client bias reduction.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n Shi et al.~\\n & Similar to our survey, this survey addresses some common challenges of synchronous federated training such as communication efficiency.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item It focuses on controlling the system architecture and scheduling algorithms in data center learning, which is not applicable to synchronuous FL.\\n \\\\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and data heterogeneity.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n Kairouz et al.~\\n & Similar to our survey, this survey addresses some common challenges of synchronous federated training such as client bias reduction and communication efficiency.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item The scope of this survey is quite board, where for example theoretical analysis of convergence and privacy-preserving mechanisms are also extensively discussed.\\n \\\\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and load balancing.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n Lim et al.~\\n & Similar to our survey, this survey addresses some common challenges of FL such as communication efficiency and participant selection. It also mentions the standard workflow of synchronous training.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item The scope of this survey is quite board, where many applications of FL in mobile edge network, e.g., cyberattack detection, are also elaborated.\\n \\\\item They do not organize system-level optimizations by the phases in the standard training workflow.\\n \\\\item Our work addresses additional challenges that are not mentioned in this survey such as client bias reduction and lightweight privacy-preserving aggregation.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n Wang et al.~\\n & Similar to our survey, this survey addresses some common challenges of FL such as communication efficiency and lightweight privacy-preserving aggregation.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item It centers on the algorithm-level optimizations as well as their theoretical guanrantees, while the challenges and directives of system-level issues are briefly introduced.\\n \\\\item Our work addresses additional challenges that are not mentioned in this survey such as participant selection and load balancing.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n Wahab et al.~\\n & Similar to our survey, this survey addresses some common challenges of FL such as participant selection and communication efficiency.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item The scope of this survey is quite board, where even training service pricing are also included.\\n \\\\item Our work contains additional exploratory work in the system community that is not mentioned in this survey including measurement studies and benchmarking suites.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n Xu et al.~\\n & Similar to our survey, this survey focuses on optimizing FL training process in the precence of client heterogeneity.\\n & \\\\hspace{-5mm}\\n \\\\begin{minipage}[b]{0.52\\\\textwidth}\\n \\\\vspace{1mm}\\n \\\\begin{itemize}\\n \\\\item Its target architecture is asynchornous and thus the intersection of its introduced techniques and ours are fairly small.\\n \\\\item Our work sorts the literature by the phases in the standard synchronous training workflow, which is distinct from the classification mechanism used in this survey.\\n \\\\end{itemize}\\n \\\\vspace{1mm}\\n \\\\end{minipage} \\\\\\\\\\n \\\\bottomrule\\n \\\\end{tabular}\\n }\\n \\\\end{center}\\n\\\\end{table*}\",\n \"id\": \"526fca83-6724-45b6-99ae-76110ef046e4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"24e6f906-e64a-4296-a2e5-eb81d7da883e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Related Surveys\\\\label{sec:conclusion_related\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Related Surveys\\\\label{sec:conclusion_related\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nThe primary goal of this survey is to help researchers design future optimization solutions. To stimulate more directives for FL practitioners, we discuss in the following some possible future directions that we derive from the literature as well as our development practice.\",\n \"id\": \"1404800d-4173-49f1-b574-96677b669f93\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"24e6f906-e64a-4296-a2e5-eb81d7da883e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Future Research Directions\\\\label{sec:conclusion_future\"\n ]\n ],\n \"subsections\": [\n \"e4f64951-7cd3-49d1-bb82-2ae75da77f76\",\n \"3dbd150d-cc33-4585-ba9e-a4c4d2278414\",\n \"0e591dcb-87fb-4bba-a561-d583bde20c52\"\n ],\n \"title\": \"Future Research Directions\\\\label{sec:conclusion_future\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 585,\n 582\n ],\n \"content\": \"}\\nWhile RL agents are commonly used in guiding client selection for their beneficial balance between exploitation and exploration~, their training costs are inherently high. Before an RL agent is ready for use, it has to learn from tens to hundreds of full-sized FL training processes, whose cost may not be justifiable in practice. Inspired by the experience from container management~, we are interested in whether certain techniques, e.g., transfer learning~, can be adopted so that an RL agent can be trained from prior FL tasks, instead of learning from scratch for every single task.\\nMoreover, existing algorithms all assume the participation scale (i.e., the number of clients training simultaneously) to be the magnitude of a hundred, because involving more clients in a round is observed to have marginal benefits under primary aggregation methods (e.g., FedAvg~). On the other hand, the number of available clients at each minute can be as many as thousands in the cross-device practice. It is still desirable to explore the selection of a larger crowd for boosting the time-to-accuracy performance.\",\n \"id\": \"e4f64951-7cd3-49d1-bb82-2ae75da77f76\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"1404800d-4173-49f1-b574-96677b669f93\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Future Research Directions\\\\label{sec:conclusion_future\"\n ],\n [\n \"subsubsection\",\n \"On the Selection Phase\\\\label{sec:conclusion_future_selection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"On the Selection Phase\\\\label{sec:conclusion_future_selection\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 662\n ],\n \"content\": \"}\\nAligned with the observations in the literature~, most prior arts consistently configure different clients. Although there exist some heterogeneity-aware efforts like load balancing where the number of batches, batch size, and number of local epochs can vary across clients (\\\\cref{sec:prior_configuration_latency}), we anticipate that the design space for heterogeneity-aware client configurations could be larger, e.g., using different compression ratios or synchronization frequencies.\\nTo achieve efficient and secure communication, it is also worth studying how to combine sparsification techniques with privacy-preserving methods. For example, model sparsification is by nature not compatible with secure multi-party computation protocols such as SecAgg~, because the coordinates of sparsified model values often vary across clients, preventing the pairwise masks from being canceled out during model aggregation as required by SecAgg. A recent work named SparseSecAgg~ attempts to tackle this issue, however, it implies the need for sharing some sparsified locations between each pair of clients, which cannot be directly extended to conventional sparsification schemes such as the top $s\\\\%$ scheme (\\\\cref{sec:prior_configuration_size}). More general reconciliation between the two techniques is still ripe for future investigation.\",\n \"id\": \"3dbd150d-cc33-4585-ba9e-a4c4d2278414\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"1404800d-4173-49f1-b574-96677b669f93\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Future Research Directions\\\\label{sec:conclusion_future\"\n ],\n [\n \"subsubsection\",\n \"On the Configuration Phase\\\\label{sec:conclusion_future_configuration\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"On the Configuration Phase\\\\label{sec:conclusion_future_configuration\"\n },\n {\n \"cite_extract_rate\": 0.5454545454545451,\n \"cites\": [\n 664,\n 663,\n 666,\n 7038,\n 665,\n 591\n ],\n \"content\": \"}\\nAs the system bottleneck is usually assumed to locate in clients instead of the server, most of the existing optimization efforts focus on improving the utility (system and statistical) of clients. It is thus interesting to investigate whether such an assumption holds in practice, especially when the scalability of the server is restricted due to rigid capabilities or limited budgets, or when its aggregation latency is not negligible due to the presence of large-scale models. \\nBesides, existing lightweight privacy-preserving aggregation methods are not compatible with robustness enforcement techniques~. Because the true values of local models are concealed from the server, it has no way of inspecting their numerical features~ or validating their performance for anomaly detection~. This raises the tension between privacy and robustness. Leveraging Trusted Execution Environment (TEE) to perform model inspection securely might be a helpful workaround. However, due to current limits on TEE's hardware capabilities , there could be a foreseeably large performance downgrade of doing so. It thus remains largely unexplored as to how to navigate the sweet point of jointly maximizing accuracy, performance, privacy, and robustness in synchronous federated training.\",\n \"id\": \"0e591dcb-87fb-4bba-a561-d583bde20c52\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"1404800d-4173-49f1-b574-96677b669f93\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Future Research Directions\\\\label{sec:conclusion_future\"\n ],\n [\n \"subsubsection\",\n \"On the Reporting Phase\\\\label{sec:conclusion_future_reporting\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"On the Reporting Phase\\\\label{sec:conclusion_future_reporting\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nWe further discuss the applicability of the above-introduced optimization techniques in different contexts.\",\n \"id\": \"02bbb984-68d7-4c50-804c-74430a998104\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"24e6f906-e64a-4296-a2e5-eb81d7da883e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Discussion\\\\label{sec:conclusion_dicussion\"\n ]\n ],\n \"subsections\": [\n \"19eb0c2c-4b7e-4dcc-9766-a74abafd9f2f\",\n \"23cebd68-fb90-4604-b70b-25f0eba0e52b\"\n ],\n \"title\": \"Discussion\\\\label{sec:conclusion_dicussion\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 590,\n 591,\n 583\n ],\n \"content\": \"}\\nFL applications are often categorized as either cross-device FL (where the participants are a mass of less capable mobile or IoT devices) or cross-silo FL (where the participants are 2-100 organizational entities)~. While the FL workflow that we base on throughout this survey is primarily proposed for cross-device FL~, it also generalizes to cross-silo settings. Hence, the scope of this survey does not preclude cross-silo FL, and hence many practical methods mentioned here should apply to both settings. For those techniques that are suitable for merely one setting, we have emphasized their limitations and stated the practical reasons behind them.\",\n \"id\": \"19eb0c2c-4b7e-4dcc-9766-a74abafd9f2f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"02bbb984-68d7-4c50-804c-74430a998104\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Discussion\\\\label{sec:conclusion_dicussion\"\n ],\n [\n \"subsubsection\",\n \"Cross-Device FL and Cross-Silo FL\\\\label{sec:conclusion_dicussion_cross\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Cross-Device FL and Cross-Silo FL\\\\label{sec:conclusion_dicussion_cross\"\n },\n {\n \"cite_extract_rate\": 0.625,\n \"cites\": [\n 669,\n 670,\n 667,\n 671,\n 668\n ],\n \"content\": \"}\\nBased on the characteristics of data distribution, FL applications can also basically be classified as horizontal FL (where clients' data have the same feature space but different samples) or vertical FL (where clients' data share the same sample space but have different features)~. Horizontal FL is typically initiated by a service provider who wants to improve the quality of an ML application to better fit the end users' data of a specific domain. For example, by combining the statistics of users' input habits, the developers of a mobile keyboard application could achieve a more intelligent prediction of the next words typed by the users, thus enhancing their experience.\\nOn the other hand, vertical FL participants are usually a few organizations who hold data in different domains while being likely to achieve win-win cooperation by knowledge sharing. For instance, a regional e-commerce company and a bank may share the same set of clients. By incorporating the knowledge of the clients' revenue and expenditure recorded at the bank as well as the purchasing traces collected by the e-commerce, they can build a more accurate prediction model on the clients' purchasing behaviors, benefiting both of their businesses.\\nGiven that there exist diverse training workflows of vertical FL and the community has not yet achieved a consensus on the use of any one~, much discussion in this survey is biased towards horizontal FL that has a widely acknowledged architecture (\\\\cref{sec:background_federated}). Still, we can provide some insights for optimizing vertical FL training in general:\\n\\\\begin{enumerate}\\n \\\\item As the data is feature-partitioned, vertical FL needs the participation of all clients in each model update attempt. Thus, optimizations on participant selection (\\\\cref{sec:prior_selection}) are not applicable to vertical FL.\\n \\\\item In terms of communication, model-agnostic compression techniques (\\\\cref{sec:prior_configuration_size}) should still apply to plaintext variants of vertical FL such as~. However, synchronization frequency reduction (\\\\cref{sec:prior_configuration_freq}) is not possible in vertical FL, as each client does not have a complete model and cannot conduct training independently.\\n \\\\item As for local training, as all participants need processing data of the same sample space, load balancing techniques that assign a different amount of data samples to each client (\\\\cref{sec:prior_configuration_latency}) are not feasible in vertical FL. Moreover, as clients inherently have different local models, adaptive optimization and bias reduction methods (\\\\cref{sec:prior_configuration_round}) from horizontal FL do not generalize well to vertical FL. \\n \\\\item Unlike horizontal FL, the forms of aggregation in vertical FL differ significantly by the model architecture. Thus, the mentioned optimizations (\\\\cref{sec:prior_reporting}) can hardly be a generic remedy.\\n\\\\end{enumerate}\\nBesides horizontal FL and vertical FL, there recently emerges a novel collaborative learning paradigm called Federated Transfer Learning (FTL)~ which can cope with more relaxed assumptions of client data distribution using transfer learning techniques~. Specifically, it deals with the cases where two clients' data not only differ in sample space but also feature space. Toward this end, FTL learns a common feature representation between the two clients and minimizes the empirical errors in predicting one client's labels by leveraging the labels of other clients. As FTL's training workflow ensembles some variants of vertical FL~, we believe that the above-mentioned insights for vertical FL also hold for FTL.\",\n \"id\": \"23cebd68-fb90-4604-b70b-25f0eba0e52b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"02bbb984-68d7-4c50-804c-74430a998104\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Discussion\\\\label{sec:conclusion_dicussion\"\n ],\n [\n \"subsubsection\",\n \"Horizontal FL and Vertical FL\\\\label{sec:conclusion_dicussion_horizontal\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Horizontal FL and Vertical FL\\\\label{sec:conclusion_dicussion_horizontal\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nIn this survey, we focus on system optimization in synchronous federated training and propose a natural taxonomy that categorizes existing solutions based on both the training phase and the type of utility at which they target. Apart from problem-driven attempts, we also include related cornerstone efforts including measurement studies and benchmarking suites. We expect this manuscript to be a \\\\textit{useful guideline} for the design and implementation of federated learning systems.\\n\\\\ifCLASSOPTIONcompsoc\\n \\\\section*{Acknowledgments}\\n\\\\else\\n \\\\section*{Acknowledgment}\\n\\\\fi\\nThe research was supported in part by RGC RIF grant R6021-20, and RGC GRF grants under the contracts 16209120, 16200221, and 16213120. We also thank the anonymous reviewers for their valuable feedback.\\n\\\\ifCLASSOPTIONcaptionsoff\\n \\\\newpage\\n\\\\fi\\n\\\\Urlmuskip=0mu plus 1mu\\\\relax\\n\\\\bibliographystyle{IEEEtran}\\n\\\\bibliography{./ref.bib}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/zjiangaj.jpg}}]{Zhifeng Jiang}\\nreceived the BEng (Hons.) degree from the Department of Computer Science and Technology at Zhejiang University (ZJU) in 2019. Since then, he has been working towards a Ph.D. degree in the Department of Computer Science and Engineering, Hong Kong University of Science and Technology (HKUST). He is mainly interested in machine learning systems, with a special focus on efficient and scalable federated learning. He was a recipient of the Best Paper Runner-up Award at IEEE ICDCS 2021. He served on the artifact evaluation committees of ACM SOSP 2021, USENIX OSDI 2022, and USENIX ATC 2022.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/weiwa.jpeg}}]{Wei Wang} (Member, IEEE) received his B.Engr. and M.Engr. degrees from the Department of Electrical Engineering, Shanghai Jiao Tong University, China, in 2007 and 2010, respectively and his Ph.D. degree from the Department of Electrical and Computer Engineering, University of Toronto, Canada, in 2015. Since 2015, he has been with the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology (HKUST), where he is currently an Associate Professor. He is also affiliated with the HKUST Big Data Institute. Dr. Wang's research interests cover the broad area of distributed systems, with focus on serverless computing, machine learning systems, and cloud resource management. He published extensively in the premier conferences and journals of his fields. His research has won the Best Paper Runner Up awards of IEEE ICDCS 2021 and USENIX ICAC 2013.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/bli.jpg}}]{Bo Li} is a Chair Professor in Department of Computer Science and Engineering, Hong Kong University of Science and Technology. He was a Cheung Kong Scholar Visiting Chair Professor in Shanghai Jiao Tong University (2010-2016), and was the Chief Technical Advisor for ChinaCache Corp. (NASDAQ:CCIH), a leading CDN provider. He made pioneering contributions in multimedia communications and the Internet video broadcast, which attracted significant investment and received the Test-of-Time Best Paper Award from IEEE INFOCOM (2015). He received 6 Best Paper Awards from IEEE including IEEE INFOCOM (2021). He was the Co-TPC Chair for IEEE INFOCOM (2004).\\nHe is a Fellow of IEEE. He received his PhD in the ECE Department, University of Massachusetts at Amherst, and his B. Eng. (summa cum laude) in the Computer Science from Tsinghua University, Beijing, China.\\n\\\\end{IEEEbiography}\\n\\\\vfill\\n\\\\newpage\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./biophoto/qyang.jpeg}}]{Qiang Yang} received the B.Sc. degree in astrophysics from Peking University, Beijing, China, in 1982, and the Ph.D. degree in computer science and the M.Sc. degree in astrophysics from the University of Maryland at College Park, College Park, MD, USA, in 1985 and 1989, respectively.\\nHe was a Faculty Member with the University of Waterloo, Waterloo, ON, Canada, from 1989 to 1995, and Simon Fraser University, Burnaby, BC, Canada, from 1995 to 2001. He was the Founding Director of the Noah’s Ark Laboratory, Huawei, Hong Kong, from 2012 to 2014, and a Co-Founder of 4Paradigm Corporation, Beijing, an AI platform company. He is currently the Head (Chief AI Officer) with the AI Department, WeBank, Shenzhen, China, and the Chair Professor with the Department of Computer Science and Engineering (CSE), The Hong Kong University of Science and Technology, Hong Kong, where he was the Former Head of the Department of CSE and the Founding Director of the Big Data Institute, Hong Kong, from 2015 to 2018. He has authored several books, including Intelligent Planning (Springer), Crafting Your Research Future (Morgan and Claypool), and Constraint-Based Design Recovery for Software Engineering (Springer). His research interests include artificial intelligence, machine learning, and data mining, with an emphasis on transfer learning, automated planning, federated learning, and case-based reasoning.\\nDr. Yang is a fellow of several international societies, including the ACM, AAAI, IAPR, and AAAS. He served as an Executive Council Member for the Association for the Advancement of AI from 2016 to 2020 and the President for the International Joint Conference on AI from 2017 to 2019. He was a recipient of several awards, including the 2004/2005 ACM KDDCUP Championship, the AAAI Innovative AI Applications Award in 2016, and the ACM SIGKDD Distinguished Service Award in 2017. He was the Founding Editor-in-Chief of the ACM Transactions on Intelligent Systems and Technology and the IEEE TRANSACTIONS ON BIG DATA.\\n\\\\end{IEEEbiography}\\n\\\\end{document}\",\n \"id\": \"a4f483b3-14e7-4959-a5a0-300b337d0549\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"24e6f906-e64a-4296-a2e5-eb81d7da883e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Towards Efficient Synchronous\\\\\\\\Federated Training: A Survey on\\\\\\\\System Optimization Strategies\"\n ],\n [\n \"section\",\n \"Related Work and Concluding Remarks\\\\label{chap:conclusion\"\n ],\n [\n \"subsection\",\n \"Conclusion\\\\label{sec:conclusion_conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\\\\label{sec:conclusion_conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":80,"string":"80"},"all_cites":{"kind":"list like","value":[585,589,584,581,587,580,583,586,7262,582,7261,588,590,591,593,595,596,600,598,599,594,592,7038,597,3469,618,617,608,605,602,603,606,613,616,612,607,610,614,604,611,615,601,4338,609,619,620,626,628,624,622,621,625,8366,627,623,630,629,631,632,634,638,635,640,636,639,641,633,637,643,646,644,645,647,7039,642,8367,648,649,652,653,650,651,656,655,41,7199,654,661,657,658,8368,659,660,662,664,663,666,665,669,670,667,671,668],"string":"[\n 585,\n 589,\n 584,\n 581,\n 587,\n 580,\n 583,\n 586,\n 7262,\n 582,\n 7261,\n 588,\n 590,\n 591,\n 593,\n 595,\n 596,\n 600,\n 598,\n 599,\n 594,\n 592,\n 7038,\n 597,\n 3469,\n 618,\n 617,\n 608,\n 605,\n 602,\n 603,\n 606,\n 613,\n 616,\n 612,\n 607,\n 610,\n 614,\n 604,\n 611,\n 615,\n 601,\n 4338,\n 609,\n 619,\n 620,\n 626,\n 628,\n 624,\n 622,\n 621,\n 625,\n 8366,\n 627,\n 623,\n 630,\n 629,\n 631,\n 632,\n 634,\n 638,\n 635,\n 640,\n 636,\n 639,\n 641,\n 633,\n 637,\n 643,\n 646,\n 644,\n 645,\n 647,\n 7039,\n 642,\n 8367,\n 648,\n 649,\n 652,\n 653,\n 650,\n 651,\n 656,\n 655,\n 41,\n 7199,\n 654,\n 661,\n 657,\n 658,\n 8368,\n 659,\n 660,\n 662,\n 664,\n 663,\n 666,\n 665,\n 669,\n 670,\n 667,\n 671,\n 668\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.9854534340463341,"string":"0.985453"}}},{"rowIdx":46,"cells":{"authors":{"kind":"list like","value":["Ji Liu","Jizhou Huang","Yang Zhou","Xuhong Li","Shilei Ji","Haoyi Xiong","Dejing Dou"],"string":"[\n \"Ji Liu\",\n \"Jizhou Huang\",\n \"Yang Zhou\",\n \"Xuhong Li\",\n \"Shilei Ji\",\n \"Haoyi Xiong\",\n \"Dejing Dou\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"From Distributed Machine Learning to Federated Learning: A Survey"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-04-29T14:15:11Z","string":"2021-04-29T14:15:11Z"},"category":{"kind":"string","value":"cs.DC"},"abstract":{"kind":"string","value":" In recent years, data and computing resources are typically distributed in the devices of end users, various regions or organizations. Because of laws or regulations, the distributed data and computing resources cannot be \\liu{aggregated or} directly shared among different regions or organizations for machine learning tasks. Federated learning emerges as an efficient approach to exploit distributed data and computing resources, so as to collaboratively train machine learning models\\liu{. At the same time, federated learning obeys} the laws and regulations and ensures data security and data privacy. In this paper, we provide a comprehensive survey of existing works for federated learning. \\liu{First, we propose a functional architecture of federated learning systems and a taxonomy of related techniques. Second, we explain the federated learning systems from four aspects: diverse types of parallelism, aggregation algorithms, data communication, and the security of federated learning systems. Third, we present four widely used federated systems based on the functional architecture. Finally, we summarize the limitations and propose future research directions.} \\keywords{Federated learning \\and Distributed system \\and Parallel computing \\and Security, Privacy} "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"]],"subsections":["bda39eab-e9c6-47ab-8897-26d79998749b","2371347b-6739-4c7c-b6c1-b23b3b02d8ed","5dfeced7-075b-4276-abd0-1ff3e9bb48a7","cd0b3f18-6523-488f-aaa8-addaf533ff9f","727148cb-019c-4e5c-9d02-e8f0c8fb671c","226ea081-07c1-4d87-ad04-18733a7f7202","4ff5262c-b6d9-4162-aba6-d016f80c7306"],"title":"root"},{"cite_extract_rate":0.47222222222222204,"cites":[3469,1315,5419,3477,591,8917,659,5422,5423,1311,5424,671,627,5420,600,582,5421],"content":"With billions of connected Internet of Things (IoT) devices , smartphones and large websites around the world, in recent years, we have witnessed huge amounts of data generated and dispersed over various mobile devices of end users, or the data centers of different organizations. As the data contain sensitive information of end users or organizations, such as facial images, location-based services, health information , or personal economic status, moving the raw data from personal devices or data centers of multiple organizations to a centralized server or data center may pose immediate or potential information leakage. Due to the concerns of data security and data privacy, legal restrictions, such as the Cybersecurity Law of the People’s Republic (CLPR) of China , the General Data Protection Regulation (GDPR) in European Union, the Personal Data Protection Act (PDP) in Singapore, the California Consumer Privacy Act (CCPA) , and the Consumer Privacy Bill of Rights (CPBR) in the United States, have been introduced and put in practice, which makes data aggregation from distributed devices, multiple regions, or organizations, almost impossible . In addition, computing and storage resources are also typically distributed in multiple regions and organizations , which cannot be aggregated in a single data center.\nFederated Learning (FL) emerges as an efficient approach to exploit the distributed resources to collaboratively train a machine learning model. FL is a distributed machine learning approach where multiple users collaboratively train a model, while keeping the raw data decentralized without being moved to a single server or data center .\nFL not only exploits the distributed resources to efficiently carry out the training process of machine learning, but also promises to provide security and privacy for the decentralized raw data. Within FL, the raw data, or the data generated based on the raw data with security processing, serves as the training data. FL only allows the intermediate data to be transferred among the distributed computing resources while avoiding the transfer of training data. \nThe distributed computing resources refer to mobile devices of end users or servers of multiple organizations.\nFL brings the code to the data, instead of bringing the data to the code, and it addresses the fundamental problems of privacy, ownership, and locality of data .\nIn this way, FL can enable multiple users to train a model while satisfying the legal data restrictions.\nTraditional centralized machine learning approaches typically gather the distributed raw data generated on different devices or organizations to a single server or a cluster with shared data storage, which may bring serious data privacy and security concerns . \nThe centralized approaches, in general, are associated with diverse challenges, including computational power and training time, and most importantly, security and privacy with respect to distributed data .\nFL differs from the centralized approach in three aspects. First, FL does not allow direct raw data communication, while the centralized approach has no restriction. Second, FL exploits the distributed computing resources in multiple regions or organizations, while the centralized approach generally only utilizes a single server or a cluster in a single region, which belongs to a single organization. Third, FL generally takes advantage of encryption or other defense techniques to ensure the data privacy or security, while the centralized approach pays little attention to these security issues .\nThe term ``federated learning'' was first introduced in 2016 , which focuses on the unbalanced and non-Independent and Identically Distributed (non-IID) data in mobile devices. \nThe concept of FL was extended to three data scenarios, i.e., horizontal, vertical, and hybrid . \nThe horizontal FL addresses the decentralized data of the same features, while the identifications are different. The vertical FL handles the decentralized data of the same identifications with different features. The hybrid FL deals with the data of different identifications and different features. Then, FL is formally defined as a machine learning approach where multiple clients collaborate in solving a machine learning problem while the raw data is stored locally and is neither exchanged nor transferred . \nAn FL system is an efficient tool to carry out FL with decentralized data and resources. Several open-source FL systems, e.g., FATE , PaddleFL , TensorflowFL , and Pysyft , are now intensively used by both research communities, e.g., healthcare , and computer visions , and by industrial groups, e.g., WeBank . Although various FL systems exist, the architecture of FL systems has common features: In particular, they share the capability to collaboratively train a machine learning model. Most FL systems are composed of four layers, i.e., presentation, user services, FL training, and infrastructure. These four layers enable FL system users to design, execute, and analyze machine learning models with distributed data.\nAlthough FL differs from the centralized machine learning approaches, it not only utilizes novel techniques designed for FL, but also takes advantage of the techniques designed for distributed machine learning. \nFL exploits parallelization techniques designed for distributed machine learning.\nFor instance, horizontal FL exploits the data parallelism, which trains multiple instances of the same model on different subsets of the training dataset . Vertical FL utilizes model parallelism to distribute parallel paths of a single model to multiple devices in order to handle the data of different features . \nMultiple aggregation algorithms are proposed to aggregate the models in distributed computing resources.\nData transfer techniques are also utilized in FL, e.g., model compression . \nAs FL promises to provide data security and data privacy, diverse defense techniques, e.g., differential privacy , homomorphic encryption , and Robustness Aggregation , are designed to address the possible attacks .\nThere have been a few surveys of FL. Some works provide a comprehensive study of FL, from the taxonomy of FL to the techniques, e.g., the efficiency, data privacy, security, and applications of FL. Some surveys focus on the data privacy and security of FL. Other surveys present the application of FL in a specific area, e.g., healthcare informatics , mobile edge networks , and neural architecture searches , and they personalize global models to work better for individual clients . However, few of them present the architecture of FL or analyze parallelization techniques in FL.\nIn this paper, we provide a survey of federated learning and the related parallelization techniques. The main contributions of this paper are:\n\\begin{itemize}\n \\item A four-layer FL system architecture, which is useful for discussing the techniques for FL. This architecture can also be a baseline for other work and can help with the assessment and comparison of FL systems.\n \\item A taxonomy of FL-related techniques, including the parallelization techniques, the aggregation algorithms, and the techniques for data communication and security, with a comparative analysis of the existing solutions.\n \\item A discussion of research issues to improve the efficiency and security of FL systems.\n\\end{itemize}\nThis paper is organized as follows. Section \\ref{sec:overview} gives an overview of the execution of FL, including the FL system architectures and basic functional architecture of FL systems. Section \\ref{sec:pexec} focuses on the techniques used for distributed training of FL and aggregation methods. Section \\ref{sec:security} presents the techniques for distributed execution, data communication and data security of FL. Section \\ref{sec:frameworks} demonstrates the existing FL frameworks.\nSection \\ref{sec:future} discusses the open issues raised for the execution of FL with distributed resources. Section \\ref{sec:con} summarizes the main findings of this study.","id":"bda39eab-e9c6-47ab-8897-26d79998749b","level":"section","origin_cites_number":36,"parent_id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:overview}\nIn this section, we introduce the basic concepts of federated learning. Then, we present the life cycle of FL models. Afterwards, we detail the functional architecture and the corresponding functionality of FL systems.","id":"2371347b-6739-4c7c-b6c1-b23b3b02d8ed","level":"section","origin_cites_number":0,"parent_id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"]],"subsections":["4f50be54-11aa-4ae8-ae5e-fb6d0e6797e6","cd1b4108-9163-49ab-bfbc-9fe67cc3a340","78f72b80-7c28-4237-9e16-d97c8c144603"],"title":"An Overview of Federated Learning"},{"cite_extract_rate":0.37037037037037,"cites":[1315,591,5426,659,5425,671,600,166,582],"content":"Machine learning is the process to automatically extract the models or patterns from data . \nThe models or patterns are expressed as machine learning models.\nA machine learning model is an ensemble of a model structure, which is typically expressed as a Directed Acyclic Graph (DAG), data processing units, e.g., activation functions in Deep Neural Networks (DNNs), and the associated parameters or hyper-parameters. \nThe input data can be processed through a machine learning model to generate the output, e.g., the prediction results or the classification results, which is the inference process.\nThe machine learning model is generated based on the training data, which is the training process.\nDuring the training process, the parameters or the model structure of the machine learning model is adjusted based on a training algorithm in order to improve the performance, e.g., the accuracy or the generalization capacity. The training algorithm is also denoted by machine learning algorithms. The duration of the training process is training time.\nAccording to whether the training data have labels, the training process of machine learning can be classified into four types , i.e., supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Supervised learning represents that a machine learning task exploits the training data composed of input features and the corresponding labels . In this paper, we focus on this type of training data. For instance, each data point in the training dataset contains $(x,y)$, where $x$ represents the input features and $y$ represents the desired output value. Unsupervised learning represents that a machine learning task exploits the training data, which only consists of input features without output values; i.e., each data point only contains $x$ and does not have $y$. Semi-supervised learning represents that one (generally small) part of the training data contains output values, while the other (generally small) part of the training does not. Reinforcement learning represents that each iteration in the training process considers its observation of the environment from the last iteration. \nWhile the training data become huge, e.g., on the order of terabyte , or when the training data is inherently distributed or too big to store on single machines , the training process is carried out using distributed resources, which is distributed machine learning. One of the important features of the distributed machine learning is that it can significantly accelerate the training speed so as to reduce the training time. Diverse parallelization techniques are used in distributed machine learning. For instance, Graphics Processing Units (GPUs) using Single Instruction Multiple Data (SIMD) and Tensor Processing Units (TPUs) using Multiple Instructions Multiple Data (MIMD) are exploited . In addition, distributed machine learning takes advantage of three types of parallelism to parallelize the training process, i.e., data parallelism , model parallelism , and pipeline parallelism . With the data parallelism approach, the training data is partitioned as many times as the number of computing resources, and all computing resources subsequently apply the same machine learning algorithm to process different chunks of the data sets . With the model parallelism approach, exact copies of the entirety of the data (the training data or the intermediate data)\nare processed by each computing resource, each of which exploits different parts of the machine learning model . The pipeline parallelism approach combines the data parallelism and the model parallelism. With this approach, each computing resource processes a part of the training data with a part of the machine learning model, while the processing, e.g., computation or communication, at each node can be parallelized . \nFL is a distributed machine learning approach where multiple users collaboratively train a model, while keeping the raw data distributed without being moved to a single server or data center .\nThe model used for FL is denoted by FL model.\nFL is first proposed to handle the unbalanced and non-Independent and Identically Distributed (non-IID) data of the same features in mobile devices .\nThen, the concept of FL is extended to the distributed data of diverse features in multiple organizations or various regions .\nFL systems are used within one or multiple phases of the life cycle of FL models. An FL system is a distributed system to manage the distributed training process with distributed resources.\nFL is a special type of distributed machine learning, which differs from other distributed machine learning approaches in the following three points.\nFirst, FL does not allow direct raw data communication, while other approaches have no restriction. \nAs the raw data are of multiple ownerships, FL approaches with this restriction can meet the requirements defined by the related laws, e.g., CLPR , GDPR , PDPA , CCPA , and CPBR .\nIn particular, the consent (GDPR Article 6) and the data minimalization principle (GDPR Article 5) limit data collection and storage to only what is consumer-consented and what is absolutely necessary for processing .\nSecond, FL exploits the distributed computing resources in multiple regions or organizations, while the other approaches generally only utilize a single server or a cluster in a single region, which belongs to a single organization. FL enables the collaboration among multiple organizations.\nThird, FL generally takes advantage of encryption or other defense techniques to ensure the data privacy or security, while the other approaches pay little attention to this security issue . FL promises to ensure the privacy and security of the raw data, as the leakage of information may incur significant financial and reputational losses. \nDuring the training process of FL, an optimization problem is solved as shown in Formula \\ref{eq:problem}. Given $n$ training dataset $\\mathcal{D} = {D_1, D_2, ..., D_n}$, where each data point $(x,y)\\sim\\mathcal{D}$, the problem of FL is to learn a function $\\widehat F$ from all possible hypotheses $\\mathcal{H}$, while minimizing the expectation of loss over the distribution of all the dataset $\\mathcal{D}$. \n\\begin{equation}\n\\label{eq:problem}\n \\widehat F = \\underset{F\\in\\mathcal{H}}{\\mathrm{argmin}} \\underset{(x,y)\\in\\mathcal{D}}{\\mathbb{E}} L(y, F(x)),\n\\end{equation}\nwhere $L(y, F(x))$ refers to the loss of $F(x)$ to the label $y$. \nDuring the training process, the Stochastic Gradient Descent (SGD) approach is generally used to minimize the loss function using Formula \\ref{eq:gd}.\n\\begin{equation}\n\\label{eq:gd}\n F_{k+1}(x)\\gets F_k(x) - \\eta_k \\nabla F_k(x),\n\\end{equation}\nwhere $F_k(x)$ refers to the learned model in the $k^{th}$ iteration, $\\nabla F_k(x)$ refers to the gradient of the model at the $k^{th}$ iteration based on the model already obtained $F_k(x)$ and the training dataset, $\\eta_k$ refers to the learning rate, and $F_{k+1}(x)$ refers to the update model of the $k^{th}$ iteration. Within each iteration, there are two phases, i.e., forward propagation and backward propagation. The forward propagation calculates the output based on the input data $x$ using the model, while the backward propagation calculates the gradients $\\nabla F_k(x)$ and updates the model.\nWhen the calculation is distributed among multiple computing resources, the gradients or models of each computing resource are aggregated using an aggregation algorithm (see details in Section \\ref{sebsec:aggre}), in order to achieve consensus of multiple models and to generate a global model. \\liu{The learning rate can be dynamically adapted using a local adaptive optimizer, e.g., Adam, and/or cross-round learning rate schedulers .}","id":"4f50be54-11aa-4ae8-ae5e-fb6d0e6797e6","level":"subsection","origin_cites_number":27,"parent_id":"2371347b-6739-4c7c-b6c1-b23b3b02d8ed","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"],["subsection","Basic Concepts"]],"subsections":[],"title":"Basic Concepts"},{"cite_extract_rate":0.8,"cites":[591,612,671,5427],"content":"The life cycle of an FL model is a description of the state transitions of an FL model from creation to completion . Lo \\textit{et al.} propose that the life cycle of an FL model consists of 8 phases: initiated, broadcast, trained, transmitted, aggregated, evaluated, deployed, and monitored. Kairouz \\textit{et al.} propose that the life cycle of an FL model includes 6 phases: problem identification, client instrumentation, simulation prototyping, federated model training, model evaluation, and deployment. However, they focus on the FL with distributed data in mobile devices. In this paper, we adopt a combination of workflow life cycle views with a few variations , condensed into four phases:\n\\begin{enumerate}\n \\item The composition phase is for the creation of an FL model, which is used to address a specific machine learning problem, e.g., classification. First, a machine learning model is created to address the problem with certain requirements, e.g., the requirement of accuracy. Then, the machine learning model is adapted to FL scenarios. For instance, if the distributed data is of different features, the machine learning model is partitioned (see details in Section \\ref{subsubsec:modelP}) to process the distributed data. \n \\item The FL training phase is for the training phase of the FL model. During this phase, a training strategy, which includes parallelism and aggregation algorithms (see details in Section \\ref{sec:pexec}), is used to update the parameters, hyper parameters, and even the structure of the network, in order to improve the accuracy and the generalization capacity of the FL model. \n \\item The FL model evaluation phase is to apply the trained FL models, in order to analyze the performance of the trained FL models on a simulation platform or a real distributed system . As a result, the FL models with the best performance are selected. If the FL models do not meet the requirements, the FL model should be modified, or the training phase should be carried out again.\n \\item The FL model deployment phase is to deploy the FL model in a real-life scenario to process the data. If the final model can be shared without restriction, there is no difference between the FL model deployment and the model generated from a traditional centralized approach. Otherwise, the deployment of the final model should consider the ownership of the corresponding parts. \n\\end{enumerate}","id":"cd1b4108-9163-49ab-bfbc-9fe67cc3a340","level":"subsection","origin_cites_number":5,"parent_id":"2371347b-6739-4c7c-b6c1-b23b3b02d8ed","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"],["subsection","FL Model Life Cycle"]],"subsections":[],"title":"FL Model Life Cycle"},{"cite_extract_rate":0,"cites":[],"content":"The functional architecture of an FL system can be layered as follows : presentation, user services, FL training, and infrastructure. Figure \\ref{fig:layers} shows this architecture. The higher layers exploit the lower layers to provide their own functionality. A user interacts with an FL system through the presentation layer and realizes independent functionalities at the user services layer. During the training phase of FL models, a Federated Learning Execution Plan (FLEP) is generated, and the corresponding distributed training is carried out at the FL training layer. The FLEP is composed of a type of parallelism, a scheduling strategy, and a fault-tolerance mechanism. The FL system manages the physical resources through the infrastructure layer for the distributed training. \n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.8\\textwidth]{figures/layers.pdf}\n \\caption{\\liu{Functional architecture of an FL system.}}\n \\label{fig:layers}\n\\end{figure}","id":"78f72b80-7c28-4237-9e16-d97c8c144603","level":"subsection","origin_cites_number":1,"parent_id":"2371347b-6739-4c7c-b6c1-b23b3b02d8ed","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"],["subsection","Functional Architecture of FL Systems"]],"subsections":["afff657f-2ca3-4a0d-b609-6427f5eee76c","8ca8a082-7841-461b-ad59-49e7d8fef8fe","6af71683-cb7e-4e75-87a8-c21ef2982fdc","cfb855a7-c817-4897-9235-30d180850302"],"title":"Functional Architecture of FL Systems"},{"cite_extract_rate":0.2,"cites":[656],"content":"The presentation layer is a User Interface (UI) for the interaction between Users and FL systems at one or multiple stages of the FL model life cycle. The UI can be textual or graphical. This interface is responsible for designing a new FL model or choosing an existing machine learning model as an FL model. In addition, this layer also supports the modules at the user services layer, e.g., shows the status of the distributed training process. The textual UI is largely used for designing FL models \\liu{based on the command line or scripts }. The models can be directly expressed using Python, with the textual interface in PaddleFL , TensorFlowFL , PySyft , and FATE . A graphic UI can make the interaction more practical, while the users can drag or drop the data processing element to design an FL model. For instance, FATE provides a Graphic UI (GUI) through a web portal. However, the graphic portal also exploits textual programming languages as inner representations of an FL model.","id":"afff657f-2ca3-4a0d-b609-6427f5eee76c","level":"subsubsection","origin_cites_number":5,"parent_id":"78f72b80-7c28-4237-9e16-d97c8c144603","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"],["subsection","Functional Architecture of FL Systems"],["subsubsection","Presentation Layer"]],"subsections":[],"title":"Presentation Layer"},{"cite_extract_rate":0.8888888888888881,"cites":[5429,591,5434,8918,5437,5435,5433,5432,7927,5428,5430,7926,8502,5436,5438,5431],"content":"The user service layer supports the expected functionalities, i.e., \\liu{monitoring and steering and log; interpretability and explainability; and graph data.} The monitoring enables the users to get the real-time status of the distributed training process. As the training process of FL models can be very long, e.g., from several hours to days , it is of much importance to track the execution status, which allows the user to verify whether the training proceeds normally. The log service is generally supported by major FL systems, which can be used to analyze the training process. In addition, the log generated during the training process can be used to debug the system or adjust the FL model. FATE provides a visual monitoring board to users through its GUI. When there are unexpected results or errors during the training process, steering enables users to adjust the training process in order to reduce the time necessary to carry out the distributed training from scratch. Most FL systems can enable the users to stop the training, while the adjustment of parameters is not fully supported by major FL systems. The interpretability of FL is to describe the internals of an FL system in a way that is understandable to humans . The explainability focuses on explaining the representation of data inside an FL model . With interpretability and explainability, the FL system can \\liu{provide} a description of the results of the trained FL model based on the training data and the distributed training process. Shapley values have been used to provide the interpretability , while both the interpretability and explainability remain open challenges as each is hard to fully support. \n\\liu{In the real world, graph data widely exist in multiple domains, and a bunch of FL approaches have been proposed to handle the decentralized graph data for community detection , financial crime , and especially knowledge graph completion .\nFL is particularly useful in the field of knowledge graph completion, as a knowledge graph could not only contain text but also images or other type of data, i.e., multimodal Knowledge Graphs ; and the completion is realized in a collaborative fashion within an FL system .\nThe decentralized graphs can be inter-graph, i.e., the decentralized data belongs to multiple graphs, or intra-graph, i.e., the decentralized data is within one big graph , while most of the existing works focus on the intra-graph situation.\nHorizontal FL techniques can be exploited on the Graph Neural Networks (GNN) with encryption techniques (see details in Section \\ref{subsubsec:infrastructure}) while the performance of FL may be much worse than that of centralized GNNs .\nThe aggregation algorithms (see details in Section \\ref{subsubsec:centralized}) are also proposed based on the FedAvg or optimal transportation . \nIn addition, decentralized aggregation algorithms (see details in Section \\ref{subsubsec:decentralized}) are also proposed to deal with the decentralized graph data for social network and drug discovery .\nWhile the fine-tuning of the FL system is time-consuming, Bayesian optimization and evolutionary optimization strategies are utilized to automatically tune the hyper-parameters and the network structure, respectively. \nGraph data can be vertically distributed where the features of the nodes are distributed across multiple data owners, and a data owner may or may not have the edges. Vertical FL exploits embeddings or autoencoders to represent the nodes, which can be transferred to train a GNN. In addition, the differential privacy (see details in Section \\ref{subsubsec:infrastructure}) is combined with the embeddings to protect the data privacy of knowledge graphs. }","id":"8ca8a082-7841-461b-ad59-49e7d8fef8fe","level":"subsubsection","origin_cites_number":18,"parent_id":"78f72b80-7c28-4237-9e16-d97c8c144603","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"],["subsection","Functional Architecture of FL Systems"],["subsubsection","User Services Layer"]],"subsections":[],"title":"User Services Layer"},{"cite_extract_rate":0.5,"cites":[590],"content":"The FL training layer carries out the distributed training process with distributed data and computing resources. This layer consists of three modules: parallelization, scheduling, and fault-tolerance. FL parallelization exploits diverse types of parallelism, e.g., data parallelism, model parallelism, and pipeline parallelism, to generate executable tasks. Through the FL scheduling module, an FL system produces a Scheduling Plan (SP) of executable tasks, which aims at fully exploiting distributed computing resources and preventing training stalling. During the training process, the SP is generally defined by a training algorithm, which aggregates the updates, i.e., gradients or models, from each computing resource in order to generate a final machine learning model. The FL fault-tolerance mechanism handles the failures or errors of task execution and the connection of distributed resources. Reactive approaches are generally exploited, e.g., using check-points, restart, and task replication . A reactive approach reduces the effect of failures after perceiving failures . An FLEP, which captures the execution directives, typically the result of compiling and optimizing the training process of FL models, is generated at this layer.","id":"6af71683-cb7e-4e75-87a8-c21ef2982fdc","level":"subsubsection","origin_cites_number":2,"parent_id":"78f72b80-7c28-4237-9e16-d97c8c144603","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"],["subsection","Functional Architecture of FL Systems"],["subsubsection","FL Training Layer"]],"subsections":[],"title":"FL Training Layer"},{"cite_extract_rate":0.6666666666666661,"cites":[617,7608],"content":"\\label{subsubsec:infrastructure}\nThe infrastructure layer provides the interaction between an FL system and the distributed resources, including the computing resources, storage resources, network resources, and data resources. This layer contains three modules: a data security module, a data transfer module, and a distributed execution module. The data security module generally exploits Differential Privacy (DP) and encryption techniques, e.g., homomorphic , to protect the raw data used during the training process. Although the raw data cannot be directly transferred, intermediate data, e.g., the gradients or models, can be communicated among distributed computing resources. The data transfer module exploits data compression techniques to improve the data transfer efficiency. \nAt this layer, the FLEP generated at the FL training layer is carried out within the distributed execution module; i.e., concrete tasks are executed in distributed computing resources.","id":"cfb855a7-c817-4897-9235-30d180850302","level":"subsubsection","origin_cites_number":3,"parent_id":"78f72b80-7c28-4237-9e16-d97c8c144603","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","An Overview of Federated Learning"],["subsection","Functional Architecture of FL Systems"],["subsubsection","Infrastructure Layer"]],"subsections":[],"title":"Infrastructure Layer"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:pexec}\nIn this section, we present the distributed training process for FL. First, we present three types of parallelism in distributed training and the application within FL. The parallelism approaches are generally implemented in the parallelization module. Then, we discuss existing aggregation algorithms for the distributed training, which is implemented in the scheduling module.","id":"5dfeced7-075b-4276-abd0-1ff3e9bb48a7","level":"section","origin_cites_number":0,"parent_id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"]],"subsections":["9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1","10053136-8ab4-4b06-aec3-a7a9b0e56bc3"],"title":"Distributed Training"},{"cite_extract_rate":0.6000000000000001,"cites":[600,1315,671],"content":"Three types of parallelism exist for distributed machine learning: data parallelism, model parallelism, and pipeline parallelism . FL can be classified to three types, i.e., horizontal, vertical, and hybrid . The horizontal FL generally exploits data parallelism, and the vertical FL typically takes advantage of model parallelism. However, the hybrid FL relies on transfer learning , which is not a parallelism approach and is out of the scope of this paper. \n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.65\\textwidth]{figures/dnn.pdf}\n \\caption{An example of a neural network.}\n \\label{fig:dnn}\n\\end{figure}\nIn this section, we take an example of a neural network as shown in Figure \\ref{fig:dnn} to explain the parallelism. In the example, we assume that the model contains three layers and seven data processing nodes (neurons), i.e., $A_1$, $A_2$, $B_1$, $B_2$, $B_3$, $C_1$, $C_2$, $D$. The arrows represent the data flow among different data processing nodes. The execution of the data processing nodes at each layer can be carried out in parallel, while the execution of different layers should be performed sequentially. The input data contains 4 data points. We assume two/three computing resources owned by two/three users. Each has a part of the input data.\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.6\\textwidth]{figures/datap.pdf}\n \\caption{Data parallelism. The forward and backward process of $I^1$ and $I^2$ is performed in computing resource 1, while that of $I^3$ and $I^4$ is performed in computing resource 2 at the same time. Then the model or gradient is transferred to calculate an average model or gradient to be sent to each computing resource for the following training.}\n \\label{fig:datap}\n\\end{figure}","id":"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1","level":"subsection","origin_cites_number":5,"parent_id":"5dfeced7-075b-4276-abd0-1ff3e9bb48a7","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Parallelism \\& FL Types"]],"subsections":["0430fcfa-cd31-4b1c-bf51-10d342c0d942","62c91cb4-aacb-47de-82b2-8f4949518fdc","1b1418f9-bcb1-4bbe-b99b-80b182bd318e"],"title":"Parallelism \\& FL Types"},{"cite_extract_rate":0.6000000000000001,"cites":[671,582,591],"content":"Data parallelism is realized by having the data processing performed in parallel at different computing resources, with the same model, on different data points. As shown in Figure \\ref{fig:datap}, data parallelism is exploited when the ensemble of data points is distributed among different computing resources. During the training process of FL, the training data is not transferred among different computing resources, while the intermediate data, e.g., the models or the gradients $\\nabla F_k(x)$ in Formula \\ref{eq:gd}, are transferred. \nThe data in each computing resource can be Independent and Identically Distributed Data (IID) or non-IID. FL focuses on the non-IID , while other distributed machine learning approaches mainly focus on IID data. \nWith the data parallelism, the FL is horizontal , i.e., the data and the calculation are horizontally distributed among multiple computing resources.\nIn addition, this parallelism generally corresponds to the cross-device FL , where a large number of devices (mobiles or edge devices) collaboratively participate in training a single global model to have good accuracy.\nWhen the number of devices is small, e.g., 2-100, and the computing resources are from diverse organizations, this parallelism also corresponds to cross-silo FL . In addition to the general data-parallel schemes for federated learning, some specific privacy-preserved distributed statistical tricks have been invented for federated sparse models~.\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.75\\textwidth]{figures/modelp.pdf}\n \\caption{Model parallelism. The dashed arrows represent inter-computing resource communication. For each input data point $I$, different parts, i.e., $I_{A_1}$ and $I_{A_2}$, are distributed at different computing resources.}\n \\label{fig:modelp}\n\\end{figure}","id":"0430fcfa-cd31-4b1c-bf51-10d342c0d942","level":"subsubsection","origin_cites_number":5,"parent_id":"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Parallelism \\& FL Types"],["subsubsection","Data Parallelism"]],"subsections":[],"title":"Data Parallelism"},{"cite_extract_rate":0.625,"cites":[600,5439,671,668,591],"content":"\\label{subsubsec:modelP}\nModel parallelism is realized by having independent data processing nodes distributed at different computing resources, so as to process the data points of specific features. \nTwo data processing nodes can be either independent, i.e., the execution of any node does not depend on the output of the other one; or dependent, i.e., there is a data dependency between them .\nAs shown in Figure \\ref{fig:modelp}, model parallelism is achieved when different parts of each data point are distributed at different computing resources. \nFor instance, the data process on Node $A_1$ and that of $A_2$ can be carried out in parallel.\nWith the model parallelism, vertical FL, where the data points and calculation are vertically distributed among multiple computing resources , is realized.\nIn this case, the original model needs to be partitioned to be distributed at different computing resources.\nTwo organizations generally apply this type of FL when each organization owns parts of the features of users and they would like to collaboratively train a model using the data of all the features, which corresponds to cross-silo FL .\nMost studies of vertical federated learning only support two parties (with or without a central coordinator) . \nFor instance, SecureGBM is proposed to train a tree-based Gradient Boosting Machine (GBM).\nIn order to support multiple parties, the idea of multi-view learning is exploited in a multi-participant, multi-class vertical federated learning framework .\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.7\\textwidth]{figures/pipelinep.pdf}\n \\caption{Pipeline parallelism. The dashed arrows represent inter computing resource communication.}\n \\label{fig:pipelinep}\n\\end{figure}","id":"62c91cb4-aacb-47de-82b2-8f4949518fdc","level":"subsubsection","origin_cites_number":8,"parent_id":"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Parallelism \\& FL Types"],["subsubsection","Model Parallelism"]],"subsections":[],"title":"Model Parallelism"},{"cite_extract_rate":0.5,"cites":[5425],"content":"Pipeline parallelism is realized by having dependent data processing nodes distributed at different computing resources . As shown in Figure \\ref{fig:pipelinep}, the data processing nodes are distributed at multiple computing resources. While data point $I^3_A$ is processed in computing resource 1, the outputs of $A_1$ and $A_2$ are processed in computing resource 2, and the outputs of $B_1$, $B_2$, and $B_3$ are processed in computing resource 3. With this type of parallelism, the dependent data processing nodes can process the data in parallel. As this parallelism may incur many inter-computing resource data transfers, it is not widely used for FL.","id":"1b1418f9-bcb1-4bbe-b99b-80b182bd318e","level":"subsubsection","origin_cites_number":2,"parent_id":"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Parallelism \\& FL Types"],["subsubsection","Pipeline Parallelism"]],"subsections":[],"title":"Pipeline Parallelism"},{"cite_extract_rate":1,"cites":[636],"content":"\\label{sebsec:aggre}\nWith the horizontal FL and data parallelism, aggregation algorithms are used to aggregate the models or gradients generated from the forward and backward propagation in each computing resource. The aggregation algorithms can be either centralized, or hierarchical, and decentralized. The centralized aggregation algorithms generally rely on a centralized server, i.e., a parameter server, to synchronize or schedule the execution of distributed computing resources, while hierarchical aggregation algorithms rely on multiple parameter servers for the model aggregation.\nThe decentralized aggregation algorithms make each computing resource equally perform the calculation based on a predefined protocol, without relying on a centralized server. \\liu{Please refer to for the details of federated optimization. The characteristics are summarized in Table \\ref{tal:aggregation}, which can be used to choose appropriate algorithms in a specific situation. }\n\\begin{table}[ht]\n\\centering\n\\caption{\\liu{Comparison among diverse types of aggregation algorithms. ``Complexity'' represents the complexity to implement the algorithms (``H'' represents high complexity, ``M'' represents medium complexity, and ``L'' represents low complexity). ``Trust'' represents whether the aggregation algorithms require that the data owners trust the centralized server. ``Imbalance'' represents whether the algorithms can address the unbalanced data. ``High-latency'' represents whether the algorithms can support the high-latency model or gradient data transfer. ``Y'' represents that the algorithms support the functionality while ``N'' represents that the algorithms do not support the functionality.}}\n\\label{tal:aggregation}\n\\begin{tabular}{|c|c|c|c|c|c|}\n\\hline\nType & Complexity & Trust & Imbalance & High-latency \\\\\n\\hline\nCentralized & L & Y & N & N \\\\\n\\hline\nHierarchical & M & Y & Y & Y \\\\\n\\hline\nDecentralized & H & N & N & Y \\\\\n\\hline\n\\end{tabular}\n\\end{table}","id":"10053136-8ab4-4b06-aec3-a7a9b0e56bc3","level":"subsection","origin_cites_number":1,"parent_id":"5dfeced7-075b-4276-abd0-1ff3e9bb48a7","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Aggregation Algorithms"]],"subsections":["e5dd26ef-00aa-4b84-aa33-7ff2605aebde","508677b5-4dbe-4b11-ad6a-7789cb9e091d","ecce2209-838c-4800-b0a9-49b94ea5037a"],"title":"Aggregation Algorithms"},{"cite_extract_rate":0.8125,"cites":[635,602,7928,5440,5443,5441,5445,8919,5442,582,5444,5446,632],"content":"\\label{subsubsec:centralized}\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.7\\textwidth]{figures/cenagg.png}\n \\caption{The architecture of centralized aggregation. ``CR'' represents computer resource. $\\omega$ represents the local parameters or the weights of the model calculated in each computing resource. $\\mathrm{g}$ represents the local gradients in backward propagation in each computing resource. $\\overline{\\omega}$ represents the global model calculated in the parameter server. $\\overline{\\mathrm{g}}$ represents the global gradients calculated in the parameter server.}\n \\label{fig:centra}\n\\end{figure}\nAs shown in Figure \\ref{fig:centra}, a single parameter server is used to calculate the average models or gradients sent from multiple computing resources (mobiles).\nThe weights of the model (model) or the gradients are calculated in each computing resource, which are transferred to a parameter server. \nThe parameter server calculates global gradients or global models according to a centralized aggregation algorithm.\nThe global gradients or global models are transferred to each computing resource for the following computation. \nThe update of the model is based on the SGD defined in Formula \\ref{eq:gd} in both computing resources, or on the parameter server.\nA bunch of centralized aggregation algorithms have been proposed.\nFederated Averaging (FedAvg) algorithm is introduced as the aggregation method in Google's implementation of an FL system. A centralized server aggregates the machine learning models from selected users. Then, a global model is generated using a weighted sum of each aggregated machine learning model. Afterward, the global model is shared with selected users, and the training process is continued in the computing resource of selected users. \nHowever, the trained model of FedAvg may be biased towards computing resources with favorable network conditions . \nWhile FedAvg is a straightforward approach, some other methods are proposed to address additional problems. \nA Federated Stochastic Block Coordinate Descent (FedBCD) algorithm is proposed to reduce the number of communication rounds by enabling multiple local updates before the model communication between a user and the server. In addition, FedBCD also considers the regularization during the training process. The training problem with regularization can be formulated as:\n\\begin{equation}\n\\label{eq:regularizer}\n \\widehat F = \\underset{F_\\theta\\in\\mathcal{H}}{\\mathrm{argmin}} \\underset{(x,y)\\sim\\mathcal{D}}{\\mathbb{E}} L(y, F_\\theta(x)) + \\lambda \\cdot \\gamma(\\theta),\n\\end{equation}\nwhere $F$, $D$, $\\mathcal{H}$ are the same as those in Formula \\ref{eq:problem}, while $\\gamma(\\cdot)$ denotes the regularizer and $\\lambda$ is the hyper-parameter. The regularization is exploited to improve the generalization capacity of the trained machine learning model. As the fairness among multiple users is important for an FL system, the Stochastic Agnostic Federated Learning (SAFL) algorithm and the FedMGDA+ algorithm are proposed to achieve fairness during the training process of FL. The fairness represents that the data distribution among multiple users can be equally considered without the influence of unrelated factors. \nFairness may also refer to two other concepts: (1) A user gets a final model according to the contribution ; and/or (2) Uniform accuracy distribution among all the distributed computing resources , which are out of the scope of this paper.\nIn addition, while the computing resources may be heterogeneous, FedProx is proposed to tackle the heterogeneity in an FL system. FedProx enables multiple iterations in each computing resource, while minimizing a cost function based on the local loss function and the global model. Furthermore, in order to address permutation of data processing nodes during the training process, Federated Matched Averaging (FedMA) is proposed. FedMA exploits an existing approach, i.e., BBP-MAP , to generate a matrix, in order to align the data processing nodes of the models from computing resources and the server. SCAFFOLD is proposed to reduce the communication rounds, using stateful variables in the distributed computing resources. Attention-augmented mechanism is exploited in Attentive Federated Aggregation (FedAttOpt) to aggregate the knowledge generated from each computing resource (client), based on the contribution of the model from each client. \nWhen the data distribution is heterogeneous among users, personalization remains an open problem. In order to address this problem, the model can be split into local layers and global layers, which has been proposed in adaptive personalized federated learning (APFL) , FedPer , and pFedMe . The local layers are trained with the decentralized data in each computing resource of users, while the global layers are trained in the computing resources of users and the server. However, it is difficult to choose a dataset and its partition among clients to measure the personalization brought by APFL or FedPer, so as to prove the improvement compared with FedAvg.\nThe attention-augmented mechanism helps reduce the communication rounds.\nIn addition, knowledge distillation can also be exploited to aggregate the models, while requiring that there is data in the centralized server .\nAll these algorithms can handle non-IID data. A comparison among the aforementioned algorithms is proposed in Table \\ref{tal:aggre}. \n\\begin{table}[ht]\n\\centering\n\\caption{Comparison among aggregation algorithms. ``Reg'' represents regularization. Heterogeneity represents that the computing resources are heterogeneous. ``Firness'' represents that the data distribution among multiple users can be equally considered without the influence of unrelated factors. ``Permutation'' refers to the permutation of data processing nodes during the training process. ``C-E'' represents communication efficient. ``S'' represents that the algorithm supports the functionality, while ``N'' represents that the algorithm does not have support. }\n\\label{tal:aggre}\n\\begin{tabular}{|c|c|c|c|c|c|}\n\\hline\nAlgorithm & Reg & Fairness & Heterogeneity & Permutation & C-E \\\\\n\\hline\nFedAvg & N & N & N & N & N \\\\\n\\hline\nFedBCD & S & N & N & N & S \\\\\n\\hline\nSAFL & N & S & N & N & N \\\\\n\\hline\nFedMGDA+ & N & S & N & N & S \\\\\n\\hline\nFedProx & N & N & S & N & S \\\\\n\\hline\nFedMA & N & N & N & S & S \\\\\n\\hline\nSCAFFOLD& N & N & N & N & S \\\\\n\\hline\nFedAttOpt& N & N & N & N & S \\\\\n\\hline\n\\end{tabular}\n\\end{table}","id":"e5dd26ef-00aa-4b84-aa33-7ff2605aebde","level":"subsubsection","origin_cites_number":16,"parent_id":"10053136-8ab4-4b06-aec3-a7a9b0e56bc3","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Aggregation Algorithms"],["subsubsection","Centralized Aggregation"]],"subsections":[],"title":"Centralized Aggregation"},{"cite_extract_rate":0.7142857142857141,"cites":[646,644,5448,5447,5449],"content":"\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.95\\textwidth]{figures/hier.png}\n \\caption{The architecture of hierarchical aggregation. ``CR'' represents computer resource. $\\omega$ represents the local parameters or the weights of the model calculated in each computing resource. $\\mathrm{g}$ represents the local gradients in backward propagation in each computing resource. $\\overline{\\omega}$ represents the region or global model calculated in each parameter server. $\\overline{\\mathrm{g}}$ represents the region or global gradients calculated in each parameter server. The region model or gradients are calculated by a region parameter server, while the global model or gradients are calculated by a global parameter server.}\n \\label{fig:hier}\n\\end{figure}\nAs shown in Figure \\ref{fig:hier}, a hierarchical architecture is also exploited using multiple parameter servers. \nA two-layer hierarchical architecture is proposed to reduce the time to transfer models between a parameter server and computing resources . The hierarchical architecture uses a global parameter server (GPS) and multiple region parameter servers. Each region parameter server (RPS) is implemented in a cell base station where the computing resources (mobiles) can be connected, with low latency. A hierarchical algorithm, i.e., Hierarchical Federated Learning (HFL) is deployed to realize the model aggregation. Within each iteration of HFL, each RPS calculates an average model using the models of the computing resources within its cluster. It sends the averaged model to the GPS, and it receives a global averaged model at every certain iteration. Afterward, it broadcasts the averaged model to all its computing resources. Some other algorithms, e.g., HierFAVG , HFEL , and LanFL , are similar to HFL, while the SPS is an edge or Local-Area Network (LAN) parameter server and the MPS is a parameter server implemented on the cloud or a Wide-Area Network (WAN). These algorithms take advantage of hierarchical architecture to reduce high-latency model or gradient data transfer, so as to accelerate the training process. In addition, by well-clustering the computing resources to groups, the hierarchical architecture is also exploited to address unbalanced data distributed among multiple computing resources , or to address data privacy .","id":"508677b5-4dbe-4b11-ad6a-7789cb9e091d","level":"subsubsection","origin_cites_number":7,"parent_id":"10053136-8ab4-4b06-aec3-a7a9b0e56bc3","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Aggregation Algorithms"],["subsubsection","Hierarchical Aggregation"]],"subsections":[],"title":"Hierarchical Aggregation"},{"cite_extract_rate":0.5,"cites":[5452,8920,5450,7930,5451,7929],"content":"\\label{subsubsec:decentralized}\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.8\\textwidth]{figures/decentral.png}\n \\caption{The architecture of decentralized aggregation. ``CR'' represents computer resource. $\\omega$ represents the local parameters, or the weights of the model calculated in each computing resource. $\\mathrm{g}$ represents the local gradients in backward propagation in each computing resource. When two computing resources are neighbors, they can communicate with each other.}\n \\label{fig:decentralized}\n\\end{figure}\nWhile collaboratively training a machine learning model with a decentralized aggregation algorithm, the computing resources can be organized with a {\\em connected} topology and can communicate with a peer-to-peer manner, as shown in Figure \\ref{fig:decentralized}. The degree and connectivity of the topology affects the communication efficiency and the convergence rate of the aggregation algorithm. \nFor a given topology, we define $w_{i,j}$, the weight to scale information flowing from node $j$ to node $i$, as follows\\vspace{-2mm}\n\\begin{align}\\label{wij}\nw_{ij}\n\\begin{cases}\n> 0 & \\mbox{if node $j$ is connected to node $i$, or $i=j$;} \\\\\n= 0 & \\mbox{otherwise.}\\vspace{-2mm}\n\\end{cases}\n\\end{align}\nWe further define the {\\bf topology matrix} $W = [w_{ij}]_{i,j=0}^{n-1} \\in \\mathbb{R}^{n\\times n}$ as the matrix to represent the topology. \nIn the remainder of this paper, \nwe assume that $W$ satisfies \n$W\\mathds{1} = \\mathds{1}$ \nand $\\mathds{1}^T W = \\mathds{1}^T$, \ni.e., both the row sum and the column sum of $W$ are equal to $1$, \nso as to guarantee that the neighborhood averaging will asymptotically approach the global averaging . \nWhen a computing resource $j$ is directly connected to computing resource $i$, i.e., $w_{i,j} \\neq 0$, computing resource $j$ is the neighbor of computing resource $i$.\n\\liu{Please note that the weight $w_{i,j}$ denotes the confidence node $i$ has in the information it receives from node $j$ , which is different from the bandwidth or data transfer capacity in a network.}\nThe centralized aggregation algorithm is a special type of decentralized aggregation with a star topology while only the centralized server communicates with its neighbors.\n\\liu{A well designed topology, e.g., an exponential graph , can improve the convergence rate, which accelerates the training speed.}\nWith the decentralized SGD (D-SGD), each computing resource maintains a local copy of \nthe global model parameters, and it updates the local copy using the models of its neighbors. \nAccording to the order to conduct neighborhood averaging and gradient descent, D-SGD has two common types of realizations: Average-With-Communication (AWC) and Average-Before-Communication (ABC) .\nAWC can overlap communication and gradient computation, while ABC needs to sequentially calculate and communicate the gradient or model. \nHowever, ABC is robust , and it converges fast in terms of iterations by exploiting its large learning rate.\nIn addition, the decentralized aggregation algorithms can be classified to Full Communication (FC) and Partial Communication (PC) according to the number of neighbors. Within the iterations of FC, each computing resource calculates an averaged model or gradient, based on all the models or gradients of the last version from all its neighbors. However, within the iterations of PC, each computing resource calculates an averaged model or gradient based on one or multiple chosen neighbors. With PC, the selection of the neighbors can be based on a gossip algorithm . For instance, a random neighbor can be selected ; the neighbors that provide benign models are selected to avoid attack .","id":"ecce2209-838c-4800-b0a9-49b94ea5037a","level":"subsubsection","origin_cites_number":12,"parent_id":"10053136-8ab4-4b06-aec3-a7a9b0e56bc3","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Distributed Training"],["subsection","Aggregation Algorithms"],["subsubsection","Decentralized Aggregation"]],"subsections":[],"title":"Decentralized Aggregation"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:security}\nAt the infrastructure layer of an FL system, there are three types of data manipulation: data security mechanisms, data transfer, and distributed data processing within the distributed execution module. \nWe first present the techniques for the distributed execution in an FL system. \nThen, we present the techniques for data transfer during the training process of an FL system.\nFinally, as data security is of much importance to an FL system , we present the techniques to protect the data security.","id":"cd0b3f18-6523-488f-aaa8-addaf533ff9f","level":"section","origin_cites_number":1,"parent_id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Data Manipulation"]],"subsections":["8048e455-33dd-411a-bdf6-5bafe4141358","09ac7749-f7de-46a4-bc72-304efd143b4a","4a182037-a6a4-48d6-aa9c-0986bac2c4b8"],"title":"Data Manipulation"},{"cite_extract_rate":0.5,"cites":[5454,623,655,659,5453],"content":"While the bandwidth within a single data center is high, e.g., InfiniBand, the High Performance Computing (HPC) libraries, e.g., Message Passing Interface (MPI) or NVIDIA Collective Communications Library (NCCL) , are widely exploited for distributed data processing . With MPI or NCCL, the gradients or models in each computing resource can be easily calculated using ring-AllReduce algorithm . However, one of the drawbacks of the HPC libraries is that they lack support for fault-tolerance, as the HPC libraries are designed for high performance servers with high quality networks. When any computing resource within the network becomes unavailable, the distributed training process may be broken.\nHowever, as an FL system is generally implemented for the collaboration of large amounts of mobile device users or different organizations, the network connection among computing resources is of moderate quality, i.e., the bandwidth is not as good as that within a single data center, and the latency is high.\nFor instance, the Internet upload speed is typically much slower than the download speed . Also, some users with unstable wireless communication channels may consequently drop out due to disconnection from the Internet .\nIn this environment, the connection between computing resources and parameter servers has a high possibility of becoming disabled. Thus, Remote Procedure Call (RPC) frameworks are widely exploited, as this kind of framework can ignore the disconnected computing resources and continue the distributed training of an FL system , e.g., PaddleFL , PySyft , or TensorflowFL .","id":"8048e455-33dd-411a-bdf6-5bafe4141358","level":"subsection","origin_cites_number":10,"parent_id":"cd0b3f18-6523-488f-aaa8-addaf533ff9f","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Data Manipulation"],["subsection","Distributed Data Processing"]],"subsections":[],"title":"Distributed Data Processing"},{"cite_extract_rate":0.769230769230769,"cites":[613,7932,623,622,7931,5456,5457,5455,625,8366],"content":"As the network connection is of moderate quality, the data transfer module mainly focuses on data compression to transfer intermediate data, e.g., gradients or models. \nSketched updates are proposed for gradient compression to accelerate the data transfer during the distributed training within a single data center .\nWith the data parallelism and centralized aggregation algorithm, before sending the intermediate data, the intermediate data can be sketched with subsampling , quantization , sparsification , or projection to lower dimensional spaces , in each computing resource, in order to reduce the cost to transfer data.\nSubsampling refers to transferring only a random subset of the intermediate data .\nQuantization methods encode each value using a fixed number of bits, so as to reduce the length of gradients or models .\nWith the sparsification approach, only selected parts of the intermediate data are transferred, while the selection is based on a threshold, e.g., the gradients larger than a threshold are selected . Then, when the intermediate data is received in the server, they are decompressed to be aggregated according to the aggregation algorithms presented in Section \\ref{subsubsec:centralized}. The convergence of the quantization approach is analyzed in , which shows that this approach can also provide good convergence rates . In addition, irrelevant intermediate data can be precluded to be transferred to the server, in order to substantially reduce the communication overhead .","id":"09ac7749-f7de-46a4-bc72-304efd143b4a","level":"subsection","origin_cites_number":13,"parent_id":"cd0b3f18-6523-488f-aaa8-addaf533ff9f","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Data Manipulation"],["subsection","Data Transfer"]],"subsections":[],"title":"Data Transfer"},{"cite_extract_rate":0,"cites":[],"content":"Data security is of much importance for data processing. The problem of data security is related to significant financial and reputational losses. For instance, Uber had to pay \\$148,000,000 to settle the investigation incurred by a breach of 600,000 drivers’ personal information in 2016 . \nData security mainly includes two aspects, i.e., data privacy and model security.\nData privacy refers to the protection of raw data to avoid raw data information leakage during or after the distributed training of FL systems.\nModel security refers to the protection of the security of trained models, in order to avoid wrong output based on the trained models incurred by malicious attacks.\nIn this section, we first present the techniques to protect data privacy. Then we present the defense methods for model security.","id":"4a182037-a6a4-48d6-aa9c-0986bac2c4b8","level":"subsection","origin_cites_number":4,"parent_id":"cd0b3f18-6523-488f-aaa8-addaf533ff9f","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Data Manipulation"],["subsection","Data Security"]],"subsections":["7a14e4c6-106a-4ec2-8926-70aed111f75a","cecfb28f-1aff-43d4-9671-2b7aa114b7c6"],"title":"Data Security"},{"cite_extract_rate":0.48571428571428504,"cites":[3469,5460,3477,892,5459,5461,5462,5439,5458,3478,7933,614,5463,7608,5441,5464,7727],"content":"The techniques to protect data privacy consist of three types: Trusted Execution Environment (TEE), encryption, Differential Privacy (DP), and anti-Generative Adversarial Network (GAN) methods. These techniques can be combined in FL systems, e.g., the combination of DP and TEE in , the combination of encryption and DP , and the combination of DP and anti-GAN .\nA TEE is an environment where the execution is secured and no information can be leaked to unauthorized users. \nIntel SGX technique has been first proposed as a secure environment while providing a set of security-related instruction codes built within Intel Central Processing Units (CPUs). \nThen, the implementation of machine learning models has been carried out in the TEE, i.e., Intel SGX, in order to enable collaborative data analysis based on machine learning algorithms while providing a security guarantee . Afterwards, the TEE has been exploited in FL systems, in order to protect the privacy of data in two ways. The first way is to put the entire training process in the TEE of each distributed computing resource to protect the data privacy during the distributed training . The second way is to use TEE to check a small part of the distributed training, while exploiting insecure computing resources, e.g., GPUs, to reduce the training time . \nAs an encryption technique, homomorphic encryption has been used to ensure the data privacy for FL systems .\nHomomorphic encryption allows specific types of computations to be carried out on encrypted input data, and to generate an encrypted result, which matches the result of the same computations on the decrypted input data. \n\\liu{Two main branches of homomorphic encryption exist, i.e., fully homomorphic encryption and partially homomorphic encryption. \nThe fully homomorphic encryption supports both addition and multiplication on ciphertext, while partially homomorphic encryption only supports either an addition or a multiplication operation on ciphertext, which corresponds to less computational flexibility and better runtime efficiency. \nBoth the fully and partially homomorphic encryptions can be exploited with the horizontal and vertical federated learning.}\nAs sharing gradients also leaks the information of training data \\liu{in horizontal federated learning} , it is of much importance to protect the privacy of the intermediate data.\nThus, the intermediate data can be encrypted using a homomorphic encryption algorithm before being sent to a parameter server . In this way, the intermediate data remain encrypted during the aggregation process while only the computing resource can decrypt the encrypted intermediate data. \nEven if the transferred encrypted intermediate data is leaked, the information of gradients or models remains safe, and the privacy of the training data is ensured.\n\\liu{In addition, partial homomorphic encryption, e.g., Paillier , is exploited in vertical federated learning .}\nHowever, the homomorphic encryption incurs significant costs in computation and communication during distributed training . In order to reduce the overhead of homomorphic encryption, a set of quantized gradients are encrypted . \nDifferential Privacy (DP) protects the data privacy by adding artificial noise to a small part of raw data, while ensuring that the modification does not substantially affect the performance of the machine learning models . \nDP is widely used in FL systems as the first step to process the raw data, and the output is the training data to be used for the distributed training . With more added noise, the privacy is better protected, i.e., there is less possibility to leak raw data information, while it takes more time to converge for the machine learning models . A trade-off between the privacy protection and the convergence performance can be made by selecting a certain number of distributed resources . However, DP may not be able to ensure the data privacy under certain attacks, e.g., Generative Adversarial Network (GAN) attacks .\nA well-trained machine learning model can leak information about the training data based on the intermediate data, e.g., gradients . GANs can be used to generate data similar to the training data based on a well-trained machine learning model in either a parameter server or a distributed computing resource . \nThe adversary can reconstruct other participating clients’ private data, even if it has no knowledge of the label information using the GANs.\nThus, during the distributed training process of FL systems, a malicious user can exploit GANs to infer the training data of other users. DP can be used to prevent the GAN-based attack . In addition, fake training data can be generated based on a GAN and original raw data, which is then used during the distributed training process to prevent the GAN-based attack .","id":"7a14e4c6-106a-4ec2-8926-70aed111f75a","level":"subsubsection","origin_cites_number":35,"parent_id":"4a182037-a6a4-48d6-aa9c-0986bac2c4b8","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Data Manipulation"],["subsection","Data Security"],["subsubsection","Data Privacy"]],"subsections":[],"title":"Data Privacy"},{"cite_extract_rate":0.923076923076923,"cites":[2673,5468,5470,5465,5451,5471,5467,7038,3415,597,5469,5466],"content":"We mainly focus on poisoning attacks in this section. \nThe objective of poisoning attacks is to reduce the accuracy of machine learning models using artificially designed data, i.e., data poisoning, or models, i.e., model poisoning, in one or several distributed computing resources during the model aggregation process (see details in in Section \\ref{subsubsec:centralized}). There are two ways to carry out poisoning attacks, i.e., data poisoning and model poisoning. \nData poisoning can be realized by modifying the features or the labels of the input data. For instance, malicious users can modify the data points of a certain class $C$ to other classes, and they can then use the modified data points to participate in the distributed training. \nThe modification of the labels is denoted by the label flipping attack.\nAs a result, the accuracy of the trained model has low accuracy in terms of Class $C$ . \nModel poisoning refers to the attacks in which the updated intermediate data, e.g., gradients or models, are poisoned before being sent to a parameter server in order to reduce the accuracy of the trained model . The goal of the model poisoning is to reduce the performance of the trained model on targeted tasks or classes, while the performance of the model remains unchanged in terms of other tasks or classes . \nData poisoning eventually realizes the model poisoning, as it enables some computing resources to update poisoned intermediate data based on the calculation of poisoned training data . \nHowever, model poisoning can be more powerful than data poisoning, as model poisoning directly influences the weights of the models and trains in a way that benefits the attack .\nBoth the data poisoning and the model poisoning rely on the backdoor attacks to modify the training data or the intermediate data . Backdoor attacks are performed by embedding the hidden instructions into machine learning models, so that the infected model performs well on benign testing samples when the backdoor is not activated, while its prediction will be changed to the attacker-specified target label when the backdoor is activated by the attacker . \nIn order to defend against these data attacks or model attacks, the malicious users should be identified by analyzing the updated intermediate data using dimensionality reduction methods, e.g., Principal Component Analysis (PCA) , anomaly detection , or interpretability techniques .\nIn addition, the model poisoning can be incurred by Byzantine failures of certain distributed computing resources . With Byzantine failures, some computing resources (bad users) are manipulated by attackers during the distributed training process, which significantly degrades the performance of the global model in terms of test error . \nIn order to make the training process robust against the Byzantine failures, the bad users can be identified by analyzing the updated intermediate data using a hidden Markov model or via secure aggregation protocols .","id":"cecfb28f-1aff-43d4-9671-2b7aa114b7c6","level":"subsubsection","origin_cites_number":13,"parent_id":"4a182037-a6a4-48d6-aa9c-0986bac2c4b8","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Data Manipulation"],["subsection","Data Security"],["subsubsection","Model Security"]],"subsections":[],"title":"Model Security"},{"cite_extract_rate":0.33333333333333304,"cites":[5420,5423],"content":"\\label{sec:frameworks}\nFL systems are widely applied in diverse domains, e.g., \nmobile service, healthcare , and finance .\nAn FL system generally exploits an FL framework, which is deployed on distributed resources.\nIn this section, we present four widely used FL frameworks: PaddleFL , TensorFlowFederated , FATE , and PySyft .","id":"727148cb-019c-4e5c-9d02-e8f0c8fb671c","level":"section","origin_cites_number":6,"parent_id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Federated Learning Frameworks"]],"subsections":["37621a76-9ea9-4868-9b62-1b2d192ce562","c27dbf26-a266-414a-aa80-dc2e4d0fee2b","d79e2f48-eba0-495e-b5ac-cb343eff8f4d","baf085e7-c8e0-4563-86bc-a6f1583fb40d","c6708387-e962-4e76-9ea7-b422245775eb"],"title":"Federated Learning Frameworks"},{"cite_extract_rate":0.33333333333333304,"cites":[5422],"content":"PaddleFL is an open source federated learning framework based on PaddlePaddle , which is supported by Baidu. At the presentation layer, PaddleFL provides a textual UI for the interaction between users and the FL system. At the User Services layer, PaddleFL provides the log and monitoring supports, and it can leverage the interpretability module of PaddlePaddle in the future. At the FL training layer, PaddleFL can realize data parallelism (horizontal FL) and model parallelism (vertical FL). It supports multiple aggregation algorithms, e.g., FedAvg, and fault-tolerance. At the infrastructure layer, PaddleFL exploits RPC for the distributed execution. PaddleFL exploits DP to protect the data security. PaddleFL is widely used in multiple domains, e.g., Natural Language Processing (NLP), Computing Vision (CV) , and recommendation.","id":"37621a76-9ea9-4868-9b62-1b2d192ce562","level":"subsection","origin_cites_number":3,"parent_id":"727148cb-019c-4e5c-9d02-e8f0c8fb671c","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Federated Learning Frameworks"],["subsection","PaddleFL"]],"subsections":[],"title":"PaddleFL"},{"cite_extract_rate":0.6666666666666661,"cites":[5419,582],"content":"TensorFlow Federated (TFF) is an open-source framework for federated learning on decentralized data, which is supported by Google. TFF also provides a textual UI through Python. TFF supports the monitoring and log functionality at the user service layer. TFF supports data parallelism (horizontal FL), multiple aggregation algorithms, and fault-tolerance of mobile devices. TFF exploits RPC for the distributed execution and DP for the protection of data privacy.\nTFF enables Android mobile users to predict the next word while using the keyboard on their mobile phones .","id":"c27dbf26-a266-414a-aa80-dc2e4d0fee2b","level":"subsection","origin_cites_number":3,"parent_id":"727148cb-019c-4e5c-9d02-e8f0c8fb671c","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Federated Learning Frameworks"],["subsection","TensorFlowFederated"]],"subsections":[],"title":"TensorFlowFederated"},{"cite_extract_rate":0,"cites":[],"content":"FATE is an open-source FL framework supported by WeBank. \nFATE provides both a graphical and textual UI. \nFATE can support the monitoring of distributed training through a web portal. \nFATE takes advantage of database management systems (DBMS) to track the execution status.\nFATE can enable horizontal (data parallelism), vertical (model parallelism), and hybrid federated learning. \nFATE exploits both the DP and HE to protect the data privacy. \nIn addition, FATE exploits RPC to perform the distributed execution.","id":"d79e2f48-eba0-495e-b5ac-cb343eff8f4d","level":"subsection","origin_cites_number":1,"parent_id":"727148cb-019c-4e5c-9d02-e8f0c8fb671c","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Federated Learning Frameworks"],["subsection","FATE"]],"subsections":[],"title":"FATE"},{"cite_extract_rate":0.33333333333333304,"cites":[656],"content":"PySyft is an open-source FL framework based on the PyTorch framework .\nPySyft is written in Python and provides a textual UI based on Python.\nPySyft mainly supports the data parallelism and model parallelism based on an aggregator or orchestrating server. The aggregator or orchestrating server sends a part of the model to participating clients to process local data and gets results for federated averaging. PySyft exploits DP and encryption techniques to protect the data security. PySyft exploits multiple communication protocols for distributed execution, e.g., RPC, websocket etc.","id":"baf085e7-c8e0-4563-86bc-a6f1583fb40d","level":"subsection","origin_cites_number":3,"parent_id":"727148cb-019c-4e5c-9d02-e8f0c8fb671c","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Federated Learning Frameworks"],["subsection","PySyft"]],"subsections":[],"title":"PySyft"},{"cite_extract_rate":0.33333333333333304,"cites":[7934],"content":"}\n\\liu{\nDiverse FL frameworks exist while each has its advantage. We summarize the characteristics of each framework in Table \\ref{tal:FLFramework}, so as to help select a proper framework for use. From the table, we can see that all the frameworks implement the centralized aggregation algorithms, while employing DP and HE for the data security. PaddleFL can exploit Paddle to realize data, model, and pipeline parallelism. FATE and TFF are based on Tensorflow as the engine, while FATE can provide Web portal UI, which is convenient for novices. PySyft is compatible with PyTorch, which can easily handle the PyTorch-based tasks, while PaddleFL is compatible with Paddle, which can easily deal with rich pre-trained models published in PaddleHub .}\n\\begin{table}[ht]\n\\centering\n\\caption{\\liu{Comparison among diverse frameworks. ``Aggregation'' represents the type of aggregation algorithms. ``Textual'' represents the textual UI, while ``Web'' represents Web portal.}}\n\\label{tal:FLFramework}\n\\begin{tabular}{|c|c|c|c|c|c|}\n\\hline\nFramework & Engine & Aggregation & UI & Parallelism & Security \\\\\n\\hline\nPaddleFL & Paddle & Centralized & textual & Data/Model/Pipeline & DP/HE \\\\\n\\hline\nTFF & TensorFlow & Centralized & textual & Data/Model & DP/HE \\\\\n\\hline\nFATE & TensorFlow & Centralized & Web & Data/Model & DP/HE \\\\\n\\hline\nPySyft & PyTorch & Centralized & textual & Data & DP/HE \\\\\n\\hline\n\\end{tabular}\n\\end{table}\n\\liu{Table \\ref{tal:FrameworkSupport} represents the support of diverse types of FL in terms of data distribution. All the frameworks support horizontal FL, while vertical FL is supported by three frameworks except TFF. PySyft cannot directly support the vertical FL, while PyVertical , which is built upon PySyft, can be used to support vertical FL with the compatibility of PyTorch models. The hybrid FL is only supported by Paddle and FATE. In addition, all the frameworks support the execution with GPU. In practice, although PaddleFL may correspond to slightly longer time, the accuracy of the trained model can be higher that of TFF and FATE, while PySyft may generate ``out of memory'' errors .}\n\\begin{table}[ht]\n\\centering\n\\caption{\\liu{Comparison among diverse frameworks for the support of diverse FL types, e.g., horizontal FL, vertical FL, and hybrid FL, and GPU. $\\|$\\checkmark$\\|$ represents that the support is not realized by itself but a close one.}}\n\\label{tal:FrameworkSupport}\n\\begin{tabular}{|c|c|c|c|c|c|}\n\\hline\n\\multicolumn{2}{|c|}{} & PaddleFL & TFF & FATE & PySyft \\\\\n\\hline\n\\multirow{3}{*}{Types} & Horizontal & \\checkmark & \\checkmark & \\checkmark & \\checkmark \\\\\n\\cline{2-6}\n& Vertical & \\checkmark & \\xmark & \\checkmark & $\\|$\\checkmark$\\|$ \\\\\n\\cline{2-6}\n& Hybrid & \\checkmark & \\xmark & \\checkmark & \\xmark \\\\\n\\hline\n\\multicolumn{2}{|c|}{GPU} & \\checkmark & \\checkmark & \\checkmark & \\checkmark \\\\\n\\hline\n\\end{tabular}\n\\end{table}","id":"c6708387-e962-4e76-9ea7-b422245775eb","level":"subsection","origin_cites_number":3,"parent_id":"727148cb-019c-4e5c-9d02-e8f0c8fb671c","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Federated Learning Frameworks"],["subsection","\\liu{Concluding Remarks"]],"subsections":[],"title":"\\liu{Concluding Remarks"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:future}\nAlthough much work has been done on the FL systems, there remain\nsome limitations, e.g., interpretability of FL, decentralized aggregation, FL on graphs, benchmarks of FL systems, and applications to distributed intelligent systems. This section\ndiscusses the limitations of the existing frameworks and proposes new research directions.","id":"226ea081-07c1-4d87-ad04-18733a7f7202","level":"section","origin_cites_number":0,"parent_id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Research Directions"]],"subsections":["9866dd12-79cc-42bf-8557-f19459ea5936","b6673d3f-9d3e-4599-b8f4-93fdf908307f","9c3427af-198a-448e-bda6-b7fbe346fc0a","35e51ec5-ea11-4e9b-a3ee-cbf4853180ca","e54c8539-ee3c-4e42-b5a8-6e639fe2d31f","31802936-db7d-403f-b909-ee8ee4d7b0c8"],"title":"Research Directions"},{"cite_extract_rate":0.4,"cites":[653,582],"content":"Several datasets exist for experiments on FL systems. For instance, Federated Extended MNIST (FEMNIST) is built by partitioning the data in Extended MNIST based on each writer.\nShakespeare is built from The Complete Works of William Shakespeare based on each speaking role. \nBoth of these datasets can be used for horizontal FL. However, no public datasets exist for vertical FL or transfer FL. In addition, no open decentralized IID or non-IID distribution of popular datasets, e.g., ImageNet , exist for FL systems.","id":"9866dd12-79cc-42bf-8557-f19459ea5936","level":"subsection","origin_cites_number":5,"parent_id":"226ea081-07c1-4d87-ad04-18733a7f7202","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Research Directions"],["subsection","Benchmarks"]],"subsections":[],"title":"Benchmarks"},{"cite_extract_rate":1,"cites":[5421,5437],"content":"Deep neural networks have excellent performance in various areas, while it is often difficult to understand the results of deep neural network models, especially within FL systems.\nShapley values have been used to provide the interpretability , while it focuses on vertical FL. When multiple users collaboratively train an FL model, it remains an open problem to evaluate the contributions of each user, which helps provide evidence for the incentive of each user. \nThe primary incentive for clients to participate in federated learning is obtaining better models , while the benefit of participating in federated learning for clients who have sufficient private data to train accurate local models is disputable. \nInterpretability can help understand the contributions of each user and provide an objective opinion on the incentive strategy within an FL system.\nIn addition, the interpretability helps domain experts to understand the relationship between data and the final trained model in critical domains, e.g., healthcare and finance.\nHowever, the interpretability within FL systems remains an open problem.","id":"b6673d3f-9d3e-4599-b8f4-93fdf908307f","level":"subsection","origin_cites_number":2,"parent_id":"226ea081-07c1-4d87-ad04-18733a7f7202","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Research Directions"],["subsection","Interpretability"]],"subsections":[],"title":"Interpretability"},{"cite_extract_rate":0,"cites":[],"content":"Current aggregation algorithms of FL systems focus on the full connection or star connection topology, while other topologies, e.g., dynamic exponential-2 graph, may help accelerate the distributed training with FL systems . \\liu{In addition, well-known graph algorithms, e.g., graph partitioning algorithms, and ad-hoc policies can be exploited to help better distribute computing resources with the topology defined in Section \\ref{subsubsec:decentralized} in order to improve the efficiency of FL systems. } While the peer-to-peer communication enables the FL with an arbitrary topology matrix, the data security under diverse attacks, e.g., data or model poisoning, GAN-based attacks, remain open problems and deserve further investigation.","id":"9c3427af-198a-448e-bda6-b7fbe346fc0a","level":"subsection","origin_cites_number":1,"parent_id":"226ea081-07c1-4d87-ad04-18733a7f7202","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Research Directions"],["subsection","Decentralized Aggregation"]],"subsections":[],"title":"Decentralized Aggregation"},{"cite_extract_rate":0.6000000000000001,"cites":[7935,5438,180],"content":"Graphs or graph neural networks (GNN) have gained increasing popularity in multiple domains, e.g., social network, knowledge graph, and recommender system. FL frameworks for graphs, i.e., GraphFL , and GNN, i.e., SGNN , have been proposed to train a model with decentralized graphs. However, the data security of FL on graphs remains an open problem. \\liu{In addition, while a multimodal knowledge graph could not only contain text but also images or other type of data , it is worth further exploration to efficiently support the multimodel knowledge graph construction within an FL system .}","id":"35e51ec5-ea11-4e9b-a3ee-cbf4853180ca","level":"subsection","origin_cites_number":5,"parent_id":"226ea081-07c1-4d87-ad04-18733a7f7202","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Research Directions"],["subsection","Federated Learning on Graphs"]],"subsections":[],"title":"Federated Learning on Graphs"},{"cite_extract_rate":0.375,"cites":[8921,5472,5473],"content":"}\n\\liu{\nAlthough FL focuses on the non-IID data, the real-world decentralized data usually exhibit an imbalanced distribution . While the imbalanced data exist in multiple areas, such as computer vision , bioinformatics, and biomedicine , learning from such data requires special attention upon data sampling , data augmentation , and loss function designs . The imbalanced data is related to diverse tasks, e.g., two-class or multi-class classification . However, an optimized approach can be proposed to address the imbalanced data within FL systems.}","id":"e54c8539-ee3c-4e42-b5a8-6e639fe2d31f","level":"subsection","origin_cites_number":8,"parent_id":"226ea081-07c1-4d87-ad04-18733a7f7202","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Research Directions"],["subsection","\\liu{Imbalanced Data"]],"subsections":[],"title":"\\liu{Imbalanced Data"},{"cite_extract_rate":0,"cites":[],"content":"Machine learning algorithms have been widely used to boost the performance of intelligent systems, while FL systems could further enhance intelligent systems~ in distributed computing environments with privacy and security ensured. An intelligent system is a group of machines that has the capacity to gather data, analyze the data, and respond to other systems or the world around. \nWith FL systems, the distributed data can be exploited to generate models of high performance so as to produce smart responses.","id":"31802936-db7d-403f-b909-ee8ee4d7b0c8","level":"subsection","origin_cites_number":3,"parent_id":"226ea081-07c1-4d87-ad04-18733a7f7202","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Research Directions"],["subsection","Applications to Distributed Intelligent Systems"]],"subsections":[],"title":"Applications to Distributed Intelligent Systems"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:con}\nIn this paper, we discussed the current state of the art of FL systems, including the functional architecture of FL systems, distributed training, and data manipulation. \nFirst, we presented an overview of FL systems. In particular, we introduced the life cycle of FL models, including four phases. Then, we presented the four-layer functional architecture of FL systems, including presentation, user services, FL training, and infrastructure, and we presented each layer in detail. \nSecond, we detailed the distributed training with two parts, i.e., parallelism and aggregation algorithms. We presented three types of parallelism, including data parallelism, model parallelism, and pipeline parallelism. We associate each parallelism to a corresponding type of FL. For instance, data parallelism is associated with the horizontal FL, which corresponds to cross-device or cross-silo FL. Model parallelism is related to vertical FL and cross-silo FL. We presented the features of different aggregation algorithms in three types: centralized aggregation, hierarchical aggregation, and decentralized aggregation. \nThird, we presented the techniques for data manipulation within FL systems. We showed that FL systems prefer RPC for the distributed execution, to handle the fault-tolerance because of moderate network connection. Intermediate data are sketched in order to compress the data, so as to reduce the data communication time. In addition, we presented the data privacy and model security attacks and corresponding defense techniques, e.g., DP, HE, TEE, and the analysis of updated intermediate data for malicious user identification. \nWe mainly introduced four FL systems: PaddleFL, TensorFlowFederated, FATE, and PySyft. The current solutions primarily focus on the horizontal FL. And we identified five research directions that deserve further investigation: benchmarks, interpretability, decentralized aggregation, FL on graphs, imbalanced data, and the applications of FL systems to distributed intelligent systems.\n\\bibliographystyle{plain}\n\\bibliography{references}\n\\end{document}","id":"4ff5262c-b6d9-4162-aba6-d016f80c7306","level":"section","origin_cites_number":0,"parent_id":"42e77c1c-0c8b-4e28-9d53-60920d7558be","prefix_titles":[["title","From Distributed Machine Learning to Federated Learning: A Survey"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ]\n ],\n \"subsections\": [\n \"bda39eab-e9c6-47ab-8897-26d79998749b\",\n \"2371347b-6739-4c7c-b6c1-b23b3b02d8ed\",\n \"5dfeced7-075b-4276-abd0-1ff3e9bb48a7\",\n \"cd0b3f18-6523-488f-aaa8-addaf533ff9f\",\n \"727148cb-019c-4e5c-9d02-e8f0c8fb671c\",\n \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"4ff5262c-b6d9-4162-aba6-d016f80c7306\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.47222222222222204,\n \"cites\": [\n 3469,\n 1315,\n 5419,\n 3477,\n 591,\n 8917,\n 659,\n 5422,\n 5423,\n 1311,\n 5424,\n 671,\n 627,\n 5420,\n 600,\n 582,\n 5421\n ],\n \"content\": \"With billions of connected Internet of Things (IoT) devices , smartphones and large websites around the world, in recent years, we have witnessed huge amounts of data generated and dispersed over various mobile devices of end users, or the data centers of different organizations. As the data contain sensitive information of end users or organizations, such as facial images, location-based services, health information , or personal economic status, moving the raw data from personal devices or data centers of multiple organizations to a centralized server or data center may pose immediate or potential information leakage. Due to the concerns of data security and data privacy, legal restrictions, such as the Cybersecurity Law of the People’s Republic (CLPR) of China , the General Data Protection Regulation (GDPR) in European Union, the Personal Data Protection Act (PDP) in Singapore, the California Consumer Privacy Act (CCPA) , and the Consumer Privacy Bill of Rights (CPBR) in the United States, have been introduced and put in practice, which makes data aggregation from distributed devices, multiple regions, or organizations, almost impossible . In addition, computing and storage resources are also typically distributed in multiple regions and organizations , which cannot be aggregated in a single data center.\\nFederated Learning (FL) emerges as an efficient approach to exploit the distributed resources to collaboratively train a machine learning model. FL is a distributed machine learning approach where multiple users collaboratively train a model, while keeping the raw data decentralized without being moved to a single server or data center .\\nFL not only exploits the distributed resources to efficiently carry out the training process of machine learning, but also promises to provide security and privacy for the decentralized raw data. Within FL, the raw data, or the data generated based on the raw data with security processing, serves as the training data. FL only allows the intermediate data to be transferred among the distributed computing resources while avoiding the transfer of training data. \\nThe distributed computing resources refer to mobile devices of end users or servers of multiple organizations.\\nFL brings the code to the data, instead of bringing the data to the code, and it addresses the fundamental problems of privacy, ownership, and locality of data .\\nIn this way, FL can enable multiple users to train a model while satisfying the legal data restrictions.\\nTraditional centralized machine learning approaches typically gather the distributed raw data generated on different devices or organizations to a single server or a cluster with shared data storage, which may bring serious data privacy and security concerns . \\nThe centralized approaches, in general, are associated with diverse challenges, including computational power and training time, and most importantly, security and privacy with respect to distributed data .\\nFL differs from the centralized approach in three aspects. First, FL does not allow direct raw data communication, while the centralized approach has no restriction. Second, FL exploits the distributed computing resources in multiple regions or organizations, while the centralized approach generally only utilizes a single server or a cluster in a single region, which belongs to a single organization. Third, FL generally takes advantage of encryption or other defense techniques to ensure the data privacy or security, while the centralized approach pays little attention to these security issues .\\nThe term ``federated learning'' was first introduced in 2016 , which focuses on the unbalanced and non-Independent and Identically Distributed (non-IID) data in mobile devices. \\nThe concept of FL was extended to three data scenarios, i.e., horizontal, vertical, and hybrid . \\nThe horizontal FL addresses the decentralized data of the same features, while the identifications are different. The vertical FL handles the decentralized data of the same identifications with different features. The hybrid FL deals with the data of different identifications and different features. Then, FL is formally defined as a machine learning approach where multiple clients collaborate in solving a machine learning problem while the raw data is stored locally and is neither exchanged nor transferred . \\nAn FL system is an efficient tool to carry out FL with decentralized data and resources. Several open-source FL systems, e.g., FATE , PaddleFL , TensorflowFL , and Pysyft , are now intensively used by both research communities, e.g., healthcare , and computer visions , and by industrial groups, e.g., WeBank . Although various FL systems exist, the architecture of FL systems has common features: In particular, they share the capability to collaboratively train a machine learning model. Most FL systems are composed of four layers, i.e., presentation, user services, FL training, and infrastructure. These four layers enable FL system users to design, execute, and analyze machine learning models with distributed data.\\nAlthough FL differs from the centralized machine learning approaches, it not only utilizes novel techniques designed for FL, but also takes advantage of the techniques designed for distributed machine learning. \\nFL exploits parallelization techniques designed for distributed machine learning.\\nFor instance, horizontal FL exploits the data parallelism, which trains multiple instances of the same model on different subsets of the training dataset . Vertical FL utilizes model parallelism to distribute parallel paths of a single model to multiple devices in order to handle the data of different features . \\nMultiple aggregation algorithms are proposed to aggregate the models in distributed computing resources.\\nData transfer techniques are also utilized in FL, e.g., model compression . \\nAs FL promises to provide data security and data privacy, diverse defense techniques, e.g., differential privacy , homomorphic encryption , and Robustness Aggregation , are designed to address the possible attacks .\\nThere have been a few surveys of FL. Some works provide a comprehensive study of FL, from the taxonomy of FL to the techniques, e.g., the efficiency, data privacy, security, and applications of FL. Some surveys focus on the data privacy and security of FL. Other surveys present the application of FL in a specific area, e.g., healthcare informatics , mobile edge networks , and neural architecture searches , and they personalize global models to work better for individual clients . However, few of them present the architecture of FL or analyze parallelization techniques in FL.\\nIn this paper, we provide a survey of federated learning and the related parallelization techniques. The main contributions of this paper are:\\n\\\\begin{itemize}\\n \\\\item A four-layer FL system architecture, which is useful for discussing the techniques for FL. This architecture can also be a baseline for other work and can help with the assessment and comparison of FL systems.\\n \\\\item A taxonomy of FL-related techniques, including the parallelization techniques, the aggregation algorithms, and the techniques for data communication and security, with a comparative analysis of the existing solutions.\\n \\\\item A discussion of research issues to improve the efficiency and security of FL systems.\\n\\\\end{itemize}\\nThis paper is organized as follows. Section \\\\ref{sec:overview} gives an overview of the execution of FL, including the FL system architectures and basic functional architecture of FL systems. Section \\\\ref{sec:pexec} focuses on the techniques used for distributed training of FL and aggregation methods. Section \\\\ref{sec:security} presents the techniques for distributed execution, data communication and data security of FL. Section \\\\ref{sec:frameworks} demonstrates the existing FL frameworks.\\nSection \\\\ref{sec:future} discusses the open issues raised for the execution of FL with distributed resources. Section \\\\ref{sec:con} summarizes the main findings of this study.\",\n \"id\": \"bda39eab-e9c6-47ab-8897-26d79998749b\",\n \"level\": \"section\",\n \"origin_cites_number\": 36,\n \"parent_id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:overview}\\nIn this section, we introduce the basic concepts of federated learning. Then, we present the life cycle of FL models. Afterwards, we detail the functional architecture and the corresponding functionality of FL systems.\",\n \"id\": \"2371347b-6739-4c7c-b6c1-b23b3b02d8ed\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ]\n ],\n \"subsections\": [\n \"4f50be54-11aa-4ae8-ae5e-fb6d0e6797e6\",\n \"cd1b4108-9163-49ab-bfbc-9fe67cc3a340\",\n \"78f72b80-7c28-4237-9e16-d97c8c144603\"\n ],\n \"title\": \"An Overview of Federated Learning\"\n },\n {\n \"cite_extract_rate\": 0.37037037037037,\n \"cites\": [\n 1315,\n 591,\n 5426,\n 659,\n 5425,\n 671,\n 600,\n 166,\n 582\n ],\n \"content\": \"Machine learning is the process to automatically extract the models or patterns from data . \\nThe models or patterns are expressed as machine learning models.\\nA machine learning model is an ensemble of a model structure, which is typically expressed as a Directed Acyclic Graph (DAG), data processing units, e.g., activation functions in Deep Neural Networks (DNNs), and the associated parameters or hyper-parameters. \\nThe input data can be processed through a machine learning model to generate the output, e.g., the prediction results or the classification results, which is the inference process.\\nThe machine learning model is generated based on the training data, which is the training process.\\nDuring the training process, the parameters or the model structure of the machine learning model is adjusted based on a training algorithm in order to improve the performance, e.g., the accuracy or the generalization capacity. The training algorithm is also denoted by machine learning algorithms. The duration of the training process is training time.\\nAccording to whether the training data have labels, the training process of machine learning can be classified into four types , i.e., supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Supervised learning represents that a machine learning task exploits the training data composed of input features and the corresponding labels . In this paper, we focus on this type of training data. For instance, each data point in the training dataset contains $(x,y)$, where $x$ represents the input features and $y$ represents the desired output value. Unsupervised learning represents that a machine learning task exploits the training data, which only consists of input features without output values; i.e., each data point only contains $x$ and does not have $y$. Semi-supervised learning represents that one (generally small) part of the training data contains output values, while the other (generally small) part of the training does not. Reinforcement learning represents that each iteration in the training process considers its observation of the environment from the last iteration. \\nWhile the training data become huge, e.g., on the order of terabyte , or when the training data is inherently distributed or too big to store on single machines , the training process is carried out using distributed resources, which is distributed machine learning. One of the important features of the distributed machine learning is that it can significantly accelerate the training speed so as to reduce the training time. Diverse parallelization techniques are used in distributed machine learning. For instance, Graphics Processing Units (GPUs) using Single Instruction Multiple Data (SIMD) and Tensor Processing Units (TPUs) using Multiple Instructions Multiple Data (MIMD) are exploited . In addition, distributed machine learning takes advantage of three types of parallelism to parallelize the training process, i.e., data parallelism , model parallelism , and pipeline parallelism . With the data parallelism approach, the training data is partitioned as many times as the number of computing resources, and all computing resources subsequently apply the same machine learning algorithm to process different chunks of the data sets . With the model parallelism approach, exact copies of the entirety of the data (the training data or the intermediate data)\\nare processed by each computing resource, each of which exploits different parts of the machine learning model . The pipeline parallelism approach combines the data parallelism and the model parallelism. With this approach, each computing resource processes a part of the training data with a part of the machine learning model, while the processing, e.g., computation or communication, at each node can be parallelized . \\nFL is a distributed machine learning approach where multiple users collaboratively train a model, while keeping the raw data distributed without being moved to a single server or data center .\\nThe model used for FL is denoted by FL model.\\nFL is first proposed to handle the unbalanced and non-Independent and Identically Distributed (non-IID) data of the same features in mobile devices .\\nThen, the concept of FL is extended to the distributed data of diverse features in multiple organizations or various regions .\\nFL systems are used within one or multiple phases of the life cycle of FL models. An FL system is a distributed system to manage the distributed training process with distributed resources.\\nFL is a special type of distributed machine learning, which differs from other distributed machine learning approaches in the following three points.\\nFirst, FL does not allow direct raw data communication, while other approaches have no restriction. \\nAs the raw data are of multiple ownerships, FL approaches with this restriction can meet the requirements defined by the related laws, e.g., CLPR , GDPR , PDPA , CCPA , and CPBR .\\nIn particular, the consent (GDPR Article 6) and the data minimalization principle (GDPR Article 5) limit data collection and storage to only what is consumer-consented and what is absolutely necessary for processing .\\nSecond, FL exploits the distributed computing resources in multiple regions or organizations, while the other approaches generally only utilize a single server or a cluster in a single region, which belongs to a single organization. FL enables the collaboration among multiple organizations.\\nThird, FL generally takes advantage of encryption or other defense techniques to ensure the data privacy or security, while the other approaches pay little attention to this security issue . FL promises to ensure the privacy and security of the raw data, as the leakage of information may incur significant financial and reputational losses. \\nDuring the training process of FL, an optimization problem is solved as shown in Formula \\\\ref{eq:problem}. Given $n$ training dataset $\\\\mathcal{D} = {D_1, D_2, ..., D_n}$, where each data point $(x,y)\\\\sim\\\\mathcal{D}$, the problem of FL is to learn a function $\\\\widehat F$ from all possible hypotheses $\\\\mathcal{H}$, while minimizing the expectation of loss over the distribution of all the dataset $\\\\mathcal{D}$. \\n\\\\begin{equation}\\n\\\\label{eq:problem}\\n \\\\widehat F = \\\\underset{F\\\\in\\\\mathcal{H}}{\\\\mathrm{argmin}} \\\\underset{(x,y)\\\\in\\\\mathcal{D}}{\\\\mathbb{E}} L(y, F(x)),\\n\\\\end{equation}\\nwhere $L(y, F(x))$ refers to the loss of $F(x)$ to the label $y$. \\nDuring the training process, the Stochastic Gradient Descent (SGD) approach is generally used to minimize the loss function using Formula \\\\ref{eq:gd}.\\n\\\\begin{equation}\\n\\\\label{eq:gd}\\n F_{k+1}(x)\\\\gets F_k(x) - \\\\eta_k \\\\nabla F_k(x),\\n\\\\end{equation}\\nwhere $F_k(x)$ refers to the learned model in the $k^{th}$ iteration, $\\\\nabla F_k(x)$ refers to the gradient of the model at the $k^{th}$ iteration based on the model already obtained $F_k(x)$ and the training dataset, $\\\\eta_k$ refers to the learning rate, and $F_{k+1}(x)$ refers to the update model of the $k^{th}$ iteration. Within each iteration, there are two phases, i.e., forward propagation and backward propagation. The forward propagation calculates the output based on the input data $x$ using the model, while the backward propagation calculates the gradients $\\\\nabla F_k(x)$ and updates the model.\\nWhen the calculation is distributed among multiple computing resources, the gradients or models of each computing resource are aggregated using an aggregation algorithm (see details in Section \\\\ref{sebsec:aggre}), in order to achieve consensus of multiple models and to generate a global model. \\\\liu{The learning rate can be dynamically adapted using a local adaptive optimizer, e.g., Adam, and/or cross-round learning rate schedulers .}\",\n \"id\": \"4f50be54-11aa-4ae8-ae5e-fb6d0e6797e6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 27,\n \"parent_id\": \"2371347b-6739-4c7c-b6c1-b23b3b02d8ed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ],\n [\n \"subsection\",\n \"Basic Concepts\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Basic Concepts\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 591,\n 612,\n 671,\n 5427\n ],\n \"content\": \"The life cycle of an FL model is a description of the state transitions of an FL model from creation to completion . Lo \\\\textit{et al.} propose that the life cycle of an FL model consists of 8 phases: initiated, broadcast, trained, transmitted, aggregated, evaluated, deployed, and monitored. Kairouz \\\\textit{et al.} propose that the life cycle of an FL model includes 6 phases: problem identification, client instrumentation, simulation prototyping, federated model training, model evaluation, and deployment. However, they focus on the FL with distributed data in mobile devices. In this paper, we adopt a combination of workflow life cycle views with a few variations , condensed into four phases:\\n\\\\begin{enumerate}\\n \\\\item The composition phase is for the creation of an FL model, which is used to address a specific machine learning problem, e.g., classification. First, a machine learning model is created to address the problem with certain requirements, e.g., the requirement of accuracy. Then, the machine learning model is adapted to FL scenarios. For instance, if the distributed data is of different features, the machine learning model is partitioned (see details in Section \\\\ref{subsubsec:modelP}) to process the distributed data. \\n \\\\item The FL training phase is for the training phase of the FL model. During this phase, a training strategy, which includes parallelism and aggregation algorithms (see details in Section \\\\ref{sec:pexec}), is used to update the parameters, hyper parameters, and even the structure of the network, in order to improve the accuracy and the generalization capacity of the FL model. \\n \\\\item The FL model evaluation phase is to apply the trained FL models, in order to analyze the performance of the trained FL models on a simulation platform or a real distributed system . As a result, the FL models with the best performance are selected. If the FL models do not meet the requirements, the FL model should be modified, or the training phase should be carried out again.\\n \\\\item The FL model deployment phase is to deploy the FL model in a real-life scenario to process the data. If the final model can be shared without restriction, there is no difference between the FL model deployment and the model generated from a traditional centralized approach. Otherwise, the deployment of the final model should consider the ownership of the corresponding parts. \\n\\\\end{enumerate}\",\n \"id\": \"cd1b4108-9163-49ab-bfbc-9fe67cc3a340\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"2371347b-6739-4c7c-b6c1-b23b3b02d8ed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ],\n [\n \"subsection\",\n \"FL Model Life Cycle\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"FL Model Life Cycle\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The functional architecture of an FL system can be layered as follows : presentation, user services, FL training, and infrastructure. Figure \\\\ref{fig:layers} shows this architecture. The higher layers exploit the lower layers to provide their own functionality. A user interacts with an FL system through the presentation layer and realizes independent functionalities at the user services layer. During the training phase of FL models, a Federated Learning Execution Plan (FLEP) is generated, and the corresponding distributed training is carried out at the FL training layer. The FLEP is composed of a type of parallelism, a scheduling strategy, and a fault-tolerance mechanism. The FL system manages the physical resources through the infrastructure layer for the distributed training. \\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.8\\\\textwidth]{figures/layers.pdf}\\n \\\\caption{\\\\liu{Functional architecture of an FL system.}}\\n \\\\label{fig:layers}\\n\\\\end{figure}\",\n \"id\": \"78f72b80-7c28-4237-9e16-d97c8c144603\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"2371347b-6739-4c7c-b6c1-b23b3b02d8ed\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ],\n [\n \"subsection\",\n \"Functional Architecture of FL Systems\"\n ]\n ],\n \"subsections\": [\n \"afff657f-2ca3-4a0d-b609-6427f5eee76c\",\n \"8ca8a082-7841-461b-ad59-49e7d8fef8fe\",\n \"6af71683-cb7e-4e75-87a8-c21ef2982fdc\",\n \"cfb855a7-c817-4897-9235-30d180850302\"\n ],\n \"title\": \"Functional Architecture of FL Systems\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 656\n ],\n \"content\": \"The presentation layer is a User Interface (UI) for the interaction between Users and FL systems at one or multiple stages of the FL model life cycle. The UI can be textual or graphical. This interface is responsible for designing a new FL model or choosing an existing machine learning model as an FL model. In addition, this layer also supports the modules at the user services layer, e.g., shows the status of the distributed training process. The textual UI is largely used for designing FL models \\\\liu{based on the command line or scripts }. The models can be directly expressed using Python, with the textual interface in PaddleFL , TensorFlowFL , PySyft , and FATE . A graphic UI can make the interaction more practical, while the users can drag or drop the data processing element to design an FL model. For instance, FATE provides a Graphic UI (GUI) through a web portal. However, the graphic portal also exploits textual programming languages as inner representations of an FL model.\",\n \"id\": \"afff657f-2ca3-4a0d-b609-6427f5eee76c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"78f72b80-7c28-4237-9e16-d97c8c144603\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ],\n [\n \"subsection\",\n \"Functional Architecture of FL Systems\"\n ],\n [\n \"subsubsection\",\n \"Presentation Layer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Presentation Layer\"\n },\n {\n \"cite_extract_rate\": 0.8888888888888881,\n \"cites\": [\n 5429,\n 591,\n 5434,\n 8918,\n 5437,\n 5435,\n 5433,\n 5432,\n 7927,\n 5428,\n 5430,\n 7926,\n 8502,\n 5436,\n 5438,\n 5431\n ],\n \"content\": \"The user service layer supports the expected functionalities, i.e., \\\\liu{monitoring and steering and log; interpretability and explainability; and graph data.} The monitoring enables the users to get the real-time status of the distributed training process. As the training process of FL models can be very long, e.g., from several hours to days , it is of much importance to track the execution status, which allows the user to verify whether the training proceeds normally. The log service is generally supported by major FL systems, which can be used to analyze the training process. In addition, the log generated during the training process can be used to debug the system or adjust the FL model. FATE provides a visual monitoring board to users through its GUI. When there are unexpected results or errors during the training process, steering enables users to adjust the training process in order to reduce the time necessary to carry out the distributed training from scratch. Most FL systems can enable the users to stop the training, while the adjustment of parameters is not fully supported by major FL systems. The interpretability of FL is to describe the internals of an FL system in a way that is understandable to humans . The explainability focuses on explaining the representation of data inside an FL model . With interpretability and explainability, the FL system can \\\\liu{provide} a description of the results of the trained FL model based on the training data and the distributed training process. Shapley values have been used to provide the interpretability , while both the interpretability and explainability remain open challenges as each is hard to fully support. \\n\\\\liu{In the real world, graph data widely exist in multiple domains, and a bunch of FL approaches have been proposed to handle the decentralized graph data for community detection , financial crime , and especially knowledge graph completion .\\nFL is particularly useful in the field of knowledge graph completion, as a knowledge graph could not only contain text but also images or other type of data, i.e., multimodal Knowledge Graphs ; and the completion is realized in a collaborative fashion within an FL system .\\nThe decentralized graphs can be inter-graph, i.e., the decentralized data belongs to multiple graphs, or intra-graph, i.e., the decentralized data is within one big graph , while most of the existing works focus on the intra-graph situation.\\nHorizontal FL techniques can be exploited on the Graph Neural Networks (GNN) with encryption techniques (see details in Section \\\\ref{subsubsec:infrastructure}) while the performance of FL may be much worse than that of centralized GNNs .\\nThe aggregation algorithms (see details in Section \\\\ref{subsubsec:centralized}) are also proposed based on the FedAvg or optimal transportation . \\nIn addition, decentralized aggregation algorithms (see details in Section \\\\ref{subsubsec:decentralized}) are also proposed to deal with the decentralized graph data for social network and drug discovery .\\nWhile the fine-tuning of the FL system is time-consuming, Bayesian optimization and evolutionary optimization strategies are utilized to automatically tune the hyper-parameters and the network structure, respectively. \\nGraph data can be vertically distributed where the features of the nodes are distributed across multiple data owners, and a data owner may or may not have the edges. Vertical FL exploits embeddings or autoencoders to represent the nodes, which can be transferred to train a GNN. In addition, the differential privacy (see details in Section \\\\ref{subsubsec:infrastructure}) is combined with the embeddings to protect the data privacy of knowledge graphs. }\",\n \"id\": \"8ca8a082-7841-461b-ad59-49e7d8fef8fe\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"78f72b80-7c28-4237-9e16-d97c8c144603\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ],\n [\n \"subsection\",\n \"Functional Architecture of FL Systems\"\n ],\n [\n \"subsubsection\",\n \"User Services Layer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"User Services Layer\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 590\n ],\n \"content\": \"The FL training layer carries out the distributed training process with distributed data and computing resources. This layer consists of three modules: parallelization, scheduling, and fault-tolerance. FL parallelization exploits diverse types of parallelism, e.g., data parallelism, model parallelism, and pipeline parallelism, to generate executable tasks. Through the FL scheduling module, an FL system produces a Scheduling Plan (SP) of executable tasks, which aims at fully exploiting distributed computing resources and preventing training stalling. During the training process, the SP is generally defined by a training algorithm, which aggregates the updates, i.e., gradients or models, from each computing resource in order to generate a final machine learning model. The FL fault-tolerance mechanism handles the failures or errors of task execution and the connection of distributed resources. Reactive approaches are generally exploited, e.g., using check-points, restart, and task replication . A reactive approach reduces the effect of failures after perceiving failures . An FLEP, which captures the execution directives, typically the result of compiling and optimizing the training process of FL models, is generated at this layer.\",\n \"id\": \"6af71683-cb7e-4e75-87a8-c21ef2982fdc\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"78f72b80-7c28-4237-9e16-d97c8c144603\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ],\n [\n \"subsection\",\n \"Functional Architecture of FL Systems\"\n ],\n [\n \"subsubsection\",\n \"FL Training Layer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"FL Training Layer\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 617,\n 7608\n ],\n \"content\": \"\\\\label{subsubsec:infrastructure}\\nThe infrastructure layer provides the interaction between an FL system and the distributed resources, including the computing resources, storage resources, network resources, and data resources. This layer contains three modules: a data security module, a data transfer module, and a distributed execution module. The data security module generally exploits Differential Privacy (DP) and encryption techniques, e.g., homomorphic , to protect the raw data used during the training process. Although the raw data cannot be directly transferred, intermediate data, e.g., the gradients or models, can be communicated among distributed computing resources. The data transfer module exploits data compression techniques to improve the data transfer efficiency. \\nAt this layer, the FLEP generated at the FL training layer is carried out within the distributed execution module; i.e., concrete tasks are executed in distributed computing resources.\",\n \"id\": \"cfb855a7-c817-4897-9235-30d180850302\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"78f72b80-7c28-4237-9e16-d97c8c144603\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"An Overview of Federated Learning\"\n ],\n [\n \"subsection\",\n \"Functional Architecture of FL Systems\"\n ],\n [\n \"subsubsection\",\n \"Infrastructure Layer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Infrastructure Layer\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:pexec}\\nIn this section, we present the distributed training process for FL. First, we present three types of parallelism in distributed training and the application within FL. The parallelism approaches are generally implemented in the parallelization module. Then, we discuss existing aggregation algorithms for the distributed training, which is implemented in the scheduling module.\",\n \"id\": \"5dfeced7-075b-4276-abd0-1ff3e9bb48a7\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ]\n ],\n \"subsections\": [\n \"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1\",\n \"10053136-8ab4-4b06-aec3-a7a9b0e56bc3\"\n ],\n \"title\": \"Distributed Training\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 600,\n 1315,\n 671\n ],\n \"content\": \"Three types of parallelism exist for distributed machine learning: data parallelism, model parallelism, and pipeline parallelism . FL can be classified to three types, i.e., horizontal, vertical, and hybrid . The horizontal FL generally exploits data parallelism, and the vertical FL typically takes advantage of model parallelism. However, the hybrid FL relies on transfer learning , which is not a parallelism approach and is out of the scope of this paper. \\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.65\\\\textwidth]{figures/dnn.pdf}\\n \\\\caption{An example of a neural network.}\\n \\\\label{fig:dnn}\\n\\\\end{figure}\\nIn this section, we take an example of a neural network as shown in Figure \\\\ref{fig:dnn} to explain the parallelism. In the example, we assume that the model contains three layers and seven data processing nodes (neurons), i.e., $A_1$, $A_2$, $B_1$, $B_2$, $B_3$, $C_1$, $C_2$, $D$. The arrows represent the data flow among different data processing nodes. The execution of the data processing nodes at each layer can be carried out in parallel, while the execution of different layers should be performed sequentially. The input data contains 4 data points. We assume two/three computing resources owned by two/three users. Each has a part of the input data.\\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.6\\\\textwidth]{figures/datap.pdf}\\n \\\\caption{Data parallelism. The forward and backward process of $I^1$ and $I^2$ is performed in computing resource 1, while that of $I^3$ and $I^4$ is performed in computing resource 2 at the same time. Then the model or gradient is transferred to calculate an average model or gradient to be sent to each computing resource for the following training.}\\n \\\\label{fig:datap}\\n\\\\end{figure}\",\n \"id\": \"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"5dfeced7-075b-4276-abd0-1ff3e9bb48a7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Parallelism \\\\& FL Types\"\n ]\n ],\n \"subsections\": [\n \"0430fcfa-cd31-4b1c-bf51-10d342c0d942\",\n \"62c91cb4-aacb-47de-82b2-8f4949518fdc\",\n \"1b1418f9-bcb1-4bbe-b99b-80b182bd318e\"\n ],\n \"title\": \"Parallelism \\\\& FL Types\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 671,\n 582,\n 591\n ],\n \"content\": \"Data parallelism is realized by having the data processing performed in parallel at different computing resources, with the same model, on different data points. As shown in Figure \\\\ref{fig:datap}, data parallelism is exploited when the ensemble of data points is distributed among different computing resources. During the training process of FL, the training data is not transferred among different computing resources, while the intermediate data, e.g., the models or the gradients $\\\\nabla F_k(x)$ in Formula \\\\ref{eq:gd}, are transferred. \\nThe data in each computing resource can be Independent and Identically Distributed Data (IID) or non-IID. FL focuses on the non-IID , while other distributed machine learning approaches mainly focus on IID data. \\nWith the data parallelism, the FL is horizontal , i.e., the data and the calculation are horizontally distributed among multiple computing resources.\\nIn addition, this parallelism generally corresponds to the cross-device FL , where a large number of devices (mobiles or edge devices) collaboratively participate in training a single global model to have good accuracy.\\nWhen the number of devices is small, e.g., 2-100, and the computing resources are from diverse organizations, this parallelism also corresponds to cross-silo FL . In addition to the general data-parallel schemes for federated learning, some specific privacy-preserved distributed statistical tricks have been invented for federated sparse models~.\\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.75\\\\textwidth]{figures/modelp.pdf}\\n \\\\caption{Model parallelism. The dashed arrows represent inter-computing resource communication. For each input data point $I$, different parts, i.e., $I_{A_1}$ and $I_{A_2}$, are distributed at different computing resources.}\\n \\\\label{fig:modelp}\\n\\\\end{figure}\",\n \"id\": \"0430fcfa-cd31-4b1c-bf51-10d342c0d942\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Parallelism \\\\& FL Types\"\n ],\n [\n \"subsubsection\",\n \"Data Parallelism\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Parallelism\"\n },\n {\n \"cite_extract_rate\": 0.625,\n \"cites\": [\n 600,\n 5439,\n 671,\n 668,\n 591\n ],\n \"content\": \"\\\\label{subsubsec:modelP}\\nModel parallelism is realized by having independent data processing nodes distributed at different computing resources, so as to process the data points of specific features. \\nTwo data processing nodes can be either independent, i.e., the execution of any node does not depend on the output of the other one; or dependent, i.e., there is a data dependency between them .\\nAs shown in Figure \\\\ref{fig:modelp}, model parallelism is achieved when different parts of each data point are distributed at different computing resources. \\nFor instance, the data process on Node $A_1$ and that of $A_2$ can be carried out in parallel.\\nWith the model parallelism, vertical FL, where the data points and calculation are vertically distributed among multiple computing resources , is realized.\\nIn this case, the original model needs to be partitioned to be distributed at different computing resources.\\nTwo organizations generally apply this type of FL when each organization owns parts of the features of users and they would like to collaboratively train a model using the data of all the features, which corresponds to cross-silo FL .\\nMost studies of vertical federated learning only support two parties (with or without a central coordinator) . \\nFor instance, SecureGBM is proposed to train a tree-based Gradient Boosting Machine (GBM).\\nIn order to support multiple parties, the idea of multi-view learning is exploited in a multi-participant, multi-class vertical federated learning framework .\\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.7\\\\textwidth]{figures/pipelinep.pdf}\\n \\\\caption{Pipeline parallelism. The dashed arrows represent inter computing resource communication.}\\n \\\\label{fig:pipelinep}\\n\\\\end{figure}\",\n \"id\": \"62c91cb4-aacb-47de-82b2-8f4949518fdc\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Parallelism \\\\& FL Types\"\n ],\n [\n \"subsubsection\",\n \"Model Parallelism\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model Parallelism\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5425\n ],\n \"content\": \"Pipeline parallelism is realized by having dependent data processing nodes distributed at different computing resources . As shown in Figure \\\\ref{fig:pipelinep}, the data processing nodes are distributed at multiple computing resources. While data point $I^3_A$ is processed in computing resource 1, the outputs of $A_1$ and $A_2$ are processed in computing resource 2, and the outputs of $B_1$, $B_2$, and $B_3$ are processed in computing resource 3. With this type of parallelism, the dependent data processing nodes can process the data in parallel. As this parallelism may incur many inter-computing resource data transfers, it is not widely used for FL.\",\n \"id\": \"1b1418f9-bcb1-4bbe-b99b-80b182bd318e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"9be6bb25-a2cc-4134-abfd-4e31f8d8b0b1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Parallelism \\\\& FL Types\"\n ],\n [\n \"subsubsection\",\n \"Pipeline Parallelism\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pipeline Parallelism\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 636\n ],\n \"content\": \"\\\\label{sebsec:aggre}\\nWith the horizontal FL and data parallelism, aggregation algorithms are used to aggregate the models or gradients generated from the forward and backward propagation in each computing resource. The aggregation algorithms can be either centralized, or hierarchical, and decentralized. The centralized aggregation algorithms generally rely on a centralized server, i.e., a parameter server, to synchronize or schedule the execution of distributed computing resources, while hierarchical aggregation algorithms rely on multiple parameter servers for the model aggregation.\\nThe decentralized aggregation algorithms make each computing resource equally perform the calculation based on a predefined protocol, without relying on a centralized server. \\\\liu{Please refer to for the details of federated optimization. The characteristics are summarized in Table \\\\ref{tal:aggregation}, which can be used to choose appropriate algorithms in a specific situation. }\\n\\\\begin{table}[ht]\\n\\\\centering\\n\\\\caption{\\\\liu{Comparison among diverse types of aggregation algorithms. ``Complexity'' represents the complexity to implement the algorithms (``H'' represents high complexity, ``M'' represents medium complexity, and ``L'' represents low complexity). ``Trust'' represents whether the aggregation algorithms require that the data owners trust the centralized server. ``Imbalance'' represents whether the algorithms can address the unbalanced data. ``High-latency'' represents whether the algorithms can support the high-latency model or gradient data transfer. ``Y'' represents that the algorithms support the functionality while ``N'' represents that the algorithms do not support the functionality.}}\\n\\\\label{tal:aggregation}\\n\\\\begin{tabular}{|c|c|c|c|c|c|}\\n\\\\hline\\nType & Complexity & Trust & Imbalance & High-latency \\\\\\\\\\n\\\\hline\\nCentralized & L & Y & N & N \\\\\\\\\\n\\\\hline\\nHierarchical & M & Y & Y & Y \\\\\\\\\\n\\\\hline\\nDecentralized & H & N & N & Y \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"10053136-8ab4-4b06-aec3-a7a9b0e56bc3\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"5dfeced7-075b-4276-abd0-1ff3e9bb48a7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Aggregation Algorithms\"\n ]\n ],\n \"subsections\": [\n \"e5dd26ef-00aa-4b84-aa33-7ff2605aebde\",\n \"508677b5-4dbe-4b11-ad6a-7789cb9e091d\",\n \"ecce2209-838c-4800-b0a9-49b94ea5037a\"\n ],\n \"title\": \"Aggregation Algorithms\"\n },\n {\n \"cite_extract_rate\": 0.8125,\n \"cites\": [\n 635,\n 602,\n 7928,\n 5440,\n 5443,\n 5441,\n 5445,\n 8919,\n 5442,\n 582,\n 5444,\n 5446,\n 632\n ],\n \"content\": \"\\\\label{subsubsec:centralized}\\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.7\\\\textwidth]{figures/cenagg.png}\\n \\\\caption{The architecture of centralized aggregation. ``CR'' represents computer resource. $\\\\omega$ represents the local parameters or the weights of the model calculated in each computing resource. $\\\\mathrm{g}$ represents the local gradients in backward propagation in each computing resource. $\\\\overline{\\\\omega}$ represents the global model calculated in the parameter server. $\\\\overline{\\\\mathrm{g}}$ represents the global gradients calculated in the parameter server.}\\n \\\\label{fig:centra}\\n\\\\end{figure}\\nAs shown in Figure \\\\ref{fig:centra}, a single parameter server is used to calculate the average models or gradients sent from multiple computing resources (mobiles).\\nThe weights of the model (model) or the gradients are calculated in each computing resource, which are transferred to a parameter server. \\nThe parameter server calculates global gradients or global models according to a centralized aggregation algorithm.\\nThe global gradients or global models are transferred to each computing resource for the following computation. \\nThe update of the model is based on the SGD defined in Formula \\\\ref{eq:gd} in both computing resources, or on the parameter server.\\nA bunch of centralized aggregation algorithms have been proposed.\\nFederated Averaging (FedAvg) algorithm is introduced as the aggregation method in Google's implementation of an FL system. A centralized server aggregates the machine learning models from selected users. Then, a global model is generated using a weighted sum of each aggregated machine learning model. Afterward, the global model is shared with selected users, and the training process is continued in the computing resource of selected users. \\nHowever, the trained model of FedAvg may be biased towards computing resources with favorable network conditions . \\nWhile FedAvg is a straightforward approach, some other methods are proposed to address additional problems. \\nA Federated Stochastic Block Coordinate Descent (FedBCD) algorithm is proposed to reduce the number of communication rounds by enabling multiple local updates before the model communication between a user and the server. In addition, FedBCD also considers the regularization during the training process. The training problem with regularization can be formulated as:\\n\\\\begin{equation}\\n\\\\label{eq:regularizer}\\n \\\\widehat F = \\\\underset{F_\\\\theta\\\\in\\\\mathcal{H}}{\\\\mathrm{argmin}} \\\\underset{(x,y)\\\\sim\\\\mathcal{D}}{\\\\mathbb{E}} L(y, F_\\\\theta(x)) + \\\\lambda \\\\cdot \\\\gamma(\\\\theta),\\n\\\\end{equation}\\nwhere $F$, $D$, $\\\\mathcal{H}$ are the same as those in Formula \\\\ref{eq:problem}, while $\\\\gamma(\\\\cdot)$ denotes the regularizer and $\\\\lambda$ is the hyper-parameter. The regularization is exploited to improve the generalization capacity of the trained machine learning model. As the fairness among multiple users is important for an FL system, the Stochastic Agnostic Federated Learning (SAFL) algorithm and the FedMGDA+ algorithm are proposed to achieve fairness during the training process of FL. The fairness represents that the data distribution among multiple users can be equally considered without the influence of unrelated factors. \\nFairness may also refer to two other concepts: (1) A user gets a final model according to the contribution ; and/or (2) Uniform accuracy distribution among all the distributed computing resources , which are out of the scope of this paper.\\nIn addition, while the computing resources may be heterogeneous, FedProx is proposed to tackle the heterogeneity in an FL system. FedProx enables multiple iterations in each computing resource, while minimizing a cost function based on the local loss function and the global model. Furthermore, in order to address permutation of data processing nodes during the training process, Federated Matched Averaging (FedMA) is proposed. FedMA exploits an existing approach, i.e., BBP-MAP , to generate a matrix, in order to align the data processing nodes of the models from computing resources and the server. SCAFFOLD is proposed to reduce the communication rounds, using stateful variables in the distributed computing resources. Attention-augmented mechanism is exploited in Attentive Federated Aggregation (FedAttOpt) to aggregate the knowledge generated from each computing resource (client), based on the contribution of the model from each client. \\nWhen the data distribution is heterogeneous among users, personalization remains an open problem. In order to address this problem, the model can be split into local layers and global layers, which has been proposed in adaptive personalized federated learning (APFL) , FedPer , and pFedMe . The local layers are trained with the decentralized data in each computing resource of users, while the global layers are trained in the computing resources of users and the server. However, it is difficult to choose a dataset and its partition among clients to measure the personalization brought by APFL or FedPer, so as to prove the improvement compared with FedAvg.\\nThe attention-augmented mechanism helps reduce the communication rounds.\\nIn addition, knowledge distillation can also be exploited to aggregate the models, while requiring that there is data in the centralized server .\\nAll these algorithms can handle non-IID data. A comparison among the aforementioned algorithms is proposed in Table \\\\ref{tal:aggre}. \\n\\\\begin{table}[ht]\\n\\\\centering\\n\\\\caption{Comparison among aggregation algorithms. ``Reg'' represents regularization. Heterogeneity represents that the computing resources are heterogeneous. ``Firness'' represents that the data distribution among multiple users can be equally considered without the influence of unrelated factors. ``Permutation'' refers to the permutation of data processing nodes during the training process. ``C-E'' represents communication efficient. ``S'' represents that the algorithm supports the functionality, while ``N'' represents that the algorithm does not have support. }\\n\\\\label{tal:aggre}\\n\\\\begin{tabular}{|c|c|c|c|c|c|}\\n\\\\hline\\nAlgorithm & Reg & Fairness & Heterogeneity & Permutation & C-E \\\\\\\\\\n\\\\hline\\nFedAvg & N & N & N & N & N \\\\\\\\\\n\\\\hline\\nFedBCD & S & N & N & N & S \\\\\\\\\\n\\\\hline\\nSAFL & N & S & N & N & N \\\\\\\\\\n\\\\hline\\nFedMGDA+ & N & S & N & N & S \\\\\\\\\\n\\\\hline\\nFedProx & N & N & S & N & S \\\\\\\\\\n\\\\hline\\nFedMA & N & N & N & S & S \\\\\\\\\\n\\\\hline\\nSCAFFOLD& N & N & N & N & S \\\\\\\\\\n\\\\hline\\nFedAttOpt& N & N & N & N & S \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"e5dd26ef-00aa-4b84-aa33-7ff2605aebde\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"10053136-8ab4-4b06-aec3-a7a9b0e56bc3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Aggregation Algorithms\"\n ],\n [\n \"subsubsection\",\n \"Centralized Aggregation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Centralized Aggregation\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 646,\n 644,\n 5448,\n 5447,\n 5449\n ],\n \"content\": \"\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.95\\\\textwidth]{figures/hier.png}\\n \\\\caption{The architecture of hierarchical aggregation. ``CR'' represents computer resource. $\\\\omega$ represents the local parameters or the weights of the model calculated in each computing resource. $\\\\mathrm{g}$ represents the local gradients in backward propagation in each computing resource. $\\\\overline{\\\\omega}$ represents the region or global model calculated in each parameter server. $\\\\overline{\\\\mathrm{g}}$ represents the region or global gradients calculated in each parameter server. The region model or gradients are calculated by a region parameter server, while the global model or gradients are calculated by a global parameter server.}\\n \\\\label{fig:hier}\\n\\\\end{figure}\\nAs shown in Figure \\\\ref{fig:hier}, a hierarchical architecture is also exploited using multiple parameter servers. \\nA two-layer hierarchical architecture is proposed to reduce the time to transfer models between a parameter server and computing resources . The hierarchical architecture uses a global parameter server (GPS) and multiple region parameter servers. Each region parameter server (RPS) is implemented in a cell base station where the computing resources (mobiles) can be connected, with low latency. A hierarchical algorithm, i.e., Hierarchical Federated Learning (HFL) is deployed to realize the model aggregation. Within each iteration of HFL, each RPS calculates an average model using the models of the computing resources within its cluster. It sends the averaged model to the GPS, and it receives a global averaged model at every certain iteration. Afterward, it broadcasts the averaged model to all its computing resources. Some other algorithms, e.g., HierFAVG , HFEL , and LanFL , are similar to HFL, while the SPS is an edge or Local-Area Network (LAN) parameter server and the MPS is a parameter server implemented on the cloud or a Wide-Area Network (WAN). These algorithms take advantage of hierarchical architecture to reduce high-latency model or gradient data transfer, so as to accelerate the training process. In addition, by well-clustering the computing resources to groups, the hierarchical architecture is also exploited to address unbalanced data distributed among multiple computing resources , or to address data privacy .\",\n \"id\": \"508677b5-4dbe-4b11-ad6a-7789cb9e091d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"10053136-8ab4-4b06-aec3-a7a9b0e56bc3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Aggregation Algorithms\"\n ],\n [\n \"subsubsection\",\n \"Hierarchical Aggregation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hierarchical Aggregation\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5452,\n 8920,\n 5450,\n 7930,\n 5451,\n 7929\n ],\n \"content\": \"\\\\label{subsubsec:decentralized}\\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.8\\\\textwidth]{figures/decentral.png}\\n \\\\caption{The architecture of decentralized aggregation. ``CR'' represents computer resource. $\\\\omega$ represents the local parameters, or the weights of the model calculated in each computing resource. $\\\\mathrm{g}$ represents the local gradients in backward propagation in each computing resource. When two computing resources are neighbors, they can communicate with each other.}\\n \\\\label{fig:decentralized}\\n\\\\end{figure}\\nWhile collaboratively training a machine learning model with a decentralized aggregation algorithm, the computing resources can be organized with a {\\\\em connected} topology and can communicate with a peer-to-peer manner, as shown in Figure \\\\ref{fig:decentralized}. The degree and connectivity of the topology affects the communication efficiency and the convergence rate of the aggregation algorithm. \\nFor a given topology, we define $w_{i,j}$, the weight to scale information flowing from node $j$ to node $i$, as follows\\\\vspace{-2mm}\\n\\\\begin{align}\\\\label{wij}\\nw_{ij}\\n\\\\begin{cases}\\n> 0 & \\\\mbox{if node $j$ is connected to node $i$, or $i=j$;} \\\\\\\\\\n= 0 & \\\\mbox{otherwise.}\\\\vspace{-2mm}\\n\\\\end{cases}\\n\\\\end{align}\\nWe further define the {\\\\bf topology matrix} $W = [w_{ij}]_{i,j=0}^{n-1} \\\\in \\\\mathbb{R}^{n\\\\times n}$ as the matrix to represent the topology. \\nIn the remainder of this paper, \\nwe assume that $W$ satisfies \\n$W\\\\mathds{1} = \\\\mathds{1}$ \\nand $\\\\mathds{1}^T W = \\\\mathds{1}^T$, \\ni.e., both the row sum and the column sum of $W$ are equal to $1$, \\nso as to guarantee that the neighborhood averaging will asymptotically approach the global averaging . \\nWhen a computing resource $j$ is directly connected to computing resource $i$, i.e., $w_{i,j} \\\\neq 0$, computing resource $j$ is the neighbor of computing resource $i$.\\n\\\\liu{Please note that the weight $w_{i,j}$ denotes the confidence node $i$ has in the information it receives from node $j$ , which is different from the bandwidth or data transfer capacity in a network.}\\nThe centralized aggregation algorithm is a special type of decentralized aggregation with a star topology while only the centralized server communicates with its neighbors.\\n\\\\liu{A well designed topology, e.g., an exponential graph , can improve the convergence rate, which accelerates the training speed.}\\nWith the decentralized SGD (D-SGD), each computing resource maintains a local copy of \\nthe global model parameters, and it updates the local copy using the models of its neighbors. \\nAccording to the order to conduct neighborhood averaging and gradient descent, D-SGD has two common types of realizations: Average-With-Communication (AWC) and Average-Before-Communication (ABC) .\\nAWC can overlap communication and gradient computation, while ABC needs to sequentially calculate and communicate the gradient or model. \\nHowever, ABC is robust , and it converges fast in terms of iterations by exploiting its large learning rate.\\nIn addition, the decentralized aggregation algorithms can be classified to Full Communication (FC) and Partial Communication (PC) according to the number of neighbors. Within the iterations of FC, each computing resource calculates an averaged model or gradient, based on all the models or gradients of the last version from all its neighbors. However, within the iterations of PC, each computing resource calculates an averaged model or gradient based on one or multiple chosen neighbors. With PC, the selection of the neighbors can be based on a gossip algorithm . For instance, a random neighbor can be selected ; the neighbors that provide benign models are selected to avoid attack .\",\n \"id\": \"ecce2209-838c-4800-b0a9-49b94ea5037a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"10053136-8ab4-4b06-aec3-a7a9b0e56bc3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Distributed Training\"\n ],\n [\n \"subsection\",\n \"Aggregation Algorithms\"\n ],\n [\n \"subsubsection\",\n \"Decentralized Aggregation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Decentralized Aggregation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:security}\\nAt the infrastructure layer of an FL system, there are three types of data manipulation: data security mechanisms, data transfer, and distributed data processing within the distributed execution module. \\nWe first present the techniques for the distributed execution in an FL system. \\nThen, we present the techniques for data transfer during the training process of an FL system.\\nFinally, as data security is of much importance to an FL system , we present the techniques to protect the data security.\",\n \"id\": \"cd0b3f18-6523-488f-aaa8-addaf533ff9f\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Data Manipulation\"\n ]\n ],\n \"subsections\": [\n \"8048e455-33dd-411a-bdf6-5bafe4141358\",\n \"09ac7749-f7de-46a4-bc72-304efd143b4a\",\n \"4a182037-a6a4-48d6-aa9c-0986bac2c4b8\"\n ],\n \"title\": \"Data Manipulation\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5454,\n 623,\n 655,\n 659,\n 5453\n ],\n \"content\": \"While the bandwidth within a single data center is high, e.g., InfiniBand, the High Performance Computing (HPC) libraries, e.g., Message Passing Interface (MPI) or NVIDIA Collective Communications Library (NCCL) , are widely exploited for distributed data processing . With MPI or NCCL, the gradients or models in each computing resource can be easily calculated using ring-AllReduce algorithm . However, one of the drawbacks of the HPC libraries is that they lack support for fault-tolerance, as the HPC libraries are designed for high performance servers with high quality networks. When any computing resource within the network becomes unavailable, the distributed training process may be broken.\\nHowever, as an FL system is generally implemented for the collaboration of large amounts of mobile device users or different organizations, the network connection among computing resources is of moderate quality, i.e., the bandwidth is not as good as that within a single data center, and the latency is high.\\nFor instance, the Internet upload speed is typically much slower than the download speed . Also, some users with unstable wireless communication channels may consequently drop out due to disconnection from the Internet .\\nIn this environment, the connection between computing resources and parameter servers has a high possibility of becoming disabled. Thus, Remote Procedure Call (RPC) frameworks are widely exploited, as this kind of framework can ignore the disconnected computing resources and continue the distributed training of an FL system , e.g., PaddleFL , PySyft , or TensorflowFL .\",\n \"id\": \"8048e455-33dd-411a-bdf6-5bafe4141358\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"cd0b3f18-6523-488f-aaa8-addaf533ff9f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Data Manipulation\"\n ],\n [\n \"subsection\",\n \"Distributed Data Processing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Distributed Data Processing\"\n },\n {\n \"cite_extract_rate\": 0.769230769230769,\n \"cites\": [\n 613,\n 7932,\n 623,\n 622,\n 7931,\n 5456,\n 5457,\n 5455,\n 625,\n 8366\n ],\n \"content\": \"As the network connection is of moderate quality, the data transfer module mainly focuses on data compression to transfer intermediate data, e.g., gradients or models. \\nSketched updates are proposed for gradient compression to accelerate the data transfer during the distributed training within a single data center .\\nWith the data parallelism and centralized aggregation algorithm, before sending the intermediate data, the intermediate data can be sketched with subsampling , quantization , sparsification , or projection to lower dimensional spaces , in each computing resource, in order to reduce the cost to transfer data.\\nSubsampling refers to transferring only a random subset of the intermediate data .\\nQuantization methods encode each value using a fixed number of bits, so as to reduce the length of gradients or models .\\nWith the sparsification approach, only selected parts of the intermediate data are transferred, while the selection is based on a threshold, e.g., the gradients larger than a threshold are selected . Then, when the intermediate data is received in the server, they are decompressed to be aggregated according to the aggregation algorithms presented in Section \\\\ref{subsubsec:centralized}. The convergence of the quantization approach is analyzed in , which shows that this approach can also provide good convergence rates . In addition, irrelevant intermediate data can be precluded to be transferred to the server, in order to substantially reduce the communication overhead .\",\n \"id\": \"09ac7749-f7de-46a4-bc72-304efd143b4a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"cd0b3f18-6523-488f-aaa8-addaf533ff9f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Data Manipulation\"\n ],\n [\n \"subsection\",\n \"Data Transfer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Transfer\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Data security is of much importance for data processing. The problem of data security is related to significant financial and reputational losses. For instance, Uber had to pay \\\\$148,000,000 to settle the investigation incurred by a breach of 600,000 drivers’ personal information in 2016 . \\nData security mainly includes two aspects, i.e., data privacy and model security.\\nData privacy refers to the protection of raw data to avoid raw data information leakage during or after the distributed training of FL systems.\\nModel security refers to the protection of the security of trained models, in order to avoid wrong output based on the trained models incurred by malicious attacks.\\nIn this section, we first present the techniques to protect data privacy. Then we present the defense methods for model security.\",\n \"id\": \"4a182037-a6a4-48d6-aa9c-0986bac2c4b8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"cd0b3f18-6523-488f-aaa8-addaf533ff9f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Data Manipulation\"\n ],\n [\n \"subsection\",\n \"Data Security\"\n ]\n ],\n \"subsections\": [\n \"7a14e4c6-106a-4ec2-8926-70aed111f75a\",\n \"cecfb28f-1aff-43d4-9671-2b7aa114b7c6\"\n ],\n \"title\": \"Data Security\"\n },\n {\n \"cite_extract_rate\": 0.48571428571428504,\n \"cites\": [\n 3469,\n 5460,\n 3477,\n 892,\n 5459,\n 5461,\n 5462,\n 5439,\n 5458,\n 3478,\n 7933,\n 614,\n 5463,\n 7608,\n 5441,\n 5464,\n 7727\n ],\n \"content\": \"The techniques to protect data privacy consist of three types: Trusted Execution Environment (TEE), encryption, Differential Privacy (DP), and anti-Generative Adversarial Network (GAN) methods. These techniques can be combined in FL systems, e.g., the combination of DP and TEE in , the combination of encryption and DP , and the combination of DP and anti-GAN .\\nA TEE is an environment where the execution is secured and no information can be leaked to unauthorized users. \\nIntel SGX technique has been first proposed as a secure environment while providing a set of security-related instruction codes built within Intel Central Processing Units (CPUs). \\nThen, the implementation of machine learning models has been carried out in the TEE, i.e., Intel SGX, in order to enable collaborative data analysis based on machine learning algorithms while providing a security guarantee . Afterwards, the TEE has been exploited in FL systems, in order to protect the privacy of data in two ways. The first way is to put the entire training process in the TEE of each distributed computing resource to protect the data privacy during the distributed training . The second way is to use TEE to check a small part of the distributed training, while exploiting insecure computing resources, e.g., GPUs, to reduce the training time . \\nAs an encryption technique, homomorphic encryption has been used to ensure the data privacy for FL systems .\\nHomomorphic encryption allows specific types of computations to be carried out on encrypted input data, and to generate an encrypted result, which matches the result of the same computations on the decrypted input data. \\n\\\\liu{Two main branches of homomorphic encryption exist, i.e., fully homomorphic encryption and partially homomorphic encryption. \\nThe fully homomorphic encryption supports both addition and multiplication on ciphertext, while partially homomorphic encryption only supports either an addition or a multiplication operation on ciphertext, which corresponds to less computational flexibility and better runtime efficiency. \\nBoth the fully and partially homomorphic encryptions can be exploited with the horizontal and vertical federated learning.}\\nAs sharing gradients also leaks the information of training data \\\\liu{in horizontal federated learning} , it is of much importance to protect the privacy of the intermediate data.\\nThus, the intermediate data can be encrypted using a homomorphic encryption algorithm before being sent to a parameter server . In this way, the intermediate data remain encrypted during the aggregation process while only the computing resource can decrypt the encrypted intermediate data. \\nEven if the transferred encrypted intermediate data is leaked, the information of gradients or models remains safe, and the privacy of the training data is ensured.\\n\\\\liu{In addition, partial homomorphic encryption, e.g., Paillier , is exploited in vertical federated learning .}\\nHowever, the homomorphic encryption incurs significant costs in computation and communication during distributed training . In order to reduce the overhead of homomorphic encryption, a set of quantized gradients are encrypted . \\nDifferential Privacy (DP) protects the data privacy by adding artificial noise to a small part of raw data, while ensuring that the modification does not substantially affect the performance of the machine learning models . \\nDP is widely used in FL systems as the first step to process the raw data, and the output is the training data to be used for the distributed training . With more added noise, the privacy is better protected, i.e., there is less possibility to leak raw data information, while it takes more time to converge for the machine learning models . A trade-off between the privacy protection and the convergence performance can be made by selecting a certain number of distributed resources . However, DP may not be able to ensure the data privacy under certain attacks, e.g., Generative Adversarial Network (GAN) attacks .\\nA well-trained machine learning model can leak information about the training data based on the intermediate data, e.g., gradients . GANs can be used to generate data similar to the training data based on a well-trained machine learning model in either a parameter server or a distributed computing resource . \\nThe adversary can reconstruct other participating clients’ private data, even if it has no knowledge of the label information using the GANs.\\nThus, during the distributed training process of FL systems, a malicious user can exploit GANs to infer the training data of other users. DP can be used to prevent the GAN-based attack . In addition, fake training data can be generated based on a GAN and original raw data, which is then used during the distributed training process to prevent the GAN-based attack .\",\n \"id\": \"7a14e4c6-106a-4ec2-8926-70aed111f75a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 35,\n \"parent_id\": \"4a182037-a6a4-48d6-aa9c-0986bac2c4b8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Data Manipulation\"\n ],\n [\n \"subsection\",\n \"Data Security\"\n ],\n [\n \"subsubsection\",\n \"Data Privacy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Privacy\"\n },\n {\n \"cite_extract_rate\": 0.923076923076923,\n \"cites\": [\n 2673,\n 5468,\n 5470,\n 5465,\n 5451,\n 5471,\n 5467,\n 7038,\n 3415,\n 597,\n 5469,\n 5466\n ],\n \"content\": \"We mainly focus on poisoning attacks in this section. \\nThe objective of poisoning attacks is to reduce the accuracy of machine learning models using artificially designed data, i.e., data poisoning, or models, i.e., model poisoning, in one or several distributed computing resources during the model aggregation process (see details in in Section \\\\ref{subsubsec:centralized}). There are two ways to carry out poisoning attacks, i.e., data poisoning and model poisoning. \\nData poisoning can be realized by modifying the features or the labels of the input data. For instance, malicious users can modify the data points of a certain class $C$ to other classes, and they can then use the modified data points to participate in the distributed training. \\nThe modification of the labels is denoted by the label flipping attack.\\nAs a result, the accuracy of the trained model has low accuracy in terms of Class $C$ . \\nModel poisoning refers to the attacks in which the updated intermediate data, e.g., gradients or models, are poisoned before being sent to a parameter server in order to reduce the accuracy of the trained model . The goal of the model poisoning is to reduce the performance of the trained model on targeted tasks or classes, while the performance of the model remains unchanged in terms of other tasks or classes . \\nData poisoning eventually realizes the model poisoning, as it enables some computing resources to update poisoned intermediate data based on the calculation of poisoned training data . \\nHowever, model poisoning can be more powerful than data poisoning, as model poisoning directly influences the weights of the models and trains in a way that benefits the attack .\\nBoth the data poisoning and the model poisoning rely on the backdoor attacks to modify the training data or the intermediate data . Backdoor attacks are performed by embedding the hidden instructions into machine learning models, so that the infected model performs well on benign testing samples when the backdoor is not activated, while its prediction will be changed to the attacker-specified target label when the backdoor is activated by the attacker . \\nIn order to defend against these data attacks or model attacks, the malicious users should be identified by analyzing the updated intermediate data using dimensionality reduction methods, e.g., Principal Component Analysis (PCA) , anomaly detection , or interpretability techniques .\\nIn addition, the model poisoning can be incurred by Byzantine failures of certain distributed computing resources . With Byzantine failures, some computing resources (bad users) are manipulated by attackers during the distributed training process, which significantly degrades the performance of the global model in terms of test error . \\nIn order to make the training process robust against the Byzantine failures, the bad users can be identified by analyzing the updated intermediate data using a hidden Markov model or via secure aggregation protocols .\",\n \"id\": \"cecfb28f-1aff-43d4-9671-2b7aa114b7c6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"4a182037-a6a4-48d6-aa9c-0986bac2c4b8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Data Manipulation\"\n ],\n [\n \"subsection\",\n \"Data Security\"\n ],\n [\n \"subsubsection\",\n \"Model Security\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model Security\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 5420,\n 5423\n ],\n \"content\": \"\\\\label{sec:frameworks}\\nFL systems are widely applied in diverse domains, e.g., \\nmobile service, healthcare , and finance .\\nAn FL system generally exploits an FL framework, which is deployed on distributed resources.\\nIn this section, we present four widely used FL frameworks: PaddleFL , TensorFlowFederated , FATE , and PySyft .\",\n \"id\": \"727148cb-019c-4e5c-9d02-e8f0c8fb671c\",\n \"level\": \"section\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Federated Learning Frameworks\"\n ]\n ],\n \"subsections\": [\n \"37621a76-9ea9-4868-9b62-1b2d192ce562\",\n \"c27dbf26-a266-414a-aa80-dc2e4d0fee2b\",\n \"d79e2f48-eba0-495e-b5ac-cb343eff8f4d\",\n \"baf085e7-c8e0-4563-86bc-a6f1583fb40d\",\n \"c6708387-e962-4e76-9ea7-b422245775eb\"\n ],\n \"title\": \"Federated Learning Frameworks\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 5422\n ],\n \"content\": \"PaddleFL is an open source federated learning framework based on PaddlePaddle , which is supported by Baidu. At the presentation layer, PaddleFL provides a textual UI for the interaction between users and the FL system. At the User Services layer, PaddleFL provides the log and monitoring supports, and it can leverage the interpretability module of PaddlePaddle in the future. At the FL training layer, PaddleFL can realize data parallelism (horizontal FL) and model parallelism (vertical FL). It supports multiple aggregation algorithms, e.g., FedAvg, and fault-tolerance. At the infrastructure layer, PaddleFL exploits RPC for the distributed execution. PaddleFL exploits DP to protect the data security. PaddleFL is widely used in multiple domains, e.g., Natural Language Processing (NLP), Computing Vision (CV) , and recommendation.\",\n \"id\": \"37621a76-9ea9-4868-9b62-1b2d192ce562\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"727148cb-019c-4e5c-9d02-e8f0c8fb671c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Federated Learning Frameworks\"\n ],\n [\n \"subsection\",\n \"PaddleFL\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"PaddleFL\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 5419,\n 582\n ],\n \"content\": \"TensorFlow Federated (TFF) is an open-source framework for federated learning on decentralized data, which is supported by Google. TFF also provides a textual UI through Python. TFF supports the monitoring and log functionality at the user service layer. TFF supports data parallelism (horizontal FL), multiple aggregation algorithms, and fault-tolerance of mobile devices. TFF exploits RPC for the distributed execution and DP for the protection of data privacy.\\nTFF enables Android mobile users to predict the next word while using the keyboard on their mobile phones .\",\n \"id\": \"c27dbf26-a266-414a-aa80-dc2e4d0fee2b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"727148cb-019c-4e5c-9d02-e8f0c8fb671c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Federated Learning Frameworks\"\n ],\n [\n \"subsection\",\n \"TensorFlowFederated\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"TensorFlowFederated\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"FATE is an open-source FL framework supported by WeBank. \\nFATE provides both a graphical and textual UI. \\nFATE can support the monitoring of distributed training through a web portal. \\nFATE takes advantage of database management systems (DBMS) to track the execution status.\\nFATE can enable horizontal (data parallelism), vertical (model parallelism), and hybrid federated learning. \\nFATE exploits both the DP and HE to protect the data privacy. \\nIn addition, FATE exploits RPC to perform the distributed execution.\",\n \"id\": \"d79e2f48-eba0-495e-b5ac-cb343eff8f4d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"727148cb-019c-4e5c-9d02-e8f0c8fb671c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Federated Learning Frameworks\"\n ],\n [\n \"subsection\",\n \"FATE\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"FATE\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 656\n ],\n \"content\": \"PySyft is an open-source FL framework based on the PyTorch framework .\\nPySyft is written in Python and provides a textual UI based on Python.\\nPySyft mainly supports the data parallelism and model parallelism based on an aggregator or orchestrating server. The aggregator or orchestrating server sends a part of the model to participating clients to process local data and gets results for federated averaging. PySyft exploits DP and encryption techniques to protect the data security. PySyft exploits multiple communication protocols for distributed execution, e.g., RPC, websocket etc.\",\n \"id\": \"baf085e7-c8e0-4563-86bc-a6f1583fb40d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"727148cb-019c-4e5c-9d02-e8f0c8fb671c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Federated Learning Frameworks\"\n ],\n [\n \"subsection\",\n \"PySyft\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"PySyft\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 7934\n ],\n \"content\": \"}\\n\\\\liu{\\nDiverse FL frameworks exist while each has its advantage. We summarize the characteristics of each framework in Table \\\\ref{tal:FLFramework}, so as to help select a proper framework for use. From the table, we can see that all the frameworks implement the centralized aggregation algorithms, while employing DP and HE for the data security. PaddleFL can exploit Paddle to realize data, model, and pipeline parallelism. FATE and TFF are based on Tensorflow as the engine, while FATE can provide Web portal UI, which is convenient for novices. PySyft is compatible with PyTorch, which can easily handle the PyTorch-based tasks, while PaddleFL is compatible with Paddle, which can easily deal with rich pre-trained models published in PaddleHub .}\\n\\\\begin{table}[ht]\\n\\\\centering\\n\\\\caption{\\\\liu{Comparison among diverse frameworks. ``Aggregation'' represents the type of aggregation algorithms. ``Textual'' represents the textual UI, while ``Web'' represents Web portal.}}\\n\\\\label{tal:FLFramework}\\n\\\\begin{tabular}{|c|c|c|c|c|c|}\\n\\\\hline\\nFramework & Engine & Aggregation & UI & Parallelism & Security \\\\\\\\\\n\\\\hline\\nPaddleFL & Paddle & Centralized & textual & Data/Model/Pipeline & DP/HE \\\\\\\\\\n\\\\hline\\nTFF & TensorFlow & Centralized & textual & Data/Model & DP/HE \\\\\\\\\\n\\\\hline\\nFATE & TensorFlow & Centralized & Web & Data/Model & DP/HE \\\\\\\\\\n\\\\hline\\nPySyft & PyTorch & Centralized & textual & Data & DP/HE \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\\n\\\\liu{Table \\\\ref{tal:FrameworkSupport} represents the support of diverse types of FL in terms of data distribution. All the frameworks support horizontal FL, while vertical FL is supported by three frameworks except TFF. PySyft cannot directly support the vertical FL, while PyVertical , which is built upon PySyft, can be used to support vertical FL with the compatibility of PyTorch models. The hybrid FL is only supported by Paddle and FATE. In addition, all the frameworks support the execution with GPU. In practice, although PaddleFL may correspond to slightly longer time, the accuracy of the trained model can be higher that of TFF and FATE, while PySyft may generate ``out of memory'' errors .}\\n\\\\begin{table}[ht]\\n\\\\centering\\n\\\\caption{\\\\liu{Comparison among diverse frameworks for the support of diverse FL types, e.g., horizontal FL, vertical FL, and hybrid FL, and GPU. $\\\\|$\\\\checkmark$\\\\|$ represents that the support is not realized by itself but a close one.}}\\n\\\\label{tal:FrameworkSupport}\\n\\\\begin{tabular}{|c|c|c|c|c|c|}\\n\\\\hline\\n\\\\multicolumn{2}{|c|}{} & PaddleFL & TFF & FATE & PySyft \\\\\\\\\\n\\\\hline\\n\\\\multirow{3}{*}{Types} & Horizontal & \\\\checkmark & \\\\checkmark & \\\\checkmark & \\\\checkmark \\\\\\\\\\n\\\\cline{2-6}\\n& Vertical & \\\\checkmark & \\\\xmark & \\\\checkmark & $\\\\|$\\\\checkmark$\\\\|$ \\\\\\\\\\n\\\\cline{2-6}\\n& Hybrid & \\\\checkmark & \\\\xmark & \\\\checkmark & \\\\xmark \\\\\\\\\\n\\\\hline\\n\\\\multicolumn{2}{|c|}{GPU} & \\\\checkmark & \\\\checkmark & \\\\checkmark & \\\\checkmark \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"c6708387-e962-4e76-9ea7-b422245775eb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"727148cb-019c-4e5c-9d02-e8f0c8fb671c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Federated Learning Frameworks\"\n ],\n [\n \"subsection\",\n \"\\\\liu{Concluding Remarks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"\\\\liu{Concluding Remarks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:future}\\nAlthough much work has been done on the FL systems, there remain\\nsome limitations, e.g., interpretability of FL, decentralized aggregation, FL on graphs, benchmarks of FL systems, and applications to distributed intelligent systems. This section\\ndiscusses the limitations of the existing frameworks and proposes new research directions.\",\n \"id\": \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Research Directions\"\n ]\n ],\n \"subsections\": [\n \"9866dd12-79cc-42bf-8557-f19459ea5936\",\n \"b6673d3f-9d3e-4599-b8f4-93fdf908307f\",\n \"9c3427af-198a-448e-bda6-b7fbe346fc0a\",\n \"35e51ec5-ea11-4e9b-a3ee-cbf4853180ca\",\n \"e54c8539-ee3c-4e42-b5a8-6e639fe2d31f\",\n \"31802936-db7d-403f-b909-ee8ee4d7b0c8\"\n ],\n \"title\": \"Research Directions\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 653,\n 582\n ],\n \"content\": \"Several datasets exist for experiments on FL systems. For instance, Federated Extended MNIST (FEMNIST) is built by partitioning the data in Extended MNIST based on each writer.\\nShakespeare is built from The Complete Works of William Shakespeare based on each speaking role. \\nBoth of these datasets can be used for horizontal FL. However, no public datasets exist for vertical FL or transfer FL. In addition, no open decentralized IID or non-IID distribution of popular datasets, e.g., ImageNet , exist for FL systems.\",\n \"id\": \"9866dd12-79cc-42bf-8557-f19459ea5936\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Research Directions\"\n ],\n [\n \"subsection\",\n \"Benchmarks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Benchmarks\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 5421,\n 5437\n ],\n \"content\": \"Deep neural networks have excellent performance in various areas, while it is often difficult to understand the results of deep neural network models, especially within FL systems.\\nShapley values have been used to provide the interpretability , while it focuses on vertical FL. When multiple users collaboratively train an FL model, it remains an open problem to evaluate the contributions of each user, which helps provide evidence for the incentive of each user. \\nThe primary incentive for clients to participate in federated learning is obtaining better models , while the benefit of participating in federated learning for clients who have sufficient private data to train accurate local models is disputable. \\nInterpretability can help understand the contributions of each user and provide an objective opinion on the incentive strategy within an FL system.\\nIn addition, the interpretability helps domain experts to understand the relationship between data and the final trained model in critical domains, e.g., healthcare and finance.\\nHowever, the interpretability within FL systems remains an open problem.\",\n \"id\": \"b6673d3f-9d3e-4599-b8f4-93fdf908307f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Research Directions\"\n ],\n [\n \"subsection\",\n \"Interpretability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Interpretability\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Current aggregation algorithms of FL systems focus on the full connection or star connection topology, while other topologies, e.g., dynamic exponential-2 graph, may help accelerate the distributed training with FL systems . \\\\liu{In addition, well-known graph algorithms, e.g., graph partitioning algorithms, and ad-hoc policies can be exploited to help better distribute computing resources with the topology defined in Section \\\\ref{subsubsec:decentralized} in order to improve the efficiency of FL systems. } While the peer-to-peer communication enables the FL with an arbitrary topology matrix, the data security under diverse attacks, e.g., data or model poisoning, GAN-based attacks, remain open problems and deserve further investigation.\",\n \"id\": \"9c3427af-198a-448e-bda6-b7fbe346fc0a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Research Directions\"\n ],\n [\n \"subsection\",\n \"Decentralized Aggregation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Decentralized Aggregation\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 7935,\n 5438,\n 180\n ],\n \"content\": \"Graphs or graph neural networks (GNN) have gained increasing popularity in multiple domains, e.g., social network, knowledge graph, and recommender system. FL frameworks for graphs, i.e., GraphFL , and GNN, i.e., SGNN , have been proposed to train a model with decentralized graphs. However, the data security of FL on graphs remains an open problem. \\\\liu{In addition, while a multimodal knowledge graph could not only contain text but also images or other type of data , it is worth further exploration to efficiently support the multimodel knowledge graph construction within an FL system .}\",\n \"id\": \"35e51ec5-ea11-4e9b-a3ee-cbf4853180ca\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Research Directions\"\n ],\n [\n \"subsection\",\n \"Federated Learning on Graphs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Federated Learning on Graphs\"\n },\n {\n \"cite_extract_rate\": 0.375,\n \"cites\": [\n 8921,\n 5472,\n 5473\n ],\n \"content\": \"}\\n\\\\liu{\\nAlthough FL focuses on the non-IID data, the real-world decentralized data usually exhibit an imbalanced distribution . While the imbalanced data exist in multiple areas, such as computer vision , bioinformatics, and biomedicine , learning from such data requires special attention upon data sampling , data augmentation , and loss function designs . The imbalanced data is related to diverse tasks, e.g., two-class or multi-class classification . However, an optimized approach can be proposed to address the imbalanced data within FL systems.}\",\n \"id\": \"e54c8539-ee3c-4e42-b5a8-6e639fe2d31f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Research Directions\"\n ],\n [\n \"subsection\",\n \"\\\\liu{Imbalanced Data\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"\\\\liu{Imbalanced Data\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Machine learning algorithms have been widely used to boost the performance of intelligent systems, while FL systems could further enhance intelligent systems~ in distributed computing environments with privacy and security ensured. An intelligent system is a group of machines that has the capacity to gather data, analyze the data, and respond to other systems or the world around. \\nWith FL systems, the distributed data can be exploited to generate models of high performance so as to produce smart responses.\",\n \"id\": \"31802936-db7d-403f-b909-ee8ee4d7b0c8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"226ea081-07c1-4d87-ad04-18733a7f7202\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Research Directions\"\n ],\n [\n \"subsection\",\n \"Applications to Distributed Intelligent Systems\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Applications to Distributed Intelligent Systems\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:con}\\nIn this paper, we discussed the current state of the art of FL systems, including the functional architecture of FL systems, distributed training, and data manipulation. \\nFirst, we presented an overview of FL systems. In particular, we introduced the life cycle of FL models, including four phases. Then, we presented the four-layer functional architecture of FL systems, including presentation, user services, FL training, and infrastructure, and we presented each layer in detail. \\nSecond, we detailed the distributed training with two parts, i.e., parallelism and aggregation algorithms. We presented three types of parallelism, including data parallelism, model parallelism, and pipeline parallelism. We associate each parallelism to a corresponding type of FL. For instance, data parallelism is associated with the horizontal FL, which corresponds to cross-device or cross-silo FL. Model parallelism is related to vertical FL and cross-silo FL. We presented the features of different aggregation algorithms in three types: centralized aggregation, hierarchical aggregation, and decentralized aggregation. \\nThird, we presented the techniques for data manipulation within FL systems. We showed that FL systems prefer RPC for the distributed execution, to handle the fault-tolerance because of moderate network connection. Intermediate data are sketched in order to compress the data, so as to reduce the data communication time. In addition, we presented the data privacy and model security attacks and corresponding defense techniques, e.g., DP, HE, TEE, and the analysis of updated intermediate data for malicious user identification. \\nWe mainly introduced four FL systems: PaddleFL, TensorFlowFederated, FATE, and PySyft. The current solutions primarily focus on the horizontal FL. And we identified five research directions that deserve further investigation: benchmarks, interpretability, decentralized aggregation, FL on graphs, imbalanced data, and the applications of FL systems to distributed intelligent systems.\\n\\\\bibliographystyle{plain}\\n\\\\bibliography{references}\\n\\\\end{document}\",\n \"id\": \"4ff5262c-b6d9-4162-aba6-d016f80c7306\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"42e77c1c-0c8b-4e28-9d53-60920d7558be\",\n \"prefix_titles\": [\n [\n \"title\",\n \"From Distributed Machine Learning to Federated Learning: A Survey\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":81,"string":"81"},"all_cites":{"kind":"list like","value":[3469,1315,5419,3477,591,8917,659,5422,5423,1311,5424,671,627,5420,600,582,5421,5426,5425,166,612,5427,656,5429,5434,8918,5437,5435,5433,5432,7927,5428,5430,7926,8502,5436,5438,5431,590,617,7608,5439,668,636,635,602,7928,5440,5443,5441,5445,8919,5442,5444,5446,632,646,644,5448,5447,5449,5452,8920,5450,7930,5451,7929,5454,623,655,5453,613,7932,622,7931,5456,5457,5455,625,8366,5460,892,5459,5461,5462,5458,3478,7933,614,5463,5464,7727,2673,5468,5470,5465,5471,5467,7038,3415,597,5469,5466,7934,653,7935,180,8921,5472,5473],"string":"[\n 3469,\n 1315,\n 5419,\n 3477,\n 591,\n 8917,\n 659,\n 5422,\n 5423,\n 1311,\n 5424,\n 671,\n 627,\n 5420,\n 600,\n 582,\n 5421,\n 5426,\n 5425,\n 166,\n 612,\n 5427,\n 656,\n 5429,\n 5434,\n 8918,\n 5437,\n 5435,\n 5433,\n 5432,\n 7927,\n 5428,\n 5430,\n 7926,\n 8502,\n 5436,\n 5438,\n 5431,\n 590,\n 617,\n 7608,\n 5439,\n 668,\n 636,\n 635,\n 602,\n 7928,\n 5440,\n 5443,\n 5441,\n 5445,\n 8919,\n 5442,\n 5444,\n 5446,\n 632,\n 646,\n 644,\n 5448,\n 5447,\n 5449,\n 5452,\n 8920,\n 5450,\n 7930,\n 5451,\n 7929,\n 5454,\n 623,\n 655,\n 5453,\n 613,\n 7932,\n 622,\n 7931,\n 5456,\n 5457,\n 5455,\n 625,\n 8366,\n 5460,\n 892,\n 5459,\n 5461,\n 5462,\n 5458,\n 3478,\n 7933,\n 614,\n 5463,\n 5464,\n 7727,\n 2673,\n 5468,\n 5470,\n 5465,\n 5471,\n 5467,\n 7038,\n 3415,\n 597,\n 5469,\n 5466,\n 7934,\n 653,\n 7935,\n 180,\n 8921,\n 5472,\n 5473\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.907732913478076,"string":"0.907733"}}},{"rowIdx":47,"cells":{"authors":{"kind":"list like","value":["Chongyi Li","Chunle Guo","Linghao Han","Jun Jiang","Ming-Ming Cheng","Jinwei Gu","Chen Change Loy"],"string":"[\n \"Chongyi Li\",\n \"Chunle Guo\",\n \"Linghao Han\",\n \"Jun Jiang\",\n \"Ming-Ming Cheng\",\n \"Jinwei Gu\",\n \"Chen Change Loy\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-04-21T19:12:19Z","string":"2021-04-21T19:12:19Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" \t\\label{sec:Abstrat} \tLow-light image enhancement (LLIE) aims at improving the perception or interpretability of an image captured in an environment with poor illumination. \tRecent advances in this area are dominated by deep learning-based solutions, where many learning strategies, network structures, loss functions, training data, etc. have been employed. \tIn this paper, we provide a comprehensive survey to cover various aspects ranging from algorithm taxonomy to unsolved open issues. \tTo examine the generalization of existing methods, we propose a low-light image and video dataset, in which the images and videos are taken by different mobile phones' cameras under diverse illumination conditions. \tBesides, for the first time, we provide a unified online platform that covers many popular LLIE methods, of which the results can be produced through a user-friendly web interface. \tIn addition to qualitative and quantitative evaluation of existing methods on publicly available and our proposed datasets, we also validate their performance in face detection in the dark. \tThis survey together with the proposed dataset and online platform could serve as a reference source for future study and promote the development of this research field. \tThe proposed platform and dataset as well as the collected methods, datasets, and evaluation metrics are publicly available and will be regularly updated. Project page: \\href{https://www.mmlab-ntu.com/project/lliv\\_survey/index.html}{https://www.mmlab-ntu.com/project/lliv\\_survey/index.html}. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"43aba542-7f6b-493f-9beb-9ae4164ca84b","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"]],"subsections":["c6721c4f-f568-49fa-a376-8b26124c3f6f","8eda68ea-596d-47b0-8c59-3c11fbc66273","3ad6854d-28fe-41e8-80bd-198011ce8b19","62b1377f-b6e1-4398-be9f-a1fa2952056b","c43fed76-06de-4d00-89a5-701cdf5d52e3","0f160045-16e9-4579-89b0-811073538246"],"title":"root"},{"cite_extract_rate":1,"cites":[1875],"content":"\\label{sec:Introduction}}\n\\IEEEPARstart{I}{mages} are often taken under sub-optimal lighting conditions, under the influence of backlit, uneven light, and dim light, due to inevitable environmental and/or technical constraints such as insufficient illumination and limited exposure time. Such images suffer from the compromised aesthetic quality and unsatisfactory transmission of information for high-level tasks such as object tracking, recognition, and detection. Figure \\ref{fig:example} shows some examples of the degradations induced by sub-optimal lighting conditions. \nLow-light enhancement enjoys a wide range of applications in different areas, including visual surveillance, autonomous driving, and computational photography. In particular, smartphone photography has become ubiquitous and prominent. Limited by the size of the camera aperture, the requirement of real-time processing, and the constraint of memory, taking photographs with a smartphone's camera in a dim environment is especially challenging. There is an exciting research arena of enhancing low-light images and videos in such applications.\n\\begin{figure}[!t]\n\t\\begin{center}\n\t\t\\begin{tabular}{c@{ }c@{ }c@{ }}\n\t\t\t\\includegraphics[width=0.3\\linewidth,height=0.2\\linewidth]{figures/Fig1_backlit.png}&\n\t\t\t\\includegraphics[width=0.3\\linewidth,height=0.2\\linewidth]{figures/Fig1_nonuniform.png}&\n\t\t\t\\includegraphics[width=0.3\\linewidth,height=0.2\\linewidth]{figures/Fig1_weak.png}\\\\\n\t\t\t(a) back lit & (b) uneven light &(c) dim light\\\\\n\t\t\t\\includegraphics[width=0.3\\linewidth,height=0.2\\linewidth]{figures/Fig1_dark.png}&\n\t\t\t\\includegraphics[width=0.3\\linewidth,height=0.2\\linewidth]{figures/0027.png}&\n\t\t\t\\includegraphics[width=0.3\\linewidth,height=0.2\\linewidth]{figures/Fig1_noise.png}\\\\\n\t\t\t(d) extremely low &\t(e) colored light & (f) boosted noise\\\\\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\vspace{-3pt}\n\t\\caption{Examples of images taken under sub-optimal lighting conditions. These images suffer from the buried scene content, reduced contrast, boosted noise, and inaccurate color. }\n\t\\label{fig:example}\n\t\\vspace{-5pt}\n\\end{figure}\n\\begin{figure*}[thb]\n\t\\centering \\centerline{\\includegraphics[width=1\\linewidth]{figures/chronology.png}}\n\t\\vspace{-2pt}\n\t\\caption{A concise milestone of deep learning-based low-light image and video enhancement methods. \\textbf{Supervised learning-based methods}: LLNet , Chen et al. , MBLLEN , Retinex-Net , LightenNet , SCIE , DeepUPE , Chen et al. , Jiang and Zheng , Wang et al. , KinD , Ren et al. , Xu et al. , Fan et al. , Lv et al. , EEMEFN~, SIDGAN. , LPNet , DLN , TBEFN , DSLR , Zhang et al. , PRIEN , and Retinex-Net . \\textbf{Reinforcement learning-based method}: DeepExposure . \\textbf{Unsupervised learning-based method}: EnlightenGAN . \\textbf{Zero-shot learning-based methods}: ExCNet , Zero-DCE , RRDNet , Zero-DCE++ , RetinexDIP , and RUAS . \\textbf{Semi-supervised learning-based method}: DRBN and DRBN .}\n\t\\label{fig:milestones}\n\t\\vspace{-4pt}\n\\end{figure*}\nTraditional methods for low-light enhancement include Histogram Equalization-based methods and Retinex model-based methods . The latter received relatively more attention. A typical Retinex model-based approach decomposes a low-light image into a reflection component and an illumination component by priors or regularizations. The estimated reflection component is treated as the enhanced result.\nSuch methods have some limitations:\n\\textbf{1)} the ideal assumption that treats the reflection component as the enhanced result does not always hold, especially given various illumination properties, which could lead to unrealistic enhancement such as loss of details and distorted colors, \\textbf{2)} the noise is usually ignored in the Retinex model, thus it is remained or amplified in the enhanced results, \\textbf{ 3)} finding an effective prior or regularization is challenging. Inaccurate prior or regularization may result in artifacts and color deviations in the enhanced results, and \\textbf{4)} the runtime is relatively long because of their complicated optimization process.\nRecent years have witnessed the compelling success of deep learning-based LLIE since the first seminal work . Deep learning-based solutions enjoy better accuracy, robustness, and speed over conventional methods, thus attracting increasing attention. A concise milestone of deep learning-based LLIE methods is shown in Figure \\ref{fig:milestones}. As shown, \nsince 2017, the number of deep learning-based solutions has grown year by year. Learning strategies used in these solutions cover Supervised Learning (SL), Reinforcement Learning (RL), Unsupervised Learning (UL), Zero-Shot Learning (ZSL), and Semi-Supervised Learning (SSL). Note that we only report some representative methods in Figure \\ref{fig:milestones}. In fact, there are more than 100 papers on deep learning-based methods from 2017 to 2021. Moreover, although some general photo enhancement methods can improve the brightness of images to some extent, we omit them in this survey as they are not designed to handle diverse low-light conditions. We concentrate on deep learning-based solutions that are specially developed for low-light image and video enhancement.\n\tDespite deep learning has dominated the research of LLIE, an in-depth and comprehensive survey on deep learning-based solutions is lacking. There are two reviews of LLIE . Wang et al. mainly reviews conventional LLIE methods while our work systematically and comprehensively reviews recent advances of deep learning-based LLIE. In comparison to Liu et al. that reviews existing LLIE algorithms, measures the machine vision performance of different methods, provides a low-light image dataset serving both low-level and high-level vision enhancement, and develops an enhanced face detector, our survey reviews the low-light image and video enhancement from different aspects and has the following unique characteristics.\n\t\\textbf{1)} Our work mainly focuses on recent advances of deep learning-based low-light image and video enhancement, where we provide in-depth analysis and discussion in various aspects, covering learning strategies, network structures, loss functions, training datasets, test datasets, evaluation metrics, model sizes, inference speed, enhancement performance, etc. Thus, this survey centers on deep learning and its applications in low-light image and video enhancement.\n\t\\textbf{2)} We propose a dataset that contains images and videos captured by different mobile phones' cameras under diverse illumination conditions to evaluate the generalization of existing methods. This new and challenging dataset is a supplement of existing low-light image and video enhancement datasets as such a dataset is lacking in this research area. Besides, we are the first, to the best of our knowledge, to compare the performance of deep learning-based low-light image enhancement methods on this kind of data. \n\t\\textbf{3)} We provide an online platform that covers many popular deep learning-based low-light image enhancement methods, where the results can be produced by a user-friendly web interface. With our platform, one without any GPUs can assess the results of different methods for any input images online, which speeds up the development of this research field and helps to create new research.\n\tWe hope that our survey could provide novel insights and inspiration to facilitate the understanding of deep learning-based LLIE, foster research on the raised open issues, and speed up the development of this research field.","id":"c6721c4f-f568-49fa-a376-8b26124c3f6f","level":"section","origin_cites_number":8,"parent_id":"43aba542-7f6b-493f-9beb-9ae4164ca84b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Solution}","id":"8eda68ea-596d-47b0-8c59-3c11fbc66273","level":"section","origin_cites_number":0,"parent_id":"43aba542-7f6b-493f-9beb-9ae4164ca84b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Deep Learning-Based LLIE"]],"subsections":["b406f65d-5d41-4498-9ddb-95caec62c1be","44126fdb-481e-4904-b2e2-6e15d767dcdd"],"title":"Deep Learning-Based LLIE"},{"cite_extract_rate":0,"cites":[],"content":"We first give a common formulation of the deep learning-based LLIE problem. For a low-light image $I\\in \\mathbb{R}^{W\\times H\\times3}$ of width $W$ and height $H$, the process can be modeled as:\n\\begin{equation}\n\t\\label{equ_1}\n\t\\widehat{R}=\\mathcal{F}(I;\\theta),\n\\end{equation}\nwhere $\\widehat{R}\\in \\mathbb{R}^{W\\times H\\times3}$ is the enhanced result and $\\mathcal{F}$ represents the network with trainable parameters $\\theta$. The purpose of deep learning is to find optimal network parameters $\\widehat{\\theta}$ that minimizes the error:\n\\begin{equation}\n\t\\label{equ_2}\n\t\\widehat{\\theta}=\\argmin_{\\theta}\\mathcal{L}(\\widehat{R},R),\n\\end{equation}\nwhere $R\\in \\mathbb{R}^{W\\times H\\times3}$ is the ground truth, and the loss function $\\mathcal{L}(\\widehat{R},R)$ drives the optimization of network.\nVarious loss functions such as supervised loss and unsupervised loss can be used. More details will be presented in Section \\ref{sec:Technical}.","id":"b406f65d-5d41-4498-9ddb-95caec62c1be","level":"subsection","origin_cites_number":0,"parent_id":"8eda68ea-596d-47b0-8c59-3c11fbc66273","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Deep Learning-Based LLIE"],["subsection","Problem Definition"]],"subsections":[],"title":"Problem Definition"},{"cite_extract_rate":1,"cites":[1875],"content":"According to different learning strategies, we categorize existing LLIE methods into supervised learning, reinforcement learning, unsupervised learning, zero-shot learning, and semi-supervised learning. \nA statistic analysis from different perspectives is presented in Figure \\ref{fig:statistic}. In what follows, we review some representative methods of each strategy. \n\\noindent\n\\textbf{Supervised Learning.} \nFor supervised learning-based LLIE methods, they are further divided into end-to-end, deep Retinex-based, and realistic data-driven methods.\nThe first deep learning-based LLIE method LLNet~ employs a variant of stacked-sparse\ndenoising autoencoder to brighten and denoise low-light images simultaneously. \nThis pioneering work inspires the usage of end-to-end networks in LLIE. \nLv et al. propose an end-to-end multi-branch enhancement network (MBLLEN). The MBLLEN improves the performance of LLIE via extracting effective feature representations by a feature extraction module, an enhancement module, and a fusion module. \nThe same authors propose other three subnetworks including an Illumination-Net, a Fusion-Net, and a Restoration-Net to further improve the performance. \nRen et al. design a more complex end-to-end network that comprises an encoder-decoder network for image content enhancement and a recurrent neural network for image edge enhancement. \nSimilar to Ren et al. , Zhu et al. propose a method called EEMEFN. The EEMEFN consists of two stages: multi-exposure fusion and edge enhancement. \nA multi-exposure fusion network, TBEFN~, is proposed for LLIE. The TBEFN estimates a transfer function in two branches, of which two enhancement results can be obtained. At last, a simple average scheme is employed to fuse these two images and further refine the result via a refinement unit. \nIn addition, pyramid network (LPNet) , residual network , and Laplacian pyramid (DSLR)\nare introduced into LLIE. These methods learn to effectively and efficiently integrate feature representations via commonly used end-to-end network structures for LLIE.\nBased on the observation that noise exhibits different levels of contrast in different frequency layers, Xu et al. proposed a frequency-based decomposition-and-enhancement network. This network recovers image contents with noise suppression in the low-frequency layer while inferring the details in the high-frequency layer. \nRecently, a progressive-recursive low-light image enhancement network is proposed, which uses a recursive unit to gradually enhance the input image.\nTo solve temporal instability when handling low-light videos, Zhang et al. propose to learn and infer motion field from a single image then enforce temporal consistency.\nIn comparison to directly learning an enhanced result in an end-to-end network, deep Retinex-based methods enjoy better enhancement performance in most cases owing to the physically explicable Retinex theory . Deep Retinex-based methods usually separately enhance the illuminance component and the reflectance components via specialized subnetworks.\nA Retinex-Net is proposed, which includes a Decom-Net that splits the input image into light-independent reflectance and structure-aware smooth illumination and an Enhance-Net that adjusts the illumination map for low-light enhancement. Recently, the Retinex-Net is extended by adding new constraints and advanced network designs for better enhancement performance .\nTo reduce the computational burden, Li et al. propose a lightweight LightenNet for weakly illuminated image enhancement, which only consists of four layers. \nThe LightenNet takes a weakly illuminated image as the input and then estimates its illumination map. Based on the Retinex theory , the enhanced image is obtained by dividing the input image by the illumination map.\nTo accurately estimate the illumination map, Wang et al. extract the global and local features to learn an image-to-illumination mapping by their proposed DeepUPE network.\nZhang et al. separately develop three subnetworks for layer decomposition, reflectance restoration, and illumination adjustment, called KinD. \nFurthermore, the authors alleviate the visual defects left in the results of KinD by a multi-scale illumination attention module. The improved KinD is called KinD++ .\nTo solve the issue that the noise is omitted in the deep Retinex-based methods,\nWang et al. propose a progressive Retinex network, where an IM-Net estimates the illumination and a NM-Net estimates the noise level. These two subnetworks work in a progressive mechanism until obtaining stable results. \nFan et al. integrate semantic segmentation and Retinex model for further improving the enhancement performance in real cases. The core idea is to use semantic prior to guide the enhancement of both the illumination component and the reflectance component. \n\tAlthough some methods can achieve decent performance, they show poor generalization capability in real low-light cases due to the usage of synthetic training data. To solve this issue, some works attempt to generate more realistic training data or capture real data.\n\tCai et al. build a multi-exposure image dataset, where the low-contrast images of different exposure levels have their corresponding high-quality reference images. \n\tEach high-quality reference image is obtained by subjectively selecting the best output from 13 results enhanced by different methods. Moreover, a frequency decomposition network is trained on the built dataset and separately enhances the high-frequency layer and the low-frequency layer via a two-stage structure. \n\tChen et al. collect a real low-light image dataset (SID) and \n\ttrain the U-Net to learn a mapping from low-light raw data to the corresponding long-exposure high-quality reference image.\n\tFurther, Chen et al. extend the SID dataset to low-light videos (DRV). The DRV contains static videos with the corresponding long-exposure ground truths. To ensure the generalization capability of processing the videos of dynamic scenes, a siamese network is proposed. \n\tTo enhance the moving objects in the dark, Jiang and Zheng~ design a co-axis optical system to capture temporally synchronized and spatially aligned low-light and well-lighted video pairs (SMOID). \n\tUnlike the DRV video dataset , the SMOID video dataset contains dynamic scenes. To learn the mapping from raw low-light video to well-lighted video, a 3D U-Net-based network is proposed.\n\tConsidering the limitations of previous low-light video datasets such as DRV dataset only containing statistic videos and SMOID dataset only having 179 video pairs, \n\tTriantafyllidou et al.~ propose a low-light video synthesis pipeline, dubbed SIDGAN. The SIDGAN can produce dynamic video data (raw-to-RGB) by a semi-supervised dual CycleGAN with intermediate domain mapping. \n\tTo train this pipeline, the real-world videos are collected from Vimeo-90K dataset . The low-light raw video data and the corresponding long-exposure images are sampled from DRV dataset . \n\tWith the synthesized training data, this work adopts the same U-Net network as Chen et al. for low-light video enhancement. \n\t\\noindent\n\t\\textbf{Reinforcement Learning.}\n\tWithout paired training data, Yu et al. learn to expose photos with reinforcement adversarial learning, named DeepExposure. Specifically, an input image is first segmented into sub-images according to exposures. For each sub-image, local exposure is learned by the policy network sequentially based on reinforcement learning. The reward evaluation function is approximated by adversarial learning. At last, each local exposure is employed to retouch the input, thus obtaining multiple retouched images under different exposures. The final result is achieved by fusing these images. \n\t\\noindent\n\t\\textbf{Unsupervised Learning.}\n\tTraining a deep model on paired data may result in overfitting and limited generalization capability. To solve this issue, an unsupervised learning method named EnligthenGAN is proposed.\n\tThe EnlightenGAN adopts an attention-guided U-Net as the generator and uses the global-local discriminators to ensure the enhanced results look like realistic normal-light images. \n\tIn addition to global and local adversarial losses, the global and local self feature preserving losses are proposed to preserve the image content before and after the enhancement. This is a key point for the stable training of such a one-path Generative Adversarial Network (GAN) structure. \n\t\\noindent\n\t\\textbf{Zero-Shot Learning.} \n\tThe supervised learning, reinforcement learning, and unsupervised learning methods either have limited generalization capability or suffer from unstable training. To remedy these issues, zero-shot learning is proposed to learn the enhancement solely from the testing images. \n\tNote that the concept of zero-shot learning in the low-level vision tasks is used to emphasize that the method does not require paired or unpaired training data, which is different from its definition in high-level visual tasks. \n\tZhang et al. propose a zero-shot learning method, called ExCNet, for back-lit image restoration. \n\tA network is first used to estimate the S-curve that best fits the input image. Once the S-curve is estimated, the input image is separated into a base layer and a detail layer using the guided filter . Then the base layer is adjusted by the estimated S-curve. Finally, the Weber contrast is used to fuse the detailed layer and the adjusted base layer. To train the ExCNet, the authors formulate the loss function as a block-based energy minimization problem. \n\tZhu et al. propose a three-branch CNN, called RRDNet, for underexposed images restoration. The RRDNet decomposes an input image into illumination, reflectance, and noise via iteratively minimizing specially designed loss functions. \n\tTo drive the zero-shot learning, a combination of Retinex reconstruction loss, texture enhancement loss, and illumination-guided noise estimation loss is proposed. \n\tZhao et al. perform Retinex decomposition via neural networks and then enhance the low-light image based on the Retinex model, called RetinexDIP. Inspired by Deep Image Prior (DIP) , RetinexDIP generates the reflectance component and illumination component of an input image by randomly sampled white noise, in which the component characteristics-related losses such as illumination smoothness are used for training.\n\tLiu et al. propose a Retinex-inspired unrolling method for LLIE, in which the cooperative architecture search is used to discover lightweight prior architectures of basic blocks and non-reference losses are used to train the network.\n\tDifferent from the image reconstruction-based methods , \n\ta deep curve estimation network, Zero-DCE , is proposed. Zero-DCE formulates the light enhancement as a task of image-specific curve estimation, which takes a low-light image as input and produces high-order curves as its output. These curves are used for pixel-wise adjustment on the dynamic range of the input to obtain an enhanced image. Further, an accelerated and lightweight version is proposed, called Zero-DCE++ . \n\tSuch curve-based methods do not require any paired or unpaired data during training. They achieve zero-reference learning via a set of non-reference loss functions. Besides, unlike the image reconstruction-based methods that need high computational resources, the image-to-curve mapping only requires lightweight networks, thus achieving a fast inference speed. \n\t\\begin{figure*}[!t]\n\t\t\\begin{center}\n\t\t\t\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }}\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture1.png}&\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture2.png}&\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture3.png}&\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture4.png}\\\\\n\t\t\t\t(a) learning strategy & (b) network structure& (c) Retinex model & (d) data format \\\\\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture5.png}&\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture6.png}&\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture7.png}&\n\t\t\t\t\\includegraphics[width=.23\\textwidth]{figures/Picture8.png}\\\\\n\t\t\t\t(e) loss function & (f) training dataset & (g) testing dataset & (h) evaluation metric \\\\\n\t\t\t\\end{tabular}\n\t\t\\end{center}\n\t\t\\vspace{-2pt}\n\t\t\\caption{A statictic analysis of deep learning-based LLIE methods, including learning strategy, network characteristic, Retinex model, data format, loss function, training dataset, testing dataset, and evaluation metric. Best viewed by zooming in.}\n\t\t\\label{fig:statistic}\n\t\t\\vspace{-4pt}\n\t\\end{figure*}\n\t\\noindent\n\t\\textbf{Semi-Supervised Learning.}\n\tTo combine the strengths of supervised learning and unsupervised learning, semi-supervised learning has been proposed in recent years.\n\tYang et al. propose a semi-supervised deep recursive band network (DRBN). The DRBN first recovers a linear band representation of an enhanced image under supervised learning, and then obtains an improved one by recomposing the given bands via a learnable linear transformation based on unsupervised adversarial learning. \n\tThe DRBN is extended by introducing Long Short Term Memory (LSTM) networks and an image quality assessment network pre-trained on an aesthetic visual analysis dataset, which achieves better enhancement performance .\n\tObserving Figure \\ref{fig:statistic}(a), we can find that supervised learning is the mainstream among deep learning-based LLIE methods, of which the percentage reaches 73\\%. This is because supervised learning is relatively easy when paired training data such as LOL , SID and diverse low-/normal-light image synthesis approaches are used. \n\tHowever, supervised learning-based methods suffer from some challenges: \n\t\\textbf{1)} collecting a large-scale paired dataset that covers diverse real-world low-light conditions is difficult, \n\t\\textbf{2)} synthetic low-light images do not accurately represent real-world illuminance conditions such as spatially varying lighting and different levels of noise, and \n\t\\textbf{3)} training a deep model on paired data may result in limited generalization to real-world images of diverse illumination properties. \n\tTherefore, some methods adopt unsupervised learning, reinforcement learning, semi-supervised learning, and zero-shot learning to bypass the challenges in supervised learning. Although these methods achieve competing performance, they still suffer from some limitations: \n\t\\textbf{1)} for unsupervised learning/semi-supervised learning methods, how to implement stable training, avoid color deviations, and build the relations of cross-domain information challenges current methods,\n\t\\textbf{2)} for reinforcement learning methods, designing an effective reward mechanism and implementing efficient and stable training are intricate, and \n\t\\textbf{3)} for zero-shot learning methods, the design of non-reference losses is non-trivial when the color preservation, artifact removal, and gradient back-propagation should be taken into account. \n\t\\begin{table*}\n\t\t\\rowcolors{1}{gray!20}{white}\n\t\t\\centering\n\t\t\\caption{\n\t\t\t{Summary of essential characteristics of representative deep learning-based methods. ``Retinex'' indicates whether the models are Retinex-based or not. ``simulated'' means the testing data are simulated by the same approach as the synthetic training data. ``self-selected'' stands for the real-world images selected by the authors. ``\\#P''represents the number of trainable parameters. ``-'' means this item is not available or not indicated in the paper.}\n\t\t}\n\t\t\\vspace{-6pt}\n\t\t\\label{table:methods}\n\t\t\\begin{threeparttable}\n\t\t\t\\resizebox{1\\textwidth}{!}{\n\t\t\t\t\\setlength\\tabcolsep{2pt}\n\t\t\t\t\\renewcommand\\arraystretch{0.98}\n\t\t\t\t\\begin{tabular}{c|r||c|c|c|c|c|c|c|c|c}\n\t\t\t\t\t\\hline\n\t\t\t\t\t&\\textbf{Method}~~~~~~~~~&\\textbf{Learning} &\\textbf{Network Structure} &\\textbf{Loss Function}\n\t\t\t\t\t&\\textbf{Training Data} & \\textbf{Testing Data} &\\textbf{Evaluation Metric} & \\textbf{Format} &\\textbf{Platform} &\\textbf{Retinex}\\\\\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\multirow{1}{*}{\\rotatebox{90}{\\textbf{2017}}}\n\t\t\t\t\t&LLNet~ &SL &SSDA &SRR loss\n\t\t\t\t\t&\\tabincell{c}{simulated by \\\\Gamma Correction \\& \\\\Gaussian Noise} &\\tabincell{c}{simulated \\\\ self-selected} &\\tabincell{c}{PSNR SSIM} &RGB &Theano &\\\\\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\multirow{1}{*}{\\rotatebox{90}{\\textbf{2018}}}\n\t\t\t\t\t&LightenNet~ &SL &four layers &$L_{2}$ loss\n\t\t\t\t\t&\\tabincell{c}{simulated by \\\\random illumination values} &\\tabincell{c}{simulated\\\\self-selected} &\\tabincell{c}{PSNR MAE\\\\SSIM \\\\User Study} &RGB &\\tabincell{c}{Caffe\\\\MATLAB}&\\checkmark\\\\\n\t\t\t\t\t&Retinex-Net~ &SL &\n\t\t\t\t\tmulti-scale network &\\tabincell{c}{$L_{1}$ loss smoothness loss \\\\invariable reflectance loss}\n\t\t\t\t\t&\\tabincell{c}{LOL\\\\simulated by \\\\adjusting histogram} &self-selected &- &RGB &TensorFlow&\\checkmark\\\\\n\t\t\t\t\t&MBLLEN~ &SL &multi-branch fusion &\\tabincell{c}{SSIM loss region loss\\\\perceptual loss}\n\t\t\t\t\t&\\tabincell{c}{simulated by \\\\Gamma Correction \\& \\\\Poisson Noise} &\\tabincell{c}{simulated\\\\ self-selected} &\\tabincell{c}{PSNR SSIM\\\\AB VIF\\\\LOE TOMI} &RGB &\\tabincell{c}{TensorFlow} & \\\\\n\t\t\t\t\t&SCIE~ &SL &frequency decomposition &\\tabincell{c}{$L_{2}$ loss $L_{1}$ loss SSIM loss}\n\t\t\t\t\t&SCIE &SCIE &\\tabincell{c}{PSNR FSIM\\\\Runtime FLOPs} &RGB&\\tabincell{c}{Caffe\\\\MATLAB}&\\\\\n\t\t\t\t\t&Chen et al.~ &SL &U-Net &$L_{1}$ loss\n\t\t\t\t\t&SID &SID &\\tabincell{c}{PSNR SSIM} &raw&TensorFlow&\\\\\n\t\t\t\t\t&Deepexposure~ &RL &\\tabincell{c}{policy network\\\\GAN} &\\tabincell{c}{deterministic policy gradient\\\\adversarial loss}\n\t\t\t\t\t&MIT-Adobe FiveK&MIT-Adobe FiveK &\\tabincell{c}{PSNR SSIM} &raw&TensorFlow &\\\\\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\multirow{1}{*}{\\rotatebox{90}{\\textbf{2019}}}\n\t\t\t\t\t&Chen et al.~ &SL &siamese network &\\tabincell{c}{$L_{1}$ loss\\\\ self-consistency loss}\n\t\t\t\t\t&DRV &DRV &\\tabincell{c}{PSNR\\\\ SSIM\\\\ MAE} &raw &TensorFlow&\\\\\n\t\t\t\t\t&Jiang and Zheng~ &SL &3D U-Net &$L_{1}$ loss\n\t\t\t\t\t&SMOID &SMOID &\\tabincell{c}{PSNR SSIM MSE} &raw &TensorFlow &\\\\\n\t\t\t\t\t&DeepUPE~ &SL &illumination map &\\tabincell{c}{$L_{1}$ loss color loss\\\\smoothness loss}\n\t\t\t\t\t&retouched image pairs&MIT-Adobe FiveK&\\tabincell{c}{PSNR SSIM\\\\User Study} &RGB &TensorFlow &\\checkmark\\\\\n\t\t\t\t\t&KinD~ &SL &\\tabincell{c}{three subnetworks\\\\U-Net} &\\tabincell{c}{reflectance similarity loss\\\\illumination smoothness loss\\\\mutual consistency loss\\\\$L_{1}$ loss $L_{2}$ loss SSIM loss\\\\texture similarity loss\\\\illumination adjustment loss}\n\t\t\t\t\t&LOL &\\tabincell{c}{LOL LIME\\\\NPE MEF} &\\tabincell{c}{PSNR SSIM\\\\LOE NIQE} &RGB &TensorFlow&\\checkmark\\\\\n\t\t\t\t\t&Wang et al.~ &SL &\\tabincell{c}{two subnetworks\\\\pointwise Conv} &$L_{1}$ loss \n\t\t\t\t\t&\\tabincell{c}{simulated by \\\\camera imaging model} &\\tabincell{c}{IP100 FNF38\\\\MPI LOL NPE} &\\tabincell{c}{PSNR SSIM\\\\NIQE} &RGB &Caffe &\\checkmark \\\\\n\t\t\t\t\t&Ren et al.~ &SL &\\tabincell{c}{U-Net like network\\\\RNN dilated Conv} &\\tabincell{c}{$L_{2}$ loss perceptual loss\\\\adversarial loss} \n\t\t\t\t\t&\\tabincell{c}{MIT-Adobe FiveK\\\\with Gamma correction \\\\ \\& Gaussion noise} &\\tabincell{c}{simulated \\\\self-selected DPED} &\\tabincell{c}{PSNR SSIM\\\\Runtime} &RGB &Caffe&\\\\\n\t\t\t\t\t&EnlightenGAN~ &UL &U-Net like network &\\tabincell{c}{adversarial loss\\\\self feature preserving loss}\n\t\t\t\t\t&unpaired real images &\\tabincell{c}{NPE LIME\\\\MEF DICM\\\\VV BBD-100K\\\\ExDARK} &\\tabincell{c}{User Study NIQE\\\\Classification} &RGB &PyTorch&\\\\\n\t\t\t\t\t&ExCNet.~ &ZSL &\n\t\t\t\t\t\\tabincell{c}{fully connected layers} &energy minimization loss &real images\n\t\t\t\t\t&$IE_{ps}D$ &\\tabincell{c}{User Study\\\\CDIQA LOD} &RGB &PyTorch&\\\\\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\multirow{1}{*}{\\rotatebox{90}{\\textbf{2020}}}\n\t\t\t\t\t&Zero-DCE~ &ZSL &U-Net like network &\\tabincell{c}{spatial consistency loss\\\\exposure control loss\\\\color constancy loss\\\\illumination smoothness loss}\n\t\t\t\t\t&SICE &\\tabincell{c}{SICE NPE\\\\LIME MEF\\\\DICM VV\\\\ DARK FACE} &\\tabincell{c}{User Study PI\\\\PNSR SSIM\\\\MAE Runtime\\\\Face detection} &RGB &PyTorch &\\\\\n\t\t\t\t\t&DRBN~ &SSL &recursive network &\\tabincell{c}{SSIM loss perceptual loss\\\\adversarial loss}\n\t\t\t\t\t&\\tabincell{c}{LOL\\\\images selected by MOS} &LOL &\\tabincell{c}{PSNR SSIM\\\\SSIM-GC} &RGB &PyTorch &\\\\\n\t\t\t\t\t&Lv et al.~&SL &U-Net like network &\\tabincell{c}{Huber loss\\\\SSIM loss\\\\perceptual loss\\\\illumination smoothness loss}\n\t\t\t\t\t&\\tabincell{c}{simulated by a\\\\ retouching module} &\\tabincell{c}{LOL SICE \\\\DeepUPE} &\\tabincell{c}{User Study PSNR\\\\SSIM VIF\\\\LOE NIQE\\\\ \\#P Runtime\\\\Face detection} &RGB&\\tabincell{c}{TensorFlow}&\\checkmark\\\\\n\t\t\t\t\t&Fan et al.~ &SL &\\tabincell{c}{four subnetworks\\\\U-Net like network\\\\feature modulation} &\\tabincell{c}{mutual smoothness loss\\\\reconstruction loss\\\\illumination smoothness loss\\\\cross entropy loss\\\\ consistency loss\\\\SSIM loss\\\\gradient loss\\\\ratio learning loss}\n\t\t\t\t\t&\\tabincell{c}{simulated by \\\\illumination adjustment, \\\\slight color distortion, \\\\and noise simulation} &\\tabincell{c}{simulated\\\\self-selected} &\\tabincell{c}{PSNR SSIM\\\\NIQE} &RGB & - &\\checkmark\\\\\n\t\t\t\t\t&Xu et al.~&SL &\\tabincell{c}{frequency decomposition\\\\U-Net like network} &\\tabincell{c}{$L_{2}$ loss\\\\perceptual loss}\n\t\t\t\t\t&SID in RGB&\\tabincell{c}{SID in RGB \\\\self-selected} &\\tabincell{c}{PSNR SSIM} &RGB &PyTorch&\\\\\n\t\t\t\t\t&EEMEFN~&SL &\\tabincell{c}{U-Net like network\\\\edge detection network} & \\tabincell{c}{$L_{1}$ loss\\\\weighted cross-entropy loss}\n\t\t\t\t\t&SID &SID &\\tabincell{c}{PSNR SSIM} &raw& \\tabincell{c}{TensorFlow\\\\PaddlePaddle}&\\\\\n\t\t\t\t\t&DLN~&SL &\\tabincell{c}{residual learning\\\\interactive factor\\\\back projection network} &\\tabincell{c}{SSIM loss\\\\total variation loss}\n\t\t\t\t\t&\\tabincell{c}{simulated by \\\\illumination adjustment, \\\\slight color distortion,\\\\and noise simulation} &\\tabincell{c}{simulated \\\\LOL} &\\tabincell{c}{User Study PSNR\\\\SSIM NIQE} &RGB &PyTorch&\\\\\n\t\t\t\t\t&LPNet~&SL &\\tabincell{c}{pyramid network} &\\tabincell{c}{$L_{1}$ loss\\\\perceptual loss\\\\luminance loss}\n\t\t\t\t\t&\\tabincell{c}{LOL SID in RGB\\\\MIT-Adobe FiveK} &\\tabincell{c}{LOL SID in RGB\\\\MIT-Adobe FiveK\\\\ MEF NPE DICM VV} &\\tabincell{c}{PSNR SSIM\\\\NIQE \\#P\\\\FLOPs Runtime} &RGB &PyTorch&\\\\\n\t\t\t\t\t&SIDGAN~ &SL &U-Net &CycleGAN loss\n\t\t\t\t\t&SIDGAN &SIDGAN &\\tabincell{c}{PSNR SSIM\\\\TPSNR TSSIM ATWE} &raw &\\tabincell{c}{TensorFlow}&\\\\\n\t\t\t\t\t&RRDNet &ZSL&three subnetworks &\\tabincell{c}{retinex reconstruction loss\\\\texture enhancement loss\\\\noise estimation loss}\n\t\t\t\t\t&-&\\tabincell{c}{NPE LIME\\\\MEF DICM} &\\tabincell{c}{NIQE CPCQI} &RGB&PyTorch &\\checkmark\\\\\n\t\t\t\t\t&TBEFN~&SL&\\tabincell{c}{three stages\\\\U-Net like network} &\\tabincell{c}{SSIM loss\\\\perceptual loss\\\\smoothness loss}\n\t\t\t\t\t&\\tabincell{c}{SCIE \\\\LOL} &\\tabincell{c}{SCIE LOL\\\\DICM MEF\\\\NPE VV} &\\tabincell{c}{PSNR SSIM\\\\NIQE Runtime\\\\ \\#P FLOPs} &RGB &TensorFlow&\\checkmark\\\\\n\t\t\t\t\t&DSLR~&SL&\\tabincell{c}{Laplacian pyramid\\\\U-Net like network} &\\tabincell{c}{$L_{2}$ loss\\\\Laplacian loss\\\\color loss}\n\t\t\t\t\t&MIT-Adobe FiveK &\\tabincell{c}{MIT-Adobe FiveK\\\\self-selected} &\\tabincell{c}{PSNR SSIM\\\\NIQMC NIQE\\\\BTMQI CaHDC} &RGB& PyTorch&\\\\\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\hline\n\t\t\t\t\t\\multirow{1}{*}{\\rotatebox{90}{\\textbf{2021}}}\n\t\t\t\t\t&RUAS~ &ZSL &neural architecture search &\\tabincell{c}{cooperative loss\\\\similar loss\\\\total variation loss}\n\t\t\t\t\t&\\tabincell{c}{LOL\\\\MIT-Adobe FiveK} &\\tabincell{c}{LOL\\\\MIT-Adobe FiveK} &\\tabincell{c}{PSNR SSIM\\\\ Runtime \\#P FLOPs} &RGB &PyTorch &\\checkmark\\\\\n\t\t\t\t\t&Zhang et al.~&SL &U-Net &\\tabincell{c}{$L_{1}$ loss\\\\ consistency loss}\n\t\t\t\t\t&\\tabincell{c}{simulated by illumination \\\\adjustmentand noise simulation} &\\tabincell{c}{simulated \\\\self-selected} &\\tabincell{c}{User Study PSNR\\\\SSIM AB\\\\MABD WE} &RGB&\\tabincell{c}{PyTorch}&\\\\\n\t\t\t\t\t&Zero-DCE++~ &ZSL &U-Net like network &\\tabincell{c}{spatial consistency loss\\\\exposure control loss\\\\color constancy loss\\\\illumination smoothness loss}\n\t\t\t\t\t&SICE &\\tabincell{c}{SICE NPE\\\\LIME MEF\\\\DICM VV\\\\ DARK FACE} &\\tabincell{c}{User Study PI\\\\PNSR SSIM \\#P\\\\MAE Runtime\\\\Face detection FLOPs} &RGB &PyTorch &\\\\\n\t\t\t\t\t&DRBN~ &SSL &recursive network &\\tabincell{c}{perceptual loss\\\\detail loss quality loss}\n\t\t\t\t\t&LOL &LOL &\\tabincell{c}{PSNR SSIM\\\\SSIM-GC} &RGB &PyTorch &\\\\\n\t\t\t\t\t&Retinex-Net~ &SL &\n\t\t\t\t\tthree subnetworks &\\tabincell{c}{$L_{1}$ loss $L_{2}$ loss\\\\SSIM loss\\\\total variation loss}\n\t\t\t\t\t&\\tabincell{c}{LOL\\\\simulated by adjusting histogram} &\\tabincell{c}{LOL simulated\\\\NPE DICM VV} &\\tabincell{c}{PNSR SSIM\\\\UQI OSS User Study} &RGB &PyTorch&\\checkmark\\\\\n\t\t\t\t\t&RetinexDIP &ZSL & encoder-decoder networks &\\tabincell{c}{reconstruction loss\\\\illumination-consistency loss\\\\reflectnce loss \\\\illumination smoothness loss}&- &\\tabincell{c}{DICM, ExDark\\\\Fusion LIME\\\\ NASA NPE VV} &\\tabincell{c}{NIQE \\\\NIQMC CPCQI} &RGB &PyTorch&\\checkmark\\\\\n\t\t\t\t\t&PRIEN~ &SL &recursive network &SSIM loss&\\tabincell{c}{MEF LOL\\\\simulated by adjusting histogram} &\\tabincell{c}{LOL LIME\\\\NPE MEF VV} &\\tabincell{c}{PNSR SSIM\\\\LOE TMQI} &RGB &PyTorch &\\\\\n\t\t\t\t\t\\hline\n\t\t\t\t\\end{tabular}\n\t\t\t}\n\t\t\\end{threeparttable}\n\t\\end{table*}","id":"44126fdb-481e-4904-b2e2-6e15d767dcdd","level":"subsection","origin_cites_number":6,"parent_id":"8eda68ea-596d-47b0-8c59-3c11fbc66273","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Deep Learning-Based LLIE"],["subsection","Learning Strategies"]],"subsections":[],"title":"Learning Strategies"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Technical}\nIn this section, we first summarize the representative deep learning-based LLIE methods in Table \\ref{table:methods}, then analyze and discuss their technical characteristics.","id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","level":"section","origin_cites_number":0,"parent_id":"43aba542-7f6b-493f-9beb-9ae4164ca84b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"]],"subsections":["68e1f02c-799a-4eae-8da7-114a11e886b0","6a2b6ca0-a22f-4003-8b94-e1d5297a6a5e","ae53f5e7-73c5-4548-9f34-2ad3930fd293","30a23ace-8ed7-4b68-a78a-02ea2fbd6679","b98d1366-433d-4a92-8dac-6386ad39398e","e339615c-7f2d-4a6a-95fe-00d1129c19c1","afdfe29b-5291-42fa-b702-0e326d794db8"],"title":"Technical Review and Discussion"},{"cite_extract_rate":0,"cites":[],"content":"Diverse network structures and designs have been used in the existing models, spanning from the basic U-Net, pyramid network, multi-stage network to frequency decomposition network. \nAfter analyzing Figure \\ref{fig:statistic}(b), it can be observed that the U-Net and U-Net-like networks are mainly adopted network structures in LLIE. \nThis is because U-Net can effectively integrate multi-scale \nfeatures and employ both low-level and high-level features. Such characteristics are essential for achieving satisfactory low-light enhancement.\nNevertheless, some key issues may be ignored in the current LLIE network structures:\n\\textbf{1)} after going through several convolutional layers, the gradients of an extremely low light image may vanish during the gradient back-propagation due to its small pixel values. This would degrade the enhancement performance and affect the convergence of network training, \n\\textbf{2)} the skip-connections used in the U-Net-like networks might introduce noise and redundant features into the final results. How to effectively filter out the noise and integrate both low-level and high-level features should be carefully considered, and \n\\textbf{3)} although some designs and components are proposed for LLIE, most of them are borrowed or modified from related low-level visual tasks. The characteristics of low-light data should be considered when designing the network structures.","id":"68e1f02c-799a-4eae-8da7-114a11e886b0","level":"subsection","origin_cites_number":0,"parent_id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"],["subsection","Network Structure"]],"subsections":[],"title":"Network Structure"},{"cite_extract_rate":0,"cites":[],"content":"As presented in Figure \\ref{fig:statistic}(c), almost 1/3 of methods combine the designs of deep networks with the Retinex theory, e.g., designing different subnetworks to estimate the components of the Retinex model and estimating the illumination map to guide the learning of networks. Despite such a combination can bridge deep learning-based and model-based methods, their respective weaknesses may be introduced into the final models: \n\\textbf{1)} the ideal assumption that the reflectance is the final enhanced result used in Retinex-based LLIE methods would still affect the final results, and \n\\textbf{2)} the risk of overfitting in deep networks still exists despite the use of Retinex theory. How to cream off the best and filter out the impurities should be carefully considered when researchers combine deep learning with the Retinex theory.","id":"6a2b6ca0-a22f-4003-8b94-e1d5297a6a5e","level":"subsection","origin_cites_number":0,"parent_id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"],["subsection","Combination of Deep Model and Retinex Theory"]],"subsections":[],"title":"Combination of Deep Model and Retinex Theory"},{"cite_extract_rate":0,"cites":[],"content":"As shown in Figure \\ref{fig:statistic}(d), RGB data format dominates most methods as it is commonly found as the final imagery form produced by smartphone cameras, Go-Pro cameras, and drone cameras. Although raw data are limited to specific sensors such as those based on Bayer patterns, the data cover wider color gamut and higher dynamic range. Hence, deep models trained on raw data usually recover clear details and high contrast, obtain vivid color, reduce the effects of noises and artifacts, and improve the brightness of extremely low-light images. In future research, a smooth transformation from raw data of different patterns to RGB format would have the potentials to combine the convenience of RGB data and the advantage of high-quality enhancement of raw data for LLIE.","id":"ae53f5e7-73c5-4548-9f34-2ad3930fd293","level":"subsection","origin_cites_number":0,"parent_id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"],["subsection","Data Format"]],"subsections":[],"title":"Data Format"},{"cite_extract_rate":0,"cites":[],"content":"In Figure\\ref{fig:statistic}(e), the commonly adopted loss functions in LLIE models include reconstruction loss ($L_{1}$, $L_{2}$, SSIM), perceptual loss, and smoothness loss. Besides, according to different demands and formulations, color loss, exposure loss, adversarial loss, etc are also adopted. We detail representative loss functions as follows.\n\\noindent\n\\textbf{Reconstruction Loss.}\n\tDifferent reconstruction losses have their advantages and disadvantages. $L_{2}$ loss tends to penalize larger errors, but is tolerant to small errors. $L_{1}$ loss preserves colors and\n\tluminance well since an error is weighted equally regardless of the local structure. SSIM loss preserves the structure and texture well. Refer to this research paper for detailed analysis.\n\\noindent\n\\textbf{Perceptual Loss.}\n\tPerceptual loss , particularly the feature reconstruction loss, is proposed to constrain the results similar to the ground truth in the feature space. The loss improves the visual quality of results. It is defined as the Euclidean distance between the feature representations of an enhanced result and those of corresponding ground truth. The feature representations are typically extracted from the VGG network pre-trained on ImageNet dataset . \n\\noindent\n\\textbf{Smoothness Loss.}\n\tTo remove noise in the enhanced results or preserve the relationship of neighboring pixels, smoothness loss (TV loss) is often used to constrain the enhanced result or the estimated illumination map.\n\\noindent\n\\textbf{Adversarial Loss.}\n\tTo encourage enhanced results to be indistinguishable from reference images, adversarial learning solves a max-min optimization problem .\n\\noindent\n\\textbf{Exposure Loss.}\n\tAs one of key non-reference losses, exposure loss measures the exposure levels of enhanced results without paired or unpaired images as reference images.\n\tThe commonly used loss functions in LLIE networks are also employed in image reconstruction networks for image super-resolution , image denoising , image detraining , and image deblurring .\n\tDifferent from these versatile losses, the specially designed exposure loss for LLIE inspires the design of non-reference losses. A non-reference loss makes a model enjoying better generalization capability.\n\tIt is an on-going research to consider image characteristics for the design of loss functions.","id":"30a23ace-8ed7-4b68-a78a-02ea2fbd6679","level":"subsection","origin_cites_number":0,"parent_id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"],["subsection","Loss Function"]],"subsections":[],"title":"Loss Function"},{"cite_extract_rate":1,"cites":[1875],"content":"Figure \\ref{fig:statistic}(f) reports the usage of a variety of paired training datasets for training low-light enhancement networks. These datasets include real-world captured datasets and synthetic datasets. We list them in Table \\ref{table:training}. \n\\noindent\n\\textbf{Simulated by Gamma Correction.}\nOwing to its nonlinearity and simplicity, \nGamma correction is used to adjust the luminance or tristimulus values in video or still image systems. It is defined by a power-law expression:\n\\begin{equation}\n\t\\label{equ_12}\n\tV_{\\text{out}}=AV_{\\text{in}}^{\\gamma},\n\\end{equation}\nwhere the input $V_{\\text{in}}$ and output $V_{\\text{out}}$ are typically in the range of [0,1]. The constant $A$ is set to 1 in the common case. The power $\\gamma$ controls the luminance of the output. Intuitively, the input is brightened when $\\gamma<$1 while the input is darkened when $\\gamma>$1. The input can be the three RGB channels of an image or the luminance-related channels such as $L$ channel in the CIELab color space and $Y$ channel in the YCbCr color space. After adjusting the luminance-related channel using Gamma correction, the corresponding channels in the color space are adjusted by equal proportion to avoid producing artifacts and color deviations. \nTo simulate images taken in real-world low-light scenes, Gaussian noise, Poisson noise, or realistic noise is added to the Gamma corrected images. The low-light image synthesized using Gamma correction can be expressed as:\n\\begin{equation}\n\t\\label{equ_13}\n\tI_{\\text{low}}=n(g(I_{\\text{in}};\\gamma)),\n\\end{equation}\nwhere $n$ represents the noise model, $g(I_{\\text{in}};\\gamma)$ represents the Gamma correction function with Gamma value $\\gamma$, $I_{\\text{in}}$ is a normal-light and high-quality image or luminance-related channel. Although this function produces low-light images of different lighting levels by changing the Gamma value $\\gamma$, \nit tends to introduce artifacts and color deviations into the synthetic low-light images due to the nonlinear adjustment. \n\\begin{table}\n\t\\centering\n\t\\setlength\\tabcolsep{5pt}\n\t\\caption{\n\t\t{Summary of paired training datasets. `Syn' represents Synthetic.}\n\t}\n\t\\vspace{-6pt}\n\t\\label{table:training}\n\t\\begin{tabular}{r|c|c|c|c}\n\t\t\\hline\n\t\t\\textbf{Name}~~~~~~~~~~~~~~~~~~~~~&\\textbf{Number}& \n\t\t\\textbf{Format} & \\textbf{Real/Syn}&\\textbf{Video} \\\\\n\t\t\\hline\n\t\tGamma Correction &+$\\infty$ &RGB \n\t\t&Syn & \\\\\n\t\tRandom Illumination &+$\\infty$ &RGB \n\t\t&Syn & \\\\\n\t\t\\hline\n\t\tLOL &500 &RGB \n\t\t&Real &\\\\\n\t\tSCIE &4,413 &RGB \n\t\t&Real &\\\\\n\t\tVE-LOL-L &2,500 &RGB \n\t\t&Real+Syn &\\\\\n\t\tMIT-Adobe FiveK &5,000 &raw \n\t\t&Real &\\\\\n\t\tSID &5,094 &raw \n\t\t&Real &\\\\\n\t\tDRV &202 &raw \n\t\t&Real &\\checkmark\\\\\n\t\tSMOID &179 &raw \n\t\t&Real &\\checkmark\\\\\n\t\t\\hline\n\t\\end{tabular}\n\t\\vspace{-4pt}\n\\end{table}\n\\noindent\n\\textbf{Simulated by Random Illumination.}\nAccording to the Retinex model, an image can be decomposed into a reflectance component and an illumination component. Assuming image content is independent of illumination component and local region in the illumination component have the same intensity, a low-light image can be obtained by \n\\begin{equation}\n\t\\label{equ_14}\n\tI_{\\text{low}}=I_{\\text{in}}L,\n\\end{equation}\nwhere $L$ is a random illumination value in the range of [0,1]. Noises can be added to the synthetic image. Such a linear function avoids artifacts, but the strong assumption requires the synthesis to operate only on image patches where local regions have the same brightness. A deep model trained on such image patches may lead to sub-optimal performance due to the negligence of context information.\n\\noindent\n\\textbf{LOL.}\nLOL is the first paired low-/normal-light image dataset taken in real scenes. The low-light images are collected by changing the exposure time and ISO. LOL contains 500 pairs of low-/normal-light images of size 400$\\times$600 saved in RGB format. \n\\noindent\n\\textbf{SCIE.}\nSCIE is a multi-exposure image dataset of low-contrast and good-contrast image pairs. It includes multi-exposure sequences of 589 indoor and outdoor scenes. Each sequence has 3 to 18 low-contrast images of different exposure levels, thus containing 4,413 multi-exposure images in total. The 589 high-quality reference images are obtained by selecting from the results of 13 representative enhancement algorithms. That is many multi-exposure images have the same high-contrast reference image.\nThe image resolutions are between 3,000$\\times$2,000 and \n6,000$\\times$4,000. The images in SCIE are saved in RGB format.\n\\noindent\n\\textbf{MIT-Adobe FiveK.}\nMIT-Adobe FiveK was collected for global tone adjustment but has been used in LLIE. This is because the input images have low light and low contrast. MIT-Adobe FiveK contains 5,000 images, each of which is retouched by 5 trained photographers towards visually pleasing renditions, akin to a postcard. The images are all in raw format. To train the networks that can handle images of RGB format, one needs to use Adobe Lightroom to pre-process the images and save them as RGB format following a dedicated pipeline\\footnote{\\url{https://github.com/nothinglo/Deep-Photo-Enhancer/issues/38\\#issuecomment-449786636}}.\nThe images are commonly resized to have a long edge of 500 pixels.\n\\noindent\n\\textbf{SID.}\nSID contains 5,094 raw short-exposure images, each with a corresponding long-exposure reference image. The number of distinct long-exposure reference images is 424. In other words, multiple short-exposure images correspond to the same long-exposure reference image. \nThe images were taken using two cameras: Sony $\\alpha$7S II and Fujifilm X-T2 in both indoor and outdoor scenes. Thus, the images have different sensor patterns (Sony camera' Bayer sensor and Fuji camera's APS-C X-Trans sensor). \nThe resolution is 4,240$\\times$2,832 for Sony and 6,000$\\times$4,000 for Fuji. Usually, the long-exposure images are processed by libraw (a raw image processing library) and saved in the RGB color space, and randomly cropped 512$\\times$512 patches for training.\n\\noindent\n\\textbf{VE-LOL.} VE-LOL consists of two subsets: paired VE-LOL-L that is used for training and evaluating LLIE methods and unpaired VE-LOL-H that is used for evaluating the effect of LLIE methods on face detection. Specifically, VE-LOL-L includes 2,500 paired images. Among them, 1,000 pairs are synthetic, while 1,500 pairs are real. VE-LOL-H includes 10,940 unpaired images, where human faces are manually annotated with bounding boxes. \n\\noindent\n\\textbf{DRV.}\nDRV contains 202 static raw videos, each of which has a corresponding long-exposure ground truth. Each video was taken at approximately 16 to 18 frames per second in a continuous shooting mode and is with up to 110 frames. The images were taken by a Sony RX100 VI camera in both indoor and outdoor scenes, thus all in raw format of Bayer pattern. The resolution is 3,672$\\times$5,496. \n\\noindent\n\\textbf{SMOID.}\nSMOID contains 179 pairs of videos taken by a co-axis optical system, each of which has 200 frames. Thus, SMOID includes 35,800 extremely low light raw data of Bayer pattern and their corresponding well-lightened RGB counterparts. SMOID consists of moving vehicles and pedestrians under different illumination conditions.\nSome issues challenge the aforementioned paired training datasets: \n\\textbf{1)} deep models trained on synthetic data may introduce artifacts and color deviations when processing real-world images and videos due to the gap between synthetic data and real data, \n\\textbf{2)} the scale and diversity of real training data are unsatisfactory, thus some methods incorporate synthetic data to augment the training data. This may lead to sub-optimal enhancement, and \n\\textbf{3)} the input images and corresponding ground truths may exist misalignment due to the effects of motion, hardware, and environment. This would affect the performance of deep networks trained using pixel-wise loss functions.","id":"b98d1366-433d-4a92-8dac-6386ad39398e","level":"subsection","origin_cites_number":2,"parent_id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"],["subsection","Training Datasets"]],"subsections":[],"title":"Training Datasets"},{"cite_extract_rate":1,"cites":[1875],"content":"In addition to the testing subsets in the paired datasets , there are several testing data collected from related works or commonly used for experimental comparisons. \nBesides, some datasets such as face detection in the dark and detection and recognition in low-light images are employed to test the effects of LLIE on high-level visual tasks. We summarize the commonly used testing datasets in Table \\ref{table:testing} and introduce the representative testing datasets as follows. \n\\begin{table}\n\t\\centering\n\t\\caption{\n\t\t{Summary of testing datasets.}\n\t}\n\t\\vspace{-6pt}\n\t\\label{table:testing}\n\t\\begin{tabular}{r|c|c|c|c}\n\t\t\\hline\n\t\t\\textbf{Name}~~~~~~~~~&\\textbf{Number}& \n\t\t\\textbf{Format} & \\textbf{Application}&\\textbf{Video} \\\\\n\t\t\\hline\n\t\tLIME &10 &RGB & & \\\\\n\t\tNPE &84 &RGB & & \\\\\n\t\tMEF &17 &RGB & &\\\\\n\t\tDICM &64 &RGB & &\\\\\n\t\tVV\\tablefootnote{\\url{https://sites.google.com/site/vonikakis/datasets}} &24 &RGB & &\\\\\n\t\t\\hline\n\t\tBBD-100K &10,000 &RGB &\\checkmark &\\checkmark\\\\\n\t\tExDARK &7,363 &RGB &\\checkmark &\\\\\n\t\tDARK FACE &6,000 &RGB &\\checkmark &\\\\\n\t\tVE-LOL-H &10,940 &RGB &\\checkmark &\\\\\n\t\t\\hline\n\t\\end{tabular}\n\t\\vspace{-4pt}\n\\end{table}\n\\noindent\n\\textbf{BBD-100K.}\nBBD-100K is the largest driving video dataset with 10,000 videos taken over 1,100-hour driving experience across many different times in the day, weather conditions, and driving scenarios, and 10 tasks annotations. The videos taken at nighttime in BBD-100K are used to validate the effects of LLIE on high-level visual tasks and the enhancement performance in real scenarios.\n\\noindent\n\\textbf{ExDARK.}\nExDARK~ dataset is built for object detection and recognition in low-light images. ExDARK dataset contains 7,363 low-light images from extremely low-light environments to twilight with 12 object classes annotated with image class labels and local object bounding boxes.\n\\noindent\n\\textbf{DARK FACE.}\nDARK FACE dataset contains 6,000 low-light images captured during the nighttime, each of which is labeled with bounding boxes of the human face.\nFrom Figure \\ref{table:testing}(g) and Table \\ref{table:methods}, we can observe that one prefers using the self-collected testing data in the experiments. The main reasons lie into three-fold: \n\\textbf{1)} besides the test partition of paired datasets, there is no acknowledged benchmark for evaluations, \\textbf{2)}\nthe commonly used test sets suffer from some shortcomings such as small scale (some test sets contain 10 images only), repeated content and illumination properties, and unknown experimental settings, and\\textbf{ 3)} some of the commonly used testing data are not originally collected for evaluating LLIE. In general, current testing datasets may lead to bias and unfair comparisons.","id":"e339615c-7f2d-4a6a-95fe-00d1129c19c1","level":"subsection","origin_cites_number":4,"parent_id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"],["subsection","Testing Datasets"]],"subsections":[],"title":"Testing Datasets"},{"cite_extract_rate":0,"cites":[],"content":"Besides human perception-based subjective evaluations, image quality assessment (IQA) metrics, including both full-reference and non-reference IQA metrics, are able to evaluate image quality objectively. In addition, user study, number of trainable parameters, FLOPs, runtime, and applications also reflect the performance of LLIE models, as shown in Fig. \\ref{fig:statistic}(h). \nWe will detail them as follows.\n\\noindent\n\\textbf{PSNR and MSE.}\nPSNR and MSE are widely used IQA metrics. They are always non-negative, and values closer to infinite (PSNR) and zero (MSE) are better.\nNevertheless, the pixel-wise PSNR and MSE may provide an inaccurate indication of the visual perception of image quality since they neglect the relation of neighboring pixels.\n\\noindent\n\\textbf{MAE.}\nMAE represents the mean absolute error, serving as a measure of errors between paired observations. \nThe smaller the MAE value is, the better similarity is.\n\\noindent\n\\textbf{SSIM.}\nSSIM is used to measure the similarity between two images. It is a perception-based model that considers image degradation as perceived change in structural information. \nThe value 1 is only reachable in the case of two identical sets of data, indicating perfect structural similarity. \n\\noindent\n\\textbf{LOE.}\nLOE represents the lightness order error that reflects the naturalness of an enhanced image.\nFor LOE, the smaller the LOE value is, the better the lightness order is preserved.\n\\noindent\n\\textbf{Application.} \nBesides improving the visual quality, one of the purposes of image enhancement is to serve high-level visual tasks. Thus, the effects of LLIE on high-level visual applications are commonly examined to validate the performance of different methods.\nThe current evaluation approaches used in LLIE need to be improved in several aspects: \n\\textbf{1)} although the PSNR, MSE, MAE, and SSIM are classic and popular metrics, they are still far from capturing real visual perception of human,\n\\textbf{2)} some metrics are not originally designed for low-light images. They are used for assessing the fidelity of image information and contrast. Using these metrics may reflect the image quality, but they are far from the real purpose of low-light enhancement,\n\\textbf{3)} metrics especially designed for low-light images are lacking, except for the LOE metric. Moreover, there is no metric for evaluating low-light video enhancement, and\n\\textbf{4)} a metric that can balance both the human vision and the machine perception is expected.","id":"afdfe29b-5291-42fa-b702-0e326d794db8","level":"subsection","origin_cites_number":0,"parent_id":"3ad6854d-28fe-41e8-80bd-198011ce8b19","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Technical Review and Discussion"],["subsection","Evaluation Metrics"]],"subsections":[],"title":"Evaluation Metrics"},{"cite_extract_rate":1,"cites":[1875],"content":"\\label{sec:evaluation}\nThis section provides empirical analysis and highlights some key challenges in deep learning-based LLIE. To facilitate the analysis, we propose a low-light image and video dataset to examine the performance of different solutions. We also develop the first online platform, where the results of LLIE models can be produced via a user-friendly web interface. In this section, we conduct extensive evaluations on several benchmarks and our proposed dataset. \nIn the experiments, we compare 13 representative RGB format-based methods, including eight supervised learning-based methods (LLNet , LightenNet , Retinex-Net , MBLLEN , KinD , KinD++ , TBEFN , DSLR ), one unsupervised learning-based method (EnlightenGAN ), one semi-supervised learning-based method (DRBN ), and three zero-shot learning-based methods (ExCNet , Zero-DCE , RRDNet ). Besides, we also compare two raw format-based methods, including SID and EEMEFN . Note that RGB format-based methods dominate LLIE. Moreover, most raw format-based methods do not release their code. Thus, we choose two representative methods to provide empirical analysis and insights. For all compared methods, we use the publicly available code to produce their results for fair comparisons. \n\\begin{table}\n\t\\centering\n\t\\caption{{Summary of LLIV-Phone dataset. LLIV-Phone dataset contains 120 videos (45,148 images) taken by 18 different mobile phones' cameras. ``\\#Video'' and ``\\#Image'' represent the number of videos and images, respectively.}}\n\t\\vspace{-6pt}\n\t\\label{table:LLIVPhone}\n\t\\begin{tabular}{r|c|c|c}\n\t\t\\hline\n\t\t\\textbf{Phone's Brand}&\\textbf{\\#Video}& \n\t\t\\textbf{\\#Image} &\\textbf{Resolution}\\\\\n\t\t\\hline\n\t\tiPhone 6s &4 &1,029 &1920$\\times$1080\\\\\n\t\tiPhone 7 &13 &6,081 &1920$\\times$1080\\\\\n\t\tiPhone7 Plus &2 &900 &1920$\\times$1080\\\\\n\t\tiPhone8 Plus &1 &489 &1280$\\times$720\\\\\n\t\tiPhone 11 &7 &2,200 &1920$\\times$1080\\\\\n\t\tiPhone 11 Pro &17 &7,739 &1920$\\times$1080\\\\\n\t\tiPhone XS &11 &2,470 &1920$\\times$1080\\\\\n\t\tiPhone XR &16 &4,997 &1920$\\times$1080\\\\\n\t\tiPhone SE &1 &455 &1920$\\times$1080\\\\\n\t\tXiaomi Mi 9 &2 &1,145 &1920$\\times$1080\\\\\n\t\tXiaomi Mi Mix 3 &6 &2,972 &1920$\\times$1080\\\\\n\t\tPixel 3&4 &1,311 &1920$\\times$1080\\\\\n\t\tPixel 4&3 &1,923 &1920$\\times$1080\\\\\n\t\tOppo R17 &6 &2,126 &1920$\\times$1080\\\\\n\t\tVivo Nex &12 &4,097 &1280$\\times$720\\\\\n\t\tLG M322&2 &761 &1920$\\times$1080\\\\\n\t\tOnePlus 5T &1 &293 &1920$\\times$1080\\\\\n\t\tHuawei Mate 20 Pro & 12&4,160 &1920$\\times$1080\\\\\n\t\t\\hline\n\t\\end{tabular}\n\t\\vspace{-4pt}\n\\end{table}","id":"62b1377f-b6e1-4398-be9f-a1fa2952056b","level":"section","origin_cites_number":3,"parent_id":"43aba542-7f6b-493f-9beb-9ae4164ca84b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Benchmarking and Empirical Analysis"]],"subsections":["d8e5b8c1-ee06-4aad-800d-232c4dd98922","7ee6e01b-208a-4fef-a9a1-3cc6a973f6c4","67f70829-d734-4fe0-baa6-b7f64f6f5b82","e142566f-157a-4609-a670-8df6dbfe9874","8ea7b52c-e5f5-46c9-8ff5-69763784bc99"],"title":"Benchmarking and Empirical Analysis"},{"cite_extract_rate":1,"cites":[1875],"content":"We propose a Low-Light Image and Video dataset, called LLIV-Phone, to comprehensively and thoroughly validate the performance of LLIE methods. LLIV-Phone is the largest and most challenging real-world testing dataset of its kind. In particular, the dataset contains 120 videos (45,148 images) taken by 18 different mobile phones' cameras including iPhone 6s, iPhone 7, iPhone7 Plus, iPhone8 Plus, iPhone 11, iPhone 11 Pro, iPhone XS, iPhone XR, iPhone SE, Xiaomi Mi 9, Xiaomi Mi Mix 3, Pixel 3, Pixel 4, Oppo R17, Vivo Nex, LG M322, OnePlus 5T, Huawei Mate 20 Pro under diverse illumination conditions (e.g., weak lighting, underexposure, moonlight, twilight, dark, extremely dark, back-lit, non-uniform light, and colored light.) in both indoor and outdoor scenes. A summary of the LLIV-Phone dataset is provided in Table \\ref{table:LLIVPhone}. \nWe present several samples of LLIV-Phone dataset in Figure \\ref{fig:dataset_samples}. \nThe LLIV-Phone dataset is available at the project page. \nThis challenging dataset is collected in real scenes and contains diverse low-light images and videos. Consequently, it is suitable for evaluating the generalization capability of different low-light image and video enhancement models. Notably, the dataset can be used as the training dataset for unsupervised learning and the reference dataset for synthesis methods to generate realistic low-light data. \n\\begin{figure}[!t]\n\t\\centering \\centerline{\\includegraphics[width=1\\linewidth]{figures/dataset_samples.png}}\n\t\\vspace{-2pt}\n\t\\caption{Several images sampled from the proposed LLIV-Phone dataset. The images and videos are taken by different devices under diverse lighting conditions and scenes.}\n\t\\label{fig:dataset_samples}\n\t\\vspace{-4pt}\n\\end{figure}\n\\begin{figure*} [!t]\n\t\\begin{center}\n\t\t\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }c@{ }c@{ }c}\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_input.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_LLNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_LightenNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_RetinexNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_MBLLEN.png}\\\\\n\t\t\t(a) input & (b) LLNet & (c) LightenNet & (d) Retinex-Net & (e) \tMBLLEN \\\\\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_KinD.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_KinD++.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_TBEFN.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_DSLR.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_EnlightenGAN.png}\\\\\n\t\t\t(f) KinD & (g) KinD++ & (h) TBEFN & (i) \tDSLR & (j) EnlightenGAN \\\\\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_DRBN.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_ExCNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_ZeroDCE.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_RRDNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/LOL_GT.png}\\\\\n\t\t\t(k) DRBN & (l)\tExCNet & (m) Zero-DCE & (n) \tRRDNet & (o) GT \\\\\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\vspace{-2pt}\n\t\\caption{Visual results of different methods on a low-light image sampled from LOL-test dataset.}\n\t\\label{fig:LOL}\n\t\\vspace{-4pt}\n\\end{figure*}\n\\begin{figure*}[!t]\n\t\\begin{center}\n\t\t\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }c@{ }c@{ }c}\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_input.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_LLNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_LightenNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_RetinexNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_MBLLEN.png}\\\\\n\t\t\t(a) input & (b) LLNet & (c) LightenNet & (d) Retinex-Net & (e) \tMBLLEN \\\\\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_KinD.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_KinD++.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_TBEFN.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_DSLR.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_EnlightenGAN.png}\\\\\n\t\t\t(f) KinD & (g) KinD++ & (h) TBEFN & (i) \tDSLR & (j) EnlightenGAN \\\\\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_DRBN.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_ExCNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_ZeroDCE.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_RRDNet.png}&\n\t\t\t\\includegraphics[width=0.18\\linewidth]{figures/5K_GT.png}\\\\\n\t\t\t(k) DRBN & (l)\tExCNet & (m) Zero-DCE & (n) \tRRDNet & (o) GT \\\\\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\vspace{-2pt}\n\t\\caption{Visual results of different methods on a low-light image sampled from MIT-Adobe FiveK-test dataset.}\n\t\\label{fig:5K}\n\t\\vspace{-4pt}\n\\end{figure*}","id":"d8e5b8c1-ee06-4aad-800d-232c4dd98922","level":"subsection","origin_cites_number":3,"parent_id":"62b1377f-b6e1-4398-be9f-a1fa2952056b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Benchmarking and Empirical Analysis"],["subsection","A New Low-Light Image and Video Dataset"]],"subsections":[],"title":"A New Low-Light Image and Video Dataset"},{"cite_extract_rate":1,"cites":[1875],"content":"Different deep models may be implemented in different platforms such as Caffe, Theano, TensorFlow, and PyTorch. As a result, different algorithms demand different configurations, GPU versions, and hardware specifications. Such requirements are prohibitive to many researchers, especially for beginners who are new to this area and may not even have GPU resources. To resolve these problems, we develop an LLIE online platform, called LLIE-Platform, which is available at \\url{http://mc.nankai.edu.cn/ll/}. \nTo the date of this submission, the LLIE-Platform covers 14 popular deep learning-based LLIE methods including LLNet , LightenNet , Retinex-Net , EnlightenGAN , MBLLEN , KinD , KinD++ , TBEFN , DSLR , DRBN , ExCNet , Zero-DCE , Zero-DCE++ , and RRDNet , where the results of any input can be produced through a user-friendly web interface. We will regularly offer new methods on this platform.\nWe wish that this LLIE-Platform could serve the\ngrowing research community by providing users a flexible interface\nto run existing deep learning-based LLIE methods and develop their own new LLIE methods. \n\\newcommand{\\addImg}[1]{\\includegraphics[width=0.14\\linewidth,height=0.2\\linewidth]{figures/#1}}\n\\begin{figure*}[!t]\n\t\\centering\n\t\\setlength\\tabcolsep{0.5pt}\n\t\\begin{tabular}{ccccccccc}\n\t\t\\addImg{Phone_input1.png} &\n\t\t\\addImg{Phone_LLNet1.png}&\n\t\t\\addImg{Phone_LightenNet1.png}&\n\t\t\\addImg{Phone_RetinexNet1.png}&\n\t\t\\addImg{Phone_MBLLEN1.png}&\n\t\t\\addImg{Phone_KinD1.png}&\n\t\t\\addImg{Phone_KinD++1.png}&\\\\\n\t\t\t(a) & (b) & (c) & (d) & (e) & (f) & (g) \\\\\n\t\t\\addImg{Phone_TBEFN1.png}&\n\t\t\\addImg{Phone_DSLR1.png}&\n\t\t\\addImg{Phone_EnlightenGAN1.png}&\n\t\t\\addImg{Phone_DRBN1.png}&\n\t\t\\addImg{Phone_ExCNet1.png}&\n\t\t\\addImg{Phone_ZeroDCE1.png}&\n\t\t\\addImg{Phone_RRDNet1.png}\\\\\n\t\t\t(h) & (i) \t & (j) & \t(k) & (l) & (m) & (n) \t \\\\\n\t\\end{tabular}\n\t\\vspace{-12pt}\n\t\\caption{Visual results of different methods on a low-light image sampled from LLIV-Phone-imgT dataset. (a) input. (b) LLNet . (c) LightenNet . (d) Retinex-Net . (e) \tMBLLEN . (f) KinD . (g) KinD++ . (h) TBEFN . (i) \tDSLR . (j) EnlightenGAN .\t(k) DRBN . (l)\tExCNet . (m) Zero-DCE . (n) \tRRDNet .}\n\t\\label{fig:Phone1}\n\t\\vspace{-4pt}\n\\end{figure*}\n\\begin{figure*}[!t]\n\t\\setlength\\tabcolsep{0.5pt}\n\t\\centering\n\t\\begin{tabular}{ccccccccc}\n\t\t\\addImg{Phone_input2.png}&\n\t\t\\addImg{Phone_LLNet2.png}&\n\t\t\\addImg{Phone_LightenNet2.png}&\n\t\t\\addImg{Phone_RetinexNet2.png}&\n\t\t\\addImg{Phone_MBLLEN2.png}&\n\t\t\\addImg{Phone_KinD2.png}&\n\t\t\\addImg{Phone_KinD++2.png}&\\\\\n\t\t\t(a) & (b) & (c) & (d) & (e) \t & (f) & (g) \\\\\n\t\t\\addImg{Phone_TBEFN2.png}&\n\t\t\\addImg{Phone_DSLR2.png}&\n\t\t\\addImg{Phone_EnlightenGAN2.png}&\n\t\t\\addImg{Phone_DRBN2.png}&\n\t\t\\addImg{Phone_ExCNet2.png}&\n\t\t\\addImg{Phone_ZeroDCE2.png}&\n\t\t\\addImg{Phone_RRDNet2.png}\\\\\n\t\t\t(h) & (i) \t & (j) & \t(k) & (l)\t & (m) & (n) \t \\\\\n\t\\end{tabular}\n\t\\vspace{-12pt}\n\t\\caption{Visual results of different methods on a low-light image sampled from LLIV-Phone-imgT dataset. (a) input. (b) LLNet . (c) LightenNet . (d) Retinex-Net . (e) \tMBLLEN . (f) KinD . (g) KinD++ . (h) TBEFN . (i) \tDSLR . (j) EnlightenGAN .\t(k) DRBN . (l)\tExCNet . (m) Zero-DCE . (n) \tRRDNet .}\n\t\\label{fig:Phone2}\n\t\\vspace{-4pt}\n\\end{figure*}\n\\begin{figure*}[!t]\n\t\\begin{center}\n\t\t\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }c@{ }}\n\t\t\t\\rotatebox{90}{~~~~~~~~~Bayer~~~~~~~~~~~~APS-C X-Trans}~&\n\t\t\t\\includegraphics[width=0.21\\linewidth]{figures/raw_input.png}&~~\n\t\t\t\\includegraphics[width=0.21\\linewidth]{figures/raw_SID.png}&~~\n\t\t\t\\includegraphics[width=0.21\\linewidth]{figures/raw_EEMEFN.png}&~~\n\t\t\t\\includegraphics[width=0.21\\linewidth]{figures/raw_GT.png}\\\\\n\t\t\t&~\t(a) inputs &~~ (b) SID &~~ (c) EEMEFN &~~ (d) GT \\\\\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\vspace{-2pt}\n\t\\caption{Visual results of different methods on two raw low-light images sampled from SID-test-Bayer and SID-test-X-Trans test datasets. The inputs are amplified for visualization.}\n\t\\label{fig:raw}\n\t\\vspace{-4pt}\n\\end{figure*}","id":"7ee6e01b-208a-4fef-a9a1-3cc6a973f6c4","level":"subsection","origin_cites_number":4,"parent_id":"62b1377f-b6e1-4398-be9f-a1fa2952056b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Benchmarking and Empirical Analysis"],["subsection","Online Evaluation Platform"]],"subsections":[],"title":"Online Evaluation Platform"},{"cite_extract_rate":1,"cites":[1875],"content":"To qualitatively and quantitatively evaluate different methods, in addition to the proposed LLIV-Phone dataset, we also adopt the commonly used LOL and MIT-Adobe FiveK datasets for RGB format-based methods, and SID dataset for raw format-based methods. More visual results can be found in the supplementary material. The comparative results on the real low-light videos taken by different mobile phones' cameras can be found at YouTube \\url{https://www.youtube.com/watch?v=Elo9TkrG5Oo&t=6s}.\nWe select five images on average from each video of the LLIV-Phone dataset, forming an image testing dataset with a total of 600 images (denoted as LLIV-Phone-imgT). Furthermore, we randomly select one video from the videos of each phone's brand of LLIV-Phone dataset, forming a video testing dataset with a total of 18 videos (denoted as LLIV-Phone-vidT). We half the resolutions of the frames in both LLIV-Phone-imgT and LLIV-Phone-vidT because some deep learning-based methods cannot process the full resolution of test images and videos. \nFor the LOL dataset, we adopt the original test set including 15 low-light images captured in real scenes for testing, denoted as LOL-test. For the MIT-Adobe FiveK dataset, we follow the protocol in Chen et al. to decode the images into PNG format and resize them to have a long edge of 512 pixels using Lightroom. We adopt the same testing dataset as Chen et al. , MIT-Adobe FiveK-test, including 500 images with the retouching results by expert C as the corresponding ground truths. For the SID dataset, we use the default test set used in EEMEFN for fair comparisons, denoted as SID-test (SID-test-Bayer and SID-test-X-Trans), which is a partial test set of SID . The SID-test-Bayer includes 93 images of the Bayer pattern while the SID-test-X-Trans includes 94 images of the APS-C X-Trans pattern. \n\\begin{table*}[t]\n\t\\caption{Quantitative comparisons on LOL-test and MIT-Adobe FiveK-test test datasets in terms of MSE ($\\times10^3$), PSNR (in dB), SSIM , and LPIPS . The best result is in {\\color{red}{red}} whereas the second and third best results are in {\\color{blue}{blue}} and {\\color{purple}{purple}} under each case, respectively.}\n\t\\vspace{-6pt}\n\t\\begin{center}\n\t\t\\begin{tabular}{c|c|c|c|c|c||c|c|c|c}\n\t\t\t\\hline\n\t\t\t\\multirow{2}{*}{\\textbf{Learning}} &\\multirow{2}{*}{\\textbf{Method}} & \\multicolumn{4}{c||}{\\textbf{LOL-test}} & \\multicolumn{4}{c}{\\textbf{MIT-Adobe FiveK-test}} \\\\\n\t\t\t\\cline{3-10}\n\t\t\t\\cline{3-10}\n\t\t\t& &\t\\textbf{MSE}$\\downarrow$ & \\textbf{PSNR} $\\uparrow$ & \\textbf{SSIM}$\\uparrow$ & \\textbf{LPIPS}$\\downarrow$ & \\textbf{MSE}$\\downarrow$ & \\textbf{PSNR} $\\uparrow$ & \\textbf{SSIM}$\\uparrow$ & \\textbf{LPIPS}$\\downarrow$ \\\\\n\t\t\t\\hline\n\t\t\t&input & 12.613 & 7.773 &0.181 & 0.560& {\\color{blue}{1.670}}& {\\color{blue}{17.824}} & {\\color{purple}{0.779}} &{\\color{blue}{0.148}} \\\\\n\t\t\t\\hline\n\t\t\t&LLNet & {\\color{red}{1.290}} & {\\color{red}{17.959}} &0.713 &0.360&4.465 & 12.177 &0.645 &0.292\\\\\n\t\t\t&LightenNet & 7.614 & 10.301 & 0.402 & 0.394 & 4.127 & 13.579 & 0.744 &{\\color{purple}{0.166}}\\\\\n\t\t\t&Retinex-Net & 1.651 & 16.774& 0.462 & 0.474 & 4.406 & 12.310 & 0.671 &0.239\\\\\n\t\t\tSL&MBLLEN &1.444&{\\color{blue}{17.902}}&0.715& 0.247& {\\color{red}{1.296}} & {\\color{red}{19.781}} & {\\color{red}{0.825}}&{\\color{red}{0.108}} \\\\\n\t\t\t&KinD & {\\color{purple}{1.431}} & 17.648& {\\color{blue}{0.779}}& {\\color{red}{0.175}} & 2.675 & 14.535 & 0.741 &0.177 \\\\\n\t\t\t&KinD++ & {\\color{blue}{1.298}}&{\\color{purple}{17.752}}& {\\color{purple}{0.760}}& {\\color{blue}{0.198}} & 7.582 & 9.732 &0.568 & 0.336 \\\\\n\t\t\t&TBEFN &1.764&17.351& {\\color{red}{0.786}}& {\\color{purple}{0.210}} & 3.865 & 12.769 &0.704 &0.178 \\\\\n\t\t\t&DSLR &3.536&15.050& 0.597&0.337 &{\\color{purple}{1.925}} & {\\color{purple}{16.632}} &{\\color{blue}{0.782}} & 0.167 \\\\\n\t\t\t\\hline\n\t\t\tUL\t& EnlightenGAN &1.998&17.483& 0.677&0.322 &3.628&13.260&0.745& 0.170 \\\\\n\t\t\t\\hline\n\t\t\tSSL\t&DRBN &2.359&15.125& 0.472&0.316 &3.314& 13.355&0.378 & 0.281 \\\\\n\t\t\t\\hline\n\t\t\t&\tExCNet &2.292&15.783& 0.515&0.373 &2.927& 13.978&0.710 &0.187 \\\\\t\t\t\n\t\t\tZSL\t&Zero-DCE &3.282&14.861& 0.589&0.335 &3.476&13.199&0.709 &0.203\\\\\t\t\t\t\n\t\t\t&\tRRDNet &6.313&11.392&0.468&0.361 &7.057&10.135&0.620 &0.303\\\\\t\t\t\t\t\t\t\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\label{Table:Quan_1}\n\t\\vspace{-4pt}\n\\end{table*}\n\\begin{table*}[t]\n\t\\caption{Quantitative comparisons on SID-test test dataset in terms of MSE ($\\times10^3$), PSNR (in dB), SSIM , and LPIPS . The best result is in {\\color{red}{red}} under each case. To compute the quantitative scores of input raw data, we use the corresponding camera ISP pipelines provided by Chen et al. to transfer raw data to RGB format.}\n\t\\vspace{-6pt}\n\t\\begin{center}\n\t\t\\begin{tabular}{c|c|c|c|c|c||c|c|c|c}\n\t\t\t\\hline\n\t\t\t\\multirow{2}{*}{\\textbf{Learning}} &\\multirow{2}{*}{\\textbf{Method}} & \\multicolumn{4}{c||}{\\textbf{SID-test--Bayer}} & \\multicolumn{4}{c}{\\textbf{SID-test--X-Trans}} \\\\\n\t\t\t\\cline{3-10}\n\t\t\t\\cline{3-10}\n\t\t\t& &\t\\textbf{MSE}$\\downarrow$ & \\textbf{PSNR} $\\uparrow$ & \\textbf{SSIM}$\\uparrow$ & \\textbf{LPIPS}$\\downarrow$ & \\textbf{MSE}$\\downarrow$ & \\textbf{PSNR} $\\uparrow$ & \\textbf{SSIM}$\\uparrow$ & \\textbf{LPIPS}$\\downarrow$ \\\\\n\t\t\t\\hline\n\t\t\t&input &5.378 &11.840 &0.063 &0.711 &4.803 &11.880 &0.075 &0.796 \\\\\n\t\t\t\\hline\n\t\t\tSL&SID &0.140 &28.614 &0.757 &0.465 &0.235 &26.663 &0.680&0.586\\\\\n\t\t\t&EEMEFN &{\\color{red}{0.126}} &{\\color{red}{29.212}} &{\\color{red}{0.768}} &{\\color{red}{0.448}}&{\\color{red}{0.191}} &{\\color{red}{27.423}} &{\\color{red}{0.695}} &{\\color{red}{0.546}}\\\\\t\t\t\t\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\label{Table:Quan_11}\n\t\\vspace{-4pt}\n\\end{table*}\n\\noindent\n\\textbf{Qualitative Comparison.}\nWe first present the results of different methods on the images sampled from LOL-test and MIT-Adobe FiveK-test datasets in Figures \\ref{fig:LOL} and \\ref{fig:5K}. \nAs shown in Figure \\ref{fig:LOL}, all methods improve the brightness and contrast of the input image. However, none of them successfully recovers the accurate color of the input image when the results are compared with the ground truth. In particular, LLNet produces blurring result. LightenNet and RRDNet produce under-exposed results while MBLLEN and ExCNet over-expose the image. KinD , KinD++ , TBEFN , DSLR , EnlightenGAN , and DRBN introduce obvious artifacts.\nIn Figure \\ref{fig:5K}, LLNet , KinD++ , TBEFN , and RRDNet produce over-exposed results. Retinex-Net , KinD++ , and RRDNet yield artifacts and blurring in the results. \nWe found that the ground truths of MIT-Adobe FiveK dataset still contain some dark regions. This is because the dataset is originally designed for global image retouching, where restoring low light regions is not the main priority in this task. \nWe also observed that the input images in LOL dataset and MIT-Adobe FiveK dataset are relatively clean from noise, which is different from real low-light scenes. \nAlthough some methods take the MIT-Adobe FiveK dataset as the training or testing dataset, we argue that this dataset is not appropriate for the task of LLIE due to its mismatched/unsatisfactory ground truth for LLIE.\nTo examine the generalization capability of different methods, we conduct comparisons on the images sampled from our LLIV-Phone-imgT dataset. The visual results of different methods are shown in Figures \\ref{fig:Phone1} and \\ref{fig:Phone2}.\nAs presented in Figure \\ref{fig:Phone1}, all methods cannot effectively improve the brightness and remove the noise of the input low-light image. Moreover, Retinex-Net , MBLLEN , and DRBN produce obvious artifacts.\nIn Figure \\ref{fig:Phone2}, all methods enhance the brightness of this input image. However, only \nMBLLEN and RRDNet obtain visually pleasing enhancement without color deviation, artifacts, and over-/under-exposure. Notably, for regions with a light source, none of the methods can brighten the image without amplifying the noise around these regions. \nTaking light sources into account for LLIE would be an interesting direction to explore.\nThe results suggest the difficulty of enhancing the images of the LLIV-Phone-imgT dataset. \nReal low-light images fail most existing LLIE methods due to the limited generalization capability of these methods. The potential reasons are the use of synthetic training data, small-scaled training data, or unrealistic assumptions such as the local illumination consistency and treating the reflectance component as the final result in the Retinex model in these methods.\nWe further present the visual comparisons of raw format-based methods in Figure \\ref{fig:raw}. As shown, the input raw data have obvious noises. Both SID and EEMEFN can effectively remove the effects of noises. In comparison to the simple U-Net structure used in SID , the more complex structure of EEMEFN obtains better brightness recovery. However, their results are far from the corresponding GT, especially for the input of APS-C X-Trans pattern.\n\\noindent\n\\textbf{Quantitative Comparison.}\nFor test sets with ground truth i.e., LOL-test, MIT-Adobe FiveK-test, and SID-test, we adopt MSE, PSNR, SSIM , and LPIPS metrics to quantitatively compare different methods. LPIPS is a deep learning-based image quality assessment metric that measures the perceptual similarity between a result and its corresponding ground truth by deep visual representations. For LPIPS, we employ the AlexNet-based model to compute the perceptual similarity. A lower LPIPS\nvalue suggests a result that is closer to the corresponding ground truth in terms of perceptual similarity. \nIn Table \\ref{Table:Quan_1} and Table \\ref{Table:Quan_11}, we show the quantitative results of RGB format-based methods and raw format-based methods, respectively.\nAs presented in Table \\ref{Table:Quan_1}, the quantitative scores of supervised learning-based methods are better than those of unsupervised learning-based, semi-supervised learning-based, and zero-shot learning-based methods on LOL-test and MIT-Adobe FiveK-test datasets. Among them, LLNet obtains the best MSE and PSNR values on the LOL-test dataset; however, its performance drops on the MIT-Adobe FiveK-test dataset. \nThis may be caused by the bias of LLNet towards the LOL dataset since it was trained using the LOL training dataset. \nFor the LOL-test dataset, TBEFN obtains the highest SSIM value while KinD achieves the lowest LPIPS value. There is no winner across these four evaluation metrics on the LOL-test dataset despite the fact that some methods were trained on the LOL training dataset. \nFor the MIT-Adobe FiveK-test dataset, MBLLEN outperforms all compared methods under the four evaluation metrics in spite of being trained on synthetic training data. Nevertheless, MBLLEN still cannot obtain the best performance on both two test datasets. \nAs presented in Table \\ref{Table:Quan_11}, both SID and EEMEFN improve the quality of input raw data. Compared with the quantitative scores of SID , EEMEFN achieves consistently better performance across different raw data patterns and evaluation metrics.\nFor LLIV-Phone-imgT test set, we use the non-reference IQA metrics, i.e., NIQE , perceptual index (PI) , LOE , and SPAQ to quantitatively compare different methods. In terms of LOE, the smaller the LOE value is, the better the lightness order is preserved. For NIQE, the smaller the NIQE value is, the better the visual quality is. A lower PI value indicates better perceptual quality. SPAQ is devised for the perceptual quality assessment of smartphone photography. A larger SPAQ value suggests better perceptual quality of smartphone photography. The quantitative results are provided in Table \\ref{Table:Quan_2}.\n\\begin{table}[ht]\n\t\\caption{Quantitative comparisons on LLIV-Phone-imgT dataset in terms of NIQE , LOE , PI , and SPAQ . The best result is in {\\color{red}{red}} whereas the second and third best results are in {\\color{blue}{blue}} and {\\color{purple}{purple}} under each case, respectively.}\n\t\\begin{center}\n\t\t\\setlength{\\tabcolsep}{1.7mm}{\n\t\t\t\\begin{tabular}{c|c|c|c|c|c}\n\t\t\t\t\\hline\n\t\t\t\t\\multirow{2}{*}{\\textbf{Learning}}&\\multirow{2}{*}{\\textbf{Method}} & \\multicolumn{4}{c}{\\textbf{LoLi-Phone-imgT}}\\\\\n\t\t\t\t\\cline{3-6}\n\t\t\t\t\\cline{3-6}\n\t\t\t\t&\t& \\textbf{NIQE}$\\downarrow$ & \\textbf{LOE} $\\downarrow$ & \\textbf{PI}$\\downarrow$ & \\textbf{SPAQ}$\\uparrow$ \\\\\n\t\t\t\t\\hline\n\t\t\t\t&\tinput &6.99 &{\\color{red}{0.00}} &5.86 &44.45 \\\\\n\t\t\t\t\\hline\n\t\t\t\t&\tLLNet &5.86 &{\\color{blue}{5.86}} &5.66 &40.56 \\\\\n\t\t\t\t&\tLightenNet &5.34 &952.33 &4.58 &45.74 \\\\\n\t\t\t\t&\t\n\t\t\t\tRetinex-Net & 5.01 &790.21 &{\\color{red}{3.48}} &{\\color{red}{50.95}} \\\\\n\t\t\t\tSL&MBLLEN &5.08 &220.63 & 4.27 &42.50 \\\\\n\t\t\t\t&\tKinD &4.97 &405.88 &4.37 &44.79 \\\\\n\t\t\t\t&\tKinD++ &{\\color{red}{4.73}}\n\t\t\t\t&681.97 &{\\color{blue}{3.99}} &{\\color{blue}{46.89}} \\\\\n\t\t\t\t&\tTBEFN &4.81 &552.91 & 4.30 &44.14 \\\\\n\t\t\t\t&\tDSLR &{\\color{blue}{4.77}} &447.98 &4.31 & 41.08\\\\\n\t\t\t\t\\hline\n\t\t\t\tUL &EnlightenGAN &{\\color{purple}{4.79}} &821.87 &{\\color{purple}{4.19}} &45.48 \\\\\n\t\t\t\t\\hline\n\t\t\t\tSSL&\tDRBN &5.80 &885.75 &5.54 &42.74 \\\\\t\n\t\t\t\t\\hline\n\t\t\t\t&\tExCNet &5.55 &723.56 &4.38 &46.74 \\\\\t\t\t\n\t\t\t\tZSL &\tZero-DCE &5.82 &307.09 & 4.76 &{\\color{purple}{46.85}} \\\\ \t\t\t\n\t\t\t\t&\tRRDNet &5.97 &{\\color{purple}{142.89}} &4.84 &45.31 \\\\ \t\t\t\t\t\t\t\n\t\t\t\t\\hline\n\t\t\\end{tabular}}\n\t\\end{center}\n\t\\label{Table:Quan_2}\n\t\\vspace{-4pt}\n\\end{table}\n\\begin{table}[t]\n\t\\caption{Quantitative comparisons on LLIV-Phone-vidT dataset in terms of average luminance variance (ALV) score. The best result is in {\\color{red}{red}} whereas the second and third best results are in {\\color{blue}{blue}} and {\\color{purple}{purple}}.}\n\t\\begin{center}\n\t\t\\begin{tabular}{c|c|c}\n\t\t\t\\hline\n\t\t\t\\multirow{2}{*}{\\textbf{Learning}}&\\multirow{2}{*}{\\textbf{Method}} & \\multicolumn{1}{c}{\\textbf{LoLi-Phone-vidT}}\\\\\n\t\t\t\\cline{3-3}\n\t\t\t\\cline{3-3}\n\t\t\t&\t& \\textbf{ALV}$\\downarrow$\\\\\n\t\t\t\\hline\n\t\t\t&\tinput &185.60 \\\\\n\t\t\t\\hline\n\t\t\t&\tLLNet &{\\color{blue}{85.72}} \\\\\n\t\t\t&\tLightenNet &643.93\\\\\n\t\t\t&\t\n\t\t\tRetinex-Net &94.05 \\\\\n\t\t\tSL&MBLLEN &113.18 \\\\\n\t\t\t&\tKinD &98.05 \\\\\n\t\t\t&\tKinD++ &115.21\\\\\n\t\t\t&\tTBEFN &{\\color{red}{58.69}} \\\\\n\t\t\t&\tDSLR &175.35 \\\\\n\t\t\t\\hline\n\t\t\tUL &EnlightenGAN &{\\color{purple}{90.69}} \\\\\n\t\t\t\\hline\n\t\t\tSSL&\tDRBN &115.04 \\\\\t\n\t\t\t\\hline\n\t\t\t&\tExCNet &1375.29 \\\\\t\t\t\n\t\t\tZSL &\tZero-DCE &117.22 \\\\ \t\t\t\n\t\t\t&\tRRDNet &147.11 \\\\ \t\t\t\t\t\t\t\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\label{Table:Quan_33}\n\t\\vspace{-4pt}\n\\end{table}\nObserving Table \\ref{Table:Quan_2}, we can find that the performance of Retinex-Net , KinD++ , and EnlightenGAN is relatively better than the other methods. Retinex-Net achieves the best PI and SPAQ scores. The scores suggest the good perceptual quality of the results enhanced by Retinex-Net . However, from Figure~\\ref{fig:Phone1}(d) and Figure~\\ref{fig:Phone2}(d), the results of Retinex-Net evidently suffer from artifacts and color deviations.\nMoreover, KinD++ attains the lowest NIQE score while the original input achieves the lowest LOE score. For the de-facto standard LOE metric, we question if the lightness order can effectively reflect the enhancement performance. Overall, the non-reference IQA metrics experience biases on the evaluations of the quality of enhanced low-light images in some cases. \nTo prepare videos in the LLIV-vidT testing set, we first discard videos without obvious objects in consecutive frames. A total of 10 videos are chosen. For each video, we select one object that appears in all frames. We then use a tracker to track the object in consecutive frames of the input video and ensure the same object appears in the bounding boxes. We discard the frames with inaccurate object tracking. The coordinates of the bounding box in each frame are collected. We employ these coordinates to crop the corresponding regions in the results enhanced by different methods and compute the average luminance variance (ALV) scores of the object in the consecutive frames as: $ALV=\\frac{1}{N}\\sum\\limits_{i=1}^{N}(L_{i}-L_{\\text{avg}})^2$, where $N$ is the number of frames of a video, $L_{i}$ represents the average luminance value of the region of bounding box in the $i$th frame, and $L_{\\text{avg}}$ denotes the average luminance value of all bounding box regions in the video. A lower ALV value suggests better temporal coherence of the enhanced video. \nThe ALV values of different methods averaged over the 10 videos of the LLIV-vidT testing set are shown in Table \\ref{Table:Quan_33}. \nThe ALV values of different methods on each video can be found in the supplementary material. Besides, we follow Jiang and Zheng to plot their luminance curves in the supplementary material.\nAs shown in Table \\ref{Table:Quan_33}, \tTBEFN obtains the best temporal coherence in terms of ALV value whereas LLNet and EnlightenGAN rank the second and third best, respectively. In contrast, the ALV value of ExCNet , as the worst performer, reaches 1375.29. This is because the performance of the zero-reference learning-based ExCNet is unstable for the enhancement of consecutive frames. ExCNet can effectively improve the brightness of some frames while it does not work well on other frames.\n\\begin{table*}[t]\n\t\\caption{Quantitative comparisons of computational complexity in terms of runtime (in second), number of trainable parameters (\\#Parameters) (in M), and FLOPs (in G). The best result is in {\\color{red}{red}} whereas the second and third best results are in {\\color{blue}{blue}} and {\\color{purple}{purple}} under each case, respectively. `-' indicates the result is not available.}\n\t\\vspace{-6pt}\n\t\\begin{center}\n\t\t\\begin{tabular}{c|c|c|c|c|c}\n\t\t\t\\hline\n\t\t\t\\textbf{Learning}&\t\\textbf{Method} \n\t\t\t& \\textbf{RunTime}$\\downarrow$ & \\textbf{\\#Parameters} $\\downarrow$ & \\textbf{FLOPs}$\\downarrow$ & \\textbf{Platform}\\\\\n\t\t\t\\hline\n\t\t\t&\tLLNet &36.270 &17.908 &4124.177 & Theano\\\\\n\t\t\t&\tLightenNet & -&{\\color{red}{0.030}} &{\\color{red}{30.540}} &MATLAB\\\\\n\t\t\t&\tRetinex-Net &0.120 &0.555 &587.470 &TensorFlow\\\\\n\t\t\tSL&\tMBLLEN &13.995&{\\color{purple}{0.450}}&301.120 &TensorFlow\\\\\n\t\t\t&\tKinD &0.148 &8.160&574.954 &TensorFlow\\\\\n\t\t\t&\tKinD++ &1.068&8.275&12238.026 &TensorFlow\\\\\n\t\t\t&\tTBEFN &{\\color{purple}{0.050}}&0.486&108.532 &TensorFlow\\\\\n\t\t\t&\tDSLR &0.074&14.931&{\\color{purple}{96.683}} &PyTorch\\\\\n\t\t\t\\hline\n\t\t\tUL&\tEnlightenGAN &{\\color{blue}{0.008}}&8.637& 273.240 &PyTorch\\\\\n\t\t\t\\hline\n\t\t\tSSL\t&\tDRBN &0.878&0.577&196.359 &PyTorch\\\\\t\n\t\t\t\\hline\n\t\t\t&\tExCNet &23.280&8.274&- &PyTorch\\\\\t\t\t\n\t\t\tZSL\t&\tZero-DCE &{\\color{red}{0.003}}&{\\color{blue}{0.079}}&{\\color{blue}{84.990}} &PyTorch \\\\\t\t\t\t\n\t\t\t&\tRRDNet &167.260 &0.128&- &PyTorch\\\\\t\t\t\t\t\t\t\t\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\label{Table:Quan_4}\n\t\\vspace{-4pt}\n\\end{table*}","id":"67f70829-d734-4fe0-baa6-b7f64f6f5b82","level":"subsection","origin_cites_number":6,"parent_id":"62b1377f-b6e1-4398-be9f-a1fa2952056b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Benchmarking and Empirical Analysis"],["subsection","Benchmarking Results"]],"subsections":[],"title":"Benchmarking Results"},{"cite_extract_rate":1,"cites":[1875],"content":"In Table \\ref{Table:Quan_4}, we compare the computational complexity of RGB format-based methods, including runtime, trainable parameters, and FLOPs averaged over 32 images of size 1200$\\times$900$\\times$3 using an NVIDIA 1080Ti GPU. We omit LightenNet for fair comparisons because only the CPU version of its code is publicly available.\nBesides, we do not report the FLOPs of ExCNet and RRDNet as the number depends on the input images (different inputs require different numbers of iterations).\nAs presented in Table \\ref{Table:Quan_4}, Zero-DCE has the shortest runtime because it only estimates several curve parameters via a lightweight network. As a result, its number of trainable parameters and FLOPs are much fewer. Moreover, the number of trainable parameters and FLOPs of LightenNet are the least among the compared methods. This is because LightenNet estimates the illumination map of input image via a tiny network of four convolutional layers. In contrast, the FLOPs of LLNet and KinD++ are extremely large, reaching 4124.177G and 12238.026G, respectively. The runtime of SSL-based ExCNet and RRDNet is long due to the time-consuming optimization process.\n\\begin{figure}[!t]\n\t\\centering \\centerline{\\includegraphics[width=1\\linewidth]{figures/face.png}}\n\t\\vspace{-2pt}\n\t\\caption{The P-R curves of face detection in the dark.}\n\t\\label{fig:face}\n\t\\vspace{-4pt}\n\\end{figure}\n\\begin{table}[!t]\n\t\\caption{Quantitative comparisons of AP under different IoU thresholds of face detection in the dark. The best result is in {\\color{red}{red}} whereas the second and third best results are in {\\color{blue}{blue}} and {\\color{purple}{purple}} under each case, respectively.}\n\t\\begin{center}\n\t\t\\begin{tabular}{c|c|c|c|c}\n\t\t\t\\hline\n\t\t\t\\multirow{2}{*}{\\textbf{Learning}}&\\multirow{2}{*}{\\textbf{Method}}&\\multicolumn{3}{c}{\\textbf{IoU thresholds}} \\\\\n\t\t\t\\cline{3-5}\n\t\t\t\\cline{3-5}\n\t\t\t&\t & \\textbf{0.5} & \\textbf{0.6} & \\textbf{0.7}\\\\\n\t\t\t\\hline\n\t\t\t&input &0.195 &0.061 &0.007\\\\\n\t\t\t\\hline\n\t\t\t&\tLLNet &0.208 &0.063&0.006\\\\\n\t\t\t&\tLightenNet &0.249 &0.085&{\\color{purple}{0.010}}\\\\\n\t\t\t&\tRetinex-Net &{\\color{blue}{0.261}} &{\\color{red}{0.101}}&{\\color{red}{0.013}}\\\\\n\t\t\tSL&\tMBLLEN &0.249 &{\\color{purple}{0.092}}&{\\color{purple}{0.010}}\\\\\n\t\t\t&\tKinD &0.235 &0.081&{\\color{purple}{0.010}}\\\\\n\t\t\t&\tKinD++ &0.251 &0.090&{\\color{blue}{0.011}}\\\\\n\t\t\t&\tTBEFN &{\\color{red}{0.268}} &{\\color{blue}{0.099}}&{\\color{blue}{0.011}}\\\\\n\t\t\t&\tDSLR &0.223 &0.067&0.007\\\\\n\t\t\t\\hline\n\t\t\tUL&\tEnlightenGAN &0.246 &0.088&{\\color{blue}{0.011}}\\\\\n\t\t\t\\hline\n\t\t\tSSL\t&\tDRBN &0.199 &0.061&0.007\\\\\t\n\t\t\t\\hline\n\t\t\t&\tExCNet &0.256 &{\\color{purple}{0.092}}&{\\color{purple}{0.010}}\\\\\t\t\t\n\t\t\tZSL\t&\tZero-DCE &{\\color{purple}{0.259}} &{\\color{purple}{0.092}}&{\\color{blue}{0.011}}\\\\\t\t\t\t\n\t\t\t&\tRRDNet &0.248 &0.083&{\\color{purple}{0.010}}\\\\\t\t\t\t\t\t\t\t\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\label{Table:Quan_8}\n\t\\vspace{-4pt}\n\\end{table}\n\\subsection {Application-Based Evaluation}\nWe investigate the performance of low-light image enhancement methods on face detection in the dark. \nFollowing the setting presented in Guo et al. , we use the DARK FACE dataset that is composed of images with faces taken in the dark. Since the bounding boxes of the test set are not publicly available, we perform the evaluation on 500 images randomly sampled from the training and validation sets.\nThe Dual Shot Face Detector (DSFD) trained on WIDER FACE dataset is used as the face detector. We feed the results of different LLIE methods to the DSFD and depict the precision-recall (P-R) curves under 0.5 IoU threshold in Figure \\ref{fig:face}. We compare the average precision (AP) under different IoU thresholds using the evaluation tool\\footnote{\\url{https://github.com/Ir1d/DARKFACE_eval_tools}} provided in DARK FACE dataset in Table \\ref{Table:Quan_8}. \nAs shown in Figure \\ref{fig:face}, all the deep learning-based solutions improve the performance of face detection in the dark, suggesting the effectiveness of deep learning-based LLIE solutions for face detection in the dark. As shown in Table \\ref{Table:Quan_8}, the AP scores of best performers under different IoU thresholds range from 0.268 to 0.013 and the AP scores of input under different IoU thresholds are very low. The results suggest that there is still room for improvement. It is noteworthy that Retinex-Net , Zero-DCE , and TBEFN achieve relatively robust performance on face detection in the dark. We show the visual results of different methods in Figure \\ref{fig:face_visual}. Although Retinex-Net performs better than other methods on the AP score, its visual result contains obvious artifacts and unnatural textures. In general, Zero-DCE obtains a good balance between the AP score and the perceptual quality for face detection in the dark. \tNote that the results of face detection in the dark are related to not only the enhanced results but also the face detector including the detector model and the training data of the detector. Here, we only take the pre-trained DSFD as an example to validate the low-light image enhancement performance of different methods to some extent.","id":"e142566f-157a-4609-a670-8df6dbfe9874","level":"subsection","origin_cites_number":3,"parent_id":"62b1377f-b6e1-4398-be9f-a1fa2952056b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Benchmarking and Empirical Analysis"],["subsection","Computational Complexity"]],"subsections":[],"title":"Computational Complexity"},{"cite_extract_rate":1,"cites":[1875],"content":"From the experimental results, we obtain several interesting observations and insights:\n\\begin{figure*}[!t]\n\t\\begin{center}\n\t\t\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }}\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_input.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_LightenNet.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_RetinexNet.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_MBLLEN.png}\\\\\n\t\t\t(a) input & (b) LightenNet & (c) Retinex-Net & (d) \tMBLLEN \\\\\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_KinD++.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_TBEFN.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_DSLR.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_EnlightenGAN.png}\\\\\n\t\t\t(e) KinD++ &\t(f) TBEFN & (g) \tDSLR \t & (h) EnlightenGAN \\\\\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_DRBN.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_ExCNet.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_ZeroDCE.png}&\n\t\t\t\\includegraphics[width=0.24\\linewidth]{figures/face_RRDNet.png}\\\\\n\t\t\t(i) DRBN & (j)\tExCNet & (k) Zero-DCE & (l) RRDNet \t \\\\\n\t\t\\end{tabular}\n\t\\end{center}\n\t\\vspace{-2pt}\n\t\\caption{Visual results of different methods on a low-light image sampled from DARK FACE dataset. Better see with zoom in for the bounding boxes of faces.}\n\t\\label{fig:face_visual}\n\t\\vspace{-4pt}\n\\end{figure*}\n1) The performance of different methods significantly varies based on the test datasets and evaluation metrics. In terms of the full-reference IQA metrics on commonly used test datasets, MBLLEN , KinD++ , and DSLR are generally better than other compared methods. For real-world low-light images taken by mobile phones, supervised learning-based Retinex-Net and KinD++ obtain better scores measured in the non-reference IQA metrics. For real-world low-light videos taken by mobile phones, TBEFN preserves the temporal coherence better. When coming to the computational efficiency, LightenNet and Zero-DCE are outstanding. From the aspect of face detection in the dark, TBEFN , Retinex-Net , and Zero-DCE rank the first three. No method always wins. Overall, Retinex-Net , , Zero-DCE , and DSLR are better choice in most cases. For raw data, EEMEFN obtains relatively better qualitative and quantitative performance than SID . However, from the visual results, EEMEFN and cannot recover the color well when compared with the corresponding ground truth.\n2) LLIV-Phone dataset fails most methods. The generalization capability of existing methods needs further improvements. It is worth noting that it is inadequate to use only the average luminance variance to evaluate the performance of different methods for low-light video enhancement. More effective and comprehensive assessment metrics would guide the development of low-light video enhancement towards the right track.\n3) Regarding learning strategies, supervised learning achieves better performance in most cases, but requires high computational resources and paired training data. In comparison, zero-shot learning is more appealing in practical applications because it does not require paired or unpaired training data. Consequently, zero-shot learning-based methods enjoy better generalization capability. However, their quantitative performance is inferior to other methods. \n4) There is a gap between visual results and quantitative IQA scores. In other words, a good visual appearance does not always yield a good IQA score. The relationships between human perception and IQA scores are worth more investigation. Pursuing better visual perception or quantitative scores depends on specific applications. For instance, to show the results to observers, more attention should be paid to visual perception. In contrast, accuracy is more important when LLIE methods are applied to face detection in the dark. Thus, more comprehensive comparisons should be performed when comparing different methods.\n5) Deep learning-based LLIE methods are beneficial to face detection in the dark. Such results further support the significance of enhancing low-light images and videos. However, in comparison to the high accuracy of face detection in normal-light images, the accuracy of face detection in the dark is extremely low, despite using LLIE methods.\n6) In comparison to RGB format-based LLIE methods, raw format-based LLIE methods usually recover details better, obtain more vivid color, and reduce noises and artifacts more effectively. This is because raw data contain more information such as wider color gamut and higher dynamic range. However, raw format-based LLIE methods are limited to specific sensors and formats such as the Bayer pattern of the Sony camera and the APS-C X-Trans pattern of the Fuji camera. In contrast, RGB format-based LLIE methods are more convenient and versatile since RGB images are commonly found as the final imagery form produced by mobile devices. However, RGB format-based LLIE methods cannot cope well with cases that exhibit low light and excessive noise.","id":"8ea7b52c-e5f5-46c9-8ff5-69763784bc99","level":"subsection","origin_cites_number":2,"parent_id":"62b1377f-b6e1-4398-be9f-a1fa2952056b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Benchmarking and Empirical Analysis"],["subsection","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":1,"cites":[1875],"content":"\\label{sec:Issue}\nIn this section, we summarize the open issues in low-light image and video enhancement as follows.\n\\noindent\n\\textbf{Generalization Capability.} Although existing methods can produce some visually pleasing results, they have limited generalization capability. For example, a method trained on MIT-Adobe FiveK dataset cannot effectively enhance the low-light images of LOL dataset . Albeit synthetic data are used to augment the diversity of training data, the models trained on the combination of real and synthetic data cannot solve this issue well. Improving the generalization capability of LLIE methods is an unsolved open issue.\n\\noindent\n\\textbf{Removing Unknown Noises.} Observing the results of existing methods on the low-light images captured by different types of phones' cameras, we can find that these methods cannot remove the noises well and even amplify the noises, especially when the types of noises are unknown. Despite some methods add Gaussian and/or Poisson noises in their training data, the noise types are different from real noises, thus the performance of these methods is unsatisfactory in real scenarios. Removing unknown noises is still unsolved.\n\\noindent\n\\textbf{Removing Unknown Artifacts.} One may enhance a low-light image downloaded from the Internet. The image may have gone through a serial of degradations such as JPEG compression or editing. Thus, the image may contain unknown artifacts. Suppressing unknown artifacts still challenges existing low-light image and video enhancement methods.\n\\noindent\n\\textbf{Correcting Uneven Illumination.} Images taken in real scenes usually exhibit uneven illumination. For example, an image captured at night has both dark regions and normal-light or over-exposed regions such as the regions of light sources. Existing methods tend to brighten both the dark regions and the light source regions, affecting the visual quality of the enhanced result. It is expected to enhance dark regions but suppress over-exposed regions. However, this open issue is not solved well in existing LLIE methods.\n\\noindent\n\\textbf{Distinguishing Semantic Regions.} Existing methods tend to enhance a low-light image without considering the semantic information of its different regions. For example, the black hair of a man in a low-light image is enhanced to be off-white as the black hair is treated as the low-light regions. An ideal enhancement method is expected to only enhance the low-light regions induced by external environments. How to distinguish semantic regions is an open issue. \n\\noindent\n\\textbf{Using Neighbouring Frames.} \n\tDespite some methods that have been proposed to enhance low-light videos, they commonly process a video frame-by-frame. How to make full use of the neighboring frames to improve the enhancement performance and speed up the processing speed is an unsolved open issue. For example, the well-lit regions of neighboring frames are used to enhance the current frame. For another example, the estimated parameters for processing neighboring frames can be reused to enhance the current frame for reducing the time of parameter estimation.","id":"c43fed76-06de-4d00-89a5-701cdf5d52e3","level":"section","origin_cites_number":1,"parent_id":"43aba542-7f6b-493f-9beb-9ae4164ca84b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Open Issues"]],"subsections":[],"title":"Open Issues"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Future}\nLow-light enhancement is a challenging research topic. As can be observed from the experiments presented in Section~\\ref{sec:evaluation} and the unsolved open issues in Section~\\ref{sec:Issue}, there is still room for improvement. We suggest potential future research directions as follows.\n\\noindent\n\\textbf{Effective Learning Strategies.}\nAs aforementioned, current LLIE models mainly adopt supervised learning that requires massive paired training data and may overfit on a specific dataset. Although some researchers attempted to introduce unsupervised learning into LLIE, the inherent relationships between LLIE and these learning strategies are not clear and their effectiveness in LLIE needs further improvements. Zero-shot learning has shown robust performance for real scenes while not requiring paired training data. The unique advantage suggests zero-shot learning as a potential research direction, especially on the formulation of zero-reference losses, deep priors, and optimization strategies. \n\\noindent\n\\textbf{Specialized Network Structures.}\nA network structure can significantly affect enhancement performance. As previously analyzed, most LLIE deep models employ U-Net or U-Net-like structures. Though they have achieved promising performance in some cases, the investigation if such an encoder-decoder network structure is most suitable for the LLIE task is still lacking.\nSome network structures require a high memory footprint and long inference time due to their\nlarge parameter space. Such network structures are unacceptable for practical applications. \nThus, it is worthwhile to investigate a more effective network structure for LLIE, considering the characteristics of low-light images such as non-uniform illumination, small pixel values, noise suppression, and color constancy. \nOne can also design more efficient network structures via taking into account the local similarity of low-light images or considering more efficient operations such as depthwise separable convolution layer and self-calibrated convolution . Neural architecture search (NAS) technique may be considered to obtain more effective and efficient LLIE network structures. Adapting the transformer architecture into LLIE may be a potential and interesting research direction.\n\\noindent\n\\textbf{Loss Functions.}\nLoss functions constrain the relationships between an input image and ground truth and drive the optimization of deep networks. In LLIE, the commonly used loss functions are borrowed from related vision tasks. \nThus, designing loss functions that are more well-suited for LLIE is desired. Recent studies have shown the possibility of using deep neural networks to approximate human visual perception of image quality . These ideas and fundamental theories could be used to guide the designs of loss functions for low-light enhancement networks. \n\\noindent\n\\textbf{Realistic Training Data.}\nAlthough there are several training datasets for LLIE, their authenticity, scales, and diversities fall behind real low-light conditions. Thus, as shown in Section \\ref{sec:evaluation}, current LLIE deep models cannot achieve satisfactory performance when encountering low-light images captured in real-world scenes.\nMore efforts are needed to study the collection of large-scale and diverse real-world paired LLIE training datasets or to generate more realistic synthetic data. \n\\noindent\n\\textbf{Standard Test Data.}\nCurrently, there is no well-accepted LLIE evaluation benchmark. Researchers prefer selecting their test data that may bias to their proposed methods. Despite some researchers leave some paired data as test data, the division of training and test partitions are mostly ad-hoc across the literature.\nConsequently, conducting a fair comparison among different methods is often laborious if not impossible. Besides, some test data are either easy to be handled or not originally collected for low-light enhancement. \nIt is desired to have a standard low-light image and video test dataset, which includes a large number of test samples with the corresponding ground truths, covering diverse scenes and challenging illumination conditions. \n\\noindent\n\\textbf{Task-Specific Evaluation Metrics.}\nThe commonly adopted evaluation metrics in LLIE can reflect the image quality to some extent. However, how to measure how good a result is enhanced by an LLIE method still challenges current IQA metrics, especially for non-reference measurements.\nThe current IQA metrics either focus on human visual perceptual such as subjective quality or emphasize machine perceptual such as the effects on high-level visual tasks. Therefore, \nmore works are expected in this research direction to make efforts on designing more accurate and task-specific evaluation metrics for LLIE.\n\\noindent\n\\textbf{Robust Generalization Capability.}\nObserving the experimental results on real-world test data, most methods fail due to their limited generalization capability. The poor generalization is caused by several factors such as synthetic training data, small-scaled training data, ineffective network structures, or unrealistic assumptions.\nIt is important to explore ways to improve the generalization.\n\\noindent\n\\textbf{Extension to Low-Light Video Enhancement.}\nUnlike the rapid development of video enhancement in other low-level vision tasks such as video deblurring , video denoising , and video super-resolution , low-light video enhancement receives less attention.\nA direct application of existing LLIE methods to videos often leads to unsatisfactory results and flickering artifacts. \nMore efforts are needed to remove visual flickering effectively, exploit the temporal information between neighboring frames, and speed up the enhancement speed. \n\\noindent\n\\textbf{Integrating Semantic Information.}\nSemantic information is crucial for low-light enhancement. It guides the networks to distinguish different regions in the process of enhancement. \nA network without access to semantic priors can easily deviate the original color of a region, e.g., turning black hair to gray color after enhancement. Therefore, integrating semantic priors into LLIE models is a promising research direction. Similar work has been done on image super-resolution and face restoration .\n\t\\ifCLASSOPTIONcompsoc\n\t\\section*{Acknowledgments}\n\t\\else\n\t\\section*{Acknowledgment}\n\t\\fi\nThis study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also partially supported by the NTU SUG and NAP grant. Chunle Guo is sponsored by CAAI-Huawei MindSpore Open Fund.\n\t\\ifCLASSOPTIONcaptionsoff\n\t\\newpage\n\t\\fi\n\t{\n\t\t\\bibliographystyle{IEEEtran}\n\t\t\\bibliography{bibliography}\n\t}\n\t\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/lichongyi.jpg}}]{Chongyi Li} is a Research Assistant Professor with the School of Computer Science and Engineering, Nanyang Technological University, Singapore. He received the Ph.D. degree from Tianjin University, China in 2018. From 2016 to 2017, he was a joint-training Ph.D. Student with Australian National University, Australia. Prior to joining NTU, he was a postdoctoral fellow with City University of Hong Kong and Nanyang Technological University from 2018 to 2021. His current research focuses on image processing, computer vision, and deep learning, particularly in the domains of image restoration and enhancement. He serves as an associate editor of the Journal of Signal, Image and Video Processing and a lead guest editor of the IEEE Journal of Oceanic Engineering.\n\t\\end{IEEEbiography}\n\t\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/guochunle.jpg}}]{Chunle Guo} received his PhD degree from Tianjin University in China under the supervision of Prof. Jichang Guo. He conducted the Ph.D. research as a Visiting Student with the School of Electronic Engineering and Computer Science, Queen Mary University of London (QMUL), UK. He continued his research as a Research Associate with the Department of Computer Science, City University of Hong Kong (CityU), from 2018 to 2019. Now he is a postdoc research fellow working with Prof. Ming-Ming Cheng at Nankai University. His research interests lie in image processing, computer vision, and deep learning.\n\t\\end{IEEEbiography}\n\t\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/linghaohan.jpg}}]{Linhao Han} is currently a master student at the College of Computer Science, Nankai\n\t\tUniversity, under the supervision of Prof. Ming-Ming Cheng. His research interests include\n\t\tdeep learning and computer vision.\n\t\\end{IEEEbiography}\n\t\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/jiangjun.jpg}}]{Jun Jiang} received the PhD degree in Color Science from Rochester Institute of Technology in 2013. He is a Senior Researcher in SenseBrain focusing on algorithm development to improve image quality on smartphone cameras. His research interest includes computational photography, low-level computer vision, and deep learning.\n\t\\end{IEEEbiography}\n\t\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/chengmingming.jpg}}]{Ming-Ming Cheng} (Senior Member, IEEE) received the Ph.D. degree from Tsinghua University in 2012. Then he did two years research fellowship with Prof. Philip Torr at Oxford. He is currently a Professor at Nankai University and leading the Media Computing Laboratory. His research interests include computer graphics, computer vision, and image processing. He received research awards, including the ACM China Rising Star Award, the IBM Global SUR Award, and the CCF-Intel Young Faculty Researcher Program. He is on the Editorial Board Member of IEEE Transactions on Image Processing (TIP).\n\t\\end{IEEEbiography}\n\t\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/jinweigu.jpg}}]{Jinwei Gu} (Senior Member, IEEE) is the R\\&D Executive Director of SenseTime USA. His current research focuses on low-level computer vision, computational photography, smart visual sensing and perception, and robotics. He obtained his Ph.D. degree in 2010 from Columbia University, and his B.S and M.S. from Tsinghua University, in 2002 and 2005 respectively. Before joining SenseTime, he was a senior research scientist in NVIDIA Research from 2015 to 2018. Prior to that, he was an assistant professor in Rochester Institute of Technology from 2010 to 2013, and a senior researcher in the media lab of Futurewei Technologies from 2013 to 2015. He is an associate editor for IEEE Transactions on Computational Imaging and an IEEE senior member since 2018.\n\t\\end{IEEEbiography}\n\t\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/cavan.jpg}}]{Chen Change Loy} (Senior Member, IEEE) is an Associate Professor with the School of Computer Science and Engineering, Nanyang Technological University, Singapore. He is also an Adjunct Associate Professor at The Chinese University of Hong Kong. He received his Ph.D. (2010) in Computer Science from the Queen Mary University of London. Prior to joining NTU, he served as a Research Assistant Professor at the MMLab of The Chinese University of Hong Kong, from 2013 to 2018. He was a postdoctoral researcher at Queen Mary University of London and Vision Semantics Limited, from 2010 to 2013. He serves as an Associate Editor of the IEEE Transactions on Pattern Analysis and Machine Intelligence and International Journal of Computer Vision. He also serves/served as an Area Chair of major conferences such as ICCV, CVPR, ECCV and AAAI. His research interests include image/video restoration and enhancement, generative tasks, and representation learning. \n\t\\end{IEEEbiography}\n\\end{document}","id":"0f160045-16e9-4579-89b0-811073538246","level":"section","origin_cites_number":0,"parent_id":"43aba542-7f6b-493f-9beb-9ae4164ca84b","prefix_titles":[["title","Low-Light Image and Video Enhancement \\\\Using Deep Learning: A Survey"],["section","Future Research Directions"]],"subsections":[],"title":"Future Research Directions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"43aba542-7f6b-493f-9beb-9ae4164ca84b\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ]\n ],\n \"subsections\": [\n \"c6721c4f-f568-49fa-a376-8b26124c3f6f\",\n \"8eda68ea-596d-47b0-8c59-3c11fbc66273\",\n \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"62b1377f-b6e1-4398-be9f-a1fa2952056b\",\n \"c43fed76-06de-4d00-89a5-701cdf5d52e3\",\n \"0f160045-16e9-4579-89b0-811073538246\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"\\\\label{sec:Introduction}}\\n\\\\IEEEPARstart{I}{mages} are often taken under sub-optimal lighting conditions, under the influence of backlit, uneven light, and dim light, due to inevitable environmental and/or technical constraints such as insufficient illumination and limited exposure time. Such images suffer from the compromised aesthetic quality and unsatisfactory transmission of information for high-level tasks such as object tracking, recognition, and detection. Figure \\\\ref{fig:example} shows some examples of the degradations induced by sub-optimal lighting conditions. \\nLow-light enhancement enjoys a wide range of applications in different areas, including visual surveillance, autonomous driving, and computational photography. In particular, smartphone photography has become ubiquitous and prominent. Limited by the size of the camera aperture, the requirement of real-time processing, and the constraint of memory, taking photographs with a smartphone's camera in a dim environment is especially challenging. There is an exciting research arena of enhancing low-light images and videos in such applications.\\n\\\\begin{figure}[!t]\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c@{ }c@{ }c@{ }}\\n\\t\\t\\t\\\\includegraphics[width=0.3\\\\linewidth,height=0.2\\\\linewidth]{figures/Fig1_backlit.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.3\\\\linewidth,height=0.2\\\\linewidth]{figures/Fig1_nonuniform.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.3\\\\linewidth,height=0.2\\\\linewidth]{figures/Fig1_weak.png}\\\\\\\\\\n\\t\\t\\t(a) back lit & (b) uneven light &(c) dim light\\\\\\\\\\n\\t\\t\\t\\\\includegraphics[width=0.3\\\\linewidth,height=0.2\\\\linewidth]{figures/Fig1_dark.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.3\\\\linewidth,height=0.2\\\\linewidth]{figures/0027.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.3\\\\linewidth,height=0.2\\\\linewidth]{figures/Fig1_noise.png}\\\\\\\\\\n\\t\\t\\t(d) extremely low &\\t(e) colored light & (f) boosted noise\\\\\\\\\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\vspace{-3pt}\\n\\t\\\\caption{Examples of images taken under sub-optimal lighting conditions. These images suffer from the buried scene content, reduced contrast, boosted noise, and inaccurate color. }\\n\\t\\\\label{fig:example}\\n\\t\\\\vspace{-5pt}\\n\\\\end{figure}\\n\\\\begin{figure*}[thb]\\n\\t\\\\centering \\\\centerline{\\\\includegraphics[width=1\\\\linewidth]{figures/chronology.png}}\\n\\t\\\\vspace{-2pt}\\n\\t\\\\caption{A concise milestone of deep learning-based low-light image and video enhancement methods. \\\\textbf{Supervised learning-based methods}: LLNet , Chen et al. , MBLLEN , Retinex-Net , LightenNet , SCIE , DeepUPE , Chen et al. , Jiang and Zheng , Wang et al. , KinD , Ren et al. , Xu et al. , Fan et al. , Lv et al. , EEMEFN~, SIDGAN. , LPNet , DLN , TBEFN , DSLR , Zhang et al. , PRIEN , and Retinex-Net . \\\\textbf{Reinforcement learning-based method}: DeepExposure . \\\\textbf{Unsupervised learning-based method}: EnlightenGAN . \\\\textbf{Zero-shot learning-based methods}: ExCNet , Zero-DCE , RRDNet , Zero-DCE++ , RetinexDIP , and RUAS . \\\\textbf{Semi-supervised learning-based method}: DRBN and DRBN .}\\n\\t\\\\label{fig:milestones}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure*}\\nTraditional methods for low-light enhancement include Histogram Equalization-based methods and Retinex model-based methods . The latter received relatively more attention. A typical Retinex model-based approach decomposes a low-light image into a reflection component and an illumination component by priors or regularizations. The estimated reflection component is treated as the enhanced result.\\nSuch methods have some limitations:\\n\\\\textbf{1)} the ideal assumption that treats the reflection component as the enhanced result does not always hold, especially given various illumination properties, which could lead to unrealistic enhancement such as loss of details and distorted colors, \\\\textbf{2)} the noise is usually ignored in the Retinex model, thus it is remained or amplified in the enhanced results, \\\\textbf{ 3)} finding an effective prior or regularization is challenging. Inaccurate prior or regularization may result in artifacts and color deviations in the enhanced results, and \\\\textbf{4)} the runtime is relatively long because of their complicated optimization process.\\nRecent years have witnessed the compelling success of deep learning-based LLIE since the first seminal work . Deep learning-based solutions enjoy better accuracy, robustness, and speed over conventional methods, thus attracting increasing attention. A concise milestone of deep learning-based LLIE methods is shown in Figure \\\\ref{fig:milestones}. As shown, \\nsince 2017, the number of deep learning-based solutions has grown year by year. Learning strategies used in these solutions cover Supervised Learning (SL), Reinforcement Learning (RL), Unsupervised Learning (UL), Zero-Shot Learning (ZSL), and Semi-Supervised Learning (SSL). Note that we only report some representative methods in Figure \\\\ref{fig:milestones}. In fact, there are more than 100 papers on deep learning-based methods from 2017 to 2021. Moreover, although some general photo enhancement methods can improve the brightness of images to some extent, we omit them in this survey as they are not designed to handle diverse low-light conditions. We concentrate on deep learning-based solutions that are specially developed for low-light image and video enhancement.\\n\\tDespite deep learning has dominated the research of LLIE, an in-depth and comprehensive survey on deep learning-based solutions is lacking. There are two reviews of LLIE . Wang et al. mainly reviews conventional LLIE methods while our work systematically and comprehensively reviews recent advances of deep learning-based LLIE. In comparison to Liu et al. that reviews existing LLIE algorithms, measures the machine vision performance of different methods, provides a low-light image dataset serving both low-level and high-level vision enhancement, and develops an enhanced face detector, our survey reviews the low-light image and video enhancement from different aspects and has the following unique characteristics.\\n\\t\\\\textbf{1)} Our work mainly focuses on recent advances of deep learning-based low-light image and video enhancement, where we provide in-depth analysis and discussion in various aspects, covering learning strategies, network structures, loss functions, training datasets, test datasets, evaluation metrics, model sizes, inference speed, enhancement performance, etc. Thus, this survey centers on deep learning and its applications in low-light image and video enhancement.\\n\\t\\\\textbf{2)} We propose a dataset that contains images and videos captured by different mobile phones' cameras under diverse illumination conditions to evaluate the generalization of existing methods. This new and challenging dataset is a supplement of existing low-light image and video enhancement datasets as such a dataset is lacking in this research area. Besides, we are the first, to the best of our knowledge, to compare the performance of deep learning-based low-light image enhancement methods on this kind of data. \\n\\t\\\\textbf{3)} We provide an online platform that covers many popular deep learning-based low-light image enhancement methods, where the results can be produced by a user-friendly web interface. With our platform, one without any GPUs can assess the results of different methods for any input images online, which speeds up the development of this research field and helps to create new research.\\n\\tWe hope that our survey could provide novel insights and inspiration to facilitate the understanding of deep learning-based LLIE, foster research on the raised open issues, and speed up the development of this research field.\",\n \"id\": \"c6721c4f-f568-49fa-a376-8b26124c3f6f\",\n \"level\": \"section\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"43aba542-7f6b-493f-9beb-9ae4164ca84b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Solution}\",\n \"id\": \"8eda68ea-596d-47b0-8c59-3c11fbc66273\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"43aba542-7f6b-493f-9beb-9ae4164ca84b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning-Based LLIE\"\n ]\n ],\n \"subsections\": [\n \"b406f65d-5d41-4498-9ddb-95caec62c1be\",\n \"44126fdb-481e-4904-b2e2-6e15d767dcdd\"\n ],\n \"title\": \"Deep Learning-Based LLIE\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"We first give a common formulation of the deep learning-based LLIE problem. For a low-light image $I\\\\in \\\\mathbb{R}^{W\\\\times H\\\\times3}$ of width $W$ and height $H$, the process can be modeled as:\\n\\\\begin{equation}\\n\\t\\\\label{equ_1}\\n\\t\\\\widehat{R}=\\\\mathcal{F}(I;\\\\theta),\\n\\\\end{equation}\\nwhere $\\\\widehat{R}\\\\in \\\\mathbb{R}^{W\\\\times H\\\\times3}$ is the enhanced result and $\\\\mathcal{F}$ represents the network with trainable parameters $\\\\theta$. The purpose of deep learning is to find optimal network parameters $\\\\widehat{\\\\theta}$ that minimizes the error:\\n\\\\begin{equation}\\n\\t\\\\label{equ_2}\\n\\t\\\\widehat{\\\\theta}=\\\\argmin_{\\\\theta}\\\\mathcal{L}(\\\\widehat{R},R),\\n\\\\end{equation}\\nwhere $R\\\\in \\\\mathbb{R}^{W\\\\times H\\\\times3}$ is the ground truth, and the loss function $\\\\mathcal{L}(\\\\widehat{R},R)$ drives the optimization of network.\\nVarious loss functions such as supervised loss and unsupervised loss can be used. More details will be presented in Section \\\\ref{sec:Technical}.\",\n \"id\": \"b406f65d-5d41-4498-9ddb-95caec62c1be\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"8eda68ea-596d-47b0-8c59-3c11fbc66273\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning-Based LLIE\"\n ],\n [\n \"subsection\",\n \"Problem Definition\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Problem Definition\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"According to different learning strategies, we categorize existing LLIE methods into supervised learning, reinforcement learning, unsupervised learning, zero-shot learning, and semi-supervised learning. \\nA statistic analysis from different perspectives is presented in Figure \\\\ref{fig:statistic}. In what follows, we review some representative methods of each strategy. \\n\\\\noindent\\n\\\\textbf{Supervised Learning.} \\nFor supervised learning-based LLIE methods, they are further divided into end-to-end, deep Retinex-based, and realistic data-driven methods.\\nThe first deep learning-based LLIE method LLNet~ employs a variant of stacked-sparse\\ndenoising autoencoder to brighten and denoise low-light images simultaneously. \\nThis pioneering work inspires the usage of end-to-end networks in LLIE. \\nLv et al. propose an end-to-end multi-branch enhancement network (MBLLEN). The MBLLEN improves the performance of LLIE via extracting effective feature representations by a feature extraction module, an enhancement module, and a fusion module. \\nThe same authors propose other three subnetworks including an Illumination-Net, a Fusion-Net, and a Restoration-Net to further improve the performance. \\nRen et al. design a more complex end-to-end network that comprises an encoder-decoder network for image content enhancement and a recurrent neural network for image edge enhancement. \\nSimilar to Ren et al. , Zhu et al. propose a method called EEMEFN. The EEMEFN consists of two stages: multi-exposure fusion and edge enhancement. \\nA multi-exposure fusion network, TBEFN~, is proposed for LLIE. The TBEFN estimates a transfer function in two branches, of which two enhancement results can be obtained. At last, a simple average scheme is employed to fuse these two images and further refine the result via a refinement unit. \\nIn addition, pyramid network (LPNet) , residual network , and Laplacian pyramid (DSLR)\\nare introduced into LLIE. These methods learn to effectively and efficiently integrate feature representations via commonly used end-to-end network structures for LLIE.\\nBased on the observation that noise exhibits different levels of contrast in different frequency layers, Xu et al. proposed a frequency-based decomposition-and-enhancement network. This network recovers image contents with noise suppression in the low-frequency layer while inferring the details in the high-frequency layer. \\nRecently, a progressive-recursive low-light image enhancement network is proposed, which uses a recursive unit to gradually enhance the input image.\\nTo solve temporal instability when handling low-light videos, Zhang et al. propose to learn and infer motion field from a single image then enforce temporal consistency.\\nIn comparison to directly learning an enhanced result in an end-to-end network, deep Retinex-based methods enjoy better enhancement performance in most cases owing to the physically explicable Retinex theory . Deep Retinex-based methods usually separately enhance the illuminance component and the reflectance components via specialized subnetworks.\\nA Retinex-Net is proposed, which includes a Decom-Net that splits the input image into light-independent reflectance and structure-aware smooth illumination and an Enhance-Net that adjusts the illumination map for low-light enhancement. Recently, the Retinex-Net is extended by adding new constraints and advanced network designs for better enhancement performance .\\nTo reduce the computational burden, Li et al. propose a lightweight LightenNet for weakly illuminated image enhancement, which only consists of four layers. \\nThe LightenNet takes a weakly illuminated image as the input and then estimates its illumination map. Based on the Retinex theory , the enhanced image is obtained by dividing the input image by the illumination map.\\nTo accurately estimate the illumination map, Wang et al. extract the global and local features to learn an image-to-illumination mapping by their proposed DeepUPE network.\\nZhang et al. separately develop three subnetworks for layer decomposition, reflectance restoration, and illumination adjustment, called KinD. \\nFurthermore, the authors alleviate the visual defects left in the results of KinD by a multi-scale illumination attention module. The improved KinD is called KinD++ .\\nTo solve the issue that the noise is omitted in the deep Retinex-based methods,\\nWang et al. propose a progressive Retinex network, where an IM-Net estimates the illumination and a NM-Net estimates the noise level. These two subnetworks work in a progressive mechanism until obtaining stable results. \\nFan et al. integrate semantic segmentation and Retinex model for further improving the enhancement performance in real cases. The core idea is to use semantic prior to guide the enhancement of both the illumination component and the reflectance component. \\n\\tAlthough some methods can achieve decent performance, they show poor generalization capability in real low-light cases due to the usage of synthetic training data. To solve this issue, some works attempt to generate more realistic training data or capture real data.\\n\\tCai et al. build a multi-exposure image dataset, where the low-contrast images of different exposure levels have their corresponding high-quality reference images. \\n\\tEach high-quality reference image is obtained by subjectively selecting the best output from 13 results enhanced by different methods. Moreover, a frequency decomposition network is trained on the built dataset and separately enhances the high-frequency layer and the low-frequency layer via a two-stage structure. \\n\\tChen et al. collect a real low-light image dataset (SID) and \\n\\ttrain the U-Net to learn a mapping from low-light raw data to the corresponding long-exposure high-quality reference image.\\n\\tFurther, Chen et al. extend the SID dataset to low-light videos (DRV). The DRV contains static videos with the corresponding long-exposure ground truths. To ensure the generalization capability of processing the videos of dynamic scenes, a siamese network is proposed. \\n\\tTo enhance the moving objects in the dark, Jiang and Zheng~ design a co-axis optical system to capture temporally synchronized and spatially aligned low-light and well-lighted video pairs (SMOID). \\n\\tUnlike the DRV video dataset , the SMOID video dataset contains dynamic scenes. To learn the mapping from raw low-light video to well-lighted video, a 3D U-Net-based network is proposed.\\n\\tConsidering the limitations of previous low-light video datasets such as DRV dataset only containing statistic videos and SMOID dataset only having 179 video pairs, \\n\\tTriantafyllidou et al.~ propose a low-light video synthesis pipeline, dubbed SIDGAN. The SIDGAN can produce dynamic video data (raw-to-RGB) by a semi-supervised dual CycleGAN with intermediate domain mapping. \\n\\tTo train this pipeline, the real-world videos are collected from Vimeo-90K dataset . The low-light raw video data and the corresponding long-exposure images are sampled from DRV dataset . \\n\\tWith the synthesized training data, this work adopts the same U-Net network as Chen et al. for low-light video enhancement. \\n\\t\\\\noindent\\n\\t\\\\textbf{Reinforcement Learning.}\\n\\tWithout paired training data, Yu et al. learn to expose photos with reinforcement adversarial learning, named DeepExposure. Specifically, an input image is first segmented into sub-images according to exposures. For each sub-image, local exposure is learned by the policy network sequentially based on reinforcement learning. The reward evaluation function is approximated by adversarial learning. At last, each local exposure is employed to retouch the input, thus obtaining multiple retouched images under different exposures. The final result is achieved by fusing these images. \\n\\t\\\\noindent\\n\\t\\\\textbf{Unsupervised Learning.}\\n\\tTraining a deep model on paired data may result in overfitting and limited generalization capability. To solve this issue, an unsupervised learning method named EnligthenGAN is proposed.\\n\\tThe EnlightenGAN adopts an attention-guided U-Net as the generator and uses the global-local discriminators to ensure the enhanced results look like realistic normal-light images. \\n\\tIn addition to global and local adversarial losses, the global and local self feature preserving losses are proposed to preserve the image content before and after the enhancement. This is a key point for the stable training of such a one-path Generative Adversarial Network (GAN) structure. \\n\\t\\\\noindent\\n\\t\\\\textbf{Zero-Shot Learning.} \\n\\tThe supervised learning, reinforcement learning, and unsupervised learning methods either have limited generalization capability or suffer from unstable training. To remedy these issues, zero-shot learning is proposed to learn the enhancement solely from the testing images. \\n\\tNote that the concept of zero-shot learning in the low-level vision tasks is used to emphasize that the method does not require paired or unpaired training data, which is different from its definition in high-level visual tasks. \\n\\tZhang et al. propose a zero-shot learning method, called ExCNet, for back-lit image restoration. \\n\\tA network is first used to estimate the S-curve that best fits the input image. Once the S-curve is estimated, the input image is separated into a base layer and a detail layer using the guided filter . Then the base layer is adjusted by the estimated S-curve. Finally, the Weber contrast is used to fuse the detailed layer and the adjusted base layer. To train the ExCNet, the authors formulate the loss function as a block-based energy minimization problem. \\n\\tZhu et al. propose a three-branch CNN, called RRDNet, for underexposed images restoration. The RRDNet decomposes an input image into illumination, reflectance, and noise via iteratively minimizing specially designed loss functions. \\n\\tTo drive the zero-shot learning, a combination of Retinex reconstruction loss, texture enhancement loss, and illumination-guided noise estimation loss is proposed. \\n\\tZhao et al. perform Retinex decomposition via neural networks and then enhance the low-light image based on the Retinex model, called RetinexDIP. Inspired by Deep Image Prior (DIP) , RetinexDIP generates the reflectance component and illumination component of an input image by randomly sampled white noise, in which the component characteristics-related losses such as illumination smoothness are used for training.\\n\\tLiu et al. propose a Retinex-inspired unrolling method for LLIE, in which the cooperative architecture search is used to discover lightweight prior architectures of basic blocks and non-reference losses are used to train the network.\\n\\tDifferent from the image reconstruction-based methods , \\n\\ta deep curve estimation network, Zero-DCE , is proposed. Zero-DCE formulates the light enhancement as a task of image-specific curve estimation, which takes a low-light image as input and produces high-order curves as its output. These curves are used for pixel-wise adjustment on the dynamic range of the input to obtain an enhanced image. Further, an accelerated and lightweight version is proposed, called Zero-DCE++ . \\n\\tSuch curve-based methods do not require any paired or unpaired data during training. They achieve zero-reference learning via a set of non-reference loss functions. Besides, unlike the image reconstruction-based methods that need high computational resources, the image-to-curve mapping only requires lightweight networks, thus achieving a fast inference speed. \\n\\t\\\\begin{figure*}[!t]\\n\\t\\t\\\\begin{center}\\n\\t\\t\\t\\\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }}\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture1.png}&\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture2.png}&\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture3.png}&\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture4.png}\\\\\\\\\\n\\t\\t\\t\\t(a) learning strategy & (b) network structure& (c) Retinex model & (d) data format \\\\\\\\\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture5.png}&\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture6.png}&\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture7.png}&\\n\\t\\t\\t\\t\\\\includegraphics[width=.23\\\\textwidth]{figures/Picture8.png}\\\\\\\\\\n\\t\\t\\t\\t(e) loss function & (f) training dataset & (g) testing dataset & (h) evaluation metric \\\\\\\\\\n\\t\\t\\t\\\\end{tabular}\\n\\t\\t\\\\end{center}\\n\\t\\t\\\\vspace{-2pt}\\n\\t\\t\\\\caption{A statictic analysis of deep learning-based LLIE methods, including learning strategy, network characteristic, Retinex model, data format, loss function, training dataset, testing dataset, and evaluation metric. Best viewed by zooming in.}\\n\\t\\t\\\\label{fig:statistic}\\n\\t\\t\\\\vspace{-4pt}\\n\\t\\\\end{figure*}\\n\\t\\\\noindent\\n\\t\\\\textbf{Semi-Supervised Learning.}\\n\\tTo combine the strengths of supervised learning and unsupervised learning, semi-supervised learning has been proposed in recent years.\\n\\tYang et al. propose a semi-supervised deep recursive band network (DRBN). The DRBN first recovers a linear band representation of an enhanced image under supervised learning, and then obtains an improved one by recomposing the given bands via a learnable linear transformation based on unsupervised adversarial learning. \\n\\tThe DRBN is extended by introducing Long Short Term Memory (LSTM) networks and an image quality assessment network pre-trained on an aesthetic visual analysis dataset, which achieves better enhancement performance .\\n\\tObserving Figure \\\\ref{fig:statistic}(a), we can find that supervised learning is the mainstream among deep learning-based LLIE methods, of which the percentage reaches 73\\\\%. This is because supervised learning is relatively easy when paired training data such as LOL , SID and diverse low-/normal-light image synthesis approaches are used. \\n\\tHowever, supervised learning-based methods suffer from some challenges: \\n\\t\\\\textbf{1)} collecting a large-scale paired dataset that covers diverse real-world low-light conditions is difficult, \\n\\t\\\\textbf{2)} synthetic low-light images do not accurately represent real-world illuminance conditions such as spatially varying lighting and different levels of noise, and \\n\\t\\\\textbf{3)} training a deep model on paired data may result in limited generalization to real-world images of diverse illumination properties. \\n\\tTherefore, some methods adopt unsupervised learning, reinforcement learning, semi-supervised learning, and zero-shot learning to bypass the challenges in supervised learning. Although these methods achieve competing performance, they still suffer from some limitations: \\n\\t\\\\textbf{1)} for unsupervised learning/semi-supervised learning methods, how to implement stable training, avoid color deviations, and build the relations of cross-domain information challenges current methods,\\n\\t\\\\textbf{2)} for reinforcement learning methods, designing an effective reward mechanism and implementing efficient and stable training are intricate, and \\n\\t\\\\textbf{3)} for zero-shot learning methods, the design of non-reference losses is non-trivial when the color preservation, artifact removal, and gradient back-propagation should be taken into account. \\n\\t\\\\begin{table*}\\n\\t\\t\\\\rowcolors{1}{gray!20}{white}\\n\\t\\t\\\\centering\\n\\t\\t\\\\caption{\\n\\t\\t\\t{Summary of essential characteristics of representative deep learning-based methods. ``Retinex'' indicates whether the models are Retinex-based or not. ``simulated'' means the testing data are simulated by the same approach as the synthetic training data. ``self-selected'' stands for the real-world images selected by the authors. ``\\\\#P''represents the number of trainable parameters. ``-'' means this item is not available or not indicated in the paper.}\\n\\t\\t}\\n\\t\\t\\\\vspace{-6pt}\\n\\t\\t\\\\label{table:methods}\\n\\t\\t\\\\begin{threeparttable}\\n\\t\\t\\t\\\\resizebox{1\\\\textwidth}{!}{\\n\\t\\t\\t\\t\\\\setlength\\\\tabcolsep{2pt}\\n\\t\\t\\t\\t\\\\renewcommand\\\\arraystretch{0.98}\\n\\t\\t\\t\\t\\\\begin{tabular}{c|r||c|c|c|c|c|c|c|c|c}\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t&\\\\textbf{Method}~~~~~~~~~&\\\\textbf{Learning} &\\\\textbf{Network Structure} &\\\\textbf{Loss Function}\\n\\t\\t\\t\\t\\t&\\\\textbf{Training Data} & \\\\textbf{Testing Data} &\\\\textbf{Evaluation Metric} & \\\\textbf{Format} &\\\\textbf{Platform} &\\\\textbf{Retinex}\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\multirow{1}{*}{\\\\rotatebox{90}{\\\\textbf{2017}}}\\n\\t\\t\\t\\t\\t&LLNet~ &SL &SSDA &SRR loss\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by \\\\\\\\Gamma Correction \\\\& \\\\\\\\Gaussian Noise} &\\\\tabincell{c}{simulated \\\\\\\\ self-selected} &\\\\tabincell{c}{PSNR SSIM} &RGB &Theano &\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\multirow{1}{*}{\\\\rotatebox{90}{\\\\textbf{2018}}}\\n\\t\\t\\t\\t\\t&LightenNet~ &SL &four layers &$L_{2}$ loss\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by \\\\\\\\random illumination values} &\\\\tabincell{c}{simulated\\\\\\\\self-selected} &\\\\tabincell{c}{PSNR MAE\\\\\\\\SSIM \\\\\\\\User Study} &RGB &\\\\tabincell{c}{Caffe\\\\\\\\MATLAB}&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&Retinex-Net~ &SL &\\n\\t\\t\\t\\t\\tmulti-scale network &\\\\tabincell{c}{$L_{1}$ loss smoothness loss \\\\\\\\invariable reflectance loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{LOL\\\\\\\\simulated by \\\\\\\\adjusting histogram} &self-selected &- &RGB &TensorFlow&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&MBLLEN~ &SL &multi-branch fusion &\\\\tabincell{c}{SSIM loss region loss\\\\\\\\perceptual loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by \\\\\\\\Gamma Correction \\\\& \\\\\\\\Poisson Noise} &\\\\tabincell{c}{simulated\\\\\\\\ self-selected} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\AB VIF\\\\\\\\LOE TOMI} &RGB &\\\\tabincell{c}{TensorFlow} & \\\\\\\\\\n\\t\\t\\t\\t\\t&SCIE~ &SL &frequency decomposition &\\\\tabincell{c}{$L_{2}$ loss $L_{1}$ loss SSIM loss}\\n\\t\\t\\t\\t\\t&SCIE &SCIE &\\\\tabincell{c}{PSNR FSIM\\\\\\\\Runtime FLOPs} &RGB&\\\\tabincell{c}{Caffe\\\\\\\\MATLAB}&\\\\\\\\\\n\\t\\t\\t\\t\\t&Chen et al.~ &SL &U-Net &$L_{1}$ loss\\n\\t\\t\\t\\t\\t&SID &SID &\\\\tabincell{c}{PSNR SSIM} &raw&TensorFlow&\\\\\\\\\\n\\t\\t\\t\\t\\t&Deepexposure~ &RL &\\\\tabincell{c}{policy network\\\\\\\\GAN} &\\\\tabincell{c}{deterministic policy gradient\\\\\\\\adversarial loss}\\n\\t\\t\\t\\t\\t&MIT-Adobe FiveK&MIT-Adobe FiveK &\\\\tabincell{c}{PSNR SSIM} &raw&TensorFlow &\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\multirow{1}{*}{\\\\rotatebox{90}{\\\\textbf{2019}}}\\n\\t\\t\\t\\t\\t&Chen et al.~ &SL &siamese network &\\\\tabincell{c}{$L_{1}$ loss\\\\\\\\ self-consistency loss}\\n\\t\\t\\t\\t\\t&DRV &DRV &\\\\tabincell{c}{PSNR\\\\\\\\ SSIM\\\\\\\\ MAE} &raw &TensorFlow&\\\\\\\\\\n\\t\\t\\t\\t\\t&Jiang and Zheng~ &SL &3D U-Net &$L_{1}$ loss\\n\\t\\t\\t\\t\\t&SMOID &SMOID &\\\\tabincell{c}{PSNR SSIM MSE} &raw &TensorFlow &\\\\\\\\\\n\\t\\t\\t\\t\\t&DeepUPE~ &SL &illumination map &\\\\tabincell{c}{$L_{1}$ loss color loss\\\\\\\\smoothness loss}\\n\\t\\t\\t\\t\\t&retouched image pairs&MIT-Adobe FiveK&\\\\tabincell{c}{PSNR SSIM\\\\\\\\User Study} &RGB &TensorFlow &\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&KinD~ &SL &\\\\tabincell{c}{three subnetworks\\\\\\\\U-Net} &\\\\tabincell{c}{reflectance similarity loss\\\\\\\\illumination smoothness loss\\\\\\\\mutual consistency loss\\\\\\\\$L_{1}$ loss $L_{2}$ loss SSIM loss\\\\\\\\texture similarity loss\\\\\\\\illumination adjustment loss}\\n\\t\\t\\t\\t\\t&LOL &\\\\tabincell{c}{LOL LIME\\\\\\\\NPE MEF} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\LOE NIQE} &RGB &TensorFlow&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&Wang et al.~ &SL &\\\\tabincell{c}{two subnetworks\\\\\\\\pointwise Conv} &$L_{1}$ loss \\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by \\\\\\\\camera imaging model} &\\\\tabincell{c}{IP100 FNF38\\\\\\\\MPI LOL NPE} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\NIQE} &RGB &Caffe &\\\\checkmark \\\\\\\\\\n\\t\\t\\t\\t\\t&Ren et al.~ &SL &\\\\tabincell{c}{U-Net like network\\\\\\\\RNN dilated Conv} &\\\\tabincell{c}{$L_{2}$ loss perceptual loss\\\\\\\\adversarial loss} \\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{MIT-Adobe FiveK\\\\\\\\with Gamma correction \\\\\\\\ \\\\& Gaussion noise} &\\\\tabincell{c}{simulated \\\\\\\\self-selected DPED} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\Runtime} &RGB &Caffe&\\\\\\\\\\n\\t\\t\\t\\t\\t&EnlightenGAN~ &UL &U-Net like network &\\\\tabincell{c}{adversarial loss\\\\\\\\self feature preserving loss}\\n\\t\\t\\t\\t\\t&unpaired real images &\\\\tabincell{c}{NPE LIME\\\\\\\\MEF DICM\\\\\\\\VV BBD-100K\\\\\\\\ExDARK} &\\\\tabincell{c}{User Study NIQE\\\\\\\\Classification} &RGB &PyTorch&\\\\\\\\\\n\\t\\t\\t\\t\\t&ExCNet.~ &ZSL &\\n\\t\\t\\t\\t\\t\\\\tabincell{c}{fully connected layers} &energy minimization loss &real images\\n\\t\\t\\t\\t\\t&$IE_{ps}D$ &\\\\tabincell{c}{User Study\\\\\\\\CDIQA LOD} &RGB &PyTorch&\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\multirow{1}{*}{\\\\rotatebox{90}{\\\\textbf{2020}}}\\n\\t\\t\\t\\t\\t&Zero-DCE~ &ZSL &U-Net like network &\\\\tabincell{c}{spatial consistency loss\\\\\\\\exposure control loss\\\\\\\\color constancy loss\\\\\\\\illumination smoothness loss}\\n\\t\\t\\t\\t\\t&SICE &\\\\tabincell{c}{SICE NPE\\\\\\\\LIME MEF\\\\\\\\DICM VV\\\\\\\\ DARK FACE} &\\\\tabincell{c}{User Study PI\\\\\\\\PNSR SSIM\\\\\\\\MAE Runtime\\\\\\\\Face detection} &RGB &PyTorch &\\\\\\\\\\n\\t\\t\\t\\t\\t&DRBN~ &SSL &recursive network &\\\\tabincell{c}{SSIM loss perceptual loss\\\\\\\\adversarial loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{LOL\\\\\\\\images selected by MOS} &LOL &\\\\tabincell{c}{PSNR SSIM\\\\\\\\SSIM-GC} &RGB &PyTorch &\\\\\\\\\\n\\t\\t\\t\\t\\t&Lv et al.~&SL &U-Net like network &\\\\tabincell{c}{Huber loss\\\\\\\\SSIM loss\\\\\\\\perceptual loss\\\\\\\\illumination smoothness loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by a\\\\\\\\ retouching module} &\\\\tabincell{c}{LOL SICE \\\\\\\\DeepUPE} &\\\\tabincell{c}{User Study PSNR\\\\\\\\SSIM VIF\\\\\\\\LOE NIQE\\\\\\\\ \\\\#P Runtime\\\\\\\\Face detection} &RGB&\\\\tabincell{c}{TensorFlow}&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&Fan et al.~ &SL &\\\\tabincell{c}{four subnetworks\\\\\\\\U-Net like network\\\\\\\\feature modulation} &\\\\tabincell{c}{mutual smoothness loss\\\\\\\\reconstruction loss\\\\\\\\illumination smoothness loss\\\\\\\\cross entropy loss\\\\\\\\ consistency loss\\\\\\\\SSIM loss\\\\\\\\gradient loss\\\\\\\\ratio learning loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by \\\\\\\\illumination adjustment, \\\\\\\\slight color distortion, \\\\\\\\and noise simulation} &\\\\tabincell{c}{simulated\\\\\\\\self-selected} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\NIQE} &RGB & - &\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&Xu et al.~&SL &\\\\tabincell{c}{frequency decomposition\\\\\\\\U-Net like network} &\\\\tabincell{c}{$L_{2}$ loss\\\\\\\\perceptual loss}\\n\\t\\t\\t\\t\\t&SID in RGB&\\\\tabincell{c}{SID in RGB \\\\\\\\self-selected} &\\\\tabincell{c}{PSNR SSIM} &RGB &PyTorch&\\\\\\\\\\n\\t\\t\\t\\t\\t&EEMEFN~&SL &\\\\tabincell{c}{U-Net like network\\\\\\\\edge detection network} & \\\\tabincell{c}{$L_{1}$ loss\\\\\\\\weighted cross-entropy loss}\\n\\t\\t\\t\\t\\t&SID &SID &\\\\tabincell{c}{PSNR SSIM} &raw& \\\\tabincell{c}{TensorFlow\\\\\\\\PaddlePaddle}&\\\\\\\\\\n\\t\\t\\t\\t\\t&DLN~&SL &\\\\tabincell{c}{residual learning\\\\\\\\interactive factor\\\\\\\\back projection network} &\\\\tabincell{c}{SSIM loss\\\\\\\\total variation loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by \\\\\\\\illumination adjustment, \\\\\\\\slight color distortion,\\\\\\\\and noise simulation} &\\\\tabincell{c}{simulated \\\\\\\\LOL} &\\\\tabincell{c}{User Study PSNR\\\\\\\\SSIM NIQE} &RGB &PyTorch&\\\\\\\\\\n\\t\\t\\t\\t\\t&LPNet~&SL &\\\\tabincell{c}{pyramid network} &\\\\tabincell{c}{$L_{1}$ loss\\\\\\\\perceptual loss\\\\\\\\luminance loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{LOL SID in RGB\\\\\\\\MIT-Adobe FiveK} &\\\\tabincell{c}{LOL SID in RGB\\\\\\\\MIT-Adobe FiveK\\\\\\\\ MEF NPE DICM VV} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\NIQE \\\\#P\\\\\\\\FLOPs Runtime} &RGB &PyTorch&\\\\\\\\\\n\\t\\t\\t\\t\\t&SIDGAN~ &SL &U-Net &CycleGAN loss\\n\\t\\t\\t\\t\\t&SIDGAN &SIDGAN &\\\\tabincell{c}{PSNR SSIM\\\\\\\\TPSNR TSSIM ATWE} &raw &\\\\tabincell{c}{TensorFlow}&\\\\\\\\\\n\\t\\t\\t\\t\\t&RRDNet &ZSL&three subnetworks &\\\\tabincell{c}{retinex reconstruction loss\\\\\\\\texture enhancement loss\\\\\\\\noise estimation loss}\\n\\t\\t\\t\\t\\t&-&\\\\tabincell{c}{NPE LIME\\\\\\\\MEF DICM} &\\\\tabincell{c}{NIQE CPCQI} &RGB&PyTorch &\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&TBEFN~&SL&\\\\tabincell{c}{three stages\\\\\\\\U-Net like network} &\\\\tabincell{c}{SSIM loss\\\\\\\\perceptual loss\\\\\\\\smoothness loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{SCIE \\\\\\\\LOL} &\\\\tabincell{c}{SCIE LOL\\\\\\\\DICM MEF\\\\\\\\NPE VV} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\NIQE Runtime\\\\\\\\ \\\\#P FLOPs} &RGB &TensorFlow&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&DSLR~&SL&\\\\tabincell{c}{Laplacian pyramid\\\\\\\\U-Net like network} &\\\\tabincell{c}{$L_{2}$ loss\\\\\\\\Laplacian loss\\\\\\\\color loss}\\n\\t\\t\\t\\t\\t&MIT-Adobe FiveK &\\\\tabincell{c}{MIT-Adobe FiveK\\\\\\\\self-selected} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\NIQMC NIQE\\\\\\\\BTMQI CaHDC} &RGB& PyTorch&\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\t\\\\multirow{1}{*}{\\\\rotatebox{90}{\\\\textbf{2021}}}\\n\\t\\t\\t\\t\\t&RUAS~ &ZSL &neural architecture search &\\\\tabincell{c}{cooperative loss\\\\\\\\similar loss\\\\\\\\total variation loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{LOL\\\\\\\\MIT-Adobe FiveK} &\\\\tabincell{c}{LOL\\\\\\\\MIT-Adobe FiveK} &\\\\tabincell{c}{PSNR SSIM\\\\\\\\ Runtime \\\\#P FLOPs} &RGB &PyTorch &\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&Zhang et al.~&SL &U-Net &\\\\tabincell{c}{$L_{1}$ loss\\\\\\\\ consistency loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{simulated by illumination \\\\\\\\adjustmentand noise simulation} &\\\\tabincell{c}{simulated \\\\\\\\self-selected} &\\\\tabincell{c}{User Study PSNR\\\\\\\\SSIM AB\\\\\\\\MABD WE} &RGB&\\\\tabincell{c}{PyTorch}&\\\\\\\\\\n\\t\\t\\t\\t\\t&Zero-DCE++~ &ZSL &U-Net like network &\\\\tabincell{c}{spatial consistency loss\\\\\\\\exposure control loss\\\\\\\\color constancy loss\\\\\\\\illumination smoothness loss}\\n\\t\\t\\t\\t\\t&SICE &\\\\tabincell{c}{SICE NPE\\\\\\\\LIME MEF\\\\\\\\DICM VV\\\\\\\\ DARK FACE} &\\\\tabincell{c}{User Study PI\\\\\\\\PNSR SSIM \\\\#P\\\\\\\\MAE Runtime\\\\\\\\Face detection FLOPs} &RGB &PyTorch &\\\\\\\\\\n\\t\\t\\t\\t\\t&DRBN~ &SSL &recursive network &\\\\tabincell{c}{perceptual loss\\\\\\\\detail loss quality loss}\\n\\t\\t\\t\\t\\t&LOL &LOL &\\\\tabincell{c}{PSNR SSIM\\\\\\\\SSIM-GC} &RGB &PyTorch &\\\\\\\\\\n\\t\\t\\t\\t\\t&Retinex-Net~ &SL &\\n\\t\\t\\t\\t\\tthree subnetworks &\\\\tabincell{c}{$L_{1}$ loss $L_{2}$ loss\\\\\\\\SSIM loss\\\\\\\\total variation loss}\\n\\t\\t\\t\\t\\t&\\\\tabincell{c}{LOL\\\\\\\\simulated by adjusting histogram} &\\\\tabincell{c}{LOL simulated\\\\\\\\NPE DICM VV} &\\\\tabincell{c}{PNSR SSIM\\\\\\\\UQI OSS User Study} &RGB &PyTorch&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&RetinexDIP &ZSL & encoder-decoder networks &\\\\tabincell{c}{reconstruction loss\\\\\\\\illumination-consistency loss\\\\\\\\reflectnce loss \\\\\\\\illumination smoothness loss}&- &\\\\tabincell{c}{DICM, ExDark\\\\\\\\Fusion LIME\\\\\\\\ NASA NPE VV} &\\\\tabincell{c}{NIQE \\\\\\\\NIQMC CPCQI} &RGB &PyTorch&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\t\\t&PRIEN~ &SL &recursive network &SSIM loss&\\\\tabincell{c}{MEF LOL\\\\\\\\simulated by adjusting histogram} &\\\\tabincell{c}{LOL LIME\\\\\\\\NPE MEF VV} &\\\\tabincell{c}{PNSR SSIM\\\\\\\\LOE TMQI} &RGB &PyTorch &\\\\\\\\\\n\\t\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\\\end{tabular}\\n\\t\\t\\t}\\n\\t\\t\\\\end{threeparttable}\\n\\t\\\\end{table*}\",\n \"id\": \"44126fdb-481e-4904-b2e2-6e15d767dcdd\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"8eda68ea-596d-47b0-8c59-3c11fbc66273\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning-Based LLIE\"\n ],\n [\n \"subsection\",\n \"Learning Strategies\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Learning Strategies\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Technical}\\nIn this section, we first summarize the representative deep learning-based LLIE methods in Table \\\\ref{table:methods}, then analyze and discuss their technical characteristics.\",\n \"id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"43aba542-7f6b-493f-9beb-9ae4164ca84b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ]\n ],\n \"subsections\": [\n \"68e1f02c-799a-4eae-8da7-114a11e886b0\",\n \"6a2b6ca0-a22f-4003-8b94-e1d5297a6a5e\",\n \"ae53f5e7-73c5-4548-9f34-2ad3930fd293\",\n \"30a23ace-8ed7-4b68-a78a-02ea2fbd6679\",\n \"b98d1366-433d-4a92-8dac-6386ad39398e\",\n \"e339615c-7f2d-4a6a-95fe-00d1129c19c1\",\n \"afdfe29b-5291-42fa-b702-0e326d794db8\"\n ],\n \"title\": \"Technical Review and Discussion\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Diverse network structures and designs have been used in the existing models, spanning from the basic U-Net, pyramid network, multi-stage network to frequency decomposition network. \\nAfter analyzing Figure \\\\ref{fig:statistic}(b), it can be observed that the U-Net and U-Net-like networks are mainly adopted network structures in LLIE. \\nThis is because U-Net can effectively integrate multi-scale \\nfeatures and employ both low-level and high-level features. Such characteristics are essential for achieving satisfactory low-light enhancement.\\nNevertheless, some key issues may be ignored in the current LLIE network structures:\\n\\\\textbf{1)} after going through several convolutional layers, the gradients of an extremely low light image may vanish during the gradient back-propagation due to its small pixel values. This would degrade the enhancement performance and affect the convergence of network training, \\n\\\\textbf{2)} the skip-connections used in the U-Net-like networks might introduce noise and redundant features into the final results. How to effectively filter out the noise and integrate both low-level and high-level features should be carefully considered, and \\n\\\\textbf{3)} although some designs and components are proposed for LLIE, most of them are borrowed or modified from related low-level visual tasks. The characteristics of low-light data should be considered when designing the network structures.\",\n \"id\": \"68e1f02c-799a-4eae-8da7-114a11e886b0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ],\n [\n \"subsection\",\n \"Network Structure\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Structure\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As presented in Figure \\\\ref{fig:statistic}(c), almost 1/3 of methods combine the designs of deep networks with the Retinex theory, e.g., designing different subnetworks to estimate the components of the Retinex model and estimating the illumination map to guide the learning of networks. Despite such a combination can bridge deep learning-based and model-based methods, their respective weaknesses may be introduced into the final models: \\n\\\\textbf{1)} the ideal assumption that the reflectance is the final enhanced result used in Retinex-based LLIE methods would still affect the final results, and \\n\\\\textbf{2)} the risk of overfitting in deep networks still exists despite the use of Retinex theory. How to cream off the best and filter out the impurities should be carefully considered when researchers combine deep learning with the Retinex theory.\",\n \"id\": \"6a2b6ca0-a22f-4003-8b94-e1d5297a6a5e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ],\n [\n \"subsection\",\n \"Combination of Deep Model and Retinex Theory\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Combination of Deep Model and Retinex Theory\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As shown in Figure \\\\ref{fig:statistic}(d), RGB data format dominates most methods as it is commonly found as the final imagery form produced by smartphone cameras, Go-Pro cameras, and drone cameras. Although raw data are limited to specific sensors such as those based on Bayer patterns, the data cover wider color gamut and higher dynamic range. Hence, deep models trained on raw data usually recover clear details and high contrast, obtain vivid color, reduce the effects of noises and artifacts, and improve the brightness of extremely low-light images. In future research, a smooth transformation from raw data of different patterns to RGB format would have the potentials to combine the convenience of RGB data and the advantage of high-quality enhancement of raw data for LLIE.\",\n \"id\": \"ae53f5e7-73c5-4548-9f34-2ad3930fd293\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ],\n [\n \"subsection\",\n \"Data Format\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Format\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In Figure\\\\ref{fig:statistic}(e), the commonly adopted loss functions in LLIE models include reconstruction loss ($L_{1}$, $L_{2}$, SSIM), perceptual loss, and smoothness loss. Besides, according to different demands and formulations, color loss, exposure loss, adversarial loss, etc are also adopted. We detail representative loss functions as follows.\\n\\\\noindent\\n\\\\textbf{Reconstruction Loss.}\\n\\tDifferent reconstruction losses have their advantages and disadvantages. $L_{2}$ loss tends to penalize larger errors, but is tolerant to small errors. $L_{1}$ loss preserves colors and\\n\\tluminance well since an error is weighted equally regardless of the local structure. SSIM loss preserves the structure and texture well. Refer to this research paper for detailed analysis.\\n\\\\noindent\\n\\\\textbf{Perceptual Loss.}\\n\\tPerceptual loss , particularly the feature reconstruction loss, is proposed to constrain the results similar to the ground truth in the feature space. The loss improves the visual quality of results. It is defined as the Euclidean distance between the feature representations of an enhanced result and those of corresponding ground truth. The feature representations are typically extracted from the VGG network pre-trained on ImageNet dataset . \\n\\\\noindent\\n\\\\textbf{Smoothness Loss.}\\n\\tTo remove noise in the enhanced results or preserve the relationship of neighboring pixels, smoothness loss (TV loss) is often used to constrain the enhanced result or the estimated illumination map.\\n\\\\noindent\\n\\\\textbf{Adversarial Loss.}\\n\\tTo encourage enhanced results to be indistinguishable from reference images, adversarial learning solves a max-min optimization problem .\\n\\\\noindent\\n\\\\textbf{Exposure Loss.}\\n\\tAs one of key non-reference losses, exposure loss measures the exposure levels of enhanced results without paired or unpaired images as reference images.\\n\\tThe commonly used loss functions in LLIE networks are also employed in image reconstruction networks for image super-resolution , image denoising , image detraining , and image deblurring .\\n\\tDifferent from these versatile losses, the specially designed exposure loss for LLIE inspires the design of non-reference losses. A non-reference loss makes a model enjoying better generalization capability.\\n\\tIt is an on-going research to consider image characteristics for the design of loss functions.\",\n \"id\": \"30a23ace-8ed7-4b68-a78a-02ea2fbd6679\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ],\n [\n \"subsection\",\n \"Loss Function\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Loss Function\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"Figure \\\\ref{fig:statistic}(f) reports the usage of a variety of paired training datasets for training low-light enhancement networks. These datasets include real-world captured datasets and synthetic datasets. We list them in Table \\\\ref{table:training}. \\n\\\\noindent\\n\\\\textbf{Simulated by Gamma Correction.}\\nOwing to its nonlinearity and simplicity, \\nGamma correction is used to adjust the luminance or tristimulus values in video or still image systems. It is defined by a power-law expression:\\n\\\\begin{equation}\\n\\t\\\\label{equ_12}\\n\\tV_{\\\\text{out}}=AV_{\\\\text{in}}^{\\\\gamma},\\n\\\\end{equation}\\nwhere the input $V_{\\\\text{in}}$ and output $V_{\\\\text{out}}$ are typically in the range of [0,1]. The constant $A$ is set to 1 in the common case. The power $\\\\gamma$ controls the luminance of the output. Intuitively, the input is brightened when $\\\\gamma<$1 while the input is darkened when $\\\\gamma>$1. The input can be the three RGB channels of an image or the luminance-related channels such as $L$ channel in the CIELab color space and $Y$ channel in the YCbCr color space. After adjusting the luminance-related channel using Gamma correction, the corresponding channels in the color space are adjusted by equal proportion to avoid producing artifacts and color deviations. \\nTo simulate images taken in real-world low-light scenes, Gaussian noise, Poisson noise, or realistic noise is added to the Gamma corrected images. The low-light image synthesized using Gamma correction can be expressed as:\\n\\\\begin{equation}\\n\\t\\\\label{equ_13}\\n\\tI_{\\\\text{low}}=n(g(I_{\\\\text{in}};\\\\gamma)),\\n\\\\end{equation}\\nwhere $n$ represents the noise model, $g(I_{\\\\text{in}};\\\\gamma)$ represents the Gamma correction function with Gamma value $\\\\gamma$, $I_{\\\\text{in}}$ is a normal-light and high-quality image or luminance-related channel. Although this function produces low-light images of different lighting levels by changing the Gamma value $\\\\gamma$, \\nit tends to introduce artifacts and color deviations into the synthetic low-light images due to the nonlinear adjustment. \\n\\\\begin{table}\\n\\t\\\\centering\\n\\t\\\\setlength\\\\tabcolsep{5pt}\\n\\t\\\\caption{\\n\\t\\t{Summary of paired training datasets. `Syn' represents Synthetic.}\\n\\t}\\n\\t\\\\vspace{-6pt}\\n\\t\\\\label{table:training}\\n\\t\\\\begin{tabular}{r|c|c|c|c}\\n\\t\\t\\\\hline\\n\\t\\t\\\\textbf{Name}~~~~~~~~~~~~~~~~~~~~~&\\\\textbf{Number}& \\n\\t\\t\\\\textbf{Format} & \\\\textbf{Real/Syn}&\\\\textbf{Video} \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tGamma Correction &+$\\\\infty$ &RGB \\n\\t\\t&Syn & \\\\\\\\\\n\\t\\tRandom Illumination &+$\\\\infty$ &RGB \\n\\t\\t&Syn & \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tLOL &500 &RGB \\n\\t\\t&Real &\\\\\\\\\\n\\t\\tSCIE &4,413 &RGB \\n\\t\\t&Real &\\\\\\\\\\n\\t\\tVE-LOL-L &2,500 &RGB \\n\\t\\t&Real+Syn &\\\\\\\\\\n\\t\\tMIT-Adobe FiveK &5,000 &raw \\n\\t\\t&Real &\\\\\\\\\\n\\t\\tSID &5,094 &raw \\n\\t\\t&Real &\\\\\\\\\\n\\t\\tDRV &202 &raw \\n\\t\\t&Real &\\\\checkmark\\\\\\\\\\n\\t\\tSMOID &179 &raw \\n\\t\\t&Real &\\\\checkmark\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\\\end{tabular}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table}\\n\\\\noindent\\n\\\\textbf{Simulated by Random Illumination.}\\nAccording to the Retinex model, an image can be decomposed into a reflectance component and an illumination component. Assuming image content is independent of illumination component and local region in the illumination component have the same intensity, a low-light image can be obtained by \\n\\\\begin{equation}\\n\\t\\\\label{equ_14}\\n\\tI_{\\\\text{low}}=I_{\\\\text{in}}L,\\n\\\\end{equation}\\nwhere $L$ is a random illumination value in the range of [0,1]. Noises can be added to the synthetic image. Such a linear function avoids artifacts, but the strong assumption requires the synthesis to operate only on image patches where local regions have the same brightness. A deep model trained on such image patches may lead to sub-optimal performance due to the negligence of context information.\\n\\\\noindent\\n\\\\textbf{LOL.}\\nLOL is the first paired low-/normal-light image dataset taken in real scenes. The low-light images are collected by changing the exposure time and ISO. LOL contains 500 pairs of low-/normal-light images of size 400$\\\\times$600 saved in RGB format. \\n\\\\noindent\\n\\\\textbf{SCIE.}\\nSCIE is a multi-exposure image dataset of low-contrast and good-contrast image pairs. It includes multi-exposure sequences of 589 indoor and outdoor scenes. Each sequence has 3 to 18 low-contrast images of different exposure levels, thus containing 4,413 multi-exposure images in total. The 589 high-quality reference images are obtained by selecting from the results of 13 representative enhancement algorithms. That is many multi-exposure images have the same high-contrast reference image.\\nThe image resolutions are between 3,000$\\\\times$2,000 and \\n6,000$\\\\times$4,000. The images in SCIE are saved in RGB format.\\n\\\\noindent\\n\\\\textbf{MIT-Adobe FiveK.}\\nMIT-Adobe FiveK was collected for global tone adjustment but has been used in LLIE. This is because the input images have low light and low contrast. MIT-Adobe FiveK contains 5,000 images, each of which is retouched by 5 trained photographers towards visually pleasing renditions, akin to a postcard. The images are all in raw format. To train the networks that can handle images of RGB format, one needs to use Adobe Lightroom to pre-process the images and save them as RGB format following a dedicated pipeline\\\\footnote{\\\\url{https://github.com/nothinglo/Deep-Photo-Enhancer/issues/38\\\\#issuecomment-449786636}}.\\nThe images are commonly resized to have a long edge of 500 pixels.\\n\\\\noindent\\n\\\\textbf{SID.}\\nSID contains 5,094 raw short-exposure images, each with a corresponding long-exposure reference image. The number of distinct long-exposure reference images is 424. In other words, multiple short-exposure images correspond to the same long-exposure reference image. \\nThe images were taken using two cameras: Sony $\\\\alpha$7S II and Fujifilm X-T2 in both indoor and outdoor scenes. Thus, the images have different sensor patterns (Sony camera' Bayer sensor and Fuji camera's APS-C X-Trans sensor). \\nThe resolution is 4,240$\\\\times$2,832 for Sony and 6,000$\\\\times$4,000 for Fuji. Usually, the long-exposure images are processed by libraw (a raw image processing library) and saved in the RGB color space, and randomly cropped 512$\\\\times$512 patches for training.\\n\\\\noindent\\n\\\\textbf{VE-LOL.} VE-LOL consists of two subsets: paired VE-LOL-L that is used for training and evaluating LLIE methods and unpaired VE-LOL-H that is used for evaluating the effect of LLIE methods on face detection. Specifically, VE-LOL-L includes 2,500 paired images. Among them, 1,000 pairs are synthetic, while 1,500 pairs are real. VE-LOL-H includes 10,940 unpaired images, where human faces are manually annotated with bounding boxes. \\n\\\\noindent\\n\\\\textbf{DRV.}\\nDRV contains 202 static raw videos, each of which has a corresponding long-exposure ground truth. Each video was taken at approximately 16 to 18 frames per second in a continuous shooting mode and is with up to 110 frames. The images were taken by a Sony RX100 VI camera in both indoor and outdoor scenes, thus all in raw format of Bayer pattern. The resolution is 3,672$\\\\times$5,496. \\n\\\\noindent\\n\\\\textbf{SMOID.}\\nSMOID contains 179 pairs of videos taken by a co-axis optical system, each of which has 200 frames. Thus, SMOID includes 35,800 extremely low light raw data of Bayer pattern and their corresponding well-lightened RGB counterparts. SMOID consists of moving vehicles and pedestrians under different illumination conditions.\\nSome issues challenge the aforementioned paired training datasets: \\n\\\\textbf{1)} deep models trained on synthetic data may introduce artifacts and color deviations when processing real-world images and videos due to the gap between synthetic data and real data, \\n\\\\textbf{2)} the scale and diversity of real training data are unsatisfactory, thus some methods incorporate synthetic data to augment the training data. This may lead to sub-optimal enhancement, and \\n\\\\textbf{3)} the input images and corresponding ground truths may exist misalignment due to the effects of motion, hardware, and environment. This would affect the performance of deep networks trained using pixel-wise loss functions.\",\n \"id\": \"b98d1366-433d-4a92-8dac-6386ad39398e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ],\n [\n \"subsection\",\n \"Training Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Training Datasets\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"In addition to the testing subsets in the paired datasets , there are several testing data collected from related works or commonly used for experimental comparisons. \\nBesides, some datasets such as face detection in the dark and detection and recognition in low-light images are employed to test the effects of LLIE on high-level visual tasks. We summarize the commonly used testing datasets in Table \\\\ref{table:testing} and introduce the representative testing datasets as follows. \\n\\\\begin{table}\\n\\t\\\\centering\\n\\t\\\\caption{\\n\\t\\t{Summary of testing datasets.}\\n\\t}\\n\\t\\\\vspace{-6pt}\\n\\t\\\\label{table:testing}\\n\\t\\\\begin{tabular}{r|c|c|c|c}\\n\\t\\t\\\\hline\\n\\t\\t\\\\textbf{Name}~~~~~~~~~&\\\\textbf{Number}& \\n\\t\\t\\\\textbf{Format} & \\\\textbf{Application}&\\\\textbf{Video} \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tLIME &10 &RGB & & \\\\\\\\\\n\\t\\tNPE &84 &RGB & & \\\\\\\\\\n\\t\\tMEF &17 &RGB & &\\\\\\\\\\n\\t\\tDICM &64 &RGB & &\\\\\\\\\\n\\t\\tVV\\\\tablefootnote{\\\\url{https://sites.google.com/site/vonikakis/datasets}} &24 &RGB & &\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tBBD-100K &10,000 &RGB &\\\\checkmark &\\\\checkmark\\\\\\\\\\n\\t\\tExDARK &7,363 &RGB &\\\\checkmark &\\\\\\\\\\n\\t\\tDARK FACE &6,000 &RGB &\\\\checkmark &\\\\\\\\\\n\\t\\tVE-LOL-H &10,940 &RGB &\\\\checkmark &\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\\\end{tabular}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table}\\n\\\\noindent\\n\\\\textbf{BBD-100K.}\\nBBD-100K is the largest driving video dataset with 10,000 videos taken over 1,100-hour driving experience across many different times in the day, weather conditions, and driving scenarios, and 10 tasks annotations. The videos taken at nighttime in BBD-100K are used to validate the effects of LLIE on high-level visual tasks and the enhancement performance in real scenarios.\\n\\\\noindent\\n\\\\textbf{ExDARK.}\\nExDARK~ dataset is built for object detection and recognition in low-light images. ExDARK dataset contains 7,363 low-light images from extremely low-light environments to twilight with 12 object classes annotated with image class labels and local object bounding boxes.\\n\\\\noindent\\n\\\\textbf{DARK FACE.}\\nDARK FACE dataset contains 6,000 low-light images captured during the nighttime, each of which is labeled with bounding boxes of the human face.\\nFrom Figure \\\\ref{table:testing}(g) and Table \\\\ref{table:methods}, we can observe that one prefers using the self-collected testing data in the experiments. The main reasons lie into three-fold: \\n\\\\textbf{1)} besides the test partition of paired datasets, there is no acknowledged benchmark for evaluations, \\\\textbf{2)}\\nthe commonly used test sets suffer from some shortcomings such as small scale (some test sets contain 10 images only), repeated content and illumination properties, and unknown experimental settings, and\\\\textbf{ 3)} some of the commonly used testing data are not originally collected for evaluating LLIE. In general, current testing datasets may lead to bias and unfair comparisons.\",\n \"id\": \"e339615c-7f2d-4a6a-95fe-00d1129c19c1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ],\n [\n \"subsection\",\n \"Testing Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Testing Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Besides human perception-based subjective evaluations, image quality assessment (IQA) metrics, including both full-reference and non-reference IQA metrics, are able to evaluate image quality objectively. In addition, user study, number of trainable parameters, FLOPs, runtime, and applications also reflect the performance of LLIE models, as shown in Fig. \\\\ref{fig:statistic}(h). \\nWe will detail them as follows.\\n\\\\noindent\\n\\\\textbf{PSNR and MSE.}\\nPSNR and MSE are widely used IQA metrics. They are always non-negative, and values closer to infinite (PSNR) and zero (MSE) are better.\\nNevertheless, the pixel-wise PSNR and MSE may provide an inaccurate indication of the visual perception of image quality since they neglect the relation of neighboring pixels.\\n\\\\noindent\\n\\\\textbf{MAE.}\\nMAE represents the mean absolute error, serving as a measure of errors between paired observations. \\nThe smaller the MAE value is, the better similarity is.\\n\\\\noindent\\n\\\\textbf{SSIM.}\\nSSIM is used to measure the similarity between two images. It is a perception-based model that considers image degradation as perceived change in structural information. \\nThe value 1 is only reachable in the case of two identical sets of data, indicating perfect structural similarity. \\n\\\\noindent\\n\\\\textbf{LOE.}\\nLOE represents the lightness order error that reflects the naturalness of an enhanced image.\\nFor LOE, the smaller the LOE value is, the better the lightness order is preserved.\\n\\\\noindent\\n\\\\textbf{Application.} \\nBesides improving the visual quality, one of the purposes of image enhancement is to serve high-level visual tasks. Thus, the effects of LLIE on high-level visual applications are commonly examined to validate the performance of different methods.\\nThe current evaluation approaches used in LLIE need to be improved in several aspects: \\n\\\\textbf{1)} although the PSNR, MSE, MAE, and SSIM are classic and popular metrics, they are still far from capturing real visual perception of human,\\n\\\\textbf{2)} some metrics are not originally designed for low-light images. They are used for assessing the fidelity of image information and contrast. Using these metrics may reflect the image quality, but they are far from the real purpose of low-light enhancement,\\n\\\\textbf{3)} metrics especially designed for low-light images are lacking, except for the LOE metric. Moreover, there is no metric for evaluating low-light video enhancement, and\\n\\\\textbf{4)} a metric that can balance both the human vision and the machine perception is expected.\",\n \"id\": \"afdfe29b-5291-42fa-b702-0e326d794db8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3ad6854d-28fe-41e8-80bd-198011ce8b19\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Technical Review and Discussion\"\n ],\n [\n \"subsection\",\n \"Evaluation Metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Evaluation Metrics\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"\\\\label{sec:evaluation}\\nThis section provides empirical analysis and highlights some key challenges in deep learning-based LLIE. To facilitate the analysis, we propose a low-light image and video dataset to examine the performance of different solutions. We also develop the first online platform, where the results of LLIE models can be produced via a user-friendly web interface. In this section, we conduct extensive evaluations on several benchmarks and our proposed dataset. \\nIn the experiments, we compare 13 representative RGB format-based methods, including eight supervised learning-based methods (LLNet , LightenNet , Retinex-Net , MBLLEN , KinD , KinD++ , TBEFN , DSLR ), one unsupervised learning-based method (EnlightenGAN ), one semi-supervised learning-based method (DRBN ), and three zero-shot learning-based methods (ExCNet , Zero-DCE , RRDNet ). Besides, we also compare two raw format-based methods, including SID and EEMEFN . Note that RGB format-based methods dominate LLIE. Moreover, most raw format-based methods do not release their code. Thus, we choose two representative methods to provide empirical analysis and insights. For all compared methods, we use the publicly available code to produce their results for fair comparisons. \\n\\\\begin{table}\\n\\t\\\\centering\\n\\t\\\\caption{{Summary of LLIV-Phone dataset. LLIV-Phone dataset contains 120 videos (45,148 images) taken by 18 different mobile phones' cameras. ``\\\\#Video'' and ``\\\\#Image'' represent the number of videos and images, respectively.}}\\n\\t\\\\vspace{-6pt}\\n\\t\\\\label{table:LLIVPhone}\\n\\t\\\\begin{tabular}{r|c|c|c}\\n\\t\\t\\\\hline\\n\\t\\t\\\\textbf{Phone's Brand}&\\\\textbf{\\\\#Video}& \\n\\t\\t\\\\textbf{\\\\#Image} &\\\\textbf{Resolution}\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tiPhone 6s &4 &1,029 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tiPhone 7 &13 &6,081 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tiPhone7 Plus &2 &900 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tiPhone8 Plus &1 &489 &1280$\\\\times$720\\\\\\\\\\n\\t\\tiPhone 11 &7 &2,200 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tiPhone 11 Pro &17 &7,739 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tiPhone XS &11 &2,470 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tiPhone XR &16 &4,997 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tiPhone SE &1 &455 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tXiaomi Mi 9 &2 &1,145 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tXiaomi Mi Mix 3 &6 &2,972 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tPixel 3&4 &1,311 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tPixel 4&3 &1,923 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tOppo R17 &6 &2,126 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tVivo Nex &12 &4,097 &1280$\\\\times$720\\\\\\\\\\n\\t\\tLG M322&2 &761 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tOnePlus 5T &1 &293 &1920$\\\\times$1080\\\\\\\\\\n\\t\\tHuawei Mate 20 Pro & 12&4,160 &1920$\\\\times$1080\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\\\end{tabular}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table}\",\n \"id\": \"62b1377f-b6e1-4398-be9f-a1fa2952056b\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"43aba542-7f6b-493f-9beb-9ae4164ca84b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Benchmarking and Empirical Analysis\"\n ]\n ],\n \"subsections\": [\n \"d8e5b8c1-ee06-4aad-800d-232c4dd98922\",\n \"7ee6e01b-208a-4fef-a9a1-3cc6a973f6c4\",\n \"67f70829-d734-4fe0-baa6-b7f64f6f5b82\",\n \"e142566f-157a-4609-a670-8df6dbfe9874\",\n \"8ea7b52c-e5f5-46c9-8ff5-69763784bc99\"\n ],\n \"title\": \"Benchmarking and Empirical Analysis\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"We propose a Low-Light Image and Video dataset, called LLIV-Phone, to comprehensively and thoroughly validate the performance of LLIE methods. LLIV-Phone is the largest and most challenging real-world testing dataset of its kind. In particular, the dataset contains 120 videos (45,148 images) taken by 18 different mobile phones' cameras including iPhone 6s, iPhone 7, iPhone7 Plus, iPhone8 Plus, iPhone 11, iPhone 11 Pro, iPhone XS, iPhone XR, iPhone SE, Xiaomi Mi 9, Xiaomi Mi Mix 3, Pixel 3, Pixel 4, Oppo R17, Vivo Nex, LG M322, OnePlus 5T, Huawei Mate 20 Pro under diverse illumination conditions (e.g., weak lighting, underexposure, moonlight, twilight, dark, extremely dark, back-lit, non-uniform light, and colored light.) in both indoor and outdoor scenes. A summary of the LLIV-Phone dataset is provided in Table \\\\ref{table:LLIVPhone}. \\nWe present several samples of LLIV-Phone dataset in Figure \\\\ref{fig:dataset_samples}. \\nThe LLIV-Phone dataset is available at the project page. \\nThis challenging dataset is collected in real scenes and contains diverse low-light images and videos. Consequently, it is suitable for evaluating the generalization capability of different low-light image and video enhancement models. Notably, the dataset can be used as the training dataset for unsupervised learning and the reference dataset for synthesis methods to generate realistic low-light data. \\n\\\\begin{figure}[!t]\\n\\t\\\\centering \\\\centerline{\\\\includegraphics[width=1\\\\linewidth]{figures/dataset_samples.png}}\\n\\t\\\\vspace{-2pt}\\n\\t\\\\caption{Several images sampled from the proposed LLIV-Phone dataset. The images and videos are taken by different devices under diverse lighting conditions and scenes.}\\n\\t\\\\label{fig:dataset_samples}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure}\\n\\\\begin{figure*} [!t]\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }c@{ }c@{ }c}\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_input.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_LLNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_LightenNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_RetinexNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_MBLLEN.png}\\\\\\\\\\n\\t\\t\\t(a) input & (b) LLNet & (c) LightenNet & (d) Retinex-Net & (e) \\tMBLLEN \\\\\\\\\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_KinD.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_KinD++.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_TBEFN.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_DSLR.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_EnlightenGAN.png}\\\\\\\\\\n\\t\\t\\t(f) KinD & (g) KinD++ & (h) TBEFN & (i) \\tDSLR & (j) EnlightenGAN \\\\\\\\\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_DRBN.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_ExCNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_ZeroDCE.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_RRDNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/LOL_GT.png}\\\\\\\\\\n\\t\\t\\t(k) DRBN & (l)\\tExCNet & (m) Zero-DCE & (n) \\tRRDNet & (o) GT \\\\\\\\\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\vspace{-2pt}\\n\\t\\\\caption{Visual results of different methods on a low-light image sampled from LOL-test dataset.}\\n\\t\\\\label{fig:LOL}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure*}\\n\\\\begin{figure*}[!t]\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }c@{ }c@{ }c}\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_input.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_LLNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_LightenNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_RetinexNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_MBLLEN.png}\\\\\\\\\\n\\t\\t\\t(a) input & (b) LLNet & (c) LightenNet & (d) Retinex-Net & (e) \\tMBLLEN \\\\\\\\\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_KinD.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_KinD++.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_TBEFN.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_DSLR.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_EnlightenGAN.png}\\\\\\\\\\n\\t\\t\\t(f) KinD & (g) KinD++ & (h) TBEFN & (i) \\tDSLR & (j) EnlightenGAN \\\\\\\\\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_DRBN.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_ExCNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_ZeroDCE.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_RRDNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.18\\\\linewidth]{figures/5K_GT.png}\\\\\\\\\\n\\t\\t\\t(k) DRBN & (l)\\tExCNet & (m) Zero-DCE & (n) \\tRRDNet & (o) GT \\\\\\\\\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\vspace{-2pt}\\n\\t\\\\caption{Visual results of different methods on a low-light image sampled from MIT-Adobe FiveK-test dataset.}\\n\\t\\\\label{fig:5K}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure*}\",\n \"id\": \"d8e5b8c1-ee06-4aad-800d-232c4dd98922\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"62b1377f-b6e1-4398-be9f-a1fa2952056b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Benchmarking and Empirical Analysis\"\n ],\n [\n \"subsection\",\n \"A New Low-Light Image and Video Dataset\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"A New Low-Light Image and Video Dataset\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"Different deep models may be implemented in different platforms such as Caffe, Theano, TensorFlow, and PyTorch. As a result, different algorithms demand different configurations, GPU versions, and hardware specifications. Such requirements are prohibitive to many researchers, especially for beginners who are new to this area and may not even have GPU resources. To resolve these problems, we develop an LLIE online platform, called LLIE-Platform, which is available at \\\\url{http://mc.nankai.edu.cn/ll/}. \\nTo the date of this submission, the LLIE-Platform covers 14 popular deep learning-based LLIE methods including LLNet , LightenNet , Retinex-Net , EnlightenGAN , MBLLEN , KinD , KinD++ , TBEFN , DSLR , DRBN , ExCNet , Zero-DCE , Zero-DCE++ , and RRDNet , where the results of any input can be produced through a user-friendly web interface. We will regularly offer new methods on this platform.\\nWe wish that this LLIE-Platform could serve the\\ngrowing research community by providing users a flexible interface\\nto run existing deep learning-based LLIE methods and develop their own new LLIE methods. \\n\\\\newcommand{\\\\addImg}[1]{\\\\includegraphics[width=0.14\\\\linewidth,height=0.2\\\\linewidth]{figures/#1}}\\n\\\\begin{figure*}[!t]\\n\\t\\\\centering\\n\\t\\\\setlength\\\\tabcolsep{0.5pt}\\n\\t\\\\begin{tabular}{ccccccccc}\\n\\t\\t\\\\addImg{Phone_input1.png} &\\n\\t\\t\\\\addImg{Phone_LLNet1.png}&\\n\\t\\t\\\\addImg{Phone_LightenNet1.png}&\\n\\t\\t\\\\addImg{Phone_RetinexNet1.png}&\\n\\t\\t\\\\addImg{Phone_MBLLEN1.png}&\\n\\t\\t\\\\addImg{Phone_KinD1.png}&\\n\\t\\t\\\\addImg{Phone_KinD++1.png}&\\\\\\\\\\n\\t\\t\\t(a) & (b) & (c) & (d) & (e) & (f) & (g) \\\\\\\\\\n\\t\\t\\\\addImg{Phone_TBEFN1.png}&\\n\\t\\t\\\\addImg{Phone_DSLR1.png}&\\n\\t\\t\\\\addImg{Phone_EnlightenGAN1.png}&\\n\\t\\t\\\\addImg{Phone_DRBN1.png}&\\n\\t\\t\\\\addImg{Phone_ExCNet1.png}&\\n\\t\\t\\\\addImg{Phone_ZeroDCE1.png}&\\n\\t\\t\\\\addImg{Phone_RRDNet1.png}\\\\\\\\\\n\\t\\t\\t(h) & (i) \\t & (j) & \\t(k) & (l) & (m) & (n) \\t \\\\\\\\\\n\\t\\\\end{tabular}\\n\\t\\\\vspace{-12pt}\\n\\t\\\\caption{Visual results of different methods on a low-light image sampled from LLIV-Phone-imgT dataset. (a) input. (b) LLNet . (c) LightenNet . (d) Retinex-Net . (e) \\tMBLLEN . (f) KinD . (g) KinD++ . (h) TBEFN . (i) \\tDSLR . (j) EnlightenGAN .\\t(k) DRBN . (l)\\tExCNet . (m) Zero-DCE . (n) \\tRRDNet .}\\n\\t\\\\label{fig:Phone1}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure*}\\n\\\\begin{figure*}[!t]\\n\\t\\\\setlength\\\\tabcolsep{0.5pt}\\n\\t\\\\centering\\n\\t\\\\begin{tabular}{ccccccccc}\\n\\t\\t\\\\addImg{Phone_input2.png}&\\n\\t\\t\\\\addImg{Phone_LLNet2.png}&\\n\\t\\t\\\\addImg{Phone_LightenNet2.png}&\\n\\t\\t\\\\addImg{Phone_RetinexNet2.png}&\\n\\t\\t\\\\addImg{Phone_MBLLEN2.png}&\\n\\t\\t\\\\addImg{Phone_KinD2.png}&\\n\\t\\t\\\\addImg{Phone_KinD++2.png}&\\\\\\\\\\n\\t\\t\\t(a) & (b) & (c) & (d) & (e) \\t & (f) & (g) \\\\\\\\\\n\\t\\t\\\\addImg{Phone_TBEFN2.png}&\\n\\t\\t\\\\addImg{Phone_DSLR2.png}&\\n\\t\\t\\\\addImg{Phone_EnlightenGAN2.png}&\\n\\t\\t\\\\addImg{Phone_DRBN2.png}&\\n\\t\\t\\\\addImg{Phone_ExCNet2.png}&\\n\\t\\t\\\\addImg{Phone_ZeroDCE2.png}&\\n\\t\\t\\\\addImg{Phone_RRDNet2.png}\\\\\\\\\\n\\t\\t\\t(h) & (i) \\t & (j) & \\t(k) & (l)\\t & (m) & (n) \\t \\\\\\\\\\n\\t\\\\end{tabular}\\n\\t\\\\vspace{-12pt}\\n\\t\\\\caption{Visual results of different methods on a low-light image sampled from LLIV-Phone-imgT dataset. (a) input. (b) LLNet . (c) LightenNet . (d) Retinex-Net . (e) \\tMBLLEN . (f) KinD . (g) KinD++ . (h) TBEFN . (i) \\tDSLR . (j) EnlightenGAN .\\t(k) DRBN . (l)\\tExCNet . (m) Zero-DCE . (n) \\tRRDNet .}\\n\\t\\\\label{fig:Phone2}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure*}\\n\\\\begin{figure*}[!t]\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }c@{ }}\\n\\t\\t\\t\\\\rotatebox{90}{~~~~~~~~~Bayer~~~~~~~~~~~~APS-C X-Trans}~&\\n\\t\\t\\t\\\\includegraphics[width=0.21\\\\linewidth]{figures/raw_input.png}&~~\\n\\t\\t\\t\\\\includegraphics[width=0.21\\\\linewidth]{figures/raw_SID.png}&~~\\n\\t\\t\\t\\\\includegraphics[width=0.21\\\\linewidth]{figures/raw_EEMEFN.png}&~~\\n\\t\\t\\t\\\\includegraphics[width=0.21\\\\linewidth]{figures/raw_GT.png}\\\\\\\\\\n\\t\\t\\t&~\\t(a) inputs &~~ (b) SID &~~ (c) EEMEFN &~~ (d) GT \\\\\\\\\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\vspace{-2pt}\\n\\t\\\\caption{Visual results of different methods on two raw low-light images sampled from SID-test-Bayer and SID-test-X-Trans test datasets. The inputs are amplified for visualization.}\\n\\t\\\\label{fig:raw}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure*}\",\n \"id\": \"7ee6e01b-208a-4fef-a9a1-3cc6a973f6c4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"62b1377f-b6e1-4398-be9f-a1fa2952056b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Benchmarking and Empirical Analysis\"\n ],\n [\n \"subsection\",\n \"Online Evaluation Platform\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Online Evaluation Platform\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"To qualitatively and quantitatively evaluate different methods, in addition to the proposed LLIV-Phone dataset, we also adopt the commonly used LOL and MIT-Adobe FiveK datasets for RGB format-based methods, and SID dataset for raw format-based methods. More visual results can be found in the supplementary material. The comparative results on the real low-light videos taken by different mobile phones' cameras can be found at YouTube \\\\url{https://www.youtube.com/watch?v=Elo9TkrG5Oo&t=6s}.\\nWe select five images on average from each video of the LLIV-Phone dataset, forming an image testing dataset with a total of 600 images (denoted as LLIV-Phone-imgT). Furthermore, we randomly select one video from the videos of each phone's brand of LLIV-Phone dataset, forming a video testing dataset with a total of 18 videos (denoted as LLIV-Phone-vidT). We half the resolutions of the frames in both LLIV-Phone-imgT and LLIV-Phone-vidT because some deep learning-based methods cannot process the full resolution of test images and videos. \\nFor the LOL dataset, we adopt the original test set including 15 low-light images captured in real scenes for testing, denoted as LOL-test. For the MIT-Adobe FiveK dataset, we follow the protocol in Chen et al. to decode the images into PNG format and resize them to have a long edge of 512 pixels using Lightroom. We adopt the same testing dataset as Chen et al. , MIT-Adobe FiveK-test, including 500 images with the retouching results by expert C as the corresponding ground truths. For the SID dataset, we use the default test set used in EEMEFN for fair comparisons, denoted as SID-test (SID-test-Bayer and SID-test-X-Trans), which is a partial test set of SID . The SID-test-Bayer includes 93 images of the Bayer pattern while the SID-test-X-Trans includes 94 images of the APS-C X-Trans pattern. \\n\\\\begin{table*}[t]\\n\\t\\\\caption{Quantitative comparisons on LOL-test and MIT-Adobe FiveK-test test datasets in terms of MSE ($\\\\times10^3$), PSNR (in dB), SSIM , and LPIPS . The best result is in {\\\\color{red}{red}} whereas the second and third best results are in {\\\\color{blue}{blue}} and {\\\\color{purple}{purple}} under each case, respectively.}\\n\\t\\\\vspace{-6pt}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c|c|c|c|c|c||c|c|c|c}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multirow{2}{*}{\\\\textbf{Learning}} &\\\\multirow{2}{*}{\\\\textbf{Method}} & \\\\multicolumn{4}{c||}{\\\\textbf{LOL-test}} & \\\\multicolumn{4}{c}{\\\\textbf{MIT-Adobe FiveK-test}} \\\\\\\\\\n\\t\\t\\t\\\\cline{3-10}\\n\\t\\t\\t\\\\cline{3-10}\\n\\t\\t\\t& &\\t\\\\textbf{MSE}$\\\\downarrow$ & \\\\textbf{PSNR} $\\\\uparrow$ & \\\\textbf{SSIM}$\\\\uparrow$ & \\\\textbf{LPIPS}$\\\\downarrow$ & \\\\textbf{MSE}$\\\\downarrow$ & \\\\textbf{PSNR} $\\\\uparrow$ & \\\\textbf{SSIM}$\\\\uparrow$ & \\\\textbf{LPIPS}$\\\\downarrow$ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&input & 12.613 & 7.773 &0.181 & 0.560& {\\\\color{blue}{1.670}}& {\\\\color{blue}{17.824}} & {\\\\color{purple}{0.779}} &{\\\\color{blue}{0.148}} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&LLNet & {\\\\color{red}{1.290}} & {\\\\color{red}{17.959}} &0.713 &0.360&4.465 & 12.177 &0.645 &0.292\\\\\\\\\\n\\t\\t\\t&LightenNet & 7.614 & 10.301 & 0.402 & 0.394 & 4.127 & 13.579 & 0.744 &{\\\\color{purple}{0.166}}\\\\\\\\\\n\\t\\t\\t&Retinex-Net & 1.651 & 16.774& 0.462 & 0.474 & 4.406 & 12.310 & 0.671 &0.239\\\\\\\\\\n\\t\\t\\tSL&MBLLEN &1.444&{\\\\color{blue}{17.902}}&0.715& 0.247& {\\\\color{red}{1.296}} & {\\\\color{red}{19.781}} & {\\\\color{red}{0.825}}&{\\\\color{red}{0.108}} \\\\\\\\\\n\\t\\t\\t&KinD & {\\\\color{purple}{1.431}} & 17.648& {\\\\color{blue}{0.779}}& {\\\\color{red}{0.175}} & 2.675 & 14.535 & 0.741 &0.177 \\\\\\\\\\n\\t\\t\\t&KinD++ & {\\\\color{blue}{1.298}}&{\\\\color{purple}{17.752}}& {\\\\color{purple}{0.760}}& {\\\\color{blue}{0.198}} & 7.582 & 9.732 &0.568 & 0.336 \\\\\\\\\\n\\t\\t\\t&TBEFN &1.764&17.351& {\\\\color{red}{0.786}}& {\\\\color{purple}{0.210}} & 3.865 & 12.769 &0.704 &0.178 \\\\\\\\\\n\\t\\t\\t&DSLR &3.536&15.050& 0.597&0.337 &{\\\\color{purple}{1.925}} & {\\\\color{purple}{16.632}} &{\\\\color{blue}{0.782}} & 0.167 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tUL\\t& EnlightenGAN &1.998&17.483& 0.677&0.322 &3.628&13.260&0.745& 0.170 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSSL\\t&DRBN &2.359&15.125& 0.472&0.316 &3.314& 13.355&0.378 & 0.281 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tExCNet &2.292&15.783& 0.515&0.373 &2.927& 13.978&0.710 &0.187 \\\\\\\\\\t\\t\\t\\n\\t\\t\\tZSL\\t&Zero-DCE &3.282&14.861& 0.589&0.335 &3.476&13.199&0.709 &0.203\\\\\\\\\\t\\t\\t\\t\\n\\t\\t\\t&\\tRRDNet &6.313&11.392&0.468&0.361 &7.057&10.135&0.620 &0.303\\\\\\\\\\t\\t\\t\\t\\t\\t\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\label{Table:Quan_1}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table*}\\n\\\\begin{table*}[t]\\n\\t\\\\caption{Quantitative comparisons on SID-test test dataset in terms of MSE ($\\\\times10^3$), PSNR (in dB), SSIM , and LPIPS . The best result is in {\\\\color{red}{red}} under each case. To compute the quantitative scores of input raw data, we use the corresponding camera ISP pipelines provided by Chen et al. to transfer raw data to RGB format.}\\n\\t\\\\vspace{-6pt}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c|c|c|c|c|c||c|c|c|c}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multirow{2}{*}{\\\\textbf{Learning}} &\\\\multirow{2}{*}{\\\\textbf{Method}} & \\\\multicolumn{4}{c||}{\\\\textbf{SID-test--Bayer}} & \\\\multicolumn{4}{c}{\\\\textbf{SID-test--X-Trans}} \\\\\\\\\\n\\t\\t\\t\\\\cline{3-10}\\n\\t\\t\\t\\\\cline{3-10}\\n\\t\\t\\t& &\\t\\\\textbf{MSE}$\\\\downarrow$ & \\\\textbf{PSNR} $\\\\uparrow$ & \\\\textbf{SSIM}$\\\\uparrow$ & \\\\textbf{LPIPS}$\\\\downarrow$ & \\\\textbf{MSE}$\\\\downarrow$ & \\\\textbf{PSNR} $\\\\uparrow$ & \\\\textbf{SSIM}$\\\\uparrow$ & \\\\textbf{LPIPS}$\\\\downarrow$ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&input &5.378 &11.840 &0.063 &0.711 &4.803 &11.880 &0.075 &0.796 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSL&SID &0.140 &28.614 &0.757 &0.465 &0.235 &26.663 &0.680&0.586\\\\\\\\\\n\\t\\t\\t&EEMEFN &{\\\\color{red}{0.126}} &{\\\\color{red}{29.212}} &{\\\\color{red}{0.768}} &{\\\\color{red}{0.448}}&{\\\\color{red}{0.191}} &{\\\\color{red}{27.423}} &{\\\\color{red}{0.695}} &{\\\\color{red}{0.546}}\\\\\\\\\\t\\t\\t\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\label{Table:Quan_11}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table*}\\n\\\\noindent\\n\\\\textbf{Qualitative Comparison.}\\nWe first present the results of different methods on the images sampled from LOL-test and MIT-Adobe FiveK-test datasets in Figures \\\\ref{fig:LOL} and \\\\ref{fig:5K}. \\nAs shown in Figure \\\\ref{fig:LOL}, all methods improve the brightness and contrast of the input image. However, none of them successfully recovers the accurate color of the input image when the results are compared with the ground truth. In particular, LLNet produces blurring result. LightenNet and RRDNet produce under-exposed results while MBLLEN and ExCNet over-expose the image. KinD , KinD++ , TBEFN , DSLR , EnlightenGAN , and DRBN introduce obvious artifacts.\\nIn Figure \\\\ref{fig:5K}, LLNet , KinD++ , TBEFN , and RRDNet produce over-exposed results. Retinex-Net , KinD++ , and RRDNet yield artifacts and blurring in the results. \\nWe found that the ground truths of MIT-Adobe FiveK dataset still contain some dark regions. This is because the dataset is originally designed for global image retouching, where restoring low light regions is not the main priority in this task. \\nWe also observed that the input images in LOL dataset and MIT-Adobe FiveK dataset are relatively clean from noise, which is different from real low-light scenes. \\nAlthough some methods take the MIT-Adobe FiveK dataset as the training or testing dataset, we argue that this dataset is not appropriate for the task of LLIE due to its mismatched/unsatisfactory ground truth for LLIE.\\nTo examine the generalization capability of different methods, we conduct comparisons on the images sampled from our LLIV-Phone-imgT dataset. The visual results of different methods are shown in Figures \\\\ref{fig:Phone1} and \\\\ref{fig:Phone2}.\\nAs presented in Figure \\\\ref{fig:Phone1}, all methods cannot effectively improve the brightness and remove the noise of the input low-light image. Moreover, Retinex-Net , MBLLEN , and DRBN produce obvious artifacts.\\nIn Figure \\\\ref{fig:Phone2}, all methods enhance the brightness of this input image. However, only \\nMBLLEN and RRDNet obtain visually pleasing enhancement without color deviation, artifacts, and over-/under-exposure. Notably, for regions with a light source, none of the methods can brighten the image without amplifying the noise around these regions. \\nTaking light sources into account for LLIE would be an interesting direction to explore.\\nThe results suggest the difficulty of enhancing the images of the LLIV-Phone-imgT dataset. \\nReal low-light images fail most existing LLIE methods due to the limited generalization capability of these methods. The potential reasons are the use of synthetic training data, small-scaled training data, or unrealistic assumptions such as the local illumination consistency and treating the reflectance component as the final result in the Retinex model in these methods.\\nWe further present the visual comparisons of raw format-based methods in Figure \\\\ref{fig:raw}. As shown, the input raw data have obvious noises. Both SID and EEMEFN can effectively remove the effects of noises. In comparison to the simple U-Net structure used in SID , the more complex structure of EEMEFN obtains better brightness recovery. However, their results are far from the corresponding GT, especially for the input of APS-C X-Trans pattern.\\n\\\\noindent\\n\\\\textbf{Quantitative Comparison.}\\nFor test sets with ground truth i.e., LOL-test, MIT-Adobe FiveK-test, and SID-test, we adopt MSE, PSNR, SSIM , and LPIPS metrics to quantitatively compare different methods. LPIPS is a deep learning-based image quality assessment metric that measures the perceptual similarity between a result and its corresponding ground truth by deep visual representations. For LPIPS, we employ the AlexNet-based model to compute the perceptual similarity. A lower LPIPS\\nvalue suggests a result that is closer to the corresponding ground truth in terms of perceptual similarity. \\nIn Table \\\\ref{Table:Quan_1} and Table \\\\ref{Table:Quan_11}, we show the quantitative results of RGB format-based methods and raw format-based methods, respectively.\\nAs presented in Table \\\\ref{Table:Quan_1}, the quantitative scores of supervised learning-based methods are better than those of unsupervised learning-based, semi-supervised learning-based, and zero-shot learning-based methods on LOL-test and MIT-Adobe FiveK-test datasets. Among them, LLNet obtains the best MSE and PSNR values on the LOL-test dataset; however, its performance drops on the MIT-Adobe FiveK-test dataset. \\nThis may be caused by the bias of LLNet towards the LOL dataset since it was trained using the LOL training dataset. \\nFor the LOL-test dataset, TBEFN obtains the highest SSIM value while KinD achieves the lowest LPIPS value. There is no winner across these four evaluation metrics on the LOL-test dataset despite the fact that some methods were trained on the LOL training dataset. \\nFor the MIT-Adobe FiveK-test dataset, MBLLEN outperforms all compared methods under the four evaluation metrics in spite of being trained on synthetic training data. Nevertheless, MBLLEN still cannot obtain the best performance on both two test datasets. \\nAs presented in Table \\\\ref{Table:Quan_11}, both SID and EEMEFN improve the quality of input raw data. Compared with the quantitative scores of SID , EEMEFN achieves consistently better performance across different raw data patterns and evaluation metrics.\\nFor LLIV-Phone-imgT test set, we use the non-reference IQA metrics, i.e., NIQE , perceptual index (PI) , LOE , and SPAQ to quantitatively compare different methods. In terms of LOE, the smaller the LOE value is, the better the lightness order is preserved. For NIQE, the smaller the NIQE value is, the better the visual quality is. A lower PI value indicates better perceptual quality. SPAQ is devised for the perceptual quality assessment of smartphone photography. A larger SPAQ value suggests better perceptual quality of smartphone photography. The quantitative results are provided in Table \\\\ref{Table:Quan_2}.\\n\\\\begin{table}[ht]\\n\\t\\\\caption{Quantitative comparisons on LLIV-Phone-imgT dataset in terms of NIQE , LOE , PI , and SPAQ . The best result is in {\\\\color{red}{red}} whereas the second and third best results are in {\\\\color{blue}{blue}} and {\\\\color{purple}{purple}} under each case, respectively.}\\n\\t\\\\begin{center}\\n\\t\\t\\\\setlength{\\\\tabcolsep}{1.7mm}{\\n\\t\\t\\t\\\\begin{tabular}{c|c|c|c|c|c}\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t\\\\multirow{2}{*}{\\\\textbf{Learning}}&\\\\multirow{2}{*}{\\\\textbf{Method}} & \\\\multicolumn{4}{c}{\\\\textbf{LoLi-Phone-imgT}}\\\\\\\\\\n\\t\\t\\t\\t\\\\cline{3-6}\\n\\t\\t\\t\\t\\\\cline{3-6}\\n\\t\\t\\t\\t&\\t& \\\\textbf{NIQE}$\\\\downarrow$ & \\\\textbf{LOE} $\\\\downarrow$ & \\\\textbf{PI}$\\\\downarrow$ & \\\\textbf{SPAQ}$\\\\uparrow$ \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t&\\tinput &6.99 &{\\\\color{red}{0.00}} &5.86 &44.45 \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t&\\tLLNet &5.86 &{\\\\color{blue}{5.86}} &5.66 &40.56 \\\\\\\\\\n\\t\\t\\t\\t&\\tLightenNet &5.34 &952.33 &4.58 &45.74 \\\\\\\\\\n\\t\\t\\t\\t&\\t\\n\\t\\t\\t\\tRetinex-Net & 5.01 &790.21 &{\\\\color{red}{3.48}} &{\\\\color{red}{50.95}} \\\\\\\\\\n\\t\\t\\t\\tSL&MBLLEN &5.08 &220.63 & 4.27 &42.50 \\\\\\\\\\n\\t\\t\\t\\t&\\tKinD &4.97 &405.88 &4.37 &44.79 \\\\\\\\\\n\\t\\t\\t\\t&\\tKinD++ &{\\\\color{red}{4.73}}\\n\\t\\t\\t\\t&681.97 &{\\\\color{blue}{3.99}} &{\\\\color{blue}{46.89}} \\\\\\\\\\n\\t\\t\\t\\t&\\tTBEFN &4.81 &552.91 & 4.30 &44.14 \\\\\\\\\\n\\t\\t\\t\\t&\\tDSLR &{\\\\color{blue}{4.77}} &447.98 &4.31 & 41.08\\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\tUL &EnlightenGAN &{\\\\color{purple}{4.79}} &821.87 &{\\\\color{purple}{4.19}} &45.48 \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\tSSL&\\tDRBN &5.80 &885.75 &5.54 &42.74 \\\\\\\\\\t\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t&\\tExCNet &5.55 &723.56 &4.38 &46.74 \\\\\\\\\\t\\t\\t\\n\\t\\t\\t\\tZSL &\\tZero-DCE &5.82 &307.09 & 4.76 &{\\\\color{purple}{46.85}} \\\\\\\\ \\t\\t\\t\\n\\t\\t\\t\\t&\\tRRDNet &5.97 &{\\\\color{purple}{142.89}} &4.84 &45.31 \\\\\\\\ \\t\\t\\t\\t\\t\\t\\t\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}}\\n\\t\\\\end{center}\\n\\t\\\\label{Table:Quan_2}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table}\\n\\\\begin{table}[t]\\n\\t\\\\caption{Quantitative comparisons on LLIV-Phone-vidT dataset in terms of average luminance variance (ALV) score. The best result is in {\\\\color{red}{red}} whereas the second and third best results are in {\\\\color{blue}{blue}} and {\\\\color{purple}{purple}}.}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c|c|c}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multirow{2}{*}{\\\\textbf{Learning}}&\\\\multirow{2}{*}{\\\\textbf{Method}} & \\\\multicolumn{1}{c}{\\\\textbf{LoLi-Phone-vidT}}\\\\\\\\\\n\\t\\t\\t\\\\cline{3-3}\\n\\t\\t\\t\\\\cline{3-3}\\n\\t\\t\\t&\\t& \\\\textbf{ALV}$\\\\downarrow$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tinput &185.60 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tLLNet &{\\\\color{blue}{85.72}} \\\\\\\\\\n\\t\\t\\t&\\tLightenNet &643.93\\\\\\\\\\n\\t\\t\\t&\\t\\n\\t\\t\\tRetinex-Net &94.05 \\\\\\\\\\n\\t\\t\\tSL&MBLLEN &113.18 \\\\\\\\\\n\\t\\t\\t&\\tKinD &98.05 \\\\\\\\\\n\\t\\t\\t&\\tKinD++ &115.21\\\\\\\\\\n\\t\\t\\t&\\tTBEFN &{\\\\color{red}{58.69}} \\\\\\\\\\n\\t\\t\\t&\\tDSLR &175.35 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tUL &EnlightenGAN &{\\\\color{purple}{90.69}} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSSL&\\tDRBN &115.04 \\\\\\\\\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tExCNet &1375.29 \\\\\\\\\\t\\t\\t\\n\\t\\t\\tZSL &\\tZero-DCE &117.22 \\\\\\\\ \\t\\t\\t\\n\\t\\t\\t&\\tRRDNet &147.11 \\\\\\\\ \\t\\t\\t\\t\\t\\t\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\label{Table:Quan_33}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table}\\nObserving Table \\\\ref{Table:Quan_2}, we can find that the performance of Retinex-Net , KinD++ , and EnlightenGAN is relatively better than the other methods. Retinex-Net achieves the best PI and SPAQ scores. The scores suggest the good perceptual quality of the results enhanced by Retinex-Net . However, from Figure~\\\\ref{fig:Phone1}(d) and Figure~\\\\ref{fig:Phone2}(d), the results of Retinex-Net evidently suffer from artifacts and color deviations.\\nMoreover, KinD++ attains the lowest NIQE score while the original input achieves the lowest LOE score. For the de-facto standard LOE metric, we question if the lightness order can effectively reflect the enhancement performance. Overall, the non-reference IQA metrics experience biases on the evaluations of the quality of enhanced low-light images in some cases. \\nTo prepare videos in the LLIV-vidT testing set, we first discard videos without obvious objects in consecutive frames. A total of 10 videos are chosen. For each video, we select one object that appears in all frames. We then use a tracker to track the object in consecutive frames of the input video and ensure the same object appears in the bounding boxes. We discard the frames with inaccurate object tracking. The coordinates of the bounding box in each frame are collected. We employ these coordinates to crop the corresponding regions in the results enhanced by different methods and compute the average luminance variance (ALV) scores of the object in the consecutive frames as: $ALV=\\\\frac{1}{N}\\\\sum\\\\limits_{i=1}^{N}(L_{i}-L_{\\\\text{avg}})^2$, where $N$ is the number of frames of a video, $L_{i}$ represents the average luminance value of the region of bounding box in the $i$th frame, and $L_{\\\\text{avg}}$ denotes the average luminance value of all bounding box regions in the video. A lower ALV value suggests better temporal coherence of the enhanced video. \\nThe ALV values of different methods averaged over the 10 videos of the LLIV-vidT testing set are shown in Table \\\\ref{Table:Quan_33}. \\nThe ALV values of different methods on each video can be found in the supplementary material. Besides, we follow Jiang and Zheng to plot their luminance curves in the supplementary material.\\nAs shown in Table \\\\ref{Table:Quan_33}, \\tTBEFN obtains the best temporal coherence in terms of ALV value whereas LLNet and EnlightenGAN rank the second and third best, respectively. In contrast, the ALV value of ExCNet , as the worst performer, reaches 1375.29. This is because the performance of the zero-reference learning-based ExCNet is unstable for the enhancement of consecutive frames. ExCNet can effectively improve the brightness of some frames while it does not work well on other frames.\\n\\\\begin{table*}[t]\\n\\t\\\\caption{Quantitative comparisons of computational complexity in terms of runtime (in second), number of trainable parameters (\\\\#Parameters) (in M), and FLOPs (in G). The best result is in {\\\\color{red}{red}} whereas the second and third best results are in {\\\\color{blue}{blue}} and {\\\\color{purple}{purple}} under each case, respectively. `-' indicates the result is not available.}\\n\\t\\\\vspace{-6pt}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c|c|c|c|c|c}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\textbf{Learning}&\\t\\\\textbf{Method} \\n\\t\\t\\t& \\\\textbf{RunTime}$\\\\downarrow$ & \\\\textbf{\\\\#Parameters} $\\\\downarrow$ & \\\\textbf{FLOPs}$\\\\downarrow$ & \\\\textbf{Platform}\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tLLNet &36.270 &17.908 &4124.177 & Theano\\\\\\\\\\n\\t\\t\\t&\\tLightenNet & -&{\\\\color{red}{0.030}} &{\\\\color{red}{30.540}} &MATLAB\\\\\\\\\\n\\t\\t\\t&\\tRetinex-Net &0.120 &0.555 &587.470 &TensorFlow\\\\\\\\\\n\\t\\t\\tSL&\\tMBLLEN &13.995&{\\\\color{purple}{0.450}}&301.120 &TensorFlow\\\\\\\\\\n\\t\\t\\t&\\tKinD &0.148 &8.160&574.954 &TensorFlow\\\\\\\\\\n\\t\\t\\t&\\tKinD++ &1.068&8.275&12238.026 &TensorFlow\\\\\\\\\\n\\t\\t\\t&\\tTBEFN &{\\\\color{purple}{0.050}}&0.486&108.532 &TensorFlow\\\\\\\\\\n\\t\\t\\t&\\tDSLR &0.074&14.931&{\\\\color{purple}{96.683}} &PyTorch\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tUL&\\tEnlightenGAN &{\\\\color{blue}{0.008}}&8.637& 273.240 &PyTorch\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSSL\\t&\\tDRBN &0.878&0.577&196.359 &PyTorch\\\\\\\\\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tExCNet &23.280&8.274&- &PyTorch\\\\\\\\\\t\\t\\t\\n\\t\\t\\tZSL\\t&\\tZero-DCE &{\\\\color{red}{0.003}}&{\\\\color{blue}{0.079}}&{\\\\color{blue}{84.990}} &PyTorch \\\\\\\\\\t\\t\\t\\t\\n\\t\\t\\t&\\tRRDNet &167.260 &0.128&- &PyTorch\\\\\\\\\\t\\t\\t\\t\\t\\t\\t\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\label{Table:Quan_4}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table*}\",\n \"id\": \"67f70829-d734-4fe0-baa6-b7f64f6f5b82\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"62b1377f-b6e1-4398-be9f-a1fa2952056b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Benchmarking and Empirical Analysis\"\n ],\n [\n \"subsection\",\n \"Benchmarking Results\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Benchmarking Results\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"In Table \\\\ref{Table:Quan_4}, we compare the computational complexity of RGB format-based methods, including runtime, trainable parameters, and FLOPs averaged over 32 images of size 1200$\\\\times$900$\\\\times$3 using an NVIDIA 1080Ti GPU. We omit LightenNet for fair comparisons because only the CPU version of its code is publicly available.\\nBesides, we do not report the FLOPs of ExCNet and RRDNet as the number depends on the input images (different inputs require different numbers of iterations).\\nAs presented in Table \\\\ref{Table:Quan_4}, Zero-DCE has the shortest runtime because it only estimates several curve parameters via a lightweight network. As a result, its number of trainable parameters and FLOPs are much fewer. Moreover, the number of trainable parameters and FLOPs of LightenNet are the least among the compared methods. This is because LightenNet estimates the illumination map of input image via a tiny network of four convolutional layers. In contrast, the FLOPs of LLNet and KinD++ are extremely large, reaching 4124.177G and 12238.026G, respectively. The runtime of SSL-based ExCNet and RRDNet is long due to the time-consuming optimization process.\\n\\\\begin{figure}[!t]\\n\\t\\\\centering \\\\centerline{\\\\includegraphics[width=1\\\\linewidth]{figures/face.png}}\\n\\t\\\\vspace{-2pt}\\n\\t\\\\caption{The P-R curves of face detection in the dark.}\\n\\t\\\\label{fig:face}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure}\\n\\\\begin{table}[!t]\\n\\t\\\\caption{Quantitative comparisons of AP under different IoU thresholds of face detection in the dark. The best result is in {\\\\color{red}{red}} whereas the second and third best results are in {\\\\color{blue}{blue}} and {\\\\color{purple}{purple}} under each case, respectively.}\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c|c|c|c|c}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multirow{2}{*}{\\\\textbf{Learning}}&\\\\multirow{2}{*}{\\\\textbf{Method}}&\\\\multicolumn{3}{c}{\\\\textbf{IoU thresholds}} \\\\\\\\\\n\\t\\t\\t\\\\cline{3-5}\\n\\t\\t\\t\\\\cline{3-5}\\n\\t\\t\\t&\\t & \\\\textbf{0.5} & \\\\textbf{0.6} & \\\\textbf{0.7}\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&input &0.195 &0.061 &0.007\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tLLNet &0.208 &0.063&0.006\\\\\\\\\\n\\t\\t\\t&\\tLightenNet &0.249 &0.085&{\\\\color{purple}{0.010}}\\\\\\\\\\n\\t\\t\\t&\\tRetinex-Net &{\\\\color{blue}{0.261}} &{\\\\color{red}{0.101}}&{\\\\color{red}{0.013}}\\\\\\\\\\n\\t\\t\\tSL&\\tMBLLEN &0.249 &{\\\\color{purple}{0.092}}&{\\\\color{purple}{0.010}}\\\\\\\\\\n\\t\\t\\t&\\tKinD &0.235 &0.081&{\\\\color{purple}{0.010}}\\\\\\\\\\n\\t\\t\\t&\\tKinD++ &0.251 &0.090&{\\\\color{blue}{0.011}}\\\\\\\\\\n\\t\\t\\t&\\tTBEFN &{\\\\color{red}{0.268}} &{\\\\color{blue}{0.099}}&{\\\\color{blue}{0.011}}\\\\\\\\\\n\\t\\t\\t&\\tDSLR &0.223 &0.067&0.007\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tUL&\\tEnlightenGAN &0.246 &0.088&{\\\\color{blue}{0.011}}\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSSL\\t&\\tDRBN &0.199 &0.061&0.007\\\\\\\\\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t&\\tExCNet &0.256 &{\\\\color{purple}{0.092}}&{\\\\color{purple}{0.010}}\\\\\\\\\\t\\t\\t\\n\\t\\t\\tZSL\\t&\\tZero-DCE &{\\\\color{purple}{0.259}} &{\\\\color{purple}{0.092}}&{\\\\color{blue}{0.011}}\\\\\\\\\\t\\t\\t\\t\\n\\t\\t\\t&\\tRRDNet &0.248 &0.083&{\\\\color{purple}{0.010}}\\\\\\\\\\t\\t\\t\\t\\t\\t\\t\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\label{Table:Quan_8}\\n\\t\\\\vspace{-4pt}\\n\\\\end{table}\\n\\\\subsection {Application-Based Evaluation}\\nWe investigate the performance of low-light image enhancement methods on face detection in the dark. \\nFollowing the setting presented in Guo et al. , we use the DARK FACE dataset that is composed of images with faces taken in the dark. Since the bounding boxes of the test set are not publicly available, we perform the evaluation on 500 images randomly sampled from the training and validation sets.\\nThe Dual Shot Face Detector (DSFD) trained on WIDER FACE dataset is used as the face detector. We feed the results of different LLIE methods to the DSFD and depict the precision-recall (P-R) curves under 0.5 IoU threshold in Figure \\\\ref{fig:face}. We compare the average precision (AP) under different IoU thresholds using the evaluation tool\\\\footnote{\\\\url{https://github.com/Ir1d/DARKFACE_eval_tools}} provided in DARK FACE dataset in Table \\\\ref{Table:Quan_8}. \\nAs shown in Figure \\\\ref{fig:face}, all the deep learning-based solutions improve the performance of face detection in the dark, suggesting the effectiveness of deep learning-based LLIE solutions for face detection in the dark. As shown in Table \\\\ref{Table:Quan_8}, the AP scores of best performers under different IoU thresholds range from 0.268 to 0.013 and the AP scores of input under different IoU thresholds are very low. The results suggest that there is still room for improvement. It is noteworthy that Retinex-Net , Zero-DCE , and TBEFN achieve relatively robust performance on face detection in the dark. We show the visual results of different methods in Figure \\\\ref{fig:face_visual}. Although Retinex-Net performs better than other methods on the AP score, its visual result contains obvious artifacts and unnatural textures. In general, Zero-DCE obtains a good balance between the AP score and the perceptual quality for face detection in the dark. \\tNote that the results of face detection in the dark are related to not only the enhanced results but also the face detector including the detector model and the training data of the detector. Here, we only take the pre-trained DSFD as an example to validate the low-light image enhancement performance of different methods to some extent.\",\n \"id\": \"e142566f-157a-4609-a670-8df6dbfe9874\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"62b1377f-b6e1-4398-be9f-a1fa2952056b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Benchmarking and Empirical Analysis\"\n ],\n [\n \"subsection\",\n \"Computational Complexity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Computational Complexity\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"From the experimental results, we obtain several interesting observations and insights:\\n\\\\begin{figure*}[!t]\\n\\t\\\\begin{center}\\n\\t\\t\\\\begin{tabular}{c@{ }c@{ }c@{ }c@{ }}\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_input.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_LightenNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_RetinexNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_MBLLEN.png}\\\\\\\\\\n\\t\\t\\t(a) input & (b) LightenNet & (c) Retinex-Net & (d) \\tMBLLEN \\\\\\\\\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_KinD++.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_TBEFN.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_DSLR.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_EnlightenGAN.png}\\\\\\\\\\n\\t\\t\\t(e) KinD++ &\\t(f) TBEFN & (g) \\tDSLR \\t & (h) EnlightenGAN \\\\\\\\\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_DRBN.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_ExCNet.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_ZeroDCE.png}&\\n\\t\\t\\t\\\\includegraphics[width=0.24\\\\linewidth]{figures/face_RRDNet.png}\\\\\\\\\\n\\t\\t\\t(i) DRBN & (j)\\tExCNet & (k) Zero-DCE & (l) RRDNet \\t \\\\\\\\\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{center}\\n\\t\\\\vspace{-2pt}\\n\\t\\\\caption{Visual results of different methods on a low-light image sampled from DARK FACE dataset. Better see with zoom in for the bounding boxes of faces.}\\n\\t\\\\label{fig:face_visual}\\n\\t\\\\vspace{-4pt}\\n\\\\end{figure*}\\n1) The performance of different methods significantly varies based on the test datasets and evaluation metrics. In terms of the full-reference IQA metrics on commonly used test datasets, MBLLEN , KinD++ , and DSLR are generally better than other compared methods. For real-world low-light images taken by mobile phones, supervised learning-based Retinex-Net and KinD++ obtain better scores measured in the non-reference IQA metrics. For real-world low-light videos taken by mobile phones, TBEFN preserves the temporal coherence better. When coming to the computational efficiency, LightenNet and Zero-DCE are outstanding. From the aspect of face detection in the dark, TBEFN , Retinex-Net , and Zero-DCE rank the first three. No method always wins. Overall, Retinex-Net , , Zero-DCE , and DSLR are better choice in most cases. For raw data, EEMEFN obtains relatively better qualitative and quantitative performance than SID . However, from the visual results, EEMEFN and cannot recover the color well when compared with the corresponding ground truth.\\n2) LLIV-Phone dataset fails most methods. The generalization capability of existing methods needs further improvements. It is worth noting that it is inadequate to use only the average luminance variance to evaluate the performance of different methods for low-light video enhancement. More effective and comprehensive assessment metrics would guide the development of low-light video enhancement towards the right track.\\n3) Regarding learning strategies, supervised learning achieves better performance in most cases, but requires high computational resources and paired training data. In comparison, zero-shot learning is more appealing in practical applications because it does not require paired or unpaired training data. Consequently, zero-shot learning-based methods enjoy better generalization capability. However, their quantitative performance is inferior to other methods. \\n4) There is a gap between visual results and quantitative IQA scores. In other words, a good visual appearance does not always yield a good IQA score. The relationships between human perception and IQA scores are worth more investigation. Pursuing better visual perception or quantitative scores depends on specific applications. For instance, to show the results to observers, more attention should be paid to visual perception. In contrast, accuracy is more important when LLIE methods are applied to face detection in the dark. Thus, more comprehensive comparisons should be performed when comparing different methods.\\n5) Deep learning-based LLIE methods are beneficial to face detection in the dark. Such results further support the significance of enhancing low-light images and videos. However, in comparison to the high accuracy of face detection in normal-light images, the accuracy of face detection in the dark is extremely low, despite using LLIE methods.\\n6) In comparison to RGB format-based LLIE methods, raw format-based LLIE methods usually recover details better, obtain more vivid color, and reduce noises and artifacts more effectively. This is because raw data contain more information such as wider color gamut and higher dynamic range. However, raw format-based LLIE methods are limited to specific sensors and formats such as the Bayer pattern of the Sony camera and the APS-C X-Trans pattern of the Fuji camera. In contrast, RGB format-based LLIE methods are more convenient and versatile since RGB images are commonly found as the final imagery form produced by mobile devices. However, RGB format-based LLIE methods cannot cope well with cases that exhibit low light and excessive noise.\",\n \"id\": \"8ea7b52c-e5f5-46c9-8ff5-69763784bc99\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"62b1377f-b6e1-4398-be9f-a1fa2952056b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Benchmarking and Empirical Analysis\"\n ],\n [\n \"subsection\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1875\n ],\n \"content\": \"\\\\label{sec:Issue}\\nIn this section, we summarize the open issues in low-light image and video enhancement as follows.\\n\\\\noindent\\n\\\\textbf{Generalization Capability.} Although existing methods can produce some visually pleasing results, they have limited generalization capability. For example, a method trained on MIT-Adobe FiveK dataset cannot effectively enhance the low-light images of LOL dataset . Albeit synthetic data are used to augment the diversity of training data, the models trained on the combination of real and synthetic data cannot solve this issue well. Improving the generalization capability of LLIE methods is an unsolved open issue.\\n\\\\noindent\\n\\\\textbf{Removing Unknown Noises.} Observing the results of existing methods on the low-light images captured by different types of phones' cameras, we can find that these methods cannot remove the noises well and even amplify the noises, especially when the types of noises are unknown. Despite some methods add Gaussian and/or Poisson noises in their training data, the noise types are different from real noises, thus the performance of these methods is unsatisfactory in real scenarios. Removing unknown noises is still unsolved.\\n\\\\noindent\\n\\\\textbf{Removing Unknown Artifacts.} One may enhance a low-light image downloaded from the Internet. The image may have gone through a serial of degradations such as JPEG compression or editing. Thus, the image may contain unknown artifacts. Suppressing unknown artifacts still challenges existing low-light image and video enhancement methods.\\n\\\\noindent\\n\\\\textbf{Correcting Uneven Illumination.} Images taken in real scenes usually exhibit uneven illumination. For example, an image captured at night has both dark regions and normal-light or over-exposed regions such as the regions of light sources. Existing methods tend to brighten both the dark regions and the light source regions, affecting the visual quality of the enhanced result. It is expected to enhance dark regions but suppress over-exposed regions. However, this open issue is not solved well in existing LLIE methods.\\n\\\\noindent\\n\\\\textbf{Distinguishing Semantic Regions.} Existing methods tend to enhance a low-light image without considering the semantic information of its different regions. For example, the black hair of a man in a low-light image is enhanced to be off-white as the black hair is treated as the low-light regions. An ideal enhancement method is expected to only enhance the low-light regions induced by external environments. How to distinguish semantic regions is an open issue. \\n\\\\noindent\\n\\\\textbf{Using Neighbouring Frames.} \\n\\tDespite some methods that have been proposed to enhance low-light videos, they commonly process a video frame-by-frame. How to make full use of the neighboring frames to improve the enhancement performance and speed up the processing speed is an unsolved open issue. For example, the well-lit regions of neighboring frames are used to enhance the current frame. For another example, the estimated parameters for processing neighboring frames can be reused to enhance the current frame for reducing the time of parameter estimation.\",\n \"id\": \"c43fed76-06de-4d00-89a5-701cdf5d52e3\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"43aba542-7f6b-493f-9beb-9ae4164ca84b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Open Issues\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Open Issues\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Future}\\nLow-light enhancement is a challenging research topic. As can be observed from the experiments presented in Section~\\\\ref{sec:evaluation} and the unsolved open issues in Section~\\\\ref{sec:Issue}, there is still room for improvement. We suggest potential future research directions as follows.\\n\\\\noindent\\n\\\\textbf{Effective Learning Strategies.}\\nAs aforementioned, current LLIE models mainly adopt supervised learning that requires massive paired training data and may overfit on a specific dataset. Although some researchers attempted to introduce unsupervised learning into LLIE, the inherent relationships between LLIE and these learning strategies are not clear and their effectiveness in LLIE needs further improvements. Zero-shot learning has shown robust performance for real scenes while not requiring paired training data. The unique advantage suggests zero-shot learning as a potential research direction, especially on the formulation of zero-reference losses, deep priors, and optimization strategies. \\n\\\\noindent\\n\\\\textbf{Specialized Network Structures.}\\nA network structure can significantly affect enhancement performance. As previously analyzed, most LLIE deep models employ U-Net or U-Net-like structures. Though they have achieved promising performance in some cases, the investigation if such an encoder-decoder network structure is most suitable for the LLIE task is still lacking.\\nSome network structures require a high memory footprint and long inference time due to their\\nlarge parameter space. Such network structures are unacceptable for practical applications. \\nThus, it is worthwhile to investigate a more effective network structure for LLIE, considering the characteristics of low-light images such as non-uniform illumination, small pixel values, noise suppression, and color constancy. \\nOne can also design more efficient network structures via taking into account the local similarity of low-light images or considering more efficient operations such as depthwise separable convolution layer and self-calibrated convolution . Neural architecture search (NAS) technique may be considered to obtain more effective and efficient LLIE network structures. Adapting the transformer architecture into LLIE may be a potential and interesting research direction.\\n\\\\noindent\\n\\\\textbf{Loss Functions.}\\nLoss functions constrain the relationships between an input image and ground truth and drive the optimization of deep networks. In LLIE, the commonly used loss functions are borrowed from related vision tasks. \\nThus, designing loss functions that are more well-suited for LLIE is desired. Recent studies have shown the possibility of using deep neural networks to approximate human visual perception of image quality . These ideas and fundamental theories could be used to guide the designs of loss functions for low-light enhancement networks. \\n\\\\noindent\\n\\\\textbf{Realistic Training Data.}\\nAlthough there are several training datasets for LLIE, their authenticity, scales, and diversities fall behind real low-light conditions. Thus, as shown in Section \\\\ref{sec:evaluation}, current LLIE deep models cannot achieve satisfactory performance when encountering low-light images captured in real-world scenes.\\nMore efforts are needed to study the collection of large-scale and diverse real-world paired LLIE training datasets or to generate more realistic synthetic data. \\n\\\\noindent\\n\\\\textbf{Standard Test Data.}\\nCurrently, there is no well-accepted LLIE evaluation benchmark. Researchers prefer selecting their test data that may bias to their proposed methods. Despite some researchers leave some paired data as test data, the division of training and test partitions are mostly ad-hoc across the literature.\\nConsequently, conducting a fair comparison among different methods is often laborious if not impossible. Besides, some test data are either easy to be handled or not originally collected for low-light enhancement. \\nIt is desired to have a standard low-light image and video test dataset, which includes a large number of test samples with the corresponding ground truths, covering diverse scenes and challenging illumination conditions. \\n\\\\noindent\\n\\\\textbf{Task-Specific Evaluation Metrics.}\\nThe commonly adopted evaluation metrics in LLIE can reflect the image quality to some extent. However, how to measure how good a result is enhanced by an LLIE method still challenges current IQA metrics, especially for non-reference measurements.\\nThe current IQA metrics either focus on human visual perceptual such as subjective quality or emphasize machine perceptual such as the effects on high-level visual tasks. Therefore, \\nmore works are expected in this research direction to make efforts on designing more accurate and task-specific evaluation metrics for LLIE.\\n\\\\noindent\\n\\\\textbf{Robust Generalization Capability.}\\nObserving the experimental results on real-world test data, most methods fail due to their limited generalization capability. The poor generalization is caused by several factors such as synthetic training data, small-scaled training data, ineffective network structures, or unrealistic assumptions.\\nIt is important to explore ways to improve the generalization.\\n\\\\noindent\\n\\\\textbf{Extension to Low-Light Video Enhancement.}\\nUnlike the rapid development of video enhancement in other low-level vision tasks such as video deblurring , video denoising , and video super-resolution , low-light video enhancement receives less attention.\\nA direct application of existing LLIE methods to videos often leads to unsatisfactory results and flickering artifacts. \\nMore efforts are needed to remove visual flickering effectively, exploit the temporal information between neighboring frames, and speed up the enhancement speed. \\n\\\\noindent\\n\\\\textbf{Integrating Semantic Information.}\\nSemantic information is crucial for low-light enhancement. It guides the networks to distinguish different regions in the process of enhancement. \\nA network without access to semantic priors can easily deviate the original color of a region, e.g., turning black hair to gray color after enhancement. Therefore, integrating semantic priors into LLIE models is a promising research direction. Similar work has been done on image super-resolution and face restoration .\\n\\t\\\\ifCLASSOPTIONcompsoc\\n\\t\\\\section*{Acknowledgments}\\n\\t\\\\else\\n\\t\\\\section*{Acknowledgment}\\n\\t\\\\fi\\nThis study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also partially supported by the NTU SUG and NAP grant. Chunle Guo is sponsored by CAAI-Huawei MindSpore Open Fund.\\n\\t\\\\ifCLASSOPTIONcaptionsoff\\n\\t\\\\newpage\\n\\t\\\\fi\\n\\t{\\n\\t\\t\\\\bibliographystyle{IEEEtran}\\n\\t\\t\\\\bibliography{bibliography}\\n\\t}\\n\\t\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/lichongyi.jpg}}]{Chongyi Li} is a Research Assistant Professor with the School of Computer Science and Engineering, Nanyang Technological University, Singapore. He received the Ph.D. degree from Tianjin University, China in 2018. From 2016 to 2017, he was a joint-training Ph.D. Student with Australian National University, Australia. Prior to joining NTU, he was a postdoctoral fellow with City University of Hong Kong and Nanyang Technological University from 2018 to 2021. His current research focuses on image processing, computer vision, and deep learning, particularly in the domains of image restoration and enhancement. He serves as an associate editor of the Journal of Signal, Image and Video Processing and a lead guest editor of the IEEE Journal of Oceanic Engineering.\\n\\t\\\\end{IEEEbiography}\\n\\t\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/guochunle.jpg}}]{Chunle Guo} received his PhD degree from Tianjin University in China under the supervision of Prof. Jichang Guo. He conducted the Ph.D. research as a Visiting Student with the School of Electronic Engineering and Computer Science, Queen Mary University of London (QMUL), UK. He continued his research as a Research Associate with the Department of Computer Science, City University of Hong Kong (CityU), from 2018 to 2019. Now he is a postdoc research fellow working with Prof. Ming-Ming Cheng at Nankai University. His research interests lie in image processing, computer vision, and deep learning.\\n\\t\\\\end{IEEEbiography}\\n\\t\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/linghaohan.jpg}}]{Linhao Han} is currently a master student at the College of Computer Science, Nankai\\n\\t\\tUniversity, under the supervision of Prof. Ming-Ming Cheng. His research interests include\\n\\t\\tdeep learning and computer vision.\\n\\t\\\\end{IEEEbiography}\\n\\t\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/jiangjun.jpg}}]{Jun Jiang} received the PhD degree in Color Science from Rochester Institute of Technology in 2013. He is a Senior Researcher in SenseBrain focusing on algorithm development to improve image quality on smartphone cameras. His research interest includes computational photography, low-level computer vision, and deep learning.\\n\\t\\\\end{IEEEbiography}\\n\\t\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/chengmingming.jpg}}]{Ming-Ming Cheng} (Senior Member, IEEE) received the Ph.D. degree from Tsinghua University in 2012. Then he did two years research fellowship with Prof. Philip Torr at Oxford. He is currently a Professor at Nankai University and leading the Media Computing Laboratory. His research interests include computer graphics, computer vision, and image processing. He received research awards, including the ACM China Rising Star Award, the IBM Global SUR Award, and the CCF-Intel Young Faculty Researcher Program. He is on the Editorial Board Member of IEEE Transactions on Image Processing (TIP).\\n\\t\\\\end{IEEEbiography}\\n\\t\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/jinweigu.jpg}}]{Jinwei Gu} (Senior Member, IEEE) is the R\\\\&D Executive Director of SenseTime USA. His current research focuses on low-level computer vision, computational photography, smart visual sensing and perception, and robotics. He obtained his Ph.D. degree in 2010 from Columbia University, and his B.S and M.S. from Tsinghua University, in 2002 and 2005 respectively. Before joining SenseTime, he was a senior research scientist in NVIDIA Research from 2015 to 2018. Prior to that, he was an assistant professor in Rochester Institute of Technology from 2010 to 2013, and a senior researcher in the media lab of Futurewei Technologies from 2013 to 2015. He is an associate editor for IEEE Transactions on Computational Imaging and an IEEE senior member since 2018.\\n\\t\\\\end{IEEEbiography}\\n\\t\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{figures/cavan.jpg}}]{Chen Change Loy} (Senior Member, IEEE) is an Associate Professor with the School of Computer Science and Engineering, Nanyang Technological University, Singapore. He is also an Adjunct Associate Professor at The Chinese University of Hong Kong. He received his Ph.D. (2010) in Computer Science from the Queen Mary University of London. Prior to joining NTU, he served as a Research Assistant Professor at the MMLab of The Chinese University of Hong Kong, from 2013 to 2018. He was a postdoctoral researcher at Queen Mary University of London and Vision Semantics Limited, from 2010 to 2013. He serves as an Associate Editor of the IEEE Transactions on Pattern Analysis and Machine Intelligence and International Journal of Computer Vision. He also serves/served as an Area Chair of major conferences such as ICCV, CVPR, ECCV and AAAI. His research interests include image/video restoration and enhancement, generative tasks, and representation learning. \\n\\t\\\\end{IEEEbiography}\\n\\\\end{document}\",\n \"id\": \"0f160045-16e9-4579-89b0-811073538246\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"43aba542-7f6b-493f-9beb-9ae4164ca84b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Low-Light Image and Video Enhancement \\\\\\\\Using Deep Learning: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Future Research Directions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":82,"string":"82"},"all_cites":{"kind":"list like","value":[1875],"string":"[\n 1875\n]","isConversational":false},"Bertopic_CD":{"kind":"null"}}},{"rowIdx":48,"cells":{"authors":{"kind":"list like","value":["Jer Shyuan Ng","Wei Yang Bryan Lim","Nguyen Cong Luong","Zehui Xiong","Alia Asheralieva","Dusit Niyato","Cyril Leung","Chunyan Miao"],"string":"[\n \"Jer Shyuan Ng\",\n \"Wei Yang Bryan Lim\",\n \"Nguyen Cong Luong\",\n \"Zehui Xiong\",\n \"Alia Asheralieva\",\n \"Dusit Niyato\",\n \"Cyril Leung\",\n \"Chunyan Miao\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey of Coded Distributed Computing"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-08-20T16:02:35Z","string":"2020-08-20T16:02:35Z"},"category":{"kind":"string","value":"cs.DC"},"abstract":{"kind":"string","value":" Distributed computing has become a common approach for large-scale computation of tasks due to benefits such as high reliability, scalability, computation speed, and cost-effectiveness. However, distributed computing faces critical issues related to communication load and straggler effects. In particular, computing nodes need to exchange intermediate results with each other in order to calculate the final result, and this significantly increases communication overheads. Furthermore, a distributed computing network may include straggling nodes that run intermittently slower. This results in a longer overall time needed to execute the computation tasks, thereby limiting the performance of distributed computing. To address these issues, coded distributed computing (CDC), i.e., a combination of coding theoretic techniques and distributed computing, has been recently proposed as a promising solution. Coding theoretic techniques have proved effective in WiFi and cellular systems to deal with channel noise. Therefore, CDC may significantly reduce communication load, alleviate the effects of stragglers, provide fault-tolerance, privacy and security. In this survey, we first introduce the fundamentals of CDC, followed by basic CDC schemes. Then, we review and analyze a number of CDC approaches proposed to reduce the communication costs, mitigate the straggler effects, and guarantee privacy and security. Furthermore, we present and discuss applications of CDC in modern computer networks. Finally, we highlight important challenges and promising research directions related to CDC. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"43fb2934-e7bf-4933-946e-becc16dd4285","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey of Coded Distributed Computing"]],"subsections":["023f3c5d-7805-45c9-b902-bdd9101e05cb","fc3d5a29-85c5-4d14-9457-330462c762f2","9ba1f087-de1d-45e2-9661-620859b12846","f06fb9ca-5252-413b-b98f-08903f106a9c","79250ecd-4bf6-4689-8a12-f801fbeb39c1","ee3d99f5-815c-4ba8-b336-968d0def830a","825a0578-3fe6-4ffc-8428-dfabbdc2d42a","8c1cb679-0c49-4078-b259-687a2df3303d","5a390405-bf12-44a0-aa67-38cecb4ab9be"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:intro}\nIn recent years, distributed computing has used for large-scale computation~ since it offers several advantages over centralized computing. First, distributed computing is able to provide computing services with high reliability and fault-tolerance. In particular, distributed computing systems can efficiently and reliably work even if some of the computing nodes, i.e., computers or workers, fail. Second, distributed computing has high computation speed as the computation load is shared among various computing nodes. Third, distributed computing systems are scalable since computing nodes can easily be added. Fourth, distributed computing is economical by using computing nodes with low-cost hardware. Likewise, distributed computing is adopted in cloud computing and other emerging services. Given the aforementioned advantages, distributed computing has been applied in numerous real-life applications such as telecommunication networks~ (e.g., telephone networks and wireless sensor networks), network applications~ (e.g., world-wide web networks, massively multiplayer online games and virtual reality communities, distributed database management systems, and network file systems), real-time process control~ (e.g., aircraft control systems), and parallel computation~ (e.g., cluster computing, grid computing, and computer graphics).\nHowever, distributed computing faces serious challenges. Let us consider one of the most common distributed computation frameworks, i.e., MapReduce~. MapReduce is a software framework and programming\nmodel for processing a computation task across large datasets using a large number of computing nodes, i.e., workers. A set of the computing nodes is referred to as a cluster or a grid. In general, the overall computation task is decomposed into three phases, i.e., the ``Map'' phase, ``Shuffle'' phase, and ``Reduce'' phase. In the Map phase, a master node splits the computation task into multiple subtasks and assigns the subtasks to the computing nodes. The computing nodes compute the subtasks according to the allocated Map functions to generate intermediate results. Then, the intermediate results are exchanged among the computing nodes, namely ``data shuffling'', during the Shuffle phase. In the Reduce phase, the computing nodes use these results to compute the final result in a distributed manner by using their allocated Reduce functions. \nDistributed computing has two major challenges. First, the computing nodes need to exchange a number of intermediate results over the network with each other in order to calculate the final result; this significantly increases communication overheads and limits the performance of distributed computing applications such as Self-Join, Terasort , and machine learning . For example, for the Hadoop cluster at Facebook, it is observed that on average, the data shuffling phase accounts for 33\\% of the overall job execution time~. 65\\% and 70\\% of the overall job execution time is spent on the Shuffle phase when running a TeraSort and Self-Join application respectively on a heterogeneous Amazon EC2 cluster~. In fact, the communication bottleneck is worse in the trainings of convolutional neural networks (CNNs), e.g., Resnet-50 and AlexNet , which include updates of millions of model parameters. Second, distributed computing is executed by a large number of computing nodes which may have very different computing and networking resources. As a result, there are straggling nodes or stragglers, i.e., computing nodes which run unintentionally slower than others, thereby increasing the overall time needed to complete the computing tasks. To address the straggler effects, traditional approaches such as work exchange~ and naive replication have been adopted for the distributed computing. However, such approaches either introduce redundancy or require coordination among the nodes that significantly increases communication costs and computation loads. This motivates the need for a novel technique that is able to more effectively and completely address the straggler effects and communication load of distributed computing. \nCoding theoretic techniques, e.g., channel coding such as low-density parity-check (LDPC)~, have been widely used in WiFi and cellular systems to combat the impact of channel noise and impairments. They have also been applied in distributed storage systems and cache networks~ to reduce storage cost and network traffic. The basic principle of the coding theoretic techniques is that redundant information, i.e., redundancy, is introduced in messages/signals before they are transmitted to a receiver. The redundancy is included in the messages in a controller manner such that it can be utilized by the receiver to correct errors caused by the channel noise. Coding theoretic techniques have been recently regarded as promising solutions to cope with the challenges in distributed computing~, . For example, coding theoretic techniques can be used to encode the Map tasks of the computing nodes such that the master node is able to recover the final result from partially finished nodes, thus alleviating the straggler effects~. Another example is that coding theoretic techniques enable coding opportunities across intermediate results of the distributed computation tasks which significantly reduces the communication load by reducing the number and the size of data transmissions among the processing nodes~. The combination of coding techniques and distributed computing is called coded distributed computing (CDC)~. Apart from reducing communication load and alleviating the effects of stragglers, CDC can provide fault-tolerance, preserve privacy~, and improve security~ in distributed computing. As a result, CDC approaches have recently received a lot of attention.\nCDC schemes can be applied in modern networks such as Network Function Virtualization (NFV) and edge computing. With the data mainly generated by end devices, e.g., Internet of Things (IoT), that have significant sensing as well as computational and storage capabilities, it is natural to perform some computations at the end devices, instead of the cloud which may not be able to handle the large amounts of data generated. As such, edge computing has been proposed as a solution to perform distributed computation tasks. In order to perform complex computations, e.g., the training of deep neural networks that involve a large number of training layers, resource-constrained devices may need to pool their resources to perform their computations collaboratively . This results in high communication costs and computation latency. CDC schemes can be implemented to overcome these challenges. Furthermore, CDC schemes can be implemented on edge computing networks that involve constantly-moving end devices , e.g., vehicles and smartphones, which imposes additional communication constraints.\nTo the best of our knowledge, although there are several surveys and books related to distributed computing, there is no survey paper on CDC. In particular, large-scale distributed computing and applications are discussed in . Surveys related to distributed computing include grid resource management systems for distributed computing~, resource allocation in high performance distributed computing~, wireless distributed computing~, and wireless grid computing~. This motivates the need for this survey on CDC. In summary, our survey has the following contributions:\n\\begin{itemize}\n\\item We describe the fundamentals of CDC. In particular, we introduce the commonly used distributed computation frameworks for the implementation of coding techniques and algorithms. We then discuss basic CDC schemes.\n\\item We review and discuss a number of CDC schemes to reduce the communication costs for distributed computing. The approaches include file allocation, coded shuffling design, and function allocation. We further analyze and compare the advantages and disadvantages of the CDC schemes. \n\\item We review, discuss, and analyze CDC schemes which mitigate the straggler effects of distributed computing. The approaches include computation load allocations, approximate coding, and exploitation of stragglers. \n\\item We review and present CDC schemes that can improve the privacy and security in distributed computing.\n\\item We analyze and provide insights into the existing approaches and solutions in the CDC literature.\n\\item We present and discuss applications of CDC in modern networks such as NFV and edge computing. \n\\item We highlight challenges and discuss promising research directions related to CDC.\n\\end{itemize}\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{structure.pdf}\n\\caption{\\small Structure of survey.}\n\\label{fig:structure}\n\\end{figure}\nFor the reader's convenience, we classify the related CDC studies according to the challenges that need to be handled. In particular, the issues are communication costs, straggler effects, and security. As such, readers who are interested in or working on related issues will benefit greatly from our insightful reviews and in-depth discussions of existing approaches, remaining/open problems, and potential solutions. The rest of this paper is organized as follows. Section~\\ref{sec:fundamentals} introduces the fundamentals of CDC. Section~\\ref{sec:schemes} presents basic CDC schemes. Section~\\ref{sec:comms} reviews CDC approaches that have been proposed to reduce communication costs. Section~\\ref{sec:stragglers} discusses CDC approaches that have been proposed to mitigate straggler effects. Section~\\ref{sec:security} presents CDC approaches that have been proposed to enhance privacy and security in distributed computing.~Section~\\ref{sec:applications} discusses applications of CDC. Section~\\ref{sec:challenges} highlights important challenges and promising research directions. The structure of the survey is presented in Figure~\\ref{fig:structure}. Section~\\ref{sec:conclusion} concludes the paper. A list of abbreviations commonly used in this paper is given in Table~\\ref{tab:table_abb}.\n\\begin{table}[h!]\n\\scriptsize\n \\caption{\\small List of common abbreviations used in this paper.}\n \\label{tab:table_abb}\n \\centering\n \\begin{tabularx}{8.7cm}{|Sl|X|}\n \\hline\n \\rowcolor{mygray}\n \\textbf{Abbreviation} & \\textbf{Description} \\\\ \\hline\nARIMA & Auto Regressive Integrated Moving Average \\\\ \\hline\nBCC & Batch Coupon's Collector\\\\ \\hline\nBGC & Bernoulli Gradient Code\\\\ \\hline\nBGW & Ben-Or, Goldwasser, and Wigderson \\\\ \\hline\nBPCC & Batch-Processing Based Coded Computing\\\\ \\hline\nC3P & Coded Cooperative Computation Protocol\\\\ \\hline\nCDC & Coded Distributed Computing\\\\ \\hline\nCNN & Convolutional Neural Network\\\\ \\hline\nCPGC & Coded Partial Gradient Computation\\\\ \\hline\nDAG & Directed Acyclic Graphs\\\\ \\hline\nDNN & Deep Neural Network \\\\ \\hline\nFRC & Fractional Repetition Coding\\\\ \\hline\nHCMM & Heterogeneous Coded Matrix Multiplication\\\\ \\hline\nIoT & Internet of Things\\\\ \\hline\nLCC & Lagrange Coded Computing\\\\ \\hline\nLDPC & Low-Density Parity-Check\\\\ \\hline\nLT & Luby Transform\\\\ \\hline\nMDS & Maximum Distance Separable\\\\ \\hline\nMMC & Multi-Message Communication\\\\ \\hline\nMPC & Multi-Party Computation \\\\ \\hline\nNFV & Network Function Virtualization \\\\ \\hline\nPCR & Polynomially Coded Regression\\\\ \\hline\nPDA & Placement Delivery Array \\\\ \\hline\nSDMM & Secure Distributed Matrix Multiplication\\\\ \\hline\nSGC & Stochastic Gradient Coding \\\\ \\hline\nSGD & Stochastic Gradient Descent\\\\ \\hline\nSVD & Singular Vector Decomposition\\\\ \\hline\nUAV & Unmanned Aerial Vehicle\\\\ \\hline\n\\end{tabularx}\n\\end{table}","id":"023f3c5d-7805-45c9-b902-bdd9101e05cb","level":"section","origin_cites_number":9,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:fundamentals}\nDistributed computing has been an important solution to large-scale, complex computation problems, which involves massive amounts of data. Various distributed computing models, e.g., cluster computing , grid computing and cloud computing , have been developed to perform the distributed computation tasks while providing high quality of services (QoS) to the users. Among the distributed computing models, cloud computing is gaining much popularity recently as it eliminates the need for users to purchase expensive hardware and software resources since the users only need to pay for the cloud services based on their usage needs in an on-demand basis. A comparison between cluster, grid and cloud computing models is summarized in Table~\\ref{tab:compare}. \nDistributed computing has been widely implemented in a variety of applications, e.g., sensor networks, healthcare applications, the development of smart cities, automated manufacturing processes and vehicular applications. In order to improve the performance of the distributed computing systems, various aspects such as resource allocation strategies , task allocation strategies , scheduling algorithms , incentive mechanisms , energy efficiency , network security and the performance modelling of the distributed computing systems have been extensively studied in the literature. However, the performance of the distributed computing systems is still limited by the high communication costs and straggler effects which lead to a longer time needed to execute the computation tasks. As a result, recent research has focused on coding techniques to overcome these implementation challenges of the distributed computing systems, the aims of which are to minimize the communication load as well as to mitigate the straggler effects.\nIn this section, we discuss commonly used distributed computation frameworks for the implementation of coding techniques and algorithms. Note that while the different computation frameworks are useful for different computing applications, we focus specifically on the MapReduce framework as the majority of the research works on CDC schemes are based on the MapReduce computation framework. We also introduce the two main lines of works in CDC, i.e., to reduce communication load and to mitigate the straggler effects, which aim to solve the challenges in distributed computing.\n\\begin{table*}[t]\n\\caption{\\small Comparison between cluster, grid and cloud computing models . }\n\\label{tab:compare}\n\\centering\n\\begin{tabular}{|l |c| c| c|}\n\\hline\n\\rowcolor{mygray}\n\\multicolumn{1}{c|}{\\textbf{Feature}} & \\textbf{Cluster} & \\textbf{Grid} & \\textbf{Cloud} \\\\ \\hline\nSize & Small to medium & Large & Small to large\\\\ \\hline\nNetwork type & Private, LAN & Private, WAN & Public, WAN\\\\ \\hline\nJob management and scheduling & Centralized & Decentralized & Both\\\\ \\hline\nCoupling & Tight & Loose/tight & Loose \\\\ \\hline\nResource reservation & Pre-reserved & Pre-reserved & On-demand\\\\ \\hline\nService-level agreement (SLA) constraint & Strict & High & High\\\\ \\hline\nResource support & Homogeneous and heterogeneous (GPU) & Heterogeneous & Heterogenous\\\\ \\hline\nVirtualization & Semi-virtualized & Semi-virtualized & Completely virtualized\\\\ \\hline\nSecurity type & Medium & High & Low \\\\ \\hline\nService-oriented architecture and heterogeneity support & Not supported & Supported & Supported\\\\ \\hline\nUser interface & Single system image & Diverse and dynamic & Single system image \\\\ \\hline\nInitial infrastructure cost & Very high & High & Low \\\\ \\hline\nSelf service and elasticisty & No & No & Yes\\\\ \\hline\nAdministrative domain & Single & Multi & Both\\\\ \\hline\n\\end{tabular}\n\\end{table*}","id":"fc3d5a29-85c5-4d14-9457-330462c762f2","level":"section","origin_cites_number":12,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Fundamentals of Coded Distributed Computing"]],"subsections":["9766cc05-52f0-40d7-8bb2-882e1f23113b","e5ec24a4-049a-49ee-b4fa-e11576083168"],"title":"Fundamentals of Coded Distributed Computing"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:mapreduce}\nWhile the distributed computation frameworks have moved beyond a simple MapReduce framework, the majority of the studies on CDC have focused on the MapReduce framework. \n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{conventional.pdf}\n\\caption{\\small Illustration of conventional MapReduce framework. The intermediate output pairs are represented by (key,frequency)[book number] and the output pairs are represented by (key,frequency).}\n\\label{fig:conventional}\n\\end{figure}\nMapReduce is a software framework and programming model that runs on a large cluster of commodity machines for the processing of large-scale datasets in a distributed computing environment. The cluster of computers is modelled as a master-worker system which consists of a single master node and multiple workers to store and analyzes massive amount of unstructured data. Due to its scalability and its ability to tolerate machines' failure , the MapReduce framework is commonly used in a wide range of applications , e.g., the analysis of web access logs, the clustering of documents, the construction of web-link graph that matches the all source URLs to a target URL and the development of machine learning algorithms. Generally, the MapReduce computation framework involves the processing of a large input file to generate multiple output pairs of which each pair consists of a key and a corresponding value. Figure~\\ref{fig:conventional} demonstrates the implementation of the conventional MapReduce framework to determine the frequency of occurrence of 4 specific words in the books, where the 4 processing nodes, i.e., the workers are to compute the 4 output pairs. There are three important phases in the MapReduce computation framework:\n\\begin{enumerate}\n\\item In the \\textbf{\\emph{Map}} phase, there are two stages, namely the allocation of Map tasks and the execution of Map tasks. Generally, a Map task is a function that generates a key-value output pair based on the allocated subfiles. Firstly, as seen in Fig.~\\ref{fig:conventional}, the master node splits the input file into 8 subfiles of smaller sizes and allocates the subfiles to the 4 workers. Secondly, each worker produces 4 intermediate key-value pairs for each allocated subfile using the map functions. Since the workers are allocated 2 subfiles each, each worker generates 8 computed intermediate results.\n\\item In the \\textbf{\\emph{Shuffle}} phase, the workers exchange their computed intermediate results to obtain the required intermediate results for the computation of the Reduce functions. In particular, in each time slot, one of the workers creates a message that contains information of the intermediate output pairs from the Map phase and transmits the message to other workers. The shuffling process continues until all workers have received the required intermediate output pairs for the Reduce phase. \n\\item In the \\textbf{\\emph{Reduce}} phase, the workers aggregate the 8 intermediate key-value pairs obtained from the Shuffle phase and compute the final result which is a smaller set of key-value pairs using the reduce functions. In particular, the reduction tasks are evenly distributed among the workers. Each reduce function is responsible for the evaluation of a key. For example, node 1 in Fig.~\\ref{fig:conventional} is responsible for the evaluation of ``Tree''. Therefore, the total number of reduce functions needed equals the total number of keys of the output, i.e., 4 reduce functions are needed to compute 4 output pairs.\n\\end{enumerate}\nApart from the MapReduce framework, there are other distributed computation frameworks that provide support for the processing of large-scale datasets such as:\n\\begin{itemize}\n\\item \\emph{Spark : }It supports applications that need to reuse a working dataset across the multiple parallel processes. These applications cannot be expressed as efficiently as acyclic data flows which are required in popular computation frameworks such as MapReduce. There are two use cases for the implementation of Spark computation framework: (i) iterative machine learning algorithms which operate on the same dataset repeatedly, and (ii) iterative data analysis tools, where different users query for a subset of data from the same dataset.\n\\item \\emph{Dryad : }By allowing the developers to construct their own communication graphs and the subroutines at the vertices through simple, high-level programming language, Dryad executes large-scale data-intensive computations over clusters consisting multiple computers. It does not require the developers to express their code in Map, Shuffle and Reduce phases in order to adopt the MapReduce framework for computations. Besides, the Dryad execution engine, which is based on the constructed data flow graph, takes care of the implementation issues of the distributed computation tasks such as the scheduling of tasks, allocation of resources and the recovery from communication and computation failures.\n\\item \\emph{CIEL : }The main characteristic of the CIEL computation framework is that it allows data-dependent data flows where the directed acyclic graphs (DAG) are built dynamically based on the execution of previous computations, rather than a statistically predetermined DAG. Instead of maximizing throughput, CIEL aims to minimize latency of individual tasks, which is very useful for the implementation of iterative algorithms, where latency grows significantly as the number of iteration increases.\n\\end{itemize}\nA comparison between the distributed computation frameworks is presented in Table~\\ref{tab:frameworks}.\n\\begin{table}[t]\n\\caption{\\small Comparison between distributed computation frameworks . }\n\\label{tab:frameworks}\n\\centering\n\\begin{tabular}{|m{2.8cm} | >{\\centering\\arraybackslash}m{1.5cm} |>{\\centering\\arraybackslash}m{1.5cm} |>{\\centering\\arraybackslash}m{1.3cm} |}\n\\hline\n\\rowcolor{mygray}\n\\multicolumn{1}{c|}{\\textbf{Feature}} & \\textbf{MapReduce } & \\textbf{Dyrad } & \\textbf{CIEL } \\\\ \\hline\nDynamic Control Flow & No & No & Yes \\\\ \\hline\nTask Dependencies & Fixed (2-stage) & Fixed (DAG) & Dynamic \\\\ \\hline\nFault Tolerant & Transparent & Transparent & Transparent \\\\ \\hline\nData Locality & Yes & Yes & Yes \\\\ \\hline\nTransparent Scaling & Yes & Yes & Yes \\\\ \\hline\n\\end{tabular}\n\\end{table}","id":"9766cc05-52f0-40d7-8bb2-882e1f23113b","level":"subsection","origin_cites_number":4,"parent_id":"fc3d5a29-85c5-4d14-9457-330462c762f2","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Fundamentals of Coded Distributed Computing"],["subsection","Coded Distributed Computation Frameworks"]],"subsections":[],"title":"Coded Distributed Computation Frameworks"},{"cite_extract_rate":0,"cites":[],"content":"There are two main lines of works in CDC. Firstly, the CDC schemes are implemented to minimize the \\emph{communication load} in distributed computing systems. Secondly, the CDC schemes aim to mitigate the \\emph{straggler effects} which cause a delay in the computation of the distributed tasks. For each of the objectives, we discuss the importance of solving these issues to improve the performance of the distributed computing systems. Then, we briefly discuss the existing solutions that have been proposed in the literature to meet these objectives. However, the current existing solutions do not adopt coding approaches. Different from the existing solutions, the CDC schemes are able to meet these objectives by introducing coded redundancy. In fact, the CDC schemes outperform the replication methods, e.g., naive replication and fork-join model , which introduce redundancy without coding techniques, in terms of time taken to execute the tasks.","id":"e5ec24a4-049a-49ee-b4fa-e11576083168","level":"subsection","origin_cites_number":2,"parent_id":"fc3d5a29-85c5-4d14-9457-330462c762f2","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Fundamentals of Coded Distributed Computing"],["subsection","Objectives of CDC Schemes"]],"subsections":["109ae9de-6803-47dc-a8d7-f4437916dc6b","274c5abc-70eb-46be-bb88-eff1a21724b3"],"title":"Objectives of CDC Schemes"},{"cite_extract_rate":0,"cites":[],"content":"} \\label{subsub:comms} Among the three phases of the MapReduce computation framework, the Shuffle phase dominates the time required to complete the computation tasks since multiple communications between the processing nodes are needed to exchange their intermediate results. For the Hadoop cluster at Facebook, it is observed that on average, the data shuffling phase accounts for 33\\% of the overall job execution time . In fact, the data shuffling phase is more time consuming when running on heterogeneous clusters with diverse computational, communication and storage capabilities. When running TeraSort , which is a conventional distributed sorting algorithm for large amount of data, and Self-Join applications on heterogeneous Amazon EC2 clusters, 65\\% and 70\\% of the overall job execution time is spent on the Shuffle phase respectively . The data shuffling process is also an important step in implementing distributed learning algorithms. In particular, to train machine learning models with distributed algorithms, it is common to shuffle the data randomly and run the algorithms iteratively for a number of times such that the processing nodes compute a different subset of the data at each iteration until there is a convergence . \\emph{For the logistic regression application which requires at least 100 iteration to converge, 42\\% of the iteration time is spent on communication }. For each time the data shuffling process is performed, the entire training dataset is communicated over the network, resulting in high communication costs which limit the performance of the distributed computing systems. \nSince the performance of the data shuffling process has a significant impact on the overall performance of the distributed computing systems, it has been extensively studied in the literature, e.g., . Various data shuffling strategies are proposed to achieve different objectives such as minimizing the job execution time, maximizing the utilization of resources and accommodating interactive workloads. While the overlap between the map computations and the shuffle communication helps to reduce the latency of the distributed computation tasks , the computing nodes require large storage capacities for buffering. An efficient and adaptive data shuffling strategy is proposed in the study of to manage the tradeoff between the accumulation of the shuffle blocks and the minimization of the utilization of memory space to reduce the overall job execution time and improve the scalability of the distributed computing systems. In , the authors propose a virtual data shuffling strategy which reduces storage space and traffic load in the network by delaying the actual movement of the data until it is needed to complete the computations in the Reduce phase. \nTo improve the performance of the data shuffling process, task scheduling algorithms such as Quincy scheduler , Hadoop Fair Scheduler and delay scheduling algorithm are also designed to allocate tasks to the workers. In the design of optimal task scheduling and task selection algorithms, the communication load can be minimized through various approaches such as by optimizing the placement of computation tasks, distributing the computing resources fairly to the nodes and maximizing the resource utilization of the systems. \nSince the task scheduling schemes are not the focus of this survey, we refer interested readers to the study of and the references therein for more detailed information on the scheduling techniques.\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{uncoded.pdf}\n\\caption{\\small Illustration of naive replication MapReduce framework.}\n\\label{fig:uncoded}\n\\end{figure}\nOne of the ways to reduce communication load in the shuffling phase is by repeating the computation tasks . Figure~\\ref{fig:uncoded} illustrates the implementation of the MapReduce framework by using the naive replication method in which 4 processing nodes are required to compute 4 output pairs, which is the same computation task illustrated in Fig.~\\ref{fig:conventional}. By simply replicating the Map tasks where each worker is required to compute more intermediate output pairs, i.e., 16 in Fig.~\\ref{fig:conventional} instead of 8 intermediate output pairs in Fig.~\\ref{fig:uncoded}, the communication load is reduced as fewer intermediate output pairs are communicated. For example, in the conventional MapReduce framework in Fig.~\\ref{fig:conventional}, node 1 needs to obtain 6 intermediate output pairs from other workers whereas in the naive replication scheme in Fig.~\\ref{fig:uncoded}, node 1 only needs to obtain 4 intermediate output pairs from other workers. \nHowever, the aforementioned non-coding methods have limits to which the communication load in the data shuffling phase can be minimized. Given that the naive replication method reduces communication load in the data shuffling phase by introducing redundancy to the systems (which is also discussed in Section~\\ref{subsec:cdccommunication}), coding techniques can be used to introduce redundancy to further minimize communication load, which will be discussed in-depth in Section~\\ref{sec:comms}.","id":"109ae9de-6803-47dc-a8d7-f4437916dc6b","level":"subsubsection","origin_cites_number":9,"parent_id":"e5ec24a4-049a-49ee-b4fa-e11576083168","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Fundamentals of Coded Distributed Computing"],["subsection","Objectives of CDC Schemes"],["subsubsection","Communication Load"]],"subsections":[],"title":"Communication Load"},{"cite_extract_rate":0,"cites":[],"content":"}In distributed computing systems, the processing nodes may have heterogeneous computing capabilities, i.e., different processing speeds. As such, another line of work in the CDC literature is to solve the bottleneck that results from a variation in time taken to complete the allocated tasks. In distributed computing systems, there are \\emph{stragglers}, which are the processing nodes that run unexpectedly slower than the average or nodes that may be disconnected from the network due to several factors such as insufficient power, contention of shared resources, imbalance work allocation and network congestion . As a result, the overall time needed to execute the tasks is determined by the slowest processing node. We briefly discuss the existing approaches (which are summarized in Table~\\ref{tab:priorstragglers}) to handle the straggler effects as follows:\n\\begin{table*}[t]\n\\caption{\\small Approaches to mitigate the straggler effects.} \n\\label{tab:priorstragglers}\n\\centering\n\\begin{tabular}{|>{\\centering\\arraybackslash}m{2.6cm} |m{6.5cm} |m{6.5cm}|}\n\\hline\n\\rowcolor{mygray}\n\\textbf{Approach} & \\multicolumn{1}{c|}{\\textbf{Key Ideas}} & \\multicolumn{1}{c|}{\\textbf{Shortcomings}} \\\\ \\hline\nStragglers Detection & Detects the straggling nodes, determines the cause of delay and implements targeted solutions & Difficult to identify the cause of delays \\\\ \\hline\nWork Stealing & Reallocates the remaining computation tasks from the slower workers to the faster workers & The capabilities of the slower workers are not maximized\\\\ \\hline\nWork Exchange & Reallocates the computation tasks every time a worker completes its tasks & Incurs high communication costs due to the feedback from the workers to the master node as well as the reallocation of data \\\\ \\hline\nNaive Replication & Introduces redundancy where each computation subtask is performed by more than one processing node & Incurs high communication costs and computation load\\\\ \\hline\nCoded Redundancy & Uses coding techniques to introduce redundancy such that the master node can recover the final result from any decodable set of workers & Still incurs high communication costs and computation load, but lower than that of naive replication \\\\ \\hline\n\\end{tabular}\n\\end{table*}\n\\begin{itemize}\n\\item \\emph{Stragglers Detection: }The most direct approach to mitigate the straggler effects is to detect the stragglers and act on them early in their lifetime. For example, Mantri detects the stragglers by identifying the tasks that are processed at a rate slower than the average. The system determines the cause of the delay and implements targeted solutions to mitigate the stragglers. The solutions include restarting the tasks allocated to the stragglers at other processing nodes, optimally allocating the tasks based on network resources as well as protecting against interim data loss by replicating the outputs of valuable tasks. \n\\item \\emph{Work Stealing : }The basic idea of the work stealing algorithm is to allow the faster processing nodes to take over the remaining computation tasks from the slower processing nodes so that the overall job execution time is minimized. By adopting this approach, the faster processing nodes operate continuously while leaving the slower processing nodes idle after their jobs are taken over by the faster processing nodes till the end of the computation session. \n\\item \\emph{Work Exchange : } By leveraging on the information of computational heterogeneity in the system, the master node first allocates the tasks to the workers based on their computational capabilities. Upon receiving the first computed result from any of the workers, the master node pauses the computation process and redistribute the remaining incomplete work to be computed by the workers. The process is performed for a number of iterations until all work is done. Since the workers need to inform the master node of the amount of work done at each time that the computation process is paused, additional communication costs are incurred. The higher communication costs are also a result of the reallocation of data to the workers. \n\\item \\emph{Naive Replication: }One of the solutions to handle stragglers in the distributed computing systems is by introducing redundancy to minimize computation latency. The computation task is replicated and executed over multiple processing nodes. Since all processing nodes are working on the same computation task, the time required to complete the computation task is determined by the fastest processing node. The partial computations of the remaining processing nodes are discarded. Experiments on Google Trace data have shown the effectiveness of the use of redundancy in minimizing computation latency by eliminating the need for the computed results by the stragglers. However, the introduction of redundancy comes at the expense of higher cost such as high communication bandwidth and high computation load . Various redundancy strategies have been analyzed to derive the limiting distribution of the state of the systems . Although the introduction of redundancy helps to reduce latency, the performance varies under different settings. In fact, in some settings, it is optimal to not use any redundancy strategy. Looking into this, the work in presents the optimal redundant-requesting policies under diverse settings.\n\\end{itemize}\nSimilar to the existing methods to reduce communication costs as discussed in the previous section (Section~\\ref{subsub:comms}), the existing methods to mitigate the straggler effects do not adopt coding approaches. Coding techniques can also be used to introduce redundancy into the systems to mitigate the straggler effects. The authors in investigate the tradeoff between latency and cost for both replication-redundancy systems and coded-redundancy systems. Coded-redundancy systems outperform the replication-redundancy systems in both latency and cost. In other words, by using coding techniques, the latency and cost incurred are lower than that of naive replication. The use of coding techniques to mitigate the straggler effects is discussed in more detail in Section~\\ref{subsec:codingtech} and Section~\\ref{sec:stragglers}.\nTo better understand the proposed CDC schemes, some of the commonly used performance metrics of the distributed computing systems are defined as follows:\n\\begin{enumerate}\n\\item \\emph{Storage Space }is defined as the total number of files stored across $K$ processing nodes, normalized by the total number of subfiles $N$ .\n\\item \\emph{Computation load }is represented by $r$, where $1\\leq r \\leq K$, is defined as the total number of Map functions computed across $K$ processing nodes, normalized by the total number of subfiles $N$ . In particular, when $r=1$, it means that each Map function is only computed by a single processing node. When $r=2$, it means that each Map function is computed by two processing nodes on average. \n\\item \\emph{Communication load }is represented by $L$, where $0\\leq L \\leq 1$, is defined as the total number of bits communicated by the $K$ processing nodes in the Shuffle phase, normalized by the total number of subfiles $N$ .\n\\end{enumerate}\nGiven that coding techniques are able to solve the aforementioned implementation challenges of the distributed computing systems, we review various proposed CDC schemes, which is the main focus of this paper. In the following section, we present a tutorial of the simple CDC schemes along these two lines of works, i.e., minimizing communication load and mitigating the straggler effects, of which is useful to better understand the related works discussed in Section~\\ref{sec:comms}, \\ref{sec:stragglers} and \\ref{sec:security}.","id":"274c5abc-70eb-46be-bb88-eff1a21724b3","level":"subsubsection","origin_cites_number":7,"parent_id":"e5ec24a4-049a-49ee-b4fa-e11576083168","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Fundamentals of Coded Distributed Computing"],["subsection","Objectives of CDC Schemes"],["subsubsection","Straggler Effects"]],"subsections":[],"title":"Straggler Effects"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:schemes}\nRecently, coding techniques have become a popular approach to solve the challenges of the distributed computing systems. As mentioned previously, there are two main lines of work in CDC: (i) to reduce the communication costs and (ii) to mitigate the straggler effects. In this section, we introduce the two basic CDC schemes which are the first works that show the effectiveness of using coding techniques to solve these two challenges separately. Then, we discuss a unified CDC scheme that characterizes the tradeoff between computation latency and communication load.","id":"9ba1f087-de1d-45e2-9661-620859b12846","level":"section","origin_cites_number":0,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"]],"subsections":["e9e65886-933d-4d54-aa7f-a7c03b8cc0f9","9f1f1f0b-c93c-42b6-815f-1a27a54d04ee","2dffdb1e-356b-4152-ad17-f178b093d653"],"title":"Coded Distributed Computing (CDC) Schemes"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:cdccommunication}\nIn the conventional MapReduce computation framework as shown in Fig.~\\ref{fig:conventional}, after the split of the input file into multiple subfiles, each subfile is mapped to only one of the processing nodes, i.e., the workers. The naive replication scheme, i.e., uncoded data shuffling scheme which relaxes this restriction, can reduce the communication costs of the system by allowing each subfile to be replicated and mapped to more than one processing nodes. In the example illustrated in Fig.~\\ref{fig:uncoded}, each subfile is repeated twice. Hence, as compared to the conventional MapReduce framework, each processing node has more Map tasks to perform in the naive replication scheme. However, by simple replication, the communication load in the Shuffle phase decreases and this gain is known as the \\emph{repetition gain}. Specifically, the communication load for uncoded schemes which include both the conventional MapReduce framework and the naive replication scheme, is denoted as follows :\n\\begin{equation}\nL_{uncoded}(r)=1-\\frac{r}{K},\n\\label{eqn:uncoded}\n\\end{equation}\nwhere $K$ is the number of processing nodes in the network. Based on Equation (\\ref{eqn:uncoded}), the communication loads achieved by the conventional MapReduce framework in Fig.~\\ref{fig:conventional} and the naive replication scheme in Fig.~\\ref{fig:uncoded} are $\\frac{3}{4}$ where $r=1$ and $\\frac{1}{2}$ where $r=2$ respectively.\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{coded.pdf}\n\\caption{\\small Illustration of Coded MapReduce framework.}\n\\label{fig:coded}\n\\end{figure}\nTo further reduce the communication load, i.e., to increase the repetition gain, the Coded MapReduce computation framework is proposed in where the Map tasks are carefully distributed among the processing nodes and the messages are encoded for transmission in the Shuffle phase by using coding theory. Figure~\\ref{fig:coded} illustrates the Coded MapReduce framework with 4 processing nodes to determine the 4 output pairs. After the Map phase, the processing node multicasts a bit-wise XOR of two computed intermediate pairs, denoted by $\\oplus$, satisfying the requirements of two other processing nodes simultaneously. For example, node 1 multicasts a bit-wise XOR of ``Bear'' and ``Fork'' to both nodes 2 and 3, which involves the transmission of only one packet of information, instead of two packets if the information is sent separately to the nodes in a unicast manner. Since the intermediate output pairs are now coded, there is an additional step of decoding before the reduce functions are applied. Given the coded ``BearFork'' information, node 2 is able to decode and recover the required ``Bear'' information by cancelling the ``Fork'' information since node 2 has also computed the same ``Fork'' information. Similarly, node 3 can recover ``Fork'' information by cancelling the ``Bear'' information. The simulation results in show that the Coded MapReduce reduces the communication load by 66\\% and 50\\% as compared to the conventional MapReduce framework and the naive replication scheme respectively.\nSince the use of coding techniques reduces both the latency and cost of the distributed computing systems , a more generalized framework known as the Coded Distributed Computing (CDC) scheme is introduced in . \nThe study of presents the fundamental inverse relationship between computation load and communication load. Specifically, the communication load in the Shuffle phase can be reduced by a factor $r$ by increasing the computation load in the Map phase by the same factor $r$, as shown in Fig.~\\ref{fig:tradeoff}. The communication load achieved by the CDC framework, $L_{coded}$, is given as follows:\n\\begin{equation}\nL_{coded}(r)=\\frac{1}{r}\\left(1-\\frac{r}{K}\\right).\n\\label{eqn:coded}\n\\end{equation}\nNote that the information-theoretic lower bound derived on the minimum communication load $L^*(r)$ equals $L_{coded}(r)$ of the CDC framework. As such, the optimal tradeoff between the computation load and communication load is characterized as follows :\n\\begin{equation}\nL^*(r)=L_{coded}(r)=\\frac{1}{r}\\left(1-\\frac{r}{K}\\right), r \\in\\mathcal{K}.\n\\end{equation}\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{tradeoff.pdf}\n\\caption{\\small Comparison of communication load between the CDC scheme and the uncoded scheme .}\n\\label{fig:tradeoff}\n\\end{figure}\nFrom Equation (\\ref{eqn:uncoded}) which applies to the uncoded computation schemes, the communication load $L$ decreases linearly as the computation load $r$ increases. However, when the number of processing nodes $K$ becomes large, there is no significant impact of increasing computation load on the communication load. On the other hand, for the proposed CDC framework, the communication load is inversely proportional to the computation load (Equation (\\ref{eqn:coded})). Even when $K$ becomes large, the increase in computation load still significantly reduces the communication load. \nSince the proposed CDC framework can be applied to any distributed computation framework with an underlying MapReduce structure, the performance of the CDC framework on TeraSort is evaluated. Experimental results on Amazon EC2 clusters show that the Coded TeraSort scheme which is a coded distributed sorting algorithm, achieves a reduction in the overall job execution time by factors of 2.16 and 3.39 with 16 processing nodes and computation loads of $r=3$ and $r=5$, respectively as compared to the uncoded TeraSort scheme.\nPreviously in , computation load is linearly dependent on the number of replicated Map tasks, i.e., load redundancy, as each processing node is assumed to compute all intermediate values for all subfiles allocated in its memory. However, the processing nodes can be selective in choosing the intermediate values to compute. As such, the load redundancy is no longer a direct measure of computation load. In other words, the storage constraints do not necessarily imply computation constraints. Building on the work in , the authors in propose an alternative tradeoff between computation load and communication load under a predefined storage constraint. In fact, the computation load is quadratic in terms of load redundancy. By taking load redundancy into consideration, an alternative computation-communication tradeoff curve is derived. In particular, this alternative tradeoff curve is especially relevant to the processing nodes that do have the sufficient resources or time to perform computations for all the allocated subfiles. Given that the processing nodes can only perform a limited amount of computations below the computation load threshold, the alternative tradeoff curve proposed in accurately defines the amount of communication load needed for the distributed computation tasks.\nSince the processing nodes are not required to compute all intermediate results that can be obtained from their locally stored data, the storage capabilities of the processing nodes should be considered in the CDC computation framework in . The study of characterizes the tradeoff between storage, computation and communication where the minimum communication load is determined given the storage and computational capabilities of the processing nodes. In particular, the optimal computation curve is obtained by characterizing the optimal storage-communication tradeoff given the minimum computation load. As a result, the triangles between the optimal communication curve and the optimal computation curve reflect the pareto-optimal surface of all achievable storage-computation-communication triples. However, as the number of processing nodes in the system increases, the number of input files required increases exponentially, resulting in an increase in the number of transmissions needed and hence high communication costs. As such, it is important to reduce the number of input files, which is discussed in Section~\\ref{subsubsec:packet} later. \n\\begin{table*}[!ht]\n\\caption{\\small Coding techniques to mitigate the straggler effects.} \n\\label{tab:stragglers}\n\\centering\n\\begin{tabular}{|m{2.0cm} |>{\\centering\\arraybackslash}m{1.5cm} |>{\\centering\\arraybackslash}m{3.3cm} |m{8cm}|}\n\\hline\n\\rowcolor{mygray}\n\\multicolumn{1}{|c|}{\\textbf{Problems}} & \\textbf{Ref.} & \\textbf{Coding Schemes} & \\multicolumn{1}{c|}{\\textbf{Key Ideas}} \\\\ \\hline\n\\multirow{8}{=}{Matrix-Vector} & & MDS Codes & Reduce the computation latency as the master node is able to recover the final result without waiting for the slowest processing node \\\\\n\\cline{2-4}\n & & LT Codes & Exploit the rateless property to generate unlimited number of encoded symbol from a finite set of source symbols\\\\\n\\cline{2-4}\n & & Short-Dot Codes & Reduce the length of dot-products computed at the processing nodes by introducing sparsity to the encoded matrices\\\\\n\\cline{2-4}\n & & s-diagonal codes & Exploit the diagonal structure of the matrices to achieve both optimal recovery threshold and optimal computation load\\\\\n\\hline\n\\multirow{16}{=}{Matrix-Matrix} & & Product Codes & Instead of encoding the matrices along one dimension as in the MDS-coded schemes, the matrices are encoded with MDS codes along both dimensions, i.e. row and column \\\\\n\\cline{2-4}\n & & Polynomial Codes & - Design the algebraic structure of the encoded matrices such that the MDS structure is found in both the encoded matrices and the intermediate computations\\\\\n & & & - Reconstruct the final results by solving the polynomial interpolation problem\\\\\n\\cline{2-4}\n & & MatDot Codes & Achieve lower recovery threshold than Polynomial Codes at the expense of higher communication costs by computing only the relevant cross-products\\\\\n\\cline{2-4}\n & & PolyDot Codes & Characterize the tradeoff between recovery threshold and communication costs where Polynomial Codes and MatDot Codes are the two extreme ends on this tradeoff curve\\\\\n\\cline{2-4}\n & & Sparse Codes & Exploit to sparsity in both input and output matrices to reduce computation load, while achieving near optimal threshold\\\\\n\\hline\n\\multirow{15}{=}{Gradient Descent} & & Fractional Repetition Coding & - Workers are divided into multiple groups and data is divided among the workers in each group \\\\\n & & & - Each partition of data is performed by more than one worker\\\\\n\\cline{2-4}\n & & Cyclic Repetition Coding & Data is allocated based on a cyclic assignment strategy\\\\\n\\cline{2-4}\n & & Cyclic MDS Codes & The entries to the columns of the encoding matrix are cyclic shifts of the entries to the first column \\\\\n\\cline{2-4}\n & & Reed-Solomon Codes & Use balanced mask matrix and chooses appropriate codewords from the RS codes to construct the encoding matrix \\\\\n\\cline{2-4}\n & & Batch Coupons Collector & - Divide the data into multiple batches which are allocated randomly to the workers for computations\\\\\n & & & - communication costs are significantly reduced as there is no need for communication between the workers and for feedback from the master node to the workers\\\\\n\\cline{2-4}\n & & Polynomially Coded Regression & Encode the data batches directly instead of the computed intermediate results\\\\\n\\hline\nConvolution & & Coded Convolution & Split both vectors into multiple parts of specified length and encodes one of the vectors with MDS codes \\\\\n\\hline\nFourier Transform & & Coded Fourier Transform & Leverage on recursive structure and the linearity of the discrete Fourier Transform operations\\\\\n\\hline\n\\end{tabular}\n\\end{table*}\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{codedcomputation.pdf}\n\\caption{\\small Illustration of coded computation with 3 workers. The master node is able to recover the final result upon receiving the computed results from any 2 workers.}\n\\label{fig:codedcomputation}\n\\end{figure}","id":"e9e65886-933d-4d54-aa7f-a7c03b8cc0f9","level":"subsection","origin_cites_number":5,"parent_id":"9ba1f087-de1d-45e2-9661-620859b12846","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"],["subsection","CDC to Minimize Communication Load"]],"subsections":[],"title":"CDC to Minimize Communication Load"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:codingtech}\nApart from reducing communication load in the Shuffle phase of the MapReduce framework, coding techniques can also be used to alleviate the straggler effects. Since matrix multiplication is one of the most basic linear operations used in distributed computing systems, a coded computation framework is proposed in to minimize computation latency of distributed matrix multiplication tasks. The coded computation framework uses erasure codes to generate redundant intermediate computations. In particular, \nthe master node encodes the equal-sized data blocks, i.e., submatrices and distributes them to the workers to compute the local functions. Upon completion, the workers transmit the computed results to the master node. The master node can recover the final result by using the decoding functions once the local computations from any of the decodable sets are completed. As seen in Fig.~\\ref{fig:codedcomputation}, the master node can recover the final result upon receiving the computed results from any 2 workers, instead of all 3 workers. As such, the total computation is not determined by the slowest straggler, but by the time when the master node receives computed results from some decodable set of indices. In this work , the authors explore the effectiveness of encoding the submatrices by using maximum distance separable (MDS) codes to mitigate the effects of stragglers.\nConsidering $K$ workers and a shifted-exponential distribution for the job execution time of the distributed algorithm, the simulation results show that the optimal repetition-coded distributed algorithm achieves a lower average job execution time when the straggling parameter is smaller than one, i.e., $\\mu<1$ but is still slower than the optimal MDS-coded distributed algorithm by a factor of $\\Theta (\\log K)$. However, the storage cost of the coded distributed algorithm is higher than that of the uncoded distributed algorithm as more data is required to be stored at the workers' sites for the coded distributed algorithm. The proposed algorithm is tested on an Amazon EC2 cluster and is compared against various parallel matrix multiplication algorithms, e.g., block matrix multiplication, column-partition matrix multiplication and row-partition matrix multiplication. The simulation results show that the proposed algorithm in performs better where the coded matrix multiplication achieves $40\\%$ and $39.5\\%$ reduction in average job execution time on clusters of m1-small and c1-medium instances with 10 workers each respectively as compared to the best of the three uncoded distributed algorithms. \nAlthough the MDS codes proposed in is able to mitigate the straggler effects, it cannot be generalized to all types of computation tasks. In order to mitigate the straggler effects of different distributed computation tasks, the coding techniques can be designed by exploiting the algebraic structures of the specific operations. An important performance metric that is introduced in the proposed CDC schemes is the recovery threshold, which refers to the worst-case required number of workers the master needs to wait to recover the final result for job completion . The smaller the recovery threshold, the shorter the computation latency. The objective is to reduce the recovery threshold so that the final result can be recovered by waiting for a smaller number of workers, thus contributing to a reduction in computation latency. Here, we discuss the coding techniques for various types of computation tasks, namely (i) matrix-vector multiplications, (ii) matrix-matrix multiplications, (iii) gradient descent, (iv) convolution and Fourier transform. Table~\\ref{tab:stragglers} summarizes the coding techniques designed for different distributed computation tasks.","id":"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee","level":"subsection","origin_cites_number":2,"parent_id":"9ba1f087-de1d-45e2-9661-620859b12846","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"],["subsection","CDC to Mitigate the Straggler Effects"]],"subsections":["8d5bcfc5-309d-4be0-81a5-8bdc9010e406","795c9420-d6c1-4f57-b9b5-1d340af6f0f3","c272fc00-2e47-4ea2-92bd-9be23a93633e","bc8c4799-cb61-4844-9fd9-6fd743d773dc"],"title":"CDC to Mitigate the Straggler Effects"},{"cite_extract_rate":0,"cites":[],"content":"Distributed matrix-vector multiplications are the building blocks of linear transformation computations which are an important step in machine learning and signal processing applications. In particular, the computation of linear transformation on high-dimensional vectors is required for popular dimensionality reduction techniques such as Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA) .\nInstead of using MDS codes that is proposed in , the authors in propose the use of Luby Transform (LT) codes to mitigate the straggler effects in distributed matrix-vector multiplications problems. Different from the works in and which use LT codes in fixed-rate settings, the rateless property of the LT codes can be exploited to generate unlimited number of encoded symbol from a finite set of source symbols. There are several advantages of using rateless codes: (i) near-ideal load balancing, (ii) negligible redundant computation, (iii) maximum straggler tolerance, and (iv) low decoding complexity. To further reduce the latency for practical implementations, blockwise communication can be used to transmit the submatrix-vector products. Instead of transmitting each encoded row-vector product separately to the master node, the workers are allowed to transmit the computed results in blocks where each block comprises a few row-vector products, reducing the number of communication rounds needed and hence minimizing the time needed to complete the computation tasks. \n The authors in propose Short-Dot codes to perform the computation of linear transforms reliably and efficiently under the presence of straggling nodes. Specifically, the processing nodes compute shorter dot products by imposing sparsity on the encoded submatrices. However, there is tradeoff between the optimal threshold recovery and the length of the dot-products where the master node needs to wait for computed results from more processing nodes when the length of the dot-products is shorter. The experimental results on the classification of hand-written digits of MNIST show that the Short-Dot codes achieve 32\\% faster expected computation time than the MDS codes .\nAlthough the Short-Dot codes can offer lower recovery threshold, the greater length of the dot-products means greater computation load for the processing nodes. With this concern, s-diagonal codes are proposed to achieve both optimal recovery threshold and optimal computation load by exploiting the diagonal structure of the encoding matrix. The computation time can be further reduced by using a low-complexity hybrid decoding algorithm which combines the peeling decoding algorithm and Gaussian elimination techniques.","id":"8d5bcfc5-309d-4be0-81a5-8bdc9010e406","level":"subsubsection","origin_cites_number":4,"parent_id":"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"],["subsection","CDC to Mitigate the Straggler Effects"],["subsubsection","Matrix-vector multiplications"]],"subsections":[],"title":"Matrix-vector multiplications"},{"cite_extract_rate":0,"cites":[],"content":"For large-scale distributed matrix-matrix multiplications, the coded computation schemes based on MDS codes are no longer suitable as the encoding and decoding processes scale with system size. Besides, the size of one of the matrices is assumed to be small enough in order to allow individual workers to perform the computations , restricting the implementation of MDS codes in large-scale multiplications. Hence, for large-scale problems, coded schemes not only need to achieve low computation time, but also require efficient encoding and decoding algorithms in order to minimize the overall job execution time. To deal with the straggler effects in high-dimensional distributed matrix multiplications, four types of coded computation schemes are proposed:\n\\begin{itemize}\n\\item \\emph{Product codes : }Product codes are implemented by building a larger code upon smaller MDS codes. Instead of encoding computations along only one dimension in MDS-coded schemes, the product codes encode computations along both dimensions, i.e., rows and columns of the matrices. When the number of backup workers increases sub-linearly with the number of subtasks, the product-coded schemes outperform the MDS-coded schemes in terms of average computation time and decoding time. In the linear regime, the one-dimensional decoding of the MDS-coded schemes is sufficient to recover the missing entries of the computation results. By allowing each row and column of the MDS constituent codes to have different code rates , the average computation time can be further reduced, contributing to a decrease in the overall job execution time. Product codes can also be used to solve higher-dimensional linear operations such as tensor operations by exploiting the tensor-structured encoding matrix . To reduce the decoding time of the product codes, efficient decoding algorithms such as Reed-Solomon codes and LDPC codes can be explored.\n\\item \\emph{Polynomial codes : }The key advantage of the polynomial codes in solving large-scale matrix multiplication problems is that they provide a lower bound to the optimal recovery threshold. For polynomial codes, the recovery threshold does not scale with the number of workers involved where as for MDS codes and the product codes, the recovery thresholds scale proportionally with the number of workers. By taking advantage of the algebraic structure of the polynomial codes, the master node can recover the final result by using polynomial interpolation algorithms, e.g., the Reed-Solomon codes, to decode the computation results from the workers. In addition to the optimal recovery threshold, the polynomial-coded schemes achieve minimum possible computation latency and communication load for distributed matrix multiplications. However, as the number of workers increases, the encoding and decoding costs are much higher than that of the product codes. Furthermore, by implementing Reed-Solomon codes, there is a limit to the number of workers that can be handled, which is not useful for practical implementations where the systems may involve up to thousands of nodes. As an extension to the polynomial codes proposed in , the entangled polynomial code that is proposed in achieves a lower recovery threshold which is only half of that achieved by the PolyDot codes , to be discussed later. Different from the polynomial codes which only allow column-wise partitioning of the matrices, the entangled polynomial codes allow arbitrary partitioning of the input matrices and evaluate only a subspace of bilinear functions such that unnecessary multiplications are avoided. The issue of numerical stability has also received attention to ensure the scalability of the polynomial-coded schemes . \n\\item \\emph{PolyDot codes : }PolyDot codes characterize the tradeoff between the recovery threshold and communication costs where the polynomial codes and the MatDot codes are special instances of this coding framework by considering two extreme ends of this tradeoff: minimizing either recovery threshold or communication costs. In particular, the MatDot codes achieve lower recovery threshold than the polynomial codes at the expense of much higher communication costs. This is achieved by only computing the relevant cross-products of the submatrices. Building on the work of PolyDot codes , the Generalized PolyDot codes are used to compute matrix-vector multiplications and achieve the same recovery threshold as the entangled polynomial codes . More importantly, the Generalized PolyDot codes can be extended for the training of large deep neural networks (DNNs), which consists of multiple non-linear layers. \n\\item \\emph{Sparse codes : }Although the polynomial codes achieve optimal recovery threshold, the computation loads of the workers increase due to the increased density of the input matrix, resulting in an increase in the overall job execution time which is not desirable. By exploiting sparsity, i.e., the number of zero entries of the encoded matrix, not only the recovery threshold is kept low, but the computation loads of the workers also decrease while maintaining a nearly linear decoding time , jointly contributing to the shorter overall job execution time. The basic idea of the algorithm proposed in is to allow the master node to find the linear combination of row vectors such that only the particular relevant sub-blocks are recovered. Then, the entire block of matrix can be recovered by aggregating the partial recovery of sub-blocks. Simulation results show that the sparse codes require an overall shortest time to complete the job as compared to other computation schemes, e.g., uncoded scheme, product codes , polynomial codes , sparse MDS codes and LT codes . A further analysis of the different components of subtasks, i.e., communication time, computation time and decoding time, shows that the sparse codes require much shorter time to decode, thus contributing significantly to the shorter overall job execution time.\n\\end{itemize}\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{frc.pdf}\n\\caption{\\small Fractional repetition coding with 6 workers and 2 stragglers.}\n\\label{fig:frc}\n\\end{figure}","id":"795c9420-d6c1-4f57-b9b5-1d340af6f0f3","level":"subsubsection","origin_cites_number":5,"parent_id":"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"],["subsection","CDC to Mitigate the Straggler Effects"],["subsubsection","Matrix-matrix multiplications"]],"subsections":[],"title":"Matrix-matrix multiplications"},{"cite_extract_rate":0,"cites":[],"content":"Apart from matrices, coding techniques can be applied to recover batch gradients of any loss function of the distributed gradient descent tasks. In , the authors have introduced the idea of gradient coding which is useful to mitigate stragglers that may slow down the computation tasks. Two gradient coding schemes are proposed, namely (i) fractional repetition coding (FRC) and (ii) cyclic repetition coding. In the FRC scheme, the workers are first divided into several groups. In each group, the data is equally divided and allocated to the workers. As a result, all groups of workers are replica of each other as shown in Fig.~\\ref{fig:frc}. Upon completing their subtasks, the workers in each group transmit the sum of partial gradients to the master node. In the cyclic repetition coding scheme, the data partitions are allocated to the workers based on a cyclic assignment strategy. The partial gradients computed by each worker are encoded by linearly combining them, of which is transmitted as a single coded message to the master node. By applying the gradient coding schemes, the distributed computation tasks do not suffer from delays incurred by the straggling nodes as the master node is able to recover the final result with the results from the non-straggling nodes. Other coding theories such as the cyclic MDS codes and the Reed-Solomon codes can be used to compute exact gradients of the distributed gradient descent problems. \nTo efficiently mitigate the straggler effects in distributed gradient descent algorithms, Batch Coupon's Collector (BCC) scheme is proposed in . In BCC, there are two important steps, namely (i) batching and (ii) coupon collecting. In batching, the training set is partitioned into batches which are distributed to the workers randomly whereas in coupon collecting, the master node collects the computed results from the workers until the results from all batches of data are received. This decentralized BCC scheme does not require any communication between the workers nodes and each worker is allocated data batches independently of other workers. As a result, it is easy to implement the BCC scheme in practical scenarios. Another important advantage of the BCC scheme is its universality. Different from other coding schemes that are designed to guarantee their robustness to a fixed number of straggling nodes, the BCC scheme does not require any prior knowledge about the straggling nodes which is more practical as it is difficult to estimate the number of straggling nodes present in the clusters. Furthermore, the BCC scheme can be easily extended to solve gradient descent problems over heterogeneous clusters where the workers have different computational and communication capabilities. The simulation results show that the BCC scheme speeds up the overall job execution time by up to 85.4\\% and 69.9\\% over the uncoded scheme and the cyclic repetition coding scheme respectively.\nThe gradient coding schemes proposed in illustrate the tradeoff between computation load and straggler tolerance. However, in non-linear learning tasks, communication costs dominate the overall job execution time as the number of iterations increases. As such, to generalize the coding schemes in , the authors in incorporate communication costs into their framework and present a fundamental tradeoff between the three parameters, namely computation load, straggler tolerance and communication costs. In particular, for a fixed computation load, the communication costs can be reduced by waiting for more workers.\nInstead of encoding the partial gradients computed based on uncoded data as seen in the studies of , coding techniques can be applied directly to the data batches to reduce the straggler effects and the overall job execution time. Considering the gradient computations for least-square regression problems, the polynomially coded regression (PCR) scheme exploits the underlying algebraic property to generate coded submatrices such that they are linear combinations of the uncoded input matrices. The master node can evaluate the final gradient by interpolating the polynomials from the computed partial gradients by the workers. Compared to the gradient coding schemes proposed in , the simulation results show that the PCR scheme achieves much lower recovery threshold and hence shorter computation and communication time, resulting in shorter time needed for overall job execution.","id":"c272fc00-2e47-4ea2-92bd-9be23a93633e","level":"subsubsection","origin_cites_number":2,"parent_id":"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"],["subsection","CDC to Mitigate the Straggler Effects"],["subsubsection","Gradient Descent"]],"subsections":[],"title":"Gradient Descent"},{"cite_extract_rate":0,"cites":[],"content":"The polynomial codes proposed in both the studies of and can be extended to the applications of distributed coded convolution based on the coded convolution scheme proposed in . The work in explores the use of MDS codes to encode the pre-specified vectors such that fast convolution is performed under deadline constraint. In addition to that, MDS codes can be used to mitigate the straggler effects in widely-implemented distributed discrete Fourier Transform operations which are used in many applications such as machine learning algorithms and signal processing frameworks.","id":"bc8c4799-cb61-4844-9fd9-6fd743d773dc","level":"subsubsection","origin_cites_number":2,"parent_id":"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"],["subsection","CDC to Mitigate the Straggler Effects"],["subsubsection","Convolution and Fourier transform"]],"subsections":[],"title":"Convolution and Fourier transform"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:unified} \nGiven the aforementioned coding schemes of and , we can observe that coding techniques are used to speed up distributed computing applications with two different approaches. On one hand, the authors in propose the ``Minimum Bandwidth Code\" that minimizes the communication load by repeating the computation tasks in the Map phase to introduce multicasting opportunities in the Shuffle phase. On the other hand, the authors in propose the ``Minimum Latency Code\" which minimizes the computation latency by encoding the Map tasks such that the master node is able to recover the final result without waiting for the straggling processing nodes.\nInspired by the aforementioned approaches, a unified coded scheme that characterizes the tradeoff between computation latency and communication load is proposed in given the computation load. The coding schemes in and are considered to be the extreme cases of this unified scheme. The unified coded scheme exploits the advantages of the two coding approaches by applying the MDS codes to the Map tasks and replicating the encoded Map tasks. Specifically, the unified coded scheme first encodes the rows of the matrix, following which the coded rows of the matrix are replicated and stored at the processing nodes in a specific pattern. Then, the processing nodes perform the computation until a certain number of the fastest processing nodes complete their tasks. To reduce communication load in the Shuffle phase, coded multitasking is used to exchange the intermediate results that are needed to recover the final results in the Reduce phase. An improvement to the latency-communication tradeoff presented in the unified coded scheme is proposed in by leveraging on the redundancy created by the repetition code. By increasing the redundancy rate of the repetition code, both the communication load in the Shuffle phase and the computation latency in the Map phase can be simultaneously improved, thus contributing to a improved latency-communication tradeoff.\nThe aforementioned initial works of coding schemes have shown their effectiveness in minimizing communication costs and alleviating the straggler effects. In the following sections, we review related works that leverage on coding techniques to address the implementation challenges of the distributed computing systems.","id":"2dffdb1e-356b-4152-ad17-f178b093d653","level":"subsection","origin_cites_number":3,"parent_id":"9ba1f087-de1d-45e2-9661-620859b12846","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Coded Distributed Computing (CDC) Schemes"],["subsection","Unified CDC Scheme"]],"subsections":[],"title":"Unified CDC Scheme"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:comms}\nWith more computing nodes that are equipped with greater capabilities to collect and process data, massive amounts of data are generated for the computations of user-defined computation tasks. Since the computations are scaled out across a large number of distributed computing nodes, large number of intermediate results need to be exchanged between the computing nodes in the Shuffle phase of the MapReduce framework to complete the computation tasks, resulting in significant data movement. Oftentimes, for the training of a model with distributed learning algorithms, data is shuffled at each iteration, contributing to high communication costs, which is a bottleneck of the distributed computing systems. As a result, there is a need to reduce the communication costs in order to speed up the distributed computation tasks. In this section, we present four approaches to reduce communication costs: \n\\begin{itemize}\n\\item \\emph{File Allocation: }In this approach, the studies aim to design an optimal file allocation strategy that considers the heterogeneities of the processing nodes' capabilities in the systems, maximizing data locality or reducing subpacketization level, which refers to the number of subfiles generated , . These different approaches work towards reducing the communication load in the distributed computing systems.\n\\item \\emph{Coded Shuffling Design: }Since data shuffling phase incurs a large proportion of the communication costs, data is encoded before it is transmitted so that the communication load can be minimized. Apart from combining coding with different techniques, e.g., compression and randomization techniques to improve the performance of the shuffling phase, the coding techniques are also designed to solve different computation problems, e.g., distributed graph computation problems , and multistage MapReduce computations .\n\\item \\emph{Consideration of Underlying Network Architecture: }Generally, the communications between the workers as well as between the workers and the master node are affected by the way that they are connected to each other. For example, server-rack architecture (Fig.~\\ref{fig:server}) is one of the most commonly used methods to connect the various servers. By taking the underlying architecture into consideration, the effectiveness of the coding implementation in reducing communication costs can be greatly improved.\n\\item \\emph{Function Allocation: }Similar to the allocation of files, the studies apply this approach on heterogeneous systems. In addition, some studies consider a cascaded system where each Reduce function is allowed to be computed at multiple processing nodes. In some cases where the data is randomly stored at the processing nodes, e.g., when the processing nodes are constantly moving, an optimal function allocation strategy is useful in reducing the number of broadcast transmissions and thus minimizing the communication load. \n\\end{itemize}","id":"f06fb9ca-5252-413b-b98f-08903f106a9c","level":"section","origin_cites_number":3,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"]],"subsections":["a30f2979-178f-4897-ab80-a68948bc9c5f","08339afb-55ae-4aba-a167-25ccaafe6521","2cb9e913-abbc-43af-837d-bdfad67a0786","1109cc8b-6b58-4ee5-ba62-ca162ae902db","71e25826-5a72-4095-8ca3-d26170c96db8"],"title":"Minimization of Communication Load"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:files}\nThe design of file allocation at each processing node is one of the major steps for the implementation of CDC scheme. There are a few approaches to an optimal file allocation strategy: (i) considering heterogeneous systems, (ii) maximizing data locality, and (iii) reducing subpacketization level.","id":"a30f2979-178f-4897-ab80-a68948bc9c5f","level":"subsection","origin_cites_number":0,"parent_id":"f06fb9ca-5252-413b-b98f-08903f106a9c","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","File Allocation"]],"subsections":["97abadeb-ed93-480e-acfa-dc9e87a6bd77","c8d7498f-a451-4053-980e-15ca23008379","73c478e0-3e12-49d2-8043-871c00017ef4"],"title":"File Allocation"},{"cite_extract_rate":0,"cites":[],"content":"}As discussed in Section~\\ref{subsec:cdccommunication}, although the CDC scheme proposed in carefully allocates the subfiles to the processing nodes in order to introduce coded multitasking opportunities, it considers a homogeneous system which may not be useful for practical implementation. In order to appropriately allocate the files to the distributed computing nodes, heterogeneous systems where the processing nodes have diverse storage, computational and communication capabilities, should be considered in determining the optimal file allocation strategy and coding scheme that minimize the communication load . \nBy leveraging on the extra storage capacity of the workers, the communication costs between the master node and the workers in the process of data shuffling are minimized. The reason is that if more data can be stored at the workers, fewer communication rounds are needed for the workers to receive shuffled data from the master node. In the extreme case, if the worker can store the entire dataset, there is no communication needed for the worker to receive shuffled data in any iteration. As a result, there is a tradeoff between the storage capacity of the workers and the communication overhead in the data shuffling process. In the data shuffling process, there are two phases, namely data delivery and storage update. Instead of a random storage placement , a deterministic and systematic storage update strategy creates more coding opportunities in transmitting data to the workers at each iteration, reducing the communication load.","id":"97abadeb-ed93-480e-acfa-dc9e87a6bd77","level":"subsubsection","origin_cites_number":3,"parent_id":"a30f2979-178f-4897-ab80-a68948bc9c5f","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","File Allocation"],["subsubsection","Considering Heterogeneous Systems"]],"subsections":[],"title":"Considering Heterogeneous Systems"},{"cite_extract_rate":0,"cites":[],"content":"}One of the important factors in determining the optimal file allocation strategy is data locality. Data locality is defined as the percentage of local tasks over the total number of Map tasks, i.e., the fraction of Map tasks that are allocated to the processing nodes having the required data for computations such that no communication is needed to obtain the data. High data locality means that less communication bandwidth is needed for the transmission of subfiles, which is required if the processing node does not have the needed subfiles for the execution of the Map tasks. In order to maximize data locality, the problem of allocation of Map tasks to different processing nodes can be tackled by solving a constrained integer optimization problem .","id":"c8d7498f-a451-4053-980e-15ca23008379","level":"subsubsection","origin_cites_number":1,"parent_id":"a30f2979-178f-4897-ab80-a68948bc9c5f","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","File Allocation"],["subsubsection","Maximizing Data Locality"]],"subsections":[],"title":"Maximizing Data Locality"},{"cite_extract_rate":0,"cites":[],"content":"} \\label{subsubsec:packet}As the number of processing nodes in the network increases, the input file needs to be split into a large number of subfiles. Specifically, the number of subfiles generated increases exponentially in the number of processing nodes . However, there is a maximum allowable subpacketization level, i.e., number of subfiles, where the dataset can only be partitioned into a limited number of packets, beyond which the communication load increases due to more transmissions required and the unevenly-sized intermediate results. Hence, there are several reasons for the reduction of subpacketization level: (i) to reduce the communication load in the Shuffle phase even when there is a large number of processing nodes, (ii) to reduce the packet overheads which increases with the number of broadcast transmissions, and (iii) to reduce the number of unevenly-mapped outputs which require zero padding. To keep the subpacketization level below the maximum allowable level, Group-based Coded MapReduce allocates the dataset based on the random groupings of the processing nodes and allows the processing nodes to cooperate in the transmission of messages. \nTo avoid splitting the input file too finely, the authors in use an appropriate resolvable design , which is based on linear error correcting codes, to determine the number of subfiles, the allocation of the subfiles to the processing nodes and the construction of the coded messages in the Shuffle phase. Building on this initial work, the authors in use the resolvable design based scheme to solve the limitation of the compressed CDC scheme that uses both compression and coding techniques. Although the compressed CDC scheme helps to reduce communication load, it requires large number of jobs to be processed simultaneously. Hence, the resolvable design based scheme is used to reduce the number of subfiles generated. Specifically, for each job in the compressed CDC scheme, the single-parity code is used to split the input file and the resolvable design based scheme is used to allocate the subfiles to the processing nodes. By aggregating the underlying functions and applying the resolvable design based scheme, multiple jobs can be processed in parallel while minimizing the execution time in the Shuffle phase, contributing to the reduction of the overall job execution time. Although the number of subfiles or number of jobs generated still increases exponentially with some of the system parameters, e.g., number of computing nodes and number of output functions, the exponent is much smaller when the resolvable design based scheme is implemented.\nIn addition to the exponential increase in the number of subfiles required, the number of output functions required also increases exponentially as the number of processing nodes in the network increases. There are other methods to reduce the number of subfiles and the number of output functions such as the hypercube computing scheme and the placement delivery array (PDA) . However, most of the CDC schemes consider non-cascaded systems, i.e., each Reduce function is computed at exactly one processing node . In , a cascaded system is considered but only two values for the number of processing nodes that perform each Reduce function are considered. By applying the concept of PDA to the distributed computation framework, the performance of the proposed computation scheme is evaluated for different number of processing nodes that compute the Reduce functions . Although the implementation of these various methods reduces the number of subfiles generated, it may come at the expense of higher communication load , .","id":"73c478e0-3e12-49d2-8043-871c00017ef4","level":"subsubsection","origin_cites_number":3,"parent_id":"a30f2979-178f-4897-ab80-a68948bc9c5f","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","File Allocation"],["subsubsection","Reducing Subpacketization Level"]],"subsections":[],"title":"Reducing Subpacketization Level"},{"cite_extract_rate":0,"cites":[],"content":"In the design of coded shuffling algorithms, we have classified the approaches into three different categories: (i) compression and randomization, (ii) coding across multiple iterations and (iii) problem-specific coding approaches.","id":"08339afb-55ae-4aba-a167-25ccaafe6521","level":"subsection","origin_cites_number":0,"parent_id":"f06fb9ca-5252-413b-b98f-08903f106a9c","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","Coded Shuffling Design"]],"subsections":["de2801c0-f5ea-4b6a-9559-1daeed4b4d16","47dbc343-daef-4549-a329-038a4aa2efda","4b18764b-6d94-4230-8ec2-e099673e37eb"],"title":"Coded Shuffling Design"},{"cite_extract_rate":0,"cites":[],"content":"To further reduce the communication costs of the distributed computation tasks, the design of the coded data shuffling scheme can incorporate different techniques to create more coded multicasting opportunities. Besides, the coded shuffling schemes are designed to minimize communication costs for different distributed computation problems such as iterative algorithms, graph computations and multistage dataflow problems. \nThe work in generates replications of the computation tasks in the Map phase in order to reduce communication load in the Shuffle phase by coding and multicasting the intermediate results. To further reduce the communication load, compression and randomization techniques can be applied to the design of the coded shuffling algorithms. \n\\begin{itemize}\n\\item \\emph{Compression Techniques: }The compressed CDC computation scheme is proposed in by jointly using two techniques, i.e., compression and coding techniques. Each processing node first computes the allocated Map tasks and generates the intermediate results. By using the compression techniques, several intermediate results of a single computation task are compressed into a single pre-combined value. The communication bandwidth needed to transmit a single pre-combined value is much smaller than that of transmitting several uncombined intermediate values since the size of the pre-combined value equals the size of only one intermediate value. With the pre-combined values from different computation tasks, the processing node codes them for multicasting to other processing nodes simultaneously. There are two advantages to this compressed CDC scheme: (i) the communication load is reduced proportional to the storage capacity of each processing node, and (ii) the communication load does not scale linearly, i.e., slower than linear, with the size of the dataset.\nIn some cases, e.g., parallel stochastic gradient descent (SGD) algorithms, instead of transmitting intermediate results, computed gradient updates are exchanged among the workers. In such cases, Quantized SGD , a compression technique, can be used to reduce communication bandwidth used during the gradient updates between the processing nodes. In each iteration, the processing nodes are allowed to adjust the number of transmitted bits by quantizing each component to a discrete set of values and encoding these quantized gradients.\n\\item \\emph{Randomization Techniques: }Instead of introducing coded multicasting opportunities to reduce the communication load in the Shuffle phase, there are other coding techniques that can be applied to increase efficiency of data shuffling. One of the ways is to perform a semi-random data shuffling and coding scheme based on pliable index coding which introduces randomization in the data shuffling process . There are two important modifications made to the conventional pliable index coding scheme , which is used to minimize the number of broadcast transmissions while satisfying users' demands. Firstly, the correlation of messages between workers is reduced. In order to do so, a message should only be transmitted to a fraction of the workers so that the same message is not held by all workers. As such, the pliable index coding problem is formulated with an objective where the goal is to minimize the number of broadcast transmissions under the constraint of a maximum number of workers that can receive the same message. Secondly, the correlation of messages between iterations is reduced. The reduction of correlation of messages prevents the workers from performing computations on the same dataset after shuffling, which may be redundant. A two-layer hierarchical structure is proposed for data shuffling. In the upper layer, the messages are partitioned into multiple groups of which each group of messages is transmitted to a fraction of workers. In the lower layer, each group of messages and the corresponding allocated workers are formulated as a constrained pliable index coding problem. Randomization occurs in two stages: (i) when the master node selects the messages in each group and transmits them to the workers, and (ii) when the workers discard old messages from their cache. Experimental results \nshow that the proposed pliable index coding requires only 12\\% of broadcast transmissions needed by an uncoded scheme, i.e., random shuffling with replacement scheme.\n\\end{itemize}","id":"de2801c0-f5ea-4b6a-9559-1daeed4b4d16","level":"subsubsection","origin_cites_number":4,"parent_id":"08339afb-55ae-4aba-a167-25ccaafe6521","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","Coded Shuffling Design"],["subsubsection","Compression and randomization"]],"subsections":[],"title":"Compression and randomization"},{"cite_extract_rate":0,"cites":[],"content":"Most works on coded iterative algorithms focus on the optimization of a single computation iteration or the minimization of communication load in a single communication round . However, multiple rounds of communications are generally required to solve distributed iterative problems. In the studies of and , the results of multiple iterations are transmitted in a single round of communication by jointly coding across several iterations of the distributed computation task. By leveraging on the computation and storage redundancy of the workers, the number of communication rounds between the master node and the workers is greatly reduced, resulting in a reduction in the communication costs. However, the computation and storage costs may not be optimal as compared to uncoded computing schemes, e.g., and , which achieve near-optimal computation and storage costs.","id":"47dbc343-daef-4549-a329-038a4aa2efda","level":"subsubsection","origin_cites_number":2,"parent_id":"08339afb-55ae-4aba-a167-25ccaafe6521","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","Coded Shuffling Design"],["subsubsection","Coding across multiple iterations"]],"subsections":[],"title":"Coding across multiple iterations"},{"cite_extract_rate":0,"cites":[],"content":"Apart from the typical distributed computation problems, the MapReduce framework can also be used to solve distributed graph computation problems, . However, for graph computing systems, the computation at each vertex is a function of the graph structure which means that each computation only needs data from its neighbouring vertices. More specifically, the communication load in the Shuffle phase depends on the connectivity probabilities of vertices in the graph, in which each vertex is only allowed to communicate reliably with a subset of random vertices . As a result, the CDC scheme proposed in (which was previously discussed in Section~\\ref{subsec:cdccommunication}) is not applicable to solving the graph computation problems. Looking into this, the authors in propose a coded scheme to solve the problem of random Erdös-Rényi graph while minimizing the communication load in the Shuffle phase. A similar inverse tradeoff curve between the computation load and average communication load is obtained by using coding techniques in solving distributed graph computations. \nMoreover, many distributed computing applications consists of multiple stages of MapReduce computations. The multistage data flow can be represented by a layered DAG in which the processing nodes, i.e., vertices in a particular computation stage are grouped into a single layer. Each vertex computes a user-defined function, i.e., Map or Reduce computation that transforms the given input files into intermediate results whereas the edges represent data flow between the processing nodes. By exploiting the redundancy of the computing nodes, coding techniques are applied to the processing nodes individually to minimize the communication costs. The proposed work in considers a uniform resource allocation strategy where the computation of each vertex is distributed across all processing nodes. However, the communication load can be further reduced by reducing the number of processing nodes that are used to compute each vertex. Given that fewer processing nodes are used to compute each vertex, the computation load performed by each processing node increases and the processing nodes have more local information, thus reducing the need for communication to obtain the required information. Therefore, a dynamic resource allocation strategy is needed to further minimize the communication load in the multistage MapReduce problems.\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{server.pdf}\n\\caption{\\small Server-rack architecture where multiple servers in each rack are connected via a Top of Rack switch and the Root switch connects multiple Top of the Rack switches.}\n\\label{fig:server}\n\\end{figure}","id":"4b18764b-6d94-4230-8ec2-e099673e37eb","level":"subsubsection","origin_cites_number":2,"parent_id":"08339afb-55ae-4aba-a167-25ccaafe6521","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","Coded Shuffling Design"],["subsubsection","Problem-specific coding approaches"]],"subsections":[],"title":"Problem-specific coding approaches"},{"cite_extract_rate":0,"cites":[],"content":"Although various techniques can be used jointly with the coding techniques to increase the efficiency of data shuffling phase, it is important to consider the underlying network architecture, i.e., how the servers are connected to each other, in designing the coded shuffling algorithms. \nIn , Hybrid Coded MapReduce is proposed by considering the server-rack architecture (Fig.~\\ref{fig:server}) in the distributed computing systems. There are two types of communications in the Shuffle phase: (i) the cross-rack communication where the data is shuffled across different rack layers, and (ii) the intra-rack communication where the data is shuffled within the rack. In the first stage where cross-rack communication takes place, the Coded MapReduce algorithm (Fig.~\\ref{fig:coded}) is used to create multicasting opportunities for the transmission of messages. In the second stage where the intra-rack communication is performed, data is shuffled in a unicast manner where no coding technique is applied. The simulation results show that the cross-rack communication costs incurred by the hybrid scheme are lower than those of both Coded MapReduce and uncoded scheme at the expense of a higher intra-rack communication costs. Although the Coded MapReduce scheme still achieves the lowest total communication costs, the overall communication costs for the hybrid scheme can be further reduced by parallelizing the intra-rack operations to provide a more accurate comparison between the different computation schemes.\nThe CDC computation scheme proposed in is useful for networks with processing nodes that are closely located with each other and connected via a common communication bus. However, in practical distributed computing networks, it is hard to implement a common-bus topology for the physically separated processing nodes. Hence, to reduce communication load for the distributed computation tasks, it is important to consider a practical data center network topology to reap the coding benefits of the CDC schemes. As such, in , the authors propose a CDC scheme based on a widely-used low-cost network topology which is the t-ary fat-tree topology . It has the characteristics of network symmetry and connections between any two processing nodes in the network. Given a practical network topology design, the proposed topological CDC scheme achieves optimal max-link communication load over all links in the topology of which the optimal tradeoff between the max-link communication load and the computation load is characterized. \nAlthough the coded shuffling algorithm proposed in reduces communication load of the data shuffling process, there are two main limitations: (i) it assumes a perfect broadcast channel between the master node and the workers, and (ii) the theoretical guarantee of number of broadcast transmissions only holds when the number of data points approaches infinity. To overcome the limitations of the coded shuffling algorithm proposed in , Ubershuffle fills the missing entries in the encoding tables by reallocating the data points between the workers. This reduces the number of transmitted encoded packets, resulting in a reduction of communication load. The performance of the UberShuffle algorithm is evaluated when applied to different problem settings. The experimental results show that in comparison to the coded shuffling algorithm in , the UberShuffle algorithm reduces the shuffling time by up to 47.2\\% and 35.7\\% when implemented with the distributed SGD algorithm for a low-rank matrix completion problem and the parallel SGD algorithm for a linear regression problem respectively.","id":"2cb9e913-abbc-43af-837d-bdfad67a0786","level":"subsection","origin_cites_number":5,"parent_id":"f06fb9ca-5252-413b-b98f-08903f106a9c","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","Consideration of Underlying Network Architecture"]],"subsections":[],"title":"Consideration of Underlying Network Architecture"},{"cite_extract_rate":0,"cites":[],"content":"In most of the studies of CDC computation framework, non-cascaded systems are considered. In other words, each Reduce function can only be computed at exactly one processing node. Considering that the processing nodes have different storage, computational and communication capabilities, a non-cascaded heterogeneous system where each Reduce function is computed exactly once is considered. In this proposed scheme, the processing nodes are allocated with different number of Reduce functions of which the processing nodes with greater storage and computational capabilities are allocated more Reduce functions. This heterogenous Reduce function allocation creates a symmetry among the multicast groups such that each processing node in a group requests the same number of intermediate outputs from the other processing nodes in the same group. The heterogeneous CDC system achieves a lower communication load than an equivalent homogeneous CDC system.\nHowever, it is desirable for the Reduce functions to be computed at multiple processing nodes in some applications, e.g., iterative algorithms in the Spark model where the output of a MapReduce procedure acts as the input to the MapReduce procedure in the next iteration. Although the work in generalizes the CDC framework by allowing each Reduce function to be computed at more than one processing node, it only applies to homogeneous systems. Similar to , heterogeneities of the systems are also considered in . However, instead of a non-cascaded system, the authors propose a more general framework of cascaded CDC on heterogeneous systems. In other words, each Reduce function is allowed to be computed at multiple processing nodes. Since the processing nodes have different storage capacities, the number of subfiles stored at each processing node differs. The processing nodes with larger storage capacities are allocated more subfiles and thus, compute more intermediate results. Instead of allocating the files and functions randomly as in the work in , this cascaded CDC scheme uses a hypercube approach to allocate the files and functions to the processing nodes. The simulation results show that for the same number of processing nodes in the network, the proposed cascaded CDC scheme achieves smaller communication load than the state-of-art schemes that consider homogeneous systems , . \nIn the study of , the allocation of the functions to the processing nodes does not depend on their capabilities but on the data stored at the nodes, of which the data placement is assumed to be random. This is very useful for applications that the processing nodes are mobile and collect data on-the-move. If the probability that the processing nodes contain data is higher than the pre-defined threshold, it is possible to allocate Reduce functions such that the processing nodes do not need to exchange their intermediate results for the computations of Reduce functions, i.e., each processing node can compute the Reduce functions based on its locally stored data. This reduces the number of broadcast transmissions in the Shuffle phase, thus minimizing the communication load.\nAlthough heterogeneities of the systems are taken into consideration in some of the works , , , they merely focus on either file allocation or function allocation. On one hand, the work in proposes an optimal file allocation strategy in consideration of the heterogeneous storage capabilities of the processing nodes. However, the Reduce functions are distributed uniformly among the processing nodes. On the other hand, the works in and propose the allocation of Reduce functions based on the computational capabilities of the processing nodes. However, the input file is split equally and distributed among the processing nodes. Considering a more generalized heterogeneous system, a joint file and function allocation strategy is proposed in to reduce the communication load in the Shuffle phase. The file allocation and function assignment strategies allocate more subfiles and Reduce functions respectively to the processing nodes with higher computational and storage capabilities. Generally, the Reduce function assignment is related to the input file allocation as the processing nodes with more input files have greater storage and computational capabilities and hence are more capable of computing more output functions. In particular, there are two proposed schemes of function assignment, i.e., computation-aware function assignment and shuffle-aware function assignment. For computation-aware function assignment strategy, the number of functions allocated is proportional to the computational capabilities of the processing nodes in order to reduce computation latency. For the shuffle-aware function assignment strategy, the functions are mostly allocated to the processing nodes with high computational capabilities so that the communication load in the Shuffle phase is minimized. The simulation results show that the communication loads achieved by both computation-aware and shuffle-aware function assignment strategies are lower than the uniform function allocation strategy. Besides, the computation-aware function assignment strategy requires fewer number of output functions as compared to the proposed schemes in and . However, the number of input files required is much larger, especially when the number of processing nodes in the system increases.\nWhile there has been great attention on the design of file allocation, coded shuffling algorithms and function allocation, all these works assume a fixed number of processing nodes in the distributed computing systems. For a given computation task that specifies the number of subfiles and the number of outputs, the allocation of functions and the data shuffling schemes can be implemented with minimum number of processing nodes. In the study of , the resource allocation problem is formulated as an optimization problem that minimizes the overall job execution time with optimal number of processing nodes used. For more practical implementation, the resource allocation strategy should consider the heterogeneity in the processing speed of the nodes, since the straggler effects cause an increase in computation latency which increases the overall job execution time.\nBy exploiting the fact that the processing nodes have time-varying computing resources, e.g., the processing nodes may have different central processing unit (CPU) power over time, an optimal computation task scheduling scheme helps to reduce the communication load. In the scheduling of tasks under dynamic computing resources, the total communication load is minimized by optimizing the number of allocated tasks and load redundancy at each time slot .\n\\begin{table*}[t]\n\\caption{\\small CDC schemes to reduce communication costs.} \n\\label{tab:communication}\n\\centering\n\\begin{tabular}{|>{\\centering\\arraybackslash}m{2cm} |>{\\centering\\arraybackslash}m{1.4cm} |>{\\centering\\arraybackslash}m{3.5cm} |m{6.3cm} |>{\\centering\\arraybackslash}m{2.3cm}|}\n\\hline\n\\rowcolor{mygray}\n\\textbf{Approach} & \\textbf{Ref.} & \\textbf{Coding Schemes} & \\multicolumn{1}{c|}{\\textbf{Key Ideas}} & \\textbf{Platform Support} \\\\ \\hline\n\\multirow{12}{=}{File Allocation} & & - & Deterministic and systematic storage update strategy & Heterogeneous\\\\\n\\cline{2-5}\n& & Hybrid Coded MapReduce & Allocates Map tasks such that data locality is maximized & Homogeneous\\\\\n\\cline{2-5}\n& & Group-based Coded MapReduce & Allocates dataset by using a group-based method in order to avoid high subpacketization level and allow processing nodes to cooperate in the transmission of messages& Homogeneous\\\\\n\\cline{2-5}\n& & Resolvable Design & Uses single-parity code to determine the number of subfiles and allocate the subfiles based on the corresponding resolvable design & Homogeneous\\\\\n\\cline{2-5}\n& & Placement Delivery Arrays & Construction of CDC schemes based on PDA which has the property of illustrating the placement and delivery phase in a single array & Homogeneous\\\\ \n\\hline\n\\multirow{18}{=}{Coded Shuffling Design} & & Compressed CDC & Pre-combines computed intermediate values of the same function, followed by coding the pre-combined packets for communication between different processing nodes & Homogeneous\\\\\n\\cline{2-5}\n& & Quantized SGD & Quantizes the component of the gradient vector to a discrete set of values and encodes the quantized gradients given their statistical properties & Homogeneous\\\\\n\\cline{2-5}\n& & Pliable Index Coding & Semi-random data shuffling scheme based on modified pliable index coding to reduce the number of communication rounds & Homogeneous\\\\\n\\cline{2-5}\n& & Cross-iteration Coded Computing & Jointly codes across multiple iterations for a single communication round & Homogeneous\\\\\n\\cline{2-5}\n& & CDC for distributed graph computing systems & Instead of communicating with all other processing nodes, each processing nodes only needs to communicate with a subset of processing nodes to obtain the required data to complete its computation tasks & Homogeneous\\\\\n\\cline{2-5}\n& & CDC for multistage dataflow & A more generalized CDC scheme is proposed to handle multistage dataflow computation tasks which are represented by layered DAGs & Homogeneous\\\\\n\\hline\n\\multirow{12}{=}{Consideration of Underlying Network Architecture} & & Hybrid Coded MapReduce & Reduces cross-rack communication at the expense of higher intra-rack communication based on the server-rack architecture & Homogeneous\\\\\n\\cline{2-5}\n & & Topological CDC & Considers t-ary fat-tree topology which is a more practical topology to connect the physically separated processing nodes in data center networks & Homogeneous\\\\\n\\cline{2-5}\n & & UberShuffle & Considers imperfect communication channel between the workers and the master node & Homogeneous\\\\\n\\cline{2-5}\n& & - & Considers multicore setup where each computing node can have multiple cores, e.g., the CPU instances of publicly available cloud infrastructure can deliver up to 128 cores & Homogeneous\\\\\n\\hline\n\\multirow{12}{=}{Function Allocation} & & - & Considers a non-cascaded system and allocates Reduce functions over a simplified heterogeneous network which comprises multiple homogeneous networks & Heterogenous \\\\\n\\cline{2-5}\n& & Cascaded CDC & Reduce functions are computed at more than one processing node and they are allocated based on the combinatorial design in & Heterogeneous\\\\\n\\cline{2-5}\n & & - & Allocates functions to maximize data locality such that the number of communication rounds required is reduced & Homogeneous\\\\\n\\cline{2-5}\n & & - & Joint file and function allocation strategy & Heterogeneous\\\\\n\\cline{2-5}\n & & - & Considers the availability of time-varying computing resources & Homogeneous\\\\\n\\hline\n\\end{tabular}\n\\end{table*}","id":"1109cc8b-6b58-4ee5-ba62-ca162ae902db","level":"subsection","origin_cites_number":6,"parent_id":"f06fb9ca-5252-413b-b98f-08903f106a9c","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","Function Allocation"]],"subsections":[],"title":"Function Allocation"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we have reviewed four approaches to minimize communication costs in distributed computing systems. For each approach, we discuss the solutions that are proposed in different works. We summarize the coding schemes to minimize communication costs in Table~\\ref{tab:communication}. The lessons learned are as follows:\n\\begin{itemize}\n\\item To handle the increasing amounts of data generated, more processing nodes are needed for the completion of distributed computations given the limited capabilities of the processing nodes. With more processing nodes connected to the network, more communication rounds of the computed intermediate results are required, resulting in higher communication costs which lead to a longer job execution time. Besides, the high communication costs in the data shuffling phase impede the implementation of efficient distributed iterative algorithms which are useful for the training of machine learning models. As such, the minimization of communication costs is a key step to achieve the objective of reducing the overall job execution time of the distributed computation tasks.\n\\item While having more processing nodes to perform the computation tasks in parallel helps to reduce the computation load of each processing node, the communication costs may increase, slowing down the entire computation process. Instead of generating infinitely large number of subfiles and distributing them among large number of processing nodes, several studies focus on determining the optimal number of subfiles to avoid the input file from splitting too finely. For example, the resolvable design based schemes and PDA approaches are adopted to split the input files. Hence, it may not be optimal to use all the processing nodes that are connected to the network. In fact, the authors in propose an optimal resource allocation scheme that determines the minimum number of processing nodes needed to achieve the minimum overall job execution time. \n\\item Coding techniques have shown to be effective in reducing communication load in the data shuffling phase at the expense of higher computation load . However, the two-dimensional tradeoff is insufficient to fully evaluate the performance of the CDC schemes. Apart from leveraging on the computational capabilities of the processing nodes, their storage capabilities can be exploited. For example, more data can be stored at processing nodes with higher storage capacities such that the number of communication rounds is reduced . Moreover, by considering the data stored at the processing nodes, the allocation of functions that maximizes data locality helps to reduce the need for communication bandwidth . Hence, instead of the two-dimensional tradeoff between computation load and communication load, the three-dimensional tradeoff between computation, communication and storage cost has to be carefully managed for the implementation of efficient CDC schemes.\n\\item For effective implementation of CDC schemes in practical distributed computing systems, the underlying architecture has to be considered. Generally, the distributed computing systems operate under the server architecture which consists of multiple racks where each rack has several servers. The Hybrid Coded MapReduce scheme reduces cross-rack communication at the expense of higher intra-rack communication. Besides, the communication channels between the master node and workers are imperfect. As a result, the theoretical gains of the coded data shuffling schemes are not achievable under practical setups. In order to design and analyze the performance of the CDC schemes, the limitations of the distributed computing systems should be taken into consideration. \n\\item Most of the CDC schemes focus on the minimization of communication costs in the data shuffling phase at the expense of reasonable increase in computation load. However, the computational overhead of the algorithm is not negligible under some settings. For example, under much faster broadcast environment, the UberShuffle algorithm incurs significant encoding and decoding costs such that the shuffling gain is offset by the high computational overhead. For future works, more practical CDC schemes can be proposed such that the communication costs are minimized while maintaining low computational cost to improve the performance of the distributed computing systems. Given the uncoded computing schemes, e.g., and , achieve near-optimal computation and storage costs, one possible research direction is to merge the communication-efficient proposed schemes with the uncoded computing schemes to reduce the computation and storage costs.\n\\item Although some of the works consider heterogeneities in the capabilities of the processing nodes to allocate files and functions, the presence of stragglers which have slower processing speeds still hinder the performance of the distributed computing systems. Therefore, we further discuss the approaches to mitigate the straggler effects in the next section.\n\\end{itemize}","id":"71e25826-5a72-4095-8ca3-d26170c96db8","level":"subsection","origin_cites_number":5,"parent_id":"f06fb9ca-5252-413b-b98f-08903f106a9c","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Minimization of Communication Load"],["subsection","Summary and Lessons Learned"]],"subsections":[],"title":"Summary and Lessons Learned"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:stragglers}\nIn distributed computing systems, processing nodes have different processing speeds and thus the time taken to complete their allocated subtasks differs from each other. Since the computation task is distributed among the processing nodes, the master node needs to wait for all processing nodes to complete their subtasks and return the computed results before recovering the final result. As such, the time taken to execute a computation task is determined by the slowest processing node. This is also known as the \\emph{straggler effects}. \nThe problem of straggler effects has been widely observed in the distributed computing systems. Previously, various methods such as straggler detection , , asynchronous execution and naive replication of jobs , have been proposed to reduce the overall time taken to execute the computation tasks. Recently, coding approaches have been shown to outperform the aforementioned methods in reducing computation latency of the distributed computing systems. In this section, we discuss three approaches to mitigate the straggler effects: \n\\begin{itemize}\n\\item \\emph{Computation Load Allocations: }Coding techniques can be implemented together with computation load allocation strategies to reduce the computation latency in the distributed computing systems. It is important to take into account the variation in the computational capabilities of the processing nodes to allocate computation load. As such, different prediction methods such as an long short-term memory (LSTM) algorithm , an Auto Regressive Integrated Moving Average (ARIMA) model and a Markov model are used to estimate the processing speeds of the nodes. Generally, the objective of the load allocation strategies is to minimize computation latency. However, in some applications where strict deadlines are given, the load allocation strategies aim to maximize timely computation throughput .\n\\item \\emph{Approximate Coding: }For some applications, e.g., location-based recommendation systems, exact solutions are not necessary. The studies explore different coding approaches to obtain approximate solutions to the problems. The approximate coding methods relax the requirement for convergence and thus reduce the number of workers that are required to return their computed results. This can avoid stragglers to make an adverse effect to the system.\n\\item \\emph{Exploitation of Stragglers: }The straggling nodes may have completed a fraction of the allocated computation tasks, of which is a waste to be ignored completely. In fact, the stragglers may not be persistent over the entire computation process and thus their partial computed results can be useful to recover the final result. In order to maximize the resource utilization of the straggling nodes, the workers are allowed to sequentially process their allocated subtasks and transmit their partial computed results continuously . However, this may come at the expense of higher communication load, which needs to be carefully managed. \n\\end{itemize}","id":"79250ecd-4bf6-4689-8a12-f801fbeb39c1","level":"section","origin_cites_number":6,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"]],"subsections":["bbc9eda4-c6aa-4f51-abdb-b74f0a6b37df","99bd280c-f0a8-4997-9dfb-23945fa49c2a","775a5f6a-bf8f-475d-ab6b-0bf91f59fd88","07a5860c-2192-4141-a8ce-fe22c5666e51"],"title":"Mitigation of Stragglers"},{"cite_extract_rate":0,"cites":[],"content":"Apart from having different straggling rates that affect the completion of tasks, the processing nodes have different capabilities, e.g., storage capacities, computing resources and communication bandwidths. To better handle the straggling nodes in the distributed computing systems, an optimal load allocation strategy that takes into account these heterogeneities is necessary to minimize the overall job execution time. Given the computation time parameters, i.e., the straggling and shift parameters of each worker, the Heterogeneous Coded Matrix Multiplication (HCMM) algorithm determines the allocation of computation load to each worker. The HCMM scheme exploits the benefits of both coding techniques and computation load allocation strategy to minimize the average computation time of the computation tasks. Given that it is difficult to derive closed-form expressions of expected computation time of the heterogeneous processing nodes, a two-step alternative problem formulation is proposed. In the first step, given a time period, the number of computed results by the workers is maximized by optimizing the load allocation. In the second step, given the derived load allocation in the first step, the time needed to ensure sufficient results are returned at a pre-defined probability is minimized. The simulation results show that when workers' computation time is assumed to follow a shifted exponential runtime distribution, HCMM reduces the average computation time by up to 71\\%, 53\\% and 39\\% over uniform uncoded, load-balanced uncoded and uniform coded load allocation schemes, respectively. In practical experiments over Amazon EC2 clusters, the combination of HCMM and LT codes outperforms the uniform uncoded, load-balanced uncoded and uniform coded load allocation schemes by up to 61\\%, 46\\% and 36\\%, respectively.\nAlthough HCMM achieves asymptotically optimal computation time, the decoding complexity is high, which suggests the opportunity to further speed up the overall computation tasks. In practical distributed computing systems, some processing nodes have the same computational capabilities, in terms of the same distributions of computation time, and thus they can be grouped together. By exploiting the group structure and heterogeneities among different groups of processing nodes , , the implementation of a combination of group codes and an optimal load allocation strategy not only approaches the optimal computation time that is achieved by the MDS codes, but also has low decoding complexity. In addition, by varying the number of allocated rows of the matrix to the workers , the computation latency can be reduced by orders of magnitude over the MDS codes with fixed computation load allocation as the number of workers increases. The load allocation strategy proposed in focuses mainly on the design of an optimal MDS code. Other efficient coding algorithms such as LT codes can also be explored in future studies.\nIn addition to the heterogeneous capabilities of the processing nodes, the amount of available resources of the processing nodes may vary over time. By always allocating computation tasks to the processing nodes with higher capabilities, delays may be incurred in completing the allocated tasks if the processing nodes start to work on the newly-allocated computation tasks in parallel. At the same time, the resources of the processing nodes with lower capabilities may be under-utilized. In order to maximize the resource utilization of the processing nodes, dynamic workload allocation algorithms which are adaptive to the time-varying capabilities of the processing nodes are proposed , , . To provide robustness against the straggling nodes, the design of the load allocation algorithms often depends on the historical data of the processing nodes such as computation time, which can be used to predict the processing speeds by using an LSTM algorithm , an ARIMA model or a Markov model . The best performing LSTM model achieves 5\\% lower prediction error than an ARIMA (1,0,0) model .\nIn the study of , the authors propose Coded Cooperative Computation Protocol (C3P) which is a dynamic and adaptive coded cooperation framework that efficiently utilizes the available resources of the processing node while minimizing the overall job execution time. Specifically, the master node determines the coded packet transmission intervals based on the responsiveness of the processing nodes. For the processing nodes which are not able to complete the tasks within the given transmission interval, they suffer from longer waiting time for the next coded packets. In comparison to the HCMM scheme which does not consider the dynamic resource heterogeneity in workers, the C3P framework achieves more than 30\\% improvement in task completion delay.\nThe dynamic and adaptive load allocation algorithms are especially useful in providing timely services with deadline constraints which are common in many IoT applications. For such applications, instead of minimizing task completion delay, the objective of the load allocation algorithms is to maximize timely computation throughput, i.e., the average number of computation tasks that are successfully completed before the given deadline . \nFor some applications that may need timely but not necessarily optimal results, it is more important to recover the final result with the highest accuracy possible by the stipulated deadlines than to solve for an exact solution. The algorithm to solve for an approximate solution requires significantly shorter computation time than that of an algorithm that solves for an exact solution . In the study of , the batch-processing based coded computing (BPCC) algorithm is proposed. The workers first partition the allocated encoded matrix into several batches, i.e., submatrices, for computations. As soon as the workers complete the computation of each batch of submatrix with a given vector, they return the partial computation results which are used to generate the approximate solution. Based on the computation time parameters, i.e., the straggling and shift parameters of the workers, the BPCC algorithm is used to optimally allocate the computation load to each worker by considering two cases: (i) negligible batching overheads, and (ii) linear batching overheads. Hence, the allocation of computation load is optimized by jointly minimizing both the overall job execution time as well as the potential overhead of batch processing. In addition to reducing computation latency, the BPCC algorithm exploits the partial computation results of all processing nodes, including the straggling nodes, which contribute to approximate solutions of higher accuracy . The simulation results show that BPCC algorithm with negligible batching overheads achieves up to 73\\%, 56\\% and 34\\% reduction in average job execution time over uniform coded, load-balanced uncoded and HCMM load allocation schemes respectively. The experimental results on Amazon EC2 clusters and Unmanned Aerial Vehicles (UAVs) based airborne computing platform also demonstrate similar results. \n\\begin{table*}[!ht]\n\\caption{\\small Approximate CDC schemes to mitigate the straggler effects.} \n\\label{tab:approximate}\n\\centering\n\\begin{tabular}{|>{\\centering\\arraybackslash}m{2.0cm} |>{\\centering\\arraybackslash}m{1.5cm} |>{\\centering\\arraybackslash}m{3.3cm} |m{8cm}|}\n\\hline\n\\rowcolor{mygray}\n\\multicolumn{1}{|c|}{\\textbf{Problems}} & \\textbf{Ref.} & \\textbf{Coding Schemes} & \\multicolumn{1}{c|}{\\textbf{Key Ideas}} \\\\ \\hline\n\\multirow{5}{=}{Matrix-Vector} & & Anytime Coding & Computations can be stopped anytime and the approximate solution is derived from the processing nodes that have completed their tasks\\\\\n\\cline{2-4}\n & & Coded Sequential Computation Scheme & A sequence of approximated problems are designed such that the time required to solve these problems is shorter than solving the original problem directly\\\\\n\\hline\n\\multirow{3}{=}{Matrix-Matrix} && CodedSketch & Use a combination of count-sketch technique and structured polynomial codes \\\\\n\\cline{2-4}\n & & OverSketch & Divide the sketched matrices into blocks for computations\\\\\n\\hline\n\\multirow{8}{=}{Gradient Descent} & & Bernoulli Gradient Codes & Use Bernoulli random variables as entries of the function assignment matrix\\\\\n\\cline{2-4}\n && Stochastic Block Codes & Interpolation between BGC and FRC \\\\\n\\cline{2-4}\n & & Stochastic Gradient Coding & Distribute data based on pair-wise balanced scheme and provide a rigorous convergence analysis of the proposed coding scheme\\\\\n\\cline{2-4}\n & & LDPC Codes & Encode the second moment of the data points\\\\\n\\cline{2-4}\n &, & Encoded Optimization & Encode both the labels and data such that the redundancy is introduced in the formulation of the optimization problems \\\\\n\\hline\n\\multirow{4}{=}{Non-Linear} & & Generalized PolyDot & Generalization of PolyDot codes and is used for the training of DNNs\\\\\n\\cline{2-4}\n & & Learning-based Approach & Design neural network architectures to learn and train the encoding and decoding functions to approximate unavailable outputs\\\\\n\\hline\n\\end{tabular}\n\\end{table*}","id":"bbc9eda4-c6aa-4f51-abdb-b74f0a6b37df","level":"subsection","origin_cites_number":5,"parent_id":"79250ecd-4bf6-4689-8a12-f801fbeb39c1","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"],["subsection","Computation Load Allocation"]],"subsections":[],"title":"Computation Load Allocation"},{"cite_extract_rate":0,"cites":[],"content":"For some applications, it is not necessary to obtain exact solutions. In this subsection, we review the coding approaches that are used to derive approximate solutions for the distributed computation tasks: (i) matrix multiplication, (ii) gradient descent, (iii) non-linear computations. Table~\\ref{tab:approximate} summarizes the approximate coding schemes for various distributed computation tasks.","id":"99bd280c-f0a8-4997-9dfb-23945fa49c2a","level":"subsection","origin_cites_number":0,"parent_id":"79250ecd-4bf6-4689-8a12-f801fbeb39c1","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"],["subsection","Approximate Coding"]],"subsections":["a171484f-77a0-4766-94ff-e7249ae60fc2","e8ed1de1-1ec4-4ea0-b627-b1ea53a5a3a3","fd439e01-9cb8-4b85-a497-db80e1e80ae6"],"title":"Approximate Coding"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsubsec:matrix} The anytime coding scheme derives an approximate solution by using the output results of the completed processing nodes at any given time. Based on singular value decomposition (SVD), the given computation task is decomposed into several subtasks of different priorities. More important subtasks are allocated more processing nodes for computations as they improve the accuracy of the approximation. To allow the users to receive useful information from time to time, approximate solutions can be transmitted to the users sequentially. This can be achieved by solving a sequence of approximated problems , instead of solving the original problem directly. \nTo further reduce the computation time of approximate matrix multiplications, sketching techniques , can be used to remove redundancy in the structure of the matrices through dimensionality reduction. However, by using sketching techniques, the recovery threshold increases as the redundancy is removed. In contrast, coding techniques reduces recovering threshold by introducing redundancy. As such, a combination of both techniques that carefully manage the tradeoff between the recovery threshold and the amount of redundancy can be implemented to minimize computation latency. In particular, count-sketch technique is combined with structured codes to mitigate the straggler effects by preserving a certain amount of redundancy, thereby achieving the optimal recovery threshold and hence computation latency , .","id":"a171484f-77a0-4766-94ff-e7249ae60fc2","level":"subsubsection","origin_cites_number":4,"parent_id":"99bd280c-f0a8-4997-9dfb-23945fa49c2a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"],["subsection","Approximate Coding"],["subsubsection","Matrix multiplications"]],"subsections":[],"title":"Matrix multiplications"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsubsec:gradient}To speed up the distributed gradient descent tasks, several approximate gradient coding schemes are proposed to approximately compute any sum of functions. Instead of constructing the gradient codes based on expander graphs which are difficult to compute due to high complexity, a more efficient and simpler Bernoulli Gradient Code (BGC) is proposed by using sparse random graphs which introduce randomness into the entries of the function assignment matrix. Since the performance of the gradient codes depends on the efficiency of the decoding algorithms, the authors also present two decoding techniques to recover the approximate solution from the outputs of the non-straggling nodes. The simulation results show that the BGC schemes can handle adversaries with polynomial-time computations but at a cost of higher decoding error than the FRC schemes . Besides, the optimal decoding algorithm always achieves a lower decoding error than that of the one-step decoding algorithm across various gradient coding schemes. A rigorous convergence analysis of the approximate gradient codes and the performance of BGC on different practical datasets such as the Amazon dataset, Covertype dataset and KC Housing dataset are presented in . Stochastic block code which is based on the stochastic block model from random graph theory, is an interpolation between FRC and BGC . On one hand, the FRC schemes achieve small reconstruction errors under random straggler selection while on the other hand, the BGC schemes are robust against polynomial-time adversarial stragglers.\nOther approximate gradient coding methods such as the Stochastic Gradient Coding (SGC) and LDPC codes are used to obtain an estimate of the true gradient. Similar to the idea of encoding data batches in the PCR schemes , the data variables in the optimization formulation can be efficiently encoded to mitigate the straggler effects in more general large-scale optimization problems such as support vector machines, linear regressions and compressed sensing , . Generally, in solving for approximate solutions to reduce computation latency of the distributed tasks, there is an inherent tradeoff between the recovery threshold and the approximation error where the recovery threshold can be reduced at the expense of larger approximation error .","id":"e8ed1de1-1ec4-4ea0-b627-b1ea53a5a3a3","level":"subsubsection","origin_cites_number":4,"parent_id":"99bd280c-f0a8-4997-9dfb-23945fa49c2a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"],["subsection","Approximate Coding"],["subsubsection","Gradient descent"]],"subsections":[],"title":"Gradient descent"},{"cite_extract_rate":0,"cites":[],"content":"By leveraging on the enormous amounts of data generated, machine learning algorithms are useful in making predictions and allowing devices to respond intuitively to user demands without human interception. Different neural network architectures are developed to make accurate inference given the dataset. Since some of the layers of the neural networks such as the max-pooling functions and the activation layer are non-linear, the overall computation of the functions are non-linear. As a result, most of the prior works on linear computations which are discussed in Sections~\\ref{subsec:codingtech}, \\ref{subsubsec:matrix} and \\ref{subsubsec:gradient}, are not applicable to the computation of the increasingly important non-linear neural networks, in which their performances are also limited by the straggling nodes. One of the few approaches that can be extended to the training of DNNs is the Generalized PolyDot codes which are used to compute matrix-vector multiplications. The Generalized PolyDot codes are used to code the linear operations at each layer of the neural networks. This coding scheme allows for errors in the training of each layer. In other words, decoding can still be performed correctly given that the amount of errors does not exceed the maximum error tolerance level. The effectiveness of coding techniques in mitigating the straggler effects of different neural network architectures such as AlexNet and Visual Geometry Group (VGG) in an IoT system is illustrated in . However, this unified coded DNN training strategy may not be relevant to the training of other neural networks which have large number of non-linear functions. As such, the authors in propose a learning-based approach for designing codes. Based on the dilated CNN and multilayer perceptrons (MLP), neural network architectures and a training methodology are proposed to learn the encoding and decoding functions. The outputs of the decoding functions are used to approximate the unavailable outputs of any differentiable non-linear function. The simulation results show that the learning-based approach to designing the encoding and decoding functions can accurately reconstruct 95.71\\%, 82.77\\% and 60.74\\% of the unavailable ResNet-18 classifier outputs on MNIST, Fashion-MNIST and CIFAR-10 datasets respectively.","id":"fd439e01-9cb8-4b85-a497-db80e1e80ae6","level":"subsubsection","origin_cites_number":2,"parent_id":"99bd280c-f0a8-4997-9dfb-23945fa49c2a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"],["subsection","Approximate Coding"],["subsubsection","Non-linear computations"]],"subsections":[],"title":"Non-linear computations"},{"cite_extract_rate":0,"cites":[],"content":"To avoid delays caused by the straggling nodes in the network, most distributed computation schemes ignore the work completed by the straggling nodes by either increasing the workload of the non-straggling nodes or by obtaining less accurate approximate solutions. However, the amount of work that has been completed by the straggling nodes, especially the non-persistent stragglers, is non-negligible and can be better utilized.\nIn order to exploit the computational capacities of these non-persistent stragglers, multi-message communication (MMC) is used where the workers are allowed to send multiple messages to the master node at each iteration. This allows the workers to transmit their partial computed results whenever they complete a fraction of the allocated task, rather than completing the entire computation task before transmitting the computed result in a single communication round. The work in considers the implementation of Lagrange Coded Computing (LCC) with MMC to minimize the average job execution time at the expense of higher communication load due to the increase in number of messages transmitted by the workers to the master node. Since the LCC scheme utilizes polynomial interpolation to recover the final result, the decoding complexity and the number of transmissions can be further reduced by increasing the number of polynomials used in decoding the computed results returned by the workers. The simulation results show that by exploiting the computing resources of the non-persistent stragglers via MMC, the average job execution time decreases as the computation load of each worker increases. However, since the communication load of the LCC-MMC scheme is constant as the computation load increases, it is suitable to be implemented only when computation time dominates the overall execution time of the distributed tasks. The total time needed to execute the distributed tasks includes the time needed for both computation and communication. Otherwise, if communication load is the cause of the bottleneck of the network, LCC without MMC should be used instead since the communication load can be reduced at the expense of higher computation load.\nGiven that MMC is allowed where the workers perform more than one round of communication for each iteration, the computation work done by the straggling nodes can be exploited by allowing sequential processing where the workers need to transmit the computation results of their completed subtask before working on the next subtask. To fully exploit the useful information provided by the straggling nodes, the hierarchical coded computation scheme is proposed in to utilize the computations from all workers. Each worker first divides the allocated computation task into layers of sub-computations, which are processed sequentially, i.e., the result of a layer of sub-computation is transmitted to the master node before the computation of the next layer starts. Since the workers have different processing speeds and the sub-computations are performed sequentially, the finishing time for each layer is different. MDS codes are used to encode the layers so that the finishing time of each layer is approximately the same. The top layers which have lower probability of erasure are encoded with higher-rate MDS codes whereas the bottom layers are encoded with lower-rate MDS codes. The simulation results show that for the same amount of tasks to be completed, the proposed hierarchical coded computation scheme achieves $1.5$ factor improvement in expected finishing time as compared to the coded computation scheme proposed in which ignores the computations of the straggling nodes.\nIn the study of , by computing the block products sequentially, the partial computation results from the straggling nodes are used to aggregate the final result. In order to preserve the sparsity of matrices for processing by the workers, instead of coding the entire matrix, the fraction of coded blocks can be specified to control the level to which coding is utilized in the solutions. There are two different approaches considered, depending on the placement of the coded blocks. When the coded blocks appear at the bottom of the workers, it is easier for the master to decode. When the coded blocks appear at the top of the nodes, it minimizes the computation load of the workers. As such, the placement can be calibrated based on the task requirements.\nWhile the coded computation schemes achieve low communication load and reduce the average job execution time for each iteration, the uncoded computation schemes have their own benefits of having no decoding complexity and allowing partial gradient updates. In order for a system to benefit from the advantages of both schemes, coded partial gradient computation (CPGC) scheme is proposed in . In the CPGC scheme, uncoded submatrices are allocated for the first computation since there is a high probability for each worker to complete their first computation task. Subsequently, coded submatrices are allocated to handle the straggling nodes. The master node is able to update the gradient parameters by using the computation results from a subset of workers and by exploiting the partial computations of the straggling nodes. As a result, the average job execution time for each iteration is reduced. \n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{mitigate.pdf}\n\\caption{\\small Approaches to mitigate the straggler effects include (a) computation load allocation by predicting the speed of the processing nodes , , (b) approximate coding and (c) exploitation of stragglers by allowing multi-message communications , .}\n\\label{fig:mitigate}\n\\end{figure}","id":"775a5f6a-bf8f-475d-ab6b-0bf91f59fd88","level":"subsection","origin_cites_number":2,"parent_id":"79250ecd-4bf6-4689-8a12-f801fbeb39c1","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"],["subsection","Exploitation of Stragglers"]],"subsections":[],"title":"Exploitation of Stragglers"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we have discussed three approaches (Fig.~\\ref{fig:mitigate}) to mitigate the straggler effects. The lessons learned are as follows:\n\\begin{itemize}\n\\item The straggler effect is a key issue to be resolved in order to reduce computation latency, hence minimizing the overall job execution time. Due to various factors such as insufficient power, contention of shared resources, imbalance work allocation and network congestions , some processing nodes may run slower than the average or even be disconnected from the network. Since the computation tasks are only completed when all processing nodes complete their computations, the time needed to complete the tasks is determined by the slowest processing node. Coding techniques have shown their effectiveness in reducing computation latency by introducing redundancy . In this section, we have explored the use of coding techniques for different distributed computation tasks, e.g., matrix-vector multiplications, matrix-matrix multiplications, linear inverse problems, iterative algorithms, convolutions and non-linear problems. While most of the research focuses on the design of encoding techniques, the decoding complexity of the codes also affects the computation latency significantly. Apart from Reed-Solomon codes and LDPC codes , more effective codes with low decoding complexity can be investigated in future studies. \n\\item Considering heterogeneities in the capabilities of the processing nodes, effective computation load allocation strategies are implemented to allocate workload to the processing nodes. We have discussed the proposed computation load allocation algorithms under different constraints, e.g., strict deadlines and time-varying computing resources. \nIn addition, different prediction methods such as the LSTM algorithm , an ARIMA model and a Markov model that predict the speeds of the processing nodes are explored. However, the stragglers may be non-persistent in nature and thus they may be useful when they are able to perform computations faster than the average rate. Hence, the load allocation based on the responsiveness of the processing nodes may be more useful in such situations. \n\\item Instead of exact solutions, it is acceptable to derive approximate solutions for some applications, e.g., location-based recommender systems. Various coding techniques to derive approximate solutions are investigated. For example, in the studies of and , sketching techniques are used with structured codes to minimize computation latency. However, there exists a tradeoff between the recovery threshold and the approximation error. For future works, an improvement to this tradeoff can be investigated.\n\\item Although the straggling nodes run slower than the average, the computations that are completed may still be useful. It is wasteful to not utilize the partial computed results of the straggling nodes. Besides, these partial computed results can help to improve the accuracy of the estimates. For example, in , the stragglers are treated as soft errors instead of erasures to minimize the mean-squared error of the iterative linear inverse solvers under a deadline constraint by using approximate weights. Outputs from all computing nodes, including the straggling nodes are used to recover estimates that are as close as possible to the convergence values when the computation deadline is brought forward or when the number of computing nodes increases. Unfortunately, as the processing nodes are required to send their partial results once they complete, more communication rounds are performed. The high communication costs may be the bottleneck of the distributed computation tasks. Given the advantages of using partial results from the straggling nodes, optimization approaches to minimize the communication costs between the master node and the workers should be explored.\n\\item The current studies in this section have proposed effective coding schemes for implementation. However, they do not consider security in the design of the coding schemes. For example, the FRC scheme achieves high accuracy in the presence of stragglers, but it is susceptible to attacks from adversarial stragglers, which turn more processing nodes into straggling nodes. Besides, other than the straggling workers, there may exist malicious or curious workers that may compromise the privacy and security of the system. Therefore, approaches to ensure secure coding are discussed in-depth in the next section. \n\\end{itemize}","id":"07a5860c-2192-4141-a8ce-fe22c5666e51","level":"subsection","origin_cites_number":5,"parent_id":"79250ecd-4bf6-4689-8a12-f801fbeb39c1","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Mitigation of Stragglers"],["subsection","Summary and Lessons Learned"]],"subsections":[],"title":"Summary and Lessons Learned"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:security}\nIn distributed computing, the data owner, master node, and workers may not belong to the same entity. For example, the data owner may wish to perform a task on a massive dataset on which intensive computations have to be performed. The computations may be divided and distributed to multiple workers on third-party computing services. However, sensitive data, e.g., in healthcare services , may be involved. In this case, \\textit{curious} workers may collude to obtain information about the raw data, whereas \\textit{malicious} workers may intentionally contribute erroneous inputs to introduce biases to the model. Besides, in some cases, the dataset does not belong to either the master node or the workers, and as such, the raw dataset has to be guarded against both parties.\nTo ensure that privacy and security are preserved during the computing tasks, conventional methods such as homomorphic encryption have been proposed in which the data is first encrypted before being shared to workers. However, the encryption techniques are found to be costly in terms of computation and storage costs . Besides, the secure multi-party computation (MPC) approaches mainly focus on the correctness and privacy of the data , whereas neglecting to reduce the complexity of computation at each workers, or the number of workers required to perform the task. Recently, coding techniques that have originally been introduced to mitigate straggling workers are increasingly utilized to provide information-theoretic privacy guarantees. Specifically, information-theoretic privacy considers the setup consisting honest but curious workers, in which collusions formed between $T$ of $N$ workers do not leak information about the training dataset. Coding schemes can be used to not only mitigate the stragglers but also the colluding curious workers and malicious workers as illustrated in Fig.~\\ref{fig:security}. For ease of exposition, we classify the related studies into two main categories in this section:\n\\begin{itemize}\n\\item \\textit{Secure Distributed Computing:} In this section, the studies aim to reduce the number of workers needed for information-theoretic privacy, i.e., where the colluding workers are unable to infer sensitive information from the dataset. For some studies, this objective is met while simultaneously preserving the efficiency of distributed computing, e.g., through providing resiliency against straggling workers . \n\\item \\textit{Secure Distributed Matrix Multiplication (SDMM):} The studies in the aforementioned category mainly focus on generic operations, e.g., addition, subtraction, multiplication, or computation of polynomial functions, whereas the studies in this category focus specifically on matrix multiplication. One key difference between the two categories is that SDMM considers the specific scenario in which \\textit{both} input matrices in the multiplication operation are private information, i.e., two-sided privacy , whereas the prior category mainly considers one-sided privacy, i.e., only one input matrix is private. In addition, a performance metric of interest in the SDMM literature is the download rate, i.e., the ratio of the size of desired result to the total amount of information downloaded by a user from the workers.\n\\end{itemize}","id":"ee3d99f5-815c-4ba8-b336-968d0def830a","level":"section","origin_cites_number":7,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Secure Coding for Distributed Computing"]],"subsections":["8d724f64-c2a5-4e3f-a151-1bdc01b1a822","d523704a-b56b-49c2-97c7-6500404f0ac2","362bc888-78c5-42b5-b72f-1d88622fe4e4"],"title":"Secure Coding for Distributed Computing"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:securedist}\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{cdc-security.pdf}\n\\caption{\\small Illustration of coding framework for the objectives of mitigating stragglers, colluding curious workers, and malicious workers.}\n\\label{fig:security}\n\\end{figure}\nIn Section~\\ref{subsec:codingtech}, we discuss that the polynomial codes proposed in have the desirable property of an optimal recovery threshold that does not scale with the number of workers involved. In consideration of this useful property, the authors in propose the polynomial sharing approach which combines the polynomial codes and the Ben-Or, Goldwasser, and Wigderson (BGW) scheme . The system model considered in this study is that the data originates from external sources, and thus has to be kept private against both the workers and master node. In contrast to the BGW approach which uses Shamir's scheme to encode the dataset, the study of proposes to encoded the dataset using the polynomial coding scheme. The authors show that the polynomial sharing approach may be applied to perform several procedures, e.g., addition, multiplication, and the computation of polynomial functions, while reducing the number of workers required to complete the task as compared to conventional MPC approaches even when workers have capacity-limited communication links. \nTypically, in conventional polynomial coding schemes, the dataset on which computations are performed is divided into multiple sub-tasks, with one sub-task encoded and assigned to each worker. In this case, faster workers that complete their task will be idle while waiting for straggling workers. To further mitigate the straggler effects, the authors in leverage on computation redundancy to propose the private asynchronous polynomial coding scheme in which a computation task is divided into several relatively smaller sub-tasks for distribution to each worker. This results in two key advantages, in addition to retaining the privacy preservation properties of polynomial coding. Firstly, the smaller sub-tasks can be successfully completed by straggling workers with limited computing capacity. Secondly, workers of the fastest groups are assigned more tasks to continue working throughout the whole duration rather than wait for the stragglers, thus reducing the computation time. \nHowever, the studies mainly utilize polynomial coding for privacy preservation which is restrictive in certain aspects, e.g., it only allows column-wise partitioning of the matrices . As such, the entangled polynomial codes are applied by as an extension to polynomial sharing, so as to further reduce the restrictions during the data sharing phase, and hence, the number of workers required to perform the same computations while meeting privacy constraints.\nWhile the studies of consider the scenario in which honest-but-curious workers are involved, workers may randomly be malicious in nature. As an illustration, a group of workers may be involved to compute gradients towards training a machine learning model. However, the gradients may be intentionally misreported by the workers to introduce biases or inaccuracies to the model . An existing approach is to perform median, rather than mean, based aggregation of the gradients to eliminate misreports that are usually outliers . However, the median based aggregation is computationally costly and faces convergence issues. As such, the study of proposes DRACO, which is based on the coding of gradients and algorithmic redundancy, i.e., each worker evaluates redundant gradients, such that the accurate gradients may be derived even in the presence of adversarial nodes. The simulation results show that DRACO is more than 3 times faster in achieving 90\\% test accuracy for gradient computations on MNIST dataset as compared to the geometric median method.\n\\begin{table}\n\\caption{\\small Comparison of BGW, LCC, Harmonic coding schemes for Gradient-type computations.}\n\\label{table:lccharmonic}\n\\centering\n\\begin{tabular}{|>{\\centering\\arraybackslash}m{2.0cm}|c|c|c|}\n\\hline\n\\rowcolor{mygray}\n & \\textbf{Sharmir} & \\textbf{LCC} & \\textbf{Harmonic} \\\\ \\hline\n Min. number of workers & $K($deg $g+1)$ & $K$deg $g$+1 & $K($deg $ g-1) +2$ \\\\ \\hline\n \\end{tabular}\n\\end{table}\nAn improvement to the studies of is done in , which proposes LCC to achieve an optimum tradeoff between resiliency against straggling workers, security against malicious workers, and information-theoretic privacy amid colluding workers. In LCC, the dataset of the master is encoded using the Lagrange polynomial to create computational redundancy. Then, the coded data is shared to the workers for computation on the encoded data, as if the coding did not take place. In comparison with the BGW MPC scheme , LCC requires more workers. However, the Lagrange polynomial based encoding leads to a reduction in the amount of randomness required to encode the data, which translates to lower storage and computation costs incurred by each worker. The LCC also outperforms the BGW based polynomial sharing in terms of communication costs, given that the polynomial sharing scheme requires a communication round for each bilinear operations. In addition, LCC is less computationally costly than DRACO , which does not utilize the algebraic structure of the encoded gradients. However, the Lagrange coding only works for computations involving arbitrary multivariate polynomial functions of the input dataset. As an extension, the study of proposes CodedPrivateML which adopts polynomial approximations to handle the non-linearities of the gradient computation when the sigmoid function is involved, such that logistic regression can be conducted on LCC-encoded data while providing information-theoretic privacy for the dataset. Given that the advantages of the LCC is preserved, the experiments conducted on Amazon EC2 clusters validate that the proposed scheme is close to 34 times faster than the BGW based MPC approaches.\nIn light of the growing popularity of machine learning, the study of proposes Harmonic coding for tasks specific to gradient-type computations, e.g., for loss function minimization in distributed model training. Harmonic coding leverages on the structure of the intermediate partial gradients computed to enable the cancellation of redundant results, such that the encoding and decoding process is more efficient. As such, for the same level of privacy constraint, Harmonic coding improves on Shamir's secret sharing scheme and LCC in terms of requiring fewer workers to compute gradient-type functions. This result is further summarized in Table \\ref{table:lccharmonic}, where we present a comparison of the minimum number of workers required for the discussed schemes. Note that $K$ refers to the number of partitions of the input dataset, $g$ refers to the fixed multivariate polynomial , and deg refers to the degree of $g$. Like LCC, Harmonic coding can also be applied universally to any gradient-type function computation. As such, the data encoding can be performed before the computing task is specified, thus further reducing the delay in computation.","id":"8d724f64-c2a5-4e3f-a151-1bdc01b1a822","level":"subsection","origin_cites_number":5,"parent_id":"ee3d99f5-815c-4ba8-b336-968d0def830a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Secure Coding for Distributed Computing"],["subsection","Secure Distributed Computing"]],"subsections":[],"title":"Secure Distributed Computing"},{"cite_extract_rate":0,"cites":[],"content":"Matrix multiplication is a key operation in many popular machine learning algorithms , e.g., principal component analysis , support vector machines , and gradient-based computations. While the reviewed studies in Section~\\ref{subsec:securedist} discuss coded computing for privacy preservation in general operations, the studies to be discussed consider tailored strategies for SDMM.\nIn and , the authors propose the use of staircase codes in place of linear secret sharing codes, e.g., Shamir's codes . As an illustration, we consider that a master encodes its data $A$ with random matrix $R$ into three secret shares before transmitting a share to each of the workers to perform matrix multiplication. When linear secret sharing codes are used, the data and random matrix are not segmented but instead encoded and transmitted as a whole (Table~\\ref{table:staircase}). As such, the master has to wait for two \\textit{full} responses from any two of three workers before being able to decode and derive the desired results. In contrast, when the staircase code is used, the data and random matrices are segmented into sub-shares before transmission to the workers. When sufficient sub-tasks have been completed by the workers, the master can then instruct the workers to cease computation. Clearly, the staircase coding approach reduces the computation cost of workers and communication costs incurred by the master node. Accordingly, the staircase coding approach can outperform the classical secret sharing code in terms of mean waiting time. With 4 workers considered, experiments conducted on the Amazon EC2 clusters show a 59\\% decrease in mean waiting time using staircase codes. \nHowever, and still consider the case of one-sided privacy, i.e., the approach is designed to keep only one of two input matrices involved in SDMM operations private. As such, several studies have shifted the focus towards applying coding techniques to the specific case in which two-sided privacy, i.e., in which both input matrices are private, is ensured. Among the first such study is that of , which applies the aligned secret sharing for two-sided privacy. Specifically, the input matrices are split into submatrices and encoded with random matrices. Then, the undesired terms are aligned such that the server only recovers the desired results and saves on communication costs. This leads to an improved download rate, i.e., the ratio of size of desired result to the amount of information downloaded by a user, over conventional secret sharing schemes. \nFollowing , the study of proposes an inductive approach to find a close-to-optimal partition of the input matrices in consideration of two metrics namely the download rate and the minimum number of required workers. The proposed scheme improves on download rate, number of tolerable colluding servers, and computational complexity as compared to the study of . Inspired by , the polynomial coding scheme is also extended to SDMM operations and specifically convolution tasks in , while preserving two-sided privacy, download rate similar to that of , and further mitigating the straggler effects. For convolution tasks, the authors leverage on the inherent property in which the sums of convolutions of sub-vectors may be used to derive the convolution result. Then, the upper and lower bounds of the recovery threshold is derived to show that an order-optimal recovery threshold is achieved, i.e., it does not scale with number of workers. \nHowever, the key weakness of and as indicated in is that the proposed theoretical results do not clarify the effect of matrix dimensions on the download rate, i.e., the download rates are derived in the case whereby matrices are simply assumed to have large dimensions but without any specifications otherwise. However, the study of found that the results for may be violated in some cases for differing relative dimensions of the input matrices. Under this context, the model proposed in allows the matrix dimensions to be specified, and a new converse bound for the two-sided security SDMM is derived.\nIn general, the encoded results of matrix multiplication are sent to the master, where interpolation is performed to obtain the multiplication results, i.e., coefficients of a polynomial. In , the encoding of the private matrices is such that the coefficients are mainly non-zero. In contrast, the study of propose the Gap Additive Secure Polynomial (GASP) codes such that there are as many zero coefficients as possible. This allows the product to be interpolated and the desired results to be derived with fewer number of evaluations performed, which implies that fewer workers are required to perform the matrix multiplication. To assign the exponents for decodability while having as many zero coefficients as possible, the authors propose the degree table to solve the combinatorial problem. In , the authors further generalize the GASP codes to be applicable to different partitions of input matrices and security parameters. The GASP codes are shown to outperform the approaches in in terms of download rate.\n\\begin{table}[!]\n\\caption{\\small Comparison of linear secret sharing codes and staircase codes in the distribution of computation tasks in a system with three workers .}\n\\label{table:staircase}\n\\centering\n\\scalebox{0.9}{\n\\begin{tabular}{ll ll l}\n\\hline\n\\rowcolor{mygray}\n\\multicolumn{1}{|l|}{} & \\multicolumn{1}{l|}{$S_1$} & \\multicolumn{1}{l|}{$S_2$} & \\multicolumn{1}{l|}{$S_3$} \\\\ \\hline\n\\multicolumn{1}{|l|}{\\begin{tabular}[c]{@{}l@{}}Linear secret\\\\ sharing code\\end{tabular}} & \\multicolumn{1}{l|}{$R$} & \\multicolumn{1}{l|}{$R+A$} & \\multicolumn{1}{l|}{$R+2A$} \\\\ \\hline\n\\multicolumn{1}{|l|}{Staircase code} & \\multicolumn{1}{l|}{\\begin{tabular}[c]{@{}l@{}}$A_1+2A_2+4R_1$\\\\ $R_1+R_2$\\end{tabular}} & \\multicolumn{1}{l|}{\\begin{tabular}[c]{@{}l@{}}$A_1+2A_2+4R_1$\\\\ $R_1+2R_2$\\end{tabular}} & \\multicolumn{1}{l|}{\\begin{tabular}[c]{@{}l@{}}$A_1+2A_2+4R_1$\\\\ $R_1+3R_2$\\end{tabular}} \\\\ \\hline\n\\end{tabular}\n}\n\\end{table}","id":"d523704a-b56b-49c2-97c7-6500404f0ac2","level":"subsection","origin_cites_number":7,"parent_id":"ee3d99f5-815c-4ba8-b336-968d0def830a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Secure Coding for Distributed Computing"],["subsection","Secure Distributed Matrix Multiplication (SDMM)"]],"subsections":[],"title":"Secure Distributed Matrix Multiplication (SDMM)"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we have discussed studies that have adopted a coded computing approach towards ensuring secure distributed computing. Then, we discuss the specific applications of SDMM, which require two-sided privacy. The summary and lessons learned are as follows:\n\\begin{itemize}\n\\item Originally proposed to mitigate stragglers in distributed computing systems, coding approaches have also been extended to ensure information theoretic privacy and for some studies, security against malicious workers . However, given that the coding techniques, e.g., polynomial codes, are designed with the intention to mitigate stragglers, the studies discussed in this section have focused on modifications to the approaches to simultaneously achieve both straggler mitigation and privacy. As was mentioned in this section, the coding approaches to ensure privacy preservation have outperformed conventional secret sharing schemes that mainly focused on data correctness and privacy. \n\\item With the recent popularity of machine learning, several studies have also focused on tailoring the coding approaches to be utilized for various machine learning tasks, e.g., entangled polynomial coding for logistic regression and harmonic coding for gradient based computations . In particular, given the importance of two-sided privacy in SDMM, we have also specifically discussed papers that focus on two-sided privacy separate from papers that only considered one-sided privacy. In general, papers that focus on taking advantage of the algebraic structure of the specific operations , e.g., convolution tasks or gradient computations , for efficient encoding and decoding usually perform better.\n\\item In most of the studies that we have discussed, the focus tends to be on information-theoretic privacy. However, weak security is key and may be explored in future works. Specifically, a system is weakly secure if attackers are unable to learn the sensitive intermediate values without having received a certain number of coded packets . This may be relevant when there exist eavesdroppers with an access to communication links between the master and workers . For example, in the study of , a data shuffling design and redundancy reduction algorithm to assign computing tasks have been proposed to ensure weak security in the system. However, the focus has been on workers, rather than eavesdroppers that may not be involved in the computations.\n\\item Given that the straggler effect is a fundamental concern in distributed computing, most papers have considered workers with heterogeneous computing capabilities. However, as discussed in , the issue of heterogeneous networks in other aspects have been under-explored in the aforementioned studies. For example, the workers may have different levels of reputation . This enables the adoption of context-sensitive solutions, e.g., data security may be guarded against new workers or workers with low reputation, whereas it may not be required for trusted workers. With this, the computation complexity and duration may be reduced in trusted nodes.\n\\end{itemize}","id":"362bc888-78c5-42b5-b72f-1d88622fe4e4","level":"subsection","origin_cites_number":4,"parent_id":"ee3d99f5-815c-4ba8-b336-968d0def830a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Secure Coding for Distributed Computing"],["subsection","Summary and Lessons Learned"]],"subsections":[],"title":"Summary and Lessons Learned"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:applications}\n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{nfv.pdf}\n\\caption{\\small Application of CDC to Network Function Virtualization (NFV) model for uplink channel decoding.}\n\\label{fig:nfv}\n\\end{figure}","id":"825a0578-3fe6-4ffc-8428-dfabbdc2d42a","level":"section","origin_cites_number":0,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","CDC Applications"]],"subsections":["6f2cf0e3-4854-4438-b7c6-37b57226ff9d","f8c9cb00-a005-4862-b9c8-df9e98e5996f","eb23bbfe-351d-4db3-8737-3055dd885cdb"],"title":"CDC Applications"},{"cite_extract_rate":0,"cites":[],"content":"Network Function Virtualization serves as an enabling technology for optimizing the 5G as well as emerging 6G networks , presenting a promising paradigm shift in the telecommunication service provisioning industry. By leveraging on virtualization technologies, NFV simplifies the management and operations of the networking services. In particular, NFV decouples the Network Functions (NFs) such as routing and baseband processing from the physical network equipment on which they operate. The NFs are mapped to the Virtual Network Functions (VNFs) that are supported by the Commercial Off-The-Shelf (COTS) physical resources which provide storage, networking and computational capabilities. As the software component in the network is decoupled from the hardware component by a virtualization layer, the VNFs can be easily implemented over the distributed network locations which have the required hardware resources. Due to the flexibility in the deployment of the VNFs, the NFV brings about significant reduction in operating and capital expenses. Moreover, the development of new networking services is faster and cheaper as the COTS resources can be instantiated easily to provide the required network connection services. Extensive literature has been carried out on the various aspects of NFV such as architectural designs , resource allocation , energy efficiency , performance improvement and security . More details can be found in .\nHowever, one of the limiting factors of the performance of NFV lies in the reliability of the COTS hardware resources . Hardware failure due to several factors such as component malfunctioning, temporary unavailability and malicious attacks affects the implementation of NFV, hindering the provision of services. Apart from the fault-tolerant virtualization strategies that are based on the diversity approach which maps VNFs onto the various virtual machines (VMs) such that the probability of a disruptive failure is minimized , coding approach can also be used to minimize computation latency in NFV. \nIn the study of , the authors consider the highly complex uplink channel decoding of the Cloud Radio Access Network (C-RAN) architecture , which is a key application of NFV. In this system, the users communicate with the cloud via Remote Radio Head (RRH). In order to ensure the reliability of channel decoding, the received data frames by the RRH are encoded by leveraging on their algebraic structures before being distributed to the different VMs as shown in Fig.~\\ref{fig:nfv}. The simulation results show that the coding approach achieves lower probability of error for decoding at the cloud than that of the diversity-based approach. However, there are several assumptions that are made in this proposed scheme. Firstly, the binary symmetric communication channel is assumed between the users and the RRH. By considering other communication channels such as the additive Gaussian noise channels, different coding techniques may be applied. Secondly, this simple framework works well for a network with three processing nodes, but its performance for larger-size networks is not guaranteed. \nConsidering the same issue of uplink channel decoding in the C-RAN architecture, the authors in propose a more generalized coded computation framework that works for any number of servers, random computing runtimes and random packet arrivals by adopting the coding approach proposed in . Given the randomness in the arrivals of data frames transmitted by the users, two queue management policies are considered: (i) per-frame decoding where one frame is decoded at any point of time, and (ii) continuous decoding where the servers start to decode the next packet of data frame upon completion of the first packet. There is an underlying tradeoff between the average decoding latency and the frame unavailability probability which is an indication of the reliability of the decoding process at the servers. The simulation results show that properly designed NFV codes are useful in achieving the desired tradeoffs by optimizing the minimum distance of the codes.\nThe idea of adopting coding approach in NFV is relatively new. Different coding techniques can be explored in the future. Besides, instead of addressing the uplink channel decoding, which is the most computationally-intensive baseband function , other network functions such as routing and security functions can be considered.","id":"6f2cf0e3-4854-4438-b7c6-37b57226ff9d","level":"subsection","origin_cites_number":6,"parent_id":"825a0578-3fe6-4ffc-8428-dfabbdc2d42a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","CDC Applications"],["subsection","Network Function Virtualization (NFV)"]],"subsections":[],"title":"Network Function Virtualization (NFV)"},{"cite_extract_rate":0,"cites":[],"content":"With the enhanced sensing capabilities of end devices, an overwhelming amount of data is produced at the edge of the network today. Traditional schemes of computation offloading to the cloud is thus unsustainable. Moreover, certain edge applications may involve end devices in remote areas that have limited connectivity. This necessitates a paradigm shift towards edge computing, in which computation is performed closer to the edge of the network where data is produced. However, resource-constrained devices may not be able to carry out complex computations individually, especially given the increasing size and complexity of state-of-the-art AI models . As such, one of the enabling technologies of edge computing is cooperative computation in which the available resources of end devices and edge nodes, e.g., road side units in vehicular networking, can be pooled together to execute computation intensive tasks collaboratively . For ease of exposition, we refer to these participating end devices and edge nodes as workers in this section.\nAs the number of devices connected to the network increases, more information needs to be exchanged among the workers, resulting in high communication load. However, the communication bandwidth is fixed and thus the network is unable to handle the high communication load, causing a bottleneck as a result. Moreover, the heterogeneous nature of workers in the edge computing paradigm, e.g., in terms of computational and communication capabilities, can lead to the straggler effects. In face of these challenges, coding techniques can be used.\nIn the study of , the \\emph{coded wireless distributed computing} (CWDC) framework is proposed. The system model consists of multiple devices, i.e., workers, involved in cooperative computation. As an illustration, a worker may have an input, e.g., an image, that has to be processed, e.g., for object recognition. Individually, a worker may not have the storage or computational capabilities to execute the task. Therefore, the inference model may be split and stored on each worker, whereas the cooperative computation of results can be implemented following the MapReduce framework as discussed in Section~\\ref{subsec:mapreduce}. An access point, e.g., a base station or a Wi-Fi router can then be utilized to facilitate the exchange of intermediate results among workers. The proposed framework achieves communication loads that are independent of the size of the network and the storage size of the workers. Moreover, the CWDC framework can be generalized and applied to different types of applications.\nIn practical distributed computing systems, the workers may have heterogeneous computational, communication, and storage capabilities. Based on a similar system model proposed in where the workers communicate with each other via an access point, the study in considers devices with heterogeneous storage capacities over wireless networks. For uplink transmission, the allocation of files is based on the scheme proposed in as previously discussed in Section~\\ref{subsec:files} whereas for downlink transmission, data is encoded at the access point for the reduction of downlink communication load. However, the achievable scheme has only been validated in a small network that consists just three processing nodes. \n\\begin{figure}\n\\centering\n\\includegraphics[width=\\linewidth]{flmodel.pdf}\n\\caption{\\small Illustration of the Coded Federated Learning (CFL) scheme in a FL system that consists of multiple data owners and a FL model owner.}\n\\label{fig:flmodel}\n\\end{figure}\nIn light of the growing popularity of machine learning model training at the edge, the study of considers the distribution of gradient descent computations across workers in the network to train a linear regression model. The proposed heterogeneous coded gradient descent (HCGD) assigns each worker with an optimal load partition, through modelling the computation delay of devices with a shifted exponential distribution. In consideration of data privacy, the authors in propose the Coded Federated Learning (CFL) approach for privacy-preserving linear regression. Federated Learning (FL) is a privacy preserving distributed machine learning paradigm proposed in , in which the sensitive data of data owners are kept locally, whereas only the model parameters are transmitted to the central server for aggregation to update the model. However, FL still suffers from the issues of straggling devices and communication inefficiency . As such, in the proposed CFL scheme, each data owner first generates parity data from its local data for transmission to the central server. At the central server, the gradients are also computed from parity data simultaenously, such that only a subset of gradients from the data owners have to be received for completion of model update. Figure~\\ref{fig:flmodel} illustrates the implementation of the CFL scheme. The CFL approach is also shown to converge almost four times faster than an uncoded approach. However, the communication and computation cost involved in generating and transmitting the parity data has not been well elaborated in the study. Clearly, a major weakness of aforementioned studies of is that they can only be applied to linear regression model training. The study of extends on the aforementioned works with the proposed CodedFedL designed to mitigate straggler mitigation in non-linear regression and classification tasks.\nIn some cases, the resource level of the devices may not be known by the network operator. To enable a dynamic and adaptive coded sub-task allocation for cooperative computation, an Automatic Repeat reQuest (ARQ) mechanism is proposed in the study of , in which devices are allocated with specific levels of packets for computation based on their responsiveness. Specifically, devices that are more responsive are assumed to have more available resources. These devices will thus be assigned with more sub-tasks for computation.\nInstead of reducing communication load, it is also important to consider communication efficiency, i.e., achieved data rates, especially in wireless networks that have limited spectral resources or networks with mutual interference among users. In order to improve spectral efficiency, the co-channel communication model is proposed in the study of which consists of two stages, i.e., the uplink multiple access stage and the downlink broadcasting stage. The communication model turns out to be equivalent to a multiple-input-multiple-output (MIMO) interference channel. Interference alignment has been an effective approach in handling mutual interference among users. The signals are precoded into the same subspaces at the unintended receivers and the desired signals are recovered at the intended receivers by using the decoding matrix. Linear coding scheme is adopted to establish the conditions for interference alignment. A low-rank optimization problem is formulated to minimize the number of channel used subject to the established interference alignment conditions. By solving this optimization problem, the achievable symmetric degree-of-freedom (DoF), which implies the extent to which interference is eliminated, can be maximized. In , an efficient difference-of-convex-functions (DC) algorithm based on a DC representation for the rank function is proposed to solve the low-rank optimization problem. The performance of the DC algorithm is evaluated in two different scenarios by varying: (i) the number of files stored in the devices, and (ii) the number of antennas that are equipped by the devices. In the first scenario, as the number of files stored in the devices increases, the achievable DoF increases. Similarly in the second scenario, as the number of antennas equipped by the devices increases, the achievable DoF increases. Furthermore, the simulation results show that the DC approach achieves higher DoF than the existing benchmark algorithms, e.g., iterative reweighted least squares (IRLS) algorithm and the nuclear norm relaxation approach. However, the proposed scheme is based on a homogeneous network, i.e., the number of files stored is the same across all devices and all devices are equipped with the same number of antennas. As an extension, heterogeneous networks can be considered. \nInstead of communicating with each other via an access point, the devices can communicate directly with each other over wireless interference channels . In particular, the transmission of data in the Shuffle phase operates over wireless interference channels. While the CDC scheme in allows communications based on time-division multiple access (TDMA) scheme which allows each processing node to transmit one coded information packet at any time, the one-shot linear scheme adopted in allows more than one processing node to transmit information simultaneously at any given time slot. The transmitted symbols are linear combinations of coded intermediate results from the processing nodes. The transmitted symbols are broadcasted to the processing nodes, following which the nodes can decode to recover the desired information. The study in characterizes an improved computation-communication tradeoff as compared to the study in . However, the proposed scheme operates under the assumption of perfect channel state information (CSI) where the CSI is available to all processing nodes. As such, the authors in propose superposition coding scheme which has better performance than that of the CDC scheme and the one-shot linear scheme under imperfect CSI condition. The study is extended to account for the presence of stragglers in the networks .\nIn general, the aforementioned studies have considered conventional implementations of the CDC scheme in which each device has to complete the computation task first, before transmitting the intermediate results, e.g., to other devices via the access point. However, given the limited computational capabilities of devices, the latency involved can still be significant. As such, the studies in and consider the batch-processing based coded computing (BPCC) framework which allows each device to return only the partially completed results to a master node in batches, even before the task is fully completed. These partial results are then utilized to generate approximations to the full solution through the singular value decomposition (SVD) approach . This is particularly useful in applications that require fast, but not necessarily optimized, results. The BPCC scheme has its effectiveness validated on the EC2 computing cluster, in which latency is proven to be reduced. Moreover, in consideration of the energy limitations of Unmanned Aerial Vehicles (UAV), the BPCC framework was proposed for UAV-based mobile edge computing to provide energy-efficient computation offloading support on-demand .","id":"f8c9cb00-a005-4862-b9c8-df9e98e5996f","level":"subsection","origin_cites_number":5,"parent_id":"825a0578-3fe6-4ffc-8428-dfabbdc2d42a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","CDC Applications"],["subsection","Edge Computing"]],"subsections":[],"title":"Edge Computing"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we have discussed applications of CDC in NFV and edge computing. The lessons learned are as follows:\n\\begin{itemize}\n\\item The convergence of the recently popular edge computing and machine learning has given rise to edge intelligence, in which the computational capabilities of edge servers and devices have been increasingly leveraged to conduct machine learning model training closer to where the data is produced. While this leads to several benefits, e.g., lower training latency and enhanced privacy preservation, the problem of straggling devices is still a bottleneck. As such, CDC approaches have been increasingly applied in this context.\n\\item One major difference between distributed computing at the edge and at computing clusters is that the end devices and edge servers, i.e., workers, are not specifically dedicated to computations. For example, the workers may only share a fraction of their processing power at time slots in which they are idle. As such, the heterogeneity in computational capabilities of workers may also be greater in this regard. In face of these challenges, optimal load partition and allocation strategies have been explored. Moreover, for the future works, the dynamics of the system may be captured using Deep Reinforcement Learning based resource allocation .\n\\item In edge intelligence, there may be several data owners involved. Moreover, the data owners can be devices with computational constraints. As such, conventional techniques of data encoding before transmission to workers for computation may not be feasible. To meet this challenge, FL has been proposed in the work of , whereas CFL and CodedFedL have been proposed to mitigate the straggler effects in FL. However, the proposed methods are highly restrictive. For example, the CFL can only be applied to linear regression problems, whereas both methods require costly computation and transmission of parity data. Moreover, they have not been implemented on typical end devices to validate the feasibility of the schemes in practical implementation. For the future works, studies on using CDC schemes in edge computing applications may adopt the approach of , in which the schemes are implemented under the context of practical hardware constraints.\n\\end{itemize}","id":"eb23bbfe-351d-4db3-8737-3055dd885cdb","level":"subsection","origin_cites_number":1,"parent_id":"825a0578-3fe6-4ffc-8428-dfabbdc2d42a","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","CDC Applications"],["subsection","Summary and Lessons Learned"]],"subsections":[],"title":"Summary and Lessons Learned"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:challenges}\nThe utilization of CDC schemes to solve the implementation challenges of the distributed computing systems is a relatively new and recent approach. There are still challenges and open issues that have yet to be addressed and this provides opportunities for new research directions in the future. We present major challenges that need to be looked into for effective implementation of CDC schemes.\n\\begin{itemize}\n\\item \\emph{Heterogeneous nodes: }As compared to traditional distributed computing clusters, heterogeneities among computing nodes are much more significant when the computing nodes are connected in the edge computing networks, e.g., smartphones and wearable devices. Many studies, e.g., , , , have considered heterogeneous systems where the processing nodes have different computational, communication and storage capabilities. For example, in , a joint file and function allocation strategy is proposed to assign jobs to the processing nodes such that the communication load is minimized. In , the computation load is allocated based on the capabilities of the processing nodes. However, other aspects of heterogeneity such as the reputation and the willingness to participate of the processing nodes are not taken into account. New allocation strategies need to consider different aspects of heterogeneities of the processing devices so that coding techniques can be implemented effectively and securely. For example, the computation tasks can be allocated to workers with higher reputation which implies higher probability of the workers in completing their allocated tasks.\n\\item \\emph{Encoding and decoding complexities: }The studies that we have discussed in Section~\\ref{sec:comms} and Section~\\ref{sec:stragglers} mainly minimize communication load in Shuffle phase and computation latency in the Map phase respectively. However, the complexities of encoding and decoding are often not evaluated. It is important to ensure low encoding and decoding complexities in order to minimize the overall job execution time. Otherwise, the speedup gain achieved by in specific phases, e.g., communication and computation phases may be offset by the high encoding and decoding complexities. For example, the UberShuffle algorithm incurs high computational overhead under the fast broadcast environment, i.e., networks with large bandwidth, such that it is not feasible for implementation even though it achieves significant shuffling gain. Hence, to better assess the performance of the CDC schemes, the complexities of the encoding and decoding methods have to be evaluated.\n\\item \\emph{Non-static computing nodes: }For the commonly used distributed computing models such as cluster computing, grid computing and cloud computing that are discussed in Section~\\ref{sec:fundamentals}, the computing nodes are static, i.e., the nodes are located at fixed location. For example, the servers are located at specific data centers. Data required for computations is transmitted over the wireless communication channels to the servers. However, as the edge devices, e.g., IoT devices, wearable devices and vehicles, have greater communication and computational capabilities, new distributed computing models such as mobile edge computing , and fog computing have been developed recently. In , the basic CDC scheme is implemented in the context of fog computing. However, the edge devices are usually mobile. The data that is processed by the edge devices depends on the locations that they visit , and hence the master node has no control of the data distribution to the workers. New coding approaches for edge and fog computing which involve moving workers can be proposed in future works. One proposed solution is to allocate Reduce functions based on the data stored at the processing nodes .\n\\item \\emph{Security concerns: }Coding techniques are able to mitigate the straggler effects while preserving privacy as shown in the studies of , and . The proposed secure coding techniques are an extension to the coding techniques that are originally proposed to mitigate the straggler effects. As mentioned in , the tradeoff between resiliency against straggling workers, security against malicious workers and information-theoretic privacy amid colluding workers needs to be carefully managed. In addition, there may exist eavesdroppers which tap on the less secure communication links between the master node and the workers. As such, more research effort can be directed towards developing weakly secure systems which prevents the eavesdroppers from retrieving sensitive information. \n\\item \\emph{Network architectures: }It is important to consider how the computing nodes are connected and communicate with each other for effective implementation of the CDC schemes. For example, in and , the authors introduce a hierarchical structure where the master communicates with multiple submaster and each submaster leads a group of computing nodes. From the studies that we have reviewed, the network architecture is only considered for the implementation of CDC schemes to reduce communication load. However, it is an important consideration when designing secure coding schemes. In practice, it may be safe for computing nodes within a group, e.g., from the same location, to share information freely with each other but not with computing nodes from another group, e.g., from different location. In addition, the communication channels are not perfect, i.e., they may not have perfect CSI or the transmitted information may have missing entries. Some networks may have limited spectral resources or suffer from mutual interference among users . As such, future research can work towards designing coding schemes that can be implemented in practical distributed computing systems. Besides a need to design effective coding schemes, there is also a need to consider the design of low-cost, easily-implementable and scalable network architectures in which the coding schemes can be applied to.\n\\item \\emph{Different computation frameworks: }Currently, most of the studies are based on the MapReduce computing model. Specifically, Coded MapReduce is proposed in by implementing coding techniques in the MapReduce framework. However, there are limitations to MapReduce model that hinder its wide adoption for all types of distributed computation tasks as explained in Section~\\ref{subsec:mapreduce}. In fact, there are other computing models such as Spark, Dryad and CIEL which support iterative algorithms, in which the feasibility of implementing coding techniques has not been explored. As such, the importance of these computing models motivates future directions such as the design of coding schemes that are specific to these computing models in order to solve any distributed computation tasks, e.g., convolution, Fourier transform and non-linear computations. \n\\item \\emph{Coding for both communication reduction and stragglers mitigation: }As we have discussed previously in Section~\\ref{sec:comms} and Section~\\ref{sec:stragglers}, coding techniques are used to either reduce communication load or mitigate the straggler effects. However, coding techniques cannot be applied to solve these implementation challenges simultaneously. As characterized in , there is a tradeoff between communication load and computation latency. However, in an ideal situation, both communication load and computation latency need to be minimized. Thus, it is important to carefully manage the tradeoff to achieve optimal performance of the distributed computing systems. For future works on CDC schemes, there is a need to improve the latency-communication tradeoff curve so that the time taken to execute the allocated tasks can be significantly reduced. In addition, instead of the two-dimensional tradeoff, the three-dimensional tradeoff between computation, communication and storage cost , which is much more challenging to manage, should be considered.\n\\item \\emph{CDC applications: }Given the advantages, CDC schemes are implemented in distributed computing applications such as the NFV and edge computing. Apart from the UAVs, CDC schemes can be extended to edge computing applications in other areas, e.g., vehicular networks, healthcare systems and industrial operations. For the studies that we have reviewed in CDC applications, the main focus lies in the implementation of coding techniques in various applications, without considering privacy and security. Given the importance of secure coding as discussed in Section~\\ref{sec:security}, secure coding techniques need to be considered in the implementation of CDC applications. Besides, application-specific issues need to be addressed. For example, in vehicular networks where the vehicles are constantly moving, the CDC schemes need to be robust to vehicles which do not have consistent access to the wireless communication channels.\n\\end{itemize}\nThe idea of using coding techniques to overcome the challenges in distributed computing systems is relatively new. For effective implementation in practical distributed computing systems, various aspects such as the heterogeneities of the computing nodes and the network architectures are worth the in-depth studies. Some promising research directions presented in this survey serve as useful guidelines and valuable references for future research in CDC.","id":"8c1cb679-0c49-4078-b259-687a2df3303d","level":"section","origin_cites_number":6,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Challenges, Open Issues and Future Works"]],"subsections":[],"title":"Challenges, Open Issues and Future Works"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclusion}\nIn this paper, we provided a tutorial of CDC schemes and a comprehensive survey on the two main lines of CDC works. We first motivated the need for CDC schemes. The current performance of distributed computing systems can be improved using coding schemes. Then, we described the fundamentals and principles of CDC schemes. We also reviewed CDC works which aim to minimize communication costs, mitigate straggler effects as well as enhance privacy and security. In addition, we discussed the implementation of CDC schemes in practical distributed computing applications. Finally, we highlighted the challenges and discussed promising research directions.\n\\bibliographystyle{ieeetr}\n\\bibliography{cdc}\n\\end{document}","id":"5a390405-bf12-44a0-aa67-38cecb4ab9be","level":"section","origin_cites_number":0,"parent_id":"43fb2934-e7bf-4933-946e-becc16dd4285","prefix_titles":[["title","A Survey of Coded Distributed Computing"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ]\n ],\n \"subsections\": [\n \"023f3c5d-7805-45c9-b902-bdd9101e05cb\",\n \"fc3d5a29-85c5-4d14-9457-330462c762f2\",\n \"9ba1f087-de1d-45e2-9661-620859b12846\",\n \"f06fb9ca-5252-413b-b98f-08903f106a9c\",\n \"79250ecd-4bf6-4689-8a12-f801fbeb39c1\",\n \"ee3d99f5-815c-4ba8-b336-968d0def830a\",\n \"825a0578-3fe6-4ffc-8428-dfabbdc2d42a\",\n \"8c1cb679-0c49-4078-b259-687a2df3303d\",\n \"5a390405-bf12-44a0-aa67-38cecb4ab9be\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:intro}\\nIn recent years, distributed computing has used for large-scale computation~ since it offers several advantages over centralized computing. First, distributed computing is able to provide computing services with high reliability and fault-tolerance. In particular, distributed computing systems can efficiently and reliably work even if some of the computing nodes, i.e., computers or workers, fail. Second, distributed computing has high computation speed as the computation load is shared among various computing nodes. Third, distributed computing systems are scalable since computing nodes can easily be added. Fourth, distributed computing is economical by using computing nodes with low-cost hardware. Likewise, distributed computing is adopted in cloud computing and other emerging services. Given the aforementioned advantages, distributed computing has been applied in numerous real-life applications such as telecommunication networks~ (e.g., telephone networks and wireless sensor networks), network applications~ (e.g., world-wide web networks, massively multiplayer online games and virtual reality communities, distributed database management systems, and network file systems), real-time process control~ (e.g., aircraft control systems), and parallel computation~ (e.g., cluster computing, grid computing, and computer graphics).\\nHowever, distributed computing faces serious challenges. Let us consider one of the most common distributed computation frameworks, i.e., MapReduce~. MapReduce is a software framework and programming\\nmodel for processing a computation task across large datasets using a large number of computing nodes, i.e., workers. A set of the computing nodes is referred to as a cluster or a grid. In general, the overall computation task is decomposed into three phases, i.e., the ``Map'' phase, ``Shuffle'' phase, and ``Reduce'' phase. In the Map phase, a master node splits the computation task into multiple subtasks and assigns the subtasks to the computing nodes. The computing nodes compute the subtasks according to the allocated Map functions to generate intermediate results. Then, the intermediate results are exchanged among the computing nodes, namely ``data shuffling'', during the Shuffle phase. In the Reduce phase, the computing nodes use these results to compute the final result in a distributed manner by using their allocated Reduce functions. \\nDistributed computing has two major challenges. First, the computing nodes need to exchange a number of intermediate results over the network with each other in order to calculate the final result; this significantly increases communication overheads and limits the performance of distributed computing applications such as Self-Join, Terasort , and machine learning . For example, for the Hadoop cluster at Facebook, it is observed that on average, the data shuffling phase accounts for 33\\\\% of the overall job execution time~. 65\\\\% and 70\\\\% of the overall job execution time is spent on the Shuffle phase when running a TeraSort and Self-Join application respectively on a heterogeneous Amazon EC2 cluster~. In fact, the communication bottleneck is worse in the trainings of convolutional neural networks (CNNs), e.g., Resnet-50 and AlexNet , which include updates of millions of model parameters. Second, distributed computing is executed by a large number of computing nodes which may have very different computing and networking resources. As a result, there are straggling nodes or stragglers, i.e., computing nodes which run unintentionally slower than others, thereby increasing the overall time needed to complete the computing tasks. To address the straggler effects, traditional approaches such as work exchange~ and naive replication have been adopted for the distributed computing. However, such approaches either introduce redundancy or require coordination among the nodes that significantly increases communication costs and computation loads. This motivates the need for a novel technique that is able to more effectively and completely address the straggler effects and communication load of distributed computing. \\nCoding theoretic techniques, e.g., channel coding such as low-density parity-check (LDPC)~, have been widely used in WiFi and cellular systems to combat the impact of channel noise and impairments. They have also been applied in distributed storage systems and cache networks~ to reduce storage cost and network traffic. The basic principle of the coding theoretic techniques is that redundant information, i.e., redundancy, is introduced in messages/signals before they are transmitted to a receiver. The redundancy is included in the messages in a controller manner such that it can be utilized by the receiver to correct errors caused by the channel noise. Coding theoretic techniques have been recently regarded as promising solutions to cope with the challenges in distributed computing~, . For example, coding theoretic techniques can be used to encode the Map tasks of the computing nodes such that the master node is able to recover the final result from partially finished nodes, thus alleviating the straggler effects~. Another example is that coding theoretic techniques enable coding opportunities across intermediate results of the distributed computation tasks which significantly reduces the communication load by reducing the number and the size of data transmissions among the processing nodes~. The combination of coding techniques and distributed computing is called coded distributed computing (CDC)~. Apart from reducing communication load and alleviating the effects of stragglers, CDC can provide fault-tolerance, preserve privacy~, and improve security~ in distributed computing. As a result, CDC approaches have recently received a lot of attention.\\nCDC schemes can be applied in modern networks such as Network Function Virtualization (NFV) and edge computing. With the data mainly generated by end devices, e.g., Internet of Things (IoT), that have significant sensing as well as computational and storage capabilities, it is natural to perform some computations at the end devices, instead of the cloud which may not be able to handle the large amounts of data generated. As such, edge computing has been proposed as a solution to perform distributed computation tasks. In order to perform complex computations, e.g., the training of deep neural networks that involve a large number of training layers, resource-constrained devices may need to pool their resources to perform their computations collaboratively . This results in high communication costs and computation latency. CDC schemes can be implemented to overcome these challenges. Furthermore, CDC schemes can be implemented on edge computing networks that involve constantly-moving end devices , e.g., vehicles and smartphones, which imposes additional communication constraints.\\nTo the best of our knowledge, although there are several surveys and books related to distributed computing, there is no survey paper on CDC. In particular, large-scale distributed computing and applications are discussed in . Surveys related to distributed computing include grid resource management systems for distributed computing~, resource allocation in high performance distributed computing~, wireless distributed computing~, and wireless grid computing~. This motivates the need for this survey on CDC. In summary, our survey has the following contributions:\\n\\\\begin{itemize}\\n\\\\item We describe the fundamentals of CDC. In particular, we introduce the commonly used distributed computation frameworks for the implementation of coding techniques and algorithms. We then discuss basic CDC schemes.\\n\\\\item We review and discuss a number of CDC schemes to reduce the communication costs for distributed computing. The approaches include file allocation, coded shuffling design, and function allocation. We further analyze and compare the advantages and disadvantages of the CDC schemes. \\n\\\\item We review, discuss, and analyze CDC schemes which mitigate the straggler effects of distributed computing. The approaches include computation load allocations, approximate coding, and exploitation of stragglers. \\n\\\\item We review and present CDC schemes that can improve the privacy and security in distributed computing.\\n\\\\item We analyze and provide insights into the existing approaches and solutions in the CDC literature.\\n\\\\item We present and discuss applications of CDC in modern networks such as NFV and edge computing. \\n\\\\item We highlight challenges and discuss promising research directions related to CDC.\\n\\\\end{itemize}\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{structure.pdf}\\n\\\\caption{\\\\small Structure of survey.}\\n\\\\label{fig:structure}\\n\\\\end{figure}\\nFor the reader's convenience, we classify the related CDC studies according to the challenges that need to be handled. In particular, the issues are communication costs, straggler effects, and security. As such, readers who are interested in or working on related issues will benefit greatly from our insightful reviews and in-depth discussions of existing approaches, remaining/open problems, and potential solutions. The rest of this paper is organized as follows. Section~\\\\ref{sec:fundamentals} introduces the fundamentals of CDC. Section~\\\\ref{sec:schemes} presents basic CDC schemes. Section~\\\\ref{sec:comms} reviews CDC approaches that have been proposed to reduce communication costs. Section~\\\\ref{sec:stragglers} discusses CDC approaches that have been proposed to mitigate straggler effects. Section~\\\\ref{sec:security} presents CDC approaches that have been proposed to enhance privacy and security in distributed computing.~Section~\\\\ref{sec:applications} discusses applications of CDC. Section~\\\\ref{sec:challenges} highlights important challenges and promising research directions. The structure of the survey is presented in Figure~\\\\ref{fig:structure}. Section~\\\\ref{sec:conclusion} concludes the paper. A list of abbreviations commonly used in this paper is given in Table~\\\\ref{tab:table_abb}.\\n\\\\begin{table}[h!]\\n\\\\scriptsize\\n \\\\caption{\\\\small List of common abbreviations used in this paper.}\\n \\\\label{tab:table_abb}\\n \\\\centering\\n \\\\begin{tabularx}{8.7cm}{|Sl|X|}\\n \\\\hline\\n \\\\rowcolor{mygray}\\n \\\\textbf{Abbreviation} & \\\\textbf{Description} \\\\\\\\ \\\\hline\\nARIMA & Auto Regressive Integrated Moving Average \\\\\\\\ \\\\hline\\nBCC & Batch Coupon's Collector\\\\\\\\ \\\\hline\\nBGC & Bernoulli Gradient Code\\\\\\\\ \\\\hline\\nBGW & Ben-Or, Goldwasser, and Wigderson \\\\\\\\ \\\\hline\\nBPCC & Batch-Processing Based Coded Computing\\\\\\\\ \\\\hline\\nC3P & Coded Cooperative Computation Protocol\\\\\\\\ \\\\hline\\nCDC & Coded Distributed Computing\\\\\\\\ \\\\hline\\nCNN & Convolutional Neural Network\\\\\\\\ \\\\hline\\nCPGC & Coded Partial Gradient Computation\\\\\\\\ \\\\hline\\nDAG & Directed Acyclic Graphs\\\\\\\\ \\\\hline\\nDNN & Deep Neural Network \\\\\\\\ \\\\hline\\nFRC & Fractional Repetition Coding\\\\\\\\ \\\\hline\\nHCMM & Heterogeneous Coded Matrix Multiplication\\\\\\\\ \\\\hline\\nIoT & Internet of Things\\\\\\\\ \\\\hline\\nLCC & Lagrange Coded Computing\\\\\\\\ \\\\hline\\nLDPC & Low-Density Parity-Check\\\\\\\\ \\\\hline\\nLT & Luby Transform\\\\\\\\ \\\\hline\\nMDS & Maximum Distance Separable\\\\\\\\ \\\\hline\\nMMC & Multi-Message Communication\\\\\\\\ \\\\hline\\nMPC & Multi-Party Computation \\\\\\\\ \\\\hline\\nNFV & Network Function Virtualization \\\\\\\\ \\\\hline\\nPCR & Polynomially Coded Regression\\\\\\\\ \\\\hline\\nPDA & Placement Delivery Array \\\\\\\\ \\\\hline\\nSDMM & Secure Distributed Matrix Multiplication\\\\\\\\ \\\\hline\\nSGC & Stochastic Gradient Coding \\\\\\\\ \\\\hline\\nSGD & Stochastic Gradient Descent\\\\\\\\ \\\\hline\\nSVD & Singular Vector Decomposition\\\\\\\\ \\\\hline\\nUAV & Unmanned Aerial Vehicle\\\\\\\\ \\\\hline\\n\\\\end{tabularx}\\n\\\\end{table}\",\n \"id\": \"023f3c5d-7805-45c9-b902-bdd9101e05cb\",\n \"level\": \"section\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:fundamentals}\\nDistributed computing has been an important solution to large-scale, complex computation problems, which involves massive amounts of data. Various distributed computing models, e.g., cluster computing , grid computing and cloud computing , have been developed to perform the distributed computation tasks while providing high quality of services (QoS) to the users. Among the distributed computing models, cloud computing is gaining much popularity recently as it eliminates the need for users to purchase expensive hardware and software resources since the users only need to pay for the cloud services based on their usage needs in an on-demand basis. A comparison between cluster, grid and cloud computing models is summarized in Table~\\\\ref{tab:compare}. \\nDistributed computing has been widely implemented in a variety of applications, e.g., sensor networks, healthcare applications, the development of smart cities, automated manufacturing processes and vehicular applications. In order to improve the performance of the distributed computing systems, various aspects such as resource allocation strategies , task allocation strategies , scheduling algorithms , incentive mechanisms , energy efficiency , network security and the performance modelling of the distributed computing systems have been extensively studied in the literature. However, the performance of the distributed computing systems is still limited by the high communication costs and straggler effects which lead to a longer time needed to execute the computation tasks. As a result, recent research has focused on coding techniques to overcome these implementation challenges of the distributed computing systems, the aims of which are to minimize the communication load as well as to mitigate the straggler effects.\\nIn this section, we discuss commonly used distributed computation frameworks for the implementation of coding techniques and algorithms. Note that while the different computation frameworks are useful for different computing applications, we focus specifically on the MapReduce framework as the majority of the research works on CDC schemes are based on the MapReduce computation framework. We also introduce the two main lines of works in CDC, i.e., to reduce communication load and to mitigate the straggler effects, which aim to solve the challenges in distributed computing.\\n\\\\begin{table*}[t]\\n\\\\caption{\\\\small Comparison between cluster, grid and cloud computing models . }\\n\\\\label{tab:compare}\\n\\\\centering\\n\\\\begin{tabular}{|l |c| c| c|}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n\\\\multicolumn{1}{c|}{\\\\textbf{Feature}} & \\\\textbf{Cluster} & \\\\textbf{Grid} & \\\\textbf{Cloud} \\\\\\\\ \\\\hline\\nSize & Small to medium & Large & Small to large\\\\\\\\ \\\\hline\\nNetwork type & Private, LAN & Private, WAN & Public, WAN\\\\\\\\ \\\\hline\\nJob management and scheduling & Centralized & Decentralized & Both\\\\\\\\ \\\\hline\\nCoupling & Tight & Loose/tight & Loose \\\\\\\\ \\\\hline\\nResource reservation & Pre-reserved & Pre-reserved & On-demand\\\\\\\\ \\\\hline\\nService-level agreement (SLA) constraint & Strict & High & High\\\\\\\\ \\\\hline\\nResource support & Homogeneous and heterogeneous (GPU) & Heterogeneous & Heterogenous\\\\\\\\ \\\\hline\\nVirtualization & Semi-virtualized & Semi-virtualized & Completely virtualized\\\\\\\\ \\\\hline\\nSecurity type & Medium & High & Low \\\\\\\\ \\\\hline\\nService-oriented architecture and heterogeneity support & Not supported & Supported & Supported\\\\\\\\ \\\\hline\\nUser interface & Single system image & Diverse and dynamic & Single system image \\\\\\\\ \\\\hline\\nInitial infrastructure cost & Very high & High & Low \\\\\\\\ \\\\hline\\nSelf service and elasticisty & No & No & Yes\\\\\\\\ \\\\hline\\nAdministrative domain & Single & Multi & Both\\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\end{table*}\",\n \"id\": \"fc3d5a29-85c5-4d14-9457-330462c762f2\",\n \"level\": \"section\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Fundamentals of Coded Distributed Computing\"\n ]\n ],\n \"subsections\": [\n \"9766cc05-52f0-40d7-8bb2-882e1f23113b\",\n \"e5ec24a4-049a-49ee-b4fa-e11576083168\"\n ],\n \"title\": \"Fundamentals of Coded Distributed Computing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:mapreduce}\\nWhile the distributed computation frameworks have moved beyond a simple MapReduce framework, the majority of the studies on CDC have focused on the MapReduce framework. \\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{conventional.pdf}\\n\\\\caption{\\\\small Illustration of conventional MapReduce framework. The intermediate output pairs are represented by (key,frequency)[book number] and the output pairs are represented by (key,frequency).}\\n\\\\label{fig:conventional}\\n\\\\end{figure}\\nMapReduce is a software framework and programming model that runs on a large cluster of commodity machines for the processing of large-scale datasets in a distributed computing environment. The cluster of computers is modelled as a master-worker system which consists of a single master node and multiple workers to store and analyzes massive amount of unstructured data. Due to its scalability and its ability to tolerate machines' failure , the MapReduce framework is commonly used in a wide range of applications , e.g., the analysis of web access logs, the clustering of documents, the construction of web-link graph that matches the all source URLs to a target URL and the development of machine learning algorithms. Generally, the MapReduce computation framework involves the processing of a large input file to generate multiple output pairs of which each pair consists of a key and a corresponding value. Figure~\\\\ref{fig:conventional} demonstrates the implementation of the conventional MapReduce framework to determine the frequency of occurrence of 4 specific words in the books, where the 4 processing nodes, i.e., the workers are to compute the 4 output pairs. There are three important phases in the MapReduce computation framework:\\n\\\\begin{enumerate}\\n\\\\item In the \\\\textbf{\\\\emph{Map}} phase, there are two stages, namely the allocation of Map tasks and the execution of Map tasks. Generally, a Map task is a function that generates a key-value output pair based on the allocated subfiles. Firstly, as seen in Fig.~\\\\ref{fig:conventional}, the master node splits the input file into 8 subfiles of smaller sizes and allocates the subfiles to the 4 workers. Secondly, each worker produces 4 intermediate key-value pairs for each allocated subfile using the map functions. Since the workers are allocated 2 subfiles each, each worker generates 8 computed intermediate results.\\n\\\\item In the \\\\textbf{\\\\emph{Shuffle}} phase, the workers exchange their computed intermediate results to obtain the required intermediate results for the computation of the Reduce functions. In particular, in each time slot, one of the workers creates a message that contains information of the intermediate output pairs from the Map phase and transmits the message to other workers. The shuffling process continues until all workers have received the required intermediate output pairs for the Reduce phase. \\n\\\\item In the \\\\textbf{\\\\emph{Reduce}} phase, the workers aggregate the 8 intermediate key-value pairs obtained from the Shuffle phase and compute the final result which is a smaller set of key-value pairs using the reduce functions. In particular, the reduction tasks are evenly distributed among the workers. Each reduce function is responsible for the evaluation of a key. For example, node 1 in Fig.~\\\\ref{fig:conventional} is responsible for the evaluation of ``Tree''. Therefore, the total number of reduce functions needed equals the total number of keys of the output, i.e., 4 reduce functions are needed to compute 4 output pairs.\\n\\\\end{enumerate}\\nApart from the MapReduce framework, there are other distributed computation frameworks that provide support for the processing of large-scale datasets such as:\\n\\\\begin{itemize}\\n\\\\item \\\\emph{Spark : }It supports applications that need to reuse a working dataset across the multiple parallel processes. These applications cannot be expressed as efficiently as acyclic data flows which are required in popular computation frameworks such as MapReduce. There are two use cases for the implementation of Spark computation framework: (i) iterative machine learning algorithms which operate on the same dataset repeatedly, and (ii) iterative data analysis tools, where different users query for a subset of data from the same dataset.\\n\\\\item \\\\emph{Dryad : }By allowing the developers to construct their own communication graphs and the subroutines at the vertices through simple, high-level programming language, Dryad executes large-scale data-intensive computations over clusters consisting multiple computers. It does not require the developers to express their code in Map, Shuffle and Reduce phases in order to adopt the MapReduce framework for computations. Besides, the Dryad execution engine, which is based on the constructed data flow graph, takes care of the implementation issues of the distributed computation tasks such as the scheduling of tasks, allocation of resources and the recovery from communication and computation failures.\\n\\\\item \\\\emph{CIEL : }The main characteristic of the CIEL computation framework is that it allows data-dependent data flows where the directed acyclic graphs (DAG) are built dynamically based on the execution of previous computations, rather than a statistically predetermined DAG. Instead of maximizing throughput, CIEL aims to minimize latency of individual tasks, which is very useful for the implementation of iterative algorithms, where latency grows significantly as the number of iteration increases.\\n\\\\end{itemize}\\nA comparison between the distributed computation frameworks is presented in Table~\\\\ref{tab:frameworks}.\\n\\\\begin{table}[t]\\n\\\\caption{\\\\small Comparison between distributed computation frameworks . }\\n\\\\label{tab:frameworks}\\n\\\\centering\\n\\\\begin{tabular}{|m{2.8cm} | >{\\\\centering\\\\arraybackslash}m{1.5cm} |>{\\\\centering\\\\arraybackslash}m{1.5cm} |>{\\\\centering\\\\arraybackslash}m{1.3cm} |}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n\\\\multicolumn{1}{c|}{\\\\textbf{Feature}} & \\\\textbf{MapReduce } & \\\\textbf{Dyrad } & \\\\textbf{CIEL } \\\\\\\\ \\\\hline\\nDynamic Control Flow & No & No & Yes \\\\\\\\ \\\\hline\\nTask Dependencies & Fixed (2-stage) & Fixed (DAG) & Dynamic \\\\\\\\ \\\\hline\\nFault Tolerant & Transparent & Transparent & Transparent \\\\\\\\ \\\\hline\\nData Locality & Yes & Yes & Yes \\\\\\\\ \\\\hline\\nTransparent Scaling & Yes & Yes & Yes \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"9766cc05-52f0-40d7-8bb2-882e1f23113b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"fc3d5a29-85c5-4d14-9457-330462c762f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Fundamentals of Coded Distributed Computing\"\n ],\n [\n \"subsection\",\n \"Coded Distributed Computation Frameworks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Coded Distributed Computation Frameworks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"There are two main lines of works in CDC. Firstly, the CDC schemes are implemented to minimize the \\\\emph{communication load} in distributed computing systems. Secondly, the CDC schemes aim to mitigate the \\\\emph{straggler effects} which cause a delay in the computation of the distributed tasks. For each of the objectives, we discuss the importance of solving these issues to improve the performance of the distributed computing systems. Then, we briefly discuss the existing solutions that have been proposed in the literature to meet these objectives. However, the current existing solutions do not adopt coding approaches. Different from the existing solutions, the CDC schemes are able to meet these objectives by introducing coded redundancy. In fact, the CDC schemes outperform the replication methods, e.g., naive replication and fork-join model , which introduce redundancy without coding techniques, in terms of time taken to execute the tasks.\",\n \"id\": \"e5ec24a4-049a-49ee-b4fa-e11576083168\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"fc3d5a29-85c5-4d14-9457-330462c762f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Fundamentals of Coded Distributed Computing\"\n ],\n [\n \"subsection\",\n \"Objectives of CDC Schemes\"\n ]\n ],\n \"subsections\": [\n \"109ae9de-6803-47dc-a8d7-f4437916dc6b\",\n \"274c5abc-70eb-46be-bb88-eff1a21724b3\"\n ],\n \"title\": \"Objectives of CDC Schemes\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"} \\\\label{subsub:comms} Among the three phases of the MapReduce computation framework, the Shuffle phase dominates the time required to complete the computation tasks since multiple communications between the processing nodes are needed to exchange their intermediate results. For the Hadoop cluster at Facebook, it is observed that on average, the data shuffling phase accounts for 33\\\\% of the overall job execution time . In fact, the data shuffling phase is more time consuming when running on heterogeneous clusters with diverse computational, communication and storage capabilities. When running TeraSort , which is a conventional distributed sorting algorithm for large amount of data, and Self-Join applications on heterogeneous Amazon EC2 clusters, 65\\\\% and 70\\\\% of the overall job execution time is spent on the Shuffle phase respectively . The data shuffling process is also an important step in implementing distributed learning algorithms. In particular, to train machine learning models with distributed algorithms, it is common to shuffle the data randomly and run the algorithms iteratively for a number of times such that the processing nodes compute a different subset of the data at each iteration until there is a convergence . \\\\emph{For the logistic regression application which requires at least 100 iteration to converge, 42\\\\% of the iteration time is spent on communication }. For each time the data shuffling process is performed, the entire training dataset is communicated over the network, resulting in high communication costs which limit the performance of the distributed computing systems. \\nSince the performance of the data shuffling process has a significant impact on the overall performance of the distributed computing systems, it has been extensively studied in the literature, e.g., . Various data shuffling strategies are proposed to achieve different objectives such as minimizing the job execution time, maximizing the utilization of resources and accommodating interactive workloads. While the overlap between the map computations and the shuffle communication helps to reduce the latency of the distributed computation tasks , the computing nodes require large storage capacities for buffering. An efficient and adaptive data shuffling strategy is proposed in the study of to manage the tradeoff between the accumulation of the shuffle blocks and the minimization of the utilization of memory space to reduce the overall job execution time and improve the scalability of the distributed computing systems. In , the authors propose a virtual data shuffling strategy which reduces storage space and traffic load in the network by delaying the actual movement of the data until it is needed to complete the computations in the Reduce phase. \\nTo improve the performance of the data shuffling process, task scheduling algorithms such as Quincy scheduler , Hadoop Fair Scheduler and delay scheduling algorithm are also designed to allocate tasks to the workers. In the design of optimal task scheduling and task selection algorithms, the communication load can be minimized through various approaches such as by optimizing the placement of computation tasks, distributing the computing resources fairly to the nodes and maximizing the resource utilization of the systems. \\nSince the task scheduling schemes are not the focus of this survey, we refer interested readers to the study of and the references therein for more detailed information on the scheduling techniques.\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{uncoded.pdf}\\n\\\\caption{\\\\small Illustration of naive replication MapReduce framework.}\\n\\\\label{fig:uncoded}\\n\\\\end{figure}\\nOne of the ways to reduce communication load in the shuffling phase is by repeating the computation tasks . Figure~\\\\ref{fig:uncoded} illustrates the implementation of the MapReduce framework by using the naive replication method in which 4 processing nodes are required to compute 4 output pairs, which is the same computation task illustrated in Fig.~\\\\ref{fig:conventional}. By simply replicating the Map tasks where each worker is required to compute more intermediate output pairs, i.e., 16 in Fig.~\\\\ref{fig:conventional} instead of 8 intermediate output pairs in Fig.~\\\\ref{fig:uncoded}, the communication load is reduced as fewer intermediate output pairs are communicated. For example, in the conventional MapReduce framework in Fig.~\\\\ref{fig:conventional}, node 1 needs to obtain 6 intermediate output pairs from other workers whereas in the naive replication scheme in Fig.~\\\\ref{fig:uncoded}, node 1 only needs to obtain 4 intermediate output pairs from other workers. \\nHowever, the aforementioned non-coding methods have limits to which the communication load in the data shuffling phase can be minimized. Given that the naive replication method reduces communication load in the data shuffling phase by introducing redundancy to the systems (which is also discussed in Section~\\\\ref{subsec:cdccommunication}), coding techniques can be used to introduce redundancy to further minimize communication load, which will be discussed in-depth in Section~\\\\ref{sec:comms}.\",\n \"id\": \"109ae9de-6803-47dc-a8d7-f4437916dc6b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"e5ec24a4-049a-49ee-b4fa-e11576083168\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Fundamentals of Coded Distributed Computing\"\n ],\n [\n \"subsection\",\n \"Objectives of CDC Schemes\"\n ],\n [\n \"subsubsection\",\n \"Communication Load\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Communication Load\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}In distributed computing systems, the processing nodes may have heterogeneous computing capabilities, i.e., different processing speeds. As such, another line of work in the CDC literature is to solve the bottleneck that results from a variation in time taken to complete the allocated tasks. In distributed computing systems, there are \\\\emph{stragglers}, which are the processing nodes that run unexpectedly slower than the average or nodes that may be disconnected from the network due to several factors such as insufficient power, contention of shared resources, imbalance work allocation and network congestion . As a result, the overall time needed to execute the tasks is determined by the slowest processing node. We briefly discuss the existing approaches (which are summarized in Table~\\\\ref{tab:priorstragglers}) to handle the straggler effects as follows:\\n\\\\begin{table*}[t]\\n\\\\caption{\\\\small Approaches to mitigate the straggler effects.} \\n\\\\label{tab:priorstragglers}\\n\\\\centering\\n\\\\begin{tabular}{|>{\\\\centering\\\\arraybackslash}m{2.6cm} |m{6.5cm} |m{6.5cm}|}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n\\\\textbf{Approach} & \\\\multicolumn{1}{c|}{\\\\textbf{Key Ideas}} & \\\\multicolumn{1}{c|}{\\\\textbf{Shortcomings}} \\\\\\\\ \\\\hline\\nStragglers Detection & Detects the straggling nodes, determines the cause of delay and implements targeted solutions & Difficult to identify the cause of delays \\\\\\\\ \\\\hline\\nWork Stealing & Reallocates the remaining computation tasks from the slower workers to the faster workers & The capabilities of the slower workers are not maximized\\\\\\\\ \\\\hline\\nWork Exchange & Reallocates the computation tasks every time a worker completes its tasks & Incurs high communication costs due to the feedback from the workers to the master node as well as the reallocation of data \\\\\\\\ \\\\hline\\nNaive Replication & Introduces redundancy where each computation subtask is performed by more than one processing node & Incurs high communication costs and computation load\\\\\\\\ \\\\hline\\nCoded Redundancy & Uses coding techniques to introduce redundancy such that the master node can recover the final result from any decodable set of workers & Still incurs high communication costs and computation load, but lower than that of naive replication \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\end{table*}\\n\\\\begin{itemize}\\n\\\\item \\\\emph{Stragglers Detection: }The most direct approach to mitigate the straggler effects is to detect the stragglers and act on them early in their lifetime. For example, Mantri detects the stragglers by identifying the tasks that are processed at a rate slower than the average. The system determines the cause of the delay and implements targeted solutions to mitigate the stragglers. The solutions include restarting the tasks allocated to the stragglers at other processing nodes, optimally allocating the tasks based on network resources as well as protecting against interim data loss by replicating the outputs of valuable tasks. \\n\\\\item \\\\emph{Work Stealing : }The basic idea of the work stealing algorithm is to allow the faster processing nodes to take over the remaining computation tasks from the slower processing nodes so that the overall job execution time is minimized. By adopting this approach, the faster processing nodes operate continuously while leaving the slower processing nodes idle after their jobs are taken over by the faster processing nodes till the end of the computation session. \\n\\\\item \\\\emph{Work Exchange : } By leveraging on the information of computational heterogeneity in the system, the master node first allocates the tasks to the workers based on their computational capabilities. Upon receiving the first computed result from any of the workers, the master node pauses the computation process and redistribute the remaining incomplete work to be computed by the workers. The process is performed for a number of iterations until all work is done. Since the workers need to inform the master node of the amount of work done at each time that the computation process is paused, additional communication costs are incurred. The higher communication costs are also a result of the reallocation of data to the workers. \\n\\\\item \\\\emph{Naive Replication: }One of the solutions to handle stragglers in the distributed computing systems is by introducing redundancy to minimize computation latency. The computation task is replicated and executed over multiple processing nodes. Since all processing nodes are working on the same computation task, the time required to complete the computation task is determined by the fastest processing node. The partial computations of the remaining processing nodes are discarded. Experiments on Google Trace data have shown the effectiveness of the use of redundancy in minimizing computation latency by eliminating the need for the computed results by the stragglers. However, the introduction of redundancy comes at the expense of higher cost such as high communication bandwidth and high computation load . Various redundancy strategies have been analyzed to derive the limiting distribution of the state of the systems . Although the introduction of redundancy helps to reduce latency, the performance varies under different settings. In fact, in some settings, it is optimal to not use any redundancy strategy. Looking into this, the work in presents the optimal redundant-requesting policies under diverse settings.\\n\\\\end{itemize}\\nSimilar to the existing methods to reduce communication costs as discussed in the previous section (Section~\\\\ref{subsub:comms}), the existing methods to mitigate the straggler effects do not adopt coding approaches. Coding techniques can also be used to introduce redundancy into the systems to mitigate the straggler effects. The authors in investigate the tradeoff between latency and cost for both replication-redundancy systems and coded-redundancy systems. Coded-redundancy systems outperform the replication-redundancy systems in both latency and cost. In other words, by using coding techniques, the latency and cost incurred are lower than that of naive replication. The use of coding techniques to mitigate the straggler effects is discussed in more detail in Section~\\\\ref{subsec:codingtech} and Section~\\\\ref{sec:stragglers}.\\nTo better understand the proposed CDC schemes, some of the commonly used performance metrics of the distributed computing systems are defined as follows:\\n\\\\begin{enumerate}\\n\\\\item \\\\emph{Storage Space }is defined as the total number of files stored across $K$ processing nodes, normalized by the total number of subfiles $N$ .\\n\\\\item \\\\emph{Computation load }is represented by $r$, where $1\\\\leq r \\\\leq K$, is defined as the total number of Map functions computed across $K$ processing nodes, normalized by the total number of subfiles $N$ . In particular, when $r=1$, it means that each Map function is only computed by a single processing node. When $r=2$, it means that each Map function is computed by two processing nodes on average. \\n\\\\item \\\\emph{Communication load }is represented by $L$, where $0\\\\leq L \\\\leq 1$, is defined as the total number of bits communicated by the $K$ processing nodes in the Shuffle phase, normalized by the total number of subfiles $N$ .\\n\\\\end{enumerate}\\nGiven that coding techniques are able to solve the aforementioned implementation challenges of the distributed computing systems, we review various proposed CDC schemes, which is the main focus of this paper. In the following section, we present a tutorial of the simple CDC schemes along these two lines of works, i.e., minimizing communication load and mitigating the straggler effects, of which is useful to better understand the related works discussed in Section~\\\\ref{sec:comms}, \\\\ref{sec:stragglers} and \\\\ref{sec:security}.\",\n \"id\": \"274c5abc-70eb-46be-bb88-eff1a21724b3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"e5ec24a4-049a-49ee-b4fa-e11576083168\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Fundamentals of Coded Distributed Computing\"\n ],\n [\n \"subsection\",\n \"Objectives of CDC Schemes\"\n ],\n [\n \"subsubsection\",\n \"Straggler Effects\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Straggler Effects\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:schemes}\\nRecently, coding techniques have become a popular approach to solve the challenges of the distributed computing systems. As mentioned previously, there are two main lines of work in CDC: (i) to reduce the communication costs and (ii) to mitigate the straggler effects. In this section, we introduce the two basic CDC schemes which are the first works that show the effectiveness of using coding techniques to solve these two challenges separately. Then, we discuss a unified CDC scheme that characterizes the tradeoff between computation latency and communication load.\",\n \"id\": \"9ba1f087-de1d-45e2-9661-620859b12846\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ]\n ],\n \"subsections\": [\n \"e9e65886-933d-4d54-aa7f-a7c03b8cc0f9\",\n \"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee\",\n \"2dffdb1e-356b-4152-ad17-f178b093d653\"\n ],\n \"title\": \"Coded Distributed Computing (CDC) Schemes\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:cdccommunication}\\nIn the conventional MapReduce computation framework as shown in Fig.~\\\\ref{fig:conventional}, after the split of the input file into multiple subfiles, each subfile is mapped to only one of the processing nodes, i.e., the workers. The naive replication scheme, i.e., uncoded data shuffling scheme which relaxes this restriction, can reduce the communication costs of the system by allowing each subfile to be replicated and mapped to more than one processing nodes. In the example illustrated in Fig.~\\\\ref{fig:uncoded}, each subfile is repeated twice. Hence, as compared to the conventional MapReduce framework, each processing node has more Map tasks to perform in the naive replication scheme. However, by simple replication, the communication load in the Shuffle phase decreases and this gain is known as the \\\\emph{repetition gain}. Specifically, the communication load for uncoded schemes which include both the conventional MapReduce framework and the naive replication scheme, is denoted as follows :\\n\\\\begin{equation}\\nL_{uncoded}(r)=1-\\\\frac{r}{K},\\n\\\\label{eqn:uncoded}\\n\\\\end{equation}\\nwhere $K$ is the number of processing nodes in the network. Based on Equation (\\\\ref{eqn:uncoded}), the communication loads achieved by the conventional MapReduce framework in Fig.~\\\\ref{fig:conventional} and the naive replication scheme in Fig.~\\\\ref{fig:uncoded} are $\\\\frac{3}{4}$ where $r=1$ and $\\\\frac{1}{2}$ where $r=2$ respectively.\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{coded.pdf}\\n\\\\caption{\\\\small Illustration of Coded MapReduce framework.}\\n\\\\label{fig:coded}\\n\\\\end{figure}\\nTo further reduce the communication load, i.e., to increase the repetition gain, the Coded MapReduce computation framework is proposed in where the Map tasks are carefully distributed among the processing nodes and the messages are encoded for transmission in the Shuffle phase by using coding theory. Figure~\\\\ref{fig:coded} illustrates the Coded MapReduce framework with 4 processing nodes to determine the 4 output pairs. After the Map phase, the processing node multicasts a bit-wise XOR of two computed intermediate pairs, denoted by $\\\\oplus$, satisfying the requirements of two other processing nodes simultaneously. For example, node 1 multicasts a bit-wise XOR of ``Bear'' and ``Fork'' to both nodes 2 and 3, which involves the transmission of only one packet of information, instead of two packets if the information is sent separately to the nodes in a unicast manner. Since the intermediate output pairs are now coded, there is an additional step of decoding before the reduce functions are applied. Given the coded ``BearFork'' information, node 2 is able to decode and recover the required ``Bear'' information by cancelling the ``Fork'' information since node 2 has also computed the same ``Fork'' information. Similarly, node 3 can recover ``Fork'' information by cancelling the ``Bear'' information. The simulation results in show that the Coded MapReduce reduces the communication load by 66\\\\% and 50\\\\% as compared to the conventional MapReduce framework and the naive replication scheme respectively.\\nSince the use of coding techniques reduces both the latency and cost of the distributed computing systems , a more generalized framework known as the Coded Distributed Computing (CDC) scheme is introduced in . \\nThe study of presents the fundamental inverse relationship between computation load and communication load. Specifically, the communication load in the Shuffle phase can be reduced by a factor $r$ by increasing the computation load in the Map phase by the same factor $r$, as shown in Fig.~\\\\ref{fig:tradeoff}. The communication load achieved by the CDC framework, $L_{coded}$, is given as follows:\\n\\\\begin{equation}\\nL_{coded}(r)=\\\\frac{1}{r}\\\\left(1-\\\\frac{r}{K}\\\\right).\\n\\\\label{eqn:coded}\\n\\\\end{equation}\\nNote that the information-theoretic lower bound derived on the minimum communication load $L^*(r)$ equals $L_{coded}(r)$ of the CDC framework. As such, the optimal tradeoff between the computation load and communication load is characterized as follows :\\n\\\\begin{equation}\\nL^*(r)=L_{coded}(r)=\\\\frac{1}{r}\\\\left(1-\\\\frac{r}{K}\\\\right), r \\\\in\\\\mathcal{K}.\\n\\\\end{equation}\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{tradeoff.pdf}\\n\\\\caption{\\\\small Comparison of communication load between the CDC scheme and the uncoded scheme .}\\n\\\\label{fig:tradeoff}\\n\\\\end{figure}\\nFrom Equation (\\\\ref{eqn:uncoded}) which applies to the uncoded computation schemes, the communication load $L$ decreases linearly as the computation load $r$ increases. However, when the number of processing nodes $K$ becomes large, there is no significant impact of increasing computation load on the communication load. On the other hand, for the proposed CDC framework, the communication load is inversely proportional to the computation load (Equation (\\\\ref{eqn:coded})). Even when $K$ becomes large, the increase in computation load still significantly reduces the communication load. \\nSince the proposed CDC framework can be applied to any distributed computation framework with an underlying MapReduce structure, the performance of the CDC framework on TeraSort is evaluated. Experimental results on Amazon EC2 clusters show that the Coded TeraSort scheme which is a coded distributed sorting algorithm, achieves a reduction in the overall job execution time by factors of 2.16 and 3.39 with 16 processing nodes and computation loads of $r=3$ and $r=5$, respectively as compared to the uncoded TeraSort scheme.\\nPreviously in , computation load is linearly dependent on the number of replicated Map tasks, i.e., load redundancy, as each processing node is assumed to compute all intermediate values for all subfiles allocated in its memory. However, the processing nodes can be selective in choosing the intermediate values to compute. As such, the load redundancy is no longer a direct measure of computation load. In other words, the storage constraints do not necessarily imply computation constraints. Building on the work in , the authors in propose an alternative tradeoff between computation load and communication load under a predefined storage constraint. In fact, the computation load is quadratic in terms of load redundancy. By taking load redundancy into consideration, an alternative computation-communication tradeoff curve is derived. In particular, this alternative tradeoff curve is especially relevant to the processing nodes that do have the sufficient resources or time to perform computations for all the allocated subfiles. Given that the processing nodes can only perform a limited amount of computations below the computation load threshold, the alternative tradeoff curve proposed in accurately defines the amount of communication load needed for the distributed computation tasks.\\nSince the processing nodes are not required to compute all intermediate results that can be obtained from their locally stored data, the storage capabilities of the processing nodes should be considered in the CDC computation framework in . The study of characterizes the tradeoff between storage, computation and communication where the minimum communication load is determined given the storage and computational capabilities of the processing nodes. In particular, the optimal computation curve is obtained by characterizing the optimal storage-communication tradeoff given the minimum computation load. As a result, the triangles between the optimal communication curve and the optimal computation curve reflect the pareto-optimal surface of all achievable storage-computation-communication triples. However, as the number of processing nodes in the system increases, the number of input files required increases exponentially, resulting in an increase in the number of transmissions needed and hence high communication costs. As such, it is important to reduce the number of input files, which is discussed in Section~\\\\ref{subsubsec:packet} later. \\n\\\\begin{table*}[!ht]\\n\\\\caption{\\\\small Coding techniques to mitigate the straggler effects.} \\n\\\\label{tab:stragglers}\\n\\\\centering\\n\\\\begin{tabular}{|m{2.0cm} |>{\\\\centering\\\\arraybackslash}m{1.5cm} |>{\\\\centering\\\\arraybackslash}m{3.3cm} |m{8cm}|}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n\\\\multicolumn{1}{|c|}{\\\\textbf{Problems}} & \\\\textbf{Ref.} & \\\\textbf{Coding Schemes} & \\\\multicolumn{1}{c|}{\\\\textbf{Key Ideas}} \\\\\\\\ \\\\hline\\n\\\\multirow{8}{=}{Matrix-Vector} & & MDS Codes & Reduce the computation latency as the master node is able to recover the final result without waiting for the slowest processing node \\\\\\\\\\n\\\\cline{2-4}\\n & & LT Codes & Exploit the rateless property to generate unlimited number of encoded symbol from a finite set of source symbols\\\\\\\\\\n\\\\cline{2-4}\\n & & Short-Dot Codes & Reduce the length of dot-products computed at the processing nodes by introducing sparsity to the encoded matrices\\\\\\\\\\n\\\\cline{2-4}\\n & & s-diagonal codes & Exploit the diagonal structure of the matrices to achieve both optimal recovery threshold and optimal computation load\\\\\\\\\\n\\\\hline\\n\\\\multirow{16}{=}{Matrix-Matrix} & & Product Codes & Instead of encoding the matrices along one dimension as in the MDS-coded schemes, the matrices are encoded with MDS codes along both dimensions, i.e. row and column \\\\\\\\\\n\\\\cline{2-4}\\n & & Polynomial Codes & - Design the algebraic structure of the encoded matrices such that the MDS structure is found in both the encoded matrices and the intermediate computations\\\\\\\\\\n & & & - Reconstruct the final results by solving the polynomial interpolation problem\\\\\\\\\\n\\\\cline{2-4}\\n & & MatDot Codes & Achieve lower recovery threshold than Polynomial Codes at the expense of higher communication costs by computing only the relevant cross-products\\\\\\\\\\n\\\\cline{2-4}\\n & & PolyDot Codes & Characterize the tradeoff between recovery threshold and communication costs where Polynomial Codes and MatDot Codes are the two extreme ends on this tradeoff curve\\\\\\\\\\n\\\\cline{2-4}\\n & & Sparse Codes & Exploit to sparsity in both input and output matrices to reduce computation load, while achieving near optimal threshold\\\\\\\\\\n\\\\hline\\n\\\\multirow{15}{=}{Gradient Descent} & & Fractional Repetition Coding & - Workers are divided into multiple groups and data is divided among the workers in each group \\\\\\\\\\n & & & - Each partition of data is performed by more than one worker\\\\\\\\\\n\\\\cline{2-4}\\n & & Cyclic Repetition Coding & Data is allocated based on a cyclic assignment strategy\\\\\\\\\\n\\\\cline{2-4}\\n & & Cyclic MDS Codes & The entries to the columns of the encoding matrix are cyclic shifts of the entries to the first column \\\\\\\\\\n\\\\cline{2-4}\\n & & Reed-Solomon Codes & Use balanced mask matrix and chooses appropriate codewords from the RS codes to construct the encoding matrix \\\\\\\\\\n\\\\cline{2-4}\\n & & Batch Coupons Collector & - Divide the data into multiple batches which are allocated randomly to the workers for computations\\\\\\\\\\n & & & - communication costs are significantly reduced as there is no need for communication between the workers and for feedback from the master node to the workers\\\\\\\\\\n\\\\cline{2-4}\\n & & Polynomially Coded Regression & Encode the data batches directly instead of the computed intermediate results\\\\\\\\\\n\\\\hline\\nConvolution & & Coded Convolution & Split both vectors into multiple parts of specified length and encodes one of the vectors with MDS codes \\\\\\\\\\n\\\\hline\\nFourier Transform & & Coded Fourier Transform & Leverage on recursive structure and the linearity of the discrete Fourier Transform operations\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table*}\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{codedcomputation.pdf}\\n\\\\caption{\\\\small Illustration of coded computation with 3 workers. The master node is able to recover the final result upon receiving the computed results from any 2 workers.}\\n\\\\label{fig:codedcomputation}\\n\\\\end{figure}\",\n \"id\": \"e9e65886-933d-4d54-aa7f-a7c03b8cc0f9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"9ba1f087-de1d-45e2-9661-620859b12846\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ],\n [\n \"subsection\",\n \"CDC to Minimize Communication Load\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"CDC to Minimize Communication Load\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:codingtech}\\nApart from reducing communication load in the Shuffle phase of the MapReduce framework, coding techniques can also be used to alleviate the straggler effects. Since matrix multiplication is one of the most basic linear operations used in distributed computing systems, a coded computation framework is proposed in to minimize computation latency of distributed matrix multiplication tasks. The coded computation framework uses erasure codes to generate redundant intermediate computations. In particular, \\nthe master node encodes the equal-sized data blocks, i.e., submatrices and distributes them to the workers to compute the local functions. Upon completion, the workers transmit the computed results to the master node. The master node can recover the final result by using the decoding functions once the local computations from any of the decodable sets are completed. As seen in Fig.~\\\\ref{fig:codedcomputation}, the master node can recover the final result upon receiving the computed results from any 2 workers, instead of all 3 workers. As such, the total computation is not determined by the slowest straggler, but by the time when the master node receives computed results from some decodable set of indices. In this work , the authors explore the effectiveness of encoding the submatrices by using maximum distance separable (MDS) codes to mitigate the effects of stragglers.\\nConsidering $K$ workers and a shifted-exponential distribution for the job execution time of the distributed algorithm, the simulation results show that the optimal repetition-coded distributed algorithm achieves a lower average job execution time when the straggling parameter is smaller than one, i.e., $\\\\mu<1$ but is still slower than the optimal MDS-coded distributed algorithm by a factor of $\\\\Theta (\\\\log K)$. However, the storage cost of the coded distributed algorithm is higher than that of the uncoded distributed algorithm as more data is required to be stored at the workers' sites for the coded distributed algorithm. The proposed algorithm is tested on an Amazon EC2 cluster and is compared against various parallel matrix multiplication algorithms, e.g., block matrix multiplication, column-partition matrix multiplication and row-partition matrix multiplication. The simulation results show that the proposed algorithm in performs better where the coded matrix multiplication achieves $40\\\\%$ and $39.5\\\\%$ reduction in average job execution time on clusters of m1-small and c1-medium instances with 10 workers each respectively as compared to the best of the three uncoded distributed algorithms. \\nAlthough the MDS codes proposed in is able to mitigate the straggler effects, it cannot be generalized to all types of computation tasks. In order to mitigate the straggler effects of different distributed computation tasks, the coding techniques can be designed by exploiting the algebraic structures of the specific operations. An important performance metric that is introduced in the proposed CDC schemes is the recovery threshold, which refers to the worst-case required number of workers the master needs to wait to recover the final result for job completion . The smaller the recovery threshold, the shorter the computation latency. The objective is to reduce the recovery threshold so that the final result can be recovered by waiting for a smaller number of workers, thus contributing to a reduction in computation latency. Here, we discuss the coding techniques for various types of computation tasks, namely (i) matrix-vector multiplications, (ii) matrix-matrix multiplications, (iii) gradient descent, (iv) convolution and Fourier transform. Table~\\\\ref{tab:stragglers} summarizes the coding techniques designed for different distributed computation tasks.\",\n \"id\": \"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"9ba1f087-de1d-45e2-9661-620859b12846\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ],\n [\n \"subsection\",\n \"CDC to Mitigate the Straggler Effects\"\n ]\n ],\n \"subsections\": [\n \"8d5bcfc5-309d-4be0-81a5-8bdc9010e406\",\n \"795c9420-d6c1-4f57-b9b5-1d340af6f0f3\",\n \"c272fc00-2e47-4ea2-92bd-9be23a93633e\",\n \"bc8c4799-cb61-4844-9fd9-6fd743d773dc\"\n ],\n \"title\": \"CDC to Mitigate the Straggler Effects\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Distributed matrix-vector multiplications are the building blocks of linear transformation computations which are an important step in machine learning and signal processing applications. In particular, the computation of linear transformation on high-dimensional vectors is required for popular dimensionality reduction techniques such as Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA) .\\nInstead of using MDS codes that is proposed in , the authors in propose the use of Luby Transform (LT) codes to mitigate the straggler effects in distributed matrix-vector multiplications problems. Different from the works in and which use LT codes in fixed-rate settings, the rateless property of the LT codes can be exploited to generate unlimited number of encoded symbol from a finite set of source symbols. There are several advantages of using rateless codes: (i) near-ideal load balancing, (ii) negligible redundant computation, (iii) maximum straggler tolerance, and (iv) low decoding complexity. To further reduce the latency for practical implementations, blockwise communication can be used to transmit the submatrix-vector products. Instead of transmitting each encoded row-vector product separately to the master node, the workers are allowed to transmit the computed results in blocks where each block comprises a few row-vector products, reducing the number of communication rounds needed and hence minimizing the time needed to complete the computation tasks. \\n The authors in propose Short-Dot codes to perform the computation of linear transforms reliably and efficiently under the presence of straggling nodes. Specifically, the processing nodes compute shorter dot products by imposing sparsity on the encoded submatrices. However, there is tradeoff between the optimal threshold recovery and the length of the dot-products where the master node needs to wait for computed results from more processing nodes when the length of the dot-products is shorter. The experimental results on the classification of hand-written digits of MNIST show that the Short-Dot codes achieve 32\\\\% faster expected computation time than the MDS codes .\\nAlthough the Short-Dot codes can offer lower recovery threshold, the greater length of the dot-products means greater computation load for the processing nodes. With this concern, s-diagonal codes are proposed to achieve both optimal recovery threshold and optimal computation load by exploiting the diagonal structure of the encoding matrix. The computation time can be further reduced by using a low-complexity hybrid decoding algorithm which combines the peeling decoding algorithm and Gaussian elimination techniques.\",\n \"id\": \"8d5bcfc5-309d-4be0-81a5-8bdc9010e406\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ],\n [\n \"subsection\",\n \"CDC to Mitigate the Straggler Effects\"\n ],\n [\n \"subsubsection\",\n \"Matrix-vector multiplications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Matrix-vector multiplications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"For large-scale distributed matrix-matrix multiplications, the coded computation schemes based on MDS codes are no longer suitable as the encoding and decoding processes scale with system size. Besides, the size of one of the matrices is assumed to be small enough in order to allow individual workers to perform the computations , restricting the implementation of MDS codes in large-scale multiplications. Hence, for large-scale problems, coded schemes not only need to achieve low computation time, but also require efficient encoding and decoding algorithms in order to minimize the overall job execution time. To deal with the straggler effects in high-dimensional distributed matrix multiplications, four types of coded computation schemes are proposed:\\n\\\\begin{itemize}\\n\\\\item \\\\emph{Product codes : }Product codes are implemented by building a larger code upon smaller MDS codes. Instead of encoding computations along only one dimension in MDS-coded schemes, the product codes encode computations along both dimensions, i.e., rows and columns of the matrices. When the number of backup workers increases sub-linearly with the number of subtasks, the product-coded schemes outperform the MDS-coded schemes in terms of average computation time and decoding time. In the linear regime, the one-dimensional decoding of the MDS-coded schemes is sufficient to recover the missing entries of the computation results. By allowing each row and column of the MDS constituent codes to have different code rates , the average computation time can be further reduced, contributing to a decrease in the overall job execution time. Product codes can also be used to solve higher-dimensional linear operations such as tensor operations by exploiting the tensor-structured encoding matrix . To reduce the decoding time of the product codes, efficient decoding algorithms such as Reed-Solomon codes and LDPC codes can be explored.\\n\\\\item \\\\emph{Polynomial codes : }The key advantage of the polynomial codes in solving large-scale matrix multiplication problems is that they provide a lower bound to the optimal recovery threshold. For polynomial codes, the recovery threshold does not scale with the number of workers involved where as for MDS codes and the product codes, the recovery thresholds scale proportionally with the number of workers. By taking advantage of the algebraic structure of the polynomial codes, the master node can recover the final result by using polynomial interpolation algorithms, e.g., the Reed-Solomon codes, to decode the computation results from the workers. In addition to the optimal recovery threshold, the polynomial-coded schemes achieve minimum possible computation latency and communication load for distributed matrix multiplications. However, as the number of workers increases, the encoding and decoding costs are much higher than that of the product codes. Furthermore, by implementing Reed-Solomon codes, there is a limit to the number of workers that can be handled, which is not useful for practical implementations where the systems may involve up to thousands of nodes. As an extension to the polynomial codes proposed in , the entangled polynomial code that is proposed in achieves a lower recovery threshold which is only half of that achieved by the PolyDot codes , to be discussed later. Different from the polynomial codes which only allow column-wise partitioning of the matrices, the entangled polynomial codes allow arbitrary partitioning of the input matrices and evaluate only a subspace of bilinear functions such that unnecessary multiplications are avoided. The issue of numerical stability has also received attention to ensure the scalability of the polynomial-coded schemes . \\n\\\\item \\\\emph{PolyDot codes : }PolyDot codes characterize the tradeoff between the recovery threshold and communication costs where the polynomial codes and the MatDot codes are special instances of this coding framework by considering two extreme ends of this tradeoff: minimizing either recovery threshold or communication costs. In particular, the MatDot codes achieve lower recovery threshold than the polynomial codes at the expense of much higher communication costs. This is achieved by only computing the relevant cross-products of the submatrices. Building on the work of PolyDot codes , the Generalized PolyDot codes are used to compute matrix-vector multiplications and achieve the same recovery threshold as the entangled polynomial codes . More importantly, the Generalized PolyDot codes can be extended for the training of large deep neural networks (DNNs), which consists of multiple non-linear layers. \\n\\\\item \\\\emph{Sparse codes : }Although the polynomial codes achieve optimal recovery threshold, the computation loads of the workers increase due to the increased density of the input matrix, resulting in an increase in the overall job execution time which is not desirable. By exploiting sparsity, i.e., the number of zero entries of the encoded matrix, not only the recovery threshold is kept low, but the computation loads of the workers also decrease while maintaining a nearly linear decoding time , jointly contributing to the shorter overall job execution time. The basic idea of the algorithm proposed in is to allow the master node to find the linear combination of row vectors such that only the particular relevant sub-blocks are recovered. Then, the entire block of matrix can be recovered by aggregating the partial recovery of sub-blocks. Simulation results show that the sparse codes require an overall shortest time to complete the job as compared to other computation schemes, e.g., uncoded scheme, product codes , polynomial codes , sparse MDS codes and LT codes . A further analysis of the different components of subtasks, i.e., communication time, computation time and decoding time, shows that the sparse codes require much shorter time to decode, thus contributing significantly to the shorter overall job execution time.\\n\\\\end{itemize}\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{frc.pdf}\\n\\\\caption{\\\\small Fractional repetition coding with 6 workers and 2 stragglers.}\\n\\\\label{fig:frc}\\n\\\\end{figure}\",\n \"id\": \"795c9420-d6c1-4f57-b9b5-1d340af6f0f3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ],\n [\n \"subsection\",\n \"CDC to Mitigate the Straggler Effects\"\n ],\n [\n \"subsubsection\",\n \"Matrix-matrix multiplications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Matrix-matrix multiplications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Apart from matrices, coding techniques can be applied to recover batch gradients of any loss function of the distributed gradient descent tasks. In , the authors have introduced the idea of gradient coding which is useful to mitigate stragglers that may slow down the computation tasks. Two gradient coding schemes are proposed, namely (i) fractional repetition coding (FRC) and (ii) cyclic repetition coding. In the FRC scheme, the workers are first divided into several groups. In each group, the data is equally divided and allocated to the workers. As a result, all groups of workers are replica of each other as shown in Fig.~\\\\ref{fig:frc}. Upon completing their subtasks, the workers in each group transmit the sum of partial gradients to the master node. In the cyclic repetition coding scheme, the data partitions are allocated to the workers based on a cyclic assignment strategy. The partial gradients computed by each worker are encoded by linearly combining them, of which is transmitted as a single coded message to the master node. By applying the gradient coding schemes, the distributed computation tasks do not suffer from delays incurred by the straggling nodes as the master node is able to recover the final result with the results from the non-straggling nodes. Other coding theories such as the cyclic MDS codes and the Reed-Solomon codes can be used to compute exact gradients of the distributed gradient descent problems. \\nTo efficiently mitigate the straggler effects in distributed gradient descent algorithms, Batch Coupon's Collector (BCC) scheme is proposed in . In BCC, there are two important steps, namely (i) batching and (ii) coupon collecting. In batching, the training set is partitioned into batches which are distributed to the workers randomly whereas in coupon collecting, the master node collects the computed results from the workers until the results from all batches of data are received. This decentralized BCC scheme does not require any communication between the workers nodes and each worker is allocated data batches independently of other workers. As a result, it is easy to implement the BCC scheme in practical scenarios. Another important advantage of the BCC scheme is its universality. Different from other coding schemes that are designed to guarantee their robustness to a fixed number of straggling nodes, the BCC scheme does not require any prior knowledge about the straggling nodes which is more practical as it is difficult to estimate the number of straggling nodes present in the clusters. Furthermore, the BCC scheme can be easily extended to solve gradient descent problems over heterogeneous clusters where the workers have different computational and communication capabilities. The simulation results show that the BCC scheme speeds up the overall job execution time by up to 85.4\\\\% and 69.9\\\\% over the uncoded scheme and the cyclic repetition coding scheme respectively.\\nThe gradient coding schemes proposed in illustrate the tradeoff between computation load and straggler tolerance. However, in non-linear learning tasks, communication costs dominate the overall job execution time as the number of iterations increases. As such, to generalize the coding schemes in , the authors in incorporate communication costs into their framework and present a fundamental tradeoff between the three parameters, namely computation load, straggler tolerance and communication costs. In particular, for a fixed computation load, the communication costs can be reduced by waiting for more workers.\\nInstead of encoding the partial gradients computed based on uncoded data as seen in the studies of , coding techniques can be applied directly to the data batches to reduce the straggler effects and the overall job execution time. Considering the gradient computations for least-square regression problems, the polynomially coded regression (PCR) scheme exploits the underlying algebraic property to generate coded submatrices such that they are linear combinations of the uncoded input matrices. The master node can evaluate the final gradient by interpolating the polynomials from the computed partial gradients by the workers. Compared to the gradient coding schemes proposed in , the simulation results show that the PCR scheme achieves much lower recovery threshold and hence shorter computation and communication time, resulting in shorter time needed for overall job execution.\",\n \"id\": \"c272fc00-2e47-4ea2-92bd-9be23a93633e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ],\n [\n \"subsection\",\n \"CDC to Mitigate the Straggler Effects\"\n ],\n [\n \"subsubsection\",\n \"Gradient Descent\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Gradient Descent\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The polynomial codes proposed in both the studies of and can be extended to the applications of distributed coded convolution based on the coded convolution scheme proposed in . The work in explores the use of MDS codes to encode the pre-specified vectors such that fast convolution is performed under deadline constraint. In addition to that, MDS codes can be used to mitigate the straggler effects in widely-implemented distributed discrete Fourier Transform operations which are used in many applications such as machine learning algorithms and signal processing frameworks.\",\n \"id\": \"bc8c4799-cb61-4844-9fd9-6fd743d773dc\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"9f1f1f0b-c93c-42b6-815f-1a27a54d04ee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ],\n [\n \"subsection\",\n \"CDC to Mitigate the Straggler Effects\"\n ],\n [\n \"subsubsection\",\n \"Convolution and Fourier transform\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Convolution and Fourier transform\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:unified} \\nGiven the aforementioned coding schemes of and , we can observe that coding techniques are used to speed up distributed computing applications with two different approaches. On one hand, the authors in propose the ``Minimum Bandwidth Code\\\" that minimizes the communication load by repeating the computation tasks in the Map phase to introduce multicasting opportunities in the Shuffle phase. On the other hand, the authors in propose the ``Minimum Latency Code\\\" which minimizes the computation latency by encoding the Map tasks such that the master node is able to recover the final result without waiting for the straggling processing nodes.\\nInspired by the aforementioned approaches, a unified coded scheme that characterizes the tradeoff between computation latency and communication load is proposed in given the computation load. The coding schemes in and are considered to be the extreme cases of this unified scheme. The unified coded scheme exploits the advantages of the two coding approaches by applying the MDS codes to the Map tasks and replicating the encoded Map tasks. Specifically, the unified coded scheme first encodes the rows of the matrix, following which the coded rows of the matrix are replicated and stored at the processing nodes in a specific pattern. Then, the processing nodes perform the computation until a certain number of the fastest processing nodes complete their tasks. To reduce communication load in the Shuffle phase, coded multitasking is used to exchange the intermediate results that are needed to recover the final results in the Reduce phase. An improvement to the latency-communication tradeoff presented in the unified coded scheme is proposed in by leveraging on the redundancy created by the repetition code. By increasing the redundancy rate of the repetition code, both the communication load in the Shuffle phase and the computation latency in the Map phase can be simultaneously improved, thus contributing to a improved latency-communication tradeoff.\\nThe aforementioned initial works of coding schemes have shown their effectiveness in minimizing communication costs and alleviating the straggler effects. In the following sections, we review related works that leverage on coding techniques to address the implementation challenges of the distributed computing systems.\",\n \"id\": \"2dffdb1e-356b-4152-ad17-f178b093d653\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"9ba1f087-de1d-45e2-9661-620859b12846\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Coded Distributed Computing (CDC) Schemes\"\n ],\n [\n \"subsection\",\n \"Unified CDC Scheme\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Unified CDC Scheme\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:comms}\\nWith more computing nodes that are equipped with greater capabilities to collect and process data, massive amounts of data are generated for the computations of user-defined computation tasks. Since the computations are scaled out across a large number of distributed computing nodes, large number of intermediate results need to be exchanged between the computing nodes in the Shuffle phase of the MapReduce framework to complete the computation tasks, resulting in significant data movement. Oftentimes, for the training of a model with distributed learning algorithms, data is shuffled at each iteration, contributing to high communication costs, which is a bottleneck of the distributed computing systems. As a result, there is a need to reduce the communication costs in order to speed up the distributed computation tasks. In this section, we present four approaches to reduce communication costs: \\n\\\\begin{itemize}\\n\\\\item \\\\emph{File Allocation: }In this approach, the studies aim to design an optimal file allocation strategy that considers the heterogeneities of the processing nodes' capabilities in the systems, maximizing data locality or reducing subpacketization level, which refers to the number of subfiles generated , . These different approaches work towards reducing the communication load in the distributed computing systems.\\n\\\\item \\\\emph{Coded Shuffling Design: }Since data shuffling phase incurs a large proportion of the communication costs, data is encoded before it is transmitted so that the communication load can be minimized. Apart from combining coding with different techniques, e.g., compression and randomization techniques to improve the performance of the shuffling phase, the coding techniques are also designed to solve different computation problems, e.g., distributed graph computation problems , and multistage MapReduce computations .\\n\\\\item \\\\emph{Consideration of Underlying Network Architecture: }Generally, the communications between the workers as well as between the workers and the master node are affected by the way that they are connected to each other. For example, server-rack architecture (Fig.~\\\\ref{fig:server}) is one of the most commonly used methods to connect the various servers. By taking the underlying architecture into consideration, the effectiveness of the coding implementation in reducing communication costs can be greatly improved.\\n\\\\item \\\\emph{Function Allocation: }Similar to the allocation of files, the studies apply this approach on heterogeneous systems. In addition, some studies consider a cascaded system where each Reduce function is allowed to be computed at multiple processing nodes. In some cases where the data is randomly stored at the processing nodes, e.g., when the processing nodes are constantly moving, an optimal function allocation strategy is useful in reducing the number of broadcast transmissions and thus minimizing the communication load. \\n\\\\end{itemize}\",\n \"id\": \"f06fb9ca-5252-413b-b98f-08903f106a9c\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ]\n ],\n \"subsections\": [\n \"a30f2979-178f-4897-ab80-a68948bc9c5f\",\n \"08339afb-55ae-4aba-a167-25ccaafe6521\",\n \"2cb9e913-abbc-43af-837d-bdfad67a0786\",\n \"1109cc8b-6b58-4ee5-ba62-ca162ae902db\",\n \"71e25826-5a72-4095-8ca3-d26170c96db8\"\n ],\n \"title\": \"Minimization of Communication Load\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:files}\\nThe design of file allocation at each processing node is one of the major steps for the implementation of CDC scheme. There are a few approaches to an optimal file allocation strategy: (i) considering heterogeneous systems, (ii) maximizing data locality, and (iii) reducing subpacketization level.\",\n \"id\": \"a30f2979-178f-4897-ab80-a68948bc9c5f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"f06fb9ca-5252-413b-b98f-08903f106a9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"File Allocation\"\n ]\n ],\n \"subsections\": [\n \"97abadeb-ed93-480e-acfa-dc9e87a6bd77\",\n \"c8d7498f-a451-4053-980e-15ca23008379\",\n \"73c478e0-3e12-49d2-8043-871c00017ef4\"\n ],\n \"title\": \"File Allocation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}As discussed in Section~\\\\ref{subsec:cdccommunication}, although the CDC scheme proposed in carefully allocates the subfiles to the processing nodes in order to introduce coded multitasking opportunities, it considers a homogeneous system which may not be useful for practical implementation. In order to appropriately allocate the files to the distributed computing nodes, heterogeneous systems where the processing nodes have diverse storage, computational and communication capabilities, should be considered in determining the optimal file allocation strategy and coding scheme that minimize the communication load . \\nBy leveraging on the extra storage capacity of the workers, the communication costs between the master node and the workers in the process of data shuffling are minimized. The reason is that if more data can be stored at the workers, fewer communication rounds are needed for the workers to receive shuffled data from the master node. In the extreme case, if the worker can store the entire dataset, there is no communication needed for the worker to receive shuffled data in any iteration. As a result, there is a tradeoff between the storage capacity of the workers and the communication overhead in the data shuffling process. In the data shuffling process, there are two phases, namely data delivery and storage update. Instead of a random storage placement , a deterministic and systematic storage update strategy creates more coding opportunities in transmitting data to the workers at each iteration, reducing the communication load.\",\n \"id\": \"97abadeb-ed93-480e-acfa-dc9e87a6bd77\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"a30f2979-178f-4897-ab80-a68948bc9c5f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"File Allocation\"\n ],\n [\n \"subsubsection\",\n \"Considering Heterogeneous Systems\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Considering Heterogeneous Systems\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}One of the important factors in determining the optimal file allocation strategy is data locality. Data locality is defined as the percentage of local tasks over the total number of Map tasks, i.e., the fraction of Map tasks that are allocated to the processing nodes having the required data for computations such that no communication is needed to obtain the data. High data locality means that less communication bandwidth is needed for the transmission of subfiles, which is required if the processing node does not have the needed subfiles for the execution of the Map tasks. In order to maximize data locality, the problem of allocation of Map tasks to different processing nodes can be tackled by solving a constrained integer optimization problem .\",\n \"id\": \"c8d7498f-a451-4053-980e-15ca23008379\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"a30f2979-178f-4897-ab80-a68948bc9c5f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"File Allocation\"\n ],\n [\n \"subsubsection\",\n \"Maximizing Data Locality\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Maximizing Data Locality\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"} \\\\label{subsubsec:packet}As the number of processing nodes in the network increases, the input file needs to be split into a large number of subfiles. Specifically, the number of subfiles generated increases exponentially in the number of processing nodes . However, there is a maximum allowable subpacketization level, i.e., number of subfiles, where the dataset can only be partitioned into a limited number of packets, beyond which the communication load increases due to more transmissions required and the unevenly-sized intermediate results. Hence, there are several reasons for the reduction of subpacketization level: (i) to reduce the communication load in the Shuffle phase even when there is a large number of processing nodes, (ii) to reduce the packet overheads which increases with the number of broadcast transmissions, and (iii) to reduce the number of unevenly-mapped outputs which require zero padding. To keep the subpacketization level below the maximum allowable level, Group-based Coded MapReduce allocates the dataset based on the random groupings of the processing nodes and allows the processing nodes to cooperate in the transmission of messages. \\nTo avoid splitting the input file too finely, the authors in use an appropriate resolvable design , which is based on linear error correcting codes, to determine the number of subfiles, the allocation of the subfiles to the processing nodes and the construction of the coded messages in the Shuffle phase. Building on this initial work, the authors in use the resolvable design based scheme to solve the limitation of the compressed CDC scheme that uses both compression and coding techniques. Although the compressed CDC scheme helps to reduce communication load, it requires large number of jobs to be processed simultaneously. Hence, the resolvable design based scheme is used to reduce the number of subfiles generated. Specifically, for each job in the compressed CDC scheme, the single-parity code is used to split the input file and the resolvable design based scheme is used to allocate the subfiles to the processing nodes. By aggregating the underlying functions and applying the resolvable design based scheme, multiple jobs can be processed in parallel while minimizing the execution time in the Shuffle phase, contributing to the reduction of the overall job execution time. Although the number of subfiles or number of jobs generated still increases exponentially with some of the system parameters, e.g., number of computing nodes and number of output functions, the exponent is much smaller when the resolvable design based scheme is implemented.\\nIn addition to the exponential increase in the number of subfiles required, the number of output functions required also increases exponentially as the number of processing nodes in the network increases. There are other methods to reduce the number of subfiles and the number of output functions such as the hypercube computing scheme and the placement delivery array (PDA) . However, most of the CDC schemes consider non-cascaded systems, i.e., each Reduce function is computed at exactly one processing node . In , a cascaded system is considered but only two values for the number of processing nodes that perform each Reduce function are considered. By applying the concept of PDA to the distributed computation framework, the performance of the proposed computation scheme is evaluated for different number of processing nodes that compute the Reduce functions . Although the implementation of these various methods reduces the number of subfiles generated, it may come at the expense of higher communication load , .\",\n \"id\": \"73c478e0-3e12-49d2-8043-871c00017ef4\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"a30f2979-178f-4897-ab80-a68948bc9c5f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"File Allocation\"\n ],\n [\n \"subsubsection\",\n \"Reducing Subpacketization Level\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Reducing Subpacketization Level\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In the design of coded shuffling algorithms, we have classified the approaches into three different categories: (i) compression and randomization, (ii) coding across multiple iterations and (iii) problem-specific coding approaches.\",\n \"id\": \"08339afb-55ae-4aba-a167-25ccaafe6521\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"f06fb9ca-5252-413b-b98f-08903f106a9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"Coded Shuffling Design\"\n ]\n ],\n \"subsections\": [\n \"de2801c0-f5ea-4b6a-9559-1daeed4b4d16\",\n \"47dbc343-daef-4549-a329-038a4aa2efda\",\n \"4b18764b-6d94-4230-8ec2-e099673e37eb\"\n ],\n \"title\": \"Coded Shuffling Design\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To further reduce the communication costs of the distributed computation tasks, the design of the coded data shuffling scheme can incorporate different techniques to create more coded multicasting opportunities. Besides, the coded shuffling schemes are designed to minimize communication costs for different distributed computation problems such as iterative algorithms, graph computations and multistage dataflow problems. \\nThe work in generates replications of the computation tasks in the Map phase in order to reduce communication load in the Shuffle phase by coding and multicasting the intermediate results. To further reduce the communication load, compression and randomization techniques can be applied to the design of the coded shuffling algorithms. \\n\\\\begin{itemize}\\n\\\\item \\\\emph{Compression Techniques: }The compressed CDC computation scheme is proposed in by jointly using two techniques, i.e., compression and coding techniques. Each processing node first computes the allocated Map tasks and generates the intermediate results. By using the compression techniques, several intermediate results of a single computation task are compressed into a single pre-combined value. The communication bandwidth needed to transmit a single pre-combined value is much smaller than that of transmitting several uncombined intermediate values since the size of the pre-combined value equals the size of only one intermediate value. With the pre-combined values from different computation tasks, the processing node codes them for multicasting to other processing nodes simultaneously. There are two advantages to this compressed CDC scheme: (i) the communication load is reduced proportional to the storage capacity of each processing node, and (ii) the communication load does not scale linearly, i.e., slower than linear, with the size of the dataset.\\nIn some cases, e.g., parallel stochastic gradient descent (SGD) algorithms, instead of transmitting intermediate results, computed gradient updates are exchanged among the workers. In such cases, Quantized SGD , a compression technique, can be used to reduce communication bandwidth used during the gradient updates between the processing nodes. In each iteration, the processing nodes are allowed to adjust the number of transmitted bits by quantizing each component to a discrete set of values and encoding these quantized gradients.\\n\\\\item \\\\emph{Randomization Techniques: }Instead of introducing coded multicasting opportunities to reduce the communication load in the Shuffle phase, there are other coding techniques that can be applied to increase efficiency of data shuffling. One of the ways is to perform a semi-random data shuffling and coding scheme based on pliable index coding which introduces randomization in the data shuffling process . There are two important modifications made to the conventional pliable index coding scheme , which is used to minimize the number of broadcast transmissions while satisfying users' demands. Firstly, the correlation of messages between workers is reduced. In order to do so, a message should only be transmitted to a fraction of the workers so that the same message is not held by all workers. As such, the pliable index coding problem is formulated with an objective where the goal is to minimize the number of broadcast transmissions under the constraint of a maximum number of workers that can receive the same message. Secondly, the correlation of messages between iterations is reduced. The reduction of correlation of messages prevents the workers from performing computations on the same dataset after shuffling, which may be redundant. A two-layer hierarchical structure is proposed for data shuffling. In the upper layer, the messages are partitioned into multiple groups of which each group of messages is transmitted to a fraction of workers. In the lower layer, each group of messages and the corresponding allocated workers are formulated as a constrained pliable index coding problem. Randomization occurs in two stages: (i) when the master node selects the messages in each group and transmits them to the workers, and (ii) when the workers discard old messages from their cache. Experimental results \\nshow that the proposed pliable index coding requires only 12\\\\% of broadcast transmissions needed by an uncoded scheme, i.e., random shuffling with replacement scheme.\\n\\\\end{itemize}\",\n \"id\": \"de2801c0-f5ea-4b6a-9559-1daeed4b4d16\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"08339afb-55ae-4aba-a167-25ccaafe6521\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"Coded Shuffling Design\"\n ],\n [\n \"subsubsection\",\n \"Compression and randomization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Compression and randomization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Most works on coded iterative algorithms focus on the optimization of a single computation iteration or the minimization of communication load in a single communication round . However, multiple rounds of communications are generally required to solve distributed iterative problems. In the studies of and , the results of multiple iterations are transmitted in a single round of communication by jointly coding across several iterations of the distributed computation task. By leveraging on the computation and storage redundancy of the workers, the number of communication rounds between the master node and the workers is greatly reduced, resulting in a reduction in the communication costs. However, the computation and storage costs may not be optimal as compared to uncoded computing schemes, e.g., and , which achieve near-optimal computation and storage costs.\",\n \"id\": \"47dbc343-daef-4549-a329-038a4aa2efda\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"08339afb-55ae-4aba-a167-25ccaafe6521\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"Coded Shuffling Design\"\n ],\n [\n \"subsubsection\",\n \"Coding across multiple iterations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Coding across multiple iterations\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Apart from the typical distributed computation problems, the MapReduce framework can also be used to solve distributed graph computation problems, . However, for graph computing systems, the computation at each vertex is a function of the graph structure which means that each computation only needs data from its neighbouring vertices. More specifically, the communication load in the Shuffle phase depends on the connectivity probabilities of vertices in the graph, in which each vertex is only allowed to communicate reliably with a subset of random vertices . As a result, the CDC scheme proposed in (which was previously discussed in Section~\\\\ref{subsec:cdccommunication}) is not applicable to solving the graph computation problems. Looking into this, the authors in propose a coded scheme to solve the problem of random Erdös-Rényi graph while minimizing the communication load in the Shuffle phase. A similar inverse tradeoff curve between the computation load and average communication load is obtained by using coding techniques in solving distributed graph computations. \\nMoreover, many distributed computing applications consists of multiple stages of MapReduce computations. The multistage data flow can be represented by a layered DAG in which the processing nodes, i.e., vertices in a particular computation stage are grouped into a single layer. Each vertex computes a user-defined function, i.e., Map or Reduce computation that transforms the given input files into intermediate results whereas the edges represent data flow between the processing nodes. By exploiting the redundancy of the computing nodes, coding techniques are applied to the processing nodes individually to minimize the communication costs. The proposed work in considers a uniform resource allocation strategy where the computation of each vertex is distributed across all processing nodes. However, the communication load can be further reduced by reducing the number of processing nodes that are used to compute each vertex. Given that fewer processing nodes are used to compute each vertex, the computation load performed by each processing node increases and the processing nodes have more local information, thus reducing the need for communication to obtain the required information. Therefore, a dynamic resource allocation strategy is needed to further minimize the communication load in the multistage MapReduce problems.\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{server.pdf}\\n\\\\caption{\\\\small Server-rack architecture where multiple servers in each rack are connected via a Top of Rack switch and the Root switch connects multiple Top of the Rack switches.}\\n\\\\label{fig:server}\\n\\\\end{figure}\",\n \"id\": \"4b18764b-6d94-4230-8ec2-e099673e37eb\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"08339afb-55ae-4aba-a167-25ccaafe6521\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"Coded Shuffling Design\"\n ],\n [\n \"subsubsection\",\n \"Problem-specific coding approaches\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Problem-specific coding approaches\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Although various techniques can be used jointly with the coding techniques to increase the efficiency of data shuffling phase, it is important to consider the underlying network architecture, i.e., how the servers are connected to each other, in designing the coded shuffling algorithms. \\nIn , Hybrid Coded MapReduce is proposed by considering the server-rack architecture (Fig.~\\\\ref{fig:server}) in the distributed computing systems. There are two types of communications in the Shuffle phase: (i) the cross-rack communication where the data is shuffled across different rack layers, and (ii) the intra-rack communication where the data is shuffled within the rack. In the first stage where cross-rack communication takes place, the Coded MapReduce algorithm (Fig.~\\\\ref{fig:coded}) is used to create multicasting opportunities for the transmission of messages. In the second stage where the intra-rack communication is performed, data is shuffled in a unicast manner where no coding technique is applied. The simulation results show that the cross-rack communication costs incurred by the hybrid scheme are lower than those of both Coded MapReduce and uncoded scheme at the expense of a higher intra-rack communication costs. Although the Coded MapReduce scheme still achieves the lowest total communication costs, the overall communication costs for the hybrid scheme can be further reduced by parallelizing the intra-rack operations to provide a more accurate comparison between the different computation schemes.\\nThe CDC computation scheme proposed in is useful for networks with processing nodes that are closely located with each other and connected via a common communication bus. However, in practical distributed computing networks, it is hard to implement a common-bus topology for the physically separated processing nodes. Hence, to reduce communication load for the distributed computation tasks, it is important to consider a practical data center network topology to reap the coding benefits of the CDC schemes. As such, in , the authors propose a CDC scheme based on a widely-used low-cost network topology which is the t-ary fat-tree topology . It has the characteristics of network symmetry and connections between any two processing nodes in the network. Given a practical network topology design, the proposed topological CDC scheme achieves optimal max-link communication load over all links in the topology of which the optimal tradeoff between the max-link communication load and the computation load is characterized. \\nAlthough the coded shuffling algorithm proposed in reduces communication load of the data shuffling process, there are two main limitations: (i) it assumes a perfect broadcast channel between the master node and the workers, and (ii) the theoretical guarantee of number of broadcast transmissions only holds when the number of data points approaches infinity. To overcome the limitations of the coded shuffling algorithm proposed in , Ubershuffle fills the missing entries in the encoding tables by reallocating the data points between the workers. This reduces the number of transmitted encoded packets, resulting in a reduction of communication load. The performance of the UberShuffle algorithm is evaluated when applied to different problem settings. The experimental results show that in comparison to the coded shuffling algorithm in , the UberShuffle algorithm reduces the shuffling time by up to 47.2\\\\% and 35.7\\\\% when implemented with the distributed SGD algorithm for a low-rank matrix completion problem and the parallel SGD algorithm for a linear regression problem respectively.\",\n \"id\": \"2cb9e913-abbc-43af-837d-bdfad67a0786\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"f06fb9ca-5252-413b-b98f-08903f106a9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"Consideration of Underlying Network Architecture\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Consideration of Underlying Network Architecture\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In most of the studies of CDC computation framework, non-cascaded systems are considered. In other words, each Reduce function can only be computed at exactly one processing node. Considering that the processing nodes have different storage, computational and communication capabilities, a non-cascaded heterogeneous system where each Reduce function is computed exactly once is considered. In this proposed scheme, the processing nodes are allocated with different number of Reduce functions of which the processing nodes with greater storage and computational capabilities are allocated more Reduce functions. This heterogenous Reduce function allocation creates a symmetry among the multicast groups such that each processing node in a group requests the same number of intermediate outputs from the other processing nodes in the same group. The heterogeneous CDC system achieves a lower communication load than an equivalent homogeneous CDC system.\\nHowever, it is desirable for the Reduce functions to be computed at multiple processing nodes in some applications, e.g., iterative algorithms in the Spark model where the output of a MapReduce procedure acts as the input to the MapReduce procedure in the next iteration. Although the work in generalizes the CDC framework by allowing each Reduce function to be computed at more than one processing node, it only applies to homogeneous systems. Similar to , heterogeneities of the systems are also considered in . However, instead of a non-cascaded system, the authors propose a more general framework of cascaded CDC on heterogeneous systems. In other words, each Reduce function is allowed to be computed at multiple processing nodes. Since the processing nodes have different storage capacities, the number of subfiles stored at each processing node differs. The processing nodes with larger storage capacities are allocated more subfiles and thus, compute more intermediate results. Instead of allocating the files and functions randomly as in the work in , this cascaded CDC scheme uses a hypercube approach to allocate the files and functions to the processing nodes. The simulation results show that for the same number of processing nodes in the network, the proposed cascaded CDC scheme achieves smaller communication load than the state-of-art schemes that consider homogeneous systems , . \\nIn the study of , the allocation of the functions to the processing nodes does not depend on their capabilities but on the data stored at the nodes, of which the data placement is assumed to be random. This is very useful for applications that the processing nodes are mobile and collect data on-the-move. If the probability that the processing nodes contain data is higher than the pre-defined threshold, it is possible to allocate Reduce functions such that the processing nodes do not need to exchange their intermediate results for the computations of Reduce functions, i.e., each processing node can compute the Reduce functions based on its locally stored data. This reduces the number of broadcast transmissions in the Shuffle phase, thus minimizing the communication load.\\nAlthough heterogeneities of the systems are taken into consideration in some of the works , , , they merely focus on either file allocation or function allocation. On one hand, the work in proposes an optimal file allocation strategy in consideration of the heterogeneous storage capabilities of the processing nodes. However, the Reduce functions are distributed uniformly among the processing nodes. On the other hand, the works in and propose the allocation of Reduce functions based on the computational capabilities of the processing nodes. However, the input file is split equally and distributed among the processing nodes. Considering a more generalized heterogeneous system, a joint file and function allocation strategy is proposed in to reduce the communication load in the Shuffle phase. The file allocation and function assignment strategies allocate more subfiles and Reduce functions respectively to the processing nodes with higher computational and storage capabilities. Generally, the Reduce function assignment is related to the input file allocation as the processing nodes with more input files have greater storage and computational capabilities and hence are more capable of computing more output functions. In particular, there are two proposed schemes of function assignment, i.e., computation-aware function assignment and shuffle-aware function assignment. For computation-aware function assignment strategy, the number of functions allocated is proportional to the computational capabilities of the processing nodes in order to reduce computation latency. For the shuffle-aware function assignment strategy, the functions are mostly allocated to the processing nodes with high computational capabilities so that the communication load in the Shuffle phase is minimized. The simulation results show that the communication loads achieved by both computation-aware and shuffle-aware function assignment strategies are lower than the uniform function allocation strategy. Besides, the computation-aware function assignment strategy requires fewer number of output functions as compared to the proposed schemes in and . However, the number of input files required is much larger, especially when the number of processing nodes in the system increases.\\nWhile there has been great attention on the design of file allocation, coded shuffling algorithms and function allocation, all these works assume a fixed number of processing nodes in the distributed computing systems. For a given computation task that specifies the number of subfiles and the number of outputs, the allocation of functions and the data shuffling schemes can be implemented with minimum number of processing nodes. In the study of , the resource allocation problem is formulated as an optimization problem that minimizes the overall job execution time with optimal number of processing nodes used. For more practical implementation, the resource allocation strategy should consider the heterogeneity in the processing speed of the nodes, since the straggler effects cause an increase in computation latency which increases the overall job execution time.\\nBy exploiting the fact that the processing nodes have time-varying computing resources, e.g., the processing nodes may have different central processing unit (CPU) power over time, an optimal computation task scheduling scheme helps to reduce the communication load. In the scheduling of tasks under dynamic computing resources, the total communication load is minimized by optimizing the number of allocated tasks and load redundancy at each time slot .\\n\\\\begin{table*}[t]\\n\\\\caption{\\\\small CDC schemes to reduce communication costs.} \\n\\\\label{tab:communication}\\n\\\\centering\\n\\\\begin{tabular}{|>{\\\\centering\\\\arraybackslash}m{2cm} |>{\\\\centering\\\\arraybackslash}m{1.4cm} |>{\\\\centering\\\\arraybackslash}m{3.5cm} |m{6.3cm} |>{\\\\centering\\\\arraybackslash}m{2.3cm}|}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n\\\\textbf{Approach} & \\\\textbf{Ref.} & \\\\textbf{Coding Schemes} & \\\\multicolumn{1}{c|}{\\\\textbf{Key Ideas}} & \\\\textbf{Platform Support} \\\\\\\\ \\\\hline\\n\\\\multirow{12}{=}{File Allocation} & & - & Deterministic and systematic storage update strategy & Heterogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & Hybrid Coded MapReduce & Allocates Map tasks such that data locality is maximized & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & Group-based Coded MapReduce & Allocates dataset by using a group-based method in order to avoid high subpacketization level and allow processing nodes to cooperate in the transmission of messages& Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & Resolvable Design & Uses single-parity code to determine the number of subfiles and allocate the subfiles based on the corresponding resolvable design & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & Placement Delivery Arrays & Construction of CDC schemes based on PDA which has the property of illustrating the placement and delivery phase in a single array & Homogeneous\\\\\\\\ \\n\\\\hline\\n\\\\multirow{18}{=}{Coded Shuffling Design} & & Compressed CDC & Pre-combines computed intermediate values of the same function, followed by coding the pre-combined packets for communication between different processing nodes & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & Quantized SGD & Quantizes the component of the gradient vector to a discrete set of values and encodes the quantized gradients given their statistical properties & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & Pliable Index Coding & Semi-random data shuffling scheme based on modified pliable index coding to reduce the number of communication rounds & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & Cross-iteration Coded Computing & Jointly codes across multiple iterations for a single communication round & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & CDC for distributed graph computing systems & Instead of communicating with all other processing nodes, each processing nodes only needs to communicate with a subset of processing nodes to obtain the required data to complete its computation tasks & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & CDC for multistage dataflow & A more generalized CDC scheme is proposed to handle multistage dataflow computation tasks which are represented by layered DAGs & Homogeneous\\\\\\\\\\n\\\\hline\\n\\\\multirow{12}{=}{Consideration of Underlying Network Architecture} & & Hybrid Coded MapReduce & Reduces cross-rack communication at the expense of higher intra-rack communication based on the server-rack architecture & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n & & Topological CDC & Considers t-ary fat-tree topology which is a more practical topology to connect the physically separated processing nodes in data center networks & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n & & UberShuffle & Considers imperfect communication channel between the workers and the master node & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n& & - & Considers multicore setup where each computing node can have multiple cores, e.g., the CPU instances of publicly available cloud infrastructure can deliver up to 128 cores & Homogeneous\\\\\\\\\\n\\\\hline\\n\\\\multirow{12}{=}{Function Allocation} & & - & Considers a non-cascaded system and allocates Reduce functions over a simplified heterogeneous network which comprises multiple homogeneous networks & Heterogenous \\\\\\\\\\n\\\\cline{2-5}\\n& & Cascaded CDC & Reduce functions are computed at more than one processing node and they are allocated based on the combinatorial design in & Heterogeneous\\\\\\\\\\n\\\\cline{2-5}\\n & & - & Allocates functions to maximize data locality such that the number of communication rounds required is reduced & Homogeneous\\\\\\\\\\n\\\\cline{2-5}\\n & & - & Joint file and function allocation strategy & Heterogeneous\\\\\\\\\\n\\\\cline{2-5}\\n & & - & Considers the availability of time-varying computing resources & Homogeneous\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table*}\",\n \"id\": \"1109cc8b-6b58-4ee5-ba62-ca162ae902db\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"f06fb9ca-5252-413b-b98f-08903f106a9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"Function Allocation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Function Allocation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we have reviewed four approaches to minimize communication costs in distributed computing systems. For each approach, we discuss the solutions that are proposed in different works. We summarize the coding schemes to minimize communication costs in Table~\\\\ref{tab:communication}. The lessons learned are as follows:\\n\\\\begin{itemize}\\n\\\\item To handle the increasing amounts of data generated, more processing nodes are needed for the completion of distributed computations given the limited capabilities of the processing nodes. With more processing nodes connected to the network, more communication rounds of the computed intermediate results are required, resulting in higher communication costs which lead to a longer job execution time. Besides, the high communication costs in the data shuffling phase impede the implementation of efficient distributed iterative algorithms which are useful for the training of machine learning models. As such, the minimization of communication costs is a key step to achieve the objective of reducing the overall job execution time of the distributed computation tasks.\\n\\\\item While having more processing nodes to perform the computation tasks in parallel helps to reduce the computation load of each processing node, the communication costs may increase, slowing down the entire computation process. Instead of generating infinitely large number of subfiles and distributing them among large number of processing nodes, several studies focus on determining the optimal number of subfiles to avoid the input file from splitting too finely. For example, the resolvable design based schemes and PDA approaches are adopted to split the input files. Hence, it may not be optimal to use all the processing nodes that are connected to the network. In fact, the authors in propose an optimal resource allocation scheme that determines the minimum number of processing nodes needed to achieve the minimum overall job execution time. \\n\\\\item Coding techniques have shown to be effective in reducing communication load in the data shuffling phase at the expense of higher computation load . However, the two-dimensional tradeoff is insufficient to fully evaluate the performance of the CDC schemes. Apart from leveraging on the computational capabilities of the processing nodes, their storage capabilities can be exploited. For example, more data can be stored at processing nodes with higher storage capacities such that the number of communication rounds is reduced . Moreover, by considering the data stored at the processing nodes, the allocation of functions that maximizes data locality helps to reduce the need for communication bandwidth . Hence, instead of the two-dimensional tradeoff between computation load and communication load, the three-dimensional tradeoff between computation, communication and storage cost has to be carefully managed for the implementation of efficient CDC schemes.\\n\\\\item For effective implementation of CDC schemes in practical distributed computing systems, the underlying architecture has to be considered. Generally, the distributed computing systems operate under the server architecture which consists of multiple racks where each rack has several servers. The Hybrid Coded MapReduce scheme reduces cross-rack communication at the expense of higher intra-rack communication. Besides, the communication channels between the master node and workers are imperfect. As a result, the theoretical gains of the coded data shuffling schemes are not achievable under practical setups. In order to design and analyze the performance of the CDC schemes, the limitations of the distributed computing systems should be taken into consideration. \\n\\\\item Most of the CDC schemes focus on the minimization of communication costs in the data shuffling phase at the expense of reasonable increase in computation load. However, the computational overhead of the algorithm is not negligible under some settings. For example, under much faster broadcast environment, the UberShuffle algorithm incurs significant encoding and decoding costs such that the shuffling gain is offset by the high computational overhead. For future works, more practical CDC schemes can be proposed such that the communication costs are minimized while maintaining low computational cost to improve the performance of the distributed computing systems. Given the uncoded computing schemes, e.g., and , achieve near-optimal computation and storage costs, one possible research direction is to merge the communication-efficient proposed schemes with the uncoded computing schemes to reduce the computation and storage costs.\\n\\\\item Although some of the works consider heterogeneities in the capabilities of the processing nodes to allocate files and functions, the presence of stragglers which have slower processing speeds still hinder the performance of the distributed computing systems. Therefore, we further discuss the approaches to mitigate the straggler effects in the next section.\\n\\\\end{itemize}\",\n \"id\": \"71e25826-5a72-4095-8ca3-d26170c96db8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"f06fb9ca-5252-413b-b98f-08903f106a9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Minimization of Communication Load\"\n ],\n [\n \"subsection\",\n \"Summary and Lessons Learned\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary and Lessons Learned\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:stragglers}\\nIn distributed computing systems, processing nodes have different processing speeds and thus the time taken to complete their allocated subtasks differs from each other. Since the computation task is distributed among the processing nodes, the master node needs to wait for all processing nodes to complete their subtasks and return the computed results before recovering the final result. As such, the time taken to execute a computation task is determined by the slowest processing node. This is also known as the \\\\emph{straggler effects}. \\nThe problem of straggler effects has been widely observed in the distributed computing systems. Previously, various methods such as straggler detection , , asynchronous execution and naive replication of jobs , have been proposed to reduce the overall time taken to execute the computation tasks. Recently, coding approaches have been shown to outperform the aforementioned methods in reducing computation latency of the distributed computing systems. In this section, we discuss three approaches to mitigate the straggler effects: \\n\\\\begin{itemize}\\n\\\\item \\\\emph{Computation Load Allocations: }Coding techniques can be implemented together with computation load allocation strategies to reduce the computation latency in the distributed computing systems. It is important to take into account the variation in the computational capabilities of the processing nodes to allocate computation load. As such, different prediction methods such as an long short-term memory (LSTM) algorithm , an Auto Regressive Integrated Moving Average (ARIMA) model and a Markov model are used to estimate the processing speeds of the nodes. Generally, the objective of the load allocation strategies is to minimize computation latency. However, in some applications where strict deadlines are given, the load allocation strategies aim to maximize timely computation throughput .\\n\\\\item \\\\emph{Approximate Coding: }For some applications, e.g., location-based recommendation systems, exact solutions are not necessary. The studies explore different coding approaches to obtain approximate solutions to the problems. The approximate coding methods relax the requirement for convergence and thus reduce the number of workers that are required to return their computed results. This can avoid stragglers to make an adverse effect to the system.\\n\\\\item \\\\emph{Exploitation of Stragglers: }The straggling nodes may have completed a fraction of the allocated computation tasks, of which is a waste to be ignored completely. In fact, the stragglers may not be persistent over the entire computation process and thus their partial computed results can be useful to recover the final result. In order to maximize the resource utilization of the straggling nodes, the workers are allowed to sequentially process their allocated subtasks and transmit their partial computed results continuously . However, this may come at the expense of higher communication load, which needs to be carefully managed. \\n\\\\end{itemize}\",\n \"id\": \"79250ecd-4bf6-4689-8a12-f801fbeb39c1\",\n \"level\": \"section\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ]\n ],\n \"subsections\": [\n \"bbc9eda4-c6aa-4f51-abdb-b74f0a6b37df\",\n \"99bd280c-f0a8-4997-9dfb-23945fa49c2a\",\n \"775a5f6a-bf8f-475d-ab6b-0bf91f59fd88\",\n \"07a5860c-2192-4141-a8ce-fe22c5666e51\"\n ],\n \"title\": \"Mitigation of Stragglers\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Apart from having different straggling rates that affect the completion of tasks, the processing nodes have different capabilities, e.g., storage capacities, computing resources and communication bandwidths. To better handle the straggling nodes in the distributed computing systems, an optimal load allocation strategy that takes into account these heterogeneities is necessary to minimize the overall job execution time. Given the computation time parameters, i.e., the straggling and shift parameters of each worker, the Heterogeneous Coded Matrix Multiplication (HCMM) algorithm determines the allocation of computation load to each worker. The HCMM scheme exploits the benefits of both coding techniques and computation load allocation strategy to minimize the average computation time of the computation tasks. Given that it is difficult to derive closed-form expressions of expected computation time of the heterogeneous processing nodes, a two-step alternative problem formulation is proposed. In the first step, given a time period, the number of computed results by the workers is maximized by optimizing the load allocation. In the second step, given the derived load allocation in the first step, the time needed to ensure sufficient results are returned at a pre-defined probability is minimized. The simulation results show that when workers' computation time is assumed to follow a shifted exponential runtime distribution, HCMM reduces the average computation time by up to 71\\\\%, 53\\\\% and 39\\\\% over uniform uncoded, load-balanced uncoded and uniform coded load allocation schemes, respectively. In practical experiments over Amazon EC2 clusters, the combination of HCMM and LT codes outperforms the uniform uncoded, load-balanced uncoded and uniform coded load allocation schemes by up to 61\\\\%, 46\\\\% and 36\\\\%, respectively.\\nAlthough HCMM achieves asymptotically optimal computation time, the decoding complexity is high, which suggests the opportunity to further speed up the overall computation tasks. In practical distributed computing systems, some processing nodes have the same computational capabilities, in terms of the same distributions of computation time, and thus they can be grouped together. By exploiting the group structure and heterogeneities among different groups of processing nodes , , the implementation of a combination of group codes and an optimal load allocation strategy not only approaches the optimal computation time that is achieved by the MDS codes, but also has low decoding complexity. In addition, by varying the number of allocated rows of the matrix to the workers , the computation latency can be reduced by orders of magnitude over the MDS codes with fixed computation load allocation as the number of workers increases. The load allocation strategy proposed in focuses mainly on the design of an optimal MDS code. Other efficient coding algorithms such as LT codes can also be explored in future studies.\\nIn addition to the heterogeneous capabilities of the processing nodes, the amount of available resources of the processing nodes may vary over time. By always allocating computation tasks to the processing nodes with higher capabilities, delays may be incurred in completing the allocated tasks if the processing nodes start to work on the newly-allocated computation tasks in parallel. At the same time, the resources of the processing nodes with lower capabilities may be under-utilized. In order to maximize the resource utilization of the processing nodes, dynamic workload allocation algorithms which are adaptive to the time-varying capabilities of the processing nodes are proposed , , . To provide robustness against the straggling nodes, the design of the load allocation algorithms often depends on the historical data of the processing nodes such as computation time, which can be used to predict the processing speeds by using an LSTM algorithm , an ARIMA model or a Markov model . The best performing LSTM model achieves 5\\\\% lower prediction error than an ARIMA (1,0,0) model .\\nIn the study of , the authors propose Coded Cooperative Computation Protocol (C3P) which is a dynamic and adaptive coded cooperation framework that efficiently utilizes the available resources of the processing node while minimizing the overall job execution time. Specifically, the master node determines the coded packet transmission intervals based on the responsiveness of the processing nodes. For the processing nodes which are not able to complete the tasks within the given transmission interval, they suffer from longer waiting time for the next coded packets. In comparison to the HCMM scheme which does not consider the dynamic resource heterogeneity in workers, the C3P framework achieves more than 30\\\\% improvement in task completion delay.\\nThe dynamic and adaptive load allocation algorithms are especially useful in providing timely services with deadline constraints which are common in many IoT applications. For such applications, instead of minimizing task completion delay, the objective of the load allocation algorithms is to maximize timely computation throughput, i.e., the average number of computation tasks that are successfully completed before the given deadline . \\nFor some applications that may need timely but not necessarily optimal results, it is more important to recover the final result with the highest accuracy possible by the stipulated deadlines than to solve for an exact solution. The algorithm to solve for an approximate solution requires significantly shorter computation time than that of an algorithm that solves for an exact solution . In the study of , the batch-processing based coded computing (BPCC) algorithm is proposed. The workers first partition the allocated encoded matrix into several batches, i.e., submatrices, for computations. As soon as the workers complete the computation of each batch of submatrix with a given vector, they return the partial computation results which are used to generate the approximate solution. Based on the computation time parameters, i.e., the straggling and shift parameters of the workers, the BPCC algorithm is used to optimally allocate the computation load to each worker by considering two cases: (i) negligible batching overheads, and (ii) linear batching overheads. Hence, the allocation of computation load is optimized by jointly minimizing both the overall job execution time as well as the potential overhead of batch processing. In addition to reducing computation latency, the BPCC algorithm exploits the partial computation results of all processing nodes, including the straggling nodes, which contribute to approximate solutions of higher accuracy . The simulation results show that BPCC algorithm with negligible batching overheads achieves up to 73\\\\%, 56\\\\% and 34\\\\% reduction in average job execution time over uniform coded, load-balanced uncoded and HCMM load allocation schemes respectively. The experimental results on Amazon EC2 clusters and Unmanned Aerial Vehicles (UAVs) based airborne computing platform also demonstrate similar results. \\n\\\\begin{table*}[!ht]\\n\\\\caption{\\\\small Approximate CDC schemes to mitigate the straggler effects.} \\n\\\\label{tab:approximate}\\n\\\\centering\\n\\\\begin{tabular}{|>{\\\\centering\\\\arraybackslash}m{2.0cm} |>{\\\\centering\\\\arraybackslash}m{1.5cm} |>{\\\\centering\\\\arraybackslash}m{3.3cm} |m{8cm}|}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n\\\\multicolumn{1}{|c|}{\\\\textbf{Problems}} & \\\\textbf{Ref.} & \\\\textbf{Coding Schemes} & \\\\multicolumn{1}{c|}{\\\\textbf{Key Ideas}} \\\\\\\\ \\\\hline\\n\\\\multirow{5}{=}{Matrix-Vector} & & Anytime Coding & Computations can be stopped anytime and the approximate solution is derived from the processing nodes that have completed their tasks\\\\\\\\\\n\\\\cline{2-4}\\n & & Coded Sequential Computation Scheme & A sequence of approximated problems are designed such that the time required to solve these problems is shorter than solving the original problem directly\\\\\\\\\\n\\\\hline\\n\\\\multirow{3}{=}{Matrix-Matrix} && CodedSketch & Use a combination of count-sketch technique and structured polynomial codes \\\\\\\\\\n\\\\cline{2-4}\\n & & OverSketch & Divide the sketched matrices into blocks for computations\\\\\\\\\\n\\\\hline\\n\\\\multirow{8}{=}{Gradient Descent} & & Bernoulli Gradient Codes & Use Bernoulli random variables as entries of the function assignment matrix\\\\\\\\\\n\\\\cline{2-4}\\n && Stochastic Block Codes & Interpolation between BGC and FRC \\\\\\\\\\n\\\\cline{2-4}\\n & & Stochastic Gradient Coding & Distribute data based on pair-wise balanced scheme and provide a rigorous convergence analysis of the proposed coding scheme\\\\\\\\\\n\\\\cline{2-4}\\n & & LDPC Codes & Encode the second moment of the data points\\\\\\\\\\n\\\\cline{2-4}\\n &, & Encoded Optimization & Encode both the labels and data such that the redundancy is introduced in the formulation of the optimization problems \\\\\\\\\\n\\\\hline\\n\\\\multirow{4}{=}{Non-Linear} & & Generalized PolyDot & Generalization of PolyDot codes and is used for the training of DNNs\\\\\\\\\\n\\\\cline{2-4}\\n & & Learning-based Approach & Design neural network architectures to learn and train the encoding and decoding functions to approximate unavailable outputs\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table*}\",\n \"id\": \"bbc9eda4-c6aa-4f51-abdb-b74f0a6b37df\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"79250ecd-4bf6-4689-8a12-f801fbeb39c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ],\n [\n \"subsection\",\n \"Computation Load Allocation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Computation Load Allocation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"For some applications, it is not necessary to obtain exact solutions. In this subsection, we review the coding approaches that are used to derive approximate solutions for the distributed computation tasks: (i) matrix multiplication, (ii) gradient descent, (iii) non-linear computations. Table~\\\\ref{tab:approximate} summarizes the approximate coding schemes for various distributed computation tasks.\",\n \"id\": \"99bd280c-f0a8-4997-9dfb-23945fa49c2a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"79250ecd-4bf6-4689-8a12-f801fbeb39c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ],\n [\n \"subsection\",\n \"Approximate Coding\"\n ]\n ],\n \"subsections\": [\n \"a171484f-77a0-4766-94ff-e7249ae60fc2\",\n \"e8ed1de1-1ec4-4ea0-b627-b1ea53a5a3a3\",\n \"fd439e01-9cb8-4b85-a497-db80e1e80ae6\"\n ],\n \"title\": \"Approximate Coding\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsubsec:matrix} The anytime coding scheme derives an approximate solution by using the output results of the completed processing nodes at any given time. Based on singular value decomposition (SVD), the given computation task is decomposed into several subtasks of different priorities. More important subtasks are allocated more processing nodes for computations as they improve the accuracy of the approximation. To allow the users to receive useful information from time to time, approximate solutions can be transmitted to the users sequentially. This can be achieved by solving a sequence of approximated problems , instead of solving the original problem directly. \\nTo further reduce the computation time of approximate matrix multiplications, sketching techniques , can be used to remove redundancy in the structure of the matrices through dimensionality reduction. However, by using sketching techniques, the recovery threshold increases as the redundancy is removed. In contrast, coding techniques reduces recovering threshold by introducing redundancy. As such, a combination of both techniques that carefully manage the tradeoff between the recovery threshold and the amount of redundancy can be implemented to minimize computation latency. In particular, count-sketch technique is combined with structured codes to mitigate the straggler effects by preserving a certain amount of redundancy, thereby achieving the optimal recovery threshold and hence computation latency , .\",\n \"id\": \"a171484f-77a0-4766-94ff-e7249ae60fc2\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"99bd280c-f0a8-4997-9dfb-23945fa49c2a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ],\n [\n \"subsection\",\n \"Approximate Coding\"\n ],\n [\n \"subsubsection\",\n \"Matrix multiplications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Matrix multiplications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsubsec:gradient}To speed up the distributed gradient descent tasks, several approximate gradient coding schemes are proposed to approximately compute any sum of functions. Instead of constructing the gradient codes based on expander graphs which are difficult to compute due to high complexity, a more efficient and simpler Bernoulli Gradient Code (BGC) is proposed by using sparse random graphs which introduce randomness into the entries of the function assignment matrix. Since the performance of the gradient codes depends on the efficiency of the decoding algorithms, the authors also present two decoding techniques to recover the approximate solution from the outputs of the non-straggling nodes. The simulation results show that the BGC schemes can handle adversaries with polynomial-time computations but at a cost of higher decoding error than the FRC schemes . Besides, the optimal decoding algorithm always achieves a lower decoding error than that of the one-step decoding algorithm across various gradient coding schemes. A rigorous convergence analysis of the approximate gradient codes and the performance of BGC on different practical datasets such as the Amazon dataset, Covertype dataset and KC Housing dataset are presented in . Stochastic block code which is based on the stochastic block model from random graph theory, is an interpolation between FRC and BGC . On one hand, the FRC schemes achieve small reconstruction errors under random straggler selection while on the other hand, the BGC schemes are robust against polynomial-time adversarial stragglers.\\nOther approximate gradient coding methods such as the Stochastic Gradient Coding (SGC) and LDPC codes are used to obtain an estimate of the true gradient. Similar to the idea of encoding data batches in the PCR schemes , the data variables in the optimization formulation can be efficiently encoded to mitigate the straggler effects in more general large-scale optimization problems such as support vector machines, linear regressions and compressed sensing , . Generally, in solving for approximate solutions to reduce computation latency of the distributed tasks, there is an inherent tradeoff between the recovery threshold and the approximation error where the recovery threshold can be reduced at the expense of larger approximation error .\",\n \"id\": \"e8ed1de1-1ec4-4ea0-b627-b1ea53a5a3a3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"99bd280c-f0a8-4997-9dfb-23945fa49c2a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ],\n [\n \"subsection\",\n \"Approximate Coding\"\n ],\n [\n \"subsubsection\",\n \"Gradient descent\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Gradient descent\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"By leveraging on the enormous amounts of data generated, machine learning algorithms are useful in making predictions and allowing devices to respond intuitively to user demands without human interception. Different neural network architectures are developed to make accurate inference given the dataset. Since some of the layers of the neural networks such as the max-pooling functions and the activation layer are non-linear, the overall computation of the functions are non-linear. As a result, most of the prior works on linear computations which are discussed in Sections~\\\\ref{subsec:codingtech}, \\\\ref{subsubsec:matrix} and \\\\ref{subsubsec:gradient}, are not applicable to the computation of the increasingly important non-linear neural networks, in which their performances are also limited by the straggling nodes. One of the few approaches that can be extended to the training of DNNs is the Generalized PolyDot codes which are used to compute matrix-vector multiplications. The Generalized PolyDot codes are used to code the linear operations at each layer of the neural networks. This coding scheme allows for errors in the training of each layer. In other words, decoding can still be performed correctly given that the amount of errors does not exceed the maximum error tolerance level. The effectiveness of coding techniques in mitigating the straggler effects of different neural network architectures such as AlexNet and Visual Geometry Group (VGG) in an IoT system is illustrated in . However, this unified coded DNN training strategy may not be relevant to the training of other neural networks which have large number of non-linear functions. As such, the authors in propose a learning-based approach for designing codes. Based on the dilated CNN and multilayer perceptrons (MLP), neural network architectures and a training methodology are proposed to learn the encoding and decoding functions. The outputs of the decoding functions are used to approximate the unavailable outputs of any differentiable non-linear function. The simulation results show that the learning-based approach to designing the encoding and decoding functions can accurately reconstruct 95.71\\\\%, 82.77\\\\% and 60.74\\\\% of the unavailable ResNet-18 classifier outputs on MNIST, Fashion-MNIST and CIFAR-10 datasets respectively.\",\n \"id\": \"fd439e01-9cb8-4b85-a497-db80e1e80ae6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"99bd280c-f0a8-4997-9dfb-23945fa49c2a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ],\n [\n \"subsection\",\n \"Approximate Coding\"\n ],\n [\n \"subsubsection\",\n \"Non-linear computations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Non-linear computations\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To avoid delays caused by the straggling nodes in the network, most distributed computation schemes ignore the work completed by the straggling nodes by either increasing the workload of the non-straggling nodes or by obtaining less accurate approximate solutions. However, the amount of work that has been completed by the straggling nodes, especially the non-persistent stragglers, is non-negligible and can be better utilized.\\nIn order to exploit the computational capacities of these non-persistent stragglers, multi-message communication (MMC) is used where the workers are allowed to send multiple messages to the master node at each iteration. This allows the workers to transmit their partial computed results whenever they complete a fraction of the allocated task, rather than completing the entire computation task before transmitting the computed result in a single communication round. The work in considers the implementation of Lagrange Coded Computing (LCC) with MMC to minimize the average job execution time at the expense of higher communication load due to the increase in number of messages transmitted by the workers to the master node. Since the LCC scheme utilizes polynomial interpolation to recover the final result, the decoding complexity and the number of transmissions can be further reduced by increasing the number of polynomials used in decoding the computed results returned by the workers. The simulation results show that by exploiting the computing resources of the non-persistent stragglers via MMC, the average job execution time decreases as the computation load of each worker increases. However, since the communication load of the LCC-MMC scheme is constant as the computation load increases, it is suitable to be implemented only when computation time dominates the overall execution time of the distributed tasks. The total time needed to execute the distributed tasks includes the time needed for both computation and communication. Otherwise, if communication load is the cause of the bottleneck of the network, LCC without MMC should be used instead since the communication load can be reduced at the expense of higher computation load.\\nGiven that MMC is allowed where the workers perform more than one round of communication for each iteration, the computation work done by the straggling nodes can be exploited by allowing sequential processing where the workers need to transmit the computation results of their completed subtask before working on the next subtask. To fully exploit the useful information provided by the straggling nodes, the hierarchical coded computation scheme is proposed in to utilize the computations from all workers. Each worker first divides the allocated computation task into layers of sub-computations, which are processed sequentially, i.e., the result of a layer of sub-computation is transmitted to the master node before the computation of the next layer starts. Since the workers have different processing speeds and the sub-computations are performed sequentially, the finishing time for each layer is different. MDS codes are used to encode the layers so that the finishing time of each layer is approximately the same. The top layers which have lower probability of erasure are encoded with higher-rate MDS codes whereas the bottom layers are encoded with lower-rate MDS codes. The simulation results show that for the same amount of tasks to be completed, the proposed hierarchical coded computation scheme achieves $1.5$ factor improvement in expected finishing time as compared to the coded computation scheme proposed in which ignores the computations of the straggling nodes.\\nIn the study of , by computing the block products sequentially, the partial computation results from the straggling nodes are used to aggregate the final result. In order to preserve the sparsity of matrices for processing by the workers, instead of coding the entire matrix, the fraction of coded blocks can be specified to control the level to which coding is utilized in the solutions. There are two different approaches considered, depending on the placement of the coded blocks. When the coded blocks appear at the bottom of the workers, it is easier for the master to decode. When the coded blocks appear at the top of the nodes, it minimizes the computation load of the workers. As such, the placement can be calibrated based on the task requirements.\\nWhile the coded computation schemes achieve low communication load and reduce the average job execution time for each iteration, the uncoded computation schemes have their own benefits of having no decoding complexity and allowing partial gradient updates. In order for a system to benefit from the advantages of both schemes, coded partial gradient computation (CPGC) scheme is proposed in . In the CPGC scheme, uncoded submatrices are allocated for the first computation since there is a high probability for each worker to complete their first computation task. Subsequently, coded submatrices are allocated to handle the straggling nodes. The master node is able to update the gradient parameters by using the computation results from a subset of workers and by exploiting the partial computations of the straggling nodes. As a result, the average job execution time for each iteration is reduced. \\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{mitigate.pdf}\\n\\\\caption{\\\\small Approaches to mitigate the straggler effects include (a) computation load allocation by predicting the speed of the processing nodes , , (b) approximate coding and (c) exploitation of stragglers by allowing multi-message communications , .}\\n\\\\label{fig:mitigate}\\n\\\\end{figure}\",\n \"id\": \"775a5f6a-bf8f-475d-ab6b-0bf91f59fd88\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"79250ecd-4bf6-4689-8a12-f801fbeb39c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ],\n [\n \"subsection\",\n \"Exploitation of Stragglers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Exploitation of Stragglers\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we have discussed three approaches (Fig.~\\\\ref{fig:mitigate}) to mitigate the straggler effects. The lessons learned are as follows:\\n\\\\begin{itemize}\\n\\\\item The straggler effect is a key issue to be resolved in order to reduce computation latency, hence minimizing the overall job execution time. Due to various factors such as insufficient power, contention of shared resources, imbalance work allocation and network congestions , some processing nodes may run slower than the average or even be disconnected from the network. Since the computation tasks are only completed when all processing nodes complete their computations, the time needed to complete the tasks is determined by the slowest processing node. Coding techniques have shown their effectiveness in reducing computation latency by introducing redundancy . In this section, we have explored the use of coding techniques for different distributed computation tasks, e.g., matrix-vector multiplications, matrix-matrix multiplications, linear inverse problems, iterative algorithms, convolutions and non-linear problems. While most of the research focuses on the design of encoding techniques, the decoding complexity of the codes also affects the computation latency significantly. Apart from Reed-Solomon codes and LDPC codes , more effective codes with low decoding complexity can be investigated in future studies. \\n\\\\item Considering heterogeneities in the capabilities of the processing nodes, effective computation load allocation strategies are implemented to allocate workload to the processing nodes. We have discussed the proposed computation load allocation algorithms under different constraints, e.g., strict deadlines and time-varying computing resources. \\nIn addition, different prediction methods such as the LSTM algorithm , an ARIMA model and a Markov model that predict the speeds of the processing nodes are explored. However, the stragglers may be non-persistent in nature and thus they may be useful when they are able to perform computations faster than the average rate. Hence, the load allocation based on the responsiveness of the processing nodes may be more useful in such situations. \\n\\\\item Instead of exact solutions, it is acceptable to derive approximate solutions for some applications, e.g., location-based recommender systems. Various coding techniques to derive approximate solutions are investigated. For example, in the studies of and , sketching techniques are used with structured codes to minimize computation latency. However, there exists a tradeoff between the recovery threshold and the approximation error. For future works, an improvement to this tradeoff can be investigated.\\n\\\\item Although the straggling nodes run slower than the average, the computations that are completed may still be useful. It is wasteful to not utilize the partial computed results of the straggling nodes. Besides, these partial computed results can help to improve the accuracy of the estimates. For example, in , the stragglers are treated as soft errors instead of erasures to minimize the mean-squared error of the iterative linear inverse solvers under a deadline constraint by using approximate weights. Outputs from all computing nodes, including the straggling nodes are used to recover estimates that are as close as possible to the convergence values when the computation deadline is brought forward or when the number of computing nodes increases. Unfortunately, as the processing nodes are required to send their partial results once they complete, more communication rounds are performed. The high communication costs may be the bottleneck of the distributed computation tasks. Given the advantages of using partial results from the straggling nodes, optimization approaches to minimize the communication costs between the master node and the workers should be explored.\\n\\\\item The current studies in this section have proposed effective coding schemes for implementation. However, they do not consider security in the design of the coding schemes. For example, the FRC scheme achieves high accuracy in the presence of stragglers, but it is susceptible to attacks from adversarial stragglers, which turn more processing nodes into straggling nodes. Besides, other than the straggling workers, there may exist malicious or curious workers that may compromise the privacy and security of the system. Therefore, approaches to ensure secure coding are discussed in-depth in the next section. \\n\\\\end{itemize}\",\n \"id\": \"07a5860c-2192-4141-a8ce-fe22c5666e51\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"79250ecd-4bf6-4689-8a12-f801fbeb39c1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Mitigation of Stragglers\"\n ],\n [\n \"subsection\",\n \"Summary and Lessons Learned\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary and Lessons Learned\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:security}\\nIn distributed computing, the data owner, master node, and workers may not belong to the same entity. For example, the data owner may wish to perform a task on a massive dataset on which intensive computations have to be performed. The computations may be divided and distributed to multiple workers on third-party computing services. However, sensitive data, e.g., in healthcare services , may be involved. In this case, \\\\textit{curious} workers may collude to obtain information about the raw data, whereas \\\\textit{malicious} workers may intentionally contribute erroneous inputs to introduce biases to the model. Besides, in some cases, the dataset does not belong to either the master node or the workers, and as such, the raw dataset has to be guarded against both parties.\\nTo ensure that privacy and security are preserved during the computing tasks, conventional methods such as homomorphic encryption have been proposed in which the data is first encrypted before being shared to workers. However, the encryption techniques are found to be costly in terms of computation and storage costs . Besides, the secure multi-party computation (MPC) approaches mainly focus on the correctness and privacy of the data , whereas neglecting to reduce the complexity of computation at each workers, or the number of workers required to perform the task. Recently, coding techniques that have originally been introduced to mitigate straggling workers are increasingly utilized to provide information-theoretic privacy guarantees. Specifically, information-theoretic privacy considers the setup consisting honest but curious workers, in which collusions formed between $T$ of $N$ workers do not leak information about the training dataset. Coding schemes can be used to not only mitigate the stragglers but also the colluding curious workers and malicious workers as illustrated in Fig.~\\\\ref{fig:security}. For ease of exposition, we classify the related studies into two main categories in this section:\\n\\\\begin{itemize}\\n\\\\item \\\\textit{Secure Distributed Computing:} In this section, the studies aim to reduce the number of workers needed for information-theoretic privacy, i.e., where the colluding workers are unable to infer sensitive information from the dataset. For some studies, this objective is met while simultaneously preserving the efficiency of distributed computing, e.g., through providing resiliency against straggling workers . \\n\\\\item \\\\textit{Secure Distributed Matrix Multiplication (SDMM):} The studies in the aforementioned category mainly focus on generic operations, e.g., addition, subtraction, multiplication, or computation of polynomial functions, whereas the studies in this category focus specifically on matrix multiplication. One key difference between the two categories is that SDMM considers the specific scenario in which \\\\textit{both} input matrices in the multiplication operation are private information, i.e., two-sided privacy , whereas the prior category mainly considers one-sided privacy, i.e., only one input matrix is private. In addition, a performance metric of interest in the SDMM literature is the download rate, i.e., the ratio of the size of desired result to the total amount of information downloaded by a user from the workers.\\n\\\\end{itemize}\",\n \"id\": \"ee3d99f5-815c-4ba8-b336-968d0def830a\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Secure Coding for Distributed Computing\"\n ]\n ],\n \"subsections\": [\n \"8d724f64-c2a5-4e3f-a151-1bdc01b1a822\",\n \"d523704a-b56b-49c2-97c7-6500404f0ac2\",\n \"362bc888-78c5-42b5-b72f-1d88622fe4e4\"\n ],\n \"title\": \"Secure Coding for Distributed Computing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:securedist}\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{cdc-security.pdf}\\n\\\\caption{\\\\small Illustration of coding framework for the objectives of mitigating stragglers, colluding curious workers, and malicious workers.}\\n\\\\label{fig:security}\\n\\\\end{figure}\\nIn Section~\\\\ref{subsec:codingtech}, we discuss that the polynomial codes proposed in have the desirable property of an optimal recovery threshold that does not scale with the number of workers involved. In consideration of this useful property, the authors in propose the polynomial sharing approach which combines the polynomial codes and the Ben-Or, Goldwasser, and Wigderson (BGW) scheme . The system model considered in this study is that the data originates from external sources, and thus has to be kept private against both the workers and master node. In contrast to the BGW approach which uses Shamir's scheme to encode the dataset, the study of proposes to encoded the dataset using the polynomial coding scheme. The authors show that the polynomial sharing approach may be applied to perform several procedures, e.g., addition, multiplication, and the computation of polynomial functions, while reducing the number of workers required to complete the task as compared to conventional MPC approaches even when workers have capacity-limited communication links. \\nTypically, in conventional polynomial coding schemes, the dataset on which computations are performed is divided into multiple sub-tasks, with one sub-task encoded and assigned to each worker. In this case, faster workers that complete their task will be idle while waiting for straggling workers. To further mitigate the straggler effects, the authors in leverage on computation redundancy to propose the private asynchronous polynomial coding scheme in which a computation task is divided into several relatively smaller sub-tasks for distribution to each worker. This results in two key advantages, in addition to retaining the privacy preservation properties of polynomial coding. Firstly, the smaller sub-tasks can be successfully completed by straggling workers with limited computing capacity. Secondly, workers of the fastest groups are assigned more tasks to continue working throughout the whole duration rather than wait for the stragglers, thus reducing the computation time. \\nHowever, the studies mainly utilize polynomial coding for privacy preservation which is restrictive in certain aspects, e.g., it only allows column-wise partitioning of the matrices . As such, the entangled polynomial codes are applied by as an extension to polynomial sharing, so as to further reduce the restrictions during the data sharing phase, and hence, the number of workers required to perform the same computations while meeting privacy constraints.\\nWhile the studies of consider the scenario in which honest-but-curious workers are involved, workers may randomly be malicious in nature. As an illustration, a group of workers may be involved to compute gradients towards training a machine learning model. However, the gradients may be intentionally misreported by the workers to introduce biases or inaccuracies to the model . An existing approach is to perform median, rather than mean, based aggregation of the gradients to eliminate misreports that are usually outliers . However, the median based aggregation is computationally costly and faces convergence issues. As such, the study of proposes DRACO, which is based on the coding of gradients and algorithmic redundancy, i.e., each worker evaluates redundant gradients, such that the accurate gradients may be derived even in the presence of adversarial nodes. The simulation results show that DRACO is more than 3 times faster in achieving 90\\\\% test accuracy for gradient computations on MNIST dataset as compared to the geometric median method.\\n\\\\begin{table}\\n\\\\caption{\\\\small Comparison of BGW, LCC, Harmonic coding schemes for Gradient-type computations.}\\n\\\\label{table:lccharmonic}\\n\\\\centering\\n\\\\begin{tabular}{|>{\\\\centering\\\\arraybackslash}m{2.0cm}|c|c|c|}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n & \\\\textbf{Sharmir} & \\\\textbf{LCC} & \\\\textbf{Harmonic} \\\\\\\\ \\\\hline\\n Min. number of workers & $K($deg $g+1)$ & $K$deg $g$+1 & $K($deg $ g-1) +2$ \\\\\\\\ \\\\hline\\n \\\\end{tabular}\\n\\\\end{table}\\nAn improvement to the studies of is done in , which proposes LCC to achieve an optimum tradeoff between resiliency against straggling workers, security against malicious workers, and information-theoretic privacy amid colluding workers. In LCC, the dataset of the master is encoded using the Lagrange polynomial to create computational redundancy. Then, the coded data is shared to the workers for computation on the encoded data, as if the coding did not take place. In comparison with the BGW MPC scheme , LCC requires more workers. However, the Lagrange polynomial based encoding leads to a reduction in the amount of randomness required to encode the data, which translates to lower storage and computation costs incurred by each worker. The LCC also outperforms the BGW based polynomial sharing in terms of communication costs, given that the polynomial sharing scheme requires a communication round for each bilinear operations. In addition, LCC is less computationally costly than DRACO , which does not utilize the algebraic structure of the encoded gradients. However, the Lagrange coding only works for computations involving arbitrary multivariate polynomial functions of the input dataset. As an extension, the study of proposes CodedPrivateML which adopts polynomial approximations to handle the non-linearities of the gradient computation when the sigmoid function is involved, such that logistic regression can be conducted on LCC-encoded data while providing information-theoretic privacy for the dataset. Given that the advantages of the LCC is preserved, the experiments conducted on Amazon EC2 clusters validate that the proposed scheme is close to 34 times faster than the BGW based MPC approaches.\\nIn light of the growing popularity of machine learning, the study of proposes Harmonic coding for tasks specific to gradient-type computations, e.g., for loss function minimization in distributed model training. Harmonic coding leverages on the structure of the intermediate partial gradients computed to enable the cancellation of redundant results, such that the encoding and decoding process is more efficient. As such, for the same level of privacy constraint, Harmonic coding improves on Shamir's secret sharing scheme and LCC in terms of requiring fewer workers to compute gradient-type functions. This result is further summarized in Table \\\\ref{table:lccharmonic}, where we present a comparison of the minimum number of workers required for the discussed schemes. Note that $K$ refers to the number of partitions of the input dataset, $g$ refers to the fixed multivariate polynomial , and deg refers to the degree of $g$. Like LCC, Harmonic coding can also be applied universally to any gradient-type function computation. As such, the data encoding can be performed before the computing task is specified, thus further reducing the delay in computation.\",\n \"id\": \"8d724f64-c2a5-4e3f-a151-1bdc01b1a822\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"ee3d99f5-815c-4ba8-b336-968d0def830a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Secure Coding for Distributed Computing\"\n ],\n [\n \"subsection\",\n \"Secure Distributed Computing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Secure Distributed Computing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Matrix multiplication is a key operation in many popular machine learning algorithms , e.g., principal component analysis , support vector machines , and gradient-based computations. While the reviewed studies in Section~\\\\ref{subsec:securedist} discuss coded computing for privacy preservation in general operations, the studies to be discussed consider tailored strategies for SDMM.\\nIn and , the authors propose the use of staircase codes in place of linear secret sharing codes, e.g., Shamir's codes . As an illustration, we consider that a master encodes its data $A$ with random matrix $R$ into three secret shares before transmitting a share to each of the workers to perform matrix multiplication. When linear secret sharing codes are used, the data and random matrix are not segmented but instead encoded and transmitted as a whole (Table~\\\\ref{table:staircase}). As such, the master has to wait for two \\\\textit{full} responses from any two of three workers before being able to decode and derive the desired results. In contrast, when the staircase code is used, the data and random matrices are segmented into sub-shares before transmission to the workers. When sufficient sub-tasks have been completed by the workers, the master can then instruct the workers to cease computation. Clearly, the staircase coding approach reduces the computation cost of workers and communication costs incurred by the master node. Accordingly, the staircase coding approach can outperform the classical secret sharing code in terms of mean waiting time. With 4 workers considered, experiments conducted on the Amazon EC2 clusters show a 59\\\\% decrease in mean waiting time using staircase codes. \\nHowever, and still consider the case of one-sided privacy, i.e., the approach is designed to keep only one of two input matrices involved in SDMM operations private. As such, several studies have shifted the focus towards applying coding techniques to the specific case in which two-sided privacy, i.e., in which both input matrices are private, is ensured. Among the first such study is that of , which applies the aligned secret sharing for two-sided privacy. Specifically, the input matrices are split into submatrices and encoded with random matrices. Then, the undesired terms are aligned such that the server only recovers the desired results and saves on communication costs. This leads to an improved download rate, i.e., the ratio of size of desired result to the amount of information downloaded by a user, over conventional secret sharing schemes. \\nFollowing , the study of proposes an inductive approach to find a close-to-optimal partition of the input matrices in consideration of two metrics namely the download rate and the minimum number of required workers. The proposed scheme improves on download rate, number of tolerable colluding servers, and computational complexity as compared to the study of . Inspired by , the polynomial coding scheme is also extended to SDMM operations and specifically convolution tasks in , while preserving two-sided privacy, download rate similar to that of , and further mitigating the straggler effects. For convolution tasks, the authors leverage on the inherent property in which the sums of convolutions of sub-vectors may be used to derive the convolution result. Then, the upper and lower bounds of the recovery threshold is derived to show that an order-optimal recovery threshold is achieved, i.e., it does not scale with number of workers. \\nHowever, the key weakness of and as indicated in is that the proposed theoretical results do not clarify the effect of matrix dimensions on the download rate, i.e., the download rates are derived in the case whereby matrices are simply assumed to have large dimensions but without any specifications otherwise. However, the study of found that the results for may be violated in some cases for differing relative dimensions of the input matrices. Under this context, the model proposed in allows the matrix dimensions to be specified, and a new converse bound for the two-sided security SDMM is derived.\\nIn general, the encoded results of matrix multiplication are sent to the master, where interpolation is performed to obtain the multiplication results, i.e., coefficients of a polynomial. In , the encoding of the private matrices is such that the coefficients are mainly non-zero. In contrast, the study of propose the Gap Additive Secure Polynomial (GASP) codes such that there are as many zero coefficients as possible. This allows the product to be interpolated and the desired results to be derived with fewer number of evaluations performed, which implies that fewer workers are required to perform the matrix multiplication. To assign the exponents for decodability while having as many zero coefficients as possible, the authors propose the degree table to solve the combinatorial problem. In , the authors further generalize the GASP codes to be applicable to different partitions of input matrices and security parameters. The GASP codes are shown to outperform the approaches in in terms of download rate.\\n\\\\begin{table}[!]\\n\\\\caption{\\\\small Comparison of linear secret sharing codes and staircase codes in the distribution of computation tasks in a system with three workers .}\\n\\\\label{table:staircase}\\n\\\\centering\\n\\\\scalebox{0.9}{\\n\\\\begin{tabular}{ll ll l}\\n\\\\hline\\n\\\\rowcolor{mygray}\\n\\\\multicolumn{1}{|l|}{} & \\\\multicolumn{1}{l|}{$S_1$} & \\\\multicolumn{1}{l|}{$S_2$} & \\\\multicolumn{1}{l|}{$S_3$} \\\\\\\\ \\\\hline\\n\\\\multicolumn{1}{|l|}{\\\\begin{tabular}[c]{@{}l@{}}Linear secret\\\\\\\\ sharing code\\\\end{tabular}} & \\\\multicolumn{1}{l|}{$R$} & \\\\multicolumn{1}{l|}{$R+A$} & \\\\multicolumn{1}{l|}{$R+2A$} \\\\\\\\ \\\\hline\\n\\\\multicolumn{1}{|l|}{Staircase code} & \\\\multicolumn{1}{l|}{\\\\begin{tabular}[c]{@{}l@{}}$A_1+2A_2+4R_1$\\\\\\\\ $R_1+R_2$\\\\end{tabular}} & \\\\multicolumn{1}{l|}{\\\\begin{tabular}[c]{@{}l@{}}$A_1+2A_2+4R_1$\\\\\\\\ $R_1+2R_2$\\\\end{tabular}} & \\\\multicolumn{1}{l|}{\\\\begin{tabular}[c]{@{}l@{}}$A_1+2A_2+4R_1$\\\\\\\\ $R_1+3R_2$\\\\end{tabular}} \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n}\\n\\\\end{table}\",\n \"id\": \"d523704a-b56b-49c2-97c7-6500404f0ac2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"ee3d99f5-815c-4ba8-b336-968d0def830a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Secure Coding for Distributed Computing\"\n ],\n [\n \"subsection\",\n \"Secure Distributed Matrix Multiplication (SDMM)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Secure Distributed Matrix Multiplication (SDMM)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we have discussed studies that have adopted a coded computing approach towards ensuring secure distributed computing. Then, we discuss the specific applications of SDMM, which require two-sided privacy. The summary and lessons learned are as follows:\\n\\\\begin{itemize}\\n\\\\item Originally proposed to mitigate stragglers in distributed computing systems, coding approaches have also been extended to ensure information theoretic privacy and for some studies, security against malicious workers . However, given that the coding techniques, e.g., polynomial codes, are designed with the intention to mitigate stragglers, the studies discussed in this section have focused on modifications to the approaches to simultaneously achieve both straggler mitigation and privacy. As was mentioned in this section, the coding approaches to ensure privacy preservation have outperformed conventional secret sharing schemes that mainly focused on data correctness and privacy. \\n\\\\item With the recent popularity of machine learning, several studies have also focused on tailoring the coding approaches to be utilized for various machine learning tasks, e.g., entangled polynomial coding for logistic regression and harmonic coding for gradient based computations . In particular, given the importance of two-sided privacy in SDMM, we have also specifically discussed papers that focus on two-sided privacy separate from papers that only considered one-sided privacy. In general, papers that focus on taking advantage of the algebraic structure of the specific operations , e.g., convolution tasks or gradient computations , for efficient encoding and decoding usually perform better.\\n\\\\item In most of the studies that we have discussed, the focus tends to be on information-theoretic privacy. However, weak security is key and may be explored in future works. Specifically, a system is weakly secure if attackers are unable to learn the sensitive intermediate values without having received a certain number of coded packets . This may be relevant when there exist eavesdroppers with an access to communication links between the master and workers . For example, in the study of , a data shuffling design and redundancy reduction algorithm to assign computing tasks have been proposed to ensure weak security in the system. However, the focus has been on workers, rather than eavesdroppers that may not be involved in the computations.\\n\\\\item Given that the straggler effect is a fundamental concern in distributed computing, most papers have considered workers with heterogeneous computing capabilities. However, as discussed in , the issue of heterogeneous networks in other aspects have been under-explored in the aforementioned studies. For example, the workers may have different levels of reputation . This enables the adoption of context-sensitive solutions, e.g., data security may be guarded against new workers or workers with low reputation, whereas it may not be required for trusted workers. With this, the computation complexity and duration may be reduced in trusted nodes.\\n\\\\end{itemize}\",\n \"id\": \"362bc888-78c5-42b5-b72f-1d88622fe4e4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"ee3d99f5-815c-4ba8-b336-968d0def830a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Secure Coding for Distributed Computing\"\n ],\n [\n \"subsection\",\n \"Summary and Lessons Learned\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary and Lessons Learned\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:applications}\\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{nfv.pdf}\\n\\\\caption{\\\\small Application of CDC to Network Function Virtualization (NFV) model for uplink channel decoding.}\\n\\\\label{fig:nfv}\\n\\\\end{figure}\",\n \"id\": \"825a0578-3fe6-4ffc-8428-dfabbdc2d42a\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"CDC Applications\"\n ]\n ],\n \"subsections\": [\n \"6f2cf0e3-4854-4438-b7c6-37b57226ff9d\",\n \"f8c9cb00-a005-4862-b9c8-df9e98e5996f\",\n \"eb23bbfe-351d-4db3-8737-3055dd885cdb\"\n ],\n \"title\": \"CDC Applications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Network Function Virtualization serves as an enabling technology for optimizing the 5G as well as emerging 6G networks , presenting a promising paradigm shift in the telecommunication service provisioning industry. By leveraging on virtualization technologies, NFV simplifies the management and operations of the networking services. In particular, NFV decouples the Network Functions (NFs) such as routing and baseband processing from the physical network equipment on which they operate. The NFs are mapped to the Virtual Network Functions (VNFs) that are supported by the Commercial Off-The-Shelf (COTS) physical resources which provide storage, networking and computational capabilities. As the software component in the network is decoupled from the hardware component by a virtualization layer, the VNFs can be easily implemented over the distributed network locations which have the required hardware resources. Due to the flexibility in the deployment of the VNFs, the NFV brings about significant reduction in operating and capital expenses. Moreover, the development of new networking services is faster and cheaper as the COTS resources can be instantiated easily to provide the required network connection services. Extensive literature has been carried out on the various aspects of NFV such as architectural designs , resource allocation , energy efficiency , performance improvement and security . More details can be found in .\\nHowever, one of the limiting factors of the performance of NFV lies in the reliability of the COTS hardware resources . Hardware failure due to several factors such as component malfunctioning, temporary unavailability and malicious attacks affects the implementation of NFV, hindering the provision of services. Apart from the fault-tolerant virtualization strategies that are based on the diversity approach which maps VNFs onto the various virtual machines (VMs) such that the probability of a disruptive failure is minimized , coding approach can also be used to minimize computation latency in NFV. \\nIn the study of , the authors consider the highly complex uplink channel decoding of the Cloud Radio Access Network (C-RAN) architecture , which is a key application of NFV. In this system, the users communicate with the cloud via Remote Radio Head (RRH). In order to ensure the reliability of channel decoding, the received data frames by the RRH are encoded by leveraging on their algebraic structures before being distributed to the different VMs as shown in Fig.~\\\\ref{fig:nfv}. The simulation results show that the coding approach achieves lower probability of error for decoding at the cloud than that of the diversity-based approach. However, there are several assumptions that are made in this proposed scheme. Firstly, the binary symmetric communication channel is assumed between the users and the RRH. By considering other communication channels such as the additive Gaussian noise channels, different coding techniques may be applied. Secondly, this simple framework works well for a network with three processing nodes, but its performance for larger-size networks is not guaranteed. \\nConsidering the same issue of uplink channel decoding in the C-RAN architecture, the authors in propose a more generalized coded computation framework that works for any number of servers, random computing runtimes and random packet arrivals by adopting the coding approach proposed in . Given the randomness in the arrivals of data frames transmitted by the users, two queue management policies are considered: (i) per-frame decoding where one frame is decoded at any point of time, and (ii) continuous decoding where the servers start to decode the next packet of data frame upon completion of the first packet. There is an underlying tradeoff between the average decoding latency and the frame unavailability probability which is an indication of the reliability of the decoding process at the servers. The simulation results show that properly designed NFV codes are useful in achieving the desired tradeoffs by optimizing the minimum distance of the codes.\\nThe idea of adopting coding approach in NFV is relatively new. Different coding techniques can be explored in the future. Besides, instead of addressing the uplink channel decoding, which is the most computationally-intensive baseband function , other network functions such as routing and security functions can be considered.\",\n \"id\": \"6f2cf0e3-4854-4438-b7c6-37b57226ff9d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"825a0578-3fe6-4ffc-8428-dfabbdc2d42a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"CDC Applications\"\n ],\n [\n \"subsection\",\n \"Network Function Virtualization (NFV)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Function Virtualization (NFV)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"With the enhanced sensing capabilities of end devices, an overwhelming amount of data is produced at the edge of the network today. Traditional schemes of computation offloading to the cloud is thus unsustainable. Moreover, certain edge applications may involve end devices in remote areas that have limited connectivity. This necessitates a paradigm shift towards edge computing, in which computation is performed closer to the edge of the network where data is produced. However, resource-constrained devices may not be able to carry out complex computations individually, especially given the increasing size and complexity of state-of-the-art AI models . As such, one of the enabling technologies of edge computing is cooperative computation in which the available resources of end devices and edge nodes, e.g., road side units in vehicular networking, can be pooled together to execute computation intensive tasks collaboratively . For ease of exposition, we refer to these participating end devices and edge nodes as workers in this section.\\nAs the number of devices connected to the network increases, more information needs to be exchanged among the workers, resulting in high communication load. However, the communication bandwidth is fixed and thus the network is unable to handle the high communication load, causing a bottleneck as a result. Moreover, the heterogeneous nature of workers in the edge computing paradigm, e.g., in terms of computational and communication capabilities, can lead to the straggler effects. In face of these challenges, coding techniques can be used.\\nIn the study of , the \\\\emph{coded wireless distributed computing} (CWDC) framework is proposed. The system model consists of multiple devices, i.e., workers, involved in cooperative computation. As an illustration, a worker may have an input, e.g., an image, that has to be processed, e.g., for object recognition. Individually, a worker may not have the storage or computational capabilities to execute the task. Therefore, the inference model may be split and stored on each worker, whereas the cooperative computation of results can be implemented following the MapReduce framework as discussed in Section~\\\\ref{subsec:mapreduce}. An access point, e.g., a base station or a Wi-Fi router can then be utilized to facilitate the exchange of intermediate results among workers. The proposed framework achieves communication loads that are independent of the size of the network and the storage size of the workers. Moreover, the CWDC framework can be generalized and applied to different types of applications.\\nIn practical distributed computing systems, the workers may have heterogeneous computational, communication, and storage capabilities. Based on a similar system model proposed in where the workers communicate with each other via an access point, the study in considers devices with heterogeneous storage capacities over wireless networks. For uplink transmission, the allocation of files is based on the scheme proposed in as previously discussed in Section~\\\\ref{subsec:files} whereas for downlink transmission, data is encoded at the access point for the reduction of downlink communication load. However, the achievable scheme has only been validated in a small network that consists just three processing nodes. \\n\\\\begin{figure}\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{flmodel.pdf}\\n\\\\caption{\\\\small Illustration of the Coded Federated Learning (CFL) scheme in a FL system that consists of multiple data owners and a FL model owner.}\\n\\\\label{fig:flmodel}\\n\\\\end{figure}\\nIn light of the growing popularity of machine learning model training at the edge, the study of considers the distribution of gradient descent computations across workers in the network to train a linear regression model. The proposed heterogeneous coded gradient descent (HCGD) assigns each worker with an optimal load partition, through modelling the computation delay of devices with a shifted exponential distribution. In consideration of data privacy, the authors in propose the Coded Federated Learning (CFL) approach for privacy-preserving linear regression. Federated Learning (FL) is a privacy preserving distributed machine learning paradigm proposed in , in which the sensitive data of data owners are kept locally, whereas only the model parameters are transmitted to the central server for aggregation to update the model. However, FL still suffers from the issues of straggling devices and communication inefficiency . As such, in the proposed CFL scheme, each data owner first generates parity data from its local data for transmission to the central server. At the central server, the gradients are also computed from parity data simultaenously, such that only a subset of gradients from the data owners have to be received for completion of model update. Figure~\\\\ref{fig:flmodel} illustrates the implementation of the CFL scheme. The CFL approach is also shown to converge almost four times faster than an uncoded approach. However, the communication and computation cost involved in generating and transmitting the parity data has not been well elaborated in the study. Clearly, a major weakness of aforementioned studies of is that they can only be applied to linear regression model training. The study of extends on the aforementioned works with the proposed CodedFedL designed to mitigate straggler mitigation in non-linear regression and classification tasks.\\nIn some cases, the resource level of the devices may not be known by the network operator. To enable a dynamic and adaptive coded sub-task allocation for cooperative computation, an Automatic Repeat reQuest (ARQ) mechanism is proposed in the study of , in which devices are allocated with specific levels of packets for computation based on their responsiveness. Specifically, devices that are more responsive are assumed to have more available resources. These devices will thus be assigned with more sub-tasks for computation.\\nInstead of reducing communication load, it is also important to consider communication efficiency, i.e., achieved data rates, especially in wireless networks that have limited spectral resources or networks with mutual interference among users. In order to improve spectral efficiency, the co-channel communication model is proposed in the study of which consists of two stages, i.e., the uplink multiple access stage and the downlink broadcasting stage. The communication model turns out to be equivalent to a multiple-input-multiple-output (MIMO) interference channel. Interference alignment has been an effective approach in handling mutual interference among users. The signals are precoded into the same subspaces at the unintended receivers and the desired signals are recovered at the intended receivers by using the decoding matrix. Linear coding scheme is adopted to establish the conditions for interference alignment. A low-rank optimization problem is formulated to minimize the number of channel used subject to the established interference alignment conditions. By solving this optimization problem, the achievable symmetric degree-of-freedom (DoF), which implies the extent to which interference is eliminated, can be maximized. In , an efficient difference-of-convex-functions (DC) algorithm based on a DC representation for the rank function is proposed to solve the low-rank optimization problem. The performance of the DC algorithm is evaluated in two different scenarios by varying: (i) the number of files stored in the devices, and (ii) the number of antennas that are equipped by the devices. In the first scenario, as the number of files stored in the devices increases, the achievable DoF increases. Similarly in the second scenario, as the number of antennas equipped by the devices increases, the achievable DoF increases. Furthermore, the simulation results show that the DC approach achieves higher DoF than the existing benchmark algorithms, e.g., iterative reweighted least squares (IRLS) algorithm and the nuclear norm relaxation approach. However, the proposed scheme is based on a homogeneous network, i.e., the number of files stored is the same across all devices and all devices are equipped with the same number of antennas. As an extension, heterogeneous networks can be considered. \\nInstead of communicating with each other via an access point, the devices can communicate directly with each other over wireless interference channels . In particular, the transmission of data in the Shuffle phase operates over wireless interference channels. While the CDC scheme in allows communications based on time-division multiple access (TDMA) scheme which allows each processing node to transmit one coded information packet at any time, the one-shot linear scheme adopted in allows more than one processing node to transmit information simultaneously at any given time slot. The transmitted symbols are linear combinations of coded intermediate results from the processing nodes. The transmitted symbols are broadcasted to the processing nodes, following which the nodes can decode to recover the desired information. The study in characterizes an improved computation-communication tradeoff as compared to the study in . However, the proposed scheme operates under the assumption of perfect channel state information (CSI) where the CSI is available to all processing nodes. As such, the authors in propose superposition coding scheme which has better performance than that of the CDC scheme and the one-shot linear scheme under imperfect CSI condition. The study is extended to account for the presence of stragglers in the networks .\\nIn general, the aforementioned studies have considered conventional implementations of the CDC scheme in which each device has to complete the computation task first, before transmitting the intermediate results, e.g., to other devices via the access point. However, given the limited computational capabilities of devices, the latency involved can still be significant. As such, the studies in and consider the batch-processing based coded computing (BPCC) framework which allows each device to return only the partially completed results to a master node in batches, even before the task is fully completed. These partial results are then utilized to generate approximations to the full solution through the singular value decomposition (SVD) approach . This is particularly useful in applications that require fast, but not necessarily optimized, results. The BPCC scheme has its effectiveness validated on the EC2 computing cluster, in which latency is proven to be reduced. Moreover, in consideration of the energy limitations of Unmanned Aerial Vehicles (UAV), the BPCC framework was proposed for UAV-based mobile edge computing to provide energy-efficient computation offloading support on-demand .\",\n \"id\": \"f8c9cb00-a005-4862-b9c8-df9e98e5996f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"825a0578-3fe6-4ffc-8428-dfabbdc2d42a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"CDC Applications\"\n ],\n [\n \"subsection\",\n \"Edge Computing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Edge Computing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we have discussed applications of CDC in NFV and edge computing. The lessons learned are as follows:\\n\\\\begin{itemize}\\n\\\\item The convergence of the recently popular edge computing and machine learning has given rise to edge intelligence, in which the computational capabilities of edge servers and devices have been increasingly leveraged to conduct machine learning model training closer to where the data is produced. While this leads to several benefits, e.g., lower training latency and enhanced privacy preservation, the problem of straggling devices is still a bottleneck. As such, CDC approaches have been increasingly applied in this context.\\n\\\\item One major difference between distributed computing at the edge and at computing clusters is that the end devices and edge servers, i.e., workers, are not specifically dedicated to computations. For example, the workers may only share a fraction of their processing power at time slots in which they are idle. As such, the heterogeneity in computational capabilities of workers may also be greater in this regard. In face of these challenges, optimal load partition and allocation strategies have been explored. Moreover, for the future works, the dynamics of the system may be captured using Deep Reinforcement Learning based resource allocation .\\n\\\\item In edge intelligence, there may be several data owners involved. Moreover, the data owners can be devices with computational constraints. As such, conventional techniques of data encoding before transmission to workers for computation may not be feasible. To meet this challenge, FL has been proposed in the work of , whereas CFL and CodedFedL have been proposed to mitigate the straggler effects in FL. However, the proposed methods are highly restrictive. For example, the CFL can only be applied to linear regression problems, whereas both methods require costly computation and transmission of parity data. Moreover, they have not been implemented on typical end devices to validate the feasibility of the schemes in practical implementation. For the future works, studies on using CDC schemes in edge computing applications may adopt the approach of , in which the schemes are implemented under the context of practical hardware constraints.\\n\\\\end{itemize}\",\n \"id\": \"eb23bbfe-351d-4db3-8737-3055dd885cdb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"825a0578-3fe6-4ffc-8428-dfabbdc2d42a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"CDC Applications\"\n ],\n [\n \"subsection\",\n \"Summary and Lessons Learned\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary and Lessons Learned\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:challenges}\\nThe utilization of CDC schemes to solve the implementation challenges of the distributed computing systems is a relatively new and recent approach. There are still challenges and open issues that have yet to be addressed and this provides opportunities for new research directions in the future. We present major challenges that need to be looked into for effective implementation of CDC schemes.\\n\\\\begin{itemize}\\n\\\\item \\\\emph{Heterogeneous nodes: }As compared to traditional distributed computing clusters, heterogeneities among computing nodes are much more significant when the computing nodes are connected in the edge computing networks, e.g., smartphones and wearable devices. Many studies, e.g., , , , have considered heterogeneous systems where the processing nodes have different computational, communication and storage capabilities. For example, in , a joint file and function allocation strategy is proposed to assign jobs to the processing nodes such that the communication load is minimized. In , the computation load is allocated based on the capabilities of the processing nodes. However, other aspects of heterogeneity such as the reputation and the willingness to participate of the processing nodes are not taken into account. New allocation strategies need to consider different aspects of heterogeneities of the processing devices so that coding techniques can be implemented effectively and securely. For example, the computation tasks can be allocated to workers with higher reputation which implies higher probability of the workers in completing their allocated tasks.\\n\\\\item \\\\emph{Encoding and decoding complexities: }The studies that we have discussed in Section~\\\\ref{sec:comms} and Section~\\\\ref{sec:stragglers} mainly minimize communication load in Shuffle phase and computation latency in the Map phase respectively. However, the complexities of encoding and decoding are often not evaluated. It is important to ensure low encoding and decoding complexities in order to minimize the overall job execution time. Otherwise, the speedup gain achieved by in specific phases, e.g., communication and computation phases may be offset by the high encoding and decoding complexities. For example, the UberShuffle algorithm incurs high computational overhead under the fast broadcast environment, i.e., networks with large bandwidth, such that it is not feasible for implementation even though it achieves significant shuffling gain. Hence, to better assess the performance of the CDC schemes, the complexities of the encoding and decoding methods have to be evaluated.\\n\\\\item \\\\emph{Non-static computing nodes: }For the commonly used distributed computing models such as cluster computing, grid computing and cloud computing that are discussed in Section~\\\\ref{sec:fundamentals}, the computing nodes are static, i.e., the nodes are located at fixed location. For example, the servers are located at specific data centers. Data required for computations is transmitted over the wireless communication channels to the servers. However, as the edge devices, e.g., IoT devices, wearable devices and vehicles, have greater communication and computational capabilities, new distributed computing models such as mobile edge computing , and fog computing have been developed recently. In , the basic CDC scheme is implemented in the context of fog computing. However, the edge devices are usually mobile. The data that is processed by the edge devices depends on the locations that they visit , and hence the master node has no control of the data distribution to the workers. New coding approaches for edge and fog computing which involve moving workers can be proposed in future works. One proposed solution is to allocate Reduce functions based on the data stored at the processing nodes .\\n\\\\item \\\\emph{Security concerns: }Coding techniques are able to mitigate the straggler effects while preserving privacy as shown in the studies of , and . The proposed secure coding techniques are an extension to the coding techniques that are originally proposed to mitigate the straggler effects. As mentioned in , the tradeoff between resiliency against straggling workers, security against malicious workers and information-theoretic privacy amid colluding workers needs to be carefully managed. In addition, there may exist eavesdroppers which tap on the less secure communication links between the master node and the workers. As such, more research effort can be directed towards developing weakly secure systems which prevents the eavesdroppers from retrieving sensitive information. \\n\\\\item \\\\emph{Network architectures: }It is important to consider how the computing nodes are connected and communicate with each other for effective implementation of the CDC schemes. For example, in and , the authors introduce a hierarchical structure where the master communicates with multiple submaster and each submaster leads a group of computing nodes. From the studies that we have reviewed, the network architecture is only considered for the implementation of CDC schemes to reduce communication load. However, it is an important consideration when designing secure coding schemes. In practice, it may be safe for computing nodes within a group, e.g., from the same location, to share information freely with each other but not with computing nodes from another group, e.g., from different location. In addition, the communication channels are not perfect, i.e., they may not have perfect CSI or the transmitted information may have missing entries. Some networks may have limited spectral resources or suffer from mutual interference among users . As such, future research can work towards designing coding schemes that can be implemented in practical distributed computing systems. Besides a need to design effective coding schemes, there is also a need to consider the design of low-cost, easily-implementable and scalable network architectures in which the coding schemes can be applied to.\\n\\\\item \\\\emph{Different computation frameworks: }Currently, most of the studies are based on the MapReduce computing model. Specifically, Coded MapReduce is proposed in by implementing coding techniques in the MapReduce framework. However, there are limitations to MapReduce model that hinder its wide adoption for all types of distributed computation tasks as explained in Section~\\\\ref{subsec:mapreduce}. In fact, there are other computing models such as Spark, Dryad and CIEL which support iterative algorithms, in which the feasibility of implementing coding techniques has not been explored. As such, the importance of these computing models motivates future directions such as the design of coding schemes that are specific to these computing models in order to solve any distributed computation tasks, e.g., convolution, Fourier transform and non-linear computations. \\n\\\\item \\\\emph{Coding for both communication reduction and stragglers mitigation: }As we have discussed previously in Section~\\\\ref{sec:comms} and Section~\\\\ref{sec:stragglers}, coding techniques are used to either reduce communication load or mitigate the straggler effects. However, coding techniques cannot be applied to solve these implementation challenges simultaneously. As characterized in , there is a tradeoff between communication load and computation latency. However, in an ideal situation, both communication load and computation latency need to be minimized. Thus, it is important to carefully manage the tradeoff to achieve optimal performance of the distributed computing systems. For future works on CDC schemes, there is a need to improve the latency-communication tradeoff curve so that the time taken to execute the allocated tasks can be significantly reduced. In addition, instead of the two-dimensional tradeoff, the three-dimensional tradeoff between computation, communication and storage cost , which is much more challenging to manage, should be considered.\\n\\\\item \\\\emph{CDC applications: }Given the advantages, CDC schemes are implemented in distributed computing applications such as the NFV and edge computing. Apart from the UAVs, CDC schemes can be extended to edge computing applications in other areas, e.g., vehicular networks, healthcare systems and industrial operations. For the studies that we have reviewed in CDC applications, the main focus lies in the implementation of coding techniques in various applications, without considering privacy and security. Given the importance of secure coding as discussed in Section~\\\\ref{sec:security}, secure coding techniques need to be considered in the implementation of CDC applications. Besides, application-specific issues need to be addressed. For example, in vehicular networks where the vehicles are constantly moving, the CDC schemes need to be robust to vehicles which do not have consistent access to the wireless communication channels.\\n\\\\end{itemize}\\nThe idea of using coding techniques to overcome the challenges in distributed computing systems is relatively new. For effective implementation in practical distributed computing systems, various aspects such as the heterogeneities of the computing nodes and the network architectures are worth the in-depth studies. Some promising research directions presented in this survey serve as useful guidelines and valuable references for future research in CDC.\",\n \"id\": \"8c1cb679-0c49-4078-b259-687a2df3303d\",\n \"level\": \"section\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Challenges, Open Issues and Future Works\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Challenges, Open Issues and Future Works\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclusion}\\nIn this paper, we provided a tutorial of CDC schemes and a comprehensive survey on the two main lines of CDC works. We first motivated the need for CDC schemes. The current performance of distributed computing systems can be improved using coding schemes. Then, we described the fundamentals and principles of CDC schemes. We also reviewed CDC works which aim to minimize communication costs, mitigate straggler effects as well as enhance privacy and security. In addition, we discussed the implementation of CDC schemes in practical distributed computing applications. Finally, we highlighted the challenges and discussed promising research directions.\\n\\\\bibliographystyle{ieeetr}\\n\\\\bibliography{cdc}\\n\\\\end{document}\",\n \"id\": \"5a390405-bf12-44a0-aa67-38cecb4ab9be\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"43fb2934-e7bf-4933-946e-becc16dd4285\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Coded Distributed Computing\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":83,"string":"83"},"all_cites":{"kind":"list like","value":[],"string":"[]","isConversational":false},"Bertopic_CD":{"kind":"null"}}},{"rowIdx":49,"cells":{"authors":{"kind":"list like","value":["Ziqiang Li","Muhammad Usman","Rentuo Tao","Pengfei Xia","Chaoyue Wang","Huanhuan Chen","Bin Li"],"string":"[\n \"Ziqiang Li\",\n \"Muhammad Usman\",\n \"Rentuo Tao\",\n \"Pengfei Xia\",\n \"Chaoyue Wang\",\n \"Huanhuan Chen\",\n \"Bin Li\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Systematic Survey of Regularization and Normalization in GANs"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-08-19T12:52:10Z","string":"2020-08-19T12:52:10Z"},"category":{"kind":"string","value":"cs.LG"},"abstract":{"kind":"string","value":" \t\tGenerative Adversarial Networks (GANs) have been widely applied in different scenarios thanks to the development of deep neural networks. The original GAN was proposed based on the non-parametric assumption of the infinite capacity of networks. However, it is still unknown whether GANs can fit the target distribution without any prior information. Due to the overconfident assumption, many issues remain unaddressed in GANs' training, such as non-convergence, mode collapses, gradient vanishing. Regularization and normalization are common methods of introducing prior information to stabilize training and improve discrimination. Although a handful number of regularization and normalization methods have been proposed for GANs, to the best of our knowledge, there exists no comprehensive survey which primarily focuses on objectives and development of these methods, apart from some in-comprehensive and limited scope studies. In this work, we conduct a comprehensive survey on the regularization and normalization techniques from different perspectives of GANs training. First, we systematically describe different perspectives of GANs training and thus obtain the different objectives of regularization and normalization. Based on these objectives, we propose a new taxonomy. Furthermore, we compare the performance of the mainstream methods on different datasets and investigate the applications of regularization and normalization techniques that have been frequently employed in state-of-the-art GANs. Finally, we highlight potential future directions of research in this domain. Code and studies related to the regularization and normalization of GANs in this work is summarized on https://github.com/iceli1007/GANs-Regularization-Review. \t"},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4410407c-7d7c-48f1-8174-a9a95319844f","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"]],"subsections":["29469c09-cd5f-4556-9382-b0a3b4de184e","5fdbf0be-ea34-468a-9943-d62d05b593ca","8469161e-625b-422b-90e0-a6cca9638a04","fdfb947c-602c-4735-9040-96d30810700e","44c32216-8be6-4e5c-aa2f-ca241ea4139f","bc57c604-a3ec-457f-ab6f-07cf5243c526","fb4135aa-ff8b-439d-9a28-a790ffddd379","b359d174-4880-4149-b0f4-5fd0e7bd75f4","f6a7c46a-3c0c-4605-a602-4104833c8854"],"title":"root"},{"cite_extract_rate":0.7213114754098361,"cites":[63,89,73,86,68,85,78,76,8317,53,60,6998,66,57,70,88,80,6996,72,54,58,62,59,7190,55,69,77,6997,82,74,56,81,87,6999,8316,84,71,61,75,67,65,83,64,79],"content":"\\setlength{\\abovecaptionskip}{0.cm}\n\tGenerative adversarial networks (GANs) have been widely used in computer vision, such as image inpainting , style transfer , text-to-image translations , and attribute editing . GANs training is a two-player zero-sum game between a generator and a discriminator, which can be understood from different perspectives: (i) \"Real \\& Fake\" , (ii) \"Fitting distribution\" , and (iii) \"Training dynamics\" . GANs training suffers from several issues, for instance: non-convergence , mode collapses , gradient vanishing , overfitting , discriminator forgetting and deficiency , and hyperparameters sensitivity . Many solutions to mitigate these issues have been proposed, focusing on designing new architectures , loss functions , optimization methods , regularization , and normalization . Among them, regularization and normalization techniques are compatible with loss functions, model structures, and tasks, which has attracted the attention of scholars. \n\tRegularization and normalization are widely applied in neural networks training to introduce prior knowledge. For supervised tasks, regularization in literature has been proposed to introduce some advantages like overfitting prevention , semi-supervised assumptions , manifold assumptions , feature selection , and low rank representation . On the other hand, normalization is advantageous for the Stochastic Gradient Descent (SGD) , accelerating convergence and improving accuracy. Unlike the icing on the cake of supervisory tasks, regularization and normalization are utilized inevitably in weak-supervised and unsupervised tasks. GANs' training is a two-player zero-sum game having a solution to Nash equilibrium. The proposal of standard GAN is based upon the non-parametric assumption of the infinite capacity of networks, an unsupervised learning task. Likewise, a good number of research studies targeting GANs training from different perspectives argue that unconstrained training causes unstable training (generator and discriminator ) and significant bias between real images and fake images (attributes domain and frequency domain ). Therefore, a large amount of prior should be introduced into GANs training through regularization and normalization\n\t\\begin{figure}\n\t\t\\centering\n\t\t\\includegraphics[scale=.52]{figs/summary_new.pdf}\n\t\t\\caption{The overview of different perspectives of GANs training and the summary of the regularization and normalization for GANs.}\n\t\t\\label{fig:overview}\n\t\\end{figure}\n\tRegularization and normalization are effectively used to stabilize training and improve the performance of GANs in existing literature . Due to diverse nature of the topic, there is a need for systematic literature survey. There exist some literature studies , however, these studies either lack comprehensive coverage of the topic, miss detailed background information and theoretical analysis, or do not correlate different methods. In this paper, based on the different perspectives of GANs training, we propose a new taxonomy, denoted as \\textbf{\"Real \\& Fake\"}, \\textbf{\"Fitting distribution\"}, \\textbf{\"Training dynamics\"}, and \\textbf{\"Other methods\"}, for a better understanding of regularization and normalization during the GANs training as depicted in Figure \\ref{fig:overview}. {\\textbf{\"Real \\& Fake\"} is the low-level (intuitive) perspective of GANs, in which GANs is considered as \"counterfeiters-police\" competition. At this level, D estimates the real probability of both real and fake samples, which is similar to the bi-classification task. Therefore, prior information and additional supervision tasks in classification task are also urgent during the training process of the discriminator. Based on these, some regularization methods, such as {\\textit{Data Augmentation and Preprocessing}}}\\footnote{Data augmentation and preprocess introduce additional data and prior, which is similar to regularization. More importantly, both consistency regularization and self-supervision need different data transformation operations. Hence, this paper also discusses some works on this.}, {\\textit{Consistency Regularization}, and {\\textit{Self-Supervision}} are proposed to mitigate overfitting, improve the representation of discriminator, and avoid discriminator forgetting by introducing additional supervised information and data; \\textbf{\"Fitting distribution\"} is the middle-level perspective of GANs. At this level, generator is considered as a distribution mapping function, and the discriminator is a distribution divergence. Among various distances, Wasserstein distance is a popular form, and Lipschitz continuity is a necessary condition for achieving Wasserstein distance. Based on these, \\textit{Gradient Penalty}, \\textit{Weight Normalization}, \\textit{Weight Regularization}, and \\textit{Gradient} \\textit{Normalization} are used to fulfill Lipschitz continuity and ensure training stability of discriminator; \\textbf{\"Training Dynamics\"} is the high-level (essential) perspective of GANs. At this level, GANs training is a two-player zero-sum game with a solution to Nash Equilibrium. To achieve theoretical local convergence, \\textit{Jacobian Regularization} needs to be used; Finally, \\textbf{\"Other methods\"} containing \\textit{Layer} \\textit{ Normalization} and \\textit{Inverse Gradient Penalty} are used for conditional generation and easing mode collapse, respectively}\n\tIn summary, we make the following contributions in this survey:\n\t\\begin{itemize}\n\t\t\\item \\textit{Comprehensive analysis of GANs training.} In this study, we analyze the GANs training from three perspectives including \"Real \\& Fake\", \"Fitting distribution\", and \"Training dynamics\". To the best of our knowledge, this survey is the first in this domain with comprehensive analysis.\n\t\t\\item \\textit{New taxonomy.} Based on the analysis of GANs training from different perspectives, we propose a novel taxonomy and contextualize the regularization and normalization of GANs comprehensively.\n\t\t\\item \\textit{ Comparison and analysis.} Following the taxonomy, we also provide quantitative and qualitative analysis and comparison for each type of regularization and normalization techniques, which has helped the researchers and practitioners navigate this space.\n\t\\end{itemize}\n\t{\\textbf{The Scope of This Survey.} This survey aims to systematically analyze the prevalent problems in the GANs training, such as non-convergence, mode collapses, gradient vanishing, and discriminator overfitting. Accordingly, different regularization and normalization technologies have been summarized. Of course, not all regularization and normalization technologies for GANs are covered in this survey. In some case, some regularization and normalization technologies are highly dependent on the task and data in the hand, we recommend looking for domain-specific regularization and normalization techniques from the following reviews: data-efficient generation , medical image generation, image super-resolution , biomedical informatics generation , Spatio- temporal data generation , text generation . Our survey is concerned with general technologies in GANs, which are not dependent on model structures, data, and task. We hope our study can provide general and universal insights of GANs for the community.}\n\tThe rest of this paper is organized as follows: Section 2 introduces the background and different training perspectives of GANs. Section 3, 4, 5, and 6\n\tdescribe regularization and normalization methods in different groups, respectively. Furthermore, we investigate the applications of regularization and normalization techniques that have been frequently employed in SOTA GANs in Section 7 and discuss the current problems and prospects for future work in Section 8.","id":"29469c09-cd5f-4556-9382-b0a3b4de184e","level":"section","origin_cites_number":61,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"","id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","level":"section","origin_cites_number":0,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"]],"subsections":["5a1f5e65-9650-482a-9b5a-2625a5dd0b96","c2fed4a0-e8f0-4ddc-ab66-70cc7bfe76f8","fd688225-4a59-4600-886f-76309f6e8ba9","4e9cd8aa-b10a-44ef-a7b7-7d2880558ce2","a915370f-85fe-4345-a91b-c8d67718061c","a397f8d2-2658-471a-8c3b-cbde88271e09","0edf29a6-bba8-43a2-a318-3f272fcb91d9"],"title":"Background and Three Perspectives of GANs Training"},{"cite_extract_rate":0.27272727272727204,"cites":[91,66,57,90,71,67],"content":"{Regularization and normalization are common and important techniques to introduce prior knowledge in neural networks.} Regularization is a technique to control the complexity of learning models. Weight decay is a typical method to minimize the square of weights together with the training loss in the training of neural networks , which can be used to improve generalization. \n\tIn Bayesian learning methods, such as relevance vector machine , probabilistic classification vector machines , and others , regularization is termed as prior distribution. Specifically, L2 regularization is equivalent to introducing Gaussian prior to the parameters, and L1 regularization is equivalent to introducing Laplace prior to the parameters. The theoretical connection between regularization and prior information has been investigated in neural network ensembles research . Regularization does not only control overfitting but also provide other characteristics like semi-supervised assumptions , manifold assumptions , feature selection , low rank representation , and consistency assumptions . {Normalization is the mapping of data to a specified range, which is advantageous for the Stochastic Gradient Descent (SGD) , accelerating convergence and improving accuracy.}","id":"5a1f5e65-9650-482a-9b5a-2625a5dd0b96","level":"subsection","origin_cites_number":22,"parent_id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"],["subsection","Regularization and Normalization"]],"subsections":[],"title":"Regularization and Normalization"},{"cite_extract_rate":0.75,"cites":[64,92,54],"content":"GANs are two-player zero-sum games, where generator (G) and discriminator ($D$) try to optimize opposing loss functions to find the global Nash equilibrium. In general, GANs can be formulated as follows:\n\t\\begin{equation}\n\t\t\\mathop{\\min}\\limits_{\\phi}\\mathop{\\max}\\limits_{\\theta} f(\\phi,\\theta)\n\t\t=\\mathop{\\min}\\limits_{\\phi}\\mathop{\\max}\\limits_{\\theta}\\mathbb{E}_{x\\sim p_{r}}[g_1(D_\\theta(x))]\n\t\t+\\mathbb{E}_{z\\sim p_z}[g_2(D_\\theta(G_\\phi(z)))],\n\t\t\\label{EQ:eqn1}\n\t\\end{equation}\n\twhere $\\phi$ and $\\theta$ are parameters of the generator $G$ and the discriminator $D$, respectively. $p_r$ and $p_z$ represent the real distribution and the latent distribution, respectively. {$g_1$and $g_2$ are different functions corresponding to different GANs. Specifically, vanilla GAN can be described as $g_1(t)=g_2(-t)=-\\log(1+e^{-t})$; $f$-GAN can be written as $g_1(t)=-e^{-t}$, $g_2(t)=1-t$; Morever, Geometric GAN and WGAN are described as $g_1(t)=g_2(-t)=-\\mathop{\\max}(0,1-t)$ and $g_1(t)$\\\\$=g_2(-t)=t$, respectively.} \n\t{Different from supervised learning, GANs training is an unsupervised learning, which leads to the urgency of regularization and normalization in the training of GANs.} In the following parts, we elaborate the training of GANs from three perspectives: low level: the perspective of \"Real \\& Fake\", middle level: the perspective of \"Fitting distribution\", and high level: the perspective of \"Training dynamics\". According to different perspectives of GANs, various regularization and normalization have been proposed in the GANs training.\n\t\\begin{comment}","id":"c2fed4a0-e8f0-4ddc-ab66-70cc7bfe76f8","level":"subsection","origin_cites_number":4,"parent_id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"],["subsection","GANs"]],"subsections":[],"title":"GANs"},{"cite_extract_rate":0.27272727272727204,"cites":[91,66,57,90,71,67],"content":"Regularization is a technique to control the complexity of learning models. Weight decay is a typical method to minimize the square of weights together with the training loss in the training of neural networks , which can be used to improve generalization. \n\tIn Bayesian learning methods, such as relevance vector machine , probabilistic classification vector machines , and others , regularization is termed as prior distribution. Specifically, L2 regularization is equivalent to introducing Gaussian prior to the parameters, and L1 regularization is equivalent to introducing Laplace prior to the parameters. The theoretical connection between regularization and prior information has been investigated in neural network ensembles research . Regularization does not only control overfitting but also provide other characteristics like semi-supervised assumptions , manifold assumptions , feature selection , low rank representation , and consistency assumptions . {Normalization is the mapping of data to a specified range, which is advantageous for the Stochastic Gradient Descent (SGD) , accelerating convergence and improving accuracy.\n\t}\n\t\\end{comment}","id":"fd688225-4a59-4600-886f-76309f6e8ba9","level":"subsection","origin_cites_number":22,"parent_id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"],["subsection","Regularization and Normalization"]],"subsections":[],"title":"Regularization and Normalization"},{"cite_extract_rate":0.5,"cites":[63,92],"content":"In low level, GANs is considered as \"counterfeiters-police\" competition, where the generator (G) can be thought of counterfeiters, trying to produce fake currency and use it undetected, while the discriminator (D) is analogous to the police, trying to detect the counterfeit currency. This competition drives both teams to upgrade their methods until the fake currency is indistinguishable from the real ones. Generally, D estimates the real probability of both real and fake samples, which is very similar to the bi-classification task, while G generates fake samples similar to real ones. Hence, the loss function in Eq (\\ref{EQ:eqn1}) is formulated as:\n\t\\begin{equation}\n\t\t\\mathop{\\min}\\limits_{\\phi}\\mathop{\\max}\\limits_{\\theta} f(\\phi,\\theta)\n\t\t=\\mathop{\\min}\\limits_{\\phi}\\mathop{\\max}\\limits_{\\theta}\\mathbb{E}_{x\\sim p_{r}}[\\log(D_\\theta(x))]\n\t\t+\\mathbb{E}_{z\\sim p_z}[\\log(1-D_\\theta(G_\\phi(z)))],\n\t\t\\label{EQ:eqnganlevel1}\n\t\\end{equation}\n\twhere $f(\\phi,\\theta)$ is a binary cross-entropy function, commonly used in binary classification problems. Eq (\\ref{EQ:eqnganlevel1}) is proposed in original GAN and can be optimized by alternate training. The training of discriminator is:\n\t\\begin{equation}\n\t\t\\mathop{\\max}\\limits_{\\theta} \\mathbb{E}_{x\\sim p_{r}}[\\log(D_\\theta(x))]\n\t\t+\\mathbb{E}_{z\\sim p_z}[\\log(1-D_\\theta(G_\\phi(z)))],\n\t\t\\label{EQ:eqndiscriminator}\n\t\\end{equation}\n\twhich is the same as a bi-classification task between real images and generated images. However, the naive binary cross-entropy function suffers from many problems, such as gradients vanishing. {Gradients vanishing is present when the difference between real and generated images as measured by discriminator is large, which leads generators cannot get optimised directions.} Accordingly, many techniques from classification like loss functions and regularization methods have been used to improve the training of discriminator. For instance, to overcome the gradients vanishing problem, Mao et al. propose the LSGANs which adopts the least\n\tsquares loss function for the discriminator. The least squares loss function moves the fake samples toward the decision boundary even though they are correctly classified. Based on this property, LSGANs is able to generate samples that are closer to real ones. The loss functions of LSGANs can be defined as follows:\n\t\\begin{equation}\n\t\t\\begin{aligned}\n\t\t\t&\\mathop{\\min}\\limits_{\\theta} \\mathbb{E}_{x \\sim p_{r}}[(D_{\\theta}({x})-b)^{2}] \n\t\t\t+ \\mathbb{E}_{{z} \\sim p_{{z}}}[(D_{\\theta}(G_{\\phi}({z}))-a)^{2}], \\\\\n\t\t\t&\\mathop{\\min}\\limits_{\\phi} \\mathbb{E}_{z \\sim p_{z}}[(D_{\\theta}(G_{\\phi}({z}))-c)^{2}],\n\t\t\\end{aligned}\n\t\\end{equation}\n\twhere $b$ and $a$ are objectives that $D$ uses for the training of real and fake samples respectively, $c$ denotes the value that $G$ wants $D$ to believe for fake sample. Gradients vanishing problem of the LSGANs only appears with $D_{\\theta}(G_{\\phi}({z}))=c$, which is hard. Furthermore, Lin et al. use SVM separating hyperplane that maximizes the margin to propose geometric GAN. Authors use the Hinge loss to train the models, which can be formulated as:\n\t\\begin{equation}\n\t\t\\begin{aligned}\n\t\t\t&\\mathop{\\min}\\limits_{\\theta} \\mathbb{E}_{x \\sim p_{r}}[(1-D_{\\theta}({x})]_{+}\n\t\t\t+ \\mathbb{E}_{{z} \\sim p_{{z}}}[1+D_{\\theta}(G_{\\phi}({z}))]_{+}, \\\\\n\t\t\t&\\mathop{\\min}\\limits_{\\phi} -\\mathbb{E}_{z \\sim p_{z}}D_{\\theta}(G_{\\phi}({z})),\n\t\t\\end{aligned}\n\t\\end{equation}\n\twhere $[x]_{+}=\\mathop{\\max}\\{0,x\\}$.\n\tThe motivation of GANs is to train the generator based on the output of the discriminator. Unlike the direct training objective of the classification task (minimizing cross-entropy loss), the objective of generator is indirect (with the help of the discriminator output). Hence, discriminator should provide a richer representation on the truth or false of samples compared to the classification task. More prior information and additional supervision tasks are urgent during the training process of the discriminator. Based on these, some regularization methods, such as {\\textit{Data Augmentation and Preprocessing}}, {\\textit{Consistency Regularization}}, and {\\textit{Self-Supervision}} are proposed to improve the stability and generalizability of discriminator.","id":"4e9cd8aa-b10a-44ef-a7b7-7d2880558ce2","level":"subsection","origin_cites_number":4,"parent_id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"],["subsection","Low Level: The Perspective of \"Real \\& Fake\""]],"subsections":[],"title":"Low Level: The Perspective of \"Real \\& Fake\""},{"cite_extract_rate":0.642857142857142,"cites":[88,54,64,93,94,95,78,8318,8317],"content":"At middle level, generator $G(z)$ is considered as a distribution mapping function that maps latent distribution $p_z(z)$ to generated distribution $P_g(x)$, and the discriminator $D(x)$ is a distribution distance that evaluates the distance between the target distribution $P_r(x)$ and the generated distribution $P_g(x)$ as illustrated in Figure \\ref{FIG}. For the optimal discriminator, the generator $G(z)$ tries to minimize the distance between $P_r(x)$ and $P_g(x)$. For instance, generator of the vanilla GAN\\footnote{Vanilla GAN, also known as standard GAN, is the first GAN model.} and $f$-GAN\\footnote{\\label{ft:5} $f$-GAN is a collective term for a type of GAN models whose discriminator minimizes $f$ divergence. $f$ divergence is the general form of KL divergence. It can be demonstrated as: $D_f(P||Q)=\\int q(x)f\\big(\\frac{p(x)}{q(x)}\\big)\\rm{d}x$, where $f$ is a mapping function from non-negative real numbers to real numbers ($\\mathbb{R}^*\\rightarrow\\mathbb{R}$) that satisfies: (1) $f(1)=0$. (2) $f$ is a convex function. To be more specific, KL divergence corresponds to $f(u)=u\\log u$ and JS divergence corresponds to $f(u)=-\\frac{u+1}{2}\\log\\frac{1+u}{2}+\\frac{u}{2}\\log u$. More details can be viewed in } are considered to minimize Jensen–Shannon (JS) divergence and $f$ divergence\\textsuperscript{\\ref{ft:5}}, respectively. When the conditions of LSGANs loss are set to $b-c=1$ and $b-a=2$, generator of the LSGAN considers the minimization of Pearson ${\\chi}^2$ divergence. Furthermore, generators of the WGAN-div\\footnote{Different from WGAN-div, WGAN minimize Wasserstein distance, not Wasserstein divergence.} and GAN-QP consider the minimization of Wasserstein divergence and Quadratic divergence, respectively.\n\t\\begin{figure}\n\t\t\\centering\n\t\t\\includegraphics[scale=0.55]{figs/GAN.pdf}\n\t\t\\caption{The framework of GANs. $P_z$ is a latent space distribution, $P_r$ and $P_g$ represent the real distribution and the generated distribution, respectively.}\n\t\t\\label{FIG}\n\t\\end{figure}\n\tGenerator is a transportation map from $z$ to $p_g(x)$. In this section, we introduce the optimal transport and the optimal transport with regular term, which leads to the form of Wasserstein GANs with gradient penalty (WGAN-GP) and Wasserstein GANs with Lipschitz penalty (WGAN-LP), respectively. Wasserstein distance is a popular and important distance in GANs and it corresponds to the optimal transport of the generator. To solve the dual problem of Wasserstein distance, Lipschitz continuity is introduced, which is the reason why gradient penalty and weight normalization techniques are proposed in the GANs training.\n\t\\begin{comment}\n\tOptimal transport was proposed in the 18th century to minimize the transportation cost while preserving the measure quantities. Given the space with probability measures $(X,\\mu)$ and $(Y,\\upsilon)$, if there is a map $T:X\\rightarrow Y$ which is measure-preserving, then for any $B\\subset Y$, having:\n\t\\begin{equation}\\label{eqn1}\n\t\t\\int_{T^{-1}(B)}\\mathrm{d}\\mu(x)=\\int_B \\mathrm{d}\\upsilon(y).\n\t\\end{equation}\n\tWriting the measure-preserving map as $T_*(\\mu)=\\upsilon$. For any $x\\in X$ and $y\\in Y$, the transportation distance is defined as $c(x,y)$, the total transportation cost is given by:\n\t\\begin{equation}\\label{eqn1}\n\t\tC(T):=\\int_X c(x,T(x)) \\mathrm{d}\\mu(x).\n\t\\end{equation}\n\tIn the 18th century, Monge et al. proposed the Optimal Mass Transportation Map that corresponds to the smallest total transportation cost: $C(T)$. The transportation cost corresponding to the optimal transportation map is called the Wasserstein distance between probability measures $\\mu$ and $\\upsilon$:\n\t\\begin{equation}\\label{eqn1}\n\t\tW_c(\\mu,\\upsilon)=\\mathop{\\min}\\limits_{T}\\left\\{\\int_X c(x,T(x)) \\mathrm{d}\\mu(x)\\ |\\ T_*(\\mu)=\\upsilon\\right\\}.\n\t\\end{equation}\n\tIn 1940s, Kantorovich proved the existence and uniqueness of the solution for Monge problem, and according to the duality of linear programming, the Kantorovich-Rubinstein (KR) duality of Wasserstein distance is given by:\n\t\\begin{equation}\\label{eqn1}\n\t\tW_c(\\mu,\\upsilon)\n\t\t=\\mathop{\\max}\\limits_{\\varphi ,\\psi}\\left\\{\\int_X \\varphi \\mathrm{d}\\mu\\ +\\int_Y \\psi \\mathrm{d}\\upsilon\\ |\\ \\varphi(x)+\\psi(y)\\leq c(x,y)\\right\\}.\n\t\\end{equation}\n\tThis dual problem is constrained, defining the c-transform: $\\psi(y)=\\varphi^c(y):=inf_x\\{c(x,y)-\\varphi(x)\\}$, and the Wasserstein distance becomes:\n\t\\begin{equation}\\label{eqn1}\n\t\tW_c(\\mu,\\upsilon)=\\mathop{\\max}\\limits_{\\varphi}\\left\\{\\int_X \\varphi \\mathrm{d}\\mu\\ +\\int_Y \\varphi^c \\mathrm{d}\\upsilon\\right\\},\n\t\\end{equation}\n\twhere $\\varphi$ is called the Kantorovich potential. It can be shown that if $c(x,y)=|x-y|$ and Kantorovich potential satisfies the 1-Lipschitz continuity, then $\\varphi^c=-\\varphi$. Kantorovich potential can be fitted by a deep neural network, which is recorded as $\\varphi_\\xi$. Wasserstein distance is:\n\t\\begin{equation}\n\t\tW_c(\\mu,\\upsilon)=\\mathop{\\max}\\limits_{||\\varphi_\\xi||_L\\leq 1}\\left\\{\\int_X \\varphi_\\xi \\mathrm{d}\\mu\\ -\\int_Y \\varphi_\\xi \\mathrm{d}\\upsilon\\right\\}.\n\t\t\\label{eqn7}\n\t\\end{equation}\n\tIf $X$ is the generated image space, $Y$ is the real sample space, $Z$ is latent space and $g_\\theta$ is the geneartor, the Wasserstein GANs (WGAN) is formulated as a min-max problem:\n\t\\begin{equation}\\label{eqn1}\n\t\t\\mathop{\\min}\\limits_{\\theta}\\mathop{\\max}\\limits_{||\\varphi_\\xi||_L\\leq 1}\\left\\{\\int_Z \\varphi_\\xi(g_\\theta(z)) \\mathrm{d}z \\ -\\int_Y \\varphi_\\xi(y) \\mathrm{d}y\\right\\}.\n\t\\end{equation}\n\tIn the optimization process, the generator and the Kantorovich potential function (discriminator) are independent of each other, optimized in a step-by-step iteration.\n\tIf $c(x,y)=\\frac{|x-y|^2}{2}$, there is a convex function $u$ that is called Brenier potential . The optimal transportation map is given by the gradient map of Brenier potential: $T(x)=\\nabla u(x)$. There exists a relationship between Kantorovich potential and Brenier potential : \n\t\\begin{equation}\\label{eqn1}\n\t\tu(x)=\\frac{|x|^2}{2}-\\varphi(x).\n\t\\end{equation}\n\tFrom the previous discussion, it is evident that the optimal transportation map (Brenier potential) corresponds to the generator, and Kantorovich potential corresponds to the discriminator. After the discriminator is optimized, the generator is directly drivable without the optimization process .\n\tThe transportation cost of Eq (3) is defined as the form of two distribution distances:\n\t\\begin{equation}\\label{eqn1}\n\t\tOT(P||Q)=\\mathop{inf}\\limits_{\\pi}\\int\\pi(x,y)c(x,y)\\mathrm{d}x\\mathrm{d}y,\n\t\\end{equation}\n\twhere $\\pi(x,y)$ is the joint distribution, satisfying $\\int_y\\pi(x,y)dy=P(x)$ and $\\int_x\\pi(x,y)dx=Q(y)$. The dual form of Eq (10) is derived as follows::\n\t\\begin{equation}\\label{eqn1}\n\t\tOT(P||Q)=\\mathop{\\max}\\limits_{\\varphi ,\\psi}\\{\\int_x \\varphi(x)P(x) \\mathrm{d}x\\ \\\\ +\\int_y\\psi(y)Q(y) \\mathrm{d}y\\ |\\ \\varphi(x)+\\psi(y)\\leq c(x,y)\\}.\n\t\\end{equation} \n\tConsidering the optimal transportation with regular terms, Peyr{\\'e} et al. added the entropic regularization for optimal transportation that transforms the dual problem into a smooth unconstrained convex problem. The regularized optimal transport is defined as:\n\t\\begin{equation}\\label{eqn1}\n\t\tOT_c(P||Q)=\\mathop{\\min}\\limits_{\\pi}\\int\\pi(x,y)c(x,y)\\mathrm{d}x\\mathrm{d}y+\\epsilon E(\\pi).\n\t\\end{equation}\n\tIf $E(\\pi)=\\int_x\\int_y\\pi(x,y)\\log(\\frac{\\pi(x,y)}{P(x)Q(y)})\\mathrm{d}x\\mathrm{d}y$, Eq (12) can be written as:\n\t\\begin{equation}\\label{eqn1}\n\t\t\\begin{split}\n\t\t\tOT_c(P||Q)=&\\mathop{\\min}\\limits_{\\pi}\\int\\pi(x,y)c(x,y)\\mathrm{d}x\\mathrm{d}y+\\epsilon \\int_x\\int_y\\pi(x,y)\\log\\left(\\frac{\\pi(x,y)}{P(x)Q(y)}\\right)\\mathrm{d}x\\mathrm{d}y\\\\\n\t\t\ts.t. \\int_y\\pi(x,y)&\\mathrm{d}y=P(x),\\int_x\\pi(x,y)\\mathrm{d}x=Q(y).\n\t\t\\end{split}\n\t\\end{equation}\n\tThe dual form of Eq (13) becomes:\n\t\\begin{equation}\n\t\t\\begin{split}\n\t\t\tOT_c(P||Q)\n\t\t\t&=\\mathop{\\max}\\limits_{\\varphi ,\\psi}\\int_x \\varphi(x)P(x) \\mathrm{d}x\\ +\\int_y\\psi(y)Q(y)\\mathrm{d}y\\\\\n\t\t\t&+\\frac{\\epsilon}{e}\\int_x\\int_y\\exp\\left(\\frac{-\\left(c(x,y)+\\varphi(x)+\\psi(y)\\right)}{\\epsilon}\\right)\\mathrm{d}x\\mathrm{d}y.\n\t\t\\end{split}\n\t\t\\label{EQ:eqn14}\n\t\\end{equation}\n\t\\end{comment}\n\tThe details of optimal transportation and optimal transportation with regular terms for WGAN and Lipschitz continuity can be found on the Section \\ref{sect:A-1} of the Supplementary Online-only Material. It is pertinent to note that \\textit{Gradient Penalty} and \\textit{Gradient} \\textit{Normalization} are two simple and effective ways to implement the Lipschitz continuity. Furthermore, demonstrates that the spectral norm and the Lipschitz constant have the same meaning. Therefore, the spectral norm can be used to represent the Lipschitz constant. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately. Hence, \\textit{Weight Normalization} and \\textit{Weight Regularization} can also be used to enable the Lipschitz continuity of the discriminator.\n\t\\begin{comment}","id":"a915370f-85fe-4345-a91b-c8d67718061c","level":"subsection","origin_cites_number":14,"parent_id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"],["subsection","Middle Level: The Perspective of \"Fitting distribution\""]],"subsections":[],"title":"Middle Level: The Perspective of \"Fitting distribution\""},{"cite_extract_rate":1,"cites":[97,96],"content":"WGAN is a popular and important generative adversarial network. From the optimal transport introduced in the last subsection, to obtain Eq (\\ref{eqn7}), the discriminator must satisfy the 1-Lipschitz continuity. This section introduces the form of the Lipschitz constant and shows that the spectral norm and the Lipschitz constant have the same meaning.\n\t1-Lipschitz continuity is represented as:\n\t\\begin{equation}\n\t||D(x_1)-D(x_2)||\\leq ||x_1-x_2||.\\footnote{Lipschitz continuity can be defined by any form of norm.}\n\t\\end{equation}\n\tGenerally, considering the K-Lipschitz for a neural network $f(x)$:\n\t\\begin{equation}\n\tf(x)=g_N\\circ\\cdots g_2\\circ g_1(x),\\footnote{$\\circ$ is the symbol for function cascade. Specifically, $h\\circ g(x)=h(g(x))$. This definition of neural network is not general, such as DenseNet and ResNet , which can not be defined like this. Therefore, we do not strictly derive the relationship between the matrix norm and Lipschitz continuity.}\n\t\\end{equation}\n\twhere $g_i(x)=\\sigma (W_i x+b_i)$. And K-Lipschitz continuity for $f(x)$ is:\n\t\\begin{equation}\n\t||f(x_1)-f(x_2)||\\leq \\mathrm{K}||x_1-x_2||,\n\t\\label{EQ:eqn17}\n\t\\end{equation}\n\twhere K is Lipschitz constant of the function $f$. Due to the consistency of Lipschitz $||h\\circ g||_{Lip}\\leq ||h||_{Lip}\\cdot||g||_{Lip}$, $g_i$ needs to satisfy the C-Lipschitz continuity ($\\mathrm{C}=\\sqrt[N]{\\mathrm{K}}$) so that $f$ satisfies the K-Lipschitz continuity:\n\t\\begin{equation}\n\t||g_i(x_1)-g_i(x_2)||\\leq \\mathrm{C}||x_1-x_2||,\n\t\\end{equation}\n\t\\begin{equation}\n\t||\\sigma(Wx_1+b)-\\sigma(Wx_2+b)||\\leq \\mathrm{C}||x_1-x_2||.\n\t\\label{eq:23}\n\t\\end{equation}\n\tWhen $x_1\\rightarrow x_2$, the Taylor expansion of Eq (\\ref{eq:23}):\n\t\\begin{equation}\n\t||\\frac{\\partial\\sigma}{\\partial x} W(x_1-x_2)||\\leq \\mathrm{C}||x_1-x_2||.\n\t\\end{equation}\n\tNormally, $\\sigma$ is a function with limited derivatives such as Sigmoid, so the $\\mathrm{C'}$-Lipschitz continuity is be written as:\n\t\\begin{equation}\n\t|| W(x_1-x_2)||\\leq \\mathrm{C'}||x_1-x_2||,\n\t\\end{equation}\n\twhere $\\mathrm{C'}$ is a limited constant, which is determined by $\\frac{\\partial\\sigma}{\\partial x}$ and $\\mathrm{C}$.\n\tSimilarly, the spectral norm of matrix is defined by:\n\t\\begin{equation}\n\t||W||_2=\\mathop{\\max}\\limits_{x\\not=0}\\frac{||Wx||}{||x||}.\n\t\\end{equation}\n\tIn this context, the spectral norm $||W||_2$ can be used to represent the Lipschitz constant $\\mathrm{C'}$. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately. Hence, \\textit{Weight Normalization} and \\textit{Weight Regularization} can also be used to enable the Lipschitz continuity of the discriminator.\n\t\\end{comment}","id":"a397f8d2-2658-471a-8c3b-cbde88271e09","level":"subsection","origin_cites_number":2,"parent_id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"],["subsection","Lipschitz Continuity and Matrix Norm"]],"subsections":[],"title":"Lipschitz Continuity and Matrix Norm"},{"cite_extract_rate":0.9,"cites":[74,98,100,54,102,99,101,64,7190],"content":"GANs training is a two-player zero-sum game with a solution to Nash Equilibrium. At high level, we analyze the convergence of GANs by understanding the optimization process. Based on these, some regularization techniques are proposed to guide the GANs model to reach the theoretical equilibrium point leading to improvement in the effectiveness of GANs.\n\tReconsidering the Eq (\\ref{EQ:eqn1}) in Section 2, the training of GANs is achieved by solving a two-player zero-sum game via Simultaneous Gradient Descent (SimGD) . The updates of the SimGD are given as:\n\t\\begin{equation} \n\t\t\\label{eq:eqn27}\n\t\t\\phi^{(k+1)}=\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)}),\\quad\n\t\t\\theta^{(k+1)}=\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)}).\n\t\\end{equation}\n\tAssuming that the objectives of GANs are convex, many research studies discuss their global convergence characteristics . However, due to the high non-convexity of deep networks, even a simple GAN does not satisfy the convexity assumption . A recent study shows that it is unrealistic to obtain approximate global convergence under the assumption of the optimal discriminator, so the community considers local convergence. It hopes that the trajectory of the dynamic system can enter a local convergence point with continuity iterations, that is, Nash equilibrium:\n\t\\begin{equation} \n\t\t\\bar{\\phi}=\\mathop{\\arg\\max}_{\\phi}-f(\\phi,\\bar{\\theta}),\\quad\n\t\t\\bar{\\theta}=\\mathop{\\arg\\max}_{\\theta}f(\\bar{\\phi},\\theta).\n\t\t\\label{EQ:eq24}\n\t\\end{equation}\n\tIf the point $(\\bar{\\phi},\\bar{\\theta})$ is called the local Nash-equilibrium, Eq (\\ref{EQ:eq24}) holds in a local neighborhood of $(\\bar{\\phi},\\bar{\\theta})$.\n\tFor this differentiable two-player zero-sum game, a vector is defined as below:\n\t\\begin{equation}\n\t\tv(\\phi,\\theta)=\n\t\t\\left(\n\t\t\\begin{aligned}\n\t\t\t-\\nabla_\\phi f(\\phi,\\theta)\\\\\n\t\t\t\\nabla_\\theta f(\\phi,\\theta)\n\t\t\\end{aligned} \n\t\t\\right).\n\t\\end{equation}\n\tThe Jacobian matrix is:\n\t\\begin{equation}\n\t\tv^{'}(\\phi,\\theta)=\n\t\t\\left(\n\t\t\\begin{aligned}\n\t\t\t-&\\nabla^2_{\\phi,\\phi} f(\\phi,\\theta)\\quad-\\nabla^2_{\\phi,\\theta} f(\\phi,\\theta)\\\\\n\t\t\t&\\nabla^2_{\\phi,\\theta} f(\\phi,\\theta)\\quad\\nabla^2_{\\theta,\\theta} f(\\phi,\\theta)\n\t\t\\end{aligned} \n\t\t\\right).\n\t\\end{equation}\n\tAccording to the propositions on the Section \\ref{sect:A-2} of the Supplementary Online-only Material, under the premise of asymptotic convergence, the local convergence of GAN is equivalent to the absolute value of all eigenvalues of the Jacobian matrix at the fixed point $(v(\\bar{\\phi},\\bar{\\theta})=0)$ being less than 1. To get this condition, \\textit{Jacobian Regularization} needs to be used.","id":"0edf29a6-bba8-43a2-a318-3f272fcb91d9","level":"subsection","origin_cites_number":10,"parent_id":"5fdbf0be-ea34-468a-9943-d62d05b593ca","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Background and Three Perspectives of GANs Training"],["subsection","High Level: The Perspective of \"Training dynamics\""]],"subsections":[],"title":"High Level: The Perspective of \"Training dynamics\""},{"cite_extract_rate":1,"cites":[6998,72],"content":"From the perspective of the \"Real \\& Fake\", generator is counterfeiter designed to deceive the discriminator, while discriminator is police designed to distinguish between real and fake samples. The motivation of GANs is to train the generator based on the loss of the discriminator. {Compared to supervised classification tasks, discriminator formally needs to perform only bi-classification tasks, which is easy to implement. Therefore, discriminator is very easy to overfit.} Furthermore, unlike the direct training objective of the classification task (Minimizing cross-entropy loss), the objective of GANs training is indirect. Hence, only one-dimensional output of the discriminator does not provide a complete representation on truth or false of samples. Some studies have shown that the present discriminators contain some significant deficiencies in the frequency domain and attribute domain , which are evidence of the lacking discrimination for discriminators. Excessive shortage of discrimination makes the generator lack incentives from the discriminator to learn useful information of the data. {In addition to discriminator overfitting and lacking discrimination of discriminators, discriminator forgetting is another challenge for GANs.} To alleviate these situations, many regularization methods and additional supervision tasks have been proposed in the literature, which can be divided into three categories: \\textit{Data Augmentation and Preprocessing}, \\textit{Consistency Regularization}, and \\textit{Self-supervision}. {All of them are based on data augmentation and are orthogonal to each other. As shown in Table \\ref{table:SOTA }, The state-of-the-art GANs always adopt two or even all of the above regularization.}","id":"8469161e-625b-422b-90e0-a6cca9638a04","level":"section","origin_cites_number":2,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""]],"subsections":["6dc08e83-31a2-4629-92a0-ea9cbb6c27a0","a5e04be5-4733-477b-9081-c52ce2a3be68","4e2e9490-faa8-4a2f-9dbd-66ae6c225118","2267f79d-3862-4c72-bdff-9670534c3017"],"title":"Regularization and Normalization of \"Real \\& Fake\""},{"cite_extract_rate":0.7894736842105261,"cites":[91,7191,111,104,112,107,106,110,109,108,103,113,61,8319,105],"content":"\\begin{figure}\n\t\t\\centering\n\t\t\\includegraphics[scale=.35]{figs/data_augmentation_framework.pdf}\n\t\t\\caption{Framework of data augmentation and preprocessing for updating D (left) and G (right). (Coming from )}\n\t\t\\label{fig:data_augmentation_framework}\n\t\\end{figure}\n\tData Augmentation plays a significant role in deep learning algorithms. It increases the diversity of the training data naturally, thus reduces the overfitting in many computer vision and graphics applications . Date augmentation adopts different data transformation techniques ({\\color{red}$T$}) to increase the number of training samples. One type of data transformation is spatial transformation of data, such as $zoomout$, $zoomin$, $translation$, $translationx$, $translationy$, $cutout$ , $cutmix$ ; The other is\n\tvisual transformation, such as $brightness$, $redness$, $greenness$, $blueness$, $mixup$ . {Furthermore, recent study is also attempting to use frequency transformation to data augmentation. }\n\tSimilarly, the performance of GANs heavily deteriorates given a limited amount of training data . {For instance, shows that Frechet Inception Distance (FID) starts to rise at some point during training and outputs of discriminator keep drifting apart during training, when training data is limited. More analysis can be found in the survey on data-efficient GANs training.} However, recent studies observe that augmenting only real images (Only applying $T$ to (i) in Figure \\ref{fig:data_augmentation_framework}), only generated images (Only applying $T$ to (ii) in Figure \\ref{fig:data_augmentation_framework}), and only discriminator (Both applying $T$ to (i) and (ii) in Figure \\ref{fig:data_augmentation_framework}) do not help with GANs training. Naturally, one problem needs to be considered: whether the overfitting exists in GANs' training? Some studies demonstrate that, even with big dataset and state of the art models, the training of GANs suffers from severe overfitting. Furthermore, in case of small training data, overfitting occurs at an early stage in the training. Recently, some studies on data augmentation for GANs training have been proposed. It is argued that the classical data augmentation approach could mislead the generator to learn the distribution of the augmented data, which could be different from that of the original data. To deal with this problem, these studies augment both real and fake samples and let gradients propagate through the augmented samples to G (Applying $T$ to (i), (ii), and (iii) in Figure \\ref{fig:data_augmentation_framework}). By adding the data augmentation to all processes of GANs training, the performance of GANs has been significantly improved. {However, this \"\\textit{Augment All}\" strategy may lead to the “leaking” of augmentations to the generated samples, which is highly undesirable. The experiments in demonstrate that as long as the probability of executing the augmentation remains below 0.8, leaks are unlikely to happen in practice.}\n\tData augmentation in GANs has remarkable achievement. However, which augmentation is most beneficial for GANs training is still an open problem. Figure 2 in shows the FID comparisons of BigGAN on CIFAR-10 dataset. For data augmentation (represented by ‘vanilla\\_rf’), the operations in spatial augmentation such as $translation$, $zoomout$, and $zoomin$, are much more effective than the operations in visual augmentation, such as $brightness$, $colorness$ ($redness$ and $greenness$), and $mixup$. The results indicate that augmentation leads to spatial changes which improves GANs performance compared with cases where visual changes are induced. It is easy to understand that generated images are significantly lacking in detail information compared to the real images, and spatial augmentation improves the ability of the generator to fit detailed textures through spatial changes. {$InstanceNoise$, resulting in images out of the natural data manifold, cannot help with improving GANs performance. Apart form applying only a limited range of augmentations, some studies explore some strong data augmentations in GANs training. For instance, Jeong et al. adopt contrastive learning to extract more useful information under stronger data augmentation beyond the existing yet limited practices. Combining adaptive strategies and 18 transformations (Both spatial and visual transformations) , even and $InstanceNoise$ only can bring performance improvement over strong GANs baselines. {Furthermore, is the first method to tackle the generative learning trilemma with denoising diffusion GANs.} Apart from these, Jiang et al. also devise an adaptive strategy to control the strength of selecting generated images to augment real data, which can further boosts the performance of GANs. Data augmentation is popular and significant in GANs training, whose achievements are attributed to improving discrimination, avoiding overfitting, and increasing the overlap between real and fake distributions.}\n\t{Different from data augmentation that increases the amount of the training data, data preprocessing only adopt prior knowledge and do some uniform data transformation before the network training. Data preprocessing is orthogonal to data augmentation, which can further enhance the performance combining the data augmentation.} Li et al. indicate that high-frequency components between real and fake images are different, which is not conducive to the GANs training. They propose two preprocessing eliminating high-frequency differences in GANs training: High-Frequency Confusion (HFC) and High-Frequency Filter (HFF). These methods are applied in places (i), (ii), and (iii) in Figure \\ref{fig:data_augmentation_framework} and improve the performance of GANs with a fraction of the cost. \n\tIn summary, both data augmentation and data preprocessing improve the performance of GANs with little cost. Data augmentation uses different transformations to improve discrimination and avoid disciminator overfitting. Furthermore, spatial augmentations achieve better performance than visual augmentations. More specifically, Zhao et al. demonstrate that hybrid augmentation with $Color + Translation + Cutout$ is especially effective and is widely used in other studies . Adaptive data augmentation (ADA) is the most popular method in GANs. Besides the data augmentation, data preprocessing is also a remarkable method.\n\t\\begin{figure}\n\t\t\\centering\n\t\t\\includegraphics[scale=.3]{figs/consistency.pdf}\n\t\t\\caption{Overview of consistency regularization, where image consistency regularization updates the G (left) and network consistency regularization updates the D (right). $H$ is the feature mapping function and {\\color{red}$T$} is different data transformation techniques.}\n\t\t\\label{fig:consistency}\n\t\\end{figure}","id":"6dc08e83-31a2-4629-92a0-ea9cbb6c27a0","level":"subsection","origin_cites_number":19,"parent_id":"8469161e-625b-422b-90e0-a6cca9638a04","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Data Augmentation and Preprocessing"]],"subsections":[],"title":"Data Augmentation and Preprocessing"},{"cite_extract_rate":1,"cites":[115,114,7000,7192],"content":"In context of semi-supervised and unsupervised learning, consistency regularization has been widely used in . It is motivated by the fact that models should produce consistent predictions given input and their semantics-preserving augmentations, such as image rotating, and adversarial attacks. It is pertinent to note that the supervision of GANs training is weak. To increase the discrimination of discriminator, some consistency regularization techniques have also been used. Due to different goals, we divide these into two parts: \\textit{Image Consistency} and \\textit{Network Consistency} as demonstrated in Figure \\ref{fig:consistency}.","id":"a5e04be5-4733-477b-9081-c52ce2a3be68","level":"subsection","origin_cites_number":4,"parent_id":"8469161e-625b-422b-90e0-a6cca9638a04","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Consistency Regularization"]],"subsections":["a0c72661-2f99-4ec9-8181-70dfdeab36dc","86d17fb0-5862-4222-a247-42d917acf482"],"title":"Consistency Regularization"},{"cite_extract_rate":0.833333333333333,"cites":[116,6998,117,121,8320,119,118,120,91,69],"content":"The purpose of GANs is to generate fake images similar to real ones. In GANs, the discriminator is generally used to distinguish real images and generated images. However, outputs of the discriminator with only one dimension hardly portray the authenticity of the image completely. To improve the representation of the discriminator, some studies extend the outputs of the discriminator, for example, relativistic discriminator , distribution discriminator , and cascading rejection . Apart from this, some studies reduce the training difficulty of discriminators by introducing prior information. Regularizing the distance between the generated and real images with different measurements, namely image consistency, is the focus of this paper. The overview of it is demonstrated in left part of Figure \\ref{fig:consistency}, where consistency regularization is used to update generator (G) and can be formulated as:\n\t\\begin{equation}\n\t\t\\mathcal{L}_C=\\mathbb{E}_{x\\sim p_r, z\\sim p_z} C(H(x),H(G(z))),\n\t\t\\label{EQ:image_consistency}\n\t\\end{equation}\n\twhere $H$ is the feature mapping function and $C$ is the consistency measurement function. Different image consistency regularization have different $H$ and $C$. For instance, Salimans et al. recommend that the generator is trained using a feature matching procedure. The objective is:\n\t\\begin{equation}\n\t\t\\mathcal{L}_C=||\\mathbb{E}_{x\\sim p_r} f(x)-\\mathbb{E}_{z\\sim p_z}f(G(z))||^2_2,\n\t\\end{equation}\n\twhere $f(x)$ denotes the intermediate layer of the discriminator. Similarly, the intermediate layer of another pre-trained classification model is an alternate option. The empirical results indicate that feature matching is indeed effective in situations where normal GAN becomes unstable. Unlike the above study which only uses mean feature matching to training generators, Mroueh et al. propose McGAN, which trains both the generator and discriminator using the mean and covariance feature matching. The objective is:\n\t\\begin{equation}\n\t\t\t\\mathcal{L}_C=\\mathcal{L}_\\mu+\\mathcal{L}_\\sigma\n\t\t\t=||\\mu(p_r)-\\mu(G(p_z))||_q+||\\sum\\left(p_r\\right)-\\sum(p(G(z)))||_k,\n\t\\end{equation}\n\twhere $\\mu(p_r)=\\mathbb{E}_{x\\sim p_r} f(x)$ and $\\sum(p_r)=\\mathbb{E}_{x\\sim p_r} f(x)\\cdot f(x)^\\mathrm{T}$ represent the mean and the covariance of the feature layer $f(x)$, respectively. Apart from statistical differences, some studies focus on the difference in frequency domain between the generated and real image. For instance, Durall et al. find that the deep generative models based on up-convolution fail to reproduce spectral distributions leading to considerable differences in the spectral distributions between real images and generated images. Thus, the spectral regularization has been proposed as follows:\n\t\\begin{equation}\n\t\t\\mathcal{L}_C=\\frac{1}{M/2-1}\\sum_{i=0}^{M/2-1}AI^{real}_i\\cdot\\log(AI^{fake}_i)\n\t\t+(1-AI^{real}_i)\\cdot\\log(1-AI^{fake}_i),\n\t\\end{equation}\n\twhere $M$ is the image size and $AI$ is the spectral representation from the Fourier transform of the images. Corresponding to Eq (\\ref{EQ:image_consistency}), $H$ and $C$ are implemented with $AI$ and cross-entropy, respectively.\n\tContrary to this, the research study uses hard example mining to improve the discriminatory of the discriminator based on the difference between real and generated samples under different metrics. Although this paradigm is different from the paradigm of image consistency regularization, both cases are motivated by obtaining generated samples similar to real images under different distance measures, so we integrate them. Chen et al. consider both downsampling strategies: downsampling with anti-aliasing and downsampling without anti-aliasing, leads to high frequencies missing in the discriminator. High frequencies missing leads to high frequency deviation between real and generated images. To mitigate this issue, authors propose SSD-GAN, which introduces an additional spectral classifier to detect frequency spectrum discrepancy between real and generated images and integrate it into the discriminator of GANs. The overall realness of sample x is represented as:\n\t\\begin{equation}\n\t\tD^{s s}(x)=\\lambda D(x)+(1-\\lambda) C(\\phi(x)),\n\t\\end{equation}\n\twhere the enhanced discriminator $D^{s s}$ consists of two modules, a vanilla discriminator $D$ that measures the spatial realness, and a spectral classifier $C$. $\\lambda$ is a hyperparameter that controls the relative importance of the spatial realness and the spectral realness. The adversarial loss of the framework can be written as:\n\t\\begin{equation}\n\t\t\\mathcal{L}_{D} =\\mathbb{E}_{x \\sim p_{\\text {data }}(x)}\\left[\\log D^{s s}(x)\\right] \n\t\t+\\mathbb{E}_{x \\sim p_{g}(x)}\\left[\\log \\left(1-D^{s s}(x)\\right)\\right].\n\t\\end{equation}\n\t{karnewar et al. introduce adversarial loss into the intermediate layer of the generator, which provides multiple and richer metrics for the training of generator.}\n\t\\begin{table}\n\t\t\\caption{The summary of the consistency regularization.}\n\t\t\\label{table:consistency}\n\t\t\\tiny\n\t\t\\centering\n\t\t\\begin{tabular}{c| c }\t\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\tMethod& Consistency regularization term $L_C$\t \\\\\n\t\t\t\\hline\n\t\t\tMean regularization &$||\\mathbb{E}_{x\\sim p_r} f(x)-\\mathbb{E}_{z\\sim p_z}f(G(z))||_q$\\\\\n\t\t\t\\hline\n\t\t\tMean and Convariance regularization &$||\\mathbb{E}_{x\\sim p_r} f(x)-\\mathbb{E}_{z\\sim p_z}f(G(z))||_q+||\\mathbb{E}_{x\\sim p_r} f(x)\\cdot f(x)^\\mathrm{T}-\\mathbb{E}_{z\\sim p_z}f(G(z))\\cdot f(G(z))^\\mathrm{T}||_k$\\\\\n\t\t\t\\hline\n\t\t\tSpectral regularization &$\\mathcal{L}_C=\\frac{1}{M/2-1}\\sum_{i=0}^{M/2-1}AI^{real}_i\\cdot\\log(AI^{fake}_i)+(1-AI^{real}_i)\\cdot\\log(1-AI^{fake}_i)$\\\\\n\t\t\t\\hline\n\t\t\tCR-GAN &$\\mathbb{E}_{x\\sim p_r}|| D(x)-D(T(x))||_2$\\\\\n\t\t\t\\hline\n\t\t\tbCR-GAN &$\\lambda_{real}\\mathbb{E}_{x\\sim p_r}|| D(x)-D(T(x))||_2+\\lambda_{fake}\\mathbb{E}_{z\\sim p_z}||D(G(z))-D(T(G(z)))||_2$\\\\\n\t\t\t\\hline\n\t\t\tzCR-GAN&$\\lambda_{dis}\\mathbb{E}_{z\\sim p_z}||D(G(z))-D(G(T(z)))||_2-\\lambda_{gen}\\mathbb{E}_{z\\sim p_z}||G(z)-G(T(z))||_2$\\\\\n\t\t\t\\bottomrule\n\t\t\\end{tabular}\n\t\\end{table}\n\tThe summary of the image consistency regularization is given in Table \\ref{table:consistency}. In summary, image consistency considers that the real images and the generated images are similar not only in the output of discriminator, but also in statistical information and frequency domain. The analysis of biases between real and generated images using different metrics will be an interesting future research direction.","id":"a0c72661-2f99-4ec9-8181-70dfdeab36dc","level":"subsubsection","origin_cites_number":12,"parent_id":"a5e04be5-4733-477b-9081-c52ce2a3be68","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Consistency Regularization"],["subsubsection","Image Consistency"]],"subsections":[],"title":"Image Consistency"},{"cite_extract_rate":1,"cites":[91,8320,69,111],"content":"Network consistency regularization can be regarded as Lipschitz continuity on semantics-preserving transformation. Specifically, we hope discriminator is insensitive to semantics-preserving transformation, which drives the discriminator to pay more attention to the authenticity of the images. For example, in the image domain, the reality of images should not change if we flip the image horizontally or translate the image by a few pixels. To resolve this, Zhang et al. propose the Consistency Regularization GAN (CR-GAN) that uses the consistency regularization on the discriminator during GANs training:\n\t\\begin{equation}\n\t\t\\mathcal{L}_C=\\mathbb{E}_{x\\sim p_r}|| D(x)-D(T(x))||_2,\n\t\\end{equation}\n\twhere $T(x)$ represents a transformation (shift, flip, cutout, etc.) of images. One key problem with the CR-GAN is that the discriminator might occur the 'mistakenly believe'. 'mistakenly believe' considers that the transformations are actual features of the target dataset, due to only applying these transformations on real images. This phenomenon is not easy to notice for certain types of transformations (e.g. image shifting and flipping). However, some types of transformations, such as cutout transformations, contain visual artifacts not belonging to real images, which effects greatly limits the choice of advanced transformations we could use. To address this issue, Zhao et al. propose Balanced Consistency Regularization (bCR-GAN) that uses regulation with respect to both real and fake images and balances the training of discriminator between real images and fake images by $\\lambda_{real}$ and $\\lambda_{fake}$:\n\t\\begin{equation}\n\t\t\\mathcal{L}_C=\\lambda_{real}\\mathbb{E}_{x\\sim p_r}|| D(x)-D(T(x))||_2\n\t\t+\\lambda_{fake}\\mathbb{E}_{z\\sim p_z}||D(G(z))-D(T(G(z)))||_2.\n\t\\end{equation}\n\tThe overview of bCR is demonstrated in right part of Figure \\ref{fig:consistency}.\n\tContrary to the methods which focus on consistency regularization with respect to transformations in image space, Zhao et al. also propose Latent Consistency Regularization (zCR) that considers the consistency regularization on transformations in latent space. Authors expect that output of the discriminator ought not to change much with respect to the small enough perturbation $\\Delta z$ and modify the discriminator loss by enforcing:\n\t\\begin{equation}\n\t\t\\mathcal{L}^D_C=\\lambda_{dis}\\mathbb{E}_{z\\sim p_z}||D(G(z))-D(G(T(z)))||_2,\n\t\\end{equation}\n\twhere $T(z)$ represents the added small perturbation noise. However, if only this loss is added into the GAN loss, mode collapse can easily appear in the training of generators. To avoid this, an inverse gradient penalty (we will describe it in section 6.2) is added to modify the loss function for generator. Hence, we modify the generator loss by enforcing:\n\t\\begin{equation}\n\t\t\\mathcal{L}^D_C=-\\lambda_{gen}\\mathbb{E}_{z\\sim p_z}||G(z)-G(T(z))||_2.\n\t\\end{equation}\n\tNaturally, putting both bCR and zCR together, Improved Consistency Regularization (ICR) is also proposed by Zhao et al. . In addition, there are some applications where cyclic consistency regularization is used for unpaired image-to-image translation . The summary of network consistency regularization is given in Table \\ref{table:consistency}. \n\tIn summary, network consistency considers the networks, especially the discriminator, to be insensitive to semantics-preserving transformation ($T$). The results in demonstrate that random shift and flip is the best way to perform image transformation on the CIFAR-10 dataset. Furthermore, the FID results with CR, bCR, zCR, and ICR (where transformation is flipping horizontally and shifting by multiple pixels) as presented in are shown in Table \\ref{table:consistency_results}. The results demonstrate that network consistency regularization can significantly improve the performance of GANs. However, which transformation is best for consistency regularization, is a question. Zhao et al compare the effect of different data transformation techniques (mentioned in Section 5.1) on bCR. Figure 2 in shows the FID results of BigGAN adding bCR (represented by 'bcr') on CIFAR-10 dataset. From the results, the best BigGAN FID 8.65 is with transformation technology $translation$ of strength $\\lambda=0.4$, outperforming the corresponding FID 10.54 reported in Zhao et al. . Moreover, spatial transforms, which retain the major content while introducing spatial variances, can substantially improve GANs performance together with bCR. While visual transforms, which retain the spatial variances, can not further improve the performance of GANs compared with data augmentation only. Furthermore, bCR with stronger transformation (larger value of $\\lambda_{aug}$) does not improve the performance of GANs, the optimal value of $\\lambda_{aug}$ is uncertain for different data transformations. \n\t\\begin{table}\n\t\t\\caption{FID scores for class conditional image generation of the network consistency regularization (Data come from ).}\n\t\t\\label{table:consistency_results}\n\t\t\\centering\n\t\t\\begin{tabular}{ccc}\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\tModels & CIFAR-10 & ImageNet \\\\\n\t\t\t\\hline SNGAN & 17.50 & 27.62 \\\\\n\t\t\tBigGAN & 14.73 & 8.73 \\\\\n\t\t\tCR-BigGAN & 11.48 & 6.66 \\\\\n\t\t\tbCR-BigGAN & 10.54 & 6.24 \\\\\n\t\t\tzCR-BigGAN & 10.19 & 5.87 \\\\\n\t\t\tICR-BigGAN & $\\mathbf{9 . 2 1}$ & $\\mathbf{5 . 3 8}$ \\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{table}","id":"86d17fb0-5862-4222-a247-42d917acf482","level":"subsubsection","origin_cites_number":4,"parent_id":"a5e04be5-4733-477b-9081-c52ce2a3be68","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Consistency Regularization"],["subsubsection","Network Consistency"]],"subsections":[],"title":"Network Consistency"},{"cite_extract_rate":1,"cites":[123,122,7000],"content":"Self-supervised learning aims to learn representations from the data itself without explicit manual supervision. Recently, some self-supervised studies provide competitive results on ImageNet classification and the representations learned from which transfer well to downstream tasks. Self-supervised learning outperforms its supervised pre-training counterpart in many tasks, such as detection and segmentation, sometimes surpassing it by large margins. This suggests that self-supervised learning obtains more representational features and significantly improve the representation of networks. Based on this, self-supervised learning is introduced into the training of GANs, and we divide them into two categories according to different self-supervision tasks: \\textit{Predictive Self-supervised Learning} and \\textit{Contrastive Self-supervised Learning}.","id":"4e2e9490-faa8-4a2f-9dbd-66ae6c225118","level":"subsection","origin_cites_number":3,"parent_id":"8469161e-625b-422b-90e0-a6cca9638a04","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Self-Supervision"]],"subsections":["b5c16876-cba3-44e8-b6be-d7b6cabbf7ee","aff71d0b-4dab-423d-9325-d61d96004987"],"title":"Self-Supervision"},{"cite_extract_rate":0.833333333333333,"cites":[126,132,8321,128,124,129,130,131,7193,103,127,55,113,111,125],"content":"Predictive self-supervised learning is a popular method to improve the representation of neural networks by introducing additional supervised tasks, such as context prediction and rotations prediction . Predictive self-supervised learning is introduced into GANs by Chen et al. to avoid discriminator forgetting. Discriminator forgetting means that the discriminator does not remember all tasks at the same time during the training process. For example, learning varying levels of detail, structure, and texture, which causes the discriminator to fail to get a comprehensive representation of the current images. \"If the outcome is your focus, then it’s easy to look for shortcuts. And ultimately shortcuts keep you from seeing the truth, it drains your spark for life. What matters most is your will to seek the truth despite the outcome.\"\\footnote{Come from \"JoJo's Bizarre Adventure:Golden Wind\" -Araki Hirohiko.}. The same is true for GANs, which are only driven by the loss of discriminator, which is easy to distinguish between real images and generated images through shortcuts, instead of the texture and structural features we need. Predictive self-supervised learning solves this problem by introducing new generalization tasks, which also helps to prevent overfitting. The overview of the predictive self-supervised learning of GANs are demonstrated on Figure \\ref{fig:predict_self_supervised}. Depending on the data transformation function $T$, we can design different self-supervised tasks. \n\t\\begin{figure}\n\t\t\\centering\n\t\t\\includegraphics[scale=.4]{figs/predict_self_supervised.pdf}\n\t\t\\caption{Overview of the predict self supervised learning of GANs, where $C$ performs the predict classification task and shares the weights with the discriminator except for the last layer, {\\color{red}$T$} is different data transformation techniques. Furthermore, the self-supervised task of generated images is used to update the generator (left) and the self-supervised task of real images is used to update the classification (right).}\n\t\t\\label{fig:predict_self_supervised}\n\t\\end{figure}\n\tChen et al. introduced the predictive self-supervision in GANs training. Authors adopt the rotation prediction as the expanding task to prevent the discriminator from forgetting. Besides, plenty of other prediction tasks have also been proposed to improve the discrimination. Huang et al. exploit the feature exchange to make the discriminator learn the proper feature structure of natural images. Baykal et al. introduce a reshuffling task to randomly arrange the structural blocks of the images, thus helping the discriminator increase its expressive capacity for spatial structure and realistic appearance. Contrary to the methods for designing tasks at the image or feature level, Patel et al. propose a self-supervised task with latent transformation detection, which identifies whether the latent transformation applied in the given pair is the same as that of the other pair. All above methods have designed different self-supervised tasks, and their loss functions can be formulated as:\n\t\\begin{equation}\n\t\t\\begin{split}\n\t\t\t\\mathcal{L}_{D,C}=-\\lambda_{r}\\mathbb{E}_{x\\sim p^T_r}\\mathbb{E}_{T_k\\sim {T}}\\log\\big( C_k(x)\\big) \\quad \\text { for } k=1, \\ldots, K,\\\\\n\t\t\t\\mathcal{L}_G=-\\lambda_{g}\\mathbb{E}_{x\\sim p^T_g}\\mathbb{E}_{T_k\\sim {T}}\\log\\big( C_k(x)\\big) \\quad \\text { for } k=1, \\ldots, K,\n\t\t\\end{split}\n\t\\end{equation}\n\twhere ${T}$ represents the different types of image transfer, such as rotation and reshuffling. Furthermore, $T_k$ represents different forms of the transfer ${T}$, such as $0^\\circ,90^\\circ,180^\\circ, 270^\\circ$ for rotation task, $K$ is the number of transformed forms, $C_k$ is the k-th output of the classifier $C$ that shares parameters with discriminator except for two different heads, $P^T_r$ and $P^T_g$ are the transformed distributions of real and generated images, respectively. For rotation conversion task , $K=4$, and the classifier $C$ predicts the rotation angle; For feature exchange task , $K=2$, and the classifier $C$ predicts whether the swap has occurred; For block reshuffling task , the image is divided into $9$ blocks and the number of the permutations is $9!$, which is unnecessarily huge. Thirty different permutations are selected in terms of the Hamming distances between the permutations in . As a result, $K$ is set to 30, and classifier $C$ predicts the Hamming distances of different permutations; For the latent transformation task, $K=2$, and the classifier $C$ predicts whether the transformations parameterized by the same $\\epsilon$ or different. {Besides, some study introduces the autoencoder task, making discriminator reconstruct the input. }\n\t\\begin{table*}\n\t\t\\renewcommand\\arraystretch{1}\n\t\t\\caption{The summary of the Self-supervision}\n\t\t\\label{table:self-supervised}\n\t\t\\setlength{\\tabcolsep}{0.5mm}\n\t\t\\centering\n\t\t\\begin{tabular}{c| c |c }\t\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\t\\tiny Method&\\tiny Description&\\tiny Types\\\\\n\t\t\t\\hline\n\t\t\t\\tiny Rotation Prediction &\\tiny Predicting the angle of rotation ($0^\\circ,90^\\circ,180^\\circ, 270^\\circ$)&\\tiny PSS\\\\\n\t\t\t\\hline\n\t\t\t\\tiny Feature Exchange Detection &\\tiny Predicting if some exchanges have occurred at the feature level (yes or not)&\\tiny PSS\\\\\n\t\t\t\\hline\n\t\t\t\\tiny Block Reshuffling Prediction &\\tiny Predicting the Hamming distances of different reshuffling in image level (Total 30 categories)&\\tiny PSS\\\\\n\t\t\t\\hline \n\t\t\t\\tiny Latent Transformation Detection &\\tiny Predicting if some exchanges have occurred at the latent space level (yes or not)&\\tiny PSS\\\\\n\t\t\t\\hline\n\t\t\t\\tiny Autoencoder &\\tiny Reconstruct the input of the discriminator&\\tiny PSS\\\\\n\t\t\t\\hline\n\t\t\t\\tiny InfoMax-GAN &\\tiny \\makecell[c]{Positive pairs: Global and local features of an image (both real and fake images)\\\\ Negative pairs: Global and local features of different images (both real fake images)}&\\tiny CSS\\\\\n\t\t\t\\hline \n\t\t\t\\tiny Cntr-GAN &\\tiny \\makecell[c]{Positive pairs: Two different data transformations of the same image (both real and fake images).\\\\ Negative pairs: Otherwise}&\\tiny CSS\\\\\n\t\t\t\\hline\n\t\t\t\\tiny\tContraD &\\tiny \\makecell[c]{Positive pairs: Two different data transformations of the same image (real images only) + Two fake images\\\\ Negative pairs: Otherwise}&\\tiny CSS\\\\\n\t\t\t\\hline\n\t\t\t\\tiny\tInsGen &\\tiny \\makecell[c]{Positive pairs: Two different data transformations (additional latent transformations for fake image)\\\\ of the same image (both real and fake images). Negative pairs: Otherwise.}&\\tiny CSS\\\\\n\t\t\t\\hline\n\t\t\t\\tiny\tFakeCLR &\\tiny \\makecell[c]{Positive pairs: Two different data transformations and additional latent transformations for fake image.\\\\ Negative pairs: Otherwise.}&\\tiny CSS\\\\\n\t\t\t\\bottomrule\n\t\t\\end{tabular}\n\t\\end{table*}\n\tThe above methods design different kinds of self-supervised prediction tasks and participate in the training of the discriminator or generator, independently, having “loophole” that, during generator learning, $G$ could exploit to minimize $\\mathcal{L}_G$ without truly learning the data distribution. To address this issue, Ngoc-TrungTran et al. introduce true or false judgment along with self-supervised prediction. The number of classification is $K+1$, while the loss function can be expressed as:\n\t\\begin{equation}\n\t\t\\begin{aligned}\n\t\t\t\\mathcal{L}_{D,C}=&-\\lambda_{r}\\Bigg(\\mathbb{E}_{x\\sim p^T_r}\\mathbb{E}_{T_k\\sim {T}}\\log\\big( C_k(x)\\big)\n\t\t\t+\\mathbb{E}_{x\\sim p^T_g}\\mathbb{E}_{T_k\\sim {T}}\\log\\big( C_{K+1}(x)\\big)\\Bigg)\\quad \\text { for } k=1, \\ldots, K,\\\\\n\t\t\t\\mathcal{L}_G=&-\\lambda_{g}\\Bigg(\\mathbb{E}_{x\\sim p^T_g}\\mathbb{E}_{T_k\\sim {T}}\\log\\big( C_k(x)\\big)\n\t\t\t-\\mathbb{E}_{x\\sim p^T_g}\\mathbb{E}_{T_k\\sim {T}}\\log\\big( C_{K+1}(x)\\big)\\Bigg)\\quad \\text { for } k=1, \\ldots, K,\n\t\t\\end{aligned}\n\t\\end{equation}\n\twhere $C_k$ is a classifier that predicts the rotation angles and $C_{K+1}$ is a classifier that predicts the truth of the images. The new self-supervised rotation-based GANs use the multi-class minimax game to avoid the mode collapse, which is better than the original predictive self-supervised paradigm.\n\tIn summary, predictive self-supervised learning improves the discrimination by designing different self-supervised prediction tasks, among them, rotation prediction is widely used () for its simplicity and practicality. The summary of different methods is illustrated in Table \\ref{table:self-supervised}. {All of these methods are juxtaposed with each other. However, few studies have used multiple self-supervised tasks simultaneously, and the use of multiple self-supervised tasks to improve GANs training is still an open problem.}","id":"b5c16876-cba3-44e8-b6be-d7b6cabbf7ee","level":"subsubsection","origin_cites_number":18,"parent_id":"4e2e9490-faa8-4a2f-9dbd-66ae6c225118","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Self-Supervision"],["subsubsection","Predictive Self-Supervised Learning (PSS)"]],"subsections":[],"title":"Predictive Self-Supervised Learning (PSS)"},{"cite_extract_rate":0.9166666666666661,"cites":[134,133,123,130,103,125,122,7000,111,61],"content":"Contrastive self-supervised Learning , as the name implies, learn representations by contrasting positive and negative examples. These techniques have resulted in empirical success in computer vision tasks with unsupervised contrastive pre-training. A handful number of studies demonstrate that self-supervised learning outperforms its supervised pre-training counterpart in many tasks, which indicates contrastive self-supervised learning leads to more expressive features. Considering two views $(v^{(1)}$ and $v^{(2)})$, contrastive self-supervised learning aims to identify whether two views are dependent or not. More specifically, it means to maximize the mutual information of positive pairs. To this end, Oord et al. propose to minimize InfoNCE loss, which turns out to maximize a lower bound of mutual information. The InfoNCE loss is defined by:\n\t\\begin{equation}\n\t\tL_{\\mathrm{NCE}}\\left({v}_{i}^{(1)} ; {v}^{(2)}, s\\right):=-\\log \\frac{\\exp \\left(s\\left({v}_{i}^{(1)}, {v}_{i}^{(2)}\\right)\\right)}{\\sum_{j=1}^{K} \\exp \\left(s\\left({v}_{i}^{(1)}, {v}_{j}^{(2)}\\right)\\right)},\n\t\\end{equation}\n\twhere $s(,)$ is the score function that measure the similarity, positive pairs ($v_i^{(1)}$ and $v_i^{(2)}$) are different views of the same sample, and negative pairs ($v_i^{(1)}$ and $v_j^{(2)}$, $(i\\neq j)$) are different views of different samples. InfoNCE loss is the cornerstone of contrastive self-supervised learning as depicted in Figure \\ref{fig:contrastive_self_supervised}. \n\t\\begin{figure}\n\t\t\\centering\n\t\t\\includegraphics[scale=.3]{figs/constrative_self_supervised.pdf}\n\t\t\\caption{Overview of the contrastive self supervised learning, where x is real of fake images, $v^{(1)}$ and $v^{(2)}$ are different views of the image x, s is the score function (Usually discriminator in GANs) that measures the similarity, the square on the middle part is the label of InfoNCE (blue, white, and black squares are labels with 1, 0, and undefined respectively.), and the square on the right part is the label of SimCLR. Obviously, SimCLR defines more negative pairs to improve the sample utilization.}\n\t\t\\label{fig:contrastive_self_supervised}\n\t\\end{figure}\n\tMany advanced self-supervised methods are implemented by modifying the views of images $v^{(1)}$, $v^{(2)}$ and score function $s(,)$. Specifically, Deep InfoMAX maximizes the mutual information between local and global features, that is, image $x$ passes through the encoder $E_{\\psi}=f_{\\psi}\\circ C_{\\psi}$, producing local feature map $C_{\\psi}(x)$ and global feature vector $E_{\\psi}(x)$. To maximize the lower bound of the InfoMax: $\\mathcal{I}\\big(C_{\\psi}(x),E_{\\psi}(x)\\big)$, the theoretical InfoMAX loss has been defined as:\n\t\\begin{equation}\n\t\t\\begin{aligned}\n\t\t\tL_{InfoMAX}(X)=&-\\mathbb{E}_{x\\in X}\\mathbb{E}_{i\\in\\mathcal{A}}\\big[\n\t\t\t\\log \\frac{\\exp \\left(g_{\\theta, \\omega}\\left(C_{\\psi}^{(i)}(x), E_{\\psi}(x)\\right)\\right)}{\\sum_{\\left(x^{\\prime}, i\\right) \\in X \\times \\mathcal{A}} \\exp \\left(g_{\\theta, \\omega}\\left(C_{\\psi}^{(i)}\\left(x^{\\prime}\\right), E_{\\psi}(x)\\right)\\right)}\\big],\\\\\n\t\t\tg_{\\theta, \\omega}&\\left(C_{\\psi}^{(i)}(x), E_{\\psi}(x)\\right)=\\phi_{\\theta}\\left(C_{\\psi}^{(i)}(x)\\right)^{T} \\phi_{\\omega}\\left(E_{\\psi}(x)\\right),\n\t\t\\end{aligned}\n\t\\end{equation}\n\twhere $X=\\{x_1,\\cdots,x_N\\}$ is a set of random images and $\\mathcal{A}=\\{0,1,\\cdots,M^2-1\\}$ represents indices of a $M\\times M$ spatial sized local feature map. Based on this, positive sample pairs are $C^{(i)}_{\\phi}(x)$ and $E_{\\phi}(x)$, and negative sample pairs are $C^{(i)}_{\\phi}(x^{'})$ and $E_{\\phi}(x)$), where $x^{'}$ is a different image from $x$. SimCLR is another popular contrast learning framework, which applies two independent transformations, namely $t_1$ and $t_2$, to obtain the different views $v^{(1)},v^{(2)}=t_1(x),t_2(x)$. The loss function of SimCLR is defined as: \n\t\\begin{equation}\n\t\t\\begin{aligned}\n\t\t\tL_{\\mathrm{SimCLR}}\\left({v}^{(1)}, {v}^{(2)}\\right)=\\frac{1}{N} \\sum_{i=1}^{N}\\left(L_{\\mathrm{NCE}}\\left({v}_{i}^{(1)} ;\\left[{v}^{(2)} ; {v}_{-i}^{(1)}\\right], s_{\\mathrm{SimCLR}}\\right)\\right),\n\t\t\\end{aligned}\n\t\\end{equation}\n\twhere ${v}_{-i}:={v} \\backslash\\left\\{{v}_{i}\\right\\}$ and $s_{\\mathrm{SimCLR}}$ is defined as:\n\t\\begin{equation}\n\t\ts_{\\text {SimCLR }}\\left({v}^{(1)}, {v}^{(2)} ; f, h\\right)=\\frac{h\\left(f\\left({v}^{(1)}\\right)\\right) \\cdot h\\left(f\\left({v}^{(2)}\\right)\\right)}{\\tau \\cdot\\left\\|h\\left(f\\left({v}^{(1)}\\right)\\right)\\right\\|_{2}\\left\\|h\\left(f\\left({v}^{(2)}\\right)\\right)\\right\\|_{2}}.\n\t\\end{equation}\n\tAs shown in Figure \\ref{fig:contrastive_self_supervised}, SimCLR defines more negative pairs to improve the sample utilization compared to InfoNCE. {However, SimCLR needs a large batch size to obtain some sufficiently rich negative samples (In , batch size is set to 4096). To alleviate the attachment of SimCLR to large batch size, MoCo introduce a negative queue to store and update negative samples.}\n\t\\begin{comment}\n\tSun et al. used the triplet matching objective as a pretext task for cGANs: pairs of images with the same category will have similar features and vice versa. Specifically, positive pairs contain the same category, such as two images with black hair; negative pairs contain different category, such as a image with black hair and a image with white hair. The loss function can be expressed as:\n\t\\begin{equation}\n\t\\begin{aligned}\n\t\\mathcal{L}_{D}=-\\lambda_{d}&\\mathop{\\mathbb{E}}\\limits_{x_a,x_p\\sim P_{X|y},x_n\\sim P_{X|y'}}\\bigg[\\log\\big(D_{mch}(x_a,x_p)\\big)\\\\\n\t&+\\log\\big(1-D_{mch}(x_a,x_n)\\big)\\bigg],\\\\\n\t\\mathcal{L}_{G}=-\\lambda_{g}&\\mathop{\\mathbb{E}}\\limits_{x_1,x_2\\sim P_{X|y},x_3\\sim P_{X|y'}}\\bigg[\\log\\big(D_{mch}(G(x_1,y),G(x_2,y))\\big)\\\\\n\t&+\\log\\big(1-D_{mch}(G(x_1,y),G(x_3,y'))\\big)\\bigg],\n\t\\end{aligned}\n\t\\end{equation}\n\twhere $y\\neq y'$, $x_a$ and $x_p$ are positive pairs of real images, while $x_a$ and $x_n$ are negative pairs of real images. \n\t\\end{comment}\n\tThe self-supervised methods mentioned above are also widely applied to the training of GANs. \n\tInspired by Deep InfoMax , Lee et al. propose InfoMax-GAN maximizing the mutual information between local and global features of real and fake images. The regularization of discriminator is expressed as:\n\t\\begin{equation}\n\t\t\\mathcal{L}_\\mathrm{InfoMax-GAN}=\\lambda_d \\{L_{\\mathrm{InfoMAX}}(x_{r})+L_{\\mathrm{InfoMAX}}(x_{f}) \\}.\n\t\\end{equation}\n\twhere $x_r$ and $x_f$ represent sets of real and fake images, respectively.\n\tInspired by SimCLR , some studies introduce different data transformation techniques to create positive and negative pairs during GANs training. Zhao et al propose Cntr-GAN, where SimCLR loss is used to regularize the discriminator on two random augmented copies of both real and fake images. The regularization of discriminator for transformation $T$ is:\n\t\\begin{equation}\n\t\t\\mathcal{L}_{\\mathrm{Cntr-GAN}}=\\lambda_d \\{L_{\\mathrm{SimCLR}}(x_{r}, T(x_{r}))+L_{\\mathrm{SimCLR}}(x_{f}, T(x_{f})) \\}.\n\t\t\\label{Eq:Cntr-GAN}\n\t\\end{equation}\n\tThey also compare the effect of different data transformation techniques (mentioned in Section 5.1) on Cntr-GAN. Figure 6 in shows the FID results of BigGAN adding SimCLR loss on CIFAR-10 dataset. The results illustrate that spatial transformations still work better than visual transformations and the best FID of 11.87 is achieved by applying adjusted SimCLR transformations with the cropping/resizing strength of 0.3. Although, regularization of auxiliary SimCLR loss improves GAN training, but does not outperform existing methods based on simple data augmentations, $e.g.$, bCR (demonstrated on Figure 2 in ). \n\tTo improve the efficiency of contrastive learning, Jeong et al. propose Contrastive Discriminator (ContraD), a way of training discriminators of GANs using improved SimCLR. Different from Cntr-GAN with SimCLR loss on both real and generated images, ContraD uses the SimCLR loss on the real images and the supervised contrastive loss on the generated images. Supervised contrastive loss adopts the contrastive between real and generated images, required as a GAN discriminator. More concretely, for two views $v^{(1)},v^{(2)}=t_1(x),t_2(x)$ with ${t}_{1}, {t}_{2} \\sim {T}$, the loss of real images are:\n\t\\begin{equation}\n\t\tL_{\\text {con }}^{+}\\left(D, h_{{r}}\\right)=L_{\\mathrm{SimCLR}}\\left({v}_{{r}}^{(1)}, {v}_{{r}}^{(2)} ; D, h_{{r}}\\right),\n\t\\end{equation}\n\twhere $h_r$ is a projection head for this loss. However, the loss for generated images, an extended version of contrastive loss to support supervised learning by allowing more than one view to be positive. More concretely, they assume all the views from fake samples have the same label against those from real samples. Formally, for each $v_i^{(1)}$, the positive views are represented by $V_i^{(2)}$ that is a subset of $v^{(2)}$. The supervised contrastive loss is defined by:\n\t\\begin{equation}\n\t\tL_{\\text {SupCon }}\\left({v}_{i}^{(1)}, {v}^{(2)}, V_{i+}^{(2)}\\right)=\n\t\t-\\frac{1}{\\left|V_{i+}^{(2)}\\right|} \\sum_{{v}_{i+}^{(2)} \\in V_{i+}^{(2)}} \\log \\frac{\\exp \\left(s_{{SimCLR}}\\left({v}_{i}^{(1)}, {v}_{i+}^{(2)}\\right)\\right)}{\\sum_{j} \\exp \\left(s_{{SimCLR}}\\left({v}_{i}^{(1)}, {v}_{j}^{(2)}\\right)\\right)}.\n\t\\end{equation}\n\tUsing the notation,the ContraD loss for fake samples are:\n\t\\begin{equation}\n\t\tL_{\\text {con }}^{-}\\left(D, h_{f}\\right)=\n\t\t\\frac{1}{N} \\sum_{i=1}^{N} L_{\\text {SupCon }}\\left({v}_{f, i},\\left[{v}_{f,-i} ; {v}_{{r}}^{(1)} ; {v}_{{r}}^{(2)}\\right],\\left[{v}_{f,-i}\\right] ; D, h_{{f}}\\right),\n\t\\end{equation}\n\twhere $v_f= t_3(G(z))$ is a random view of fake samples ($t_3\\sim T$), and $v_{-i}=v\\backslash \\{v_i\\}$ is subset of $v$ that does not contain $v_i$. It is pertinent to note that authors also use an independent projection header $h_f$ in this loss instead of $h_r$ in $L_{\\text {con }}^{+}\\left(D, h_{{r}}\\right)$. {The adopted supervised contrastive learning on the fake images introduce the real/fake information into contrastive learning, which improves the efficiency of contrastive learning, thus improve the discrimination of the discriminator.}\n\tTo sum up, ContraD learns its contrastive representation by minimizing the following regularization loss:\n\t\\begin{equation}\n\t\tL_{\\text {con }}\\left(D, h_{{r}}, h_{{f}}\\right)=L_{\\text {con }}^{+}\\left(D, h_{{r}}\\right)+\\lambda_{\\text {con }} L_{{con}}^{-}\\left(D, h_{{f}}\\right).\n\t\\end{equation}\n\tThe experimental results show that ContraD consistently improves the performance of GANs compared to other methods, such as Cntr-GAN, DiffAug, bCR, and CR. However, ContraD with different data transformations is not discussed further. \n\t{The achievement of above SimCLR-based contrastive learning methods depends on the sufficiently large batch size. However, large-scale GANs training often has only a small batch size for limited computational resources. Therefore, MoCo-based contrastive learning method, namely InsGen , has been introduced into GANs training. InsGen follows the MoCo-v2 to store the various negative samples with an extra queue. Furthermore, it also introduces a latent space augmentation for fake images. Combining with ADA and MoCo-based contrastive learning, InsGen has achieved state-of-the-art performance on a variety of datasets and training settings. Recently, Li et al. identify that only latent space augmentation for fake images brings the major performance improvement and contrastive learning in real images causes performance drop on limited data generation (DE-GANs). Based on this, they propose FakeCLR, which only applies contrastive learning on perturbed fake samples and devises three related training techniques. The experimental results manifest the new state of the arts in both few-shot generation and limited-data generation.}\n\tIn summary, contrastive self-supervised learning designs different positive and negative pairs and maximizes the mutual information of positive pairs according to the InfoNCE loss. Different from classification and segmentation tasks, two types of samples (real and fake images) exist for generating adversarial networks, which add more possibilities to the definition of positive and negative pairs. In the future, score-based contrastive lea rning may be proposed during the training of GANs. The summary of contrastive self-supervised regularization techniques of GANs is given in Table \\ref{table:self-supervised}.","id":"aff71d0b-4dab-423d-9325-d61d96004987","level":"subsubsection","origin_cites_number":12,"parent_id":"4e2e9490-faa8-4a2f-9dbd-66ae6c225118","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Self-Supervision"],["subsubsection","Contrastive Self-Supervised Learning (CSS)"]],"subsections":[],"title":"Contrastive Self-Supervised Learning (CSS)"},{"cite_extract_rate":1,"cites":[130,108,125],"content":"{According to the perspective of \\textbf{\"Real $\\&$ Fake\"}, many regularization and normalization technologies inspired from supervised learning have been proposed to GANs training. The key point of them is improving the representation and generalizability of the discriminator. \\textit{Data Augmentation and Preprocessing} is a basic operation containing many types such as spatial augmentation, visual augmentation, frequency augmentation, and noise augmentation. Among them, combining adaptive strategies and all augmentation has achieved the most remarkable achievement and has been employed as default operations in most GANs training. \\textit{Consistency Regularization} and \\textit{Self-supervision} are designed additional tasks based on data augmentation, which further improve the efficiency of data augmentation and extract more useful information under stronger data augmentation beyond the existing yet limited practices. Currently, combining contrastive self-supervised learning with adaptive data augmentation has achieved state of the art in GANs training.}","id":"2267f79d-3862-4c72-bdff-9670534c3017","level":"subsection","origin_cites_number":3,"parent_id":"8469161e-625b-422b-90e0-a6cca9638a04","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Real \\& Fake\""],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":1,"cites":[64,135,136,6997],"content":"From the perspective of \"Fitting distribution\", generator is considered as a distribution mapping function and the optimal discriminator is considered to be the distribution divergence. Wasserstein distance is a popular and important in GANs, and it corresponds to the optimal transport of the generator. To solve the dual problem of Wasserstein distance, Lipschitz continuity is introduced into the GANs training. The Wasserstein-based GANs (WGAN and WGAN-GP) have achieved remarkable results during the training. However, some studies suggest that the success of WGAN-GP is not due to the Wasserstein distance and the Lipschitz constraint of discriminator may improve the performance and stability of GANs training regardless of the statistical distance used as a loss function. Therefore, the Lipschitz continuity of discriminator is an essential condition during GANs training. {Weight clipping is a simple and the first solution to enforce a Lipschitz constraint, which clamps the weights of discriminator to a fixed box after each gradient update.} Furthermore, {gradient penalty}, {weight normalization}, and {weight regularization} are widely applied in GANs training for fulfilling Lipschitz continuity as summarized in subsequent subsections.","id":"fdfb947c-602c-4735-9040-96d30810700e","level":"section","origin_cites_number":4,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "]],"subsections":["1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c","16b503b7-d01a-45ea-9b9c-0986b73e7409","3e0461ca-996a-4c2c-8c3e-dbd291acb92b","ee52e377-012a-4b04-8fa5-1df1ed551949"],"title":"Regularization and Normalization of \"Fitting distribution\" "},{"cite_extract_rate":1,"cites":[137,136,88,138,93,135,8317,6997],"content":"Gradient penalty is a simple and direct way to fulfill Lipschitz continuity. Specifically, K-Lipschitz continuity of the function $f$ can be accessed by $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla f(\\hat{x})||_2-{\\rm K})^2$. According to the optimal transport theory mentioned on the Section \\ref{sect:A-1} of the Supplementary Online-only Material, gradient penalty can be used for the approximation of $W_c(\\mu,\\upsilon)$ in WGANs, named WGAN-GP . Specifically, WGAN-GP fulfills the 1-Lipschitz continuity of the discriminator by $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_2-{\\rm 1})^2$, which limits the gradient of the discriminator to 1. Although WGAN-GP solves the instability of GANs training to some extent, the assumption of optimal transport is a constrained linear programming problem. Overly strict restriction reduces the exploratory of the discriminator. In contrast, the optimal transport with the regular term mentioned is an unconstrained optimization problem. Like optimal transport corresponds to 1-Lipschitz continuity, the optimal transport with the regular term corresponds to k-Lipschitz continuity (${\\rm k}\\leq1$) of the discriminator, named WGAN-LP , which is implemented by $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||_2-1\\}\\right)^2\\right]$. WGAN-LP achieves better performance by using a weaker regularization term which enforces the Lipschitz constraint of the discriminator. \n\tWGAN-GP and WGAN-LP introduce Wasserstein distance into GANs framework. Due to the gap between limited input samples and the strict Lipschitz constraint on the whole input sample domain, the approximation of the Wasserstein distance is a challenging task. To this end, WGAN-div introduces a Wasserstein divergence into GANs training. The objective of WGAN-div can be smoothly derived as:\n\t\\begin{equation}\n\t\t\\mathbb{E}_{y\\sim q(y)}[\\varphi(y)]-\\mathbb{E}_{x\\sim p(x)}[\\varphi(x)]+k \\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[ ||\\varphi(\\hat{x})||^p\\right].\n\t\\end{equation}\n\tThe objective of WGAN-div is similar to WGAN-GP and WGAN-LP. It can be considered as achieving 0-Lipschitz continuity of discriminator by adopting $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[||\\nabla D_{\\theta}(\\hat{x})||^p\\right]$. \n\tGenerally, Wasserstein distance and Wasserstein divergence are reliable ways of measuring the difference between fake and real data distribution, which leads to the stable training of WGAN-based algorithms. However, a recent study shows that the c-transform method achieves better estimation of Wasserstein divergence but leads to worse performance compared to the gradient penalty method. The results demonstrate that the success of WGAN-based methodologies cannot truly be attributed to approximate the Wasserstein distance and the gradient penalty methods improve the performance indeed. Furthermore, some studies also demonstrate that gradient penalty methods of discriminator, such as 1-GP, k-GP (${\\rm k}\\leq 1$), and 0-GP stabilize the training and improve the performance of GANs remarkably regardless of the loss functions. Based on these observations, stabilizing GANs training using gradient penalty is widely applied in the research community for various losses of GANs. In the rest of this section, we discuss gradient penalty methods regardless of the loss function by dividing them into three parts: \\textit{1-GP}: $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||-{\\rm 1})^p$, \\textit{k-GP (${\\rm k}\\leq1$)}: $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||-1\\}\\right)^p\\right]$, and \\textit{0-GP}: $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[||\\nabla D_{\\theta}(\\hat{x})||^p\\right]$, where $\\pi$ is the distribution of different image space (entire image space or part of image space) and $||\\cdot||$ represents the norm of the gradient. Generally, the loss function of the discriminator with GP can be formulated as:\n\t\\begin{equation}\\label{GP}\n\t\t\\mathcal{L}_{D}=f(\\phi,\\theta)+\\lambda\\mathcal{L}_{GP},\n\t\\end{equation}\n\twhere $f(\\phi,\\theta)$ is the uniform loss function defined in Eq (\\ref{EQ:eqn1}) and $\\mathcal{L}_{GP}$ is the gradient penalty regularization.","id":"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c","level":"subsection","origin_cites_number":8,"parent_id":"fdfb947c-602c-4735-9040-96d30810700e","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","Gradient Penalty"]],"subsections":["258a5900-7326-4a5a-9ab5-4d4abbbbc5a5","4a0b3b8a-89ed-4329-a9ef-855c13a16be6","349a6a29-8d68-4bed-a022-d5ef2efaccc2"],"title":"Gradient Penalty"},{"cite_extract_rate":0.833333333333333,"cites":[141,140,139,8317,6997],"content":"Gulrajani et al. used 1-GP in WGAN-GP to train GANs. WGAN-GP uses the 2-norm gradient penalty across the entire image domain, which can be formulated as: \n\t\\begin{equation}\\label{wgan-gp}\n\t\t\\mathcal{L}_{GP}=\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_2-1)^2,\n\t\\end{equation}\n\twhere $\\pi$ is the distribution of entire image space approximated by the interpolation of real distribution ($p_r$) and generated distribution ($p_g$): $\\pi=t\\cdot p_r+(1-t)\\cdot p_g$ for $t\\sim U[0,1]$. Although, WGAN-GP stabilizes the training of GANs to a great extent, the overly strict gradient penalty limits the exploratory of discriminator. To loosen the penalty, many efforts of $\\pi$, $||\\cdot||$, and gradient direction are proposed. \n\tTo relax the image distribution, Kodali et al. track the training process of GANs and find that the decrease of the Inception Score (IS) is accompanied by a sudden change of the discriminator’s gradient around the real images. Authors propose DRAGAN by restricting the Lipschitz constant around the real images $\\pi=p_r+\\epsilon$, where $\\epsilon\\sim N_d(0,cI)$.\n\tIn order to relax the gradient direction, Zhou et al. argue that restricting the global Lipschitz constant is unnecessary. Therefore, only maximum gradient is necessary to be penalized:\n\t\\begin{equation}\n\t\t\\mathcal{L}_{GP}=\\left(\\mathop{\\max}\\limits_{\\hat{x}\\sim\\pi}||\\nabla D_{\\theta}(\\hat{x})||_2-1\\right)^2,\n\t\\end{equation}\n\twhere $\\pi=t\\cdot p_r+(1-t)\\cdot p_g$; Furthermore, inspired by Virtual Adversarial Training (VAT) , D{\\'a}vid et al. propose a method, called Adversarial Lipschitz Regularization (ALR), which restricts the 1-Lipschitz continuity at $\\pi=p_r\\cup p_g$ in the direction of adversarial perturbation. {Adversarial perturbation direction is the most unstable direction, Restricting the 1-Lipschitz continuity to the adversarial direction means restricting only the largest Lipschitz constant, which is simpler and more efficient than the previous method.} The proposed ALP shows the SOTA performance in terms of Inception Score and Fréchet Inception Distance among non-progressive growing methods trained on CIFAR-10 dataset.\n\tContrary to the methods which penalize the gradient in Euclidean space, Adler et al. extended the $L_p(p=2)$ space with gradient penalty to Banach space that contains the $L_p$ space and Sobolev space. For the Banach space B, the Banach norm $||.||_{B^*}$ is defined as: \n\t\\begin{equation}\n\t\t||x^*||_{B^*}=\\sup\\limits_{x\\in B}\\frac{x^*(x)}{||x||_B}.\n\t\\end{equation}\n\tThus, the gradient penalty of Banach wasserstein GAN can be expressed as:\n\t\\begin{equation}\n\t\t\\mathcal{L}_{GP}=\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_{B^*}-1)^2,\n\t\\end{equation}\n\twhere $\\pi=t\\cdot p_r+(1-t)\\cdot p_g$. {Banach wasserstein GAN expands the Lipschitz continuity into Banach space containing both $L_p$ space and Sobolev space, which has more restriction than wasserstein GAN.}","id":"258a5900-7326-4a5a-9ab5-4d4abbbbc5a5","level":"subsubsection","origin_cites_number":6,"parent_id":"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","Gradient Penalty"],["subsubsection","1-GP"]],"subsections":[],"title":"1-GP"},{"cite_extract_rate":1,"cites":[88,142,8317],"content":"\\leq1$)}\n\t{k-GP (${\\rm k}\\leq1$) was first tested by Gulrajani et al. and named one sided gradient penalty. It uses the 2-norm gradient penalty across the entire image domain, which is formulated as: }\n\t\\begin{equation}\\label{wgan-lp}\n\t\t\\mathcal{L}_{GP}=\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||_2-1\\}\\right)^2\\right],\n\t\\end{equation}\n\twhere $\\pi$ is the distribution of entire image space approximated by the interpolation of real distribution ($p_r$) and generated distribution ($p_g$): $\\pi=t\\cdot p_r+(1-t)\\cdot p_g$ for $t\\sim U[0,1]$. Inspired by the optimal transport with the regular term, Petzka et al. also used k-GP (${\\rm k}\\leq1$) to training GANs named WGAN-LP. Furthermore, Xu et al show a more general dual form of the Wasserstein distance compared to KR duality (mentioned in section 2.4), named Sobolev duality, which relaxes the Lipschitz constraint but still maintains the favorable gradient\n\tproperty of the Wasserstein distance. Authors also show that the KR duality is a special case of the proposed Sobolev duality. Based on the Sobolev duality, the relaxed gradient penalty of the proposed SWGAN is formulated as:\n\t\\begin{equation}\\label{wgan-sp}\n\t\t\\mathcal{L}_{GP}=\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||^2-1\\}\\right)^2\\right],\n\t\\end{equation}\n\twhere $\\pi=t\\cdot p_r+(1-t)\\cdot p_g$ for $t\\sim U[0,1]$. It is clear that above three method have the same form of gradient penalty. Interestingly, different relaxation methods yield the same form of regularization.","id":"4a0b3b8a-89ed-4329-a9ef-855c13a16be6","level":"subsubsection","origin_cites_number":3,"parent_id":"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","Gradient Penalty"],["subsubsection","k-GP (${\\rm k"]],"subsections":[],"title":"k-GP (${\\rm k"},{"cite_extract_rate":0.8571428571428571,"cites":[142,136,143,8317,60,7002,88,7001,99,94,6997,140,8322,141,92,64,93,135],"content":"\\begin{table}\n\t\t\\caption{The Gradient penalty of the Discriminator. $\\mu$ and $v$ are real and generated distribution, respectively.}\n\t\t\\footnotesize\n\t\t\\label{table:gradient penalty}\n\t\t\\centering\n\t\t\\begin{tabular}{c | c | c | c }\t\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\t{ Method}& { $\\mathcal{L}_{GP}$}& $\\pi$&Lipschitz continuity\t \\\\\n\t\t\t\\hline\n\t\t\tWGAN-GP &$\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_2-1)^2$&$t\\cdot p_r+(1-t)\\cdot p_g$&$||D_{\\theta}||_{Lip}\\to1$\\\\\n\t\t\t\\hline\n\t\t\tDRAGAN &$\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_2-1)^2$&$p_r+\\epsilon$&$||D_{\\theta}||_{Lip}\\to1$\\\\\n\t\t\t\\hline \n\t\t\tMax-GP &$\\left(\\mathop{\\max}\\limits_{\\hat{x}\\sim\\pi}||\\nabla D_{\\theta}(\\hat{x})||_2-1\\right)^2$&$t\\cdot p_r+(1-t)\\cdot p_g$&$||D_{\\theta}||_{Lip}\\to1$\\\\\n\t\t\t\\hline\n\t\t\tALP &$\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_2-1)^2$&$p_r\\cup p_g$&$||D_{\\theta}||_{ALP-Lip}\\to1$\\\\\n\t\t\t\\hline\n\t\t\tBanach-GP &$\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_{B^*}-1)^2$&$t\\cdot p_r+(1-t)\\cdot p_g$&$||D_{\\theta}||_{Lip}\\to1$\\\\\n\t\t\t\\hline\n\t\t\tWGAN-LP &$\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||_2-1\\}\\right)^2\\right]$&$t\\cdot p_r+(1-t)\\cdot p_g$&$||D_{\\theta}||_{Lip}\\leq1$\\\\\n\t\t\t\\hline\n\t\t\tSWGAN &$\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||^2-1\\}\\right)^2\\right]$&$t\\cdot p_r+(1-t)\\cdot p_g$&$||D_{\\theta}||_{Lip}\\leq1$\\\\\n\t\t\t\\hline\n\t\t\tzc-GP &$\\mathbb{E}_{\\hat{x}\\sim\\pi}||\\nabla D_{\\theta}(\\hat{x})||^2_2$&$p_r\\cup p_g$&$||D_{\\theta}||_{Lip}\\to0$\\\\\n\t\t\t\\hline\n\t\t\tGAN-QP &$\\mathcal{L}_{GP}=\\mathbb{E}_{x_r,x_g\\sim\\pi}\\frac{\\left(D_{\\theta}(x_r)-D_{\\theta}(x_f)\\right)^2}{||x_r-x_f||}$&$\\pi=p_r\\cdot p_g$&$\\frac{\\left(D_{\\theta}(x_r)-D_{\\theta}(x_f)\\right)^2}{||x_r-x_f||}\\to 0$\\\\\n\t\t\t\\hline\n\t\t\tZP-Max &$\\mathop{\\max}\\limits_{\\hat{x}\\sim\\pi}||\\nabla D_{\\theta}(\\hat{x})||^2_2$&$t\\cdot p_r+(1-t)\\cdot p_g$&$||D_{\\theta}||_{Lip}\\to0$\\\\\n\t\t\t\\hline\n\t\t\tZP &$\\mathbb{E}_{\\hat{x}\\sim\\pi}||\\nabla D_{\\theta}(\\hat{x})||^2_2$&$t\\cdot p_r+(1-t)\\cdot p_g$&$||D_{\\theta}||_{Lip}\\to0$\\\\\n\t\t\t\\bottomrule\n\t\t\\end{tabular}\n\t\\end{table} \n\tWu et al. used 0-GP, and proposed Wasserstein divergence. According to , Wasserstein divergence is solved by minimizing:\n\t\\begin{equation}\n\t\t\\label{0-gp}\n\t\t\\mathcal{L}_{GP}=\\mathbb{E}_{\\hat{x}\\sim\\pi}||\\nabla D_{\\theta}(\\hat{x})||^2,\n\t\\end{equation}\n\twhere $\\pi$ is both the real distribution ($p_r$) and the generated distribution ($p_g$): $\\pi=p_r\\cup p_g$. Furthermore, Mescheder et al. also demonstrate that the optimization of unregularized GAN is not always locally convergent and some simplified zero centered gradient penalty (zc-GP) techniques, implemented by minimizing Eq (\\ref{0-gp}), can be used to achieve local convergence of GANs. Li et al. introduce the adversarial training to discriminator training, which is turned out to be an adaptive 0-GP. \n\tBesides, some other 0-GP methods are derived by different theoretical derivations. For instance, Su et al. propose a Quadratic Potential (QP) for GANs training with the following formulation:\n\t\\begin{equation}\n\t\t\\mathcal{L}_{GP}=\\mathbb{E}_{x_r,x_g\\sim\\pi}\\frac{\\left(D_{\\theta}(x_r)-D_{\\theta}(x_f)\\right)^2}{||x_r-x_f||},\n\t\\end{equation}\n\twhere $\\pi$ is the joint distribution of the real and generated distributions: $\\pi=p_r\\cdot p_g$; Zhang et al. combine a Total Variational (TV) regularizing term into the training of GANs, that is $|D_{\\theta}(x_r)-D_{\\theta}(x_f)-\\delta|$. According to , the TV term can be approximated by Eq (\\ref{0-gp}), which is exhilarating; Zhou et al. propose the Lipschitz GANs, with the maximum of the gradients penalty for guaranteeing the gradient informativeness:\n\t\\begin{equation}\n\t\t\\mathcal{L}_{GP}=\\mathop{\\max}\\limits_{\\hat{x}\\sim\\pi}||\\nabla f(\\hat{x})||^2_2,\n\t\\end{equation}\n\twhere $\\pi=t\\cdot p_r+(1-t)\\cdot p_g$; Thanh-Tung et al. also propose the 0-GP with gradients penalty at $\\pi=t\\cdot p_r+(1-t)\\cdot p_g$:\n\t\\begin{equation}\n\t\t\\mathcal{L}_{GP}=\\mathbb{E}_{\\hat{x}\\sim\\pi}||\\nabla f(\\hat{x})||^2_2.\n\t\\end{equation}\n\tIn summary, gradient penalty techniques are widely used in the GANs training to achieve Lipschitz continuity of discriminator. As shown in Table \\ref{table:gradient penalty}, many techniques are proposed based on different theories and phenomena. But to the best of our knowledge, there is no fair and comprehensive work comparing the performance of these gradient penalty methods.\n\t\\begin{comment}\n\tThe summary of the gradient penalty of the discriminator is shown in TABLE \\ref{table:gradient penalty}. Based on the summary and the 1-Lipschitz continuity in WGAN-GP, the gradient penalty has been improved in two directions.\n\t\\begin{enumerate}\n\t\\item[*] Limit the Lipschitz constant so that let the $||D||_{Lip}\\leq 1$ or $||D||_{Lip}\\to 0$.\n\t\\item[*] Explore the suitable scope for Lipschitz continuity.\n\t\\end{enumerate}\n\t\\end{comment}\n\tTo compare the performance of various methods intuitively, a comparative experiment on CIFAR-10 and CIFAR-100 datasets is conducted\\footnote{The base framework comes from wgan-gp in \\url{https://github.com/kwotsin/mimicry}}. The results of FID for various gradient penalty methods with different loss functions are shown in Table \\ref{table: experiments results}. The results validate the conclusion in studies , that the Lipschitz constraint of discriminator may improve the performance and stability of GANs training regardless of the statistical distance used as a loss function. All gradient penalty methods improve the performance of GANs upon all three loss functions. Among them, zc-GP obtains the best performance and is widely used in SOTA methods as illustrated in Table \\ref{table:SOTA }.\n\t\\begin{comment}\n\t\\begin{table}\n\t\\caption{The experimental results of different Gradient penalty}\n\t\\label{table: experiments results}\n\t\\setlength{\\tabcolsep}{1mm}\n\t\\centering\n\t\\begin{tabular}{c c c }\t\n\t\\toprule\n\t\\midrule\n\t{ Method}&Inception Score& FID \\\\\n\t\\hline\n\tWGAN-GP &7.869&16.62\\\\\n\tWGAN-LP &7.98&15.85\\\\\n\tWGAN-ALP &8.247&14.19\\\\\n\tWGAN-GP-Max &7.956&18.43\\\\\n\tWGAN-ZP-Max &7.908&17.97\\\\\n\tWGAN-ZP-Sample &8.013&15.87\\\\\n\tWGAN-ZP &7.957&16.08\\\\\n\t\\bottomrule\n\t\\end{tabular}\n\t\\end{table} \n\t\\end{comment}\n\t\\begin{table*}\n\t\t\\tiny\n\t\t\\caption{ FID results on the CIFAR-10 and CIFAR-100 datasets for various gradient penalty methods with different GAN losses.}\n\t\t\\centering\n\t\t\\label{table: experiments results}\n\t\t\\begin{tabular}{cccccccccc}\n\t\t\t\\toprule\n\t\t\t\\multirow{2}{*}{Dataset}&\\multirow{2}{*}{Loss} & \\multicolumn{8}{c}{\\makecell*[c]{Gradient Penalty Methods}} \\\\ \n\t\t\t\\cline{3-10} \n\t\t\t&& \\makecell*[c]{None}&GP&DRAGAN &MAX-GP &LP&zc-GP &ZP-MAX &ZP \\\\ \n\t\t\t\\midrule\n\t\t\t\\multirow{3}{*}{CIFAR-10}&\n\t\t\tGAN& 42.41& 23.45&20.98& 26.65&22.9&\\textbf{19.39}&24.38&23.96 \\\\\n\t\t\t&WGAN&290&30.38& 29.53&37.21&28.31&\\textbf{26.99}&31.28&30.19\\\\\n\t\t\t&Hinge&58.34&21.19&21.77&25.4&20.79&\\textbf{18.75}&23.1&22.58\\\\\n\t\t\t\\cline{1-10} \n\t\t\t\\specialrule{0em}{2pt}{2pt}\n\t\t\t\\multirow{3}{*}{CIFAR-100}&\n\t\t\tGAN&44.5&25.76& 25.37&24.29&23.82&\\textbf{21.81}&26.27&25.38 \\\\\n\t\t\t&WGAN&244&32.28&31.93&38.71&32.19&\\textbf{29.12}&39.75&37.8 \\\\\n\t\t\t&Hinge&59.43&25.13&25.42&28.34&23.67&\\textbf{21.55}&26.19&26.06\\\\\n\t\t\t\\bottomrule\n\t\t\\end{tabular}\n\t\\end{table*}","id":"349a6a29-8d68-4bed-a022-d5ef2efaccc2","level":"subsubsection","origin_cites_number":21,"parent_id":"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","Gradient Penalty"],["subsubsection","0-GP"]],"subsections":[],"title":"0-GP"},{"cite_extract_rate":0,"cites":[],"content":"WGAN is a popular and important generative adversarial network. From the optimal transport introduced on the Section \\ref{sect:A-1} of the Supplementary Online-only Material, to obtain $W_c(\\mu,\\upsilon)=\\mathop{\\max}\\limits_{||\\varphi_\\xi||_L\\leq 1}\\left\\{\\int_X \\varphi_\\xi \\mathrm{d}\\mu\\ -\\int_Y \\varphi_\\xi \\mathrm{d}\\upsilon\\right\\}$, the discriminator must satisfy the 1-Lipschitz continuity. According to the Section \\ref{sect:A-3} in the Supplementary Online-only Material, the spectral norm $||W||_2$ can be used to represent the Lipschitz constant $\\mathrm{C'}$. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately. Hence, \\textit{Weight Normalization} and \\textit{Weight Regularization} can also be used to enable the Lipschitz continuity of the discriminator.","id":"16b503b7-d01a-45ea-9b9c-0986b73e7409","level":"subsection","origin_cites_number":0,"parent_id":"fdfb947c-602c-4735-9040-96d30810700e","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","Weight Normalization and Weight Regularization"]],"subsections":["930186fd-4bd2-4fae-a134-426d00072853","9c9a2b69-11c0-4e4b-90b7-101fced8cb32"],"title":"Weight Normalization and Weight Regularization"},{"cite_extract_rate":0.5,"cites":[78,144],"content":"Spectral norm of the weight and the Lipschitz constant express the same concept. Therefore, weight normalization is another method to achieve Lipschitz continuity. {More important, weight normalization methods are Non-sampling-based, which don’t have the lack of support problem in contrast to gradient penalties.} Spectral normalization of the weight limits the Lipschitz constant to 1. Certainly, upper bound of the spectral norm can be used to normalize the weights, achieving $k (k\\leq1)$ Lipschitz continuity. The following lemmas put forward some upper bounds of the spectral norm.\n\t\\newenvironment{lemma3}{{\\indent\\it \\textbf{Lemma 3.1:}}}{\\hfill \\par}\n\t\\begin{lemma3}\n\t\t\\textit{\n\t\t\tIf $\\lambda_1\\leq\\lambda_2\\leq\\cdots\\leq\\lambda_M$\n\t\t\tare the eigenvalues of the $W^\\top W$, then the spectral norm$||W||_2=\\sqrt{\\lambda_M}$; The Frobenius norm$||W||_F=\\sqrt{\\sum_{i=1}^{M}\\lambda_i}$ \n\t\t}\n\t\\end{lemma3}\n\t\\newenvironment{lemma4}{{\\indent\\it \\textbf{Lemma 3.2:}}}{\\hfill \\par}\n\t\\newenvironment{lemma5}{{\\indent\\it \\textbf{Lemma 3.3:}}}{\\hfill \\par}\n\t\\newenvironment{proof3}{{\\indent\\it Proof 3.1:}}{\\hfill $\\square$\\par}\n\t\\newenvironment{proof4}{{\\indent\\it Proof 3.2:}}{\\hfill $\\square$\\par}\n\t\\newenvironment{proof5}{{\\indent\\it Proof 3.3:}}{\\hfill $\\square$\\par}\n\t\\begin{proof3}\n\t\tSee and \n\t\\end{proof3}\n\t\\begin{lemma4}\n\t\t\\textit{\n\t\t\tFor a $n\\times m$ matrix, $||W||_1=\\mathop{\\max}\\limits_{j}\\sum_{i=1}^{n}|a_{i,j}|$, $ ||W||_\\infty=\\mathop{\\max}\\limits_{i}\\sum_{j=1}^{m}|a_{i,j}|$, then $||W||_2\\leq\\sqrt{||W||_1||W||_\\infty}$\n\t\t}\n\t\\end{lemma4}\n\t\\begin{proof4}\n\t\tSee \n\t\\end{proof4}\n\t\\begin{lemma5}\n\t\t\\textit{\n\t\t\tFor a $n\\times m$ matrix, \n\t\t\t$||W||_F=\\sqrt{\\left(\\sum_{j=1}^{m}\\sum_{i=1}^{n}|a_{i,j}|^2\\right)}$, then $||W||_2\\leq||W||_F$\n\t\t}\n\t\\end{lemma5}\n\t\\begin{proof5}\n\t\tSee \n\t\\end{proof5}\n\t1-Lipschitz continuity can be expressed by the spectral normalization. Miyato et al. control the Lipschitz constant through spectral normalization $W_\\sigma=\\frac{W}{||W||_2}$ of each layer for D, leading to a better result than WGAN-GP. {Practically, the power iteration method is used as a fast approximation for the spectral norm ($||W||_2$).} Similarly, according to the optimal transport with regular term, Lipschitz constant of discriminator should be less than or equal to 1. Correspondingly, upper bound of the spectral norm can be utilized to normalize the weight ($||W_\\sigma||_2\\leq1$), achieving $k (k\\leq1)$ Lipschitz continuity. In terms of Lemma 3.2 and Lemma 3.3, $\\sqrt{||W||_1||W||_\\infty}$ and Frobenius norm ($||W||_F$) are simple upper bound of the spectral norm ($||W||_2$) and can be used to normalize the weight. For example, Zhang et al. use the $\\sqrt{||W||_1||W||_\\infty}$, seeking for an approximation of the spectral norm that is easy to calculate. Miyato et al. explain that the Frobenius norm is a restriction on all eigenvalues. It is different from the spectral norm, which only constrains the maximum eigenvalue. Authors conjecture that Frobenius normalization affects the network's ability to express, but no experiments are reported to compare it with the spectral normalization. Liu et al. find that the mode collapse is often accompanied by the collapse of the eigenvalue of the discriminator. Because the spectral normalization only limits the maximum eigenvalue, and the eigenvalue collapse means the remaining eigenvalues suddenly decrease. Therefore, authors adopt the following methods to prevent the collapse of the eigenvalues:\n\t\\begin{equation}\n\t\tW_{\\sigma}=\\frac{W+\\nabla W}{||W||_2}=\\frac{W}{||W||_2}+\\frac{\\nabla W}{||W||_2}.\n\t\\end{equation}\n\tThe results demonstrate that this method effectively prevents mode collapses. Although the experiments are reported in this study, but it misses theoratical proofs. Therefore the relationship between the matrix eigenvalues and GAN performance is not clear.\n\tFew researches focus on weight normalization as demonstrated in Table \\ref{table:norm normalization}. Among these studies, spectral normalization is widely applied in some SOTA methods, as demonstrated in Section 7.\n\t\\begin{table}\n\t\\caption{The summary of the weight normalization and wight regularization.}\n\t\\label{table:norm normalization}\n\t\\centering\n\t\\begin{tabular}{c | c | c }\t\n\t\t\\toprule\n\t\t\\midrule\n\t\t{ Method}&Implementation &Motivation\t \\\\\n\t\t\\hline\n\t\tSpectral normalization (SN) &$W_\\sigma=W/||W||_2$&$||D||_{Lip}\\to1$\\\\\n\t\t\\hline\n\t\tF normalization &$W_\\sigma=W/||W||_F$&$||D||_{Lip}\\leq1$\\\\\n\t\t\\hline\n\t\tMixed normalization &$W_\\sigma=W/\\sqrt{||W||_1||W||_\\infty}$&$||D||_{Lip}\\leq1$\\\\\n\t\t\\hline\n\t\tSpectral increment normalization &$W_\\sigma=W/||W||_2+\\nabla W/||W||_2$&$||D||_{Lip}\\to1$\\\\\n\t\t\\bottomrule\n\t\\end{tabular}\n\\end{table}","id":"930186fd-4bd2-4fae-a134-426d00072853","level":"subsubsection","origin_cites_number":4,"parent_id":"16b503b7-d01a-45ea-9b9c-0986b73e7409","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","Weight Normalization and Weight Regularization"],["subsubsection","Weight Normalization"]],"subsections":[],"title":"Weight Normalization"},{"cite_extract_rate":0.5,"cites":[65],"content":"Compared with spectral normalization similar to 1-GP, spectral regularization is similar to the 0-GP. Kurach et al. use the $\\mathcal{L}_R=||W||_2$ to regularize the loss function. Zhou et al. also use the $L_P$-norm ($P=1,F,\\infty$) to regularize the discriminator. However, these studies have worse performance than weight normalization and did not catch much attention among researchers.","id":"9c9a2b69-11c0-4e4b-90b7-101fced8cb32","level":"subsubsection","origin_cites_number":2,"parent_id":"16b503b7-d01a-45ea-9b9c-0986b73e7409","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","Weight Normalization and Weight Regularization"],["subsubsection","Weight Regularization"]],"subsections":[],"title":"Weight Regularization"},{"cite_extract_rate":1,"cites":[145,146],"content":"}\n\tGradient normalization is also a popular method to impose the Lipschitz constraint on the discriminator. As we all know, 1-Lipschitz constraint can be implemented by $\\mathop{\\min}\\mathbb{E}_{{x}\\sim\\pi}(||\\nabla_{x} D_{\\theta}({x})||_2-1)^2$ to let the gradient of the discriminator ($||\\nabla_{{x}} D_{\\theta}({x})||_2$) equal to 1. Therefore, control the Lipischitz constant through gradient normalization $\\hat{D}_{\\theta}(x)=\\frac{D_{\\theta}(x)}{||\\nabla_{{x}}D_{\\theta}({x})||_2}$ for $D_{\\theta}$. Accordingly, the gradient of $\\hat{D}_{\\theta}$ can be represented as $||\\nabla_{{x}} \\hat {D}_{\\theta}({x})||_2=||\\nabla_{{x}}\\left(\\frac{D_{\\theta}(x)}{||\\nabla_{{x}}D_{\\theta}({x})||_2}\\right)||_2$, equaling to 1. To ensure the boundedness, different studies have different implementation. For instance, adopts $\\hat{D}_{\\theta}(x)=\\frac{D_{\\theta}(x)}{||\\nabla_{{x}}D_{\\theta}({x})||_2+D_{\\theta}(x)}$ and adopts $\\hat{D}_{\\theta}(x)=\\frac{D_{\\theta}(x)}{||\\nabla_{{x}}D_{\\theta}({x})||_2+\\epsilon}$. Extensive experiments demonstrate that both implementation of gradient normalization attain significant performance gains comparing to gradient penalty, weight normalization, and weight regularization.","id":"3e0461ca-996a-4c2c-8c3e-dbd291acb92b","level":"subsection","origin_cites_number":2,"parent_id":"fdfb947c-602c-4735-9040-96d30810700e","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","{Gradient Normalization"]],"subsections":[],"title":"{Gradient Normalization"},{"cite_extract_rate":1,"cites":[145],"content":"}\n\t{As mentioned above, weight clipping, gradient penalty, weight regularization, weight normalization, and gradient normalization all could enable Lipschitz continuity of the discriminator. However, what are the advantages and disadvantages of investigated techniques? Imposing the Lipschitz constraint on the discriminator can be characterized by three properties . 1) \\textit{model-} or \\textit{module-wise} \\textit{constraint}. Model-wise constraint is defined as methods that constraint objective depends on full model, while module-wise constraint is defined as methods that constraint objective depends on layers. Generally, model-wise constraint is better since module-wise constraint is strict, which limits\n\t\tthe layer capacities and reduces the power of discriminator. 2) \\textit{sampling-based} or \\textit{non-sampling} \\textit{-based} \\textit{constraint}. Sampling-based constraint is defined as requiring sampling data during usage, while non-sampling-based constraint depends on the model, not data sampling. Generally, non-sampling-based constraint performs better since Lipschitz constraint should be fulfilled on the entire data manifold, not only sampling data. 3) \\textit{Hard} or \\textit{soft constraint}. {The accurate constraint of Lipschitz continuity is defined as hard constraint and the converse to be soft constraint. Hard constraint has achieved the exact Lipschitz continuity through limiting the spectral norm, which is expected to perform better. While soft constraint only obtain the Lipschitz continuity approximatively through optimization.} Table \\ref{table:lipschitz continuity} summarizes the properties of different technologies, from which gradient normalization is a model-wise, non-sampling-based, and hard constraint method.}\n\t\\begin{table}\n\t\t\\caption{Summary of different regularization and normalization technologies for imposing Lipschitz continuity.}\n\t\t\\label{table:lipschitz continuity}\n\t\t\\centering\n\t\t\\begin{tabular}{c | c | c | c }\t\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\tMethod& Model-wise&Non-sampling-based&Hard\\\\\n\t\t\t\\hline\n\t\t\tWeight Clipping&&\\checkmark&\\\\\n\t\t\t\\hline\n\t\t\tGradient Penalty&\\checkmark&&\\\\\n\t\t\t\\hline\n\t\t\tWeight Regularization&&\\checkmark&\\\\\n\t\t\t\\hline\n\t\t\tWeight Normalization&&\\checkmark&\\checkmark\\\\\n\t\t\t\\hline\n\t\t\tGradient Normalization&\\checkmark&\\checkmark&\\checkmark\\\\\n\t\t\t\\bottomrule\n\t\t\\end{tabular}\n\t\\end{table}","id":"ee52e377-012a-4b04-8fa5-1df1ed551949","level":"subsection","origin_cites_number":1,"parent_id":"fdfb947c-602c-4735-9040-96d30810700e","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Fitting distribution\" "],["subsection","{Summary"]],"subsections":[],"title":"{Summary"},{"cite_extract_rate":1,"cites":[98,54,99,7190,8323],"content":"Assuming the objectives of GANs are convex-concave, some studies have proposed the global convergence of GANs . However, these theoretical convergence analyses are only applicable to the GANs with the optimal discriminator. Therefore, some studies focus on analyzing the local convergence of GANs. According to Nagarajan et al. and Mescheder et al. , under some assumptions, GANs dynamics are locally convergent. However, if these assumptions are not satisfied, especially if the data distributions are not continuous, GANs dynamics do not always converge locally unless some regularization techniques are used.\n\tWe review {Jacobian regularization} techniques in this section, which minimize the Jacobian matrix to achieve local convergence. With the same motivation, Mescheder et al. propose a simplified gradient penalties method, named zero-centered gradient penalties (zc-GP), that guarantees the local convergence under suitable assumptions. Since it is similar to 0-GP, we cover it in Section 4.","id":"44c32216-8be6-4e5c-aa2f-ca241ea4139f","level":"section","origin_cites_number":5,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Training dynamics\""]],"subsections":["f5ab6a67-88dd-45b6-aeae-1b85b061e943","8377346c-065d-4bc5-adf5-275fad74664d"],"title":"Regularization and Normalization of \"Training dynamics\""},{"cite_extract_rate":1,"cites":[74,149,7003,98,148,99,147,7190],"content":"In \\textit{Proposition 2.2} of Section 2: absolute values of all eigenvalues of the Jacobian matrix ($v^{'}(\\phi,\\theta)$) are expected to be less than 1 at the fixed point, which is equivalent to the real part of the eigenvalue being negative. Additionally, the learning rate must be relatively low . To meet these requirements, Mescheder et al. used the Consensus Optimization (ConOpt) to make the real part of the eigenvalue negative. Its regularized updates are:\n\t\\begin{equation} \\label{eqn1}\n\t\t\\begin{split}\n\t\t\t\\phi^{(k+1)}=\\phi^{(k)}+h\\nabla_\\phi\\left(- f(\\phi^{(k)},\\theta ^{(k)})-\\gamma L(\\phi^k,\\theta^k)\\right),\\\\\n\t\t\t\\theta^{(k+1)}=\\theta^{(k)}+h\\nabla_\\theta\\left(f(\\phi^{(k)},\\theta ^{(k)})-\\gamma L(\\phi^k,\\theta^k)\\right),\n\t\t\\end{split}\n\t\\end{equation}\n\twhere $L(\\phi^k,\\theta^k)=\\frac{1}{2}||v(\\phi^k,\\theta^k)||^2=\\frac{1}{2}\\left(||\\nabla_\\phi f(\\phi^k,\\theta^k)||^2+||\\nabla_\\theta f(\\phi^k,\\theta^k)||^2\\right)$ is the regularization of the Jacobian matrix.\n\tApart from , Nagaraja et al. also analyze the relationship between local convergence of GANs and all eigenvalues of the Jacobian\n\tof the gradient vector field. Authors prove the local convergence for absolutely continuous generator and data distributions under certain regularity assumptions. This requires the loss function of the GANs to be strictly concave, which is not the case for some GANs. Based on this, a simple regularization technology that regularized the generator using the gradient of the discriminator is proposed by Nagaraja et al. . The regularized updates for the generator can be expressed as:\n\t\\begin{equation} \\label{eqn1}\n\t\t\\phi^{(k+1)}=\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\phi||\\nabla_\\theta f(\\phi^{k},\\theta^{k})||^2.\n\t\\end{equation}\n\tHerein, the update of the discriminator is similar to SimGD. Furthermore, Nie et al. propose a method that only regularizes the discriminator. The regularized update of the discriminator in this case is given by:\n\t\\begin{equation} \\label{eqn1}\n\t\t\\theta^{(k+1)}=\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\theta||\\nabla_\\phi f(\\phi^{k},\\theta^{k})||^2.\n\t\\end{equation}\n\tThe update of the generator is the same as SimGD. Nie et al. propose JAcobian REgularization (JARE) that regularizes both the generator and the discriminator. The regularized updates for the generator and the discriminator are:\n\t\\begin{equation} \\label{eqn1}\n\t\t\\begin{split}\n\t\t\t\\phi^{(k+1)}=\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\phi||\\nabla_\\theta f(\\phi^{k},\\theta^{k})||^2,\\\\\n\t\t\t\\theta^{(k+1)}=\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\theta||\\nabla_\\phi f(\\phi^{k},\\theta^{k})||^2.\n\t\t\\end{split}\n\t\\end{equation}\n\tThe key difference between JARE and ConOpt is that JARE does not contain the Hessians $\\nabla^2_{\\phi,\\phi}f(\\phi^{k},\\theta^k)$ and $\\nabla^2_{\\theta,\\theta}f(\\phi^{k},\\theta^k)$ in the regularization term. \n\t{There are several Jacobian regularization methods that have been proposed to deal with the training instabilities of GANs. What is the difference of them? Nie et al. consider a simple toy example to analyse the convergence of GANs. There may exist two factors of the Jacobian in the GANs dynamics simultaneously that destroy the GANs training: (i) the Phase Factor, i.e., the Jacobian has complex eigenvalues with a large imaginary-to-real ratio; (ii) the Conditioning Factor, i.e., the Jacobian is ill-conditioned. According to the toy example,\n\t\tOnly Regularizing Generator , Only Regularizing Discriminator , and ConOpt could only alleviate the impact of the Phase Factor but not alleviating the impact of the Conditioning Factor. However, JARE can address both factors by construction}.\\footnote{Intuitively, a reason for not introducing Hessians in JARE is to avoid the risk of reversing the gradient flows, which may diverge the GAN training dynamics (see Appendix C in for a detailed explanation).}\n\tThe above discussions of local convergence during GANs training involve a premise: absolutely continuous data and generator distributions. Indeed, the assumption of absolute continuity is not true for common cases of GANs, where both distributions, specially the data distribution, may lie on lower-dimensional manifolds . More generally, Mescheder et al. extend the convergence proof by to the case where the generator and data distribution do not locally have the same support. Based on this, a simplified zero-centered gradient penalties (zc-GP) method is proposed, which guarantees the local convergence under suitable assumptions. Zc-GP is obtained from the training dynamics, which is similar to 0-GP methods mentioned in Section 4.\n\t{Furthermore, there are also other literature studies analyze the training of GANs through other tools. For instance, introduces a novel algorithm, competitive gradient descent (CGD), that is a natural extension of gradient descent to the competitive setting. Different from gradient descent ascent (GDA) in Eq (\\ref{eq:eqn27}), CGD does not need to reduce the stepsize to match the increase of the interactions to avoid divergence. Specifically, CGD introduces an equilibrium term that lets each player prefer strategies that are less vulnerable to the actions of the other player. also elucidates the cause of undesirable convergence of GDA is leader's (discriminator) gradient step takes the system away from the ridge, which has undesirable convergence properties and requires using very small learning rates to converge. To mitigate this, Follow-the-Ridge (FR) term ($\\mathbf{H}_{\\mathbf{\\theta} \\mathbf{\\theta}}^{-1} \\mathbf{H}_{\\mathbf{\\theta} \\mathbf{\\phi}} \\nabla_{\\mathbf{\\phi}} f\\left(\\mathbf{\\phi}^{(k)}, \\mathbf{\\theta}^{(k)}\\right)$) has been added to the updating of the discriminator. studies the continuous-time dynamics induced by GANs training. In this perspective, instabilities in training GANs arise from the integration error in discretizing the continuous dynamics. It treats GANs training as solving ODEs and shows that higher-order solvers lead to better convergence.}","id":"f5ab6a67-88dd-45b6-aeae-1b85b061e943","level":"subsection","origin_cites_number":8,"parent_id":"44c32216-8be6-4e5c-aa2f-ca241ea4139f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Training dynamics\""],["subsection","Jacobian Regularization"]],"subsections":[],"title":"Jacobian Regularization"},{"cite_extract_rate":0.8,"cites":[74,98,99,7190],"content":"In summary, Jacobian regularization techniques are obtained from the training dynamics of GANs, which are used for achieving local convergence and stabilizing training. The summary of the Jacobian regularization methods is demonstrated in Table \\ref{table:Jacobian regularization}. Jacobian regularization is similar to the Gradient penalty in terms of update form. In general, zc-GP is used in many SOTA methods, as demonstrated in Section 7.\n\t\\begin{table}\n\t\t\\caption{The summary of the Jacobian regularization.}\n\t\t\\tiny\n\t\t\\label{table:Jacobian regularization}\n\t\t\\centering\n\t\t\\begin{tabular}{c|c|c}\t\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\tMethod& regularized updates of generator ($\\phi^{(k+1)}$)®ularized updates of discriminator ($\\theta^{(k+1)}$)\t \\\\\n\t\t\t\\hline\n\t\t\tSimGD &$\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})$&$\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})$\\\\\n\t\t\t\\hline\n\t\t\tConOpt &$\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\phi||v(\\phi^{(k)},\\theta^{(k)})||^2$&$\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\theta||v(\\phi^{(k)},\\theta^{(k)})||^2$\\\\\n\t\t\t\\hline\n\t\t\tGenerator &$\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\phi||\\nabla_\\theta f(\\phi^{(k)},\\theta^{(k)})||^2$&$\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})$\\\\\n\t\t\t\\hline\n\t\t\tDiscriminator &$\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})$&$\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\theta||\\nabla_\\phi f(\\phi^{(k)},\\theta^{(k)})||^2 $\\\\\n\t\t\t\\hline\n\t\t\tJARE &$\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\phi||\\nabla_\\theta f(\\phi^{(k)},\\theta^{(k)})||^2$&$\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\theta||\\nabla_\\phi f(\\phi^{(k)},\\theta^{(k)})||^2 $\\\\\n\t\t\t\\hline\n\t\t\tzc-GP &$\\phi^{(k)}-h\\nabla_\\phi f(\\phi^{(k)},\\theta ^{(k)})$&$\\theta^{(k)}+h\\nabla_\\theta f(\\phi^{(k)},\\theta ^{(k)})-\\frac{1}{2}h\\gamma\\nabla_\\theta||\\nabla D_{\\theta}(x)||^2 $\\\\\n\t\t\t\\bottomrule\n\t\t\\end{tabular}\n\t\\end{table}","id":"8377346c-065d-4bc5-adf5-275fad74664d","level":"subsection","origin_cites_number":5,"parent_id":"44c32216-8be6-4e5c-aa2f-ca241ea4139f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Training dynamics\""],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0,"cites":[],"content":"In addition to the three groups mentioned above, this section discusses and summarizes the remaining regularization and normalization techniques, namely, \\textit{Layer Normalization} and \\textit{Inverse Gradient Penalty}. \\textit{Layer Normalization} consists of unconditional-based layer normalization and conditional-based layer normalization, the former inspired by supervised learning is used to accelerate training, but its impact on the GANs is small and sometimes \n\tdrops the performance, while the latter is used in the conditional generation and significantly improves the performance of conditional generation; On the other hand, \\textit{Inverse Gradient Penalty} mitigates mode collapse by maximizing the Lipschitz constant of the generator.","id":"bc57c604-a3ec-457f-ab6f-07cf5243c526","level":"section","origin_cites_number":0,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Other methods\""]],"subsections":["253f0a0c-e301-40a7-a3f4-7aa82d059919","ed3fad77-b260-419c-962b-ecc924abe2e8"],"title":"Regularization and Normalization of \"Other methods\""},{"cite_extract_rate":1,"cites":[71],"content":"Data in machine learning is expected to be independent and identically distributed ($i.i.d$). However, in terms of deep learning, because of the Internal Covariate Shift (ICS) , inputs of each neuron do not satisfy the $i.i.d$, making the training of the deep neural networks hard and unstable. Layer normalization\\footnote{\\label{ft:10} layer normalization is different from the Layer Normalization (LN), where layer normalization is a general term for a class of methods such as BN, LN. } has been proposed to avoid such problems. The general form of the layer normalization is (the difference between the normalization methods lies in the choice of $h$ and the calculation of $\\mathbb{E}[h]$ and $var[h]$):\n\t\\begin{equation}\n\t\th_N=\\frac{x-\\mathbb{E}[h]}{\\sqrt{var[h]+\\epsilon}}\\cdot\\gamma+\\beta.\n\t\\end{equation}\n\tFor GANs, the layer normalization is divided into two parts: {\\textit{unconditional-based layer normalization}} and {\\textit{conditional-based layer normalization}}. Unconditional-based layer normalizations are used for unconditional generation similar to the other deep neural networks. On the other hand, conditional-based layer normalizations are used for the generator of the conditional generation, where the shift and scale parameters ($\\gamma, \\beta$) depend on the condition information, as given below:\n\t\\begin{equation}\n\t\th_N=\\frac{x-\\mathbb{E}[h]}{\\sqrt{var[h]+\\epsilon}}\\cdot\\gamma(c)+\\beta(c).\n\t\\end{equation}","id":"253f0a0c-e301-40a7-a3f4-7aa82d059919","level":"subsection","origin_cites_number":1,"parent_id":"bc57c604-a3ec-457f-ab6f-07cf5243c526","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Other methods\""],["subsection","Layer Normalization"]],"subsections":["e30468f0-4365-4dc0-8ffe-61832a49d364","f33ea61f-e2e1-4dfc-9f9a-3c4da4ad1e64"],"title":"Layer Normalization"},{"cite_extract_rate":1,"cites":[154,155,57,8324,153,156,150,152,151,7194,157,78,71],"content":"Unconditional-based layer normalization is used for both the generator and discriminator with the same motivation as in other deep neural networks. Ioffe et al. proposed the first normalization for neural networks, namely, Batch Normalization (BN). Batch normalization adopts the data of the mini-batch to compute the mean and variance, making the data distribution of each mini-batch approximately the same. Miyato et al. used the BN in GANs. BN normalizes at the mini-batch level, which destroys the difference between pixels during the generation on account of image generation being a pixel-level task. {Therefore, Batch Norm can be less applicable to style transfer and can’t be used with gradient penalty\n\t\tmethods since the gradient would be dependent on multiple inputs.} Contrary to BN which normalizes the same channel with different images, Layer Normalization\\textsuperscript{\\ref {ft:10}} (LN) normalizes different channels of a single image that also destroys the diversity between channels for the pixel-by-pixel generative model . Instance Normalization (IN) has also been proposed for style transformation that is adopted for a single channel of a single image. Moreover, Group Normalization (GN) sits between LN and IN, which first divides the channel into many groups, and normalizes different groups of a single image. Compared to normalization of input of neural networks in BN, LN, IN and GN, Weight Normalization (WN) normalizes the weight matrix of neural networks. Miyato et al. also used this normalization in GANs. \n\tIn summary, unconditional-based layer normalization in GANs is similar to other neural networks. The related summaries are shown in Table \\ref{table:layer normalization}. To the best of our knowledge, no study compares the performance of these methods, therefore, we demonstrate the FID results\\footnote{The base framework comes from the SNGAN in \\url{https://github.com/kwotsin/mimicry}} for different normalization methods on CIFAR-10 and CIFAR-100 datasets in Table \\ref{table:normalization}. Among them, LN and GN obtained better performance than the most popular normalization method: Spectral normalization (mentioned in Section 4.2) and other methods significantly affect the stability of GANs training.\n\t\\begin{table*}\n\t\t\\caption{The summary of the layer normalization}\n\t\t\\label{table:layer normalization}\n\t\t\\scriptsize\n\t\t\\centering\n\t\t\\begin{tabular}{c|c | c | c }\t\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\tMethod&{ Reference}& Classification&Inputs of $\\gamma(c)$ and $\\beta(c)$\t \\\\\n\t\t\t\\hline\n\t\t\tBatch Normalization (BN)&2018 & unconditional-based&-\\\\\n\t\t\t\\hline\n\t\t\tLayer Normalization (LN)&2018 & unconditional-based&-\\\\\n\t\t\t\\hline\n\t\t\tInstance Normalization (IN)&2018 & unconditional-based&-\\\\\n\t\t\t\\hline\n\t\t\tGroup Normalization (GN)&2018 & unconditional-based&-\\\\\n\t\t\t\\hline\n\t\t\tWeight Normalization (WN) &2018 & unconditional-based&-\\\\\n\t\t\t\\hline\n\t\t\tConditional Batch Normalization (CBN)&2018 &conditional-based&class label\\\\\n\t\t\t\\hline\n\t\t\tAdaptive Instance Normalization (AdaIN)&2017 ,2019 &conditional-based&target images\\\\\n\t\t\t\\hline\n\t\t\tSpatially-adaptive (de) Normalization (SPADE)&2019 &conditional-based&sematic segmentation map\\\\\n\t\t\t\\hline\n\t\t\tAttentive Normalization(AN)&2020 &conditional-based&self\\\\\n\t\t\t\\bottomrule\n\t\t\\end{tabular}\n\t\\end{table*}","id":"e30468f0-4365-4dc0-8ffe-61832a49d364","level":"subsubsection","origin_cites_number":13,"parent_id":"253f0a0c-e301-40a7-a3f4-7aa82d059919","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Other methods\""],["subsection","Layer Normalization"],["subsubsection","Unconditional-based layer Normalization"]],"subsections":[],"title":"Unconditional-based layer Normalization"},{"cite_extract_rate":1,"cites":[154,8324,150,151,157,7194],"content":"Conditional-based layer normalization is only used for the generator of the conditional generation. It aims to introduce conditional information to each layer of the generator, which helps to improve the quality of the generated images. $\\gamma(c)$ and $\\beta(c)$ in Eq (44) are calculated with different features or class labels as input to the neural network in different methods. Miyato et al. and Zhang et al. used the Conditional Batch Normalization (CBN) to encode class labels, thereby improving the quality of conditional generation. Huang et al. and Karras et al. used the Adaptive Instance Normalization (AdaIN) with target images to improve the accuracy of style transfer. Park et al. used the Spatially-Adaptive (de) Normalization (SPADE) with semantic segmentation image to incorporate semantic information into all layers. Wang et al. used the Attentive Normalization (AN) to model long-range dependent attention, which is similar to self-attention GAN .\n\tIn summary, the main difference between these conditional-based normalizations is the content of conditional inputs (c in Eq (45)). As the information of inputs is gradually enriched, the performance of conditional generation is gradually improved. The related summaries are shown in Table \\ref{table:layer normalization}.","id":"f33ea61f-e2e1-4dfc-9f9a-3c4da4ad1e64","level":"subsubsection","origin_cites_number":6,"parent_id":"253f0a0c-e301-40a7-a3f4-7aa82d059919","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Other methods\""],["subsection","Layer Normalization"],["subsubsection","Conditional-based layer Normalization"]],"subsections":[],"title":"Conditional-based layer Normalization"},{"cite_extract_rate":1,"cites":[56,159,78,160,158],"content":"Mode collapse is a common phenomenon in GANs' training, that is, changes in the latent space do not cause changes in the generated images. Geometrically, the phenomenon means that all the tangent vectors of the manifold are no longer independent of each other - some tangent vectors either disappear or become linearly correlated with each other. Intuitively, we can solve this problem by maximizing the Lipschitz constant of the generator, which is opposite of the gradient penalty of the discriminator described in the previous section. Based on this, inverse gradient penalty of the generator has been proposed. Concretely, under the little perturbation of the latent space, the generator needs to produce different images. Yang et al. use it in conditional generation, especially for tasks that are rich in conditional information, such as inpainting and super-resolution. \n\t\\begin{equation}\n\t\t\\mathop{\\max}\\limits_{G}\\mathcal{L}_z(G)=\\mathop{\\max}\\mathbb{E}_{z_1,z_2}\\left[\\mathop{\\min}\\left(\\frac{||G(y,z_1)-G(y,z_2)||}{||z_1-z_2||},\\tau\\right)\\right],\n\t\\end{equation}\n\twhere $y$ is the class label and $\\tau$ is the bound to ensure numerical stability. Unlike the intuition-based study described above, Odena et al. demonstrate that the decreasing of singular value in the Jacobian matrix of the generator is the main reason for the mode collapse during GANs training. Furthermore, the singular value can be approximated by the gradient, so Jacobian clamping is used to limit singular values to $[\\lambda_{min},\\lambda_{\\max}]$. The loss is expressed as:\n\t\\begin{equation}\n\t\t\\mathop{\\min}\\limits_{G}\\mathcal{L}_z(G)=\\big(\\mathop{\\max}(Q,\\lambda_{\\max})-\\lambda_{\\max}\\big)^2\n\t\t+\\big(\\mathop{\\min}(Q,\\lambda_{\\min})-\\lambda_{\\min}\\big)^2,\n\t\\end{equation}\n\twhere $Q=||G(z)-G(z')||/||z-z'||$. \n\t{In summary, the above two methods are similar and mitigate the model collapse of generator to some extent. The key point is to improve the sensitivity of the generator to latent space. In addition to the above methods used to implement inverse gradient penalty for generators, some studies adopt orthogonal regularization to enforce amenability to truncation by conditioning G to be\n\t\tsmooth. Accordingly, the full space of $z$ will map to good output samples. Introducing orthogonality condition is a direct method:}\n\t\\begin{equation}\n\t\tR(W)=\\beta\\left\\|W^{\\top} W -I\\right\\|_{\\mathrm{F}}^{2},\n\t\\end{equation}\n\t{where W is a weight matrix and $\\beta$ is a hyperparameter. However, this regularization is too limiting . Therefore, a relaxed constraint has been designed by . Brock et al. apply Off-Diagonal Orthogonal Regularization (Off-Diagonal OR) to the generator directly enforcing the orthogonality condition:}\n\t\\begin{equation}\n\t\tR_o(W)=\\beta\\left\\|W^{\\top} W \\odot(\\mathbf{1}-I)\\right\\|_{\\mathrm{F}}^{2},\n\t\\end{equation}\n\t{where $\\mathbf{1}$ denotes a matrix with all elements set to 1. The Off-Diagonal OR makes G smooth so that the entire space of $z$ will map to good output samples. Orthogonality regularization is different from spectral normalization . Orthogonality regularization destroys the information about the spectrum by setting all the singular values to one, while spectral normalization only makes the maximum singular be one.\n\t}\n\t\\begin{table}\n\t\t\\caption{ FID results for different normalization methods on CIFAR-10 and CIFAR-100 datasets. ( The structure is the same as SNGAN except that the discriminator uses different normalization methods).\\tiny}\n\t\t\\label{table:normalization}\n\t\t\\centering\n\t\t\\begin{tabular}{ccc}\n\t\t\t\\toprule\n\t\t\t\\midrule\n\t\t\tmethods & CIFAR-10 & CIFAR-100 \\\\\n\t\t\t\\hline\n\t\t\tNone & 40.91 & 45.44 \\\\\n\t\t\tBN & 37.63 & 44.45\\\\\n\t\t\tLN & \\textbf{19.21} & 21.15 \\\\\n\t\t\tIN & 34.14 & 43.64\\\\\n\t\t\tGN & 19.31 & $\\textbf{20.80}$ \\\\\n\t\t\tWN & 24.28 & 29.96 \\\\\n\t\t\tSN&19.75&22.89\\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{table}","id":"ed3fad77-b260-419c-962b-ecc924abe2e8","level":"subsection","origin_cites_number":5,"parent_id":"bc57c604-a3ec-457f-ab6f-07cf5243c526","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Regularization and Normalization of \"Other methods\""],["subsection","Inverse Gradient Penalty"]],"subsections":[],"title":"Inverse Gradient Penalty"},{"cite_extract_rate":1,"cites":[56,164,8324,108,99,130,161,162,163,62,157,78,8317],"content":"{In this section, to provide a side view to the selection of regularization and normalization techniques, we investigate the applications of regularization and normalization techniques frequently employed in state-of-the-art and popular GANs. We select six methods (One per year from 2017-2022) categorized into two classes according to different tasks: Unconditional Generation and Conditional Generation. The selected methods and analysis are shown in Table \\ref{table:SOTA }. PGGAN is a popular GAN model in recent years, which grows the size of both the generator and discriminator progressively. PGGAN empowers high-resolution image generation. Since PGGAN was proposed in 2017, only some simple regularization techniques were applied: WGAN-GP , BN , and LN ; BigGAN is a popular conditional generative adversarial networks, which uses many regularization and normalization techniques, such as zc-GP, SN , Off-Diagonal OR , and CBN ; AutoGAN is the first study introducing the Neural architecture search (NAS) to GANs. It defines the search space for the generator architecture and adopts Inception score as the reward to discover the best architecture. The main focus of AutoGAN is architecture, so AutoGAN only comprises SN ; StyleGAN2 is the most popular architecture of GANs, which produces photorealistic images with large varieties and is widely used in image generation, such as Image Completion , Image-to-Image Translation . StyleGAN2-ADA proposes a novelty adaptive data augmentation methods. Combining StyleGAN2 and adaptive data augmentation, StyleGAN2-ADA obtains impressive performance in image generation, particularly in data-effficient generation. Furthermore, InsGen combines StyleGAN2-ADA with contrastive learning, acquiring state of the art on many generation tasks and datasets. Recently, StyleGAN-XL scales StyleGAN to large diverse datasets and sets a new state-of-the-art on large-scale image synthesis. In summary, many regularization and normalization techniques have been used in state-of-the-art GANs with zc-GP and SN being more attractive to researchers. Data augmentation is a striking method and orthogonal to other ongoing researches on training, architecture, and regularization. Therefore, popular augmentation strategies, such as ADA, have been employed as default operations GANs training. Furthermore, self-supervision has been used to further improve the performance of GANs, which is also orthogonal to other methods.}\n\t\\begin{table}\n\t\t\\caption{The applications of the Regularization and Normalization techniques used in SOTA GANs.}\n\t\t\\label{table:SOTA }\n\t\t\\centering\n\t\t\\footnotesize\n\t\t\\begin{tabular}{c | c | c |c | c | c|c }\t\n\t\t\t\\hline\n\t\t\t\\hline\n\t\t\t{ Method}& Task&\\makecell[c]{Gradient \\\\Penalty}&\\makecell[c]{Data augmentation\\\\ and preprocessing}&\\makecell{Self\\\\supervision}&\\makecell[c]{Weight\\\\ normalization }&\\makecell[c]{Layer\\\\ normalization}\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{PGGAN\\\\(2017)}&\\makecell{Unconditinal\\\\ Generation}&WGAN-GP&\\makecell{None}&None&\\makecell{None}&BN: G,LN: D\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{BigGAN\\\\ (2018 )}&\\makecell{Conditinal \\\\Generation}&zc-GP&None&None&\\makecell{SN: G, D}&CBN\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{AutoGAN\\\\ (2019 )}&\\makecell{Unconditinal\\\\ Generation}&None&None&None&\\makecell{SN:D}&None\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{StyleGAN2-ADA\\\\(2020)}&\\makecell{Unconditinal \\\\Generation}&zc-GP&\\makecell{Adaptive}&None&None&IN\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{InsGen\\\\(2021)}&\\makecell{Unconditinal \\\\Generation}&zc-GP&\\makecell{Adaptive}&contrastive&None&IN\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{StyleGAN-XL\\\\(2022)}&\\makecell{Conditinal \\\\Generation}&None&\\makecell{Translation\\\\Cutout}&None&SN:D&IN\\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{table}","id":"fb4135aa-ff8b-439d-9a28-a790ffddd379","level":"section","origin_cites_number":13,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Applications of Regularization and Normalization in SOTA GANs"]],"subsections":[],"title":"Applications of Regularization and Normalization in SOTA GANs"},{"cite_extract_rate":0,"cites":[],"content":"","id":"b359d174-4880-4149-b0f4-5fd0e7bd75f4","level":"section","origin_cites_number":0,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Summary and Outlook"]],"subsections":["9cf8af1d-75ca-445d-aa5a-fb3169fa5585","c7f682e9-c24a-41ee-8bfd-ac2766fa59c0"],"title":"Summary and Outlook"},{"cite_extract_rate":0.8888888888888881,"cites":[144,88,7001,99,8322,78,8317,6997],"content":"Recently, significant achievements of GANs have been made in generation tasks and the network has been widely used in many computer vision tasks, such as image inpainting, style transfer, text-to-image translations, and attribute editing. However, due to the overconfident assumptions, the training faces many challenges, such as non-convergence, mode collapse, gradient vanishing, and overfitting. To mitigate these problems, many solutions focus on designing new architectures, new loss functions, new optimization methods, and regularization and normalization techniques. \n\tIn this paper, we study GANs training from three perspectives and propose a new taxonomy, denoted as \\textbf{\"Training dynamics\"}, \\textbf{\"Fitting distribution\"}, \\textbf{\"Real \\& Fake\"}, and \\textbf{\"Other methods\"}, to survey the different regularization and normalization techniques during GANs training. Our study provides a systematic and comprehensive analysis of the reviewed methods to serve researchers of the community. In addition, we also demonstrate the motivation and objectives of different methods and compare the performance of some popular methods in a fair manner quantitatively, which has implications for future research in selecting their research topics or developing their approaches.\n\t\\begin{comment}\n\tThis paper summarizes the regularization and normalization of generative adversarial networks and explains the problem from three aspects: \n\t\\textbf{Optimal transport and Lipschitz continuity:} First, optimal transport and optimal transport with the regular term are introduced. According to the duality form, WGAN-GP and WGAN-LP are proposed to make the discriminator satisfy 1-Lipschitz continuity and k-Lipschitz continuity ($k\\leq1$), respectively. Next, many gradient penalty methods have been analyzed. These methods achieved certain improvements as compared with WGAN-GP in two prespecitves: \\textbf{(1) To limit the Lipschitz constant to a small value}, such as 0 ; \\textbf{(2) To find the right restricted space}. The restricted space of WGAN-GP and WGAN-LP is the interpolation between the real images and fake images. Some works restricted it at the real images or fake images . The relationship between the Lipschitz constant and the spectral norm of the matrix is also derived in this paper. It indicates that the normalization of the spectral norm can also be used to achieve 1-Lipschitz continuity. Furthermore, the Frobenius norm is proved to be the upper bound of the spectral norm. The results of WGAN-LP can be obtained by normalizing the Frobenius norm. Moreover, the work analyzed the relationship between the mode collapse of the GAN and the eigenvalue distribution of the discriminator weight matrix, ending up in avoiding the mode collapse by limiting the eigenvalue collapse. However, there is a lack of the specific reason or sound proof.\n\t\\textbf{ Training dynamics:} GANs is a two-player zero-sum game. Due to the nonlinearity of the neural network, it is difficult to find the global convergence solution. In terms of the local convergence, the Jacobian regularization needs to be added. Few research studies are found in this area.\n\t\\textbf{ Representation:} As mentioned earlier, training of GANs is a semi-supervised task. More prior information is needed to improve the presentative ability of the network. Conditional layer normalization is used to better encode condition information and reduce the difficulty of conditional generation. Consistent regularization, data augmentation and self-supervision use unsupervised methods to improve the supervision information in GANs training, and use additional labels and tasks to improve GANs performance. We suggest that this area is promising and more studies can be conducted in this specific direction.\n\t\\end{comment}","id":"9cf8af1d-75ca-445d-aa5a-fb3169fa5585","level":"subsection","origin_cites_number":9,"parent_id":"b359d174-4880-4149-b0f4-5fd0e7bd75f4","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Summary and Outlook"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":1,"cites":[165],"content":"By reviewing the regularization and normalization of GANs, the following questions and thoughts are proposed based on different perspectives of GANs training:\n\t\\begin{itemize}\n\t\t\\item [1)] \n\t\t\\textit{What is a good distance metric, and which divergence should be used in GANs training?} The priority in the training process of GANs is to find a suitable divergence to measure the distance between the generated distribution and the true distribution. Wasserstein divergence is important for the training of GANs. However, it is uncertain whether the next proposed divergence performs better.\n\t\t\\item [2)]\n\t\t\\textit{What is the main difference between real images and generated images?} During the training of unconstrained and unprioritized GANs, if we can quantitatively represent the difference between real images and generated images from different perspectives, the efficient regularization methods can be designed based on this.\n\t\t\\item [3)]\n\t\t\\textit{How to avoid real images forgetting\\footnote{Real images forgetting is caused by not introducing real images while training the generator, which is different from discriminator forgetting.}?} As acknowledged, real images do not directly participate in the training of the generator, thus the discriminator needs to remember the characteristics of the real images to optimize the generator indirectly. We call this the real images forgetting. We conjecture that real images forgetting may exist, and which may increase the difficulty of GANs training. Some works might serve as basis to prove this hypothesis and propose effective solutions.\n\t\t\\item [4)]\n\t\tRecent studies show that discriminator suffers from overfitting and discriminator forgetting. It is a common problem of neural networks, which is caused by the shortcut of the loss driven method. Some new methods, such as contrastive learning, representation learining, can be proposed to improve the generalization of the discriminator.\n\t\t\\item [5)]\n\t\tRecently, diffusion model acquires the impressive performance in image generation. One possible reason for success is the phased training strategy in diffusion model . Inspired by this, some strategies to reduce the difficulty of GANs training may be proposed.\n\t\\end{itemize}\n\t\\section*{Acknowledgment}\n\tThe work is partially supported by the National Natural Science Foundation of China under Grand No.U19B2044, No.61836011 and No.91746209. We are very grateful to the help of Jianlin Su, whose blog is \\url{https://spaces.ac.cn/tag/GAN/}.\n\t\\bibliographystyle{ACM-Reference-Format}\n\t\\bibliography{bare_jrnl_transmag}\n\t\\appendix\n\t\\clearpage","id":"c7f682e9-c24a-41ee-8bfd-ac2766fa59c0","level":"subsection","origin_cites_number":1,"parent_id":"b359d174-4880-4149-b0f4-5fd0e7bd75f4","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Summary and Outlook"],["subsection","Outlook"]],"subsections":[],"title":"Outlook"},{"cite_extract_rate":0,"cites":[],"content":"","id":"f6a7c46a-3c0c-4605-a602-4104833c8854","level":"section","origin_cites_number":0,"parent_id":"4410407c-7d7c-48f1-8174-a9a95319844f","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Supplementary Online-only Material"]],"subsections":["9df0baaf-f220-4a8d-b36a-27ebd583cf6c","375c251d-5d04-454e-83a6-2140b14015f5","d07e4dc9-6fdd-4534-85ee-925acd678910"],"title":"Supplementary Online-only Material"},{"cite_extract_rate":0.42857142857142805,"cites":[88,95,8318],"content":"\\label{sect:A-1}\n\tOptimal transport was proposed in the 18th century to minimize the transportation cost while preserving the measure quantities. Given the space with probability measures $(X,\\mu)$ and $(Y,\\upsilon)$, if there is a map $T:X\\rightarrow Y$ which is measure-preserving, then for any $B\\subset Y$, having:\n\t\\begin{equation}\\label{eqn1}\n\t\t\\int_{T^{-1}(B)}\\mathrm{d}\\mu(x)=\\int_B \\mathrm{d}\\upsilon(y).\n\t\\end{equation}\n\tWriting the measure-preserving map as $T_*(\\mu)=\\upsilon$. For any $x\\in X$ and $y\\in Y$, the transportation distance is defined as $c(x,y)$, the total transportation cost is given by:\n\t\\begin{equation}\\label{eqn1}\n\t\tC(T):=\\int_X c(x,T(x)) \\mathrm{d}\\mu(x).\n\t\\end{equation}\n\tIn the 18th century, Monge et al. proposed the Optimal Mass Transportation Map that corresponds to the smallest total transportation cost: $C(T)$. The transportation cost corresponding to the optimal transportation map is called the Wasserstein distance between probability measures $\\mu$ and $\\upsilon$:\n\t\\begin{equation}\\label{eqn3}\n\t\tW_c(\\mu,\\upsilon)=\\mathop{\\min}\\limits_{T}\\left\\{\\int_X c(x,T(x)) \\mathrm{d}\\mu(x)\\ |\\ T_*(\\mu)=\\upsilon\\right\\}.\n\t\\end{equation}\n\tIn 1940s, Kantorovich proved the existence and uniqueness of the solution for Monge problem, and according to the duality of linear programming, the Kantorovich-Rubinstein (KR) duality of Wasserstein distance is given by:\n\t\\begin{equation}\\label{eqn1}\n\t\tW_c(\\mu,\\upsilon)\n\t\t=\\mathop{\\max}\\limits_{\\varphi ,\\psi}\\left\\{\\int_X \\varphi \\mathrm{d}\\mu\\ +\\int_Y \\psi \\mathrm{d}\\upsilon\\ |\\ \\varphi(x)+\\psi(y)\\leq c(x,y)\\right\\}.\n\t\\end{equation}\n\tThis dual problem is constrained, defining the c-transform: $\\psi(y)=\\varphi^c(y):=inf_x\\{c(x,y)-\\varphi(x)\\}$, and the Wasserstein distance becomes:\n\t\\begin{equation}\\label{eqn1}\n\t\tW_c(\\mu,\\upsilon)=\\mathop{\\max}\\limits_{\\varphi}\\left\\{\\int_X \\varphi \\mathrm{d}\\mu\\ +\\int_Y \\varphi^c \\mathrm{d}\\upsilon\\right\\},\n\t\\end{equation}\n\twhere $\\varphi$ is called the Kantorovich potential. It can be shown that if $c(x,y)=|x-y|$ and Kantorovich potential satisfies the 1-Lipschitz continuity, then $\\varphi^c=-\\varphi$. Kantorovich potential can be fitted by a deep neural network, which is recorded as $\\varphi_\\xi$. Wasserstein distance is:\n\t\\begin{equation}\n\t\tW_c(\\mu,\\upsilon)=\\mathop{\\max}\\limits_{||\\varphi_\\xi||_L\\leq 1}\\left\\{\\int_X \\varphi_\\xi \\mathrm{d}\\mu\\ -\\int_Y \\varphi_\\xi \\mathrm{d}\\upsilon\\right\\}.\n\t\t\\label{eqn7}\n\t\\end{equation}\n\tIf $X$ is the generated image space, $Y$ is the real sample space, $Z$ is latent space and $g_\\theta$ is the geneartor, the Wasserstein GANs (WGAN) is formulated as a min-max problem:\n\t\\begin{equation}\\label{eqn1}\n\t\t\\mathop{\\min}\\limits_{\\theta}\\mathop{\\max}\\limits_{||\\varphi_\\xi||_L\\leq 1}\\left\\{\\int_Z \\varphi_\\xi(g_\\theta(z)) \\mathrm{d}z \\ -\\int_Y \\varphi_\\xi(y) \\mathrm{d}y\\right\\}.\n\t\\end{equation}\n\tIn the optimization process, the generator and the Kantorovich potential function (discriminator) are independent of each other, optimized in a step-by-step iteration.\n\tIf $c(x,y)=\\frac{|x-y|^2}{2}$, there is a convex function $u$ that is called Brenier potential . The optimal transportation map is given by the gradient map of Brenier potential: $T(x)=\\nabla u(x)$. There exists a relationship between Kantorovich potential and Brenier potential : \n\t\\begin{equation}\\label{eqn1}\n\t\tu(x)=\\frac{|x|^2}{2}-\\varphi(x).\n\t\\end{equation}\n\tFrom the previous discussion, it is evident that the optimal transportation map (Brenier potential) corresponds to the generator, and Kantorovich potential corresponds to the discriminator. After the discriminator is optimized, the generator is directly drivable without the optimization process .\n\tThe transportation cost of Eq (\\ref{eqn3}) is defined as the form of two distribution distances:\n\t\\begin{equation}\\label{eqn10}\n\t\tOT(P||Q)=\\mathop{inf}\\limits_{\\pi}\\int\\pi(x,y)c(x,y)\\mathrm{d}x\\mathrm{d}y,\n\t\\end{equation}\n\twhere $\\pi(x,y)$ is the joint distribution, satisfying $\\int_y\\pi(x,y)dy=P(x)$ and $\\int_x\\pi(x,y)dx=Q(y)$. The dual form of Eq (\\ref{eqn10}) is derived as follows::\n\t\\begin{equation}\\label{eqn1}\n\t\tOT(P||Q)=\\mathop{\\max}\\limits_{\\varphi ,\\psi}\\{\\int_x \\varphi(x)P(x) \\mathrm{d}x\\ \\\\ +\\int_y\\psi(y)Q(y) \\mathrm{d}y\\ |\\ \\varphi(x)+\\psi(y)\\leq c(x,y)\\}.\n\t\\end{equation} \n\tConsidering the optimal transportation with regular terms, Peyr{\\'e} et al. added the entropic regularization for optimal transportation that transforms the dual problem into a smooth unconstrained convex problem. The regularized optimal transport is defined as:\n\t\\begin{equation}\\label{eqn12}\n\t\tOT_c(P||Q)=\\mathop{\\min}\\limits_{\\pi}\\int\\pi(x,y)c(x,y)\\mathrm{d}x\\mathrm{d}y+\\epsilon E(\\pi).\n\t\\end{equation}\n\tIf $E(\\pi)=\\int_x\\int_y\\pi(x,y)\\log(\\frac{\\pi(x,y)}{P(x)Q(y)})\\mathrm{d}x\\mathrm{d}y$, Eq (\\ref{eqn12}) can be written as:\n\t\\begin{equation}\\label{eqn13}\n\t\t\\begin{split}\n\t\t\tOT_c(P||Q)=&\\mathop{\\min}\\limits_{\\pi}\\int\\pi(x,y)c(x,y)\\mathrm{d}x\\mathrm{d}y+\\epsilon \\int_x\\int_y\\pi(x,y)\\log\\left(\\frac{\\pi(x,y)}{P(x)Q(y)}\\right)\\mathrm{d}x\\mathrm{d}y\\\\\n\t\t\ts.t. \\int_y\\pi(x,y)&\\mathrm{d}y=P(x),\\int_x\\pi(x,y)\\mathrm{d}x=Q(y).\n\t\t\\end{split}\n\t\\end{equation}\n\tThe dual form of Eq (\\ref{eqn13}) becomes:\n\t\\begin{equation}\n\t\t\\begin{split}\n\t\t\tOT_c(P||Q)\n\t\t\t&=\\mathop{\\max}\\limits_{\\varphi ,\\psi}\\int_x \\varphi(x)P(x) \\mathrm{d}x\\ +\\int_y\\psi(y)Q(y)\\mathrm{d}y\\\\\n\t\t\t&+\\frac{\\epsilon}{e}\\int_x\\int_y\\exp\\left(\\frac{-\\left(c(x,y)+\\varphi(x)+\\psi(y)\\right)}{\\epsilon}\\right)\\mathrm{d}x\\mathrm{d}y.\n\t\t\\end{split}\n\t\t\\label{EQ:eqn14}\n\t\\end{equation}\n\tPetzka et al. set $c(x,y)=||x-y||_2$ in Eq (\\ref{EQ:eqn14}), and the dual form of optimal transport with the regular term can be expressed as:\n\t\\begin{equation}\n\t\t\\mathop{\\sup}\\limits_{\\varphi,\\psi}\\{\\mathbb{E}_{x\\sim{p(x)}}[\\varphi(x)]-\\mathbb{E}_{y\\sim q(y)}[\\psi(y)]\n\t\t-\\frac{4}{\\epsilon}\\int\\int\\mathop{\\max}\\{0,(\\varphi(x)-\\psi(y)-||x-y||_2)\\}^2\\mathrm{d}p(x)\\mathrm{d}q(y) \\}.\n\t\t\\label{EQ:eqn32}\n\t\\end{equation}\n\tSimilar to dealing with a single function, one can replace $\\varphi = \\psi$ in Eq (\\ref{EQ:eqn32}), which leads to the objective of minimum:\n\t\\begin{equation}\\label{eqn1}\n\t\t\\mathbb{E}_{y\\sim q(y)}[\\varphi(y)]-\\mathbb{E}_{x\\sim p(x)}[\\varphi(x)]\n\t\t+\\frac{4}{\\epsilon}\\int\\int\\mathop{\\max}\\{0,(\\varphi(x)-\\varphi(y)-||x-y||_2)\\}^2\\mathrm{d}p(x)\\mathrm{d}q(y).\n\t\\end{equation}","id":"9df0baaf-f220-4a8d-b36a-27ebd583cf6c","level":"subsection","origin_cites_number":7,"parent_id":"f6a7c46a-3c0c-4605-a602-4104833c8854","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Supplementary Online-only Material"],["subsection","Optimal Transport and Lipschitz Continuity"]],"subsections":[],"title":"Optimal Transport and Lipschitz Continuity"},{"cite_extract_rate":0.5,"cites":[7190],"content":"\\label{sect:A-2}\n\t\\newenvironment{lemma1}{{\\indent\\it \\textbf{Proposition 2.1:}}}{\\hfill \\par}\n\t\\begin{lemma1}\n\t\t\\textit{For zero-sum games, $v^{'}$is negative semi-definite for any local Nash-equilibrium. Conversely, if $v(\\bar{x})=0$ and $v^{'}$is negative definite, then $\\bar{x}$ is a local Nash-equilibrium.\n\t\t}\n\t\\end{lemma1}\n\t\\newenvironment{proof1}{{\\indent\\it Proof 2.1:}}{\\hfill $\\square$\\par}\n\t\\begin{proof1}\n\t\tRefer to \n\t\\end{proof1}\n\t\\textit{Proposition 2.1} gives the conditions for the local convergence of GANs, which is converted into the negative semi-definite problem of the Jacobian matrix. Negative semi-definite of the Jacobian matrix corresponds to its eigenvalue less than or equal to 0. If the eigenvalue of the Jacobian matrix at a certain point is a negative real number, the training process can converge; but if the eigenvalue is complex and the real part of the eigenvalue is small and the imaginary part is relatively large, the training process is difficult to converge unless the learning rate is very small. \n\t\\newenvironment{lemma2}{{\\indent\\it \\textbf{Proposition 2.2:}}}{\\hfill \\par}\n\t\\begin{lemma2}\n\t\t\\textit{\n\t\t\tLet $F:\\Omega\\rightarrow\\Omega$ be a continuously differentiable function on an open subset $\\Omega$ of $R^n$ and let $\\bar{x}\\in\\Omega$ be so that: 1. $F(\\bar{x})=\\bar{x}$ and 2. the absolute values of the eigenvalues of the Jacobian $F^{'}(x)$ are all smaller than 1.}\n\t\t\\textit{There is an open neighborhood $U$ of $\\bar{x}$ so that for all $x_0\\in U$, the iterates $F^{(k)}(x_0)$ converge to $\\bar{x}$. The rate of convergence is at least linear. More precisely, the error $||F^{(k)}(x_0)-\\bar{x}||$ is in $\\mathcal{O}(|\\lambda_{max}|^k)$ for $k\\rightarrow\\infty$ where $\\lambda_{max}$ is the eigenvalue of $F^{'}(\\bar{x})$ with the largest absolute value.}\n\t\\end{lemma2}\n\t\\newenvironment{proof2}{{\\indent\\it Proof 2.2:}}{\\hfill $\\square$\\par}\n\t\\begin{proof2}\n\t\tRefer to Section 3 in and Proposition 4.4.1 in .\n\t\\end{proof2}","id":"375c251d-5d04-454e-83a6-2140b14015f5","level":"subsection","origin_cites_number":2,"parent_id":"f6a7c46a-3c0c-4605-a602-4104833c8854","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Supplementary Online-only Material"],["subsection","Some Propositions of Training Dynamic in GANs"]],"subsections":[],"title":"Some Propositions of Training Dynamic in GANs"},{"cite_extract_rate":1,"cites":[97,96],"content":"\\label{sect:A-3}\n\t1-Lipschitz continuity is represented as:\n\t\\begin{equation}\n\t\t||D(x_1)-D(x_2)||\\leq ||x_1-x_2||.\\footnote{Lipschitz continuity can be defined by any form of norm.}\n\t\\end{equation}\n\tGenerally, considering the K-Lipschitz for a neural network $f(x)$:\n\t\\begin{equation}\n\t\tf(x)=g_N\\circ\\cdots g_2\\circ g_1(x),\\footnote{$\\circ$ is the symbol for function cascade. Specifically, $h\\circ g(x)=h(g(x))$. This definition of neural network is not general, such as DenseNet and ResNet , which can not be defined like this. Therefore, we do not strictly derive the relationship between the matrix norm and Lipschitz continuity.}\n\t\\end{equation}\n\twhere $g_i(x)=\\sigma (W_i x+b_i)$. And K-Lipschitz continuity for $f(x)$ is:\n\t\\begin{equation}\n\t\t||f(x_1)-f(x_2)||\\leq \\mathrm{K}||x_1-x_2||,\n\t\t\\label{EQ:eqn17}\n\t\\end{equation}\n\twhere K is Lipschitz constant of the function $f$. Due to the consistency of Lipschitz $||h\\circ g||_{Lip}\\leq ||h||_{Lip}\\cdot||g||_{Lip}$, $g_i$ needs to satisfy the C-Lipschitz continuity ($\\mathrm{C}=\\sqrt[N]{\\mathrm{K}}$) so that $f$ satisfies the K-Lipschitz continuity:\n\t\\begin{equation}\n\t\t||g_i(x_1)-g_i(x_2)||\\leq \\mathrm{C}||x_1-x_2||,\n\t\\end{equation}\n\t\\begin{equation}\n\t\t||\\sigma(Wx_1+b)-\\sigma(Wx_2+b)||\\leq \\mathrm{C}||x_1-x_2||.\n\t\t\\label{eq:23}\n\t\\end{equation}\n\tWhen $x_1\\rightarrow x_2$, the Taylor expansion of Eq (\\ref{eq:23}):\n\t\\begin{equation}\n\t\t||\\frac{\\partial\\sigma}{\\partial x} W(x_1-x_2)||\\leq \\mathrm{C}||x_1-x_2||.\n\t\\end{equation}\n\tNormally, $\\sigma$ is a function with limited derivatives such as Sigmoid, so the $\\mathrm{C'}$-Lipschitz continuity is be written as:\n\t\\begin{equation}\n\t\t|| W(x_1-x_2)||\\leq \\mathrm{C'}||x_1-x_2||,\n\t\\end{equation}\n\twhere $\\mathrm{C'}$ is a limited constant, which is determined by $\\frac{\\partial\\sigma}{\\partial x}$ and $\\mathrm{C}$.\n\tSimilarly, the spectral norm of matrix is defined by:\n\t\\begin{equation}\n\t\t||W||_2=\\mathop{\\max}\\limits_{x\\not=0}\\frac{||Wx||}{||x||}.\n\t\\end{equation}\n\tIn this context, the spectral norm $||W||_2$ can be used to represent the Lipschitz constant $\\mathrm{C'}$. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately.\t\n\\end{document}\n\\endinput","id":"d07e4dc9-6fdd-4534-85ee-925acd678910","level":"subsection","origin_cites_number":2,"parent_id":"f6a7c46a-3c0c-4605-a602-4104833c8854","prefix_titles":[["title","A Systematic Survey of Regularization and Normalization in GANs"],["section","Supplementary Online-only Material"],["subsection","Spectral Norm and the Lipschitz Constant"]],"subsections":[],"title":"Spectral Norm and the Lipschitz Constant"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ]\n ],\n \"subsections\": [\n \"29469c09-cd5f-4556-9382-b0a3b4de184e\",\n \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"8469161e-625b-422b-90e0-a6cca9638a04\",\n \"fdfb947c-602c-4735-9040-96d30810700e\",\n \"44c32216-8be6-4e5c-aa2f-ca241ea4139f\",\n \"bc57c604-a3ec-457f-ab6f-07cf5243c526\",\n \"fb4135aa-ff8b-439d-9a28-a790ffddd379\",\n \"b359d174-4880-4149-b0f4-5fd0e7bd75f4\",\n \"f6a7c46a-3c0c-4605-a602-4104833c8854\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.7213114754098361,\n \"cites\": [\n 63,\n 89,\n 73,\n 86,\n 68,\n 85,\n 78,\n 76,\n 8317,\n 53,\n 60,\n 6998,\n 66,\n 57,\n 70,\n 88,\n 80,\n 6996,\n 72,\n 54,\n 58,\n 62,\n 59,\n 7190,\n 55,\n 69,\n 77,\n 6997,\n 82,\n 74,\n 56,\n 81,\n 87,\n 6999,\n 8316,\n 84,\n 71,\n 61,\n 75,\n 67,\n 65,\n 83,\n 64,\n 79\n ],\n \"content\": \"\\\\setlength{\\\\abovecaptionskip}{0.cm}\\n\\tGenerative adversarial networks (GANs) have been widely used in computer vision, such as image inpainting , style transfer , text-to-image translations , and attribute editing . GANs training is a two-player zero-sum game between a generator and a discriminator, which can be understood from different perspectives: (i) \\\"Real \\\\& Fake\\\" , (ii) \\\"Fitting distribution\\\" , and (iii) \\\"Training dynamics\\\" . GANs training suffers from several issues, for instance: non-convergence , mode collapses , gradient vanishing , overfitting , discriminator forgetting and deficiency , and hyperparameters sensitivity . Many solutions to mitigate these issues have been proposed, focusing on designing new architectures , loss functions , optimization methods , regularization , and normalization . Among them, regularization and normalization techniques are compatible with loss functions, model structures, and tasks, which has attracted the attention of scholars. \\n\\tRegularization and normalization are widely applied in neural networks training to introduce prior knowledge. For supervised tasks, regularization in literature has been proposed to introduce some advantages like overfitting prevention , semi-supervised assumptions , manifold assumptions , feature selection , and low rank representation . On the other hand, normalization is advantageous for the Stochastic Gradient Descent (SGD) , accelerating convergence and improving accuracy. Unlike the icing on the cake of supervisory tasks, regularization and normalization are utilized inevitably in weak-supervised and unsupervised tasks. GANs' training is a two-player zero-sum game having a solution to Nash equilibrium. The proposal of standard GAN is based upon the non-parametric assumption of the infinite capacity of networks, an unsupervised learning task. Likewise, a good number of research studies targeting GANs training from different perspectives argue that unconstrained training causes unstable training (generator and discriminator ) and significant bias between real images and fake images (attributes domain and frequency domain ). Therefore, a large amount of prior should be introduced into GANs training through regularization and normalization\\n\\t\\\\begin{figure}\\n\\t\\t\\\\centering\\n\\t\\t\\\\includegraphics[scale=.52]{figs/summary_new.pdf}\\n\\t\\t\\\\caption{The overview of different perspectives of GANs training and the summary of the regularization and normalization for GANs.}\\n\\t\\t\\\\label{fig:overview}\\n\\t\\\\end{figure}\\n\\tRegularization and normalization are effectively used to stabilize training and improve the performance of GANs in existing literature . Due to diverse nature of the topic, there is a need for systematic literature survey. There exist some literature studies , however, these studies either lack comprehensive coverage of the topic, miss detailed background information and theoretical analysis, or do not correlate different methods. In this paper, based on the different perspectives of GANs training, we propose a new taxonomy, denoted as \\\\textbf{\\\"Real \\\\& Fake\\\"}, \\\\textbf{\\\"Fitting distribution\\\"}, \\\\textbf{\\\"Training dynamics\\\"}, and \\\\textbf{\\\"Other methods\\\"}, for a better understanding of regularization and normalization during the GANs training as depicted in Figure \\\\ref{fig:overview}. {\\\\textbf{\\\"Real \\\\& Fake\\\"} is the low-level (intuitive) perspective of GANs, in which GANs is considered as \\\"counterfeiters-police\\\" competition. At this level, D estimates the real probability of both real and fake samples, which is similar to the bi-classification task. Therefore, prior information and additional supervision tasks in classification task are also urgent during the training process of the discriminator. Based on these, some regularization methods, such as {\\\\textit{Data Augmentation and Preprocessing}}}\\\\footnote{Data augmentation and preprocess introduce additional data and prior, which is similar to regularization. More importantly, both consistency regularization and self-supervision need different data transformation operations. Hence, this paper also discusses some works on this.}, {\\\\textit{Consistency Regularization}, and {\\\\textit{Self-Supervision}} are proposed to mitigate overfitting, improve the representation of discriminator, and avoid discriminator forgetting by introducing additional supervised information and data; \\\\textbf{\\\"Fitting distribution\\\"} is the middle-level perspective of GANs. At this level, generator is considered as a distribution mapping function, and the discriminator is a distribution divergence. Among various distances, Wasserstein distance is a popular form, and Lipschitz continuity is a necessary condition for achieving Wasserstein distance. Based on these, \\\\textit{Gradient Penalty}, \\\\textit{Weight Normalization}, \\\\textit{Weight Regularization}, and \\\\textit{Gradient} \\\\textit{Normalization} are used to fulfill Lipschitz continuity and ensure training stability of discriminator; \\\\textbf{\\\"Training Dynamics\\\"} is the high-level (essential) perspective of GANs. At this level, GANs training is a two-player zero-sum game with a solution to Nash Equilibrium. To achieve theoretical local convergence, \\\\textit{Jacobian Regularization} needs to be used; Finally, \\\\textbf{\\\"Other methods\\\"} containing \\\\textit{Layer} \\\\textit{ Normalization} and \\\\textit{Inverse Gradient Penalty} are used for conditional generation and easing mode collapse, respectively}\\n\\tIn summary, we make the following contributions in this survey:\\n\\t\\\\begin{itemize}\\n\\t\\t\\\\item \\\\textit{Comprehensive analysis of GANs training.} In this study, we analyze the GANs training from three perspectives including \\\"Real \\\\& Fake\\\", \\\"Fitting distribution\\\", and \\\"Training dynamics\\\". To the best of our knowledge, this survey is the first in this domain with comprehensive analysis.\\n\\t\\t\\\\item \\\\textit{New taxonomy.} Based on the analysis of GANs training from different perspectives, we propose a novel taxonomy and contextualize the regularization and normalization of GANs comprehensively.\\n\\t\\t\\\\item \\\\textit{ Comparison and analysis.} Following the taxonomy, we also provide quantitative and qualitative analysis and comparison for each type of regularization and normalization techniques, which has helped the researchers and practitioners navigate this space.\\n\\t\\\\end{itemize}\\n\\t{\\\\textbf{The Scope of This Survey.} This survey aims to systematically analyze the prevalent problems in the GANs training, such as non-convergence, mode collapses, gradient vanishing, and discriminator overfitting. Accordingly, different regularization and normalization technologies have been summarized. Of course, not all regularization and normalization technologies for GANs are covered in this survey. In some case, some regularization and normalization technologies are highly dependent on the task and data in the hand, we recommend looking for domain-specific regularization and normalization techniques from the following reviews: data-efficient generation , medical image generation, image super-resolution , biomedical informatics generation , Spatio- temporal data generation , text generation . Our survey is concerned with general technologies in GANs, which are not dependent on model structures, data, and task. We hope our study can provide general and universal insights of GANs for the community.}\\n\\tThe rest of this paper is organized as follows: Section 2 introduces the background and different training perspectives of GANs. Section 3, 4, 5, and 6\\n\\tdescribe regularization and normalization methods in different groups, respectively. Furthermore, we investigate the applications of regularization and normalization techniques that have been frequently employed in SOTA GANs in Section 7 and discuss the current problems and prospects for future work in Section 8.\",\n \"id\": \"29469c09-cd5f-4556-9382-b0a3b4de184e\",\n \"level\": \"section\",\n \"origin_cites_number\": 61,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ]\n ],\n \"subsections\": [\n \"5a1f5e65-9650-482a-9b5a-2625a5dd0b96\",\n \"c2fed4a0-e8f0-4ddc-ab66-70cc7bfe76f8\",\n \"fd688225-4a59-4600-886f-76309f6e8ba9\",\n \"4e9cd8aa-b10a-44ef-a7b7-7d2880558ce2\",\n \"a915370f-85fe-4345-a91b-c8d67718061c\",\n \"a397f8d2-2658-471a-8c3b-cbde88271e09\",\n \"0edf29a6-bba8-43a2-a318-3f272fcb91d9\"\n ],\n \"title\": \"Background and Three Perspectives of GANs Training\"\n },\n {\n \"cite_extract_rate\": 0.27272727272727204,\n \"cites\": [\n 91,\n 66,\n 57,\n 90,\n 71,\n 67\n ],\n \"content\": \"{Regularization and normalization are common and important techniques to introduce prior knowledge in neural networks.} Regularization is a technique to control the complexity of learning models. Weight decay is a typical method to minimize the square of weights together with the training loss in the training of neural networks , which can be used to improve generalization. \\n\\tIn Bayesian learning methods, such as relevance vector machine , probabilistic classification vector machines , and others , regularization is termed as prior distribution. Specifically, L2 regularization is equivalent to introducing Gaussian prior to the parameters, and L1 regularization is equivalent to introducing Laplace prior to the parameters. The theoretical connection between regularization and prior information has been investigated in neural network ensembles research . Regularization does not only control overfitting but also provide other characteristics like semi-supervised assumptions , manifold assumptions , feature selection , low rank representation , and consistency assumptions . {Normalization is the mapping of data to a specified range, which is advantageous for the Stochastic Gradient Descent (SGD) , accelerating convergence and improving accuracy.}\",\n \"id\": \"5a1f5e65-9650-482a-9b5a-2625a5dd0b96\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ],\n [\n \"subsection\",\n \"Regularization and Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Regularization and Normalization\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 64,\n 92,\n 54\n ],\n \"content\": \"GANs are two-player zero-sum games, where generator (G) and discriminator ($D$) try to optimize opposing loss functions to find the global Nash equilibrium. In general, GANs can be formulated as follows:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathop{\\\\min}\\\\limits_{\\\\phi}\\\\mathop{\\\\max}\\\\limits_{\\\\theta} f(\\\\phi,\\\\theta)\\n\\t\\t=\\\\mathop{\\\\min}\\\\limits_{\\\\phi}\\\\mathop{\\\\max}\\\\limits_{\\\\theta}\\\\mathbb{E}_{x\\\\sim p_{r}}[g_1(D_\\\\theta(x))]\\n\\t\\t+\\\\mathbb{E}_{z\\\\sim p_z}[g_2(D_\\\\theta(G_\\\\phi(z)))],\\n\\t\\t\\\\label{EQ:eqn1}\\n\\t\\\\end{equation}\\n\\twhere $\\\\phi$ and $\\\\theta$ are parameters of the generator $G$ and the discriminator $D$, respectively. $p_r$ and $p_z$ represent the real distribution and the latent distribution, respectively. {$g_1$and $g_2$ are different functions corresponding to different GANs. Specifically, vanilla GAN can be described as $g_1(t)=g_2(-t)=-\\\\log(1+e^{-t})$; $f$-GAN can be written as $g_1(t)=-e^{-t}$, $g_2(t)=1-t$; Morever, Geometric GAN and WGAN are described as $g_1(t)=g_2(-t)=-\\\\mathop{\\\\max}(0,1-t)$ and $g_1(t)$\\\\\\\\$=g_2(-t)=t$, respectively.} \\n\\t{Different from supervised learning, GANs training is an unsupervised learning, which leads to the urgency of regularization and normalization in the training of GANs.} In the following parts, we elaborate the training of GANs from three perspectives: low level: the perspective of \\\"Real \\\\& Fake\\\", middle level: the perspective of \\\"Fitting distribution\\\", and high level: the perspective of \\\"Training dynamics\\\". According to different perspectives of GANs, various regularization and normalization have been proposed in the GANs training.\\n\\t\\\\begin{comment}\",\n \"id\": \"c2fed4a0-e8f0-4ddc-ab66-70cc7bfe76f8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ],\n [\n \"subsection\",\n \"GANs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GANs\"\n },\n {\n \"cite_extract_rate\": 0.27272727272727204,\n \"cites\": [\n 91,\n 66,\n 57,\n 90,\n 71,\n 67\n ],\n \"content\": \"Regularization is a technique to control the complexity of learning models. Weight decay is a typical method to minimize the square of weights together with the training loss in the training of neural networks , which can be used to improve generalization. \\n\\tIn Bayesian learning methods, such as relevance vector machine , probabilistic classification vector machines , and others , regularization is termed as prior distribution. Specifically, L2 regularization is equivalent to introducing Gaussian prior to the parameters, and L1 regularization is equivalent to introducing Laplace prior to the parameters. The theoretical connection between regularization and prior information has been investigated in neural network ensembles research . Regularization does not only control overfitting but also provide other characteristics like semi-supervised assumptions , manifold assumptions , feature selection , low rank representation , and consistency assumptions . {Normalization is the mapping of data to a specified range, which is advantageous for the Stochastic Gradient Descent (SGD) , accelerating convergence and improving accuracy.\\n\\t}\\n\\t\\\\end{comment}\",\n \"id\": \"fd688225-4a59-4600-886f-76309f6e8ba9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ],\n [\n \"subsection\",\n \"Regularization and Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Regularization and Normalization\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 63,\n 92\n ],\n \"content\": \"In low level, GANs is considered as \\\"counterfeiters-police\\\" competition, where the generator (G) can be thought of counterfeiters, trying to produce fake currency and use it undetected, while the discriminator (D) is analogous to the police, trying to detect the counterfeit currency. This competition drives both teams to upgrade their methods until the fake currency is indistinguishable from the real ones. Generally, D estimates the real probability of both real and fake samples, which is very similar to the bi-classification task, while G generates fake samples similar to real ones. Hence, the loss function in Eq (\\\\ref{EQ:eqn1}) is formulated as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathop{\\\\min}\\\\limits_{\\\\phi}\\\\mathop{\\\\max}\\\\limits_{\\\\theta} f(\\\\phi,\\\\theta)\\n\\t\\t=\\\\mathop{\\\\min}\\\\limits_{\\\\phi}\\\\mathop{\\\\max}\\\\limits_{\\\\theta}\\\\mathbb{E}_{x\\\\sim p_{r}}[\\\\log(D_\\\\theta(x))]\\n\\t\\t+\\\\mathbb{E}_{z\\\\sim p_z}[\\\\log(1-D_\\\\theta(G_\\\\phi(z)))],\\n\\t\\t\\\\label{EQ:eqnganlevel1}\\n\\t\\\\end{equation}\\n\\twhere $f(\\\\phi,\\\\theta)$ is a binary cross-entropy function, commonly used in binary classification problems. Eq (\\\\ref{EQ:eqnganlevel1}) is proposed in original GAN and can be optimized by alternate training. The training of discriminator is:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathop{\\\\max}\\\\limits_{\\\\theta} \\\\mathbb{E}_{x\\\\sim p_{r}}[\\\\log(D_\\\\theta(x))]\\n\\t\\t+\\\\mathbb{E}_{z\\\\sim p_z}[\\\\log(1-D_\\\\theta(G_\\\\phi(z)))],\\n\\t\\t\\\\label{EQ:eqndiscriminator}\\n\\t\\\\end{equation}\\n\\twhich is the same as a bi-classification task between real images and generated images. However, the naive binary cross-entropy function suffers from many problems, such as gradients vanishing. {Gradients vanishing is present when the difference between real and generated images as measured by discriminator is large, which leads generators cannot get optimised directions.} Accordingly, many techniques from classification like loss functions and regularization methods have been used to improve the training of discriminator. For instance, to overcome the gradients vanishing problem, Mao et al. propose the LSGANs which adopts the least\\n\\tsquares loss function for the discriminator. The least squares loss function moves the fake samples toward the decision boundary even though they are correctly classified. Based on this property, LSGANs is able to generate samples that are closer to real ones. The loss functions of LSGANs can be defined as follows:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{aligned}\\n\\t\\t\\t&\\\\mathop{\\\\min}\\\\limits_{\\\\theta} \\\\mathbb{E}_{x \\\\sim p_{r}}[(D_{\\\\theta}({x})-b)^{2}] \\n\\t\\t\\t+ \\\\mathbb{E}_{{z} \\\\sim p_{{z}}}[(D_{\\\\theta}(G_{\\\\phi}({z}))-a)^{2}], \\\\\\\\\\n\\t\\t\\t&\\\\mathop{\\\\min}\\\\limits_{\\\\phi} \\\\mathbb{E}_{z \\\\sim p_{z}}[(D_{\\\\theta}(G_{\\\\phi}({z}))-c)^{2}],\\n\\t\\t\\\\end{aligned}\\n\\t\\\\end{equation}\\n\\twhere $b$ and $a$ are objectives that $D$ uses for the training of real and fake samples respectively, $c$ denotes the value that $G$ wants $D$ to believe for fake sample. Gradients vanishing problem of the LSGANs only appears with $D_{\\\\theta}(G_{\\\\phi}({z}))=c$, which is hard. Furthermore, Lin et al. use SVM separating hyperplane that maximizes the margin to propose geometric GAN. Authors use the Hinge loss to train the models, which can be formulated as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{aligned}\\n\\t\\t\\t&\\\\mathop{\\\\min}\\\\limits_{\\\\theta} \\\\mathbb{E}_{x \\\\sim p_{r}}[(1-D_{\\\\theta}({x})]_{+}\\n\\t\\t\\t+ \\\\mathbb{E}_{{z} \\\\sim p_{{z}}}[1+D_{\\\\theta}(G_{\\\\phi}({z}))]_{+}, \\\\\\\\\\n\\t\\t\\t&\\\\mathop{\\\\min}\\\\limits_{\\\\phi} -\\\\mathbb{E}_{z \\\\sim p_{z}}D_{\\\\theta}(G_{\\\\phi}({z})),\\n\\t\\t\\\\end{aligned}\\n\\t\\\\end{equation}\\n\\twhere $[x]_{+}=\\\\mathop{\\\\max}\\\\{0,x\\\\}$.\\n\\tThe motivation of GANs is to train the generator based on the output of the discriminator. Unlike the direct training objective of the classification task (minimizing cross-entropy loss), the objective of generator is indirect (with the help of the discriminator output). Hence, discriminator should provide a richer representation on the truth or false of samples compared to the classification task. More prior information and additional supervision tasks are urgent during the training process of the discriminator. Based on these, some regularization methods, such as {\\\\textit{Data Augmentation and Preprocessing}}, {\\\\textit{Consistency Regularization}}, and {\\\\textit{Self-Supervision}} are proposed to improve the stability and generalizability of discriminator.\",\n \"id\": \"4e9cd8aa-b10a-44ef-a7b7-7d2880558ce2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ],\n [\n \"subsection\",\n \"Low Level: The Perspective of \\\"Real \\\\& Fake\\\"\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Low Level: The Perspective of \\\"Real \\\\& Fake\\\"\"\n },\n {\n \"cite_extract_rate\": 0.642857142857142,\n \"cites\": [\n 88,\n 54,\n 64,\n 93,\n 94,\n 95,\n 78,\n 8318,\n 8317\n ],\n \"content\": \"At middle level, generator $G(z)$ is considered as a distribution mapping function that maps latent distribution $p_z(z)$ to generated distribution $P_g(x)$, and the discriminator $D(x)$ is a distribution distance that evaluates the distance between the target distribution $P_r(x)$ and the generated distribution $P_g(x)$ as illustrated in Figure \\\\ref{FIG}. For the optimal discriminator, the generator $G(z)$ tries to minimize the distance between $P_r(x)$ and $P_g(x)$. For instance, generator of the vanilla GAN\\\\footnote{Vanilla GAN, also known as standard GAN, is the first GAN model.} and $f$-GAN\\\\footnote{\\\\label{ft:5} $f$-GAN is a collective term for a type of GAN models whose discriminator minimizes $f$ divergence. $f$ divergence is the general form of KL divergence. It can be demonstrated as: $D_f(P||Q)=\\\\int q(x)f\\\\big(\\\\frac{p(x)}{q(x)}\\\\big)\\\\rm{d}x$, where $f$ is a mapping function from non-negative real numbers to real numbers ($\\\\mathbb{R}^*\\\\rightarrow\\\\mathbb{R}$) that satisfies: (1) $f(1)=0$. (2) $f$ is a convex function. To be more specific, KL divergence corresponds to $f(u)=u\\\\log u$ and JS divergence corresponds to $f(u)=-\\\\frac{u+1}{2}\\\\log\\\\frac{1+u}{2}+\\\\frac{u}{2}\\\\log u$. More details can be viewed in } are considered to minimize Jensen–Shannon (JS) divergence and $f$ divergence\\\\textsuperscript{\\\\ref{ft:5}}, respectively. When the conditions of LSGANs loss are set to $b-c=1$ and $b-a=2$, generator of the LSGAN considers the minimization of Pearson ${\\\\chi}^2$ divergence. Furthermore, generators of the WGAN-div\\\\footnote{Different from WGAN-div, WGAN minimize Wasserstein distance, not Wasserstein divergence.} and GAN-QP consider the minimization of Wasserstein divergence and Quadratic divergence, respectively.\\n\\t\\\\begin{figure}\\n\\t\\t\\\\centering\\n\\t\\t\\\\includegraphics[scale=0.55]{figs/GAN.pdf}\\n\\t\\t\\\\caption{The framework of GANs. $P_z$ is a latent space distribution, $P_r$ and $P_g$ represent the real distribution and the generated distribution, respectively.}\\n\\t\\t\\\\label{FIG}\\n\\t\\\\end{figure}\\n\\tGenerator is a transportation map from $z$ to $p_g(x)$. In this section, we introduce the optimal transport and the optimal transport with regular term, which leads to the form of Wasserstein GANs with gradient penalty (WGAN-GP) and Wasserstein GANs with Lipschitz penalty (WGAN-LP), respectively. Wasserstein distance is a popular and important distance in GANs and it corresponds to the optimal transport of the generator. To solve the dual problem of Wasserstein distance, Lipschitz continuity is introduced, which is the reason why gradient penalty and weight normalization techniques are proposed in the GANs training.\\n\\t\\\\begin{comment}\\n\\tOptimal transport was proposed in the 18th century to minimize the transportation cost while preserving the measure quantities. Given the space with probability measures $(X,\\\\mu)$ and $(Y,\\\\upsilon)$, if there is a map $T:X\\\\rightarrow Y$ which is measure-preserving, then for any $B\\\\subset Y$, having:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\t\\\\int_{T^{-1}(B)}\\\\mathrm{d}\\\\mu(x)=\\\\int_B \\\\mathrm{d}\\\\upsilon(y).\\n\\t\\\\end{equation}\\n\\tWriting the measure-preserving map as $T_*(\\\\mu)=\\\\upsilon$. For any $x\\\\in X$ and $y\\\\in Y$, the transportation distance is defined as $c(x,y)$, the total transportation cost is given by:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tC(T):=\\\\int_X c(x,T(x)) \\\\mathrm{d}\\\\mu(x).\\n\\t\\\\end{equation}\\n\\tIn the 18th century, Monge et al. proposed the Optimal Mass Transportation Map that corresponds to the smallest total transportation cost: $C(T)$. The transportation cost corresponding to the optimal transportation map is called the Wasserstein distance between probability measures $\\\\mu$ and $\\\\upsilon$:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)=\\\\mathop{\\\\min}\\\\limits_{T}\\\\left\\\\{\\\\int_X c(x,T(x)) \\\\mathrm{d}\\\\mu(x)\\\\ |\\\\ T_*(\\\\mu)=\\\\upsilon\\\\right\\\\}.\\n\\t\\\\end{equation}\\n\\tIn 1940s, Kantorovich proved the existence and uniqueness of the solution for Monge problem, and according to the duality of linear programming, the Kantorovich-Rubinstein (KR) duality of Wasserstein distance is given by:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)\\n\\t\\t=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi ,\\\\psi}\\\\left\\\\{\\\\int_X \\\\varphi \\\\mathrm{d}\\\\mu\\\\ +\\\\int_Y \\\\psi \\\\mathrm{d}\\\\upsilon\\\\ |\\\\ \\\\varphi(x)+\\\\psi(y)\\\\leq c(x,y)\\\\right\\\\}.\\n\\t\\\\end{equation}\\n\\tThis dual problem is constrained, defining the c-transform: $\\\\psi(y)=\\\\varphi^c(y):=inf_x\\\\{c(x,y)-\\\\varphi(x)\\\\}$, and the Wasserstein distance becomes:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi}\\\\left\\\\{\\\\int_X \\\\varphi \\\\mathrm{d}\\\\mu\\\\ +\\\\int_Y \\\\varphi^c \\\\mathrm{d}\\\\upsilon\\\\right\\\\},\\n\\t\\\\end{equation}\\n\\twhere $\\\\varphi$ is called the Kantorovich potential. It can be shown that if $c(x,y)=|x-y|$ and Kantorovich potential satisfies the 1-Lipschitz continuity, then $\\\\varphi^c=-\\\\varphi$. Kantorovich potential can be fitted by a deep neural network, which is recorded as $\\\\varphi_\\\\xi$. Wasserstein distance is:\\n\\t\\\\begin{equation}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)=\\\\mathop{\\\\max}\\\\limits_{||\\\\varphi_\\\\xi||_L\\\\leq 1}\\\\left\\\\{\\\\int_X \\\\varphi_\\\\xi \\\\mathrm{d}\\\\mu\\\\ -\\\\int_Y \\\\varphi_\\\\xi \\\\mathrm{d}\\\\upsilon\\\\right\\\\}.\\n\\t\\t\\\\label{eqn7}\\n\\t\\\\end{equation}\\n\\tIf $X$ is the generated image space, $Y$ is the real sample space, $Z$ is latent space and $g_\\\\theta$ is the geneartor, the Wasserstein GANs (WGAN) is formulated as a min-max problem:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\t\\\\mathop{\\\\min}\\\\limits_{\\\\theta}\\\\mathop{\\\\max}\\\\limits_{||\\\\varphi_\\\\xi||_L\\\\leq 1}\\\\left\\\\{\\\\int_Z \\\\varphi_\\\\xi(g_\\\\theta(z)) \\\\mathrm{d}z \\\\ -\\\\int_Y \\\\varphi_\\\\xi(y) \\\\mathrm{d}y\\\\right\\\\}.\\n\\t\\\\end{equation}\\n\\tIn the optimization process, the generator and the Kantorovich potential function (discriminator) are independent of each other, optimized in a step-by-step iteration.\\n\\tIf $c(x,y)=\\\\frac{|x-y|^2}{2}$, there is a convex function $u$ that is called Brenier potential . The optimal transportation map is given by the gradient map of Brenier potential: $T(x)=\\\\nabla u(x)$. There exists a relationship between Kantorovich potential and Brenier potential : \\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tu(x)=\\\\frac{|x|^2}{2}-\\\\varphi(x).\\n\\t\\\\end{equation}\\n\\tFrom the previous discussion, it is evident that the optimal transportation map (Brenier potential) corresponds to the generator, and Kantorovich potential corresponds to the discriminator. After the discriminator is optimized, the generator is directly drivable without the optimization process .\\n\\tThe transportation cost of Eq (3) is defined as the form of two distribution distances:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tOT(P||Q)=\\\\mathop{inf}\\\\limits_{\\\\pi}\\\\int\\\\pi(x,y)c(x,y)\\\\mathrm{d}x\\\\mathrm{d}y,\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi(x,y)$ is the joint distribution, satisfying $\\\\int_y\\\\pi(x,y)dy=P(x)$ and $\\\\int_x\\\\pi(x,y)dx=Q(y)$. The dual form of Eq (10) is derived as follows::\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tOT(P||Q)=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi ,\\\\psi}\\\\{\\\\int_x \\\\varphi(x)P(x) \\\\mathrm{d}x\\\\ \\\\\\\\ +\\\\int_y\\\\psi(y)Q(y) \\\\mathrm{d}y\\\\ |\\\\ \\\\varphi(x)+\\\\psi(y)\\\\leq c(x,y)\\\\}.\\n\\t\\\\end{equation} \\n\\tConsidering the optimal transportation with regular terms, Peyr{\\\\'e} et al. added the entropic regularization for optimal transportation that transforms the dual problem into a smooth unconstrained convex problem. The regularized optimal transport is defined as:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tOT_c(P||Q)=\\\\mathop{\\\\min}\\\\limits_{\\\\pi}\\\\int\\\\pi(x,y)c(x,y)\\\\mathrm{d}x\\\\mathrm{d}y+\\\\epsilon E(\\\\pi).\\n\\t\\\\end{equation}\\n\\tIf $E(\\\\pi)=\\\\int_x\\\\int_y\\\\pi(x,y)\\\\log(\\\\frac{\\\\pi(x,y)}{P(x)Q(y)})\\\\mathrm{d}x\\\\mathrm{d}y$, Eq (12) can be written as:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\t\\\\begin{split}\\n\\t\\t\\tOT_c(P||Q)=&\\\\mathop{\\\\min}\\\\limits_{\\\\pi}\\\\int\\\\pi(x,y)c(x,y)\\\\mathrm{d}x\\\\mathrm{d}y+\\\\epsilon \\\\int_x\\\\int_y\\\\pi(x,y)\\\\log\\\\left(\\\\frac{\\\\pi(x,y)}{P(x)Q(y)}\\\\right)\\\\mathrm{d}x\\\\mathrm{d}y\\\\\\\\\\n\\t\\t\\ts.t. \\\\int_y\\\\pi(x,y)&\\\\mathrm{d}y=P(x),\\\\int_x\\\\pi(x,y)\\\\mathrm{d}x=Q(y).\\n\\t\\t\\\\end{split}\\n\\t\\\\end{equation}\\n\\tThe dual form of Eq (13) becomes:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{split}\\n\\t\\t\\tOT_c(P||Q)\\n\\t\\t\\t&=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi ,\\\\psi}\\\\int_x \\\\varphi(x)P(x) \\\\mathrm{d}x\\\\ +\\\\int_y\\\\psi(y)Q(y)\\\\mathrm{d}y\\\\\\\\\\n\\t\\t\\t&+\\\\frac{\\\\epsilon}{e}\\\\int_x\\\\int_y\\\\exp\\\\left(\\\\frac{-\\\\left(c(x,y)+\\\\varphi(x)+\\\\psi(y)\\\\right)}{\\\\epsilon}\\\\right)\\\\mathrm{d}x\\\\mathrm{d}y.\\n\\t\\t\\\\end{split}\\n\\t\\t\\\\label{EQ:eqn14}\\n\\t\\\\end{equation}\\n\\t\\\\end{comment}\\n\\tThe details of optimal transportation and optimal transportation with regular terms for WGAN and Lipschitz continuity can be found on the Section \\\\ref{sect:A-1} of the Supplementary Online-only Material. It is pertinent to note that \\\\textit{Gradient Penalty} and \\\\textit{Gradient} \\\\textit{Normalization} are two simple and effective ways to implement the Lipschitz continuity. Furthermore, demonstrates that the spectral norm and the Lipschitz constant have the same meaning. Therefore, the spectral norm can be used to represent the Lipschitz constant. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately. Hence, \\\\textit{Weight Normalization} and \\\\textit{Weight Regularization} can also be used to enable the Lipschitz continuity of the discriminator.\\n\\t\\\\begin{comment}\",\n \"id\": \"a915370f-85fe-4345-a91b-c8d67718061c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ],\n [\n \"subsection\",\n \"Middle Level: The Perspective of \\\"Fitting distribution\\\"\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Middle Level: The Perspective of \\\"Fitting distribution\\\"\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 97,\n 96\n ],\n \"content\": \"WGAN is a popular and important generative adversarial network. From the optimal transport introduced in the last subsection, to obtain Eq (\\\\ref{eqn7}), the discriminator must satisfy the 1-Lipschitz continuity. This section introduces the form of the Lipschitz constant and shows that the spectral norm and the Lipschitz constant have the same meaning.\\n\\t1-Lipschitz continuity is represented as:\\n\\t\\\\begin{equation}\\n\\t||D(x_1)-D(x_2)||\\\\leq ||x_1-x_2||.\\\\footnote{Lipschitz continuity can be defined by any form of norm.}\\n\\t\\\\end{equation}\\n\\tGenerally, considering the K-Lipschitz for a neural network $f(x)$:\\n\\t\\\\begin{equation}\\n\\tf(x)=g_N\\\\circ\\\\cdots g_2\\\\circ g_1(x),\\\\footnote{$\\\\circ$ is the symbol for function cascade. Specifically, $h\\\\circ g(x)=h(g(x))$. This definition of neural network is not general, such as DenseNet and ResNet , which can not be defined like this. Therefore, we do not strictly derive the relationship between the matrix norm and Lipschitz continuity.}\\n\\t\\\\end{equation}\\n\\twhere $g_i(x)=\\\\sigma (W_i x+b_i)$. And K-Lipschitz continuity for $f(x)$ is:\\n\\t\\\\begin{equation}\\n\\t||f(x_1)-f(x_2)||\\\\leq \\\\mathrm{K}||x_1-x_2||,\\n\\t\\\\label{EQ:eqn17}\\n\\t\\\\end{equation}\\n\\twhere K is Lipschitz constant of the function $f$. Due to the consistency of Lipschitz $||h\\\\circ g||_{Lip}\\\\leq ||h||_{Lip}\\\\cdot||g||_{Lip}$, $g_i$ needs to satisfy the C-Lipschitz continuity ($\\\\mathrm{C}=\\\\sqrt[N]{\\\\mathrm{K}}$) so that $f$ satisfies the K-Lipschitz continuity:\\n\\t\\\\begin{equation}\\n\\t||g_i(x_1)-g_i(x_2)||\\\\leq \\\\mathrm{C}||x_1-x_2||,\\n\\t\\\\end{equation}\\n\\t\\\\begin{equation}\\n\\t||\\\\sigma(Wx_1+b)-\\\\sigma(Wx_2+b)||\\\\leq \\\\mathrm{C}||x_1-x_2||.\\n\\t\\\\label{eq:23}\\n\\t\\\\end{equation}\\n\\tWhen $x_1\\\\rightarrow x_2$, the Taylor expansion of Eq (\\\\ref{eq:23}):\\n\\t\\\\begin{equation}\\n\\t||\\\\frac{\\\\partial\\\\sigma}{\\\\partial x} W(x_1-x_2)||\\\\leq \\\\mathrm{C}||x_1-x_2||.\\n\\t\\\\end{equation}\\n\\tNormally, $\\\\sigma$ is a function with limited derivatives such as Sigmoid, so the $\\\\mathrm{C'}$-Lipschitz continuity is be written as:\\n\\t\\\\begin{equation}\\n\\t|| W(x_1-x_2)||\\\\leq \\\\mathrm{C'}||x_1-x_2||,\\n\\t\\\\end{equation}\\n\\twhere $\\\\mathrm{C'}$ is a limited constant, which is determined by $\\\\frac{\\\\partial\\\\sigma}{\\\\partial x}$ and $\\\\mathrm{C}$.\\n\\tSimilarly, the spectral norm of matrix is defined by:\\n\\t\\\\begin{equation}\\n\\t||W||_2=\\\\mathop{\\\\max}\\\\limits_{x\\\\not=0}\\\\frac{||Wx||}{||x||}.\\n\\t\\\\end{equation}\\n\\tIn this context, the spectral norm $||W||_2$ can be used to represent the Lipschitz constant $\\\\mathrm{C'}$. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately. Hence, \\\\textit{Weight Normalization} and \\\\textit{Weight Regularization} can also be used to enable the Lipschitz continuity of the discriminator.\\n\\t\\\\end{comment}\",\n \"id\": \"a397f8d2-2658-471a-8c3b-cbde88271e09\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ],\n [\n \"subsection\",\n \"Lipschitz Continuity and Matrix Norm\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Lipschitz Continuity and Matrix Norm\"\n },\n {\n \"cite_extract_rate\": 0.9,\n \"cites\": [\n 74,\n 98,\n 100,\n 54,\n 102,\n 99,\n 101,\n 64,\n 7190\n ],\n \"content\": \"GANs training is a two-player zero-sum game with a solution to Nash Equilibrium. At high level, we analyze the convergence of GANs by understanding the optimization process. Based on these, some regularization techniques are proposed to guide the GANs model to reach the theoretical equilibrium point leading to improvement in the effectiveness of GANs.\\n\\tReconsidering the Eq (\\\\ref{EQ:eqn1}) in Section 2, the training of GANs is achieved by solving a two-player zero-sum game via Simultaneous Gradient Descent (SimGD) . The updates of the SimGD are given as:\\n\\t\\\\begin{equation} \\n\\t\\t\\\\label{eq:eqn27}\\n\\t\\t\\\\phi^{(k+1)}=\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)}),\\\\quad\\n\\t\\t\\\\theta^{(k+1)}=\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)}).\\n\\t\\\\end{equation}\\n\\tAssuming that the objectives of GANs are convex, many research studies discuss their global convergence characteristics . However, due to the high non-convexity of deep networks, even a simple GAN does not satisfy the convexity assumption . A recent study shows that it is unrealistic to obtain approximate global convergence under the assumption of the optimal discriminator, so the community considers local convergence. It hopes that the trajectory of the dynamic system can enter a local convergence point with continuity iterations, that is, Nash equilibrium:\\n\\t\\\\begin{equation} \\n\\t\\t\\\\bar{\\\\phi}=\\\\mathop{\\\\arg\\\\max}_{\\\\phi}-f(\\\\phi,\\\\bar{\\\\theta}),\\\\quad\\n\\t\\t\\\\bar{\\\\theta}=\\\\mathop{\\\\arg\\\\max}_{\\\\theta}f(\\\\bar{\\\\phi},\\\\theta).\\n\\t\\t\\\\label{EQ:eq24}\\n\\t\\\\end{equation}\\n\\tIf the point $(\\\\bar{\\\\phi},\\\\bar{\\\\theta})$ is called the local Nash-equilibrium, Eq (\\\\ref{EQ:eq24}) holds in a local neighborhood of $(\\\\bar{\\\\phi},\\\\bar{\\\\theta})$.\\n\\tFor this differentiable two-player zero-sum game, a vector is defined as below:\\n\\t\\\\begin{equation}\\n\\t\\tv(\\\\phi,\\\\theta)=\\n\\t\\t\\\\left(\\n\\t\\t\\\\begin{aligned}\\n\\t\\t\\t-\\\\nabla_\\\\phi f(\\\\phi,\\\\theta)\\\\\\\\\\n\\t\\t\\t\\\\nabla_\\\\theta f(\\\\phi,\\\\theta)\\n\\t\\t\\\\end{aligned} \\n\\t\\t\\\\right).\\n\\t\\\\end{equation}\\n\\tThe Jacobian matrix is:\\n\\t\\\\begin{equation}\\n\\t\\tv^{'}(\\\\phi,\\\\theta)=\\n\\t\\t\\\\left(\\n\\t\\t\\\\begin{aligned}\\n\\t\\t\\t-&\\\\nabla^2_{\\\\phi,\\\\phi} f(\\\\phi,\\\\theta)\\\\quad-\\\\nabla^2_{\\\\phi,\\\\theta} f(\\\\phi,\\\\theta)\\\\\\\\\\n\\t\\t\\t&\\\\nabla^2_{\\\\phi,\\\\theta} f(\\\\phi,\\\\theta)\\\\quad\\\\nabla^2_{\\\\theta,\\\\theta} f(\\\\phi,\\\\theta)\\n\\t\\t\\\\end{aligned} \\n\\t\\t\\\\right).\\n\\t\\\\end{equation}\\n\\tAccording to the propositions on the Section \\\\ref{sect:A-2} of the Supplementary Online-only Material, under the premise of asymptotic convergence, the local convergence of GAN is equivalent to the absolute value of all eigenvalues of the Jacobian matrix at the fixed point $(v(\\\\bar{\\\\phi},\\\\bar{\\\\theta})=0)$ being less than 1. To get this condition, \\\\textit{Jacobian Regularization} needs to be used.\",\n \"id\": \"0edf29a6-bba8-43a2-a318-3f272fcb91d9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"5fdbf0be-ea34-468a-9943-d62d05b593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Background and Three Perspectives of GANs Training\"\n ],\n [\n \"subsection\",\n \"High Level: The Perspective of \\\"Training dynamics\\\"\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"High Level: The Perspective of \\\"Training dynamics\\\"\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6998,\n 72\n ],\n \"content\": \"From the perspective of the \\\"Real \\\\& Fake\\\", generator is counterfeiter designed to deceive the discriminator, while discriminator is police designed to distinguish between real and fake samples. The motivation of GANs is to train the generator based on the loss of the discriminator. {Compared to supervised classification tasks, discriminator formally needs to perform only bi-classification tasks, which is easy to implement. Therefore, discriminator is very easy to overfit.} Furthermore, unlike the direct training objective of the classification task (Minimizing cross-entropy loss), the objective of GANs training is indirect. Hence, only one-dimensional output of the discriminator does not provide a complete representation on truth or false of samples. Some studies have shown that the present discriminators contain some significant deficiencies in the frequency domain and attribute domain , which are evidence of the lacking discrimination for discriminators. Excessive shortage of discrimination makes the generator lack incentives from the discriminator to learn useful information of the data. {In addition to discriminator overfitting and lacking discrimination of discriminators, discriminator forgetting is another challenge for GANs.} To alleviate these situations, many regularization methods and additional supervision tasks have been proposed in the literature, which can be divided into three categories: \\\\textit{Data Augmentation and Preprocessing}, \\\\textit{Consistency Regularization}, and \\\\textit{Self-supervision}. {All of them are based on data augmentation and are orthogonal to each other. As shown in Table \\\\ref{table:SOTA }, The state-of-the-art GANs always adopt two or even all of the above regularization.}\",\n \"id\": \"8469161e-625b-422b-90e0-a6cca9638a04\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ]\n ],\n \"subsections\": [\n \"6dc08e83-31a2-4629-92a0-ea9cbb6c27a0\",\n \"a5e04be5-4733-477b-9081-c52ce2a3be68\",\n \"4e2e9490-faa8-4a2f-9dbd-66ae6c225118\",\n \"2267f79d-3862-4c72-bdff-9670534c3017\"\n ],\n \"title\": \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n },\n {\n \"cite_extract_rate\": 0.7894736842105261,\n \"cites\": [\n 91,\n 7191,\n 111,\n 104,\n 112,\n 107,\n 106,\n 110,\n 109,\n 108,\n 103,\n 113,\n 61,\n 8319,\n 105\n ],\n \"content\": \"\\\\begin{figure}\\n\\t\\t\\\\centering\\n\\t\\t\\\\includegraphics[scale=.35]{figs/data_augmentation_framework.pdf}\\n\\t\\t\\\\caption{Framework of data augmentation and preprocessing for updating D (left) and G (right). (Coming from )}\\n\\t\\t\\\\label{fig:data_augmentation_framework}\\n\\t\\\\end{figure}\\n\\tData Augmentation plays a significant role in deep learning algorithms. It increases the diversity of the training data naturally, thus reduces the overfitting in many computer vision and graphics applications . Date augmentation adopts different data transformation techniques ({\\\\color{red}$T$}) to increase the number of training samples. One type of data transformation is spatial transformation of data, such as $zoomout$, $zoomin$, $translation$, $translationx$, $translationy$, $cutout$ , $cutmix$ ; The other is\\n\\tvisual transformation, such as $brightness$, $redness$, $greenness$, $blueness$, $mixup$ . {Furthermore, recent study is also attempting to use frequency transformation to data augmentation. }\\n\\tSimilarly, the performance of GANs heavily deteriorates given a limited amount of training data . {For instance, shows that Frechet Inception Distance (FID) starts to rise at some point during training and outputs of discriminator keep drifting apart during training, when training data is limited. More analysis can be found in the survey on data-efficient GANs training.} However, recent studies observe that augmenting only real images (Only applying $T$ to (i) in Figure \\\\ref{fig:data_augmentation_framework}), only generated images (Only applying $T$ to (ii) in Figure \\\\ref{fig:data_augmentation_framework}), and only discriminator (Both applying $T$ to (i) and (ii) in Figure \\\\ref{fig:data_augmentation_framework}) do not help with GANs training. Naturally, one problem needs to be considered: whether the overfitting exists in GANs' training? Some studies demonstrate that, even with big dataset and state of the art models, the training of GANs suffers from severe overfitting. Furthermore, in case of small training data, overfitting occurs at an early stage in the training. Recently, some studies on data augmentation for GANs training have been proposed. It is argued that the classical data augmentation approach could mislead the generator to learn the distribution of the augmented data, which could be different from that of the original data. To deal with this problem, these studies augment both real and fake samples and let gradients propagate through the augmented samples to G (Applying $T$ to (i), (ii), and (iii) in Figure \\\\ref{fig:data_augmentation_framework}). By adding the data augmentation to all processes of GANs training, the performance of GANs has been significantly improved. {However, this \\\"\\\\textit{Augment All}\\\" strategy may lead to the “leaking” of augmentations to the generated samples, which is highly undesirable. The experiments in demonstrate that as long as the probability of executing the augmentation remains below 0.8, leaks are unlikely to happen in practice.}\\n\\tData augmentation in GANs has remarkable achievement. However, which augmentation is most beneficial for GANs training is still an open problem. Figure 2 in shows the FID comparisons of BigGAN on CIFAR-10 dataset. For data augmentation (represented by ‘vanilla\\\\_rf’), the operations in spatial augmentation such as $translation$, $zoomout$, and $zoomin$, are much more effective than the operations in visual augmentation, such as $brightness$, $colorness$ ($redness$ and $greenness$), and $mixup$. The results indicate that augmentation leads to spatial changes which improves GANs performance compared with cases where visual changes are induced. It is easy to understand that generated images are significantly lacking in detail information compared to the real images, and spatial augmentation improves the ability of the generator to fit detailed textures through spatial changes. {$InstanceNoise$, resulting in images out of the natural data manifold, cannot help with improving GANs performance. Apart form applying only a limited range of augmentations, some studies explore some strong data augmentations in GANs training. For instance, Jeong et al. adopt contrastive learning to extract more useful information under stronger data augmentation beyond the existing yet limited practices. Combining adaptive strategies and 18 transformations (Both spatial and visual transformations) , even and $InstanceNoise$ only can bring performance improvement over strong GANs baselines. {Furthermore, is the first method to tackle the generative learning trilemma with denoising diffusion GANs.} Apart from these, Jiang et al. also devise an adaptive strategy to control the strength of selecting generated images to augment real data, which can further boosts the performance of GANs. Data augmentation is popular and significant in GANs training, whose achievements are attributed to improving discrimination, avoiding overfitting, and increasing the overlap between real and fake distributions.}\\n\\t{Different from data augmentation that increases the amount of the training data, data preprocessing only adopt prior knowledge and do some uniform data transformation before the network training. Data preprocessing is orthogonal to data augmentation, which can further enhance the performance combining the data augmentation.} Li et al. indicate that high-frequency components between real and fake images are different, which is not conducive to the GANs training. They propose two preprocessing eliminating high-frequency differences in GANs training: High-Frequency Confusion (HFC) and High-Frequency Filter (HFF). These methods are applied in places (i), (ii), and (iii) in Figure \\\\ref{fig:data_augmentation_framework} and improve the performance of GANs with a fraction of the cost. \\n\\tIn summary, both data augmentation and data preprocessing improve the performance of GANs with little cost. Data augmentation uses different transformations to improve discrimination and avoid disciminator overfitting. Furthermore, spatial augmentations achieve better performance than visual augmentations. More specifically, Zhao et al. demonstrate that hybrid augmentation with $Color + Translation + Cutout$ is especially effective and is widely used in other studies . Adaptive data augmentation (ADA) is the most popular method in GANs. Besides the data augmentation, data preprocessing is also a remarkable method.\\n\\t\\\\begin{figure}\\n\\t\\t\\\\centering\\n\\t\\t\\\\includegraphics[scale=.3]{figs/consistency.pdf}\\n\\t\\t\\\\caption{Overview of consistency regularization, where image consistency regularization updates the G (left) and network consistency regularization updates the D (right). $H$ is the feature mapping function and {\\\\color{red}$T$} is different data transformation techniques.}\\n\\t\\t\\\\label{fig:consistency}\\n\\t\\\\end{figure}\",\n \"id\": \"6dc08e83-31a2-4629-92a0-ea9cbb6c27a0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"8469161e-625b-422b-90e0-a6cca9638a04\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Data Augmentation and Preprocessing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Augmentation and Preprocessing\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 115,\n 114,\n 7000,\n 7192\n ],\n \"content\": \"In context of semi-supervised and unsupervised learning, consistency regularization has been widely used in . It is motivated by the fact that models should produce consistent predictions given input and their semantics-preserving augmentations, such as image rotating, and adversarial attacks. It is pertinent to note that the supervision of GANs training is weak. To increase the discrimination of discriminator, some consistency regularization techniques have also been used. Due to different goals, we divide these into two parts: \\\\textit{Image Consistency} and \\\\textit{Network Consistency} as demonstrated in Figure \\\\ref{fig:consistency}.\",\n \"id\": \"a5e04be5-4733-477b-9081-c52ce2a3be68\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"8469161e-625b-422b-90e0-a6cca9638a04\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Consistency Regularization\"\n ]\n ],\n \"subsections\": [\n \"a0c72661-2f99-4ec9-8181-70dfdeab36dc\",\n \"86d17fb0-5862-4222-a247-42d917acf482\"\n ],\n \"title\": \"Consistency Regularization\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 116,\n 6998,\n 117,\n 121,\n 8320,\n 119,\n 118,\n 120,\n 91,\n 69\n ],\n \"content\": \"The purpose of GANs is to generate fake images similar to real ones. In GANs, the discriminator is generally used to distinguish real images and generated images. However, outputs of the discriminator with only one dimension hardly portray the authenticity of the image completely. To improve the representation of the discriminator, some studies extend the outputs of the discriminator, for example, relativistic discriminator , distribution discriminator , and cascading rejection . Apart from this, some studies reduce the training difficulty of discriminators by introducing prior information. Regularizing the distance between the generated and real images with different measurements, namely image consistency, is the focus of this paper. The overview of it is demonstrated in left part of Figure \\\\ref{fig:consistency}, where consistency regularization is used to update generator (G) and can be formulated as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_C=\\\\mathbb{E}_{x\\\\sim p_r, z\\\\sim p_z} C(H(x),H(G(z))),\\n\\t\\t\\\\label{EQ:image_consistency}\\n\\t\\\\end{equation}\\n\\twhere $H$ is the feature mapping function and $C$ is the consistency measurement function. Different image consistency regularization have different $H$ and $C$. For instance, Salimans et al. recommend that the generator is trained using a feature matching procedure. The objective is:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_C=||\\\\mathbb{E}_{x\\\\sim p_r} f(x)-\\\\mathbb{E}_{z\\\\sim p_z}f(G(z))||^2_2,\\n\\t\\\\end{equation}\\n\\twhere $f(x)$ denotes the intermediate layer of the discriminator. Similarly, the intermediate layer of another pre-trained classification model is an alternate option. The empirical results indicate that feature matching is indeed effective in situations where normal GAN becomes unstable. Unlike the above study which only uses mean feature matching to training generators, Mroueh et al. propose McGAN, which trains both the generator and discriminator using the mean and covariance feature matching. The objective is:\\n\\t\\\\begin{equation}\\n\\t\\t\\t\\\\mathcal{L}_C=\\\\mathcal{L}_\\\\mu+\\\\mathcal{L}_\\\\sigma\\n\\t\\t\\t=||\\\\mu(p_r)-\\\\mu(G(p_z))||_q+||\\\\sum\\\\left(p_r\\\\right)-\\\\sum(p(G(z)))||_k,\\n\\t\\\\end{equation}\\n\\twhere $\\\\mu(p_r)=\\\\mathbb{E}_{x\\\\sim p_r} f(x)$ and $\\\\sum(p_r)=\\\\mathbb{E}_{x\\\\sim p_r} f(x)\\\\cdot f(x)^\\\\mathrm{T}$ represent the mean and the covariance of the feature layer $f(x)$, respectively. Apart from statistical differences, some studies focus on the difference in frequency domain between the generated and real image. For instance, Durall et al. find that the deep generative models based on up-convolution fail to reproduce spectral distributions leading to considerable differences in the spectral distributions between real images and generated images. Thus, the spectral regularization has been proposed as follows:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_C=\\\\frac{1}{M/2-1}\\\\sum_{i=0}^{M/2-1}AI^{real}_i\\\\cdot\\\\log(AI^{fake}_i)\\n\\t\\t+(1-AI^{real}_i)\\\\cdot\\\\log(1-AI^{fake}_i),\\n\\t\\\\end{equation}\\n\\twhere $M$ is the image size and $AI$ is the spectral representation from the Fourier transform of the images. Corresponding to Eq (\\\\ref{EQ:image_consistency}), $H$ and $C$ are implemented with $AI$ and cross-entropy, respectively.\\n\\tContrary to this, the research study uses hard example mining to improve the discriminatory of the discriminator based on the difference between real and generated samples under different metrics. Although this paradigm is different from the paradigm of image consistency regularization, both cases are motivated by obtaining generated samples similar to real images under different distance measures, so we integrate them. Chen et al. consider both downsampling strategies: downsampling with anti-aliasing and downsampling without anti-aliasing, leads to high frequencies missing in the discriminator. High frequencies missing leads to high frequency deviation between real and generated images. To mitigate this issue, authors propose SSD-GAN, which introduces an additional spectral classifier to detect frequency spectrum discrepancy between real and generated images and integrate it into the discriminator of GANs. The overall realness of sample x is represented as:\\n\\t\\\\begin{equation}\\n\\t\\tD^{s s}(x)=\\\\lambda D(x)+(1-\\\\lambda) C(\\\\phi(x)),\\n\\t\\\\end{equation}\\n\\twhere the enhanced discriminator $D^{s s}$ consists of two modules, a vanilla discriminator $D$ that measures the spatial realness, and a spectral classifier $C$. $\\\\lambda$ is a hyperparameter that controls the relative importance of the spatial realness and the spectral realness. The adversarial loss of the framework can be written as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_{D} =\\\\mathbb{E}_{x \\\\sim p_{\\\\text {data }}(x)}\\\\left[\\\\log D^{s s}(x)\\\\right] \\n\\t\\t+\\\\mathbb{E}_{x \\\\sim p_{g}(x)}\\\\left[\\\\log \\\\left(1-D^{s s}(x)\\\\right)\\\\right].\\n\\t\\\\end{equation}\\n\\t{karnewar et al. introduce adversarial loss into the intermediate layer of the generator, which provides multiple and richer metrics for the training of generator.}\\n\\t\\\\begin{table}\\n\\t\\t\\\\caption{The summary of the consistency regularization.}\\n\\t\\t\\\\label{table:consistency}\\n\\t\\t\\\\tiny\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{c| c }\\t\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\tMethod& Consistency regularization term $L_C$\\t \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tMean regularization &$||\\\\mathbb{E}_{x\\\\sim p_r} f(x)-\\\\mathbb{E}_{z\\\\sim p_z}f(G(z))||_q$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tMean and Convariance regularization &$||\\\\mathbb{E}_{x\\\\sim p_r} f(x)-\\\\mathbb{E}_{z\\\\sim p_z}f(G(z))||_q+||\\\\mathbb{E}_{x\\\\sim p_r} f(x)\\\\cdot f(x)^\\\\mathrm{T}-\\\\mathbb{E}_{z\\\\sim p_z}f(G(z))\\\\cdot f(G(z))^\\\\mathrm{T}||_k$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSpectral regularization &$\\\\mathcal{L}_C=\\\\frac{1}{M/2-1}\\\\sum_{i=0}^{M/2-1}AI^{real}_i\\\\cdot\\\\log(AI^{fake}_i)+(1-AI^{real}_i)\\\\cdot\\\\log(1-AI^{fake}_i)$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tCR-GAN &$\\\\mathbb{E}_{x\\\\sim p_r}|| D(x)-D(T(x))||_2$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tbCR-GAN &$\\\\lambda_{real}\\\\mathbb{E}_{x\\\\sim p_r}|| D(x)-D(T(x))||_2+\\\\lambda_{fake}\\\\mathbb{E}_{z\\\\sim p_z}||D(G(z))-D(T(G(z)))||_2$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tzCR-GAN&$\\\\lambda_{dis}\\\\mathbb{E}_{z\\\\sim p_z}||D(G(z))-D(G(T(z)))||_2-\\\\lambda_{gen}\\\\mathbb{E}_{z\\\\sim p_z}||G(z)-G(T(z))||_2$\\\\\\\\\\n\\t\\t\\t\\\\bottomrule\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table}\\n\\tThe summary of the image consistency regularization is given in Table \\\\ref{table:consistency}. In summary, image consistency considers that the real images and the generated images are similar not only in the output of discriminator, but also in statistical information and frequency domain. The analysis of biases between real and generated images using different metrics will be an interesting future research direction.\",\n \"id\": \"a0c72661-2f99-4ec9-8181-70dfdeab36dc\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"a5e04be5-4733-477b-9081-c52ce2a3be68\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Consistency Regularization\"\n ],\n [\n \"subsubsection\",\n \"Image Consistency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Image Consistency\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 91,\n 8320,\n 69,\n 111\n ],\n \"content\": \"Network consistency regularization can be regarded as Lipschitz continuity on semantics-preserving transformation. Specifically, we hope discriminator is insensitive to semantics-preserving transformation, which drives the discriminator to pay more attention to the authenticity of the images. For example, in the image domain, the reality of images should not change if we flip the image horizontally or translate the image by a few pixels. To resolve this, Zhang et al. propose the Consistency Regularization GAN (CR-GAN) that uses the consistency regularization on the discriminator during GANs training:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_C=\\\\mathbb{E}_{x\\\\sim p_r}|| D(x)-D(T(x))||_2,\\n\\t\\\\end{equation}\\n\\twhere $T(x)$ represents a transformation (shift, flip, cutout, etc.) of images. One key problem with the CR-GAN is that the discriminator might occur the 'mistakenly believe'. 'mistakenly believe' considers that the transformations are actual features of the target dataset, due to only applying these transformations on real images. This phenomenon is not easy to notice for certain types of transformations (e.g. image shifting and flipping). However, some types of transformations, such as cutout transformations, contain visual artifacts not belonging to real images, which effects greatly limits the choice of advanced transformations we could use. To address this issue, Zhao et al. propose Balanced Consistency Regularization (bCR-GAN) that uses regulation with respect to both real and fake images and balances the training of discriminator between real images and fake images by $\\\\lambda_{real}$ and $\\\\lambda_{fake}$:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_C=\\\\lambda_{real}\\\\mathbb{E}_{x\\\\sim p_r}|| D(x)-D(T(x))||_2\\n\\t\\t+\\\\lambda_{fake}\\\\mathbb{E}_{z\\\\sim p_z}||D(G(z))-D(T(G(z)))||_2.\\n\\t\\\\end{equation}\\n\\tThe overview of bCR is demonstrated in right part of Figure \\\\ref{fig:consistency}.\\n\\tContrary to the methods which focus on consistency regularization with respect to transformations in image space, Zhao et al. also propose Latent Consistency Regularization (zCR) that considers the consistency regularization on transformations in latent space. Authors expect that output of the discriminator ought not to change much with respect to the small enough perturbation $\\\\Delta z$ and modify the discriminator loss by enforcing:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}^D_C=\\\\lambda_{dis}\\\\mathbb{E}_{z\\\\sim p_z}||D(G(z))-D(G(T(z)))||_2,\\n\\t\\\\end{equation}\\n\\twhere $T(z)$ represents the added small perturbation noise. However, if only this loss is added into the GAN loss, mode collapse can easily appear in the training of generators. To avoid this, an inverse gradient penalty (we will describe it in section 6.2) is added to modify the loss function for generator. Hence, we modify the generator loss by enforcing:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}^D_C=-\\\\lambda_{gen}\\\\mathbb{E}_{z\\\\sim p_z}||G(z)-G(T(z))||_2.\\n\\t\\\\end{equation}\\n\\tNaturally, putting both bCR and zCR together, Improved Consistency Regularization (ICR) is also proposed by Zhao et al. . In addition, there are some applications where cyclic consistency regularization is used for unpaired image-to-image translation . The summary of network consistency regularization is given in Table \\\\ref{table:consistency}. \\n\\tIn summary, network consistency considers the networks, especially the discriminator, to be insensitive to semantics-preserving transformation ($T$). The results in demonstrate that random shift and flip is the best way to perform image transformation on the CIFAR-10 dataset. Furthermore, the FID results with CR, bCR, zCR, and ICR (where transformation is flipping horizontally and shifting by multiple pixels) as presented in are shown in Table \\\\ref{table:consistency_results}. The results demonstrate that network consistency regularization can significantly improve the performance of GANs. However, which transformation is best for consistency regularization, is a question. Zhao et al compare the effect of different data transformation techniques (mentioned in Section 5.1) on bCR. Figure 2 in shows the FID results of BigGAN adding bCR (represented by 'bcr') on CIFAR-10 dataset. From the results, the best BigGAN FID 8.65 is with transformation technology $translation$ of strength $\\\\lambda=0.4$, outperforming the corresponding FID 10.54 reported in Zhao et al. . Moreover, spatial transforms, which retain the major content while introducing spatial variances, can substantially improve GANs performance together with bCR. While visual transforms, which retain the spatial variances, can not further improve the performance of GANs compared with data augmentation only. Furthermore, bCR with stronger transformation (larger value of $\\\\lambda_{aug}$) does not improve the performance of GANs, the optimal value of $\\\\lambda_{aug}$ is uncertain for different data transformations. \\n\\t\\\\begin{table}\\n\\t\\t\\\\caption{FID scores for class conditional image generation of the network consistency regularization (Data come from ).}\\n\\t\\t\\\\label{table:consistency_results}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{ccc}\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\tModels & CIFAR-10 & ImageNet \\\\\\\\\\n\\t\\t\\t\\\\hline SNGAN & 17.50 & 27.62 \\\\\\\\\\n\\t\\t\\tBigGAN & 14.73 & 8.73 \\\\\\\\\\n\\t\\t\\tCR-BigGAN & 11.48 & 6.66 \\\\\\\\\\n\\t\\t\\tbCR-BigGAN & 10.54 & 6.24 \\\\\\\\\\n\\t\\t\\tzCR-BigGAN & 10.19 & 5.87 \\\\\\\\\\n\\t\\t\\tICR-BigGAN & $\\\\mathbf{9 . 2 1}$ & $\\\\mathbf{5 . 3 8}$ \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table}\",\n \"id\": \"86d17fb0-5862-4222-a247-42d917acf482\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"a5e04be5-4733-477b-9081-c52ce2a3be68\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Consistency Regularization\"\n ],\n [\n \"subsubsection\",\n \"Network Consistency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Consistency\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 123,\n 122,\n 7000\n ],\n \"content\": \"Self-supervised learning aims to learn representations from the data itself without explicit manual supervision. Recently, some self-supervised studies provide competitive results on ImageNet classification and the representations learned from which transfer well to downstream tasks. Self-supervised learning outperforms its supervised pre-training counterpart in many tasks, such as detection and segmentation, sometimes surpassing it by large margins. This suggests that self-supervised learning obtains more representational features and significantly improve the representation of networks. Based on this, self-supervised learning is introduced into the training of GANs, and we divide them into two categories according to different self-supervision tasks: \\\\textit{Predictive Self-supervised Learning} and \\\\textit{Contrastive Self-supervised Learning}.\",\n \"id\": \"4e2e9490-faa8-4a2f-9dbd-66ae6c225118\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"8469161e-625b-422b-90e0-a6cca9638a04\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Self-Supervision\"\n ]\n ],\n \"subsections\": [\n \"b5c16876-cba3-44e8-b6be-d7b6cabbf7ee\",\n \"aff71d0b-4dab-423d-9325-d61d96004987\"\n ],\n \"title\": \"Self-Supervision\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 126,\n 132,\n 8321,\n 128,\n 124,\n 129,\n 130,\n 131,\n 7193,\n 103,\n 127,\n 55,\n 113,\n 111,\n 125\n ],\n \"content\": \"Predictive self-supervised learning is a popular method to improve the representation of neural networks by introducing additional supervised tasks, such as context prediction and rotations prediction . Predictive self-supervised learning is introduced into GANs by Chen et al. to avoid discriminator forgetting. Discriminator forgetting means that the discriminator does not remember all tasks at the same time during the training process. For example, learning varying levels of detail, structure, and texture, which causes the discriminator to fail to get a comprehensive representation of the current images. \\\"If the outcome is your focus, then it’s easy to look for shortcuts. And ultimately shortcuts keep you from seeing the truth, it drains your spark for life. What matters most is your will to seek the truth despite the outcome.\\\"\\\\footnote{Come from \\\"JoJo's Bizarre Adventure:Golden Wind\\\" -Araki Hirohiko.}. The same is true for GANs, which are only driven by the loss of discriminator, which is easy to distinguish between real images and generated images through shortcuts, instead of the texture and structural features we need. Predictive self-supervised learning solves this problem by introducing new generalization tasks, which also helps to prevent overfitting. The overview of the predictive self-supervised learning of GANs are demonstrated on Figure \\\\ref{fig:predict_self_supervised}. Depending on the data transformation function $T$, we can design different self-supervised tasks. \\n\\t\\\\begin{figure}\\n\\t\\t\\\\centering\\n\\t\\t\\\\includegraphics[scale=.4]{figs/predict_self_supervised.pdf}\\n\\t\\t\\\\caption{Overview of the predict self supervised learning of GANs, where $C$ performs the predict classification task and shares the weights with the discriminator except for the last layer, {\\\\color{red}$T$} is different data transformation techniques. Furthermore, the self-supervised task of generated images is used to update the generator (left) and the self-supervised task of real images is used to update the classification (right).}\\n\\t\\t\\\\label{fig:predict_self_supervised}\\n\\t\\\\end{figure}\\n\\tChen et al. introduced the predictive self-supervision in GANs training. Authors adopt the rotation prediction as the expanding task to prevent the discriminator from forgetting. Besides, plenty of other prediction tasks have also been proposed to improve the discrimination. Huang et al. exploit the feature exchange to make the discriminator learn the proper feature structure of natural images. Baykal et al. introduce a reshuffling task to randomly arrange the structural blocks of the images, thus helping the discriminator increase its expressive capacity for spatial structure and realistic appearance. Contrary to the methods for designing tasks at the image or feature level, Patel et al. propose a self-supervised task with latent transformation detection, which identifies whether the latent transformation applied in the given pair is the same as that of the other pair. All above methods have designed different self-supervised tasks, and their loss functions can be formulated as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{split}\\n\\t\\t\\t\\\\mathcal{L}_{D,C}=-\\\\lambda_{r}\\\\mathbb{E}_{x\\\\sim p^T_r}\\\\mathbb{E}_{T_k\\\\sim {T}}\\\\log\\\\big( C_k(x)\\\\big) \\\\quad \\\\text { for } k=1, \\\\ldots, K,\\\\\\\\\\n\\t\\t\\t\\\\mathcal{L}_G=-\\\\lambda_{g}\\\\mathbb{E}_{x\\\\sim p^T_g}\\\\mathbb{E}_{T_k\\\\sim {T}}\\\\log\\\\big( C_k(x)\\\\big) \\\\quad \\\\text { for } k=1, \\\\ldots, K,\\n\\t\\t\\\\end{split}\\n\\t\\\\end{equation}\\n\\twhere ${T}$ represents the different types of image transfer, such as rotation and reshuffling. Furthermore, $T_k$ represents different forms of the transfer ${T}$, such as $0^\\\\circ,90^\\\\circ,180^\\\\circ, 270^\\\\circ$ for rotation task, $K$ is the number of transformed forms, $C_k$ is the k-th output of the classifier $C$ that shares parameters with discriminator except for two different heads, $P^T_r$ and $P^T_g$ are the transformed distributions of real and generated images, respectively. For rotation conversion task , $K=4$, and the classifier $C$ predicts the rotation angle; For feature exchange task , $K=2$, and the classifier $C$ predicts whether the swap has occurred; For block reshuffling task , the image is divided into $9$ blocks and the number of the permutations is $9!$, which is unnecessarily huge. Thirty different permutations are selected in terms of the Hamming distances between the permutations in . As a result, $K$ is set to 30, and classifier $C$ predicts the Hamming distances of different permutations; For the latent transformation task, $K=2$, and the classifier $C$ predicts whether the transformations parameterized by the same $\\\\epsilon$ or different. {Besides, some study introduces the autoencoder task, making discriminator reconstruct the input. }\\n\\t\\\\begin{table*}\\n\\t\\t\\\\renewcommand\\\\arraystretch{1}\\n\\t\\t\\\\caption{The summary of the Self-supervision}\\n\\t\\t\\\\label{table:self-supervised}\\n\\t\\t\\\\setlength{\\\\tabcolsep}{0.5mm}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{c| c |c }\\t\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\t\\\\tiny Method&\\\\tiny Description&\\\\tiny Types\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny Rotation Prediction &\\\\tiny Predicting the angle of rotation ($0^\\\\circ,90^\\\\circ,180^\\\\circ, 270^\\\\circ$)&\\\\tiny PSS\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny Feature Exchange Detection &\\\\tiny Predicting if some exchanges have occurred at the feature level (yes or not)&\\\\tiny PSS\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny Block Reshuffling Prediction &\\\\tiny Predicting the Hamming distances of different reshuffling in image level (Total 30 categories)&\\\\tiny PSS\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t\\\\tiny Latent Transformation Detection &\\\\tiny Predicting if some exchanges have occurred at the latent space level (yes or not)&\\\\tiny PSS\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny Autoencoder &\\\\tiny Reconstruct the input of the discriminator&\\\\tiny PSS\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny InfoMax-GAN &\\\\tiny \\\\makecell[c]{Positive pairs: Global and local features of an image (both real and fake images)\\\\\\\\ Negative pairs: Global and local features of different images (both real fake images)}&\\\\tiny CSS\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t\\\\tiny Cntr-GAN &\\\\tiny \\\\makecell[c]{Positive pairs: Two different data transformations of the same image (both real and fake images).\\\\\\\\ Negative pairs: Otherwise}&\\\\tiny CSS\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny\\tContraD &\\\\tiny \\\\makecell[c]{Positive pairs: Two different data transformations of the same image (real images only) + Two fake images\\\\\\\\ Negative pairs: Otherwise}&\\\\tiny CSS\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny\\tInsGen &\\\\tiny \\\\makecell[c]{Positive pairs: Two different data transformations (additional latent transformations for fake image)\\\\\\\\ of the same image (both real and fake images). Negative pairs: Otherwise.}&\\\\tiny CSS\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\tiny\\tFakeCLR &\\\\tiny \\\\makecell[c]{Positive pairs: Two different data transformations and additional latent transformations for fake image.\\\\\\\\ Negative pairs: Otherwise.}&\\\\tiny CSS\\\\\\\\\\n\\t\\t\\t\\\\bottomrule\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table*}\\n\\tThe above methods design different kinds of self-supervised prediction tasks and participate in the training of the discriminator or generator, independently, having “loophole” that, during generator learning, $G$ could exploit to minimize $\\\\mathcal{L}_G$ without truly learning the data distribution. To address this issue, Ngoc-TrungTran et al. introduce true or false judgment along with self-supervised prediction. The number of classification is $K+1$, while the loss function can be expressed as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{aligned}\\n\\t\\t\\t\\\\mathcal{L}_{D,C}=&-\\\\lambda_{r}\\\\Bigg(\\\\mathbb{E}_{x\\\\sim p^T_r}\\\\mathbb{E}_{T_k\\\\sim {T}}\\\\log\\\\big( C_k(x)\\\\big)\\n\\t\\t\\t+\\\\mathbb{E}_{x\\\\sim p^T_g}\\\\mathbb{E}_{T_k\\\\sim {T}}\\\\log\\\\big( C_{K+1}(x)\\\\big)\\\\Bigg)\\\\quad \\\\text { for } k=1, \\\\ldots, K,\\\\\\\\\\n\\t\\t\\t\\\\mathcal{L}_G=&-\\\\lambda_{g}\\\\Bigg(\\\\mathbb{E}_{x\\\\sim p^T_g}\\\\mathbb{E}_{T_k\\\\sim {T}}\\\\log\\\\big( C_k(x)\\\\big)\\n\\t\\t\\t-\\\\mathbb{E}_{x\\\\sim p^T_g}\\\\mathbb{E}_{T_k\\\\sim {T}}\\\\log\\\\big( C_{K+1}(x)\\\\big)\\\\Bigg)\\\\quad \\\\text { for } k=1, \\\\ldots, K,\\n\\t\\t\\\\end{aligned}\\n\\t\\\\end{equation}\\n\\twhere $C_k$ is a classifier that predicts the rotation angles and $C_{K+1}$ is a classifier that predicts the truth of the images. The new self-supervised rotation-based GANs use the multi-class minimax game to avoid the mode collapse, which is better than the original predictive self-supervised paradigm.\\n\\tIn summary, predictive self-supervised learning improves the discrimination by designing different self-supervised prediction tasks, among them, rotation prediction is widely used () for its simplicity and practicality. The summary of different methods is illustrated in Table \\\\ref{table:self-supervised}. {All of these methods are juxtaposed with each other. However, few studies have used multiple self-supervised tasks simultaneously, and the use of multiple self-supervised tasks to improve GANs training is still an open problem.}\",\n \"id\": \"b5c16876-cba3-44e8-b6be-d7b6cabbf7ee\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"4e2e9490-faa8-4a2f-9dbd-66ae6c225118\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Self-Supervision\"\n ],\n [\n \"subsubsection\",\n \"Predictive Self-Supervised Learning (PSS)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Predictive Self-Supervised Learning (PSS)\"\n },\n {\n \"cite_extract_rate\": 0.9166666666666661,\n \"cites\": [\n 134,\n 133,\n 123,\n 130,\n 103,\n 125,\n 122,\n 7000,\n 111,\n 61\n ],\n \"content\": \"Contrastive self-supervised Learning , as the name implies, learn representations by contrasting positive and negative examples. These techniques have resulted in empirical success in computer vision tasks with unsupervised contrastive pre-training. A handful number of studies demonstrate that self-supervised learning outperforms its supervised pre-training counterpart in many tasks, which indicates contrastive self-supervised learning leads to more expressive features. Considering two views $(v^{(1)}$ and $v^{(2)})$, contrastive self-supervised learning aims to identify whether two views are dependent or not. More specifically, it means to maximize the mutual information of positive pairs. To this end, Oord et al. propose to minimize InfoNCE loss, which turns out to maximize a lower bound of mutual information. The InfoNCE loss is defined by:\\n\\t\\\\begin{equation}\\n\\t\\tL_{\\\\mathrm{NCE}}\\\\left({v}_{i}^{(1)} ; {v}^{(2)}, s\\\\right):=-\\\\log \\\\frac{\\\\exp \\\\left(s\\\\left({v}_{i}^{(1)}, {v}_{i}^{(2)}\\\\right)\\\\right)}{\\\\sum_{j=1}^{K} \\\\exp \\\\left(s\\\\left({v}_{i}^{(1)}, {v}_{j}^{(2)}\\\\right)\\\\right)},\\n\\t\\\\end{equation}\\n\\twhere $s(,)$ is the score function that measure the similarity, positive pairs ($v_i^{(1)}$ and $v_i^{(2)}$) are different views of the same sample, and negative pairs ($v_i^{(1)}$ and $v_j^{(2)}$, $(i\\\\neq j)$) are different views of different samples. InfoNCE loss is the cornerstone of contrastive self-supervised learning as depicted in Figure \\\\ref{fig:contrastive_self_supervised}. \\n\\t\\\\begin{figure}\\n\\t\\t\\\\centering\\n\\t\\t\\\\includegraphics[scale=.3]{figs/constrative_self_supervised.pdf}\\n\\t\\t\\\\caption{Overview of the contrastive self supervised learning, where x is real of fake images, $v^{(1)}$ and $v^{(2)}$ are different views of the image x, s is the score function (Usually discriminator in GANs) that measures the similarity, the square on the middle part is the label of InfoNCE (blue, white, and black squares are labels with 1, 0, and undefined respectively.), and the square on the right part is the label of SimCLR. Obviously, SimCLR defines more negative pairs to improve the sample utilization.}\\n\\t\\t\\\\label{fig:contrastive_self_supervised}\\n\\t\\\\end{figure}\\n\\tMany advanced self-supervised methods are implemented by modifying the views of images $v^{(1)}$, $v^{(2)}$ and score function $s(,)$. Specifically, Deep InfoMAX maximizes the mutual information between local and global features, that is, image $x$ passes through the encoder $E_{\\\\psi}=f_{\\\\psi}\\\\circ C_{\\\\psi}$, producing local feature map $C_{\\\\psi}(x)$ and global feature vector $E_{\\\\psi}(x)$. To maximize the lower bound of the InfoMax: $\\\\mathcal{I}\\\\big(C_{\\\\psi}(x),E_{\\\\psi}(x)\\\\big)$, the theoretical InfoMAX loss has been defined as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{aligned}\\n\\t\\t\\tL_{InfoMAX}(X)=&-\\\\mathbb{E}_{x\\\\in X}\\\\mathbb{E}_{i\\\\in\\\\mathcal{A}}\\\\big[\\n\\t\\t\\t\\\\log \\\\frac{\\\\exp \\\\left(g_{\\\\theta, \\\\omega}\\\\left(C_{\\\\psi}^{(i)}(x), E_{\\\\psi}(x)\\\\right)\\\\right)}{\\\\sum_{\\\\left(x^{\\\\prime}, i\\\\right) \\\\in X \\\\times \\\\mathcal{A}} \\\\exp \\\\left(g_{\\\\theta, \\\\omega}\\\\left(C_{\\\\psi}^{(i)}\\\\left(x^{\\\\prime}\\\\right), E_{\\\\psi}(x)\\\\right)\\\\right)}\\\\big],\\\\\\\\\\n\\t\\t\\tg_{\\\\theta, \\\\omega}&\\\\left(C_{\\\\psi}^{(i)}(x), E_{\\\\psi}(x)\\\\right)=\\\\phi_{\\\\theta}\\\\left(C_{\\\\psi}^{(i)}(x)\\\\right)^{T} \\\\phi_{\\\\omega}\\\\left(E_{\\\\psi}(x)\\\\right),\\n\\t\\t\\\\end{aligned}\\n\\t\\\\end{equation}\\n\\twhere $X=\\\\{x_1,\\\\cdots,x_N\\\\}$ is a set of random images and $\\\\mathcal{A}=\\\\{0,1,\\\\cdots,M^2-1\\\\}$ represents indices of a $M\\\\times M$ spatial sized local feature map. Based on this, positive sample pairs are $C^{(i)}_{\\\\phi}(x)$ and $E_{\\\\phi}(x)$, and negative sample pairs are $C^{(i)}_{\\\\phi}(x^{'})$ and $E_{\\\\phi}(x)$), where $x^{'}$ is a different image from $x$. SimCLR is another popular contrast learning framework, which applies two independent transformations, namely $t_1$ and $t_2$, to obtain the different views $v^{(1)},v^{(2)}=t_1(x),t_2(x)$. The loss function of SimCLR is defined as: \\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{aligned}\\n\\t\\t\\tL_{\\\\mathrm{SimCLR}}\\\\left({v}^{(1)}, {v}^{(2)}\\\\right)=\\\\frac{1}{N} \\\\sum_{i=1}^{N}\\\\left(L_{\\\\mathrm{NCE}}\\\\left({v}_{i}^{(1)} ;\\\\left[{v}^{(2)} ; {v}_{-i}^{(1)}\\\\right], s_{\\\\mathrm{SimCLR}}\\\\right)\\\\right),\\n\\t\\t\\\\end{aligned}\\n\\t\\\\end{equation}\\n\\twhere ${v}_{-i}:={v} \\\\backslash\\\\left\\\\{{v}_{i}\\\\right\\\\}$ and $s_{\\\\mathrm{SimCLR}}$ is defined as:\\n\\t\\\\begin{equation}\\n\\t\\ts_{\\\\text {SimCLR }}\\\\left({v}^{(1)}, {v}^{(2)} ; f, h\\\\right)=\\\\frac{h\\\\left(f\\\\left({v}^{(1)}\\\\right)\\\\right) \\\\cdot h\\\\left(f\\\\left({v}^{(2)}\\\\right)\\\\right)}{\\\\tau \\\\cdot\\\\left\\\\|h\\\\left(f\\\\left({v}^{(1)}\\\\right)\\\\right)\\\\right\\\\|_{2}\\\\left\\\\|h\\\\left(f\\\\left({v}^{(2)}\\\\right)\\\\right)\\\\right\\\\|_{2}}.\\n\\t\\\\end{equation}\\n\\tAs shown in Figure \\\\ref{fig:contrastive_self_supervised}, SimCLR defines more negative pairs to improve the sample utilization compared to InfoNCE. {However, SimCLR needs a large batch size to obtain some sufficiently rich negative samples (In , batch size is set to 4096). To alleviate the attachment of SimCLR to large batch size, MoCo introduce a negative queue to store and update negative samples.}\\n\\t\\\\begin{comment}\\n\\tSun et al. used the triplet matching objective as a pretext task for cGANs: pairs of images with the same category will have similar features and vice versa. Specifically, positive pairs contain the same category, such as two images with black hair; negative pairs contain different category, such as a image with black hair and a image with white hair. The loss function can be expressed as:\\n\\t\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\\\mathcal{L}_{D}=-\\\\lambda_{d}&\\\\mathop{\\\\mathbb{E}}\\\\limits_{x_a,x_p\\\\sim P_{X|y},x_n\\\\sim P_{X|y'}}\\\\bigg[\\\\log\\\\big(D_{mch}(x_a,x_p)\\\\big)\\\\\\\\\\n\\t&+\\\\log\\\\big(1-D_{mch}(x_a,x_n)\\\\big)\\\\bigg],\\\\\\\\\\n\\t\\\\mathcal{L}_{G}=-\\\\lambda_{g}&\\\\mathop{\\\\mathbb{E}}\\\\limits_{x_1,x_2\\\\sim P_{X|y},x_3\\\\sim P_{X|y'}}\\\\bigg[\\\\log\\\\big(D_{mch}(G(x_1,y),G(x_2,y))\\\\big)\\\\\\\\\\n\\t&+\\\\log\\\\big(1-D_{mch}(G(x_1,y),G(x_3,y'))\\\\big)\\\\bigg],\\n\\t\\\\end{aligned}\\n\\t\\\\end{equation}\\n\\twhere $y\\\\neq y'$, $x_a$ and $x_p$ are positive pairs of real images, while $x_a$ and $x_n$ are negative pairs of real images. \\n\\t\\\\end{comment}\\n\\tThe self-supervised methods mentioned above are also widely applied to the training of GANs. \\n\\tInspired by Deep InfoMax , Lee et al. propose InfoMax-GAN maximizing the mutual information between local and global features of real and fake images. The regularization of discriminator is expressed as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_\\\\mathrm{InfoMax-GAN}=\\\\lambda_d \\\\{L_{\\\\mathrm{InfoMAX}}(x_{r})+L_{\\\\mathrm{InfoMAX}}(x_{f}) \\\\}.\\n\\t\\\\end{equation}\\n\\twhere $x_r$ and $x_f$ represent sets of real and fake images, respectively.\\n\\tInspired by SimCLR , some studies introduce different data transformation techniques to create positive and negative pairs during GANs training. Zhao et al propose Cntr-GAN, where SimCLR loss is used to regularize the discriminator on two random augmented copies of both real and fake images. The regularization of discriminator for transformation $T$ is:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_{\\\\mathrm{Cntr-GAN}}=\\\\lambda_d \\\\{L_{\\\\mathrm{SimCLR}}(x_{r}, T(x_{r}))+L_{\\\\mathrm{SimCLR}}(x_{f}, T(x_{f})) \\\\}.\\n\\t\\t\\\\label{Eq:Cntr-GAN}\\n\\t\\\\end{equation}\\n\\tThey also compare the effect of different data transformation techniques (mentioned in Section 5.1) on Cntr-GAN. Figure 6 in shows the FID results of BigGAN adding SimCLR loss on CIFAR-10 dataset. The results illustrate that spatial transformations still work better than visual transformations and the best FID of 11.87 is achieved by applying adjusted SimCLR transformations with the cropping/resizing strength of 0.3. Although, regularization of auxiliary SimCLR loss improves GAN training, but does not outperform existing methods based on simple data augmentations, $e.g.$, bCR (demonstrated on Figure 2 in ). \\n\\tTo improve the efficiency of contrastive learning, Jeong et al. propose Contrastive Discriminator (ContraD), a way of training discriminators of GANs using improved SimCLR. Different from Cntr-GAN with SimCLR loss on both real and generated images, ContraD uses the SimCLR loss on the real images and the supervised contrastive loss on the generated images. Supervised contrastive loss adopts the contrastive between real and generated images, required as a GAN discriminator. More concretely, for two views $v^{(1)},v^{(2)}=t_1(x),t_2(x)$ with ${t}_{1}, {t}_{2} \\\\sim {T}$, the loss of real images are:\\n\\t\\\\begin{equation}\\n\\t\\tL_{\\\\text {con }}^{+}\\\\left(D, h_{{r}}\\\\right)=L_{\\\\mathrm{SimCLR}}\\\\left({v}_{{r}}^{(1)}, {v}_{{r}}^{(2)} ; D, h_{{r}}\\\\right),\\n\\t\\\\end{equation}\\n\\twhere $h_r$ is a projection head for this loss. However, the loss for generated images, an extended version of contrastive loss to support supervised learning by allowing more than one view to be positive. More concretely, they assume all the views from fake samples have the same label against those from real samples. Formally, for each $v_i^{(1)}$, the positive views are represented by $V_i^{(2)}$ that is a subset of $v^{(2)}$. The supervised contrastive loss is defined by:\\n\\t\\\\begin{equation}\\n\\t\\tL_{\\\\text {SupCon }}\\\\left({v}_{i}^{(1)}, {v}^{(2)}, V_{i+}^{(2)}\\\\right)=\\n\\t\\t-\\\\frac{1}{\\\\left|V_{i+}^{(2)}\\\\right|} \\\\sum_{{v}_{i+}^{(2)} \\\\in V_{i+}^{(2)}} \\\\log \\\\frac{\\\\exp \\\\left(s_{{SimCLR}}\\\\left({v}_{i}^{(1)}, {v}_{i+}^{(2)}\\\\right)\\\\right)}{\\\\sum_{j} \\\\exp \\\\left(s_{{SimCLR}}\\\\left({v}_{i}^{(1)}, {v}_{j}^{(2)}\\\\right)\\\\right)}.\\n\\t\\\\end{equation}\\n\\tUsing the notation,the ContraD loss for fake samples are:\\n\\t\\\\begin{equation}\\n\\t\\tL_{\\\\text {con }}^{-}\\\\left(D, h_{f}\\\\right)=\\n\\t\\t\\\\frac{1}{N} \\\\sum_{i=1}^{N} L_{\\\\text {SupCon }}\\\\left({v}_{f, i},\\\\left[{v}_{f,-i} ; {v}_{{r}}^{(1)} ; {v}_{{r}}^{(2)}\\\\right],\\\\left[{v}_{f,-i}\\\\right] ; D, h_{{f}}\\\\right),\\n\\t\\\\end{equation}\\n\\twhere $v_f= t_3(G(z))$ is a random view of fake samples ($t_3\\\\sim T$), and $v_{-i}=v\\\\backslash \\\\{v_i\\\\}$ is subset of $v$ that does not contain $v_i$. It is pertinent to note that authors also use an independent projection header $h_f$ in this loss instead of $h_r$ in $L_{\\\\text {con }}^{+}\\\\left(D, h_{{r}}\\\\right)$. {The adopted supervised contrastive learning on the fake images introduce the real/fake information into contrastive learning, which improves the efficiency of contrastive learning, thus improve the discrimination of the discriminator.}\\n\\tTo sum up, ContraD learns its contrastive representation by minimizing the following regularization loss:\\n\\t\\\\begin{equation}\\n\\t\\tL_{\\\\text {con }}\\\\left(D, h_{{r}}, h_{{f}}\\\\right)=L_{\\\\text {con }}^{+}\\\\left(D, h_{{r}}\\\\right)+\\\\lambda_{\\\\text {con }} L_{{con}}^{-}\\\\left(D, h_{{f}}\\\\right).\\n\\t\\\\end{equation}\\n\\tThe experimental results show that ContraD consistently improves the performance of GANs compared to other methods, such as Cntr-GAN, DiffAug, bCR, and CR. However, ContraD with different data transformations is not discussed further. \\n\\t{The achievement of above SimCLR-based contrastive learning methods depends on the sufficiently large batch size. However, large-scale GANs training often has only a small batch size for limited computational resources. Therefore, MoCo-based contrastive learning method, namely InsGen , has been introduced into GANs training. InsGen follows the MoCo-v2 to store the various negative samples with an extra queue. Furthermore, it also introduces a latent space augmentation for fake images. Combining with ADA and MoCo-based contrastive learning, InsGen has achieved state-of-the-art performance on a variety of datasets and training settings. Recently, Li et al. identify that only latent space augmentation for fake images brings the major performance improvement and contrastive learning in real images causes performance drop on limited data generation (DE-GANs). Based on this, they propose FakeCLR, which only applies contrastive learning on perturbed fake samples and devises three related training techniques. The experimental results manifest the new state of the arts in both few-shot generation and limited-data generation.}\\n\\tIn summary, contrastive self-supervised learning designs different positive and negative pairs and maximizes the mutual information of positive pairs according to the InfoNCE loss. Different from classification and segmentation tasks, two types of samples (real and fake images) exist for generating adversarial networks, which add more possibilities to the definition of positive and negative pairs. In the future, score-based contrastive lea rning may be proposed during the training of GANs. The summary of contrastive self-supervised regularization techniques of GANs is given in Table \\\\ref{table:self-supervised}.\",\n \"id\": \"aff71d0b-4dab-423d-9325-d61d96004987\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"4e2e9490-faa8-4a2f-9dbd-66ae6c225118\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Self-Supervision\"\n ],\n [\n \"subsubsection\",\n \"Contrastive Self-Supervised Learning (CSS)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Contrastive Self-Supervised Learning (CSS)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 130,\n 108,\n 125\n ],\n \"content\": \"{According to the perspective of \\\\textbf{\\\"Real $\\\\&$ Fake\\\"}, many regularization and normalization technologies inspired from supervised learning have been proposed to GANs training. The key point of them is improving the representation and generalizability of the discriminator. \\\\textit{Data Augmentation and Preprocessing} is a basic operation containing many types such as spatial augmentation, visual augmentation, frequency augmentation, and noise augmentation. Among them, combining adaptive strategies and all augmentation has achieved the most remarkable achievement and has been employed as default operations in most GANs training. \\\\textit{Consistency Regularization} and \\\\textit{Self-supervision} are designed additional tasks based on data augmentation, which further improve the efficiency of data augmentation and extract more useful information under stronger data augmentation beyond the existing yet limited practices. Currently, combining contrastive self-supervised learning with adaptive data augmentation has achieved state of the art in GANs training.}\",\n \"id\": \"2267f79d-3862-4c72-bdff-9670534c3017\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"8469161e-625b-422b-90e0-a6cca9638a04\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Real \\\\& Fake\\\"\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 64,\n 135,\n 136,\n 6997\n ],\n \"content\": \"From the perspective of \\\"Fitting distribution\\\", generator is considered as a distribution mapping function and the optimal discriminator is considered to be the distribution divergence. Wasserstein distance is a popular and important in GANs, and it corresponds to the optimal transport of the generator. To solve the dual problem of Wasserstein distance, Lipschitz continuity is introduced into the GANs training. The Wasserstein-based GANs (WGAN and WGAN-GP) have achieved remarkable results during the training. However, some studies suggest that the success of WGAN-GP is not due to the Wasserstein distance and the Lipschitz constraint of discriminator may improve the performance and stability of GANs training regardless of the statistical distance used as a loss function. Therefore, the Lipschitz continuity of discriminator is an essential condition during GANs training. {Weight clipping is a simple and the first solution to enforce a Lipschitz constraint, which clamps the weights of discriminator to a fixed box after each gradient update.} Furthermore, {gradient penalty}, {weight normalization}, and {weight regularization} are widely applied in GANs training for fulfilling Lipschitz continuity as summarized in subsequent subsections.\",\n \"id\": \"fdfb947c-602c-4735-9040-96d30810700e\",\n \"level\": \"section\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ]\n ],\n \"subsections\": [\n \"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c\",\n \"16b503b7-d01a-45ea-9b9c-0986b73e7409\",\n \"3e0461ca-996a-4c2c-8c3e-dbd291acb92b\",\n \"ee52e377-012a-4b04-8fa5-1df1ed551949\"\n ],\n \"title\": \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 137,\n 136,\n 88,\n 138,\n 93,\n 135,\n 8317,\n 6997\n ],\n \"content\": \"Gradient penalty is a simple and direct way to fulfill Lipschitz continuity. Specifically, K-Lipschitz continuity of the function $f$ can be accessed by $\\\\mathop{\\\\min}\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla f(\\\\hat{x})||_2-{\\\\rm K})^2$. According to the optimal transport theory mentioned on the Section \\\\ref{sect:A-1} of the Supplementary Online-only Material, gradient penalty can be used for the approximation of $W_c(\\\\mu,\\\\upsilon)$ in WGANs, named WGAN-GP . Specifically, WGAN-GP fulfills the 1-Lipschitz continuity of the discriminator by $\\\\mathop{\\\\min}\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-{\\\\rm 1})^2$, which limits the gradient of the discriminator to 1. Although WGAN-GP solves the instability of GANs training to some extent, the assumption of optimal transport is a constrained linear programming problem. Overly strict restriction reduces the exploratory of the discriminator. In contrast, the optimal transport with the regular term mentioned is an unconstrained optimization problem. Like optimal transport corresponds to 1-Lipschitz continuity, the optimal transport with the regular term corresponds to k-Lipschitz continuity (${\\\\rm k}\\\\leq1$) of the discriminator, named WGAN-LP , which is implemented by $\\\\mathop{\\\\min}\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[\\\\left(\\\\mathop{\\\\max}\\\\{0,||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1\\\\}\\\\right)^2\\\\right]$. WGAN-LP achieves better performance by using a weaker regularization term which enforces the Lipschitz constraint of the discriminator. \\n\\tWGAN-GP and WGAN-LP introduce Wasserstein distance into GANs framework. Due to the gap between limited input samples and the strict Lipschitz constraint on the whole input sample domain, the approximation of the Wasserstein distance is a challenging task. To this end, WGAN-div introduces a Wasserstein divergence into GANs training. The objective of WGAN-div can be smoothly derived as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathbb{E}_{y\\\\sim q(y)}[\\\\varphi(y)]-\\\\mathbb{E}_{x\\\\sim p(x)}[\\\\varphi(x)]+k \\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[ ||\\\\varphi(\\\\hat{x})||^p\\\\right].\\n\\t\\\\end{equation}\\n\\tThe objective of WGAN-div is similar to WGAN-GP and WGAN-LP. It can be considered as achieving 0-Lipschitz continuity of discriminator by adopting $\\\\mathop{\\\\min}\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^p\\\\right]$. \\n\\tGenerally, Wasserstein distance and Wasserstein divergence are reliable ways of measuring the difference between fake and real data distribution, which leads to the stable training of WGAN-based algorithms. However, a recent study shows that the c-transform method achieves better estimation of Wasserstein divergence but leads to worse performance compared to the gradient penalty method. The results demonstrate that the success of WGAN-based methodologies cannot truly be attributed to approximate the Wasserstein distance and the gradient penalty methods improve the performance indeed. Furthermore, some studies also demonstrate that gradient penalty methods of discriminator, such as 1-GP, k-GP (${\\\\rm k}\\\\leq 1$), and 0-GP stabilize the training and improve the performance of GANs remarkably regardless of the loss functions. Based on these observations, stabilizing GANs training using gradient penalty is widely applied in the research community for various losses of GANs. In the rest of this section, we discuss gradient penalty methods regardless of the loss function by dividing them into three parts: \\\\textit{1-GP}: $\\\\mathop{\\\\min}\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||-{\\\\rm 1})^p$, \\\\textit{k-GP (${\\\\rm k}\\\\leq1$)}: $\\\\mathop{\\\\min}\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[\\\\left(\\\\mathop{\\\\max}\\\\{0,||\\\\nabla D_{\\\\theta}(\\\\hat{x})||-1\\\\}\\\\right)^p\\\\right]$, and \\\\textit{0-GP}: $\\\\mathop{\\\\min}\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^p\\\\right]$, where $\\\\pi$ is the distribution of different image space (entire image space or part of image space) and $||\\\\cdot||$ represents the norm of the gradient. Generally, the loss function of the discriminator with GP can be formulated as:\\n\\t\\\\begin{equation}\\\\label{GP}\\n\\t\\t\\\\mathcal{L}_{D}=f(\\\\phi,\\\\theta)+\\\\lambda\\\\mathcal{L}_{GP},\\n\\t\\\\end{equation}\\n\\twhere $f(\\\\phi,\\\\theta)$ is the uniform loss function defined in Eq (\\\\ref{EQ:eqn1}) and $\\\\mathcal{L}_{GP}$ is the gradient penalty regularization.\",\n \"id\": \"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"fdfb947c-602c-4735-9040-96d30810700e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"Gradient Penalty\"\n ]\n ],\n \"subsections\": [\n \"258a5900-7326-4a5a-9ab5-4d4abbbbc5a5\",\n \"4a0b3b8a-89ed-4329-a9ef-855c13a16be6\",\n \"349a6a29-8d68-4bed-a022-d5ef2efaccc2\"\n ],\n \"title\": \"Gradient Penalty\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 141,\n 140,\n 139,\n 8317,\n 6997\n ],\n \"content\": \"Gulrajani et al. used 1-GP in WGAN-GP to train GANs. WGAN-GP uses the 2-norm gradient penalty across the entire image domain, which can be formulated as: \\n\\t\\\\begin{equation}\\\\label{wgan-gp}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1)^2,\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi$ is the distribution of entire image space approximated by the interpolation of real distribution ($p_r$) and generated distribution ($p_g$): $\\\\pi=t\\\\cdot p_r+(1-t)\\\\cdot p_g$ for $t\\\\sim U[0,1]$. Although, WGAN-GP stabilizes the training of GANs to a great extent, the overly strict gradient penalty limits the exploratory of discriminator. To loosen the penalty, many efforts of $\\\\pi$, $||\\\\cdot||$, and gradient direction are proposed. \\n\\tTo relax the image distribution, Kodali et al. track the training process of GANs and find that the decrease of the Inception Score (IS) is accompanied by a sudden change of the discriminator’s gradient around the real images. Authors propose DRAGAN by restricting the Lipschitz constant around the real images $\\\\pi=p_r+\\\\epsilon$, where $\\\\epsilon\\\\sim N_d(0,cI)$.\\n\\tIn order to relax the gradient direction, Zhou et al. argue that restricting the global Lipschitz constant is unnecessary. Therefore, only maximum gradient is necessary to be penalized:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\left(\\\\mathop{\\\\max}\\\\limits_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1\\\\right)^2,\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi=t\\\\cdot p_r+(1-t)\\\\cdot p_g$; Furthermore, inspired by Virtual Adversarial Training (VAT) , D{\\\\'a}vid et al. propose a method, called Adversarial Lipschitz Regularization (ALR), which restricts the 1-Lipschitz continuity at $\\\\pi=p_r\\\\cup p_g$ in the direction of adversarial perturbation. {Adversarial perturbation direction is the most unstable direction, Restricting the 1-Lipschitz continuity to the adversarial direction means restricting only the largest Lipschitz constant, which is simpler and more efficient than the previous method.} The proposed ALP shows the SOTA performance in terms of Inception Score and Fréchet Inception Distance among non-progressive growing methods trained on CIFAR-10 dataset.\\n\\tContrary to the methods which penalize the gradient in Euclidean space, Adler et al. extended the $L_p(p=2)$ space with gradient penalty to Banach space that contains the $L_p$ space and Sobolev space. For the Banach space B, the Banach norm $||.||_{B^*}$ is defined as: \\n\\t\\\\begin{equation}\\n\\t\\t||x^*||_{B^*}=\\\\sup\\\\limits_{x\\\\in B}\\\\frac{x^*(x)}{||x||_B}.\\n\\t\\\\end{equation}\\n\\tThus, the gradient penalty of Banach wasserstein GAN can be expressed as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_{B^*}-1)^2,\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi=t\\\\cdot p_r+(1-t)\\\\cdot p_g$. {Banach wasserstein GAN expands the Lipschitz continuity into Banach space containing both $L_p$ space and Sobolev space, which has more restriction than wasserstein GAN.}\",\n \"id\": \"258a5900-7326-4a5a-9ab5-4d4abbbbc5a5\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"Gradient Penalty\"\n ],\n [\n \"subsubsection\",\n \"1-GP\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"1-GP\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 88,\n 142,\n 8317\n ],\n \"content\": \"\\\\leq1$)}\\n\\t{k-GP (${\\\\rm k}\\\\leq1$) was first tested by Gulrajani et al. and named one sided gradient penalty. It uses the 2-norm gradient penalty across the entire image domain, which is formulated as: }\\n\\t\\\\begin{equation}\\\\label{wgan-lp}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[\\\\left(\\\\mathop{\\\\max}\\\\{0,||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1\\\\}\\\\right)^2\\\\right],\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi$ is the distribution of entire image space approximated by the interpolation of real distribution ($p_r$) and generated distribution ($p_g$): $\\\\pi=t\\\\cdot p_r+(1-t)\\\\cdot p_g$ for $t\\\\sim U[0,1]$. Inspired by the optimal transport with the regular term, Petzka et al. also used k-GP (${\\\\rm k}\\\\leq1$) to training GANs named WGAN-LP. Furthermore, Xu et al show a more general dual form of the Wasserstein distance compared to KR duality (mentioned in section 2.4), named Sobolev duality, which relaxes the Lipschitz constraint but still maintains the favorable gradient\\n\\tproperty of the Wasserstein distance. Authors also show that the KR duality is a special case of the proposed Sobolev duality. Based on the Sobolev duality, the relaxed gradient penalty of the proposed SWGAN is formulated as:\\n\\t\\\\begin{equation}\\\\label{wgan-sp}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[\\\\left(\\\\mathop{\\\\max}\\\\{0,||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^2-1\\\\}\\\\right)^2\\\\right],\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi=t\\\\cdot p_r+(1-t)\\\\cdot p_g$ for $t\\\\sim U[0,1]$. It is clear that above three method have the same form of gradient penalty. Interestingly, different relaxation methods yield the same form of regularization.\",\n \"id\": \"4a0b3b8a-89ed-4329-a9ef-855c13a16be6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"Gradient Penalty\"\n ],\n [\n \"subsubsection\",\n \"k-GP (${\\\\rm k\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"k-GP (${\\\\rm k\"\n },\n {\n \"cite_extract_rate\": 0.8571428571428571,\n \"cites\": [\n 142,\n 136,\n 143,\n 8317,\n 60,\n 7002,\n 88,\n 7001,\n 99,\n 94,\n 6997,\n 140,\n 8322,\n 141,\n 92,\n 64,\n 93,\n 135\n ],\n \"content\": \"\\\\begin{table}\\n\\t\\t\\\\caption{The Gradient penalty of the Discriminator. $\\\\mu$ and $v$ are real and generated distribution, respectively.}\\n\\t\\t\\\\footnotesize\\n\\t\\t\\\\label{table:gradient penalty}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{c | c | c | c }\\t\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\t{ Method}& { $\\\\mathcal{L}_{GP}$}& $\\\\pi$&Lipschitz continuity\\t \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tWGAN-GP &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1)^2$&$t\\\\cdot p_r+(1-t)\\\\cdot p_g$&$||D_{\\\\theta}||_{Lip}\\\\to1$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tDRAGAN &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1)^2$&$p_r+\\\\epsilon$&$||D_{\\\\theta}||_{Lip}\\\\to1$\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMax-GP &$\\\\left(\\\\mathop{\\\\max}\\\\limits_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1\\\\right)^2$&$t\\\\cdot p_r+(1-t)\\\\cdot p_g$&$||D_{\\\\theta}||_{Lip}\\\\to1$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tALP &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1)^2$&$p_r\\\\cup p_g$&$||D_{\\\\theta}||_{ALP-Lip}\\\\to1$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tBanach-GP &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}(||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_{B^*}-1)^2$&$t\\\\cdot p_r+(1-t)\\\\cdot p_g$&$||D_{\\\\theta}||_{Lip}\\\\to1$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tWGAN-LP &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[\\\\left(\\\\mathop{\\\\max}\\\\{0,||\\\\nabla D_{\\\\theta}(\\\\hat{x})||_2-1\\\\}\\\\right)^2\\\\right]$&$t\\\\cdot p_r+(1-t)\\\\cdot p_g$&$||D_{\\\\theta}||_{Lip}\\\\leq1$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSWGAN &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}\\\\left[\\\\left(\\\\mathop{\\\\max}\\\\{0,||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^2-1\\\\}\\\\right)^2\\\\right]$&$t\\\\cdot p_r+(1-t)\\\\cdot p_g$&$||D_{\\\\theta}||_{Lip}\\\\leq1$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tzc-GP &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^2_2$&$p_r\\\\cup p_g$&$||D_{\\\\theta}||_{Lip}\\\\to0$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tGAN-QP &$\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{x_r,x_g\\\\sim\\\\pi}\\\\frac{\\\\left(D_{\\\\theta}(x_r)-D_{\\\\theta}(x_f)\\\\right)^2}{||x_r-x_f||}$&$\\\\pi=p_r\\\\cdot p_g$&$\\\\frac{\\\\left(D_{\\\\theta}(x_r)-D_{\\\\theta}(x_f)\\\\right)^2}{||x_r-x_f||}\\\\to 0$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tZP-Max &$\\\\mathop{\\\\max}\\\\limits_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^2_2$&$t\\\\cdot p_r+(1-t)\\\\cdot p_g$&$||D_{\\\\theta}||_{Lip}\\\\to0$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tZP &$\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^2_2$&$t\\\\cdot p_r+(1-t)\\\\cdot p_g$&$||D_{\\\\theta}||_{Lip}\\\\to0$\\\\\\\\\\n\\t\\t\\t\\\\bottomrule\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table} \\n\\tWu et al. used 0-GP, and proposed Wasserstein divergence. According to , Wasserstein divergence is solved by minimizing:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\label{0-gp}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla D_{\\\\theta}(\\\\hat{x})||^2,\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi$ is both the real distribution ($p_r$) and the generated distribution ($p_g$): $\\\\pi=p_r\\\\cup p_g$. Furthermore, Mescheder et al. also demonstrate that the optimization of unregularized GAN is not always locally convergent and some simplified zero centered gradient penalty (zc-GP) techniques, implemented by minimizing Eq (\\\\ref{0-gp}), can be used to achieve local convergence of GANs. Li et al. introduce the adversarial training to discriminator training, which is turned out to be an adaptive 0-GP. \\n\\tBesides, some other 0-GP methods are derived by different theoretical derivations. For instance, Su et al. propose a Quadratic Potential (QP) for GANs training with the following formulation:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{x_r,x_g\\\\sim\\\\pi}\\\\frac{\\\\left(D_{\\\\theta}(x_r)-D_{\\\\theta}(x_f)\\\\right)^2}{||x_r-x_f||},\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi$ is the joint distribution of the real and generated distributions: $\\\\pi=p_r\\\\cdot p_g$; Zhang et al. combine a Total Variational (TV) regularizing term into the training of GANs, that is $|D_{\\\\theta}(x_r)-D_{\\\\theta}(x_f)-\\\\delta|$. According to , the TV term can be approximated by Eq (\\\\ref{0-gp}), which is exhilarating; Zhou et al. propose the Lipschitz GANs, with the maximum of the gradients penalty for guaranteeing the gradient informativeness:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathop{\\\\max}\\\\limits_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla f(\\\\hat{x})||^2_2,\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi=t\\\\cdot p_r+(1-t)\\\\cdot p_g$; Thanh-Tung et al. also propose the 0-GP with gradients penalty at $\\\\pi=t\\\\cdot p_r+(1-t)\\\\cdot p_g$:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathcal{L}_{GP}=\\\\mathbb{E}_{\\\\hat{x}\\\\sim\\\\pi}||\\\\nabla f(\\\\hat{x})||^2_2.\\n\\t\\\\end{equation}\\n\\tIn summary, gradient penalty techniques are widely used in the GANs training to achieve Lipschitz continuity of discriminator. As shown in Table \\\\ref{table:gradient penalty}, many techniques are proposed based on different theories and phenomena. But to the best of our knowledge, there is no fair and comprehensive work comparing the performance of these gradient penalty methods.\\n\\t\\\\begin{comment}\\n\\tThe summary of the gradient penalty of the discriminator is shown in TABLE \\\\ref{table:gradient penalty}. Based on the summary and the 1-Lipschitz continuity in WGAN-GP, the gradient penalty has been improved in two directions.\\n\\t\\\\begin{enumerate}\\n\\t\\\\item[*] Limit the Lipschitz constant so that let the $||D||_{Lip}\\\\leq 1$ or $||D||_{Lip}\\\\to 0$.\\n\\t\\\\item[*] Explore the suitable scope for Lipschitz continuity.\\n\\t\\\\end{enumerate}\\n\\t\\\\end{comment}\\n\\tTo compare the performance of various methods intuitively, a comparative experiment on CIFAR-10 and CIFAR-100 datasets is conducted\\\\footnote{The base framework comes from wgan-gp in \\\\url{https://github.com/kwotsin/mimicry}}. The results of FID for various gradient penalty methods with different loss functions are shown in Table \\\\ref{table: experiments results}. The results validate the conclusion in studies , that the Lipschitz constraint of discriminator may improve the performance and stability of GANs training regardless of the statistical distance used as a loss function. All gradient penalty methods improve the performance of GANs upon all three loss functions. Among them, zc-GP obtains the best performance and is widely used in SOTA methods as illustrated in Table \\\\ref{table:SOTA }.\\n\\t\\\\begin{comment}\\n\\t\\\\begin{table}\\n\\t\\\\caption{The experimental results of different Gradient penalty}\\n\\t\\\\label{table: experiments results}\\n\\t\\\\setlength{\\\\tabcolsep}{1mm}\\n\\t\\\\centering\\n\\t\\\\begin{tabular}{c c c }\\t\\n\\t\\\\toprule\\n\\t\\\\midrule\\n\\t{ Method}&Inception Score& FID \\\\\\\\\\n\\t\\\\hline\\n\\tWGAN-GP &7.869&16.62\\\\\\\\\\n\\tWGAN-LP &7.98&15.85\\\\\\\\\\n\\tWGAN-ALP &8.247&14.19\\\\\\\\\\n\\tWGAN-GP-Max &7.956&18.43\\\\\\\\\\n\\tWGAN-ZP-Max &7.908&17.97\\\\\\\\\\n\\tWGAN-ZP-Sample &8.013&15.87\\\\\\\\\\n\\tWGAN-ZP &7.957&16.08\\\\\\\\\\n\\t\\\\bottomrule\\n\\t\\\\end{tabular}\\n\\t\\\\end{table} \\n\\t\\\\end{comment}\\n\\t\\\\begin{table*}\\n\\t\\t\\\\tiny\\n\\t\\t\\\\caption{ FID results on the CIFAR-10 and CIFAR-100 datasets for various gradient penalty methods with different GAN losses.}\\n\\t\\t\\\\centering\\n\\t\\t\\\\label{table: experiments results}\\n\\t\\t\\\\begin{tabular}{cccccccccc}\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\multirow{2}{*}{Dataset}&\\\\multirow{2}{*}{Loss} & \\\\multicolumn{8}{c}{\\\\makecell*[c]{Gradient Penalty Methods}} \\\\\\\\ \\n\\t\\t\\t\\\\cline{3-10} \\n\\t\\t\\t&& \\\\makecell*[c]{None}&GP&DRAGAN &MAX-GP &LP&zc-GP &ZP-MAX &ZP \\\\\\\\ \\n\\t\\t\\t\\\\midrule\\n\\t\\t\\t\\\\multirow{3}{*}{CIFAR-10}&\\n\\t\\t\\tGAN& 42.41& 23.45&20.98& 26.65&22.9&\\\\textbf{19.39}&24.38&23.96 \\\\\\\\\\n\\t\\t\\t&WGAN&290&30.38& 29.53&37.21&28.31&\\\\textbf{26.99}&31.28&30.19\\\\\\\\\\n\\t\\t\\t&Hinge&58.34&21.19&21.77&25.4&20.79&\\\\textbf{18.75}&23.1&22.58\\\\\\\\\\n\\t\\t\\t\\\\cline{1-10} \\n\\t\\t\\t\\\\specialrule{0em}{2pt}{2pt}\\n\\t\\t\\t\\\\multirow{3}{*}{CIFAR-100}&\\n\\t\\t\\tGAN&44.5&25.76& 25.37&24.29&23.82&\\\\textbf{21.81}&26.27&25.38 \\\\\\\\\\n\\t\\t\\t&WGAN&244&32.28&31.93&38.71&32.19&\\\\textbf{29.12}&39.75&37.8 \\\\\\\\\\n\\t\\t\\t&Hinge&59.43&25.13&25.42&28.34&23.67&\\\\textbf{21.55}&26.19&26.06\\\\\\\\\\n\\t\\t\\t\\\\bottomrule\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table*}\",\n \"id\": \"349a6a29-8d68-4bed-a022-d5ef2efaccc2\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"Gradient Penalty\"\n ],\n [\n \"subsubsection\",\n \"0-GP\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"0-GP\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"WGAN is a popular and important generative adversarial network. From the optimal transport introduced on the Section \\\\ref{sect:A-1} of the Supplementary Online-only Material, to obtain $W_c(\\\\mu,\\\\upsilon)=\\\\mathop{\\\\max}\\\\limits_{||\\\\varphi_\\\\xi||_L\\\\leq 1}\\\\left\\\\{\\\\int_X \\\\varphi_\\\\xi \\\\mathrm{d}\\\\mu\\\\ -\\\\int_Y \\\\varphi_\\\\xi \\\\mathrm{d}\\\\upsilon\\\\right\\\\}$, the discriminator must satisfy the 1-Lipschitz continuity. According to the Section \\\\ref{sect:A-3} in the Supplementary Online-only Material, the spectral norm $||W||_2$ can be used to represent the Lipschitz constant $\\\\mathrm{C'}$. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately. Hence, \\\\textit{Weight Normalization} and \\\\textit{Weight Regularization} can also be used to enable the Lipschitz continuity of the discriminator.\",\n \"id\": \"16b503b7-d01a-45ea-9b9c-0986b73e7409\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"fdfb947c-602c-4735-9040-96d30810700e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"Weight Normalization and Weight Regularization\"\n ]\n ],\n \"subsections\": [\n \"930186fd-4bd2-4fae-a134-426d00072853\",\n \"9c9a2b69-11c0-4e4b-90b7-101fced8cb32\"\n ],\n \"title\": \"Weight Normalization and Weight Regularization\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 78,\n 144\n ],\n \"content\": \"Spectral norm of the weight and the Lipschitz constant express the same concept. Therefore, weight normalization is another method to achieve Lipschitz continuity. {More important, weight normalization methods are Non-sampling-based, which don’t have the lack of support problem in contrast to gradient penalties.} Spectral normalization of the weight limits the Lipschitz constant to 1. Certainly, upper bound of the spectral norm can be used to normalize the weights, achieving $k (k\\\\leq1)$ Lipschitz continuity. The following lemmas put forward some upper bounds of the spectral norm.\\n\\t\\\\newenvironment{lemma3}{{\\\\indent\\\\it \\\\textbf{Lemma 3.1:}}}{\\\\hfill \\\\par}\\n\\t\\\\begin{lemma3}\\n\\t\\t\\\\textit{\\n\\t\\t\\tIf $\\\\lambda_1\\\\leq\\\\lambda_2\\\\leq\\\\cdots\\\\leq\\\\lambda_M$\\n\\t\\t\\tare the eigenvalues of the $W^\\\\top W$, then the spectral norm$||W||_2=\\\\sqrt{\\\\lambda_M}$; The Frobenius norm$||W||_F=\\\\sqrt{\\\\sum_{i=1}^{M}\\\\lambda_i}$ \\n\\t\\t}\\n\\t\\\\end{lemma3}\\n\\t\\\\newenvironment{lemma4}{{\\\\indent\\\\it \\\\textbf{Lemma 3.2:}}}{\\\\hfill \\\\par}\\n\\t\\\\newenvironment{lemma5}{{\\\\indent\\\\it \\\\textbf{Lemma 3.3:}}}{\\\\hfill \\\\par}\\n\\t\\\\newenvironment{proof3}{{\\\\indent\\\\it Proof 3.1:}}{\\\\hfill $\\\\square$\\\\par}\\n\\t\\\\newenvironment{proof4}{{\\\\indent\\\\it Proof 3.2:}}{\\\\hfill $\\\\square$\\\\par}\\n\\t\\\\newenvironment{proof5}{{\\\\indent\\\\it Proof 3.3:}}{\\\\hfill $\\\\square$\\\\par}\\n\\t\\\\begin{proof3}\\n\\t\\tSee and \\n\\t\\\\end{proof3}\\n\\t\\\\begin{lemma4}\\n\\t\\t\\\\textit{\\n\\t\\t\\tFor a $n\\\\times m$ matrix, $||W||_1=\\\\mathop{\\\\max}\\\\limits_{j}\\\\sum_{i=1}^{n}|a_{i,j}|$, $ ||W||_\\\\infty=\\\\mathop{\\\\max}\\\\limits_{i}\\\\sum_{j=1}^{m}|a_{i,j}|$, then $||W||_2\\\\leq\\\\sqrt{||W||_1||W||_\\\\infty}$\\n\\t\\t}\\n\\t\\\\end{lemma4}\\n\\t\\\\begin{proof4}\\n\\t\\tSee \\n\\t\\\\end{proof4}\\n\\t\\\\begin{lemma5}\\n\\t\\t\\\\textit{\\n\\t\\t\\tFor a $n\\\\times m$ matrix, \\n\\t\\t\\t$||W||_F=\\\\sqrt{\\\\left(\\\\sum_{j=1}^{m}\\\\sum_{i=1}^{n}|a_{i,j}|^2\\\\right)}$, then $||W||_2\\\\leq||W||_F$\\n\\t\\t}\\n\\t\\\\end{lemma5}\\n\\t\\\\begin{proof5}\\n\\t\\tSee \\n\\t\\\\end{proof5}\\n\\t1-Lipschitz continuity can be expressed by the spectral normalization. Miyato et al. control the Lipschitz constant through spectral normalization $W_\\\\sigma=\\\\frac{W}{||W||_2}$ of each layer for D, leading to a better result than WGAN-GP. {Practically, the power iteration method is used as a fast approximation for the spectral norm ($||W||_2$).} Similarly, according to the optimal transport with regular term, Lipschitz constant of discriminator should be less than or equal to 1. Correspondingly, upper bound of the spectral norm can be utilized to normalize the weight ($||W_\\\\sigma||_2\\\\leq1$), achieving $k (k\\\\leq1)$ Lipschitz continuity. In terms of Lemma 3.2 and Lemma 3.3, $\\\\sqrt{||W||_1||W||_\\\\infty}$ and Frobenius norm ($||W||_F$) are simple upper bound of the spectral norm ($||W||_2$) and can be used to normalize the weight. For example, Zhang et al. use the $\\\\sqrt{||W||_1||W||_\\\\infty}$, seeking for an approximation of the spectral norm that is easy to calculate. Miyato et al. explain that the Frobenius norm is a restriction on all eigenvalues. It is different from the spectral norm, which only constrains the maximum eigenvalue. Authors conjecture that Frobenius normalization affects the network's ability to express, but no experiments are reported to compare it with the spectral normalization. Liu et al. find that the mode collapse is often accompanied by the collapse of the eigenvalue of the discriminator. Because the spectral normalization only limits the maximum eigenvalue, and the eigenvalue collapse means the remaining eigenvalues suddenly decrease. Therefore, authors adopt the following methods to prevent the collapse of the eigenvalues:\\n\\t\\\\begin{equation}\\n\\t\\tW_{\\\\sigma}=\\\\frac{W+\\\\nabla W}{||W||_2}=\\\\frac{W}{||W||_2}+\\\\frac{\\\\nabla W}{||W||_2}.\\n\\t\\\\end{equation}\\n\\tThe results demonstrate that this method effectively prevents mode collapses. Although the experiments are reported in this study, but it misses theoratical proofs. Therefore the relationship between the matrix eigenvalues and GAN performance is not clear.\\n\\tFew researches focus on weight normalization as demonstrated in Table \\\\ref{table:norm normalization}. Among these studies, spectral normalization is widely applied in some SOTA methods, as demonstrated in Section 7.\\n\\t\\\\begin{table}\\n\\t\\\\caption{The summary of the weight normalization and wight regularization.}\\n\\t\\\\label{table:norm normalization}\\n\\t\\\\centering\\n\\t\\\\begin{tabular}{c | c | c }\\t\\n\\t\\t\\\\toprule\\n\\t\\t\\\\midrule\\n\\t\\t{ Method}&Implementation &Motivation\\t \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tSpectral normalization (SN) &$W_\\\\sigma=W/||W||_2$&$||D||_{Lip}\\\\to1$\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tF normalization &$W_\\\\sigma=W/||W||_F$&$||D||_{Lip}\\\\leq1$\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tMixed normalization &$W_\\\\sigma=W/\\\\sqrt{||W||_1||W||_\\\\infty}$&$||D||_{Lip}\\\\leq1$\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tSpectral increment normalization &$W_\\\\sigma=W/||W||_2+\\\\nabla W/||W||_2$&$||D||_{Lip}\\\\to1$\\\\\\\\\\n\\t\\t\\\\bottomrule\\n\\t\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"930186fd-4bd2-4fae-a134-426d00072853\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"16b503b7-d01a-45ea-9b9c-0986b73e7409\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"Weight Normalization and Weight Regularization\"\n ],\n [\n \"subsubsection\",\n \"Weight Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Weight Normalization\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 65\n ],\n \"content\": \"Compared with spectral normalization similar to 1-GP, spectral regularization is similar to the 0-GP. Kurach et al. use the $\\\\mathcal{L}_R=||W||_2$ to regularize the loss function. Zhou et al. also use the $L_P$-norm ($P=1,F,\\\\infty$) to regularize the discriminator. However, these studies have worse performance than weight normalization and did not catch much attention among researchers.\",\n \"id\": \"9c9a2b69-11c0-4e4b-90b7-101fced8cb32\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"16b503b7-d01a-45ea-9b9c-0986b73e7409\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"Weight Normalization and Weight Regularization\"\n ],\n [\n \"subsubsection\",\n \"Weight Regularization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Weight Regularization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 145,\n 146\n ],\n \"content\": \"}\\n\\tGradient normalization is also a popular method to impose the Lipschitz constraint on the discriminator. As we all know, 1-Lipschitz constraint can be implemented by $\\\\mathop{\\\\min}\\\\mathbb{E}_{{x}\\\\sim\\\\pi}(||\\\\nabla_{x} D_{\\\\theta}({x})||_2-1)^2$ to let the gradient of the discriminator ($||\\\\nabla_{{x}} D_{\\\\theta}({x})||_2$) equal to 1. Therefore, control the Lipischitz constant through gradient normalization $\\\\hat{D}_{\\\\theta}(x)=\\\\frac{D_{\\\\theta}(x)}{||\\\\nabla_{{x}}D_{\\\\theta}({x})||_2}$ for $D_{\\\\theta}$. Accordingly, the gradient of $\\\\hat{D}_{\\\\theta}$ can be represented as $||\\\\nabla_{{x}} \\\\hat {D}_{\\\\theta}({x})||_2=||\\\\nabla_{{x}}\\\\left(\\\\frac{D_{\\\\theta}(x)}{||\\\\nabla_{{x}}D_{\\\\theta}({x})||_2}\\\\right)||_2$, equaling to 1. To ensure the boundedness, different studies have different implementation. For instance, adopts $\\\\hat{D}_{\\\\theta}(x)=\\\\frac{D_{\\\\theta}(x)}{||\\\\nabla_{{x}}D_{\\\\theta}({x})||_2+D_{\\\\theta}(x)}$ and adopts $\\\\hat{D}_{\\\\theta}(x)=\\\\frac{D_{\\\\theta}(x)}{||\\\\nabla_{{x}}D_{\\\\theta}({x})||_2+\\\\epsilon}$. Extensive experiments demonstrate that both implementation of gradient normalization attain significant performance gains comparing to gradient penalty, weight normalization, and weight regularization.\",\n \"id\": \"3e0461ca-996a-4c2c-8c3e-dbd291acb92b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"fdfb947c-602c-4735-9040-96d30810700e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"{Gradient Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Gradient Normalization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 145\n ],\n \"content\": \"}\\n\\t{As mentioned above, weight clipping, gradient penalty, weight regularization, weight normalization, and gradient normalization all could enable Lipschitz continuity of the discriminator. However, what are the advantages and disadvantages of investigated techniques? Imposing the Lipschitz constraint on the discriminator can be characterized by three properties . 1) \\\\textit{model-} or \\\\textit{module-wise} \\\\textit{constraint}. Model-wise constraint is defined as methods that constraint objective depends on full model, while module-wise constraint is defined as methods that constraint objective depends on layers. Generally, model-wise constraint is better since module-wise constraint is strict, which limits\\n\\t\\tthe layer capacities and reduces the power of discriminator. 2) \\\\textit{sampling-based} or \\\\textit{non-sampling} \\\\textit{-based} \\\\textit{constraint}. Sampling-based constraint is defined as requiring sampling data during usage, while non-sampling-based constraint depends on the model, not data sampling. Generally, non-sampling-based constraint performs better since Lipschitz constraint should be fulfilled on the entire data manifold, not only sampling data. 3) \\\\textit{Hard} or \\\\textit{soft constraint}. {The accurate constraint of Lipschitz continuity is defined as hard constraint and the converse to be soft constraint. Hard constraint has achieved the exact Lipschitz continuity through limiting the spectral norm, which is expected to perform better. While soft constraint only obtain the Lipschitz continuity approximatively through optimization.} Table \\\\ref{table:lipschitz continuity} summarizes the properties of different technologies, from which gradient normalization is a model-wise, non-sampling-based, and hard constraint method.}\\n\\t\\\\begin{table}\\n\\t\\t\\\\caption{Summary of different regularization and normalization technologies for imposing Lipschitz continuity.}\\n\\t\\t\\\\label{table:lipschitz continuity}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{c | c | c | c }\\t\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\tMethod& Model-wise&Non-sampling-based&Hard\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tWeight Clipping&&\\\\checkmark&\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tGradient Penalty&\\\\checkmark&&\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tWeight Regularization&&\\\\checkmark&\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tWeight Normalization&&\\\\checkmark&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tGradient Normalization&\\\\checkmark&\\\\checkmark&\\\\checkmark\\\\\\\\\\n\\t\\t\\t\\\\bottomrule\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table}\",\n \"id\": \"ee52e377-012a-4b04-8fa5-1df1ed551949\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"fdfb947c-602c-4735-9040-96d30810700e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Fitting distribution\\\" \"\n ],\n [\n \"subsection\",\n \"{Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Summary\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 98,\n 54,\n 99,\n 7190,\n 8323\n ],\n \"content\": \"Assuming the objectives of GANs are convex-concave, some studies have proposed the global convergence of GANs . However, these theoretical convergence analyses are only applicable to the GANs with the optimal discriminator. Therefore, some studies focus on analyzing the local convergence of GANs. According to Nagarajan et al. and Mescheder et al. , under some assumptions, GANs dynamics are locally convergent. However, if these assumptions are not satisfied, especially if the data distributions are not continuous, GANs dynamics do not always converge locally unless some regularization techniques are used.\\n\\tWe review {Jacobian regularization} techniques in this section, which minimize the Jacobian matrix to achieve local convergence. With the same motivation, Mescheder et al. propose a simplified gradient penalties method, named zero-centered gradient penalties (zc-GP), that guarantees the local convergence under suitable assumptions. Since it is similar to 0-GP, we cover it in Section 4.\",\n \"id\": \"44c32216-8be6-4e5c-aa2f-ca241ea4139f\",\n \"level\": \"section\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Training dynamics\\\"\"\n ]\n ],\n \"subsections\": [\n \"f5ab6a67-88dd-45b6-aeae-1b85b061e943\",\n \"8377346c-065d-4bc5-adf5-275fad74664d\"\n ],\n \"title\": \"Regularization and Normalization of \\\"Training dynamics\\\"\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 74,\n 149,\n 7003,\n 98,\n 148,\n 99,\n 147,\n 7190\n ],\n \"content\": \"In \\\\textit{Proposition 2.2} of Section 2: absolute values of all eigenvalues of the Jacobian matrix ($v^{'}(\\\\phi,\\\\theta)$) are expected to be less than 1 at the fixed point, which is equivalent to the real part of the eigenvalue being negative. Additionally, the learning rate must be relatively low . To meet these requirements, Mescheder et al. used the Consensus Optimization (ConOpt) to make the real part of the eigenvalue negative. Its regularized updates are:\\n\\t\\\\begin{equation} \\\\label{eqn1}\\n\\t\\t\\\\begin{split}\\n\\t\\t\\t\\\\phi^{(k+1)}=\\\\phi^{(k)}+h\\\\nabla_\\\\phi\\\\left(- f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\gamma L(\\\\phi^k,\\\\theta^k)\\\\right),\\\\\\\\\\n\\t\\t\\t\\\\theta^{(k+1)}=\\\\theta^{(k)}+h\\\\nabla_\\\\theta\\\\left(f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\gamma L(\\\\phi^k,\\\\theta^k)\\\\right),\\n\\t\\t\\\\end{split}\\n\\t\\\\end{equation}\\n\\twhere $L(\\\\phi^k,\\\\theta^k)=\\\\frac{1}{2}||v(\\\\phi^k,\\\\theta^k)||^2=\\\\frac{1}{2}\\\\left(||\\\\nabla_\\\\phi f(\\\\phi^k,\\\\theta^k)||^2+||\\\\nabla_\\\\theta f(\\\\phi^k,\\\\theta^k)||^2\\\\right)$ is the regularization of the Jacobian matrix.\\n\\tApart from , Nagaraja et al. also analyze the relationship between local convergence of GANs and all eigenvalues of the Jacobian\\n\\tof the gradient vector field. Authors prove the local convergence for absolutely continuous generator and data distributions under certain regularity assumptions. This requires the loss function of the GANs to be strictly concave, which is not the case for some GANs. Based on this, a simple regularization technology that regularized the generator using the gradient of the discriminator is proposed by Nagaraja et al. . The regularized updates for the generator can be expressed as:\\n\\t\\\\begin{equation} \\\\label{eqn1}\\n\\t\\t\\\\phi^{(k+1)}=\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\phi||\\\\nabla_\\\\theta f(\\\\phi^{k},\\\\theta^{k})||^2.\\n\\t\\\\end{equation}\\n\\tHerein, the update of the discriminator is similar to SimGD. Furthermore, Nie et al. propose a method that only regularizes the discriminator. The regularized update of the discriminator in this case is given by:\\n\\t\\\\begin{equation} \\\\label{eqn1}\\n\\t\\t\\\\theta^{(k+1)}=\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\theta||\\\\nabla_\\\\phi f(\\\\phi^{k},\\\\theta^{k})||^2.\\n\\t\\\\end{equation}\\n\\tThe update of the generator is the same as SimGD. Nie et al. propose JAcobian REgularization (JARE) that regularizes both the generator and the discriminator. The regularized updates for the generator and the discriminator are:\\n\\t\\\\begin{equation} \\\\label{eqn1}\\n\\t\\t\\\\begin{split}\\n\\t\\t\\t\\\\phi^{(k+1)}=\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\phi||\\\\nabla_\\\\theta f(\\\\phi^{k},\\\\theta^{k})||^2,\\\\\\\\\\n\\t\\t\\t\\\\theta^{(k+1)}=\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\theta||\\\\nabla_\\\\phi f(\\\\phi^{k},\\\\theta^{k})||^2.\\n\\t\\t\\\\end{split}\\n\\t\\\\end{equation}\\n\\tThe key difference between JARE and ConOpt is that JARE does not contain the Hessians $\\\\nabla^2_{\\\\phi,\\\\phi}f(\\\\phi^{k},\\\\theta^k)$ and $\\\\nabla^2_{\\\\theta,\\\\theta}f(\\\\phi^{k},\\\\theta^k)$ in the regularization term. \\n\\t{There are several Jacobian regularization methods that have been proposed to deal with the training instabilities of GANs. What is the difference of them? Nie et al. consider a simple toy example to analyse the convergence of GANs. There may exist two factors of the Jacobian in the GANs dynamics simultaneously that destroy the GANs training: (i) the Phase Factor, i.e., the Jacobian has complex eigenvalues with a large imaginary-to-real ratio; (ii) the Conditioning Factor, i.e., the Jacobian is ill-conditioned. According to the toy example,\\n\\t\\tOnly Regularizing Generator , Only Regularizing Discriminator , and ConOpt could only alleviate the impact of the Phase Factor but not alleviating the impact of the Conditioning Factor. However, JARE can address both factors by construction}.\\\\footnote{Intuitively, a reason for not introducing Hessians in JARE is to avoid the risk of reversing the gradient flows, which may diverge the GAN training dynamics (see Appendix C in for a detailed explanation).}\\n\\tThe above discussions of local convergence during GANs training involve a premise: absolutely continuous data and generator distributions. Indeed, the assumption of absolute continuity is not true for common cases of GANs, where both distributions, specially the data distribution, may lie on lower-dimensional manifolds . More generally, Mescheder et al. extend the convergence proof by to the case where the generator and data distribution do not locally have the same support. Based on this, a simplified zero-centered gradient penalties (zc-GP) method is proposed, which guarantees the local convergence under suitable assumptions. Zc-GP is obtained from the training dynamics, which is similar to 0-GP methods mentioned in Section 4.\\n\\t{Furthermore, there are also other literature studies analyze the training of GANs through other tools. For instance, introduces a novel algorithm, competitive gradient descent (CGD), that is a natural extension of gradient descent to the competitive setting. Different from gradient descent ascent (GDA) in Eq (\\\\ref{eq:eqn27}), CGD does not need to reduce the stepsize to match the increase of the interactions to avoid divergence. Specifically, CGD introduces an equilibrium term that lets each player prefer strategies that are less vulnerable to the actions of the other player. also elucidates the cause of undesirable convergence of GDA is leader's (discriminator) gradient step takes the system away from the ridge, which has undesirable convergence properties and requires using very small learning rates to converge. To mitigate this, Follow-the-Ridge (FR) term ($\\\\mathbf{H}_{\\\\mathbf{\\\\theta} \\\\mathbf{\\\\theta}}^{-1} \\\\mathbf{H}_{\\\\mathbf{\\\\theta} \\\\mathbf{\\\\phi}} \\\\nabla_{\\\\mathbf{\\\\phi}} f\\\\left(\\\\mathbf{\\\\phi}^{(k)}, \\\\mathbf{\\\\theta}^{(k)}\\\\right)$) has been added to the updating of the discriminator. studies the continuous-time dynamics induced by GANs training. In this perspective, instabilities in training GANs arise from the integration error in discretizing the continuous dynamics. It treats GANs training as solving ODEs and shows that higher-order solvers lead to better convergence.}\",\n \"id\": \"f5ab6a67-88dd-45b6-aeae-1b85b061e943\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"44c32216-8be6-4e5c-aa2f-ca241ea4139f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Training dynamics\\\"\"\n ],\n [\n \"subsection\",\n \"Jacobian Regularization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Jacobian Regularization\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 74,\n 98,\n 99,\n 7190\n ],\n \"content\": \"In summary, Jacobian regularization techniques are obtained from the training dynamics of GANs, which are used for achieving local convergence and stabilizing training. The summary of the Jacobian regularization methods is demonstrated in Table \\\\ref{table:Jacobian regularization}. Jacobian regularization is similar to the Gradient penalty in terms of update form. In general, zc-GP is used in many SOTA methods, as demonstrated in Section 7.\\n\\t\\\\begin{table}\\n\\t\\t\\\\caption{The summary of the Jacobian regularization.}\\n\\t\\t\\\\tiny\\n\\t\\t\\\\label{table:Jacobian regularization}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{c|c|c}\\t\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\tMethod& regularized updates of generator ($\\\\phi^{(k+1)}$)®ularized updates of discriminator ($\\\\theta^{(k+1)}$)\\t \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSimGD &$\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})$&$\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tConOpt &$\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\phi||v(\\\\phi^{(k)},\\\\theta^{(k)})||^2$&$\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\theta||v(\\\\phi^{(k)},\\\\theta^{(k)})||^2$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tGenerator &$\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\phi||\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta^{(k)})||^2$&$\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})$\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tDiscriminator &$\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})$&$\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\theta||\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta^{(k)})||^2 $\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tJARE &$\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\phi||\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta^{(k)})||^2$&$\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\theta||\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta^{(k)})||^2 $\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tzc-GP &$\\\\phi^{(k)}-h\\\\nabla_\\\\phi f(\\\\phi^{(k)},\\\\theta ^{(k)})$&$\\\\theta^{(k)}+h\\\\nabla_\\\\theta f(\\\\phi^{(k)},\\\\theta ^{(k)})-\\\\frac{1}{2}h\\\\gamma\\\\nabla_\\\\theta||\\\\nabla D_{\\\\theta}(x)||^2 $\\\\\\\\\\n\\t\\t\\t\\\\bottomrule\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table}\",\n \"id\": \"8377346c-065d-4bc5-adf5-275fad74664d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"44c32216-8be6-4e5c-aa2f-ca241ea4139f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Training dynamics\\\"\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In addition to the three groups mentioned above, this section discusses and summarizes the remaining regularization and normalization techniques, namely, \\\\textit{Layer Normalization} and \\\\textit{Inverse Gradient Penalty}. \\\\textit{Layer Normalization} consists of unconditional-based layer normalization and conditional-based layer normalization, the former inspired by supervised learning is used to accelerate training, but its impact on the GANs is small and sometimes \\n\\tdrops the performance, while the latter is used in the conditional generation and significantly improves the performance of conditional generation; On the other hand, \\\\textit{Inverse Gradient Penalty} mitigates mode collapse by maximizing the Lipschitz constant of the generator.\",\n \"id\": \"bc57c604-a3ec-457f-ab6f-07cf5243c526\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Other methods\\\"\"\n ]\n ],\n \"subsections\": [\n \"253f0a0c-e301-40a7-a3f4-7aa82d059919\",\n \"ed3fad77-b260-419c-962b-ecc924abe2e8\"\n ],\n \"title\": \"Regularization and Normalization of \\\"Other methods\\\"\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 71\n ],\n \"content\": \"Data in machine learning is expected to be independent and identically distributed ($i.i.d$). However, in terms of deep learning, because of the Internal Covariate Shift (ICS) , inputs of each neuron do not satisfy the $i.i.d$, making the training of the deep neural networks hard and unstable. Layer normalization\\\\footnote{\\\\label{ft:10} layer normalization is different from the Layer Normalization (LN), where layer normalization is a general term for a class of methods such as BN, LN. } has been proposed to avoid such problems. The general form of the layer normalization is (the difference between the normalization methods lies in the choice of $h$ and the calculation of $\\\\mathbb{E}[h]$ and $var[h]$):\\n\\t\\\\begin{equation}\\n\\t\\th_N=\\\\frac{x-\\\\mathbb{E}[h]}{\\\\sqrt{var[h]+\\\\epsilon}}\\\\cdot\\\\gamma+\\\\beta.\\n\\t\\\\end{equation}\\n\\tFor GANs, the layer normalization is divided into two parts: {\\\\textit{unconditional-based layer normalization}} and {\\\\textit{conditional-based layer normalization}}. Unconditional-based layer normalizations are used for unconditional generation similar to the other deep neural networks. On the other hand, conditional-based layer normalizations are used for the generator of the conditional generation, where the shift and scale parameters ($\\\\gamma, \\\\beta$) depend on the condition information, as given below:\\n\\t\\\\begin{equation}\\n\\t\\th_N=\\\\frac{x-\\\\mathbb{E}[h]}{\\\\sqrt{var[h]+\\\\epsilon}}\\\\cdot\\\\gamma(c)+\\\\beta(c).\\n\\t\\\\end{equation}\",\n \"id\": \"253f0a0c-e301-40a7-a3f4-7aa82d059919\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"bc57c604-a3ec-457f-ab6f-07cf5243c526\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Other methods\\\"\"\n ],\n [\n \"subsection\",\n \"Layer Normalization\"\n ]\n ],\n \"subsections\": [\n \"e30468f0-4365-4dc0-8ffe-61832a49d364\",\n \"f33ea61f-e2e1-4dfc-9f9a-3c4da4ad1e64\"\n ],\n \"title\": \"Layer Normalization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 154,\n 155,\n 57,\n 8324,\n 153,\n 156,\n 150,\n 152,\n 151,\n 7194,\n 157,\n 78,\n 71\n ],\n \"content\": \"Unconditional-based layer normalization is used for both the generator and discriminator with the same motivation as in other deep neural networks. Ioffe et al. proposed the first normalization for neural networks, namely, Batch Normalization (BN). Batch normalization adopts the data of the mini-batch to compute the mean and variance, making the data distribution of each mini-batch approximately the same. Miyato et al. used the BN in GANs. BN normalizes at the mini-batch level, which destroys the difference between pixels during the generation on account of image generation being a pixel-level task. {Therefore, Batch Norm can be less applicable to style transfer and can’t be used with gradient penalty\\n\\t\\tmethods since the gradient would be dependent on multiple inputs.} Contrary to BN which normalizes the same channel with different images, Layer Normalization\\\\textsuperscript{\\\\ref {ft:10}} (LN) normalizes different channels of a single image that also destroys the diversity between channels for the pixel-by-pixel generative model . Instance Normalization (IN) has also been proposed for style transformation that is adopted for a single channel of a single image. Moreover, Group Normalization (GN) sits between LN and IN, which first divides the channel into many groups, and normalizes different groups of a single image. Compared to normalization of input of neural networks in BN, LN, IN and GN, Weight Normalization (WN) normalizes the weight matrix of neural networks. Miyato et al. also used this normalization in GANs. \\n\\tIn summary, unconditional-based layer normalization in GANs is similar to other neural networks. The related summaries are shown in Table \\\\ref{table:layer normalization}. To the best of our knowledge, no study compares the performance of these methods, therefore, we demonstrate the FID results\\\\footnote{The base framework comes from the SNGAN in \\\\url{https://github.com/kwotsin/mimicry}} for different normalization methods on CIFAR-10 and CIFAR-100 datasets in Table \\\\ref{table:normalization}. Among them, LN and GN obtained better performance than the most popular normalization method: Spectral normalization (mentioned in Section 4.2) and other methods significantly affect the stability of GANs training.\\n\\t\\\\begin{table*}\\n\\t\\t\\\\caption{The summary of the layer normalization}\\n\\t\\t\\\\label{table:layer normalization}\\n\\t\\t\\\\scriptsize\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{c|c | c | c }\\t\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\tMethod&{ Reference}& Classification&Inputs of $\\\\gamma(c)$ and $\\\\beta(c)$\\t \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tBatch Normalization (BN)&2018 & unconditional-based&-\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tLayer Normalization (LN)&2018 & unconditional-based&-\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tInstance Normalization (IN)&2018 & unconditional-based&-\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tGroup Normalization (GN)&2018 & unconditional-based&-\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tWeight Normalization (WN) &2018 & unconditional-based&-\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tConditional Batch Normalization (CBN)&2018 &conditional-based&class label\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tAdaptive Instance Normalization (AdaIN)&2017 ,2019 &conditional-based&target images\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tSpatially-adaptive (de) Normalization (SPADE)&2019 &conditional-based&sematic segmentation map\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tAttentive Normalization(AN)&2020 &conditional-based&self\\\\\\\\\\n\\t\\t\\t\\\\bottomrule\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table*}\",\n \"id\": \"e30468f0-4365-4dc0-8ffe-61832a49d364\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"253f0a0c-e301-40a7-a3f4-7aa82d059919\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Other methods\\\"\"\n ],\n [\n \"subsection\",\n \"Layer Normalization\"\n ],\n [\n \"subsubsection\",\n \"Unconditional-based layer Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Unconditional-based layer Normalization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 154,\n 8324,\n 150,\n 151,\n 157,\n 7194\n ],\n \"content\": \"Conditional-based layer normalization is only used for the generator of the conditional generation. It aims to introduce conditional information to each layer of the generator, which helps to improve the quality of the generated images. $\\\\gamma(c)$ and $\\\\beta(c)$ in Eq (44) are calculated with different features or class labels as input to the neural network in different methods. Miyato et al. and Zhang et al. used the Conditional Batch Normalization (CBN) to encode class labels, thereby improving the quality of conditional generation. Huang et al. and Karras et al. used the Adaptive Instance Normalization (AdaIN) with target images to improve the accuracy of style transfer. Park et al. used the Spatially-Adaptive (de) Normalization (SPADE) with semantic segmentation image to incorporate semantic information into all layers. Wang et al. used the Attentive Normalization (AN) to model long-range dependent attention, which is similar to self-attention GAN .\\n\\tIn summary, the main difference between these conditional-based normalizations is the content of conditional inputs (c in Eq (45)). As the information of inputs is gradually enriched, the performance of conditional generation is gradually improved. The related summaries are shown in Table \\\\ref{table:layer normalization}.\",\n \"id\": \"f33ea61f-e2e1-4dfc-9f9a-3c4da4ad1e64\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"253f0a0c-e301-40a7-a3f4-7aa82d059919\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Other methods\\\"\"\n ],\n [\n \"subsection\",\n \"Layer Normalization\"\n ],\n [\n \"subsubsection\",\n \"Conditional-based layer Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conditional-based layer Normalization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 56,\n 159,\n 78,\n 160,\n 158\n ],\n \"content\": \"Mode collapse is a common phenomenon in GANs' training, that is, changes in the latent space do not cause changes in the generated images. Geometrically, the phenomenon means that all the tangent vectors of the manifold are no longer independent of each other - some tangent vectors either disappear or become linearly correlated with each other. Intuitively, we can solve this problem by maximizing the Lipschitz constant of the generator, which is opposite of the gradient penalty of the discriminator described in the previous section. Based on this, inverse gradient penalty of the generator has been proposed. Concretely, under the little perturbation of the latent space, the generator needs to produce different images. Yang et al. use it in conditional generation, especially for tasks that are rich in conditional information, such as inpainting and super-resolution. \\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathop{\\\\max}\\\\limits_{G}\\\\mathcal{L}_z(G)=\\\\mathop{\\\\max}\\\\mathbb{E}_{z_1,z_2}\\\\left[\\\\mathop{\\\\min}\\\\left(\\\\frac{||G(y,z_1)-G(y,z_2)||}{||z_1-z_2||},\\\\tau\\\\right)\\\\right],\\n\\t\\\\end{equation}\\n\\twhere $y$ is the class label and $\\\\tau$ is the bound to ensure numerical stability. Unlike the intuition-based study described above, Odena et al. demonstrate that the decreasing of singular value in the Jacobian matrix of the generator is the main reason for the mode collapse during GANs training. Furthermore, the singular value can be approximated by the gradient, so Jacobian clamping is used to limit singular values to $[\\\\lambda_{min},\\\\lambda_{\\\\max}]$. The loss is expressed as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathop{\\\\min}\\\\limits_{G}\\\\mathcal{L}_z(G)=\\\\big(\\\\mathop{\\\\max}(Q,\\\\lambda_{\\\\max})-\\\\lambda_{\\\\max}\\\\big)^2\\n\\t\\t+\\\\big(\\\\mathop{\\\\min}(Q,\\\\lambda_{\\\\min})-\\\\lambda_{\\\\min}\\\\big)^2,\\n\\t\\\\end{equation}\\n\\twhere $Q=||G(z)-G(z')||/||z-z'||$. \\n\\t{In summary, the above two methods are similar and mitigate the model collapse of generator to some extent. The key point is to improve the sensitivity of the generator to latent space. In addition to the above methods used to implement inverse gradient penalty for generators, some studies adopt orthogonal regularization to enforce amenability to truncation by conditioning G to be\\n\\t\\tsmooth. Accordingly, the full space of $z$ will map to good output samples. Introducing orthogonality condition is a direct method:}\\n\\t\\\\begin{equation}\\n\\t\\tR(W)=\\\\beta\\\\left\\\\|W^{\\\\top} W -I\\\\right\\\\|_{\\\\mathrm{F}}^{2},\\n\\t\\\\end{equation}\\n\\t{where W is a weight matrix and $\\\\beta$ is a hyperparameter. However, this regularization is too limiting . Therefore, a relaxed constraint has been designed by . Brock et al. apply Off-Diagonal Orthogonal Regularization (Off-Diagonal OR) to the generator directly enforcing the orthogonality condition:}\\n\\t\\\\begin{equation}\\n\\t\\tR_o(W)=\\\\beta\\\\left\\\\|W^{\\\\top} W \\\\odot(\\\\mathbf{1}-I)\\\\right\\\\|_{\\\\mathrm{F}}^{2},\\n\\t\\\\end{equation}\\n\\t{where $\\\\mathbf{1}$ denotes a matrix with all elements set to 1. The Off-Diagonal OR makes G smooth so that the entire space of $z$ will map to good output samples. Orthogonality regularization is different from spectral normalization . Orthogonality regularization destroys the information about the spectrum by setting all the singular values to one, while spectral normalization only makes the maximum singular be one.\\n\\t}\\n\\t\\\\begin{table}\\n\\t\\t\\\\caption{ FID results for different normalization methods on CIFAR-10 and CIFAR-100 datasets. ( The structure is the same as SNGAN except that the discriminator uses different normalization methods).\\\\tiny}\\n\\t\\t\\\\label{table:normalization}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{ccc}\\n\\t\\t\\t\\\\toprule\\n\\t\\t\\t\\\\midrule\\n\\t\\t\\tmethods & CIFAR-10 & CIFAR-100 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tNone & 40.91 & 45.44 \\\\\\\\\\n\\t\\t\\tBN & 37.63 & 44.45\\\\\\\\\\n\\t\\t\\tLN & \\\\textbf{19.21} & 21.15 \\\\\\\\\\n\\t\\t\\tIN & 34.14 & 43.64\\\\\\\\\\n\\t\\t\\tGN & 19.31 & $\\\\textbf{20.80}$ \\\\\\\\\\n\\t\\t\\tWN & 24.28 & 29.96 \\\\\\\\\\n\\t\\t\\tSN&19.75&22.89\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table}\",\n \"id\": \"ed3fad77-b260-419c-962b-ecc924abe2e8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"bc57c604-a3ec-457f-ab6f-07cf5243c526\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Regularization and Normalization of \\\"Other methods\\\"\"\n ],\n [\n \"subsection\",\n \"Inverse Gradient Penalty\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Inverse Gradient Penalty\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 56,\n 164,\n 8324,\n 108,\n 99,\n 130,\n 161,\n 162,\n 163,\n 62,\n 157,\n 78,\n 8317\n ],\n \"content\": \"{In this section, to provide a side view to the selection of regularization and normalization techniques, we investigate the applications of regularization and normalization techniques frequently employed in state-of-the-art and popular GANs. We select six methods (One per year from 2017-2022) categorized into two classes according to different tasks: Unconditional Generation and Conditional Generation. The selected methods and analysis are shown in Table \\\\ref{table:SOTA }. PGGAN is a popular GAN model in recent years, which grows the size of both the generator and discriminator progressively. PGGAN empowers high-resolution image generation. Since PGGAN was proposed in 2017, only some simple regularization techniques were applied: WGAN-GP , BN , and LN ; BigGAN is a popular conditional generative adversarial networks, which uses many regularization and normalization techniques, such as zc-GP, SN , Off-Diagonal OR , and CBN ; AutoGAN is the first study introducing the Neural architecture search (NAS) to GANs. It defines the search space for the generator architecture and adopts Inception score as the reward to discover the best architecture. The main focus of AutoGAN is architecture, so AutoGAN only comprises SN ; StyleGAN2 is the most popular architecture of GANs, which produces photorealistic images with large varieties and is widely used in image generation, such as Image Completion , Image-to-Image Translation . StyleGAN2-ADA proposes a novelty adaptive data augmentation methods. Combining StyleGAN2 and adaptive data augmentation, StyleGAN2-ADA obtains impressive performance in image generation, particularly in data-effficient generation. Furthermore, InsGen combines StyleGAN2-ADA with contrastive learning, acquiring state of the art on many generation tasks and datasets. Recently, StyleGAN-XL scales StyleGAN to large diverse datasets and sets a new state-of-the-art on large-scale image synthesis. In summary, many regularization and normalization techniques have been used in state-of-the-art GANs with zc-GP and SN being more attractive to researchers. Data augmentation is a striking method and orthogonal to other ongoing researches on training, architecture, and regularization. Therefore, popular augmentation strategies, such as ADA, have been employed as default operations GANs training. Furthermore, self-supervision has been used to further improve the performance of GANs, which is also orthogonal to other methods.}\\n\\t\\\\begin{table}\\n\\t\\t\\\\caption{The applications of the Regularization and Normalization techniques used in SOTA GANs.}\\n\\t\\t\\\\label{table:SOTA }\\n\\t\\t\\\\centering\\n\\t\\t\\\\footnotesize\\n\\t\\t\\\\begin{tabular}{c | c | c |c | c | c|c }\\t\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{ Method}& Task&\\\\makecell[c]{Gradient \\\\\\\\Penalty}&\\\\makecell[c]{Data augmentation\\\\\\\\ and preprocessing}&\\\\makecell{Self\\\\\\\\supervision}&\\\\makecell[c]{Weight\\\\\\\\ normalization }&\\\\makecell[c]{Layer\\\\\\\\ normalization}\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\makecell{PGGAN\\\\\\\\(2017)}&\\\\makecell{Unconditinal\\\\\\\\ Generation}&WGAN-GP&\\\\makecell{None}&None&\\\\makecell{None}&BN: G,LN: D\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\makecell{BigGAN\\\\\\\\ (2018 )}&\\\\makecell{Conditinal \\\\\\\\Generation}&zc-GP&None&None&\\\\makecell{SN: G, D}&CBN\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\makecell{AutoGAN\\\\\\\\ (2019 )}&\\\\makecell{Unconditinal\\\\\\\\ Generation}&None&None&None&\\\\makecell{SN:D}&None\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\makecell{StyleGAN2-ADA\\\\\\\\(2020)}&\\\\makecell{Unconditinal \\\\\\\\Generation}&zc-GP&\\\\makecell{Adaptive}&None&None&IN\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\makecell{InsGen\\\\\\\\(2021)}&\\\\makecell{Unconditinal \\\\\\\\Generation}&zc-GP&\\\\makecell{Adaptive}&contrastive&None&IN\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\makecell{StyleGAN-XL\\\\\\\\(2022)}&\\\\makecell{Conditinal \\\\\\\\Generation}&None&\\\\makecell{Translation\\\\\\\\Cutout}&None&SN:D&IN\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{table}\",\n \"id\": \"fb4135aa-ff8b-439d-9a28-a790ffddd379\",\n \"level\": \"section\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Applications of Regularization and Normalization in SOTA GANs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Applications of Regularization and Normalization in SOTA GANs\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"b359d174-4880-4149-b0f4-5fd0e7bd75f4\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Summary and Outlook\"\n ]\n ],\n \"subsections\": [\n \"9cf8af1d-75ca-445d-aa5a-fb3169fa5585\",\n \"c7f682e9-c24a-41ee-8bfd-ac2766fa59c0\"\n ],\n \"title\": \"Summary and Outlook\"\n },\n {\n \"cite_extract_rate\": 0.8888888888888881,\n \"cites\": [\n 144,\n 88,\n 7001,\n 99,\n 8322,\n 78,\n 8317,\n 6997\n ],\n \"content\": \"Recently, significant achievements of GANs have been made in generation tasks and the network has been widely used in many computer vision tasks, such as image inpainting, style transfer, text-to-image translations, and attribute editing. However, due to the overconfident assumptions, the training faces many challenges, such as non-convergence, mode collapse, gradient vanishing, and overfitting. To mitigate these problems, many solutions focus on designing new architectures, new loss functions, new optimization methods, and regularization and normalization techniques. \\n\\tIn this paper, we study GANs training from three perspectives and propose a new taxonomy, denoted as \\\\textbf{\\\"Training dynamics\\\"}, \\\\textbf{\\\"Fitting distribution\\\"}, \\\\textbf{\\\"Real \\\\& Fake\\\"}, and \\\\textbf{\\\"Other methods\\\"}, to survey the different regularization and normalization techniques during GANs training. Our study provides a systematic and comprehensive analysis of the reviewed methods to serve researchers of the community. In addition, we also demonstrate the motivation and objectives of different methods and compare the performance of some popular methods in a fair manner quantitatively, which has implications for future research in selecting their research topics or developing their approaches.\\n\\t\\\\begin{comment}\\n\\tThis paper summarizes the regularization and normalization of generative adversarial networks and explains the problem from three aspects: \\n\\t\\\\textbf{Optimal transport and Lipschitz continuity:} First, optimal transport and optimal transport with the regular term are introduced. According to the duality form, WGAN-GP and WGAN-LP are proposed to make the discriminator satisfy 1-Lipschitz continuity and k-Lipschitz continuity ($k\\\\leq1$), respectively. Next, many gradient penalty methods have been analyzed. These methods achieved certain improvements as compared with WGAN-GP in two prespecitves: \\\\textbf{(1) To limit the Lipschitz constant to a small value}, such as 0 ; \\\\textbf{(2) To find the right restricted space}. The restricted space of WGAN-GP and WGAN-LP is the interpolation between the real images and fake images. Some works restricted it at the real images or fake images . The relationship between the Lipschitz constant and the spectral norm of the matrix is also derived in this paper. It indicates that the normalization of the spectral norm can also be used to achieve 1-Lipschitz continuity. Furthermore, the Frobenius norm is proved to be the upper bound of the spectral norm. The results of WGAN-LP can be obtained by normalizing the Frobenius norm. Moreover, the work analyzed the relationship between the mode collapse of the GAN and the eigenvalue distribution of the discriminator weight matrix, ending up in avoiding the mode collapse by limiting the eigenvalue collapse. However, there is a lack of the specific reason or sound proof.\\n\\t\\\\textbf{ Training dynamics:} GANs is a two-player zero-sum game. Due to the nonlinearity of the neural network, it is difficult to find the global convergence solution. In terms of the local convergence, the Jacobian regularization needs to be added. Few research studies are found in this area.\\n\\t\\\\textbf{ Representation:} As mentioned earlier, training of GANs is a semi-supervised task. More prior information is needed to improve the presentative ability of the network. Conditional layer normalization is used to better encode condition information and reduce the difficulty of conditional generation. Consistent regularization, data augmentation and self-supervision use unsupervised methods to improve the supervision information in GANs training, and use additional labels and tasks to improve GANs performance. We suggest that this area is promising and more studies can be conducted in this specific direction.\\n\\t\\\\end{comment}\",\n \"id\": \"9cf8af1d-75ca-445d-aa5a-fb3169fa5585\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"b359d174-4880-4149-b0f4-5fd0e7bd75f4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Summary and Outlook\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 165\n ],\n \"content\": \"By reviewing the regularization and normalization of GANs, the following questions and thoughts are proposed based on different perspectives of GANs training:\\n\\t\\\\begin{itemize}\\n\\t\\t\\\\item [1)] \\n\\t\\t\\\\textit{What is a good distance metric, and which divergence should be used in GANs training?} The priority in the training process of GANs is to find a suitable divergence to measure the distance between the generated distribution and the true distribution. Wasserstein divergence is important for the training of GANs. However, it is uncertain whether the next proposed divergence performs better.\\n\\t\\t\\\\item [2)]\\n\\t\\t\\\\textit{What is the main difference between real images and generated images?} During the training of unconstrained and unprioritized GANs, if we can quantitatively represent the difference between real images and generated images from different perspectives, the efficient regularization methods can be designed based on this.\\n\\t\\t\\\\item [3)]\\n\\t\\t\\\\textit{How to avoid real images forgetting\\\\footnote{Real images forgetting is caused by not introducing real images while training the generator, which is different from discriminator forgetting.}?} As acknowledged, real images do not directly participate in the training of the generator, thus the discriminator needs to remember the characteristics of the real images to optimize the generator indirectly. We call this the real images forgetting. We conjecture that real images forgetting may exist, and which may increase the difficulty of GANs training. Some works might serve as basis to prove this hypothesis and propose effective solutions.\\n\\t\\t\\\\item [4)]\\n\\t\\tRecent studies show that discriminator suffers from overfitting and discriminator forgetting. It is a common problem of neural networks, which is caused by the shortcut of the loss driven method. Some new methods, such as contrastive learning, representation learining, can be proposed to improve the generalization of the discriminator.\\n\\t\\t\\\\item [5)]\\n\\t\\tRecently, diffusion model acquires the impressive performance in image generation. One possible reason for success is the phased training strategy in diffusion model . Inspired by this, some strategies to reduce the difficulty of GANs training may be proposed.\\n\\t\\\\end{itemize}\\n\\t\\\\section*{Acknowledgment}\\n\\tThe work is partially supported by the National Natural Science Foundation of China under Grand No.U19B2044, No.61836011 and No.91746209. We are very grateful to the help of Jianlin Su, whose blog is \\\\url{https://spaces.ac.cn/tag/GAN/}.\\n\\t\\\\bibliographystyle{ACM-Reference-Format}\\n\\t\\\\bibliography{bare_jrnl_transmag}\\n\\t\\\\appendix\\n\\t\\\\clearpage\",\n \"id\": \"c7f682e9-c24a-41ee-8bfd-ac2766fa59c0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"b359d174-4880-4149-b0f4-5fd0e7bd75f4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Summary and Outlook\"\n ],\n [\n \"subsection\",\n \"Outlook\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Outlook\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"f6a7c46a-3c0c-4605-a602-4104833c8854\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4410407c-7d7c-48f1-8174-a9a95319844f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Supplementary Online-only Material\"\n ]\n ],\n \"subsections\": [\n \"9df0baaf-f220-4a8d-b36a-27ebd583cf6c\",\n \"375c251d-5d04-454e-83a6-2140b14015f5\",\n \"d07e4dc9-6fdd-4534-85ee-925acd678910\"\n ],\n \"title\": \"Supplementary Online-only Material\"\n },\n {\n \"cite_extract_rate\": 0.42857142857142805,\n \"cites\": [\n 88,\n 95,\n 8318\n ],\n \"content\": \"\\\\label{sect:A-1}\\n\\tOptimal transport was proposed in the 18th century to minimize the transportation cost while preserving the measure quantities. Given the space with probability measures $(X,\\\\mu)$ and $(Y,\\\\upsilon)$, if there is a map $T:X\\\\rightarrow Y$ which is measure-preserving, then for any $B\\\\subset Y$, having:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\t\\\\int_{T^{-1}(B)}\\\\mathrm{d}\\\\mu(x)=\\\\int_B \\\\mathrm{d}\\\\upsilon(y).\\n\\t\\\\end{equation}\\n\\tWriting the measure-preserving map as $T_*(\\\\mu)=\\\\upsilon$. For any $x\\\\in X$ and $y\\\\in Y$, the transportation distance is defined as $c(x,y)$, the total transportation cost is given by:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tC(T):=\\\\int_X c(x,T(x)) \\\\mathrm{d}\\\\mu(x).\\n\\t\\\\end{equation}\\n\\tIn the 18th century, Monge et al. proposed the Optimal Mass Transportation Map that corresponds to the smallest total transportation cost: $C(T)$. The transportation cost corresponding to the optimal transportation map is called the Wasserstein distance between probability measures $\\\\mu$ and $\\\\upsilon$:\\n\\t\\\\begin{equation}\\\\label{eqn3}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)=\\\\mathop{\\\\min}\\\\limits_{T}\\\\left\\\\{\\\\int_X c(x,T(x)) \\\\mathrm{d}\\\\mu(x)\\\\ |\\\\ T_*(\\\\mu)=\\\\upsilon\\\\right\\\\}.\\n\\t\\\\end{equation}\\n\\tIn 1940s, Kantorovich proved the existence and uniqueness of the solution for Monge problem, and according to the duality of linear programming, the Kantorovich-Rubinstein (KR) duality of Wasserstein distance is given by:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)\\n\\t\\t=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi ,\\\\psi}\\\\left\\\\{\\\\int_X \\\\varphi \\\\mathrm{d}\\\\mu\\\\ +\\\\int_Y \\\\psi \\\\mathrm{d}\\\\upsilon\\\\ |\\\\ \\\\varphi(x)+\\\\psi(y)\\\\leq c(x,y)\\\\right\\\\}.\\n\\t\\\\end{equation}\\n\\tThis dual problem is constrained, defining the c-transform: $\\\\psi(y)=\\\\varphi^c(y):=inf_x\\\\{c(x,y)-\\\\varphi(x)\\\\}$, and the Wasserstein distance becomes:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi}\\\\left\\\\{\\\\int_X \\\\varphi \\\\mathrm{d}\\\\mu\\\\ +\\\\int_Y \\\\varphi^c \\\\mathrm{d}\\\\upsilon\\\\right\\\\},\\n\\t\\\\end{equation}\\n\\twhere $\\\\varphi$ is called the Kantorovich potential. It can be shown that if $c(x,y)=|x-y|$ and Kantorovich potential satisfies the 1-Lipschitz continuity, then $\\\\varphi^c=-\\\\varphi$. Kantorovich potential can be fitted by a deep neural network, which is recorded as $\\\\varphi_\\\\xi$. Wasserstein distance is:\\n\\t\\\\begin{equation}\\n\\t\\tW_c(\\\\mu,\\\\upsilon)=\\\\mathop{\\\\max}\\\\limits_{||\\\\varphi_\\\\xi||_L\\\\leq 1}\\\\left\\\\{\\\\int_X \\\\varphi_\\\\xi \\\\mathrm{d}\\\\mu\\\\ -\\\\int_Y \\\\varphi_\\\\xi \\\\mathrm{d}\\\\upsilon\\\\right\\\\}.\\n\\t\\t\\\\label{eqn7}\\n\\t\\\\end{equation}\\n\\tIf $X$ is the generated image space, $Y$ is the real sample space, $Z$ is latent space and $g_\\\\theta$ is the geneartor, the Wasserstein GANs (WGAN) is formulated as a min-max problem:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\t\\\\mathop{\\\\min}\\\\limits_{\\\\theta}\\\\mathop{\\\\max}\\\\limits_{||\\\\varphi_\\\\xi||_L\\\\leq 1}\\\\left\\\\{\\\\int_Z \\\\varphi_\\\\xi(g_\\\\theta(z)) \\\\mathrm{d}z \\\\ -\\\\int_Y \\\\varphi_\\\\xi(y) \\\\mathrm{d}y\\\\right\\\\}.\\n\\t\\\\end{equation}\\n\\tIn the optimization process, the generator and the Kantorovich potential function (discriminator) are independent of each other, optimized in a step-by-step iteration.\\n\\tIf $c(x,y)=\\\\frac{|x-y|^2}{2}$, there is a convex function $u$ that is called Brenier potential . The optimal transportation map is given by the gradient map of Brenier potential: $T(x)=\\\\nabla u(x)$. There exists a relationship between Kantorovich potential and Brenier potential : \\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tu(x)=\\\\frac{|x|^2}{2}-\\\\varphi(x).\\n\\t\\\\end{equation}\\n\\tFrom the previous discussion, it is evident that the optimal transportation map (Brenier potential) corresponds to the generator, and Kantorovich potential corresponds to the discriminator. After the discriminator is optimized, the generator is directly drivable without the optimization process .\\n\\tThe transportation cost of Eq (\\\\ref{eqn3}) is defined as the form of two distribution distances:\\n\\t\\\\begin{equation}\\\\label{eqn10}\\n\\t\\tOT(P||Q)=\\\\mathop{inf}\\\\limits_{\\\\pi}\\\\int\\\\pi(x,y)c(x,y)\\\\mathrm{d}x\\\\mathrm{d}y,\\n\\t\\\\end{equation}\\n\\twhere $\\\\pi(x,y)$ is the joint distribution, satisfying $\\\\int_y\\\\pi(x,y)dy=P(x)$ and $\\\\int_x\\\\pi(x,y)dx=Q(y)$. The dual form of Eq (\\\\ref{eqn10}) is derived as follows::\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\tOT(P||Q)=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi ,\\\\psi}\\\\{\\\\int_x \\\\varphi(x)P(x) \\\\mathrm{d}x\\\\ \\\\\\\\ +\\\\int_y\\\\psi(y)Q(y) \\\\mathrm{d}y\\\\ |\\\\ \\\\varphi(x)+\\\\psi(y)\\\\leq c(x,y)\\\\}.\\n\\t\\\\end{equation} \\n\\tConsidering the optimal transportation with regular terms, Peyr{\\\\'e} et al. added the entropic regularization for optimal transportation that transforms the dual problem into a smooth unconstrained convex problem. The regularized optimal transport is defined as:\\n\\t\\\\begin{equation}\\\\label{eqn12}\\n\\t\\tOT_c(P||Q)=\\\\mathop{\\\\min}\\\\limits_{\\\\pi}\\\\int\\\\pi(x,y)c(x,y)\\\\mathrm{d}x\\\\mathrm{d}y+\\\\epsilon E(\\\\pi).\\n\\t\\\\end{equation}\\n\\tIf $E(\\\\pi)=\\\\int_x\\\\int_y\\\\pi(x,y)\\\\log(\\\\frac{\\\\pi(x,y)}{P(x)Q(y)})\\\\mathrm{d}x\\\\mathrm{d}y$, Eq (\\\\ref{eqn12}) can be written as:\\n\\t\\\\begin{equation}\\\\label{eqn13}\\n\\t\\t\\\\begin{split}\\n\\t\\t\\tOT_c(P||Q)=&\\\\mathop{\\\\min}\\\\limits_{\\\\pi}\\\\int\\\\pi(x,y)c(x,y)\\\\mathrm{d}x\\\\mathrm{d}y+\\\\epsilon \\\\int_x\\\\int_y\\\\pi(x,y)\\\\log\\\\left(\\\\frac{\\\\pi(x,y)}{P(x)Q(y)}\\\\right)\\\\mathrm{d}x\\\\mathrm{d}y\\\\\\\\\\n\\t\\t\\ts.t. \\\\int_y\\\\pi(x,y)&\\\\mathrm{d}y=P(x),\\\\int_x\\\\pi(x,y)\\\\mathrm{d}x=Q(y).\\n\\t\\t\\\\end{split}\\n\\t\\\\end{equation}\\n\\tThe dual form of Eq (\\\\ref{eqn13}) becomes:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\begin{split}\\n\\t\\t\\tOT_c(P||Q)\\n\\t\\t\\t&=\\\\mathop{\\\\max}\\\\limits_{\\\\varphi ,\\\\psi}\\\\int_x \\\\varphi(x)P(x) \\\\mathrm{d}x\\\\ +\\\\int_y\\\\psi(y)Q(y)\\\\mathrm{d}y\\\\\\\\\\n\\t\\t\\t&+\\\\frac{\\\\epsilon}{e}\\\\int_x\\\\int_y\\\\exp\\\\left(\\\\frac{-\\\\left(c(x,y)+\\\\varphi(x)+\\\\psi(y)\\\\right)}{\\\\epsilon}\\\\right)\\\\mathrm{d}x\\\\mathrm{d}y.\\n\\t\\t\\\\end{split}\\n\\t\\t\\\\label{EQ:eqn14}\\n\\t\\\\end{equation}\\n\\tPetzka et al. set $c(x,y)=||x-y||_2$ in Eq (\\\\ref{EQ:eqn14}), and the dual form of optimal transport with the regular term can be expressed as:\\n\\t\\\\begin{equation}\\n\\t\\t\\\\mathop{\\\\sup}\\\\limits_{\\\\varphi,\\\\psi}\\\\{\\\\mathbb{E}_{x\\\\sim{p(x)}}[\\\\varphi(x)]-\\\\mathbb{E}_{y\\\\sim q(y)}[\\\\psi(y)]\\n\\t\\t-\\\\frac{4}{\\\\epsilon}\\\\int\\\\int\\\\mathop{\\\\max}\\\\{0,(\\\\varphi(x)-\\\\psi(y)-||x-y||_2)\\\\}^2\\\\mathrm{d}p(x)\\\\mathrm{d}q(y) \\\\}.\\n\\t\\t\\\\label{EQ:eqn32}\\n\\t\\\\end{equation}\\n\\tSimilar to dealing with a single function, one can replace $\\\\varphi = \\\\psi$ in Eq (\\\\ref{EQ:eqn32}), which leads to the objective of minimum:\\n\\t\\\\begin{equation}\\\\label{eqn1}\\n\\t\\t\\\\mathbb{E}_{y\\\\sim q(y)}[\\\\varphi(y)]-\\\\mathbb{E}_{x\\\\sim p(x)}[\\\\varphi(x)]\\n\\t\\t+\\\\frac{4}{\\\\epsilon}\\\\int\\\\int\\\\mathop{\\\\max}\\\\{0,(\\\\varphi(x)-\\\\varphi(y)-||x-y||_2)\\\\}^2\\\\mathrm{d}p(x)\\\\mathrm{d}q(y).\\n\\t\\\\end{equation}\",\n \"id\": \"9df0baaf-f220-4a8d-b36a-27ebd583cf6c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"f6a7c46a-3c0c-4605-a602-4104833c8854\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Supplementary Online-only Material\"\n ],\n [\n \"subsection\",\n \"Optimal Transport and Lipschitz Continuity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Optimal Transport and Lipschitz Continuity\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7190\n ],\n \"content\": \"\\\\label{sect:A-2}\\n\\t\\\\newenvironment{lemma1}{{\\\\indent\\\\it \\\\textbf{Proposition 2.1:}}}{\\\\hfill \\\\par}\\n\\t\\\\begin{lemma1}\\n\\t\\t\\\\textit{For zero-sum games, $v^{'}$is negative semi-definite for any local Nash-equilibrium. Conversely, if $v(\\\\bar{x})=0$ and $v^{'}$is negative definite, then $\\\\bar{x}$ is a local Nash-equilibrium.\\n\\t\\t}\\n\\t\\\\end{lemma1}\\n\\t\\\\newenvironment{proof1}{{\\\\indent\\\\it Proof 2.1:}}{\\\\hfill $\\\\square$\\\\par}\\n\\t\\\\begin{proof1}\\n\\t\\tRefer to \\n\\t\\\\end{proof1}\\n\\t\\\\textit{Proposition 2.1} gives the conditions for the local convergence of GANs, which is converted into the negative semi-definite problem of the Jacobian matrix. Negative semi-definite of the Jacobian matrix corresponds to its eigenvalue less than or equal to 0. If the eigenvalue of the Jacobian matrix at a certain point is a negative real number, the training process can converge; but if the eigenvalue is complex and the real part of the eigenvalue is small and the imaginary part is relatively large, the training process is difficult to converge unless the learning rate is very small. \\n\\t\\\\newenvironment{lemma2}{{\\\\indent\\\\it \\\\textbf{Proposition 2.2:}}}{\\\\hfill \\\\par}\\n\\t\\\\begin{lemma2}\\n\\t\\t\\\\textit{\\n\\t\\t\\tLet $F:\\\\Omega\\\\rightarrow\\\\Omega$ be a continuously differentiable function on an open subset $\\\\Omega$ of $R^n$ and let $\\\\bar{x}\\\\in\\\\Omega$ be so that: 1. $F(\\\\bar{x})=\\\\bar{x}$ and 2. the absolute values of the eigenvalues of the Jacobian $F^{'}(x)$ are all smaller than 1.}\\n\\t\\t\\\\textit{There is an open neighborhood $U$ of $\\\\bar{x}$ so that for all $x_0\\\\in U$, the iterates $F^{(k)}(x_0)$ converge to $\\\\bar{x}$. The rate of convergence is at least linear. More precisely, the error $||F^{(k)}(x_0)-\\\\bar{x}||$ is in $\\\\mathcal{O}(|\\\\lambda_{max}|^k)$ for $k\\\\rightarrow\\\\infty$ where $\\\\lambda_{max}$ is the eigenvalue of $F^{'}(\\\\bar{x})$ with the largest absolute value.}\\n\\t\\\\end{lemma2}\\n\\t\\\\newenvironment{proof2}{{\\\\indent\\\\it Proof 2.2:}}{\\\\hfill $\\\\square$\\\\par}\\n\\t\\\\begin{proof2}\\n\\t\\tRefer to Section 3 in and Proposition 4.4.1 in .\\n\\t\\\\end{proof2}\",\n \"id\": \"375c251d-5d04-454e-83a6-2140b14015f5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"f6a7c46a-3c0c-4605-a602-4104833c8854\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Supplementary Online-only Material\"\n ],\n [\n \"subsection\",\n \"Some Propositions of Training Dynamic in GANs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Some Propositions of Training Dynamic in GANs\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 97,\n 96\n ],\n \"content\": \"\\\\label{sect:A-3}\\n\\t1-Lipschitz continuity is represented as:\\n\\t\\\\begin{equation}\\n\\t\\t||D(x_1)-D(x_2)||\\\\leq ||x_1-x_2||.\\\\footnote{Lipschitz continuity can be defined by any form of norm.}\\n\\t\\\\end{equation}\\n\\tGenerally, considering the K-Lipschitz for a neural network $f(x)$:\\n\\t\\\\begin{equation}\\n\\t\\tf(x)=g_N\\\\circ\\\\cdots g_2\\\\circ g_1(x),\\\\footnote{$\\\\circ$ is the symbol for function cascade. Specifically, $h\\\\circ g(x)=h(g(x))$. This definition of neural network is not general, such as DenseNet and ResNet , which can not be defined like this. Therefore, we do not strictly derive the relationship between the matrix norm and Lipschitz continuity.}\\n\\t\\\\end{equation}\\n\\twhere $g_i(x)=\\\\sigma (W_i x+b_i)$. And K-Lipschitz continuity for $f(x)$ is:\\n\\t\\\\begin{equation}\\n\\t\\t||f(x_1)-f(x_2)||\\\\leq \\\\mathrm{K}||x_1-x_2||,\\n\\t\\t\\\\label{EQ:eqn17}\\n\\t\\\\end{equation}\\n\\twhere K is Lipschitz constant of the function $f$. Due to the consistency of Lipschitz $||h\\\\circ g||_{Lip}\\\\leq ||h||_{Lip}\\\\cdot||g||_{Lip}$, $g_i$ needs to satisfy the C-Lipschitz continuity ($\\\\mathrm{C}=\\\\sqrt[N]{\\\\mathrm{K}}$) so that $f$ satisfies the K-Lipschitz continuity:\\n\\t\\\\begin{equation}\\n\\t\\t||g_i(x_1)-g_i(x_2)||\\\\leq \\\\mathrm{C}||x_1-x_2||,\\n\\t\\\\end{equation}\\n\\t\\\\begin{equation}\\n\\t\\t||\\\\sigma(Wx_1+b)-\\\\sigma(Wx_2+b)||\\\\leq \\\\mathrm{C}||x_1-x_2||.\\n\\t\\t\\\\label{eq:23}\\n\\t\\\\end{equation}\\n\\tWhen $x_1\\\\rightarrow x_2$, the Taylor expansion of Eq (\\\\ref{eq:23}):\\n\\t\\\\begin{equation}\\n\\t\\t||\\\\frac{\\\\partial\\\\sigma}{\\\\partial x} W(x_1-x_2)||\\\\leq \\\\mathrm{C}||x_1-x_2||.\\n\\t\\\\end{equation}\\n\\tNormally, $\\\\sigma$ is a function with limited derivatives such as Sigmoid, so the $\\\\mathrm{C'}$-Lipschitz continuity is be written as:\\n\\t\\\\begin{equation}\\n\\t\\t|| W(x_1-x_2)||\\\\leq \\\\mathrm{C'}||x_1-x_2||,\\n\\t\\\\end{equation}\\n\\twhere $\\\\mathrm{C'}$ is a limited constant, which is determined by $\\\\frac{\\\\partial\\\\sigma}{\\\\partial x}$ and $\\\\mathrm{C}$.\\n\\tSimilarly, the spectral norm of matrix is defined by:\\n\\t\\\\begin{equation}\\n\\t\\t||W||_2=\\\\mathop{\\\\max}\\\\limits_{x\\\\not=0}\\\\frac{||Wx||}{||x||}.\\n\\t\\\\end{equation}\\n\\tIn this context, the spectral norm $||W||_2$ can be used to represent the Lipschitz constant $\\\\mathrm{C'}$. The Lipschitz continuity is achieved by normalizing the spectral norm of the weight, approximately.\\t\\n\\\\end{document}\\n\\\\endinput\",\n \"id\": \"d07e4dc9-6fdd-4534-85ee-925acd678910\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"f6a7c46a-3c0c-4605-a602-4104833c8854\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Systematic Survey of Regularization and Normalization in GANs\"\n ],\n [\n \"section\",\n \"Supplementary Online-only Material\"\n ],\n [\n \"subsection\",\n \"Spectral Norm and the Lipschitz Constant\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Spectral Norm and the Lipschitz Constant\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":84,"string":"84"},"all_cites":{"kind":"list like","value":[63,89,73,86,68,85,78,76,8317,53,60,6998,66,57,70,88,80,6996,72,54,58,62,59,7190,55,69,77,6997,82,74,56,81,87,6999,8316,84,71,61,75,67,65,83,64,79,91,90,92,93,94,95,8318,97,96,98,100,102,99,101,7191,111,104,112,107,106,110,109,108,103,113,8319,105,115,114,7000,7192,116,117,121,8320,119,118,120,123,122,126,132,8321,128,124,129,130,131,7193,127,125,134,133,135,136,137,138,141,140,139,142,143,7002,7001,8322,144,145,146,8323,149,7003,148,147,154,155,8324,153,156,150,152,151,7194,157,159,160,158,164,161,162,163,165],"string":"[\n 63,\n 89,\n 73,\n 86,\n 68,\n 85,\n 78,\n 76,\n 8317,\n 53,\n 60,\n 6998,\n 66,\n 57,\n 70,\n 88,\n 80,\n 6996,\n 72,\n 54,\n 58,\n 62,\n 59,\n 7190,\n 55,\n 69,\n 77,\n 6997,\n 82,\n 74,\n 56,\n 81,\n 87,\n 6999,\n 8316,\n 84,\n 71,\n 61,\n 75,\n 67,\n 65,\n 83,\n 64,\n 79,\n 91,\n 90,\n 92,\n 93,\n 94,\n 95,\n 8318,\n 97,\n 96,\n 98,\n 100,\n 102,\n 99,\n 101,\n 7191,\n 111,\n 104,\n 112,\n 107,\n 106,\n 110,\n 109,\n 108,\n 103,\n 113,\n 8319,\n 105,\n 115,\n 114,\n 7000,\n 7192,\n 116,\n 117,\n 121,\n 8320,\n 119,\n 118,\n 120,\n 123,\n 122,\n 126,\n 132,\n 8321,\n 128,\n 124,\n 129,\n 130,\n 131,\n 7193,\n 127,\n 125,\n 134,\n 133,\n 135,\n 136,\n 137,\n 138,\n 141,\n 140,\n 139,\n 142,\n 143,\n 7002,\n 7001,\n 8322,\n 144,\n 145,\n 146,\n 8323,\n 149,\n 7003,\n 148,\n 147,\n 154,\n 155,\n 8324,\n 153,\n 156,\n 150,\n 152,\n 151,\n 7194,\n 157,\n 159,\n 160,\n 158,\n 164,\n 161,\n 162,\n 163,\n 165\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.85650187118258,"string":"0.856502"}}},{"rowIdx":50,"cells":{"authors":{"kind":"list like","value":["Thanh Thi Nguyen","Quoc Viet Hung Nguyen","Dung Tien Nguyen","Duc Thanh Nguyen","Thien Huynh-The","Saeid Nahavandi","Thanh Tam Nguyen","Quoc-Viet Pham","Cuong M. Nguyen"],"string":"[\n \"Thanh Thi Nguyen\",\n \"Quoc Viet Hung Nguyen\",\n \"Dung Tien Nguyen\",\n \"Duc Thanh Nguyen\",\n \"Thien Huynh-The\",\n \"Saeid Nahavandi\",\n \"Thanh Tam Nguyen\",\n \"Quoc-Viet Pham\",\n \"Cuong M. Nguyen\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Deep Learning for Deepfakes Creation and Detection: A Survey"},"year":{"kind":"string","value":"2019"},"date":{"kind":"timestamp","value":"2019-09-25T16:03:45Z","string":"2019-09-25T16:03:45Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" Deep learning has been successfully applied to solve various complex problems ranging from big data analytics to computer vision and human-level control. Deep learning advances however have also been employed to create software that can cause threats to privacy, democracy and national security. One of those deep learning-powered applications recently emerged is deepfake. Deepfake algorithms can create fake images and videos that humans cannot distinguish them from authentic ones. The proposal of technologies that can automatically detect and assess the integrity of digital visual media is therefore indispensable. This paper presents a survey of algorithms used to create deepfakes and, more importantly, methods proposed to detect deepfakes in the literature to date. We present extensive discussions on challenges, research trends and directions related to deepfake technologies. By reviewing the background of deepfakes and state-of-the-art deepfake detection methods, this study provides a comprehensive overview of deepfake techniques and facilitates the development of new and more robust methods to deal with the increasingly challenging deepfakes. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"44777335-fb2f-4b5d-83e4-47fe65116410","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"]],"subsections":["c7a70291-a1e9-476f-8254-e897ee781284","b11ed0ba-b192-4a4a-b063-d32fbaab94d4","7513ea5e-21ef-4144-b033-7ba34feb94da","af5e6522-a34d-4deb-8b41-532983a76f29","6f55771a-c73a-46f3-b17e-1720217d26ca"],"title":"root"},{"cite_extract_rate":0.5,"cites":[5133,5128,5132,5130,5131,5680,5129,1602,986,1256,4939],"content":"\\label{sec1}\nIn a narrow definition, deepfakes (stemming from ``deep learning\" and ``fake\") are created by techniques that can superimpose face images of a target person onto a video of a source person to make a video of the target person doing or saying things the source person does. This constitutes a category of deepfakes, namely \\emph{face-swap}. In a broader definition, deepfakes are artificial intelligence-synthesized content that can also fall into two other categories, i.e., \\emph{lip-sync} and \\emph{puppet-master}. Lip-sync deepfakes refer to videos that are modified to make the mouth movements consistent with an audio recording. Puppet-master deepfakes include videos of a target person (puppet) who is animated following the facial expressions, eye and head movements of another person (master) sitting in front of a camera .\nWhile some deepfakes can be created by traditional visual effects or computer-graphics approaches, the recent common underlying mechanism for deepfake creation is deep learning models such as autoencoders and generative adversarial networks (GANs), which have been applied widely in the computer vision domain . These models are used to examine facial expressions and movements of a person and synthesize facial images of another person making analogous expressions and movements . Deepfake methods normally require a large amount of image and video data to train models to create photo-realistic images and videos. As public figures such as celebrities and politicians may have a large number of videos and images available online, they are initial targets of deepfakes. Deepfakes were used to swap faces of celebrities or politicians to bodies in porn images and videos. The first deepfake video emerged in 2017 where face of a celebrity was swapped to the face of a porn actor. It is threatening to world security when deepfake methods can be employed to create videos of world leaders with fake speeches for falsification purposes . Deepfakes therefore can be abused to cause political or religion tensions between countries, to fool public and affect results in election campaigns, or create chaos in financial markets by creating fake news . It can be even used to generate fake satellite images of the Earth to contain objects that do not really exist to confuse military analysts, e.g., creating a fake bridge across a river although there is no such a bridge in reality. This can mislead a troop who have been guided to cross the bridge in a battle .\nAs the democratization of creating realistic digital humans has positive implications, there is also positive use of deepfakes such as their applications in visual effects, digital avatars, snapchat filters, creating voices of those who have lost theirs or updating episodes of movies without reshooting them . Deepfakes can have creative or productive impacts in photography, video games, virtual reality, movie productions, and entertainment, e.g., realistic video dubbing of foreign films, education through the reanimation of historical figures, virtually trying on clothes while shopping, and so on . However, the number of malicious uses of deepfakes largely dominates that of the positive ones. The development of advanced deep neural networks and the availability of large amount of data have made the forged images and videos almost indistinguishable to humans and even to sophisticated computer algorithms. The process of creating those manipulated images and videos is also much simpler today as it needs as little as an identity photo or a short video of a target individual. Less and less effort is required to produce a stunningly convincing tempered footage. Recent advances can even create a deepfake with just a still image . Deepfakes therefore can be a threat affecting not only public figures but also ordinary people. For example, a voice deepfake was used to scam a CEO out of \\$243,000 . A recent release of a software called DeepNude shows more disturbing threats as it can transform a person to a non-consensual porn . Likewise, the Chinese app Zao has gone viral lately as less-skilled users can swap their faces onto bodies of movie stars and insert themselves into well-known movies and TV clips . These forms of falsification create a huge threat to violation of privacy and identity, and affect many aspects of human lives.\n\\begin{figure}[ht]\n\\centering\n\\includegraphics[width=0.85\\columnwidth]{Fig1.pdf}\n\\caption{Number of papers related to deepfakes in years from 2016 to 2021, obtained from https://app.dimensions.ai at the end of 2021 with the search keyword ``deepfake\" applied to full text of scholarly papers.}\n\\label{fig0} \n\\end{figure}\nFinding the truth in digital domain therefore has become increasingly critical. It is even more challenging when dealing with deepfakes as they are majorly used to serve malicious purposes and almost anyone can create deepfakes these days using existing deepfake tools. Thus far, there have been numerous methods proposed to detect deepfakes . Most of them are based on deep learning, and thus a battle between malicious and positive uses of deep learning methods has been arising. To address the threat of face-swapping technology or deepfakes, the United States Defense Advanced Research Projects Agency (DARPA) initiated a research scheme in media forensics (named Media Forensics or MediFor) to accelerate the development of fake digital visual media detection methods . Recently, Facebook Inc. teaming up with Microsoft Corp and the Partnership on AI coalition have launched the Deepfake Detection Challenge to catalyse more research and development in detecting and preventing deepfakes from being used to mislead viewers . Data obtained from https://app.dimensions.ai at the end of 2021 show that the number of deepfake papers has increased significantly in recent years (Fig. \\ref{fig0}). Although the obtained numbers of deepfake papers may be lower than actual numbers but the research trend of this topic is obviously increasing.\nThere have been existing survey papers about creating and detecting deepfakes, presented in . For example, focused on reenactment approaches (i.e., to change a target’s expression, mouth, pose, gaze or body), and replacement approaches (i.e., to replace a target’s face by swap or transfer methods). separated detection approaches into conventional methods (e.g., blind methods without using any external data for training, one-class sensor-based and model-based methods, and supervised methods with handcrafted features) and deep learning-based approaches (e.g., CNN models). categorized both creation and detection methods based on the way deepfakes are created, including entire face synthesis, identity swap, attribute manipulation, and expression swap. On the other hand, we carry out the survey with a different perspective and taxonomy. We categorize the deepfake detection methods based on the data type, i.e., images or videos, as presented in Fig. \\ref{fig2}. With fake image detection methods, we focus on the features that are used, i.e., whether they are handcrafted features or deep features. With fake video detection methods, two main subcategories are identified based on whether the method uses temporal features across frames or visual artifacts within a video frame. We also discuss extensively the challenges, research trends and directions on deepfake detection and multimedia forensics problems.\n\\begin{figure}[!t]\n\\centering\n\\includegraphics[width=1.0\\columnwidth]{Fig2.pdf}\n\\caption{Categories of reviewed papers relevant to deepfake detection methods where we divide papers into two major groups, i.e., fake image detection and face video detection.}\n\\label{fig2} \n\\end{figure}","id":"c7a70291-a1e9-476f-8254-e897ee781284","level":"section","origin_cites_number":22,"parent_id":"44777335-fb2f-4b5d-83e4-47fe65116410","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0.6666666666666661,"cites":[7893,5134,62,7896,5138,5136,7895,96,5137,150,157,7894,5135,7892],"content":"\\label{sec2}\nDeepfakes have become popular due to the quality of tampered videos and also the easy-to-use ability of their applications to a wide range of users with various computer skills from professional to novice. These applications are mostly developed based on deep learning techniques. Deep learning is well known for its capability of representing complex and high-dimensional data. One variant of the deep networks with that capability is deep autoencoders, which have been widely applied for dimensionality reduction and image compression . The first attempt of deepfake creation was FakeApp, developed by a Reddit user using autoencoder-decoder pairing structure . In that method, the autoencoder extracts latent features of face images and the decoder is used to reconstruct the face images. To swap faces between source images and target images, there is a need of two encoder-decoder pairs where each pair is used to train on an image set, and the encoder's parameters are shared between two network pairs. In other words, two pairs have the same encoder network. This strategy enables the common encoder to find and learn the similarity between two sets of face images, which are relatively unchallenging because faces normally have similar features such as eyes, nose, mouth positions. Fig. \\ref{fig1} shows a deepfake creation process where the feature set of face A is connected with the decoder B to reconstruct face B from the original face A. This approach is applied in several works such as DeepFaceLab , DFaker , DeepFake\\_tf (tensorflow-based deepfakes) .\n\\begin{table*}\n\\centering\n\\begin{scriptsize}\n\\caption{Summary of notable deepfake tools}\n\\label{table1}\n\\begin{tabular}{p{0.10\\textwidth} p{0.29\\textwidth} p{0.54\\textwidth}}\n\\hline\n\\textbf{Tools} & \\textbf{Links} & \\textbf{Key Features}\\\\\n\\hline\nFaceswap\n&https://github.com/deepfakes/faceswap\n&- Using two encoder-decoder pairs.\\newline\n- Parameters of the encoder are shared.\\\\\n\\hline\nFaceswap-GAN\n&https://github.com/shaoanlu/faceswap-GAN\n&Adversarial loss and perceptual loss (VGGface) are added to an auto-encoder architecture.\\\\\n\\hline\nFew-Shot Face Translation\n& https://github.com/shaoanlu/fewshot-face-translation-GAN\n& - Use a pre-trained face recognition model to extract latent embeddings for GAN processing.\\newline\n- Incorporate semantic priors obtained by modules from FUNIT and SPADE .\\\\\n\\hline\nDeepFaceLab\n&https://github.com/iperov/DeepFaceLab\n&- Expand from the Faceswap method with new models, e.g. H64, H128, LIAEF128, SAE .\\newline\n- Support multiple face extraction modes, e.g. S3FD, MTCNN, dlib, or manual .\\\\\n\\hline\nDFaker\n&https://github.com/dfaker/df\n&- DSSIM loss function is used to reconstruct face.\\newline\n- Implemented based on Keras library.\\\\\n\\hline\nDeepFake\\_tf\n&https://github.com/StromWine/DeepFake\\_tf\n&Similar to DFaker but implemented based on tensorflow.\\\\\n\\hline\nAvatarMe\n& https://github.com/lattas/AvatarMe\n& -\tReconstruct 3D faces from arbitrary ``in-the-wild\" images.\\newline\n- Can reconstruct authentic 4K by 6K-resolution 3D faces from a single low-resolution image .\\\\\n\\hline\nMarioNETte\n& https://hyperconnect.github.io/MarioNETte\n& -\tA few-shot face reenactment framework that preserves the target identity.\\newline\n- No additional fine-tuning phase is needed for identity adaptation .\\\\\n\\hline\nDiscoFaceGAN\n& https://github.com/microsoft/DiscoFaceGAN\n& -\tGenerate face images of virtual people with independent latent variables of identity, expression, pose, and illumination.\\newline\n- Embed 3D priors into adversarial learning .\\\\\n\\hline\nStyleRig\n& https://gvv.mpi-inf.mpg.de/projects/StyleRig\n& -\tCreate portrait images of faces with a rig-like control over a pretrained and fixed StyleGAN via 3D morphable face models.\\newline\n- Self-supervised without manual annotations .\\\\\n\\hline\nFaceShifter\n& https://lingzhili.com/FaceShifterPage\n& -\tFace swapping in high-fidelity by exploiting and integrating the target attributes.\\newline\n- Can be applied to any new face pairs without requiring subject specific training .\\\\\n\\hline\nFSGAN\n& https://github.com/YuvalNirkin/fsgan\n& -\tA face swapping and reenactment model that can be applied to pairs of faces without requiring training on those faces.\\newline\n- Adjust to both pose and expression variations .\\\\\n\\hline\nStyleGAN\n& https://github.com/NVlabs/stylegan\n& -\tA new generator architecture for GANs is proposed based on style transfer literature.\\newline\n- The new architecture leads to automatic, unsupervised separation of high-level attributes and enables intuitive, scale-specific control of the synthesis of images .\\\\\n\\hline\nFace2Face & https://justusthies.github.io/posts/face2face/ & - Real-time facial reenactment of monocular target video sequence, e.g. Youtube video.\\newline\n- Animate the facial expressions of the target video by a source actor and re-render the manipulated output video in a photo-realistic fashion .\\\\\n\\hline\nNeural Textures & https://github.com/SSRSGJYD/NeuralTexture & \n- Feature maps that are learned as part of the scene capture process and stored as maps on top of {3D} mesh proxies.\\newline\n- Can coherently re-render or manipulate existing video content in both static and dynamic environments at real-time rates .\\\\\n\\hline\nTransformable Bottleneck Networks\n& https://github.com/kyleolsz/TB-Networks\n& -\tA method for fine-grained 3D manipulation of image content.\\newline\n- Apply spatial transformations in CNN models using a transformable bottleneck framework .\\\\\n\\hline\n``Do as I Do\" Motion \\newline Transfer\n& github.com/carolineec/EverybodyDanceNow\n& - Automatically transfer the motion from a source to a target person by learning a video-to-video translation.\\newline\n- Can create a motion-synchronized dancing video with multiple subjects .\\\\\n\\hline\nNeural Voice Puppetry\n& https://justusthies.github.io/posts/neural-voice-puppetry\n& - A method for audio-driven facial video synthesis.\\newline\n- Synthesize videos of a talking head from an audio sequence of another person using 3D face representation. .\\\\\n\\hline\n\\end{tabular}\n\\end{scriptsize}\n\\end{table*}\n\\begin{figure}[t]\n\\centering\n\\includegraphics[width=0.92\\columnwidth]{Fig3.pdf}\n\\caption{A deepfake creation model using two encoder-decoder pairs. Two networks use the same encoder but different decoders for training process (top). An image of face A is encoded with the common encoder and decoded with decoder B to create a deepfake (bottom). The reconstructed image (in the bottom) is the face B with the mouth shape of face A. Face B originally has the mouth of an upside-down heart while the reconstructed face B has the mouth of a conventional heart.}\n\\label{fig1} \n\\end{figure} \nBy adding adversarial loss and perceptual loss implemented in VGGFace to the encoder-decoder architecture, an improved version of deepfakes based on the generative adversarial network , i.e., faceswap-GAN, was proposed in . The VGGFace perceptual loss is added to make eye movements to be more realistic and consistent with input faces and help to smooth out artifacts in segmentation mask, leading to higher quality output videos. This model facilitates the creation of outputs with 64x64, 128x128, and 256x256 resolutions. In addition, the multi-task convolutional neural network (CNN) from the FaceNet implementation is used to make face detection more stable and face alignment more reliable. The CycleGAN is utilized for generative network implementation in this model.\n\\begin{figure}[!ht]\n\\centering\n\\includegraphics[width=0.55\\columnwidth]{Fig4.pdf}\n\\caption{The GAN architecture consisting of a generator and a discriminator, and each can be implemented by a neural network. The entire system can be trained with backpropagation that allows both networks to improve their capabilities.}\n\\label{figGAN} \n\\end{figure} \nA conventional GAN model comprises two neural networks: a generator and a discriminator as depicted in Fig. \\ref{figGAN}. Given a dataset of real images $x$ having a distribution of $p_{data}$, the aim of the generator $G$ is to produce images $G(z)$ similar to real images $x$ with $z$ being noise signals having a distribution of $p_z$. The aim of the discriminator $G$ is to correctly classify images generated by $G$ and real images $x$. The discriminator $D$ is trained to improve its classification capability, i.e., to maximize $D(x)$, which represents the probability that $x$ is a real image rather than a fake image generated by $G$. On the other hand, $G$ is trained to minimize the probability that its outputs are classified by $D$ as synthetic images, i.e., to minimize $1-D(G(z))$. This is a minimax game between two players $D$ and $G$ that can be described by the following value function :\n\\begin{multline}\n \\min_G \\max_D V(D,G)=\\EX_{x\\sim p_{data}(x)}[\\log D(x)] \\\\ + \\EX_{z\\sim p_z(z)}[\\log (1-D(G(z)))]\n\\end{multline}\nAfter sufficient training, both networks improve their capabilities, i.e., the generator $G$ is able to produce images that are really similar to real images while the discriminator $D$ is highly capable of distinguishing fake images from real ones.\nTable \\ref{table1} presents a summary of popular deepfake tools and their typical features. Among them, a prominent method for face synthesis based on a GAN model, namely StyleGAN, was introduced in . StyleGAN is motivated by style transfer with a special generator network architecture that is able to create realistic face images. In a traditional GAN model, e.g., the progressive growing of GAN (PGGAN) , the signal noise (latent code) is fed to the input layer of a feedforward network that represents the generator. In StyleGAN, there are two networks constructed and linked together, a mapping network $f$ and a synthesis network $g$. The latent code $z \\in Z$ is first converted to $w \\in W$ (where $W$ is an intermediate latent space) through a non-linear function $f:Z\\rightarrow W$, which is characterized by a neural network (i.e., the mapping network) consisting of several fully connected layers. Using an affine tranformation, the intermediate representation $w$ is specialized to styles $y=(y_s,y_b)$ that will be fed to the adaptive instance normalization (AdaIN) operations, specified as: \n\\begin{equation}\n \\mathrm{AdaIN}(x_i,y)=y_{s,i}\\frac{x_i-\\mu(x_i)}{\\sigma(x_i)}+y_{b,i}\n\\end{equation}\nwhere each feature map $x_i$ is normalized separately. The StyleGAN generator architecture allows controlling the image synthesis by modifying the styles via different scales. In addition, instead of using one random latent code during training, this method uses two latent codes to generate a given proportion of images. More specifically, two latent codes $z_1$ and $z_2$ are fed to the mapping network to create respectively $w_1$ and $w_2$ that control the styles by applying $w_1$ before and $w_2$ after the crossover point. Fig. \\ref{figMixing} demonstrates examples of images created by mixing two latent codes at three different scales where each subset of styles controls separate meaningful high-level attributes of the image. In other words, the generator architecture of StyleGAN is able to learn separation of high-level attributes (e.g., pose and identity when trained on human faces) and enables intuitive, scale-specific control of the face synthesis.\n\\begin{figure}[h]\n\\centering\n\\includegraphics[width=0.9\\columnwidth]{Fig5.pdf}\n\\caption{Examples of mixing styles using StyleGAN: the output images are generated by copying a specified subset of styles from source B and taking the rest from source A. a) Copying coarse styles from source B will generate images that have high-level aspects from source B and all colors and finer facial features from source A; b) if copying the styles of middle resolutions from B, the output images will have smaller scale facial features from B and preserve the pose, general face shape, and eyeglasses from A; c) if copying the fine styles from source B, the generated images will have the color scheme and microstructure of source B .}\n\\label{figMixing} \n\\end{figure}","id":"b11ed0ba-b192-4a4a-b063-d32fbaab94d4","level":"section","origin_cites_number":21,"parent_id":"44777335-fb2f-4b5d-83e4-47fe65116410","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Creation"]],"subsections":[],"title":"Deepfake Creation"},{"cite_extract_rate":0.33333333333333304,"cites":[1244,5139,5134],"content":"\\label{sec:3}\nDeepfake detection is normally deemed a binary classification problem where classifiers are used to classify between authentic videos and tampered ones. This kind of methods requires a large database of real and fake videos to train classification models. The number of fake videos is increasingly available, but it is still limited in terms of setting a benchmark for validating various detection methods. To address this issue, Korshunov and Marcel produced a notable deepfake dataset consisting of 620 videos based on the GAN model using the open source code Faceswap-GAN . Videos from the publicly available VidTIMIT database were used to generate low and high quality deepfake videos, which can effectively mimic the facial expressions, mouth movements, and eye blinking. These videos were then used to test various deepfake detection methods. Test results show that the popular face recognition systems based on VGG and Facenet are unable to detect deepfakes effectively. Other methods such as lip-syncing approaches and image quality metrics with support vector machine (SVM) produce very high error rate when applied to detect deepfake videos from this newly produced dataset. This raises concerns about the critical need of future development of more robust methods that can detect deepfakes from genuine.\nThis section presents a survey of deepfake detection methods where we group them into two major categories: fake image detection methods and fake video detection ones (Fig. \\ref{fig2}). The latter is distinguished into two smaller groups: \\emph{visual artifacts} within single video frame-based methods and \\emph{temporal features} across frames-based ones. Whilst most of the methods based on temporal features use deep learning \\emph{recurrent} classification models, the methods use visual artifacts within video frame can be implemented by either deep or shallow classifiers.","id":"7513ea5e-21ef-4144-b033-7ba34feb94da","level":"section","origin_cites_number":9,"parent_id":"44777335-fb2f-4b5d-83e4-47fe65116410","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"]],"subsections":["f1d90192-c843-4231-8ac1-201458f556a0","406b6759-531d-4af3-af75-8057a4f625fc"],"title":"Deepfake Detection"},{"cite_extract_rate":0.33333333333333304,"cites":[5140],"content":"Deepfakes are increasingly detrimental to privacy, society security and democracy . Methods for detecting deepfakes have been proposed as soon as this threat was introduced. Early attempts were based on handcrafted features obtained from artifacts and inconsistencies of the fake image synthesis process. Recent methods, e.g., , have commonly applied deep learning to automatically extract salient and discriminative features to detect deepfakes.","id":"f1d90192-c843-4231-8ac1-201458f556a0","level":"subsection","origin_cites_number":3,"parent_id":"7513ea5e-21ef-4144-b033-7ba34feb94da","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Image Detection"]],"subsections":["5d2250d1-1cae-45db-a279-c36ed6840045","87acc6e6-8b7c-49ef-9ce2-768ae70fea85"],"title":"Fake Image Detection"},{"cite_extract_rate":0.2,"cites":[7897,5141],"content":"Most works on detection of GAN generated images do not consider the generalization capability of the detection models although the development of GAN is ongoing, and many new extensions of GAN are frequently introduced. Xuan et al. used an image preprocessing step, e.g., Gaussian blur and Gaussian noise, to remove low level high frequency clues of GAN images. This increases the pixel level statistical similarity between real images and fake images and allows the forensic classifier to learn more intrinsic and meaningful features, which has better generalization capability than previous image forensics methods or image steganalysis networks .\nZhang et al. used the bag of words method to extract a set of compact features and fed it into various classifiers such as SVM , random forest (RF) and multi-layer perceptrons (MLP) for discriminating swapped face images from the genuine. Among deep learning-generated images, those synthesised by GAN models are probably most difficult to detect as they are realistic and high-quality based on GAN's capability to learn distribution of the complex input data and generate new outputs with similar input distribution. \nOn the other hand, Agarwal and Varshney cast the GAN-based deepfake detection as a hypothesis testing problem where a statistical framework was introduced using the information-theoretic study of authentication . The minimum distance between distributions of legitimate images and images generated by a particular GAN is defined, namely the oracle error. The analytic results show that this distance increases when the GAN is less accurate, and in this case, it is easier to detect deepfakes. In case of high-resolution image inputs, an extremely accurate GAN is required to generate fake images that are hard to detect by this method.","id":"5d2250d1-1cae-45db-a279-c36ed6840045","level":"subsubsection","origin_cites_number":10,"parent_id":"f1d90192-c843-4231-8ac1-201458f556a0","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Image Detection"],["subsubsection","Handcrafted Features-based Methods"]],"subsections":[],"title":"Handcrafted Features-based Methods"},{"cite_extract_rate":0.551724137931034,"cites":[1603,895,56,62,5142,63,5143,7194,8317,3212,1003,485,96,1600,78,7898],"content":"Face swapping has a number of compelling applications in video compositing, transfiguration in portraits, and especially in identity protection as it can replace faces in photographs by ones from a collection of stock images. However, it is also one of the techniques that cyber attackers employ to penetrate identification or authentication systems to gain illegitimate access. The use of deep learning such as CNN and GAN has made swapped face images more challenging for forensics models as it can preserve pose, facial expression and lighting of the photographs .\nHsu et al. introduced a two-phase deep learning method for detection of deepfake images. The first phase is a feature extractor based on the common fake feature network (CFFN) where the Siamese network architecture presented in is used. The CFFN encompasses several dense units with each unit including different numbers of dense blocks to improve the representative capability for the input images. \nDiscriminative features between the fake and real images are extracted through the CFFN learning process based on the use of pairwise information, which is the label of each pair of two input images. If the two images are of the same type, i.e., fake-fake or real-real, the pairwise label is $1$. In contrast, if they are of different types, i.e., fake-real, the pairwise label is $0$. The CFFN-based discriminative features are then fed to a neural network classifier to distinguish deceptive images from genuine. The proposed method is validated for both fake face and fake general image detection. On the one hand, the face dataset is obtained from CelebA , containing 10,177 identities and 202,599 aligned face images of various poses and background clutter. Five GAN variants are used to generate fake images with size of 64x64, including deep convolutional GAN (DCGAN) , Wasserstein GAN (WGAN) , WGAN with gradient penalty (WGAN-GP) , least squares GAN , and PGGAN . A total of 385,198 training images and 10,000 test images of both real and fake ones are obtained for validating the proposed method. On the other hand, the general dataset is extracted from the ILSVRC12 . The large scale GAN training model for high fidelity natural image synthesis (BIGGAN) , self-attention GAN and spectral normalization GAN are used to generate fake images with size of 128x128. The training set consists of 600,000 fake and real images whilst the test set includes 10,000 images of both types. Experimental results show the superior performance of the proposed method against its competing methods such as those introduced in .\nLikewise, proposed a CNN model, namely SCnet, to detect deepfake images, which are generated by the Glow-based facial forgery tool . The fake images synthesized by the Glow model have the facial expression maliciously tampered. These images are hyper-realistic with perfect visual qualities, but they still have subtle or noticeable manipulation traces, which are exploited by the SCnet. The SCnet is able to automatically learn high-level forensics features of image data thanks to a hierarchical feature extraction block, which is formed by stacking four convolutional layers. Each layer learns a new set of feature maps from the previous layer, with each convolutional operation is defined by:\n\\begin{equation}\n f_j^{(n)} = \\sum_{i=1}^i f_i^{(n-1)}*\\omega_{ij}^{(n)} + b_j^{(n)}\n\\end{equation}\nwhere $f_j^{(n)}$ is the $j^{th}$ feature map of the $n^{th}$ layer, $\\omega_{ij}^{(n)}$ is the weight of the $i^{th}$ channel of the $j^{th}$ convolutional kernel in the $n^{th}$ layer, and $b_j^{(n)}$ is the bias term of the $j^{th}$ convolutional kernel in the $n^{th}$ layer.\nThe proposed approach is evaluated using a dataset consisting of 321,378 face images, which are created by applying the Glow model to the CelebA face image dataset . Evaluation results show that the SCnet model obtains higher accuracy and better generalization than the Meso-4 model proposed in .\n\\begin{figure*}[!ht]\n\\centering\n\\includegraphics[width=1.75\\columnwidth]{Fig6.pdf}\n\\caption{A two-step process for face manipulation detection where the preprocessing step aims to detect, crop and align faces on a sequence of frames and the second step distinguishes manipulated and authentic face images by combining convolutional neural network (CNN) and recurrent neural network (RNN) .}\n\\label{fig3} \n\\end{figure*}\nRecently, proposed a method for deepfake detection using self-consistency of local source features, which are content-independent, spatially-local information of images. These features could come from either imaging pipelines, encoding methods or image synthesis approaches. The hypothesis is that a modified image would have different source features at different locations, while an original image will have the same source features across locations. These source features, represented in the form of down-sampled feature maps, are extracted by a CNN model using a special representation learning method called pairwise self-consistency learning. This learning method aims to penalize pairs of feature vectors that refer to locations from the same image for having a low cosine similarity score. At the same time, it also penalizes the pairs from different images for having a high similarity score. The learned feature maps are then fed to a classification method for deepfake detection. This proposed approach is evaluated on seven popular datasets, including FaceForensics++ , DeepfakeDetection , Celeb-DF-v1 \\& Celeb-DF-v2 , Deepfake Detection Challenge (DFDC) , DFDC Preview , and DeeperForensics-1.0 . Experimental results demonstrate that the proposed approach is superior to state-of-the-art methods. It however may have a limitation when dealing with fake images that are generated by methods that directly output the whole images whose source features are consistent across all positions within each image.","id":"87acc6e6-8b7c-49ef-9ce2-768ae70fea85","level":"subsubsection","origin_cites_number":29,"parent_id":"f1d90192-c843-4231-8ac1-201458f556a0","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Image Detection"],["subsubsection","Deep Features-based Methods"]],"subsections":[],"title":"Deep Features-based Methods"},{"cite_extract_rate":1,"cites":[1603],"content":"Most image detection methods cannot be used for videos because of the strong degradation of the frame data after video compression . Furthermore, videos have temporal characteristics that are varied among sets of frames and they are thus challenging for methods designed to detect only still fake images. This subsection focuses on deepfake video detection methods and categorizes them into two smaller groups: methods that employ temporal features and those that explore visual artifacts within frames.","id":"406b6759-531d-4af3-af75-8057a4f625fc","level":"subsection","origin_cites_number":1,"parent_id":"7513ea5e-21ef-4144-b033-7ba34feb94da","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Video Detection"]],"subsections":["f76aa108-55fe-4a3b-9495-cbb64d02a169","89bc57b3-3d0e-4686-9c89-9b9c741caaac"],"title":"Fake Video Detection"},{"cite_extract_rate":0.5454545454545451,"cites":[5142,1600,97,243,3758,96],"content":"Based on the observation that temporal coherence is not enforced effectively in the synthesis process of deepfakes, Sabir et al. leveraged the use of spatio-temporal features of video streams to detect deepfakes. Video manipulation is carried out on a frame-by-frame basis so that low level artifacts produced by face manipulations are believed to further manifest themselves as temporal artifacts with inconsistencies across frames. A recurrent convolutional model (RCN) was proposed based on the integration of the convolutional network DenseNet and the gated recurrent unit cells to exploit temporal discrepancies across frames (see Fig. \\ref{fig3}). The proposed method is tested on the FaceForensics++ dataset, which includes 1,000 videos , and shows promising results. \nLikewise, highlighted that deepfake videos contain intra-frame inconsistencies and temporal inconsistencies between frames. They then proposed the temporal-aware pipeline method that uses CNN and long short term memory (LSTM) to detect deepfake videos. CNN is employed to extract frame-level features, which are then fed into the LSTM to create a temporal sequence descriptor. A fully-connected network is finally used for classifying doctored videos from real ones based on the sequence descriptor as illustrated in Fig. \\ref{fig4}. An accuracy of greater than 97\\% was obtained using a dataset of 600 videos, including 300 deepfake videos collected from multiple video-hosting websites and 300 pristine videos randomly selected from the Hollywood human actions dataset in .\n\\begin{figure}[!ht]\n\\centering\n\\includegraphics[width=1\\columnwidth]{Fig7.pdf}\n\\caption{A deepfake detection method using convolutional neural network (CNN) and long short term memory (LSTM) to extract temporal features of a given video sequence, which are represented via the sequence descriptor. The detection network consisting of fully-connected layers is employed to take the sequence descriptor as input and calculate probabilities of the frame sequence belonging to either authentic or deepfake class .}\n\\label{fig4} \n\\end{figure}\nOn the other hand, the use of a physiological signal, eye blinking, to detect deepfakes was proposed in Li et al. based on the observation that a person in deepfakes has a lot less frequent blinking than that in untampered videos. A healthy adult human would normally blink somewhere between 2 to 10 seconds, and each blink would take 0.1 and 0.4 seconds. Deepfake algorithms, however, often use face images available online for training, which normally show people with open eyes, i.e., very few images published on the internet show people with closed eyes. Thus, without having access to images of people blinking, deepfake algorithms do not have the capability to generate fake faces that can blink normally. In other words, blinking rates in deepfakes are much lower than those in normal videos. To discriminate real and fake videos, Li et al. crop eye areas in the videos and distribute them into long-term recurrent convolutional networks (LRCN) for dynamic state prediction. The LRCN consists of a feature extractor based on CNN, a sequence learning based on long short term memory (LSTM), and a state prediction based on a fully connected layer to predict probability of eye open and close state. The eye blinking shows strong temporal dependencies and thus the implementation of LSTM helps to capture these temporal patterns effectively.\nRecently, proposed the use of optical flow to gauge the information along the temporal axis of a frame sequence for video deepfake detection. The optical flow is a vector field calculated on two temporal-distinct frames of a video that can describe the movement of objects in a scene. The optical flow fields are expected to be different between synthetically created frames and naturally generated ones . Unnatural movements of lips, eyes, or of the entire faces inserted into deepfake videos would introduce distinctive motion patterns when compared with pristine ones. Based on this assumption, features consisting of optical flow fields are fed into a CNN model for discriminating between deepfakes and original videos. More specifically, the ResNet50 architecture is implemented as a CNN model for experiments. The results obtained using the FaceForensics++ dataset show that this approach is comparable with state-of-the-art methods in terms of classification accuracy. A combination of this kind of feature with frame-based features is also experimented, which results in an improved deepfake detection performance. This demonstrates the usefulness of optical flow fields in capturing the inconsistencies on the temporal axis of video frames for deepfake detection.","id":"f76aa108-55fe-4a3b-9495-cbb64d02a169","level":"subsubsection","origin_cites_number":11,"parent_id":"406b6759-531d-4af3-af75-8057a4f625fc","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Video Detection"],["subsubsection","Temporal Features across Video Frames"]],"subsections":[],"title":"Temporal Features across Video Frames"},{"cite_extract_rate":0,"cites":[],"content":"As can be noticed in the previous subsection, the methods using temporal patterns across video frames are mostly based on deep recurrent network models to detect deepfake videos. This subsection investigates the other approach that normally decomposes videos into frames and explores visual artifacts within single frames to obtain discriminant features. These features are then distributed into either a deep or shallow classifier to differentiate between fake and authentic videos. We thus group methods in this subsection based on the types of classifiers, i.e. either deep or shallow.","id":"89bc57b3-3d0e-4686-9c89-9b9c741caaac","level":"subsubsection","origin_cites_number":0,"parent_id":"406b6759-531d-4af3-af75-8057a4f625fc","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Video Detection"],["subsubsection","Visual Artifacts within Video Frame"]],"subsections":["c8afcbd9-52a4-4801-8c8d-0c5d0d477915","b02a12ee-cc0c-427d-b853-a4c928ab1f56"],"title":"Visual Artifacts within Video Frame"},{"cite_extract_rate":0.642857142857142,"cites":[5144,1603,5145,7899,97,514,5146,5147,306],"content":"Deepfake videos are normally created with limited resolutions, which require an affine face warping approach (i.e., scaling, rotation and shearing) to match the configuration of the original ones. Because of the resolution inconsistency between the warped face area and the surrounding context, this process leaves artifacts that can be detected by CNN models such as VGG16 , ResNet50, ResNet101 and ResNet152 . A deep learning method to detect deepfakes based on the artifacts observed during the face warping step of the deepfake generation algorithms was proposed in . The proposed method is evaluated on two deepfake datasets, namely the UADFV and DeepfakeTIMIT. The UADFV dataset contains 49 real videos and 49 fake videos with 32,752 frames in total. The DeepfakeTIMIT dataset includes a set of low quality videos of 64 x 64 size and another set of high quality videos of 128 x 128 with totally 10,537 pristine images and 34,023 fabricated images extracted from 320 videos for each quality set. Performance of the proposed method is compared with other prevalent methods such as two deepfake detection MesoNet methods, i.e. Meso-4 and MesoInception-4 , HeadPose , and the face tampering detection method two-stream NN . Advantage of the proposed method is that it needs not to generate deepfake videos as negative examples before training the detection models. Instead, the negative examples are generated dynamically by extracting the face region of the original image and aligning it into multiple scales before applying Gaussian blur to a scaled image of random pick and warping back to the original image. This reduces a large amount of time and computational resources compared to other methods, which require deepfakes are generated in advance.\nNguyen et al. proposed the use of capsule networks for detecting manipulated images and videos. The capsule network was initially introduced to address limitations of CNNs when applied to inverse graphics tasks, which aim to find physical processes used to produce images of the world . The recent development of capsule network based on dynamic routing algorithm demonstrates its ability to describe the hierarchical pose relationships between object parts. This development is employed as a component in a pipeline for detecting fabricated images and videos as demonstrated in Fig. \\ref{fig5}. A dynamic routing algorithm is deployed to route the outputs of the three capsules to the output capsules through a number of iterations to separate between fake and real images. The method is evaluated through four datasets covering a wide range of forged image and video attacks. They include the well-known Idiap Research Institute replay-attack dataset , the deepfake face swapping dataset created by Afchar et al. , the facial reenactment FaceForensics dataset , produced by the Face2Face method , and the fully computer-generated image dataset generated by Rahmouni et al. . The proposed method yields the best performance compared to its competing methods in all of these datasets. This shows the potential of the capsule network in building a general detection system that can work effectively for various forged image and video attacks.\n\\begin{figure}[!ht]\n\\centering\n\\includegraphics[width=1\\columnwidth]{Fig8.pdf}\n\\caption{Capsule network takes features obtained from the VGG-19 network to distinguish fake images or videos from the real ones (top). The pre-processing step detects face region and scales it to the size of 128x128 before VGG-19 is used to extract latent features for the capsule network, which comprises three primary capsules and two output capsules, one for real and one for fake images (bottom). The statistical pooling constitutes an important part of capsule network that deals with forgery detection .}\n\\label{fig5} \n\\end{figure}","id":"c8afcbd9-52a4-4801-8c8d-0c5d0d477915","level":"paragraph","origin_cites_number":14,"parent_id":"89bc57b3-3d0e-4686-9c89-9b9c741caaac","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Video Detection"],["subsubsection","Visual Artifacts within Video Frame"],["paragraph","Deep classifiers"]],"subsections":[],"title":"Deep classifiers"},{"cite_extract_rate":0.512820512820512,"cites":[1255,56,7899,5148,5149,5151,485,5145,895,97,514,7900,5134,5153,306,7022,96,5152,5144,5156,1600,5155,1216,62,5142,1003,150,1603,7897,5150,5147,63,7194,8317,3212,59,5154,157,78,243],"content":"Deepfake detection methods mostly rely on the artifacts or inconsistency of intrinsic features between fake and real images or videos. Yang et al. proposed a detection method by observing the differences between 3D head poses comprising head orientation and position, which are estimated based on 68 facial landmarks of the central face region. The 3D head poses are examined because there is a shortcoming in the deepfake face generation pipeline. The extracted features are fed into an SVM classifier to obtain the detection results. Experiments on two datasets show the great performance of the proposed approach against its competing methods. The first dataset, namely UADFV, consists of 49 deep fake videos and their respective real videos . The second dataset comprises 241 real images and 252 deep fake images, which is a subset of data used in the DARPA MediFor GAN Image/Video Challenge . Likewise, a method to exploit artifacts of deepfakes and face manipulations based on visual features of eyes, teeth and facial contours was studied in . The visual artifacts arise from lacking global consistency, wrong or imprecise estimation of the incident illumination, or imprecise estimation of the underlying geometry. For deepfakes detection, missing reflections and missing details in the eye and teeth areas are exploited as well as texture features extracted from the facial region based on facial landmarks. Accordingly, the eye feature vector, teeth feature vector and features extracted from the full-face crop are used. After extracting the features, two classifiers including logistic regression and small neural network are employed to classify the deepfakes from real videos. Experiments carried out on a video dataset downloaded from YouTube show the best result of 0.851 in terms of the area under the receiver operating characteristics curve. The proposed method however has a disadvantage that requires images meeting certain prerequisite such as open eyes or visual teeth.\nThe use of photo response non uniformity (PRNU) analysis was proposed in to detect deepfakes from authentic ones. PRNU is a component of sensor pattern noise, which is attributed to the manufacturing imperfection of silicon wafers and the inconsistent sensitivity of pixels to light because of the variation of the physical characteristics of the silicon wafers. The PRNU analysis is widely used in image forensics and advocated to use in because the swapped face is supposed to alter the local PRNU pattern in the facial area of video frames. The videos are converted into frames, which are cropped to the questioned facial region. The cropped frames are then separated sequentially into eight groups where an average PRNU pattern is computed for each group. Normalised cross correlation scores are calculated for comparisons of PRNU patterns among these groups. A test dataset was created, consisting of 10 authentic videos and 16 manipulated videos, where the fake videos were produced from the genuine ones by the DeepFaceLab tool . The analysis shows a significant statistical difference in terms of mean normalised cross correlation scores between deepfakes and the genuine. This analysis therefore suggests that PRNU has a potential in deepfake detection although a larger dataset would need to be tested. \nWhen seeing a video or image with suspicion, users normally want to search for its origin. However, there is currently no feasibility for such a tool. Hasan and Salah proposed the use of blockchain and smart contracts to help users detect deepfake videos based on the assumption that videos are only real when their sources are traceable. Each video is associated with a smart contract that links to its parent video and each parent video has a link to its child in a hierarchical structure. Through this chain, users can credibly trace back to the original smart contract associated with pristine video even if the video has been copied multiple times. An important attribute of the smart contract is the unique hashes of the interplanetary file system, which is used to store video and its metadata in a decentralized and content-addressable manner . The smart contract's key features and functionalities are tested against several common security challenges such as distributed denial of services, replay and man in the middle attacks to ensure the solution meeting security requirements. This approach is generic, and it can be extended to other types of digital content, e.g., images, audios and manuscripts.\n\\begin{table*}[h!]\n\\centering\n\\caption{Summary of prominent deepfake detection methods}\n\\label{table2}\n\\begin{scriptsize}\n\\begin{tabularx}{\\linewidth}{p{0.11\\textwidth} p{0.09\\textwidth} p{0.33\\textwidth} p{0.04\\textwidth} p{0.32\\textwidth}}\n\\toprule\n\\textbf{Methods} & \\textbf{Classifiers/\\newline Techniques} & \\textbf{Key Features} & \\textbf{Dealing with} & \\textbf{Datasets Used}\\\\\n\\midrule\nEye blinking \n&LRCN\n&- Use LRCN to learn the temporal patterns of eye blinking.\\newline\n- Based on the observation that blinking frequency of deepfakes is much smaller than normal.\n&Videos\n&Consist of 49 interview and presentation videos, and their corresponding generated deepfakes.\\\\\n\\hline\nIntra-frame and temporal inconsistencies \n&CNN and LSTM\n&CNN is employed to extract frame-level features, which are distributed to LSTM to construct sequence descriptor useful for classification.\t\n&Videos\n&A collection of 600 videos obtained from multiple websites.\\\\\n\\hline\nUsing face warping artifacts \n&VGG16 ,\\newline ResNet models \n&Artifacts are discovered using CNN models based on resolution inconsistency between the warped face area and the surrounding context.\n&Videos\n&- UADFV , containing 49 real videos and 49 fake videos with 32752 frames in total.\\newline\n- DeepfakeTIMIT \\\\\n\\hline\nMesoNet \n&CNN\n&- Two deep networks, i.e. Meso-4 and MesoInception-4 are introduced to examine deepfake videos at the mesoscopic analysis level.\\newline\n- Accuracy obtained on deepfake and FaceForensics datasets are 98\\% and 95\\% respectively.\n&Videos\n&Two datasets: deepfake one constituted from online videos and the FaceForensics one created by the Face2Face approach .\\\\\n\\hline\nEye, teach and facial texture \n&Logistic regression and neural network (NN)\n&- Exploit facial texture differences, and missing reflections and details in eye and teeth areas of deepfakes.\\newline\n- Logistic regression and NN are used for classifying.\t\n&Videos\n&A video dataset downloaded from YouTube.\\\\\n\\hline\nSpatio-temporal features with RCN \n&RCN\n&Temporal discrepancies across frames are explored using RCN that integrates convolutional network DenseNet and the gated recurrent unit cells \n&Videos\n&FaceForensics++ dataset, including 1,000 videos .\\\\\n\\hline\nSpatio-temporal features with LSTM \n& Convolutional bidirectional recurrent LSTM network\n& - An XceptionNet CNN is used for facial feature extraction while audio embeddings are obtained by stacking multiple convolution modules.\\newline\n- Two loss functions, i.e. cross-entropy and Kullback-Leibler divergence, are used.\n& Videos\n& FaceForensics++ and Celeb-DF (5,639 deepfake videos) datasets and the ASVSpoof 2019 Logical Access audio dataset .\\\\\n\\hline\nAnalysis of PRNU \n&PRNU\n&- Analysis of noise patterns of light sensitive sensors of digital cameras due to their factory defects.\\newline\n- Explore the differences of PRNU patterns between the authentic and deepfake videos because face swapping is believed to alter the local PRNU patterns.\n&Videos\t\n&Created by the authors, including 10 authentic and 16 deepfake videos using DeepFaceLab .\\\\\n\\hline\nPhoneme-viseme mismatches \n& CNN\n& - Exploit the mismatches between the dynamics of the mouth shape, i.e. visemes, with a spoken phoneme.\\newline\n- Focus on sounds associated with the M, B and P phonemes as they require complete mouth closure while deepfakes often incorrectly synthesize it.\n& Videos\n& Four in-the-wild lip-sync deepfakes from Instagram and YouTube (www.instagram.com/bill\\_posters\\_uk and youtu.be/VWMEDacz3L4) and others are created using synthesis techniques, i.e. Audio-to-Video (A2V) and Text-to-Video (T2V) .\\\\\n\\hline\nUsing attribution-based confidence (ABC) metric \n& ResNet50 model , pre-trained on VGGFace2 \n& - The ABC metric is used to detect deepfake videos without accessing to training data.\\newline\n- ABC values obtained for original videos are greater than 0.94 while those of deepfakes have low ABC values.\n& Videos\n& VidTIMIT and two other original datasets obtained from the COHFACE (https://www.idiap.ch/dataset/cohface) and from YouTube. datasets from COHFACE and YouTube are used to generate two deepfake datasets by commercial website https://deepfakesweb.com and another deepfake dataset is DeepfakeTIMIT .\\\\\n\\hline\nUsing appearance and behaviour \n& Rules based on facial and behavioural features.\n& Temporal, behavioral biometric based on facial expressions and head movements are learned using ResNet-101 while static facial biometric is obtained using VGG .\n& Videos\n& The world leaders dataset , FaceForensics++ , Google/Jigsaw deepfake detection dataset , DFDC and Celeb-DF .\\\\\n\\hline\nFakeCatcher \n& CNN\n& Extract biological signals in portrait videos and use them as an implicit descriptor of authenticity because they are not spatially and temporally well-preserved in deepfakes.\n& Videos\n& UADFV , FaceForensics , FaceForensics++ , Celeb-DF , and a new dataset of 142 videos, independent of the generative model, resolution, compression, content, and context.\\\\\n\\hline\nEmotion audio-visual affective cues \n& Siamese network \n& Modality and emotion embedding vectors for the face and speech are extracted for deepfake detection.\n& Videos\n& DeepfakeTIMIT and DFDC .\\\\\n\\hline\nHead poses \n&SVM\n&- Features are extracted using 68 landmarks of the face region.\\newline\n- Use SVM to classify using the extracted features.\t\n&Videos/\\newline Images\n&- UADFV consists of 49 deep fake videos and their respective real videos.\\newline\n- 241 real images and 252 deep fake images from DARPA MediFor GAN Image/Video Challenge.\\\\\n\\hline\nCapsule-forensics \n&Capsule networks\n&- Latent features extracted by VGG-19 network are fed into the capsule network for classification.\\newline\n- A dynamic routing algorithm is used to route the outputs of three convolutional capsules to two output capsules, one for fake and another for real images, through a number of iterations.\t\n&Videos/\\newline Images\n&Four datasets: the Idiap Research Institute replay-attack , deepfake face swapping by , facial reenactment FaceForensics , and fully computer-generated image set using .\\\\\n\\hline\n\\end{tabularx}\n\\end{scriptsize}\n\\end{table*}\n\\begin{table*}[h!]\n\\centering\n\\begin{scriptsize}\n\\begin{tabularx}{\\linewidth}{p{0.1\\textwidth} p{0.1\\textwidth} p{0.33\\textwidth} p{0.04\\textwidth} p{0.32\\textwidth}}\n\\toprule\n\\textbf{Methods} & \\textbf{Classifiers/\\newline Techniques} & \\textbf{Key Features} & \\textbf{Dealing with} & \\textbf{Datasets Used}\\\\\n\\midrule\nPreprocessing combined with deep network \n&DCGAN,\nWGAN-GP and PGGAN.\n&- Enhance generalization ability of deep learning models to detect GAN generated images.\\newline\n- Remove low level features of fake images.\\newline\n- Force deep networks to focus more on pixel level similarity between fake and real images to improve generalization ability.\n&Images\n&- Real dataset: CelebA-HQ , including high quality face images of 1024x1024 resolution.\\newline\n- Fake datasets: generated by DCGAN , WGAN-GP and PGGAN .\\\\\n\\hline\nAnalyzing convolutional traces \n& KNN, SVM, and linear discriminant analysis (LDA)\n& Using expectation-maximization algorithm to extract local features pertaining to convolutional generative process of GAN-based image deepfake generators.\n& Images\n& Authentic images from CelebA and corresponding deepfakes are created by five different GANs (group-wise deep whitening-and-coloring transformation GDWCT , StarGAN ,\nAttGAN , StyleGAN , StyleGAN2 ).\\\\\n\\hline\nBag of words and shallow classifiers \t\n&SVM, RF, MLP\n&Extract discriminant features using bag of words method and feed these features into SVM, RF and MLP for binary classification: innocent vs fabricated.\n&Images\n&The well-known LFW face database , containing 13,223 images with resolution of 250x250.\\\\\n\\hline\nPairwise learning \n&CNN concatenated to CFFN\n&Two-phase procedure: feature extraction using CFFN based on the Siamese network architecture and classification using CNN.\n&Images\n&- Face images: real ones from CelebA , and fake ones generated by DCGAN , WGAN , WGAN-GP , least squares GAN , and PGGAN .\\newline\n- General images: real ones from ILSVRC12 , and fake ones generated by BIGGAN , self-attention GAN and spectral normalization GAN .\\\\\n\\hline\nDefenses against adversarial perturbations in deepfakes \n& VGG and ResNet \n& - Introduce adversarial perturbations to enhance deepfakes and fool deepfake detectors.\\newline\n- Improve accuracy of deepfake detectors using Lipschitz regularization and deep image prior techniques.\n& Images\n& 5,000 real images from CelebA and 5,000 fake images created by the “Few-Shot Face Translation GAN” method .\\\\\n\\hline\nFace X-ray \n& CNN\n& - Try to locate the blending boundary between the target and original faces instead of capturing the synthesized artifacts of specific manipulations.\\newline\n- Can be trained without fake images.\n& Images\n& FaceForensics++ , DeepfakeDetection (DFD) , DFDC and Celeb-DF .\\\\\n\\hline\nUsing common artifacts of CNN-generated images \n& ResNet-50 \npre-trained with ImageNet \n& Train the classifier using a large number of fake images generated by a high-performing unconditional GAN model, i.e., PGGAN and evaluate how well the classifier generalizes to other CNN-synthesized images.\n& Images\n& A new dataset of CNN-generated images, namely ForenSynths, consisting of\nsynthesized images from 11 models such as StyleGAN , super-resolution methods and FaceForensics++ .\\\\\n\\hline\nUsing convolutional traces on GAN-based images \n& KNN, SVM, and LDA\n& Training the expectation-maximization algorithm to detect\nand extract discriminative features via a fingerprint that represents the convolutional traces left by GANs during image generation.\n& Images\n& A dataset of images generated by ten GAN models, including CycleGAN , StarGAN , AttGAN , GDWCT , StyleGAN , StyleGAN2 , PGGAN , FaceForensics++ , IMLE , and\nSPADE .\\\\\n\\hline\nUsing deep features extracted by CNN \n& A new CNN model, namely SCnet\n& The CNN-based SCnet is able to automatically learn high-level forensics features of image data thanks to a hierarchical feature extraction block, which is formed by stacking four convolutional layers.\n& Images\n& A dataset of 321,378 face images, created by applying the Glow model to the CelebA face image dataset .\n\\\\\n\\bottomrule\n\\end{tabularx}\n\\end{scriptsize}\n\\end{table*}","id":"b02a12ee-cc0c-427d-b853-a4c928ab1f56","level":"paragraph","origin_cites_number":78,"parent_id":"89bc57b3-3d0e-4686-9c89-9b9c741caaac","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Deepfake Detection"],["subsection","Fake Video Detection"],["subsubsection","Visual Artifacts within Video Frame"],["paragraph","Shallow classifiers"]],"subsections":[],"title":"Shallow classifiers"},{"cite_extract_rate":0.4,"cites":[7901,5160,5157,5158,5156,5161,5162,5159],"content":"\\label{sec:4}\nWith the support of deep learning, deepfakes can be created easier than ever before. The spread of these fake contents is also quicker thanks to the development of social media platforms . Sometimes deepfakes do not need to be spread to massive audience to cause detrimental effects. People who create deepfakes with malicious purpose only need to deliver them to target audiences as part of their sabotage strategy without using social media. For example, this approach can be utilized by intelligence services trying to influence decisions made by important people such as politicians, leading to national and international security threats . Catching the deepfake alarming problem, research community has focused on developing deepfake detection algorithms and numerous results have been reported. This paper has reviewed the state-of-the-art methods and a summary of typical approaches is provided in Table \\ref{table2}. It is noticeable that a battle between those who use advanced machine learning to create deepfakes with those who make effort to detect deepfakes is growing.\nDeepfakes' quality has been increasing and the performance of detection methods needs to be improved accordingly. The inspiration is that what AI has broken can be fixed by AI as well . Detection methods are still in their early stage and various methods have been proposed and evaluated but using fragmented datasets. An approach to improve performance of detection methods is to create a growing updated benchmark dataset of deepfakes to validate the ongoing development of detection methods. This will facilitate the training process of detection models, especially those based on deep learning, which requires a large training set . \nImproving performance of deepfake detection methods is important, especially in cross-forgery and cross-dataset scenarios. Most detection models are designed and evaluated in the same-forgery and in-dataset experiments, which do not ensure their generalization capability. Some previous studies have addressed this issue, e.g., in , but more work needs to be done in this direction. A model trained on a specific forgery needs to be able to work against another unknown one because potential deepfake types are not normally known in the real-world scenarios. Likewise, current detection methods mostly focus on drawbacks of the deepfake generation pipelines, i.e., finding weakness of the competitors to attack them. This kind of information and knowledge is not always available in adversarial environments where attackers commonly attempt not to reveal such deepfake creation technologies. Recent works on adversarial perturbation attacks to fool DNN-based detectors make the deepfake detection task more difficult . These are real challenges for detection method development and a future study needs to focus on introducing more robust, scalable and generalizable methods.\nAnother research direction is to integrate detection methods into distribution platforms such as social media to increase its effectiveness in dealing with the widespread impact of deepfakes. The screening or filtering mechanism using effective detection methods can be implemented on these platforms to ease the deepfakes detection . Legal requirements can be made for tech companies who own these platforms to remove deepfakes quickly to reduce its impacts. In addition, watermarking tools can also be integrated into devices that people use to make digital contents to create immutable metadata for storing originality details such as time and location of multimedia contents as well as their untampered attestment . This integration is difficult to implement but a solution for this could be the use of the disruptive blockchain technology. The blockchain has been used effectively in many areas and there are very few studies so far addressing the deepfake detection problems based on this technology. As it can create a chain of unique unchangeable blocks of metadata, it is a great tool for digital provenance solution. The integration of blockchain technologies to this problem has demonstrated certain results but this research direction is far from mature.\nUsing detection methods to spot deepfakes is crucial, but understanding the real intent of people publishing deepfakes is even more important. This requires the judgement of users based on social context in which deepfake is discovered, e.g. who distributed it and what they said about it . This is critical as deepfakes are getting more and more photorealistic and it is highly anticipated that detection software will be lagging behind deepfake creation technology. A study on social context of deepfakes to assist users in such judgement is thus worth performing. \nVideos and photographics have been widely used as evidences in police investigation and justice cases. They may be introduced as evidences in a court of law by digital media forensics experts who have background in computer or law enforcement and experience in collecting, examining and analysing digital information. The development of machine learning and AI technologies might have been used to modify these digital contents and thus the experts' opinions may not be enough to authenticate these evidences because even experts are unable to discern manipulated contents. This aspect needs to take into account in courtrooms nowadays when images and videos are used as evidences to convict perpetrators because of the existence of a wide range of digital manipulation methods . The digital media forensics results therefore must be proved to be valid and reliable before they can be used in courts. This requires careful documentation for each step of the forensics process and how the results are reached. Machine learning and AI algorithms can be used to support the determination of the authenticity of digital media and have obtained accurate and reliable results, e.g., , but most of these algorithms are unexplainable. This creates a huge hurdle for the applications of AI in forensics problems because not only the forensics experts oftentimes do not have expertise in computer algorithms, but the computer professionals also cannot explain the results properly as most of these algorithms are black box models . This is more critical as the most recent models with the most accurate results are based on deep learning methods consisting of many neural network parameters. Researchers have recently attempted to create white box and explainable detection methods. An example is the approach proposed by in which they use discrete cosine transform statistics to detect so-called specific GAN frequencies to differentiate between real images and deepfakes. Through the analysis of particular frequency statistics, that method can be used to mathematically explain whether a multimedia content is a deepfake and why it is. More research must be conducted in this area and explainable AI in computer vision therefore is a research direction that is needed to promote and utilize the advances and advantages of AI and machine learning in digital media forensics.","id":"af5e6522-a34d-4deb-8b41-532983a76f29","level":"section","origin_cites_number":20,"parent_id":"44777335-fb2f-4b5d-83e4-47fe65116410","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Discussions and Future Research Directions"]],"subsections":[],"title":"Discussions and Future Research Directions"},{"cite_extract_rate":0,"cites":[],"content":"Deepfakes have begun to erode trust of people in media contents as seeing them is no longer commensurate with believing in them. They could cause distress and negative effects to those targeted, heighten disinformation and hate speech, and even could stimulate political tension, inflame the public, violence or war. This is especially critical nowadays as the technologies for creating deepfakes are increasingly approachable and social media platforms can spread those fake contents quickly. This survey provides a timely overview of deepfake creation and detection methods and presents a broad discussion on challenges, potential trends, and future directions in this area. This study therefore will be valuable for the artificial intelligence research community to develop effective methods for tackling deepfakes.\n\\section*{Declaration of Competing Interest}\nAuthors declare no conflict of interest.\n\\bibliography{cas-refs}\n\\end{document}\n\\endinput","id":"6f55771a-c73a-46f3-b17e-1720217d26ca","level":"section","origin_cites_number":0,"parent_id":"44777335-fb2f-4b5d-83e4-47fe65116410","prefix_titles":[["title","Deep Learning for Deepfakes Creation and Detection: A Survey"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"44777335-fb2f-4b5d-83e4-47fe65116410\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ]\n ],\n \"subsections\": [\n \"c7a70291-a1e9-476f-8254-e897ee781284\",\n \"b11ed0ba-b192-4a4a-b063-d32fbaab94d4\",\n \"7513ea5e-21ef-4144-b033-7ba34feb94da\",\n \"af5e6522-a34d-4deb-8b41-532983a76f29\",\n \"6f55771a-c73a-46f3-b17e-1720217d26ca\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5133,\n 5128,\n 5132,\n 5130,\n 5131,\n 5680,\n 5129,\n 1602,\n 986,\n 1256,\n 4939\n ],\n \"content\": \"\\\\label{sec1}\\nIn a narrow definition, deepfakes (stemming from ``deep learning\\\" and ``fake\\\") are created by techniques that can superimpose face images of a target person onto a video of a source person to make a video of the target person doing or saying things the source person does. This constitutes a category of deepfakes, namely \\\\emph{face-swap}. In a broader definition, deepfakes are artificial intelligence-synthesized content that can also fall into two other categories, i.e., \\\\emph{lip-sync} and \\\\emph{puppet-master}. Lip-sync deepfakes refer to videos that are modified to make the mouth movements consistent with an audio recording. Puppet-master deepfakes include videos of a target person (puppet) who is animated following the facial expressions, eye and head movements of another person (master) sitting in front of a camera .\\nWhile some deepfakes can be created by traditional visual effects or computer-graphics approaches, the recent common underlying mechanism for deepfake creation is deep learning models such as autoencoders and generative adversarial networks (GANs), which have been applied widely in the computer vision domain . These models are used to examine facial expressions and movements of a person and synthesize facial images of another person making analogous expressions and movements . Deepfake methods normally require a large amount of image and video data to train models to create photo-realistic images and videos. As public figures such as celebrities and politicians may have a large number of videos and images available online, they are initial targets of deepfakes. Deepfakes were used to swap faces of celebrities or politicians to bodies in porn images and videos. The first deepfake video emerged in 2017 where face of a celebrity was swapped to the face of a porn actor. It is threatening to world security when deepfake methods can be employed to create videos of world leaders with fake speeches for falsification purposes . Deepfakes therefore can be abused to cause political or religion tensions between countries, to fool public and affect results in election campaigns, or create chaos in financial markets by creating fake news . It can be even used to generate fake satellite images of the Earth to contain objects that do not really exist to confuse military analysts, e.g., creating a fake bridge across a river although there is no such a bridge in reality. This can mislead a troop who have been guided to cross the bridge in a battle .\\nAs the democratization of creating realistic digital humans has positive implications, there is also positive use of deepfakes such as their applications in visual effects, digital avatars, snapchat filters, creating voices of those who have lost theirs or updating episodes of movies without reshooting them . Deepfakes can have creative or productive impacts in photography, video games, virtual reality, movie productions, and entertainment, e.g., realistic video dubbing of foreign films, education through the reanimation of historical figures, virtually trying on clothes while shopping, and so on . However, the number of malicious uses of deepfakes largely dominates that of the positive ones. The development of advanced deep neural networks and the availability of large amount of data have made the forged images and videos almost indistinguishable to humans and even to sophisticated computer algorithms. The process of creating those manipulated images and videos is also much simpler today as it needs as little as an identity photo or a short video of a target individual. Less and less effort is required to produce a stunningly convincing tempered footage. Recent advances can even create a deepfake with just a still image . Deepfakes therefore can be a threat affecting not only public figures but also ordinary people. For example, a voice deepfake was used to scam a CEO out of \\\\$243,000 . A recent release of a software called DeepNude shows more disturbing threats as it can transform a person to a non-consensual porn . Likewise, the Chinese app Zao has gone viral lately as less-skilled users can swap their faces onto bodies of movie stars and insert themselves into well-known movies and TV clips . These forms of falsification create a huge threat to violation of privacy and identity, and affect many aspects of human lives.\\n\\\\begin{figure}[ht]\\n\\\\centering\\n\\\\includegraphics[width=0.85\\\\columnwidth]{Fig1.pdf}\\n\\\\caption{Number of papers related to deepfakes in years from 2016 to 2021, obtained from https://app.dimensions.ai at the end of 2021 with the search keyword ``deepfake\\\" applied to full text of scholarly papers.}\\n\\\\label{fig0} \\n\\\\end{figure}\\nFinding the truth in digital domain therefore has become increasingly critical. It is even more challenging when dealing with deepfakes as they are majorly used to serve malicious purposes and almost anyone can create deepfakes these days using existing deepfake tools. Thus far, there have been numerous methods proposed to detect deepfakes . Most of them are based on deep learning, and thus a battle between malicious and positive uses of deep learning methods has been arising. To address the threat of face-swapping technology or deepfakes, the United States Defense Advanced Research Projects Agency (DARPA) initiated a research scheme in media forensics (named Media Forensics or MediFor) to accelerate the development of fake digital visual media detection methods . Recently, Facebook Inc. teaming up with Microsoft Corp and the Partnership on AI coalition have launched the Deepfake Detection Challenge to catalyse more research and development in detecting and preventing deepfakes from being used to mislead viewers . Data obtained from https://app.dimensions.ai at the end of 2021 show that the number of deepfake papers has increased significantly in recent years (Fig. \\\\ref{fig0}). Although the obtained numbers of deepfake papers may be lower than actual numbers but the research trend of this topic is obviously increasing.\\nThere have been existing survey papers about creating and detecting deepfakes, presented in . For example, focused on reenactment approaches (i.e., to change a target’s expression, mouth, pose, gaze or body), and replacement approaches (i.e., to replace a target’s face by swap or transfer methods). separated detection approaches into conventional methods (e.g., blind methods without using any external data for training, one-class sensor-based and model-based methods, and supervised methods with handcrafted features) and deep learning-based approaches (e.g., CNN models). categorized both creation and detection methods based on the way deepfakes are created, including entire face synthesis, identity swap, attribute manipulation, and expression swap. On the other hand, we carry out the survey with a different perspective and taxonomy. We categorize the deepfake detection methods based on the data type, i.e., images or videos, as presented in Fig. \\\\ref{fig2}. With fake image detection methods, we focus on the features that are used, i.e., whether they are handcrafted features or deep features. With fake video detection methods, two main subcategories are identified based on whether the method uses temporal features across frames or visual artifacts within a video frame. We also discuss extensively the challenges, research trends and directions on deepfake detection and multimedia forensics problems.\\n\\\\begin{figure}[!t]\\n\\\\centering\\n\\\\includegraphics[width=1.0\\\\columnwidth]{Fig2.pdf}\\n\\\\caption{Categories of reviewed papers relevant to deepfake detection methods where we divide papers into two major groups, i.e., fake image detection and face video detection.}\\n\\\\label{fig2} \\n\\\\end{figure}\",\n \"id\": \"c7a70291-a1e9-476f-8254-e897ee781284\",\n \"level\": \"section\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"44777335-fb2f-4b5d-83e4-47fe65116410\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 7893,\n 5134,\n 62,\n 7896,\n 5138,\n 5136,\n 7895,\n 96,\n 5137,\n 150,\n 157,\n 7894,\n 5135,\n 7892\n ],\n \"content\": \"\\\\label{sec2}\\nDeepfakes have become popular due to the quality of tampered videos and also the easy-to-use ability of their applications to a wide range of users with various computer skills from professional to novice. These applications are mostly developed based on deep learning techniques. Deep learning is well known for its capability of representing complex and high-dimensional data. One variant of the deep networks with that capability is deep autoencoders, which have been widely applied for dimensionality reduction and image compression . The first attempt of deepfake creation was FakeApp, developed by a Reddit user using autoencoder-decoder pairing structure . In that method, the autoencoder extracts latent features of face images and the decoder is used to reconstruct the face images. To swap faces between source images and target images, there is a need of two encoder-decoder pairs where each pair is used to train on an image set, and the encoder's parameters are shared between two network pairs. In other words, two pairs have the same encoder network. This strategy enables the common encoder to find and learn the similarity between two sets of face images, which are relatively unchallenging because faces normally have similar features such as eyes, nose, mouth positions. Fig. \\\\ref{fig1} shows a deepfake creation process where the feature set of face A is connected with the decoder B to reconstruct face B from the original face A. This approach is applied in several works such as DeepFaceLab , DFaker , DeepFake\\\\_tf (tensorflow-based deepfakes) .\\n\\\\begin{table*}\\n\\\\centering\\n\\\\begin{scriptsize}\\n\\\\caption{Summary of notable deepfake tools}\\n\\\\label{table1}\\n\\\\begin{tabular}{p{0.10\\\\textwidth} p{0.29\\\\textwidth} p{0.54\\\\textwidth}}\\n\\\\hline\\n\\\\textbf{Tools} & \\\\textbf{Links} & \\\\textbf{Key Features}\\\\\\\\\\n\\\\hline\\nFaceswap\\n&https://github.com/deepfakes/faceswap\\n&- Using two encoder-decoder pairs.\\\\newline\\n- Parameters of the encoder are shared.\\\\\\\\\\n\\\\hline\\nFaceswap-GAN\\n&https://github.com/shaoanlu/faceswap-GAN\\n&Adversarial loss and perceptual loss (VGGface) are added to an auto-encoder architecture.\\\\\\\\\\n\\\\hline\\nFew-Shot Face Translation\\n& https://github.com/shaoanlu/fewshot-face-translation-GAN\\n& - Use a pre-trained face recognition model to extract latent embeddings for GAN processing.\\\\newline\\n- Incorporate semantic priors obtained by modules from FUNIT and SPADE .\\\\\\\\\\n\\\\hline\\nDeepFaceLab\\n&https://github.com/iperov/DeepFaceLab\\n&- Expand from the Faceswap method with new models, e.g. H64, H128, LIAEF128, SAE .\\\\newline\\n- Support multiple face extraction modes, e.g. S3FD, MTCNN, dlib, or manual .\\\\\\\\\\n\\\\hline\\nDFaker\\n&https://github.com/dfaker/df\\n&- DSSIM loss function is used to reconstruct face.\\\\newline\\n- Implemented based on Keras library.\\\\\\\\\\n\\\\hline\\nDeepFake\\\\_tf\\n&https://github.com/StromWine/DeepFake\\\\_tf\\n&Similar to DFaker but implemented based on tensorflow.\\\\\\\\\\n\\\\hline\\nAvatarMe\\n& https://github.com/lattas/AvatarMe\\n& -\\tReconstruct 3D faces from arbitrary ``in-the-wild\\\" images.\\\\newline\\n- Can reconstruct authentic 4K by 6K-resolution 3D faces from a single low-resolution image .\\\\\\\\\\n\\\\hline\\nMarioNETte\\n& https://hyperconnect.github.io/MarioNETte\\n& -\\tA few-shot face reenactment framework that preserves the target identity.\\\\newline\\n- No additional fine-tuning phase is needed for identity adaptation .\\\\\\\\\\n\\\\hline\\nDiscoFaceGAN\\n& https://github.com/microsoft/DiscoFaceGAN\\n& -\\tGenerate face images of virtual people with independent latent variables of identity, expression, pose, and illumination.\\\\newline\\n- Embed 3D priors into adversarial learning .\\\\\\\\\\n\\\\hline\\nStyleRig\\n& https://gvv.mpi-inf.mpg.de/projects/StyleRig\\n& -\\tCreate portrait images of faces with a rig-like control over a pretrained and fixed StyleGAN via 3D morphable face models.\\\\newline\\n- Self-supervised without manual annotations .\\\\\\\\\\n\\\\hline\\nFaceShifter\\n& https://lingzhili.com/FaceShifterPage\\n& -\\tFace swapping in high-fidelity by exploiting and integrating the target attributes.\\\\newline\\n- Can be applied to any new face pairs without requiring subject specific training .\\\\\\\\\\n\\\\hline\\nFSGAN\\n& https://github.com/YuvalNirkin/fsgan\\n& -\\tA face swapping and reenactment model that can be applied to pairs of faces without requiring training on those faces.\\\\newline\\n- Adjust to both pose and expression variations .\\\\\\\\\\n\\\\hline\\nStyleGAN\\n& https://github.com/NVlabs/stylegan\\n& -\\tA new generator architecture for GANs is proposed based on style transfer literature.\\\\newline\\n- The new architecture leads to automatic, unsupervised separation of high-level attributes and enables intuitive, scale-specific control of the synthesis of images .\\\\\\\\\\n\\\\hline\\nFace2Face & https://justusthies.github.io/posts/face2face/ & - Real-time facial reenactment of monocular target video sequence, e.g. Youtube video.\\\\newline\\n- Animate the facial expressions of the target video by a source actor and re-render the manipulated output video in a photo-realistic fashion .\\\\\\\\\\n\\\\hline\\nNeural Textures & https://github.com/SSRSGJYD/NeuralTexture & \\n- Feature maps that are learned as part of the scene capture process and stored as maps on top of {3D} mesh proxies.\\\\newline\\n- Can coherently re-render or manipulate existing video content in both static and dynamic environments at real-time rates .\\\\\\\\\\n\\\\hline\\nTransformable Bottleneck Networks\\n& https://github.com/kyleolsz/TB-Networks\\n& -\\tA method for fine-grained 3D manipulation of image content.\\\\newline\\n- Apply spatial transformations in CNN models using a transformable bottleneck framework .\\\\\\\\\\n\\\\hline\\n``Do as I Do\\\" Motion \\\\newline Transfer\\n& github.com/carolineec/EverybodyDanceNow\\n& - Automatically transfer the motion from a source to a target person by learning a video-to-video translation.\\\\newline\\n- Can create a motion-synchronized dancing video with multiple subjects .\\\\\\\\\\n\\\\hline\\nNeural Voice Puppetry\\n& https://justusthies.github.io/posts/neural-voice-puppetry\\n& - A method for audio-driven facial video synthesis.\\\\newline\\n- Synthesize videos of a talking head from an audio sequence of another person using 3D face representation. .\\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{scriptsize}\\n\\\\end{table*}\\n\\\\begin{figure}[t]\\n\\\\centering\\n\\\\includegraphics[width=0.92\\\\columnwidth]{Fig3.pdf}\\n\\\\caption{A deepfake creation model using two encoder-decoder pairs. Two networks use the same encoder but different decoders for training process (top). An image of face A is encoded with the common encoder and decoded with decoder B to create a deepfake (bottom). The reconstructed image (in the bottom) is the face B with the mouth shape of face A. Face B originally has the mouth of an upside-down heart while the reconstructed face B has the mouth of a conventional heart.}\\n\\\\label{fig1} \\n\\\\end{figure} \\nBy adding adversarial loss and perceptual loss implemented in VGGFace to the encoder-decoder architecture, an improved version of deepfakes based on the generative adversarial network , i.e., faceswap-GAN, was proposed in . The VGGFace perceptual loss is added to make eye movements to be more realistic and consistent with input faces and help to smooth out artifacts in segmentation mask, leading to higher quality output videos. This model facilitates the creation of outputs with 64x64, 128x128, and 256x256 resolutions. In addition, the multi-task convolutional neural network (CNN) from the FaceNet implementation is used to make face detection more stable and face alignment more reliable. The CycleGAN is utilized for generative network implementation in this model.\\n\\\\begin{figure}[!ht]\\n\\\\centering\\n\\\\includegraphics[width=0.55\\\\columnwidth]{Fig4.pdf}\\n\\\\caption{The GAN architecture consisting of a generator and a discriminator, and each can be implemented by a neural network. The entire system can be trained with backpropagation that allows both networks to improve their capabilities.}\\n\\\\label{figGAN} \\n\\\\end{figure} \\nA conventional GAN model comprises two neural networks: a generator and a discriminator as depicted in Fig. \\\\ref{figGAN}. Given a dataset of real images $x$ having a distribution of $p_{data}$, the aim of the generator $G$ is to produce images $G(z)$ similar to real images $x$ with $z$ being noise signals having a distribution of $p_z$. The aim of the discriminator $G$ is to correctly classify images generated by $G$ and real images $x$. The discriminator $D$ is trained to improve its classification capability, i.e., to maximize $D(x)$, which represents the probability that $x$ is a real image rather than a fake image generated by $G$. On the other hand, $G$ is trained to minimize the probability that its outputs are classified by $D$ as synthetic images, i.e., to minimize $1-D(G(z))$. This is a minimax game between two players $D$ and $G$ that can be described by the following value function :\\n\\\\begin{multline}\\n \\\\min_G \\\\max_D V(D,G)=\\\\EX_{x\\\\sim p_{data}(x)}[\\\\log D(x)] \\\\\\\\ + \\\\EX_{z\\\\sim p_z(z)}[\\\\log (1-D(G(z)))]\\n\\\\end{multline}\\nAfter sufficient training, both networks improve their capabilities, i.e., the generator $G$ is able to produce images that are really similar to real images while the discriminator $D$ is highly capable of distinguishing fake images from real ones.\\nTable \\\\ref{table1} presents a summary of popular deepfake tools and their typical features. Among them, a prominent method for face synthesis based on a GAN model, namely StyleGAN, was introduced in . StyleGAN is motivated by style transfer with a special generator network architecture that is able to create realistic face images. In a traditional GAN model, e.g., the progressive growing of GAN (PGGAN) , the signal noise (latent code) is fed to the input layer of a feedforward network that represents the generator. In StyleGAN, there are two networks constructed and linked together, a mapping network $f$ and a synthesis network $g$. The latent code $z \\\\in Z$ is first converted to $w \\\\in W$ (where $W$ is an intermediate latent space) through a non-linear function $f:Z\\\\rightarrow W$, which is characterized by a neural network (i.e., the mapping network) consisting of several fully connected layers. Using an affine tranformation, the intermediate representation $w$ is specialized to styles $y=(y_s,y_b)$ that will be fed to the adaptive instance normalization (AdaIN) operations, specified as: \\n\\\\begin{equation}\\n \\\\mathrm{AdaIN}(x_i,y)=y_{s,i}\\\\frac{x_i-\\\\mu(x_i)}{\\\\sigma(x_i)}+y_{b,i}\\n\\\\end{equation}\\nwhere each feature map $x_i$ is normalized separately. The StyleGAN generator architecture allows controlling the image synthesis by modifying the styles via different scales. In addition, instead of using one random latent code during training, this method uses two latent codes to generate a given proportion of images. More specifically, two latent codes $z_1$ and $z_2$ are fed to the mapping network to create respectively $w_1$ and $w_2$ that control the styles by applying $w_1$ before and $w_2$ after the crossover point. Fig. \\\\ref{figMixing} demonstrates examples of images created by mixing two latent codes at three different scales where each subset of styles controls separate meaningful high-level attributes of the image. In other words, the generator architecture of StyleGAN is able to learn separation of high-level attributes (e.g., pose and identity when trained on human faces) and enables intuitive, scale-specific control of the face synthesis.\\n\\\\begin{figure}[h]\\n\\\\centering\\n\\\\includegraphics[width=0.9\\\\columnwidth]{Fig5.pdf}\\n\\\\caption{Examples of mixing styles using StyleGAN: the output images are generated by copying a specified subset of styles from source B and taking the rest from source A. a) Copying coarse styles from source B will generate images that have high-level aspects from source B and all colors and finer facial features from source A; b) if copying the styles of middle resolutions from B, the output images will have smaller scale facial features from B and preserve the pose, general face shape, and eyeglasses from A; c) if copying the fine styles from source B, the generated images will have the color scheme and microstructure of source B .}\\n\\\\label{figMixing} \\n\\\\end{figure}\",\n \"id\": \"b11ed0ba-b192-4a4a-b063-d32fbaab94d4\",\n \"level\": \"section\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"44777335-fb2f-4b5d-83e4-47fe65116410\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Creation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deepfake Creation\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 1244,\n 5139,\n 5134\n ],\n \"content\": \"\\\\label{sec:3}\\nDeepfake detection is normally deemed a binary classification problem where classifiers are used to classify between authentic videos and tampered ones. This kind of methods requires a large database of real and fake videos to train classification models. The number of fake videos is increasingly available, but it is still limited in terms of setting a benchmark for validating various detection methods. To address this issue, Korshunov and Marcel produced a notable deepfake dataset consisting of 620 videos based on the GAN model using the open source code Faceswap-GAN . Videos from the publicly available VidTIMIT database were used to generate low and high quality deepfake videos, which can effectively mimic the facial expressions, mouth movements, and eye blinking. These videos were then used to test various deepfake detection methods. Test results show that the popular face recognition systems based on VGG and Facenet are unable to detect deepfakes effectively. Other methods such as lip-syncing approaches and image quality metrics with support vector machine (SVM) produce very high error rate when applied to detect deepfake videos from this newly produced dataset. This raises concerns about the critical need of future development of more robust methods that can detect deepfakes from genuine.\\nThis section presents a survey of deepfake detection methods where we group them into two major categories: fake image detection methods and fake video detection ones (Fig. \\\\ref{fig2}). The latter is distinguished into two smaller groups: \\\\emph{visual artifacts} within single video frame-based methods and \\\\emph{temporal features} across frames-based ones. Whilst most of the methods based on temporal features use deep learning \\\\emph{recurrent} classification models, the methods use visual artifacts within video frame can be implemented by either deep or shallow classifiers.\",\n \"id\": \"7513ea5e-21ef-4144-b033-7ba34feb94da\",\n \"level\": \"section\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"44777335-fb2f-4b5d-83e4-47fe65116410\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ]\n ],\n \"subsections\": [\n \"f1d90192-c843-4231-8ac1-201458f556a0\",\n \"406b6759-531d-4af3-af75-8057a4f625fc\"\n ],\n \"title\": \"Deepfake Detection\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 5140\n ],\n \"content\": \"Deepfakes are increasingly detrimental to privacy, society security and democracy . Methods for detecting deepfakes have been proposed as soon as this threat was introduced. Early attempts were based on handcrafted features obtained from artifacts and inconsistencies of the fake image synthesis process. Recent methods, e.g., , have commonly applied deep learning to automatically extract salient and discriminative features to detect deepfakes.\",\n \"id\": \"f1d90192-c843-4231-8ac1-201458f556a0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"7513ea5e-21ef-4144-b033-7ba34feb94da\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Image Detection\"\n ]\n ],\n \"subsections\": [\n \"5d2250d1-1cae-45db-a279-c36ed6840045\",\n \"87acc6e6-8b7c-49ef-9ce2-768ae70fea85\"\n ],\n \"title\": \"Fake Image Detection\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 7897,\n 5141\n ],\n \"content\": \"Most works on detection of GAN generated images do not consider the generalization capability of the detection models although the development of GAN is ongoing, and many new extensions of GAN are frequently introduced. Xuan et al. used an image preprocessing step, e.g., Gaussian blur and Gaussian noise, to remove low level high frequency clues of GAN images. This increases the pixel level statistical similarity between real images and fake images and allows the forensic classifier to learn more intrinsic and meaningful features, which has better generalization capability than previous image forensics methods or image steganalysis networks .\\nZhang et al. used the bag of words method to extract a set of compact features and fed it into various classifiers such as SVM , random forest (RF) and multi-layer perceptrons (MLP) for discriminating swapped face images from the genuine. Among deep learning-generated images, those synthesised by GAN models are probably most difficult to detect as they are realistic and high-quality based on GAN's capability to learn distribution of the complex input data and generate new outputs with similar input distribution. \\nOn the other hand, Agarwal and Varshney cast the GAN-based deepfake detection as a hypothesis testing problem where a statistical framework was introduced using the information-theoretic study of authentication . The minimum distance between distributions of legitimate images and images generated by a particular GAN is defined, namely the oracle error. The analytic results show that this distance increases when the GAN is less accurate, and in this case, it is easier to detect deepfakes. In case of high-resolution image inputs, an extremely accurate GAN is required to generate fake images that are hard to detect by this method.\",\n \"id\": \"5d2250d1-1cae-45db-a279-c36ed6840045\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"f1d90192-c843-4231-8ac1-201458f556a0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Image Detection\"\n ],\n [\n \"subsubsection\",\n \"Handcrafted Features-based Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Handcrafted Features-based Methods\"\n },\n {\n \"cite_extract_rate\": 0.551724137931034,\n \"cites\": [\n 1603,\n 895,\n 56,\n 62,\n 5142,\n 63,\n 5143,\n 7194,\n 8317,\n 3212,\n 1003,\n 485,\n 96,\n 1600,\n 78,\n 7898\n ],\n \"content\": \"Face swapping has a number of compelling applications in video compositing, transfiguration in portraits, and especially in identity protection as it can replace faces in photographs by ones from a collection of stock images. However, it is also one of the techniques that cyber attackers employ to penetrate identification or authentication systems to gain illegitimate access. The use of deep learning such as CNN and GAN has made swapped face images more challenging for forensics models as it can preserve pose, facial expression and lighting of the photographs .\\nHsu et al. introduced a two-phase deep learning method for detection of deepfake images. The first phase is a feature extractor based on the common fake feature network (CFFN) where the Siamese network architecture presented in is used. The CFFN encompasses several dense units with each unit including different numbers of dense blocks to improve the representative capability for the input images. \\nDiscriminative features between the fake and real images are extracted through the CFFN learning process based on the use of pairwise information, which is the label of each pair of two input images. If the two images are of the same type, i.e., fake-fake or real-real, the pairwise label is $1$. In contrast, if they are of different types, i.e., fake-real, the pairwise label is $0$. The CFFN-based discriminative features are then fed to a neural network classifier to distinguish deceptive images from genuine. The proposed method is validated for both fake face and fake general image detection. On the one hand, the face dataset is obtained from CelebA , containing 10,177 identities and 202,599 aligned face images of various poses and background clutter. Five GAN variants are used to generate fake images with size of 64x64, including deep convolutional GAN (DCGAN) , Wasserstein GAN (WGAN) , WGAN with gradient penalty (WGAN-GP) , least squares GAN , and PGGAN . A total of 385,198 training images and 10,000 test images of both real and fake ones are obtained for validating the proposed method. On the other hand, the general dataset is extracted from the ILSVRC12 . The large scale GAN training model for high fidelity natural image synthesis (BIGGAN) , self-attention GAN and spectral normalization GAN are used to generate fake images with size of 128x128. The training set consists of 600,000 fake and real images whilst the test set includes 10,000 images of both types. Experimental results show the superior performance of the proposed method against its competing methods such as those introduced in .\\nLikewise, proposed a CNN model, namely SCnet, to detect deepfake images, which are generated by the Glow-based facial forgery tool . The fake images synthesized by the Glow model have the facial expression maliciously tampered. These images are hyper-realistic with perfect visual qualities, but they still have subtle or noticeable manipulation traces, which are exploited by the SCnet. The SCnet is able to automatically learn high-level forensics features of image data thanks to a hierarchical feature extraction block, which is formed by stacking four convolutional layers. Each layer learns a new set of feature maps from the previous layer, with each convolutional operation is defined by:\\n\\\\begin{equation}\\n f_j^{(n)} = \\\\sum_{i=1}^i f_i^{(n-1)}*\\\\omega_{ij}^{(n)} + b_j^{(n)}\\n\\\\end{equation}\\nwhere $f_j^{(n)}$ is the $j^{th}$ feature map of the $n^{th}$ layer, $\\\\omega_{ij}^{(n)}$ is the weight of the $i^{th}$ channel of the $j^{th}$ convolutional kernel in the $n^{th}$ layer, and $b_j^{(n)}$ is the bias term of the $j^{th}$ convolutional kernel in the $n^{th}$ layer.\\nThe proposed approach is evaluated using a dataset consisting of 321,378 face images, which are created by applying the Glow model to the CelebA face image dataset . Evaluation results show that the SCnet model obtains higher accuracy and better generalization than the Meso-4 model proposed in .\\n\\\\begin{figure*}[!ht]\\n\\\\centering\\n\\\\includegraphics[width=1.75\\\\columnwidth]{Fig6.pdf}\\n\\\\caption{A two-step process for face manipulation detection where the preprocessing step aims to detect, crop and align faces on a sequence of frames and the second step distinguishes manipulated and authentic face images by combining convolutional neural network (CNN) and recurrent neural network (RNN) .}\\n\\\\label{fig3} \\n\\\\end{figure*}\\nRecently, proposed a method for deepfake detection using self-consistency of local source features, which are content-independent, spatially-local information of images. These features could come from either imaging pipelines, encoding methods or image synthesis approaches. The hypothesis is that a modified image would have different source features at different locations, while an original image will have the same source features across locations. These source features, represented in the form of down-sampled feature maps, are extracted by a CNN model using a special representation learning method called pairwise self-consistency learning. This learning method aims to penalize pairs of feature vectors that refer to locations from the same image for having a low cosine similarity score. At the same time, it also penalizes the pairs from different images for having a high similarity score. The learned feature maps are then fed to a classification method for deepfake detection. This proposed approach is evaluated on seven popular datasets, including FaceForensics++ , DeepfakeDetection , Celeb-DF-v1 \\\\& Celeb-DF-v2 , Deepfake Detection Challenge (DFDC) , DFDC Preview , and DeeperForensics-1.0 . Experimental results demonstrate that the proposed approach is superior to state-of-the-art methods. It however may have a limitation when dealing with fake images that are generated by methods that directly output the whole images whose source features are consistent across all positions within each image.\",\n \"id\": \"87acc6e6-8b7c-49ef-9ce2-768ae70fea85\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 29,\n \"parent_id\": \"f1d90192-c843-4231-8ac1-201458f556a0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Image Detection\"\n ],\n [\n \"subsubsection\",\n \"Deep Features-based Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Features-based Methods\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1603\n ],\n \"content\": \"Most image detection methods cannot be used for videos because of the strong degradation of the frame data after video compression . Furthermore, videos have temporal characteristics that are varied among sets of frames and they are thus challenging for methods designed to detect only still fake images. This subsection focuses on deepfake video detection methods and categorizes them into two smaller groups: methods that employ temporal features and those that explore visual artifacts within frames.\",\n \"id\": \"406b6759-531d-4af3-af75-8057a4f625fc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"7513ea5e-21ef-4144-b033-7ba34feb94da\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Video Detection\"\n ]\n ],\n \"subsections\": [\n \"f76aa108-55fe-4a3b-9495-cbb64d02a169\",\n \"89bc57b3-3d0e-4686-9c89-9b9c741caaac\"\n ],\n \"title\": \"Fake Video Detection\"\n },\n {\n \"cite_extract_rate\": 0.5454545454545451,\n \"cites\": [\n 5142,\n 1600,\n 97,\n 243,\n 3758,\n 96\n ],\n \"content\": \"Based on the observation that temporal coherence is not enforced effectively in the synthesis process of deepfakes, Sabir et al. leveraged the use of spatio-temporal features of video streams to detect deepfakes. Video manipulation is carried out on a frame-by-frame basis so that low level artifacts produced by face manipulations are believed to further manifest themselves as temporal artifacts with inconsistencies across frames. A recurrent convolutional model (RCN) was proposed based on the integration of the convolutional network DenseNet and the gated recurrent unit cells to exploit temporal discrepancies across frames (see Fig. \\\\ref{fig3}). The proposed method is tested on the FaceForensics++ dataset, which includes 1,000 videos , and shows promising results. \\nLikewise, highlighted that deepfake videos contain intra-frame inconsistencies and temporal inconsistencies between frames. They then proposed the temporal-aware pipeline method that uses CNN and long short term memory (LSTM) to detect deepfake videos. CNN is employed to extract frame-level features, which are then fed into the LSTM to create a temporal sequence descriptor. A fully-connected network is finally used for classifying doctored videos from real ones based on the sequence descriptor as illustrated in Fig. \\\\ref{fig4}. An accuracy of greater than 97\\\\% was obtained using a dataset of 600 videos, including 300 deepfake videos collected from multiple video-hosting websites and 300 pristine videos randomly selected from the Hollywood human actions dataset in .\\n\\\\begin{figure}[!ht]\\n\\\\centering\\n\\\\includegraphics[width=1\\\\columnwidth]{Fig7.pdf}\\n\\\\caption{A deepfake detection method using convolutional neural network (CNN) and long short term memory (LSTM) to extract temporal features of a given video sequence, which are represented via the sequence descriptor. The detection network consisting of fully-connected layers is employed to take the sequence descriptor as input and calculate probabilities of the frame sequence belonging to either authentic or deepfake class .}\\n\\\\label{fig4} \\n\\\\end{figure}\\nOn the other hand, the use of a physiological signal, eye blinking, to detect deepfakes was proposed in Li et al. based on the observation that a person in deepfakes has a lot less frequent blinking than that in untampered videos. A healthy adult human would normally blink somewhere between 2 to 10 seconds, and each blink would take 0.1 and 0.4 seconds. Deepfake algorithms, however, often use face images available online for training, which normally show people with open eyes, i.e., very few images published on the internet show people with closed eyes. Thus, without having access to images of people blinking, deepfake algorithms do not have the capability to generate fake faces that can blink normally. In other words, blinking rates in deepfakes are much lower than those in normal videos. To discriminate real and fake videos, Li et al. crop eye areas in the videos and distribute them into long-term recurrent convolutional networks (LRCN) for dynamic state prediction. The LRCN consists of a feature extractor based on CNN, a sequence learning based on long short term memory (LSTM), and a state prediction based on a fully connected layer to predict probability of eye open and close state. The eye blinking shows strong temporal dependencies and thus the implementation of LSTM helps to capture these temporal patterns effectively.\\nRecently, proposed the use of optical flow to gauge the information along the temporal axis of a frame sequence for video deepfake detection. The optical flow is a vector field calculated on two temporal-distinct frames of a video that can describe the movement of objects in a scene. The optical flow fields are expected to be different between synthetically created frames and naturally generated ones . Unnatural movements of lips, eyes, or of the entire faces inserted into deepfake videos would introduce distinctive motion patterns when compared with pristine ones. Based on this assumption, features consisting of optical flow fields are fed into a CNN model for discriminating between deepfakes and original videos. More specifically, the ResNet50 architecture is implemented as a CNN model for experiments. The results obtained using the FaceForensics++ dataset show that this approach is comparable with state-of-the-art methods in terms of classification accuracy. A combination of this kind of feature with frame-based features is also experimented, which results in an improved deepfake detection performance. This demonstrates the usefulness of optical flow fields in capturing the inconsistencies on the temporal axis of video frames for deepfake detection.\",\n \"id\": \"f76aa108-55fe-4a3b-9495-cbb64d02a169\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"406b6759-531d-4af3-af75-8057a4f625fc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Video Detection\"\n ],\n [\n \"subsubsection\",\n \"Temporal Features across Video Frames\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Temporal Features across Video Frames\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As can be noticed in the previous subsection, the methods using temporal patterns across video frames are mostly based on deep recurrent network models to detect deepfake videos. This subsection investigates the other approach that normally decomposes videos into frames and explores visual artifacts within single frames to obtain discriminant features. These features are then distributed into either a deep or shallow classifier to differentiate between fake and authentic videos. We thus group methods in this subsection based on the types of classifiers, i.e. either deep or shallow.\",\n \"id\": \"89bc57b3-3d0e-4686-9c89-9b9c741caaac\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"406b6759-531d-4af3-af75-8057a4f625fc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Video Detection\"\n ],\n [\n \"subsubsection\",\n \"Visual Artifacts within Video Frame\"\n ]\n ],\n \"subsections\": [\n \"c8afcbd9-52a4-4801-8c8d-0c5d0d477915\",\n \"b02a12ee-cc0c-427d-b853-a4c928ab1f56\"\n ],\n \"title\": \"Visual Artifacts within Video Frame\"\n },\n {\n \"cite_extract_rate\": 0.642857142857142,\n \"cites\": [\n 5144,\n 1603,\n 5145,\n 7899,\n 97,\n 514,\n 5146,\n 5147,\n 306\n ],\n \"content\": \"Deepfake videos are normally created with limited resolutions, which require an affine face warping approach (i.e., scaling, rotation and shearing) to match the configuration of the original ones. Because of the resolution inconsistency between the warped face area and the surrounding context, this process leaves artifacts that can be detected by CNN models such as VGG16 , ResNet50, ResNet101 and ResNet152 . A deep learning method to detect deepfakes based on the artifacts observed during the face warping step of the deepfake generation algorithms was proposed in . The proposed method is evaluated on two deepfake datasets, namely the UADFV and DeepfakeTIMIT. The UADFV dataset contains 49 real videos and 49 fake videos with 32,752 frames in total. The DeepfakeTIMIT dataset includes a set of low quality videos of 64 x 64 size and another set of high quality videos of 128 x 128 with totally 10,537 pristine images and 34,023 fabricated images extracted from 320 videos for each quality set. Performance of the proposed method is compared with other prevalent methods such as two deepfake detection MesoNet methods, i.e. Meso-4 and MesoInception-4 , HeadPose , and the face tampering detection method two-stream NN . Advantage of the proposed method is that it needs not to generate deepfake videos as negative examples before training the detection models. Instead, the negative examples are generated dynamically by extracting the face region of the original image and aligning it into multiple scales before applying Gaussian blur to a scaled image of random pick and warping back to the original image. This reduces a large amount of time and computational resources compared to other methods, which require deepfakes are generated in advance.\\nNguyen et al. proposed the use of capsule networks for detecting manipulated images and videos. The capsule network was initially introduced to address limitations of CNNs when applied to inverse graphics tasks, which aim to find physical processes used to produce images of the world . The recent development of capsule network based on dynamic routing algorithm demonstrates its ability to describe the hierarchical pose relationships between object parts. This development is employed as a component in a pipeline for detecting fabricated images and videos as demonstrated in Fig. \\\\ref{fig5}. A dynamic routing algorithm is deployed to route the outputs of the three capsules to the output capsules through a number of iterations to separate between fake and real images. The method is evaluated through four datasets covering a wide range of forged image and video attacks. They include the well-known Idiap Research Institute replay-attack dataset , the deepfake face swapping dataset created by Afchar et al. , the facial reenactment FaceForensics dataset , produced by the Face2Face method , and the fully computer-generated image dataset generated by Rahmouni et al. . The proposed method yields the best performance compared to its competing methods in all of these datasets. This shows the potential of the capsule network in building a general detection system that can work effectively for various forged image and video attacks.\\n\\\\begin{figure}[!ht]\\n\\\\centering\\n\\\\includegraphics[width=1\\\\columnwidth]{Fig8.pdf}\\n\\\\caption{Capsule network takes features obtained from the VGG-19 network to distinguish fake images or videos from the real ones (top). The pre-processing step detects face region and scales it to the size of 128x128 before VGG-19 is used to extract latent features for the capsule network, which comprises three primary capsules and two output capsules, one for real and one for fake images (bottom). The statistical pooling constitutes an important part of capsule network that deals with forgery detection .}\\n\\\\label{fig5} \\n\\\\end{figure}\",\n \"id\": \"c8afcbd9-52a4-4801-8c8d-0c5d0d477915\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"89bc57b3-3d0e-4686-9c89-9b9c741caaac\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Video Detection\"\n ],\n [\n \"subsubsection\",\n \"Visual Artifacts within Video Frame\"\n ],\n [\n \"paragraph\",\n \"Deep classifiers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep classifiers\"\n },\n {\n \"cite_extract_rate\": 0.512820512820512,\n \"cites\": [\n 1255,\n 56,\n 7899,\n 5148,\n 5149,\n 5151,\n 485,\n 5145,\n 895,\n 97,\n 514,\n 7900,\n 5134,\n 5153,\n 306,\n 7022,\n 96,\n 5152,\n 5144,\n 5156,\n 1600,\n 5155,\n 1216,\n 62,\n 5142,\n 1003,\n 150,\n 1603,\n 7897,\n 5150,\n 5147,\n 63,\n 7194,\n 8317,\n 3212,\n 59,\n 5154,\n 157,\n 78,\n 243\n ],\n \"content\": \"Deepfake detection methods mostly rely on the artifacts or inconsistency of intrinsic features between fake and real images or videos. Yang et al. proposed a detection method by observing the differences between 3D head poses comprising head orientation and position, which are estimated based on 68 facial landmarks of the central face region. The 3D head poses are examined because there is a shortcoming in the deepfake face generation pipeline. The extracted features are fed into an SVM classifier to obtain the detection results. Experiments on two datasets show the great performance of the proposed approach against its competing methods. The first dataset, namely UADFV, consists of 49 deep fake videos and their respective real videos . The second dataset comprises 241 real images and 252 deep fake images, which is a subset of data used in the DARPA MediFor GAN Image/Video Challenge . Likewise, a method to exploit artifacts of deepfakes and face manipulations based on visual features of eyes, teeth and facial contours was studied in . The visual artifacts arise from lacking global consistency, wrong or imprecise estimation of the incident illumination, or imprecise estimation of the underlying geometry. For deepfakes detection, missing reflections and missing details in the eye and teeth areas are exploited as well as texture features extracted from the facial region based on facial landmarks. Accordingly, the eye feature vector, teeth feature vector and features extracted from the full-face crop are used. After extracting the features, two classifiers including logistic regression and small neural network are employed to classify the deepfakes from real videos. Experiments carried out on a video dataset downloaded from YouTube show the best result of 0.851 in terms of the area under the receiver operating characteristics curve. The proposed method however has a disadvantage that requires images meeting certain prerequisite such as open eyes or visual teeth.\\nThe use of photo response non uniformity (PRNU) analysis was proposed in to detect deepfakes from authentic ones. PRNU is a component of sensor pattern noise, which is attributed to the manufacturing imperfection of silicon wafers and the inconsistent sensitivity of pixels to light because of the variation of the physical characteristics of the silicon wafers. The PRNU analysis is widely used in image forensics and advocated to use in because the swapped face is supposed to alter the local PRNU pattern in the facial area of video frames. The videos are converted into frames, which are cropped to the questioned facial region. The cropped frames are then separated sequentially into eight groups where an average PRNU pattern is computed for each group. Normalised cross correlation scores are calculated for comparisons of PRNU patterns among these groups. A test dataset was created, consisting of 10 authentic videos and 16 manipulated videos, where the fake videos were produced from the genuine ones by the DeepFaceLab tool . The analysis shows a significant statistical difference in terms of mean normalised cross correlation scores between deepfakes and the genuine. This analysis therefore suggests that PRNU has a potential in deepfake detection although a larger dataset would need to be tested. \\nWhen seeing a video or image with suspicion, users normally want to search for its origin. However, there is currently no feasibility for such a tool. Hasan and Salah proposed the use of blockchain and smart contracts to help users detect deepfake videos based on the assumption that videos are only real when their sources are traceable. Each video is associated with a smart contract that links to its parent video and each parent video has a link to its child in a hierarchical structure. Through this chain, users can credibly trace back to the original smart contract associated with pristine video even if the video has been copied multiple times. An important attribute of the smart contract is the unique hashes of the interplanetary file system, which is used to store video and its metadata in a decentralized and content-addressable manner . The smart contract's key features and functionalities are tested against several common security challenges such as distributed denial of services, replay and man in the middle attacks to ensure the solution meeting security requirements. This approach is generic, and it can be extended to other types of digital content, e.g., images, audios and manuscripts.\\n\\\\begin{table*}[h!]\\n\\\\centering\\n\\\\caption{Summary of prominent deepfake detection methods}\\n\\\\label{table2}\\n\\\\begin{scriptsize}\\n\\\\begin{tabularx}{\\\\linewidth}{p{0.11\\\\textwidth} p{0.09\\\\textwidth} p{0.33\\\\textwidth} p{0.04\\\\textwidth} p{0.32\\\\textwidth}}\\n\\\\toprule\\n\\\\textbf{Methods} & \\\\textbf{Classifiers/\\\\newline Techniques} & \\\\textbf{Key Features} & \\\\textbf{Dealing with} & \\\\textbf{Datasets Used}\\\\\\\\\\n\\\\midrule\\nEye blinking \\n&LRCN\\n&- Use LRCN to learn the temporal patterns of eye blinking.\\\\newline\\n- Based on the observation that blinking frequency of deepfakes is much smaller than normal.\\n&Videos\\n&Consist of 49 interview and presentation videos, and their corresponding generated deepfakes.\\\\\\\\\\n\\\\hline\\nIntra-frame and temporal inconsistencies \\n&CNN and LSTM\\n&CNN is employed to extract frame-level features, which are distributed to LSTM to construct sequence descriptor useful for classification.\\t\\n&Videos\\n&A collection of 600 videos obtained from multiple websites.\\\\\\\\\\n\\\\hline\\nUsing face warping artifacts \\n&VGG16 ,\\\\newline ResNet models \\n&Artifacts are discovered using CNN models based on resolution inconsistency between the warped face area and the surrounding context.\\n&Videos\\n&- UADFV , containing 49 real videos and 49 fake videos with 32752 frames in total.\\\\newline\\n- DeepfakeTIMIT \\\\\\\\\\n\\\\hline\\nMesoNet \\n&CNN\\n&- Two deep networks, i.e. Meso-4 and MesoInception-4 are introduced to examine deepfake videos at the mesoscopic analysis level.\\\\newline\\n- Accuracy obtained on deepfake and FaceForensics datasets are 98\\\\% and 95\\\\% respectively.\\n&Videos\\n&Two datasets: deepfake one constituted from online videos and the FaceForensics one created by the Face2Face approach .\\\\\\\\\\n\\\\hline\\nEye, teach and facial texture \\n&Logistic regression and neural network (NN)\\n&- Exploit facial texture differences, and missing reflections and details in eye and teeth areas of deepfakes.\\\\newline\\n- Logistic regression and NN are used for classifying.\\t\\n&Videos\\n&A video dataset downloaded from YouTube.\\\\\\\\\\n\\\\hline\\nSpatio-temporal features with RCN \\n&RCN\\n&Temporal discrepancies across frames are explored using RCN that integrates convolutional network DenseNet and the gated recurrent unit cells \\n&Videos\\n&FaceForensics++ dataset, including 1,000 videos .\\\\\\\\\\n\\\\hline\\nSpatio-temporal features with LSTM \\n& Convolutional bidirectional recurrent LSTM network\\n& - An XceptionNet CNN is used for facial feature extraction while audio embeddings are obtained by stacking multiple convolution modules.\\\\newline\\n- Two loss functions, i.e. cross-entropy and Kullback-Leibler divergence, are used.\\n& Videos\\n& FaceForensics++ and Celeb-DF (5,639 deepfake videos) datasets and the ASVSpoof 2019 Logical Access audio dataset .\\\\\\\\\\n\\\\hline\\nAnalysis of PRNU \\n&PRNU\\n&- Analysis of noise patterns of light sensitive sensors of digital cameras due to their factory defects.\\\\newline\\n- Explore the differences of PRNU patterns between the authentic and deepfake videos because face swapping is believed to alter the local PRNU patterns.\\n&Videos\\t\\n&Created by the authors, including 10 authentic and 16 deepfake videos using DeepFaceLab .\\\\\\\\\\n\\\\hline\\nPhoneme-viseme mismatches \\n& CNN\\n& - Exploit the mismatches between the dynamics of the mouth shape, i.e. visemes, with a spoken phoneme.\\\\newline\\n- Focus on sounds associated with the M, B and P phonemes as they require complete mouth closure while deepfakes often incorrectly synthesize it.\\n& Videos\\n& Four in-the-wild lip-sync deepfakes from Instagram and YouTube (www.instagram.com/bill\\\\_posters\\\\_uk and youtu.be/VWMEDacz3L4) and others are created using synthesis techniques, i.e. Audio-to-Video (A2V) and Text-to-Video (T2V) .\\\\\\\\\\n\\\\hline\\nUsing attribution-based confidence (ABC) metric \\n& ResNet50 model , pre-trained on VGGFace2 \\n& - The ABC metric is used to detect deepfake videos without accessing to training data.\\\\newline\\n- ABC values obtained for original videos are greater than 0.94 while those of deepfakes have low ABC values.\\n& Videos\\n& VidTIMIT and two other original datasets obtained from the COHFACE (https://www.idiap.ch/dataset/cohface) and from YouTube. datasets from COHFACE and YouTube are used to generate two deepfake datasets by commercial website https://deepfakesweb.com and another deepfake dataset is DeepfakeTIMIT .\\\\\\\\\\n\\\\hline\\nUsing appearance and behaviour \\n& Rules based on facial and behavioural features.\\n& Temporal, behavioral biometric based on facial expressions and head movements are learned using ResNet-101 while static facial biometric is obtained using VGG .\\n& Videos\\n& The world leaders dataset , FaceForensics++ , Google/Jigsaw deepfake detection dataset , DFDC and Celeb-DF .\\\\\\\\\\n\\\\hline\\nFakeCatcher \\n& CNN\\n& Extract biological signals in portrait videos and use them as an implicit descriptor of authenticity because they are not spatially and temporally well-preserved in deepfakes.\\n& Videos\\n& UADFV , FaceForensics , FaceForensics++ , Celeb-DF , and a new dataset of 142 videos, independent of the generative model, resolution, compression, content, and context.\\\\\\\\\\n\\\\hline\\nEmotion audio-visual affective cues \\n& Siamese network \\n& Modality and emotion embedding vectors for the face and speech are extracted for deepfake detection.\\n& Videos\\n& DeepfakeTIMIT and DFDC .\\\\\\\\\\n\\\\hline\\nHead poses \\n&SVM\\n&- Features are extracted using 68 landmarks of the face region.\\\\newline\\n- Use SVM to classify using the extracted features.\\t\\n&Videos/\\\\newline Images\\n&- UADFV consists of 49 deep fake videos and their respective real videos.\\\\newline\\n- 241 real images and 252 deep fake images from DARPA MediFor GAN Image/Video Challenge.\\\\\\\\\\n\\\\hline\\nCapsule-forensics \\n&Capsule networks\\n&- Latent features extracted by VGG-19 network are fed into the capsule network for classification.\\\\newline\\n- A dynamic routing algorithm is used to route the outputs of three convolutional capsules to two output capsules, one for fake and another for real images, through a number of iterations.\\t\\n&Videos/\\\\newline Images\\n&Four datasets: the Idiap Research Institute replay-attack , deepfake face swapping by , facial reenactment FaceForensics , and fully computer-generated image set using .\\\\\\\\\\n\\\\hline\\n\\\\end{tabularx}\\n\\\\end{scriptsize}\\n\\\\end{table*}\\n\\\\begin{table*}[h!]\\n\\\\centering\\n\\\\begin{scriptsize}\\n\\\\begin{tabularx}{\\\\linewidth}{p{0.1\\\\textwidth} p{0.1\\\\textwidth} p{0.33\\\\textwidth} p{0.04\\\\textwidth} p{0.32\\\\textwidth}}\\n\\\\toprule\\n\\\\textbf{Methods} & \\\\textbf{Classifiers/\\\\newline Techniques} & \\\\textbf{Key Features} & \\\\textbf{Dealing with} & \\\\textbf{Datasets Used}\\\\\\\\\\n\\\\midrule\\nPreprocessing combined with deep network \\n&DCGAN,\\nWGAN-GP and PGGAN.\\n&- Enhance generalization ability of deep learning models to detect GAN generated images.\\\\newline\\n- Remove low level features of fake images.\\\\newline\\n- Force deep networks to focus more on pixel level similarity between fake and real images to improve generalization ability.\\n&Images\\n&- Real dataset: CelebA-HQ , including high quality face images of 1024x1024 resolution.\\\\newline\\n- Fake datasets: generated by DCGAN , WGAN-GP and PGGAN .\\\\\\\\\\n\\\\hline\\nAnalyzing convolutional traces \\n& KNN, SVM, and linear discriminant analysis (LDA)\\n& Using expectation-maximization algorithm to extract local features pertaining to convolutional generative process of GAN-based image deepfake generators.\\n& Images\\n& Authentic images from CelebA and corresponding deepfakes are created by five different GANs (group-wise deep whitening-and-coloring transformation GDWCT , StarGAN ,\\nAttGAN , StyleGAN , StyleGAN2 ).\\\\\\\\\\n\\\\hline\\nBag of words and shallow classifiers \\t\\n&SVM, RF, MLP\\n&Extract discriminant features using bag of words method and feed these features into SVM, RF and MLP for binary classification: innocent vs fabricated.\\n&Images\\n&The well-known LFW face database , containing 13,223 images with resolution of 250x250.\\\\\\\\\\n\\\\hline\\nPairwise learning \\n&CNN concatenated to CFFN\\n&Two-phase procedure: feature extraction using CFFN based on the Siamese network architecture and classification using CNN.\\n&Images\\n&- Face images: real ones from CelebA , and fake ones generated by DCGAN , WGAN , WGAN-GP , least squares GAN , and PGGAN .\\\\newline\\n- General images: real ones from ILSVRC12 , and fake ones generated by BIGGAN , self-attention GAN and spectral normalization GAN .\\\\\\\\\\n\\\\hline\\nDefenses against adversarial perturbations in deepfakes \\n& VGG and ResNet \\n& - Introduce adversarial perturbations to enhance deepfakes and fool deepfake detectors.\\\\newline\\n- Improve accuracy of deepfake detectors using Lipschitz regularization and deep image prior techniques.\\n& Images\\n& 5,000 real images from CelebA and 5,000 fake images created by the “Few-Shot Face Translation GAN” method .\\\\\\\\\\n\\\\hline\\nFace X-ray \\n& CNN\\n& - Try to locate the blending boundary between the target and original faces instead of capturing the synthesized artifacts of specific manipulations.\\\\newline\\n- Can be trained without fake images.\\n& Images\\n& FaceForensics++ , DeepfakeDetection (DFD) , DFDC and Celeb-DF .\\\\\\\\\\n\\\\hline\\nUsing common artifacts of CNN-generated images \\n& ResNet-50 \\npre-trained with ImageNet \\n& Train the classifier using a large number of fake images generated by a high-performing unconditional GAN model, i.e., PGGAN and evaluate how well the classifier generalizes to other CNN-synthesized images.\\n& Images\\n& A new dataset of CNN-generated images, namely ForenSynths, consisting of\\nsynthesized images from 11 models such as StyleGAN , super-resolution methods and FaceForensics++ .\\\\\\\\\\n\\\\hline\\nUsing convolutional traces on GAN-based images \\n& KNN, SVM, and LDA\\n& Training the expectation-maximization algorithm to detect\\nand extract discriminative features via a fingerprint that represents the convolutional traces left by GANs during image generation.\\n& Images\\n& A dataset of images generated by ten GAN models, including CycleGAN , StarGAN , AttGAN , GDWCT , StyleGAN , StyleGAN2 , PGGAN , FaceForensics++ , IMLE , and\\nSPADE .\\\\\\\\\\n\\\\hline\\nUsing deep features extracted by CNN \\n& A new CNN model, namely SCnet\\n& The CNN-based SCnet is able to automatically learn high-level forensics features of image data thanks to a hierarchical feature extraction block, which is formed by stacking four convolutional layers.\\n& Images\\n& A dataset of 321,378 face images, created by applying the Glow model to the CelebA face image dataset .\\n\\\\\\\\\\n\\\\bottomrule\\n\\\\end{tabularx}\\n\\\\end{scriptsize}\\n\\\\end{table*}\",\n \"id\": \"b02a12ee-cc0c-427d-b853-a4c928ab1f56\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 78,\n \"parent_id\": \"89bc57b3-3d0e-4686-9c89-9b9c741caaac\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Deepfake Detection\"\n ],\n [\n \"subsection\",\n \"Fake Video Detection\"\n ],\n [\n \"subsubsection\",\n \"Visual Artifacts within Video Frame\"\n ],\n [\n \"paragraph\",\n \"Shallow classifiers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Shallow classifiers\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 7901,\n 5160,\n 5157,\n 5158,\n 5156,\n 5161,\n 5162,\n 5159\n ],\n \"content\": \"\\\\label{sec:4}\\nWith the support of deep learning, deepfakes can be created easier than ever before. The spread of these fake contents is also quicker thanks to the development of social media platforms . Sometimes deepfakes do not need to be spread to massive audience to cause detrimental effects. People who create deepfakes with malicious purpose only need to deliver them to target audiences as part of their sabotage strategy without using social media. For example, this approach can be utilized by intelligence services trying to influence decisions made by important people such as politicians, leading to national and international security threats . Catching the deepfake alarming problem, research community has focused on developing deepfake detection algorithms and numerous results have been reported. This paper has reviewed the state-of-the-art methods and a summary of typical approaches is provided in Table \\\\ref{table2}. It is noticeable that a battle between those who use advanced machine learning to create deepfakes with those who make effort to detect deepfakes is growing.\\nDeepfakes' quality has been increasing and the performance of detection methods needs to be improved accordingly. The inspiration is that what AI has broken can be fixed by AI as well . Detection methods are still in their early stage and various methods have been proposed and evaluated but using fragmented datasets. An approach to improve performance of detection methods is to create a growing updated benchmark dataset of deepfakes to validate the ongoing development of detection methods. This will facilitate the training process of detection models, especially those based on deep learning, which requires a large training set . \\nImproving performance of deepfake detection methods is important, especially in cross-forgery and cross-dataset scenarios. Most detection models are designed and evaluated in the same-forgery and in-dataset experiments, which do not ensure their generalization capability. Some previous studies have addressed this issue, e.g., in , but more work needs to be done in this direction. A model trained on a specific forgery needs to be able to work against another unknown one because potential deepfake types are not normally known in the real-world scenarios. Likewise, current detection methods mostly focus on drawbacks of the deepfake generation pipelines, i.e., finding weakness of the competitors to attack them. This kind of information and knowledge is not always available in adversarial environments where attackers commonly attempt not to reveal such deepfake creation technologies. Recent works on adversarial perturbation attacks to fool DNN-based detectors make the deepfake detection task more difficult . These are real challenges for detection method development and a future study needs to focus on introducing more robust, scalable and generalizable methods.\\nAnother research direction is to integrate detection methods into distribution platforms such as social media to increase its effectiveness in dealing with the widespread impact of deepfakes. The screening or filtering mechanism using effective detection methods can be implemented on these platforms to ease the deepfakes detection . Legal requirements can be made for tech companies who own these platforms to remove deepfakes quickly to reduce its impacts. In addition, watermarking tools can also be integrated into devices that people use to make digital contents to create immutable metadata for storing originality details such as time and location of multimedia contents as well as their untampered attestment . This integration is difficult to implement but a solution for this could be the use of the disruptive blockchain technology. The blockchain has been used effectively in many areas and there are very few studies so far addressing the deepfake detection problems based on this technology. As it can create a chain of unique unchangeable blocks of metadata, it is a great tool for digital provenance solution. The integration of blockchain technologies to this problem has demonstrated certain results but this research direction is far from mature.\\nUsing detection methods to spot deepfakes is crucial, but understanding the real intent of people publishing deepfakes is even more important. This requires the judgement of users based on social context in which deepfake is discovered, e.g. who distributed it and what they said about it . This is critical as deepfakes are getting more and more photorealistic and it is highly anticipated that detection software will be lagging behind deepfake creation technology. A study on social context of deepfakes to assist users in such judgement is thus worth performing. \\nVideos and photographics have been widely used as evidences in police investigation and justice cases. They may be introduced as evidences in a court of law by digital media forensics experts who have background in computer or law enforcement and experience in collecting, examining and analysing digital information. The development of machine learning and AI technologies might have been used to modify these digital contents and thus the experts' opinions may not be enough to authenticate these evidences because even experts are unable to discern manipulated contents. This aspect needs to take into account in courtrooms nowadays when images and videos are used as evidences to convict perpetrators because of the existence of a wide range of digital manipulation methods . The digital media forensics results therefore must be proved to be valid and reliable before they can be used in courts. This requires careful documentation for each step of the forensics process and how the results are reached. Machine learning and AI algorithms can be used to support the determination of the authenticity of digital media and have obtained accurate and reliable results, e.g., , but most of these algorithms are unexplainable. This creates a huge hurdle for the applications of AI in forensics problems because not only the forensics experts oftentimes do not have expertise in computer algorithms, but the computer professionals also cannot explain the results properly as most of these algorithms are black box models . This is more critical as the most recent models with the most accurate results are based on deep learning methods consisting of many neural network parameters. Researchers have recently attempted to create white box and explainable detection methods. An example is the approach proposed by in which they use discrete cosine transform statistics to detect so-called specific GAN frequencies to differentiate between real images and deepfakes. Through the analysis of particular frequency statistics, that method can be used to mathematically explain whether a multimedia content is a deepfake and why it is. More research must be conducted in this area and explainable AI in computer vision therefore is a research direction that is needed to promote and utilize the advances and advantages of AI and machine learning in digital media forensics.\",\n \"id\": \"af5e6522-a34d-4deb-8b41-532983a76f29\",\n \"level\": \"section\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"44777335-fb2f-4b5d-83e4-47fe65116410\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Discussions and Future Research Directions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussions and Future Research Directions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Deepfakes have begun to erode trust of people in media contents as seeing them is no longer commensurate with believing in them. They could cause distress and negative effects to those targeted, heighten disinformation and hate speech, and even could stimulate political tension, inflame the public, violence or war. This is especially critical nowadays as the technologies for creating deepfakes are increasingly approachable and social media platforms can spread those fake contents quickly. This survey provides a timely overview of deepfake creation and detection methods and presents a broad discussion on challenges, potential trends, and future directions in this area. This study therefore will be valuable for the artificial intelligence research community to develop effective methods for tackling deepfakes.\\n\\\\section*{Declaration of Competing Interest}\\nAuthors declare no conflict of interest.\\n\\\\bibliography{cas-refs}\\n\\\\end{document}\\n\\\\endinput\",\n \"id\": \"6f55771a-c73a-46f3-b17e-1720217d26ca\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"44777335-fb2f-4b5d-83e4-47fe65116410\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Deepfakes Creation and Detection: A Survey\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":6,"string":"6"},"all_cites":{"kind":"list like","value":[5133,5128,5132,5130,5131,5680,5129,1602,986,1256,4939,7893,5134,62,7896,5138,5136,7895,96,5137,150,157,7894,5135,7892,1244,5139,5140,7897,5141,1603,895,56,5142,63,5143,7194,8317,3212,1003,485,1600,78,7898,97,243,3758,5144,5145,7899,514,5146,5147,306,1255,5148,5149,5151,7900,5153,7022,5152,5156,5155,1216,5150,59,5154,7901,5160,5157,5158,5161,5162,5159],"string":"[\n 5133,\n 5128,\n 5132,\n 5130,\n 5131,\n 5680,\n 5129,\n 1602,\n 986,\n 1256,\n 4939,\n 7893,\n 5134,\n 62,\n 7896,\n 5138,\n 5136,\n 7895,\n 96,\n 5137,\n 150,\n 157,\n 7894,\n 5135,\n 7892,\n 1244,\n 5139,\n 5140,\n 7897,\n 5141,\n 1603,\n 895,\n 56,\n 5142,\n 63,\n 5143,\n 7194,\n 8317,\n 3212,\n 1003,\n 485,\n 1600,\n 78,\n 7898,\n 97,\n 243,\n 3758,\n 5144,\n 5145,\n 7899,\n 514,\n 5146,\n 5147,\n 306,\n 1255,\n 5148,\n 5149,\n 5151,\n 7900,\n 5153,\n 7022,\n 5152,\n 5156,\n 5155,\n 1216,\n 5150,\n 59,\n 5154,\n 7901,\n 5160,\n 5157,\n 5158,\n 5161,\n 5162,\n 5159\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.057985542879675,"string":"1.057986"}}},{"rowIdx":51,"cells":{"authors":{"kind":"list like","value":["Stefan Escaida Navarro","Stephan Mühlbacher-Karrer","Hosam Alagi","Hubert Zangl","Keisuke Koyama","Björn Hein","Christian Duriez","Joshua R. Smith"],"string":"[\n \"Stefan Escaida Navarro\",\n \"Stephan Mühlbacher-Karrer\",\n \"Hosam Alagi\",\n \"Hubert Zangl\",\n \"Keisuke Koyama\",\n \"Björn Hein\",\n \"Christian Duriez\",\n \"Joshua R. Smith\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-08-16T16:28:26Z","string":"2021-08-16T16:28:26Z"},"category":{"kind":"string","value":"cs.RO"},"abstract":{"kind":"string","value":" Proximity perception is a technology that has the potential to play an essential role in the future of robotics. It can fulfill the promise of safe, robust, and autonomous systems in industry and everyday life, alongside humans, as well as in remote locations in space and underwater. In this survey paper, we cover the developments of this field from the early days up to the present, with a focus on human-centered robotics. Here, proximity sensors are typically deployed in two scenarios: first, on the exterior of manipulator arms to support safety and interaction functionality, and second, on the inside of grippers or hands to support grasping and exploration. Starting from this observation, we propose a categorization for the approaches found in the literature. To provide a basis for understanding these approaches, we devote effort to present the technologies and different measuring principles that were developed over the years, also providing a summary in form of a table. Then, we show the diversity of applications that have been presented in the literature. Finally, we give an overview of the most important trends that will shape the future of this domain. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4528ef20-9412-4b01-95cc-07b7f164956d","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "]],"subsections":["01d49bee-621f-47ab-9ac0-165e2dba2afa","7c679291-40be-4834-b2fe-70234a403984","702ec4ea-e939-40cf-abd8-8e5caea8edc6","55e0e327-ca41-4cdf-b82e-23bda770d2ce","d1273b1a-f5ad-41b8-be89-3d4b86e92fbe","46c2dad5-012b-443e-b703-dee40d0ff9dc"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Intro}\n\\IEEEPARstart{I}{n} the current robotics research landscape, a lot of effort is still dedicated to overcoming the challenges posed by unstructured environments. Areas, such as medicine, health care, agriculture, Industry\\ 4.0, and exploration endeavors in space as well as underwater are awaiting to profit from the robotics technologies currently in development. Furthermore, in terms of unstructured environments or situations, one of the main challenges is to develop robotics technologies that enable a safe and reliable interaction with the human. At the same time, the functionality and intelligence provided by the robot system must justify the investment, meaning its autonomous behavior, oftentimes in environments made for humans, must contribute real value. A hallmark of robust and efficient robot behavior is that task execution does not need to be interrupted in the presence of emergent events. A technology that is capable of addressing these challenges is proximity perception. It has been developed over the years, with first, impactful applications being shown in the late 1980s and early 1990s, which were sparked by seminal developments in robot control. Proximity perception is complementary to the main robotics perceptive modalities of vision and touch. Its use is often motivated by closing the perception gap left by occlusions and blind spots as well as by dealing with pose uncertainty of the robot with respect to objects and its environment. Therefore, one of the big challenges in this domain is to find sensor designs that can coexist with the main existing modalities of vision and touch. \n\\begin{figure}[t]\n \\centering\n \\includegraphics[width=0.45\\textwidth]{./Figures/Intro_AT-I.png}\n \\caption{A robot skin with proximity sensing capability can be deployed in scenarios with intense human-robot interaction and collaboration. The skin helps closing the gap between vision-based perception and tactile/force perception.}\n\\label{fig:OcclusionHRI}\n\\end{figure}\nIn human-centered robotics, the typical applications of proximity perception can be broadly divided into two categories: the first one is pertaining a sensitive skin covering the links of a robot manipulator for safety and interaction functionality, which we call \\emph{\\gls{ATI}}. The second one is where a robot gripper or hand is equipped with sensors to support grasping and exploration tasks, which we call \\emph{\\gls{ATII}}. In Fig.~\\ref{fig:OcclusionHRI}, a typical scenario of human-robot interaction and collaboration is illustrated (\\gls{ATI}). As the human approaches the robot, the view of the camera monitoring the robot and its workspace will become increasingly occluded. A tactile skin covering the robot is not adequate to handle this perception gap in general. This is because detecting the human or the environment only when contact is established, implies operating the robot at very low velocities, thus undermining the purpose of installing such a system in the first place. To address scenarios like these, a sensitive skin with proximity perception capabilities has been proposed by several authors. In Fig.~\\ref{fig:Intro_Preshaping}, a typical scenario for grasping supported by proximity perception shown (\\gls{ATII}). Since the robotic hand can detect an object's surface before touch is established, a \\emph{pre-touch} closed-loop control can be implemented to adjust the hand posture during this phase. This is called \\emph{reactive preshaping} and can also have a diversity of use-cases. In Fig.~\\ref{fig:Intro_Preshaping}, the three typical phases of this procedure are shown. Proximity perception also has the potential to play an important role in robotic solutions that are compliant with norms and standards, such as ISO/TS 15066 for the operation of collaborative robots.\n\\begin{figure}\n \\centering\n \\includegraphics[width=0.45\\textwidth]{Figures/Intro_AT-II.jpg}\n \\caption{Reactive preshaping to an object based on proximity perception has three characteristic phases: (1) object detection by vision and approaching of the hand to the object, (2) detection of the object by the hand's proximity sensors, start of closed-loop control and occlusion of the object in the camera view, and (3) finalized preshaping control, where the fingers and palm of the hand are aligned with the object.}\n\\label{fig:Intro_Preshaping}\n\\end{figure}\nIn this paper, we want to provide an up-to-date perspective on the field of proximity perception in human-centered robotics as well as an introduction to the principles and technologies developed. Proximity perception in areas such as autonomous vehicles or \\gls{UAVs} usually aim at autonomous driving or flying and thus address larger distances and speeds and avoidance of contact and interaction with objects and humans. Recent surveys that include discussions on proximity perception in these domains are~ (millimeter wave Radar), (sensor fusion), both for autonomous driving, and~ for indoor localization of \\gls{UAVs}. In human-centered robotics, proximity perception is related to short distances between humans and robots and aiming for improving human-robot interaction as well as safety. Nonetheless, many ideas presented in this paper can be valid in the automotive domain and for \\gls{UAVs} as well, especially those about the sensing principles and their applicability. We hope to give the readers a starting point to understand the principles and applications of proximity perception that have been developed in human-centered robotics. Furthermore, we want to provide a perspective on what, in our opinion, are the important trends that will shape the developments within the next years.\nThe main contributions of this paper are as follows:\n\\begin{itemize}\n \\item We provide an introduction to the concept of proximity perception and an overview of the possible use-cases. We provide a categorization according to the application types and the complexity of the implemented behavior (Sec.~\\ref{sec:ProximitySensor}). We use this categorization throughout the paper to organize the different works found in the literature.\n \\item We give an introduction to the working principles of the most important proximity sensor designs (capacitive, optical, radar, etc.) and give a review of the related work in the context of robotics (Sec.~\\ref{sec:PhysicalWorkingPrinciples}).\n \\item We cover what use-cases for proximity perception have been studied in robotics research and industrial robotics. Here, we give a detailed account of the two most important basic applications, \\gls{ATI} and \\gls{ATII} (Sec.~\\ref{sec:Applications}, see also Figs.~\\ref{fig:OcclusionHRI},~\\ref{fig:Intro_Preshaping}, and~\\ref{fig:ApplicationsAndBehaviors}). We start by giving a historical account and cover the basic forms of behavior possible to the more advanced, cognitive approaches.\n \\item We provide a systematic comparison of the technologies from the field summarized in Table~\\ref{tab:ComparisonYear}\n \\item Finally, we project the current developments into the future and finish the paper with concluding remarks (Sec.~\\ref{sec:FuturePerspectives} and \\ref{sec:SummaryAndConclusions}).\n\\end{itemize}","id":"01d49bee-621f-47ab-9ac0-165e2dba2afa","level":"section","origin_cites_number":3,"parent_id":"4528ef20-9412-4b01-95cc-07b7f164956d","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:ProximitySensor}","id":"7c679291-40be-4834-b2fe-70234a403984","level":"section","origin_cites_number":0,"parent_id":"4528ef20-9412-4b01-95cc-07b7f164956d","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Proximity Sensor: Characterization, Applications and Safety Considerations"]],"subsections":["a3bf09b4-0e28-4364-af23-7c860d5e9658","4c0a7015-4171-43a9-86c9-7a4825899568","fdd92cab-0437-4f63-850b-5ab4cab257fd"],"title":"Proximity Sensor: Characterization, Applications and Safety Considerations"},{"cite_extract_rate":0,"cites":[],"content":"Providing a concise characterization of proximity sensors is challenging. One thing common to all proximity sensors is that they detect objects without physical contact. However, this alone does not distinguish them from cameras, which is problematic, as both modalities are considered to be complementary. To address this, we propose a series of attributes that generally characterize proximity sensor designs. At the same time, not all of the attributes need to be present at once in a particular case. Thus, proximity sensors provide non-contact detection of objects \\emph{and} more often than not\n\\begin{itemize}\n \\item use active measurement principles, i.\\,e. they probe the nearby environment to detect an object's presence, \n item provide limited sensing range and even small detection ranges can be considered to be useful, \n \\item are \\emph{skinlike}, i.\\,e. they can be deployed on surfaces such as robot arm segments as well as fingers where they can form a network of sensing elements,\n \\item are suitable for being highly integrated into the sensory-motor functionality of the robot, enabling reflex-like behaviors due to low latency measurements,\n \\item are used to handle occlusions in vision systems, i.\\,e.\\ are complementary to vision,\n \\item are used to supervise approaching objects which are bound to enter in contact with the robot, i.\\,e.\\ are complementary to tactile sensing.\n\\end{itemize}\nIn Fig.~\\ref{fig:SensingRange}, we propose a definition for the sensing range of a proximity sensor. Any detection distance below $\\SI{50}{\\centi\\meter}$ can be considered to be within the proximity range. This limit is not strict, but in \\gls{HRI} and \\gls{HRC}, this is an approximate distance at which visual occlusions begin to become problematic. As discussed later in Sec.~\\ref{subsec:SafetyConsiderations}, this is a similar range in which monitoring of separation distance is relevant for compliance with safety standards. At larger distances, i.\\,e.\\ mid-range and long-range perception, other technologies (LIDAR, long-range stereo vision, etc.) can provide better performance in workspace surveillance or for providing \\gls{HRI} functionality. This is especially true here because the requirements on reactivity can be relaxed at larger distances. Furthermore, it is interesting to consider the contributions by anthropologist Edward T.\\ Hall, who describes the \\emph{intimate space} of humans as part of his studies on \\emph{proxemics}~. The intimate space starts at a distance of typically $\\SI{45}{\\centi\\meter}$, which is also close to the range proposed above. Therefore, proximity sensing is easy to understand from the perspective of humans, as they can intuitively relate this perception to the ``intimate space\" of the robot by analogy. \nFinally, a distinction can also be made for a range below $\\SI{10}{\\centi\\meter}$ that we call \\emph{pre-touch}-range. This is the type of sensing that precedes contact interactions, for instance during grasping. Here it is especially important to have uninterrupted sensing until contact. Some sensor designs might not feature a long detection range, but the sensing capabilities provided are still useful for closed-loop control of finger and hand posture, which is executed until touch is established. A more in-depth discussion of the available proximity sensing technologies is provided in Sec.~\\ref{sec:PhysicalWorkingPrinciples}. In Fig.~\\ref{fig:LargeScaleModularSkin}, an example of a modern humanoid robot covered in a multi-modal skin is shown, displaying many of the characteristics discussed in this section. \n\\begin{figure}\n \\centering\n \\footnotesize\n \\fontsize{8}{10}\\selectfont\n \\def\\svgwidth{.4\\textwidth}\n \\input{Figures/Generated/Characterization_SensingRange_svg-tex.pdf_tex}\n \\caption{Definition of the proximity sensing range.}\n \\label{fig:SensingRange}\n\\end{figure}","id":"a3bf09b4-0e28-4364-af23-7c860d5e9658","level":"subsection","origin_cites_number":1,"parent_id":"7c679291-40be-4834-b2fe-70234a403984","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Proximity Sensor: Characterization, Applications and Safety Considerations"],["subsection","Characterization"]],"subsections":[],"title":"Characterization"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:ApplicationsAndBehaviorTypes}\nTo talk about proximity sensors in human-centered robotics as a whole, it is useful to consider first a categorization of the possible applications and desired behaviors. In the introduction, we already mentioned that a broad classification of applications into two categories is possible: the ones relating to safety and \\gls{HRI}/\\gls{HRC} and the ones relating to preshaping and grasping (Figs.~\\ref{fig:OcclusionHRI} and~\\ref{fig:Intro_Preshaping}). Beyond this, automated behaviors based on proximity sensors can be organized according to their conceptual complexity and how instantaneous their effect is on the movement of the robot. One example for a low-complexity behavior is a \\emph{safety stop}, i.\\,e.\\ enabling the brakes of the robot based on a sensor signal surpassing a threshold value. This behavior is closely tied to the update-rate of the sensors and the low-level robot controller. In that sense, it can be called \\emph{reactive} or \\emph{reflex-like}. Modern collaborative robots, e.\\,g.\\ the Franka Emika Panda~ or the KUKA LBR iiwa~ have control loop cycles of $t_{cl}=1~ms$. Thus, the closer the response time of the proximity sensor is to $t_r<t_{cl}$, the better. An example of high-complexity behavior is object exploration. It involves managing an object model as well as a planner to complete this model with purposeful exploration steps, resulting in a robot behavior that is executed in several phases and over a longer time compared to the basic control loop cycle times. This behavior is also characterized by being executed at different layers, reaching, as mentioned, up to the planning and cognitive components in the robot's architecture. Fig.~\\ref{fig:ApplicationsAndBehaviors} illustrates the categorizing of applications and behaviors we propose as well as providing some examples (not an exhaustive list). As a result, we have a broad classification of applications into two types, \\gls{ATI} (left) and \\gls{ATII} (right), and behaviors into two types, \\gls{BTI} (bottom) and \\gls{BTII} (top). In general, \\gls{BTI} will appear as subsystems of \\gls{BTII}.\n\\begin{figure}\n \\centering\n \\footnotesize\n \\includegraphics[width=0.97\\linewidth]{Figures/ApplicationsAndBehaviorsV5.pdf}\n \\caption{Categorization of some use-cases for proximity sensors in robotics according to the possible application types (\\gls{ATI} and \\gls{ATII}) and behavior types (\\gls{BTI} and \\gls{BTII})}.\n \\label{fig:ApplicationsAndBehaviors}\n\\end{figure}\n\\begin{figure}\n \\centering\n \\fontsize{8}{10}\\selectfont\n \\def\\svgwidth{0.48\\textwidth}\n \\input{Figures/Generated/Humanoid_Cheng+Modules_svg-tex.pdf_tex}\n \\caption{Left: Researchers at the Technical University of Munich, Chair for Cognitive Systems, have developed a multi-modal, modular skin, which they have deployed on their robot H-1. Right: A single cell of the multi-modal sensor. (Copyright left picture. A. Eckert / TU M\\\"unchen. Right picture \\copyright 2019 IEEE) . Both reprinted with kind permission of the authors.)}\n \\label{fig:LargeScaleModularSkin}\n\\end{figure}","id":"4c0a7015-4171-43a9-86c9-7a4825899568","level":"subsection","origin_cites_number":3,"parent_id":"7c679291-40be-4834-b2fe-70234a403984","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Proximity Sensor: Characterization, Applications and Safety Considerations"],["subsection","Application and Behavior Types"]],"subsections":[],"title":"Application and Behavior Types"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:SafetyConsiderations}\nFrom the safety perspective, a proximity sensor deployed on a collaborative robot in an industrial environment has to fulfill the requirements of the ISO/TS 15066 complying with the ISO 10218 for robots and robotic devices, as well as the performance level defined in the safety of machinery ISO 13849. Proximity sensing hardware is very well suited to operate a collaborative robot in the \\emph{speed and separation monitoring mode} as defined in the ISO/TS 15066 to monitor the protective separation distance $S_p$ between a human and a robot's surface. According to the standard, the following equation has to be fulfilled during the operation mode:\n\\begin{equation}\n \\label{eq:separationDistance}\n \\small\n \\begin{aligned}\n S_{p}(t=t_0) = v_{h} (T_{r}+T_{s}) + v_{r} T_{r} + S_{s} + C_{i} + Z_{d} +Z_{r},\\\\\n \\end{aligned}\n\\end{equation}\nwhere $v_h$ is the human speed (if not monitored, $v_h=\n\\SI{1.6}{\\meter\\per\\second}$)\n$T_r$ is the reaction time of the robot, $T_s$ is the robot stopping time, $v_r$ is the robot speed, $S_s$ is the robot stopping distance, $C_i$ is the intrusion distance, $Z_d$ is the position uncertainty of the human and $Z_r$ is the position uncertainty of the robot.\nTo provide an illustrative example for a state of the art collaborative robot, we look at the UR10e series. It has a control loop cycle of $\\SI{500}{\\hertz}$ and the safety parameters can be configured $v_r =\\SI{5}{\\meter\\per\\second}$ (max end-effector speed)\n$T_r = \\SI{4}{\\milli\\second}$ (two control loop cycles), $T_s = \\SI{100}{\\milli\\second} $, $S_s = \\SI{50}{\\milli\\meter}$, $Z_r = \\SI{0.05}{\\milli\\meter}$. This configuration results in separation distance of $S_p=\\SI{0.236}{\\meter}$ excluding the uncertainty of the position of the human and the intrusion distance, as this depends on the sensor parameters monitoring the area.","id":"fdd92cab-0437-4f63-850b-5ab4cab257fd","level":"subsection","origin_cites_number":0,"parent_id":"7c679291-40be-4834-b2fe-70234a403984","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Proximity Sensor: Characterization, Applications and Safety Considerations"],["subsection","Safety Considerations and Norm Compliance"]],"subsections":[],"title":"Safety Considerations and Norm Compliance"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:PhysicalWorkingPrinciples}\nIn this section, we will give an introduction to the main physical principles available to implement proximity sensing. The idea is to be able to ease the process of reviewing articles by starting with an explanation of the basics. This will also help us in the systematization in Table~\\ref{tab:ComparisonYear}. There, we use the abbreviations introduced in this section. Here, we will already do a review of some representative works in the field focusing on how the proximity sensing technology is implemented. This section is closely linked to Sec.~\\ref{sec:Applications}, where the details of the implemented applications are discussed. We try to cross-reference the most relevant relationships. However, favoring readability, cross-referencing is not exhaustive.\n\\input{Content/03.01-CapacitiveSensing.tex}\n\\input{Content/03.02-OpticalSensing.tex}\n\\input{Content/03.03-Radar.tex}\n\\input{Content/03.04-OtherSensingPrinciples.tex}\n\\input{Content/03.05-Multi-Modal}","id":"702ec4ea-e939-40cf-abd8-8e5caea8edc6","level":"section","origin_cites_number":0,"parent_id":"4528ef20-9412-4b01-95cc-07b7f164956d","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Measurement Principles for Proximity Sensing"]],"subsections":[],"title":"Measurement Principles for Proximity Sensing"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:Applications}\nIn this section, we will review the contributions from the field focusing on the applications and methods presented. We will follow the organization presented in Sec.~\\ref{subsec:ApplicationsAndBehaviorTypes} and Fig.~\\ref{fig:ApplicationsAndBehaviors}, i.\\,e.\\ focusing separately on \\gls{ATI} \nand \\gls{ATII} and going from low-complexity behaviors (\\gls{BTI}) to high-complexity behaviors (\\gls{BTII}).\n\\input{Content/04.01-CollisionAvoidance.tex}\n\\input{Content/04.02-Preshaping.tex}\n\\input{Content/04.03-HigherLevelApplications.tex}\n\\input{Content/04.04-IndustrialTechnologiesAndSolutions.tex}","id":"55e0e327-ca41-4cdf-b82e-23bda770d2ce","level":"section","origin_cites_number":0,"parent_id":"4528ef20-9412-4b01-95cc-07b7f164956d","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Applications and Methods in the Research and Industry Domains"]],"subsections":[],"title":"Applications and Methods in the Research and Industry Domains"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:FuturePerspectives}\nIn this section, we provide an outlook for the domain of human-centered proximity perception in robotics. We have highlighted what we think are \\emph{grand challenges} at the end of each the three sub-sections.","id":"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe","level":"section","origin_cites_number":0,"parent_id":"4528ef20-9412-4b01-95cc-07b7f164956d","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Future Perspectives: Grand Challenges"]],"subsections":["017a9c7a-b440-4a26-8ec2-d03025b12bb0","b968a456-2525-49a5-a737-0c5965cfb1de","c6615d1d-30c1-4597-a3db-191781a41f9f"],"title":"Future Perspectives: Grand Challenges"},{"cite_extract_rate":0.25,"cites":[2680,2679],"content":"\\label{subsec:HRIInIndustryAndServiceRobotics}\nProximity perception technology is mature enough to be deployed under strict safety requirements, as discussed in Sec.~\\ref{subsec:IndustrialTechnologies}. However, today, collaborative automation still struggles to be an interesting value proposition, i.\\,e.\\ providing a large increase in efficiency that will justify the cost of investing in this technology. A value proposition that is more likely to be attractive in the near future is that of \\emph{fenceless automation}. Here, humans and robots coexist in the same space but do not necessarily share a task. Value is added for instance by the fact that real-estate on the shop-floor is saved or because robotic automation is now possible in spaces that were previously considered to be too small. Nonetheless, safety certification is still a major challenge that is preventing wide-spread use of proximity perception. Today, technologies such as the ones discussed in ~\\ref{subsec:IndustrialTechnologies} need to be certified on a solution level. Certification at the modular sensor level is not yet possible. Therefore, technologies such as radar (see Sec.~\\ref{subsec:Radar}) can be attractive alternatives for HRI in industrial automation. Radar is more likely to achieve a safety rating on modular level soon, also profiting from all the prior experience coming from the developments in autonomous driving.\nAreas such as medical robotics face similar issues for commercialization. The process of certifying the solutions is long and costly. However, we think that the market for automated solutions in health care or service robotics based on proximity perception is there, especially in scenarios where the human is very close to the robot, preventing the sole use of cameras. In this survey, we discussed examples such as grasping of moving objects and handover~, dressing and washing of patient , or assistive robotics based on teleoperation . Overall, we think that medical and service robotics is a promising field for proximity perception, but more research is needed to establish use-cases and the corresponding methods before commercial exploitation appears on the horizon.Furthermore, the study of highly redundant robotic systems that use proximity sensing is scarce. Meanwhile, the exploitation of kinematic redundancy based on proximity sensor feeds is natural, as discussed in Secs.~\\ref{subsubsec:EarlyJacobian} and \\ref{subsubsec:RecentJacobian}. This includes the use of proximity perception in unusual areas, such as on the legs or feet of robots. More research, like the one done in the group of Prof.\\ Cheng~, is needed. In summary, we can say the following about the challenges in these domains:\n\\begin{itemize}\n \\item Lowering the difficulties in achieving safety ratings for proximity sensing technologies will significantly expand the market for fenceless or collaborative automation. An effective procedure for safety certification, elaborated by industrial stakeholders and certification organizations, is needed.\n \\item Modularized technologies, such as radar chips, can have significant advantages in a certification procedure because solutions can be based on safety rated components.\n \\item In the service, medical and active assisted living domains proximity perception allows for new ways of interaction between humans and robots yet more research is needed before commercially viable applications are established.\n \\item A better understanding of how highly redundant robotic systems can profit from proximity perception is needed.\n \\item The potential of proximity perception for increasing robustness of grasping of moving objects, for instance during handover tasks or in teleoperation, needs to be established further, for example with conclusive user studies.\n\\end{itemize}","id":"017a9c7a-b440-4a26-8ec2-d03025b12bb0","level":"subsection","origin_cites_number":8,"parent_id":"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Future Perspectives: Grand Challenges"],["subsection","Human-robot interaction in the industry and service domains"]],"subsections":[],"title":"Human-robot interaction in the industry and service domains"},{"cite_extract_rate":0.6000000000000001,"cites":[2681,2683,2682],"content":"\\label{subsec:CognitiveRobotics}\n\\begin{figure}\n \\centering\n \\includegraphics[width=0.95\\linewidth]{Figures/Illustration_pointcloudV3.pdf}\n \\caption{Proximity perception can play an important role in multi-modal exploration: a) Shows a typical scene with objects as captured by an RGBD-camera. b) A perspective shift reveals considerable gaps in the visual perception, having occluded elements that could be explored contactless and contact-based exploration steps, c) and d) respectively. A multi-modal cognitive model controls the exploration and aggregates the information to the object model, which includes geometric, material property, fill-level information as well as potential grasping regions obtained from the exploration steps in c) and d). This aggregated information is illustrated in e). The yellow and red colors on the mug (in c)-d)) illustrate regions that are extracted from the proximity and touch-based perception, respectively. }\n\\label{fig:futureperspectives}\n\\end{figure}\nOne of the hallmarks of cognitive robotics is active perception. It is the paradigm in which perception is enabled by purposeful actions, i.\\,e.\\ exploration). The active perception principle has been studied in many robotics research papers, as discussed in a survey by Bohg et al.~. Machine learning is essential in active perception for tasks such as object pose estimation or scene labeling, which can be based on proximity perception, e.\\,g.~. Currently, a trend in active perception is the combined perception of vision and touch. We think that extending this trend to include proximity sensing will be a relevant research question in the near future. Therefore, we will see haptic exploration strategies that will be able to explore occluded regions of the workspace combining touch-based and touch-less exploratory movements. We have illustrated this possible multi-modal exploration approach in Fig.~\\ref{fig:futureperspectives}. Some aspects of it have already been addressed in the literature discussed in this paper, e.\\,g.~. Another important domain in this area is the multi-modal modeling of the human for safe interaction and collaboration. In summary, regarding the major challenges we can say that:\n\\begin{itemize}\n \\item As we mentioned throughout the paper, there is a need for proximity sensing technologies that can easily be combined with vision and/or touch so they can be deployed together. This will make it attractive to include proximity sensing in established and novel active perception/cognitive approaches.\n \\item It can be expected that the trend regarding sim-to-real learning is going to prevail for the next years. Therefore, it is a challenge to implement realistic simulation models for the different measurement principles discussed in Sec.~\\ref{sec:PhysicalWorkingPrinciples}. In most cases, current approaches to model them are not viable in terms of their temporal performance. Overcoming this is an important challenge.\n\\end{itemize}","id":"b968a456-2525-49a5-a737-0c5965cfb1de","level":"subsection","origin_cites_number":5,"parent_id":"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Future Perspectives: Grand Challenges"],["subsection","Cognitive Robotics"]],"subsections":[],"title":"Cognitive Robotics"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsec:SoftRobotics}\nWe think that a good portion of the topics relevant for proximity perception will eventually transfer to Soft Robotics. In this way, soft manipulators equipped with proximity sensors can perform collision avoidance in an analogous way as described in Sec.~\\ref{subsec:CollisionAvoidance}, i.\\,e.\\ respecting task hierarchies. Similarly, soft robots will be able to execute reactive preshaping, grasping, and exploration tasks using proximity sensors. Arguably more so than in other areas, in Soft Robotics it is of interest to study how to purposefully engage in contacts to achieve the desired task. Proximity perception can help in finding and reaching desired contact states. Overall, the support of simulation frameworks for model-based control and sensing, such as SOFA~, will also be relevant. Therefore, regarding the major challenges in this domain, we can say that:\n\\begin{itemize}\n \\item New control methods must be found for soft robots to integrate the information provided by the proximity sensors and adapt the actuation strategy adequately.\n \\item Integration of proximity sensors in deformable structures represents an important challenge. A significant cross-talk between the global deformation and the detection of tactile and proximity events has to be expected. As stated in the previous section, appropriate models for proximity sensors are needed for sim-to-real learning or for interpreting their signals correctly under deformation using interactive simulations~. \n \\item Except for capacitive sensing, the realization of proximity sensors having deformable or stretchable sensing elements is challenging. Shrinking the size of rigid components, such as ICs, within a stretchable substrate may still allow the realization of deformable circuits, as described e.\\,g. in .\n\\end{itemize}","id":"c6615d1d-30c1-4597-a3db-191781a41f9f","level":"subsection","origin_cites_number":4,"parent_id":"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Future Perspectives: Grand Challenges"],["subsection","Soft Robotics"]],"subsections":[],"title":"Soft Robotics"},{"cite_extract_rate":0.037735849056603,"cites":[2680,2679,2681,8577],"content":"\\label{sec:SummaryAndConclusions}\nIn this paper, we have given an overview of the main aspects of proximity sensing in today's robotic landscape. Considering that the field has not had a significant formalization over the years, we provide a basic scheme for categorization of the robotic applications and technologies based on proximity sensors and we propose a set of traits that characterize proximity sensors in human-centered robotics. We give an account of the existing technologies and the main measurement principles reported by authors since the early 1970s and have organized the technologies in Table~\\ref{tab:ComparisonYear}, including characteristics such as sensing range, update rate, sensing element size, etc. We then proceeded to detail how the technologies have been used for implementing applications such as collision avoidance and human-robot interaction (\\gls{ATI}) as well as preshaping and grasping (\\gls{ATII}). We start with the seminal developments of the early years in these domains and cover the progress up to today (2021). The tight integration into the sensory-motor functionality has received constant attention over the years in order to realize highly reactive behavior of robotic systems (\\gls{BTI}). Meanwhile, as the area of robotics progresses as a whole, we report that more and more approaches begin to have cognitive aspects in them (\\gls{BTII}). \\gls{BTII} includes areas such as teleoperation, where the human is in the loop, autonomous object exploration and even bio-inspired approaches that mimic weakly electric fish. \nFinally, Sec.~\\ref{sec:FuturePerspectives} is dedicated to summarizing our projections for the field regarding the grand challenges we have identified. We think that as the technology of proximity sensors is reaching the maturity to coexist with tactile and visual perception in terms of integration, costs, and norm conformity, it will be adopted for a variety of solutions, especially those involving interaction with humans. \n\\section*{Acknowledgment}\nThis work was supported by the Region Hauts-de-France, the project COMOROS (ANR-17-ERC2-0029), the European Regional Development Fund (ERDF) and the project Inventor (I-SITE ULNE, le programme d’Investissements d’Avenir, M\\'etropole Europ\\'eenne de Lille). \nThis work has received funding from the \"K\\\"arntner Wirtschaftsf\\\"orderung Fonds\" (KWF) and the \"European Regional Development Fund\" (EFRE) within the CapSize project 26616/30969/44253. \nThis work also was supported by the Federal Ministry of Education and Research (Bundesministerium f\\\"ur Bildung und Forschung, BMBF) within the project (Verbundprojekt-Nr.: 16SV7823K:). \n The authors would like to thank cartoonist Adriana Filippini for the illustrations. \n\\ifCLASSOPTIONcaptionsoff\n \\newpage\n\\fi\n\\bibliographystyle{IEEEtran}\n\\bibliography{ReviewProximityPerception}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/escaidanavarro}}]{Stefan Escaida Navarro}\nreceived his diploma in computer science (Dipl.-Informatiker) from Karlsruhe Institute of Technology (KIT), Germany in 2010. He obtained is PhD degree (Dr.-Ing.) from the same institution in 2016. During his doctoral phase he researched on haptic object recognition and grasping as well as the technology and the applications for capacitive tactile proximity sensors, such as preshaping, collision avoidance and teleoperation. He is currently a postdoctoral researcher at Inria Lille - Nord Europe, France, where he is working on model-based force and shape sensing for Soft Robotics. \n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/muehlbacherkarrer}}]{Stephan M\\\"{u}hlbacher-Karrer}\nreceived the Dipl.Ing. degree in telematics from the Graz University of Technology (TUG), Graz, Austria, in 2012, where he focused on Autonomous Robots and Telecommunications. In 2017 he received his Dr. tech. degree in information and communications engineering from the Alpen-Adria-Universit\\\"{ä}t Klagenfurt, Klagenfurt, Austria, focusing on capacitive sensing for robotic applications.\nFrom 2012 to 2013, he was with Gigatronik Ingolstadt GmbH, Gaimersheim, Germany, working\nfor Audi Electronics Venture GmbH, Gaimersheim, where he was involved in the pre-development for autonomous driving. Since 2017 he is a senior researcher with JOANNEUM RESEARCH ROBOTICS, Institute for Robotics and Mechatronics in Klagenfurt Austria. His current research interests include near field sensor technology, signal processing, and autonomous robots.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/alagi}}]{Hosam Alagi}\nreceived his M.Sc. degree in electrical engineering and information technology at the Karlsruhe Institute of Technology in Germany with a specialization in System-on-Chip. Since August 2015 he is working as a research assistant at the Institute for Anthropomatics and Robotics (IAR) at Intelligent Process Automation and Robotics Lab (IPR). Focusing on human-robot interaction, he develops capacitive multi-modal sensors and evaluates methods that allow robots to perceive their near environment. Within the scope of the PhD Studies both the sensors and sensor applications are being implemented, addressing the industrial and domestic domains.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/hein}}]{Bj\\\"{o}rn Hein}\nstudied electrical engineering with a focus on control theory and received his PhD in 2003 concerning automatic collision-free path planning at the Karlsruhe Institute of Technology (KIT). May 2010 he finished his post-doctoral lecture qualification for computer science with stress on human-robot interaction. 2012-2018 he was professor for Interaction Technologies for Robot Systems at the Institute for Anthropomatics and Robotics (IAR) - Intelligent Process Automation and Robotics Lab (IPR). Currently he holds a professorship at the University of Applied Sciences in Karlsruhe. His research focus comprises: algorithms for collision free motion planning and path optimization, methods for intuitive and automatic programming of robots, human-robot interaction, novel sensor-technologies for enhancing capabilities of robot systems.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/zangl}}]{Hubert Zangl} received the Dipl.Ing. degree in telematics, the Dr. Tech. degree in electrical engineering, and the Venia Docendi degree in sensors and instrumentation from the Graz University of Technology (TUG), Graz, Austria, in 2001, 2005, and 2009, respectively.\nFrom 2010 to 2013, he was an Associate Professor of Sensors and Instrumentation with the Institute of Electrical Measurement and Measurement Signal Processing, TUG. Since 2013, he has been a Professor chairing Sensors and Actuators with the Institute of Smart System Technologies, Alpen-Adria-Universität Klagenfurt, Klagenfurt, Austria. \nHis current research interests include the design and optimization of smart sensors and actuators, robustness, and reliability of sensors and actuators, sensor signal processing, autarkic wireless sensors, and energy harvesting with target applications in the field of IoT and robotics.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/koyama}}]{Keisuke Koyama} received the B.E., M.E. and Ph.D. in Eng. degrees in mechanical engineering and intelligent Systems from the University of Electro-Communications (UEC), \n in 2013, 2015 and 2017, respectively. He was a research fellow of Japan Society for the Promotion of Science (JSPS) from 2015 to 2017. \n He researched on high-speed proximity sensor for pre-grasping and integrated control of multi-degree-of-freedom robot hand and arm. \n From 2017 to 2019, he was project Assistant Professor, Department of Information Physics and Computing, Graduate School of Information Science and Technology, The University of Tokyo. \n Since 2019, he has been Assistant Professor, Department of Systems Innovation, Graduate School of Engineering Science, Osaka University. \n And, he has been Visiting Researcher, Department of Information Physics and Computing, Graduate School of Information Science and Technology, The University of Tokyo. \n His current research focuses on high-speed, high-precision proximity sensing for high-speed robotic manipulation and assembly. \n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/christian_duriez}}]{Christian Duriez} is Research Director at Inria Lille - Nord Europe.\nHe received the engineering degree from the Institut Catholique d’Arts et Métiers of Lille, France and a PhD degree in robotics from University of Evry and CEA in France in 2004. He had a postdoctoral position at the CIMIT SimGroup in Boston. He arrived at INRIA in 2006 and worked on interactive simulation of deformable objects and haptic rendering with focus on surgical simulation. He is now the head of DEFROST team, created in January 2015. In 2018 he was invited researcher at Stanford University. His research topics are Soft Robot models and control, Fast Finite Element Methods, simulation of contact response, new algorithms for haptics… He participated to the creation of the open-source SOFA framework. He was also one of the founders of the start-up company InSimo.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/smith}}]{Joshua R. Smith} is the Milton and Delia Zeutschel Professor, jointly appointed in the Allen School of Computer Science and Engineering, and the Department of Electrical Engineering, at the University of Washington, Seattle. He received BA degrees in Computer Science and Philosophy from Williams College in 1991, an MA in Physics from the University of Cambridge in 1997, and SM and PhD degrees in 1995 and 1999 from the MIT Media Lab's Physics and Media Group, where he began working on Electric Field Sensing for input device applications. He began applying these techniques to robotics while he was a researcher at Intel Labs Seattle, between 2004 and 2010. In recent years his group at UW has developed a number of new approaches to proximity perception for robotics.\n\\end{IEEEbiography}\n\\newpage\n\\markboth{APPENDIX to Proximity Perception in Human-centered Robotics - A Survey on Sensing Systems and Applications}\n{Shell \\MakeLowercase{\\textit{et al.}}: Bare Demo of IEEEtran.cls for IEEE Journals}\n\\onecolumn\n\\setcounter{page}{1}\n\\begin{center}\n\\Huge{Appendix} \\\\ \\vspace{.5cm}\n\\end{center}\n\\printglossaries\n\\clearpage\n\\newpage\n\\begin{landscape}\n\\begin{center}\n\\large{Overview - Sorted by year - Table I}\n\\end{center}\n\\tiny\n\\begin{doclongtable}{p{0.3cm}p{1.3cm}p{0.8cm}p{0.8cm}p{1.1cm}p{1.3cm}p{1.1cm}p{1.3cm}p{2.1cm}p{1.3cm}p{2.7cm}p{3.5cm}p{1.3cm}} \n\\toprule\n{} & Measurement Principle & Reference & Year & Reported Min. Working Distance [mm] & Reported Max. Range [mm] & Field of View [degree] & Measurement Rate [Hz] & Sensing Element Dimension [$mm$ / $mm^2$ / $mm^3$] & Multiple-Obstacles & Commercially Available Core Components & Categorization & Basis reference \\\\\n\\midrule\n\\endhead\n\\midrule\n\\multicolumn{13}{r}{{Continued on next page}} \\\\\n\\midrule\n\\endfoot\n\\bottomrule\n\\endlastfoot\n1 & O-Tri & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Manipulation and Grasping & - \\\\\n2 & O-Tri & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Grasping, Teleoperation & \\\\\n4 & O-Tri & & 1984 & n/s & v1: focal point 4.5 +-0.56 cm, v2: focal poin... & 13 & 1000 & 110 & no & no & AT-II, measure distance, orientation and curva... & - \\\\\n7 & O-RLI & & 1988 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\n6 & C-M & & 1988 & 40 & 100 & n/s & n/s & 80x210 & not tested & no & AT-I Obstacle Detection & - \\\\\n8 & O-RLI & & 1989 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\n9 & A-US & & 1990 & n/s & 200-300 & 180 & n/s & ~20x20x20 & n/s & n/s & AT-II & - \\\\\n10 & C-M & & 1991 & 0 & 304.8 human and alu, 127 grafite lead & n/s & n/s & 355,6x190,5 & no & n/s & AT-I, Collision Avoidence & - \\\\\n13 & O-RLI & & 1992 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & OD8810 and SFH205 & AT-I, Collision Avoidence & - \\\\\n14 & O-RLI & & 1992 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidence & - \\\\\n12 & C-M & & 1992 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & - \\\\\n16 & O-RLI & & 1993 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidance & \\\\\n17 & O-RLI & & 1993 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\n15 & C-M & & 1993 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\n19 & C-M & & 1994 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\n18 & A-US & & 1994 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Contour Following & - \\\\\n20 & O-RLI & & 1996 & 5 & 75 & 100 (+-50) & 300 & ~5x5x13 & no & STM & AT-II, Manipulation and Grasping & - \\\\\n22 & O-RLI & & 1997 & n/s & 40 & 60 (+-30) & 10000 & ~32 & no & no & AT-II, Manipulation and Grasping & - \\\\\n24 & O-RLI & & 1997 & n/s & 200 & 130 (+-65) & n/s & 2020-08-07 00:00:00 & no & no & AT-II, Object Tracking for Grasping & - \\\\\n23 & A-US,O-RLI & & 1997 & 25 (A-US), 50 (O-RLI) & 3000 (A-US), 450 (O-RLI) & n/s & 40 (A-US), 100 (O-RLI) & 63.5x44.5x20 & yes & no & AT-I, Sensor Skin & - \\\\\n25 & O-BB & & 2000 & n/s & n/s & n/s & n/s & n/s & no & no & AT-II, BT-I, Sensing for Preshaping & - \\\\\n27 & C-SE & & 2006 & 0 & n/s & n/s & 200 & 15x5 & no & MC33794 & AT-I, Sensor Skin & - \\\\\n29 & O-RLI & & 2007 & 0 & 50 & 360 & 48 & 1 & yes & no & AT-I, BT-I, Contour Following & - \\\\\n28 & C-M & & 2007 & n/s & 50 & 180 & n/s & ~50 & no & no & AT-II, Sensing for Preshaping & - \\\\\n31 & C-M & & 2008 & n/s & n/s & n/s & 500 & n/s & n/s & no & AT-II, Material recognition & - \\\\\n32 & C-M & & 2008 & n/s & 190 & n/s & 30 & ~50 & n/s & no & AT-II, Proximity Servoing & - \\\\\n33 & O-Tri & & 2009 & 10 & 60 & n/s & 224 & 90 x 90 & yes & TCRT1000 & AT-I, AT-II, Sensor Array & - \\\\\n36 & O-RLI & & 2009 & 2 & 40 & 4 alnaog sensors per finger = 90 & n/s & 6x3.7x3.7 & n/s & Vishay TCND5000 & AT-II, Preshaping and Grasping & - \\\\\n35 & C-M & & 2009 & n/s & 170 & n/s & n/s & 22x22 & no & no & AT-I, AT-II, Sensor & - \\\\\n38 & C-SE, C-M & & 2010 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & Preshaping, Grasping, Exploration, Teleoperati... & - \\\\\n37 & C-M & & 2010 & 0 & 150 & n/s & 20 (update cycle for control) & Fingertip of Barret Hand & not tested & no & AT-II, Preshaping, Grasping and Co-manipulation & - \\\\\n40 & C-M & & 2010 & n/s & n/s & n/s & n/s & n/s & yes & BOSCH (complete solution) & AT-I, Sensor & - \\\\\n42 & C-M & & 2010 & 0 & 160 & 180 & 25000 & 220x150 & no & no & Automotive Seat Occupancy & - \\\\\n43 & O-RLI & & 2011 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Modular Multi-Modal Skin & - \\\\\n44 & C-M & & 2011 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, bioinspired & - \\\\\n46 & C-M & & 2011 & 0 & 2000 & 360 & n/s & 30 x 570 & yes & AD7143 & Automotive parking & - \\\\\n48 & O-RLI & & 2012 & 1 & 10 & narrow & n/s & 15x15 & no & ADNS-9500 & AT-II, Slip Detection, Object Rreconstruction,... & - \\\\\n49 & A-US & & 2012 & 20 & 400 & 15 & 40 & 45 x 20 x 15 & no & HC-SR04 & AT-II legged climbing robot & - \\\\\n47 & A-S & & 2012 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & - \\\\\n57 & O-Tri & & 2013 & 200 & 1500 & n/s & 250 (20 sensors) & ~20 x 40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & - \\\\\n52 & O-RLI & & 2013 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & - \\\\\n59 & O-RLI & & 2013 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping & - \\\\\n54 & C-SE, C-M & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Multi-Modal Sensor System & - \\\\\n55 & C-SE & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Hand and Object Tracking, Collision Pre... & \\\\\n51 & C-M & & 2013 & 0 & $>$100 & 180 & 6250 & 1500x150x0.35 & no & no & AT-I, BT-I, Reactive Collision Avoidance & - \\\\\n60 & C-M & & 2013 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, Bio-inspired & \\\\\n61 & A-S & & 2013 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & \\\\\n63 & O-Tri & & 2014 & 200 & 1500 & n/s & 250 (20 sensors) & ~20x40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & \\\\\n62 & C-SE & & 2014 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Preshaping, Grasping and Exploration & \\\\\n65 & C-M & & 2014 & 0 & 90 & n/s & n/s & 5x5 & yes & no & Sensor System & - \\\\\n66 & O-RLI & & 2015 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & - \\\\\n67 & O-RLI & & 2015 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II & \\\\\n72 & O-RLI & & 2015 & n/s & 300 & n/s & 1000 & 100x100 & not considered & no & AT-I, AT-II & - \\\\\n73 & O-BB & & 2015 & 0 & $>$84 & n/s & n/s & 32x19x6.5 & no & - & AT-II, Preshaping and Grasping & - \\\\\n68 & C-M & & 2015 & 0 & ~250 & 360 & n/s & whole body & no & no & AT-I, AT-II, Bio-Inspired & - \\\\\n69 & C-M & & 2015 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Teleoperation, Exploration & \\\\\n70 & C-M & & 2015 & 5 & 50 & Cross Section & n/s & 200x200 & yes & no & AT-I, BT-II Object detection, tomography & - \\\\\n71 & C-M & & 2015 & 0 & n/s & 180 & quasi-simultaneously & Meka H2 Finger & no & no & At-II, BT-II Active Object Categorization, Gra... & - \\\\\n83 & R & & 2016 & n/s & 500 (for given resolution) & 58 (E) 57 (H) & n/s & 13,2 (antenna diameter) & yes & no & BT-II, Object exploration & - \\\\\n76 & O-RLI & & 2016 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & \\\\\n77 & O-RLI & & 2016 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\n81 & O-RLI & & 2016 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & \\\\\n78 & C-SE, I & & 2016 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\n80 & C-SE, C-M & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\n75 & C-SE & & 2016 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-I, Contour Following & \\\\\n79 & C-SE & & 2016 & 0 & 350 & n/s & 40 & n/s & yes & MRK-Systeme (complete solution) & AT-II, HRI & - \\\\\n74 & C-M & & 2016 & 0 & 50/100 & 180 & n/s & 100x150 & yes & no & AT-II, Gesture Control, Grasping \\& Object Man... & - \\\\\n82 & C-M & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Preshaping and Grasping & \\\\\n86 & O-ToF, C-SE & & 2017 & 10 & 100 & 180 for whole fipgertip (6 module) & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\n85 & O-ToF & & 2017 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Preshaping and Grasping & \\\\\n84 & O-RLI & & 2017 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, BT-II & \\\\\n88 & C-SE, I & & 2017 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\n87 & A-US & & 2017 & n/s & 6000 & 180 & \\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\n96 & R & & 2018 & n/s & n/s & n/s & n/s & n/s & yes & yes & AT-II, Grasping & - \\\\\n97 & O-Tri & & 2018 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II Dynamic grasping & - \\\\\n98 & O-ToF & & 2018 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\n100 & O-ToF & & 2018 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x & AT-II, Teleoperation & \\\\\n91 & O-RLI, O-ToF & & 2018 & 0 & 150 & n/s & 1000, 100 & 3.2 x 1.9 x 1.1, 4.8 x 2.8 x 1 & no & EE-SY 1200, VL6180x & AT-II, Preshaping and Grasping & - \\\\\n90 & O-RLI & & 2018 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, Object Exploration & \\\\\n93 & O-RLI & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\n94 & O-RLI & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\n95 & C-SE & & 2018 & 0 & 100 & n/s & 200 & 115 x 85 x 1 & n/s & MPR121 & AT-II, Contour Following, Health Care & - \\\\\n89 & C-M & & 2018 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-II, Material Recognition & \\\\\n92 & C-M & & 2018 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Material Recognition & \\\\\n99 & A-US & & 2018 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Teleoperation & - \\\\\n104 & R & & 2019 & 300 & n/s & n/s & n/s & 300 (waveguide length) & yes & no & AT-I & - \\\\\n110 & O-Tri & & 2019 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II, Preshaping and Grasping & \\\\\n102 & O-ToF & & 2019 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Multi-Modal Sensor System & \\\\\n108 & O-RLI & & 2019 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Whole-Body Control & \\\\\n109 & O-RLI & & 2019 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\n101 & C-SE, O-ToF & & 2019 & 10 & 100 & 180 for whole & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-I, Multi-Modal Skin & - \\\\\n103 & C-SE & & 2019 & 0 & n/s & n/s & 100 & 30 x 30 & no & Teensy 3.2 & AT-II, Contour Following, Health Care & - \\\\\n106 & C-SE & & 2019 & 0 & 300 & n/s & 125 & 50 x 50 to 100 x 100 & yes & FOGALE Robotics (complete solution) & AT-I, Collision Avoidance & - \\\\\n111 & C-SE & & 2019 & 0 & 50 & 180 & n/s & 5 x 5 & no & no & AT-II, Grasping in Harsh environments, Industr... & - \\\\\n107 & C-M, O-ToF & & 2019 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I, Collision Avoidance & \\\\\n105 & A-US & & 2019 & 0 & ~8 & n/s & n/s & ~30x14x14 & no & no & AT-II, Material Recognition & - \\\\\n113 & R & & 2020 & n/s & $>$1200 & 160 & 40 & 9x13 & yes & yes & AT-I, BT-II & - \\\\\n118 & O-Tof, O-RLI & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x, MAX30105 & AT-II, Object Exploration & - \\\\\n116 & O-ToF & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & \\\\\n119 & O-RLI & & 2020 & n/s & n/s & n/s & n/s & 7.1 x 2.75 x 2.7 & no & Broadcom Limited, HSDL-9100-021 & AT-II, Teleoperation, VR, Transcutaneous Elect... & - \\\\\n115 & C-SE, O-ToF & & 2020 & 0 & 350 & 180 & 50 & 27x27 & no & no & AT-II, HRI & - \\\\\n117 & C-SE, I & & 2020 & 0 & 300 & n/s & 100 & 100 x 100 x 2.75 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\n114 & C-SE & & 2020 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Teleoperation, Tactile Feedback & \\\\\n121 & C-SE & & 2020 & 0 & 350 & 180 & \\~108 & \\~175 x 80 & no & AD7147 & AT-I, Collision Avoidance & \\\\\n112 & C-M, ToF & & 2020 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I Collision Avoidance & \\\\\n120 & A-US & & 2020 & n/s & 6000 & 180 & \\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\n\\caption{Overview - Sorted by year \\label{tab:ComparisonYear}}\n\\end{doclongtable}\n\\toprule\n & & Year & Reported Min. Working Distance [mm] & Reported Max. Range [mm] & Field of View [degree] & Measurement Rate [Hz] & Sensing Element Dimension [$mm$ / $mm^2$ / $mm^3$] & Multiple-Obstacles & Commercially Available Core Components & Categorization & Basis reference \\\\\nMeasurement Principle & Reference & & & & & & & & & & \\\\\n\\midrule\n\\endhead\n\\midrule\n\\multicolumn{12}{r}{{Continued on next page}} \\\\\n\\midrule\n\\endfoot\n\\bottomrule\n\\endlastfoot\n\\multirow{4}{*}{R} & & 2016 & n/s & 500 (for given resolution) & 58 (E) 57 (H) & n/s & 13,2 (antenna diameter) & yes & no & BT-II, Object exploration & - \\\\\n & & 2018 & n/s & n/s & n/s & n/s & n/s & yes & yes & AT-II, Grasping & - \\\\\n & & 2019 & 300 & n/s & n/s & n/s & 300 (waveguide length) & yes & no & AT-I & - \\\\\n & & 2020 & n/s & $>$1200 & 160 & 40 & 9x13 & yes & yes & AT-I, BT-II & - \\\\\n\\cline{1-12}\n\\multirow{8}{*}{O-Tri} & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Manipulation and Grasping & - \\\\\n & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Grasping, Teleoperation & \\\\\n & & 1984 & n/s & v1: focal point 4.5 +-0.56 cm, v2: focal poin... & 13 & 1000 & 110 & no & no & AT-II, measure distance, orientation and curva... & - \\\\\n & & 2009 & 10 & 60 & n/s & 224 & 90 x 90 & yes & TCRT1000 & AT-I, AT-II, Sensor Array & - \\\\\n & & 2013 & 200 & 1500 & n/s & 250 (20 sensors) & ~20 x 40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & - \\\\\n & & 2014 & 200 & 1500 & n/s & 250 (20 sensors) & ~20x40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & \\\\\n & & 2018 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II Dynamic grasping & - \\\\\n & & 2019 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II, Preshaping and Grasping & \\\\\n\\cline{1-12}\nO-Tof, O-RLI & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x, MAX30105 & AT-II, Object Exploration & - \\\\\nO-ToF, C-SE & & 2017 & 10 & 100 & 180 for whole fipgertip (6 module) & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\n\\multirow{5}{*}{O-ToF} & & 2017 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Preshaping and Grasping & \\\\\n & & 2018 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\n & & 2018 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x & AT-II, Teleoperation & \\\\\n & & 2019 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Multi-Modal Sensor System & \\\\\n & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & \\\\\n\\cline{1-12}\nO-RLI, O-ToF & & 2018 & 0 & 150 & n/s & 1000, 100 & 3.2 x 1.9 x 1.1, 4.8 x 2.8 x 1 & no & EE-SY 1200, VL6180x & AT-II, Preshaping and Grasping & - \\\\\n\\multirow{28}{*}{O-RLI} & & 1988 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\n & & 1989 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\n & & 1992 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & OD8810 and SFH205 & AT-I, Collision Avoidence & - \\\\\n & & 1992 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidence & - \\\\\n & & 1993 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidance & \\\\\n & & 1993 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\n & & 1996 & 5 & 75 & 100 (+-50) & 300 & ~5x5x13 & no & STM & AT-II, Manipulation and Grasping & - \\\\\n & & 1997 & n/s & 40 & 60 (+-30) & 10000 & ~32 & no & no & AT-II, Manipulation and Grasping & - \\\\\n & & 1997 & n/s & 200 & 130 (+-65) & n/s & 2020-08-07 00:00:00 & no & no & AT-II, Object Tracking for Grasping & - \\\\\n & & 2007 & 0 & 50 & 360 & 48 & 1 & yes & no & AT-I, BT-I, Contour Following & - \\\\\n & & 2009 & 2 & 40 & 4 alnaog sensors per finger = 90 & n/s & 6x3.7x3.7 & n/s & Vishay TCND5000 & AT-II, Preshaping and Grasping & - \\\\\n & & 2011 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Modular Multi-Modal Skin & - \\\\\n & & 2012 & 1 & 10 & narrow & n/s & 15x15 & no & ADNS-9500 & AT-II, Slip Detection, Object Rreconstruction,... & - \\\\\n & & 2013 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & - \\\\\n & & 2013 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping & - \\\\\n & & 2015 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & - \\\\\n & & 2015 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II & \\\\\n & & 2015 & n/s & 300 & n/s & 1000 & 100x100 & not considered & no & AT-I, AT-II & - \\\\\n & & 2016 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & \\\\\n & & 2016 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\n & & 2016 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & \\\\\n & & 2017 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, BT-II & \\\\\n & & 2018 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, Object Exploration & \\\\\n & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\n & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\n & & 2019 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Whole-Body Control & \\\\\n & & 2019 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\n & & 2020 & n/s & n/s & n/s & n/s & 7.1 x 2.75 x 2.7 & no & Broadcom Limited, HSDL-9100-021 & AT-II, Teleoperation, VR, Transcutaneous Elect... & - \\\\\n\\cline{1-12}\n\\multirow{2}{*}{O-BB} & & 2000 & n/s & n/s & n/s & n/s & n/s & no & no & AT-II, BT-I, Sensing for Preshaping & - \\\\\n & & 2015 & 0 & $>$84 & n/s & n/s & 32x19x6.5 & no & - & AT-II, Preshaping and Grasping & - \\\\\n\\cline{1-12}\n\\multirow{2}{*}{C-SE, O-ToF} & & 2019 & 10 & 100 & 180 for whole & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-I, Multi-Modal Skin & - \\\\\n & & 2020 & 0 & 350 & 180 & 50 & 27x27 & no & no & AT-II, HRI & - \\\\\n\\cline{1-12}\n\\multirow{3}{*}{C-SE, I} & & 2016 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\n & & 2017 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\n & & 2020 & 0 & 300 & n/s & 100 & 100 x 100 x 2.75 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\n\\cline{1-12}\n\\multirow{3}{*}{C-SE, C-M} & & 2010 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & Preshaping, Grasping, Exploration, Teleoperati... & - \\\\\n & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Multi-Modal Sensor System & - \\\\\n & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\n\\cline{1-12}\n\\multirow{11}{*}{C-SE} & & 2006 & 0 & n/s & n/s & 200 & 15x5 & no & MC33794 & AT-I, Sensor Skin & - \\\\\n & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Hand and Object Tracking, Collision Pre... & \\\\\n & & 2014 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Preshaping, Grasping and Exploration & \\\\\n & & 2016 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-I, Contour Following & \\\\\n & & 2016 & 0 & 350 & n/s & 40 & n/s & yes & MRK-Systeme (complete solution) & AT-II, HRI & - \\\\\n & & 2018 & 0 & 100 & n/s & 200 & 115 x 85 x 1 & n/s & MPR121 & AT-II, Contour Following, Health Care & - \\\\\n & & 2019 & 0 & n/s & n/s & 100 & 30 x 30 & no & Teensy 3.2 & AT-II, Contour Following, Health Care & - \\\\\n & & 2019 & 0 & 300 & n/s & 125 & 50 x 50 to 100 x 100 & yes & FOGALE Robotics (complete solution) & AT-I, Collision Avoidance & - \\\\\n & & 2019 & 0 & 50 & 180 & n/s & 5 x 5 & no & no & AT-II, Grasping in Harsh environments, Industr... & - \\\\\n & & 2020 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Teleoperation, Tactile Feedback & \\\\\n & & 2020 & 0 & 350 & 180 & \\~108 & \\~175 x 80 & no & AD7147 & AT-I, Collision Avoidance & \\\\\n\\cline{1-12}\nC-M, ToF & & 2020 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I Collision Avoidance & \\\\\nC-M, O-ToF & & 2019 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I, Collision Avoidance & \\\\\n\\multirow{25}{*}{C-M} & & 1988 & 40 & 100 & n/s & n/s & 80x210 & not tested & no & AT-I Obstacle Detection & - \\\\\n & & 1991 & 0 & 304.8 human and alu, 127 grafite lead & n/s & n/s & 355,6x190,5 & no & n/s & AT-I, Collision Avoidence & - \\\\\n & & 1992 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & - \\\\\n & & 1993 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\n & & 1994 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\n & & 2007 & n/s & 50 & 180 & n/s & ~50 & no & no & AT-II, Sensing for Preshaping & - \\\\\n & & 2008 & n/s & n/s & n/s & 500 & n/s & n/s & no & AT-II, Material recognition & - \\\\\n & & 2008 & n/s & 190 & n/s & 30 & ~50 & n/s & no & AT-II, Proximity Servoing & - \\\\\n & & 2009 & n/s & 170 & n/s & n/s & 22x22 & no & no & AT-I, AT-II, Sensor & - \\\\\n & & 2010 & 0 & 150 & n/s & 20 (update cycle for control) & Fingertip of Barret Hand & not tested & no & AT-II, Preshaping, Grasping and Co-manipulation & - \\\\\n & & 2010 & n/s & n/s & n/s & n/s & n/s & yes & BOSCH (complete solution) & AT-I, Sensor & - \\\\\n & & 2010 & 0 & 160 & 180 & 25000 & 220x150 & no & no & Automotive Seat Occupancy & - \\\\\n & & 2011 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, bioinspired & - \\\\\n & & 2011 & 0 & 2000 & 360 & n/s & 30 x 570 & yes & AD7143 & Automotive parking & - \\\\\n & & 2013 & 0 & $>$100 & 180 & 6250 & 1500x150x0.35 & no & no & AT-I, BT-I, Reactive Collision Avoidance & - \\\\\n & & 2013 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, Bio-inspired & \\\\\n & & 2014 & 0 & 90 & n/s & n/s & 5x5 & yes & no & Sensor System & - \\\\\n & & 2015 & 0 & ~250 & 360 & n/s & whole body & no & no & AT-I, AT-II, Bio-Inspired & - \\\\\n & & 2015 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Teleoperation, Exploration & \\\\\n & & 2015 & 5 & 50 & Cross Section & n/s & 200x200 & yes & no & AT-I, BT-II Object detection, tomography & - \\\\\n & & 2015 & 0 & n/s & 180 & quasi-simultaneously & Meka H2 Finger & no & no & At-II, BT-II Active Object Categorization, Gra... & - \\\\\n & & 2016 & 0 & 50/100 & 180 & n/s & 100x150 & yes & no & AT-II, Gesture Control, Grasping \\& Object Man... & - \\\\\n & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Preshaping and Grasping & \\\\\n & & 2018 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-II, Material Recognition & \\\\\n & & 2018 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Material Recognition & \\\\\n\\cline{1-12}\nA-US,O-RLI & & 1997 & 25 (A-US), 50 (O-RLI) & 3000 (A-US), 450 (O-RLI) & n/s & 40 (A-US), 100 (O-RLI) & 63.5x44.5x20 & yes & no & AT-I, Sensor Skin & - \\\\\n\\multirow{7}{*}{A-US} & & 1990 & n/s & 200-300 & 180 & n/s & ~20x20x20 & n/s & n/s & AT-II & - \\\\\n & & 1994 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Contour Following & - \\\\\n & & 2012 & 20 & 400 & 15 & 40 & 45 x 20 x 15 & no & HC-SR04 & AT-II legged climbing robot & - \\\\\n & & 2017 & n/s & 6000 & 180 & \\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\n & & 2018 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Teleoperation & - \\\\\n & & 2019 & 0 & ~8 & n/s & n/s & ~30x14x14 & no & no & AT-II, Material Recognition & - \\\\\n & & 2020 & n/s & 6000 & 180 & \\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\n\\cline{1-12}\n\\multirow{2}{*}{A-S} & & 2012 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & - \\\\\n & & 2013 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & \\\\\n\\cline{1-12}\n\\multirow{947}{*}{-} & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\n\\end{longtable}\n\\end{landscape}\n\\tableofcontents\n\\end{document}","id":"46c2dad5-012b-443e-b703-dee40d0ff9dc","level":"section","origin_cites_number":106,"parent_id":"4528ef20-9412-4b01-95cc-07b7f164956d","prefix_titles":[["title","Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications "],["section","Summary and Conclusions"]],"subsections":[],"title":"Summary and Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4528ef20-9412-4b01-95cc-07b7f164956d\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ]\n ],\n \"subsections\": [\n \"01d49bee-621f-47ab-9ac0-165e2dba2afa\",\n \"7c679291-40be-4834-b2fe-70234a403984\",\n \"702ec4ea-e939-40cf-abd8-8e5caea8edc6\",\n \"55e0e327-ca41-4cdf-b82e-23bda770d2ce\",\n \"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe\",\n \"46c2dad5-012b-443e-b703-dee40d0ff9dc\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Intro}\\n\\\\IEEEPARstart{I}{n} the current robotics research landscape, a lot of effort is still dedicated to overcoming the challenges posed by unstructured environments. Areas, such as medicine, health care, agriculture, Industry\\\\ 4.0, and exploration endeavors in space as well as underwater are awaiting to profit from the robotics technologies currently in development. Furthermore, in terms of unstructured environments or situations, one of the main challenges is to develop robotics technologies that enable a safe and reliable interaction with the human. At the same time, the functionality and intelligence provided by the robot system must justify the investment, meaning its autonomous behavior, oftentimes in environments made for humans, must contribute real value. A hallmark of robust and efficient robot behavior is that task execution does not need to be interrupted in the presence of emergent events. A technology that is capable of addressing these challenges is proximity perception. It has been developed over the years, with first, impactful applications being shown in the late 1980s and early 1990s, which were sparked by seminal developments in robot control. Proximity perception is complementary to the main robotics perceptive modalities of vision and touch. Its use is often motivated by closing the perception gap left by occlusions and blind spots as well as by dealing with pose uncertainty of the robot with respect to objects and its environment. Therefore, one of the big challenges in this domain is to find sensor designs that can coexist with the main existing modalities of vision and touch. \\n\\\\begin{figure}[t]\\n \\\\centering\\n \\\\includegraphics[width=0.45\\\\textwidth]{./Figures/Intro_AT-I.png}\\n \\\\caption{A robot skin with proximity sensing capability can be deployed in scenarios with intense human-robot interaction and collaboration. The skin helps closing the gap between vision-based perception and tactile/force perception.}\\n\\\\label{fig:OcclusionHRI}\\n\\\\end{figure}\\nIn human-centered robotics, the typical applications of proximity perception can be broadly divided into two categories: the first one is pertaining a sensitive skin covering the links of a robot manipulator for safety and interaction functionality, which we call \\\\emph{\\\\gls{ATI}}. The second one is where a robot gripper or hand is equipped with sensors to support grasping and exploration tasks, which we call \\\\emph{\\\\gls{ATII}}. In Fig.~\\\\ref{fig:OcclusionHRI}, a typical scenario of human-robot interaction and collaboration is illustrated (\\\\gls{ATI}). As the human approaches the robot, the view of the camera monitoring the robot and its workspace will become increasingly occluded. A tactile skin covering the robot is not adequate to handle this perception gap in general. This is because detecting the human or the environment only when contact is established, implies operating the robot at very low velocities, thus undermining the purpose of installing such a system in the first place. To address scenarios like these, a sensitive skin with proximity perception capabilities has been proposed by several authors. In Fig.~\\\\ref{fig:Intro_Preshaping}, a typical scenario for grasping supported by proximity perception shown (\\\\gls{ATII}). Since the robotic hand can detect an object's surface before touch is established, a \\\\emph{pre-touch} closed-loop control can be implemented to adjust the hand posture during this phase. This is called \\\\emph{reactive preshaping} and can also have a diversity of use-cases. In Fig.~\\\\ref{fig:Intro_Preshaping}, the three typical phases of this procedure are shown. Proximity perception also has the potential to play an important role in robotic solutions that are compliant with norms and standards, such as ISO/TS 15066 for the operation of collaborative robots.\\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=0.45\\\\textwidth]{Figures/Intro_AT-II.jpg}\\n \\\\caption{Reactive preshaping to an object based on proximity perception has three characteristic phases: (1) object detection by vision and approaching of the hand to the object, (2) detection of the object by the hand's proximity sensors, start of closed-loop control and occlusion of the object in the camera view, and (3) finalized preshaping control, where the fingers and palm of the hand are aligned with the object.}\\n\\\\label{fig:Intro_Preshaping}\\n\\\\end{figure}\\nIn this paper, we want to provide an up-to-date perspective on the field of proximity perception in human-centered robotics as well as an introduction to the principles and technologies developed. Proximity perception in areas such as autonomous vehicles or \\\\gls{UAVs} usually aim at autonomous driving or flying and thus address larger distances and speeds and avoidance of contact and interaction with objects and humans. Recent surveys that include discussions on proximity perception in these domains are~ (millimeter wave Radar), (sensor fusion), both for autonomous driving, and~ for indoor localization of \\\\gls{UAVs}. In human-centered robotics, proximity perception is related to short distances between humans and robots and aiming for improving human-robot interaction as well as safety. Nonetheless, many ideas presented in this paper can be valid in the automotive domain and for \\\\gls{UAVs} as well, especially those about the sensing principles and their applicability. We hope to give the readers a starting point to understand the principles and applications of proximity perception that have been developed in human-centered robotics. Furthermore, we want to provide a perspective on what, in our opinion, are the important trends that will shape the developments within the next years.\\nThe main contributions of this paper are as follows:\\n\\\\begin{itemize}\\n \\\\item We provide an introduction to the concept of proximity perception and an overview of the possible use-cases. We provide a categorization according to the application types and the complexity of the implemented behavior (Sec.~\\\\ref{sec:ProximitySensor}). We use this categorization throughout the paper to organize the different works found in the literature.\\n \\\\item We give an introduction to the working principles of the most important proximity sensor designs (capacitive, optical, radar, etc.) and give a review of the related work in the context of robotics (Sec.~\\\\ref{sec:PhysicalWorkingPrinciples}).\\n \\\\item We cover what use-cases for proximity perception have been studied in robotics research and industrial robotics. Here, we give a detailed account of the two most important basic applications, \\\\gls{ATI} and \\\\gls{ATII} (Sec.~\\\\ref{sec:Applications}, see also Figs.~\\\\ref{fig:OcclusionHRI},~\\\\ref{fig:Intro_Preshaping}, and~\\\\ref{fig:ApplicationsAndBehaviors}). We start by giving a historical account and cover the basic forms of behavior possible to the more advanced, cognitive approaches.\\n \\\\item We provide a systematic comparison of the technologies from the field summarized in Table~\\\\ref{tab:ComparisonYear}\\n \\\\item Finally, we project the current developments into the future and finish the paper with concluding remarks (Sec.~\\\\ref{sec:FuturePerspectives} and \\\\ref{sec:SummaryAndConclusions}).\\n\\\\end{itemize}\",\n \"id\": \"01d49bee-621f-47ab-9ac0-165e2dba2afa\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"4528ef20-9412-4b01-95cc-07b7f164956d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:ProximitySensor}\",\n \"id\": \"7c679291-40be-4834-b2fe-70234a403984\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4528ef20-9412-4b01-95cc-07b7f164956d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Proximity Sensor: Characterization, Applications and Safety Considerations\"\n ]\n ],\n \"subsections\": [\n \"a3bf09b4-0e28-4364-af23-7c860d5e9658\",\n \"4c0a7015-4171-43a9-86c9-7a4825899568\",\n \"fdd92cab-0437-4f63-850b-5ab4cab257fd\"\n ],\n \"title\": \"Proximity Sensor: Characterization, Applications and Safety Considerations\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Providing a concise characterization of proximity sensors is challenging. One thing common to all proximity sensors is that they detect objects without physical contact. However, this alone does not distinguish them from cameras, which is problematic, as both modalities are considered to be complementary. To address this, we propose a series of attributes that generally characterize proximity sensor designs. At the same time, not all of the attributes need to be present at once in a particular case. Thus, proximity sensors provide non-contact detection of objects \\\\emph{and} more often than not\\n\\\\begin{itemize}\\n \\\\item use active measurement principles, i.\\\\,e. they probe the nearby environment to detect an object's presence, \\n item provide limited sensing range and even small detection ranges can be considered to be useful, \\n \\\\item are \\\\emph{skinlike}, i.\\\\,e. they can be deployed on surfaces such as robot arm segments as well as fingers where they can form a network of sensing elements,\\n \\\\item are suitable for being highly integrated into the sensory-motor functionality of the robot, enabling reflex-like behaviors due to low latency measurements,\\n \\\\item are used to handle occlusions in vision systems, i.\\\\,e.\\\\ are complementary to vision,\\n \\\\item are used to supervise approaching objects which are bound to enter in contact with the robot, i.\\\\,e.\\\\ are complementary to tactile sensing.\\n\\\\end{itemize}\\nIn Fig.~\\\\ref{fig:SensingRange}, we propose a definition for the sensing range of a proximity sensor. Any detection distance below $\\\\SI{50}{\\\\centi\\\\meter}$ can be considered to be within the proximity range. This limit is not strict, but in \\\\gls{HRI} and \\\\gls{HRC}, this is an approximate distance at which visual occlusions begin to become problematic. As discussed later in Sec.~\\\\ref{subsec:SafetyConsiderations}, this is a similar range in which monitoring of separation distance is relevant for compliance with safety standards. At larger distances, i.\\\\,e.\\\\ mid-range and long-range perception, other technologies (LIDAR, long-range stereo vision, etc.) can provide better performance in workspace surveillance or for providing \\\\gls{HRI} functionality. This is especially true here because the requirements on reactivity can be relaxed at larger distances. Furthermore, it is interesting to consider the contributions by anthropologist Edward T.\\\\ Hall, who describes the \\\\emph{intimate space} of humans as part of his studies on \\\\emph{proxemics}~. The intimate space starts at a distance of typically $\\\\SI{45}{\\\\centi\\\\meter}$, which is also close to the range proposed above. Therefore, proximity sensing is easy to understand from the perspective of humans, as they can intuitively relate this perception to the ``intimate space\\\" of the robot by analogy. \\nFinally, a distinction can also be made for a range below $\\\\SI{10}{\\\\centi\\\\meter}$ that we call \\\\emph{pre-touch}-range. This is the type of sensing that precedes contact interactions, for instance during grasping. Here it is especially important to have uninterrupted sensing until contact. Some sensor designs might not feature a long detection range, but the sensing capabilities provided are still useful for closed-loop control of finger and hand posture, which is executed until touch is established. A more in-depth discussion of the available proximity sensing technologies is provided in Sec.~\\\\ref{sec:PhysicalWorkingPrinciples}. In Fig.~\\\\ref{fig:LargeScaleModularSkin}, an example of a modern humanoid robot covered in a multi-modal skin is shown, displaying many of the characteristics discussed in this section. \\n\\\\begin{figure}\\n \\\\centering\\n \\\\footnotesize\\n \\\\fontsize{8}{10}\\\\selectfont\\n \\\\def\\\\svgwidth{.4\\\\textwidth}\\n \\\\input{Figures/Generated/Characterization_SensingRange_svg-tex.pdf_tex}\\n \\\\caption{Definition of the proximity sensing range.}\\n \\\\label{fig:SensingRange}\\n\\\\end{figure}\",\n \"id\": \"a3bf09b4-0e28-4364-af23-7c860d5e9658\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"7c679291-40be-4834-b2fe-70234a403984\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Proximity Sensor: Characterization, Applications and Safety Considerations\"\n ],\n [\n \"subsection\",\n \"Characterization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Characterization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:ApplicationsAndBehaviorTypes}\\nTo talk about proximity sensors in human-centered robotics as a whole, it is useful to consider first a categorization of the possible applications and desired behaviors. In the introduction, we already mentioned that a broad classification of applications into two categories is possible: the ones relating to safety and \\\\gls{HRI}/\\\\gls{HRC} and the ones relating to preshaping and grasping (Figs.~\\\\ref{fig:OcclusionHRI} and~\\\\ref{fig:Intro_Preshaping}). Beyond this, automated behaviors based on proximity sensors can be organized according to their conceptual complexity and how instantaneous their effect is on the movement of the robot. One example for a low-complexity behavior is a \\\\emph{safety stop}, i.\\\\,e.\\\\ enabling the brakes of the robot based on a sensor signal surpassing a threshold value. This behavior is closely tied to the update-rate of the sensors and the low-level robot controller. In that sense, it can be called \\\\emph{reactive} or \\\\emph{reflex-like}. Modern collaborative robots, e.\\\\,g.\\\\ the Franka Emika Panda~ or the KUKA LBR iiwa~ have control loop cycles of $t_{cl}=1~ms$. Thus, the closer the response time of the proximity sensor is to $t_r<t_{cl}$, the better. An example of high-complexity behavior is object exploration. It involves managing an object model as well as a planner to complete this model with purposeful exploration steps, resulting in a robot behavior that is executed in several phases and over a longer time compared to the basic control loop cycle times. This behavior is also characterized by being executed at different layers, reaching, as mentioned, up to the planning and cognitive components in the robot's architecture. Fig.~\\\\ref{fig:ApplicationsAndBehaviors} illustrates the categorizing of applications and behaviors we propose as well as providing some examples (not an exhaustive list). As a result, we have a broad classification of applications into two types, \\\\gls{ATI} (left) and \\\\gls{ATII} (right), and behaviors into two types, \\\\gls{BTI} (bottom) and \\\\gls{BTII} (top). In general, \\\\gls{BTI} will appear as subsystems of \\\\gls{BTII}.\\n\\\\begin{figure}\\n \\\\centering\\n \\\\footnotesize\\n \\\\includegraphics[width=0.97\\\\linewidth]{Figures/ApplicationsAndBehaviorsV5.pdf}\\n \\\\caption{Categorization of some use-cases for proximity sensors in robotics according to the possible application types (\\\\gls{ATI} and \\\\gls{ATII}) and behavior types (\\\\gls{BTI} and \\\\gls{BTII})}.\\n \\\\label{fig:ApplicationsAndBehaviors}\\n\\\\end{figure}\\n\\\\begin{figure}\\n \\\\centering\\n \\\\fontsize{8}{10}\\\\selectfont\\n \\\\def\\\\svgwidth{0.48\\\\textwidth}\\n \\\\input{Figures/Generated/Humanoid_Cheng+Modules_svg-tex.pdf_tex}\\n \\\\caption{Left: Researchers at the Technical University of Munich, Chair for Cognitive Systems, have developed a multi-modal, modular skin, which they have deployed on their robot H-1. Right: A single cell of the multi-modal sensor. (Copyright left picture. A. Eckert / TU M\\\\\\\"unchen. Right picture \\\\copyright 2019 IEEE) . Both reprinted with kind permission of the authors.)}\\n \\\\label{fig:LargeScaleModularSkin}\\n\\\\end{figure}\",\n \"id\": \"4c0a7015-4171-43a9-86c9-7a4825899568\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"7c679291-40be-4834-b2fe-70234a403984\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Proximity Sensor: Characterization, Applications and Safety Considerations\"\n ],\n [\n \"subsection\",\n \"Application and Behavior Types\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Application and Behavior Types\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:SafetyConsiderations}\\nFrom the safety perspective, a proximity sensor deployed on a collaborative robot in an industrial environment has to fulfill the requirements of the ISO/TS 15066 complying with the ISO 10218 for robots and robotic devices, as well as the performance level defined in the safety of machinery ISO 13849. Proximity sensing hardware is very well suited to operate a collaborative robot in the \\\\emph{speed and separation monitoring mode} as defined in the ISO/TS 15066 to monitor the protective separation distance $S_p$ between a human and a robot's surface. According to the standard, the following equation has to be fulfilled during the operation mode:\\n\\\\begin{equation}\\n \\\\label{eq:separationDistance}\\n \\\\small\\n \\\\begin{aligned}\\n S_{p}(t=t_0) = v_{h} (T_{r}+T_{s}) + v_{r} T_{r} + S_{s} + C_{i} + Z_{d} +Z_{r},\\\\\\\\\\n \\\\end{aligned}\\n\\\\end{equation}\\nwhere $v_h$ is the human speed (if not monitored, $v_h=\\n\\\\SI{1.6}{\\\\meter\\\\per\\\\second}$)\\n$T_r$ is the reaction time of the robot, $T_s$ is the robot stopping time, $v_r$ is the robot speed, $S_s$ is the robot stopping distance, $C_i$ is the intrusion distance, $Z_d$ is the position uncertainty of the human and $Z_r$ is the position uncertainty of the robot.\\nTo provide an illustrative example for a state of the art collaborative robot, we look at the UR10e series. It has a control loop cycle of $\\\\SI{500}{\\\\hertz}$ and the safety parameters can be configured $v_r =\\\\SI{5}{\\\\meter\\\\per\\\\second}$ (max end-effector speed)\\n$T_r = \\\\SI{4}{\\\\milli\\\\second}$ (two control loop cycles), $T_s = \\\\SI{100}{\\\\milli\\\\second} $, $S_s = \\\\SI{50}{\\\\milli\\\\meter}$, $Z_r = \\\\SI{0.05}{\\\\milli\\\\meter}$. This configuration results in separation distance of $S_p=\\\\SI{0.236}{\\\\meter}$ excluding the uncertainty of the position of the human and the intrusion distance, as this depends on the sensor parameters monitoring the area.\",\n \"id\": \"fdd92cab-0437-4f63-850b-5ab4cab257fd\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"7c679291-40be-4834-b2fe-70234a403984\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Proximity Sensor: Characterization, Applications and Safety Considerations\"\n ],\n [\n \"subsection\",\n \"Safety Considerations and Norm Compliance\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Safety Considerations and Norm Compliance\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:PhysicalWorkingPrinciples}\\nIn this section, we will give an introduction to the main physical principles available to implement proximity sensing. The idea is to be able to ease the process of reviewing articles by starting with an explanation of the basics. This will also help us in the systematization in Table~\\\\ref{tab:ComparisonYear}. There, we use the abbreviations introduced in this section. Here, we will already do a review of some representative works in the field focusing on how the proximity sensing technology is implemented. This section is closely linked to Sec.~\\\\ref{sec:Applications}, where the details of the implemented applications are discussed. We try to cross-reference the most relevant relationships. However, favoring readability, cross-referencing is not exhaustive.\\n\\\\input{Content/03.01-CapacitiveSensing.tex}\\n\\\\input{Content/03.02-OpticalSensing.tex}\\n\\\\input{Content/03.03-Radar.tex}\\n\\\\input{Content/03.04-OtherSensingPrinciples.tex}\\n\\\\input{Content/03.05-Multi-Modal}\",\n \"id\": \"702ec4ea-e939-40cf-abd8-8e5caea8edc6\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4528ef20-9412-4b01-95cc-07b7f164956d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Measurement Principles for Proximity Sensing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Measurement Principles for Proximity Sensing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:Applications}\\nIn this section, we will review the contributions from the field focusing on the applications and methods presented. We will follow the organization presented in Sec.~\\\\ref{subsec:ApplicationsAndBehaviorTypes} and Fig.~\\\\ref{fig:ApplicationsAndBehaviors}, i.\\\\,e.\\\\ focusing separately on \\\\gls{ATI} \\nand \\\\gls{ATII} and going from low-complexity behaviors (\\\\gls{BTI}) to high-complexity behaviors (\\\\gls{BTII}).\\n\\\\input{Content/04.01-CollisionAvoidance.tex}\\n\\\\input{Content/04.02-Preshaping.tex}\\n\\\\input{Content/04.03-HigherLevelApplications.tex}\\n\\\\input{Content/04.04-IndustrialTechnologiesAndSolutions.tex}\",\n \"id\": \"55e0e327-ca41-4cdf-b82e-23bda770d2ce\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4528ef20-9412-4b01-95cc-07b7f164956d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Applications and Methods in the Research and Industry Domains\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Applications and Methods in the Research and Industry Domains\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:FuturePerspectives}\\nIn this section, we provide an outlook for the domain of human-centered proximity perception in robotics. We have highlighted what we think are \\\\emph{grand challenges} at the end of each the three sub-sections.\",\n \"id\": \"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4528ef20-9412-4b01-95cc-07b7f164956d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Future Perspectives: Grand Challenges\"\n ]\n ],\n \"subsections\": [\n \"017a9c7a-b440-4a26-8ec2-d03025b12bb0\",\n \"b968a456-2525-49a5-a737-0c5965cfb1de\",\n \"c6615d1d-30c1-4597-a3db-191781a41f9f\"\n ],\n \"title\": \"Future Perspectives: Grand Challenges\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 2680,\n 2679\n ],\n \"content\": \"\\\\label{subsec:HRIInIndustryAndServiceRobotics}\\nProximity perception technology is mature enough to be deployed under strict safety requirements, as discussed in Sec.~\\\\ref{subsec:IndustrialTechnologies}. However, today, collaborative automation still struggles to be an interesting value proposition, i.\\\\,e.\\\\ providing a large increase in efficiency that will justify the cost of investing in this technology. A value proposition that is more likely to be attractive in the near future is that of \\\\emph{fenceless automation}. Here, humans and robots coexist in the same space but do not necessarily share a task. Value is added for instance by the fact that real-estate on the shop-floor is saved or because robotic automation is now possible in spaces that were previously considered to be too small. Nonetheless, safety certification is still a major challenge that is preventing wide-spread use of proximity perception. Today, technologies such as the ones discussed in ~\\\\ref{subsec:IndustrialTechnologies} need to be certified on a solution level. Certification at the modular sensor level is not yet possible. Therefore, technologies such as radar (see Sec.~\\\\ref{subsec:Radar}) can be attractive alternatives for HRI in industrial automation. Radar is more likely to achieve a safety rating on modular level soon, also profiting from all the prior experience coming from the developments in autonomous driving.\\nAreas such as medical robotics face similar issues for commercialization. The process of certifying the solutions is long and costly. However, we think that the market for automated solutions in health care or service robotics based on proximity perception is there, especially in scenarios where the human is very close to the robot, preventing the sole use of cameras. In this survey, we discussed examples such as grasping of moving objects and handover~, dressing and washing of patient , or assistive robotics based on teleoperation . Overall, we think that medical and service robotics is a promising field for proximity perception, but more research is needed to establish use-cases and the corresponding methods before commercial exploitation appears on the horizon.Furthermore, the study of highly redundant robotic systems that use proximity sensing is scarce. Meanwhile, the exploitation of kinematic redundancy based on proximity sensor feeds is natural, as discussed in Secs.~\\\\ref{subsubsec:EarlyJacobian} and \\\\ref{subsubsec:RecentJacobian}. This includes the use of proximity perception in unusual areas, such as on the legs or feet of robots. More research, like the one done in the group of Prof.\\\\ Cheng~, is needed. In summary, we can say the following about the challenges in these domains:\\n\\\\begin{itemize}\\n \\\\item Lowering the difficulties in achieving safety ratings for proximity sensing technologies will significantly expand the market for fenceless or collaborative automation. An effective procedure for safety certification, elaborated by industrial stakeholders and certification organizations, is needed.\\n \\\\item Modularized technologies, such as radar chips, can have significant advantages in a certification procedure because solutions can be based on safety rated components.\\n \\\\item In the service, medical and active assisted living domains proximity perception allows for new ways of interaction between humans and robots yet more research is needed before commercially viable applications are established.\\n \\\\item A better understanding of how highly redundant robotic systems can profit from proximity perception is needed.\\n \\\\item The potential of proximity perception for increasing robustness of grasping of moving objects, for instance during handover tasks or in teleoperation, needs to be established further, for example with conclusive user studies.\\n\\\\end{itemize}\",\n \"id\": \"017a9c7a-b440-4a26-8ec2-d03025b12bb0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Future Perspectives: Grand Challenges\"\n ],\n [\n \"subsection\",\n \"Human-robot interaction in the industry and service domains\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Human-robot interaction in the industry and service domains\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 2681,\n 2683,\n 2682\n ],\n \"content\": \"\\\\label{subsec:CognitiveRobotics}\\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=0.95\\\\linewidth]{Figures/Illustration_pointcloudV3.pdf}\\n \\\\caption{Proximity perception can play an important role in multi-modal exploration: a) Shows a typical scene with objects as captured by an RGBD-camera. b) A perspective shift reveals considerable gaps in the visual perception, having occluded elements that could be explored contactless and contact-based exploration steps, c) and d) respectively. A multi-modal cognitive model controls the exploration and aggregates the information to the object model, which includes geometric, material property, fill-level information as well as potential grasping regions obtained from the exploration steps in c) and d). This aggregated information is illustrated in e). The yellow and red colors on the mug (in c)-d)) illustrate regions that are extracted from the proximity and touch-based perception, respectively. }\\n\\\\label{fig:futureperspectives}\\n\\\\end{figure}\\nOne of the hallmarks of cognitive robotics is active perception. It is the paradigm in which perception is enabled by purposeful actions, i.\\\\,e.\\\\ exploration). The active perception principle has been studied in many robotics research papers, as discussed in a survey by Bohg et al.~. Machine learning is essential in active perception for tasks such as object pose estimation or scene labeling, which can be based on proximity perception, e.\\\\,g.~. Currently, a trend in active perception is the combined perception of vision and touch. We think that extending this trend to include proximity sensing will be a relevant research question in the near future. Therefore, we will see haptic exploration strategies that will be able to explore occluded regions of the workspace combining touch-based and touch-less exploratory movements. We have illustrated this possible multi-modal exploration approach in Fig.~\\\\ref{fig:futureperspectives}. Some aspects of it have already been addressed in the literature discussed in this paper, e.\\\\,g.~. Another important domain in this area is the multi-modal modeling of the human for safe interaction and collaboration. In summary, regarding the major challenges we can say that:\\n\\\\begin{itemize}\\n \\\\item As we mentioned throughout the paper, there is a need for proximity sensing technologies that can easily be combined with vision and/or touch so they can be deployed together. This will make it attractive to include proximity sensing in established and novel active perception/cognitive approaches.\\n \\\\item It can be expected that the trend regarding sim-to-real learning is going to prevail for the next years. Therefore, it is a challenge to implement realistic simulation models for the different measurement principles discussed in Sec.~\\\\ref{sec:PhysicalWorkingPrinciples}. In most cases, current approaches to model them are not viable in terms of their temporal performance. Overcoming this is an important challenge.\\n\\\\end{itemize}\",\n \"id\": \"b968a456-2525-49a5-a737-0c5965cfb1de\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Future Perspectives: Grand Challenges\"\n ],\n [\n \"subsection\",\n \"Cognitive Robotics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Cognitive Robotics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsec:SoftRobotics}\\nWe think that a good portion of the topics relevant for proximity perception will eventually transfer to Soft Robotics. In this way, soft manipulators equipped with proximity sensors can perform collision avoidance in an analogous way as described in Sec.~\\\\ref{subsec:CollisionAvoidance}, i.\\\\,e.\\\\ respecting task hierarchies. Similarly, soft robots will be able to execute reactive preshaping, grasping, and exploration tasks using proximity sensors. Arguably more so than in other areas, in Soft Robotics it is of interest to study how to purposefully engage in contacts to achieve the desired task. Proximity perception can help in finding and reaching desired contact states. Overall, the support of simulation frameworks for model-based control and sensing, such as SOFA~, will also be relevant. Therefore, regarding the major challenges in this domain, we can say that:\\n\\\\begin{itemize}\\n \\\\item New control methods must be found for soft robots to integrate the information provided by the proximity sensors and adapt the actuation strategy adequately.\\n \\\\item Integration of proximity sensors in deformable structures represents an important challenge. A significant cross-talk between the global deformation and the detection of tactile and proximity events has to be expected. As stated in the previous section, appropriate models for proximity sensors are needed for sim-to-real learning or for interpreting their signals correctly under deformation using interactive simulations~. \\n \\\\item Except for capacitive sensing, the realization of proximity sensors having deformable or stretchable sensing elements is challenging. Shrinking the size of rigid components, such as ICs, within a stretchable substrate may still allow the realization of deformable circuits, as described e.\\\\,g. in .\\n\\\\end{itemize}\",\n \"id\": \"c6615d1d-30c1-4597-a3db-191781a41f9f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"d1273b1a-f5ad-41b8-be89-3d4b86e92fbe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Future Perspectives: Grand Challenges\"\n ],\n [\n \"subsection\",\n \"Soft Robotics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Soft Robotics\"\n },\n {\n \"cite_extract_rate\": 0.037735849056603,\n \"cites\": [\n 2680,\n 2679,\n 2681,\n 8577\n ],\n \"content\": \"\\\\label{sec:SummaryAndConclusions}\\nIn this paper, we have given an overview of the main aspects of proximity sensing in today's robotic landscape. Considering that the field has not had a significant formalization over the years, we provide a basic scheme for categorization of the robotic applications and technologies based on proximity sensors and we propose a set of traits that characterize proximity sensors in human-centered robotics. We give an account of the existing technologies and the main measurement principles reported by authors since the early 1970s and have organized the technologies in Table~\\\\ref{tab:ComparisonYear}, including characteristics such as sensing range, update rate, sensing element size, etc. We then proceeded to detail how the technologies have been used for implementing applications such as collision avoidance and human-robot interaction (\\\\gls{ATI}) as well as preshaping and grasping (\\\\gls{ATII}). We start with the seminal developments of the early years in these domains and cover the progress up to today (2021). The tight integration into the sensory-motor functionality has received constant attention over the years in order to realize highly reactive behavior of robotic systems (\\\\gls{BTI}). Meanwhile, as the area of robotics progresses as a whole, we report that more and more approaches begin to have cognitive aspects in them (\\\\gls{BTII}). \\\\gls{BTII} includes areas such as teleoperation, where the human is in the loop, autonomous object exploration and even bio-inspired approaches that mimic weakly electric fish. \\nFinally, Sec.~\\\\ref{sec:FuturePerspectives} is dedicated to summarizing our projections for the field regarding the grand challenges we have identified. We think that as the technology of proximity sensors is reaching the maturity to coexist with tactile and visual perception in terms of integration, costs, and norm conformity, it will be adopted for a variety of solutions, especially those involving interaction with humans. \\n\\\\section*{Acknowledgment}\\nThis work was supported by the Region Hauts-de-France, the project COMOROS (ANR-17-ERC2-0029), the European Regional Development Fund (ERDF) and the project Inventor (I-SITE ULNE, le programme d’Investissements d’Avenir, M\\\\'etropole Europ\\\\'eenne de Lille). \\nThis work has received funding from the \\\"K\\\\\\\"arntner Wirtschaftsf\\\\\\\"orderung Fonds\\\" (KWF) and the \\\"European Regional Development Fund\\\" (EFRE) within the CapSize project 26616/30969/44253. \\nThis work also was supported by the Federal Ministry of Education and Research (Bundesministerium f\\\\\\\"ur Bildung und Forschung, BMBF) within the project (Verbundprojekt-Nr.: 16SV7823K:). \\n The authors would like to thank cartoonist Adriana Filippini for the illustrations. \\n\\\\ifCLASSOPTIONcaptionsoff\\n \\\\newpage\\n\\\\fi\\n\\\\bibliographystyle{IEEEtran}\\n\\\\bibliography{ReviewProximityPerception}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/escaidanavarro}}]{Stefan Escaida Navarro}\\nreceived his diploma in computer science (Dipl.-Informatiker) from Karlsruhe Institute of Technology (KIT), Germany in 2010. He obtained is PhD degree (Dr.-Ing.) from the same institution in 2016. During his doctoral phase he researched on haptic object recognition and grasping as well as the technology and the applications for capacitive tactile proximity sensors, such as preshaping, collision avoidance and teleoperation. He is currently a postdoctoral researcher at Inria Lille - Nord Europe, France, where he is working on model-based force and shape sensing for Soft Robotics. \\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/muehlbacherkarrer}}]{Stephan M\\\\\\\"{u}hlbacher-Karrer}\\nreceived the Dipl.Ing. degree in telematics from the Graz University of Technology (TUG), Graz, Austria, in 2012, where he focused on Autonomous Robots and Telecommunications. In 2017 he received his Dr. tech. degree in information and communications engineering from the Alpen-Adria-Universit\\\\\\\"{ä}t Klagenfurt, Klagenfurt, Austria, focusing on capacitive sensing for robotic applications.\\nFrom 2012 to 2013, he was with Gigatronik Ingolstadt GmbH, Gaimersheim, Germany, working\\nfor Audi Electronics Venture GmbH, Gaimersheim, where he was involved in the pre-development for autonomous driving. Since 2017 he is a senior researcher with JOANNEUM RESEARCH ROBOTICS, Institute for Robotics and Mechatronics in Klagenfurt Austria. His current research interests include near field sensor technology, signal processing, and autonomous robots.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/alagi}}]{Hosam Alagi}\\nreceived his M.Sc. degree in electrical engineering and information technology at the Karlsruhe Institute of Technology in Germany with a specialization in System-on-Chip. Since August 2015 he is working as a research assistant at the Institute for Anthropomatics and Robotics (IAR) at Intelligent Process Automation and Robotics Lab (IPR). Focusing on human-robot interaction, he develops capacitive multi-modal sensors and evaluates methods that allow robots to perceive their near environment. Within the scope of the PhD Studies both the sensors and sensor applications are being implemented, addressing the industrial and domestic domains.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/hein}}]{Bj\\\\\\\"{o}rn Hein}\\nstudied electrical engineering with a focus on control theory and received his PhD in 2003 concerning automatic collision-free path planning at the Karlsruhe Institute of Technology (KIT). May 2010 he finished his post-doctoral lecture qualification for computer science with stress on human-robot interaction. 2012-2018 he was professor for Interaction Technologies for Robot Systems at the Institute for Anthropomatics and Robotics (IAR) - Intelligent Process Automation and Robotics Lab (IPR). Currently he holds a professorship at the University of Applied Sciences in Karlsruhe. His research focus comprises: algorithms for collision free motion planning and path optimization, methods for intuitive and automatic programming of robots, human-robot interaction, novel sensor-technologies for enhancing capabilities of robot systems.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/zangl}}]{Hubert Zangl} received the Dipl.Ing. degree in telematics, the Dr. Tech. degree in electrical engineering, and the Venia Docendi degree in sensors and instrumentation from the Graz University of Technology (TUG), Graz, Austria, in 2001, 2005, and 2009, respectively.\\nFrom 2010 to 2013, he was an Associate Professor of Sensors and Instrumentation with the Institute of Electrical Measurement and Measurement Signal Processing, TUG. Since 2013, he has been a Professor chairing Sensors and Actuators with the Institute of Smart System Technologies, Alpen-Adria-Universität Klagenfurt, Klagenfurt, Austria. \\nHis current research interests include the design and optimization of smart sensors and actuators, robustness, and reliability of sensors and actuators, sensor signal processing, autarkic wireless sensors, and energy harvesting with target applications in the field of IoT and robotics.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/koyama}}]{Keisuke Koyama} received the B.E., M.E. and Ph.D. in Eng. degrees in mechanical engineering and intelligent Systems from the University of Electro-Communications (UEC), \\n in 2013, 2015 and 2017, respectively. He was a research fellow of Japan Society for the Promotion of Science (JSPS) from 2015 to 2017. \\n He researched on high-speed proximity sensor for pre-grasping and integrated control of multi-degree-of-freedom robot hand and arm. \\n From 2017 to 2019, he was project Assistant Professor, Department of Information Physics and Computing, Graduate School of Information Science and Technology, The University of Tokyo. \\n Since 2019, he has been Assistant Professor, Department of Systems Innovation, Graduate School of Engineering Science, Osaka University. \\n And, he has been Visiting Researcher, Department of Information Physics and Computing, Graduate School of Information Science and Technology, The University of Tokyo. \\n His current research focuses on high-speed, high-precision proximity sensing for high-speed robotic manipulation and assembly. \\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/christian_duriez}}]{Christian Duriez} is Research Director at Inria Lille - Nord Europe.\\nHe received the engineering degree from the Institut Catholique d’Arts et Métiers of Lille, France and a PhD degree in robotics from University of Evry and CEA in France in 2004. He had a postdoctoral position at the CIMIT SimGroup in Boston. He arrived at INRIA in 2006 and worked on interactive simulation of deformable objects and haptic rendering with focus on surgical simulation. He is now the head of DEFROST team, created in January 2015. In 2018 he was invited researcher at Stanford University. His research topics are Soft Robot models and control, Fast Finite Element Methods, simulation of contact response, new algorithms for haptics… He participated to the creation of the open-source SOFA framework. He was also one of the founders of the start-up company InSimo.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Profiles/smith}}]{Joshua R. Smith} is the Milton and Delia Zeutschel Professor, jointly appointed in the Allen School of Computer Science and Engineering, and the Department of Electrical Engineering, at the University of Washington, Seattle. He received BA degrees in Computer Science and Philosophy from Williams College in 1991, an MA in Physics from the University of Cambridge in 1997, and SM and PhD degrees in 1995 and 1999 from the MIT Media Lab's Physics and Media Group, where he began working on Electric Field Sensing for input device applications. He began applying these techniques to robotics while he was a researcher at Intel Labs Seattle, between 2004 and 2010. In recent years his group at UW has developed a number of new approaches to proximity perception for robotics.\\n\\\\end{IEEEbiography}\\n\\\\newpage\\n\\\\markboth{APPENDIX to Proximity Perception in Human-centered Robotics - A Survey on Sensing Systems and Applications}\\n{Shell \\\\MakeLowercase{\\\\textit{et al.}}: Bare Demo of IEEEtran.cls for IEEE Journals}\\n\\\\onecolumn\\n\\\\setcounter{page}{1}\\n\\\\begin{center}\\n\\\\Huge{Appendix} \\\\\\\\ \\\\vspace{.5cm}\\n\\\\end{center}\\n\\\\printglossaries\\n\\\\clearpage\\n\\\\newpage\\n\\\\begin{landscape}\\n\\\\begin{center}\\n\\\\large{Overview - Sorted by year - Table I}\\n\\\\end{center}\\n\\\\tiny\\n\\\\begin{doclongtable}{p{0.3cm}p{1.3cm}p{0.8cm}p{0.8cm}p{1.1cm}p{1.3cm}p{1.1cm}p{1.3cm}p{2.1cm}p{1.3cm}p{2.7cm}p{3.5cm}p{1.3cm}} \\n\\\\toprule\\n{} & Measurement Principle & Reference & Year & Reported Min. Working Distance [mm] & Reported Max. Range [mm] & Field of View [degree] & Measurement Rate [Hz] & Sensing Element Dimension [$mm$ / $mm^2$ / $mm^3$] & Multiple-Obstacles & Commercially Available Core Components & Categorization & Basis reference \\\\\\\\\\n\\\\midrule\\n\\\\endhead\\n\\\\midrule\\n\\\\multicolumn{13}{r}{{Continued on next page}} \\\\\\\\\\n\\\\midrule\\n\\\\endfoot\\n\\\\bottomrule\\n\\\\endlastfoot\\n1 & O-Tri & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Manipulation and Grasping & - \\\\\\\\\\n2 & O-Tri & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Grasping, Teleoperation & \\\\\\\\\\n4 & O-Tri & & 1984 & n/s & v1: focal point 4.5 +-0.56 cm, v2: focal poin... & 13 & 1000 & 110 & no & no & AT-II, measure distance, orientation and curva... & - \\\\\\\\\\n7 & O-RLI & & 1988 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\\\\\\n6 & C-M & & 1988 & 40 & 100 & n/s & n/s & 80x210 & not tested & no & AT-I Obstacle Detection & - \\\\\\\\\\n8 & O-RLI & & 1989 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\\\\\\n9 & A-US & & 1990 & n/s & 200-300 & 180 & n/s & ~20x20x20 & n/s & n/s & AT-II & - \\\\\\\\\\n10 & C-M & & 1991 & 0 & 304.8 human and alu, 127 grafite lead & n/s & n/s & 355,6x190,5 & no & n/s & AT-I, Collision Avoidence & - \\\\\\\\\\n13 & O-RLI & & 1992 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & OD8810 and SFH205 & AT-I, Collision Avoidence & - \\\\\\\\\\n14 & O-RLI & & 1992 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidence & - \\\\\\\\\\n12 & C-M & & 1992 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & - \\\\\\\\\\n16 & O-RLI & & 1993 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidance & \\\\\\\\\\n17 & O-RLI & & 1993 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\\\\\\n15 & C-M & & 1993 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\\\\\\n19 & C-M & & 1994 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\\\\\\n18 & A-US & & 1994 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Contour Following & - \\\\\\\\\\n20 & O-RLI & & 1996 & 5 & 75 & 100 (+-50) & 300 & ~5x5x13 & no & STM & AT-II, Manipulation and Grasping & - \\\\\\\\\\n22 & O-RLI & & 1997 & n/s & 40 & 60 (+-30) & 10000 & ~32 & no & no & AT-II, Manipulation and Grasping & - \\\\\\\\\\n24 & O-RLI & & 1997 & n/s & 200 & 130 (+-65) & n/s & 2020-08-07 00:00:00 & no & no & AT-II, Object Tracking for Grasping & - \\\\\\\\\\n23 & A-US,O-RLI & & 1997 & 25 (A-US), 50 (O-RLI) & 3000 (A-US), 450 (O-RLI) & n/s & 40 (A-US), 100 (O-RLI) & 63.5x44.5x20 & yes & no & AT-I, Sensor Skin & - \\\\\\\\\\n25 & O-BB & & 2000 & n/s & n/s & n/s & n/s & n/s & no & no & AT-II, BT-I, Sensing for Preshaping & - \\\\\\\\\\n27 & C-SE & & 2006 & 0 & n/s & n/s & 200 & 15x5 & no & MC33794 & AT-I, Sensor Skin & - \\\\\\\\\\n29 & O-RLI & & 2007 & 0 & 50 & 360 & 48 & 1 & yes & no & AT-I, BT-I, Contour Following & - \\\\\\\\\\n28 & C-M & & 2007 & n/s & 50 & 180 & n/s & ~50 & no & no & AT-II, Sensing for Preshaping & - \\\\\\\\\\n31 & C-M & & 2008 & n/s & n/s & n/s & 500 & n/s & n/s & no & AT-II, Material recognition & - \\\\\\\\\\n32 & C-M & & 2008 & n/s & 190 & n/s & 30 & ~50 & n/s & no & AT-II, Proximity Servoing & - \\\\\\\\\\n33 & O-Tri & & 2009 & 10 & 60 & n/s & 224 & 90 x 90 & yes & TCRT1000 & AT-I, AT-II, Sensor Array & - \\\\\\\\\\n36 & O-RLI & & 2009 & 2 & 40 & 4 alnaog sensors per finger = 90 & n/s & 6x3.7x3.7 & n/s & Vishay TCND5000 & AT-II, Preshaping and Grasping & - \\\\\\\\\\n35 & C-M & & 2009 & n/s & 170 & n/s & n/s & 22x22 & no & no & AT-I, AT-II, Sensor & - \\\\\\\\\\n38 & C-SE, C-M & & 2010 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & Preshaping, Grasping, Exploration, Teleoperati... & - \\\\\\\\\\n37 & C-M & & 2010 & 0 & 150 & n/s & 20 (update cycle for control) & Fingertip of Barret Hand & not tested & no & AT-II, Preshaping, Grasping and Co-manipulation & - \\\\\\\\\\n40 & C-M & & 2010 & n/s & n/s & n/s & n/s & n/s & yes & BOSCH (complete solution) & AT-I, Sensor & - \\\\\\\\\\n42 & C-M & & 2010 & 0 & 160 & 180 & 25000 & 220x150 & no & no & Automotive Seat Occupancy & - \\\\\\\\\\n43 & O-RLI & & 2011 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Modular Multi-Modal Skin & - \\\\\\\\\\n44 & C-M & & 2011 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, bioinspired & - \\\\\\\\\\n46 & C-M & & 2011 & 0 & 2000 & 360 & n/s & 30 x 570 & yes & AD7143 & Automotive parking & - \\\\\\\\\\n48 & O-RLI & & 2012 & 1 & 10 & narrow & n/s & 15x15 & no & ADNS-9500 & AT-II, Slip Detection, Object Rreconstruction,... & - \\\\\\\\\\n49 & A-US & & 2012 & 20 & 400 & 15 & 40 & 45 x 20 x 15 & no & HC-SR04 & AT-II legged climbing robot & - \\\\\\\\\\n47 & A-S & & 2012 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & - \\\\\\\\\\n57 & O-Tri & & 2013 & 200 & 1500 & n/s & 250 (20 sensors) & ~20 x 40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & - \\\\\\\\\\n52 & O-RLI & & 2013 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & - \\\\\\\\\\n59 & O-RLI & & 2013 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping & - \\\\\\\\\\n54 & C-SE, C-M & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n55 & C-SE & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Hand and Object Tracking, Collision Pre... & \\\\\\\\\\n51 & C-M & & 2013 & 0 & $>$100 & 180 & 6250 & 1500x150x0.35 & no & no & AT-I, BT-I, Reactive Collision Avoidance & - \\\\\\\\\\n60 & C-M & & 2013 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, Bio-inspired & \\\\\\\\\\n61 & A-S & & 2013 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & \\\\\\\\\\n63 & O-Tri & & 2014 & 200 & 1500 & n/s & 250 (20 sensors) & ~20x40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & \\\\\\\\\\n62 & C-SE & & 2014 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Preshaping, Grasping and Exploration & \\\\\\\\\\n65 & C-M & & 2014 & 0 & 90 & n/s & n/s & 5x5 & yes & no & Sensor System & - \\\\\\\\\\n66 & O-RLI & & 2015 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n67 & O-RLI & & 2015 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II & \\\\\\\\\\n72 & O-RLI & & 2015 & n/s & 300 & n/s & 1000 & 100x100 & not considered & no & AT-I, AT-II & - \\\\\\\\\\n73 & O-BB & & 2015 & 0 & $>$84 & n/s & n/s & 32x19x6.5 & no & - & AT-II, Preshaping and Grasping & - \\\\\\\\\\n68 & C-M & & 2015 & 0 & ~250 & 360 & n/s & whole body & no & no & AT-I, AT-II, Bio-Inspired & - \\\\\\\\\\n69 & C-M & & 2015 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Teleoperation, Exploration & \\\\\\\\\\n70 & C-M & & 2015 & 5 & 50 & Cross Section & n/s & 200x200 & yes & no & AT-I, BT-II Object detection, tomography & - \\\\\\\\\\n71 & C-M & & 2015 & 0 & n/s & 180 & quasi-simultaneously & Meka H2 Finger & no & no & At-II, BT-II Active Object Categorization, Gra... & - \\\\\\\\\\n83 & R & & 2016 & n/s & 500 (for given resolution) & 58 (E) 57 (H) & n/s & 13,2 (antenna diameter) & yes & no & BT-II, Object exploration & - \\\\\\\\\\n76 & O-RLI & & 2016 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & \\\\\\\\\\n77 & O-RLI & & 2016 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\\\\\\n81 & O-RLI & & 2016 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & \\\\\\\\\\n78 & C-SE, I & & 2016 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n80 & C-SE, C-M & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n75 & C-SE & & 2016 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-I, Contour Following & \\\\\\\\\\n79 & C-SE & & 2016 & 0 & 350 & n/s & 40 & n/s & yes & MRK-Systeme (complete solution) & AT-II, HRI & - \\\\\\\\\\n74 & C-M & & 2016 & 0 & 50/100 & 180 & n/s & 100x150 & yes & no & AT-II, Gesture Control, Grasping \\\\& Object Man... & - \\\\\\\\\\n82 & C-M & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Preshaping and Grasping & \\\\\\\\\\n86 & O-ToF, C-SE & & 2017 & 10 & 100 & 180 for whole fipgertip (6 module) & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\\\\\\n85 & O-ToF & & 2017 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Preshaping and Grasping & \\\\\\\\\\n84 & O-RLI & & 2017 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, BT-II & \\\\\\\\\\n88 & C-SE, I & & 2017 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\\\\\\n87 & A-US & & 2017 & n/s & 6000 & 180 & \\\\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\\\\\\n96 & R & & 2018 & n/s & n/s & n/s & n/s & n/s & yes & yes & AT-II, Grasping & - \\\\\\\\\\n97 & O-Tri & & 2018 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II Dynamic grasping & - \\\\\\\\\\n98 & O-ToF & & 2018 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\\\\\\n100 & O-ToF & & 2018 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x & AT-II, Teleoperation & \\\\\\\\\\n91 & O-RLI, O-ToF & & 2018 & 0 & 150 & n/s & 1000, 100 & 3.2 x 1.9 x 1.1, 4.8 x 2.8 x 1 & no & EE-SY 1200, VL6180x & AT-II, Preshaping and Grasping & - \\\\\\\\\\n90 & O-RLI & & 2018 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, Object Exploration & \\\\\\\\\\n93 & O-RLI & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\\\\\\n94 & O-RLI & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\\\\\\n95 & C-SE & & 2018 & 0 & 100 & n/s & 200 & 115 x 85 x 1 & n/s & MPR121 & AT-II, Contour Following, Health Care & - \\\\\\\\\\n89 & C-M & & 2018 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-II, Material Recognition & \\\\\\\\\\n92 & C-M & & 2018 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Material Recognition & \\\\\\\\\\n99 & A-US & & 2018 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Teleoperation & - \\\\\\\\\\n104 & R & & 2019 & 300 & n/s & n/s & n/s & 300 (waveguide length) & yes & no & AT-I & - \\\\\\\\\\n110 & O-Tri & & 2019 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II, Preshaping and Grasping & \\\\\\\\\\n102 & O-ToF & & 2019 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Multi-Modal Sensor System & \\\\\\\\\\n108 & O-RLI & & 2019 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Whole-Body Control & \\\\\\\\\\n109 & O-RLI & & 2019 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\\\\\\n101 & C-SE, O-ToF & & 2019 & 10 & 100 & 180 for whole & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-I, Multi-Modal Skin & - \\\\\\\\\\n103 & C-SE & & 2019 & 0 & n/s & n/s & 100 & 30 x 30 & no & Teensy 3.2 & AT-II, Contour Following, Health Care & - \\\\\\\\\\n106 & C-SE & & 2019 & 0 & 300 & n/s & 125 & 50 x 50 to 100 x 100 & yes & FOGALE Robotics (complete solution) & AT-I, Collision Avoidance & - \\\\\\\\\\n111 & C-SE & & 2019 & 0 & 50 & 180 & n/s & 5 x 5 & no & no & AT-II, Grasping in Harsh environments, Industr... & - \\\\\\\\\\n107 & C-M, O-ToF & & 2019 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I, Collision Avoidance & \\\\\\\\\\n105 & A-US & & 2019 & 0 & ~8 & n/s & n/s & ~30x14x14 & no & no & AT-II, Material Recognition & - \\\\\\\\\\n113 & R & & 2020 & n/s & $>$1200 & 160 & 40 & 9x13 & yes & yes & AT-I, BT-II & - \\\\\\\\\\n118 & O-Tof, O-RLI & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x, MAX30105 & AT-II, Object Exploration & - \\\\\\\\\\n116 & O-ToF & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & \\\\\\\\\\n119 & O-RLI & & 2020 & n/s & n/s & n/s & n/s & 7.1 x 2.75 x 2.7 & no & Broadcom Limited, HSDL-9100-021 & AT-II, Teleoperation, VR, Transcutaneous Elect... & - \\\\\\\\\\n115 & C-SE, O-ToF & & 2020 & 0 & 350 & 180 & 50 & 27x27 & no & no & AT-II, HRI & - \\\\\\\\\\n117 & C-SE, I & & 2020 & 0 & 300 & n/s & 100 & 100 x 100 x 2.75 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\\\\\\n114 & C-SE & & 2020 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Teleoperation, Tactile Feedback & \\\\\\\\\\n121 & C-SE & & 2020 & 0 & 350 & 180 & \\\\~108 & \\\\~175 x 80 & no & AD7147 & AT-I, Collision Avoidance & \\\\\\\\\\n112 & C-M, ToF & & 2020 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I Collision Avoidance & \\\\\\\\\\n120 & A-US & & 2020 & n/s & 6000 & 180 & \\\\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\\\\\\n\\\\caption{Overview - Sorted by year \\\\label{tab:ComparisonYear}}\\n\\\\end{doclongtable}\\n\\\\toprule\\n & & Year & Reported Min. Working Distance [mm] & Reported Max. Range [mm] & Field of View [degree] & Measurement Rate [Hz] & Sensing Element Dimension [$mm$ / $mm^2$ / $mm^3$] & Multiple-Obstacles & Commercially Available Core Components & Categorization & Basis reference \\\\\\\\\\nMeasurement Principle & Reference & & & & & & & & & & \\\\\\\\\\n\\\\midrule\\n\\\\endhead\\n\\\\midrule\\n\\\\multicolumn{12}{r}{{Continued on next page}} \\\\\\\\\\n\\\\midrule\\n\\\\endfoot\\n\\\\bottomrule\\n\\\\endlastfoot\\n\\\\multirow{4}{*}{R} & & 2016 & n/s & 500 (for given resolution) & 58 (E) 57 (H) & n/s & 13,2 (antenna diameter) & yes & no & BT-II, Object exploration & - \\\\\\\\\\n & & 2018 & n/s & n/s & n/s & n/s & n/s & yes & yes & AT-II, Grasping & - \\\\\\\\\\n & & 2019 & 300 & n/s & n/s & n/s & 300 (waveguide length) & yes & no & AT-I & - \\\\\\\\\\n & & 2020 & n/s & $>$1200 & 160 & 40 & 9x13 & yes & yes & AT-I, BT-II & - \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{8}{*}{O-Tri} & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Manipulation and Grasping & - \\\\\\\\\\n & & 1973 & 2 & 20 & n/s & n/s & 3x5x5 & not considered & no & AT-II, Grasping, Teleoperation & \\\\\\\\\\n & & 1984 & n/s & v1: focal point 4.5 +-0.56 cm, v2: focal poin... & 13 & 1000 & 110 & no & no & AT-II, measure distance, orientation and curva... & - \\\\\\\\\\n & & 2009 & 10 & 60 & n/s & 224 & 90 x 90 & yes & TCRT1000 & AT-I, AT-II, Sensor Array & - \\\\\\\\\\n & & 2013 & 200 & 1500 & n/s & 250 (20 sensors) & ~20 x 40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & - \\\\\\\\\\n & & 2014 & 200 & 1500 & n/s & 250 (20 sensors) & ~20x40 & not tested & Sharp GP2Y0A02YK & AT-I, Collision Avoidance & \\\\\\\\\\n & & 2018 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II Dynamic grasping & - \\\\\\\\\\n & & 2019 & 2.85 & 20 & 90 & 1000 & 18 x 28.5 x 38.5 & not tested & VSMY1850, TEMD7500X01 & AT-II, Preshaping and Grasping & \\\\\\\\\\n\\\\cline{1-12}\\nO-Tof, O-RLI & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x, MAX30105 & AT-II, Object Exploration & - \\\\\\\\\\nO-ToF, C-SE & & 2017 & 10 & 100 & 180 for whole fipgertip (6 module) & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\\\\\\n\\\\multirow{5}{*}{O-ToF} & & 2017 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Preshaping and Grasping & \\\\\\\\\\n & & 2018 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & - \\\\\\\\\\n & & 2018 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x & AT-II, Teleoperation & \\\\\\\\\\n & & 2019 & 0 & 70 & 42 & 10 & 4.8 x 2.8 x 1 & not tested & VL6180x ToF sensor & AT-II, Multi-Modal Sensor System & \\\\\\\\\\n & & 2020 & 5 & 200 & - & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-II, Object Exploration & \\\\\\\\\\n\\\\cline{1-12}\\nO-RLI, O-ToF & & 2018 & 0 & 150 & n/s & 1000, 100 & 3.2 x 1.9 x 1.1, 4.8 x 2.8 x 1 & no & EE-SY 1200, VL6180x & AT-II, Preshaping and Grasping & - \\\\\\\\\\n\\\\multirow{28}{*}{O-RLI} & & 1988 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\\\\\\n & & 1989 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\\\\\\n & & 1992 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & OD8810 and SFH205 & AT-I, Collision Avoidence & - \\\\\\\\\\n & & 1992 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidence & - \\\\\\\\\\n & & 1993 & n/s & 200 & n/s & 30 whole skin & n/s & n/s & no & AT-I, Collision Avoidance & \\\\\\\\\\n & & 1993 & n/s & 250 & 60 & 16 whole skin & min. one sensor pair & yes & no & AT-I, Collision Avoidence & \\\\\\\\\\n & & 1996 & 5 & 75 & 100 (+-50) & 300 & ~5x5x13 & no & STM & AT-II, Manipulation and Grasping & - \\\\\\\\\\n & & 1997 & n/s & 40 & 60 (+-30) & 10000 & ~32 & no & no & AT-II, Manipulation and Grasping & - \\\\\\\\\\n & & 1997 & n/s & 200 & 130 (+-65) & n/s & 2020-08-07 00:00:00 & no & no & AT-II, Object Tracking for Grasping & - \\\\\\\\\\n & & 2007 & 0 & 50 & 360 & 48 & 1 & yes & no & AT-I, BT-I, Contour Following & - \\\\\\\\\\n & & 2009 & 2 & 40 & 4 alnaog sensors per finger = 90 & n/s & 6x3.7x3.7 & n/s & Vishay TCND5000 & AT-II, Preshaping and Grasping & - \\\\\\\\\\n & & 2011 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Modular Multi-Modal Skin & - \\\\\\\\\\n & & 2012 & 1 & 10 & narrow & n/s & 15x15 & no & ADNS-9500 & AT-II, Slip Detection, Object Rreconstruction,... & - \\\\\\\\\\n & & 2013 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & - \\\\\\\\\\n & & 2013 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping & - \\\\\\\\\\n & & 2015 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n & & 2015 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II & \\\\\\\\\\n & & 2015 & n/s & 300 & n/s & 1000 & 100x100 & not considered & no & AT-I, AT-II & - \\\\\\\\\\n & & 2016 & 0 & 20 & n/s & ~1000 & ~3 x 1 & no & KEYENCE fiber optical converter & AT-II, Multi-Modal Sensor System & \\\\\\\\\\n & & 2016 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\\\\\\n & & 2016 & 10 & 400 & 10 & ~50 & simulated & yes & TCND5000,GP2D120 & AT-I, AT-II, Teleoperation, Shared Autonomy, S... & \\\\\\\\\\n & & 2017 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, BT-II & \\\\\\\\\\n & & 2018 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-II, Object Exploration & \\\\\\\\\\n & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\\\\\\n & & 2018 & 5 & 200 possible, 60-70 in their work due to PDMS & 60 & 20 (whole skin) & 4x4 & not tested & Vishay VCNL4010 & AT-I, AT-II, Preshaping, Grasping and Gesture ... & - \\\\\\\\\\n & & 2019 & n/s & 3 & n/s (180) & 1000 & 3x4 & n/s & GP2S60 & AT-I, AT-II, Whole-Body Control & \\\\\\\\\\n & & 2019 & 0 & 50 & 90 & 1000 & 22x24x40 & n/s & EE-SY1200 & AT-II, Preshaping and Grasping & \\\\\\\\\\n & & 2020 & n/s & n/s & n/s & n/s & 7.1 x 2.75 x 2.7 & no & Broadcom Limited, HSDL-9100-021 & AT-II, Teleoperation, VR, Transcutaneous Elect... & - \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{2}{*}{O-BB} & & 2000 & n/s & n/s & n/s & n/s & n/s & no & no & AT-II, BT-I, Sensing for Preshaping & - \\\\\\\\\\n & & 2015 & 0 & $>$84 & n/s & n/s & 32x19x6.5 & no & - & AT-II, Preshaping and Grasping & - \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{2}{*}{C-SE, O-ToF} & & 2019 & 10 & 100 & 180 for whole & 30 & 4.8 x 2.8 x 1 & n/s & VL6180x & AT-I, Multi-Modal Skin & - \\\\\\\\\\n & & 2020 & 0 & 350 & 180 & 50 & 27x27 & no & no & AT-II, HRI & - \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{3}{*}{C-SE, I} & & 2016 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n & & 2017 & n/s & 150 & n/s & 5 & 30x30 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\\\\\\n & & 2020 & 0 & 300 & n/s & 100 & 100 x 100 x 2.75 & n/s & no & AT-I, AT-II, Multi-Modal Sensor System & \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{3}{*}{C-SE, C-M} & & 2010 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & Preshaping, Grasping, Exploration, Teleoperati... & - \\\\\\\\\\n & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-I, AT-II, Multi-Modal Sensor System & - \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{11}{*}{C-SE} & & 2006 & 0 & n/s & n/s & 200 & 15x5 & no & MC33794 & AT-I, Sensor Skin & - \\\\\\\\\\n & & 2013 & n/s & 100 & n/s & 30-100 & 40 x 40 & yes & no & AT-II, Hand and Object Tracking, Collision Pre... & \\\\\\\\\\n & & 2014 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Preshaping, Grasping and Exploration & \\\\\\\\\\n & & 2016 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-I, Contour Following & \\\\\\\\\\n & & 2016 & 0 & 350 & n/s & 40 & n/s & yes & MRK-Systeme (complete solution) & AT-II, HRI & - \\\\\\\\\\n & & 2018 & 0 & 100 & n/s & 200 & 115 x 85 x 1 & n/s & MPR121 & AT-II, Contour Following, Health Care & - \\\\\\\\\\n & & 2019 & 0 & n/s & n/s & 100 & 30 x 30 & no & Teensy 3.2 & AT-II, Contour Following, Health Care & - \\\\\\\\\\n & & 2019 & 0 & 300 & n/s & 125 & 50 x 50 to 100 x 100 & yes & FOGALE Robotics (complete solution) & AT-I, Collision Avoidance & - \\\\\\\\\\n & & 2019 & 0 & 50 & 180 & n/s & 5 x 5 & no & no & AT-II, Grasping in Harsh environments, Industr... & - \\\\\\\\\\n & & 2020 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Teleoperation, Tactile Feedback & \\\\\\\\\\n & & 2020 & 0 & 350 & 180 & \\\\~108 & \\\\~175 x 80 & no & AD7147 & AT-I, Collision Avoidance & \\\\\\\\\\n\\\\cline{1-12}\\nC-M, ToF & & 2020 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I Collision Avoidance & \\\\\\\\\\nC-M, O-ToF & & 2019 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-I, Collision Avoidance & \\\\\\\\\\n\\\\multirow{25}{*}{C-M} & & 1988 & 40 & 100 & n/s & n/s & 80x210 & not tested & no & AT-I Obstacle Detection & - \\\\\\\\\\n & & 1991 & 0 & 304.8 human and alu, 127 grafite lead & n/s & n/s & 355,6x190,5 & no & n/s & AT-I, Collision Avoidence & - \\\\\\\\\\n & & 1992 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & - \\\\\\\\\\n & & 1993 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\\\\\\n & & 1994 & n/s & 400 & n/s & 100 & n/s & not tested & n/s & AT-I, Collision Avoidance & \\\\\\\\\\n & & 2007 & n/s & 50 & 180 & n/s & ~50 & no & no & AT-II, Sensing for Preshaping & - \\\\\\\\\\n & & 2008 & n/s & n/s & n/s & 500 & n/s & n/s & no & AT-II, Material recognition & - \\\\\\\\\\n & & 2008 & n/s & 190 & n/s & 30 & ~50 & n/s & no & AT-II, Proximity Servoing & - \\\\\\\\\\n & & 2009 & n/s & 170 & n/s & n/s & 22x22 & no & no & AT-I, AT-II, Sensor & - \\\\\\\\\\n & & 2010 & 0 & 150 & n/s & 20 (update cycle for control) & Fingertip of Barret Hand & not tested & no & AT-II, Preshaping, Grasping and Co-manipulation & - \\\\\\\\\\n & & 2010 & n/s & n/s & n/s & n/s & n/s & yes & BOSCH (complete solution) & AT-I, Sensor & - \\\\\\\\\\n & & 2010 & 0 & 160 & 180 & 25000 & 220x150 & no & no & Automotive Seat Occupancy & - \\\\\\\\\\n & & 2011 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, bioinspired & - \\\\\\\\\\n & & 2011 & 0 & 2000 & 360 & n/s & 30 x 570 & yes & AD7143 & Automotive parking & - \\\\\\\\\\n & & 2013 & 0 & $>$100 & 180 & 6250 & 1500x150x0.35 & no & no & AT-I, BT-I, Reactive Collision Avoidance & - \\\\\\\\\\n & & 2013 & 0 & ~25 & 360 & ~1000 & whole body & n/s & no & AT-I, AT-II, Bio-inspired & \\\\\\\\\\n & & 2014 & 0 & 90 & n/s & n/s & 5x5 & yes & no & Sensor System & - \\\\\\\\\\n & & 2015 & 0 & ~250 & 360 & n/s & whole body & no & no & AT-I, AT-II, Bio-Inspired & - \\\\\\\\\\n & & 2015 & n/s & 100 & n/s & 30-100 & 40x40 & yes & no & AT-II, Teleoperation, Exploration & \\\\\\\\\\n & & 2015 & 5 & 50 & Cross Section & n/s & 200x200 & yes & no & AT-I, BT-II Object detection, tomography & - \\\\\\\\\\n & & 2015 & 0 & n/s & 180 & quasi-simultaneously & Meka H2 Finger & no & no & At-II, BT-II Active Object Categorization, Gra... & - \\\\\\\\\\n & & 2016 & 0 & 50/100 & 180 & n/s & 100x150 & yes & no & AT-II, Gesture Control, Grasping \\\\& Object Man... & - \\\\\\\\\\n & & 2016 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Preshaping and Grasping & \\\\\\\\\\n & & 2018 & n/s & 500 & 180 & Capacitive: 1000, ToF: limited bei the i2C 400k & 40x40x4 & yes & no & AT-II, Material Recognition & \\\\\\\\\\n & & 2018 & 0 & 100 & n/s & 22-380 & 20x20 to 40x40 & yes, but only x/y & no & AT-II, Material Recognition & \\\\\\\\\\n\\\\cline{1-12}\\nA-US,O-RLI & & 1997 & 25 (A-US), 50 (O-RLI) & 3000 (A-US), 450 (O-RLI) & n/s & 40 (A-US), 100 (O-RLI) & 63.5x44.5x20 & yes & no & AT-I, Sensor Skin & - \\\\\\\\\\n\\\\multirow{7}{*}{A-US} & & 1990 & n/s & 200-300 & 180 & n/s & ~20x20x20 & n/s & n/s & AT-II & - \\\\\\\\\\n & & 1994 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Contour Following & - \\\\\\\\\\n & & 2012 & 20 & 400 & 15 & 40 & 45 x 20 x 15 & no & HC-SR04 & AT-II legged climbing robot & - \\\\\\\\\\n & & 2017 & n/s & 6000 & 180 & \\\\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\\\\\\n & & 2018 & n/s & n/s & n/s & n/s & n/s & n/s & n/s & AT-II, Teleoperation & - \\\\\\\\\\n & & 2019 & 0 & ~8 & n/s & n/s & ~30x14x14 & no & no & AT-II, Material Recognition & - \\\\\\\\\\n & & 2020 & n/s & 6000 & 180 & \\\\~30 & 5.7 x 4.6 & yes & no & AT-II, Navigation & - \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{2}{*}{A-S} & & 2012 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & - \\\\\\\\\\n & & 2013 & n/s & 2020-04-03 00:00:00 & n/s & 20 & 6 & no & yes & AT-II, Reactive Grasping, Object Exploration & \\\\\\\\\\n\\\\cline{1-12}\\n\\\\multirow{947}{*}{-} & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n & - & - & - & - & - & - & - & - & - & - & - \\\\\\\\\\n\\\\end{longtable}\\n\\\\end{landscape}\\n\\\\tableofcontents\\n\\\\end{document}\",\n \"id\": \"46c2dad5-012b-443e-b703-dee40d0ff9dc\",\n \"level\": \"section\",\n \"origin_cites_number\": 106,\n \"parent_id\": \"4528ef20-9412-4b01-95cc-07b7f164956d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Proximity Perception in Human-Centered Robotics: A Survey on Sensing Systems and Applications \"\n ],\n [\n \"section\",\n \"Summary and Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary and Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":85,"string":"85"},"all_cites":{"kind":"list like","value":[2680,2679,2681,2683,2682,8577],"string":"[\n 2680,\n 2679,\n 2681,\n 2683,\n 2682,\n 8577\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.294859771856887,"string":"1.29486"}}},{"rowIdx":52,"cells":{"authors":{"kind":"list like","value":["Tianyang Lin","Yuxin Wang","Xiangyang Liu","Xipeng Qiu"],"string":"[\n \"Tianyang Lin\",\n \"Yuxin Wang\",\n \"Xiangyang Liu\",\n \"Xipeng Qiu\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey of Transformers"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-06-08T17:43:08Z","string":"2021-06-08T17:43:08Z"},"category":{"kind":"string","value":"cs.LG"},"abstract":{"kind":"string","value":" Transformers have achieved great success in many artificial intelligence fields, such as natural language processing, computer vision, and audio processing. Therefore, it is natural to attract lots of interest from academic and industry researchers. Up to the present, a great variety of Transformer variants (a.k.a. X-formers) have been proposed, however, a systematic and comprehensive literature review on these Transformer variants is still missing. In this survey, we provide a comprehensive review of various X-formers. We first briefly introduce the vanilla Transformer and then propose a new taxonomy of X-formers. Next, we introduce the various X-formers from three perspectives: architectural modification, pre-training, and applications. Finally, we outline some potential directions for future research. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"456ca844-322a-4fcf-90ea-531b14db9d9e","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey of Transformers"]],"subsections":["48308b43-2cfe-4236-84fb-c7cd67559d21","07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62","aa12aeb0-6a41-4c9b-bfb0-68b4522245cd","57f59954-45c7-48ac-baf2-dc92c147a85e","7a323139-7d7b-422d-a6c6-3617457ad2f6","04795308-eff0-4fbb-962f-a8077d36f894","b7522af4-45ec-452c-b804-11b6882e6225","b3d753a5-00c0-4eb4-b313-01e490dd9d3d","ec81ede4-3b9c-44b8-96e2-94e0f2e4f760"],"title":"root"},{"cite_extract_rate":0.7272727272727271,"cites":[7360,1447,1445,7052,2401,732,1446,38],"content":"\\label{sec:intro}\nTransformer~ is a prominent deep learning model that has been widely adopted in various fields, such as natural language processing (NLP), computer vision (CV) and speech processing. Transformer was originally proposed as a sequence-to-sequence model~ for machine translation. Later works show that Transformer-based pre-trained models (PTMs)~ can achieve \\textit{state-of-the-art} performances on various tasks. As a consequence, Transformer has become the go-to architecture in NLP, especially for PTMs. In addition to language related applications, Transformer has also been adopted in CV~, audio processing~ and even other disciplines, such as chemistry~ and life sciences~.\nDue to the success, a variety of Transformer variants (a.k.a. X-formers) have been proposed over the past few years.\nThese X-formers improve the vanilla Transformer from different perspectives.\n\\begin{enumerate}\n \\item \\textit{Model Efficiency}. A key challenge of applying Transformer is its inefficiency at processing long sequences mainly due to the computation and memory complexity of the self-attention module.\nThe improvement methods include lightweight attention (e.g. sparse attention variants) and Divide-and-conquer methods (e.g., recurrent and hierarchical mechanism).\n \\item \\textit{Model Generalization}. Since the transformer is a flexible architecture and makes few assumptions on the structural bias of input data, it is hard to train on small-scale data. The improvement methods include introducing structural bias or regularization, pre-training on large-scale unlabeled data, etc.\n \\item \\textit{Model Adaptation}. This line of work aims to adapt the Transformer to specific downstream tasks and applications.\n\\end{enumerate}\nIn this survey, we aim to provide a comprehensive review of the Transformer and its variants. Although we can organize X-formers on the basis of the perspectives mentioned above, many existing X-formers may address one or several issues. For example, sparse attention variants not only reduce the computational complexity but also introduce structural prior on input data to alleviate the overfitting problem on small datasets. Therefore, it is more methodical to categorize the various existing X-formers and propose a new taxonomy mainly according to their ways to improve the vanilla Transformer: architecture modification, pre-training, and applications.\nConsidering the audience of this survey may be from different domains, we mainly focus on the general architecture variants and just briefly discuss the specific variants on pre-training and applications.\nThe rest of the survey is organized as follows. Sec.~\\ref{sec:background} introduces the architecture and the key components of Transformer. Sec.~\\ref{sec:taxonomy} clarifies the categorization of Transformer variants. Sec.~\\ref{sec:attention}$\\sim$\\ref{sec:other_module} review the module-level modifications, including attention module, position encoding, layer normalization and feed-forward layer. Sec.~\\ref{sec:beyond} reviews the architecture-level variants. Sec.~\\ref{sec:ptm} introduces some of the representative Transformer-based PTMs. Sec.~\\ref{sec:app} introduces the application of Transformer to various different fields. Sec.~\\ref{sec:discussion} discusses some aspects of Transformer that researchers might find intriguing and summarizes the paper.","id":"48308b43-2cfe-4236-84fb-c7cd67559d21","level":"section","origin_cites_number":11,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:background}","id":"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62","level":"section","origin_cites_number":0,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Background"]],"subsections":["cbbc7151-fd9c-4d11-a9dd-a0790a09698d","89fa15e6-1537-4f12-a5ac-bdd5d071bfac","3ddd6e3c-d4a1-4bce-b510-16dfc75582ed","f0fa958a-db02-433c-858f-f7024645f38d"],"title":"Background"},{"cite_extract_rate":1,"cites":[97,57,38],"content":"\\label{sec:vanilla_xformer}\nThe vanilla Transformer~ is a sequence-to-sequence model and consists of an encoder and a decoder, each of which is a stack of $L$ identical blocks. Each \\textit{encoder block} is mainly composed of a multi-head self-attention module and a position-wise feed-forward network (FFN).\nFor building a deeper model, a residual connection~ is employed around each module, followed by Layer Normalization~ module.\nCompared to the encoder blocks, decoder blocks additionally insert cross-attention modules between the multi-head self-attention modules and the position-wise FFNs. Furthermore, the self-attention modules in the decoder are adapted to prevent each position from attending to subsequent positions.\nThe overall architecture of the vanilla Transformer is shown in Fig. \\ref{fig:xformer_arch}.\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.7\\linewidth]{assets/xformer.pdf}\n \\caption{Overview of vanilla Transformer architecture}\n \\label{fig:xformer_arch}\n\\end{figure}\nIn the following subsection, we shall introduce the key modules of the vanilla Transformer.","id":"cbbc7151-fd9c-4d11-a9dd-a0790a09698d","level":"subsection","origin_cites_number":3,"parent_id":"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Vanilla Transformer"]],"subsections":["625a964c-dc0c-4228-a0bd-e5fba16df1f2","fdfb7e39-e1e9-4079-9037-b8bc17e8bb70","0c6acb39-f7c8-4d89-9969-16b6e42547a1","64d291c3-febe-419e-831a-4ec33ca28ce7"],"title":"Vanilla Transformer"},{"cite_extract_rate":0,"cites":[],"content":"Transformer adopts attention mechanism with Query-Key-Value (QKV) model. Given the packed matrix representations of queries $\\bQ\\in\\mathbb{R}^{N\\times D_k}$, keys $\\bK\\in\\mathbb{R}^{M\\times D_k}$, and values $\\bV\\in\\mathbb{R}^{M\\times D_v}$, the scaled dot-product attention used by Transformer is given by\\footnote{if not stated otherwise, we use row-major notations throughout this survey (e.g., the $i$-th row in $\\bQ$ is the query $\\bq_i$) and all the vectors are row vectors by default.}\n\\begin{equation}\\label{attention}\n \\mathrm{Attention}(\\bQ, \\bK, \\bV)=\\mathrm{softmax}\\left(\\frac{\\bQ\\bK^\\top}{\\sqrt{D_k}}\\right)\\bV=\\bA\\bV,\n\\end{equation}\nwhere $N$ and $M$ denote the lengths of queries and keys (or values); $D_k$ and $D_v$ denote the dimensions of keys (or queries) and values; $\\bA=\\mathrm{softmax}\\left(\\frac{\\bQ\\bK^\\top}{\\sqrt{D_k}}\\right)$ is often called \\textit{attention matrix}; softmax is applied in a row-wise manner. The dot-products of queries and keys are divided by $\\sqrt{D_k}$ to alleviate gradient vanishing problem of the softmax function.\nInstead of simply applying a single attention function, Transformer uses multi-head attention, where the $D_m$-dimensional original queries, keys and values are projected into $D_k$, $D_k$ and $D_v$ dimensions, respectively, with $H$ different sets of learned projections. For each of the projected queries, keys and values, and output is computed with attention according to Eq. \\eqref{attention}. The model then concatenates all the outputs and projects them back to a $D_m$-dimensional representation.\n\\begin{align}\n \\mathrm{MultiHeadAttn}(\\bQ,\\bK,\\bV)&=\\mathrm{Concat}(\\mathrm{head}_1,\\cdots,\\mathrm{head}_H)\\bW^O\\label{eq:multihead},\\\\\n \\mathrm{where}\\ \\mathrm{head}_i&=\\mathrm{Attention}(\\bQ \\bW_i^Q,\\bK \\bW_i^K, \\bV \\bW_i^V)\\label{eq:headi}.\n\\end{align}\nIn Transformer, there are three types of attention in terms of the source of queries and key-value pairs:\n\\begin{itemize}\n \\item \\textit{Self-attention}. In Transformer encoder, we set $\\bQ=\\bK=\\bV=\\bX$ in Eq. \\eqref{eq:multihead}, where $\\bX$ is the outputs of the previous layer.\n \\item \\textit{Masked Self-attention}. In the Transformer decoder, the self-attention is restricted such that queries at each position can only attend to all key-value pairs up to and including that position. To enable parallel training, this is typically done by applying a mask function to the unnormalized attention matrix $\\hat{\\bA}=\\exp(\\frac{\\bQ\\bK^\\top}{\\sqrt{D_k}})$, where the illegal positions are masked out by setting $\\hat{A}_{ij}=-\\infty \\textrm{ if } i<j$. This kind of self-attention is often referred to as autoregressive or causal attention\\footnote{This term seems to be borrowed from the \\textit{causal system}, where the output depends on past and current inputs but not future inputs.}.\n \\item \\textit{Cross-attention}. The queries are projected from the outputs of the previous (decoder) layer, whereas the keys and values are projected using the outputs of the encoder.\n\\end{itemize}","id":"625a964c-dc0c-4228-a0bd-e5fba16df1f2","level":"subsubsection","origin_cites_number":0,"parent_id":"cbbc7151-fd9c-4d11-a9dd-a0790a09698d","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Vanilla Transformer"],["subsubsection","Attention Modules"]],"subsections":[],"title":"Attention Modules"},{"cite_extract_rate":0,"cites":[],"content":"The position-wise FFN\\footnote{The parameters are shared across different positions, thus the position-wise FFN can also be understood as two convolution layers with kernel size of 1.} is a fully connected feed-forward module that operates separately and identically on each position\n\\begin{equation}\n \\mathrm{FFN}(\\bH') = \\mathrm{ReLU}(\\bH'\\bW^1 + \\bb^1)\\bW^2+\\bb^2,\n\\end{equation}\nwhere $\\bH'$ is the outputs of previous layer, and $\\bW^1\\in\\mathbb{R}^{D_m\\times D_f},\\bW^2\\in\\mathbb{R}^{D_f\\times D_m},\\bb^1\\in\\mathbb{R}^{D_f},\\bb^2\\in\\mathbb{R}^{D_m}$ are trainable parameters. Typically the intermediate dimension $D_{f}$ of the FFN is set to be larger than $D_m$.","id":"fdfb7e39-e1e9-4079-9037-b8bc17e8bb70","level":"subsubsection","origin_cites_number":0,"parent_id":"cbbc7151-fd9c-4d11-a9dd-a0790a09698d","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Vanilla Transformer"],["subsubsection","Position-wise FFN"]],"subsections":[],"title":"Position-wise FFN"},{"cite_extract_rate":1,"cites":[97,57],"content":"In order to build a deep model, Transformer employs a residual connection~ around each module, followed by Layer Normalization~. For instance, each Transformer encoder block may be written as\n\\begin{align}\n \\bH'&=\\mathrm{LayerNorm}(\\mathrm{SelfAttention}(\\bX)+\\bX)\\\\\n \\bH&=\\mathrm{LayerNorm}(\\mathrm{FFN}(\\bH')+\\bH'),\n\\end{align}\nwhere $\\mathrm{SelfAttention}(\\cdot)$ denotes self attention module and $\\mathrm{LayerNorm}(\\cdot)$ denotes the layer normalization operation.","id":"0c6acb39-f7c8-4d89-9969-16b6e42547a1","level":"subsubsection","origin_cites_number":2,"parent_id":"cbbc7151-fd9c-4d11-a9dd-a0790a09698d","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Vanilla Transformer"],["subsubsection","Residual Connection and Normalization"]],"subsections":[],"title":"Residual Connection and Normalization"},{"cite_extract_rate":0,"cites":[],"content":"Since Transformer doesn't introduce recurrence or convolution, it is ignorant of positional information (especially for the encoder). Thus additional positional representation (Detailed discussion in Sec.~\\ref{sec:pos_rep}) is needed to model the ordering of tokens.","id":"64d291c3-febe-419e-831a-4ec33ca28ce7","level":"subsubsection","origin_cites_number":0,"parent_id":"cbbc7151-fd9c-4d11-a9dd-a0790a09698d","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Vanilla Transformer"],["subsubsection","Position Encodings"]],"subsections":[],"title":"Position Encodings"},{"cite_extract_rate":0,"cites":[],"content":"Generally, the Transformer architecture can be used in three different ways:\n\\begin{itemize}\n \\item \\textit{Encoder-Decoder}. The full Transformer architecture as introduced in Sec.~\\ref{sec:vanilla_xformer} is used. This is typically used in sequence-to-sequence modeling (e.g., neural machine translation).\n \\item \\textit{Encoder only}. Only the encoder is used and the outputs of the encoder are utilized as a representation for the input sequence. This is usually used for classification or sequence labeling problems.\n \\item \\textit{Decoder only}. Only the decoder is used, where the encoder-decoder cross-attention module is also removed. This is typically used for sequence generation, such as language modeling.\n\\end{itemize}","id":"89fa15e6-1537-4f12-a5ac-bdd5d071bfac","level":"subsection","origin_cites_number":0,"parent_id":"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Model Usage"]],"subsections":[],"title":"Model Usage"},{"cite_extract_rate":1,"cites":[38],"content":"\\label{sec:model-analysis}\nTo illustrate the computation time and parameter requirements of the Transformer, we analyze the two core components of the Transformer (i.e., the self-attention module and the position-wise FFN) in Table \\ref{tab:complexity_attn_ffn}. We assume that the hidden dimension $D_m$ of the model is $D$, and that the input sequence length is $T$. The intermediate dimension of FFN is set to $4D$ and the dimension of keys and values are set to $D/H$ as in .\n\\begin{table}[htbp]\n \\caption{Complexity and parameter counts of self-attention and position-wise FFN}\n \\label{tab:complexity_attn_ffn}\n \\centering\n \\begin{tabular}{c|c|c}\n \\hline\n Module & Complexity & \\#Parameters\\\\\n \\hline\n self-attention & $\\mathcal O(T^2\\cdot D)$ & $4D^2$\\\\\n \\hline\n position-wise FFN & $\\mathcal O(T\\cdot D^2)$ & $8D^2$\\\\\n \\hline\n \\end{tabular}\n\\end{table}\nWhen the input sequences are short, the hidden dimension $D$ dominates the complexity of self-attention and position-wise FFN. The bottleneck of Transformer thus lies in FFN. However, as the input sequences grow longer, the sequence length $T$ gradually dominates the complexity of these modules, in which case self-attention becomes the bottleneck of Transformer. Furthermore, the computation of self-attention requires that a $T\\times T$ attention distribution matrix is stored, which makes the computation of Transformer infeasible for long-sequence scenarios (e.g., long text documents and pixel-level modeling of high-resolution images). One shall see that the goal of increasing the efficiency of Transformer generally leads to the long-sequence compatibility of self-attention, as well as the computation and parameter efficiency of position-wise FFN for ordinary settings.","id":"3ddd6e3c-d4a1-4bce-b510-16dfc75582ed","level":"subsection","origin_cites_number":1,"parent_id":"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Model Analysis"]],"subsections":[],"title":"Model Analysis"},{"cite_extract_rate":0,"cites":[],"content":"","id":"f0fa958a-db02-433c-858f-f7024645f38d","level":"subsection","origin_cites_number":0,"parent_id":"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Comparing Transformer to Other Network Types"]],"subsections":["037a9c67-64a3-4bbe-b197-a41922561a14","cf87fcdd-fc72-4f50-9754-c59538a65d3f"],"title":"Comparing Transformer to Other Network Types"},{"cite_extract_rate":1,"cites":[38],"content":"As a central piece of Transformer, self-attention comes with a flexible mechanism to deal with variable-length inputs. It can be understood as a fully connected layer where the weights are dynamically generated from pairwise relations from inputs. Table \\ref{tab:op_complexities} compares the complexity, sequential operations, and maximum path length\\footnote{The maximum length of the paths forward and backward signals have to traverse to get from any input position to arbitrary output position. Shorter length implies a better potential for learning long-range dependencies.} of self-attention with three commonly used layer types. We summarize the advantages of self-attention as follows:\n\\begin{enumerate}\n \\item It has the same maximum path length as fully connected layers, making it suitable for long-range dependencies modeling. Compared to fully connected layers, it is more parameter-efficient and more flexible in handling variable-length inputs.\n \\item Due to the limited receptive field of convolutional layers, one typically needs to stack a deep network to have a global receptive field. On the other hand, the constant maximum path length enables self-attention to model long-range dependencies with a constant number of layers.\n \\item The constant sequential operations and maximum path length make self-attention more parallelizable and better at long-range modeling than recurrent layers.\n\\end{enumerate}\n\\begin{table}[ht]\n \\caption{Per-layer complexity, minimum number of sequential operations and maximum path lengths for different layer types. $T$ is the sequence length, $D$ is the representation dimension and $K$ is the kernel size of convolutions~.}\n\\label{tab:op_complexities}\n\\begin{center}\n\\vspace{-1mm}\n\\begin{tabular}{lccc}\n\\toprule\nLayer Type & Complexity & Sequential & Maximum Path Length \\\\\n &per Layer & Operations & \\\\\n\\hline\n\\rule{0pt}{2.0ex}Self-Attention & $\\mathcal O(T^2 \\cdot D)$ & $\\mathcal O(1)$ & $\\mathcal O(1)$ \\\\\nFully Connected & $\\mathcal O(T^2 \\cdot D^2)$ & $\\mathcal O(1)$ & $\\mathcal O(1)$ \\\\\nConvolutional & $\\mathcal O(K \\cdot T \\cdot D^2)$ & $\\mathcal O(1)$ & $\\mathcal O(\\log_K(T))$ \\\\\nRecurrent & $\\mathcal O(T \\cdot D^2)$ & $\\mathcal O(T)$ & $\\mathcal O(T)$ \\\\\n\\bottomrule\n\\end{tabular}\n\\end{center}\n\\end{table}","id":"037a9c67-64a3-4bbe-b197-a41922561a14","level":"subsubsection","origin_cites_number":1,"parent_id":"f0fa958a-db02-433c-858f-f7024645f38d","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Comparing Transformer to Other Network Types"],["subsubsection","Analysis of Self-Attention"]],"subsections":[],"title":"Analysis of Self-Attention"},{"cite_extract_rate":1,"cites":[208,553],"content":"Transformer is often compared against convolutional and recurrent networks. Convolutional networks are known to impose the inductive biases of translation invariance and locality with shared local kernel functions. Similarly, recurrent networks carry the inductive biases of temporal invariance and locality via their Markovian structure~. On the other hand, the Transformer architecture makes few assumptions about structural information of data. This makes Transformer a universal and flexible architecture. As a side effect, the lack of structural bias makes Transformer prone to overfitting for small-scale data.\nAnother closely related network type is Graph Neural Networks (GNNs) with message passing~. Transformer can be viewed as a GNN defined over a complete directed graph (with self-loop) where each input is a node in the graph. The key difference between Transformer and GNNs is that Transformer introduces no prior knowledge over how input data are structured — the message passing process in Transformer solely depends on similarity measures over the content.","id":"cf87fcdd-fc72-4f50-9754-c59538a65d3f","level":"subsubsection","origin_cites_number":2,"parent_id":"f0fa958a-db02-433c-858f-f7024645f38d","prefix_titles":[["title","A Survey of Transformers"],["section","Background"],["subsection","Comparing Transformer to Other Network Types"],["subsubsection","In Terms of Inductive Bias"]],"subsections":[],"title":"In Terms of Inductive Bias"},{"cite_extract_rate":0.808333333333333,"cites":[1474,1493,1457,1482,1460,1491,7360,7040,1484,1494,1464,1488,1470,1452,1466,1490,793,1445,1502,7054,1467,1463,8384,1456,1480,794,1475,1465,7361,1500,1459,7053,1458,1479,7365,1496,1133,8454,768,7367,7333,1489,1451,1469,9,1476,1453,1495,1487,1462,1478,679,1483,1492,7369,7,1473,1446,7370,8456,826,7362,7364,7363,1468,38,1486,7298,1498,1497,1501,7371,8457,7368,1455,707,8455,1471,1481,1182,732,1272,7339,1477,1449,7273,1461,1454,1448,1485,7366,798,1499,1450,1472],"content":"\\label{sec:taxonomy}\nA wide variety of models have been proposed so far based on the vanilla Transformer from three perspectives: types of architecture modification, pre-training methods, and applications.\nFig. \\ref{fig:xformer_taxonomy} gives an illustrations of our categorization of Transformer variants.\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=1.0\\linewidth]{assets/taxonomy.pdf}\n \\caption{Categorization of Transformer variants.}\n \\label{fig:xformer_taxonomy}\n\\end{figure}\nFig. \\ref{taxonomy_of_xformer} illustrates our taxonomy and some representative models.\n\\tikzstyle{leaf}=[mybox,minimum height=1em,\nfill=hidden-orange!40, text width=20em, text=black,align=left,font=\\tiny,\ninner xsep=2pt,\ninner ysep=1pt,\n]\n\\begin{figure*}[tp]\n \\centering\n\\begin{forest}\n forked edges,\n for tree={\n grow=east,\n reversed=true,\n anchor=base west,\n parent anchor=east,\n child anchor=west,\n base=left,\n font=\\small,\n rectangle,\n draw=hiddendraw,\n rounded corners,align=left,\n minimum width=2.5em,\ns sep=3pt,\ninner xsep=2pt,\ninner ysep=1pt,\nver/.style={rotate=90, child anchor=north, parent anchor=south, anchor=center},\n },\n where level=1{text width=2.7em,font=\\scriptsize,}{},\n where level=2{text width=3em,font=\\tiny}{},\n where level=3{text width=3em,font=\\tiny}{},\n [X-formers, ver\n [Module\\\\Level\n [Attention\n [Sparse\n [Star-Transformer{,} Longformer{,} ETC{,} BigBird{,} Sparse Transformer\\\\\n BP-Transformer{,} Image Transformer{,} Axial Transformer\n ,leaf,text width=21.5em]\n [Routing Transformer{,} Reformer{,} SAC{,} Sparse Sinkhorn Attention\n ,leaf,text width=21.5em]\n ]\n [Linearized\n [Linear Transformer{,} Performer{,} RFA{,} Delta Net\n ,leaf,text width=18.5em]\n ]\n [Prototype\n [Clustered Attention{,} Informer\n ,leaf,text width=12em]\n ]\n [Memory\\\\Compress\n [MCA{,} Set Transformer{,} Linformer\n ,leaf,text width=12em]\n ]\n [Low-rank\n [Low-rank Attention{,} CSALR{,} Nystr{\\\"{o}}mformer~\n ,leaf,text width=15em]\n ]\n [Prior\\\\Attention\n [Local Transformer{,} Gaussian Transformer\n ,leaf,text width=15em]\n [Predictive Attention Transformer{,} Realformer{,} Lazyformer\n ,leaf,text width=18.5em]\n [CAMTL\n ,leaf,text width=4em]\n [Average Attention{,} Hard-Coded Gaussian Attention{,} Synthesizer\n ,leaf]\n ]\n [Multi-head\n [{,} {,} Talking-head Attention\\\\\n Collaborative MHA\n ,leaf,text width=20em]\n [Adaptive Attention Span{,} Multi-Scale Transformer\n ,leaf,text width=20em]\n [Dynamic Routing\n ,leaf,text width=6.5em]\n ]\n ]\n [Position\\\\ Encoding\n [Absolute\n [BERT{,} {,} FLOATER\n ,leaf,text width=11em]\n ]\n [Relative\n [{,} Music Transformer{,} T5{,} Transformer-XL\\\\ DeBERTa\n ,leaf,text width=20em]\n ]\n [Other Rep.\n [TUPE{,} Roformer\n ,leaf,text width=7em]\n ]\n [Implicit Rep.\n [Complex Embedding{,} R-Transformer {,} CPE\n ,leaf,text width=16em]\n ]\n ]\n [LayerNorm\n [Placement\n [post-LN{,} pre-LN\n ,leaf,text width=16em]\n ]\n [Substitutes\n [AdaNorm{,} scaled $\\ell_2$ normalization{,} PowerNorm\n ,leaf,text width=16em]\n ]\n [Norm-free\n [ReZero-Transformer\n ,leaf,text width=9.5em]\n ]\n ]\n [FFN\n [Activ. Func.\n [Swish{,} GELU{,} GLU\n ,leaf,text width=9.5em]\n ]\n [Enlarge\\\\Capacity\n [Product-key Memory{,} Gshard{,} Switch Transformer{,}\\\\ Expert Prototyping{,} Hash Layer\n ,leaf,text width=16em]\n ]\n [Dropping\n [All-Attention layer{,} \n ,leaf,text width=11em]\n ]\n ]\n ]\n [Arch.\\\\Level\n [Lighweight\n [Lite Transformer{,} Funnel Transformer{,} DeLighT\n ,leaf,text width=16em]\n ]\n [Connectivity\n [Realformer{,} Predictive Attention Transformer{,} Transparent Attention\\\\\n Feedback Transformer~\n ,leaf,text width=24.5em]\n ]\n [ACT\n [UT{,} Conditional Computation Transformer{,} DeeBERT{,} PABEE{,} {,} \\\\\n ,leaf,text width=24.5em]\n ]\n [Divide \\& \\\\\n Conquer\n [Recurrence\n [Transformer-XL{,} Compressive Transformer{,} Memformer\\\\ {,} ERNIE-Doc\n ,leaf,text width=20em]\n ]\n [Hierarchy\n [{,} HIBERT{,} {,} Hi-Transformer\\\\ TENER{,} TNT\n ,leaf,text width=20em]\n ]\n ]\n [Alt. Arch.\n [ET{,} Macaron Transformer{,} Sandwich Transformer{,} MAN{,} DARTSformer\n ,leaf,text width=24.5em]\n ]\n ]\n [Pre-Train\n [Encoder\n [BERT{,} RoBERTa{,} BigBird,leaf,text width=11em]\n ]\n [Decoder\n [GPT{,} GPT-2{,} GPT-3,leaf,text width=11em]\n ]\n [Enc.Dec.\n [ BART{,} T5{,} Switch Transformer\n ,leaf,text width=11em]\n ]\n ]\n [App.\n [NLP\n [BERT{,}ET{,} Transformer-XL{,}Compressive Transformer{,} TENER\n ,leaf,text width=22em]\n ]\n [CV\n [Image Transformer{,} DETR{,} ViT{,} Swin Transformer{,} ViViT\n ,leaf,text width=22em]\n ]\n [Audio\n [Speech Transformer{,} Streaming Transformer{,} Reformer-TTS{,} Music Transformer\n ,leaf,text width=25em]\n ]\n [Multimodal\n [VisualBERT{,} VLBERT{,} VideoBERT{,} M6{,} Chimera{,} DALL-E{,} CogView\n ,leaf,text width=25em]\n ]\n ]\n ]\n\\end{forest}\n\\caption{Taxonomy of Transformers}\n\\label{taxonomy_of_xformer}\n\\end{figure*}\nIn this survey, we focus on reviewing the works on architecture modifications.\nSince the attention module is the key component of Transformer, we solely describe the attention-related variants in Sec.~\\ref{sec:attention} and introduce the other module-level variants in Sec.~\\ref{sec:other_module}. Then Sec.~\\ref{sec:beyond} describes the other architecture-level variants. Finally, we briefly review the works on pre-training in Sec.~\\ref{sec:ptm} and applications in Sec.~\\ref{sec:app}.\nThere are some comprehensive surveys on the latter two categories of work, such as pre-trained models (PTMs)~ and visual Transformers.","id":"aa12aeb0-6a41-4c9b-bfb0-68b4522245cd","level":"section","origin_cites_number":120,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Taxonomy of Transformers"]],"subsections":[],"title":"Taxonomy of Transformers"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:attention}\nSelf-attention plays an important role in Transformer, but there are two challenges in practical applications.\n\\begin{enumerate}\n \\item \\textit{Complexity}. As discussion in Sec.~\\ref{sec:model-analysis}, the complexity of self-attention is $\\mathcal O(T^2\\cdot D)$. Therefore, the attention module becomes a bottleneck when dealing with long sequences.\n \\item \\textit{Structural prior}. Self-attention does no assume any structural bias over inputs. Even the order information is also needed to be learned from training data. Therefore, Transformer (w/o pre-training) is usually easy to overfit on small or moderate-size data.\n\\end{enumerate}\nThe improvements on attention mechanism can be divided into several directions:\n\\begin{enumerate}\n \\item \\textit{Sparse Attention}. This line of work introduces sparsity bias into the attention mechanism, leading to reduced complexity.\n \\item \\textit{Linearized Attention}. This line of work disentangles the attention matrix with kernel feature maps. The attention is then computed in reversed order to achieve linear complexity.\n \\item \\textit{Prototype and Memory Compression}. This class of methods reduces the number of queries or key-value memory pairs to reduce the size of the attention matrix.\n \\item \\textit{Low-rank Self-Attention}. This line of work capture the low-rank property of self-attention.\n \\item \\textit{Attention with Prior}. The line of research explores supplementing or substituting standard attention with prior attention distributions.\n \\item \\textit{Improved Multi-Head Mechanism}. The line of studies explores different alternative multi-head mechanisms.\n\\end{enumerate}\nWe will describe these attention variants at length in the rest of this section.","id":"57f59954-45c7-48ac-baf2-dc92c147a85e","level":"section","origin_cites_number":0,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"]],"subsections":["c3955c2b-4114-49e4-bbb8-30061d5f27e8","a4f89d55-c38a-442a-8744-5625c30e01ff","99d263f7-48ac-4be9-9f96-b1cd4380de1b","9a2c3eb1-bf66-4ddc-822d-bcc1121a8120","afe04fc5-1717-4f24-b061-7af855828533","4c615d59-27b6-48ea-a7bd-4c6d27bd91b5"],"title":"Attention"},{"cite_extract_rate":1,"cites":[793],"content":"\\label{sec:sparseattn}\nIn the standard self-attention mechanism, every token needs to attend to all other tokens.\nHowever, it is observed that for the trained Transformers the learned attention matrix $\\bA$ is often very sparse across most data points~.\nTherefore, it is possible to reduce computation complexity by incorporating structural bias to limit the number of query-key pairs that each query attends to. Under this limitation, we just compute the similarity score of the query-key pairs according to pre-defined patterns\n\\begin{equation}\\label{Sparse attention:position-based}\n \\mathrm{\\hat{\\bA}}_{ij} = \\begin{cases} \\bq_i\\bk_{j}^\\top & \\text{if token }i\\text{ attends to token }j,\\\\\n -\\infty & \\text{if token }i\\text{ does not attend to token }j,\n \\end{cases}\n\\end{equation}\\label{eq:sparseattn}\nwhere $\\mathrm{\\hat{\\bA}}$ is un-normalized attention matrix. In implementation the $-\\infty$ item is usually not stored in memory so as to decrease memory footprint.\nFrom another perspective, the standard attention can be regarded as a complete bipartite graph where each query receives information from all memory nodes and updates its representation.\nThe sparse attention can be considered as a sparse graph where some of the connections between nodes are removed.\nBased on the metrics of determining the sparse connection, we categorize these approaches into two classes: \\textit{position-based} and \\textit{content-based} sparse attention.","id":"c3955c2b-4114-49e4-bbb8-30061d5f27e8","level":"subsection","origin_cites_number":1,"parent_id":"57f59954-45c7-48ac-baf2-dc92c147a85e","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Sparse Attention"]],"subsections":["766f8fe1-2a61-4664-a9be-83f0177cc298","6429c3b2-0cb9-415c-9f1d-1f7bd028a380"],"title":"Sparse Attention"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:pos_based}\nIn position-based sparse attention, the attention matrix is limited according to some pre-defined patterns.\nAlthough these sparse patterns vary in different forms, we find that some of them can be decomposed into some atomic sparse patterns.\nWe first identify some atomic sparse patterns and then describe how these patterns are composed in some existing work. Finally, we introduce some extended sparse patterns for specific data types.","id":"766f8fe1-2a61-4664-a9be-83f0177cc298","level":"subsubsection","origin_cites_number":0,"parent_id":"c3955c2b-4114-49e4-bbb8-30061d5f27e8","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Sparse Attention"],["subsubsection","Position-based Sparse Attention"]],"subsections":["3abb14c2-1133-4417-a3ba-50491a89789a","cf0fe507-a0b3-499f-9145-a52ed5faa99e","8a0b3727-3354-4028-870f-983c49224baa"],"title":"Position-based Sparse Attention"},{"cite_extract_rate":1,"cites":[4722],"content":"There are mainly five types of atomic sparse attention patterns, as shown in Fig. \\ref{fig:atomic_sparse_attentions}.\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[global\\label{fig:global}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/atomic/global.pdf}\n}\n\\subfigure[band\\label{fig:band}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/atomic/band.pdf}\n}\n\\subfigure[dilated\\label{fig:dilated}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/atomic/dilated.pdf}\n}\n\\subfigure[random\\label{fig:random}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/atomic/random.pdf}\n}\n\\subfigure[block local\\label{fig:block_local}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/atomic/block_local.pdf}\n}\n\\end{center}\n\\caption{Some representative atomic sparse attention patterns. The colored squares means corresponding attention scores are calculated and a blank square means the attention score is discarded.}\\label{fig:atomic_sparse_attentions}\n\\end{figure}\n\\begin{enumerate}\n \\item \\textit{Global Attention}. To alleviate the degradation of the ability to model the long-range dependencies in sparse attention, one can add some global nodes\\footnote{In practice, these global nodes can be selected from the sequence (internal global nodes) or virtual nodes with trainable parameters (external global nodes).} as the hub for information propagation between nodes. These global nodes can attend all nodes in the sequence and the whole sequence attend to these global nodes, as illustrated in Fig. \\ref{fig:global}.\n \\item \\textit{Band Attention}(a.k.a \\textit{sliding window attention} or \\textit{local attention}). Since most data come with a strong property of locality, it is natural to restrict each query to attend to its neighbor nodes. A widely adopted class of such sparse pattern is band attention, in which the attention matrix is a band matrix as illustrated in Fig. \\ref{fig:band}.\n \\item \\textit{Dilated Attention}. Analogous to dilated CNNs~, one can potentially increase the receptive field of the band attention without increasing computation complexity by using a dilated window with gaps of dilation $w_d\\ge 1$, as depicted in Fig. \\ref{fig:dilated}. This can be easily extended to \\textit{strided attention}, where the window size is not limited but the dilation $w_d$ is set to a large value.\n \\item \\textit{Random Attention}. To increase the ability of non-local interactions, a few edges are randomly sampled for each query, as illustrated in Fig. \\ref{fig:random}. This is based on the observation that random graphs (e.g., Erd\\H os–R\\'enyi random graph) can have similar spectral properties with complete graphs that leads to a fast mixing time for random walking on graphs.\n \\item \\textit{Block Local Attention}. This class of attention segments input sequence into several non-overlapping query blocks, each of which is associated with a local memory block. All the queries in a query block attend to only the keys in the corresponding memory block. Fig. \\ref{fig:block_local} depicts a commonly used case where the memory blocks are identical to their corresponding query blocks.\n\\end{enumerate}","id":"3abb14c2-1133-4417-a3ba-50491a89789a","level":"paragraph","origin_cites_number":1,"parent_id":"766f8fe1-2a61-4664-a9be-83f0177cc298","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Sparse Attention"],["subsubsection","Position-based Sparse Attention"],["paragraph","4.1.1.1 Atomic Sparse Attention"]],"subsections":[],"title":"4.1.1.1 Atomic Sparse Attention"},{"cite_extract_rate":1,"cites":[7298,793,7371,1499,7363,134],"content":"Existing sparse attentions are often composed of more than one of the above atomic patterns. Fig.~\\ref{fig:sparse_attn} illustrates some representative compound sparse attention patterns.\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[Star-Transformer\\label{fig:starxformer}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/compound/star.pdf}\n}\\quad\n\\quad\n\\subfigure[Longformer\\label{fig:longformer}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/compound/longformer.pdf}\n}\\quad\n\\subfigure[ETC\\label{fig:etc}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/compound/etc.pdf}\n}\\quad\n\\subfigure[BigBird\\label{fig:bigbird}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/compound/bigbird.pdf}\n}\\quad\n\\caption{Some representative compound sparse attention patterns. The red boxes indicate sequence boundaries.}\\label{fig:sparse_attn}\n\\end{center}\n\\end{figure}\nStar-Transformer~ uses a combination of band attention and global attention. Specifically, Star-Transformer just includes only a global node and a band attention with the width of 3, in which any pair of non-adjacent nodes are connected through a shared global node and adjacent nodes are connected directly with each other. This kind of sparse pattern forms a star-shaped graph among nodes.\nLongformer~ uses a combination of band attention and internal global-node attention. The global nodes are chosen to be \\texttt{[CLS]} token for classification and all question tokens for Question Answering tasks. They also replace some of the band attention heads in upper layers with dilated window attention to increase the receptive field without increasing computation.\nAs a concurrent work to Longformer~, Extended Transformer Construction (ETC)~ utilizes combination of band attention and external global-node attention. ETC also includes a masking mechanism to handle structured inputs and adapt Contrastive Predictive Coding (CPC)~ for pre-training.\nIn addition to the band and global attention, BigBird~ uses additional random attention to approximate full attention. Their theoretical analysis also reveals that the usage of a sparse encoder and sparse decoder can simulate any Turing Machine, which explains the success of those sparse attention models.\nSparse Transformer~ uses a factorized attention where different sparse patterns are designed for different types of data. For data with a periodic structure (e.g., images), it uses a composition of band attention and strided attention. Whereas for data without a periodic structure (e.g., text), it uses a composition of block local attention combined with global attention, where global nodes are from fixed positions in the input sequence.","id":"cf0fe507-a0b3-499f-9145-a52ed5faa99e","level":"paragraph","origin_cites_number":6,"parent_id":"766f8fe1-2a61-4664-a9be-83f0177cc298","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Sparse Attention"],["subsubsection","Position-based Sparse Attention"],["paragraph","4.1.1.2 Compound Sparse Attention"]],"subsections":[],"title":"4.1.1.2 Compound Sparse Attention"},{"cite_extract_rate":0.6666666666666661,"cites":[1484,1459],"content":"Apart from the above patterns, some existing studies have explored extended sparse patterns for specific data types.\nFor text data, BP-Transformer~ constructs a binary tree where all tokens are leaf nodes and the internal nodes are span nodes containing many tokens. The edges in this graph are constructed so that each leaf node is connected to its neighbor leaf nodes and higher-level span nodes containing tokens from a longer distance. This approach can be seen as an extension of global attention, where global nodes are hierarchically organized and any pair of tokens are connected with paths in the binary tree. An abstract view of this method is illustrated in Fig. \\ref{fig:bpt}.\nThere are also some extensions for vision data. Image Transformer~ explores two types of attention: (1) flattening image pixels in raster-scan order and then applying block local sparse attention. (2) 2D block local attention, where query blocks and memory blocks are arranged directly in 2D plate, as depicted in Fig. \\ref{fig:block_local_2d}. As another example of sparse pattern on vision data, Axial Transformer~ applies independent attention modules over each axis of the image. Each attention module mixes information along one axis while keeping information along the other axis independent, as illustrated in Fig. \\ref{fig:axial}. This can be understood as horizontally and vertically flattening image pixels in raster-scan order and then applying strided attention with gaps of image width and height, respectively.\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[BPT\\label{fig:bpt}]{\n\\includegraphics[width=0.40\\linewidth]{assets/sparse_variants/other/bpt.pdf}\n}\\quad\n\\subfigure[block local (2D)\\label{fig:block_local_2d}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/other/block_local.pdf}\n}\\quad\n\\subfigure[axial (2D)\\label{fig:axial}]{\n\\includegraphics[width=0.18\\linewidth]{assets/sparse_variants/other/axial.pdf}\n}\n\\caption{Other types of sparse attentions. The red box indicates the query position, and the orange nodes/squares means corresponding tokens are attended to by the query.}\\label{fig:other_sparse_attn}\n\\end{center}\n\\end{figure}","id":"8a0b3727-3354-4028-870f-983c49224baa","level":"paragraph","origin_cites_number":3,"parent_id":"766f8fe1-2a61-4664-a9be-83f0177cc298","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Sparse Attention"],["subsubsection","Position-based Sparse Attention"],["paragraph","4.1.1.3 Extended Sparse Attention"]],"subsections":[],"title":"4.1.1.3 Extended Sparse Attention"},{"cite_extract_rate":1,"cites":[1453,1473,7362,794],"content":"Another line of work creates a sparse graph based on input content, i.e., the sparse connections are conditioned on inputs.\nA straightforward way of constructing a content-based sparse graph is to select those keys that are likely to have large similarity scores with the given query. To efficiently construct the sparse graph, we can recur to Maximum Inner Product Search (MIPS) problem, where one tries to find the keys with maximum dot product with a query without computing all dot product terms. Routing Transformer~ uses k-means clustering to cluster both queries $\\{\\bq_i\\}_{i=1}^T$ and keys $\\{\\bk_i\\}_{i=}^T$ on the same set of centroid vectors $\\{\\mathbf\\mu_i\\}_{i=1}^k$. Each query only attends to the keys that belong to the same cluster.\nDuring training, the cluster centroid vectors are updated using the exponentially moving average of vectors assigned to it, divided by the exponentially moving average of cluster counts:\n\\begin{align}\n\\tilde\\mu&\\leftarrow \\lambda\\tilde\\mu+(1-\\lambda)\\left(\\sum_{i:\\mu(\\bq_i)=\\mu}\\bq_i+\\sum_{j:\\mu(\\bk_j)=\\mu}\\bk_j\\right),\\\\\nc_\\mu&\\leftarrow \\lambda c_\\mu+(1-\\lambda)|\\mu|,\\\\\n\\mu&\\leftarrow \\frac{\\tilde\\mu}{c_\\mu},\n\\end{align}\nwhere $|\\mu|$ denotes the number of vectors currently in cluster $\\mu$ and $\\lambda\\in(0,1)$ is a hyperparameter.\nLet $\\mathcal{P}_i$ denote the set of indices of keys that the $i$-th query attend to. $\\mathcal{P}_i$ in Routing Transformer is defined as\n\\begin{equation}\n \\mathcal{P}_i=\\{j: \\mu(\\bq_i)=\\mu(\\bk_j) \\}.\n\\end{equation}\nReformer~ uses locality-sensitive hashing (LSH) to select key-value pairs for each query. The proposed LSH attention allows each token to attend only to the tokens within the same hashing bucket. The basic idea is to use an LSH function to hash queries and keys into several buckets, with similar items fall in the same bucket with high probability. Specifically, they use the random matrix method for the LSH function. Let $b$ be the number of buckets, given a random matrix $R$ of size $[D_k,b/2]$, the LSH function is computed by :\n\\begin{equation}\n h(x)=\\argmax([xR;-xR]).\n\\end{equation}\nThe LSH attention allows the $i$-th query to attend only to key-value pairs with indices\n\\begin{equation}\n \\mathcal{P}_i=\\{j: h(\\bq_i)=h(\\bk_j) \\}.\n\\end{equation}\nSparse Adaptive Connection (SAC)~ views the input sequence as a graph and learns to construct attention edges to improve task-specific performances using an adaptive sparse connection. SAC uses an LSTM edge predictor to construct edges between tokens. With no ground truth for edges, the edge predictor is trained with reinforcement learning.\nSparse Sinkhorn Attention~ first splits queries and keys into several blocks and assigns a key block to each query block. Each query is only allowed to attend to the keys in the key block that is assigned to its corresponding query block. The assignment of key blocks is controlled by a sorting network, which uses Sinkhorn normalization to produce a doubly stochastic matrix as the permutation matrix representing the assignment. They use this content-based block sparse attention along with block local attention introduced in Sec.~\\ref{sec:pos_based} to enhance the ability of the model to model locality.","id":"6429c3b2-0cb9-415c-9f1d-1f7bd028a380","level":"subsubsection","origin_cites_number":4,"parent_id":"c3955c2b-4114-49e4-bbb8-30061d5f27e8","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Sparse Attention"],["subsubsection","Content-based Sparse Attention"]],"subsections":[],"title":"Content-based Sparse Attention"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:kernel}\nAssuming $\\bQ,\\bK,\\bV\\in \\mathbb{R}^{T\\times D}$, the complexity of computing $\\softmax(\\bQ\\bK^\\top)\\bV$ is quadratic w.r.t. sequence length $T$, as illustrated in Fig. \\ref{fig:standard_complexity}. If $\\softmax(\\bQ\\bK^\\top)$ can be disentangled into $\\bQ'\\bK'^\\top$, we can compute $\\bQ'\\bK'^\\top\\bV$ in reversed order (i.e., $\\bQ'\\left(\\textcolor[rgb]{0,0,1}{\\bK'^\\top\\bV}\\right)$), leading to a complexity of $\\mathcal{O}(T)$.\nLet $\\hat{\\bA} = \\exp(\\bQ\\bK^\\top)$ denote un-normalized attention matrix, and $\\exp(\\cdot)$ is applied element-wise, the regular attention can be rewritten as $\\bZ={\\bD}^{-1}\\hat{\\bA}\\bV$, where $\\bD = \\mathrm{diag}(\\hat{\\bA}\\mathbf{1}_T^\\top)$; $\\mathbf{1}_T^\\top$ is the all-ones column vector of length $T$; $\\mathrm{diag}(\\cdot)$ is a diagonal\nmatrix with the input vector as the diagonal.\nLinearized attention is a class of methods that approximate or replace the unnormalized attention matrix $\\exp(\\bQ\\bK^\\top)$ with $\\phi(\\bQ)\\phi(\\bK)^\\top$, where $\\phi$ is a feature map that is applied in row-wise manner. Hence the computation of unnormalized attention matrix can be linearized by computing $\\phi(\\bQ)\\left(\\textcolor[rgb]{0,0,1}{\\phi(\\bK)^\\top\\bV}\\right)$\\footnote{Similarly, the partition term $\\bD$ can be computed with $\\phi(\\bQ)\\left(\\textcolor[rgb]{0,0,1}{\\phi(\\bK)^\\top\\mathbf{1}_T^\\top}\\right)$ in linear time.}, as illustrated in Fig. \\ref{fig:linearized_complexity}.\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[standard self-attention]{\\label{fig:standard_complexity}\n\\includegraphics[width=0.45\\textwidth]{assets/kernel/vanilla.pdf}\n}\\hspace{.2in}\n\\subfigure[linearized self-attention]{\\label{fig:linearized_complexity}\n\\includegraphics[width=0.45\\textwidth]{assets/kernel/kernel.pdf}\n}\n\\caption{Illustration of complexity difference between standard self-attention and linearized self-attention.}\n\\end{center}\n\\end{figure}\nTo gain further insights into linearized attention, we derive the formulation in vector form. We consider a general form of attention\n\\begin{equation}\n \\bz_i=\\sum_j \\frac{\\mathrm{sim}(\\bq_i, \\bk_j)}{\\sum_{j'}\\mathrm{sim}(\\bq_i, \\bk_{j'})}\\bv_j,\n\\end{equation}\nwhere $\\mathrm{sim}(\\cdot,\\cdot)$ is a scoring function measuring similarity between input vectors. In vanilla Transformer, the scoring function is the exponential of inner product $\\exp(\\langle\\cdot,\\cdot\\rangle)$. A natural choice of $\\mathrm{sim}(\\cdot,\\cdot)$ is a kernel function $\\mathcal{K}(\\bx,\\by)=\\phi(\\bx)\\phi(\\by)^\\top$, which leads to\n\\begin{align}\n \\bz_i&=\\sum_j \\frac{\\phi(\\bq_i) \\phi(\\bk_j)^\\top}{\\sum_{j'} \\phi(\\bq_i) \\phi(\\bk_{j'})^\\top}\\bv_j\\\\\n &=\\frac{\\phi(\\bq_i) \\textcolor[rgb]{0,0,1}{\\sum_j\\phi(\\bk_j)\\otimes\\bv_j}}{\\phi(\\bq_i)\\textcolor[rgb]{0,0,1}{\\sum_{j'} \\phi(\\bk_{j'})^\\top}},\\label{eq:linearized}\n\\end{align}\nwhere $\\otimes$ denotes outer product of vectors. Based on this formulation, attention can be linearized by first computing the highlighted terms $\\sum_j\\phi(\\bk_j)\\otimes\\bv_j$ and$\\sum_{j'} \\phi(\\bk_{j'})^\\top$. This could be especially beneficial for autoregressive attention, as the cumulative sums $\\bS_i=\\sum_{j=1}^{i} \\phi(\\bk_j)\\otimes\\bv_j$ and $\\bu_i=\\sum_{j=1}^{i} \\phi(\\bk_j)$ can be computed from $\\bS_{i-1}$ and $\\bu_{i-1}$ in constant time. The effectively enables Transformer decoders to run like RNNs.\nAn interpretation of Eq. \\eqref{eq:linearized} is that the model maintains a \\textit{memory matrix} by aggregating \\textit{associations} represented by outer products of (feature mapped) keys and values, and then retrieve a value by multiplying the memory matrix with feature mapped query with proper normalization. There are two key components in this approach: (1) feature map $\\phi(\\cdot)$, and (2) aggregation rule.","id":"a4f89d55-c38a-442a-8744-5625c30e01ff","level":"subsection","origin_cites_number":0,"parent_id":"57f59954-45c7-48ac-baf2-dc92c147a85e","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Linearized Attention"]],"subsections":["c219036a-ca77-49bd-b9c6-4ac269a7d3d6","22632f15-ca9f-4bd6-a9a4-e425db6125fa"],"title":"Linearized Attention"},{"cite_extract_rate":0.6666666666666661,"cites":[8384,1494,1182,798],"content":"Linear Transformer~ propose to use a simple feature map $\\phi_i(\\bx)=\\mathrm{elu}(x_i)+1$. This feature map does not aim to approximate dot product attention, but is empirically proved to perform on par with the standard Transformer.\nPerformer~ uses random feature maps that approximate the scoring function of Transformer. The random feature maps take functions $f_1,\\cdots,f_l:\\mathbb R\\rightarrow \\mathbb R$ and $h:\\mathbb{R}^D\\rightarrow \\mathbb R$.\n\\begin{equation}\n \\phi(\\bx) = \\frac{h(\\bx)}{\\sqrt{m}}[f_1(\\omega_1^\\top \\bx),\\cdots,f_m(\\omega_m^\\top \\bx),\\cdots,f_l(\\omega_1^\\top \\bx),\\cdots,f_l(\\omega_m^\\top \\bx)],\n\\end{equation}\nwhere $\\omega_1,\\cdots,\\omega_m\\stackrel{\\text{iid}}{\\sim} \\mathcal D$ are drawn from some distribution $\\mathcal D\\in\\mathcal P(\\mathbb{R}^D)$.\nThe first version of Performer~ is inspired from the random Fourier feature map~ that was originally used to approximate Gaussian kernel. It uses trigonometric functions with $h(\\bx)=\\exp(\\frac{\\|\\bx\\|^2}{2}), l=2, f_1=\\sin, f_2=\\cos$. This approach has also been used in Random Feature Attention (RFA)~, with the difference that $h(\\bx)$ is set to $1$ as the queries and keys are $\\ell_2$-normalized before applying the feature map.\nAlthough the trigonometric random feature map leads to an unbiased approximation, it does not guarantee non-negative attention scores and thus could lead to unstable behaviors and abnormal behaviors. To mitigate this issue, the second version of Performer~ proposes positive random feature maps, which uses $h(\\bx)=\\exp(-\\frac{\\|\\bx\\|^2}{2}),l=1,f_1=\\exp$ and thus guarantees unbiased and non-negative approximation of dot-product attention. This approach is more stable than and reports better approximation results.\nIn addition to using random feature maps to approximate standard dot product attention, and also explore approximating order-1 arc-cosine kernel with $h(\\bx)=1,l=1,f_1=\\mathrm{ReLU}$. This feature map has been show to be effective in various tasks including machine translation and protein sequence modeling.\n design a feature map that aims at facilitating orthogonality in feature space. Specifically, given an input $\\bx\\in\\mathbb{R}^D$, the feature map $\\phi:\\mathbb{R}^D\\rightarrow \\mathbb{R}^{2\\nu D}$ is defined by the partial function\n\\begin{equation}\n \\phi_{i+2(j-1)D}(\\bx)=\\mathrm{ReLU}([\\bx,-\\bx])_i\\mathrm{ReLU}([\\bx,-\\bx])_{i+j}\\quad\\text{for }i=1,\\cdots,2D,j=1,\\cdots,\\nu.\n\\end{equation}","id":"c219036a-ca77-49bd-b9c6-4ac269a7d3d6","level":"subsubsection","origin_cites_number":6,"parent_id":"a4f89d55-c38a-442a-8744-5625c30e01ff","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Linearized Attention"],["subsubsection","Feature Maps"]],"subsections":[],"title":"Feature Maps"},{"cite_extract_rate":0.8,"cites":[8384,1494,1182,798],"content":"In Eq. \\eqref{eq:linearized} the associations $\\{\\phi(\\bk)_j\\otimes\\bv_j\\}$ are aggregated into the memory matrix by simple summation. This is adopted by several studies~. However, it could be more beneficial for the network to selectively drop associations as new associations are added to the memory matrix.\nRFA~ introduces a gating mechanism to the summation to model local dependency in sequence data. Specifically, when adding a new association to the memory matrix $\\bS$, at a particular time step, they weigh $\\bS$ by a learnable, input-dependent scalar $g$, and the new association by $(1-g)$ (and a similar mechanism to $\\bu$). With this modification, history associations are exponentially decayed and recent context is favored in each timestep.\n argue that simple summation limits the capacity of the memory matrix and thus propose to enlarge the capacity in a write-and-remove fashion. Specifically, given a new input key-value pair $(\\bk_i, \\bv_i)$, the model first retrieve the value $\\bar{\\bv}_i$ currently associated with $\\bk_i$ using matrix multiplication. It then writes to the memory matrix a convex combination of $\\bar{\\bv}_i$ and $\\bv_i$, using a input-dependent gating scalar $g$, and removes the association $\\bar{\\bv}_i$. They also propose \\textit{sum normalization} (normalizing $\\phi(\\bq_i),\\phi(\\bk_i)$ by the sum of their components before updating the memory matrix) instead of normalizing with the denominator in Eq. \\eqref{eq:linearized} for this aggregation rule.","id":"22632f15-ca9f-4bd6-a9a4-e425db6125fa","level":"subsubsection","origin_cites_number":5,"parent_id":"a4f89d55-c38a-442a-8744-5625c30e01ff","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Linearized Attention"],["subsubsection","Aggregation Rule"]],"subsections":[],"title":"Aggregation Rule"},{"cite_extract_rate":0,"cites":[],"content":"Apart from using sparse attention or kernel-based linearized attention, one could also reduce the complexity of attention by reducing the number of queries or key-value pairs, which leads to \\textit{query prototyping} and \\textit{memory compression}\\footnote{The key-value pairs are often referred to as a key-value memory (hence the name memory compression).} methods, respectively.","id":"99d263f7-48ac-4be9-9f96-b1cd4380de1b","level":"subsection","origin_cites_number":0,"parent_id":"57f59954-45c7-48ac-baf2-dc92c147a85e","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Query Prototyping and Memory Compression"]],"subsections":["ad89ae14-36aa-4aeb-ab1d-65a0637088d3","5fdbbc99-3354-4d28-a5a0-c76985e76cdb"],"title":"Query Prototyping and Memory Compression"},{"cite_extract_rate":1,"cites":[1479,1472],"content":"In query prototyping, several prototypes of queries serve as the main source to compute attention distributions. The model either copies the distributions to the positions of represented queries or filling those positions with discrete uniform distributions. Fig. \\ref{fig:query_prototype} illustrates the computing flow of query prototyping.\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[Query prototyping]{\\label{fig:query_prototype}\n\\includegraphics[height=1.5in]{assets/prototype/query_prototype.pdf}\n}\\hspace{.2in}\n\\subfigure[Memory compression]{\\label{fig:compressed_mem}\n\\includegraphics[height=1.5in]{assets/prototype/mem_compress.pdf}\n}\n\\caption{Query prototyping and memory compression.}\\label{fig:prototype_mem_compress}\n\\end{center}\n\\end{figure}\nClustered Attention~ groups queries into several clusters and then computes attention distributions for cluster centroids. All queries in a cluster share the attention distribution calculated with the corresponding centroid.\nInformer~ selects prototypes from queries using explicit query sparsity measurement, which is derived from an approximation of the Kullback-Leibler divergence between the query's attention distribution and the discrete uniform distribution. Attention distributions are then only calculated for the top-$u$ queries under query sparsity measurement. The rest of the queries are assigned with discrete uniform distributions.","id":"ad89ae14-36aa-4aeb-ab1d-65a0637088d3","level":"subsubsection","origin_cites_number":2,"parent_id":"99d263f7-48ac-4be9-9f96-b1cd4380de1b","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Query Prototyping and Memory Compression"],["subsubsection","Attention with Prototype Queries"]],"subsections":[],"title":"Attention with Prototype Queries"},{"cite_extract_rate":0.833333333333333,"cites":[7333,1503,1133,1504,1482],"content":"\\label{sec:compressed_mem}\nApart from decreasing the number of queries with query prototyping, one can also reduce the complexity by reducing the number of the key-value pairs before applying the attention mechanism, as depicted in Fig. \\ref{fig:compressed_mem}.\n propose Memory Compressed Attention (MCA) that reduces the number of keys and values using a strided convolution. This modification is used as a complement to local attention proposed in the same work (as discussed in Sec.~\\ref{sec:sparseattn}), in that it can capture global context. The mechanism reduces the number of keys and values by a factor of kernel size $k$ and thus allowing to process significantly longer sequences than vanilla Transformer given the same computation resources.\nSet Transformer~ and Luna~ use a number of external trainable global nodes to summarize information from inputs and then the summarized representations serve as a compressed memory that the inputs attend to. This reduces the quadratic complexity of self-attention to linear complexity w.r.t. sequence length.\nLinformer~ utilizes linear projections to project keys and values from length $n$ to a smaller length $n_k$. This also reduces the complexity of self-attention to linear. The drawback of this approach is that an input sequence length has to be assumed and hence it cannot be used in autoregressive attention.\nPoolingformer~ adopts two-level attention that combines a sliding window attention and a compressed memory attention. The compressed memory module is used after the sliding window attention to increase the receptive field. They explore a few different pooling operations as the compression operation to compress the number of keys and values, including max pooling and pooling with Dynamic Convolution~.","id":"5fdbbc99-3354-4d28-a5a0-c76985e76cdb","level":"subsubsection","origin_cites_number":6,"parent_id":"99d263f7-48ac-4be9-9f96-b1cd4380de1b","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Query Prototyping and Memory Compression"],["subsubsection","Attention with Compressed Key-Value Memory"]],"subsections":[],"title":"Attention with Compressed Key-Value Memory"},{"cite_extract_rate":0.5,"cites":[7333],"content":"Some empirical and theoretical analyses~ report the self-attention matrix $\\bA \\in \\mathbb{R}^{T\\times T}$ is often low-rank\\footnote{The rank of $\\bA$ is far lower than input length $T$.}. The implications of this property are twofold: (1) The low-rank property could be explicitly modeled with parameterization; (2) The self-attention matrix could be replaced by a low-rank approximation.","id":"9a2c3eb1-bf66-4ddc-822d-bcc1121a8120","level":"subsection","origin_cites_number":2,"parent_id":"57f59954-45c7-48ac-baf2-dc92c147a85e","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Low-rank Self-Attention"]],"subsections":["8570c9f2-0cf3-47a3-befe-ca2093b97f7a","fc87c410-c4a6-46de-806f-b6f8720b2eb4"],"title":"Low-rank Self-Attention"},{"cite_extract_rate":0,"cites":[],"content":"The fact that the rank of the attention matrix is less than sequence length implies that, for scenarios where the inputs are typically short, setting $D_k>T$ would be more than an over-parameterization and lead to overfitting. It is thus reasonable to limit the dimension of $D_k$ to explicitly model the low-rank property as an inductive bias. decompose self-attention matrix into a low-rank attention module with small $D_k$ that captures long-range non-local interactions, and a band attention module that captures local dependencies.","id":"8570c9f2-0cf3-47a3-befe-ca2093b97f7a","level":"subsubsection","origin_cites_number":1,"parent_id":"9a2c3eb1-bf66-4ddc-822d-bcc1121a8120","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Low-rank Self-Attention"],["subsubsection","Low-rank Parameterization"]],"subsections":[],"title":"Low-rank Parameterization"},{"cite_extract_rate":0.33333333333333304,"cites":[1494],"content":"Another implication of the low-rank property of the attention matrix is that one can use a low-rank matrix approximation to reduce the complexity of self-attention. A closely related methodology is the low-rank approximation of kernel matrices. We believe some existing works are inspired by kernel approximation.\nSome of the aforementioned linearized attention methods in Sec.~\\ref{sec:kernel} are inspired from kernel approximation with random feature maps. For example, Performer~ follows the Random Fourier feature map originally proposed to approximate Gaussian kernels. The method first decomposes the attention distribution matrix $\\bA$ into $\\bC_Q\\bG\\bC_K$ where $\\bG$ is a Gaussian kernel matrix and the random feature map is used to approximate $\\bG$.\nAnother line of work follow the idea of Nystr\\\"om method. These Nystr\\\"om-based methods~ first select $m$ landmark nodes from the $T$ inputs with down-sampling methods (e.g., strided average pooling). Let $\\tilde\\bQ,\\tilde\\bK$ be the selected landmark queries and keys, then the follow approximation is used in the attention computation\n\\begin{equation}\\label{eq:nystrom}\n \\tilde{\\bA}=\\mathrm{softmax}\\left(\\bQ\\tilde{\\bK}^\\top\\right)\\left(\\mathrm{softmax}\\left(\\tilde\\bQ\\tilde{\\bK}^\\top\\right)\\right)^{-1}\\mathrm{softmax}\\left(\\tilde\\bQ\\bK^\\top\\right).\n\\end{equation}\nNote that $\\bM^{-1}=\\left(\\mathrm{softmax}\\left(\\tilde\\bQ\\tilde{\\bK}^\\top\\right)\\right)^{-1}$ in Eq. \\eqref{eq:nystrom} does not always exist. To mitigate this issue, CSALR~ adds an identity matrix to $\\bM$ to make sure that the inverse always exists. Nystr{\\\"{o}}mformer~ uses the Moore-Penrose pseudoinverse of $\\bM$ instead of the inverse so that the approximation can be made for cases where $\\bM$ is singular.","id":"fc87c410-c4a6-46de-806f-b6f8720b2eb4","level":"subsubsection","origin_cites_number":3,"parent_id":"9a2c3eb1-bf66-4ddc-822d-bcc1121a8120","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Low-rank Self-Attention"],["subsubsection","Low-rank Approximation"]],"subsections":[],"title":"Low-rank Approximation"},{"cite_extract_rate":0,"cites":[],"content":"\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=0.5\\linewidth]{assets/attention_w_prior.pdf}\n \\caption{Attention with prior. This type of model fuse generated attention scores with pior attention scores, producing the final attention scores for attention computation.}\n \\label{fig:attention_w_prior}\n\\end{figure}\nAttention mechanism generally outputs an expected attended value as a weighted sum of vectors, where the weights are an attention distribution over the values. Traditionally, the distribution is generated from inputs (e.g., $\\mathrm{softmax}(\\bQ\\bK^\\top)$ in vanilla Transformer). As a generalized case, attention distribution can also come from other sources, which we refer to as \\textit{prior}. Prior attention distribution can be a supplement or substitute for distribution generated from inputs. We abstract this formulation of attention as \\textit{attention with prior}, as depicted in Fig. \\ref{fig:attention_w_prior}. In most cases, the fusion of two attention distribution can be done by computing a weighted sum of the scores corresponding to the prior and generated attention before applying softmax.","id":"afe04fc5-1717-4f24-b061-7af855828533","level":"subsection","origin_cites_number":0,"parent_id":"57f59954-45c7-48ac-baf2-dc92c147a85e","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Attention with Prior"]],"subsections":["65f37434-480a-4748-b65e-d8966645bec9","af364707-c13f-44cb-9e4e-2366287a19b6","a795db35-abb3-4e07-ad32-5c47de43f06f","55c07bfe-1106-456b-a01c-8edd7f25017b"],"title":"Attention with Prior"},{"cite_extract_rate":0.5,"cites":[1489],"content":"Some types of data (e.g., text) can exhibit a strong preference for the locality. This property can be explicitly encoded as a prior attention. A simple method would be to use a Gaussian distribution over positions. Specifically, one could multiply the generated attention distribution with some Gaussian density and then renormalize, which is equivalent to adding to the generated attention scores $\\bA$ a bias term $\\bG$, where higher $G_{ij}$ indicates a higher prior probability that the $i$-th input attend to the $j$-th input.\n proposes to first predict a central position $p_i$ for each $\\bq_i$ using a simple feed-forward network. The Gaussian bias is then defined to be\n\\begin{equation}\n G_{ij}=-\\frac{(j-p_i)^2}{2\\sigma^2},\n\\end{equation}\nwhere $\\sigma$ denotes standard deviation for the Gaussian and can be determined as a hyperparameter or predicted from inputs.\nGaussian Transformer~ assumes the central position to be $i$ for each $\\bq_i$ and defines the bias to bes\n\\begin{equation}\n G_{ij}=-|w(i-j)^2+b|,\n\\end{equation}\nwhere $w\\ge 0, b\\le 0$ are scalar parameters that controls the deviation and reduce the weight for central position, respectively.","id":"65f37434-480a-4748-b65e-d8966645bec9","level":"subsubsection","origin_cites_number":2,"parent_id":"afe04fc5-1717-4f24-b061-7af855828533","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Attention with Prior"],["subsubsection","Prior that Models locality"]],"subsections":[],"title":"Prior that Models locality"},{"cite_extract_rate":0.6666666666666661,"cites":[1497,1492],"content":"\\label{sec:prior_prev_module}\nIn Transformer architecture, it is often observed the attention distributions are similar in adjacent layers. It is thus natural to provide attention distribution from previous layer as a prior for attention computation. The final attention scores can be defined as\n\\begin{equation}\\label{crosslayer}\n \\hat{\\bA}^{(l)} = w_1\\cdot\\bA^{(l)}+w_2\\cdot g(\\bA^{(l-1)}),\n\\end{equation}\nwhere $\\bA^{(l)}$ denotes the attention scores of the $l$-th layer, $w_1,w_2\\in\\mathbb R$ are weight applied to the scores from adjacent layers, and $g:\\mathbb{R}^{n\\times n}\\rightarrow \\mathbb{R}^{n\\times n}$ is a function that translate previous scores to the prior to be applied.\nPredictive Attention Transformer~ proposes to apply a 2D-convolutional layer to previous attention scores and compute the final attention scores as a convex combination of the generated attention scores and the convolved scores. This is equivalent to setting $w_1=\\alpha,w_2=1-\\alpha$ and $g(\\cdot)$ to be a convolutional layer in Eq. \\eqref{crosslayer}. They experiment training such a model from scratch and finetune after adapting the pre-trained BERT model, and both sets of experiments show improvements over baseline models.\nRealformer~ uses adds the previous attention scores directly to the generated attention scores, thus resembles a residual skip connection on attention maps. It's equivalent to setting $w_1=w_2=1$ and $g(\\cdot)$ to be identity map in Eq. \\eqref{crosslayer}. They conduct pre-training experiments on this model. The results show that this model outperforms the baseline BERT model in multiple datasets and surpasses the baseline model even when pre-training budgets are significantly lower.\nAs an extreme case, Lazyformer~ proposes to share attention maps between a number of adjacent layers. This is equivalent to setting $g(\\cdot)$ to identity and switch the settings of $w_1=0,w_2=1$ and $w_1=1,w_2=0$ alternatingly. The benefit of this approach is that the attention maps are computed only once and reused several times in the succeeding layers, thus reducing the computation cost. Their pre-training experiments show that the resulting model remains effective while being much more efficient to compute.","id":"af364707-c13f-44cb-9e4e-2366287a19b6","level":"subsubsection","origin_cites_number":3,"parent_id":"afe04fc5-1717-4f24-b061-7af855828533","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Attention with Prior"],["subsubsection","Prior from Lower Modules"]],"subsections":[],"title":"Prior from Lower Modules"},{"cite_extract_rate":1,"cites":[725,1505],"content":"Adapters are task-dependent, trainale modules that are attached in specific locations of a pre-trained network for cross-task efficient parameter sharing~. propose a Conditionally Adaptive Multi-Task Learning (CAMTL) framework that uses a trainable attention prior $M(\\bz_i)$ that depends on task encoding $\\bz_i\\in\\mathbb{R}^{D_z}$\n\\begin{equation}\n M(\\bz_i)=\\bigoplus_{j=1}^m A'_j(\\bz_i),\\quad A'_j(\\bz_i)=A_j\\gamma_i(\\bz_i)+\\beta_i(\\bz_i),\n\\end{equation}\nwhere $\\bigoplus$ denotes direct sum, $A_j\\in \\mathbb{R}^{(n/m)\\times(n/m)}$ are trainable parameters, and $\\gamma_j,\\beta_j:\\mathbb{R}^{D_z}\\rightarrow \\mathbb{R}^{(n/m)\\times(n/m)}$ are are Feature Wise\nLinear Modulation functions~. A maximum sequence length $n_{max}$ is specified in implementation. The prior is formulated as a block diagonal matrix and added to the attention scores of upper layers in pre-trained Transformers to serve as an adapter for parameter-efficient multi-task inductive knowledge transfer.","id":"a795db35-abb3-4e07-ad32-5c47de43f06f","level":"subsubsection","origin_cites_number":2,"parent_id":"afe04fc5-1717-4f24-b061-7af855828533","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Attention with Prior"],["subsubsection","Prior as Multi-task Adapters"]],"subsections":[],"title":"Prior as Multi-task Adapters"},{"cite_extract_rate":0.8,"cites":[1489,7364,7366,1476],"content":"Some works have explored using an attention distribution that is independent of pair-wise interaction between inputs. In other words, their models exploit only a prior attention distribution.\n design an efficient Transformer decoder variant called average attention network that uses a discrete uniform distribution as the sole source of attention distribution. The values are thus aggregated as a cumulative-average of all values. To improve the expressiveness of the network, they further adds a feed-forward\ngating layer on top of the average attention module. The advantage of this approach is that the adapted Transformer decoder can train in a parallel manner as usual Transformers do and decode like an RNN, thus avoiding the $\\mathcal O(T^2)$ complexity in decoding.\n utilize a Gaussian distribution as the hardcoded attention distribution for attention calculation. The intuition is very similar to and in that attention distribution should be focused on a certain local window. Distinctively, they drop the generated attention completely and use only the Gaussian distribution for attention computation. In this approach, the mean (central position) and variance are designed to be hyperparameters. The experiments show that the hardcoded attention, when applied only to self-attention, can achieve comparable performance to the baseline model in machine translation tasks.\nSynthesizer~ proposes to replace generated attention scores with: (1) a learnable, randomly initialized attention scores, and (2) attention scores output by a feed-forward network that is only conditioned on the querying input itself. The experiments on machine translation and language modeling show that these variants can achieve competitive performance with vanilla Transformer. It is not explained why these variants work but the empirical results are intriguing.","id":"55c07bfe-1106-456b-a01c-8edd7f25017b","level":"subsubsection","origin_cites_number":5,"parent_id":"afe04fc5-1717-4f24-b061-7af855828533","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Attention with Prior"],["subsubsection","Attention with Only Prior"]],"subsections":[],"title":"Attention with Only Prior"},{"cite_extract_rate":0,"cites":[],"content":"Multi-head attention is appealing for the ability to jointly attend to information from different representation subspaces at different positions. However, there is no mechanism to guarantee that different attention heads indeed capture\ndistinct features.","id":"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5","level":"subsection","origin_cites_number":0,"parent_id":"57f59954-45c7-48ac-baf2-dc92c147a85e","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Improved Multi-Head Mechanism"]],"subsections":["dcd668ac-e997-46ab-8e9e-c34fd7ac742a","6b049103-9bb4-4d9c-bd95-dceaba44ed6d","c68f2053-1c0a-4bd7-8d8a-92474761c343","be224101-c6fc-4828-9ad0-f1f79c3097ee"],"title":"Improved Multi-Head Mechanism"},{"cite_extract_rate":1,"cites":[1474,7372,1460,1498,7369,38],"content":"A basic motivation for using multi-head attention is to allow the model to jointly attend to information from different representation subspaces at different positions~. However, in vanilla Transformer there is no explicit mechanism to guarantee different behavior across attention heads, nor is there any mechanism for heads to interact with each other. A line of work is dedicated to improving multi-head mechanism by introducing incorporating more sophisticated mechanisms that guide the behavior of different attention heads or allow interaction across attention heads.\n introduce an auxiliary disagreement regularization term into loss function to encourage diversity among different attention heads. Two regularization terms are respectively to maximize cosine distances of the input subspaces and output representations, while the last one is to disperse\nthe positions attended by multiple heads with element-wise multiplication of the corresponding attention matrices.\nSeveral probing works have revealed that pre-trained Transformer models exhibit certain patterns of self-attention that are of little linguistic backing. As a representative work, identify several simple attention patterns in BERT. For instance, many of the attention heads simply pay attention to special BERT tokens \\texttt{[CLS]} and \\texttt{[SEP]}. As a result, some constraints can be introduced to boost the training of Transformer models. To this end, propose to use an auxiliary loss, which is defined to be the Frobenius norm between attention distribution maps and predefined attention patterns.\nTalking-head Attention~ uses a talking head mechanism that linearly projects the generated attention scores from $h_k$ to $h$ heads, applies softmax in that space, and then projects to $h_v$ heads for value aggregation. The motivation is to encourage the model to move information between attention heads in a learnable fashion.\nCollaborative Multi-head Attention~ uses shared query and key projection $\\bW^Q$ and $\\bW^K$ and a mixing vector $\\bm_i$ for the $i$-th head to filter from the projection parameters such that Eq. \\eqref{eq:headi} is adapted to\n\\begin{equation}\n \\mathrm{head}_i=\\mathrm{Attention}(\\bQ\\bW^Q\\mathrm{diag}(\\bm_i),\\bK\\bW^K, \\bV\\bW_i^V),\n\\end{equation}\nwhere $\\bW^Q$ and $\\bW^K$ are shared by all the attention heads.","id":"dcd668ac-e997-46ab-8e9e-c34fd7ac742a","level":"subsubsection","origin_cites_number":6,"parent_id":"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Improved Multi-Head Mechanism"],["subsubsection","Head Behavior Modeling"]],"subsections":[],"title":"Head Behavior Modeling"},{"cite_extract_rate":0.5,"cites":[1465],"content":"Vanilla attention adopts full attention spans assume, where a query can attend to all of the key-value pairs. However, it is often observed that some heads focus their attention distribution mainly in a local context while some other heads attend to broader contexts. It could thus be beneficial to restrict the attention spans:\n\\begin{itemize}\n \\item \\textit{Locality}. Restricting attention spans induce explicit local constraints. This is advantageous in cases where locality is an important prior.\n \\item \\textit{Efficiency}. If implemented appropriately, such a model can scale to very long sequences without introducing additional memory footprint and computational time.\n\\end{itemize}\nRestricting attention spans can be expressed as multiplying each attention distribution value with a mask value and then re-normalize, where the mask can be expressed as a non-increasing function that maps a distance to a value in $[0,1]$. A vanilla attention assigns a mask value of $1$ for all distances, as depicted in Fig. \\ref{span_mask}(a).\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[mask function for vanilla attention]{\n\\includegraphics[width=0.3\\linewidth]{assets/span_mask/full_mask.pdf}\n}\n\\subfigure[mask function for adaptive span]{\n\\includegraphics[width=0.3\\linewidth]{assets/span_mask/adaptive_mask.pdf}\n}\n\\subfigure[mask function for fixed span]{\n\\includegraphics[width=0.3\\linewidth]{assets/span_mask/scale.pdf}\n}\n\\caption{Three types of span masking function $m(x)$. The horizontal axis represents distance $x$ and vertical axis the mask value.}\\label{span_mask}\n\\end{center}\n\\end{figure}\n propose to use a learnable attention span, as depicted in Fig. \\ref{span_mask}(b) . The mask is parameterized by a learnable scalar $z$ and a hyperparameter $R$. The experiments on character-level language modeling show that the adaptive-span models outperform baseline models while having significantly fewer FLOPS. It is also observed that lower layers generally have smaller learned spans and higher layers otherwise. This indicates that the model can learn a hierarchical composition of features.\nMulti-Scale Transformer~ proposes to use a fixed attention span, with different heads in different layers using a different max span. The fixed attention span is depicted in Fig. \\ref{span_mask}(c). The attention is restricted within a fixed window which is controlled by a \\textit{scale} value $w$. They design the scales from an intuitive linguistic perspective and empirical observation from BERT such that higher layers tend to have more large scales (e.g., large span size), and lower layers should be confined with a smaller scale. Their experiments on several tasks show that the model can outperform baseline models while accelerating inference on long sequences.","id":"6b049103-9bb4-4d9c-bd95-dceaba44ed6d","level":"subsubsection","origin_cites_number":2,"parent_id":"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Improved Multi-Head Mechanism"],["subsubsection","Multi-head with Restricted Spans"]],"subsections":[],"title":"Multi-head with Restricted Spans"},{"cite_extract_rate":0.8,"cites":[306,1452,8455,38],"content":"After each attention head computes its output representation, the vanilla multi-head attention~ concatenates these representation and then apply a linear transformation to the concatenated representation to obtain the final output representation, as formulated in Eq. \\eqref{eq:multihead}. Combining Eq. \\eqref{attention}\\eqref{eq:multihead} and \\eqref{eq:headi}, one can see that this \\textit{concatenate-and-project} formulation is equivalent to summation over $H$ re-parameterized attention outputs. To this end, we first divide $\\bW^O\\in\\mathbb{R}^{D_m\\times D_m}$ into $H$ blocks\n\\begin{equation}\n \\bW^O=[\\bW_1^O;\\bW_2^O;\\cdots;\\bW_H^O],\n\\end{equation}\nwhere each $\\bW_i^O$ is of dimension $D_v\\times D_m$. It's thus easy to see that multi-head attention can be reformulated as\n\\begin{equation}\n \\mathrm{MultiHeadAttn}(Q,K,V)= \\sum_{i=1}^H \\mathrm{Attention}(Q \\bW_i^Q,K \\bW_i^K, V\\textcolor[rgb]{0,0,1} {\\bW_i^V\\bW_i^O}).\n\\end{equation}\nOne might argue that this simple \\textit{aggregate-by-summation} paradigm does not fully exploit the expressiveness of multi-head attention and that it is more desirable to use a more complex aggregation.\n propose to use routing methods, originally proposed for capsule networks~, to further aggregate information produced by different attention heads. The outputs of attention heads are first transformed into input capsules, then output capsules are obtained after the iterative routing process. The output capsules are then concatenated as a final output of multi-head attention. These two works both utilizes two routing mechanisms, namely \\textit{dynamic routing} and \\textit{EM routing}. One would notice that iterative routing introduces additional parameters and computational overhead. empirically show that applying the routing mechanism only to the lower layers can best balance the translation performance and computational efficiency.","id":"c68f2053-1c0a-4bd7-8d8a-92474761c343","level":"subsubsection","origin_cites_number":5,"parent_id":"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Improved Multi-Head Mechanism"],["subsubsection","Multi-head with Refined Aggregation"]],"subsections":[],"title":"Multi-head with Refined Aggregation"},{"cite_extract_rate":1,"cites":[1507,1506],"content":"Several other modifications to the multi-head mechanism have been proposed to improve multi-head attention.\n propose multi-query attention, where key-value pairs are shared among attention heads (i.e., to use only one key projection and one value projection for all attention heads). The advantage of this method is that it reduces the memory bandwidth requirements for decoding and results in a model that is faster to decode, while incurring only minor quality degradation from the baseline.\n establish that small attention key size can affect its ability to represent arbitrary distribution. They thus propose to disentangle head size from the number of heads $h$, as opposed to the common practice that sets the head size to be $D_m/h$. It is observed empirically that setting attention head size to be input sequence length is beneficial.","id":"be224101-c6fc-4828-9ad0-f1f79c3097ee","level":"subsubsection","origin_cites_number":2,"parent_id":"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5","prefix_titles":[["title","A Survey of Transformers"],["section","Attention"],["subsection","Improved Multi-Head Mechanism"],["subsubsection","Other Modifications"]],"subsections":[],"title":"Other Modifications"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:other_module}","id":"7a323139-7d7b-422d-a6c6-3617457ad2f6","level":"section","origin_cites_number":0,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"]],"subsections":["3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc","ba9ac74a-b1ed-471e-a9a5-9097079a09db","e915275c-0207-4658-9d41-438a8b5b3e51"],"title":"Other Module-level Modifications"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:pos_rep}\n\\begin{definition}[permutation equivariant function]\nLet $\\Pi_n$ be the set of all permutations of indices $\\{1,2,\\cdots, T\\}$. A function $f:\\mathcal{X}^T\\rightarrow \\mathcal{Y}^T$ is said to be \\textit{permutation equivariant} if and only if for any $\\pi\\in\\Pi_T$\n\\begin{equation}\n f(\\pi x)=\\pi f(x).\n\\end{equation}\n\\end{definition}\nIt is easy to verify that Convolution and Recurrence networks are not permutation equivariant. However, both self-attention modules and position-wise feed-forward layers in Transformer are permutation equivariant, which could be a problem when it comes to modeling problems other than \\textit{set-input} problems where the structure of inputs is needed. For example, when modeling sequences of text, the ordering of words matters and it's thus crucial to properly encode the positions of words in Transformer architecture. Therefore, additional mechanisms are required to inject positional information into Transformers. A common design is to first represent positional information using vectors and then infuse the vectors to the model as an additional input.","id":"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc","level":"subsection","origin_cites_number":0,"parent_id":"7a323139-7d7b-422d-a6c6-3617457ad2f6","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position Representations"]],"subsections":["ccc63091-167f-4bc8-97e9-c63a8adbf741","9cdcf5bd-ef00-4ae9-a8b3-6d8cd2240795","268772ad-4cd8-43e9-9608-b7ac7de2e35c","866b7a57-7c1b-41c7-8a8f-c03a0681cbeb","90d9b719-8d67-48fa-a44c-ce04b5d77df6"],"title":"Position Representations"},{"cite_extract_rate":0.555555555555555,"cites":[1493,790,7,1450,38],"content":"In vanilla Transformer~, positional information is encoded as absolute sinusoidal position encodings.For each position index $t$, the encoding is a vector $\\bp_t = \\mathrm{PE}(t)\\in\\mathbb{R}^{D_m}$, of which every element is a sinusoidal ($\\sin$/$\\cos$) function of the index with pre-defined frequency.\n\\begin{equation}\\label{eq:PE}\n \\mathrm{PE}(t)_i = \\begin{cases} \\sin(\\omega_i t) & \\text{if }i\\text{ is even},\\\\\n \\cos(\\omega_i t) & \\text{if }i\\text{ is odd},\n \\end{cases}\n\\end{equation}\nwhere $\\omega_i$ is the hand-crafted frequency for each dimension. The position encoding of each position in the sequence is then added to the token embeddings and fed to Transformer.\nAnother way of representing absolute positions is to learn a set of positional embeddings for each position~. Compared to hand-crafted position representation, learned embeddings are more flexible in that position representation can adapt to tasks through back-propagation. But the number of embeddings is limited up to a maximum sequence length determined before training, which makes this approach no longer \\textit{inductive}, i.e., not able to handle sequences longer than sequences seen in the training time.\n propose to use sinusoidal position representation, but with each frequency $\\omega_i$ (in Eq. \\eqref{eq:PE}) learned from data. This approach retains inductiveness but is more flexible than hand-crafted sinusoidal encoding. FLOATER~ frames positional representation as a continuous dynamical system and adopts Neural ODE to enable end-to-end training with backpropagation. This method is inductive and flexible while being parameter efficient compared to a fully learnable approach.\nThe Vanilla approach to incorporating absolute position representations is to add position encodings/embeddings to token embeddings. However, as the input signals propagate through\nthe layers, the positional information might get lost in the upper layers. Later works find it beneficial to add position representations to inputs to each Transformer layer~.","id":"ccc63091-167f-4bc8-97e9-c63a8adbf741","level":"subsubsection","origin_cites_number":9,"parent_id":"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position Representations"],["subsubsection","Absolute Position Representations"]],"subsections":[],"title":"Absolute Position Representations"},{"cite_extract_rate":0.8571428571428571,"cites":[1456,1508,1481,7370,7054,9],"content":"Another line of works focuses on representing positional relationships between tokens instead of positions of individual tokens. The intuition is that in self-attention, pairwise positional relationships between input elements (direction and distance) could be more beneficial than positions of elements. Methods following this principles are called relative positional representation. propose to add a learnable relative position embedding to keys of attention mechanism\n\\begin{align}\n \\bk_j' &= \\bk_j + \\br_{ij},\\ \\text{for }i=1,\\cdots, n,\\\\\n \\br_{ij} &= \\bR_{\\mathrm{clip}(i-j)}\\label{rij},\\\\\n \\mathrm{clip}(x) &= \\max(-K, \\min(x, K))\\label{eq:shaw_clip},\n\\end{align}\nwhere $\\br_{ij}\\in\\mathbb{R}^{D_k}$ is the relative position embedding for relation between position $i$ and $j$ and $K$ is the largest offset that determines the number of embeddingg. Typically $K$ is set to a length that can accommodate most input sequences. As a special case, InDIGO~ sets $K$ to $3$ for their specially designed framework for non-autoregressive generation. As an incremental effort, Music Transformer~ further introduce a mechanism to reduce the intermediate memory requirements for this approach. Similar to this approach, T5~ adopt a simplified form of relative position embeddings where each embedding is only a learnable scalar that is added to the corresponding score used for computing the attention weights.\nTransformer-XL~ use a sinusoidal encoding to represent positional relationships but fuses contents and position information by redesign the computation of attention scores\\footnote{the scaling factor is omitted without loss of generality.}\n\\begin{equation}\n \\bA_{ij}=\\bq_i \\bk_j^\\top+\\bq_i\\left(\\bR_{i-j}\\bW^{K,R}\\right)^\\top +\\bu^1\\bk_j^\\top+\\bu^2\\left(\\bR_{i-j}\\bW_{K,R}\\right)^\\top,\n\\end{equation}\nwhere $\\bW^{K,R}\\in\\mathbb{R}^{D_m\\times D_k},\\bu^1,\\bu^2\\in\\mathbb{R}^{D_k}$ are learnable parameters and $\\bR$ is a sinusoidal encoding matrix similar to position encoding in vanilla Transformer. Then softmax function is applied to scores $\\bA$ to provide attention weights. Note that the learnable sinusoidal encoding is also a drop-in replacement to hand-crafted $\\bR$.\nDeBERTa~ utilizes position embeddings like and applies the embeddings to the model in a disentangled style similar to Transformer-XL~\n\\begin{equation}\n \\bA_{ij}=\\bq_i \\bk_j^\\top+\\bq_i\\left(\\br_{ij}\\bW^{K,R}\\right)^\\top +\\bk_j\\left(\\br_{ij}\\bW^{Q,R}\\right)^\\top,\n\\end{equation}\nwhere $\\bW^{K,R},\\bW^{Q,R}\\in\\mathbb{R}^{D_m\\times D_k}$ are learnable parameters and $\\br_{ij}$ is the learnable relative positional embedding as in Eq. \\eqref{rij}. The first term is interpreted as a content-to-content attention, and the latter two terms are interpreted as (relative) content-to-position and position-to-content attention, respectively.","id":"9cdcf5bd-ef00-4ae9-a8b3-6d8cd2240795","level":"subsubsection","origin_cites_number":7,"parent_id":"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position Representations"],["subsubsection","Relative Position Representations"]],"subsections":[],"title":"Relative Position Representations"},{"cite_extract_rate":1,"cites":[1454,8457],"content":"Some research studies have explored using hybrid positional representations that contains both absolute and relative positional information. Transformer with Untied Position Encoding (TUPE)~ re-designs the computation of attention scores as a combination of a content-to-content term, an absolute position-to-position term and a bias term representing relative positional relationships\n\\begin{equation}\n \\bA_{ij}=\\bq_i \\bk_j^\\top+\\left(\\bp_i\\bW^{Q,P}\\right)\\left(\\bp_{j}\\bW^{K,P}\\right)^\\top +b_{j-i},\n\\end{equation}\nwhere $\\bW^{K,P},\\bW^{Q,P}\\in\\mathbb{R}^{D_m\\times D_k}$ are learnable parameters, $\\bp_i, \\bp_j$ are the position embeddings for positions $i,j$, and $b_{j-i}$ is a learnable scalar relative position embedding.\nOne can also design a single set of positional representations that express both absolute and relative information. Roformer~ uses Rotary Position Embedding (RoPE) to represent the position of a token by multiplying the affine-transformed embedding of the $t$-th input $x_t$ by a rotatory matrix $\\bR_{\\Theta,t}$\n\\begin{align}\n \\bq_t=\\bx_t\\bW^Q\\bR_{\\Theta,t}\\quad & \\bk_t=\\bx_t\\bW^K\\bR_{\\Theta,t},\\\\\n \\bR_{\\Theta,t}&=\\bigoplus_{j=1}^{D_k/2} \\bM(t,\\theta_j),\n\\end{align}\nwhere $\\bigoplus$ denotes \\textit{direct sum} of matrices. Each $\\bM(t,\\theta_j)$ is a 2-D clockwise rotatory matrix of angle $t\\cdot \\theta_j$\n\\begin{equation}\n \\bM(t,\\theta_j)=\\left[\\begin{matrix} \\cos(t\\cdot \\theta_j)&\\sin(t\\cdot \\theta_j)\\\\\n -\\sin(t\\cdot \\theta_j)&\\cos(t\\cdot \\theta_j)\\end{matrix}\n \\right].\n\\end{equation}\nThe key advantage of this formulation is that the induced representation is translation invariant, i.e., the attention score of $(\\bq_i,\\bk_j)$ is only related to their relative position offset\n\\begin{equation}\n \\bq_i\\bk_j^\\top=\\left(\\bx_i\\bW^Q\\right)\\bR_{\\Theta,\\textcolor[rgb]{0,0,1}{j-i}}\\left(\\bx_j\\bW^K\\right)^\\top.\n\\end{equation}\nIn practice, the embedding matrix multiplication can be implemented by two element-wise multiplication for lower memory footprint. The RoPE uses the form of absolute embedding but can capture relative positional relations. This approach is compatible with linearized attention in Sec.~\\ref{sec:kernel}.","id":"268772ad-4cd8-43e9-9608-b7ac7de2e35c","level":"subsubsection","origin_cites_number":2,"parent_id":"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position Representations"],["subsubsection","Other Representations"]],"subsections":[],"title":"Other Representations"},{"cite_extract_rate":1,"cites":[1480,1493,1509,1471],"content":"Instead of explicitly introducing additional positional encodings, propose to encode positional information in word embeddings, by generalizing embedding to continuous (complex-valued) functions over positions.\nR-Transformer~ model locality of sequential data with a local RNN. Specifically, inputs to each block of R-Transformer are first fed to a local RNN and then to multi-Head self-attention module. The RNN structure introduces ordering information and captures local dependencies as a complement to self-attention.\nConditional positional encoding (CPE)~ generate conditional position encodings at each layer for ViT with a 2-D convolution with zero-paddings. The intuition behind this approach is that convolution networks can implicitly encode absolute positional information with zero-paddings~.","id":"866b7a57-7c1b-41c7-8a8f-c03a0681cbeb","level":"subsubsection","origin_cites_number":4,"parent_id":"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position Representations"],["subsubsection","Position Representations without Explicit Encoding"]],"subsections":[],"title":"Position Representations without Explicit Encoding"},{"cite_extract_rate":0.33333333333333304,"cites":[1510],"content":"It is worth noticing that masked self-attention is not permutation equivariant~. Thus a model that exploits only the decoder of Transformer has the potential of sensing positional information without incorporating explicit positional representation. This is confirmed by some empirical results on language modeling tasks~, where the authors find that removing position encodings even improves performance.","id":"90d9b719-8d67-48fa-a44c-ce04b5d77df6","level":"subsubsection","origin_cites_number":3,"parent_id":"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position Representations"],["subsubsection","Position Representation on Transformer Decoders"]],"subsections":[],"title":"Position Representation on Transformer Decoders"},{"cite_extract_rate":0,"cites":[],"content":"Layer Normalization (LN), along with residual connection, is considered as a mechanism to stabilizing training of deep networks (e.g., alleviating ill-posed gradients and model degeneration).\nThere are some studies that are dedicated to analyzing and improving LN module.","id":"ba9ac74a-b1ed-471e-a9a5-9097079a09db","level":"subsection","origin_cites_number":0,"parent_id":"7a323139-7d7b-422d-a6c6-3617457ad2f6","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Layer Normalization"]],"subsections":["9c98bd5d-1c97-4d09-acae-d8ba70e8c480","0329ae93-9d4a-453b-86ad-768aed04ebd3","29a8b774-a21f-46a2-8095-eac026b86ef8"],"title":"Layer Normalization"},{"cite_extract_rate":1,"cites":[1496,1511,1448,793,1464,38],"content":"\\ifx \\smv \\undefined\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[post-LN]{\n\\includegraphics[height=2.2in]{assets/layernorm/post_ln.pdf}\n}\\hspace{.6in}\n\\subfigure[pre-LN]{\n\\includegraphics[height=2.2in]{assets/layernorm/pre_ln.pdf}\n}\n\\caption{Comparison of Transformer Encoder with pre-LN and post-LN.}\\label{fig:pre_post_ln}\n\\end{center}\n\\end{figure}\n\\fi\nIn vanilla Transformer, the LN layer lies between the residual blocks, called post-LN~. Later Transformer implementations~ place the LN layer inside the residual connection before the attention or FFN, with an additional LN after the final layer to control the magnitude of final outputs, which is referred to as pre-LN\\footnote{To the best of our knowledge, this approach is adopted since \\texttt{v1.1.7} in the \\texttt{Tensor2Tensor} implementation~.}. The pre-LN has been adopted by numerous following research studies and implementations, e.g., .\nThe difference between pre-LN and post-LN is shown in Fig. \\ref{fig:pre_post_ln}.\n theoretically investigate the gradients of Transformers and find that the gradients near the output layer are large at initialization in post-LN Transformers, which could be the reason why post-LN Transformers without learning rate warm-up~\\footnote{Learning rate warm-up refers to starting optimization with an extremely small learning rate and then gradually increasing it to a pre-defined maximum value in a certain number of iterations.} leads to unstable training, whereas pre-LN Transformers do not suffer from the same problem. They thus deduce and empirically verify that warm-up stage can be safely removed for pre-LN Transformers.\nAlthough Post-LN often results in unstable training and divergence, it usually outperforms pre-LN variants after convergence~. Similar to , conduct theoretical and empirical analysis and find that post-LN encoders do not suffer from gradient imbalance. They thus conjecture that the gradient issue is not the direct cause of unstable post-LN Transformer training and further identify the \\textit{amplification effect} in post-LN Transformers — at initialization, the heavier dependency on residual branch leads to a larger output shift in post-LN Transformers, thus resulting in unstable training. In light of this finding, they introduce additional parameters to post-LN Transformers to control residual dependencies of Post-LN. These parameters are initialized according to activation variations of sample data so that the output shift of post-LN Transformers is not amplified. This approach ensures and boosts convergence of post-LN Transformers and reaches better performance than pre-LN Transformers.","id":"9c98bd5d-1c97-4d09-acae-d8ba70e8c480","level":"subsubsection","origin_cites_number":8,"parent_id":"ba9ac74a-b1ed-471e-a9a5-9097079a09db","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Layer Normalization"],["subsubsection","Placement of Layer Normalization"]],"subsections":[],"title":"Placement of Layer Normalization"},{"cite_extract_rate":1,"cites":[1466,1467,1491,71],"content":" empirically observe that the learnable parameters in the LN module do not work in most experiments, and even increase the risk of overfitting. They further conclude from controlled experiments that the forward normalization is not the reason why LN works for Transformer. From analysis and experiments, it is concluded that the derivatives of the mean and variance re-center and re-scale the gradients and play a significant role in LN. They thus propose \\textit{AdaNorm}, a normalization technique without learnable parameters\n\\begin{align}\n \\bz &= C(1-k\\by)\\odot\\by,\\\\\n \\by &=\\frac{\\bx-\\mu}{\\sigma},\n\\end{align}\nwhere $C,k$ are hyperparameters and $\\odot$ denotes element-wise multiplication. $\\mu$ and $\\sigma$ are the mean and standard deviation of input $\\bx$, respectively.\n propose to replace the LN module with \\textit{scaled $\\ell_2$ normalization}. Given any input $\\bx$ of $d$-dimension, their approach project it onto a $d-1$-sphere of learned radius $g$\n\\begin{equation}\n \\bz=g\\frac{\\bx}{\\|\\bx\\|},\n\\end{equation}\nwhere $g$ is a learnable scalar. It is more parameter efficient compared to normal LN and is shown to be effective in machine translation datasets, especially in low-resource settings.\n discuss why Batch Normalization (BN)~ performs poorly in Transformer for text data and conclude that BN's significant performance degradation stems from the instabilities associated with its batch statistics. They thus propose PowerNorm (PN) that has three modifications over BN: (1) it relaxes the zero-mean normalization; (2) it uses the quadratic mean\nof the signal, instead of the variance; (3) it uses running statistics for the\nquadratic mean, instead of using per-batch statistics. Specifically, for the $t$-th iteration, the PN computes the outputs as\n\\begin{align}\n \\bz^{(t)} &= \\gamma\\odot \\by^{(t)}+\\beta,\\\\\n \\by^{(t)} &=\\frac{\\bx^{(t)}}{\\psi^{(t-1)}},\\\\\n (\\psi^{(t)})^2&=\\alpha(\\psi^{(t-1)})^2+(1-\\alpha)\\left(\\frac{1}{|B|}\\sum_{i=1}^{|B|}(\\bx_i^{(t)})^2\\right),\n\\end{align}\nwhere $0<\\alpha<1$ is the moving average coefficient and $\\gamma, \\beta$ are the learnable parameters as in BN formulation.","id":"0329ae93-9d4a-453b-86ad-768aed04ebd3","level":"subsubsection","origin_cites_number":4,"parent_id":"ba9ac74a-b1ed-471e-a9a5-9097079a09db","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Layer Normalization"],["subsubsection","Substitutes of Layer Normalization"]],"subsections":[],"title":"Substitutes of Layer Normalization"},{"cite_extract_rate":1,"cites":[1495],"content":"Besides LN, there is another mechanism to construct deeper neural network. ReZero~ replace LN module with a learnable residual connection. For each module $F(\\cdot)$, ReZero re-scales $F(\\cdot)$ in the residual formulation:\n\\begin{equation}\n \\bH'=\\bH+\\alpha\\cdot F(\\bH),\n\\end{equation}\nwhere $\\alpha$ is a learnable parameter with zero-initialization.\nReplacing LN in Transformer with ReZero mechanism is verified to induce better dynamic isometry for input signals and leads to faster convergence.","id":"29a8b774-a21f-46a2-8095-eac026b86ef8","level":"subsubsection","origin_cites_number":1,"parent_id":"ba9ac74a-b1ed-471e-a9a5-9097079a09db","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Layer Normalization"],["subsubsection","Normalization-free Transformer"]],"subsections":[],"title":"Normalization-free Transformer"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:ffn}\nDespite its simplicity, the position-wise feed-forward network (FFN) layers are important for a Transformer to achieve good performance. observe that simply stacking self-attention modules causes a \\textit{rank collapse} problem, leading to token-uniformity inductive bias, and that the feed-forward layer is one of the important building blocks that mitigate this issue.\nVarious works have explored modifications on the FFN module.","id":"e915275c-0207-4658-9d41-438a8b5b3e51","level":"subsection","origin_cites_number":1,"parent_id":"7a323139-7d7b-422d-a6c6-3617457ad2f6","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position-wise FFN"]],"subsections":["d96aa288-77f3-4bf6-931e-abd30aeafccb","e9f737ba-3b14-4121-b536-765cb8c260d1","d1496356-82a7-4f73-8a03-62f59057e605"],"title":"Position-wise FFN"},{"cite_extract_rate":0.875,"cites":[1481,1512,7055,1455,7,7365,38],"content":"The vanilla Transformer~ adopts the Rectified Linear Units (ReLU) activation for non-linearity in between the two FFN layers. Over time, several studies have explored different activation other than ReLU.\n try to replace ReLU in Transformer with Swish function $f(x)=x\\mathrm{sigmoid}(\\beta x)$ and observe that it consistently improve performance on WMT 2014 English$\\rightarrow$German dataset.\nGPT~ replace ReLU with Gaussian Error Linear Unit (GELU)~ on language pre-training. It becomes the default practice for many pre-trained language models ~.\n explore using Gated Linear Units (GLU)~ and its variants as a drop-in replacement for ReLU in FFN. Their pre-training experiments show that the GLU variants consistently improve vanilla Transformer with ReLU activation. Note that GLU introduces extra parameters and the experiments are conducted with the intermediate dimension of FFN reduced to match the parameter count with baseline.","id":"d96aa288-77f3-4bf6-931e-abd30aeafccb","level":"subsubsection","origin_cites_number":8,"parent_id":"e915275c-0207-4658-9d41-438a8b5b3e51","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position-wise FFN"],["subsubsection","Activation Function in FFN"]],"subsections":[],"title":"Activation Function in FFN"},{"cite_extract_rate":0.833333333333333,"cites":[680,1490,1451,8454,707],"content":"Several works have focused on expanding FFNs in order for a larger model capacity. The basic idea is to replace FFNs with similar structures with much more parameters.\n replace some of the FFNs with the product-key memory layers. A product-key memory is composed of three components: a query network, a key selection module containing two\nsets of sub-keys, and a value lookup table. The model first projects an input to a latent space using the query network, and then compares the generated query to keys that are Cartesian product of the two sets of sub-keys from key selection module to get $k$ nearest neighbors, and finally finds the corresponding values in a value lookup table using the $k$ nearest keys and aggregates them to produce the final output. This process resembles the attention mechanism, in that the generated query attends to a large number of global key-value pairs. They thus propose a multi-head mechanism for the key-product memory to further enlarge the capacity of this module. The experiments on large-scale language modeling suggest that this mechanism significantly improves performance with negligible computational overhead.\nSeveral studies exploits the idea of Mixture-of-Experts (MoE) to increase the capacity of FFNs. Gshard uses sparsely-gated MoE layers to replace FFNs in Transformer. Each MoE layer consists of several FFNs (each called an expert) that are the same structure as position-wise FFNs in vanilla Transformer. The output of the layer is a weighted sum of the outputs of the FFNs, using gate values computed by a routing function $g(\\cdot)$. They design a learnable routing function that assigns tokens to experts, with auxiliary loss to satisfy balanced loads between experts and efficiency at the scale of length such that the experts can be distributed across multiple devices. For each forward pass of the MoE layer, only the experts with top-$k$ gate values are activated.\nInstead of using $k$ experts for each forward pass, Switch Transformer~ proposes to route using only a single expert with the largest gate value, leading to a much smaller computational footprint. The authors also design an auxiliary loss to encourage load balance between experts. It is reported to speed up pre-training by a large margin compared to the non-MoE counterpart while having a similar number of FLOPS.\n propose to replace top-$k$ routing with expert prototyping strategy. Specifically, the proposed strategy splits experts into $k$ different groups and applies top-1 routing within each group. The outputs of prototype groups are combined linearly to form the final output of the MoE layer. This strategy is proved to improve the model quality while maintaining constant computational costs.\nAs opposed to using a learnable routing function for expert assignment, design \\textit{hash layers} where tokens are hashed into a fixed number of buckets, each bucket corresponding to an expert. This approach requires no routing parameters or any auxiliary loss function, while showing competitive results with existing methods such as Switch Transformer~.","id":"e9f737ba-3b14-4121-b536-765cb8c260d1","level":"subsubsection","origin_cites_number":6,"parent_id":"e915275c-0207-4658-9d41-438a8b5b3e51","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position-wise FFN"],["subsubsection","Adapting FFN for Larger Capacity"]],"subsections":[],"title":"Adapting FFN for Larger Capacity"},{"cite_extract_rate":1,"cites":[1463,7053],"content":"Notably, one might argue that under some circumstances, FFN layers can be dropped completely, resulting in a simplified network.\n demonstrate that replacing the ReLU activation with Softmax and dropping the bias term in FFN effectively turns FFN into an attention module where position-wise inputs attend to a global key-value memory of $D_{\\mathrm{ffn}}$ slots. They thus propose to drop the FFN module and add to the attention module a set of global key-value pairs, which are learnable parameters concatenated with key and values generated by inputs. This approach simplifies the structure of the network with no loss of performance.\n empirically show that FFNs in the decoder of Transformer, despite its large number of parameters, is not efficient and can be removed safely with only slight or no loss of performance. This approach significantly boosts the training and inference speed.","id":"d1496356-82a7-4f73-8a03-62f59057e605","level":"subsubsection","origin_cites_number":2,"parent_id":"e915275c-0207-4658-9d41-438a8b5b3e51","prefix_titles":[["title","A Survey of Transformers"],["section","Other Module-level Modifications"],["subsection","Position-wise FFN"],["subsubsection","Dropping FFN Layers"]],"subsections":[],"title":"Dropping FFN Layers"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:beyond}\nIn this section, we introduce the X-formers that modify the vanilla Transformer beyond modules.","id":"04795308-eff0-4fbb-962f-a8077d36f894","level":"section","origin_cites_number":0,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"]],"subsections":["f397ed71-a760-4154-8eaa-b1b20d5004b2","736cefee-90d8-4e53-a49e-f031442dc1ba","690a01cd-27c7-4e1a-baef-83bbdeff8220","9b2cdbdc-bac6-48ec-96d0-b5576360be13","be720c58-062b-41bd-b6e9-732058f0f671"],"title":"Architecture-level Variants"},{"cite_extract_rate":0.5,"cites":[1468,1488],"content":"Apart from the efforts made at the module level to alleviate computation overheads, there are several attempts to adapt Transformer to be lightweight by modifications at a higher level.\nSimilar to low-rank self-attention~ that decomposes attention into a locality-constrained attention and a low-rank global attention, Lite Transformer~ proposes to replace each attention module in Transformer with a two-branch structure, where one branch uses attention to capture long-range contexts while the other branch uses depth-wise convolution and linear layers to capture local dependencies. The architecture is lightweight both in terms of model size and computation, and is thus more suitable for mobile devices.\nFunnel Transformer~ utilizes a funnel-like encoder architecture where the length of the hidden sequence is gradually reduced using pooling along the sequence dimension, and then recovered using up-sampling. The architecture effectively reduces the FLOPs and memory compared to the vanilla Transformer encoder. Naturally, one can use this architecture to build a deeper or wider model using the same computation resources.\nDeLighT~ replaces the standard Transformer block with \\texttt{DeLighT} block, which consists of three sub-modules: (1) a ``expand-and-reduce'' \\texttt{DeLighT} transformation module to learn wider representations with low computation requirements; (2) a single-head self-attention to learn pair-wise interaction; (3) a lightweight ``reduce-and-expand'' FFN (as opposed to vanilla Transformer that first expands the dimension of hidden representations and then reduces them back to $D_m$). They also propose a block-wise scaling strategy that allows for shallower and narrower blocks near the input and wider and deeper blocks near the output. The induced network is much deeper than the vanilla Transformer but with fewer parameters and operations.","id":"f397ed71-a760-4154-8eaa-b1b20d5004b2","level":"subsection","origin_cites_number":4,"parent_id":"04795308-eff0-4fbb-962f-a8077d36f894","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Adapting Transformer to Be Lightweight"]],"subsections":[],"title":"Adapting Transformer to Be Lightweight"},{"cite_extract_rate":0.75,"cites":[1457,1486,1492],"content":"In vanilla Transformer, each block takes outputs from the previous block as inputs and outputs a sequence of hidden representations. One might be interested in creating more paths along which input signals can run through the networks. In Sec.~\\ref{sec:prior_prev_module}, we introduced Realformer~ and Predictive Attention Transformer~ that reuses attention distributions from previous block to guide attention of current block. This can be seen as creating a forward path between adjacent Transformer blocks.\nIn a deep Transformer encoder-decoder model, the cross-attention modules in the decoder only utilize the final outputs of the encoder, therefore the error signal will have to traverse along the depth of the encoder. This makes Transformer more susceptible to optimization issues (e.g., vanishing gradients). Transparent Attention~ uses a weighted sum of encoder representations at all encoder layers (including the embedding layer) in each cross-attention module. For the $j$-th decoder block, the cross-attention module is modified to attend to\n\\begin{equation}\\label{eq:transparent}\n \\tilde{\\bH}^{(j)}=\\sum_{i=0}^{N}\\frac{\\exp(w_{ij})}{\\sum_{k=0}^N \\exp(w_{kj})}\\bH^{(i)},\n\\end{equation}\nwhere each $w_{ij}$ is a trainable parameter. This effectively shortens the path from each layer in the encoder to the error signal and thus eases the optimization of deeper Transformer models.\nAnother issue associated with vanilla Transformer is that each position can only attend to history representations from lower layers. Feedback Transformer~ proposes to add a feedback mechanism to Transformer decoder, where each position attends to a weighted sum of history representations from all layers\n\\begin{equation}\n \\tilde{\\bh}_i=\\sum_{l=0}^{N}\\frac{\\exp(w_{l})}{\\sum_{k=0}^N \\exp(w_{k})}\\bh_i^{(l)}.\n\\end{equation}","id":"736cefee-90d8-4e53-a49e-f031442dc1ba","level":"subsection","origin_cites_number":4,"parent_id":"04795308-eff0-4fbb-962f-a8077d36f894","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Strengthening Cross-Block Connectivity"]],"subsections":[],"title":"Strengthening Cross-Block Connectivity"},{"cite_extract_rate":0.833333333333333,"cites":[683,7273,1485,1449,1478],"content":"Vanilla Transformer, like most neural models, utilizes a fixed (learned) computation procedure to process each input. An intriguing and promising modification is to make computation time conditioned on the inputs, i.e., to introduce Adaptive Computation Time (ACT)~ into Transformer models. Such modifications potentially give rise to the following advantages:\n\\begin{itemize}\n \\item Feature refinement for hard examples. For data that are hard to process, a shallow representation might not be adequate to fulfill the task at hand. It would be more ideal to apply more computations to acquire a deeper and more refined representation.\n \\item Efficiency for easy examples. When processing easy examples, a shallow representation might be enough for the task. In this case, it would be beneficial if the network can learn to extract features using reduced computation time.\n\\end{itemize}\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[dynamic halting]{\\label{fig:ut}\\hspace{2em}\n\\includegraphics[height=1.2in]{assets/ACT/ut.pdf}\\hspace{2em}\n}\n\\subfigure[conditional skipping]{\\label{fig:skip}\\hspace{2em}\n\\includegraphics[height=1.2in]{assets/ACT/skip.pdf}\\hspace{2em}\n}\n\\subfigure[early exit]{\\label{fig:exit}\\hspace{2em}\n\\includegraphics[height=1.2in]{assets/ACT/exit.pdf}\n}\n\\caption{Three typical ACT paradigms.}\\label{fig:act}\n\\end{center}\n\\end{figure}\nUniversal Transformer (UT)~ incorporates a recurrence-over-depth mechanism that iteratively refines representations for all symbols using a module that is shared over depth, as illustrated in Fig. \\ref{fig:ut}. It also adds a per-position dynamic halting mechanism that calculates a halting probability for each symbol at every time step. If a symbol's halting probability is greater than a predefined threshold, then the symbol's representation will remain unchanged for subsequent timesteps. The recurrence is stopped when all symbols halt or when a predefined maximum step is reached.\nConditional Computation Transformer (CCT)~ adds a gating module at each self-attention and feed-forward layer to decide whether to skip the current layer, as illustrated in Fig. \\ref{fig:skip}. The authors also introduce an auxiliary loss that encourages the model to adjust the gating modules to match the practical computation cost to the available computation budget.\nSimilar to the dynamic halting mechanism used in UT, there is a line of work dedicated to adapting the number of layers to each input in order to achieve a good speed-accuracy trade-off, which is called \\textit{early exit} mechanism, as illustrated in Fig. \\ref{fig:exit}. A commonly used technique is to add an internal classifier at each layer and jointly train all classifiers. The core of these methods is the criteria used to decide whether to exit at each layer. DeeBERT~ uses the entropy of the output probability distribution of the current layer to determine whether to exit. PABEE~ counts the number of times that the predictions remain unchanged to decide whether to exit. design a window-based uncertainty criterion to achieve token-level partial exiting for sequence labeling tasks. introduces a voting-based exiting strategy that considers at each layer predictions of all the past internal classifiers to infer the correct label and to decide whether to exit.","id":"690a01cd-27c7-4e1a-baef-83bbdeff8220","level":"subsection","origin_cites_number":6,"parent_id":"04795308-eff0-4fbb-962f-a8077d36f894","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Adaptive Computation Time"]],"subsections":[],"title":"Adaptive Computation Time"},{"cite_extract_rate":0,"cites":[],"content":"The quadratic complexity of self-attention on sequences length can significantly limit the performance\nof some downstream tasks. For example, language modeling usually needs long-range context. Apart from the techniques introduced in Sec.~\\ref{sec:attention}, another effective way of dealing with long sequences is to use \\textit{divide-and-conquer} strategy, i.e., to decompose an input sequence into finer segments that can be efficiently processed by Transformer or Transformer modules. We identify two representative class of methods, \\textit{recurrent} and \\textit{hierarchical} Transformers, as illustrated in Fig. \\ref{fig:divide_and_conquer}. These techniques can be understood as a wrapper for the Transformer model in which Transformer acts as an elementary component that is reused to process different input segments.\n\\begin{figure}[htbp]\n\\begin{center}\n\\subfigure[Recurrent Transformer\\label{fig:recurrence}]{\n\\includegraphics[width=0.4\\linewidth]{assets/divide_and_conquer/recurrence.pdf}\n}\\quad\n\\subfigure[Hierarchical Transformer\\label{fig:hierarchy}]{\n\\includegraphics[width=0.4\\linewidth]{assets/divide_and_conquer/hierarchy.pdf}\n}\n\\caption{Illustrations of recurrent and hierarchical Transformers.}\\label{fig:divide_and_conquer}\n\\end{center}\n\\end{figure}","id":"9b2cdbdc-bac6-48ec-96d0-b5576360be13","level":"subsection","origin_cites_number":0,"parent_id":"04795308-eff0-4fbb-962f-a8077d36f894","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Transformers with Divide-and-Conquer Strategies"]],"subsections":["88924fec-4dc9-4d83-b667-d6862e359751","6244ba7c-c68b-4615-be13-f855c4dbca70"],"title":"Transformers with Divide-and-Conquer Strategies"},{"cite_extract_rate":0.5,"cites":[1500,7370,1487],"content":"In recurrent Transformers, a cache memory is maintained to incorporate the history information. While processing a segment of text, the network reads from the cache as an additional input. After the processing is done, the network writes to the memory by simply copying hidden states or using more complex mechanisms. The abstract process is illustrated in Fig. \\ref{fig:recurrence}.\nTransformer-XL~ address the limitation of a fixed length context by caching representations from the previous segment and reuse it as an extended context when the model processes the current segment. For the $l$-th layer and the $(\\tau+1)$-th segment, the input representation $\\bH_{\\tau+1}^{(l-1)}$ is concatenated with the representation $\\bH_\\tau^{(l-1)}$ from previous segment to produce the keys and values\n\\begin{align}\n \\tilde{\\bH}_{\\tau+1}^{(l)}&=[\\mathrm{SG}(\\bH_\\tau^{(l-1)})\\circ \\bH_{\\tau+1}^{(l-1)}]\\label{eq:xformerxl},\\\\\n \\bK_{\\tau+1}^{(l)}, \\bV_{\\tau+1}^{(l)}&=\\tilde{\\bH}_{\\tau+1}^{(l)}\\bW^K, \\tilde{\\bH}_{\\tau+1}^{(l)}\\bW^V,\n\\end{align}\nwhere $\\bH_\\tau^{(0)}$ is defined as the word embedding sequence, $\\mathrm{SG}(\\cdot)$ denotes stop-gradient operation and $[\\bX\\circ\\bY]$ denotes concatenating the two vector sequences along the time dimension. This approach extends the maximum context length by $L\\times N_{\\text{mem}}$ where $L$ is the number of layers and $N_{\\text{mem}}$ is the length of cached memory sequence.\nCompressive Transformer~ extends this idea further by extending the cache with two levels of memory. In Transformer-XL, the activations from the previous segment are cached as a memory that is used to augment the current segment, and activations from older segments are discarded. Compressive Transformer, on the other hand, applies a compression operation (e.g., Convolution, Pooling, etc.) on older activations and stores them in the compressed memory. In order to avoid the expensive backpropagating-through-time (BPTT) from training compression sub-network with gradients from the loss, they propose to use local loss functions where original memories are constructed from the compressed memories. This approach further extends the theoretical maximum history context length from $L\\times N_{\\text{mem}}$ of Transformer-XL to $L\\times (N_{\\text{mem}}+c\\times N_{\\text{cm}})$, where $c$ is the compression rate and $N_{\\text{cm}}$ is the length of compressed memory.\n Memformer~ extends the recurrence mechanism from decoder-only architecture to an encoder-decoder architecture. They introduce to the encoder a memory cross attention similar to the cross attention in vanilla Transformer to allow the Transformer encoder to attend to the memory. They also introduce a memory slot attention on top of the encoder output to explicitly write the memory for the next segment. To avoid BPTT over a long range of timesteps, they propose Memory\nReplay Back-Propagation (MRBP) algorithm, which replays\nthe memory at each timestep to accomplish gradient back-propagation over long unrolls.\n propose a simple fine-tuning mechanism to add recurrence to a pre-trained language model (e.g., GPT-2~).\n\\ifx \\smv \\undefined\nThey first compress the representations produced by the $\\tau$-th segment into one single vector representation, using a weighted average of pooled representations from each layer $l\\in\\{1,\\cdots,L\\}$\n\\begin{equation}\n \\bz_\\tau=\\sum_{l=1}^Lw_l\\sum_{j=1}^{T_\\tau}\\bh_j^{(l)},\n\\end{equation}\nwhere $T_\\tau$ denotes the sequence length of the $\\tau$-th segment, $w_l=\\mathrm{softmax}(\\mathbf\\alpha)_l$ is the weight softmax-normalized from learnable parameters $\\mathbf\\alpha=[\\alpha_1,\\cdots,\\alpha_L]$. This compressed representation is then fed to a feed-forward network to produce the memory state $\\bh_{\\text{prev},\\tau}$ for the $\\tau$-th segment, which is then prepended to the key-value inputs of a specific attention layer. This approach effectively extends the context length of a pre-trained language model, without significant change of the architecture of the original model.\n\\fi\nERNIE-Doc~ proposes an enhanced recurrence mechanism based on the recurrence mechanism used in Transformer-XL, by replacing the memory with the history representations from the $l$-th layer.\n\\ifx \\smv \\undefined\n\\begin{align}\n \\tilde{\\bH}_{\\tau+1}^{(l)}&=[\\mathrm{SG}(\\textcolor[rgb]{0,0,1}{\\bH_\\tau^{(l)}})\\circ \\bH_{\\tau+1}^{(l-1)}],\n\\end{align}\nas opposed to using representations from the $(l-1)$-th layer in Eq. \\eqref{eq:xformerxl}. This modification essentially leads to a larger effective context length.\n\\fi","id":"88924fec-4dc9-4d83-b667-d6862e359751","level":"subsubsection","origin_cites_number":6,"parent_id":"9b2cdbdc-bac6-48ec-96d0-b5576360be13","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Transformers with Divide-and-Conquer Strategies"],["subsubsection","Recurrent Transformers"]],"subsections":[],"title":"Recurrent Transformers"},{"cite_extract_rate":0,"cites":[],"content":"Hierarchical Transformer decomposes inputs hierarchically into elements of finer granularity. Low-level features are first fed to a Transformer encoder, producing output representations that are then aggregated (using pooling or other operations) to form a high-level feature, which is then processed by a high-level Transformer. This class of methods can be understood as a process of hierarchical abstraction. The overview of this approach is depicted in Fig. \\ref{fig:hierarchy}. The advantages of this approach are twofold: (1) Hierarchical modeling allows the model to handle long inputs with limited resources; (2) It has the potential to generate richer representations that are beneficial to tasks.","id":"6244ba7c-c68b-4615-be13-f855c4dbca70","level":"subsubsection","origin_cites_number":0,"parent_id":"9b2cdbdc-bac6-48ec-96d0-b5576360be13","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Transformers with Divide-and-Conquer Strategies"],["subsubsection","Hierarchical Transformers"]],"subsections":["10bc35f7-356c-4cac-a874-ee859573c18e","153b332f-732b-4172-8553-f1c9626b599f"],"title":"Hierarchical Transformers"},{"cite_extract_rate":0.6666666666666661,"cites":[1482,1477],"content":"For tasks with inherently long input length, one can use hierarchical Transformers for effective modeling of long-range dependencies. For document-level machine translation tasks, introduce dependencies on the previous sentences from both the source and target sides when translating a sentence. They use an attention mechanism as the aggregation operation to summarize low-level information. For document summarization, HIBERT~ encodes a document of text by first learn sentence representations for all sentences and then use these sentence representations to encode document-level representations that are then used to generate the summary. The model uses the last hidden representation (corresponding to the \\texttt{EOS} token) as the representation for each sentence. propose a similar hierarchical Transformer for multi-document summarization where the extracted low-level representations are aggregated using an attention layer with a global trainable query node and low-level representations as the source of key-value pairs. Hi-Transformer~ first utilizes a sentence Transformer and a document Transformer to hierarchically learn document context-aware sentence representations. The document context-aware sentence representations are then fed to another sentence Transformer to further improve the sentence context modeling.","id":"10bc35f7-356c-4cac-a874-ee859573c18e","level":"paragraph","origin_cites_number":3,"parent_id":"6244ba7c-c68b-4615-be13-f855c4dbca70","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Transformers with Divide-and-Conquer Strategies"],["subsubsection","Hierarchical Transformers"],["paragraph","6.5.2.1 Hierarchical for long sequence inputs"]],"subsections":[],"title":"6.5.2.1 Hierarchical for long sequence inputs"},{"cite_extract_rate":1,"cites":[732,7368,7367],"content":"One might also be interested in using hierarchical models to acquire richer representations that are beneficial to the tasks at hand. For example, TENER~ uses a low-level Transformer encoder to encode character features, which is then concatenated with word embeddings as the inputs to the high-level Transformer encoder. This incorporates more features and alleviates the problems of data sparsity and out-of-vocabulary (OOV). Recently emerging Vision Transformer~ divides an input image into several patches that serve as the basic input elements of Transformer, which potentially loses intrinsic pixel-level information within patches. To address this issue, Transformer in Transformer (TNT)~ uses at each layer an inner Transformer block that transforms pixel representations and an outer Transformer block that takes fused vectors of patch representations and pixel representations as input.","id":"153b332f-732b-4172-8553-f1c9626b599f","level":"paragraph","origin_cites_number":3,"parent_id":"6244ba7c-c68b-4615-be13-f855c4dbca70","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Transformers with Divide-and-Conquer Strategies"],["subsubsection","Hierarchical Transformers"],["paragraph","6.5.2.2 Hierarchical for richer representations"]],"subsections":[],"title":"6.5.2.2 Hierarchical for richer representations"},{"cite_extract_rate":1,"cites":[8456,1469,544,1475,1483,1461],"content":"Despite the success of Transformer architecture, one might question whether the current Transformer architecture is optimal. Interestingly, several studies have explored alternative architectures for Transformer.\n interpret Transformer as a numerical Ordinary Differential Equation (ODE) solver for a convection-diffusion equation in a multi-particle dynamic system and design Macaron Transformer, which replaces each Transformer block with a \\textit{FFN-attention-FFN} variant.\nSandwich Transformer~ explores reorganizing attention modules and FFN modules such that attention modules are mainly located in lower layers and FFN modules in upper layers. The induced model improves perplexity on multiple language modeling benchmarks, without increasing parameters, memory or training time.\nMask Attention Network (MAN)~ prepends a dynamic mask attention module to the self-attention module in each Transformer block. The mask is conditioned on token representations, the relative distance between tokens and head indices. The proposed dynamic mask attention is shown to effectively model locality in text data and the induced model consistently outperforms the baseline model in machine translation and abstractive summarization.\nNotably, there's a line of work that uses Neural Architecture Search (NAS) to search for alternative Transformer architectures. The Evolved Transformer (ET)~ employs evolution-based architecture search with the standard Transformer architecture seeding the initial population. The searched model demonstrates consistent improvement\nover Transformer on several language tasks. As another representative work, DARTSformer applies differentiable architecture search (DARTS)~, combined with a multi-split reversible network and a backpropagation-with-reconstruction algorithm for memory efficiency. The resulting model consistently outperforms standard Transformer and compares favorably to larger ET models, with a significantly reduced search cost.","id":"be720c58-062b-41bd-b6e9-732058f0f671","level":"subsection","origin_cites_number":6,"parent_id":"04795308-eff0-4fbb-962f-a8077d36f894","prefix_titles":[["title","A Survey of Transformers"],["section","Architecture-level Variants"],["subsection","Exploring Alternative Architecture"]],"subsections":[],"title":"Exploring Alternative Architecture"},{"cite_extract_rate":0.7000000000000001,"cites":[826,1445,679,707,9,1499,7],"content":"\\label{sec:ptm}\nAs a key difference from convolutional networks and recurrent networks that inherently incorporates the inductive bias of locality, Transformer does not make any assumption about how the data is structured. On the one hand, this effectively makes Transformer a very universal architecture that has the potential of capturing dependencies of different ranges. On the other hand, this makes Transformer prone to overfitting when the data is limited. One way to alleviate this issue is to introduce inductive bias into the model.\nRecent studies suggest that Transformer models that are pre-trained on large corpora can learn universal language representations that are beneficial for downstream tasks~. The models are pre-trained using various self-supervised objectives, e.g., predicting a masked word given its context. After pre-training a model, one can simply fine-tune it on downstream datasets, instead of training a model from scratch. To illustrate typical ways of using Transformers in pre-training, we identify some of the pre-trained Transformers and categorize them as follows.\n\\begin{itemize}\n \\item \\textit{Encoder only}. A line of work uses the Transformer encoder as its backbone architecture. BERT~ is a representative PTM that is typically used for natural language understanding tasks. It utilizes Masked Language Modeling (MLM) and Next Sentence Prediction (NSP) as the self-supervised training objective. RoBERTa~ further adapts the training of BERT and removes the NSP objective as it is found to hurt performance on downstream tasks.\n \\item \\textit{Decoder only}. Several studies focus on pre-training Transformer decoders on language modeling. For example, the Generative Pre-trained Transformer (GPT) series (i.e., GPT~, GPT-2~, and GPT-3~) is dedicated to scaling pre-trained Transformer decoders and has recently illustrated that a large-scale PTM can achieve impressive few-shot performance with the task and examples fed to the model as constructed prompts~.\n \\item \\textit{Encoder-Decoder}. There are also PTMs that adopt Transformer encoder-decoder as the overall architecture. BART~ extends the denoising objective of BERT to encoder-decoder architecture. The benefit of using an encoder-decoder architecture is that the inducing model is equipped with the ability to perform both natural language understanding and generation. T5~ adopts similar architecture and was one of the earliest studies that use task-specific text prefix in downstream tasks.\n\\end{itemize}\nSome of the Transformer architecture variants can also be applied to Transformer-based PTMs. For instance, BigBird~ introduced in Sec.~\\ref{sec:sparseattn} is a encoder-based PTM that uses compound position-based sparse attention to enable long sequence inputs. GPT-3~ uses alternating dense and locally banded sparse attention (which was also introduced in Sec.~\\ref{sec:sparseattn}) in self-attention modules. Switch Transformer~ is an encoder-based PTM that replaces FFN layers with mixture-of-experts layers and can increase parameter count while keeping the FLOPs per example constant.","id":"b7522af4-45ec-452c-b804-11b6882e6225","level":"section","origin_cites_number":10,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Pre-trained Transformers"]],"subsections":[],"title":"Pre-trained Transformers"},{"cite_extract_rate":0.7674418604651161,"cites":[7361,1500,7368,7373,1458,1446,7056,7360,7040,7370,1470,768,732,38,1272,1514,7339,1447,8459,8458,1502,1469,9,1515,1501,1461,1453,1456,1462,1503,1516,1513],"content":"\\label{sec:app}\nTransformer was originally designed for machine translation but has been widely adopted in various fields besides NLP, including CV and audio processing, due to its flexible architecture.\n(1) \\textit{Natural Language Processing}. Transformer and its variants have been extensively explored and applied in NLP tasks, e.g., machine translation~, language modeling~ and named entity recognition~. Massive effort has been dedicated to pre-training Transformer models on large-scale text corpora, which we believe is one of the major reasons of Transformer's wide application in NLP.\n(2) \\textit{Computer Vision}. Transformer have also been adapted for various vision tasks, e.g., image classification~, object detection~, image generation~ and video processing~. and provide reviews on existing work of visual Transformers. We encourage readers to refer to these surveys for further understand the current research progress on Transformers in CV.\n(3) \\textit{Audio Applications}. Transformer can also be extended for audio-related applications, e.g., speech recognition~, speech synthesis~, speech enhancement~ and music generation~.\n(4) \\textit{Multimodal Applications}. Owing to its flexible architecture, Transformer has also been applied in various multimodal scenarios, e.g., visual question answering~, visual commonsense reasoning~, caption generation~, speech-to-text translation~ and text-to-image generation~.","id":"b3d753a5-00c0-4eb4-b313-01e490dd9d3d","level":"section","origin_cites_number":43,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Applications of Transformer"]],"subsections":[],"title":"Applications of Transformer"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:discussion}\nIn this survey, we conduct a comprehensive overview of X-formers and propose a new taxonomy. Most of the existing works improve Transformer from different perspectives, such as efficiency, generalization, and applications. The improvements include incorporating structural prior, designing lightweight architecture, pre-training, and so on.\nAlthough X-formers have proven their power for various tasks, challenges still exist. Besides the current concerns (e.g. efficiency and generalization), the further improvements of Transformer may lie in the following directions:\n(1) \\textit{Theoretical Analysis}. The architecture of Transformer has been demonstrated to be capable of supporting large-scale training datasets with enough parameters. Many works show that Transformer has a larger capacity than CNNs and RNNs and hence has the ability to handle a huge amount of training data.\nWhen Transformer is trained on sufficient data, it usually has better performances than CNNs or RNNs.\nAn intuitive explanation is that Transformer has few prior assumptions on the data structure and therefore is more flexible than CNNs and RNNs. However, the theoretical reason is unclear and we need some theoretical analysis of Transformer ability.\n(2) \\textit{Better Global Interaction Mechanism beyond Attention}. A main advantage of Transformer is the use of the attention mechanism to model the global dependencies among nodes within input data. However, many studies have shown that full attention is unnecessary for most nodes. It is, to some degree, inefficient to indistinguishably calculate attention for all nodes. Therefore, there is still plenty of room for improvements in efficiently modeling global interactions. On the one hand, the self-attention module can be regarded as a fully-connected neural network with dynamical connection weights, which aggregates non-local information with dynamic routing. Therefore, other dynamic routing mechanisms are alternative approaches worth exploring. On the other hand, the global interaction can also be modeled by other types of neural networks, such as memory-enhanced models.\n(3) \\textit{Unified Framework for Multimodal Data}. In many application scenarios, integrating multimodal data is useful and necessary to boost the task performance. Moreover, the general AI also needs the ability to capture the semantic relations across different modalities.\nSince Transformer achieves great success on text, image, video, and audio, we have a chance to build a unified framework and better capture the inherent connections among multimodal data. However, the design of the intra-modal and cross-modal attention still remains to be improved.\nFinally, we wish this survey to be a hands-on reference for better understanding the current research progress on Transformers and help readers to further improve Transformers for various applications.\n\\bibliographystyle{ACM-Reference-Format}\n\\bibliography{references}\n\\end{document}","id":"ec81ede4-3b9c-44b8-96e2-94e0f2e4f760","level":"section","origin_cites_number":0,"parent_id":"456ca844-322a-4fcf-90ea-531b14db9d9e","prefix_titles":[["title","A Survey of Transformers"],["section","Conclusion and Future Directions"]],"subsections":[],"title":"Conclusion and Future Directions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ]\n ],\n \"subsections\": [\n \"48308b43-2cfe-4236-84fb-c7cd67559d21\",\n \"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62\",\n \"aa12aeb0-6a41-4c9b-bfb0-68b4522245cd\",\n \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"7a323139-7d7b-422d-a6c6-3617457ad2f6\",\n \"04795308-eff0-4fbb-962f-a8077d36f894\",\n \"b7522af4-45ec-452c-b804-11b6882e6225\",\n \"b3d753a5-00c0-4eb4-b313-01e490dd9d3d\",\n \"ec81ede4-3b9c-44b8-96e2-94e0f2e4f760\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.7272727272727271,\n \"cites\": [\n 7360,\n 1447,\n 1445,\n 7052,\n 2401,\n 732,\n 1446,\n 38\n ],\n \"content\": \"\\\\label{sec:intro}\\nTransformer~ is a prominent deep learning model that has been widely adopted in various fields, such as natural language processing (NLP), computer vision (CV) and speech processing. Transformer was originally proposed as a sequence-to-sequence model~ for machine translation. Later works show that Transformer-based pre-trained models (PTMs)~ can achieve \\\\textit{state-of-the-art} performances on various tasks. As a consequence, Transformer has become the go-to architecture in NLP, especially for PTMs. In addition to language related applications, Transformer has also been adopted in CV~, audio processing~ and even other disciplines, such as chemistry~ and life sciences~.\\nDue to the success, a variety of Transformer variants (a.k.a. X-formers) have been proposed over the past few years.\\nThese X-formers improve the vanilla Transformer from different perspectives.\\n\\\\begin{enumerate}\\n \\\\item \\\\textit{Model Efficiency}. A key challenge of applying Transformer is its inefficiency at processing long sequences mainly due to the computation and memory complexity of the self-attention module.\\nThe improvement methods include lightweight attention (e.g. sparse attention variants) and Divide-and-conquer methods (e.g., recurrent and hierarchical mechanism).\\n \\\\item \\\\textit{Model Generalization}. Since the transformer is a flexible architecture and makes few assumptions on the structural bias of input data, it is hard to train on small-scale data. The improvement methods include introducing structural bias or regularization, pre-training on large-scale unlabeled data, etc.\\n \\\\item \\\\textit{Model Adaptation}. This line of work aims to adapt the Transformer to specific downstream tasks and applications.\\n\\\\end{enumerate}\\nIn this survey, we aim to provide a comprehensive review of the Transformer and its variants. Although we can organize X-formers on the basis of the perspectives mentioned above, many existing X-formers may address one or several issues. For example, sparse attention variants not only reduce the computational complexity but also introduce structural prior on input data to alleviate the overfitting problem on small datasets. Therefore, it is more methodical to categorize the various existing X-formers and propose a new taxonomy mainly according to their ways to improve the vanilla Transformer: architecture modification, pre-training, and applications.\\nConsidering the audience of this survey may be from different domains, we mainly focus on the general architecture variants and just briefly discuss the specific variants on pre-training and applications.\\nThe rest of the survey is organized as follows. Sec.~\\\\ref{sec:background} introduces the architecture and the key components of Transformer. Sec.~\\\\ref{sec:taxonomy} clarifies the categorization of Transformer variants. Sec.~\\\\ref{sec:attention}$\\\\sim$\\\\ref{sec:other_module} review the module-level modifications, including attention module, position encoding, layer normalization and feed-forward layer. Sec.~\\\\ref{sec:beyond} reviews the architecture-level variants. Sec.~\\\\ref{sec:ptm} introduces some of the representative Transformer-based PTMs. Sec.~\\\\ref{sec:app} introduces the application of Transformer to various different fields. Sec.~\\\\ref{sec:discussion} discusses some aspects of Transformer that researchers might find intriguing and summarizes the paper.\",\n \"id\": \"48308b43-2cfe-4236-84fb-c7cd67559d21\",\n \"level\": \"section\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:background}\",\n \"id\": \"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ]\n ],\n \"subsections\": [\n \"cbbc7151-fd9c-4d11-a9dd-a0790a09698d\",\n \"89fa15e6-1537-4f12-a5ac-bdd5d071bfac\",\n \"3ddd6e3c-d4a1-4bce-b510-16dfc75582ed\",\n \"f0fa958a-db02-433c-858f-f7024645f38d\"\n ],\n \"title\": \"Background\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 97,\n 57,\n 38\n ],\n \"content\": \"\\\\label{sec:vanilla_xformer}\\nThe vanilla Transformer~ is a sequence-to-sequence model and consists of an encoder and a decoder, each of which is a stack of $L$ identical blocks. Each \\\\textit{encoder block} is mainly composed of a multi-head self-attention module and a position-wise feed-forward network (FFN).\\nFor building a deeper model, a residual connection~ is employed around each module, followed by Layer Normalization~ module.\\nCompared to the encoder blocks, decoder blocks additionally insert cross-attention modules between the multi-head self-attention modules and the position-wise FFNs. Furthermore, the self-attention modules in the decoder are adapted to prevent each position from attending to subsequent positions.\\nThe overall architecture of the vanilla Transformer is shown in Fig. \\\\ref{fig:xformer_arch}.\\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.7\\\\linewidth]{assets/xformer.pdf}\\n \\\\caption{Overview of vanilla Transformer architecture}\\n \\\\label{fig:xformer_arch}\\n\\\\end{figure}\\nIn the following subsection, we shall introduce the key modules of the vanilla Transformer.\",\n \"id\": \"cbbc7151-fd9c-4d11-a9dd-a0790a09698d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Vanilla Transformer\"\n ]\n ],\n \"subsections\": [\n \"625a964c-dc0c-4228-a0bd-e5fba16df1f2\",\n \"fdfb7e39-e1e9-4079-9037-b8bc17e8bb70\",\n \"0c6acb39-f7c8-4d89-9969-16b6e42547a1\",\n \"64d291c3-febe-419e-831a-4ec33ca28ce7\"\n ],\n \"title\": \"Vanilla Transformer\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Transformer adopts attention mechanism with Query-Key-Value (QKV) model. Given the packed matrix representations of queries $\\\\bQ\\\\in\\\\mathbb{R}^{N\\\\times D_k}$, keys $\\\\bK\\\\in\\\\mathbb{R}^{M\\\\times D_k}$, and values $\\\\bV\\\\in\\\\mathbb{R}^{M\\\\times D_v}$, the scaled dot-product attention used by Transformer is given by\\\\footnote{if not stated otherwise, we use row-major notations throughout this survey (e.g., the $i$-th row in $\\\\bQ$ is the query $\\\\bq_i$) and all the vectors are row vectors by default.}\\n\\\\begin{equation}\\\\label{attention}\\n \\\\mathrm{Attention}(\\\\bQ, \\\\bK, \\\\bV)=\\\\mathrm{softmax}\\\\left(\\\\frac{\\\\bQ\\\\bK^\\\\top}{\\\\sqrt{D_k}}\\\\right)\\\\bV=\\\\bA\\\\bV,\\n\\\\end{equation}\\nwhere $N$ and $M$ denote the lengths of queries and keys (or values); $D_k$ and $D_v$ denote the dimensions of keys (or queries) and values; $\\\\bA=\\\\mathrm{softmax}\\\\left(\\\\frac{\\\\bQ\\\\bK^\\\\top}{\\\\sqrt{D_k}}\\\\right)$ is often called \\\\textit{attention matrix}; softmax is applied in a row-wise manner. The dot-products of queries and keys are divided by $\\\\sqrt{D_k}$ to alleviate gradient vanishing problem of the softmax function.\\nInstead of simply applying a single attention function, Transformer uses multi-head attention, where the $D_m$-dimensional original queries, keys and values are projected into $D_k$, $D_k$ and $D_v$ dimensions, respectively, with $H$ different sets of learned projections. For each of the projected queries, keys and values, and output is computed with attention according to Eq. \\\\eqref{attention}. The model then concatenates all the outputs and projects them back to a $D_m$-dimensional representation.\\n\\\\begin{align}\\n \\\\mathrm{MultiHeadAttn}(\\\\bQ,\\\\bK,\\\\bV)&=\\\\mathrm{Concat}(\\\\mathrm{head}_1,\\\\cdots,\\\\mathrm{head}_H)\\\\bW^O\\\\label{eq:multihead},\\\\\\\\\\n \\\\mathrm{where}\\\\ \\\\mathrm{head}_i&=\\\\mathrm{Attention}(\\\\bQ \\\\bW_i^Q,\\\\bK \\\\bW_i^K, \\\\bV \\\\bW_i^V)\\\\label{eq:headi}.\\n\\\\end{align}\\nIn Transformer, there are three types of attention in terms of the source of queries and key-value pairs:\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Self-attention}. In Transformer encoder, we set $\\\\bQ=\\\\bK=\\\\bV=\\\\bX$ in Eq. \\\\eqref{eq:multihead}, where $\\\\bX$ is the outputs of the previous layer.\\n \\\\item \\\\textit{Masked Self-attention}. In the Transformer decoder, the self-attention is restricted such that queries at each position can only attend to all key-value pairs up to and including that position. To enable parallel training, this is typically done by applying a mask function to the unnormalized attention matrix $\\\\hat{\\\\bA}=\\\\exp(\\\\frac{\\\\bQ\\\\bK^\\\\top}{\\\\sqrt{D_k}})$, where the illegal positions are masked out by setting $\\\\hat{A}_{ij}=-\\\\infty \\\\textrm{ if } i<j$. This kind of self-attention is often referred to as autoregressive or causal attention\\\\footnote{This term seems to be borrowed from the \\\\textit{causal system}, where the output depends on past and current inputs but not future inputs.}.\\n \\\\item \\\\textit{Cross-attention}. The queries are projected from the outputs of the previous (decoder) layer, whereas the keys and values are projected using the outputs of the encoder.\\n\\\\end{itemize}\",\n \"id\": \"625a964c-dc0c-4228-a0bd-e5fba16df1f2\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"cbbc7151-fd9c-4d11-a9dd-a0790a09698d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Vanilla Transformer\"\n ],\n [\n \"subsubsection\",\n \"Attention Modules\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attention Modules\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The position-wise FFN\\\\footnote{The parameters are shared across different positions, thus the position-wise FFN can also be understood as two convolution layers with kernel size of 1.} is a fully connected feed-forward module that operates separately and identically on each position\\n\\\\begin{equation}\\n \\\\mathrm{FFN}(\\\\bH') = \\\\mathrm{ReLU}(\\\\bH'\\\\bW^1 + \\\\bb^1)\\\\bW^2+\\\\bb^2,\\n\\\\end{equation}\\nwhere $\\\\bH'$ is the outputs of previous layer, and $\\\\bW^1\\\\in\\\\mathbb{R}^{D_m\\\\times D_f},\\\\bW^2\\\\in\\\\mathbb{R}^{D_f\\\\times D_m},\\\\bb^1\\\\in\\\\mathbb{R}^{D_f},\\\\bb^2\\\\in\\\\mathbb{R}^{D_m}$ are trainable parameters. Typically the intermediate dimension $D_{f}$ of the FFN is set to be larger than $D_m$.\",\n \"id\": \"fdfb7e39-e1e9-4079-9037-b8bc17e8bb70\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"cbbc7151-fd9c-4d11-a9dd-a0790a09698d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Vanilla Transformer\"\n ],\n [\n \"subsubsection\",\n \"Position-wise FFN\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Position-wise FFN\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 97,\n 57\n ],\n \"content\": \"In order to build a deep model, Transformer employs a residual connection~ around each module, followed by Layer Normalization~. For instance, each Transformer encoder block may be written as\\n\\\\begin{align}\\n \\\\bH'&=\\\\mathrm{LayerNorm}(\\\\mathrm{SelfAttention}(\\\\bX)+\\\\bX)\\\\\\\\\\n \\\\bH&=\\\\mathrm{LayerNorm}(\\\\mathrm{FFN}(\\\\bH')+\\\\bH'),\\n\\\\end{align}\\nwhere $\\\\mathrm{SelfAttention}(\\\\cdot)$ denotes self attention module and $\\\\mathrm{LayerNorm}(\\\\cdot)$ denotes the layer normalization operation.\",\n \"id\": \"0c6acb39-f7c8-4d89-9969-16b6e42547a1\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"cbbc7151-fd9c-4d11-a9dd-a0790a09698d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Vanilla Transformer\"\n ],\n [\n \"subsubsection\",\n \"Residual Connection and Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Residual Connection and Normalization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Since Transformer doesn't introduce recurrence or convolution, it is ignorant of positional information (especially for the encoder). Thus additional positional representation (Detailed discussion in Sec.~\\\\ref{sec:pos_rep}) is needed to model the ordering of tokens.\",\n \"id\": \"64d291c3-febe-419e-831a-4ec33ca28ce7\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"cbbc7151-fd9c-4d11-a9dd-a0790a09698d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Vanilla Transformer\"\n ],\n [\n \"subsubsection\",\n \"Position Encodings\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Position Encodings\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Generally, the Transformer architecture can be used in three different ways:\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Encoder-Decoder}. The full Transformer architecture as introduced in Sec.~\\\\ref{sec:vanilla_xformer} is used. This is typically used in sequence-to-sequence modeling (e.g., neural machine translation).\\n \\\\item \\\\textit{Encoder only}. Only the encoder is used and the outputs of the encoder are utilized as a representation for the input sequence. This is usually used for classification or sequence labeling problems.\\n \\\\item \\\\textit{Decoder only}. Only the decoder is used, where the encoder-decoder cross-attention module is also removed. This is typically used for sequence generation, such as language modeling.\\n\\\\end{itemize}\",\n \"id\": \"89fa15e6-1537-4f12-a5ac-bdd5d071bfac\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Model Usage\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model Usage\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 38\n ],\n \"content\": \"\\\\label{sec:model-analysis}\\nTo illustrate the computation time and parameter requirements of the Transformer, we analyze the two core components of the Transformer (i.e., the self-attention module and the position-wise FFN) in Table \\\\ref{tab:complexity_attn_ffn}. We assume that the hidden dimension $D_m$ of the model is $D$, and that the input sequence length is $T$. The intermediate dimension of FFN is set to $4D$ and the dimension of keys and values are set to $D/H$ as in .\\n\\\\begin{table}[htbp]\\n \\\\caption{Complexity and parameter counts of self-attention and position-wise FFN}\\n \\\\label{tab:complexity_attn_ffn}\\n \\\\centering\\n \\\\begin{tabular}{c|c|c}\\n \\\\hline\\n Module & Complexity & \\\\#Parameters\\\\\\\\\\n \\\\hline\\n self-attention & $\\\\mathcal O(T^2\\\\cdot D)$ & $4D^2$\\\\\\\\\\n \\\\hline\\n position-wise FFN & $\\\\mathcal O(T\\\\cdot D^2)$ & $8D^2$\\\\\\\\\\n \\\\hline\\n \\\\end{tabular}\\n\\\\end{table}\\nWhen the input sequences are short, the hidden dimension $D$ dominates the complexity of self-attention and position-wise FFN. The bottleneck of Transformer thus lies in FFN. However, as the input sequences grow longer, the sequence length $T$ gradually dominates the complexity of these modules, in which case self-attention becomes the bottleneck of Transformer. Furthermore, the computation of self-attention requires that a $T\\\\times T$ attention distribution matrix is stored, which makes the computation of Transformer infeasible for long-sequence scenarios (e.g., long text documents and pixel-level modeling of high-resolution images). One shall see that the goal of increasing the efficiency of Transformer generally leads to the long-sequence compatibility of self-attention, as well as the computation and parameter efficiency of position-wise FFN for ordinary settings.\",\n \"id\": \"3ddd6e3c-d4a1-4bce-b510-16dfc75582ed\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Model Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model Analysis\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"f0fa958a-db02-433c-858f-f7024645f38d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"07bae4f8-4d29-4fbd-aec3-ca99cfb7bc62\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Comparing Transformer to Other Network Types\"\n ]\n ],\n \"subsections\": [\n \"037a9c67-64a3-4bbe-b197-a41922561a14\",\n \"cf87fcdd-fc72-4f50-9754-c59538a65d3f\"\n ],\n \"title\": \"Comparing Transformer to Other Network Types\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 38\n ],\n \"content\": \"As a central piece of Transformer, self-attention comes with a flexible mechanism to deal with variable-length inputs. It can be understood as a fully connected layer where the weights are dynamically generated from pairwise relations from inputs. Table \\\\ref{tab:op_complexities} compares the complexity, sequential operations, and maximum path length\\\\footnote{The maximum length of the paths forward and backward signals have to traverse to get from any input position to arbitrary output position. Shorter length implies a better potential for learning long-range dependencies.} of self-attention with three commonly used layer types. We summarize the advantages of self-attention as follows:\\n\\\\begin{enumerate}\\n \\\\item It has the same maximum path length as fully connected layers, making it suitable for long-range dependencies modeling. Compared to fully connected layers, it is more parameter-efficient and more flexible in handling variable-length inputs.\\n \\\\item Due to the limited receptive field of convolutional layers, one typically needs to stack a deep network to have a global receptive field. On the other hand, the constant maximum path length enables self-attention to model long-range dependencies with a constant number of layers.\\n \\\\item The constant sequential operations and maximum path length make self-attention more parallelizable and better at long-range modeling than recurrent layers.\\n\\\\end{enumerate}\\n\\\\begin{table}[ht]\\n \\\\caption{Per-layer complexity, minimum number of sequential operations and maximum path lengths for different layer types. $T$ is the sequence length, $D$ is the representation dimension and $K$ is the kernel size of convolutions~.}\\n\\\\label{tab:op_complexities}\\n\\\\begin{center}\\n\\\\vspace{-1mm}\\n\\\\begin{tabular}{lccc}\\n\\\\toprule\\nLayer Type & Complexity & Sequential & Maximum Path Length \\\\\\\\\\n &per Layer & Operations & \\\\\\\\\\n\\\\hline\\n\\\\rule{0pt}{2.0ex}Self-Attention & $\\\\mathcal O(T^2 \\\\cdot D)$ & $\\\\mathcal O(1)$ & $\\\\mathcal O(1)$ \\\\\\\\\\nFully Connected & $\\\\mathcal O(T^2 \\\\cdot D^2)$ & $\\\\mathcal O(1)$ & $\\\\mathcal O(1)$ \\\\\\\\\\nConvolutional & $\\\\mathcal O(K \\\\cdot T \\\\cdot D^2)$ & $\\\\mathcal O(1)$ & $\\\\mathcal O(\\\\log_K(T))$ \\\\\\\\\\nRecurrent & $\\\\mathcal O(T \\\\cdot D^2)$ & $\\\\mathcal O(T)$ & $\\\\mathcal O(T)$ \\\\\\\\\\n\\\\bottomrule\\n\\\\end{tabular}\\n\\\\end{center}\\n\\\\end{table}\",\n \"id\": \"037a9c67-64a3-4bbe-b197-a41922561a14\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"f0fa958a-db02-433c-858f-f7024645f38d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Comparing Transformer to Other Network Types\"\n ],\n [\n \"subsubsection\",\n \"Analysis of Self-Attention\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Analysis of Self-Attention\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 208,\n 553\n ],\n \"content\": \"Transformer is often compared against convolutional and recurrent networks. Convolutional networks are known to impose the inductive biases of translation invariance and locality with shared local kernel functions. Similarly, recurrent networks carry the inductive biases of temporal invariance and locality via their Markovian structure~. On the other hand, the Transformer architecture makes few assumptions about structural information of data. This makes Transformer a universal and flexible architecture. As a side effect, the lack of structural bias makes Transformer prone to overfitting for small-scale data.\\nAnother closely related network type is Graph Neural Networks (GNNs) with message passing~. Transformer can be viewed as a GNN defined over a complete directed graph (with self-loop) where each input is a node in the graph. The key difference between Transformer and GNNs is that Transformer introduces no prior knowledge over how input data are structured — the message passing process in Transformer solely depends on similarity measures over the content.\",\n \"id\": \"cf87fcdd-fc72-4f50-9754-c59538a65d3f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"f0fa958a-db02-433c-858f-f7024645f38d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Comparing Transformer to Other Network Types\"\n ],\n [\n \"subsubsection\",\n \"In Terms of Inductive Bias\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"In Terms of Inductive Bias\"\n },\n {\n \"cite_extract_rate\": 0.808333333333333,\n \"cites\": [\n 1474,\n 1493,\n 1457,\n 1482,\n 1460,\n 1491,\n 7360,\n 7040,\n 1484,\n 1494,\n 1464,\n 1488,\n 1470,\n 1452,\n 1466,\n 1490,\n 793,\n 1445,\n 1502,\n 7054,\n 1467,\n 1463,\n 8384,\n 1456,\n 1480,\n 794,\n 1475,\n 1465,\n 7361,\n 1500,\n 1459,\n 7053,\n 1458,\n 1479,\n 7365,\n 1496,\n 1133,\n 8454,\n 768,\n 7367,\n 7333,\n 1489,\n 1451,\n 1469,\n 9,\n 1476,\n 1453,\n 1495,\n 1487,\n 1462,\n 1478,\n 679,\n 1483,\n 1492,\n 7369,\n 7,\n 1473,\n 1446,\n 7370,\n 8456,\n 826,\n 7362,\n 7364,\n 7363,\n 1468,\n 38,\n 1486,\n 7298,\n 1498,\n 1497,\n 1501,\n 7371,\n 8457,\n 7368,\n 1455,\n 707,\n 8455,\n 1471,\n 1481,\n 1182,\n 732,\n 1272,\n 7339,\n 1477,\n 1449,\n 7273,\n 1461,\n 1454,\n 1448,\n 1485,\n 7366,\n 798,\n 1499,\n 1450,\n 1472\n ],\n \"content\": \"\\\\label{sec:taxonomy}\\nA wide variety of models have been proposed so far based on the vanilla Transformer from three perspectives: types of architecture modification, pre-training methods, and applications.\\nFig. \\\\ref{fig:xformer_taxonomy} gives an illustrations of our categorization of Transformer variants.\\n\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=1.0\\\\linewidth]{assets/taxonomy.pdf}\\n \\\\caption{Categorization of Transformer variants.}\\n \\\\label{fig:xformer_taxonomy}\\n\\\\end{figure}\\nFig. \\\\ref{taxonomy_of_xformer} illustrates our taxonomy and some representative models.\\n\\\\tikzstyle{leaf}=[mybox,minimum height=1em,\\nfill=hidden-orange!40, text width=20em, text=black,align=left,font=\\\\tiny,\\ninner xsep=2pt,\\ninner ysep=1pt,\\n]\\n\\\\begin{figure*}[tp]\\n \\\\centering\\n\\\\begin{forest}\\n forked edges,\\n for tree={\\n grow=east,\\n reversed=true,\\n anchor=base west,\\n parent anchor=east,\\n child anchor=west,\\n base=left,\\n font=\\\\small,\\n rectangle,\\n draw=hiddendraw,\\n rounded corners,align=left,\\n minimum width=2.5em,\\ns sep=3pt,\\ninner xsep=2pt,\\ninner ysep=1pt,\\nver/.style={rotate=90, child anchor=north, parent anchor=south, anchor=center},\\n },\\n where level=1{text width=2.7em,font=\\\\scriptsize,}{},\\n where level=2{text width=3em,font=\\\\tiny}{},\\n where level=3{text width=3em,font=\\\\tiny}{},\\n [X-formers, ver\\n [Module\\\\\\\\Level\\n [Attention\\n [Sparse\\n [Star-Transformer{,} Longformer{,} ETC{,} BigBird{,} Sparse Transformer\\\\\\\\\\n BP-Transformer{,} Image Transformer{,} Axial Transformer\\n ,leaf,text width=21.5em]\\n [Routing Transformer{,} Reformer{,} SAC{,} Sparse Sinkhorn Attention\\n ,leaf,text width=21.5em]\\n ]\\n [Linearized\\n [Linear Transformer{,} Performer{,} RFA{,} Delta Net\\n ,leaf,text width=18.5em]\\n ]\\n [Prototype\\n [Clustered Attention{,} Informer\\n ,leaf,text width=12em]\\n ]\\n [Memory\\\\\\\\Compress\\n [MCA{,} Set Transformer{,} Linformer\\n ,leaf,text width=12em]\\n ]\\n [Low-rank\\n [Low-rank Attention{,} CSALR{,} Nystr{\\\\\\\"{o}}mformer~\\n ,leaf,text width=15em]\\n ]\\n [Prior\\\\\\\\Attention\\n [Local Transformer{,} Gaussian Transformer\\n ,leaf,text width=15em]\\n [Predictive Attention Transformer{,} Realformer{,} Lazyformer\\n ,leaf,text width=18.5em]\\n [CAMTL\\n ,leaf,text width=4em]\\n [Average Attention{,} Hard-Coded Gaussian Attention{,} Synthesizer\\n ,leaf]\\n ]\\n [Multi-head\\n [{,} {,} Talking-head Attention\\\\\\\\\\n Collaborative MHA\\n ,leaf,text width=20em]\\n [Adaptive Attention Span{,} Multi-Scale Transformer\\n ,leaf,text width=20em]\\n [Dynamic Routing\\n ,leaf,text width=6.5em]\\n ]\\n ]\\n [Position\\\\\\\\ Encoding\\n [Absolute\\n [BERT{,} {,} FLOATER\\n ,leaf,text width=11em]\\n ]\\n [Relative\\n [{,} Music Transformer{,} T5{,} Transformer-XL\\\\\\\\ DeBERTa\\n ,leaf,text width=20em]\\n ]\\n [Other Rep.\\n [TUPE{,} Roformer\\n ,leaf,text width=7em]\\n ]\\n [Implicit Rep.\\n [Complex Embedding{,} R-Transformer {,} CPE\\n ,leaf,text width=16em]\\n ]\\n ]\\n [LayerNorm\\n [Placement\\n [post-LN{,} pre-LN\\n ,leaf,text width=16em]\\n ]\\n [Substitutes\\n [AdaNorm{,} scaled $\\\\ell_2$ normalization{,} PowerNorm\\n ,leaf,text width=16em]\\n ]\\n [Norm-free\\n [ReZero-Transformer\\n ,leaf,text width=9.5em]\\n ]\\n ]\\n [FFN\\n [Activ. Func.\\n [Swish{,} GELU{,} GLU\\n ,leaf,text width=9.5em]\\n ]\\n [Enlarge\\\\\\\\Capacity\\n [Product-key Memory{,} Gshard{,} Switch Transformer{,}\\\\\\\\ Expert Prototyping{,} Hash Layer\\n ,leaf,text width=16em]\\n ]\\n [Dropping\\n [All-Attention layer{,} \\n ,leaf,text width=11em]\\n ]\\n ]\\n ]\\n [Arch.\\\\\\\\Level\\n [Lighweight\\n [Lite Transformer{,} Funnel Transformer{,} DeLighT\\n ,leaf,text width=16em]\\n ]\\n [Connectivity\\n [Realformer{,} Predictive Attention Transformer{,} Transparent Attention\\\\\\\\\\n Feedback Transformer~\\n ,leaf,text width=24.5em]\\n ]\\n [ACT\\n [UT{,} Conditional Computation Transformer{,} DeeBERT{,} PABEE{,} {,} \\\\\\\\\\n ,leaf,text width=24.5em]\\n ]\\n [Divide \\\\& \\\\\\\\\\n Conquer\\n [Recurrence\\n [Transformer-XL{,} Compressive Transformer{,} Memformer\\\\\\\\ {,} ERNIE-Doc\\n ,leaf,text width=20em]\\n ]\\n [Hierarchy\\n [{,} HIBERT{,} {,} Hi-Transformer\\\\\\\\ TENER{,} TNT\\n ,leaf,text width=20em]\\n ]\\n ]\\n [Alt. Arch.\\n [ET{,} Macaron Transformer{,} Sandwich Transformer{,} MAN{,} DARTSformer\\n ,leaf,text width=24.5em]\\n ]\\n ]\\n [Pre-Train\\n [Encoder\\n [BERT{,} RoBERTa{,} BigBird,leaf,text width=11em]\\n ]\\n [Decoder\\n [GPT{,} GPT-2{,} GPT-3,leaf,text width=11em]\\n ]\\n [Enc.Dec.\\n [ BART{,} T5{,} Switch Transformer\\n ,leaf,text width=11em]\\n ]\\n ]\\n [App.\\n [NLP\\n [BERT{,}ET{,} Transformer-XL{,}Compressive Transformer{,} TENER\\n ,leaf,text width=22em]\\n ]\\n [CV\\n [Image Transformer{,} DETR{,} ViT{,} Swin Transformer{,} ViViT\\n ,leaf,text width=22em]\\n ]\\n [Audio\\n [Speech Transformer{,} Streaming Transformer{,} Reformer-TTS{,} Music Transformer\\n ,leaf,text width=25em]\\n ]\\n [Multimodal\\n [VisualBERT{,} VLBERT{,} VideoBERT{,} M6{,} Chimera{,} DALL-E{,} CogView\\n ,leaf,text width=25em]\\n ]\\n ]\\n ]\\n\\\\end{forest}\\n\\\\caption{Taxonomy of Transformers}\\n\\\\label{taxonomy_of_xformer}\\n\\\\end{figure*}\\nIn this survey, we focus on reviewing the works on architecture modifications.\\nSince the attention module is the key component of Transformer, we solely describe the attention-related variants in Sec.~\\\\ref{sec:attention} and introduce the other module-level variants in Sec.~\\\\ref{sec:other_module}. Then Sec.~\\\\ref{sec:beyond} describes the other architecture-level variants. Finally, we briefly review the works on pre-training in Sec.~\\\\ref{sec:ptm} and applications in Sec.~\\\\ref{sec:app}.\\nThere are some comprehensive surveys on the latter two categories of work, such as pre-trained models (PTMs)~ and visual Transformers.\",\n \"id\": \"aa12aeb0-6a41-4c9b-bfb0-68b4522245cd\",\n \"level\": \"section\",\n \"origin_cites_number\": 120,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Taxonomy of Transformers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Taxonomy of Transformers\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:attention}\\nSelf-attention plays an important role in Transformer, but there are two challenges in practical applications.\\n\\\\begin{enumerate}\\n \\\\item \\\\textit{Complexity}. As discussion in Sec.~\\\\ref{sec:model-analysis}, the complexity of self-attention is $\\\\mathcal O(T^2\\\\cdot D)$. Therefore, the attention module becomes a bottleneck when dealing with long sequences.\\n \\\\item \\\\textit{Structural prior}. Self-attention does no assume any structural bias over inputs. Even the order information is also needed to be learned from training data. Therefore, Transformer (w/o pre-training) is usually easy to overfit on small or moderate-size data.\\n\\\\end{enumerate}\\nThe improvements on attention mechanism can be divided into several directions:\\n\\\\begin{enumerate}\\n \\\\item \\\\textit{Sparse Attention}. This line of work introduces sparsity bias into the attention mechanism, leading to reduced complexity.\\n \\\\item \\\\textit{Linearized Attention}. This line of work disentangles the attention matrix with kernel feature maps. The attention is then computed in reversed order to achieve linear complexity.\\n \\\\item \\\\textit{Prototype and Memory Compression}. This class of methods reduces the number of queries or key-value memory pairs to reduce the size of the attention matrix.\\n \\\\item \\\\textit{Low-rank Self-Attention}. This line of work capture the low-rank property of self-attention.\\n \\\\item \\\\textit{Attention with Prior}. The line of research explores supplementing or substituting standard attention with prior attention distributions.\\n \\\\item \\\\textit{Improved Multi-Head Mechanism}. The line of studies explores different alternative multi-head mechanisms.\\n\\\\end{enumerate}\\nWe will describe these attention variants at length in the rest of this section.\",\n \"id\": \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ]\n ],\n \"subsections\": [\n \"c3955c2b-4114-49e4-bbb8-30061d5f27e8\",\n \"a4f89d55-c38a-442a-8744-5625c30e01ff\",\n \"99d263f7-48ac-4be9-9f96-b1cd4380de1b\",\n \"9a2c3eb1-bf66-4ddc-822d-bcc1121a8120\",\n \"afe04fc5-1717-4f24-b061-7af855828533\",\n \"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5\"\n ],\n \"title\": \"Attention\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 793\n ],\n \"content\": \"\\\\label{sec:sparseattn}\\nIn the standard self-attention mechanism, every token needs to attend to all other tokens.\\nHowever, it is observed that for the trained Transformers the learned attention matrix $\\\\bA$ is often very sparse across most data points~.\\nTherefore, it is possible to reduce computation complexity by incorporating structural bias to limit the number of query-key pairs that each query attends to. Under this limitation, we just compute the similarity score of the query-key pairs according to pre-defined patterns\\n\\\\begin{equation}\\\\label{Sparse attention:position-based}\\n \\\\mathrm{\\\\hat{\\\\bA}}_{ij} = \\\\begin{cases} \\\\bq_i\\\\bk_{j}^\\\\top & \\\\text{if token }i\\\\text{ attends to token }j,\\\\\\\\\\n -\\\\infty & \\\\text{if token }i\\\\text{ does not attend to token }j,\\n \\\\end{cases}\\n\\\\end{equation}\\\\label{eq:sparseattn}\\nwhere $\\\\mathrm{\\\\hat{\\\\bA}}$ is un-normalized attention matrix. In implementation the $-\\\\infty$ item is usually not stored in memory so as to decrease memory footprint.\\nFrom another perspective, the standard attention can be regarded as a complete bipartite graph where each query receives information from all memory nodes and updates its representation.\\nThe sparse attention can be considered as a sparse graph where some of the connections between nodes are removed.\\nBased on the metrics of determining the sparse connection, we categorize these approaches into two classes: \\\\textit{position-based} and \\\\textit{content-based} sparse attention.\",\n \"id\": \"c3955c2b-4114-49e4-bbb8-30061d5f27e8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Sparse Attention\"\n ]\n ],\n \"subsections\": [\n \"766f8fe1-2a61-4664-a9be-83f0177cc298\",\n \"6429c3b2-0cb9-415c-9f1d-1f7bd028a380\"\n ],\n \"title\": \"Sparse Attention\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:pos_based}\\nIn position-based sparse attention, the attention matrix is limited according to some pre-defined patterns.\\nAlthough these sparse patterns vary in different forms, we find that some of them can be decomposed into some atomic sparse patterns.\\nWe first identify some atomic sparse patterns and then describe how these patterns are composed in some existing work. Finally, we introduce some extended sparse patterns for specific data types.\",\n \"id\": \"766f8fe1-2a61-4664-a9be-83f0177cc298\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"c3955c2b-4114-49e4-bbb8-30061d5f27e8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Sparse Attention\"\n ],\n [\n \"subsubsection\",\n \"Position-based Sparse Attention\"\n ]\n ],\n \"subsections\": [\n \"3abb14c2-1133-4417-a3ba-50491a89789a\",\n \"cf0fe507-a0b3-499f-9145-a52ed5faa99e\",\n \"8a0b3727-3354-4028-870f-983c49224baa\"\n ],\n \"title\": \"Position-based Sparse Attention\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4722\n ],\n \"content\": \"There are mainly five types of atomic sparse attention patterns, as shown in Fig. \\\\ref{fig:atomic_sparse_attentions}.\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[global\\\\label{fig:global}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/atomic/global.pdf}\\n}\\n\\\\subfigure[band\\\\label{fig:band}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/atomic/band.pdf}\\n}\\n\\\\subfigure[dilated\\\\label{fig:dilated}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/atomic/dilated.pdf}\\n}\\n\\\\subfigure[random\\\\label{fig:random}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/atomic/random.pdf}\\n}\\n\\\\subfigure[block local\\\\label{fig:block_local}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/atomic/block_local.pdf}\\n}\\n\\\\end{center}\\n\\\\caption{Some representative atomic sparse attention patterns. The colored squares means corresponding attention scores are calculated and a blank square means the attention score is discarded.}\\\\label{fig:atomic_sparse_attentions}\\n\\\\end{figure}\\n\\\\begin{enumerate}\\n \\\\item \\\\textit{Global Attention}. To alleviate the degradation of the ability to model the long-range dependencies in sparse attention, one can add some global nodes\\\\footnote{In practice, these global nodes can be selected from the sequence (internal global nodes) or virtual nodes with trainable parameters (external global nodes).} as the hub for information propagation between nodes. These global nodes can attend all nodes in the sequence and the whole sequence attend to these global nodes, as illustrated in Fig. \\\\ref{fig:global}.\\n \\\\item \\\\textit{Band Attention}(a.k.a \\\\textit{sliding window attention} or \\\\textit{local attention}). Since most data come with a strong property of locality, it is natural to restrict each query to attend to its neighbor nodes. A widely adopted class of such sparse pattern is band attention, in which the attention matrix is a band matrix as illustrated in Fig. \\\\ref{fig:band}.\\n \\\\item \\\\textit{Dilated Attention}. Analogous to dilated CNNs~, one can potentially increase the receptive field of the band attention without increasing computation complexity by using a dilated window with gaps of dilation $w_d\\\\ge 1$, as depicted in Fig. \\\\ref{fig:dilated}. This can be easily extended to \\\\textit{strided attention}, where the window size is not limited but the dilation $w_d$ is set to a large value.\\n \\\\item \\\\textit{Random Attention}. To increase the ability of non-local interactions, a few edges are randomly sampled for each query, as illustrated in Fig. \\\\ref{fig:random}. This is based on the observation that random graphs (e.g., Erd\\\\H os–R\\\\'enyi random graph) can have similar spectral properties with complete graphs that leads to a fast mixing time for random walking on graphs.\\n \\\\item \\\\textit{Block Local Attention}. This class of attention segments input sequence into several non-overlapping query blocks, each of which is associated with a local memory block. All the queries in a query block attend to only the keys in the corresponding memory block. Fig. \\\\ref{fig:block_local} depicts a commonly used case where the memory blocks are identical to their corresponding query blocks.\\n\\\\end{enumerate}\",\n \"id\": \"3abb14c2-1133-4417-a3ba-50491a89789a\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"766f8fe1-2a61-4664-a9be-83f0177cc298\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Sparse Attention\"\n ],\n [\n \"subsubsection\",\n \"Position-based Sparse Attention\"\n ],\n [\n \"paragraph\",\n \"4.1.1.1 Atomic Sparse Attention\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"4.1.1.1 Atomic Sparse Attention\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7298,\n 793,\n 7371,\n 1499,\n 7363,\n 134\n ],\n \"content\": \"Existing sparse attentions are often composed of more than one of the above atomic patterns. Fig.~\\\\ref{fig:sparse_attn} illustrates some representative compound sparse attention patterns.\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[Star-Transformer\\\\label{fig:starxformer}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/compound/star.pdf}\\n}\\\\quad\\n\\\\quad\\n\\\\subfigure[Longformer\\\\label{fig:longformer}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/compound/longformer.pdf}\\n}\\\\quad\\n\\\\subfigure[ETC\\\\label{fig:etc}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/compound/etc.pdf}\\n}\\\\quad\\n\\\\subfigure[BigBird\\\\label{fig:bigbird}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/compound/bigbird.pdf}\\n}\\\\quad\\n\\\\caption{Some representative compound sparse attention patterns. The red boxes indicate sequence boundaries.}\\\\label{fig:sparse_attn}\\n\\\\end{center}\\n\\\\end{figure}\\nStar-Transformer~ uses a combination of band attention and global attention. Specifically, Star-Transformer just includes only a global node and a band attention with the width of 3, in which any pair of non-adjacent nodes are connected through a shared global node and adjacent nodes are connected directly with each other. This kind of sparse pattern forms a star-shaped graph among nodes.\\nLongformer~ uses a combination of band attention and internal global-node attention. The global nodes are chosen to be \\\\texttt{[CLS]} token for classification and all question tokens for Question Answering tasks. They also replace some of the band attention heads in upper layers with dilated window attention to increase the receptive field without increasing computation.\\nAs a concurrent work to Longformer~, Extended Transformer Construction (ETC)~ utilizes combination of band attention and external global-node attention. ETC also includes a masking mechanism to handle structured inputs and adapt Contrastive Predictive Coding (CPC)~ for pre-training.\\nIn addition to the band and global attention, BigBird~ uses additional random attention to approximate full attention. Their theoretical analysis also reveals that the usage of a sparse encoder and sparse decoder can simulate any Turing Machine, which explains the success of those sparse attention models.\\nSparse Transformer~ uses a factorized attention where different sparse patterns are designed for different types of data. For data with a periodic structure (e.g., images), it uses a composition of band attention and strided attention. Whereas for data without a periodic structure (e.g., text), it uses a composition of block local attention combined with global attention, where global nodes are from fixed positions in the input sequence.\",\n \"id\": \"cf0fe507-a0b3-499f-9145-a52ed5faa99e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"766f8fe1-2a61-4664-a9be-83f0177cc298\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Sparse Attention\"\n ],\n [\n \"subsubsection\",\n \"Position-based Sparse Attention\"\n ],\n [\n \"paragraph\",\n \"4.1.1.2 Compound Sparse Attention\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"4.1.1.2 Compound Sparse Attention\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 1484,\n 1459\n ],\n \"content\": \"Apart from the above patterns, some existing studies have explored extended sparse patterns for specific data types.\\nFor text data, BP-Transformer~ constructs a binary tree where all tokens are leaf nodes and the internal nodes are span nodes containing many tokens. The edges in this graph are constructed so that each leaf node is connected to its neighbor leaf nodes and higher-level span nodes containing tokens from a longer distance. This approach can be seen as an extension of global attention, where global nodes are hierarchically organized and any pair of tokens are connected with paths in the binary tree. An abstract view of this method is illustrated in Fig. \\\\ref{fig:bpt}.\\nThere are also some extensions for vision data. Image Transformer~ explores two types of attention: (1) flattening image pixels in raster-scan order and then applying block local sparse attention. (2) 2D block local attention, where query blocks and memory blocks are arranged directly in 2D plate, as depicted in Fig. \\\\ref{fig:block_local_2d}. As another example of sparse pattern on vision data, Axial Transformer~ applies independent attention modules over each axis of the image. Each attention module mixes information along one axis while keeping information along the other axis independent, as illustrated in Fig. \\\\ref{fig:axial}. This can be understood as horizontally and vertically flattening image pixels in raster-scan order and then applying strided attention with gaps of image width and height, respectively.\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[BPT\\\\label{fig:bpt}]{\\n\\\\includegraphics[width=0.40\\\\linewidth]{assets/sparse_variants/other/bpt.pdf}\\n}\\\\quad\\n\\\\subfigure[block local (2D)\\\\label{fig:block_local_2d}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/other/block_local.pdf}\\n}\\\\quad\\n\\\\subfigure[axial (2D)\\\\label{fig:axial}]{\\n\\\\includegraphics[width=0.18\\\\linewidth]{assets/sparse_variants/other/axial.pdf}\\n}\\n\\\\caption{Other types of sparse attentions. The red box indicates the query position, and the orange nodes/squares means corresponding tokens are attended to by the query.}\\\\label{fig:other_sparse_attn}\\n\\\\end{center}\\n\\\\end{figure}\",\n \"id\": \"8a0b3727-3354-4028-870f-983c49224baa\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"766f8fe1-2a61-4664-a9be-83f0177cc298\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Sparse Attention\"\n ],\n [\n \"subsubsection\",\n \"Position-based Sparse Attention\"\n ],\n [\n \"paragraph\",\n \"4.1.1.3 Extended Sparse Attention\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"4.1.1.3 Extended Sparse Attention\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1453,\n 1473,\n 7362,\n 794\n ],\n \"content\": \"Another line of work creates a sparse graph based on input content, i.e., the sparse connections are conditioned on inputs.\\nA straightforward way of constructing a content-based sparse graph is to select those keys that are likely to have large similarity scores with the given query. To efficiently construct the sparse graph, we can recur to Maximum Inner Product Search (MIPS) problem, where one tries to find the keys with maximum dot product with a query without computing all dot product terms. Routing Transformer~ uses k-means clustering to cluster both queries $\\\\{\\\\bq_i\\\\}_{i=1}^T$ and keys $\\\\{\\\\bk_i\\\\}_{i=}^T$ on the same set of centroid vectors $\\\\{\\\\mathbf\\\\mu_i\\\\}_{i=1}^k$. Each query only attends to the keys that belong to the same cluster.\\nDuring training, the cluster centroid vectors are updated using the exponentially moving average of vectors assigned to it, divided by the exponentially moving average of cluster counts:\\n\\\\begin{align}\\n\\\\tilde\\\\mu&\\\\leftarrow \\\\lambda\\\\tilde\\\\mu+(1-\\\\lambda)\\\\left(\\\\sum_{i:\\\\mu(\\\\bq_i)=\\\\mu}\\\\bq_i+\\\\sum_{j:\\\\mu(\\\\bk_j)=\\\\mu}\\\\bk_j\\\\right),\\\\\\\\\\nc_\\\\mu&\\\\leftarrow \\\\lambda c_\\\\mu+(1-\\\\lambda)|\\\\mu|,\\\\\\\\\\n\\\\mu&\\\\leftarrow \\\\frac{\\\\tilde\\\\mu}{c_\\\\mu},\\n\\\\end{align}\\nwhere $|\\\\mu|$ denotes the number of vectors currently in cluster $\\\\mu$ and $\\\\lambda\\\\in(0,1)$ is a hyperparameter.\\nLet $\\\\mathcal{P}_i$ denote the set of indices of keys that the $i$-th query attend to. $\\\\mathcal{P}_i$ in Routing Transformer is defined as\\n\\\\begin{equation}\\n \\\\mathcal{P}_i=\\\\{j: \\\\mu(\\\\bq_i)=\\\\mu(\\\\bk_j) \\\\}.\\n\\\\end{equation}\\nReformer~ uses locality-sensitive hashing (LSH) to select key-value pairs for each query. The proposed LSH attention allows each token to attend only to the tokens within the same hashing bucket. The basic idea is to use an LSH function to hash queries and keys into several buckets, with similar items fall in the same bucket with high probability. Specifically, they use the random matrix method for the LSH function. Let $b$ be the number of buckets, given a random matrix $R$ of size $[D_k,b/2]$, the LSH function is computed by :\\n\\\\begin{equation}\\n h(x)=\\\\argmax([xR;-xR]).\\n\\\\end{equation}\\nThe LSH attention allows the $i$-th query to attend only to key-value pairs with indices\\n\\\\begin{equation}\\n \\\\mathcal{P}_i=\\\\{j: h(\\\\bq_i)=h(\\\\bk_j) \\\\}.\\n\\\\end{equation}\\nSparse Adaptive Connection (SAC)~ views the input sequence as a graph and learns to construct attention edges to improve task-specific performances using an adaptive sparse connection. SAC uses an LSTM edge predictor to construct edges between tokens. With no ground truth for edges, the edge predictor is trained with reinforcement learning.\\nSparse Sinkhorn Attention~ first splits queries and keys into several blocks and assigns a key block to each query block. Each query is only allowed to attend to the keys in the key block that is assigned to its corresponding query block. The assignment of key blocks is controlled by a sorting network, which uses Sinkhorn normalization to produce a doubly stochastic matrix as the permutation matrix representing the assignment. They use this content-based block sparse attention along with block local attention introduced in Sec.~\\\\ref{sec:pos_based} to enhance the ability of the model to model locality.\",\n \"id\": \"6429c3b2-0cb9-415c-9f1d-1f7bd028a380\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"c3955c2b-4114-49e4-bbb8-30061d5f27e8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Sparse Attention\"\n ],\n [\n \"subsubsection\",\n \"Content-based Sparse Attention\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Content-based Sparse Attention\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:kernel}\\nAssuming $\\\\bQ,\\\\bK,\\\\bV\\\\in \\\\mathbb{R}^{T\\\\times D}$, the complexity of computing $\\\\softmax(\\\\bQ\\\\bK^\\\\top)\\\\bV$ is quadratic w.r.t. sequence length $T$, as illustrated in Fig. \\\\ref{fig:standard_complexity}. If $\\\\softmax(\\\\bQ\\\\bK^\\\\top)$ can be disentangled into $\\\\bQ'\\\\bK'^\\\\top$, we can compute $\\\\bQ'\\\\bK'^\\\\top\\\\bV$ in reversed order (i.e., $\\\\bQ'\\\\left(\\\\textcolor[rgb]{0,0,1}{\\\\bK'^\\\\top\\\\bV}\\\\right)$), leading to a complexity of $\\\\mathcal{O}(T)$.\\nLet $\\\\hat{\\\\bA} = \\\\exp(\\\\bQ\\\\bK^\\\\top)$ denote un-normalized attention matrix, and $\\\\exp(\\\\cdot)$ is applied element-wise, the regular attention can be rewritten as $\\\\bZ={\\\\bD}^{-1}\\\\hat{\\\\bA}\\\\bV$, where $\\\\bD = \\\\mathrm{diag}(\\\\hat{\\\\bA}\\\\mathbf{1}_T^\\\\top)$; $\\\\mathbf{1}_T^\\\\top$ is the all-ones column vector of length $T$; $\\\\mathrm{diag}(\\\\cdot)$ is a diagonal\\nmatrix with the input vector as the diagonal.\\nLinearized attention is a class of methods that approximate or replace the unnormalized attention matrix $\\\\exp(\\\\bQ\\\\bK^\\\\top)$ with $\\\\phi(\\\\bQ)\\\\phi(\\\\bK)^\\\\top$, where $\\\\phi$ is a feature map that is applied in row-wise manner. Hence the computation of unnormalized attention matrix can be linearized by computing $\\\\phi(\\\\bQ)\\\\left(\\\\textcolor[rgb]{0,0,1}{\\\\phi(\\\\bK)^\\\\top\\\\bV}\\\\right)$\\\\footnote{Similarly, the partition term $\\\\bD$ can be computed with $\\\\phi(\\\\bQ)\\\\left(\\\\textcolor[rgb]{0,0,1}{\\\\phi(\\\\bK)^\\\\top\\\\mathbf{1}_T^\\\\top}\\\\right)$ in linear time.}, as illustrated in Fig. \\\\ref{fig:linearized_complexity}.\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[standard self-attention]{\\\\label{fig:standard_complexity}\\n\\\\includegraphics[width=0.45\\\\textwidth]{assets/kernel/vanilla.pdf}\\n}\\\\hspace{.2in}\\n\\\\subfigure[linearized self-attention]{\\\\label{fig:linearized_complexity}\\n\\\\includegraphics[width=0.45\\\\textwidth]{assets/kernel/kernel.pdf}\\n}\\n\\\\caption{Illustration of complexity difference between standard self-attention and linearized self-attention.}\\n\\\\end{center}\\n\\\\end{figure}\\nTo gain further insights into linearized attention, we derive the formulation in vector form. We consider a general form of attention\\n\\\\begin{equation}\\n \\\\bz_i=\\\\sum_j \\\\frac{\\\\mathrm{sim}(\\\\bq_i, \\\\bk_j)}{\\\\sum_{j'}\\\\mathrm{sim}(\\\\bq_i, \\\\bk_{j'})}\\\\bv_j,\\n\\\\end{equation}\\nwhere $\\\\mathrm{sim}(\\\\cdot,\\\\cdot)$ is a scoring function measuring similarity between input vectors. In vanilla Transformer, the scoring function is the exponential of inner product $\\\\exp(\\\\langle\\\\cdot,\\\\cdot\\\\rangle)$. A natural choice of $\\\\mathrm{sim}(\\\\cdot,\\\\cdot)$ is a kernel function $\\\\mathcal{K}(\\\\bx,\\\\by)=\\\\phi(\\\\bx)\\\\phi(\\\\by)^\\\\top$, which leads to\\n\\\\begin{align}\\n \\\\bz_i&=\\\\sum_j \\\\frac{\\\\phi(\\\\bq_i) \\\\phi(\\\\bk_j)^\\\\top}{\\\\sum_{j'} \\\\phi(\\\\bq_i) \\\\phi(\\\\bk_{j'})^\\\\top}\\\\bv_j\\\\\\\\\\n &=\\\\frac{\\\\phi(\\\\bq_i) \\\\textcolor[rgb]{0,0,1}{\\\\sum_j\\\\phi(\\\\bk_j)\\\\otimes\\\\bv_j}}{\\\\phi(\\\\bq_i)\\\\textcolor[rgb]{0,0,1}{\\\\sum_{j'} \\\\phi(\\\\bk_{j'})^\\\\top}},\\\\label{eq:linearized}\\n\\\\end{align}\\nwhere $\\\\otimes$ denotes outer product of vectors. Based on this formulation, attention can be linearized by first computing the highlighted terms $\\\\sum_j\\\\phi(\\\\bk_j)\\\\otimes\\\\bv_j$ and$\\\\sum_{j'} \\\\phi(\\\\bk_{j'})^\\\\top$. This could be especially beneficial for autoregressive attention, as the cumulative sums $\\\\bS_i=\\\\sum_{j=1}^{i} \\\\phi(\\\\bk_j)\\\\otimes\\\\bv_j$ and $\\\\bu_i=\\\\sum_{j=1}^{i} \\\\phi(\\\\bk_j)$ can be computed from $\\\\bS_{i-1}$ and $\\\\bu_{i-1}$ in constant time. The effectively enables Transformer decoders to run like RNNs.\\nAn interpretation of Eq. \\\\eqref{eq:linearized} is that the model maintains a \\\\textit{memory matrix} by aggregating \\\\textit{associations} represented by outer products of (feature mapped) keys and values, and then retrieve a value by multiplying the memory matrix with feature mapped query with proper normalization. There are two key components in this approach: (1) feature map $\\\\phi(\\\\cdot)$, and (2) aggregation rule.\",\n \"id\": \"a4f89d55-c38a-442a-8744-5625c30e01ff\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Linearized Attention\"\n ]\n ],\n \"subsections\": [\n \"c219036a-ca77-49bd-b9c6-4ac269a7d3d6\",\n \"22632f15-ca9f-4bd6-a9a4-e425db6125fa\"\n ],\n \"title\": \"Linearized Attention\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 8384,\n 1494,\n 1182,\n 798\n ],\n \"content\": \"Linear Transformer~ propose to use a simple feature map $\\\\phi_i(\\\\bx)=\\\\mathrm{elu}(x_i)+1$. This feature map does not aim to approximate dot product attention, but is empirically proved to perform on par with the standard Transformer.\\nPerformer~ uses random feature maps that approximate the scoring function of Transformer. The random feature maps take functions $f_1,\\\\cdots,f_l:\\\\mathbb R\\\\rightarrow \\\\mathbb R$ and $h:\\\\mathbb{R}^D\\\\rightarrow \\\\mathbb R$.\\n\\\\begin{equation}\\n \\\\phi(\\\\bx) = \\\\frac{h(\\\\bx)}{\\\\sqrt{m}}[f_1(\\\\omega_1^\\\\top \\\\bx),\\\\cdots,f_m(\\\\omega_m^\\\\top \\\\bx),\\\\cdots,f_l(\\\\omega_1^\\\\top \\\\bx),\\\\cdots,f_l(\\\\omega_m^\\\\top \\\\bx)],\\n\\\\end{equation}\\nwhere $\\\\omega_1,\\\\cdots,\\\\omega_m\\\\stackrel{\\\\text{iid}}{\\\\sim} \\\\mathcal D$ are drawn from some distribution $\\\\mathcal D\\\\in\\\\mathcal P(\\\\mathbb{R}^D)$.\\nThe first version of Performer~ is inspired from the random Fourier feature map~ that was originally used to approximate Gaussian kernel. It uses trigonometric functions with $h(\\\\bx)=\\\\exp(\\\\frac{\\\\|\\\\bx\\\\|^2}{2}), l=2, f_1=\\\\sin, f_2=\\\\cos$. This approach has also been used in Random Feature Attention (RFA)~, with the difference that $h(\\\\bx)$ is set to $1$ as the queries and keys are $\\\\ell_2$-normalized before applying the feature map.\\nAlthough the trigonometric random feature map leads to an unbiased approximation, it does not guarantee non-negative attention scores and thus could lead to unstable behaviors and abnormal behaviors. To mitigate this issue, the second version of Performer~ proposes positive random feature maps, which uses $h(\\\\bx)=\\\\exp(-\\\\frac{\\\\|\\\\bx\\\\|^2}{2}),l=1,f_1=\\\\exp$ and thus guarantees unbiased and non-negative approximation of dot-product attention. This approach is more stable than and reports better approximation results.\\nIn addition to using random feature maps to approximate standard dot product attention, and also explore approximating order-1 arc-cosine kernel with $h(\\\\bx)=1,l=1,f_1=\\\\mathrm{ReLU}$. This feature map has been show to be effective in various tasks including machine translation and protein sequence modeling.\\n design a feature map that aims at facilitating orthogonality in feature space. Specifically, given an input $\\\\bx\\\\in\\\\mathbb{R}^D$, the feature map $\\\\phi:\\\\mathbb{R}^D\\\\rightarrow \\\\mathbb{R}^{2\\\\nu D}$ is defined by the partial function\\n\\\\begin{equation}\\n \\\\phi_{i+2(j-1)D}(\\\\bx)=\\\\mathrm{ReLU}([\\\\bx,-\\\\bx])_i\\\\mathrm{ReLU}([\\\\bx,-\\\\bx])_{i+j}\\\\quad\\\\text{for }i=1,\\\\cdots,2D,j=1,\\\\cdots,\\\\nu.\\n\\\\end{equation}\",\n \"id\": \"c219036a-ca77-49bd-b9c6-4ac269a7d3d6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"a4f89d55-c38a-442a-8744-5625c30e01ff\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Linearized Attention\"\n ],\n [\n \"subsubsection\",\n \"Feature Maps\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Feature Maps\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 8384,\n 1494,\n 1182,\n 798\n ],\n \"content\": \"In Eq. \\\\eqref{eq:linearized} the associations $\\\\{\\\\phi(\\\\bk)_j\\\\otimes\\\\bv_j\\\\}$ are aggregated into the memory matrix by simple summation. This is adopted by several studies~. However, it could be more beneficial for the network to selectively drop associations as new associations are added to the memory matrix.\\nRFA~ introduces a gating mechanism to the summation to model local dependency in sequence data. Specifically, when adding a new association to the memory matrix $\\\\bS$, at a particular time step, they weigh $\\\\bS$ by a learnable, input-dependent scalar $g$, and the new association by $(1-g)$ (and a similar mechanism to $\\\\bu$). With this modification, history associations are exponentially decayed and recent context is favored in each timestep.\\n argue that simple summation limits the capacity of the memory matrix and thus propose to enlarge the capacity in a write-and-remove fashion. Specifically, given a new input key-value pair $(\\\\bk_i, \\\\bv_i)$, the model first retrieve the value $\\\\bar{\\\\bv}_i$ currently associated with $\\\\bk_i$ using matrix multiplication. It then writes to the memory matrix a convex combination of $\\\\bar{\\\\bv}_i$ and $\\\\bv_i$, using a input-dependent gating scalar $g$, and removes the association $\\\\bar{\\\\bv}_i$. They also propose \\\\textit{sum normalization} (normalizing $\\\\phi(\\\\bq_i),\\\\phi(\\\\bk_i)$ by the sum of their components before updating the memory matrix) instead of normalizing with the denominator in Eq. \\\\eqref{eq:linearized} for this aggregation rule.\",\n \"id\": \"22632f15-ca9f-4bd6-a9a4-e425db6125fa\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"a4f89d55-c38a-442a-8744-5625c30e01ff\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Linearized Attention\"\n ],\n [\n \"subsubsection\",\n \"Aggregation Rule\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Aggregation Rule\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Apart from using sparse attention or kernel-based linearized attention, one could also reduce the complexity of attention by reducing the number of queries or key-value pairs, which leads to \\\\textit{query prototyping} and \\\\textit{memory compression}\\\\footnote{The key-value pairs are often referred to as a key-value memory (hence the name memory compression).} methods, respectively.\",\n \"id\": \"99d263f7-48ac-4be9-9f96-b1cd4380de1b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Query Prototyping and Memory Compression\"\n ]\n ],\n \"subsections\": [\n \"ad89ae14-36aa-4aeb-ab1d-65a0637088d3\",\n \"5fdbbc99-3354-4d28-a5a0-c76985e76cdb\"\n ],\n \"title\": \"Query Prototyping and Memory Compression\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1479,\n 1472\n ],\n \"content\": \"In query prototyping, several prototypes of queries serve as the main source to compute attention distributions. The model either copies the distributions to the positions of represented queries or filling those positions with discrete uniform distributions. Fig. \\\\ref{fig:query_prototype} illustrates the computing flow of query prototyping.\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[Query prototyping]{\\\\label{fig:query_prototype}\\n\\\\includegraphics[height=1.5in]{assets/prototype/query_prototype.pdf}\\n}\\\\hspace{.2in}\\n\\\\subfigure[Memory compression]{\\\\label{fig:compressed_mem}\\n\\\\includegraphics[height=1.5in]{assets/prototype/mem_compress.pdf}\\n}\\n\\\\caption{Query prototyping and memory compression.}\\\\label{fig:prototype_mem_compress}\\n\\\\end{center}\\n\\\\end{figure}\\nClustered Attention~ groups queries into several clusters and then computes attention distributions for cluster centroids. All queries in a cluster share the attention distribution calculated with the corresponding centroid.\\nInformer~ selects prototypes from queries using explicit query sparsity measurement, which is derived from an approximation of the Kullback-Leibler divergence between the query's attention distribution and the discrete uniform distribution. Attention distributions are then only calculated for the top-$u$ queries under query sparsity measurement. The rest of the queries are assigned with discrete uniform distributions.\",\n \"id\": \"ad89ae14-36aa-4aeb-ab1d-65a0637088d3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"99d263f7-48ac-4be9-9f96-b1cd4380de1b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Query Prototyping and Memory Compression\"\n ],\n [\n \"subsubsection\",\n \"Attention with Prototype Queries\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attention with Prototype Queries\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 7333,\n 1503,\n 1133,\n 1504,\n 1482\n ],\n \"content\": \"\\\\label{sec:compressed_mem}\\nApart from decreasing the number of queries with query prototyping, one can also reduce the complexity by reducing the number of the key-value pairs before applying the attention mechanism, as depicted in Fig. \\\\ref{fig:compressed_mem}.\\n propose Memory Compressed Attention (MCA) that reduces the number of keys and values using a strided convolution. This modification is used as a complement to local attention proposed in the same work (as discussed in Sec.~\\\\ref{sec:sparseattn}), in that it can capture global context. The mechanism reduces the number of keys and values by a factor of kernel size $k$ and thus allowing to process significantly longer sequences than vanilla Transformer given the same computation resources.\\nSet Transformer~ and Luna~ use a number of external trainable global nodes to summarize information from inputs and then the summarized representations serve as a compressed memory that the inputs attend to. This reduces the quadratic complexity of self-attention to linear complexity w.r.t. sequence length.\\nLinformer~ utilizes linear projections to project keys and values from length $n$ to a smaller length $n_k$. This also reduces the complexity of self-attention to linear. The drawback of this approach is that an input sequence length has to be assumed and hence it cannot be used in autoregressive attention.\\nPoolingformer~ adopts two-level attention that combines a sliding window attention and a compressed memory attention. The compressed memory module is used after the sliding window attention to increase the receptive field. They explore a few different pooling operations as the compression operation to compress the number of keys and values, including max pooling and pooling with Dynamic Convolution~.\",\n \"id\": \"5fdbbc99-3354-4d28-a5a0-c76985e76cdb\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"99d263f7-48ac-4be9-9f96-b1cd4380de1b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Query Prototyping and Memory Compression\"\n ],\n [\n \"subsubsection\",\n \"Attention with Compressed Key-Value Memory\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attention with Compressed Key-Value Memory\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7333\n ],\n \"content\": \"Some empirical and theoretical analyses~ report the self-attention matrix $\\\\bA \\\\in \\\\mathbb{R}^{T\\\\times T}$ is often low-rank\\\\footnote{The rank of $\\\\bA$ is far lower than input length $T$.}. The implications of this property are twofold: (1) The low-rank property could be explicitly modeled with parameterization; (2) The self-attention matrix could be replaced by a low-rank approximation.\",\n \"id\": \"9a2c3eb1-bf66-4ddc-822d-bcc1121a8120\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Low-rank Self-Attention\"\n ]\n ],\n \"subsections\": [\n \"8570c9f2-0cf3-47a3-befe-ca2093b97f7a\",\n \"fc87c410-c4a6-46de-806f-b6f8720b2eb4\"\n ],\n \"title\": \"Low-rank Self-Attention\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The fact that the rank of the attention matrix is less than sequence length implies that, for scenarios where the inputs are typically short, setting $D_k>T$ would be more than an over-parameterization and lead to overfitting. It is thus reasonable to limit the dimension of $D_k$ to explicitly model the low-rank property as an inductive bias. decompose self-attention matrix into a low-rank attention module with small $D_k$ that captures long-range non-local interactions, and a band attention module that captures local dependencies.\",\n \"id\": \"8570c9f2-0cf3-47a3-befe-ca2093b97f7a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"9a2c3eb1-bf66-4ddc-822d-bcc1121a8120\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Low-rank Self-Attention\"\n ],\n [\n \"subsubsection\",\n \"Low-rank Parameterization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Low-rank Parameterization\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 1494\n ],\n \"content\": \"Another implication of the low-rank property of the attention matrix is that one can use a low-rank matrix approximation to reduce the complexity of self-attention. A closely related methodology is the low-rank approximation of kernel matrices. We believe some existing works are inspired by kernel approximation.\\nSome of the aforementioned linearized attention methods in Sec.~\\\\ref{sec:kernel} are inspired from kernel approximation with random feature maps. For example, Performer~ follows the Random Fourier feature map originally proposed to approximate Gaussian kernels. The method first decomposes the attention distribution matrix $\\\\bA$ into $\\\\bC_Q\\\\bG\\\\bC_K$ where $\\\\bG$ is a Gaussian kernel matrix and the random feature map is used to approximate $\\\\bG$.\\nAnother line of work follow the idea of Nystr\\\\\\\"om method. These Nystr\\\\\\\"om-based methods~ first select $m$ landmark nodes from the $T$ inputs with down-sampling methods (e.g., strided average pooling). Let $\\\\tilde\\\\bQ,\\\\tilde\\\\bK$ be the selected landmark queries and keys, then the follow approximation is used in the attention computation\\n\\\\begin{equation}\\\\label{eq:nystrom}\\n \\\\tilde{\\\\bA}=\\\\mathrm{softmax}\\\\left(\\\\bQ\\\\tilde{\\\\bK}^\\\\top\\\\right)\\\\left(\\\\mathrm{softmax}\\\\left(\\\\tilde\\\\bQ\\\\tilde{\\\\bK}^\\\\top\\\\right)\\\\right)^{-1}\\\\mathrm{softmax}\\\\left(\\\\tilde\\\\bQ\\\\bK^\\\\top\\\\right).\\n\\\\end{equation}\\nNote that $\\\\bM^{-1}=\\\\left(\\\\mathrm{softmax}\\\\left(\\\\tilde\\\\bQ\\\\tilde{\\\\bK}^\\\\top\\\\right)\\\\right)^{-1}$ in Eq. \\\\eqref{eq:nystrom} does not always exist. To mitigate this issue, CSALR~ adds an identity matrix to $\\\\bM$ to make sure that the inverse always exists. Nystr{\\\\\\\"{o}}mformer~ uses the Moore-Penrose pseudoinverse of $\\\\bM$ instead of the inverse so that the approximation can be made for cases where $\\\\bM$ is singular.\",\n \"id\": \"fc87c410-c4a6-46de-806f-b6f8720b2eb4\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"9a2c3eb1-bf66-4ddc-822d-bcc1121a8120\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Low-rank Self-Attention\"\n ],\n [\n \"subsubsection\",\n \"Low-rank Approximation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Low-rank Approximation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\begin{figure}[htbp]\\n \\\\centering\\n \\\\includegraphics[width=0.5\\\\linewidth]{assets/attention_w_prior.pdf}\\n \\\\caption{Attention with prior. This type of model fuse generated attention scores with pior attention scores, producing the final attention scores for attention computation.}\\n \\\\label{fig:attention_w_prior}\\n\\\\end{figure}\\nAttention mechanism generally outputs an expected attended value as a weighted sum of vectors, where the weights are an attention distribution over the values. Traditionally, the distribution is generated from inputs (e.g., $\\\\mathrm{softmax}(\\\\bQ\\\\bK^\\\\top)$ in vanilla Transformer). As a generalized case, attention distribution can also come from other sources, which we refer to as \\\\textit{prior}. Prior attention distribution can be a supplement or substitute for distribution generated from inputs. We abstract this formulation of attention as \\\\textit{attention with prior}, as depicted in Fig. \\\\ref{fig:attention_w_prior}. In most cases, the fusion of two attention distribution can be done by computing a weighted sum of the scores corresponding to the prior and generated attention before applying softmax.\",\n \"id\": \"afe04fc5-1717-4f24-b061-7af855828533\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Attention with Prior\"\n ]\n ],\n \"subsections\": [\n \"65f37434-480a-4748-b65e-d8966645bec9\",\n \"af364707-c13f-44cb-9e4e-2366287a19b6\",\n \"a795db35-abb3-4e07-ad32-5c47de43f06f\",\n \"55c07bfe-1106-456b-a01c-8edd7f25017b\"\n ],\n \"title\": \"Attention with Prior\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1489\n ],\n \"content\": \"Some types of data (e.g., text) can exhibit a strong preference for the locality. This property can be explicitly encoded as a prior attention. A simple method would be to use a Gaussian distribution over positions. Specifically, one could multiply the generated attention distribution with some Gaussian density and then renormalize, which is equivalent to adding to the generated attention scores $\\\\bA$ a bias term $\\\\bG$, where higher $G_{ij}$ indicates a higher prior probability that the $i$-th input attend to the $j$-th input.\\n proposes to first predict a central position $p_i$ for each $\\\\bq_i$ using a simple feed-forward network. The Gaussian bias is then defined to be\\n\\\\begin{equation}\\n G_{ij}=-\\\\frac{(j-p_i)^2}{2\\\\sigma^2},\\n\\\\end{equation}\\nwhere $\\\\sigma$ denotes standard deviation for the Gaussian and can be determined as a hyperparameter or predicted from inputs.\\nGaussian Transformer~ assumes the central position to be $i$ for each $\\\\bq_i$ and defines the bias to bes\\n\\\\begin{equation}\\n G_{ij}=-|w(i-j)^2+b|,\\n\\\\end{equation}\\nwhere $w\\\\ge 0, b\\\\le 0$ are scalar parameters that controls the deviation and reduce the weight for central position, respectively.\",\n \"id\": \"65f37434-480a-4748-b65e-d8966645bec9\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"afe04fc5-1717-4f24-b061-7af855828533\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Attention with Prior\"\n ],\n [\n \"subsubsection\",\n \"Prior that Models locality\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Prior that Models locality\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 1497,\n 1492\n ],\n \"content\": \"\\\\label{sec:prior_prev_module}\\nIn Transformer architecture, it is often observed the attention distributions are similar in adjacent layers. It is thus natural to provide attention distribution from previous layer as a prior for attention computation. The final attention scores can be defined as\\n\\\\begin{equation}\\\\label{crosslayer}\\n \\\\hat{\\\\bA}^{(l)} = w_1\\\\cdot\\\\bA^{(l)}+w_2\\\\cdot g(\\\\bA^{(l-1)}),\\n\\\\end{equation}\\nwhere $\\\\bA^{(l)}$ denotes the attention scores of the $l$-th layer, $w_1,w_2\\\\in\\\\mathbb R$ are weight applied to the scores from adjacent layers, and $g:\\\\mathbb{R}^{n\\\\times n}\\\\rightarrow \\\\mathbb{R}^{n\\\\times n}$ is a function that translate previous scores to the prior to be applied.\\nPredictive Attention Transformer~ proposes to apply a 2D-convolutional layer to previous attention scores and compute the final attention scores as a convex combination of the generated attention scores and the convolved scores. This is equivalent to setting $w_1=\\\\alpha,w_2=1-\\\\alpha$ and $g(\\\\cdot)$ to be a convolutional layer in Eq. \\\\eqref{crosslayer}. They experiment training such a model from scratch and finetune after adapting the pre-trained BERT model, and both sets of experiments show improvements over baseline models.\\nRealformer~ uses adds the previous attention scores directly to the generated attention scores, thus resembles a residual skip connection on attention maps. It's equivalent to setting $w_1=w_2=1$ and $g(\\\\cdot)$ to be identity map in Eq. \\\\eqref{crosslayer}. They conduct pre-training experiments on this model. The results show that this model outperforms the baseline BERT model in multiple datasets and surpasses the baseline model even when pre-training budgets are significantly lower.\\nAs an extreme case, Lazyformer~ proposes to share attention maps between a number of adjacent layers. This is equivalent to setting $g(\\\\cdot)$ to identity and switch the settings of $w_1=0,w_2=1$ and $w_1=1,w_2=0$ alternatingly. The benefit of this approach is that the attention maps are computed only once and reused several times in the succeeding layers, thus reducing the computation cost. Their pre-training experiments show that the resulting model remains effective while being much more efficient to compute.\",\n \"id\": \"af364707-c13f-44cb-9e4e-2366287a19b6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"afe04fc5-1717-4f24-b061-7af855828533\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Attention with Prior\"\n ],\n [\n \"subsubsection\",\n \"Prior from Lower Modules\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Prior from Lower Modules\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 725,\n 1505\n ],\n \"content\": \"Adapters are task-dependent, trainale modules that are attached in specific locations of a pre-trained network for cross-task efficient parameter sharing~. propose a Conditionally Adaptive Multi-Task Learning (CAMTL) framework that uses a trainable attention prior $M(\\\\bz_i)$ that depends on task encoding $\\\\bz_i\\\\in\\\\mathbb{R}^{D_z}$\\n\\\\begin{equation}\\n M(\\\\bz_i)=\\\\bigoplus_{j=1}^m A'_j(\\\\bz_i),\\\\quad A'_j(\\\\bz_i)=A_j\\\\gamma_i(\\\\bz_i)+\\\\beta_i(\\\\bz_i),\\n\\\\end{equation}\\nwhere $\\\\bigoplus$ denotes direct sum, $A_j\\\\in \\\\mathbb{R}^{(n/m)\\\\times(n/m)}$ are trainable parameters, and $\\\\gamma_j,\\\\beta_j:\\\\mathbb{R}^{D_z}\\\\rightarrow \\\\mathbb{R}^{(n/m)\\\\times(n/m)}$ are are Feature Wise\\nLinear Modulation functions~. A maximum sequence length $n_{max}$ is specified in implementation. The prior is formulated as a block diagonal matrix and added to the attention scores of upper layers in pre-trained Transformers to serve as an adapter for parameter-efficient multi-task inductive knowledge transfer.\",\n \"id\": \"a795db35-abb3-4e07-ad32-5c47de43f06f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"afe04fc5-1717-4f24-b061-7af855828533\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Attention with Prior\"\n ],\n [\n \"subsubsection\",\n \"Prior as Multi-task Adapters\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Prior as Multi-task Adapters\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 1489,\n 7364,\n 7366,\n 1476\n ],\n \"content\": \"Some works have explored using an attention distribution that is independent of pair-wise interaction between inputs. In other words, their models exploit only a prior attention distribution.\\n design an efficient Transformer decoder variant called average attention network that uses a discrete uniform distribution as the sole source of attention distribution. The values are thus aggregated as a cumulative-average of all values. To improve the expressiveness of the network, they further adds a feed-forward\\ngating layer on top of the average attention module. The advantage of this approach is that the adapted Transformer decoder can train in a parallel manner as usual Transformers do and decode like an RNN, thus avoiding the $\\\\mathcal O(T^2)$ complexity in decoding.\\n utilize a Gaussian distribution as the hardcoded attention distribution for attention calculation. The intuition is very similar to and in that attention distribution should be focused on a certain local window. Distinctively, they drop the generated attention completely and use only the Gaussian distribution for attention computation. In this approach, the mean (central position) and variance are designed to be hyperparameters. The experiments show that the hardcoded attention, when applied only to self-attention, can achieve comparable performance to the baseline model in machine translation tasks.\\nSynthesizer~ proposes to replace generated attention scores with: (1) a learnable, randomly initialized attention scores, and (2) attention scores output by a feed-forward network that is only conditioned on the querying input itself. The experiments on machine translation and language modeling show that these variants can achieve competitive performance with vanilla Transformer. It is not explained why these variants work but the empirical results are intriguing.\",\n \"id\": \"55c07bfe-1106-456b-a01c-8edd7f25017b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"afe04fc5-1717-4f24-b061-7af855828533\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Attention with Prior\"\n ],\n [\n \"subsubsection\",\n \"Attention with Only Prior\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attention with Only Prior\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Multi-head attention is appealing for the ability to jointly attend to information from different representation subspaces at different positions. However, there is no mechanism to guarantee that different attention heads indeed capture\\ndistinct features.\",\n \"id\": \"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"57f59954-45c7-48ac-baf2-dc92c147a85e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Improved Multi-Head Mechanism\"\n ]\n ],\n \"subsections\": [\n \"dcd668ac-e997-46ab-8e9e-c34fd7ac742a\",\n \"6b049103-9bb4-4d9c-bd95-dceaba44ed6d\",\n \"c68f2053-1c0a-4bd7-8d8a-92474761c343\",\n \"be224101-c6fc-4828-9ad0-f1f79c3097ee\"\n ],\n \"title\": \"Improved Multi-Head Mechanism\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1474,\n 7372,\n 1460,\n 1498,\n 7369,\n 38\n ],\n \"content\": \"A basic motivation for using multi-head attention is to allow the model to jointly attend to information from different representation subspaces at different positions~. However, in vanilla Transformer there is no explicit mechanism to guarantee different behavior across attention heads, nor is there any mechanism for heads to interact with each other. A line of work is dedicated to improving multi-head mechanism by introducing incorporating more sophisticated mechanisms that guide the behavior of different attention heads or allow interaction across attention heads.\\n introduce an auxiliary disagreement regularization term into loss function to encourage diversity among different attention heads. Two regularization terms are respectively to maximize cosine distances of the input subspaces and output representations, while the last one is to disperse\\nthe positions attended by multiple heads with element-wise multiplication of the corresponding attention matrices.\\nSeveral probing works have revealed that pre-trained Transformer models exhibit certain patterns of self-attention that are of little linguistic backing. As a representative work, identify several simple attention patterns in BERT. For instance, many of the attention heads simply pay attention to special BERT tokens \\\\texttt{[CLS]} and \\\\texttt{[SEP]}. As a result, some constraints can be introduced to boost the training of Transformer models. To this end, propose to use an auxiliary loss, which is defined to be the Frobenius norm between attention distribution maps and predefined attention patterns.\\nTalking-head Attention~ uses a talking head mechanism that linearly projects the generated attention scores from $h_k$ to $h$ heads, applies softmax in that space, and then projects to $h_v$ heads for value aggregation. The motivation is to encourage the model to move information between attention heads in a learnable fashion.\\nCollaborative Multi-head Attention~ uses shared query and key projection $\\\\bW^Q$ and $\\\\bW^K$ and a mixing vector $\\\\bm_i$ for the $i$-th head to filter from the projection parameters such that Eq. \\\\eqref{eq:headi} is adapted to\\n\\\\begin{equation}\\n \\\\mathrm{head}_i=\\\\mathrm{Attention}(\\\\bQ\\\\bW^Q\\\\mathrm{diag}(\\\\bm_i),\\\\bK\\\\bW^K, \\\\bV\\\\bW_i^V),\\n\\\\end{equation}\\nwhere $\\\\bW^Q$ and $\\\\bW^K$ are shared by all the attention heads.\",\n \"id\": \"dcd668ac-e997-46ab-8e9e-c34fd7ac742a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Improved Multi-Head Mechanism\"\n ],\n [\n \"subsubsection\",\n \"Head Behavior Modeling\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Head Behavior Modeling\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1465\n ],\n \"content\": \"Vanilla attention adopts full attention spans assume, where a query can attend to all of the key-value pairs. However, it is often observed that some heads focus their attention distribution mainly in a local context while some other heads attend to broader contexts. It could thus be beneficial to restrict the attention spans:\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Locality}. Restricting attention spans induce explicit local constraints. This is advantageous in cases where locality is an important prior.\\n \\\\item \\\\textit{Efficiency}. If implemented appropriately, such a model can scale to very long sequences without introducing additional memory footprint and computational time.\\n\\\\end{itemize}\\nRestricting attention spans can be expressed as multiplying each attention distribution value with a mask value and then re-normalize, where the mask can be expressed as a non-increasing function that maps a distance to a value in $[0,1]$. A vanilla attention assigns a mask value of $1$ for all distances, as depicted in Fig. \\\\ref{span_mask}(a).\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[mask function for vanilla attention]{\\n\\\\includegraphics[width=0.3\\\\linewidth]{assets/span_mask/full_mask.pdf}\\n}\\n\\\\subfigure[mask function for adaptive span]{\\n\\\\includegraphics[width=0.3\\\\linewidth]{assets/span_mask/adaptive_mask.pdf}\\n}\\n\\\\subfigure[mask function for fixed span]{\\n\\\\includegraphics[width=0.3\\\\linewidth]{assets/span_mask/scale.pdf}\\n}\\n\\\\caption{Three types of span masking function $m(x)$. The horizontal axis represents distance $x$ and vertical axis the mask value.}\\\\label{span_mask}\\n\\\\end{center}\\n\\\\end{figure}\\n propose to use a learnable attention span, as depicted in Fig. \\\\ref{span_mask}(b) . The mask is parameterized by a learnable scalar $z$ and a hyperparameter $R$. The experiments on character-level language modeling show that the adaptive-span models outperform baseline models while having significantly fewer FLOPS. It is also observed that lower layers generally have smaller learned spans and higher layers otherwise. This indicates that the model can learn a hierarchical composition of features.\\nMulti-Scale Transformer~ proposes to use a fixed attention span, with different heads in different layers using a different max span. The fixed attention span is depicted in Fig. \\\\ref{span_mask}(c). The attention is restricted within a fixed window which is controlled by a \\\\textit{scale} value $w$. They design the scales from an intuitive linguistic perspective and empirical observation from BERT such that higher layers tend to have more large scales (e.g., large span size), and lower layers should be confined with a smaller scale. Their experiments on several tasks show that the model can outperform baseline models while accelerating inference on long sequences.\",\n \"id\": \"6b049103-9bb4-4d9c-bd95-dceaba44ed6d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Improved Multi-Head Mechanism\"\n ],\n [\n \"subsubsection\",\n \"Multi-head with Restricted Spans\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-head with Restricted Spans\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 306,\n 1452,\n 8455,\n 38\n ],\n \"content\": \"After each attention head computes its output representation, the vanilla multi-head attention~ concatenates these representation and then apply a linear transformation to the concatenated representation to obtain the final output representation, as formulated in Eq. \\\\eqref{eq:multihead}. Combining Eq. \\\\eqref{attention}\\\\eqref{eq:multihead} and \\\\eqref{eq:headi}, one can see that this \\\\textit{concatenate-and-project} formulation is equivalent to summation over $H$ re-parameterized attention outputs. To this end, we first divide $\\\\bW^O\\\\in\\\\mathbb{R}^{D_m\\\\times D_m}$ into $H$ blocks\\n\\\\begin{equation}\\n \\\\bW^O=[\\\\bW_1^O;\\\\bW_2^O;\\\\cdots;\\\\bW_H^O],\\n\\\\end{equation}\\nwhere each $\\\\bW_i^O$ is of dimension $D_v\\\\times D_m$. It's thus easy to see that multi-head attention can be reformulated as\\n\\\\begin{equation}\\n \\\\mathrm{MultiHeadAttn}(Q,K,V)= \\\\sum_{i=1}^H \\\\mathrm{Attention}(Q \\\\bW_i^Q,K \\\\bW_i^K, V\\\\textcolor[rgb]{0,0,1} {\\\\bW_i^V\\\\bW_i^O}).\\n\\\\end{equation}\\nOne might argue that this simple \\\\textit{aggregate-by-summation} paradigm does not fully exploit the expressiveness of multi-head attention and that it is more desirable to use a more complex aggregation.\\n propose to use routing methods, originally proposed for capsule networks~, to further aggregate information produced by different attention heads. The outputs of attention heads are first transformed into input capsules, then output capsules are obtained after the iterative routing process. The output capsules are then concatenated as a final output of multi-head attention. These two works both utilizes two routing mechanisms, namely \\\\textit{dynamic routing} and \\\\textit{EM routing}. One would notice that iterative routing introduces additional parameters and computational overhead. empirically show that applying the routing mechanism only to the lower layers can best balance the translation performance and computational efficiency.\",\n \"id\": \"c68f2053-1c0a-4bd7-8d8a-92474761c343\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Improved Multi-Head Mechanism\"\n ],\n [\n \"subsubsection\",\n \"Multi-head with Refined Aggregation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-head with Refined Aggregation\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1507,\n 1506\n ],\n \"content\": \"Several other modifications to the multi-head mechanism have been proposed to improve multi-head attention.\\n propose multi-query attention, where key-value pairs are shared among attention heads (i.e., to use only one key projection and one value projection for all attention heads). The advantage of this method is that it reduces the memory bandwidth requirements for decoding and results in a model that is faster to decode, while incurring only minor quality degradation from the baseline.\\n establish that small attention key size can affect its ability to represent arbitrary distribution. They thus propose to disentangle head size from the number of heads $h$, as opposed to the common practice that sets the head size to be $D_m/h$. It is observed empirically that setting attention head size to be input sequence length is beneficial.\",\n \"id\": \"be224101-c6fc-4828-9ad0-f1f79c3097ee\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4c615d59-27b6-48ea-a7bd-4c6d27bd91b5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Attention\"\n ],\n [\n \"subsection\",\n \"Improved Multi-Head Mechanism\"\n ],\n [\n \"subsubsection\",\n \"Other Modifications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Other Modifications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:other_module}\",\n \"id\": \"7a323139-7d7b-422d-a6c6-3617457ad2f6\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ]\n ],\n \"subsections\": [\n \"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc\",\n \"ba9ac74a-b1ed-471e-a9a5-9097079a09db\",\n \"e915275c-0207-4658-9d41-438a8b5b3e51\"\n ],\n \"title\": \"Other Module-level Modifications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:pos_rep}\\n\\\\begin{definition}[permutation equivariant function]\\nLet $\\\\Pi_n$ be the set of all permutations of indices $\\\\{1,2,\\\\cdots, T\\\\}$. A function $f:\\\\mathcal{X}^T\\\\rightarrow \\\\mathcal{Y}^T$ is said to be \\\\textit{permutation equivariant} if and only if for any $\\\\pi\\\\in\\\\Pi_T$\\n\\\\begin{equation}\\n f(\\\\pi x)=\\\\pi f(x).\\n\\\\end{equation}\\n\\\\end{definition}\\nIt is easy to verify that Convolution and Recurrence networks are not permutation equivariant. However, both self-attention modules and position-wise feed-forward layers in Transformer are permutation equivariant, which could be a problem when it comes to modeling problems other than \\\\textit{set-input} problems where the structure of inputs is needed. For example, when modeling sequences of text, the ordering of words matters and it's thus crucial to properly encode the positions of words in Transformer architecture. Therefore, additional mechanisms are required to inject positional information into Transformers. A common design is to first represent positional information using vectors and then infuse the vectors to the model as an additional input.\",\n \"id\": \"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"7a323139-7d7b-422d-a6c6-3617457ad2f6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position Representations\"\n ]\n ],\n \"subsections\": [\n \"ccc63091-167f-4bc8-97e9-c63a8adbf741\",\n \"9cdcf5bd-ef00-4ae9-a8b3-6d8cd2240795\",\n \"268772ad-4cd8-43e9-9608-b7ac7de2e35c\",\n \"866b7a57-7c1b-41c7-8a8f-c03a0681cbeb\",\n \"90d9b719-8d67-48fa-a44c-ce04b5d77df6\"\n ],\n \"title\": \"Position Representations\"\n },\n {\n \"cite_extract_rate\": 0.555555555555555,\n \"cites\": [\n 1493,\n 790,\n 7,\n 1450,\n 38\n ],\n \"content\": \"In vanilla Transformer~, positional information is encoded as absolute sinusoidal position encodings.For each position index $t$, the encoding is a vector $\\\\bp_t = \\\\mathrm{PE}(t)\\\\in\\\\mathbb{R}^{D_m}$, of which every element is a sinusoidal ($\\\\sin$/$\\\\cos$) function of the index with pre-defined frequency.\\n\\\\begin{equation}\\\\label{eq:PE}\\n \\\\mathrm{PE}(t)_i = \\\\begin{cases} \\\\sin(\\\\omega_i t) & \\\\text{if }i\\\\text{ is even},\\\\\\\\\\n \\\\cos(\\\\omega_i t) & \\\\text{if }i\\\\text{ is odd},\\n \\\\end{cases}\\n\\\\end{equation}\\nwhere $\\\\omega_i$ is the hand-crafted frequency for each dimension. The position encoding of each position in the sequence is then added to the token embeddings and fed to Transformer.\\nAnother way of representing absolute positions is to learn a set of positional embeddings for each position~. Compared to hand-crafted position representation, learned embeddings are more flexible in that position representation can adapt to tasks through back-propagation. But the number of embeddings is limited up to a maximum sequence length determined before training, which makes this approach no longer \\\\textit{inductive}, i.e., not able to handle sequences longer than sequences seen in the training time.\\n propose to use sinusoidal position representation, but with each frequency $\\\\omega_i$ (in Eq. \\\\eqref{eq:PE}) learned from data. This approach retains inductiveness but is more flexible than hand-crafted sinusoidal encoding. FLOATER~ frames positional representation as a continuous dynamical system and adopts Neural ODE to enable end-to-end training with backpropagation. This method is inductive and flexible while being parameter efficient compared to a fully learnable approach.\\nThe Vanilla approach to incorporating absolute position representations is to add position encodings/embeddings to token embeddings. However, as the input signals propagate through\\nthe layers, the positional information might get lost in the upper layers. Later works find it beneficial to add position representations to inputs to each Transformer layer~.\",\n \"id\": \"ccc63091-167f-4bc8-97e9-c63a8adbf741\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position Representations\"\n ],\n [\n \"subsubsection\",\n \"Absolute Position Representations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Absolute Position Representations\"\n },\n {\n \"cite_extract_rate\": 0.8571428571428571,\n \"cites\": [\n 1456,\n 1508,\n 1481,\n 7370,\n 7054,\n 9\n ],\n \"content\": \"Another line of works focuses on representing positional relationships between tokens instead of positions of individual tokens. The intuition is that in self-attention, pairwise positional relationships between input elements (direction and distance) could be more beneficial than positions of elements. Methods following this principles are called relative positional representation. propose to add a learnable relative position embedding to keys of attention mechanism\\n\\\\begin{align}\\n \\\\bk_j' &= \\\\bk_j + \\\\br_{ij},\\\\ \\\\text{for }i=1,\\\\cdots, n,\\\\\\\\\\n \\\\br_{ij} &= \\\\bR_{\\\\mathrm{clip}(i-j)}\\\\label{rij},\\\\\\\\\\n \\\\mathrm{clip}(x) &= \\\\max(-K, \\\\min(x, K))\\\\label{eq:shaw_clip},\\n\\\\end{align}\\nwhere $\\\\br_{ij}\\\\in\\\\mathbb{R}^{D_k}$ is the relative position embedding for relation between position $i$ and $j$ and $K$ is the largest offset that determines the number of embeddingg. Typically $K$ is set to a length that can accommodate most input sequences. As a special case, InDIGO~ sets $K$ to $3$ for their specially designed framework for non-autoregressive generation. As an incremental effort, Music Transformer~ further introduce a mechanism to reduce the intermediate memory requirements for this approach. Similar to this approach, T5~ adopt a simplified form of relative position embeddings where each embedding is only a learnable scalar that is added to the corresponding score used for computing the attention weights.\\nTransformer-XL~ use a sinusoidal encoding to represent positional relationships but fuses contents and position information by redesign the computation of attention scores\\\\footnote{the scaling factor is omitted without loss of generality.}\\n\\\\begin{equation}\\n \\\\bA_{ij}=\\\\bq_i \\\\bk_j^\\\\top+\\\\bq_i\\\\left(\\\\bR_{i-j}\\\\bW^{K,R}\\\\right)^\\\\top +\\\\bu^1\\\\bk_j^\\\\top+\\\\bu^2\\\\left(\\\\bR_{i-j}\\\\bW_{K,R}\\\\right)^\\\\top,\\n\\\\end{equation}\\nwhere $\\\\bW^{K,R}\\\\in\\\\mathbb{R}^{D_m\\\\times D_k},\\\\bu^1,\\\\bu^2\\\\in\\\\mathbb{R}^{D_k}$ are learnable parameters and $\\\\bR$ is a sinusoidal encoding matrix similar to position encoding in vanilla Transformer. Then softmax function is applied to scores $\\\\bA$ to provide attention weights. Note that the learnable sinusoidal encoding is also a drop-in replacement to hand-crafted $\\\\bR$.\\nDeBERTa~ utilizes position embeddings like and applies the embeddings to the model in a disentangled style similar to Transformer-XL~\\n\\\\begin{equation}\\n \\\\bA_{ij}=\\\\bq_i \\\\bk_j^\\\\top+\\\\bq_i\\\\left(\\\\br_{ij}\\\\bW^{K,R}\\\\right)^\\\\top +\\\\bk_j\\\\left(\\\\br_{ij}\\\\bW^{Q,R}\\\\right)^\\\\top,\\n\\\\end{equation}\\nwhere $\\\\bW^{K,R},\\\\bW^{Q,R}\\\\in\\\\mathbb{R}^{D_m\\\\times D_k}$ are learnable parameters and $\\\\br_{ij}$ is the learnable relative positional embedding as in Eq. \\\\eqref{rij}. The first term is interpreted as a content-to-content attention, and the latter two terms are interpreted as (relative) content-to-position and position-to-content attention, respectively.\",\n \"id\": \"9cdcf5bd-ef00-4ae9-a8b3-6d8cd2240795\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position Representations\"\n ],\n [\n \"subsubsection\",\n \"Relative Position Representations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Relative Position Representations\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1454,\n 8457\n ],\n \"content\": \"Some research studies have explored using hybrid positional representations that contains both absolute and relative positional information. Transformer with Untied Position Encoding (TUPE)~ re-designs the computation of attention scores as a combination of a content-to-content term, an absolute position-to-position term and a bias term representing relative positional relationships\\n\\\\begin{equation}\\n \\\\bA_{ij}=\\\\bq_i \\\\bk_j^\\\\top+\\\\left(\\\\bp_i\\\\bW^{Q,P}\\\\right)\\\\left(\\\\bp_{j}\\\\bW^{K,P}\\\\right)^\\\\top +b_{j-i},\\n\\\\end{equation}\\nwhere $\\\\bW^{K,P},\\\\bW^{Q,P}\\\\in\\\\mathbb{R}^{D_m\\\\times D_k}$ are learnable parameters, $\\\\bp_i, \\\\bp_j$ are the position embeddings for positions $i,j$, and $b_{j-i}$ is a learnable scalar relative position embedding.\\nOne can also design a single set of positional representations that express both absolute and relative information. Roformer~ uses Rotary Position Embedding (RoPE) to represent the position of a token by multiplying the affine-transformed embedding of the $t$-th input $x_t$ by a rotatory matrix $\\\\bR_{\\\\Theta,t}$\\n\\\\begin{align}\\n \\\\bq_t=\\\\bx_t\\\\bW^Q\\\\bR_{\\\\Theta,t}\\\\quad & \\\\bk_t=\\\\bx_t\\\\bW^K\\\\bR_{\\\\Theta,t},\\\\\\\\\\n \\\\bR_{\\\\Theta,t}&=\\\\bigoplus_{j=1}^{D_k/2} \\\\bM(t,\\\\theta_j),\\n\\\\end{align}\\nwhere $\\\\bigoplus$ denotes \\\\textit{direct sum} of matrices. Each $\\\\bM(t,\\\\theta_j)$ is a 2-D clockwise rotatory matrix of angle $t\\\\cdot \\\\theta_j$\\n\\\\begin{equation}\\n \\\\bM(t,\\\\theta_j)=\\\\left[\\\\begin{matrix} \\\\cos(t\\\\cdot \\\\theta_j)&\\\\sin(t\\\\cdot \\\\theta_j)\\\\\\\\\\n -\\\\sin(t\\\\cdot \\\\theta_j)&\\\\cos(t\\\\cdot \\\\theta_j)\\\\end{matrix}\\n \\\\right].\\n\\\\end{equation}\\nThe key advantage of this formulation is that the induced representation is translation invariant, i.e., the attention score of $(\\\\bq_i,\\\\bk_j)$ is only related to their relative position offset\\n\\\\begin{equation}\\n \\\\bq_i\\\\bk_j^\\\\top=\\\\left(\\\\bx_i\\\\bW^Q\\\\right)\\\\bR_{\\\\Theta,\\\\textcolor[rgb]{0,0,1}{j-i}}\\\\left(\\\\bx_j\\\\bW^K\\\\right)^\\\\top.\\n\\\\end{equation}\\nIn practice, the embedding matrix multiplication can be implemented by two element-wise multiplication for lower memory footprint. The RoPE uses the form of absolute embedding but can capture relative positional relations. This approach is compatible with linearized attention in Sec.~\\\\ref{sec:kernel}.\",\n \"id\": \"268772ad-4cd8-43e9-9608-b7ac7de2e35c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position Representations\"\n ],\n [\n \"subsubsection\",\n \"Other Representations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Other Representations\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1480,\n 1493,\n 1509,\n 1471\n ],\n \"content\": \"Instead of explicitly introducing additional positional encodings, propose to encode positional information in word embeddings, by generalizing embedding to continuous (complex-valued) functions over positions.\\nR-Transformer~ model locality of sequential data with a local RNN. Specifically, inputs to each block of R-Transformer are first fed to a local RNN and then to multi-Head self-attention module. The RNN structure introduces ordering information and captures local dependencies as a complement to self-attention.\\nConditional positional encoding (CPE)~ generate conditional position encodings at each layer for ViT with a 2-D convolution with zero-paddings. The intuition behind this approach is that convolution networks can implicitly encode absolute positional information with zero-paddings~.\",\n \"id\": \"866b7a57-7c1b-41c7-8a8f-c03a0681cbeb\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position Representations\"\n ],\n [\n \"subsubsection\",\n \"Position Representations without Explicit Encoding\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Position Representations without Explicit Encoding\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 1510\n ],\n \"content\": \"It is worth noticing that masked self-attention is not permutation equivariant~. Thus a model that exploits only the decoder of Transformer has the potential of sensing positional information without incorporating explicit positional representation. This is confirmed by some empirical results on language modeling tasks~, where the authors find that removing position encodings even improves performance.\",\n \"id\": \"90d9b719-8d67-48fa-a44c-ce04b5d77df6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"3f67f9f0-d57d-471a-bcb5-cbf64e5bcabc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position Representations\"\n ],\n [\n \"subsubsection\",\n \"Position Representation on Transformer Decoders\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Position Representation on Transformer Decoders\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Layer Normalization (LN), along with residual connection, is considered as a mechanism to stabilizing training of deep networks (e.g., alleviating ill-posed gradients and model degeneration).\\nThere are some studies that are dedicated to analyzing and improving LN module.\",\n \"id\": \"ba9ac74a-b1ed-471e-a9a5-9097079a09db\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"7a323139-7d7b-422d-a6c6-3617457ad2f6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Layer Normalization\"\n ]\n ],\n \"subsections\": [\n \"9c98bd5d-1c97-4d09-acae-d8ba70e8c480\",\n \"0329ae93-9d4a-453b-86ad-768aed04ebd3\",\n \"29a8b774-a21f-46a2-8095-eac026b86ef8\"\n ],\n \"title\": \"Layer Normalization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1496,\n 1511,\n 1448,\n 793,\n 1464,\n 38\n ],\n \"content\": \"\\\\ifx \\\\smv \\\\undefined\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[post-LN]{\\n\\\\includegraphics[height=2.2in]{assets/layernorm/post_ln.pdf}\\n}\\\\hspace{.6in}\\n\\\\subfigure[pre-LN]{\\n\\\\includegraphics[height=2.2in]{assets/layernorm/pre_ln.pdf}\\n}\\n\\\\caption{Comparison of Transformer Encoder with pre-LN and post-LN.}\\\\label{fig:pre_post_ln}\\n\\\\end{center}\\n\\\\end{figure}\\n\\\\fi\\nIn vanilla Transformer, the LN layer lies between the residual blocks, called post-LN~. Later Transformer implementations~ place the LN layer inside the residual connection before the attention or FFN, with an additional LN after the final layer to control the magnitude of final outputs, which is referred to as pre-LN\\\\footnote{To the best of our knowledge, this approach is adopted since \\\\texttt{v1.1.7} in the \\\\texttt{Tensor2Tensor} implementation~.}. The pre-LN has been adopted by numerous following research studies and implementations, e.g., .\\nThe difference between pre-LN and post-LN is shown in Fig. \\\\ref{fig:pre_post_ln}.\\n theoretically investigate the gradients of Transformers and find that the gradients near the output layer are large at initialization in post-LN Transformers, which could be the reason why post-LN Transformers without learning rate warm-up~\\\\footnote{Learning rate warm-up refers to starting optimization with an extremely small learning rate and then gradually increasing it to a pre-defined maximum value in a certain number of iterations.} leads to unstable training, whereas pre-LN Transformers do not suffer from the same problem. They thus deduce and empirically verify that warm-up stage can be safely removed for pre-LN Transformers.\\nAlthough Post-LN often results in unstable training and divergence, it usually outperforms pre-LN variants after convergence~. Similar to , conduct theoretical and empirical analysis and find that post-LN encoders do not suffer from gradient imbalance. They thus conjecture that the gradient issue is not the direct cause of unstable post-LN Transformer training and further identify the \\\\textit{amplification effect} in post-LN Transformers — at initialization, the heavier dependency on residual branch leads to a larger output shift in post-LN Transformers, thus resulting in unstable training. In light of this finding, they introduce additional parameters to post-LN Transformers to control residual dependencies of Post-LN. These parameters are initialized according to activation variations of sample data so that the output shift of post-LN Transformers is not amplified. This approach ensures and boosts convergence of post-LN Transformers and reaches better performance than pre-LN Transformers.\",\n \"id\": \"9c98bd5d-1c97-4d09-acae-d8ba70e8c480\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"ba9ac74a-b1ed-471e-a9a5-9097079a09db\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Layer Normalization\"\n ],\n [\n \"subsubsection\",\n \"Placement of Layer Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Placement of Layer Normalization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1466,\n 1467,\n 1491,\n 71\n ],\n \"content\": \" empirically observe that the learnable parameters in the LN module do not work in most experiments, and even increase the risk of overfitting. They further conclude from controlled experiments that the forward normalization is not the reason why LN works for Transformer. From analysis and experiments, it is concluded that the derivatives of the mean and variance re-center and re-scale the gradients and play a significant role in LN. They thus propose \\\\textit{AdaNorm}, a normalization technique without learnable parameters\\n\\\\begin{align}\\n \\\\bz &= C(1-k\\\\by)\\\\odot\\\\by,\\\\\\\\\\n \\\\by &=\\\\frac{\\\\bx-\\\\mu}{\\\\sigma},\\n\\\\end{align}\\nwhere $C,k$ are hyperparameters and $\\\\odot$ denotes element-wise multiplication. $\\\\mu$ and $\\\\sigma$ are the mean and standard deviation of input $\\\\bx$, respectively.\\n propose to replace the LN module with \\\\textit{scaled $\\\\ell_2$ normalization}. Given any input $\\\\bx$ of $d$-dimension, their approach project it onto a $d-1$-sphere of learned radius $g$\\n\\\\begin{equation}\\n \\\\bz=g\\\\frac{\\\\bx}{\\\\|\\\\bx\\\\|},\\n\\\\end{equation}\\nwhere $g$ is a learnable scalar. It is more parameter efficient compared to normal LN and is shown to be effective in machine translation datasets, especially in low-resource settings.\\n discuss why Batch Normalization (BN)~ performs poorly in Transformer for text data and conclude that BN's significant performance degradation stems from the instabilities associated with its batch statistics. They thus propose PowerNorm (PN) that has three modifications over BN: (1) it relaxes the zero-mean normalization; (2) it uses the quadratic mean\\nof the signal, instead of the variance; (3) it uses running statistics for the\\nquadratic mean, instead of using per-batch statistics. Specifically, for the $t$-th iteration, the PN computes the outputs as\\n\\\\begin{align}\\n \\\\bz^{(t)} &= \\\\gamma\\\\odot \\\\by^{(t)}+\\\\beta,\\\\\\\\\\n \\\\by^{(t)} &=\\\\frac{\\\\bx^{(t)}}{\\\\psi^{(t-1)}},\\\\\\\\\\n (\\\\psi^{(t)})^2&=\\\\alpha(\\\\psi^{(t-1)})^2+(1-\\\\alpha)\\\\left(\\\\frac{1}{|B|}\\\\sum_{i=1}^{|B|}(\\\\bx_i^{(t)})^2\\\\right),\\n\\\\end{align}\\nwhere $0<\\\\alpha<1$ is the moving average coefficient and $\\\\gamma, \\\\beta$ are the learnable parameters as in BN formulation.\",\n \"id\": \"0329ae93-9d4a-453b-86ad-768aed04ebd3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"ba9ac74a-b1ed-471e-a9a5-9097079a09db\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Layer Normalization\"\n ],\n [\n \"subsubsection\",\n \"Substitutes of Layer Normalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Substitutes of Layer Normalization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1495\n ],\n \"content\": \"Besides LN, there is another mechanism to construct deeper neural network. ReZero~ replace LN module with a learnable residual connection. For each module $F(\\\\cdot)$, ReZero re-scales $F(\\\\cdot)$ in the residual formulation:\\n\\\\begin{equation}\\n \\\\bH'=\\\\bH+\\\\alpha\\\\cdot F(\\\\bH),\\n\\\\end{equation}\\nwhere $\\\\alpha$ is a learnable parameter with zero-initialization.\\nReplacing LN in Transformer with ReZero mechanism is verified to induce better dynamic isometry for input signals and leads to faster convergence.\",\n \"id\": \"29a8b774-a21f-46a2-8095-eac026b86ef8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"ba9ac74a-b1ed-471e-a9a5-9097079a09db\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Layer Normalization\"\n ],\n [\n \"subsubsection\",\n \"Normalization-free Transformer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Normalization-free Transformer\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:ffn}\\nDespite its simplicity, the position-wise feed-forward network (FFN) layers are important for a Transformer to achieve good performance. observe that simply stacking self-attention modules causes a \\\\textit{rank collapse} problem, leading to token-uniformity inductive bias, and that the feed-forward layer is one of the important building blocks that mitigate this issue.\\nVarious works have explored modifications on the FFN module.\",\n \"id\": \"e915275c-0207-4658-9d41-438a8b5b3e51\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"7a323139-7d7b-422d-a6c6-3617457ad2f6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position-wise FFN\"\n ]\n ],\n \"subsections\": [\n \"d96aa288-77f3-4bf6-931e-abd30aeafccb\",\n \"e9f737ba-3b14-4121-b536-765cb8c260d1\",\n \"d1496356-82a7-4f73-8a03-62f59057e605\"\n ],\n \"title\": \"Position-wise FFN\"\n },\n {\n \"cite_extract_rate\": 0.875,\n \"cites\": [\n 1481,\n 1512,\n 7055,\n 1455,\n 7,\n 7365,\n 38\n ],\n \"content\": \"The vanilla Transformer~ adopts the Rectified Linear Units (ReLU) activation for non-linearity in between the two FFN layers. Over time, several studies have explored different activation other than ReLU.\\n try to replace ReLU in Transformer with Swish function $f(x)=x\\\\mathrm{sigmoid}(\\\\beta x)$ and observe that it consistently improve performance on WMT 2014 English$\\\\rightarrow$German dataset.\\nGPT~ replace ReLU with Gaussian Error Linear Unit (GELU)~ on language pre-training. It becomes the default practice for many pre-trained language models ~.\\n explore using Gated Linear Units (GLU)~ and its variants as a drop-in replacement for ReLU in FFN. Their pre-training experiments show that the GLU variants consistently improve vanilla Transformer with ReLU activation. Note that GLU introduces extra parameters and the experiments are conducted with the intermediate dimension of FFN reduced to match the parameter count with baseline.\",\n \"id\": \"d96aa288-77f3-4bf6-931e-abd30aeafccb\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"e915275c-0207-4658-9d41-438a8b5b3e51\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position-wise FFN\"\n ],\n [\n \"subsubsection\",\n \"Activation Function in FFN\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Activation Function in FFN\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 680,\n 1490,\n 1451,\n 8454,\n 707\n ],\n \"content\": \"Several works have focused on expanding FFNs in order for a larger model capacity. The basic idea is to replace FFNs with similar structures with much more parameters.\\n replace some of the FFNs with the product-key memory layers. A product-key memory is composed of three components: a query network, a key selection module containing two\\nsets of sub-keys, and a value lookup table. The model first projects an input to a latent space using the query network, and then compares the generated query to keys that are Cartesian product of the two sets of sub-keys from key selection module to get $k$ nearest neighbors, and finally finds the corresponding values in a value lookup table using the $k$ nearest keys and aggregates them to produce the final output. This process resembles the attention mechanism, in that the generated query attends to a large number of global key-value pairs. They thus propose a multi-head mechanism for the key-product memory to further enlarge the capacity of this module. The experiments on large-scale language modeling suggest that this mechanism significantly improves performance with negligible computational overhead.\\nSeveral studies exploits the idea of Mixture-of-Experts (MoE) to increase the capacity of FFNs. Gshard uses sparsely-gated MoE layers to replace FFNs in Transformer. Each MoE layer consists of several FFNs (each called an expert) that are the same structure as position-wise FFNs in vanilla Transformer. The output of the layer is a weighted sum of the outputs of the FFNs, using gate values computed by a routing function $g(\\\\cdot)$. They design a learnable routing function that assigns tokens to experts, with auxiliary loss to satisfy balanced loads between experts and efficiency at the scale of length such that the experts can be distributed across multiple devices. For each forward pass of the MoE layer, only the experts with top-$k$ gate values are activated.\\nInstead of using $k$ experts for each forward pass, Switch Transformer~ proposes to route using only a single expert with the largest gate value, leading to a much smaller computational footprint. The authors also design an auxiliary loss to encourage load balance between experts. It is reported to speed up pre-training by a large margin compared to the non-MoE counterpart while having a similar number of FLOPS.\\n propose to replace top-$k$ routing with expert prototyping strategy. Specifically, the proposed strategy splits experts into $k$ different groups and applies top-1 routing within each group. The outputs of prototype groups are combined linearly to form the final output of the MoE layer. This strategy is proved to improve the model quality while maintaining constant computational costs.\\nAs opposed to using a learnable routing function for expert assignment, design \\\\textit{hash layers} where tokens are hashed into a fixed number of buckets, each bucket corresponding to an expert. This approach requires no routing parameters or any auxiliary loss function, while showing competitive results with existing methods such as Switch Transformer~.\",\n \"id\": \"e9f737ba-3b14-4121-b536-765cb8c260d1\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"e915275c-0207-4658-9d41-438a8b5b3e51\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position-wise FFN\"\n ],\n [\n \"subsubsection\",\n \"Adapting FFN for Larger Capacity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adapting FFN for Larger Capacity\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1463,\n 7053\n ],\n \"content\": \"Notably, one might argue that under some circumstances, FFN layers can be dropped completely, resulting in a simplified network.\\n demonstrate that replacing the ReLU activation with Softmax and dropping the bias term in FFN effectively turns FFN into an attention module where position-wise inputs attend to a global key-value memory of $D_{\\\\mathrm{ffn}}$ slots. They thus propose to drop the FFN module and add to the attention module a set of global key-value pairs, which are learnable parameters concatenated with key and values generated by inputs. This approach simplifies the structure of the network with no loss of performance.\\n empirically show that FFNs in the decoder of Transformer, despite its large number of parameters, is not efficient and can be removed safely with only slight or no loss of performance. This approach significantly boosts the training and inference speed.\",\n \"id\": \"d1496356-82a7-4f73-8a03-62f59057e605\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"e915275c-0207-4658-9d41-438a8b5b3e51\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Other Module-level Modifications\"\n ],\n [\n \"subsection\",\n \"Position-wise FFN\"\n ],\n [\n \"subsubsection\",\n \"Dropping FFN Layers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dropping FFN Layers\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:beyond}\\nIn this section, we introduce the X-formers that modify the vanilla Transformer beyond modules.\",\n \"id\": \"04795308-eff0-4fbb-962f-a8077d36f894\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ]\n ],\n \"subsections\": [\n \"f397ed71-a760-4154-8eaa-b1b20d5004b2\",\n \"736cefee-90d8-4e53-a49e-f031442dc1ba\",\n \"690a01cd-27c7-4e1a-baef-83bbdeff8220\",\n \"9b2cdbdc-bac6-48ec-96d0-b5576360be13\",\n \"be720c58-062b-41bd-b6e9-732058f0f671\"\n ],\n \"title\": \"Architecture-level Variants\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1468,\n 1488\n ],\n \"content\": \"Apart from the efforts made at the module level to alleviate computation overheads, there are several attempts to adapt Transformer to be lightweight by modifications at a higher level.\\nSimilar to low-rank self-attention~ that decomposes attention into a locality-constrained attention and a low-rank global attention, Lite Transformer~ proposes to replace each attention module in Transformer with a two-branch structure, where one branch uses attention to capture long-range contexts while the other branch uses depth-wise convolution and linear layers to capture local dependencies. The architecture is lightweight both in terms of model size and computation, and is thus more suitable for mobile devices.\\nFunnel Transformer~ utilizes a funnel-like encoder architecture where the length of the hidden sequence is gradually reduced using pooling along the sequence dimension, and then recovered using up-sampling. The architecture effectively reduces the FLOPs and memory compared to the vanilla Transformer encoder. Naturally, one can use this architecture to build a deeper or wider model using the same computation resources.\\nDeLighT~ replaces the standard Transformer block with \\\\texttt{DeLighT} block, which consists of three sub-modules: (1) a ``expand-and-reduce'' \\\\texttt{DeLighT} transformation module to learn wider representations with low computation requirements; (2) a single-head self-attention to learn pair-wise interaction; (3) a lightweight ``reduce-and-expand'' FFN (as opposed to vanilla Transformer that first expands the dimension of hidden representations and then reduces them back to $D_m$). They also propose a block-wise scaling strategy that allows for shallower and narrower blocks near the input and wider and deeper blocks near the output. The induced network is much deeper than the vanilla Transformer but with fewer parameters and operations.\",\n \"id\": \"f397ed71-a760-4154-8eaa-b1b20d5004b2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"04795308-eff0-4fbb-962f-a8077d36f894\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Adapting Transformer to Be Lightweight\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adapting Transformer to Be Lightweight\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 1457,\n 1486,\n 1492\n ],\n \"content\": \"In vanilla Transformer, each block takes outputs from the previous block as inputs and outputs a sequence of hidden representations. One might be interested in creating more paths along which input signals can run through the networks. In Sec.~\\\\ref{sec:prior_prev_module}, we introduced Realformer~ and Predictive Attention Transformer~ that reuses attention distributions from previous block to guide attention of current block. This can be seen as creating a forward path between adjacent Transformer blocks.\\nIn a deep Transformer encoder-decoder model, the cross-attention modules in the decoder only utilize the final outputs of the encoder, therefore the error signal will have to traverse along the depth of the encoder. This makes Transformer more susceptible to optimization issues (e.g., vanishing gradients). Transparent Attention~ uses a weighted sum of encoder representations at all encoder layers (including the embedding layer) in each cross-attention module. For the $j$-th decoder block, the cross-attention module is modified to attend to\\n\\\\begin{equation}\\\\label{eq:transparent}\\n \\\\tilde{\\\\bH}^{(j)}=\\\\sum_{i=0}^{N}\\\\frac{\\\\exp(w_{ij})}{\\\\sum_{k=0}^N \\\\exp(w_{kj})}\\\\bH^{(i)},\\n\\\\end{equation}\\nwhere each $w_{ij}$ is a trainable parameter. This effectively shortens the path from each layer in the encoder to the error signal and thus eases the optimization of deeper Transformer models.\\nAnother issue associated with vanilla Transformer is that each position can only attend to history representations from lower layers. Feedback Transformer~ proposes to add a feedback mechanism to Transformer decoder, where each position attends to a weighted sum of history representations from all layers\\n\\\\begin{equation}\\n \\\\tilde{\\\\bh}_i=\\\\sum_{l=0}^{N}\\\\frac{\\\\exp(w_{l})}{\\\\sum_{k=0}^N \\\\exp(w_{k})}\\\\bh_i^{(l)}.\\n\\\\end{equation}\",\n \"id\": \"736cefee-90d8-4e53-a49e-f031442dc1ba\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"04795308-eff0-4fbb-962f-a8077d36f894\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Strengthening Cross-Block Connectivity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Strengthening Cross-Block Connectivity\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 683,\n 7273,\n 1485,\n 1449,\n 1478\n ],\n \"content\": \"Vanilla Transformer, like most neural models, utilizes a fixed (learned) computation procedure to process each input. An intriguing and promising modification is to make computation time conditioned on the inputs, i.e., to introduce Adaptive Computation Time (ACT)~ into Transformer models. Such modifications potentially give rise to the following advantages:\\n\\\\begin{itemize}\\n \\\\item Feature refinement for hard examples. For data that are hard to process, a shallow representation might not be adequate to fulfill the task at hand. It would be more ideal to apply more computations to acquire a deeper and more refined representation.\\n \\\\item Efficiency for easy examples. When processing easy examples, a shallow representation might be enough for the task. In this case, it would be beneficial if the network can learn to extract features using reduced computation time.\\n\\\\end{itemize}\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[dynamic halting]{\\\\label{fig:ut}\\\\hspace{2em}\\n\\\\includegraphics[height=1.2in]{assets/ACT/ut.pdf}\\\\hspace{2em}\\n}\\n\\\\subfigure[conditional skipping]{\\\\label{fig:skip}\\\\hspace{2em}\\n\\\\includegraphics[height=1.2in]{assets/ACT/skip.pdf}\\\\hspace{2em}\\n}\\n\\\\subfigure[early exit]{\\\\label{fig:exit}\\\\hspace{2em}\\n\\\\includegraphics[height=1.2in]{assets/ACT/exit.pdf}\\n}\\n\\\\caption{Three typical ACT paradigms.}\\\\label{fig:act}\\n\\\\end{center}\\n\\\\end{figure}\\nUniversal Transformer (UT)~ incorporates a recurrence-over-depth mechanism that iteratively refines representations for all symbols using a module that is shared over depth, as illustrated in Fig. \\\\ref{fig:ut}. It also adds a per-position dynamic halting mechanism that calculates a halting probability for each symbol at every time step. If a symbol's halting probability is greater than a predefined threshold, then the symbol's representation will remain unchanged for subsequent timesteps. The recurrence is stopped when all symbols halt or when a predefined maximum step is reached.\\nConditional Computation Transformer (CCT)~ adds a gating module at each self-attention and feed-forward layer to decide whether to skip the current layer, as illustrated in Fig. \\\\ref{fig:skip}. The authors also introduce an auxiliary loss that encourages the model to adjust the gating modules to match the practical computation cost to the available computation budget.\\nSimilar to the dynamic halting mechanism used in UT, there is a line of work dedicated to adapting the number of layers to each input in order to achieve a good speed-accuracy trade-off, which is called \\\\textit{early exit} mechanism, as illustrated in Fig. \\\\ref{fig:exit}. A commonly used technique is to add an internal classifier at each layer and jointly train all classifiers. The core of these methods is the criteria used to decide whether to exit at each layer. DeeBERT~ uses the entropy of the output probability distribution of the current layer to determine whether to exit. PABEE~ counts the number of times that the predictions remain unchanged to decide whether to exit. design a window-based uncertainty criterion to achieve token-level partial exiting for sequence labeling tasks. introduces a voting-based exiting strategy that considers at each layer predictions of all the past internal classifiers to infer the correct label and to decide whether to exit.\",\n \"id\": \"690a01cd-27c7-4e1a-baef-83bbdeff8220\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"04795308-eff0-4fbb-962f-a8077d36f894\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Adaptive Computation Time\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adaptive Computation Time\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The quadratic complexity of self-attention on sequences length can significantly limit the performance\\nof some downstream tasks. For example, language modeling usually needs long-range context. Apart from the techniques introduced in Sec.~\\\\ref{sec:attention}, another effective way of dealing with long sequences is to use \\\\textit{divide-and-conquer} strategy, i.e., to decompose an input sequence into finer segments that can be efficiently processed by Transformer or Transformer modules. We identify two representative class of methods, \\\\textit{recurrent} and \\\\textit{hierarchical} Transformers, as illustrated in Fig. \\\\ref{fig:divide_and_conquer}. These techniques can be understood as a wrapper for the Transformer model in which Transformer acts as an elementary component that is reused to process different input segments.\\n\\\\begin{figure}[htbp]\\n\\\\begin{center}\\n\\\\subfigure[Recurrent Transformer\\\\label{fig:recurrence}]{\\n\\\\includegraphics[width=0.4\\\\linewidth]{assets/divide_and_conquer/recurrence.pdf}\\n}\\\\quad\\n\\\\subfigure[Hierarchical Transformer\\\\label{fig:hierarchy}]{\\n\\\\includegraphics[width=0.4\\\\linewidth]{assets/divide_and_conquer/hierarchy.pdf}\\n}\\n\\\\caption{Illustrations of recurrent and hierarchical Transformers.}\\\\label{fig:divide_and_conquer}\\n\\\\end{center}\\n\\\\end{figure}\",\n \"id\": \"9b2cdbdc-bac6-48ec-96d0-b5576360be13\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"04795308-eff0-4fbb-962f-a8077d36f894\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Transformers with Divide-and-Conquer Strategies\"\n ]\n ],\n \"subsections\": [\n \"88924fec-4dc9-4d83-b667-d6862e359751\",\n \"6244ba7c-c68b-4615-be13-f855c4dbca70\"\n ],\n \"title\": \"Transformers with Divide-and-Conquer Strategies\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1500,\n 7370,\n 1487\n ],\n \"content\": \"In recurrent Transformers, a cache memory is maintained to incorporate the history information. While processing a segment of text, the network reads from the cache as an additional input. After the processing is done, the network writes to the memory by simply copying hidden states or using more complex mechanisms. The abstract process is illustrated in Fig. \\\\ref{fig:recurrence}.\\nTransformer-XL~ address the limitation of a fixed length context by caching representations from the previous segment and reuse it as an extended context when the model processes the current segment. For the $l$-th layer and the $(\\\\tau+1)$-th segment, the input representation $\\\\bH_{\\\\tau+1}^{(l-1)}$ is concatenated with the representation $\\\\bH_\\\\tau^{(l-1)}$ from previous segment to produce the keys and values\\n\\\\begin{align}\\n \\\\tilde{\\\\bH}_{\\\\tau+1}^{(l)}&=[\\\\mathrm{SG}(\\\\bH_\\\\tau^{(l-1)})\\\\circ \\\\bH_{\\\\tau+1}^{(l-1)}]\\\\label{eq:xformerxl},\\\\\\\\\\n \\\\bK_{\\\\tau+1}^{(l)}, \\\\bV_{\\\\tau+1}^{(l)}&=\\\\tilde{\\\\bH}_{\\\\tau+1}^{(l)}\\\\bW^K, \\\\tilde{\\\\bH}_{\\\\tau+1}^{(l)}\\\\bW^V,\\n\\\\end{align}\\nwhere $\\\\bH_\\\\tau^{(0)}$ is defined as the word embedding sequence, $\\\\mathrm{SG}(\\\\cdot)$ denotes stop-gradient operation and $[\\\\bX\\\\circ\\\\bY]$ denotes concatenating the two vector sequences along the time dimension. This approach extends the maximum context length by $L\\\\times N_{\\\\text{mem}}$ where $L$ is the number of layers and $N_{\\\\text{mem}}$ is the length of cached memory sequence.\\nCompressive Transformer~ extends this idea further by extending the cache with two levels of memory. In Transformer-XL, the activations from the previous segment are cached as a memory that is used to augment the current segment, and activations from older segments are discarded. Compressive Transformer, on the other hand, applies a compression operation (e.g., Convolution, Pooling, etc.) on older activations and stores them in the compressed memory. In order to avoid the expensive backpropagating-through-time (BPTT) from training compression sub-network with gradients from the loss, they propose to use local loss functions where original memories are constructed from the compressed memories. This approach further extends the theoretical maximum history context length from $L\\\\times N_{\\\\text{mem}}$ of Transformer-XL to $L\\\\times (N_{\\\\text{mem}}+c\\\\times N_{\\\\text{cm}})$, where $c$ is the compression rate and $N_{\\\\text{cm}}$ is the length of compressed memory.\\n Memformer~ extends the recurrence mechanism from decoder-only architecture to an encoder-decoder architecture. They introduce to the encoder a memory cross attention similar to the cross attention in vanilla Transformer to allow the Transformer encoder to attend to the memory. They also introduce a memory slot attention on top of the encoder output to explicitly write the memory for the next segment. To avoid BPTT over a long range of timesteps, they propose Memory\\nReplay Back-Propagation (MRBP) algorithm, which replays\\nthe memory at each timestep to accomplish gradient back-propagation over long unrolls.\\n propose a simple fine-tuning mechanism to add recurrence to a pre-trained language model (e.g., GPT-2~).\\n\\\\ifx \\\\smv \\\\undefined\\nThey first compress the representations produced by the $\\\\tau$-th segment into one single vector representation, using a weighted average of pooled representations from each layer $l\\\\in\\\\{1,\\\\cdots,L\\\\}$\\n\\\\begin{equation}\\n \\\\bz_\\\\tau=\\\\sum_{l=1}^Lw_l\\\\sum_{j=1}^{T_\\\\tau}\\\\bh_j^{(l)},\\n\\\\end{equation}\\nwhere $T_\\\\tau$ denotes the sequence length of the $\\\\tau$-th segment, $w_l=\\\\mathrm{softmax}(\\\\mathbf\\\\alpha)_l$ is the weight softmax-normalized from learnable parameters $\\\\mathbf\\\\alpha=[\\\\alpha_1,\\\\cdots,\\\\alpha_L]$. This compressed representation is then fed to a feed-forward network to produce the memory state $\\\\bh_{\\\\text{prev},\\\\tau}$ for the $\\\\tau$-th segment, which is then prepended to the key-value inputs of a specific attention layer. This approach effectively extends the context length of a pre-trained language model, without significant change of the architecture of the original model.\\n\\\\fi\\nERNIE-Doc~ proposes an enhanced recurrence mechanism based on the recurrence mechanism used in Transformer-XL, by replacing the memory with the history representations from the $l$-th layer.\\n\\\\ifx \\\\smv \\\\undefined\\n\\\\begin{align}\\n \\\\tilde{\\\\bH}_{\\\\tau+1}^{(l)}&=[\\\\mathrm{SG}(\\\\textcolor[rgb]{0,0,1}{\\\\bH_\\\\tau^{(l)}})\\\\circ \\\\bH_{\\\\tau+1}^{(l-1)}],\\n\\\\end{align}\\nas opposed to using representations from the $(l-1)$-th layer in Eq. \\\\eqref{eq:xformerxl}. This modification essentially leads to a larger effective context length.\\n\\\\fi\",\n \"id\": \"88924fec-4dc9-4d83-b667-d6862e359751\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"9b2cdbdc-bac6-48ec-96d0-b5576360be13\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Transformers with Divide-and-Conquer Strategies\"\n ],\n [\n \"subsubsection\",\n \"Recurrent Transformers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Recurrent Transformers\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Hierarchical Transformer decomposes inputs hierarchically into elements of finer granularity. Low-level features are first fed to a Transformer encoder, producing output representations that are then aggregated (using pooling or other operations) to form a high-level feature, which is then processed by a high-level Transformer. This class of methods can be understood as a process of hierarchical abstraction. The overview of this approach is depicted in Fig. \\\\ref{fig:hierarchy}. The advantages of this approach are twofold: (1) Hierarchical modeling allows the model to handle long inputs with limited resources; (2) It has the potential to generate richer representations that are beneficial to tasks.\",\n \"id\": \"6244ba7c-c68b-4615-be13-f855c4dbca70\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"9b2cdbdc-bac6-48ec-96d0-b5576360be13\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Transformers with Divide-and-Conquer Strategies\"\n ],\n [\n \"subsubsection\",\n \"Hierarchical Transformers\"\n ]\n ],\n \"subsections\": [\n \"10bc35f7-356c-4cac-a874-ee859573c18e\",\n \"153b332f-732b-4172-8553-f1c9626b599f\"\n ],\n \"title\": \"Hierarchical Transformers\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 1482,\n 1477\n ],\n \"content\": \"For tasks with inherently long input length, one can use hierarchical Transformers for effective modeling of long-range dependencies. For document-level machine translation tasks, introduce dependencies on the previous sentences from both the source and target sides when translating a sentence. They use an attention mechanism as the aggregation operation to summarize low-level information. For document summarization, HIBERT~ encodes a document of text by first learn sentence representations for all sentences and then use these sentence representations to encode document-level representations that are then used to generate the summary. The model uses the last hidden representation (corresponding to the \\\\texttt{EOS} token) as the representation for each sentence. propose a similar hierarchical Transformer for multi-document summarization where the extracted low-level representations are aggregated using an attention layer with a global trainable query node and low-level representations as the source of key-value pairs. Hi-Transformer~ first utilizes a sentence Transformer and a document Transformer to hierarchically learn document context-aware sentence representations. The document context-aware sentence representations are then fed to another sentence Transformer to further improve the sentence context modeling.\",\n \"id\": \"10bc35f7-356c-4cac-a874-ee859573c18e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"6244ba7c-c68b-4615-be13-f855c4dbca70\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Transformers with Divide-and-Conquer Strategies\"\n ],\n [\n \"subsubsection\",\n \"Hierarchical Transformers\"\n ],\n [\n \"paragraph\",\n \"6.5.2.1 Hierarchical for long sequence inputs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"6.5.2.1 Hierarchical for long sequence inputs\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 732,\n 7368,\n 7367\n ],\n \"content\": \"One might also be interested in using hierarchical models to acquire richer representations that are beneficial to the tasks at hand. For example, TENER~ uses a low-level Transformer encoder to encode character features, which is then concatenated with word embeddings as the inputs to the high-level Transformer encoder. This incorporates more features and alleviates the problems of data sparsity and out-of-vocabulary (OOV). Recently emerging Vision Transformer~ divides an input image into several patches that serve as the basic input elements of Transformer, which potentially loses intrinsic pixel-level information within patches. To address this issue, Transformer in Transformer (TNT)~ uses at each layer an inner Transformer block that transforms pixel representations and an outer Transformer block that takes fused vectors of patch representations and pixel representations as input.\",\n \"id\": \"153b332f-732b-4172-8553-f1c9626b599f\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"6244ba7c-c68b-4615-be13-f855c4dbca70\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Transformers with Divide-and-Conquer Strategies\"\n ],\n [\n \"subsubsection\",\n \"Hierarchical Transformers\"\n ],\n [\n \"paragraph\",\n \"6.5.2.2 Hierarchical for richer representations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"6.5.2.2 Hierarchical for richer representations\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 8456,\n 1469,\n 544,\n 1475,\n 1483,\n 1461\n ],\n \"content\": \"Despite the success of Transformer architecture, one might question whether the current Transformer architecture is optimal. Interestingly, several studies have explored alternative architectures for Transformer.\\n interpret Transformer as a numerical Ordinary Differential Equation (ODE) solver for a convection-diffusion equation in a multi-particle dynamic system and design Macaron Transformer, which replaces each Transformer block with a \\\\textit{FFN-attention-FFN} variant.\\nSandwich Transformer~ explores reorganizing attention modules and FFN modules such that attention modules are mainly located in lower layers and FFN modules in upper layers. The induced model improves perplexity on multiple language modeling benchmarks, without increasing parameters, memory or training time.\\nMask Attention Network (MAN)~ prepends a dynamic mask attention module to the self-attention module in each Transformer block. The mask is conditioned on token representations, the relative distance between tokens and head indices. The proposed dynamic mask attention is shown to effectively model locality in text data and the induced model consistently outperforms the baseline model in machine translation and abstractive summarization.\\nNotably, there's a line of work that uses Neural Architecture Search (NAS) to search for alternative Transformer architectures. The Evolved Transformer (ET)~ employs evolution-based architecture search with the standard Transformer architecture seeding the initial population. The searched model demonstrates consistent improvement\\nover Transformer on several language tasks. As another representative work, DARTSformer applies differentiable architecture search (DARTS)~, combined with a multi-split reversible network and a backpropagation-with-reconstruction algorithm for memory efficiency. The resulting model consistently outperforms standard Transformer and compares favorably to larger ET models, with a significantly reduced search cost.\",\n \"id\": \"be720c58-062b-41bd-b6e9-732058f0f671\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"04795308-eff0-4fbb-962f-a8077d36f894\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Architecture-level Variants\"\n ],\n [\n \"subsection\",\n \"Exploring Alternative Architecture\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Exploring Alternative Architecture\"\n },\n {\n \"cite_extract_rate\": 0.7000000000000001,\n \"cites\": [\n 826,\n 1445,\n 679,\n 707,\n 9,\n 1499,\n 7\n ],\n \"content\": \"\\\\label{sec:ptm}\\nAs a key difference from convolutional networks and recurrent networks that inherently incorporates the inductive bias of locality, Transformer does not make any assumption about how the data is structured. On the one hand, this effectively makes Transformer a very universal architecture that has the potential of capturing dependencies of different ranges. On the other hand, this makes Transformer prone to overfitting when the data is limited. One way to alleviate this issue is to introduce inductive bias into the model.\\nRecent studies suggest that Transformer models that are pre-trained on large corpora can learn universal language representations that are beneficial for downstream tasks~. The models are pre-trained using various self-supervised objectives, e.g., predicting a masked word given its context. After pre-training a model, one can simply fine-tune it on downstream datasets, instead of training a model from scratch. To illustrate typical ways of using Transformers in pre-training, we identify some of the pre-trained Transformers and categorize them as follows.\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Encoder only}. A line of work uses the Transformer encoder as its backbone architecture. BERT~ is a representative PTM that is typically used for natural language understanding tasks. It utilizes Masked Language Modeling (MLM) and Next Sentence Prediction (NSP) as the self-supervised training objective. RoBERTa~ further adapts the training of BERT and removes the NSP objective as it is found to hurt performance on downstream tasks.\\n \\\\item \\\\textit{Decoder only}. Several studies focus on pre-training Transformer decoders on language modeling. For example, the Generative Pre-trained Transformer (GPT) series (i.e., GPT~, GPT-2~, and GPT-3~) is dedicated to scaling pre-trained Transformer decoders and has recently illustrated that a large-scale PTM can achieve impressive few-shot performance with the task and examples fed to the model as constructed prompts~.\\n \\\\item \\\\textit{Encoder-Decoder}. There are also PTMs that adopt Transformer encoder-decoder as the overall architecture. BART~ extends the denoising objective of BERT to encoder-decoder architecture. The benefit of using an encoder-decoder architecture is that the inducing model is equipped with the ability to perform both natural language understanding and generation. T5~ adopts similar architecture and was one of the earliest studies that use task-specific text prefix in downstream tasks.\\n\\\\end{itemize}\\nSome of the Transformer architecture variants can also be applied to Transformer-based PTMs. For instance, BigBird~ introduced in Sec.~\\\\ref{sec:sparseattn} is a encoder-based PTM that uses compound position-based sparse attention to enable long sequence inputs. GPT-3~ uses alternating dense and locally banded sparse attention (which was also introduced in Sec.~\\\\ref{sec:sparseattn}) in self-attention modules. Switch Transformer~ is an encoder-based PTM that replaces FFN layers with mixture-of-experts layers and can increase parameter count while keeping the FLOPs per example constant.\",\n \"id\": \"b7522af4-45ec-452c-b804-11b6882e6225\",\n \"level\": \"section\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Pre-trained Transformers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pre-trained Transformers\"\n },\n {\n \"cite_extract_rate\": 0.7674418604651161,\n \"cites\": [\n 7361,\n 1500,\n 7368,\n 7373,\n 1458,\n 1446,\n 7056,\n 7360,\n 7040,\n 7370,\n 1470,\n 768,\n 732,\n 38,\n 1272,\n 1514,\n 7339,\n 1447,\n 8459,\n 8458,\n 1502,\n 1469,\n 9,\n 1515,\n 1501,\n 1461,\n 1453,\n 1456,\n 1462,\n 1503,\n 1516,\n 1513\n ],\n \"content\": \"\\\\label{sec:app}\\nTransformer was originally designed for machine translation but has been widely adopted in various fields besides NLP, including CV and audio processing, due to its flexible architecture.\\n(1) \\\\textit{Natural Language Processing}. Transformer and its variants have been extensively explored and applied in NLP tasks, e.g., machine translation~, language modeling~ and named entity recognition~. Massive effort has been dedicated to pre-training Transformer models on large-scale text corpora, which we believe is one of the major reasons of Transformer's wide application in NLP.\\n(2) \\\\textit{Computer Vision}. Transformer have also been adapted for various vision tasks, e.g., image classification~, object detection~, image generation~ and video processing~. and provide reviews on existing work of visual Transformers. We encourage readers to refer to these surveys for further understand the current research progress on Transformers in CV.\\n(3) \\\\textit{Audio Applications}. Transformer can also be extended for audio-related applications, e.g., speech recognition~, speech synthesis~, speech enhancement~ and music generation~.\\n(4) \\\\textit{Multimodal Applications}. Owing to its flexible architecture, Transformer has also been applied in various multimodal scenarios, e.g., visual question answering~, visual commonsense reasoning~, caption generation~, speech-to-text translation~ and text-to-image generation~.\",\n \"id\": \"b3d753a5-00c0-4eb4-b313-01e490dd9d3d\",\n \"level\": \"section\",\n \"origin_cites_number\": 43,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Applications of Transformer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Applications of Transformer\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:discussion}\\nIn this survey, we conduct a comprehensive overview of X-formers and propose a new taxonomy. Most of the existing works improve Transformer from different perspectives, such as efficiency, generalization, and applications. The improvements include incorporating structural prior, designing lightweight architecture, pre-training, and so on.\\nAlthough X-formers have proven their power for various tasks, challenges still exist. Besides the current concerns (e.g. efficiency and generalization), the further improvements of Transformer may lie in the following directions:\\n(1) \\\\textit{Theoretical Analysis}. The architecture of Transformer has been demonstrated to be capable of supporting large-scale training datasets with enough parameters. Many works show that Transformer has a larger capacity than CNNs and RNNs and hence has the ability to handle a huge amount of training data.\\nWhen Transformer is trained on sufficient data, it usually has better performances than CNNs or RNNs.\\nAn intuitive explanation is that Transformer has few prior assumptions on the data structure and therefore is more flexible than CNNs and RNNs. However, the theoretical reason is unclear and we need some theoretical analysis of Transformer ability.\\n(2) \\\\textit{Better Global Interaction Mechanism beyond Attention}. A main advantage of Transformer is the use of the attention mechanism to model the global dependencies among nodes within input data. However, many studies have shown that full attention is unnecessary for most nodes. It is, to some degree, inefficient to indistinguishably calculate attention for all nodes. Therefore, there is still plenty of room for improvements in efficiently modeling global interactions. On the one hand, the self-attention module can be regarded as a fully-connected neural network with dynamical connection weights, which aggregates non-local information with dynamic routing. Therefore, other dynamic routing mechanisms are alternative approaches worth exploring. On the other hand, the global interaction can also be modeled by other types of neural networks, such as memory-enhanced models.\\n(3) \\\\textit{Unified Framework for Multimodal Data}. In many application scenarios, integrating multimodal data is useful and necessary to boost the task performance. Moreover, the general AI also needs the ability to capture the semantic relations across different modalities.\\nSince Transformer achieves great success on text, image, video, and audio, we have a chance to build a unified framework and better capture the inherent connections among multimodal data. However, the design of the intra-modal and cross-modal attention still remains to be improved.\\nFinally, we wish this survey to be a hands-on reference for better understanding the current research progress on Transformers and help readers to further improve Transformers for various applications.\\n\\\\bibliographystyle{ACM-Reference-Format}\\n\\\\bibliography{references}\\n\\\\end{document}\",\n \"id\": \"ec81ede4-3b9c-44b8-96e2-94e0f2e4f760\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"456ca844-322a-4fcf-90ea-531b14db9d9e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Transformers\"\n ],\n [\n \"section\",\n \"Conclusion and Future Directions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion and Future Directions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":86,"string":"86"},"all_cites":{"kind":"list like","value":[7360,1447,1445,7052,2401,732,1446,38,97,57,208,553,1474,1493,1457,1482,1460,1491,7040,1484,1494,1464,1488,1470,1452,1466,1490,793,1502,7054,1467,1463,8384,1456,1480,794,1475,1465,7361,1500,1459,7053,1458,1479,7365,1496,1133,8454,768,7367,7333,1489,1451,1469,9,1476,1453,1495,1487,1462,1478,679,1483,1492,7369,7,1473,7370,8456,826,7362,7364,7363,1468,1486,7298,1498,1497,1501,7371,8457,7368,1455,707,8455,1471,1481,1182,1272,7339,1477,1449,7273,1461,1454,1448,1485,7366,798,1499,1450,1472,4722,134,1503,1504,725,1505,7372,306,1507,1506,790,1508,1509,1510,1511,71,1512,7055,680,683,544,7373,7056,1514,8459,8458,1515,1516,1513],"string":"[\n 7360,\n 1447,\n 1445,\n 7052,\n 2401,\n 732,\n 1446,\n 38,\n 97,\n 57,\n 208,\n 553,\n 1474,\n 1493,\n 1457,\n 1482,\n 1460,\n 1491,\n 7040,\n 1484,\n 1494,\n 1464,\n 1488,\n 1470,\n 1452,\n 1466,\n 1490,\n 793,\n 1502,\n 7054,\n 1467,\n 1463,\n 8384,\n 1456,\n 1480,\n 794,\n 1475,\n 1465,\n 7361,\n 1500,\n 1459,\n 7053,\n 1458,\n 1479,\n 7365,\n 1496,\n 1133,\n 8454,\n 768,\n 7367,\n 7333,\n 1489,\n 1451,\n 1469,\n 9,\n 1476,\n 1453,\n 1495,\n 1487,\n 1462,\n 1478,\n 679,\n 1483,\n 1492,\n 7369,\n 7,\n 1473,\n 7370,\n 8456,\n 826,\n 7362,\n 7364,\n 7363,\n 1468,\n 1486,\n 7298,\n 1498,\n 1497,\n 1501,\n 7371,\n 8457,\n 7368,\n 1455,\n 707,\n 8455,\n 1471,\n 1481,\n 1182,\n 1272,\n 7339,\n 1477,\n 1449,\n 7273,\n 1461,\n 1454,\n 1448,\n 1485,\n 7366,\n 798,\n 1499,\n 1450,\n 1472,\n 4722,\n 134,\n 1503,\n 1504,\n 725,\n 1505,\n 7372,\n 306,\n 1507,\n 1506,\n 790,\n 1508,\n 1509,\n 1510,\n 1511,\n 71,\n 1512,\n 7055,\n 680,\n 683,\n 544,\n 7373,\n 7056,\n 1514,\n 8459,\n 8458,\n 1515,\n 1516,\n 1513\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.148346752986971,"string":"1.148347"}}},{"rowIdx":53,"cells":{"authors":{"kind":"list like","value":["Ana Paula Chaves","Marco Aurelio Gerosa"],"string":"[\n \"Ana Paula Chaves\",\n \"Marco Aurelio Gerosa\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"},"year":{"kind":"string","value":"2019"},"date":{"kind":"timestamp","value":"2019-04-04T18:43:31Z","string":"2019-04-04T18:43:31Z"},"category":{"kind":"string","value":"cs.HC"},"abstract":{"kind":"string","value":" Chatbots' growing popularity has brought new challenges to HCI, having changed the patterns of human interactions with computers. The increasing need to approximate conversational interaction styles raises expectations for chatbots to present social behaviors that are habitual in human-human communication. In this survey, we argue that chatbots should be enriched with social characteristics that cohere with users' expectations, ultimately avoiding frustration and dissatisfaction. We bring together the literature on disembodied, text-based chatbots to derive a conceptual model of social characteristics for chatbots. We analyzed 56 papers from various domains to understand how social characteristics can benefit human-chatbot interactions and identify the challenges and strategies to designing them. Additionally, we discussed how characteristics may influence one another. Our results provide relevant opportunities to both researchers and designers to advance human-chatbot interactions. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"]],"subsections":["3e165ffa-29cf-4605-ba88-0e6c84a411e5","ad5096ab-28b5-4a2f-b84b-74bcbea02a02","5950bb1f-b30c-410b-94c0-7eba5ce91d72","616598c0-a6ef-4e0d-bae2-49c648e87606","cefcc373-1755-478a-bdc0-b708ace17a7a","e47b0d6a-5496-42d1-b93d-9f8d302b471a","1430d8e4-a5c1-4f7d-ae08-3a9fcec59bb4"],"title":"root"},{"cite_extract_rate":0.026315789473684,"cites":[6262],"content":"\\label{sec:introduction}\nChatbots are computer programs that interact with users in natural language . The origin of the chatbot concept dates back to 1950 . ELIZA and A.L.I.C.E. are examples of early chatbot technologies, where the main goal was to mimic human conversations. Over the years, the chatbot concept has evolved. Today, chatbots can have distinct and diverse characteristics, which has resulted in several synonyms, such as multimodal agents, chatterbots, and conversational interfaces. In this survey, we use the term ``\\textit{chatbot}'' to refer to \\textit{a disembodied conversational agent that holds a natural language conversation via text-based environment to engage the user in either a general-purpose or task-oriented conversation.}\nChatbots are changing the patterns of interactions between humans and computers . Many instant messenger tools, such as Skype, Facebook Messenger, and Telegram provide platforms to develop and deploy chatbots, which either engage with users in general conversations or help them solve domain specific tasks . As messaging tools increasingly become platforms, traditional websites and apps are providing space for this new form of human-computer interaction (HCI) . For example, in the 2018 F8 Conference, Facebook announced that it had 300K active chatbots on Facebook Messenger . The BotList \\footnote{https://botlist.co/} website indexes thousands of chatbots for education, entertainment, games, health, productivity, travel, fun, and several other categories. The growth of chatbot technology is changing how companies engage with their customers , students engage with their learning groups , and patients self-monitor the progress of their treatment , among many other applications.\nHowever, chatbots still fail to meet users' expectations . While many studies on chatbot design focus on improving chatbots' functional performance and accuracy (see, e.g., ), the literature has consistently suggested that chatbots' interactional goals should also include social capabilities . According to the Media Equation theory , people naturally respond to social situations when interacting with computers . As chatbots are designed to interact with users in a way that mimics person-to-person conversations, new challenges in HCI arise . Neururer and colleagues (\\citeyear{neururer2018perceptions}) state that making a conversational agent acceptable to users is primarily a social, not only technical, problem. Studies on chatbots have shown that people prefer agents who: conform to gender stereotypes associated with tasks ; self-disclose and show reciprocity when recommending ; and demonstrate a positive attitude and mood . When chatbots do not meet these expectations, the user may experience frustration and dissatisfaction . In contrast, designing overly humanized agents results in uncanny feelings and increased expectations , which also negatively impacts the interaction. The challenge remains as to what social characteristics are relevant for improving chatbots' communication and social skills and in which domains they have shown to be beneficial.\nAlthough chatbots' social characteristics have been explored in the literature, this knowledge is spread across several domains in which chatbots have been studied, such as customer services, education, finances, and travel. In the HCI domain, some studies focus on investigating the social aspects of human-chatbot interactions (see, e.g., ). However, most studies focus on a single or small set of characteristics (e.g., ); in other studies, the social characteristics emerged as secondary, exploratory results (e.g., ). It has become difficult to find evidence regarding what characteristics are important for designing a particular chatbot, and what research opportunities exist in the field.\nWhilst the literature has extensively reviewed the technical aspects of chatbots design (e.g., ), a lack of studies brings together the social characteristics that influence the way users perceive and behave toward chatbots. To fill this gap, this survey compiles research initiatives for understanding the impact of chatbots' social characteristics on the interaction. We bring together literature that is spread across several research areas. From our analysis of 56 scientific studies, we derive a conceptual model of social characteristics, aiming to help researchers and designers identify the subset of characteristics that are relevant to their context and how adopting--or neglecting--a particular characteristic may influence the way humans perceive the chatbots. The research question that guided our investigation was: \\textbf{What chatbot social characteristics benefit human interaction and what are the challenges and strategies associated with them?}\nTo answer this question, we discuss why designing a chatbot with a particular characteristic can enrich the human-chatbot interaction. Our results provide insight into whether the characteristic is desirable for a particular chatbot, so that designers can make informed decisions by selecting the appropriate subset of characteristics, as well as inspire researchers' further investigations. In addition, we discuss the influence of humanness and the conversational context on users' perceptions as well as the interrelationship among the identified characteristics. We stated 22 propositions about how social characteristics may influence one another. In the next section, we present an overview of the studies included in this survey.","id":"3e165ffa-29cf-4605-ba88-0e6c84a411e5","level":"section","origin_cites_number":38,"parent_id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0.16666666666666602,"cites":[6263,6262],"content":"\\label{sec:overview}\nThe literature presents no coherent definition of chatbots; thus, to find relevant studies we used a search string that includes the synonyms \\textit{``chatbots, chatterbots, conversational agents, conversational interfaces, conversational systems, conversation systems, dialogue systems, digital assistants, intelligent assistants, conversational user interfaces, and conversational UI''}. We explicitly left out studies that relate to embodiment (\\textit{``ECA, multimodal, robots, eye-gaze, gesture''}), and speech input mode (\\textit{``speech-based, speech-recognition, voice-based''}). The search string did not include the term ``social bots'', because it refers to chatbots that produce content for social networks such as Twitter . We did not include ``personal assistants'' either, since this term consistently refers to commercially available, voice-based assistants such as Google Assistant, Amazon's Alexa, Apple's Siri, and Microsoft's Cortana.\nWe decided to not include terms that relate to social characteristics/traits, because most studies do not explicitly label their results as such. Additionally, to find studies from a variety of domains (not only computing), we used Google Scholar to search for papers. We started with a set of about one thousand papers. In the first round, we reduced this amount by about half based on the titles. We also removed papers for which full text was not available or written in a language other than English. We started the second round with 464 studies. In this round, we excluded short/position papers, theses, book chapters, and documents published in non-scientific venues. We also read the abstracts and removed all the papers that focused on technical aspects, rather than social ones. The third and last round started with 104 papers. After reading the full texts, we removed the papers that either did not highlight social characteristics or represented previous, ongoing research of another complete study from the same authors. The datasets with all analyzed studies (original list of results and exclusions per round) is available in .\nAfter filtering the search results, we had 56 remaining studies. Most of the selected studies are recent publications (less than 10 years old). The publication venues include the domains of human-computer interactions (25 papers), learning and education (8 papers), information and interactive systems (8 papers), virtual agents (5 papers), artificial intelligence (3 papers), and natural language processing (2 papers). We also found papers from health, literature \\& culture, computer systems, communication, and humanities (1 paper each). Most papers (59\\%) focus on task-oriented chatbots. General purpose chatbots reflect 33\\% of the surveyed studies. Most general purpose chatbots (16 out of 19) are designed to handle topic-unrestricted conversations. The most representative specific domain is education, with 9 papers, followed by customer services, with 5 papers. See the supplementary materials for the complete list of topics. \nMost surveyed studies adopted real chatbots (35 out of 56); among them, 18 studies analyze logs of conversations or users' perceptions of third-party chatbots such as Cleverbot (e.g., ), Talkbot (e.g., ), and Woebot . Nine studies introduce a self-developed architecture and/or dialogue management (e.g. ). In another nine studies, the chatbots were designed for research purposes using third-party platforms for chatbot development and deployment, such as IBM's Watson service (e.g., ) and Microsoft's Bot Framework as well as pattern-matching packages such as Artificial Intelligence Markup Language (AIML) . When a chatbot was simulated (11 studies), Wizard of Oz (WoZ) is the most used technique (9 studies). In a WoZ study, participants believe to be interacting with a chatbot when, in fact, a person (or ``wizard'') pretends to be the automated system . Eight studies do not address a particular chatbot. See the supplementary materials for details.\nWe analyzed the papers by searching for chatbot behavior or attributed characteristics that influence the way users perceive it and behave toward it. Noticeably, the characteristics and categories are seldom explicitly highlighted to in the literature, so the conceptual model was derived using a qualitative coding process inspired by methods such as Grounded Theory (open coding stage). For each study (\\textit{document}), we selected relevant statements from the paper (\\textit{quotes}) and labeled them as a characteristic (\\textit{code}). These steps were performed by one researcher (the first author). After coding all the studies, a second researcher (the second author) reviewed the produced set of characteristics and both researchers participated in discussion sessions to identify characteristics that could be merged, renamed, or removed. At the end, the characteristics were grouped into the categories, depending on whether the characteristic relates to the chatbot's virtual representation, conversational behavior, or social protocols. Finally, the quotes for each characteristic were labeled as references to benefits, challenges, or strategies.\nWe derived a total of 11 social characteristics, and grouped them into three categories, as depicted in Table \\ref{tab:conceptualmodel}: \\textbf{conversational intelligence}, \\textbf{social intelligence}, and \\textbf{personification}. The next section describes the derived conceptual model.\n\\begin{landscape}\n\\begin{table}[ht]\n\\scriptsize\n\\centering\n\\caption{Conceptual model of chatbots social characteristics. Eleven characteristics found in the literature were grouped into three main categories, depending on whether the social characteristic relates to the chatbot's conversational behavior, social protocols, or virtual representation. For each characteristic, we point out the benefits, challenges, and strategies of implementing the characteristic in chatbots as reported in the literature.}\n\\vspace{1mm}\n\\begin{tabular}{c|p{2.5cm}|p{6cm}|p{6cm}|p{6cm}|}\n\\cline{2-5}\n & \\textbf{\\shortstack{\\\\Social\\\\Characteristics}}& \\multicolumn{1}{c|}{\\textbf{Benefits}} & \\multicolumn{1}{c|}{\\textbf{Challenges}} & \\multicolumn{1}{c|}{\\textbf{Strategies}} \\\\ \\hline\n\\multicolumn{1}{|c|}{\\multirow{14}{*}{\\rotatebox{90}{\\textbf{\\shortstack{Conversational\\\\Intelligence}}}}} & \\multirow{6}{*}{\\textit{Proactivity}} & \\textbf{[B1]} to provide additional information & \\textbf{[C1]} timing and relevance & \\textbf{[S1]} to leverage conversational context \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to inspire users and to keep the conversation alive & \\textbf{[C2]} privacy & \\textbf{[S2]} to select a topic randomly \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B3]} to recover from a failure & \\textbf{[C3]} users' perception of being controlled & \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B4]} to improve conversation productivity & & \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B5]} to guide and engage users & & \\\\ \\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{4}{*}{\\textit{Conscientiousness}} & \\textbf{[B1]} to keep the conversation on track & \\textbf{[C1]} to handle task complexity & \\textbf{[S1]} conversational flow \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to demonstrate understanding & \\textbf{[C2]} to harden the conversation & \\textbf{[S2]} visual elements \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B3]} to hold a continuous conversation & \\textbf{[C3]} to keep the user aware of the chatbot's context & \\textbf{[S3]} confirmation messages \\\\ \\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{4}{*}{\\textit{Communicability}} & \\textbf{[B1]} to unveil functionalities & \\textbf{[C1]} to provide business integration & \\textbf{[S1]} to clarify the purpose of the chatbot \\\\ \n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to manage the users' expectations & \\textbf{[C2]} to keep visual elements consistent with textual inputs & \\textbf{[S2]} to advertise the functionality and suggest the next step \\\\\n\\multicolumn{1}{|l|}{} & & & & \\textbf{[S3]} to provide a help functionality \\\\\\hline\n\\multicolumn{1}{|l|}{\\multirow{19}{*}{\\rotatebox{90}{\\textbf{\\shortstack{Social\\\\Intelligence}}}} }& \\multirow{6}{*}{\\textit{Damage control}} & \\textbf{[B1]} to appropriately respond to harassment & \\textbf{[C1]} to deal with unfriendly users & \\textbf{[S1]} emotional reactions \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to deal with testing & \\textbf{[C2]} to identify abusive utterances & \\textbf{[S2]} authoritative reactions \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B3]} to deal with lack of knowledge & \\textbf{[C3]} to fit the response to the context & \\textbf{[S3]} to ignore the user's utterance and change the topic \\\\\n\\multicolumn{1}{|l|}{} & & & & \\textbf{[S4]} \\textit{conscientiousness} and \\textit{communicability} \\\\\n\\multicolumn{1}{|l|}{} & & & & \\textbf{[S5]} to predict users' satisfaction \\\\\\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{2}{*}{\\textit{Thoroughness}} & \\textbf{[B1]} to increase human-likeness & \\textbf{[C1]} to decide how much to say & Not identified \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to increase believability & \\textbf{[C2]} to be consistent & \\\\ \\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{2}{*}{\\textit{Manners}} & \\textbf{[B1]} to increase human-likeness & \\textbf{[C1]} to deal with face-threatening acts & \\textbf{[S1]} to engage in small talk \\\\\n\\multicolumn{1}{|l|}{} & & & \\textbf{[C2]} to end a conversation gracefully & \\textbf{[S2]} to adhere turn-taking protocols \\\\ \\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{3}{*}{\\textit{Moral agency}} & \\textbf{[B1]} to avoid stereotyping & \\textbf{[C1]} to avoid alienation & Not identified \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to enrich interpersonal relationships & \\textbf{[C2]} to build unbiased training data and algorithms & \\\\ \\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{3}{*}{\\textit{\\shortstack{Emotional\\\\intelligence}}} & \\textbf{[B1]} to enrich interpersonal relationships & \\textbf{[C1]} to regulate affective reactions & \\textbf{[S1]} to use social-emotional utterances \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to increase engagement & & \\textbf{[S2]} to manifest \\textit{conscientiousness} \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B3]} to increase believability & & \\textbf{[S3]} reciprocity and self-disclosure \\\\ \\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{3}{*}{\\textit{Personalization}} & \\textbf{[B1]} to enrich interpersonal relationships & \\textbf{[C1]} privacy & \\textbf{[S1]} to learn from and about the user \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to provide unique services & & \\textbf{[S2]} to provide customizable agents \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B3]} to reduce interactional breakdowns & & \\textbf{[S3]} visual elements \\\\ \\hline\n\\multicolumn{1}{|l|}{\\multirow{5}{*}{\\rotatebox{90}{\\textbf{\\shortstack{Personifi-\\\\cation}}}} }& \\multirow{3}{*}{\\textit{Identity}} & \\textbf{[B1]} to increase engagement & \\textbf{[C1]} to avoid negative stereotypes & \\textbf{[S1]} to design and elaborate on a persona \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to increase human-likeness & \\textbf{[C2]} to balance the \\textit{identity} and the technical capabilities & \\\\\\cline{2-5}\n\\multicolumn{1}{|l|}{} & \\multirow{2}{*}{\\textit{Personality}} & \\textbf{[B1]} to increase believability & \\textbf{[C1]} to adapt humor to the users' culture & \\textbf{[S1]} to use appropriate language \\\\\n\\multicolumn{1}{|l|}{} & & \\textbf{[B2]} to enrich interpersonal relationships & \\textbf{[C2]} to balance the \\textit{personality} traits & \\textbf{[S2]} to have a sense of humor \\\\\n\\hline\n\\end{tabular}\n\\label{tab:conceptualmodel}\n\\end{table}\n\\end{landscape}","id":"ad5096ab-28b5-4a2f-b84b-74bcbea02a02","level":"section","origin_cites_number":12,"parent_id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Overview of the surveyed literature"]],"subsections":[],"title":"Overview of the surveyed literature"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:socialcharacteristics}\nThis section describes the identified social characteristics grouped into categories. As Table \\ref{tab:conceptualmodel} depicts, the category \\textbf{conversational intelligence} includes characteristics that help the chatbot manage interactions. \\textbf{Social intelligence} focuses on habitual social protocols, while \\textbf{personification} refers to the chatbot's perceived identity and personality representations. We also grouped the social characteristics based on the domain in which they were investigated (Table \\ref{tab:domain-analysis}). In the following subsections, we describe the identified social characteristics as well as the domains of study. Then, we summarize the relationship among the characteristics in Section \\ref{sec:interrelationship}. For each category, a table with an overview of the surveyed studies is provided in the supplementary materials. The supplementary materials also include tables for each social characteristic, listing the studies associated with the domains of study and reported benefits, challenges, and strategies. Finally, the supplementary materials also highlight five constructs that can be used to assess whether social characteristics reach the intended design goals.","id":"5950bb1f-b30c-410b-94c0-7eba5ce91d72","level":"section","origin_cites_number":0,"parent_id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"]],"subsections":["9ad3c98e-3389-49e5-a57e-7e261302fc58","6e94027f-a398-4b65-8b3b-d03e95d73fa6","fb12c3f5-18c5-4fb9-8465-0d7950752ebc"],"title":"Chatbots Social Characteristics"},{"cite_extract_rate":0.035087719298245,"cites":[6262,7476],"content":"\\label{sec:conversationalintelligence}\n\\textbf{Conversational intelligence} enables the chatbot to actively participate in the conversation and to demonstrate awareness of the topic discussed, the evolving conversational context, and the flow of the dialogue. Therefore, \\textbf{conversational Intelligence} refers to the ability of a chatbot to effectively converse beyond the technical capability of achieving a conversational goal . In this section, we discuss social characteristics related to \\textbf{conversational intelligence}, namely: \\textit{proactivity} (18 studies), \\textit{conscientiousness} (11 studies), and \\textit{communicability} (6 studies). Most of the studies rely on data that comes from the log of the conversations, interviews, and questionnaires. The questionnaires are mostly Likert-scales, and some of them include subjective feedback. Most studies analyzed the interaction with real chatbots, although Wizard of Oz (WoZ) settings are also common. Only two papers did not evaluate a particular type of interaction because they were based on a literature review and surveys with chatbot users in general . Only five studies applied only quantitative methods, while seven focused on qualitative methods. The majority of the studies (15) applied mixed methods (both quantitative and qualitative). See the supplementary materials for details.\n\\begin{table}[!btp]\n\\scriptsize\n\\centering\n\\caption{Studied social characteristics per domain. Research in open domain chatbots has reported most of the social characteristics, except for \\textit{Communicability}. In he task-oriented domains, some characteristics are largely influenced by the topic (e.g., \\textit{Moral agency} and \\textit{Personality}), while others are more generally applied (e.g., \\textit{Manners} and \\textit{Damage control}).}\n\\begin{tabular}{p{1.6cm}p{5cm}p{6.5cm}}\n\\hline\n\\textbf{Domain} & \\textbf{Social Characteristics} & \\textbf{Studies} \\\\\n\\hline\n\\multirow{2}{*}{\\shortstack[l]{Open\\\\domain}} & \\textit{Proactivity, Conscientiousness, Damage control, Thoroughness, Manners, Moral agency, Emotional intelligence, Personalization, Identity, Personality} & \\\\ \n\\hline\nEthnography & \\textit{Proactivity, Conscientiousness, Thoroughness, Personalization} & \\\\ \n\\hline\n\\multirow{2}{*}{\\shortstack[l]{Task\\\\management}} & \\textit{Proactivity, Damage control, Manners, Personalization, Identity} & \\\\ \n\\hline\nTourism & \\textit{Proactivity, Thoroughness, Manners} & \\\\ \n\\hline\nBusiness & \\textit{Proactivity, Personalization} & \\\\ \n\\hline\nInformation search & \\textit{Proactivity, Damage control, Manners, Emotional intelligence} & \\\\ \n\\hline\nDecision-making & \\textit{Proactivity, Damage control, Manners} & \\\\\n\\hline\nHealth-care & \\textit{Proactivity, Emotional intelligence} & \\\\ \n\\hline\n\\shortstack{Credibility\\\\assessment} & \\textit{Proactivity, Conscientiousness} & \\\\ \n\\hline\nEducation & \\textit{Proactivity, Conscientiousness, Damage control, Thoroughness, Manners, Emotional intelligence, Identity, Personality} & \\\\ \n\\hline\n\\multirow{2}{*}{\\shortstack[l]{Financial\\\\services}} & \\textit{Conscientiousness, Communicability, Damage control, Thoroughness, Personalization, Identity} & \\\\ \n\\hline\n\\multirow{2}{*}{\\shortstack[l]{Customer\\\\services}} & \\textit{Conscientiousness, Communicability, Damage control, Thoroughness, Manners, Emotional intelligence, Personalization, Identity} & \\\\ \n\\hline\nE-commerce & \\textit{Conscientiousness, Manners} & \\\\ \n\\hline\nNews & \\textit{Communicability} & \\\\ \n\\hline\n\\multirow{2}{*}{\\shortstack[l]{Human\\\\resources}} & \\textit{Communicability, Damage control, Manners, Identity} & \\\\ \n\\hline\n\\multirow{2}{*}{\\shortstack[l]{Virtual\\\\assistant}} & \\textit{Thoroughness, Emotional intelligence, Personalization, Identity} & \\\\ \n\\hline\nGaming & \\textit{Thoroughness, Emotional intelligence, Personality} & \\\\ \n\\hline\nRace-talk & \\textit{Moral agency, Identity} & \\\\ \n\\hline\nHumorous talk & \\textit{Personality} & \\\\ \n\\hline\nNot defined & \\textit{Proactivity, Conscientiousness, Communicability, Damage control, Personalization, Identity, Personality} & \\\\\n\\hline\n\\end{tabular}\n\\label{tab:domain-analysis}\n\\end{table}","id":"9ad3c98e-3389-49e5-a57e-7e261302fc58","level":"subsection","origin_cites_number":57,"parent_id":"5950bb1f-b30c-410b-94c0-7eba5ce91d72","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Conversational Intelligence"]],"subsections":["30b70a69-50cb-4695-9975-c797ccdef4e2","8d7c33ab-b84e-4247-bfcd-ac939fd2d82f","d1a8a8a3-347c-4cde-8b5a-76ed8060b75f"],"title":"Conversational Intelligence"},{"cite_extract_rate":0.09090909090909001,"cites":[6262,7476],"content":"\\label{sec:proactivity}\n\\textit{Proactivity} is the capability of a system to \nautonomously act on the user's behalf and thereby reduce the amount of human effort to complete a task . In human-chatbot conversations, a proactive behavior enables a chatbot to share initiative with the user, contributing to the conversation in a more natural way . Chatbots may manifest \\textit{proactivity} when they initiate exchanges, suggests new topics, provide additional information, or formulate follow-up questions. In this survey, we found 18 papers that report either chatbots with proactive behavior or implications of manifesting a proactive behavior. \\textit{Proactivity} (also addressed as ``\\textit{intervention mode}'') was explicitly addressed in seven studies . In most of the studies, however, \\textit{proactivity} emerged either as an exploratory result, mostly from post-intervention interviews and user's feedback , or as a strategy to attend to domain-specific requirements (e.g., monitoring, and guidance) . \\textit{Proactivity} was mostly investigated in open domain and education chatbots (four studies each). In open domain chatbots, \\textit{proactivity} helps improve engagement by introducing new topics to keep the conversation alive . Educational chatbots rely on \\textit{proactivity} to prompt students to think, share, and collaborate (e.g., ). \\textit{Proactivity} was also observed in eight other task-oriented domains, including task management and information searches as can be observed in the supplementary materials.\nThe surveyed literature evidences several benefits of chatbot \\textit{proactivity} in chatbots: \n\\textbf{[B1] to provide additional, useful information:} literature reveals that \\textit{proactivity} in chatbots adds value to interactions . Investigating evaluation criteria for chatbots, asked users of a general purpose chatbot to rate the chatbots' naturalness and report in what areas they excel. Both statistical and qualitative results confirm that taking the lead and suggesting specialized information about the conversation theme correlate to chatbots' naturalness. corroborates this result; in post-intervention interviews, ten out of 14 users mentioned they preferred a chatbot that takes the lead and volunteers additional information, such as useful links and song playlists. In a WoZ study, investigated whether proactive interventions of a chatbot contribute to a collaborative search in a group chat. The chatbot either elicits or infers needed information from the collaborative chat and proactively intervenes in the conversation by sharing useful search results. The intervention modes were not significantly different from each other, but both intervention modes resulted in a statistically significant increase of enjoyment and decrease of effort when compared to the same task with no chatbot interventions. Moreover, in a post-intervention, open-ended question, 16 out of 98 participants self-reported positive perceptions about the provided additional information.\n\\textbf{[B2] to inspire users, and keep the conversation alive: }proactively suggesting and encouraging new topics have been shown useful to both inspire users and keep the conversation alive . Participants in the study conducted by self-reported that the chatbot's suggestions helped them to get started (7 mentions) and gave them ideas about search topics (4 mentions). After iteratively evaluating prototypes for a chatbot in an educational scenario, concluded that proactively initiating topics makes the dialogue more fun and reveals topics the chatbot can talk about. The refined prototype also proactively maintains the engagement by posing a follow-up when the student had not provided an answer to the question. hypothesized that including follow-up questions based on the content of previous messages would result in higher perceived partner engagement. The hypothesis was supported, with participants in the dynamic condition rating the chatbot as more engaging. In an ethnographic data collection , users included photos in their responses to add information about their experience; 85\\% of these photos were proactively prompted by the chatbot. This result shows that prompting the user for more information stimulates them to expand their entries. also observed that chatbots' proactive messages provided insights about the chatbots' knowledge, which potentially helped the conversation to continue. In this paper, we call the strategies to convey the chatbot's knowledge and capabilities as \\textit{communicability}, and we discuss it in Section \\ref{sec:communicability}.\n\\textbf{[B3] to recover the chatbot from a failure: }in and , \\textit{proactivity} is employed to naturally recover from a failure. In both studies, the approach was to introduce a new topic when the chatbot failed to understand the user or could not find an answer, preventing the chatbot from getting stuck and keeping the conversation alive. Additionally, in , the chatbot inserted new topics when users are either abusive or non-sensical. We call the strategies to handle failure and abusive behavior as \\textit{damage control}, and we discuss this characteristic in Section \\ref{sec:damagecontrol}.\n\\textbf{[B4] to improve conversation productivity: }in task-oriented interactions, such as searching or shopping, \\textit{proactivity} can improve the conversation's productivity . In interviews with first-time users of chatbots, found that chatbots should ask follow-up questions to resolve and maintain the context of the conversation and reduce the time searching before achieving the goal. found similar results for collaborative searches; 28 out of 98 participants self-reported that chatbot's proactive interventions saved collaborators time.\n\\textbf{[B5] to guide and engage users: }in particular domains, \\textit{proactivity} helps chatbots to either guide users or establish and monitor users' goals. In , the chatbot assigns a goal to the user and proactively prompts motivational messages and reminders to keep the user engaged in the treatment. suggest that a decision-making coach chatbot needs to lead the interaction toward guiding the user to a decision. In ethnographic data collection , the chatbot prompts proactive messages that guide the users on what information they need to report. evaluates a chatbot that manage tasks in a workplace. Proactive messages are used to check whether the team member has concluded the tasks, and then report the outcome to the other stakeholders. In the educational context, \\textit{proactivity} is used to develop tutors that engage the students and facilitate learning. In , the tutor chatbot was designed to provide examples of how other students explained a topic. The network analysis of the learner's textual inputs shows that students used more key terms and provided more important messages when receiving feedback about other group members. In , , and the chatbots prompt utterances to encourage the students to reason about a topic. In all three studies, the chatbot condition provided better learning outcomes and increased students' engagement in the discussions.\nThe surveyed papers also highlight challenges in providing proactive interactions, such as timing and relevance, privacy, and the user's perception of being controlled.\n\\textbf{[C1] timing and relevance:} untimely and irrelevant proactive messages may compromise the success of the interaction. states that untimely turn-taking behavior was perceived as annoying, negatively affecting emotional engagement. and reported that \\textit{proactivity} can be disruptive. investigated \\textit{proactivity} in a workspace environment, hypothesizing that the perceived interruption of agent \\textit{proactivity} negatively affects users' opinion's. The hypothesis was supported, and the authors found that influencing the sense of interruption was the general aversion to unsolicited messages, regardless of whether they came from a chatbot or a colleague. showed that proactively introducing new topics resulted in a high number of ignored messages. The analysis of the conversation log reviewed that either the new topics were not relevant, or it was not the proper time to start a new topic. also reported annoyance when a chatbot introduces repetitive topics.\n\\textbf{[C2] privacy:} in a work-related, group chat, observed privacy concerns regarding the chatbot ``reading'' the employees' conversations to proactively act. During a semi-structured interview, researchers presented a mockup of the chatbot to employees from two different enterprises and collected perceptions of usefulness, intrusiveness, and privacy. Employees reported feeling that the chatbot represented their supervisors' interests, which conveyed as sense of workplace surveillance. Privacy concerns may result in under-motivated users, discomfort about disclosing information, and lack of engagement .\n\\textbf{[C3] user's perception of being controlled:} \\textit{proactivity} can be annoying when the chatbot conveys the impression of trying to control the user. report that seven out of 13 participants experienced irritation with the chatbot; one of the most frequent reasons was due to the chatbot directing them to specific places. For the task management chatbot, reported to have adapted the follow-up questions approach to pose questions in a time negotiated with the user. In a previous implementation, the chatbot checked the status of the task twice a day, which participants considered too frequent and annoying.\nThe surveyed literature also reveals two strategies to provide \\textit{proactivity}: leveraging the conversational context and randomly selecting a topic. \\textbf{[S1] Leveraging the conversational context} is the most frequent strategy , in which proactive messages relate to contextual information provided in the conversation to increase the usefulness of interventions . argue that general purpose, emotionally aware chatbots should recognize users' interests and intents from the conversational context to proactively offer comfort and relevant services. In , the chatbot leverages conversational context to suggest new topics and propose to share documents or links to assist employees in a work-related group chat. The chatbots studied by introduce new topics based on keywords from previous utterances posted in the chat. According to , leveraging the context can also help smoothly guide the user to a target topic. Researchers in the chatbots domain can refer to the emerging literature on Ambient Intelligence (see, e.g., ~) to understand how contextual knowledge can be leveraged to convey proactivity. One surveyed paper proposes a chatbot that \\textbf{[S2] selects a topic randomly} but also observes that the lack of context is a major problem of this approach. Contextualized proactive interventions also suggest that the chatbot should be attentive to the conversation, which conveys \\textit{conscientiousness}, as discussed in the next section.","id":"30b70a69-50cb-4695-9975-c797ccdef4e2","level":"subsubsection","origin_cites_number":22,"parent_id":"9ad3c98e-3389-49e5-a57e-7e261302fc58","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Conversational Intelligence"],["subsubsection","Proactivity"]],"subsections":[],"title":"Proactivity"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conscientiousness}\n\\textit{Conscientiousness} is a chatbot's capacity to demonstrate attentiveness to the conversation at hand . It enables a chatbot to follow the conversational flow, show understanding of the context, and interpret each utterance as a meaningful part of the whole conversation . In this survey, we found 11 papers that reported findings related to \\textit{conscientiousness} for chatbot design. Four studies explicitly investigated influences of \\textit{conscientiousness} for chatbots . In the remaining studies, \\textit{conscientiousness} emerged in exploratory findings. In , \\textit{conscientiousness} emerged from the analysis of conversational logs, while elicited \\textit{conscientiousness} aspects as a requirement for chatbot design when surveying the literature on customer service chatbots. In the remaining studies , \\textit{conscientiousness} issues were self-reported by the users in post-intervention interviews and subjective feedback to open-ended questions. \\textit{Conscientiousness} was investigated in three studies from the education domain, where the tutor chatbot is expected to be attentive and relevant as well as to control the conversation flow to keep the students focused on the topic (see e.g., ). \\textit{Conscientiousness} was also investigated in seven other domains, including open domain interactions and customer and financial services. See the supplementary materials for a complete list of domains of studies that report \\textit{conscientiousness} as a social characteristic for chatbots.\nThe surveyed literature evidenced benefits of designing conscientious chatbots:\n\\textbf{[B1] to provide meaningful answers:} some chatbots use simplistic approaches, like pattern matching rules based on keywords or template phrases applied in the last user utterance, to find the most appropriate response . However, as the chatbot does not interpret the meaning and the intent of users' utterances, the best-selected response may still sound irrelevant to the conversation . As shown by , when a chatbot does not interpret the meaning of users' utterances, users show frustration and the chatbot's credibility is compromised. This argument is supported by when studying chatbots to facilitate collaborative learning. The authors proposed a chatbot that promotes Academically Productive Talk moves. Exploratory results show that the chatbot performed inappropriate interventions, which was perceived as a lack of attention to the conversation. In , participants complained that some chatbots seemed ``\\textit{completely scripted,}'' ignoring user's inputs that did not fall into the script. In this case, users needed to adapt their inputs to match the chatbot script to be understood, which resulted in dissatisfaction. Besides avoiding frustration, showed that \\textit{conscientiousness} also influences chatbots' perceived humanness and social presence. The authors invited participants to interact with a chatbot that shows an image and asks the users to describe it. To perform the task, the chatbot could either ask the same generic follow-up questions each time (nonrelevant condition) or respond with a follow-up question related to the last participant's input (relevant condition), demonstrating attention to the information provided by the user. Statistical analysis of survey rates supported that the relevant condition increased chatbots' perceived humanness and social presence. In , to motivate the students to communicate in a second language, the proposed chatbot interprets users' input to detect breakdowns and react using an appropriate communication strategy. In interviews, participants reported ``\\textit{appreciating the help they got from the chatbot to understand and express what they have got to say}'' . This study was later extended to the context of ECAs (see for details).\n\\textbf{[B2] to hold a continuous conversation: }a conversation with a chatbot should maintain a ``\\textit{sense of continuity over time}'' to demonstrate that the chatbot is exerting effort to track the conversation. To do so, it is essential to maintain the topic. When evaluating the naturalness of a chatbot, found that maintaining a theme is convincing, while failure to do so is unconvincing. Furthermore, based on the literature on customer support chatbots, argue that comfortably conversing on any topic related to the service offering is a requirement for task-oriented chatbots. reviewed popular chatbots for practicing second languages. The author showed that most chatbots in this field cannot hold continuous conversations, since they are developed to answer the user's last input. Therefore, they did not have the sense of topic, which resulted in instances of inappropriate response. While the chatbots could change the topic, they could not sustain it afterward. Showing \\textit{conscientiousness} also requires the chatbot to understand and track the context, which is particularly important in task-oriented scenarios. In , first-time users stressed positive experiences with chatbots that retained information from previous turns. Two participants also expected the chatbots to retain this context across sessions, thus reducing the need for the user's additional input per interaction. Keeping the context across sessions was highlighted as a strategy to convey \\textit{personalization} and empathy (see Sections \\ref{sec:personalization} and \\ref{sec:emotionalintelligence}).\n\\textbf{[B3] to steer the conversation toward a productive direction:} in task-oriented interactions, a chatbot should understand the purpose of the interaction and strive to conduct the conversation toward this goal in an efficient, productive way . show that productivity is the key motivating factor for using chatbots (68\\% of the participants mentioned it as the main reason for using chatbots). First-time users in self-reported that interacting with chatbots should be more productive than using websites, phone apps, and search engines. In this sense, compared the user experience when interacting with a chatbot for solving either simple or complex tasks in a financial context. The authors found that, for complex tasks, to keep the conversation on track the user must be aware of the next steps or why something is happening. In the educational context, proposed a dialogue management approach based on communication strategies to enrich a chatbot with the capability to express its meaning when faced with challenges. Statistical results show that the communication strategies, combined with affective backchannel (which is detailed in Section \\ref{sec:emotionalintelligence}), are effective in motivating students to communicate and maintain the task flow. Thirty-two participants out of 40 reported that they preferred to interact with a chatbot with these characteristics. Noticeably, the chatbot's attentiveness to the interactional goal may not be evident to the user if the chatbot passively waits for the user to control the interaction. Thus, \\textit{conscientiousness} relates to proactive ability, as discussed in Section \\ref{sec:proactivity}. \nNevertheless, challenges in designing conscientious chatbots are also evident in the literature:\n\\textbf{[C1] to handle task complexity: }as the complexity of tasks increases, more turns are required to achieve a goal; hence, more mistakes may be made. This argument was supported by both and , where the complexity of the task compromised the experience and satisfaction in using the chatbot. also highlights that complex tasks require more effort to correct eventual mistakes. Therefore, it is an open challenge to design flexible workflows, where the chatbot recovers from failures and keeps the interaction moving productively toward the goal, despite potential misunderstandings . \nRecovering from failure \nis discussed in Section \\ref{sec:damagecontrol}.\n\\textbf{[C2] to harden the conversation:} aiming to ensure the conversational structure--and to hide natural language limitations--chatbots are designed to restrict free-text inputs from the user . However, limiting the choices of interaction may convey a lack of attention to the users' inputs. In , one participant mentioned the feeling of ``\\textit{going through a form or a fixed menu.}'' According to , participants consider the chatbot's understanding of free-text input as a criterion to determine whether it can be considered a chatbot, since chatbots are supposed to chat. In the ethnographic data collection study, reported that eight out of ten participants described the interaction using pre-set responses as too restrictive, although they fulfilled the purpose of nudging participants to report their activities. Thus, the challenge lies in how to leverage the benefits of suggesting predefined inputs without limiting conversational capabilities.\n\\textbf{[C3] to keep the user aware of the chatbot's context:} a chatbot should provide a way to inform the user of the current context, especially for complex tasks. According to , context can be inferred from explicit user input, or assumed based on data from previous interactions. In both cases, the user and chatbot should be on the same page about the chatbot's contextual state , providing the users the opportunity to clarify possible misunderstandings . highlighted that participants reported negative experience when finding ``mismatch[es] between [a] chatbot's real context and their assumptions of the chatbot context.'' We identified three strategies used to provide understanding to a chatbot:\n\\textbf{[S1] conversation workflow: }designing a conversational blueprint helps to conduct the conversation strictly and productively to the goal . However, argue that the workflow should be flexible enough to handle both multi-turn and one-turn question-answer interactions; it also should be unambiguous, such that users can efficiently achieve their goals. In addition, discuss that the workflow should make it easy to fix mistakes; otherwise, the users need to restart the workflow, which leads to frustration. In , the conversation workflow included communicative strategies to detect a learner's breakdowns and pitfalls. In that study, when the student does not respond, the chatbot uses a comprehension-check question to detect whether the student understood what was said. Then, it reacts to the user's input by adopting one of the proposed communication strategies (e.g., asking for repetition or simplifying the previous sentence). The conversation workflow could also allow the chatbot to be proactive\n. For example, participants in suggested that proactive follow-up questions would anticipate the resolution of the context, reducing the effort required from the user to achieve the goal.\n\\textbf{[S2] visual elements: }user-interface resources--such as quick replies, cards, and carousels\n--are used to structure the conversation and reduce issues regarding understanding . Using these resources, the chatbot shows the next possible utterances and conveys the conversational workflow in a step-by-step manner . Visual elements are also used to show the user what the chatbot can (or cannot) do. This is another conversational characteristic, which will be discussed in the \nSection \\ref{sec:communicability}.\n\\textbf{[S3] context window:} to keep the user aware of the chatbot's current context, developed a chatbot for shopping that shows a context window on the side of the conversation. In this window, the user can click on specific attributes and change them to fix inconsistencies. A survey showed that the chatbot outperformed a default chatbot (without the context window) for mental demand and effort constructs. However, when the chatbots are built into third-party apps (e.g., Facebook Messenger), an extra window may not be possible.\n\\textbf{[S4] confirmation messages: }a conversation workflow may include confirmation messages to convey the chatbots' context to the user . In , when trying to block a stolen credit card, a confirmation message is used to verify the given personal data. In , confirmation messages are used as a communicative strategy to check whether the system's understanding about a particular utterance matches what the user actually meant. Balancing the number of confirmation messages (see ) and the right moment to introduce them into the conversation flow is still under-investigated.\nThe surveyed literature supports that failing to demonstrate understanding about the users' individual utterances, the conversational context, and the interactional goals results in frustration and loss of credibility. However, most of the results are exploratory; a lack of studies investigate the extent to which the provided strategies influence users' behaviors and perceptions. In the field of human-human communication, the cooperative principle has been extensively applied to understand conversational partners' expectations during a conversation. Additionally, recent studies have shown that this principle influences chatbots' perceived humanness . Particularly, the maxims of relation (the ability to be relevant) and manner (the ability to be clear and orderly) can contribute to the design of \\textit{conscientious} chatbots. In addition, \\textit{conscientiousness} is by itself a \\textit{personality} trait; the more \\textit{conscientiousness} a chatbot manifests, the more it can be perceived as attentive, organized, and efficient. The relationship between \\textit{conscientiousness} and \\textit{personality} is highlighted in Section \\ref{sec:interrelationship}.","id":"8d7c33ab-b84e-4247-bfcd-ac939fd2d82f","level":"subsubsection","origin_cites_number":17,"parent_id":"9ad3c98e-3389-49e5-a57e-7e261302fc58","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Conversational Intelligence"],["subsubsection","Conscientiousness"]],"subsections":[],"title":"Conscientiousness"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:communicability}\nInteractive software is communicative by its nature, since users achieve their goals \nby exchanging messages with the system . In this context, \\textit{communicability} is defined as the capacity of a software to convey to users its underlying design intent and interactive principles . Providing \\textit{communicability} helps users to interpret the codes used by designers to convey the interactional possibilities embedded in the software , which improves system learnability . In the chatbot context, \\textit{communicability} is, therefore, the capability of a chatbot to convey its features to users . The problema around chatbots' \\textit{communicability} lies in the nature of the interface: instead of buttons, menus, and links, chatbots unveil their capabilities through conversational turns, one sentence at a time , bringing new challenges to the system learnability field. The lack of \\textit{communicability} may lead users to give up on using the chatbot when they cannot understand the available functionalities and how to use them .\nIn this survey, we found six papers that describe \\textit{communicability}, although investigating \\textit{communicability} is the main purpose of only one . Conversational logs revealed \\textit{communicability} needs in three studies ; in the other three studies \\textit{communicability} issues were self-reported by the users in post-intervention interviews and subjective feedback. \\textit{Communicability} was investigated in the customer services domain (two studies), where chatbots guide customers through available functionalities . The remaining task-oriented domains in which \\textit{communicability} was investigated are financial services, news, and human resources. We did not identify studies on open domain chatbots that report on \\textit{communicability}, since in open domain interactions users are free to talk about varying topics and guidance is less a concern.\nThe surveyed literature reports two main benefits of \\textit{communicability} for chatbots:\n\\textbf{[B1] to unveil functionalities: }while interacting with chatbots, users may not know that a desired functionality is available or how to use it . Most participants in 's study mentioned that they did not understand the functionalities of at least one of the chatbots and none of them mentioned searching for the functionalities in other sources (e.g., Google search or the chatbot website) rather than exploring options during the interaction. In a study about playful interactions in a work environment, observed that 22\\% of the participants explicitly asked the chatbot about its capabilities (e.g., \\textit{``what can you do?''}), and 1.8\\% of all the users' messages were ability-check questions. In a study about hotel chatbots, verified that 63\\% of the conversations were initiated by clicking an option displayed in the chatbot welcome message. A semiotic inspection of news-related chatbots evidenced that \\textit{communicability} strategies are effective at providing clues about the chatbot's features and ideas about what to do and how.\n\\textbf{[B2] to manage users' expectations:} observed that when first-time users do not understand chatbots' capabilities and limitations, they have high expectations and, consequently, end up more frustrated when the chatbots fail. Some participants \nblamed themselves for not knowing how to communicate and gave up. \nIn , quantitative results \nevidenced that ability-check questions can be considered signals of users struggling with functional affordances. Users posed ability-check questions after encountering errors as a means of establishing a common ground between the chatbot's capabilities and their own expectations. According to the authors, ability-check questions helped users to understand the system and reduce uncertainty . In , users also demonstrated the importance of understanding chatbots' capabilities in advance. Since the tasks related to financial support, users expected the chatbot to validate the personal data provided and to provide feedback after completing the task (e.g., explaining how long it would take for the credit card to be blocked). Therefore, \\textit{communicability} helps users gain a sense of which type of messages or functionalities a chatbot can handle.\nThe surveyed literature also highlights two challenges of providing \\textit{communicability}:\n\\textbf{[C1] to provide business integration: }communicating chatbots' functionalities should be performed as much as possible within the chat interface . Chatbots often act as an intermediary between users and services. In this case, to overcome technical challenges, chatbots answer users' inputs with links to external sources, where the request will be addressed. First-time users expressed dissatisfaction with this strategy in . Six participants complained that the chatbot redirected them to external websites. According to , business integration is a requirement for designing chatbots, so that the chatbot can solve the users' requests without transferring the interaction to another user interface.\n\\textbf{[C2] to keep visual elements consistent with textual inputs:} in the semiotic engineering evaluation, observed that some chatbots responded differently depending on whether the user accesses a visual element in the user-interface or types the desired functionality in the text-input box, even if both input modes result in the same utterance. This inconsistency produces the user's feeling of misinterpreting the affordances, which has a negative impact on the system's learnability.\nAs an outcome of the semiotic inspection process, present a set of strategies to provide \\textit{communicability}. Some of them are also emphasized in other studies, as follows:\n\\textbf{[S1] to clarify the purpose of the chatbot:} First-time users in highlighted that a clarification about the chatbots' purpose should be placed in the introductory message. found similar inference from the literature on customer services chatbots. The authors argue that providing an opening message with insights into the chatbots' capabilities is a requirement for chatbots design. In addition, a chatbot could give a short tour throughout the main functionalities at the beginning of the first sessions .\n\\textbf{[S2] to advertise the functionality and suggest the next step: }when the chatbot is not able to answer the user, or when it notices that the user is silent, it may suggest available features to stimulate the user to engage \n. In , six participants mentioned that they appreciated when the chatbot suggested responses, for example by saying \\textit{``try a few of these commands: ...''} . shows that chatbots use visual elements, such as cards, carousel, and menus (persistent or not) to show contextualized clues about the next answer, which both fulfills \\textit{communicability} purpose and spares users from having to type.\n\\textbf{[S3] to provide a help functionality: }chatbots should recognize a \\textit{``help''} input from the user, so it can provide instructions on how to proceed . reported that users highlighted this functionality as useful for the reviewed chatbots. Also, results from show that chatbots should be able to answer ability-check questions (e.g., ``\\textit{what can you do?}'' or ``\\textit{can you do [functionality]?}'').\nThe literature states the importance of communicating chatbots' functionality to the success of the interaction. Failing to provide \\textit{communicability} \nleads the users to frustration and they often give up when they do not know how to proceed. The literature on interactive systems has highlighted system learnability as the most fundamental component for usability , and an easy-to-learn system should lead the user to perform well, even during their initial interactions. Thus, researchers in the chatbots domain can leverage the vast literature on system learnability to identify metrics and evaluation methodologies, as well as propose new forms of \\textit{communicability} strategies that reduce the learnability issues in chatbot interactions. \\textit{Communicability} may also be used as a strategy to avoid mistakes (\\textit{damage control}), which will be discussed in Section \\ref{sec:damagecontrol}.\n\\vspace{-1.5mm}\n\\begin{framed} \\small\n\\vspace{-2mm}\nIn summary, the \\textbf{conversational intelligence }category includes characteristics that help a chatbot to perform a proactive, attentive, and informative role in the interaction. Acknowledging that functional aspects play a crucial role in the ability of a chatbot to converse intelligently, the survey focuses on the characteristics that go beyond the functional aspects, looking through the lens of social perceptions. The highlighted benefits relate to how a chatbot manages the conversation to make it productive, interesting, and neat. To achieve that, designers and researchers should care for the timing and relevance of provided information, privacy, interactional flexibility, and consistency.\n\\vspace{-2mm}\n\\end{framed}\n\\vspace{-2mm}","id":"d1a8a8a3-347c-4cde-8b5a-76ed8060b75f","level":"subsubsection","origin_cites_number":10,"parent_id":"9ad3c98e-3389-49e5-a57e-7e261302fc58","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Conversational Intelligence"],["subsubsection","Communicability"]],"subsections":[],"title":"Communicability"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:socialintelligence}\n\\textbf{Social Intelligence} refers to the ability of an individual to produce adequate social behavior for the purpose of achieving desired goals . In the HCI domain, the Media Equation theory posits that people react to computers as social actors. Hence, when developing chatbots, it is necessary to account for the socially acceptable protocols for conversational interactions . Chatbots should be able to respond to social cues during the conversation, accept differences, and manage conflicts as well as be empathic and demonstrate caring , which ultimately increase chatbots' authenticity . In this section, we discuss the social characteristics related to \\textbf{social intelligence}, namely: \\textit{damage control} (12 papers), \\textit{thoroughness} (13 papers), \\textit{manners} (10 papers), \\textit{moral agency} (6 papers),\\textit{ emotional intelligence} (14 papers), and \\textit{personalization} (11 papers). Although the focus of the investigations is diverse, we found more studies where the focus of the investigation relates to a specific social characteristic, particularly \\textit{moral agency} and \\textit{emotional intelligence}, when compared to the \\textbf{conversational intelligence} category. Regarding the adopted research methods, 12 studies applied qualitative methods and 10 focuses on quantitative methods. 16 studies reported both qualitative and quantitative results.","id":"6e94027f-a398-4b65-8b3b-d03e95d73fa6","level":"subsection","origin_cites_number":6,"parent_id":"5950bb1f-b30c-410b-94c0-7eba5ce91d72","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Social Intelligence"]],"subsections":["028808f1-c630-47f4-8ddb-e59065ee5178","f4e8fa92-1e99-453b-a644-c67f9ee2f76c","b7e4439b-1760-43b1-b0dc-0c9d26d8356e","f7519054-562c-44e0-9cb2-d8defc3722e3","e08fd9d3-0a36-4724-8257-1d5f35b4d5dc","69b481db-bee0-4219-a694-6a1c1caa8255"],"title":"Social Intelligence"},{"cite_extract_rate":0.05,"cites":[6262],"content":"\\label{sec:damagecontrol}\n\\textit{Damage control} is the ability of a chatbot to deal with either conflict or failure situations. Although the Media Equation theory argues that humans socially respond to computers as they respond to other people , the literature has shown that interactions with conversational agents are not quite equal to human-human interactions . When talking to a chatbot, humans are more likely to harass , test the agent's capabilities and knowledge , and feel disappointed with mistakes . When a chatbot does not respond appropriately, it may encourage the abusive behavior or disappoint the user , which ultimately leads the conversation to fail . Thus, it is necessary to enrich chatbots with the ability to recover from failures and handle inappropriate talk in a socially acceptable manner .\nIn this survey, we found 12 studies that discuss \\textit{damage control} as a relevant characteristic for chatbots, two of which focus on conflict situations, such as testing and flaming . In the remaining studies , needs for \\textit{damage control} emerged from the analysis of conversational logs and users' feedback. \\textit{Damage control} was mostly investigated for open domain (two studies) and customer service chatbots (three studies). In open domain interactions, users are free to wander among topics and testing or flaming tends to be more frequent . In the customer services context, the chatbot needs to avoid disappointing the user, as frustration may negatively reflect the business that the agent represents . \\textit{Damage control} was also identified in other six domains (see the supplementary materials), such as task management and financial services.\nThe surveyed literature highlights the following benefits of providing \\textit{damage control} in chatbots:\n\\textbf{[B1] to appropriately respond to harassment:} chatbots are more likely to targets of profanity than humans would . When analyzing conversation logs from hotel chatbots, observed that 4\\% of the conversations contained vulgar, indecent, and insulting vocabulary, and 2.8\\% of all statements were abusive. Qualitative evaluation reveals that the longer the conversations last, the more users are encouraged to go beyond the chatbots main functions. In addition, sexual expressions represented 1.8\\% of all statements. A similar number was found in a study with general purpose chatbots . When analyzing a corpus from the Amazon Alexa Prize 2017, the researchers estimated that about 4\\% of the conversations included sexually explicit utterances. used utterances from this corpus to harass a set of state-of-art chatbots and analyze the responses. The results show that chatbots respond to harassment in a variety of ways, including nonsensical, negative, and positive responses. However, the authors highlight that the responses should align with the chatbot's goal to avoid encouraging the behavior or reinforcing stereotypes.\n\\textbf{[B2] to deal with testing:} abusive behavior is often used to test chatbots' social reactions . During the evaluation of a virtual guide to the university campus , a participant answered the chatbot's introductory greeting with ``\\textit{moron,}'' likely hoping to see how the chatbot would answer. argue that handling this type of testing helps the chatbots to establish limits and resolve social positioning. Other forms of testing were highlighted in , including sending random letters, repeated greetings, laughs and acknowledgments, and posing comments and questions about the chatbot's intellectual capabilities. When analyzing conversations with a task management chatbot, observed that casually testing the chatbots' ``intelligence'' is a manifestation of satisfaction seeking. In , first-time users appreciated when the chatbot successfully performed tasks when the user expected the chatbot to fail, which shows that satisfaction is influenced by the ability to provide a clever response when the user tests the chatbot.\n\\textbf{[B3] to deal with lack of knowledge: }chatbots often fail in a conversation due to lack of either linguistic or world knowledge . \\textit{Damage control} enables the chatbot to admit the lack of knowledge or cover up cleverly . When analyzing the log of a task management chatbot, found out that the chatbot failed to answer 10\\% of the exchanged messages. The authors suggest that the chatbot should be designed to handle novel scenarios when the current knowledge is not enough to answer the requests. In some task-oriented chatbots, though, the failure may not be caused by a novel scenario, but by an off-topic utterance. In the educational context, observed that students posted off-topic utterances when they did not know what topics they could talk about, which led the chatbot to fail rather than help the users to understand its knowledge. In task-oriented scenarios, the lack of linguistic knowledge may lead the chatbots to get lost in the conversational workflow , compromising the success of the interaction. demonstrated that dialogue-reference errors (e.g., user's attempt to correct a previous answer or jumping back to an earlier question) are one of the major reasons for failed dialogues and they mostly result from chatbots' misunderstandings.\nThe literature also reveals some challenges to provide \\textit{damage control}:\n\\textbf{[C1] to deal with unfriendly users:} argue that users who want to test and find the system's borders are likely to never have a meaningful conversation with the chatbot no matter how sophisticated it is. Thus, there is a limit to which \\textit{damage control} strategies will be effective to avoid testing and abuse. In , the authors observed human tendencies to dominate, be rude, and infer stupidity, which they call ``\\textit{unfriendly partners.}'' After an intervention where users interacted with a chatbot for decision-making coaching, evaluated participants' self-perceived work and cooperation with the system. The qualitative results show that cooperative users are significantly more likely to give a higher rating for overall evaluation and decision efficiency. The qualitative analysis of the conversation log reveals that a few interactions failed because the users' motivations were curiosity and mischief rather than trying to solve the decision problem.\n\\textbf{[C2] to identify abusive utterances: }several chatbots are trained on ``clean'' data. Because they do not understand profanity or abuse, they may not recognize a statement as harassment, which makes it difficult to adopt answering strategies . shows that data-driven chatbots often provide non-coherent responses to harassment. Sometimes, these responses conveyed the impression of flirtatious or counter-aggression. Providing means to identify an abusive utterance is important to adopting \\textit{damage control} strategies.\n\\textbf{[C3] to fit the response to the context:} argue that humans negotiate conflict and social positioning well before reaching abuse. In human-chatbot interactions, however, predicting users' behavior toward the chatbots in a specific context to develop the appropriate behavior is a challenge. \\textit{Damage control} strategies need to be adapted to both the social situation and the intensity of the conflict. For example, showed that being evasive about sexual statements may convey the impression of flirtatiousness, which would not be an acceptable behavior for a customer assistant or a tutor chatbot. In contrast, adult chatbots are supposed to flirt, so encouraging behaviors are expected in some situations. argue that when the chatbot is not accepted as part of the social group it represents, it is discredited by the user, leading the interaction to fail. In addition, designing chatbots with too strong of reactions may lead to ethical concerns . For , choosing between peaceful or aggressive reactions in conflict situations is optional for socially intelligent individuals. Enriching chatbots with the ability to choose between the options is a challenge.\n\\textit{Damage control} strategies depend on the type of failure and the target benefit, as following:\n\\textbf{[S1] emotional reactions:} suggest that when faced with abuse, a chatbot could be seen to take offense and respond in kind or act hurt. The authors argue that humans might feel inhibited about hurting the pretended feelings of a machine if the machine is willing to hurt a human's feelings too . If escalating the aggressive behavior is not appropriate, the chatbot could withdraw from the conversation to demonstrate that the user's behavior is not acceptable. In , the authors discuss that users appeared to be uncomfortable and annoyed whenever the chatbot pointed out any defect in the user or reacted to aggression, as this behavior conflicted with the user's perceived power relations. This strategy is also applied in , where abusive behavior may lead the chatbot to stop responding until the student changes the topic. categorized responses from state-of-the-art conversational systems in a pool of emotional reactions, both positive and negative. The reactions include humorous responses, chastising and retaliation, and evasive responses, as well as flirtation and play-along utterances. To provide an emotional reaction, \\textit{emotional intelligence} is also required. This category is presented in Section \\ref{sec:emotionalintelligence}.\n\\textbf{[S2] authoritative reactions: }when facing testing or abuse, chatbots can communicate consequences \nor call on the authority of others . In , although the wizard acting as a chatbot was conscientiously working as a campus guide, she answered a bogus caller with \\textit{``This is the University of Melbourne. Sorry, how can I help you?''} The authors suggest that the wizard was calling on the authority of the university to handle the conflict, where being part of a recognized institution places the chatbot in a stronger social group. In , when students recurrently harass the chatbot, the chatbot informs the student that further abuse will be reported to the (human) teacher (although the paper does not clarify whether the problem is, in fact, escalated to a human). and also suggest that chatbots could redirect users' problematic requests to a human attendant in order to avoid conflict situations.\n\\textbf{[S3] to ignore the user's utterance and change the topic:} argue that ignoring abuse and testing is not a good strategy because it could encourage more extreme behaviors. It also positions the chatbot as an inferior individual, which is particularly harmful in scenarios where the chatbot should demonstrate a more prominent or authoritative social role (e.g., a tutor). However, this strategy has been found in some studies to handle lack of knowledge. When iteratively developing a chatbot for an educational context, proposed initiating a new topic in one out of four user's utterances that the chatbot did not understand.\n\\textbf{[S4] \\textit{conscientiousness} and \\textit{communicability}:} successfully implementing \\textit{conscientiousness} and \\textit{communicability} may prevent errors; hence, strategies to provide these characteristics can also be used for \\textit{damage control}. In , when users utter out-of-scope statements, the chatbot could make it clear what topics are appropriate to the situation. For task-oriented scenarios, where the conversation should evolve toward a goal, argue that the chatbot can clarify the purpose of the offered service when facing abusive behavior, bringing the user back to the task. showed that describing chatbot's capabilities after failures in the dialogue was appreciated by first-time users. In situations where the conversational workflow is susceptible to failure, discuss that posting confirmation messages avoids trapping the users in the wrong conversation path. Participants in also suggested back buttons as a strategy to fix mistakes in the workflow. In addition, the exploratory results about the user interface showed that having visual elements such as quick replies prevents errors, since they keep the users aware of what to ask and the chatbot is more likely to know how to respond .\n\\textbf{[S5] to predict users' satisfaction:} chatbots should perceive both explicit and implicit feedback about users' (dis)satisfaction .\nTo address this challenge, invited participants to send a ``\\textit{\\#fail}'' statement to express dissatisfaction. The results show that 42.4\\% of the users did it at least once, and the number of complaints and flaming for the proposed chatbot was significantly lower than the baseline. However, the amount of implicit feedback was also significant, which advocates for predicting user's satisfaction from the conversation. The most powerful conversational act that predicted user satisfaction in that study was the agent ability-check questions, see discussion in \\textit{Communicability} section) and the explicit feedback \\textit{\\#fail}, although closings and off-topic requests were also significant in predicting frustration. Although these results are promising, more investigation is needed to identify other potential predictors of users' satisfaction in real-time, in order to provide appropriate reactions.\n\\textit{Damage control} strategies have different levels of severity. Deciding what strategy is adequate to the intensity of the conflict is crucial . The strategies can escalate in severity if the conflict is not solved. For example, uses a sequence of clarification, suggesting a new topic, and asking a question about the new topic. In case of abuse, the chatbot refers to an authority after two attempts at changing the topic.\nAccording to , humans also fail in conversations; they misunderstand what their partner says and do not know things that are assumed common knowledge by others. Hence, it is unlikely that chatbots interactions will evolve to be conflict-free. That said, \\textit{damage control} intends to avoid escalating the conflicts and manifesting an unexpected behavior. In this sense, politeness can be used as a strategy to minimize the effect of lack of knowledge (see Section \\ref{sec:manners}), managing the conversation despite the possible mistakes. Regarding interpersonal conflicts, the strategies are in line with the theory on human-human communication, which includes non-negotiation, emotional appeal, personal rejection, and emphatic understanding . Further research on \\textit{damage control} can evaluate the adoption of human-human strategies in human-chatbot communication.","id":"028808f1-c630-47f4-8ddb-e59065ee5178","level":"subsubsection","origin_cites_number":20,"parent_id":"6e94027f-a398-4b65-8b3b-d03e95d73fa6","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Social Intelligence"],["subsubsection","Damage control"]],"subsections":[],"title":"Damage control"},{"cite_extract_rate":0.1,"cites":[6264,6265],"content":"\\label{sec:thoroughness}\n\\textit{Thoroughness} is the ability of a chatbot to be precise regarding how it uses language . In traditional user interfaces, user communication takes place using visual affordances, such as buttons, menus, or links. In a conversational interface, language is the main tool to achieve the communicative goal. Thus, chatbots should coherently use language that portrays the expected style . When the chatbot uses inconsistent language, or unexpected patterns of language (e.g., excessive formality), the conversation may sound strange to the user, leading to frustration. \nWe found 13 papers that report the importance of \\textit{thoroughness} in chatbot design, three of which investigate how patterns of language influence users' perceptions and behavior toward the chatbots . and suggest design principles that include concerns about language choices. Log of conversations revealed issues regarding \\textit{thoroughness} in two studies . In the remaining papers, \\textit{thoroughness} emerged from interviews and users' subjective feedback . \\textit{Thoroughness} is mainly reported for open domain (five studies) and customer service chatbots (two studies), where the interactions are expected to be natural and credible to succeed . \\textit{Thoroughness} was also reported in another six domains of studies, such as financial services and education (see the supplementary materials).\nWe found two benefits of providing \\textit{thoroughness}:\n\\textbf{[B1] to increase human-likeness:} chatbot utterances are often pre-recorded by the chatbot designer . On the one hand, this approach produces high quality utterances; on the other hand, it reduces flexibility, since the chatbot is not able to adapt the tone of the conversation based on individual users and conversational context. When analyzing interactions with a customer representative chatbot, observed that the chatbot proposed synonyms to keywords, and the repetition of this vocabulary led the users to imitate it. observed a similar tendency to matching language style. The authors compared human-human conversations with human-chatbots conversations regarding language use. They found that people use, indeed, fewer words per message and a more limited vocabulary with chatbots. However, a deeper investigation revealed that the human interlocutors were actually matching the patterns of language use with the chatbot, who sent fewer words per message. When interacting with a chatbot that uses many emojis and letter reduplication , participants reported a draining experience, since the chatbot's energy was too high to match. These outcomes show that adapting the language to the interlocutor is a common behavior for humans, and so chatbots would benefit from manifesting it. In addition to the interlocutor, chatbots should adapt their language use to the context in which they are implemented and adopt appropriate linguistic register . In the customer services domain, state that chatbots are expected to fulfill the role of a human; hence, they should produce language that corresponds to the represented service provider. In the financial scenario , some participants complained about the use of emojis in a situation of urgency (blocking a stolen credit card).\n\\textbf{[B2] to increase believability: }because people associate social qualities with machines , chatbots are deemed sub-standard when users see them ``\\textit{acting as a machine}'' . When analyzing chatbots' naturalness, found that the formal grammatical and syntactical abilities of a chatbot are the biggest discriminators between good and poor chatbots (the other factors being \\textit{conscientiousness}, \\textit{manners}, and \\textit{proactivity}). The authors highlight that chatbots should use consistent grammar and spelling. discusses how, even with English as Second Language (ESL) learners, basic grammar errors, such as pronoun confusion, diminish the value of the chatbot. In addition, states that believable chatbots also need to display unique characters through linguistic choices. In this sense, demonstrated that \\textit{personality} can be expressed by language patterns. The authors proposed a computational framework to produce utterances to manifest a target \\textit{personality}. The utterances were rated by experts in \\textit{personality} evaluation and statistically compared against utterances produced by humans who manifest the target \\textit{personality}. The outcomes show that a single utterance can manifest a believable \\textit{personality} when using the appropriate linguistic form. Participants in described some interactions as ``\\textit{robotic}'' if the chatbot repeated the keywords in the answers, reducing the interaction's naturalness. Similarly, in , participants complained about the ``\\textit{inflexibility}'' of the pre-defined, handcrafted chatbot's responses and expressed the desire for it to talk ``\\textit{more as a person.}''\nRegarding the challenges, the surveyed literature shows the following:\n\\textbf{[C1] to decide how much to say: } in , some participants described the chatbot's utterances as not having enough detail, or being too generic; however, most of them appreciated finding answers in a sentence rather than in a paragraph. Similarly, argue that simple questions should not be too detailed, while important transactions require more information. In three studies , participants complained about information overload and inefficiency caused by big blocks of texts. Balancing the granularity of information with the sentence length is a challenge to overcome.\n\\textbf{[C2] to be consistent:} chatbots should not combine different language styles. For example, in , most users found it strange that emojis were combined with a certain level of formal contact. \nWhen analyzing the critical incidents about an open-domain interaction, found that participants criticized chatbots when they used more formal language or unusual vocabulary, since general-purpose chatbots focus on casual interactions.\nDespite the highlighted benefits, \nwe did not find strategies to provide \\textit{thoroughness}. proposed a rule-based architecture where the language choices consider the agent's \\textit{personality}, emotional state, and beliefs about the social relationship among the interlocutors. However, they did not provide evidence of whether the proposed models produced the expected outcome. Although the literature in computational linguistics has proposed algorithms and statistical models to manipulate language style and matching (see e.g., ), to the best of our knowledge, these strategies have not been evaluated in the context of chatbots' social interactions.\nThis section shows that linguistic choices influence users' perceptions of chatbots. The computer-mediated communication (CMC) field has a vast literature that shows language variation according to the media and its effect on social perceptions (see e.g. ). Additionally, the cooperative principle , particularly the maxim of quantity (the ability to give the appropriate amount of information), provides theoretical basis for the challenge of deciding how much to talk. Regarding adaptation and believability, researchers in sociolinguistic fields have shown that language choices are influenced by personal style, dialect, genre, and register. For chatbots, the results presented in are promising, demonstrating that automatically generated language can manifest recognizable traits. Thus, further research on chatbot's \\textit{thoroughness} could leverage CMC and linguistics theories to provide strategies that lead language to accomplish its purpose for a particular interactional context.","id":"f4e8fa92-1e99-453b-a644-c67f9ee2f76c","level":"subsubsection","origin_cites_number":20,"parent_id":"6e94027f-a398-4b65-8b3b-d03e95d73fa6","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Social Intelligence"],["subsubsection","Thoroughness"]],"subsections":[],"title":"Thoroughness"},{"cite_extract_rate":0.105263157894736,"cites":[6266,6262],"content":"\\label{sec:manners}\n\\textit{Manners} refer to the ability of a chatbot to manifest polite behavior and conversational habits . Although individuals with different personalities and from different cultures may have different notions of what is considered polite (see e.g., ), politeness can be more generally applied as rapport management , where interlocutors strive to control the harmony between people in discourse. A chatbot can manifest \\textit{manners} by adopting speech acts such as greetings, apologies, and closings ; minimizing impositions , and making interactions more personal . \\textit{Manners} potentially reduces the feeling of annoyance and frustration that may lead the interaction to fail .\nWe identified ten studies that report \\textit{manners}, one of which directly investigates this characteristic . In some studies , \\textit{manners} were observed in the analysis of conversational logs, where participants talked to the chatbot in polite, human-like ways. Users' feedback and interviews revealed users' expectations regarding chatbots politeness and personal behavior . We identified studies reporting \\textit{manners} in nine different domains, with only open domain appearing twice. The list includes education, information search, and task management, among others. See the supplementary materials for the complete list of domains of studies that report \\textit{manners} as a social characteristic for chatbots.\nThe main benefit of providing \\textit{manners} is \\textbf{[B1] to increase human-likeness}. \\textit{Manners} is highlighted in the literature as a way to generate a more natural, convincing interaction in chatbot conversations . In an in-the-wild data collection, observed that 93\\% of the participants used polite words (e.g., ``\\textit{thanks}'' or ``\\textit{please}'') with a task management chatbot at least once, and 20\\% always spoke politely to the chatbot. Unfortunately, the chatbot evaluated in that study was not prepared to handle these protocols and ultimately failed to understand. When identifying incidents from their own conversational logs with a chatbot , several participants identified greetings as a human-seeming characteristic. The users also found convincing when the chatbot appropriately reacts to social cues statements, such as \\textit{``how are you?''}-types of utterances. Using this result, later suggested that greetings, apologies, social niceties, and introductions are significant constructs to measure chatbots' naturalness. In , the chatbot used exclamation marks at some point and frequently offered sentences available on the website, in a vaguely human-like matter. In the feedback, participants described the chatbot as rude, impolite, and cheeky. \nThe surveyed literature highlights two challenges to convey \\textit{manners}:\n\\textbf{[C1] to deal with face-threatening acts:} Face-Threatening Acts (FTA) are speech acts that threaten, either positive or negatively, the ``face'' of an interlocutor . Politeness strategies in human-human interactions are adopted to counteract the threat when an FTA needs to be performed . In , the authors discuss that the wizard performing the role of the chatbot used several politeness strategies to counteract face threats. For instance, when she did not recognize a destination, instead of providing a list of possible destinations, she stimulated the user to keep talking until they volunteered the information. In chatbot design, by contrast, providing a list of options to choose is a common strategy. For example, in , the chatbot was designed to present the user with a list of pending tasks when it did not know what task the user was reporting as completed, although the authors acknowledged that it resulted in an unnatural interaction. Although adopting politeness strategies is natural for humans and people usually do not consciously think about them, implementing them for chatbots is challenging due to the complexity of identifying face-threatening acts. For example, in the decision-making coach scenario, observed that users tend to utter straightforward and direct agreements while most of the disagreements contained modifiers that weakened their disagreement. The adoption of politeness strategies to deal with face-threatening acts is still under-investigated in the chatbot literature.\n\\textbf{[C2] to end a conversation gracefully:} discuss that first-time users expected human-like conversational etiquette from the chatbots, specifically introductory phrases and concluding phrases. Although several chatbots perform well in the introduction, the concluding phrases are less explored. Most of the participants reported being annoyed with chatbots that do not end a conversation . also highlight that chatbots need to know when the conversation ends. In that scenario, the chatbot could recognize a closing statement (the user explicitly says \\textit{``thank you''} or \\textit{``bye''}); however, it would not end the conversation otherwise. Users that stated a decision, but kept receiving more information from the chatbot, reported feeling confused and undecided afterward. Thus, recognizing the right moment to end the conversation is a challenge to overcome.\nThe strategies highlighted in the surveyed literature for providing \\textit{manners} are the following:\n\\textbf{[S1] to engage in small talk:} and point out that even task-oriented chatbots engage in small talk. When categorizing the utterances from the conversational log, the authors found a significant number of messages about the agent status (e.g., \\textit{``what are you doing?}''), opening and closing sentences as well as acknowledgment statements (\\textit{``ok,''} \\textit{``got it''}). also observed that first-time users included small talk in the introductory phrases. According to , these are common behaviors in human-human chat interface, and chatbots would likely benefit from anticipating these habitual behaviors and reproducing them. However, particularly for task-oriented chatbots, it is important to control the small talk to avoid off-topic conversations and harassment, as discussed in Sections \\ref{sec:damagecontrol} and \\ref{sec:conscientiousness}.\n\\textbf{[S2] to adhere to turn-taking protocols:} suggest that chatbots should adopt turn-taking protocols to know when to talk. Participants who received frequent follow-up questions from the task management chatbot about their pending tasks perceived the chatbot as invasive. Literature in chatbot development proposes techniques to improve chatbots' turn-taking capabilities (see e.g., ), which can be explored as a means of improving chatbots' perceived \\textit{manners}.\nAlthough the literature emphasizes that \\textit{manners} are important to approximate chatbot interactions to human conversational protocols, this social characteristic is under-investigated in the literature. Conversational acts such as greetings and apologies are often adopted (e.g., ), but there is a lack of studies on the rationality around the strategies with politeness models used in human-human social interactions . In addition, the literature points to needs for personal conversations (e.g., addressing the users by name), but we did not find studies that focus on this type of strategy. CMC is by itself more impersonal than face-to-face conversation ; even so, current online communication media has been successfully used to initiate, develop, and maintain interpersonal relationships . Researchers can learn from human behaviors in CMC and adopt similar strategies to produce more personal conversations.","id":"b7e4439b-1760-43b1-b0dc-0c9d26d8356e","level":"subsubsection","origin_cites_number":19,"parent_id":"6e94027f-a398-4b65-8b3b-d03e95d73fa6","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Social Intelligence"],["subsubsection","Manners"]],"subsections":[],"title":"Manners"},{"cite_extract_rate":0.066666666666666,"cites":[7476],"content":"\\label{sec:moralagency}\nMachine \\textit{moral agency} refers to the ability of a technology to act based on social notions of right and wrong . The lack of this ability may lead to cases such as Tay, Microsoft's Twitter chatbot that became racist, sexist, and harassing in a few hours . The case raised concerns on what makes an artificial agent (im)moral. Whether machines can be considered (moral) agents is widely discussed in the literature (see e.g., ). In this survey, the goal is not to argue about criteria to define a chatbot as moral, but to discuss the benefits of manifesting a perceived agency and the implications of disregarding chatbots' moral behavior. Hence, for the purpose of this survey, \\textit{moral agency} is a manifested behavior that may be inferred by a human as morality and agency .\nWe found six papers that address \\textit{moral agency}. developed and validated a metric for perceived \\textit{moral agency} in conversational interfaces, including chatbots. In four studies, the authors investigated the ability of chatbots to handle conversations where the persistence of gender and racial stereotypes may occur. In , \\textit{moral agency} is discussed as a secondary result, where the authors discuss the impact of biased responses on emotional connection. \\textit{Moral agency} was observed in only two domains of studies: open domain (four studies) and race-talk (two studies), which shows that this characteristic is primarily relevant when the conversational topic may raise moral concerns, which ultimately requires ethical behavior from the conversational partners.\nThe two main reported benefits of manifesting perceived \\textit{moral agency} are the following:\n\\textbf{[B1] to avoid stereotyping:} chatbots are often designed with anthropomorphized characteristics (see Section \\ref{sec:personification}), including gender, age, and ethnicity identities. Although the chatbot's personification is more evident in embodied conversational agents, text-based chatbots may also be assessed by their social representation, which risks building or reinforcing stereotypes . and argue that chatbots are often developed using language registers and cultural references of the dominant culture. In addition, a static image (or avatar) representing the agent may convey social grouping . When the chatbot is positioned in a minority \\textit{identity} group, it exposes the image of that group to judgment and flaming, which is frequent in chatbot interactions . For example, discusses the controversies caused by a chatbot designed to answer questions about Caribbean Aboriginal culture: its representation as a Caribbean Amerindian individual created an unintended context for stereotyping, where users projected the chatbot's behavior as a standard for people from the represented population. Another example is the differences in sexual discourse between male- and female-presenting chatbots. found that female-presenting chatbots are the object of implicit and explicit sexual attention and swear words more often than male-presenting chatbots. show that sex talks with the male chatbot were rarely coercive or violent; his sexual preference was often questioned, though, and he was frequently propositioned by reported male users. In contrast, the female character received violent sexual statements, and was threatened with rape five times in the analyzed corpora. In , when the avatars were presented as black adults, references to race often deteriorated into racist attacks. Manifesting moral agency may, thus, prevent obnoxious user interactions. In addition, moral agency may prevent the chatbot itself from being biased or disrespectful to humans. argue that the lack of context about the world does not redeem the chatbot from the necessity of being respectful with all the social groups.\n\\textbf{[B2] to enrich interpersonal relationships:} in a study on how interlocutors perceive conversational agents' \\textit{moral agency}, hypothesized that perceived morality may influence a range of motivations, dynamics, and effects of human-machine interactions. Based on this claim, the authors evaluated whether goodwill, trustworthiness, willingness to engage, and relational certainty in future interactions are constructs to measure perceived \\textit{moral agency}. Statistical results showed that all the constructs correlate with morality, which suggests that manifesting \\textit{moral agency} can enrich interpersonal relationships with chatbots. Similarly, suggest that to produce interpersonal responses, chatbots should be aware of inappropriate information and avoid generating biased responses.\nHowever, the surveyed literature also reveals challenges of manifesting \\textit{moral agency}:\n\\textbf{[C1] to avoid alienation:} in order to prevent a chatbot from reproducing hate speech or abusive talk, most chatbots are built over ``clean'' data, where specific words are removed from their dictionary . These chatbots have no knowledge of those words and their meaning. Although this strategy is useful to prevent unwanted behavior, it does not manifest agency, but rather alienates the chatbot of the topic. show that the lack of understanding about sex-talk does not prevent the studied chatbot from harsh verbal abuse, or even from being perceived as encouraging such abuse. From , one can notice that the absence of racist specific words did not prevent the chatbot Zo from uttering discriminatory exchanges. As a consequence, manifesting \\textit{moral agency} requires a broader understanding of the world, rather than alienation, which is an open challenge.\n\\textbf{[C2] to build unbiased algorithms and training data:} as extensively discussed in , machine learning algorithms and corpus-based language generation are biased toward the available training datasets. Hence, \\textit{moral agency} relies on data that is biased in its nature, producing unsatisfactory results from an ethical perspective. In , the authors propose a framework for developing social chatbots. The authors highlight that the core chat module should follow ethical design to generate unbiased, non-discriminative responses, but they do not discuss specific strategies for that. Building unbiased training datasets and learning algorithms that connect the outputs with individual, real-world experiences, therefore, are challenges to overcome.\nDespite the relevance of \\textit{moral agency} to the development of socially intelligent chatbots, we did not find strategies to address the issues. advocate for developing diversity-conscious databases and learning algorithms that account for ethical concerns; however, the paper focuses on outlining the main research branches and calls on the community of designers to adopt new strategies. As discussed in this section, research on perceived \\textit{moral agency} is still necessary in order to develop chatbots whose social behavior is inclusive and respectful. In the field of embodied agents, mind perception theory has been investigated as a means to improve interactions through agency and emotion . Future investigations in chatbots interactions could leverage this theory to understand when and the extent to which perceived moral agency improves communication with chatbots.","id":"f7519054-562c-44e0-9cb2-d8defc3722e3","level":"subsubsection","origin_cites_number":15,"parent_id":"6e94027f-a398-4b65-8b3b-d03e95d73fa6","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Social Intelligence"],["subsubsection","Moral agency"]],"subsections":[],"title":"Moral agency"},{"cite_extract_rate":0.11764705882352901,"cites":[7476,2047],"content":"\\label{sec:emotionalintelligence}\n\\textit{Emotional intelligence} is a subset of social intelligence that allows an individual to appraise and express feelings, regulate affective reactions, and harness emotions to solve a problem . Although chatbots do not have genuine emotions , there are considerable discussions about the role of manifesting emotional cues in chatbots . An emotionally intelligent chatbot can recognize and influence users' feelings and demonstrate respect, empathy, and understanding, improving the human-chatbot relationship .\nWe identified 14 studies that report \\textit{emotional intelligence}. Unlikely the previously discussed categories, most studies on \\textit{emotional intelligence} focused on understanding the effects of chatbots' empathy and emotional self-disclosure . Only three papers highlighted \\textit{emotional intelligence} as an exploratory outcome , where needs for \\textit{emotional intelligence} emerged from participants' subjective feedback and post-intervention surveys. \\textit{Emotional intelligence} is mainly investigated in domains where topics may involve the disclosure of feelings (e.g., in open domain interactions ) and expressions of empathy and understanding are appropriate (e.g., health care, gaming, education). See the supplementary materials for the domains of study investigating \\textit{emotional intelligence}.\nThe main reported benefits of developing emotionally intelligent chatbots are the following:\n\\textbf{[B1] to enrich interpersonal relationships:} the perception that the chatbot understands one's feelings may create a sense of belonging and acceptance . propose that chatbots for second language studies should use congratulatory, encouraging, sympathetic, and reassuring utterances to create a friendly atmosphere to the learner. The authors statistically demonstrated that affective backchannel, combined with communicative strategies (see Section \\ref{sec:conscientiousness}) significantly increased learners' confidence and desire to communicate, while reducing anxiety. In another educational study, evaluated the impact of chatbot's affective moves on friendliness and achieving social belonging. Qualitative results show that affective moves significantly improve the perception of amicability, and marginally increased social belonging. According to , when a chatbot's emotional reaction triggers a social response from the user, the chatbot has achieved group membership and users' sympathy. proposed a chatbot that uses empathic and self-oriented emotional expressions to keep users engaged in quiz-style dialogue. The survey results revealed that empathic expressions significantly improved user satisfaction. In addition, the empathic expressions also improved the user ratings of the peer agent regarding intimacy, compassion, amiability, and encouragement. Although did not find effect of chatbot's self-disclosure on emotional connection, found that self-disclosure and reciprocity significantly improved trust and interactional enjoyment. In , seven participants reported that the best thing about their experience with the therapist chatbot was perceived empathy. Five participants highlighted that the chatbot demonstrated attention to their users' feelings. In addition, the users referred the chatbot as \\textit{``he,'' ``a friend,''} \\textit{``a fun little dude,''} which demonstrates that empathy emerged from the personification of the chatbot. In another mental health care study, found that humans are twice as likely to mirror negative sentiment from a chatbot than from a human, which is a relevant implication for therapeutic interactions. In , participants reported that some content is embarrassing to ask another human, thus, talking to a chatbot would be easier due to the lack of judgement. measured users' experience in conversations with a chatbot compared to a human partner as well as the amount of intimacy disclosure and cognitive reappraisal. Participants in the chatbots condition experienced as many emotional, relational, and psychological benefits as participants who disclosed to a human partner. \n\\textbf{[B2] to increase engagement:} argue that longer conversations (10+ turns) are needed to fulfill the needs of affection and belonging. Therefore, the authors defined conversation-turns per session as a success metric for chatbots, where usefulness and emotional understanding are combined. In , empathic utterances for the quiz-style interaction significantly increase the number of users' messages per hint for both answers and non-answer utterances (such as feedback about the success/failure). This result shows that empathic utterances encouraged the users to engage and utter non-answers statements. compared the possibility of emotional connection between a classical chatbot and a pretend chatbot, simulated in a WoZ experiment. Quantitative results showed that the WoZ condition was more engaging, since it resulted in conversations that lasted longer, with a higher number of turns. The analysis of the conversational logs revealed the positive effect of the chatbot manifesting social cues and empathic signs as well as touching on personal topics.\n\\textbf{[B3] to increase believability:} argue that adapting chatbots' language to their current emotional state, along with their \\textit{personality} and social role awareness, results in more believable interactions. The authors propose that conversation acts should reflect the pretended emotional status of the agent; the extent to which the acts impact on emotion depends however on the agent's \\textit{personality} (e.g., its temperament or tolerance). \\textit{Personality} is an anthropomorphic characteristic and is discussed in Section \\ref{sec:personality}.\nAlthough \\textit{emotional intelligence} is the goal of several studies, \\textbf{[C1] regulating affective reactions} is still a challenge. The chatbot presented in was designed to mimic the patterns of affective moves in human-human interactions. Nevertheless, the chatbot has shown an only marginally significant increase in social belonging, when compared to the same interaction with a human partner. Conversational logs revealed that the human tutor performed a significantly higher number of affective moves in that context. In , the chatbot was designed to present emotive-like cues, such as exclamation marks, and interjections. The participants negatively rated the degree of emotion in the chatbot's responses. In , the energetic chatbot was reported as having an enthusiasm too high to match. In contrast, the chatbot was described as an \\textit{``emotional buddy''} was reported as being \\textit{``overly caring.'' } state that chatbot's empathic utterances may be seen as pre-programmed and inauthentic. Although their results revealed that the partners' identity (chatbot vs. person) had no effect in the perceived relational and emotional experience, the chatbot condition was a WoZ setup. The wizards were blind to whether users thought they were talking to a chatbot or a person, which reveals that identity does not matter if the challenge of regulating emotions is overcome.\nThe chatbots literature also report some strategies to manifest \\textit{emotional intelligence}:\n\\textbf{[S1] using social-emotional utterances:} affective utterances toward the user are a common strategy to demonstrate \\textit{emotional intelligence}. , , and suggest that affective utterances improve the interpersonal relationship with a tutor chatbot. In , the authors propose affective backchannel utterances (congratulatory, encouraging, sympathetic, and reassuring) to motivate the user to communicate in a second language. The tutor chatbot proposed in uses solidarity, tension release, and agreement utterances to promote its social belonging and acceptance in group chats. propose empathic utterances to express opinion about the difficulty or ease of a quiz, and feedback on success and failure.\n\\textbf{[S2] to manifest \\textit{conscientiousness}:} demonstrating \\textit{conscientiousness} may affect the emotional connection between humans and chatbots. In , participants reported the rise of affection when the chatbot remembered something they had said before, even if it was just the user's name. Keeping track of the conversation was reported as an empathic behavior and resulted in mutual affection. argue that a chatbot needs to combine usefulness with emotion by asking questions that help to clarify the users' intentions. They provide an example where a user asks the time, and the chatbot answer \\textit{``Cannot sleep?''} as an attempt to guide the conversation to a more engaging direction. Adopting this strategy requires the chatbot to handle message understanding, emotion and sentiment tracking, session context modeling, and user profiling .\n\\textbf{[S3] reciprocity and self-disclosure:} hypothesized that a high level of self-disclosure and reciprocity in communication with chatbots would increase trust, intimacy, and enjoyment, ultimately improving user satisfaction and intention to use. They performed a WoZ, where the assumed chatbot was designed to recommend movies. Results demonstrated that reciprocity and self-disclosure are strong predictors of rapport and user satisfaction. In contrast, did not find any effect of self-oriented emotional expressions in the users' satisfaction or engagement. More research is needed to understand the extent to which this strategy produces positive impact on the interaction.\nThe literature shows that \\textit{emotional intelligence} is widely investigated, with particular interest from education and mental health care domains. Using emotional utterances in a personalized, context relevant way is still a challenge. Researchers in chatbots \\textit{emotional intelligence} can learn from \\textit{emotional intelligence} theory to adapt the chatbots utterances to match the emotions expressed in the dynamic context. Adaption to the dynamic context also improves the sense of personalized interactions, which is discussed in the next section.","id":"e08fd9d3-0a36-4724-8257-1d5f35b4d5dc","level":"subsubsection","origin_cites_number":17,"parent_id":"6e94027f-a398-4b65-8b3b-d03e95d73fa6","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Social Intelligence"],["subsubsection","Emotional Intelligence"]],"subsections":[],"title":"Emotional Intelligence"},{"cite_extract_rate":0.1875,"cites":[6262,650,7476],"content":"\\label{sec:personalization}\n\\textit{Personalization} refers to the ability of a technology to adapt its functionality, interface, information access, and content to increase its personal relevance to an individual or a category of individuals . In the chatbot domain, \\textit{personalization} may increase the agents' social intelligence, since it allows a chatbot to be aware of situational context and dynamically adapt its features to better suit individual needs . Grounded in robots and artificial agents' literature, argue that \\textit{personalization} can improve rapport and cooperation, ultimately increasing engagement with chatbots. Although some studies (see e.g., ) also relate \\textit{personalization} to the attribution of personal qualities such as \\textit{personality}, we discuss personal qualities in the \\textbf{Personification} category. In this section, we focus on the ability to adapt the interface, content, and behavior to the users' preferences, needs, and situational context.\nWe found 11 studies that report \\textit{personalization}. Three studies pose \\textit{personalization} as a research goal . In most of the studies, though, \\textit{personalization} was observed in exploratory findings. In six studies, \\textit{personalization} emerged from the analysis of interviews and participants' self-reported feedback . In two studies , needs for \\textit{personalization} emerged from the conversational logs. \\textit{Personalization} was investigated in seven different domains, including open domain and task management (two studies each). In open domain interactions, personalization is derived from remembering information from previous interactions, such as personal preferences and users' details . In task-oriented contexts, such as task management, personalization aims to increase the relevance of services to particular users . See the supplementary materials for details.\nThe surveyed literature highlighted three benefits of providing personalized interactions:\n\\textbf{[B1] to enrich interpersonal relationships:} state that personalizing the amount of personal information a chatbot can access and store is required to establish a relation of trust and reciprocity in workplace environments. In , interviews with 12 participants generated a total of 59 statements about how learning from experience promotes chatbots' authenticity. argue that chatbots whose focus is engagement need to personalize the generation of responses for different users' backgrounds, personal interests, and needs in order to serve their needs for communication, affection, and social belonging. In , participants expressed the desire for the chatbot to provide different answers to different users. Although has found no significant effect of \\textit{personalization} on the user experience with the financial assistant chatbot, the study applies \\textit{personalization} as the ability to provide empathic responses according to the users' issues, where \\textit{emotional intelligence} plays a role. Interpersonal relationship can also be enriched by adapting the chatbots' language to match the user's context, energy, and formality; the ability to appropriately use language is discussed in Section \\ref{sec:thoroughness}.\n\\textbf{[B2] to provide unique services:} providing \\textit{personalization} increases the value of provided information . In the ethnography data collection study , eight participants reported dissatisfaction with the chatbot's generic guidance to specific places. Participants self-reported that the chatbot should use their current location to direct them to places more conveniently located, and ask for participants interests and preferences to direct them to areas that meet their needs. When exploring how teammates used a task-assignment chatbot, found that the use of the chatbot varied depending on the participants' levels of hierarchy. Similarly, qualitative analysis of perceived interruption in a workplace chat suggests that interruption is likely associated with users' general aversion to unsolicited messages at work. Hence, the authors argue that chatbot's messages should be personalized to the user's general preference. also found that users with low social-agent orientation emphasize the utilitarian value of the system, while users with high social-agent orientation see the system as a humanized assistant. This outcome support to the need to personalize the interaction to individual user's mental models. In , participants reported preference for a chatbot that remembers their details, likes and dislikes, and preferences, and voluntarily uses the information to make recommendations. In , two participants also expected chatbots to retain context from previous interactions to improve recommendations.\n\\textbf{[B3] to reduce interactional breakdowns:} in HCI, \\textit{personalization} is used to customize the interface toward user familiarity . When evaluating\nvisual elements (such as quick replies) compared to typing the responses, observed that users that start the interaction by clicking an option are more likely to continue the conversation if the next exchange also has visual elements as optional affordances. In contrast, users who typed are more likely to abandon the conversation when they face options to click. Thus, chatbots should adapt their interface to users' preferred input methods. In , one participant suggested that the choice of text color and font size should be customizable. also observed that participants faced difficulties with small letters, and concluded that adapting the interface to provide accessibility also needs to be considered.\nAccording to the surveyed literature, the main challenge regarding \\textit{personalization} is \\textbf{[C1] privacy}. To enrich the efficiency and productivity of the interaction, a chatbot needs to have memory of previous interactions as well as learn user's preferences and disclosed personal information . However, as and suggest, collecting personal data may lead to privacy concerns. Thus, chatbots should showcase transparent purpose and ethical standards . also suggest that there should be a way to inform a chatbot that something in the conversation is private. Similarly, participants in reported that personal data \nand social media content may be inappropriate topics for chatbots because they can be sensitive\n. These concerns may be reduced if a chatbot demonstrates care about privacy .\nThe reported strategies to provide \\textit{personalization} in chatbots interactions are the following:\n\\textbf{[S1] to learn from and about the user:} state that chatbots should present strategies to learn from cultural, behavioral, personal, conversational, and contextual interaction data. For example, the authors suggest using Facebook profile information to build knowledge about users' personal information. also suggest that the chatbot should remember user's preferences disclosed in previous conversations. In , the authors propose an architecture where responses are generated based on a \\textit{personalization} rank that applies users' feedback about their general interests and preferences. When evaluating the user's experience with a virtual assistant chatbot, found 16 mentions to personalized interactions, where participants demonstrated the need for a chatbot to be aware of their personal quirks and anticipate their needs.\n\\textbf{[S2] to provide customizable agents:} suggest that users should be able to choose the level of the chatbot's attributes, for example, the agent's look and persona. By doing so, users with low social-agent orientation could use a non-humanized interface, which would better represent their initial perspective. This differentiation could be the first signal to personalize further conversation, such as focusing on more productive or playful interactions. Regarding chatbots' learning capabilities, in , interviews with potential users revealed that users should be able to manage what information the chatbot knows about them and decide whether the chatbot can learn from previous interactions or not. If the user prefers a more generic chatbot, then it would not store personal data, potentially increasing the engagement with more resistant users. raise the possibility of offering an ``incognito'' mode for chatbots or asking the chatbot to forget what was said in previous utterances.\n\\textbf{[S3] visual elements:} adopted quick replies as a means for the chatbot to tailor its subsequent questions to the specific experience the participant had reported. As discussed in Section \\ref{sec:conscientiousness}, quick replies may be seen as restrictive from an interactional perspective; however, conversation logs showed that the tailored questions prompted the users to report more detail about their experience, which is important in ethnography research. \nBoth the benefits and strategies identified in the literature are in line with the types of \\textit{personalization} proposed by . Therefore, further investigations in \\textit{personalization} can leverage knowledge from interactive systems (e.g., ) to adapt \\textit{personalization} strategies and manage the privacy concern.\n\\begin{framed} \\small \n\\vspace{-2mm}\nIn summary, the \\textbf{social intelligence} category includes characteristics that help a chatbot to manifest an adequate social behavior, by managing conflicts, using appropriate language, displaying \\textit{manners} and \\textit{moral agency}, sharing emotions, and handling personalized interactions. The benefits relate to resolving social positioning and recovering from failures, as well as increasing believability, human-likeness, engagement, and rapport. To achieve that, designers and researchers should care about privacy, emotional regulation issues, language consistency, and identification of failures and inappropriate content.\n\\vspace{-2mm}\n\\end{framed}","id":"69b481db-bee0-4219-a694-6a1c1caa8255","level":"subsubsection","origin_cites_number":16,"parent_id":"6e94027f-a398-4b65-8b3b-d03e95d73fa6","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Social Intelligence"],["subsubsection","Personalization"]],"subsections":[],"title":"Personalization"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:personification}\nIn this section, we discuss the influence of \\textit{identity} projection on human-chatbot interaction. \\textbf{Personification} refers to assigning personal traits to non-human agents, including physical appearance, and emotional states . In the HCI field, researchers argue that using a personified character in the user interface is a natural way to support the interaction . Indeed, the literature shows that (i) users can be induced to behave as if computers were humans, even when they consciously know better ; and (ii) the more human-like a computer representation is, the more social people's responses will be .\nChatbots are, by definition, designed to have at least one human-like trait: (human) natural language. Although research on \\textbf{personification} is more common in the Embodied Conversational Agents field, claim that chatbot embodiment can be created through narrative without any visual help. According to , talking to a machine affords it a new \\textit{identity}. In this section, we divided the social characteristics that reflect \\textbf{personification} into \\textit{identity} (16 papers) and \\textit{personality} (12 papers). In this category, we found several studies where part of the main investigation relates to the social characteristics. Six studies applied quantitative methods, whereas the majority reported qualitative (10) or mixed (11) methods. See the supplementary materials for details.","id":"fb12c3f5-18c5-4fb9-8465-0d7950752ebc","level":"subsection","origin_cites_number":6,"parent_id":"5950bb1f-b30c-410b-94c0-7eba5ce91d72","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Personification"]],"subsections":["3da905ba-ca45-4a2d-9608-fe81f0daa556","eec7d496-87cc-419a-b26e-00faf7d9c1fd"],"title":"Personification"},{"cite_extract_rate":0.08333333333333301,"cites":[6262,7476],"content":"\\label{sec:identity}\n\\textit{Identity} refers to the ability of an individual to demonstrate belonging to a particular social group . Although chatbots lack the agency to choose what social group to which they want to belong, designers attribute \\textit{identity} to them, intentionally or not, when they define the way a chatbot talks or behaves . The \\textit{identity} of a partner (even if only perceived) gives rise to new processes, expectations, and effects that influence the outcomes of the interaction . Aspects that convey the chatbots' \\textit{identity} include gender, age, language style, and name. Additionally, chatbots may have anthropomorphic, zoomorphic, or robotic representations. Some authors include \\textit{identity} aspects in the definition of \\textit{personality} (see, e.g., ). We distinguish these two characteristics, where \\textit{identity} refers to appearance and cultural traits while \\textit{personality} focuses on behavioral traits.\nWe found 16 studies that discuss \\textit{identity} issues, ten of which include \\textit{identity} as part of their main investigation . In two studies, the authors argue for the impact of \\textit{identity} on the interaction based on the literature . In three studies , qualitative analysis of conversational logs revealed that participants put effort into understanding aspects of the chatbots' \\textit{identity}. \\textit{Identity} concerns were primarily investigated for open domain and customer service chatbots (four studies each). In open domain interactions, \\textit{identity} is explored as a means of building common ground . In the case of customer services, \\textit{identity} helps manifest credibility and trust . Other domains include race-talk, education, and gaming. The complete list appears in the supplementary materials.\nThe identified benefits of attributing \\textit{identity} to a chatbot are the following:\n\\textbf{[B1] to increase engagement: }when evaluating signals of playful interactions, found that agent-oriented conversations (asking about an agent's traits and status) are consistent with the tendency to anthropomorphize the agent and engage in chit-chat. In the educational scenario, also observed questions about agents' appearance, intellectual capacities, and sexual orientation, although the researchers considered these questions inappropriate for the context of tutoring chatbots. When comparing human-like vs. machine-like language style, greetings, and framing, noticed that using informal language, having a human name, and using greetings associated with human communication resulted in significantly higher scores for adjectives like likeable, friendly, and personal. In addition, framing the agent as ``intelligent'' also had a slight influence on users' scores.\n\\textbf{[B2] to increase human-likeness:} some attributes may convey a perceived human-likeness. showed that using human-like language style, name, and greetings resulted in significantly higher scores for naturalness. The chatbot's framing influenced the outcomes when combined with other anthropomorphic clues. When evaluating different typefaces for a financial adviser chatbot, found that users perceive machine-like typefaces as more chatbot-like, although they did not find strong evidence of handwriting-like typefaces conveying humanness.\nThe surveyed literature also highlights challenges regarding \\textit{identity}:\n\\textbf{[C1] to avoid negative stereotypes: }when engaging in a conversation, interlocutors base their behavior on common ground (joint knowledge, background facts, assumptions, and beliefs that participants have of each other) (see ). Common ground reflects stereotypical attributions that chatbots should be able to manage as the conversation evolves . In , the authors discuss that chatbots for company representations are often personified as attractive human-like women acting as spokespeople for their companies, while men chatbots tend to have a more important position, such as a virtual CEO. state that the agent self-disclosure of gender \\textit{identity} opens possibilities to sex talk. The authors observed that the conversations mirror stereotyped male/female encounters, and the ambiguity of the chatbot's gender may influence the exploration of homosexuality. However, fewer instances were observed of sex-talk with the chatbot personified as a robot, which demonstrates that the gender \\textit{identity} may lead to the stereotypical attributions. When evaluating the effect of gender identity on disinhibition, showed that people spoke more often about physical appearance and used more swear and sexual words with the female-presenting chatbot, and racist attacks were observed in interactions with black-representing chatbots. The conversation logs from also show instances of users attacking the chatbot persona (a static avatar of a woman pointing to the conversation box). and also highlight that race identity conveys not only the physical appearance, but all the socio-cultural expectations about the represented group (see discussion in Section \\ref{sec:moralagency}). Hence, designers should care about the impact of attributing an \nidentity to chatbots in order to avoid reinforcing negative stereotypes.\n\\textbf{[C2] to balance the \\textit{identity} and the technical capabilities:} the literature comparing embodied vs. disembodied conversational agents yields contradictory results regarding the relevance of a human representation. For example, in the context of general-purpose interactions, show that people exert more effort toward establishing common ground with the agent when they are represented as fully human; in contrast, when evaluating a website assistant chatbot, show that simpler text-based chatbots with no visual, human \\textit{identity} resulted in less of an uncanny effect and reduced negative affect. Overly humanized agents create a higher expectation for users, which eventually leads to more frustration when the chatbot fails . When arguing about why chatbots fail, advocate for balancing human versus robotic aspects, where \\textit{``too human''} representations may lead to off-topic conversations, and overly robotic interactions may lack personal touch and flexibility. When arguing about the social presence conveyed by deceptive chatbots, state that extreme anthropomorphic features may generate cognitive dissonance. The challenge thus lies in designing a chatbot that provides appropriate \\textit{identity} cues, which correspond to their capabilities and communicative purpose, in order to convey the right expectation and minimize discomfort from over-personification.\nRegarding the strategies, the surveyed literature suggests \\textbf{[S1] to design and elaborate on a persona}. Chatbots should have a comprehensive persona and answer agent-oriented conversations with a consistent description of itself . For example, discuss that Eliza, the psychotherapist chatbot, and Parry, a paranoid chatbot, have behaviors consistent with the stereotypes associated with the professional and personal identities, respectively. suggest that designers should explicitly build signals of the chatbot personification (either machine- or human-like), so the users can have the right expectation about the interaction. When \\textit{identity} aspects are not explicit, users try to establish common ground. In and , much of the small talk with the chatbot related to the chatbot's traits and status. In , the authors observed many instances of Alice's self-references to ``\\textit{her}'' artificial nature. These references triggered the users to reflect on their human-condition (self-categorization process), resulting in exchanges about their species (either informational or confrontational). observed similar results, as participants engaged in conversations about the artificial nature of the agent. Providing the chatbot with the ability to describe its personal \\textit{identity} helps to establish the common ground, and hence, enrich the interpersonal relationship .\nChatbots may be designed to deceive users about their actual \\textit{identity}, pretending to be a human . In this case, the more human the chatbot sounds, the more successful it will be. In many cases, however, there is no need to engage in deception and the chatbots can be designed to represent an elaborated persona. \nResearchers can explore social identity theory related to ingroup bias, power relations, homogeneity, and stereotyping, in order to design chatbots with \\textit{identity} traits that reflect their expected social position .","id":"3da905ba-ca45-4a2d-9608-fe81f0daa556","level":"subsubsection","origin_cites_number":24,"parent_id":"fb12c3f5-18c5-4fb9-8465-0d7950752ebc","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Personification"],["subsubsection","Identity"]],"subsections":[],"title":"Identity"},{"cite_extract_rate":0.043478260869565,"cites":[7476],"content":"\\label{sec:personality}\n\\textit{Personality} refers to personal traits that help to predict someone's thoughts, feelings, and behaviors . The most accepted set of traits is called the Five-Factor model (or Big Five model) ), which describes \\textit{personality} across five dimensions (extraversion, agreeableness, conscientiousness, neuroticism, and openness). However, \\textit{personality} can also refer to other dynamic, behavioral characteristics, such as temperament and sense of humor . In the chatbots domain, \\textit{personality} refers to the set of traits that determines the agent's interaction style, describes its character, and allows the end-user to understand its general behavior . Chatbots with consistent \\textit{personality} are more predictable and trustworthy . According to , unpredictable swings in chatbots' attitudes can disorient users and create a strong sense of discomfort. Thus, \\textit{personality} ensures that a chatbot displays behaviors that stand in agreement with the users' expectations in a particular context .\nWe found 12 studies that report \\textit{personality} issues for chatbots. In some studies, \\textit{personality} was investigated in reference to the Big Five model , while two studies focused on sense of humor . Three studies investigated the impact of the \\textit{personality} of tutor chatbots on students' engagement . compared users' preferences regarding pre-defined personalities. In the remaining studies, \\textit{personality} concerns emerged from the qualitative analysis of the interviews, users' subjective feedback, and literature reviews . \\textit{Personality} was mostly investigated in open domain (five studies) and education (three studies) chatbots. The other two domains were gaming and humorous chatbots, which are both playful agents. Hence, \\textit{personality} is relevant when believability and attitude play a role in the interaction (e.g., in open domain) and when the chatbots' attitude may increase users' mental comfort when performing a task , such as in educational contexts.\nThe surveyed literature revealed two benefits of attributing \\textit{personality} to chatbots:\n\\textbf{[B1] to increase believability:} states that chatbots should have a \\textit{personality}, defined by the Five Factor model, plus characteristics such as temperament and tolerance, in order to build utterances using linguistic choices that cohere with these attributions. When evaluating a humorous chatbot, compared the naturalness of the chatbot's inputs and the chatbot's human-likeness compared to a non-humorous chatbot. The humorous chatbot scored significantly higher in both constructs. also showed that sense of humor humanizes the interactions, since humor was one of the factors that influenced naturalness for the WoZ condition. demonstrated that manipulating language to manifest a target \\textit{personality} produced moderately natural utterances, with a mean rating of 4.59 out of 7 for the \\textit{personality} model utterances.\n\\textbf{[B2] to enrich interpersonal relationships:} chatbots \\textit{personality} can make the interaction more enjoyable . In , the second most frequent motivation for using chatbots, noted by 20\\% of the participants, was entertainment. The authors argue that a chatbot's capacity to be fun is important even when the main purpose is productivity; according to participants, the chatbot's ``\\textit{fun tips}'' enrich the user experience. This result is consistent with the experience of first-time users , where participants relate better with chatbots who have consistent \\textit{personality}. show how witty banter and casual, enthusiastic conversations help make the interaction effortless. In addition, a few participants enjoyed the chatbot with a caring \\textit{personality}, who was described as a good listener. In and , the authors argue that a consistent \\textit{personality} helps the chatbot to gain the users' confidence and trust. state that tutor chatbots should display appropriate posture, conduct, and representation, which include being encouraging, expressive, and polite. Accordingly, other studies highlight students' desire for chatbots with positive agreeableness and extraversion . Outcomes consistently suggest that students prefer a chatbot that is not overly polite, but has some attitude. Agreeableness seems to play a critical role, helping the students to feel encouraged and deal with difficulties. Notably, agreeableness requires \\textit{emotional intelligence} to be warm and sympathetic in appropriate circumstances (see Section \\ref{sec:emotionalintelligence}).\nThe reviewed literature pointed out two challenges regarding \\textit{personality}:\n\\textbf{[C1] to adapt humor to the users' culture:} sense of humor is highly shaped by cultural environment . discusses a Japanese chatbot who uses puns to create funny conversations. The authors state that puns are one of the main humor genres in that culture. However, puns are restricted to a specific culture and language, with low portability. Thus, the design challenge lies in personalizing chatbots' sense of humor to the target users' culture and interests or, alternatively, designing cross-cultural forms of humor. The ability to adapt to the context and users' preference is discussed in Section \\ref{sec:personalization}.\n\\textbf{[C2] to balance \\textit{personality} traits:} observed that users prefer a proactive, productive, witty chatbot. Yet, they also would like them to be caring, encouraging, and exciting. In , the researchers intentionally generated utterances to reflect extreme personalities; as a result, they observed that some utterances sounded unnatural because a human's \\textit{personality} is a continuous phenomenon, rather than a discrete one. also points out that, although \\textit{personality} is consistent, moods and states of mind constantly vary. Thus, balancing the predictability of the \\textit{personality} and the expected variation is a challenge to overcome.\nWe also identified strategies to design chatbots that manifest \\textit{personality}:\n\\textbf{[S1] to use appropriate language: } and suggest that the chatbot language should be consistently influenced by its \\textit{personality}. Both studies propose that chatbots' architecture should include a persona-based model that encodes the \\textit{personality} and influences the response generation. proposed a framework to shows that it is possible to automatically manipulate language features to manifest a particular \\textit{personality} based on the Big Five model. The Big Five model is a relevant tool because it can be assessed using validated psychological instruments . Using this model to represent the \\textit{personality} of chatbots was also suggested by and . discussed that a chatbot's \\textit{personality} should match its domain. Participants expected the language used by the news chatbot to be professional, while they expected the shopping chatbot to be casual and humorous. The ability to use consistent language is discussed in Section \\ref{sec:thoroughness}.\n\\textbf{[S2] to have a sense of humor:} literature highlights humor as a positive \\textit{personality} trait . In , ten participants mentioned enjoyment when the chatbots provided humorous and highly diverse responses. The authors found occurrences of the participants asking for jokes and expressing delight when the request was supported. present similar results when arguing that humor is important even for task-oriented chatbots when the user is usually seeking for productivity. For casual conversations, highlight that timely, relevant, and clever wit is a desired \\textit{personality} trait. In , the joker chatbot was perceived as more human-like, knowledgeable, and funny, and participants felt more engaged. \n\\textit{Personality} for artificial agents has been studied for some time in the Artificial Intelligence field . Thus, further investigations on chatbots' \\textit{personality} can leverage models for \\textit{personality} and evaluate how they contribute to believability and rapport building.\n\\begin{framed}\\small\n\\vspace{-2mm}\nIn summary, the \\textbf{personification} category includes characteristics that help a chatbot to manifest personal and behavioral traits. The \nbenefits relate to increasing believability, human-likeness, engagement, and interpersonal relationship, which is in line with the benefits of \\textbf{social intelligence}. However, unlike the \\textbf{social intelligence} category, designers and researchers should focus on attributing recognizable \\textit{identity} and \\textit{personality} traits that are consistent with users' expectations and the chatbot's capabilities. In addition, it is important to care about adaptation to users' cultural context and reducing the effects of negative stereotypes.\n\\vspace{-2mm}\n\\end{framed}","id":"eec7d496-87cc-419a-b26e-00faf7d9c1fd","level":"subsubsection","origin_cites_number":23,"parent_id":"fb12c3f5-18c5-4fb9-8465-0d7950752ebc","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Chatbots Social Characteristics"],["subsection","Personification"],["subsubsection","Personality"]],"subsections":[],"title":"Personality"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we discuss cross-cutting aspects of the results, namely the influence of the chatbots' perceived humanness, the conversational domains, and the relationships among the characteristics.","id":"616598c0-a6ef-4e0d-bae2-49c648e87606","level":"section","origin_cites_number":0,"parent_id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Discussion"]],"subsections":["212bc0e6-0de8-4c65-af3e-ab937c5ebe0e","3bce4038-12ed-4398-a390-053fbed73d7f","267f99ec-a3fd-4ef2-ae14-9b0b6a98c88c"],"title":"Discussion"},{"cite_extract_rate":0.16666666666666602,"cites":[6267,7476],"content":"The social characteristics identified in the survey align well with characteristics that are present in human-human interactions. On the one hand, this survey shows the benefits of considering these characteristics when developing chatbots, which is supported by the Media Equation theory ; we identified the domains in which each characteristic was studied. On the other hand, previous studies have also shown that developing overly humanized agents results in high expectations and uncanny feelings , which was also pointed out in the surveyed literature as a challenge to conveying particular characteristics, such as \\textit{identity} and \\textit{personality}.\nAn explanation for these contrasts is that interactivity is ``dependent on the identity of the source with whom/which we are carrying out the message exchange''~, i.e., people direct their messages to an artificial agent. In interactions with chatbots, the artificial agent represents the social role usually associated with a human, for example, a tutor~, a healthcare provider~, a salesperson~, a hotel concierge~, or even a friend . The idea of assigning a social role to a chatbot does not imply deceiving people into thinking the software is human. Still, as the chatbots enrich their communication and social skills, their perceived social role may approach human profiles.","id":"212bc0e6-0de8-4c65-af3e-ab937c5ebe0e","level":"subsection","origin_cites_number":12,"parent_id":"616598c0-a6ef-4e0d-bae2-49c648e87606","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Discussion"],["subsection","A note on chatbots' humanness"]],"subsections":[],"title":"A note on chatbots' humanness"},{"cite_extract_rate":0,"cites":[],"content":"As Table \\ref{tab:domain-analysis} depicts\\footnote{see more details in the supplementary materials}, almost all social characteristics (10 out of 11) were found in studies related to conversations not restricted to a specific domain (open domain): manners, moral agency, and emotional intelligence are particularly relevant, since many conversations involve personal content (e.g., preferences, personal life, and habits); damage control is necessary since interactions are more susceptible to testing or flaming; identity, personality, and thoroughness are adopted to increase believability and consistency; and conscientiousness, proactivity, and personalization foster user engagement and a coherent conversational flow. The only characteristic that was not investigated in an open domain was communicability. As the conversations in this context mimic human social interactions, conveying the chatbots' capabilities is less of a concern.\nThe reported characteristics vary across task-oriented domains. Education and customer services literature report the greatest number of characteristics, but this result might be influenced by the maturity of the research in chatbots for these domains. As we pointed out in Section \\ref{sec:overview}, education and customer services are the task-oriented domains most reported in the literature. We found conscientiousness, damage control, thoroughness, manners, emotional intelligence, and identity in studies for both domains. However, manners and emotional intelligence have a different goal in these domains. In the education context, these characteristics are designed to encourage students, especially in a situation of failure, in which the chatbot should be comforting and sensitive. This function aligns with other domains, such as healthcare. The education domain also reports needs for personality, so the chatbot can be recognized as either an instructor or a student, and proactivity, so the chatbot can motivate students to participate in the interactions. Personality is also reported in other domains in which the chatbots' character influences the interactions, such as gaming and humorous talk. Proactivity is also consistently reported in domains in which the chatbot provides guidance, such as coaching, health, ethnography, and assessment interviews.\nIn customer services, emotional intelligence and manners are used as a means to manage customers' frustrations with either a product or service. This function is also reported in other domains, such as virtual assistants and information search. Additionally, communicability is important to convey the services provided by the chatbot, and personalization helps to provide services that align with a particular customer, which is observed in domains such as ethnography, task management, virtual assistants, and financial services. Noticeably, personalization mostly co-occurs with proactivity, since users expects proactive messages to be relevant to their particular interactional context.\nMoral agency is the only characteristic that does not show up across several domains. It was reported only in race-talk, and open domain. This highlights that moral agency is a concern when the topic involves content that might lead to inappropriate responses. However, as this survey focuses on characteristics reported in the literature, we believe that it might be influenced by the approach adopted by the researchers: it makes sense to study moral agency in the context of gender or race conversations. We argue, though, that this characteristic might be relevant in domains that were not reported here, but that involve interactions that may lead to biased conversations.","id":"3bce4038-12ed-4398-a390-053fbed73d7f","level":"subsection","origin_cites_number":0,"parent_id":"616598c0-a6ef-4e0d-bae2-49c648e87606","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Discussion"],["subsection","The importance of conversational context"]],"subsections":[],"title":"The importance of conversational context"},{"cite_extract_rate":0.038461538461538006,"cites":[7476],"content":"\\label{sec:interrelationship}\nIn Section \\ref{sec:socialcharacteristics}, we organized the social characteristics into discrete groups and discussed several instances of characteristics influencing each other, or being used as a strategy to manifest one another. In this section, we describe these relations in a theoretical framework, depicted in Figure \\ref{fig:relationship}. Rather than providing a comprehensive map, the goal here is to make explicit the relationships found in the surveyed chatbots literature, which underlines the complexity of developing chatbots with appropriate social behaviors. In Figure \\ref{fig:relationship}, boxes represent the social characteristics and the colors group them into their respective categories. The axes represent the 22 propositions we derived from the literature.\n\\begin{figure}[!hb]\n \\centering\n \\includegraphics[width=8cm]{interrelationship-survey}\n \\caption{Interrelationship among social characteristics}\n \\label{fig:relationship}\n\\end{figure}\nAccording to the surveyed literature, \\textit{proactivity} influences perceived \\textit{personality} \\textbf{(P1)} , since recommending and initiating topics may manifest higher levels of extraversion. When the proactive messages are based on the context, \\textit{proactivity} increases perceived \\textit{conscientiousness} \\textbf{(P2)} , since the proactive messages may demonstrate attention to the topic. \\textit{Proactivity} supports \\textit{communicability} \\textbf{(P3)} , since a chatbot can proactively communicate its knowledge and provide guidance; in addition, \\textit{proactivity} supports \\textit{damage control} \\textbf{(P4)} , since a chatbot can introduce new topics when the user either is misunderstood, tries to break the system, or sends an inappropriate message.\n\\textit{Conscientiousness} is itself a dimension of the Big Five model; hence, \\textit{conscientiousness} influences the perceived \\textit{personality} \\textbf{(P5)} . Higher levels of context management, goal-orientation, and understanding increase the chatbots' perceived efficiency, organization, and commitment . \\textit{Conscientiousness} manifests\\textit{ emotional intelligence} \\textbf{(P6)} since retaining information from previous turns and being able to recall them demonstrates empathy . In addition, \\textit{conscientiousness} manifests \\textit{personalization} \\textbf{(P7)} because a chatbot can remember individual preferences within and across sessions.\n\\textit{Emotional intelligence} influences perceived \\textit{personality} \\textbf{(P8)}, since chatbots' \\textit{personality} traits affect the intensity of emotional reactions . Agreeableness is demonstrated through consistent warm reactions such as encouraging and motivating \n. Some \\textit{personality} traits \nrequire \\textit{personalization} \\textbf{(P9)} to adapt to the interlocutors' culture and interests . Besides, \\textit{personalization} benefits \\textit{identity} \\textbf{(P10)}, since the users' social-agent orientation may require a chatbot to adapt the level of engagement in small talk and the agent's visual representation . \\textit{Personalization} also improves \\textit{emotional intelligence} \\textbf{(P11)}, since a chatbot should dynamically regulate the affective reactions to the interlocutor . \\textit{Emotional intelligence} improves perceived \\textit{manners} \\textbf{(P12)}, since the lack of \\textit{emotional intelligence} may lead to the perception of impoliteness .\n\\textit{Conscientiousness} facilitates \\textit{damage control} \\textbf{(P13)}, since the attention to the \nworkflow and context may increase the ability to recover from a failure without restarting the workflow . \\textit{Communicability} facilitates \\textit{damage control} \\textbf{(P14)}, since it teaches the user how to communicate, reducing the numbers of mistakes . In addition, suggesting how to interact can reduce frustration after failure scenarios .\n\\textit{Personalization} manifests \\textit{thoroughness} \\textbf{(P15)} , since chatbots can adapt their language use to the conversational context and the interlocutor's expectations. When the context requires dynamic variation , \\textit{thoroughness} may reveal traits of the chatbot's \\textit{identity} \\textbf{(P16)} . As demonstrated by , \\textit{thoroughness} also reveals \\textit{personality} \\textbf{(P17)}.\n\\textit{Manners} influence \\textit{conscientiousness} \\textbf{(P18)} , since they can be applied as a strategy to politely refuse off-topic requests and to keep the conversation on track. \\textit{Manners} also influence \\textit{damage control} \\textbf{(P19)} , because they can help a chatbot prevent verbal abuse and reduce the negative effect of lack of knowledge. Both \\textit{moral agency} \\textbf{(P20)} and \\textit{emotional intelligence} \\textbf{(P21)} improve \\textit{damage control} because they provide\nthe ability to appropriately respond to abuse and testing . \\textit{Identity} influences \\textit{moral agency} \\textbf{(P22)}, since \\textit{identity} representations require the ability to prevent a chatbot from building or reinforcing negative stereotypes .","id":"267f99ec-a3fd-4ef2-ae14-9b0b6a98c88c","level":"subsection","origin_cites_number":26,"parent_id":"616598c0-a6ef-4e0d-bae2-49c648e87606","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Discussion"],["subsection","Interrelationships among the characteristics"]],"subsections":[],"title":"Interrelationships among the characteristics"},{"cite_extract_rate":0.11111111111111101,"cites":[8990,6268],"content":"\\label{sec:relatedsurveys}\nPrevious studies have reviewed the literature on chatbots. Several surveys discuss chatbots' urgency and their potential application for particular domains, which include education , business , health , information retrieval and e-commerce . Other surveys focus on technical design techniques , such as language generation models, knowledge management, and architectural challenges. Although discuss the social capabilities of chatbots, the survey focuses on the potential of available open source technologies to support these skills, highlighting technical hurdles rather than social ones.\nWe found three surveys that include insights about social characteristics of chatbots, although none of them focus on this theme. investigated chatbots that ``\\textit{mimic conversation rather than understand it},'' and review the main technologies and ideas that support their design, while focuses on identifying best practices for developing script-based chatbots. review the literature on quality issues and attributes for chatbots. The supplementary materials include a table that shows the social characteristics covered by each survey. These related surveys also point out technical characteristics and attributes that are outside the scope of this survey.","id":"cefcc373-1755-478a-bdc0-b708ace17a7a","level":"section","origin_cites_number":18,"parent_id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Related Surveys"]],"subsections":[],"title":"Related Surveys"},{"cite_extract_rate":0.16666666666666602,"cites":[6269],"content":"\\label{sec:limitations}\nThis research has some limitations. Firstly, since this survey focused on disembodied, text-based chatbots, the literature on embodied and speech-based conversational agents was excluded. We acknowledge that studies that include these attributes can have relevant social characteristics for chatbots, especially for characteristics that could be highly influenced by physical representations, tone, accent, and so forth (e.g. identity, politeness, and thoroughness). However, embodiment and speech could also bring new challenges (e.g., speech-recognition or eye-gazing), which are beyond the scope of this study and could potentially impact users' experiences with chatbots. Additionally, this decision may have caused the under-representation of certain research domains. For example, there is an established research branch on conversational agents in the Ambient Intelligence discipline~ that primarily investigates multimodal interactions, and hence is not represented in this survey. Another example is the Software Engineering domain (see, e.g.,~), where chatbots have been widely investigated, but focused on the functional, rather than social aspects of the interactions.\nSecondly, since the definition of chatbot is not consolidated in the literature and chatbots have been studied across several different domains, some studies that include social aspects of chatbots may not have been found. To account for that, we adopted several synonyms in our research string and used Google Scholar as search engine, which provides a fairly comprehensive indexing of the literature in more domains. Finally, the conceptual model of social characteristics was derived through a coding process inspired by qualitative methods, such as Grounded Theory. Like any qualitative coding methods, it relies on the researchers' subjective assessment. To mitigate this threat, the researchers discussed the social characteristics and categories during in-person meetings until reaching consensus, and both the conceptual framework and the relationships amongst the characteristics were derived based on outcomes the surveyed studies explicitly reported.","id":"e47b0d6a-5496-42d1-b93d-9f8d302b471a","level":"section","origin_cites_number":6,"parent_id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Limitations"]],"subsections":[],"title":"Limitations"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclusion}\nIn this survey, we investigated the literature on disembodied, text-based chatbots to answer the question \\textit{``What chatbot social characteristics benefit human interactions and what are the challenges and strategies associated with them?''} Our main contribution is the conceptual model of social characteristics, from which we can derive conclusions about several research opportunities. Firstly, we point out several challenges to overcome in order to design chatbots that manifest each characteristic. Secondly, further research may focus on assessing the extent to which the identified benefits are perceived by users and influence their satisfaction. Finally, further investigations may propose new strategies to manifest particular characteristics. In this sense, we highlight that we could not identify strategies to manifest moral agency and thoroughness, although strategies for several other characteristics are also under-investigated.\nWe reported domains where social characteristics were primarily investigated. We showed that some characteristics are largely influenced by the domain (e.g., \\textit{moral agency} and \\textit{communicability}), while others are more generally applied (e.g., \\textit{manners} and \\textit{damage control}). We also discussed the relationship among the characteristics through 22 propositions derived from the surveyed literature, which underline the complexity of developing chatbots with appropriate social behaviors. Although the propositions and social characteristics do not intend to be comprehensive, they cover important aspects of human-human communication. Social sciences such as sociolinguistics and communication studies have a good deal to contribute to the design of chatbots. In the description of the characteristics, we pointed out theories that could guide these investigations, such as the cooperative principle , social identity , personalization , politeness , and mind perception theories . For example, \\textit{conscientiousness}, \\textit{thoroughness}, and \\textit{proactivity} might be covered by the cooperative principle , and social identity theory might explain the relevance of \\textit{identity}, \\textit{personality}, and \\textit{moral agency} for conversational technologies. Our results provide important references for helping designers and researchers find opportunities to advance the human-chatbot interactions field.\n\\section*{Disclosure statement}\nNo potential conflict of interest was reported by the authors.\n\\bibliographystyle{apacite}\n\\bibliography{refs}\n\\section*{About the Authors}\n\\textbf{Ana Paula Chaves} is a Ph.D. Candidate at the Northern Arizona University and a Faculty at the Federal University of Technology - Paraná, Campus Campo Mourão, Brazil. She researches social aspects of human-chatbot interactions as well as technologies for tourism. More information at http://anachaves.pro.br\n\\textbf{Marco Aurelio Gerosa} is an Associate Professor at the Northern Arizona University. He researches Software Engineering and CSCW, having served on the program committee of important conferences, such as FSE, CSCW, SANER, and MSR. He is an ACM and IEEE senior member. For more information, visit http://www.marcoagerosa.com\n\\end{document}","id":"1430d8e4-a5c1-4f7d-ae08-3a9fcec59bb4","level":"section","origin_cites_number":6,"parent_id":"48e1448c-68e1-4eb0-9088-8c4cfe65bb83","prefix_titles":[["title","How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ]\n ],\n \"subsections\": [\n \"3e165ffa-29cf-4605-ba88-0e6c84a411e5\",\n \"ad5096ab-28b5-4a2f-b84b-74bcbea02a02\",\n \"5950bb1f-b30c-410b-94c0-7eba5ce91d72\",\n \"616598c0-a6ef-4e0d-bae2-49c648e87606\",\n \"cefcc373-1755-478a-bdc0-b708ace17a7a\",\n \"e47b0d6a-5496-42d1-b93d-9f8d302b471a\",\n \"1430d8e4-a5c1-4f7d-ae08-3a9fcec59bb4\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.026315789473684,\n \"cites\": [\n 6262\n ],\n \"content\": \"\\\\label{sec:introduction}\\nChatbots are computer programs that interact with users in natural language . The origin of the chatbot concept dates back to 1950 . ELIZA and A.L.I.C.E. are examples of early chatbot technologies, where the main goal was to mimic human conversations. Over the years, the chatbot concept has evolved. Today, chatbots can have distinct and diverse characteristics, which has resulted in several synonyms, such as multimodal agents, chatterbots, and conversational interfaces. In this survey, we use the term ``\\\\textit{chatbot}'' to refer to \\\\textit{a disembodied conversational agent that holds a natural language conversation via text-based environment to engage the user in either a general-purpose or task-oriented conversation.}\\nChatbots are changing the patterns of interactions between humans and computers . Many instant messenger tools, such as Skype, Facebook Messenger, and Telegram provide platforms to develop and deploy chatbots, which either engage with users in general conversations or help them solve domain specific tasks . As messaging tools increasingly become platforms, traditional websites and apps are providing space for this new form of human-computer interaction (HCI) . For example, in the 2018 F8 Conference, Facebook announced that it had 300K active chatbots on Facebook Messenger . The BotList \\\\footnote{https://botlist.co/} website indexes thousands of chatbots for education, entertainment, games, health, productivity, travel, fun, and several other categories. The growth of chatbot technology is changing how companies engage with their customers , students engage with their learning groups , and patients self-monitor the progress of their treatment , among many other applications.\\nHowever, chatbots still fail to meet users' expectations . While many studies on chatbot design focus on improving chatbots' functional performance and accuracy (see, e.g., ), the literature has consistently suggested that chatbots' interactional goals should also include social capabilities . According to the Media Equation theory , people naturally respond to social situations when interacting with computers . As chatbots are designed to interact with users in a way that mimics person-to-person conversations, new challenges in HCI arise . Neururer and colleagues (\\\\citeyear{neururer2018perceptions}) state that making a conversational agent acceptable to users is primarily a social, not only technical, problem. Studies on chatbots have shown that people prefer agents who: conform to gender stereotypes associated with tasks ; self-disclose and show reciprocity when recommending ; and demonstrate a positive attitude and mood . When chatbots do not meet these expectations, the user may experience frustration and dissatisfaction . In contrast, designing overly humanized agents results in uncanny feelings and increased expectations , which also negatively impacts the interaction. The challenge remains as to what social characteristics are relevant for improving chatbots' communication and social skills and in which domains they have shown to be beneficial.\\nAlthough chatbots' social characteristics have been explored in the literature, this knowledge is spread across several domains in which chatbots have been studied, such as customer services, education, finances, and travel. In the HCI domain, some studies focus on investigating the social aspects of human-chatbot interactions (see, e.g., ). However, most studies focus on a single or small set of characteristics (e.g., ); in other studies, the social characteristics emerged as secondary, exploratory results (e.g., ). It has become difficult to find evidence regarding what characteristics are important for designing a particular chatbot, and what research opportunities exist in the field.\\nWhilst the literature has extensively reviewed the technical aspects of chatbots design (e.g., ), a lack of studies brings together the social characteristics that influence the way users perceive and behave toward chatbots. To fill this gap, this survey compiles research initiatives for understanding the impact of chatbots' social characteristics on the interaction. We bring together literature that is spread across several research areas. From our analysis of 56 scientific studies, we derive a conceptual model of social characteristics, aiming to help researchers and designers identify the subset of characteristics that are relevant to their context and how adopting--or neglecting--a particular characteristic may influence the way humans perceive the chatbots. The research question that guided our investigation was: \\\\textbf{What chatbot social characteristics benefit human interaction and what are the challenges and strategies associated with them?}\\nTo answer this question, we discuss why designing a chatbot with a particular characteristic can enrich the human-chatbot interaction. Our results provide insight into whether the characteristic is desirable for a particular chatbot, so that designers can make informed decisions by selecting the appropriate subset of characteristics, as well as inspire researchers' further investigations. In addition, we discuss the influence of humanness and the conversational context on users' perceptions as well as the interrelationship among the identified characteristics. We stated 22 propositions about how social characteristics may influence one another. In the next section, we present an overview of the studies included in this survey.\",\n \"id\": \"3e165ffa-29cf-4605-ba88-0e6c84a411e5\",\n \"level\": \"section\",\n \"origin_cites_number\": 38,\n \"parent_id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 6263,\n 6262\n ],\n \"content\": \"\\\\label{sec:overview}\\nThe literature presents no coherent definition of chatbots; thus, to find relevant studies we used a search string that includes the synonyms \\\\textit{``chatbots, chatterbots, conversational agents, conversational interfaces, conversational systems, conversation systems, dialogue systems, digital assistants, intelligent assistants, conversational user interfaces, and conversational UI''}. We explicitly left out studies that relate to embodiment (\\\\textit{``ECA, multimodal, robots, eye-gaze, gesture''}), and speech input mode (\\\\textit{``speech-based, speech-recognition, voice-based''}). The search string did not include the term ``social bots'', because it refers to chatbots that produce content for social networks such as Twitter . We did not include ``personal assistants'' either, since this term consistently refers to commercially available, voice-based assistants such as Google Assistant, Amazon's Alexa, Apple's Siri, and Microsoft's Cortana.\\nWe decided to not include terms that relate to social characteristics/traits, because most studies do not explicitly label their results as such. Additionally, to find studies from a variety of domains (not only computing), we used Google Scholar to search for papers. We started with a set of about one thousand papers. In the first round, we reduced this amount by about half based on the titles. We also removed papers for which full text was not available or written in a language other than English. We started the second round with 464 studies. In this round, we excluded short/position papers, theses, book chapters, and documents published in non-scientific venues. We also read the abstracts and removed all the papers that focused on technical aspects, rather than social ones. The third and last round started with 104 papers. After reading the full texts, we removed the papers that either did not highlight social characteristics or represented previous, ongoing research of another complete study from the same authors. The datasets with all analyzed studies (original list of results and exclusions per round) is available in .\\nAfter filtering the search results, we had 56 remaining studies. Most of the selected studies are recent publications (less than 10 years old). The publication venues include the domains of human-computer interactions (25 papers), learning and education (8 papers), information and interactive systems (8 papers), virtual agents (5 papers), artificial intelligence (3 papers), and natural language processing (2 papers). We also found papers from health, literature \\\\& culture, computer systems, communication, and humanities (1 paper each). Most papers (59\\\\%) focus on task-oriented chatbots. General purpose chatbots reflect 33\\\\% of the surveyed studies. Most general purpose chatbots (16 out of 19) are designed to handle topic-unrestricted conversations. The most representative specific domain is education, with 9 papers, followed by customer services, with 5 papers. See the supplementary materials for the complete list of topics. \\nMost surveyed studies adopted real chatbots (35 out of 56); among them, 18 studies analyze logs of conversations or users' perceptions of third-party chatbots such as Cleverbot (e.g., ), Talkbot (e.g., ), and Woebot . Nine studies introduce a self-developed architecture and/or dialogue management (e.g. ). In another nine studies, the chatbots were designed for research purposes using third-party platforms for chatbot development and deployment, such as IBM's Watson service (e.g., ) and Microsoft's Bot Framework as well as pattern-matching packages such as Artificial Intelligence Markup Language (AIML) . When a chatbot was simulated (11 studies), Wizard of Oz (WoZ) is the most used technique (9 studies). In a WoZ study, participants believe to be interacting with a chatbot when, in fact, a person (or ``wizard'') pretends to be the automated system . Eight studies do not address a particular chatbot. See the supplementary materials for details.\\nWe analyzed the papers by searching for chatbot behavior or attributed characteristics that influence the way users perceive it and behave toward it. Noticeably, the characteristics and categories are seldom explicitly highlighted to in the literature, so the conceptual model was derived using a qualitative coding process inspired by methods such as Grounded Theory (open coding stage). For each study (\\\\textit{document}), we selected relevant statements from the paper (\\\\textit{quotes}) and labeled them as a characteristic (\\\\textit{code}). These steps were performed by one researcher (the first author). After coding all the studies, a second researcher (the second author) reviewed the produced set of characteristics and both researchers participated in discussion sessions to identify characteristics that could be merged, renamed, or removed. At the end, the characteristics were grouped into the categories, depending on whether the characteristic relates to the chatbot's virtual representation, conversational behavior, or social protocols. Finally, the quotes for each characteristic were labeled as references to benefits, challenges, or strategies.\\nWe derived a total of 11 social characteristics, and grouped them into three categories, as depicted in Table \\\\ref{tab:conceptualmodel}: \\\\textbf{conversational intelligence}, \\\\textbf{social intelligence}, and \\\\textbf{personification}. The next section describes the derived conceptual model.\\n\\\\begin{landscape}\\n\\\\begin{table}[ht]\\n\\\\scriptsize\\n\\\\centering\\n\\\\caption{Conceptual model of chatbots social characteristics. Eleven characteristics found in the literature were grouped into three main categories, depending on whether the social characteristic relates to the chatbot's conversational behavior, social protocols, or virtual representation. For each characteristic, we point out the benefits, challenges, and strategies of implementing the characteristic in chatbots as reported in the literature.}\\n\\\\vspace{1mm}\\n\\\\begin{tabular}{c|p{2.5cm}|p{6cm}|p{6cm}|p{6cm}|}\\n\\\\cline{2-5}\\n & \\\\textbf{\\\\shortstack{\\\\\\\\Social\\\\\\\\Characteristics}}& \\\\multicolumn{1}{c|}{\\\\textbf{Benefits}} & \\\\multicolumn{1}{c|}{\\\\textbf{Challenges}} & \\\\multicolumn{1}{c|}{\\\\textbf{Strategies}} \\\\\\\\ \\\\hline\\n\\\\multicolumn{1}{|c|}{\\\\multirow{14}{*}{\\\\rotatebox{90}{\\\\textbf{\\\\shortstack{Conversational\\\\\\\\Intelligence}}}}} & \\\\multirow{6}{*}{\\\\textit{Proactivity}} & \\\\textbf{[B1]} to provide additional information & \\\\textbf{[C1]} timing and relevance & \\\\textbf{[S1]} to leverage conversational context \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to inspire users and to keep the conversation alive & \\\\textbf{[C2]} privacy & \\\\textbf{[S2]} to select a topic randomly \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B3]} to recover from a failure & \\\\textbf{[C3]} users' perception of being controlled & \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B4]} to improve conversation productivity & & \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B5]} to guide and engage users & & \\\\\\\\ \\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{4}{*}{\\\\textit{Conscientiousness}} & \\\\textbf{[B1]} to keep the conversation on track & \\\\textbf{[C1]} to handle task complexity & \\\\textbf{[S1]} conversational flow \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to demonstrate understanding & \\\\textbf{[C2]} to harden the conversation & \\\\textbf{[S2]} visual elements \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B3]} to hold a continuous conversation & \\\\textbf{[C3]} to keep the user aware of the chatbot's context & \\\\textbf{[S3]} confirmation messages \\\\\\\\ \\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{4}{*}{\\\\textit{Communicability}} & \\\\textbf{[B1]} to unveil functionalities & \\\\textbf{[C1]} to provide business integration & \\\\textbf{[S1]} to clarify the purpose of the chatbot \\\\\\\\ \\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to manage the users' expectations & \\\\textbf{[C2]} to keep visual elements consistent with textual inputs & \\\\textbf{[S2]} to advertise the functionality and suggest the next step \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & & & \\\\textbf{[S3]} to provide a help functionality \\\\\\\\\\\\hline\\n\\\\multicolumn{1}{|l|}{\\\\multirow{19}{*}{\\\\rotatebox{90}{\\\\textbf{\\\\shortstack{Social\\\\\\\\Intelligence}}}} }& \\\\multirow{6}{*}{\\\\textit{Damage control}} & \\\\textbf{[B1]} to appropriately respond to harassment & \\\\textbf{[C1]} to deal with unfriendly users & \\\\textbf{[S1]} emotional reactions \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to deal with testing & \\\\textbf{[C2]} to identify abusive utterances & \\\\textbf{[S2]} authoritative reactions \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B3]} to deal with lack of knowledge & \\\\textbf{[C3]} to fit the response to the context & \\\\textbf{[S3]} to ignore the user's utterance and change the topic \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & & & \\\\textbf{[S4]} \\\\textit{conscientiousness} and \\\\textit{communicability} \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & & & \\\\textbf{[S5]} to predict users' satisfaction \\\\\\\\\\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{2}{*}{\\\\textit{Thoroughness}} & \\\\textbf{[B1]} to increase human-likeness & \\\\textbf{[C1]} to decide how much to say & Not identified \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to increase believability & \\\\textbf{[C2]} to be consistent & \\\\\\\\ \\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{2}{*}{\\\\textit{Manners}} & \\\\textbf{[B1]} to increase human-likeness & \\\\textbf{[C1]} to deal with face-threatening acts & \\\\textbf{[S1]} to engage in small talk \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & & \\\\textbf{[C2]} to end a conversation gracefully & \\\\textbf{[S2]} to adhere turn-taking protocols \\\\\\\\ \\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{3}{*}{\\\\textit{Moral agency}} & \\\\textbf{[B1]} to avoid stereotyping & \\\\textbf{[C1]} to avoid alienation & Not identified \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to enrich interpersonal relationships & \\\\textbf{[C2]} to build unbiased training data and algorithms & \\\\\\\\ \\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{3}{*}{\\\\textit{\\\\shortstack{Emotional\\\\\\\\intelligence}}} & \\\\textbf{[B1]} to enrich interpersonal relationships & \\\\textbf{[C1]} to regulate affective reactions & \\\\textbf{[S1]} to use social-emotional utterances \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to increase engagement & & \\\\textbf{[S2]} to manifest \\\\textit{conscientiousness} \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B3]} to increase believability & & \\\\textbf{[S3]} reciprocity and self-disclosure \\\\\\\\ \\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{3}{*}{\\\\textit{Personalization}} & \\\\textbf{[B1]} to enrich interpersonal relationships & \\\\textbf{[C1]} privacy & \\\\textbf{[S1]} to learn from and about the user \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to provide unique services & & \\\\textbf{[S2]} to provide customizable agents \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B3]} to reduce interactional breakdowns & & \\\\textbf{[S3]} visual elements \\\\\\\\ \\\\hline\\n\\\\multicolumn{1}{|l|}{\\\\multirow{5}{*}{\\\\rotatebox{90}{\\\\textbf{\\\\shortstack{Personifi-\\\\\\\\cation}}}} }& \\\\multirow{3}{*}{\\\\textit{Identity}} & \\\\textbf{[B1]} to increase engagement & \\\\textbf{[C1]} to avoid negative stereotypes & \\\\textbf{[S1]} to design and elaborate on a persona \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to increase human-likeness & \\\\textbf{[C2]} to balance the \\\\textit{identity} and the technical capabilities & \\\\\\\\\\\\cline{2-5}\\n\\\\multicolumn{1}{|l|}{} & \\\\multirow{2}{*}{\\\\textit{Personality}} & \\\\textbf{[B1]} to increase believability & \\\\textbf{[C1]} to adapt humor to the users' culture & \\\\textbf{[S1]} to use appropriate language \\\\\\\\\\n\\\\multicolumn{1}{|l|}{} & & \\\\textbf{[B2]} to enrich interpersonal relationships & \\\\textbf{[C2]} to balance the \\\\textit{personality} traits & \\\\textbf{[S2]} to have a sense of humor \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\label{tab:conceptualmodel}\\n\\\\end{table}\\n\\\\end{landscape}\",\n \"id\": \"ad5096ab-28b5-4a2f-b84b-74bcbea02a02\",\n \"level\": \"section\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Overview of the surveyed literature\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Overview of the surveyed literature\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:socialcharacteristics}\\nThis section describes the identified social characteristics grouped into categories. As Table \\\\ref{tab:conceptualmodel} depicts, the category \\\\textbf{conversational intelligence} includes characteristics that help the chatbot manage interactions. \\\\textbf{Social intelligence} focuses on habitual social protocols, while \\\\textbf{personification} refers to the chatbot's perceived identity and personality representations. We also grouped the social characteristics based on the domain in which they were investigated (Table \\\\ref{tab:domain-analysis}). In the following subsections, we describe the identified social characteristics as well as the domains of study. Then, we summarize the relationship among the characteristics in Section \\\\ref{sec:interrelationship}. For each category, a table with an overview of the surveyed studies is provided in the supplementary materials. The supplementary materials also include tables for each social characteristic, listing the studies associated with the domains of study and reported benefits, challenges, and strategies. Finally, the supplementary materials also highlight five constructs that can be used to assess whether social characteristics reach the intended design goals.\",\n \"id\": \"5950bb1f-b30c-410b-94c0-7eba5ce91d72\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ]\n ],\n \"subsections\": [\n \"9ad3c98e-3389-49e5-a57e-7e261302fc58\",\n \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"fb12c3f5-18c5-4fb9-8465-0d7950752ebc\"\n ],\n \"title\": \"Chatbots Social Characteristics\"\n },\n {\n \"cite_extract_rate\": 0.035087719298245,\n \"cites\": [\n 6262,\n 7476\n ],\n \"content\": \"\\\\label{sec:conversationalintelligence}\\n\\\\textbf{Conversational intelligence} enables the chatbot to actively participate in the conversation and to demonstrate awareness of the topic discussed, the evolving conversational context, and the flow of the dialogue. Therefore, \\\\textbf{conversational Intelligence} refers to the ability of a chatbot to effectively converse beyond the technical capability of achieving a conversational goal . In this section, we discuss social characteristics related to \\\\textbf{conversational intelligence}, namely: \\\\textit{proactivity} (18 studies), \\\\textit{conscientiousness} (11 studies), and \\\\textit{communicability} (6 studies). Most of the studies rely on data that comes from the log of the conversations, interviews, and questionnaires. The questionnaires are mostly Likert-scales, and some of them include subjective feedback. Most studies analyzed the interaction with real chatbots, although Wizard of Oz (WoZ) settings are also common. Only two papers did not evaluate a particular type of interaction because they were based on a literature review and surveys with chatbot users in general . Only five studies applied only quantitative methods, while seven focused on qualitative methods. The majority of the studies (15) applied mixed methods (both quantitative and qualitative). See the supplementary materials for details.\\n\\\\begin{table}[!btp]\\n\\\\scriptsize\\n\\\\centering\\n\\\\caption{Studied social characteristics per domain. Research in open domain chatbots has reported most of the social characteristics, except for \\\\textit{Communicability}. In he task-oriented domains, some characteristics are largely influenced by the topic (e.g., \\\\textit{Moral agency} and \\\\textit{Personality}), while others are more generally applied (e.g., \\\\textit{Manners} and \\\\textit{Damage control}).}\\n\\\\begin{tabular}{p{1.6cm}p{5cm}p{6.5cm}}\\n\\\\hline\\n\\\\textbf{Domain} & \\\\textbf{Social Characteristics} & \\\\textbf{Studies} \\\\\\\\\\n\\\\hline\\n\\\\multirow{2}{*}{\\\\shortstack[l]{Open\\\\\\\\domain}} & \\\\textit{Proactivity, Conscientiousness, Damage control, Thoroughness, Manners, Moral agency, Emotional intelligence, Personalization, Identity, Personality} & \\\\\\\\ \\n\\\\hline\\nEthnography & \\\\textit{Proactivity, Conscientiousness, Thoroughness, Personalization} & \\\\\\\\ \\n\\\\hline\\n\\\\multirow{2}{*}{\\\\shortstack[l]{Task\\\\\\\\management}} & \\\\textit{Proactivity, Damage control, Manners, Personalization, Identity} & \\\\\\\\ \\n\\\\hline\\nTourism & \\\\textit{Proactivity, Thoroughness, Manners} & \\\\\\\\ \\n\\\\hline\\nBusiness & \\\\textit{Proactivity, Personalization} & \\\\\\\\ \\n\\\\hline\\nInformation search & \\\\textit{Proactivity, Damage control, Manners, Emotional intelligence} & \\\\\\\\ \\n\\\\hline\\nDecision-making & \\\\textit{Proactivity, Damage control, Manners} & \\\\\\\\\\n\\\\hline\\nHealth-care & \\\\textit{Proactivity, Emotional intelligence} & \\\\\\\\ \\n\\\\hline\\n\\\\shortstack{Credibility\\\\\\\\assessment} & \\\\textit{Proactivity, Conscientiousness} & \\\\\\\\ \\n\\\\hline\\nEducation & \\\\textit{Proactivity, Conscientiousness, Damage control, Thoroughness, Manners, Emotional intelligence, Identity, Personality} & \\\\\\\\ \\n\\\\hline\\n\\\\multirow{2}{*}{\\\\shortstack[l]{Financial\\\\\\\\services}} & \\\\textit{Conscientiousness, Communicability, Damage control, Thoroughness, Personalization, Identity} & \\\\\\\\ \\n\\\\hline\\n\\\\multirow{2}{*}{\\\\shortstack[l]{Customer\\\\\\\\services}} & \\\\textit{Conscientiousness, Communicability, Damage control, Thoroughness, Manners, Emotional intelligence, Personalization, Identity} & \\\\\\\\ \\n\\\\hline\\nE-commerce & \\\\textit{Conscientiousness, Manners} & \\\\\\\\ \\n\\\\hline\\nNews & \\\\textit{Communicability} & \\\\\\\\ \\n\\\\hline\\n\\\\multirow{2}{*}{\\\\shortstack[l]{Human\\\\\\\\resources}} & \\\\textit{Communicability, Damage control, Manners, Identity} & \\\\\\\\ \\n\\\\hline\\n\\\\multirow{2}{*}{\\\\shortstack[l]{Virtual\\\\\\\\assistant}} & \\\\textit{Thoroughness, Emotional intelligence, Personalization, Identity} & \\\\\\\\ \\n\\\\hline\\nGaming & \\\\textit{Thoroughness, Emotional intelligence, Personality} & \\\\\\\\ \\n\\\\hline\\nRace-talk & \\\\textit{Moral agency, Identity} & \\\\\\\\ \\n\\\\hline\\nHumorous talk & \\\\textit{Personality} & \\\\\\\\ \\n\\\\hline\\nNot defined & \\\\textit{Proactivity, Conscientiousness, Communicability, Damage control, Personalization, Identity, Personality} & \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\label{tab:domain-analysis}\\n\\\\end{table}\",\n \"id\": \"9ad3c98e-3389-49e5-a57e-7e261302fc58\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 57,\n \"parent_id\": \"5950bb1f-b30c-410b-94c0-7eba5ce91d72\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Conversational Intelligence\"\n ]\n ],\n \"subsections\": [\n \"30b70a69-50cb-4695-9975-c797ccdef4e2\",\n \"8d7c33ab-b84e-4247-bfcd-ac939fd2d82f\",\n \"d1a8a8a3-347c-4cde-8b5a-76ed8060b75f\"\n ],\n \"title\": \"Conversational Intelligence\"\n },\n {\n \"cite_extract_rate\": 0.09090909090909001,\n \"cites\": [\n 6262,\n 7476\n ],\n \"content\": \"\\\\label{sec:proactivity}\\n\\\\textit{Proactivity} is the capability of a system to \\nautonomously act on the user's behalf and thereby reduce the amount of human effort to complete a task . In human-chatbot conversations, a proactive behavior enables a chatbot to share initiative with the user, contributing to the conversation in a more natural way . Chatbots may manifest \\\\textit{proactivity} when they initiate exchanges, suggests new topics, provide additional information, or formulate follow-up questions. In this survey, we found 18 papers that report either chatbots with proactive behavior or implications of manifesting a proactive behavior. \\\\textit{Proactivity} (also addressed as ``\\\\textit{intervention mode}'') was explicitly addressed in seven studies . In most of the studies, however, \\\\textit{proactivity} emerged either as an exploratory result, mostly from post-intervention interviews and user's feedback , or as a strategy to attend to domain-specific requirements (e.g., monitoring, and guidance) . \\\\textit{Proactivity} was mostly investigated in open domain and education chatbots (four studies each). In open domain chatbots, \\\\textit{proactivity} helps improve engagement by introducing new topics to keep the conversation alive . Educational chatbots rely on \\\\textit{proactivity} to prompt students to think, share, and collaborate (e.g., ). \\\\textit{Proactivity} was also observed in eight other task-oriented domains, including task management and information searches as can be observed in the supplementary materials.\\nThe surveyed literature evidences several benefits of chatbot \\\\textit{proactivity} in chatbots: \\n\\\\textbf{[B1] to provide additional, useful information:} literature reveals that \\\\textit{proactivity} in chatbots adds value to interactions . Investigating evaluation criteria for chatbots, asked users of a general purpose chatbot to rate the chatbots' naturalness and report in what areas they excel. Both statistical and qualitative results confirm that taking the lead and suggesting specialized information about the conversation theme correlate to chatbots' naturalness. corroborates this result; in post-intervention interviews, ten out of 14 users mentioned they preferred a chatbot that takes the lead and volunteers additional information, such as useful links and song playlists. In a WoZ study, investigated whether proactive interventions of a chatbot contribute to a collaborative search in a group chat. The chatbot either elicits or infers needed information from the collaborative chat and proactively intervenes in the conversation by sharing useful search results. The intervention modes were not significantly different from each other, but both intervention modes resulted in a statistically significant increase of enjoyment and decrease of effort when compared to the same task with no chatbot interventions. Moreover, in a post-intervention, open-ended question, 16 out of 98 participants self-reported positive perceptions about the provided additional information.\\n\\\\textbf{[B2] to inspire users, and keep the conversation alive: }proactively suggesting and encouraging new topics have been shown useful to both inspire users and keep the conversation alive . Participants in the study conducted by self-reported that the chatbot's suggestions helped them to get started (7 mentions) and gave them ideas about search topics (4 mentions). After iteratively evaluating prototypes for a chatbot in an educational scenario, concluded that proactively initiating topics makes the dialogue more fun and reveals topics the chatbot can talk about. The refined prototype also proactively maintains the engagement by posing a follow-up when the student had not provided an answer to the question. hypothesized that including follow-up questions based on the content of previous messages would result in higher perceived partner engagement. The hypothesis was supported, with participants in the dynamic condition rating the chatbot as more engaging. In an ethnographic data collection , users included photos in their responses to add information about their experience; 85\\\\% of these photos were proactively prompted by the chatbot. This result shows that prompting the user for more information stimulates them to expand their entries. also observed that chatbots' proactive messages provided insights about the chatbots' knowledge, which potentially helped the conversation to continue. In this paper, we call the strategies to convey the chatbot's knowledge and capabilities as \\\\textit{communicability}, and we discuss it in Section \\\\ref{sec:communicability}.\\n\\\\textbf{[B3] to recover the chatbot from a failure: }in and , \\\\textit{proactivity} is employed to naturally recover from a failure. In both studies, the approach was to introduce a new topic when the chatbot failed to understand the user or could not find an answer, preventing the chatbot from getting stuck and keeping the conversation alive. Additionally, in , the chatbot inserted new topics when users are either abusive or non-sensical. We call the strategies to handle failure and abusive behavior as \\\\textit{damage control}, and we discuss this characteristic in Section \\\\ref{sec:damagecontrol}.\\n\\\\textbf{[B4] to improve conversation productivity: }in task-oriented interactions, such as searching or shopping, \\\\textit{proactivity} can improve the conversation's productivity . In interviews with first-time users of chatbots, found that chatbots should ask follow-up questions to resolve and maintain the context of the conversation and reduce the time searching before achieving the goal. found similar results for collaborative searches; 28 out of 98 participants self-reported that chatbot's proactive interventions saved collaborators time.\\n\\\\textbf{[B5] to guide and engage users: }in particular domains, \\\\textit{proactivity} helps chatbots to either guide users or establish and monitor users' goals. In , the chatbot assigns a goal to the user and proactively prompts motivational messages and reminders to keep the user engaged in the treatment. suggest that a decision-making coach chatbot needs to lead the interaction toward guiding the user to a decision. In ethnographic data collection , the chatbot prompts proactive messages that guide the users on what information they need to report. evaluates a chatbot that manage tasks in a workplace. Proactive messages are used to check whether the team member has concluded the tasks, and then report the outcome to the other stakeholders. In the educational context, \\\\textit{proactivity} is used to develop tutors that engage the students and facilitate learning. In , the tutor chatbot was designed to provide examples of how other students explained a topic. The network analysis of the learner's textual inputs shows that students used more key terms and provided more important messages when receiving feedback about other group members. In , , and the chatbots prompt utterances to encourage the students to reason about a topic. In all three studies, the chatbot condition provided better learning outcomes and increased students' engagement in the discussions.\\nThe surveyed papers also highlight challenges in providing proactive interactions, such as timing and relevance, privacy, and the user's perception of being controlled.\\n\\\\textbf{[C1] timing and relevance:} untimely and irrelevant proactive messages may compromise the success of the interaction. states that untimely turn-taking behavior was perceived as annoying, negatively affecting emotional engagement. and reported that \\\\textit{proactivity} can be disruptive. investigated \\\\textit{proactivity} in a workspace environment, hypothesizing that the perceived interruption of agent \\\\textit{proactivity} negatively affects users' opinion's. The hypothesis was supported, and the authors found that influencing the sense of interruption was the general aversion to unsolicited messages, regardless of whether they came from a chatbot or a colleague. showed that proactively introducing new topics resulted in a high number of ignored messages. The analysis of the conversation log reviewed that either the new topics were not relevant, or it was not the proper time to start a new topic. also reported annoyance when a chatbot introduces repetitive topics.\\n\\\\textbf{[C2] privacy:} in a work-related, group chat, observed privacy concerns regarding the chatbot ``reading'' the employees' conversations to proactively act. During a semi-structured interview, researchers presented a mockup of the chatbot to employees from two different enterprises and collected perceptions of usefulness, intrusiveness, and privacy. Employees reported feeling that the chatbot represented their supervisors' interests, which conveyed as sense of workplace surveillance. Privacy concerns may result in under-motivated users, discomfort about disclosing information, and lack of engagement .\\n\\\\textbf{[C3] user's perception of being controlled:} \\\\textit{proactivity} can be annoying when the chatbot conveys the impression of trying to control the user. report that seven out of 13 participants experienced irritation with the chatbot; one of the most frequent reasons was due to the chatbot directing them to specific places. For the task management chatbot, reported to have adapted the follow-up questions approach to pose questions in a time negotiated with the user. In a previous implementation, the chatbot checked the status of the task twice a day, which participants considered too frequent and annoying.\\nThe surveyed literature also reveals two strategies to provide \\\\textit{proactivity}: leveraging the conversational context and randomly selecting a topic. \\\\textbf{[S1] Leveraging the conversational context} is the most frequent strategy , in which proactive messages relate to contextual information provided in the conversation to increase the usefulness of interventions . argue that general purpose, emotionally aware chatbots should recognize users' interests and intents from the conversational context to proactively offer comfort and relevant services. In , the chatbot leverages conversational context to suggest new topics and propose to share documents or links to assist employees in a work-related group chat. The chatbots studied by introduce new topics based on keywords from previous utterances posted in the chat. According to , leveraging the context can also help smoothly guide the user to a target topic. Researchers in the chatbots domain can refer to the emerging literature on Ambient Intelligence (see, e.g., ~) to understand how contextual knowledge can be leveraged to convey proactivity. One surveyed paper proposes a chatbot that \\\\textbf{[S2] selects a topic randomly} but also observes that the lack of context is a major problem of this approach. Contextualized proactive interventions also suggest that the chatbot should be attentive to the conversation, which conveys \\\\textit{conscientiousness}, as discussed in the next section.\",\n \"id\": \"30b70a69-50cb-4695-9975-c797ccdef4e2\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"9ad3c98e-3389-49e5-a57e-7e261302fc58\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Conversational Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Proactivity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Proactivity\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conscientiousness}\\n\\\\textit{Conscientiousness} is a chatbot's capacity to demonstrate attentiveness to the conversation at hand . It enables a chatbot to follow the conversational flow, show understanding of the context, and interpret each utterance as a meaningful part of the whole conversation . In this survey, we found 11 papers that reported findings related to \\\\textit{conscientiousness} for chatbot design. Four studies explicitly investigated influences of \\\\textit{conscientiousness} for chatbots . In the remaining studies, \\\\textit{conscientiousness} emerged in exploratory findings. In , \\\\textit{conscientiousness} emerged from the analysis of conversational logs, while elicited \\\\textit{conscientiousness} aspects as a requirement for chatbot design when surveying the literature on customer service chatbots. In the remaining studies , \\\\textit{conscientiousness} issues were self-reported by the users in post-intervention interviews and subjective feedback to open-ended questions. \\\\textit{Conscientiousness} was investigated in three studies from the education domain, where the tutor chatbot is expected to be attentive and relevant as well as to control the conversation flow to keep the students focused on the topic (see e.g., ). \\\\textit{Conscientiousness} was also investigated in seven other domains, including open domain interactions and customer and financial services. See the supplementary materials for a complete list of domains of studies that report \\\\textit{conscientiousness} as a social characteristic for chatbots.\\nThe surveyed literature evidenced benefits of designing conscientious chatbots:\\n\\\\textbf{[B1] to provide meaningful answers:} some chatbots use simplistic approaches, like pattern matching rules based on keywords or template phrases applied in the last user utterance, to find the most appropriate response . However, as the chatbot does not interpret the meaning and the intent of users' utterances, the best-selected response may still sound irrelevant to the conversation . As shown by , when a chatbot does not interpret the meaning of users' utterances, users show frustration and the chatbot's credibility is compromised. This argument is supported by when studying chatbots to facilitate collaborative learning. The authors proposed a chatbot that promotes Academically Productive Talk moves. Exploratory results show that the chatbot performed inappropriate interventions, which was perceived as a lack of attention to the conversation. In , participants complained that some chatbots seemed ``\\\\textit{completely scripted,}'' ignoring user's inputs that did not fall into the script. In this case, users needed to adapt their inputs to match the chatbot script to be understood, which resulted in dissatisfaction. Besides avoiding frustration, showed that \\\\textit{conscientiousness} also influences chatbots' perceived humanness and social presence. The authors invited participants to interact with a chatbot that shows an image and asks the users to describe it. To perform the task, the chatbot could either ask the same generic follow-up questions each time (nonrelevant condition) or respond with a follow-up question related to the last participant's input (relevant condition), demonstrating attention to the information provided by the user. Statistical analysis of survey rates supported that the relevant condition increased chatbots' perceived humanness and social presence. In , to motivate the students to communicate in a second language, the proposed chatbot interprets users' input to detect breakdowns and react using an appropriate communication strategy. In interviews, participants reported ``\\\\textit{appreciating the help they got from the chatbot to understand and express what they have got to say}'' . This study was later extended to the context of ECAs (see for details).\\n\\\\textbf{[B2] to hold a continuous conversation: }a conversation with a chatbot should maintain a ``\\\\textit{sense of continuity over time}'' to demonstrate that the chatbot is exerting effort to track the conversation. To do so, it is essential to maintain the topic. When evaluating the naturalness of a chatbot, found that maintaining a theme is convincing, while failure to do so is unconvincing. Furthermore, based on the literature on customer support chatbots, argue that comfortably conversing on any topic related to the service offering is a requirement for task-oriented chatbots. reviewed popular chatbots for practicing second languages. The author showed that most chatbots in this field cannot hold continuous conversations, since they are developed to answer the user's last input. Therefore, they did not have the sense of topic, which resulted in instances of inappropriate response. While the chatbots could change the topic, they could not sustain it afterward. Showing \\\\textit{conscientiousness} also requires the chatbot to understand and track the context, which is particularly important in task-oriented scenarios. In , first-time users stressed positive experiences with chatbots that retained information from previous turns. Two participants also expected the chatbots to retain this context across sessions, thus reducing the need for the user's additional input per interaction. Keeping the context across sessions was highlighted as a strategy to convey \\\\textit{personalization} and empathy (see Sections \\\\ref{sec:personalization} and \\\\ref{sec:emotionalintelligence}).\\n\\\\textbf{[B3] to steer the conversation toward a productive direction:} in task-oriented interactions, a chatbot should understand the purpose of the interaction and strive to conduct the conversation toward this goal in an efficient, productive way . show that productivity is the key motivating factor for using chatbots (68\\\\% of the participants mentioned it as the main reason for using chatbots). First-time users in self-reported that interacting with chatbots should be more productive than using websites, phone apps, and search engines. In this sense, compared the user experience when interacting with a chatbot for solving either simple or complex tasks in a financial context. The authors found that, for complex tasks, to keep the conversation on track the user must be aware of the next steps or why something is happening. In the educational context, proposed a dialogue management approach based on communication strategies to enrich a chatbot with the capability to express its meaning when faced with challenges. Statistical results show that the communication strategies, combined with affective backchannel (which is detailed in Section \\\\ref{sec:emotionalintelligence}), are effective in motivating students to communicate and maintain the task flow. Thirty-two participants out of 40 reported that they preferred to interact with a chatbot with these characteristics. Noticeably, the chatbot's attentiveness to the interactional goal may not be evident to the user if the chatbot passively waits for the user to control the interaction. Thus, \\\\textit{conscientiousness} relates to proactive ability, as discussed in Section \\\\ref{sec:proactivity}. \\nNevertheless, challenges in designing conscientious chatbots are also evident in the literature:\\n\\\\textbf{[C1] to handle task complexity: }as the complexity of tasks increases, more turns are required to achieve a goal; hence, more mistakes may be made. This argument was supported by both and , where the complexity of the task compromised the experience and satisfaction in using the chatbot. also highlights that complex tasks require more effort to correct eventual mistakes. Therefore, it is an open challenge to design flexible workflows, where the chatbot recovers from failures and keeps the interaction moving productively toward the goal, despite potential misunderstandings . \\nRecovering from failure \\nis discussed in Section \\\\ref{sec:damagecontrol}.\\n\\\\textbf{[C2] to harden the conversation:} aiming to ensure the conversational structure--and to hide natural language limitations--chatbots are designed to restrict free-text inputs from the user . However, limiting the choices of interaction may convey a lack of attention to the users' inputs. In , one participant mentioned the feeling of ``\\\\textit{going through a form or a fixed menu.}'' According to , participants consider the chatbot's understanding of free-text input as a criterion to determine whether it can be considered a chatbot, since chatbots are supposed to chat. In the ethnographic data collection study, reported that eight out of ten participants described the interaction using pre-set responses as too restrictive, although they fulfilled the purpose of nudging participants to report their activities. Thus, the challenge lies in how to leverage the benefits of suggesting predefined inputs without limiting conversational capabilities.\\n\\\\textbf{[C3] to keep the user aware of the chatbot's context:} a chatbot should provide a way to inform the user of the current context, especially for complex tasks. According to , context can be inferred from explicit user input, or assumed based on data from previous interactions. In both cases, the user and chatbot should be on the same page about the chatbot's contextual state , providing the users the opportunity to clarify possible misunderstandings . highlighted that participants reported negative experience when finding ``mismatch[es] between [a] chatbot's real context and their assumptions of the chatbot context.'' We identified three strategies used to provide understanding to a chatbot:\\n\\\\textbf{[S1] conversation workflow: }designing a conversational blueprint helps to conduct the conversation strictly and productively to the goal . However, argue that the workflow should be flexible enough to handle both multi-turn and one-turn question-answer interactions; it also should be unambiguous, such that users can efficiently achieve their goals. In addition, discuss that the workflow should make it easy to fix mistakes; otherwise, the users need to restart the workflow, which leads to frustration. In , the conversation workflow included communicative strategies to detect a learner's breakdowns and pitfalls. In that study, when the student does not respond, the chatbot uses a comprehension-check question to detect whether the student understood what was said. Then, it reacts to the user's input by adopting one of the proposed communication strategies (e.g., asking for repetition or simplifying the previous sentence). The conversation workflow could also allow the chatbot to be proactive\\n. For example, participants in suggested that proactive follow-up questions would anticipate the resolution of the context, reducing the effort required from the user to achieve the goal.\\n\\\\textbf{[S2] visual elements: }user-interface resources--such as quick replies, cards, and carousels\\n--are used to structure the conversation and reduce issues regarding understanding . Using these resources, the chatbot shows the next possible utterances and conveys the conversational workflow in a step-by-step manner . Visual elements are also used to show the user what the chatbot can (or cannot) do. This is another conversational characteristic, which will be discussed in the \\nSection \\\\ref{sec:communicability}.\\n\\\\textbf{[S3] context window:} to keep the user aware of the chatbot's current context, developed a chatbot for shopping that shows a context window on the side of the conversation. In this window, the user can click on specific attributes and change them to fix inconsistencies. A survey showed that the chatbot outperformed a default chatbot (without the context window) for mental demand and effort constructs. However, when the chatbots are built into third-party apps (e.g., Facebook Messenger), an extra window may not be possible.\\n\\\\textbf{[S4] confirmation messages: }a conversation workflow may include confirmation messages to convey the chatbots' context to the user . In , when trying to block a stolen credit card, a confirmation message is used to verify the given personal data. In , confirmation messages are used as a communicative strategy to check whether the system's understanding about a particular utterance matches what the user actually meant. Balancing the number of confirmation messages (see ) and the right moment to introduce them into the conversation flow is still under-investigated.\\nThe surveyed literature supports that failing to demonstrate understanding about the users' individual utterances, the conversational context, and the interactional goals results in frustration and loss of credibility. However, most of the results are exploratory; a lack of studies investigate the extent to which the provided strategies influence users' behaviors and perceptions. In the field of human-human communication, the cooperative principle has been extensively applied to understand conversational partners' expectations during a conversation. Additionally, recent studies have shown that this principle influences chatbots' perceived humanness . Particularly, the maxims of relation (the ability to be relevant) and manner (the ability to be clear and orderly) can contribute to the design of \\\\textit{conscientious} chatbots. In addition, \\\\textit{conscientiousness} is by itself a \\\\textit{personality} trait; the more \\\\textit{conscientiousness} a chatbot manifests, the more it can be perceived as attentive, organized, and efficient. The relationship between \\\\textit{conscientiousness} and \\\\textit{personality} is highlighted in Section \\\\ref{sec:interrelationship}.\",\n \"id\": \"8d7c33ab-b84e-4247-bfcd-ac939fd2d82f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"9ad3c98e-3389-49e5-a57e-7e261302fc58\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Conversational Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Conscientiousness\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conscientiousness\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:communicability}\\nInteractive software is communicative by its nature, since users achieve their goals \\nby exchanging messages with the system . In this context, \\\\textit{communicability} is defined as the capacity of a software to convey to users its underlying design intent and interactive principles . Providing \\\\textit{communicability} helps users to interpret the codes used by designers to convey the interactional possibilities embedded in the software , which improves system learnability . In the chatbot context, \\\\textit{communicability} is, therefore, the capability of a chatbot to convey its features to users . The problema around chatbots' \\\\textit{communicability} lies in the nature of the interface: instead of buttons, menus, and links, chatbots unveil their capabilities through conversational turns, one sentence at a time , bringing new challenges to the system learnability field. The lack of \\\\textit{communicability} may lead users to give up on using the chatbot when they cannot understand the available functionalities and how to use them .\\nIn this survey, we found six papers that describe \\\\textit{communicability}, although investigating \\\\textit{communicability} is the main purpose of only one . Conversational logs revealed \\\\textit{communicability} needs in three studies ; in the other three studies \\\\textit{communicability} issues were self-reported by the users in post-intervention interviews and subjective feedback. \\\\textit{Communicability} was investigated in the customer services domain (two studies), where chatbots guide customers through available functionalities . The remaining task-oriented domains in which \\\\textit{communicability} was investigated are financial services, news, and human resources. We did not identify studies on open domain chatbots that report on \\\\textit{communicability}, since in open domain interactions users are free to talk about varying topics and guidance is less a concern.\\nThe surveyed literature reports two main benefits of \\\\textit{communicability} for chatbots:\\n\\\\textbf{[B1] to unveil functionalities: }while interacting with chatbots, users may not know that a desired functionality is available or how to use it . Most participants in 's study mentioned that they did not understand the functionalities of at least one of the chatbots and none of them mentioned searching for the functionalities in other sources (e.g., Google search or the chatbot website) rather than exploring options during the interaction. In a study about playful interactions in a work environment, observed that 22\\\\% of the participants explicitly asked the chatbot about its capabilities (e.g., \\\\textit{``what can you do?''}), and 1.8\\\\% of all the users' messages were ability-check questions. In a study about hotel chatbots, verified that 63\\\\% of the conversations were initiated by clicking an option displayed in the chatbot welcome message. A semiotic inspection of news-related chatbots evidenced that \\\\textit{communicability} strategies are effective at providing clues about the chatbot's features and ideas about what to do and how.\\n\\\\textbf{[B2] to manage users' expectations:} observed that when first-time users do not understand chatbots' capabilities and limitations, they have high expectations and, consequently, end up more frustrated when the chatbots fail. Some participants \\nblamed themselves for not knowing how to communicate and gave up. \\nIn , quantitative results \\nevidenced that ability-check questions can be considered signals of users struggling with functional affordances. Users posed ability-check questions after encountering errors as a means of establishing a common ground between the chatbot's capabilities and their own expectations. According to the authors, ability-check questions helped users to understand the system and reduce uncertainty . In , users also demonstrated the importance of understanding chatbots' capabilities in advance. Since the tasks related to financial support, users expected the chatbot to validate the personal data provided and to provide feedback after completing the task (e.g., explaining how long it would take for the credit card to be blocked). Therefore, \\\\textit{communicability} helps users gain a sense of which type of messages or functionalities a chatbot can handle.\\nThe surveyed literature also highlights two challenges of providing \\\\textit{communicability}:\\n\\\\textbf{[C1] to provide business integration: }communicating chatbots' functionalities should be performed as much as possible within the chat interface . Chatbots often act as an intermediary between users and services. In this case, to overcome technical challenges, chatbots answer users' inputs with links to external sources, where the request will be addressed. First-time users expressed dissatisfaction with this strategy in . Six participants complained that the chatbot redirected them to external websites. According to , business integration is a requirement for designing chatbots, so that the chatbot can solve the users' requests without transferring the interaction to another user interface.\\n\\\\textbf{[C2] to keep visual elements consistent with textual inputs:} in the semiotic engineering evaluation, observed that some chatbots responded differently depending on whether the user accesses a visual element in the user-interface or types the desired functionality in the text-input box, even if both input modes result in the same utterance. This inconsistency produces the user's feeling of misinterpreting the affordances, which has a negative impact on the system's learnability.\\nAs an outcome of the semiotic inspection process, present a set of strategies to provide \\\\textit{communicability}. Some of them are also emphasized in other studies, as follows:\\n\\\\textbf{[S1] to clarify the purpose of the chatbot:} First-time users in highlighted that a clarification about the chatbots' purpose should be placed in the introductory message. found similar inference from the literature on customer services chatbots. The authors argue that providing an opening message with insights into the chatbots' capabilities is a requirement for chatbots design. In addition, a chatbot could give a short tour throughout the main functionalities at the beginning of the first sessions .\\n\\\\textbf{[S2] to advertise the functionality and suggest the next step: }when the chatbot is not able to answer the user, or when it notices that the user is silent, it may suggest available features to stimulate the user to engage \\n. In , six participants mentioned that they appreciated when the chatbot suggested responses, for example by saying \\\\textit{``try a few of these commands: ...''} . shows that chatbots use visual elements, such as cards, carousel, and menus (persistent or not) to show contextualized clues about the next answer, which both fulfills \\\\textit{communicability} purpose and spares users from having to type.\\n\\\\textbf{[S3] to provide a help functionality: }chatbots should recognize a \\\\textit{``help''} input from the user, so it can provide instructions on how to proceed . reported that users highlighted this functionality as useful for the reviewed chatbots. Also, results from show that chatbots should be able to answer ability-check questions (e.g., ``\\\\textit{what can you do?}'' or ``\\\\textit{can you do [functionality]?}'').\\nThe literature states the importance of communicating chatbots' functionality to the success of the interaction. Failing to provide \\\\textit{communicability} \\nleads the users to frustration and they often give up when they do not know how to proceed. The literature on interactive systems has highlighted system learnability as the most fundamental component for usability , and an easy-to-learn system should lead the user to perform well, even during their initial interactions. Thus, researchers in the chatbots domain can leverage the vast literature on system learnability to identify metrics and evaluation methodologies, as well as propose new forms of \\\\textit{communicability} strategies that reduce the learnability issues in chatbot interactions. \\\\textit{Communicability} may also be used as a strategy to avoid mistakes (\\\\textit{damage control}), which will be discussed in Section \\\\ref{sec:damagecontrol}.\\n\\\\vspace{-1.5mm}\\n\\\\begin{framed} \\\\small\\n\\\\vspace{-2mm}\\nIn summary, the \\\\textbf{conversational intelligence }category includes characteristics that help a chatbot to perform a proactive, attentive, and informative role in the interaction. Acknowledging that functional aspects play a crucial role in the ability of a chatbot to converse intelligently, the survey focuses on the characteristics that go beyond the functional aspects, looking through the lens of social perceptions. The highlighted benefits relate to how a chatbot manages the conversation to make it productive, interesting, and neat. To achieve that, designers and researchers should care for the timing and relevance of provided information, privacy, interactional flexibility, and consistency.\\n\\\\vspace{-2mm}\\n\\\\end{framed}\\n\\\\vspace{-2mm}\",\n \"id\": \"d1a8a8a3-347c-4cde-8b5a-76ed8060b75f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"9ad3c98e-3389-49e5-a57e-7e261302fc58\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Conversational Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Communicability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Communicability\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:socialintelligence}\\n\\\\textbf{Social Intelligence} refers to the ability of an individual to produce adequate social behavior for the purpose of achieving desired goals . In the HCI domain, the Media Equation theory posits that people react to computers as social actors. Hence, when developing chatbots, it is necessary to account for the socially acceptable protocols for conversational interactions . Chatbots should be able to respond to social cues during the conversation, accept differences, and manage conflicts as well as be empathic and demonstrate caring , which ultimately increase chatbots' authenticity . In this section, we discuss the social characteristics related to \\\\textbf{social intelligence}, namely: \\\\textit{damage control} (12 papers), \\\\textit{thoroughness} (13 papers), \\\\textit{manners} (10 papers), \\\\textit{moral agency} (6 papers),\\\\textit{ emotional intelligence} (14 papers), and \\\\textit{personalization} (11 papers). Although the focus of the investigations is diverse, we found more studies where the focus of the investigation relates to a specific social characteristic, particularly \\\\textit{moral agency} and \\\\textit{emotional intelligence}, when compared to the \\\\textbf{conversational intelligence} category. Regarding the adopted research methods, 12 studies applied qualitative methods and 10 focuses on quantitative methods. 16 studies reported both qualitative and quantitative results.\",\n \"id\": \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"5950bb1f-b30c-410b-94c0-7eba5ce91d72\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Social Intelligence\"\n ]\n ],\n \"subsections\": [\n \"028808f1-c630-47f4-8ddb-e59065ee5178\",\n \"f4e8fa92-1e99-453b-a644-c67f9ee2f76c\",\n \"b7e4439b-1760-43b1-b0dc-0c9d26d8356e\",\n \"f7519054-562c-44e0-9cb2-d8defc3722e3\",\n \"e08fd9d3-0a36-4724-8257-1d5f35b4d5dc\",\n \"69b481db-bee0-4219-a694-6a1c1caa8255\"\n ],\n \"title\": \"Social Intelligence\"\n },\n {\n \"cite_extract_rate\": 0.05,\n \"cites\": [\n 6262\n ],\n \"content\": \"\\\\label{sec:damagecontrol}\\n\\\\textit{Damage control} is the ability of a chatbot to deal with either conflict or failure situations. Although the Media Equation theory argues that humans socially respond to computers as they respond to other people , the literature has shown that interactions with conversational agents are not quite equal to human-human interactions . When talking to a chatbot, humans are more likely to harass , test the agent's capabilities and knowledge , and feel disappointed with mistakes . When a chatbot does not respond appropriately, it may encourage the abusive behavior or disappoint the user , which ultimately leads the conversation to fail . Thus, it is necessary to enrich chatbots with the ability to recover from failures and handle inappropriate talk in a socially acceptable manner .\\nIn this survey, we found 12 studies that discuss \\\\textit{damage control} as a relevant characteristic for chatbots, two of which focus on conflict situations, such as testing and flaming . In the remaining studies , needs for \\\\textit{damage control} emerged from the analysis of conversational logs and users' feedback. \\\\textit{Damage control} was mostly investigated for open domain (two studies) and customer service chatbots (three studies). In open domain interactions, users are free to wander among topics and testing or flaming tends to be more frequent . In the customer services context, the chatbot needs to avoid disappointing the user, as frustration may negatively reflect the business that the agent represents . \\\\textit{Damage control} was also identified in other six domains (see the supplementary materials), such as task management and financial services.\\nThe surveyed literature highlights the following benefits of providing \\\\textit{damage control} in chatbots:\\n\\\\textbf{[B1] to appropriately respond to harassment:} chatbots are more likely to targets of profanity than humans would . When analyzing conversation logs from hotel chatbots, observed that 4\\\\% of the conversations contained vulgar, indecent, and insulting vocabulary, and 2.8\\\\% of all statements were abusive. Qualitative evaluation reveals that the longer the conversations last, the more users are encouraged to go beyond the chatbots main functions. In addition, sexual expressions represented 1.8\\\\% of all statements. A similar number was found in a study with general purpose chatbots . When analyzing a corpus from the Amazon Alexa Prize 2017, the researchers estimated that about 4\\\\% of the conversations included sexually explicit utterances. used utterances from this corpus to harass a set of state-of-art chatbots and analyze the responses. The results show that chatbots respond to harassment in a variety of ways, including nonsensical, negative, and positive responses. However, the authors highlight that the responses should align with the chatbot's goal to avoid encouraging the behavior or reinforcing stereotypes.\\n\\\\textbf{[B2] to deal with testing:} abusive behavior is often used to test chatbots' social reactions . During the evaluation of a virtual guide to the university campus , a participant answered the chatbot's introductory greeting with ``\\\\textit{moron,}'' likely hoping to see how the chatbot would answer. argue that handling this type of testing helps the chatbots to establish limits and resolve social positioning. Other forms of testing were highlighted in , including sending random letters, repeated greetings, laughs and acknowledgments, and posing comments and questions about the chatbot's intellectual capabilities. When analyzing conversations with a task management chatbot, observed that casually testing the chatbots' ``intelligence'' is a manifestation of satisfaction seeking. In , first-time users appreciated when the chatbot successfully performed tasks when the user expected the chatbot to fail, which shows that satisfaction is influenced by the ability to provide a clever response when the user tests the chatbot.\\n\\\\textbf{[B3] to deal with lack of knowledge: }chatbots often fail in a conversation due to lack of either linguistic or world knowledge . \\\\textit{Damage control} enables the chatbot to admit the lack of knowledge or cover up cleverly . When analyzing the log of a task management chatbot, found out that the chatbot failed to answer 10\\\\% of the exchanged messages. The authors suggest that the chatbot should be designed to handle novel scenarios when the current knowledge is not enough to answer the requests. In some task-oriented chatbots, though, the failure may not be caused by a novel scenario, but by an off-topic utterance. In the educational context, observed that students posted off-topic utterances when they did not know what topics they could talk about, which led the chatbot to fail rather than help the users to understand its knowledge. In task-oriented scenarios, the lack of linguistic knowledge may lead the chatbots to get lost in the conversational workflow , compromising the success of the interaction. demonstrated that dialogue-reference errors (e.g., user's attempt to correct a previous answer or jumping back to an earlier question) are one of the major reasons for failed dialogues and they mostly result from chatbots' misunderstandings.\\nThe literature also reveals some challenges to provide \\\\textit{damage control}:\\n\\\\textbf{[C1] to deal with unfriendly users:} argue that users who want to test and find the system's borders are likely to never have a meaningful conversation with the chatbot no matter how sophisticated it is. Thus, there is a limit to which \\\\textit{damage control} strategies will be effective to avoid testing and abuse. In , the authors observed human tendencies to dominate, be rude, and infer stupidity, which they call ``\\\\textit{unfriendly partners.}'' After an intervention where users interacted with a chatbot for decision-making coaching, evaluated participants' self-perceived work and cooperation with the system. The qualitative results show that cooperative users are significantly more likely to give a higher rating for overall evaluation and decision efficiency. The qualitative analysis of the conversation log reveals that a few interactions failed because the users' motivations were curiosity and mischief rather than trying to solve the decision problem.\\n\\\\textbf{[C2] to identify abusive utterances: }several chatbots are trained on ``clean'' data. Because they do not understand profanity or abuse, they may not recognize a statement as harassment, which makes it difficult to adopt answering strategies . shows that data-driven chatbots often provide non-coherent responses to harassment. Sometimes, these responses conveyed the impression of flirtatious or counter-aggression. Providing means to identify an abusive utterance is important to adopting \\\\textit{damage control} strategies.\\n\\\\textbf{[C3] to fit the response to the context:} argue that humans negotiate conflict and social positioning well before reaching abuse. In human-chatbot interactions, however, predicting users' behavior toward the chatbots in a specific context to develop the appropriate behavior is a challenge. \\\\textit{Damage control} strategies need to be adapted to both the social situation and the intensity of the conflict. For example, showed that being evasive about sexual statements may convey the impression of flirtatiousness, which would not be an acceptable behavior for a customer assistant or a tutor chatbot. In contrast, adult chatbots are supposed to flirt, so encouraging behaviors are expected in some situations. argue that when the chatbot is not accepted as part of the social group it represents, it is discredited by the user, leading the interaction to fail. In addition, designing chatbots with too strong of reactions may lead to ethical concerns . For , choosing between peaceful or aggressive reactions in conflict situations is optional for socially intelligent individuals. Enriching chatbots with the ability to choose between the options is a challenge.\\n\\\\textit{Damage control} strategies depend on the type of failure and the target benefit, as following:\\n\\\\textbf{[S1] emotional reactions:} suggest that when faced with abuse, a chatbot could be seen to take offense and respond in kind or act hurt. The authors argue that humans might feel inhibited about hurting the pretended feelings of a machine if the machine is willing to hurt a human's feelings too . If escalating the aggressive behavior is not appropriate, the chatbot could withdraw from the conversation to demonstrate that the user's behavior is not acceptable. In , the authors discuss that users appeared to be uncomfortable and annoyed whenever the chatbot pointed out any defect in the user or reacted to aggression, as this behavior conflicted with the user's perceived power relations. This strategy is also applied in , where abusive behavior may lead the chatbot to stop responding until the student changes the topic. categorized responses from state-of-the-art conversational systems in a pool of emotional reactions, both positive and negative. The reactions include humorous responses, chastising and retaliation, and evasive responses, as well as flirtation and play-along utterances. To provide an emotional reaction, \\\\textit{emotional intelligence} is also required. This category is presented in Section \\\\ref{sec:emotionalintelligence}.\\n\\\\textbf{[S2] authoritative reactions: }when facing testing or abuse, chatbots can communicate consequences \\nor call on the authority of others . In , although the wizard acting as a chatbot was conscientiously working as a campus guide, she answered a bogus caller with \\\\textit{``This is the University of Melbourne. Sorry, how can I help you?''} The authors suggest that the wizard was calling on the authority of the university to handle the conflict, where being part of a recognized institution places the chatbot in a stronger social group. In , when students recurrently harass the chatbot, the chatbot informs the student that further abuse will be reported to the (human) teacher (although the paper does not clarify whether the problem is, in fact, escalated to a human). and also suggest that chatbots could redirect users' problematic requests to a human attendant in order to avoid conflict situations.\\n\\\\textbf{[S3] to ignore the user's utterance and change the topic:} argue that ignoring abuse and testing is not a good strategy because it could encourage more extreme behaviors. It also positions the chatbot as an inferior individual, which is particularly harmful in scenarios where the chatbot should demonstrate a more prominent or authoritative social role (e.g., a tutor). However, this strategy has been found in some studies to handle lack of knowledge. When iteratively developing a chatbot for an educational context, proposed initiating a new topic in one out of four user's utterances that the chatbot did not understand.\\n\\\\textbf{[S4] \\\\textit{conscientiousness} and \\\\textit{communicability}:} successfully implementing \\\\textit{conscientiousness} and \\\\textit{communicability} may prevent errors; hence, strategies to provide these characteristics can also be used for \\\\textit{damage control}. In , when users utter out-of-scope statements, the chatbot could make it clear what topics are appropriate to the situation. For task-oriented scenarios, where the conversation should evolve toward a goal, argue that the chatbot can clarify the purpose of the offered service when facing abusive behavior, bringing the user back to the task. showed that describing chatbot's capabilities after failures in the dialogue was appreciated by first-time users. In situations where the conversational workflow is susceptible to failure, discuss that posting confirmation messages avoids trapping the users in the wrong conversation path. Participants in also suggested back buttons as a strategy to fix mistakes in the workflow. In addition, the exploratory results about the user interface showed that having visual elements such as quick replies prevents errors, since they keep the users aware of what to ask and the chatbot is more likely to know how to respond .\\n\\\\textbf{[S5] to predict users' satisfaction:} chatbots should perceive both explicit and implicit feedback about users' (dis)satisfaction .\\nTo address this challenge, invited participants to send a ``\\\\textit{\\\\#fail}'' statement to express dissatisfaction. The results show that 42.4\\\\% of the users did it at least once, and the number of complaints and flaming for the proposed chatbot was significantly lower than the baseline. However, the amount of implicit feedback was also significant, which advocates for predicting user's satisfaction from the conversation. The most powerful conversational act that predicted user satisfaction in that study was the agent ability-check questions, see discussion in \\\\textit{Communicability} section) and the explicit feedback \\\\textit{\\\\#fail}, although closings and off-topic requests were also significant in predicting frustration. Although these results are promising, more investigation is needed to identify other potential predictors of users' satisfaction in real-time, in order to provide appropriate reactions.\\n\\\\textit{Damage control} strategies have different levels of severity. Deciding what strategy is adequate to the intensity of the conflict is crucial . The strategies can escalate in severity if the conflict is not solved. For example, uses a sequence of clarification, suggesting a new topic, and asking a question about the new topic. In case of abuse, the chatbot refers to an authority after two attempts at changing the topic.\\nAccording to , humans also fail in conversations; they misunderstand what their partner says and do not know things that are assumed common knowledge by others. Hence, it is unlikely that chatbots interactions will evolve to be conflict-free. That said, \\\\textit{damage control} intends to avoid escalating the conflicts and manifesting an unexpected behavior. In this sense, politeness can be used as a strategy to minimize the effect of lack of knowledge (see Section \\\\ref{sec:manners}), managing the conversation despite the possible mistakes. Regarding interpersonal conflicts, the strategies are in line with the theory on human-human communication, which includes non-negotiation, emotional appeal, personal rejection, and emphatic understanding . Further research on \\\\textit{damage control} can evaluate the adoption of human-human strategies in human-chatbot communication.\",\n \"id\": \"028808f1-c630-47f4-8ddb-e59065ee5178\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Social Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Damage control\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Damage control\"\n },\n {\n \"cite_extract_rate\": 0.1,\n \"cites\": [\n 6264,\n 6265\n ],\n \"content\": \"\\\\label{sec:thoroughness}\\n\\\\textit{Thoroughness} is the ability of a chatbot to be precise regarding how it uses language . In traditional user interfaces, user communication takes place using visual affordances, such as buttons, menus, or links. In a conversational interface, language is the main tool to achieve the communicative goal. Thus, chatbots should coherently use language that portrays the expected style . When the chatbot uses inconsistent language, or unexpected patterns of language (e.g., excessive formality), the conversation may sound strange to the user, leading to frustration. \\nWe found 13 papers that report the importance of \\\\textit{thoroughness} in chatbot design, three of which investigate how patterns of language influence users' perceptions and behavior toward the chatbots . and suggest design principles that include concerns about language choices. Log of conversations revealed issues regarding \\\\textit{thoroughness} in two studies . In the remaining papers, \\\\textit{thoroughness} emerged from interviews and users' subjective feedback . \\\\textit{Thoroughness} is mainly reported for open domain (five studies) and customer service chatbots (two studies), where the interactions are expected to be natural and credible to succeed . \\\\textit{Thoroughness} was also reported in another six domains of studies, such as financial services and education (see the supplementary materials).\\nWe found two benefits of providing \\\\textit{thoroughness}:\\n\\\\textbf{[B1] to increase human-likeness:} chatbot utterances are often pre-recorded by the chatbot designer . On the one hand, this approach produces high quality utterances; on the other hand, it reduces flexibility, since the chatbot is not able to adapt the tone of the conversation based on individual users and conversational context. When analyzing interactions with a customer representative chatbot, observed that the chatbot proposed synonyms to keywords, and the repetition of this vocabulary led the users to imitate it. observed a similar tendency to matching language style. The authors compared human-human conversations with human-chatbots conversations regarding language use. They found that people use, indeed, fewer words per message and a more limited vocabulary with chatbots. However, a deeper investigation revealed that the human interlocutors were actually matching the patterns of language use with the chatbot, who sent fewer words per message. When interacting with a chatbot that uses many emojis and letter reduplication , participants reported a draining experience, since the chatbot's energy was too high to match. These outcomes show that adapting the language to the interlocutor is a common behavior for humans, and so chatbots would benefit from manifesting it. In addition to the interlocutor, chatbots should adapt their language use to the context in which they are implemented and adopt appropriate linguistic register . In the customer services domain, state that chatbots are expected to fulfill the role of a human; hence, they should produce language that corresponds to the represented service provider. In the financial scenario , some participants complained about the use of emojis in a situation of urgency (blocking a stolen credit card).\\n\\\\textbf{[B2] to increase believability: }because people associate social qualities with machines , chatbots are deemed sub-standard when users see them ``\\\\textit{acting as a machine}'' . When analyzing chatbots' naturalness, found that the formal grammatical and syntactical abilities of a chatbot are the biggest discriminators between good and poor chatbots (the other factors being \\\\textit{conscientiousness}, \\\\textit{manners}, and \\\\textit{proactivity}). The authors highlight that chatbots should use consistent grammar and spelling. discusses how, even with English as Second Language (ESL) learners, basic grammar errors, such as pronoun confusion, diminish the value of the chatbot. In addition, states that believable chatbots also need to display unique characters through linguistic choices. In this sense, demonstrated that \\\\textit{personality} can be expressed by language patterns. The authors proposed a computational framework to produce utterances to manifest a target \\\\textit{personality}. The utterances were rated by experts in \\\\textit{personality} evaluation and statistically compared against utterances produced by humans who manifest the target \\\\textit{personality}. The outcomes show that a single utterance can manifest a believable \\\\textit{personality} when using the appropriate linguistic form. Participants in described some interactions as ``\\\\textit{robotic}'' if the chatbot repeated the keywords in the answers, reducing the interaction's naturalness. Similarly, in , participants complained about the ``\\\\textit{inflexibility}'' of the pre-defined, handcrafted chatbot's responses and expressed the desire for it to talk ``\\\\textit{more as a person.}''\\nRegarding the challenges, the surveyed literature shows the following:\\n\\\\textbf{[C1] to decide how much to say: } in , some participants described the chatbot's utterances as not having enough detail, or being too generic; however, most of them appreciated finding answers in a sentence rather than in a paragraph. Similarly, argue that simple questions should not be too detailed, while important transactions require more information. In three studies , participants complained about information overload and inefficiency caused by big blocks of texts. Balancing the granularity of information with the sentence length is a challenge to overcome.\\n\\\\textbf{[C2] to be consistent:} chatbots should not combine different language styles. For example, in , most users found it strange that emojis were combined with a certain level of formal contact. \\nWhen analyzing the critical incidents about an open-domain interaction, found that participants criticized chatbots when they used more formal language or unusual vocabulary, since general-purpose chatbots focus on casual interactions.\\nDespite the highlighted benefits, \\nwe did not find strategies to provide \\\\textit{thoroughness}. proposed a rule-based architecture where the language choices consider the agent's \\\\textit{personality}, emotional state, and beliefs about the social relationship among the interlocutors. However, they did not provide evidence of whether the proposed models produced the expected outcome. Although the literature in computational linguistics has proposed algorithms and statistical models to manipulate language style and matching (see e.g., ), to the best of our knowledge, these strategies have not been evaluated in the context of chatbots' social interactions.\\nThis section shows that linguistic choices influence users' perceptions of chatbots. The computer-mediated communication (CMC) field has a vast literature that shows language variation according to the media and its effect on social perceptions (see e.g. ). Additionally, the cooperative principle , particularly the maxim of quantity (the ability to give the appropriate amount of information), provides theoretical basis for the challenge of deciding how much to talk. Regarding adaptation and believability, researchers in sociolinguistic fields have shown that language choices are influenced by personal style, dialect, genre, and register. For chatbots, the results presented in are promising, demonstrating that automatically generated language can manifest recognizable traits. Thus, further research on chatbot's \\\\textit{thoroughness} could leverage CMC and linguistics theories to provide strategies that lead language to accomplish its purpose for a particular interactional context.\",\n \"id\": \"f4e8fa92-1e99-453b-a644-c67f9ee2f76c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Social Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Thoroughness\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Thoroughness\"\n },\n {\n \"cite_extract_rate\": 0.105263157894736,\n \"cites\": [\n 6266,\n 6262\n ],\n \"content\": \"\\\\label{sec:manners}\\n\\\\textit{Manners} refer to the ability of a chatbot to manifest polite behavior and conversational habits . Although individuals with different personalities and from different cultures may have different notions of what is considered polite (see e.g., ), politeness can be more generally applied as rapport management , where interlocutors strive to control the harmony between people in discourse. A chatbot can manifest \\\\textit{manners} by adopting speech acts such as greetings, apologies, and closings ; minimizing impositions , and making interactions more personal . \\\\textit{Manners} potentially reduces the feeling of annoyance and frustration that may lead the interaction to fail .\\nWe identified ten studies that report \\\\textit{manners}, one of which directly investigates this characteristic . In some studies , \\\\textit{manners} were observed in the analysis of conversational logs, where participants talked to the chatbot in polite, human-like ways. Users' feedback and interviews revealed users' expectations regarding chatbots politeness and personal behavior . We identified studies reporting \\\\textit{manners} in nine different domains, with only open domain appearing twice. The list includes education, information search, and task management, among others. See the supplementary materials for the complete list of domains of studies that report \\\\textit{manners} as a social characteristic for chatbots.\\nThe main benefit of providing \\\\textit{manners} is \\\\textbf{[B1] to increase human-likeness}. \\\\textit{Manners} is highlighted in the literature as a way to generate a more natural, convincing interaction in chatbot conversations . In an in-the-wild data collection, observed that 93\\\\% of the participants used polite words (e.g., ``\\\\textit{thanks}'' or ``\\\\textit{please}'') with a task management chatbot at least once, and 20\\\\% always spoke politely to the chatbot. Unfortunately, the chatbot evaluated in that study was not prepared to handle these protocols and ultimately failed to understand. When identifying incidents from their own conversational logs with a chatbot , several participants identified greetings as a human-seeming characteristic. The users also found convincing when the chatbot appropriately reacts to social cues statements, such as \\\\textit{``how are you?''}-types of utterances. Using this result, later suggested that greetings, apologies, social niceties, and introductions are significant constructs to measure chatbots' naturalness. In , the chatbot used exclamation marks at some point and frequently offered sentences available on the website, in a vaguely human-like matter. In the feedback, participants described the chatbot as rude, impolite, and cheeky. \\nThe surveyed literature highlights two challenges to convey \\\\textit{manners}:\\n\\\\textbf{[C1] to deal with face-threatening acts:} Face-Threatening Acts (FTA) are speech acts that threaten, either positive or negatively, the ``face'' of an interlocutor . Politeness strategies in human-human interactions are adopted to counteract the threat when an FTA needs to be performed . In , the authors discuss that the wizard performing the role of the chatbot used several politeness strategies to counteract face threats. For instance, when she did not recognize a destination, instead of providing a list of possible destinations, she stimulated the user to keep talking until they volunteered the information. In chatbot design, by contrast, providing a list of options to choose is a common strategy. For example, in , the chatbot was designed to present the user with a list of pending tasks when it did not know what task the user was reporting as completed, although the authors acknowledged that it resulted in an unnatural interaction. Although adopting politeness strategies is natural for humans and people usually do not consciously think about them, implementing them for chatbots is challenging due to the complexity of identifying face-threatening acts. For example, in the decision-making coach scenario, observed that users tend to utter straightforward and direct agreements while most of the disagreements contained modifiers that weakened their disagreement. The adoption of politeness strategies to deal with face-threatening acts is still under-investigated in the chatbot literature.\\n\\\\textbf{[C2] to end a conversation gracefully:} discuss that first-time users expected human-like conversational etiquette from the chatbots, specifically introductory phrases and concluding phrases. Although several chatbots perform well in the introduction, the concluding phrases are less explored. Most of the participants reported being annoyed with chatbots that do not end a conversation . also highlight that chatbots need to know when the conversation ends. In that scenario, the chatbot could recognize a closing statement (the user explicitly says \\\\textit{``thank you''} or \\\\textit{``bye''}); however, it would not end the conversation otherwise. Users that stated a decision, but kept receiving more information from the chatbot, reported feeling confused and undecided afterward. Thus, recognizing the right moment to end the conversation is a challenge to overcome.\\nThe strategies highlighted in the surveyed literature for providing \\\\textit{manners} are the following:\\n\\\\textbf{[S1] to engage in small talk:} and point out that even task-oriented chatbots engage in small talk. When categorizing the utterances from the conversational log, the authors found a significant number of messages about the agent status (e.g., \\\\textit{``what are you doing?}''), opening and closing sentences as well as acknowledgment statements (\\\\textit{``ok,''} \\\\textit{``got it''}). also observed that first-time users included small talk in the introductory phrases. According to , these are common behaviors in human-human chat interface, and chatbots would likely benefit from anticipating these habitual behaviors and reproducing them. However, particularly for task-oriented chatbots, it is important to control the small talk to avoid off-topic conversations and harassment, as discussed in Sections \\\\ref{sec:damagecontrol} and \\\\ref{sec:conscientiousness}.\\n\\\\textbf{[S2] to adhere to turn-taking protocols:} suggest that chatbots should adopt turn-taking protocols to know when to talk. Participants who received frequent follow-up questions from the task management chatbot about their pending tasks perceived the chatbot as invasive. Literature in chatbot development proposes techniques to improve chatbots' turn-taking capabilities (see e.g., ), which can be explored as a means of improving chatbots' perceived \\\\textit{manners}.\\nAlthough the literature emphasizes that \\\\textit{manners} are important to approximate chatbot interactions to human conversational protocols, this social characteristic is under-investigated in the literature. Conversational acts such as greetings and apologies are often adopted (e.g., ), but there is a lack of studies on the rationality around the strategies with politeness models used in human-human social interactions . In addition, the literature points to needs for personal conversations (e.g., addressing the users by name), but we did not find studies that focus on this type of strategy. CMC is by itself more impersonal than face-to-face conversation ; even so, current online communication media has been successfully used to initiate, develop, and maintain interpersonal relationships . Researchers can learn from human behaviors in CMC and adopt similar strategies to produce more personal conversations.\",\n \"id\": \"b7e4439b-1760-43b1-b0dc-0c9d26d8356e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Social Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Manners\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Manners\"\n },\n {\n \"cite_extract_rate\": 0.066666666666666,\n \"cites\": [\n 7476\n ],\n \"content\": \"\\\\label{sec:moralagency}\\nMachine \\\\textit{moral agency} refers to the ability of a technology to act based on social notions of right and wrong . The lack of this ability may lead to cases such as Tay, Microsoft's Twitter chatbot that became racist, sexist, and harassing in a few hours . The case raised concerns on what makes an artificial agent (im)moral. Whether machines can be considered (moral) agents is widely discussed in the literature (see e.g., ). In this survey, the goal is not to argue about criteria to define a chatbot as moral, but to discuss the benefits of manifesting a perceived agency and the implications of disregarding chatbots' moral behavior. Hence, for the purpose of this survey, \\\\textit{moral agency} is a manifested behavior that may be inferred by a human as morality and agency .\\nWe found six papers that address \\\\textit{moral agency}. developed and validated a metric for perceived \\\\textit{moral agency} in conversational interfaces, including chatbots. In four studies, the authors investigated the ability of chatbots to handle conversations where the persistence of gender and racial stereotypes may occur. In , \\\\textit{moral agency} is discussed as a secondary result, where the authors discuss the impact of biased responses on emotional connection. \\\\textit{Moral agency} was observed in only two domains of studies: open domain (four studies) and race-talk (two studies), which shows that this characteristic is primarily relevant when the conversational topic may raise moral concerns, which ultimately requires ethical behavior from the conversational partners.\\nThe two main reported benefits of manifesting perceived \\\\textit{moral agency} are the following:\\n\\\\textbf{[B1] to avoid stereotyping:} chatbots are often designed with anthropomorphized characteristics (see Section \\\\ref{sec:personification}), including gender, age, and ethnicity identities. Although the chatbot's personification is more evident in embodied conversational agents, text-based chatbots may also be assessed by their social representation, which risks building or reinforcing stereotypes . and argue that chatbots are often developed using language registers and cultural references of the dominant culture. In addition, a static image (or avatar) representing the agent may convey social grouping . When the chatbot is positioned in a minority \\\\textit{identity} group, it exposes the image of that group to judgment and flaming, which is frequent in chatbot interactions . For example, discusses the controversies caused by a chatbot designed to answer questions about Caribbean Aboriginal culture: its representation as a Caribbean Amerindian individual created an unintended context for stereotyping, where users projected the chatbot's behavior as a standard for people from the represented population. Another example is the differences in sexual discourse between male- and female-presenting chatbots. found that female-presenting chatbots are the object of implicit and explicit sexual attention and swear words more often than male-presenting chatbots. show that sex talks with the male chatbot were rarely coercive or violent; his sexual preference was often questioned, though, and he was frequently propositioned by reported male users. In contrast, the female character received violent sexual statements, and was threatened with rape five times in the analyzed corpora. In , when the avatars were presented as black adults, references to race often deteriorated into racist attacks. Manifesting moral agency may, thus, prevent obnoxious user interactions. In addition, moral agency may prevent the chatbot itself from being biased or disrespectful to humans. argue that the lack of context about the world does not redeem the chatbot from the necessity of being respectful with all the social groups.\\n\\\\textbf{[B2] to enrich interpersonal relationships:} in a study on how interlocutors perceive conversational agents' \\\\textit{moral agency}, hypothesized that perceived morality may influence a range of motivations, dynamics, and effects of human-machine interactions. Based on this claim, the authors evaluated whether goodwill, trustworthiness, willingness to engage, and relational certainty in future interactions are constructs to measure perceived \\\\textit{moral agency}. Statistical results showed that all the constructs correlate with morality, which suggests that manifesting \\\\textit{moral agency} can enrich interpersonal relationships with chatbots. Similarly, suggest that to produce interpersonal responses, chatbots should be aware of inappropriate information and avoid generating biased responses.\\nHowever, the surveyed literature also reveals challenges of manifesting \\\\textit{moral agency}:\\n\\\\textbf{[C1] to avoid alienation:} in order to prevent a chatbot from reproducing hate speech or abusive talk, most chatbots are built over ``clean'' data, where specific words are removed from their dictionary . These chatbots have no knowledge of those words and their meaning. Although this strategy is useful to prevent unwanted behavior, it does not manifest agency, but rather alienates the chatbot of the topic. show that the lack of understanding about sex-talk does not prevent the studied chatbot from harsh verbal abuse, or even from being perceived as encouraging such abuse. From , one can notice that the absence of racist specific words did not prevent the chatbot Zo from uttering discriminatory exchanges. As a consequence, manifesting \\\\textit{moral agency} requires a broader understanding of the world, rather than alienation, which is an open challenge.\\n\\\\textbf{[C2] to build unbiased algorithms and training data:} as extensively discussed in , machine learning algorithms and corpus-based language generation are biased toward the available training datasets. Hence, \\\\textit{moral agency} relies on data that is biased in its nature, producing unsatisfactory results from an ethical perspective. In , the authors propose a framework for developing social chatbots. The authors highlight that the core chat module should follow ethical design to generate unbiased, non-discriminative responses, but they do not discuss specific strategies for that. Building unbiased training datasets and learning algorithms that connect the outputs with individual, real-world experiences, therefore, are challenges to overcome.\\nDespite the relevance of \\\\textit{moral agency} to the development of socially intelligent chatbots, we did not find strategies to address the issues. advocate for developing diversity-conscious databases and learning algorithms that account for ethical concerns; however, the paper focuses on outlining the main research branches and calls on the community of designers to adopt new strategies. As discussed in this section, research on perceived \\\\textit{moral agency} is still necessary in order to develop chatbots whose social behavior is inclusive and respectful. In the field of embodied agents, mind perception theory has been investigated as a means to improve interactions through agency and emotion . Future investigations in chatbots interactions could leverage this theory to understand when and the extent to which perceived moral agency improves communication with chatbots.\",\n \"id\": \"f7519054-562c-44e0-9cb2-d8defc3722e3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Social Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Moral agency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Moral agency\"\n },\n {\n \"cite_extract_rate\": 0.11764705882352901,\n \"cites\": [\n 7476,\n 2047\n ],\n \"content\": \"\\\\label{sec:emotionalintelligence}\\n\\\\textit{Emotional intelligence} is a subset of social intelligence that allows an individual to appraise and express feelings, regulate affective reactions, and harness emotions to solve a problem . Although chatbots do not have genuine emotions , there are considerable discussions about the role of manifesting emotional cues in chatbots . An emotionally intelligent chatbot can recognize and influence users' feelings and demonstrate respect, empathy, and understanding, improving the human-chatbot relationship .\\nWe identified 14 studies that report \\\\textit{emotional intelligence}. Unlikely the previously discussed categories, most studies on \\\\textit{emotional intelligence} focused on understanding the effects of chatbots' empathy and emotional self-disclosure . Only three papers highlighted \\\\textit{emotional intelligence} as an exploratory outcome , where needs for \\\\textit{emotional intelligence} emerged from participants' subjective feedback and post-intervention surveys. \\\\textit{Emotional intelligence} is mainly investigated in domains where topics may involve the disclosure of feelings (e.g., in open domain interactions ) and expressions of empathy and understanding are appropriate (e.g., health care, gaming, education). See the supplementary materials for the domains of study investigating \\\\textit{emotional intelligence}.\\nThe main reported benefits of developing emotionally intelligent chatbots are the following:\\n\\\\textbf{[B1] to enrich interpersonal relationships:} the perception that the chatbot understands one's feelings may create a sense of belonging and acceptance . propose that chatbots for second language studies should use congratulatory, encouraging, sympathetic, and reassuring utterances to create a friendly atmosphere to the learner. The authors statistically demonstrated that affective backchannel, combined with communicative strategies (see Section \\\\ref{sec:conscientiousness}) significantly increased learners' confidence and desire to communicate, while reducing anxiety. In another educational study, evaluated the impact of chatbot's affective moves on friendliness and achieving social belonging. Qualitative results show that affective moves significantly improve the perception of amicability, and marginally increased social belonging. According to , when a chatbot's emotional reaction triggers a social response from the user, the chatbot has achieved group membership and users' sympathy. proposed a chatbot that uses empathic and self-oriented emotional expressions to keep users engaged in quiz-style dialogue. The survey results revealed that empathic expressions significantly improved user satisfaction. In addition, the empathic expressions also improved the user ratings of the peer agent regarding intimacy, compassion, amiability, and encouragement. Although did not find effect of chatbot's self-disclosure on emotional connection, found that self-disclosure and reciprocity significantly improved trust and interactional enjoyment. In , seven participants reported that the best thing about their experience with the therapist chatbot was perceived empathy. Five participants highlighted that the chatbot demonstrated attention to their users' feelings. In addition, the users referred the chatbot as \\\\textit{``he,'' ``a friend,''} \\\\textit{``a fun little dude,''} which demonstrates that empathy emerged from the personification of the chatbot. In another mental health care study, found that humans are twice as likely to mirror negative sentiment from a chatbot than from a human, which is a relevant implication for therapeutic interactions. In , participants reported that some content is embarrassing to ask another human, thus, talking to a chatbot would be easier due to the lack of judgement. measured users' experience in conversations with a chatbot compared to a human partner as well as the amount of intimacy disclosure and cognitive reappraisal. Participants in the chatbots condition experienced as many emotional, relational, and psychological benefits as participants who disclosed to a human partner. \\n\\\\textbf{[B2] to increase engagement:} argue that longer conversations (10+ turns) are needed to fulfill the needs of affection and belonging. Therefore, the authors defined conversation-turns per session as a success metric for chatbots, where usefulness and emotional understanding are combined. In , empathic utterances for the quiz-style interaction significantly increase the number of users' messages per hint for both answers and non-answer utterances (such as feedback about the success/failure). This result shows that empathic utterances encouraged the users to engage and utter non-answers statements. compared the possibility of emotional connection between a classical chatbot and a pretend chatbot, simulated in a WoZ experiment. Quantitative results showed that the WoZ condition was more engaging, since it resulted in conversations that lasted longer, with a higher number of turns. The analysis of the conversational logs revealed the positive effect of the chatbot manifesting social cues and empathic signs as well as touching on personal topics.\\n\\\\textbf{[B3] to increase believability:} argue that adapting chatbots' language to their current emotional state, along with their \\\\textit{personality} and social role awareness, results in more believable interactions. The authors propose that conversation acts should reflect the pretended emotional status of the agent; the extent to which the acts impact on emotion depends however on the agent's \\\\textit{personality} (e.g., its temperament or tolerance). \\\\textit{Personality} is an anthropomorphic characteristic and is discussed in Section \\\\ref{sec:personality}.\\nAlthough \\\\textit{emotional intelligence} is the goal of several studies, \\\\textbf{[C1] regulating affective reactions} is still a challenge. The chatbot presented in was designed to mimic the patterns of affective moves in human-human interactions. Nevertheless, the chatbot has shown an only marginally significant increase in social belonging, when compared to the same interaction with a human partner. Conversational logs revealed that the human tutor performed a significantly higher number of affective moves in that context. In , the chatbot was designed to present emotive-like cues, such as exclamation marks, and interjections. The participants negatively rated the degree of emotion in the chatbot's responses. In , the energetic chatbot was reported as having an enthusiasm too high to match. In contrast, the chatbot was described as an \\\\textit{``emotional buddy''} was reported as being \\\\textit{``overly caring.'' } state that chatbot's empathic utterances may be seen as pre-programmed and inauthentic. Although their results revealed that the partners' identity (chatbot vs. person) had no effect in the perceived relational and emotional experience, the chatbot condition was a WoZ setup. The wizards were blind to whether users thought they were talking to a chatbot or a person, which reveals that identity does not matter if the challenge of regulating emotions is overcome.\\nThe chatbots literature also report some strategies to manifest \\\\textit{emotional intelligence}:\\n\\\\textbf{[S1] using social-emotional utterances:} affective utterances toward the user are a common strategy to demonstrate \\\\textit{emotional intelligence}. , , and suggest that affective utterances improve the interpersonal relationship with a tutor chatbot. In , the authors propose affective backchannel utterances (congratulatory, encouraging, sympathetic, and reassuring) to motivate the user to communicate in a second language. The tutor chatbot proposed in uses solidarity, tension release, and agreement utterances to promote its social belonging and acceptance in group chats. propose empathic utterances to express opinion about the difficulty or ease of a quiz, and feedback on success and failure.\\n\\\\textbf{[S2] to manifest \\\\textit{conscientiousness}:} demonstrating \\\\textit{conscientiousness} may affect the emotional connection between humans and chatbots. In , participants reported the rise of affection when the chatbot remembered something they had said before, even if it was just the user's name. Keeping track of the conversation was reported as an empathic behavior and resulted in mutual affection. argue that a chatbot needs to combine usefulness with emotion by asking questions that help to clarify the users' intentions. They provide an example where a user asks the time, and the chatbot answer \\\\textit{``Cannot sleep?''} as an attempt to guide the conversation to a more engaging direction. Adopting this strategy requires the chatbot to handle message understanding, emotion and sentiment tracking, session context modeling, and user profiling .\\n\\\\textbf{[S3] reciprocity and self-disclosure:} hypothesized that a high level of self-disclosure and reciprocity in communication with chatbots would increase trust, intimacy, and enjoyment, ultimately improving user satisfaction and intention to use. They performed a WoZ, where the assumed chatbot was designed to recommend movies. Results demonstrated that reciprocity and self-disclosure are strong predictors of rapport and user satisfaction. In contrast, did not find any effect of self-oriented emotional expressions in the users' satisfaction or engagement. More research is needed to understand the extent to which this strategy produces positive impact on the interaction.\\nThe literature shows that \\\\textit{emotional intelligence} is widely investigated, with particular interest from education and mental health care domains. Using emotional utterances in a personalized, context relevant way is still a challenge. Researchers in chatbots \\\\textit{emotional intelligence} can learn from \\\\textit{emotional intelligence} theory to adapt the chatbots utterances to match the emotions expressed in the dynamic context. Adaption to the dynamic context also improves the sense of personalized interactions, which is discussed in the next section.\",\n \"id\": \"e08fd9d3-0a36-4724-8257-1d5f35b4d5dc\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Social Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Emotional Intelligence\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Emotional Intelligence\"\n },\n {\n \"cite_extract_rate\": 0.1875,\n \"cites\": [\n 6262,\n 650,\n 7476\n ],\n \"content\": \"\\\\label{sec:personalization}\\n\\\\textit{Personalization} refers to the ability of a technology to adapt its functionality, interface, information access, and content to increase its personal relevance to an individual or a category of individuals . In the chatbot domain, \\\\textit{personalization} may increase the agents' social intelligence, since it allows a chatbot to be aware of situational context and dynamically adapt its features to better suit individual needs . Grounded in robots and artificial agents' literature, argue that \\\\textit{personalization} can improve rapport and cooperation, ultimately increasing engagement with chatbots. Although some studies (see e.g., ) also relate \\\\textit{personalization} to the attribution of personal qualities such as \\\\textit{personality}, we discuss personal qualities in the \\\\textbf{Personification} category. In this section, we focus on the ability to adapt the interface, content, and behavior to the users' preferences, needs, and situational context.\\nWe found 11 studies that report \\\\textit{personalization}. Three studies pose \\\\textit{personalization} as a research goal . In most of the studies, though, \\\\textit{personalization} was observed in exploratory findings. In six studies, \\\\textit{personalization} emerged from the analysis of interviews and participants' self-reported feedback . In two studies , needs for \\\\textit{personalization} emerged from the conversational logs. \\\\textit{Personalization} was investigated in seven different domains, including open domain and task management (two studies each). In open domain interactions, personalization is derived from remembering information from previous interactions, such as personal preferences and users' details . In task-oriented contexts, such as task management, personalization aims to increase the relevance of services to particular users . See the supplementary materials for details.\\nThe surveyed literature highlighted three benefits of providing personalized interactions:\\n\\\\textbf{[B1] to enrich interpersonal relationships:} state that personalizing the amount of personal information a chatbot can access and store is required to establish a relation of trust and reciprocity in workplace environments. In , interviews with 12 participants generated a total of 59 statements about how learning from experience promotes chatbots' authenticity. argue that chatbots whose focus is engagement need to personalize the generation of responses for different users' backgrounds, personal interests, and needs in order to serve their needs for communication, affection, and social belonging. In , participants expressed the desire for the chatbot to provide different answers to different users. Although has found no significant effect of \\\\textit{personalization} on the user experience with the financial assistant chatbot, the study applies \\\\textit{personalization} as the ability to provide empathic responses according to the users' issues, where \\\\textit{emotional intelligence} plays a role. Interpersonal relationship can also be enriched by adapting the chatbots' language to match the user's context, energy, and formality; the ability to appropriately use language is discussed in Section \\\\ref{sec:thoroughness}.\\n\\\\textbf{[B2] to provide unique services:} providing \\\\textit{personalization} increases the value of provided information . In the ethnography data collection study , eight participants reported dissatisfaction with the chatbot's generic guidance to specific places. Participants self-reported that the chatbot should use their current location to direct them to places more conveniently located, and ask for participants interests and preferences to direct them to areas that meet their needs. When exploring how teammates used a task-assignment chatbot, found that the use of the chatbot varied depending on the participants' levels of hierarchy. Similarly, qualitative analysis of perceived interruption in a workplace chat suggests that interruption is likely associated with users' general aversion to unsolicited messages at work. Hence, the authors argue that chatbot's messages should be personalized to the user's general preference. also found that users with low social-agent orientation emphasize the utilitarian value of the system, while users with high social-agent orientation see the system as a humanized assistant. This outcome support to the need to personalize the interaction to individual user's mental models. In , participants reported preference for a chatbot that remembers their details, likes and dislikes, and preferences, and voluntarily uses the information to make recommendations. In , two participants also expected chatbots to retain context from previous interactions to improve recommendations.\\n\\\\textbf{[B3] to reduce interactional breakdowns:} in HCI, \\\\textit{personalization} is used to customize the interface toward user familiarity . When evaluating\\nvisual elements (such as quick replies) compared to typing the responses, observed that users that start the interaction by clicking an option are more likely to continue the conversation if the next exchange also has visual elements as optional affordances. In contrast, users who typed are more likely to abandon the conversation when they face options to click. Thus, chatbots should adapt their interface to users' preferred input methods. In , one participant suggested that the choice of text color and font size should be customizable. also observed that participants faced difficulties with small letters, and concluded that adapting the interface to provide accessibility also needs to be considered.\\nAccording to the surveyed literature, the main challenge regarding \\\\textit{personalization} is \\\\textbf{[C1] privacy}. To enrich the efficiency and productivity of the interaction, a chatbot needs to have memory of previous interactions as well as learn user's preferences and disclosed personal information . However, as and suggest, collecting personal data may lead to privacy concerns. Thus, chatbots should showcase transparent purpose and ethical standards . also suggest that there should be a way to inform a chatbot that something in the conversation is private. Similarly, participants in reported that personal data \\nand social media content may be inappropriate topics for chatbots because they can be sensitive\\n. These concerns may be reduced if a chatbot demonstrates care about privacy .\\nThe reported strategies to provide \\\\textit{personalization} in chatbots interactions are the following:\\n\\\\textbf{[S1] to learn from and about the user:} state that chatbots should present strategies to learn from cultural, behavioral, personal, conversational, and contextual interaction data. For example, the authors suggest using Facebook profile information to build knowledge about users' personal information. also suggest that the chatbot should remember user's preferences disclosed in previous conversations. In , the authors propose an architecture where responses are generated based on a \\\\textit{personalization} rank that applies users' feedback about their general interests and preferences. When evaluating the user's experience with a virtual assistant chatbot, found 16 mentions to personalized interactions, where participants demonstrated the need for a chatbot to be aware of their personal quirks and anticipate their needs.\\n\\\\textbf{[S2] to provide customizable agents:} suggest that users should be able to choose the level of the chatbot's attributes, for example, the agent's look and persona. By doing so, users with low social-agent orientation could use a non-humanized interface, which would better represent their initial perspective. This differentiation could be the first signal to personalize further conversation, such as focusing on more productive or playful interactions. Regarding chatbots' learning capabilities, in , interviews with potential users revealed that users should be able to manage what information the chatbot knows about them and decide whether the chatbot can learn from previous interactions or not. If the user prefers a more generic chatbot, then it would not store personal data, potentially increasing the engagement with more resistant users. raise the possibility of offering an ``incognito'' mode for chatbots or asking the chatbot to forget what was said in previous utterances.\\n\\\\textbf{[S3] visual elements:} adopted quick replies as a means for the chatbot to tailor its subsequent questions to the specific experience the participant had reported. As discussed in Section \\\\ref{sec:conscientiousness}, quick replies may be seen as restrictive from an interactional perspective; however, conversation logs showed that the tailored questions prompted the users to report more detail about their experience, which is important in ethnography research. \\nBoth the benefits and strategies identified in the literature are in line with the types of \\\\textit{personalization} proposed by . Therefore, further investigations in \\\\textit{personalization} can leverage knowledge from interactive systems (e.g., ) to adapt \\\\textit{personalization} strategies and manage the privacy concern.\\n\\\\begin{framed} \\\\small \\n\\\\vspace{-2mm}\\nIn summary, the \\\\textbf{social intelligence} category includes characteristics that help a chatbot to manifest an adequate social behavior, by managing conflicts, using appropriate language, displaying \\\\textit{manners} and \\\\textit{moral agency}, sharing emotions, and handling personalized interactions. The benefits relate to resolving social positioning and recovering from failures, as well as increasing believability, human-likeness, engagement, and rapport. To achieve that, designers and researchers should care about privacy, emotional regulation issues, language consistency, and identification of failures and inappropriate content.\\n\\\\vspace{-2mm}\\n\\\\end{framed}\",\n \"id\": \"69b481db-bee0-4219-a694-6a1c1caa8255\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"6e94027f-a398-4b65-8b3b-d03e95d73fa6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Social Intelligence\"\n ],\n [\n \"subsubsection\",\n \"Personalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Personalization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:personification}\\nIn this section, we discuss the influence of \\\\textit{identity} projection on human-chatbot interaction. \\\\textbf{Personification} refers to assigning personal traits to non-human agents, including physical appearance, and emotional states . In the HCI field, researchers argue that using a personified character in the user interface is a natural way to support the interaction . Indeed, the literature shows that (i) users can be induced to behave as if computers were humans, even when they consciously know better ; and (ii) the more human-like a computer representation is, the more social people's responses will be .\\nChatbots are, by definition, designed to have at least one human-like trait: (human) natural language. Although research on \\\\textbf{personification} is more common in the Embodied Conversational Agents field, claim that chatbot embodiment can be created through narrative without any visual help. According to , talking to a machine affords it a new \\\\textit{identity}. In this section, we divided the social characteristics that reflect \\\\textbf{personification} into \\\\textit{identity} (16 papers) and \\\\textit{personality} (12 papers). In this category, we found several studies where part of the main investigation relates to the social characteristics. Six studies applied quantitative methods, whereas the majority reported qualitative (10) or mixed (11) methods. See the supplementary materials for details.\",\n \"id\": \"fb12c3f5-18c5-4fb9-8465-0d7950752ebc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"5950bb1f-b30c-410b-94c0-7eba5ce91d72\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Personification\"\n ]\n ],\n \"subsections\": [\n \"3da905ba-ca45-4a2d-9608-fe81f0daa556\",\n \"eec7d496-87cc-419a-b26e-00faf7d9c1fd\"\n ],\n \"title\": \"Personification\"\n },\n {\n \"cite_extract_rate\": 0.08333333333333301,\n \"cites\": [\n 6262,\n 7476\n ],\n \"content\": \"\\\\label{sec:identity}\\n\\\\textit{Identity} refers to the ability of an individual to demonstrate belonging to a particular social group . Although chatbots lack the agency to choose what social group to which they want to belong, designers attribute \\\\textit{identity} to them, intentionally or not, when they define the way a chatbot talks or behaves . The \\\\textit{identity} of a partner (even if only perceived) gives rise to new processes, expectations, and effects that influence the outcomes of the interaction . Aspects that convey the chatbots' \\\\textit{identity} include gender, age, language style, and name. Additionally, chatbots may have anthropomorphic, zoomorphic, or robotic representations. Some authors include \\\\textit{identity} aspects in the definition of \\\\textit{personality} (see, e.g., ). We distinguish these two characteristics, where \\\\textit{identity} refers to appearance and cultural traits while \\\\textit{personality} focuses on behavioral traits.\\nWe found 16 studies that discuss \\\\textit{identity} issues, ten of which include \\\\textit{identity} as part of their main investigation . In two studies, the authors argue for the impact of \\\\textit{identity} on the interaction based on the literature . In three studies , qualitative analysis of conversational logs revealed that participants put effort into understanding aspects of the chatbots' \\\\textit{identity}. \\\\textit{Identity} concerns were primarily investigated for open domain and customer service chatbots (four studies each). In open domain interactions, \\\\textit{identity} is explored as a means of building common ground . In the case of customer services, \\\\textit{identity} helps manifest credibility and trust . Other domains include race-talk, education, and gaming. The complete list appears in the supplementary materials.\\nThe identified benefits of attributing \\\\textit{identity} to a chatbot are the following:\\n\\\\textbf{[B1] to increase engagement: }when evaluating signals of playful interactions, found that agent-oriented conversations (asking about an agent's traits and status) are consistent with the tendency to anthropomorphize the agent and engage in chit-chat. In the educational scenario, also observed questions about agents' appearance, intellectual capacities, and sexual orientation, although the researchers considered these questions inappropriate for the context of tutoring chatbots. When comparing human-like vs. machine-like language style, greetings, and framing, noticed that using informal language, having a human name, and using greetings associated with human communication resulted in significantly higher scores for adjectives like likeable, friendly, and personal. In addition, framing the agent as ``intelligent'' also had a slight influence on users' scores.\\n\\\\textbf{[B2] to increase human-likeness:} some attributes may convey a perceived human-likeness. showed that using human-like language style, name, and greetings resulted in significantly higher scores for naturalness. The chatbot's framing influenced the outcomes when combined with other anthropomorphic clues. When evaluating different typefaces for a financial adviser chatbot, found that users perceive machine-like typefaces as more chatbot-like, although they did not find strong evidence of handwriting-like typefaces conveying humanness.\\nThe surveyed literature also highlights challenges regarding \\\\textit{identity}:\\n\\\\textbf{[C1] to avoid negative stereotypes: }when engaging in a conversation, interlocutors base their behavior on common ground (joint knowledge, background facts, assumptions, and beliefs that participants have of each other) (see ). Common ground reflects stereotypical attributions that chatbots should be able to manage as the conversation evolves . In , the authors discuss that chatbots for company representations are often personified as attractive human-like women acting as spokespeople for their companies, while men chatbots tend to have a more important position, such as a virtual CEO. state that the agent self-disclosure of gender \\\\textit{identity} opens possibilities to sex talk. The authors observed that the conversations mirror stereotyped male/female encounters, and the ambiguity of the chatbot's gender may influence the exploration of homosexuality. However, fewer instances were observed of sex-talk with the chatbot personified as a robot, which demonstrates that the gender \\\\textit{identity} may lead to the stereotypical attributions. When evaluating the effect of gender identity on disinhibition, showed that people spoke more often about physical appearance and used more swear and sexual words with the female-presenting chatbot, and racist attacks were observed in interactions with black-representing chatbots. The conversation logs from also show instances of users attacking the chatbot persona (a static avatar of a woman pointing to the conversation box). and also highlight that race identity conveys not only the physical appearance, but all the socio-cultural expectations about the represented group (see discussion in Section \\\\ref{sec:moralagency}). Hence, designers should care about the impact of attributing an \\nidentity to chatbots in order to avoid reinforcing negative stereotypes.\\n\\\\textbf{[C2] to balance the \\\\textit{identity} and the technical capabilities:} the literature comparing embodied vs. disembodied conversational agents yields contradictory results regarding the relevance of a human representation. For example, in the context of general-purpose interactions, show that people exert more effort toward establishing common ground with the agent when they are represented as fully human; in contrast, when evaluating a website assistant chatbot, show that simpler text-based chatbots with no visual, human \\\\textit{identity} resulted in less of an uncanny effect and reduced negative affect. Overly humanized agents create a higher expectation for users, which eventually leads to more frustration when the chatbot fails . When arguing about why chatbots fail, advocate for balancing human versus robotic aspects, where \\\\textit{``too human''} representations may lead to off-topic conversations, and overly robotic interactions may lack personal touch and flexibility. When arguing about the social presence conveyed by deceptive chatbots, state that extreme anthropomorphic features may generate cognitive dissonance. The challenge thus lies in designing a chatbot that provides appropriate \\\\textit{identity} cues, which correspond to their capabilities and communicative purpose, in order to convey the right expectation and minimize discomfort from over-personification.\\nRegarding the strategies, the surveyed literature suggests \\\\textbf{[S1] to design and elaborate on a persona}. Chatbots should have a comprehensive persona and answer agent-oriented conversations with a consistent description of itself . For example, discuss that Eliza, the psychotherapist chatbot, and Parry, a paranoid chatbot, have behaviors consistent with the stereotypes associated with the professional and personal identities, respectively. suggest that designers should explicitly build signals of the chatbot personification (either machine- or human-like), so the users can have the right expectation about the interaction. When \\\\textit{identity} aspects are not explicit, users try to establish common ground. In and , much of the small talk with the chatbot related to the chatbot's traits and status. In , the authors observed many instances of Alice's self-references to ``\\\\textit{her}'' artificial nature. These references triggered the users to reflect on their human-condition (self-categorization process), resulting in exchanges about their species (either informational or confrontational). observed similar results, as participants engaged in conversations about the artificial nature of the agent. Providing the chatbot with the ability to describe its personal \\\\textit{identity} helps to establish the common ground, and hence, enrich the interpersonal relationship .\\nChatbots may be designed to deceive users about their actual \\\\textit{identity}, pretending to be a human . In this case, the more human the chatbot sounds, the more successful it will be. In many cases, however, there is no need to engage in deception and the chatbots can be designed to represent an elaborated persona. \\nResearchers can explore social identity theory related to ingroup bias, power relations, homogeneity, and stereotyping, in order to design chatbots with \\\\textit{identity} traits that reflect their expected social position .\",\n \"id\": \"3da905ba-ca45-4a2d-9608-fe81f0daa556\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 24,\n \"parent_id\": \"fb12c3f5-18c5-4fb9-8465-0d7950752ebc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Personification\"\n ],\n [\n \"subsubsection\",\n \"Identity\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Identity\"\n },\n {\n \"cite_extract_rate\": 0.043478260869565,\n \"cites\": [\n 7476\n ],\n \"content\": \"\\\\label{sec:personality}\\n\\\\textit{Personality} refers to personal traits that help to predict someone's thoughts, feelings, and behaviors . The most accepted set of traits is called the Five-Factor model (or Big Five model) ), which describes \\\\textit{personality} across five dimensions (extraversion, agreeableness, conscientiousness, neuroticism, and openness). However, \\\\textit{personality} can also refer to other dynamic, behavioral characteristics, such as temperament and sense of humor . In the chatbots domain, \\\\textit{personality} refers to the set of traits that determines the agent's interaction style, describes its character, and allows the end-user to understand its general behavior . Chatbots with consistent \\\\textit{personality} are more predictable and trustworthy . According to , unpredictable swings in chatbots' attitudes can disorient users and create a strong sense of discomfort. Thus, \\\\textit{personality} ensures that a chatbot displays behaviors that stand in agreement with the users' expectations in a particular context .\\nWe found 12 studies that report \\\\textit{personality} issues for chatbots. In some studies, \\\\textit{personality} was investigated in reference to the Big Five model , while two studies focused on sense of humor . Three studies investigated the impact of the \\\\textit{personality} of tutor chatbots on students' engagement . compared users' preferences regarding pre-defined personalities. In the remaining studies, \\\\textit{personality} concerns emerged from the qualitative analysis of the interviews, users' subjective feedback, and literature reviews . \\\\textit{Personality} was mostly investigated in open domain (five studies) and education (three studies) chatbots. The other two domains were gaming and humorous chatbots, which are both playful agents. Hence, \\\\textit{personality} is relevant when believability and attitude play a role in the interaction (e.g., in open domain) and when the chatbots' attitude may increase users' mental comfort when performing a task , such as in educational contexts.\\nThe surveyed literature revealed two benefits of attributing \\\\textit{personality} to chatbots:\\n\\\\textbf{[B1] to increase believability:} states that chatbots should have a \\\\textit{personality}, defined by the Five Factor model, plus characteristics such as temperament and tolerance, in order to build utterances using linguistic choices that cohere with these attributions. When evaluating a humorous chatbot, compared the naturalness of the chatbot's inputs and the chatbot's human-likeness compared to a non-humorous chatbot. The humorous chatbot scored significantly higher in both constructs. also showed that sense of humor humanizes the interactions, since humor was one of the factors that influenced naturalness for the WoZ condition. demonstrated that manipulating language to manifest a target \\\\textit{personality} produced moderately natural utterances, with a mean rating of 4.59 out of 7 for the \\\\textit{personality} model utterances.\\n\\\\textbf{[B2] to enrich interpersonal relationships:} chatbots \\\\textit{personality} can make the interaction more enjoyable . In , the second most frequent motivation for using chatbots, noted by 20\\\\% of the participants, was entertainment. The authors argue that a chatbot's capacity to be fun is important even when the main purpose is productivity; according to participants, the chatbot's ``\\\\textit{fun tips}'' enrich the user experience. This result is consistent with the experience of first-time users , where participants relate better with chatbots who have consistent \\\\textit{personality}. show how witty banter and casual, enthusiastic conversations help make the interaction effortless. In addition, a few participants enjoyed the chatbot with a caring \\\\textit{personality}, who was described as a good listener. In and , the authors argue that a consistent \\\\textit{personality} helps the chatbot to gain the users' confidence and trust. state that tutor chatbots should display appropriate posture, conduct, and representation, which include being encouraging, expressive, and polite. Accordingly, other studies highlight students' desire for chatbots with positive agreeableness and extraversion . Outcomes consistently suggest that students prefer a chatbot that is not overly polite, but has some attitude. Agreeableness seems to play a critical role, helping the students to feel encouraged and deal with difficulties. Notably, agreeableness requires \\\\textit{emotional intelligence} to be warm and sympathetic in appropriate circumstances (see Section \\\\ref{sec:emotionalintelligence}).\\nThe reviewed literature pointed out two challenges regarding \\\\textit{personality}:\\n\\\\textbf{[C1] to adapt humor to the users' culture:} sense of humor is highly shaped by cultural environment . discusses a Japanese chatbot who uses puns to create funny conversations. The authors state that puns are one of the main humor genres in that culture. However, puns are restricted to a specific culture and language, with low portability. Thus, the design challenge lies in personalizing chatbots' sense of humor to the target users' culture and interests or, alternatively, designing cross-cultural forms of humor. The ability to adapt to the context and users' preference is discussed in Section \\\\ref{sec:personalization}.\\n\\\\textbf{[C2] to balance \\\\textit{personality} traits:} observed that users prefer a proactive, productive, witty chatbot. Yet, they also would like them to be caring, encouraging, and exciting. In , the researchers intentionally generated utterances to reflect extreme personalities; as a result, they observed that some utterances sounded unnatural because a human's \\\\textit{personality} is a continuous phenomenon, rather than a discrete one. also points out that, although \\\\textit{personality} is consistent, moods and states of mind constantly vary. Thus, balancing the predictability of the \\\\textit{personality} and the expected variation is a challenge to overcome.\\nWe also identified strategies to design chatbots that manifest \\\\textit{personality}:\\n\\\\textbf{[S1] to use appropriate language: } and suggest that the chatbot language should be consistently influenced by its \\\\textit{personality}. Both studies propose that chatbots' architecture should include a persona-based model that encodes the \\\\textit{personality} and influences the response generation. proposed a framework to shows that it is possible to automatically manipulate language features to manifest a particular \\\\textit{personality} based on the Big Five model. The Big Five model is a relevant tool because it can be assessed using validated psychological instruments . Using this model to represent the \\\\textit{personality} of chatbots was also suggested by and . discussed that a chatbot's \\\\textit{personality} should match its domain. Participants expected the language used by the news chatbot to be professional, while they expected the shopping chatbot to be casual and humorous. The ability to use consistent language is discussed in Section \\\\ref{sec:thoroughness}.\\n\\\\textbf{[S2] to have a sense of humor:} literature highlights humor as a positive \\\\textit{personality} trait . In , ten participants mentioned enjoyment when the chatbots provided humorous and highly diverse responses. The authors found occurrences of the participants asking for jokes and expressing delight when the request was supported. present similar results when arguing that humor is important even for task-oriented chatbots when the user is usually seeking for productivity. For casual conversations, highlight that timely, relevant, and clever wit is a desired \\\\textit{personality} trait. In , the joker chatbot was perceived as more human-like, knowledgeable, and funny, and participants felt more engaged. \\n\\\\textit{Personality} for artificial agents has been studied for some time in the Artificial Intelligence field . Thus, further investigations on chatbots' \\\\textit{personality} can leverage models for \\\\textit{personality} and evaluate how they contribute to believability and rapport building.\\n\\\\begin{framed}\\\\small\\n\\\\vspace{-2mm}\\nIn summary, the \\\\textbf{personification} category includes characteristics that help a chatbot to manifest personal and behavioral traits. The \\nbenefits relate to increasing believability, human-likeness, engagement, and interpersonal relationship, which is in line with the benefits of \\\\textbf{social intelligence}. However, unlike the \\\\textbf{social intelligence} category, designers and researchers should focus on attributing recognizable \\\\textit{identity} and \\\\textit{personality} traits that are consistent with users' expectations and the chatbot's capabilities. In addition, it is important to care about adaptation to users' cultural context and reducing the effects of negative stereotypes.\\n\\\\vspace{-2mm}\\n\\\\end{framed}\",\n \"id\": \"eec7d496-87cc-419a-b26e-00faf7d9c1fd\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"fb12c3f5-18c5-4fb9-8465-0d7950752ebc\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Chatbots Social Characteristics\"\n ],\n [\n \"subsection\",\n \"Personification\"\n ],\n [\n \"subsubsection\",\n \"Personality\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Personality\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we discuss cross-cutting aspects of the results, namely the influence of the chatbots' perceived humanness, the conversational domains, and the relationships among the characteristics.\",\n \"id\": \"616598c0-a6ef-4e0d-bae2-49c648e87606\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [\n \"212bc0e6-0de8-4c65-af3e-ab937c5ebe0e\",\n \"3bce4038-12ed-4398-a390-053fbed73d7f\",\n \"267f99ec-a3fd-4ef2-ae14-9b0b6a98c88c\"\n ],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 6267,\n 7476\n ],\n \"content\": \"The social characteristics identified in the survey align well with characteristics that are present in human-human interactions. On the one hand, this survey shows the benefits of considering these characteristics when developing chatbots, which is supported by the Media Equation theory ; we identified the domains in which each characteristic was studied. On the other hand, previous studies have also shown that developing overly humanized agents results in high expectations and uncanny feelings , which was also pointed out in the surveyed literature as a challenge to conveying particular characteristics, such as \\\\textit{identity} and \\\\textit{personality}.\\nAn explanation for these contrasts is that interactivity is ``dependent on the identity of the source with whom/which we are carrying out the message exchange''~, i.e., people direct their messages to an artificial agent. In interactions with chatbots, the artificial agent represents the social role usually associated with a human, for example, a tutor~, a healthcare provider~, a salesperson~, a hotel concierge~, or even a friend . The idea of assigning a social role to a chatbot does not imply deceiving people into thinking the software is human. Still, as the chatbots enrich their communication and social skills, their perceived social role may approach human profiles.\",\n \"id\": \"212bc0e6-0de8-4c65-af3e-ab937c5ebe0e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"616598c0-a6ef-4e0d-bae2-49c648e87606\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Discussion\"\n ],\n [\n \"subsection\",\n \"A note on chatbots' humanness\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"A note on chatbots' humanness\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As Table \\\\ref{tab:domain-analysis} depicts\\\\footnote{see more details in the supplementary materials}, almost all social characteristics (10 out of 11) were found in studies related to conversations not restricted to a specific domain (open domain): manners, moral agency, and emotional intelligence are particularly relevant, since many conversations involve personal content (e.g., preferences, personal life, and habits); damage control is necessary since interactions are more susceptible to testing or flaming; identity, personality, and thoroughness are adopted to increase believability and consistency; and conscientiousness, proactivity, and personalization foster user engagement and a coherent conversational flow. The only characteristic that was not investigated in an open domain was communicability. As the conversations in this context mimic human social interactions, conveying the chatbots' capabilities is less of a concern.\\nThe reported characteristics vary across task-oriented domains. Education and customer services literature report the greatest number of characteristics, but this result might be influenced by the maturity of the research in chatbots for these domains. As we pointed out in Section \\\\ref{sec:overview}, education and customer services are the task-oriented domains most reported in the literature. We found conscientiousness, damage control, thoroughness, manners, emotional intelligence, and identity in studies for both domains. However, manners and emotional intelligence have a different goal in these domains. In the education context, these characteristics are designed to encourage students, especially in a situation of failure, in which the chatbot should be comforting and sensitive. This function aligns with other domains, such as healthcare. The education domain also reports needs for personality, so the chatbot can be recognized as either an instructor or a student, and proactivity, so the chatbot can motivate students to participate in the interactions. Personality is also reported in other domains in which the chatbots' character influences the interactions, such as gaming and humorous talk. Proactivity is also consistently reported in domains in which the chatbot provides guidance, such as coaching, health, ethnography, and assessment interviews.\\nIn customer services, emotional intelligence and manners are used as a means to manage customers' frustrations with either a product or service. This function is also reported in other domains, such as virtual assistants and information search. Additionally, communicability is important to convey the services provided by the chatbot, and personalization helps to provide services that align with a particular customer, which is observed in domains such as ethnography, task management, virtual assistants, and financial services. Noticeably, personalization mostly co-occurs with proactivity, since users expects proactive messages to be relevant to their particular interactional context.\\nMoral agency is the only characteristic that does not show up across several domains. It was reported only in race-talk, and open domain. This highlights that moral agency is a concern when the topic involves content that might lead to inappropriate responses. However, as this survey focuses on characteristics reported in the literature, we believe that it might be influenced by the approach adopted by the researchers: it makes sense to study moral agency in the context of gender or race conversations. We argue, though, that this characteristic might be relevant in domains that were not reported here, but that involve interactions that may lead to biased conversations.\",\n \"id\": \"3bce4038-12ed-4398-a390-053fbed73d7f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"616598c0-a6ef-4e0d-bae2-49c648e87606\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Discussion\"\n ],\n [\n \"subsection\",\n \"The importance of conversational context\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"The importance of conversational context\"\n },\n {\n \"cite_extract_rate\": 0.038461538461538006,\n \"cites\": [\n 7476\n ],\n \"content\": \"\\\\label{sec:interrelationship}\\nIn Section \\\\ref{sec:socialcharacteristics}, we organized the social characteristics into discrete groups and discussed several instances of characteristics influencing each other, or being used as a strategy to manifest one another. In this section, we describe these relations in a theoretical framework, depicted in Figure \\\\ref{fig:relationship}. Rather than providing a comprehensive map, the goal here is to make explicit the relationships found in the surveyed chatbots literature, which underlines the complexity of developing chatbots with appropriate social behaviors. In Figure \\\\ref{fig:relationship}, boxes represent the social characteristics and the colors group them into their respective categories. The axes represent the 22 propositions we derived from the literature.\\n\\\\begin{figure}[!hb]\\n \\\\centering\\n \\\\includegraphics[width=8cm]{interrelationship-survey}\\n \\\\caption{Interrelationship among social characteristics}\\n \\\\label{fig:relationship}\\n\\\\end{figure}\\nAccording to the surveyed literature, \\\\textit{proactivity} influences perceived \\\\textit{personality} \\\\textbf{(P1)} , since recommending and initiating topics may manifest higher levels of extraversion. When the proactive messages are based on the context, \\\\textit{proactivity} increases perceived \\\\textit{conscientiousness} \\\\textbf{(P2)} , since the proactive messages may demonstrate attention to the topic. \\\\textit{Proactivity} supports \\\\textit{communicability} \\\\textbf{(P3)} , since a chatbot can proactively communicate its knowledge and provide guidance; in addition, \\\\textit{proactivity} supports \\\\textit{damage control} \\\\textbf{(P4)} , since a chatbot can introduce new topics when the user either is misunderstood, tries to break the system, or sends an inappropriate message.\\n\\\\textit{Conscientiousness} is itself a dimension of the Big Five model; hence, \\\\textit{conscientiousness} influences the perceived \\\\textit{personality} \\\\textbf{(P5)} . Higher levels of context management, goal-orientation, and understanding increase the chatbots' perceived efficiency, organization, and commitment . \\\\textit{Conscientiousness} manifests\\\\textit{ emotional intelligence} \\\\textbf{(P6)} since retaining information from previous turns and being able to recall them demonstrates empathy . In addition, \\\\textit{conscientiousness} manifests \\\\textit{personalization} \\\\textbf{(P7)} because a chatbot can remember individual preferences within and across sessions.\\n\\\\textit{Emotional intelligence} influences perceived \\\\textit{personality} \\\\textbf{(P8)}, since chatbots' \\\\textit{personality} traits affect the intensity of emotional reactions . Agreeableness is demonstrated through consistent warm reactions such as encouraging and motivating \\n. Some \\\\textit{personality} traits \\nrequire \\\\textit{personalization} \\\\textbf{(P9)} to adapt to the interlocutors' culture and interests . Besides, \\\\textit{personalization} benefits \\\\textit{identity} \\\\textbf{(P10)}, since the users' social-agent orientation may require a chatbot to adapt the level of engagement in small talk and the agent's visual representation . \\\\textit{Personalization} also improves \\\\textit{emotional intelligence} \\\\textbf{(P11)}, since a chatbot should dynamically regulate the affective reactions to the interlocutor . \\\\textit{Emotional intelligence} improves perceived \\\\textit{manners} \\\\textbf{(P12)}, since the lack of \\\\textit{emotional intelligence} may lead to the perception of impoliteness .\\n\\\\textit{Conscientiousness} facilitates \\\\textit{damage control} \\\\textbf{(P13)}, since the attention to the \\nworkflow and context may increase the ability to recover from a failure without restarting the workflow . \\\\textit{Communicability} facilitates \\\\textit{damage control} \\\\textbf{(P14)}, since it teaches the user how to communicate, reducing the numbers of mistakes . In addition, suggesting how to interact can reduce frustration after failure scenarios .\\n\\\\textit{Personalization} manifests \\\\textit{thoroughness} \\\\textbf{(P15)} , since chatbots can adapt their language use to the conversational context and the interlocutor's expectations. When the context requires dynamic variation , \\\\textit{thoroughness} may reveal traits of the chatbot's \\\\textit{identity} \\\\textbf{(P16)} . As demonstrated by , \\\\textit{thoroughness} also reveals \\\\textit{personality} \\\\textbf{(P17)}.\\n\\\\textit{Manners} influence \\\\textit{conscientiousness} \\\\textbf{(P18)} , since they can be applied as a strategy to politely refuse off-topic requests and to keep the conversation on track. \\\\textit{Manners} also influence \\\\textit{damage control} \\\\textbf{(P19)} , because they can help a chatbot prevent verbal abuse and reduce the negative effect of lack of knowledge. Both \\\\textit{moral agency} \\\\textbf{(P20)} and \\\\textit{emotional intelligence} \\\\textbf{(P21)} improve \\\\textit{damage control} because they provide\\nthe ability to appropriately respond to abuse and testing . \\\\textit{Identity} influences \\\\textit{moral agency} \\\\textbf{(P22)}, since \\\\textit{identity} representations require the ability to prevent a chatbot from building or reinforcing negative stereotypes .\",\n \"id\": \"267f99ec-a3fd-4ef2-ae14-9b0b6a98c88c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 26,\n \"parent_id\": \"616598c0-a6ef-4e0d-bae2-49c648e87606\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Discussion\"\n ],\n [\n \"subsection\",\n \"Interrelationships among the characteristics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Interrelationships among the characteristics\"\n },\n {\n \"cite_extract_rate\": 0.11111111111111101,\n \"cites\": [\n 8990,\n 6268\n ],\n \"content\": \"\\\\label{sec:relatedsurveys}\\nPrevious studies have reviewed the literature on chatbots. Several surveys discuss chatbots' urgency and their potential application for particular domains, which include education , business , health , information retrieval and e-commerce . Other surveys focus on technical design techniques , such as language generation models, knowledge management, and architectural challenges. Although discuss the social capabilities of chatbots, the survey focuses on the potential of available open source technologies to support these skills, highlighting technical hurdles rather than social ones.\\nWe found three surveys that include insights about social characteristics of chatbots, although none of them focus on this theme. investigated chatbots that ``\\\\textit{mimic conversation rather than understand it},'' and review the main technologies and ideas that support their design, while focuses on identifying best practices for developing script-based chatbots. review the literature on quality issues and attributes for chatbots. The supplementary materials include a table that shows the social characteristics covered by each survey. These related surveys also point out technical characteristics and attributes that are outside the scope of this survey.\",\n \"id\": \"cefcc373-1755-478a-bdc0-b708ace17a7a\",\n \"level\": \"section\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Related Surveys\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Related Surveys\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 6269\n ],\n \"content\": \"\\\\label{sec:limitations}\\nThis research has some limitations. Firstly, since this survey focused on disembodied, text-based chatbots, the literature on embodied and speech-based conversational agents was excluded. We acknowledge that studies that include these attributes can have relevant social characteristics for chatbots, especially for characteristics that could be highly influenced by physical representations, tone, accent, and so forth (e.g. identity, politeness, and thoroughness). However, embodiment and speech could also bring new challenges (e.g., speech-recognition or eye-gazing), which are beyond the scope of this study and could potentially impact users' experiences with chatbots. Additionally, this decision may have caused the under-representation of certain research domains. For example, there is an established research branch on conversational agents in the Ambient Intelligence discipline~ that primarily investigates multimodal interactions, and hence is not represented in this survey. Another example is the Software Engineering domain (see, e.g.,~), where chatbots have been widely investigated, but focused on the functional, rather than social aspects of the interactions.\\nSecondly, since the definition of chatbot is not consolidated in the literature and chatbots have been studied across several different domains, some studies that include social aspects of chatbots may not have been found. To account for that, we adopted several synonyms in our research string and used Google Scholar as search engine, which provides a fairly comprehensive indexing of the literature in more domains. Finally, the conceptual model of social characteristics was derived through a coding process inspired by qualitative methods, such as Grounded Theory. Like any qualitative coding methods, it relies on the researchers' subjective assessment. To mitigate this threat, the researchers discussed the social characteristics and categories during in-person meetings until reaching consensus, and both the conceptual framework and the relationships amongst the characteristics were derived based on outcomes the surveyed studies explicitly reported.\",\n \"id\": \"e47b0d6a-5496-42d1-b93d-9f8d302b471a\",\n \"level\": \"section\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Limitations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Limitations\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclusion}\\nIn this survey, we investigated the literature on disembodied, text-based chatbots to answer the question \\\\textit{``What chatbot social characteristics benefit human interactions and what are the challenges and strategies associated with them?''} Our main contribution is the conceptual model of social characteristics, from which we can derive conclusions about several research opportunities. Firstly, we point out several challenges to overcome in order to design chatbots that manifest each characteristic. Secondly, further research may focus on assessing the extent to which the identified benefits are perceived by users and influence their satisfaction. Finally, further investigations may propose new strategies to manifest particular characteristics. In this sense, we highlight that we could not identify strategies to manifest moral agency and thoroughness, although strategies for several other characteristics are also under-investigated.\\nWe reported domains where social characteristics were primarily investigated. We showed that some characteristics are largely influenced by the domain (e.g., \\\\textit{moral agency} and \\\\textit{communicability}), while others are more generally applied (e.g., \\\\textit{manners} and \\\\textit{damage control}). We also discussed the relationship among the characteristics through 22 propositions derived from the surveyed literature, which underline the complexity of developing chatbots with appropriate social behaviors. Although the propositions and social characteristics do not intend to be comprehensive, they cover important aspects of human-human communication. Social sciences such as sociolinguistics and communication studies have a good deal to contribute to the design of chatbots. In the description of the characteristics, we pointed out theories that could guide these investigations, such as the cooperative principle , social identity , personalization , politeness , and mind perception theories . For example, \\\\textit{conscientiousness}, \\\\textit{thoroughness}, and \\\\textit{proactivity} might be covered by the cooperative principle , and social identity theory might explain the relevance of \\\\textit{identity}, \\\\textit{personality}, and \\\\textit{moral agency} for conversational technologies. Our results provide important references for helping designers and researchers find opportunities to advance the human-chatbot interactions field.\\n\\\\section*{Disclosure statement}\\nNo potential conflict of interest was reported by the authors.\\n\\\\bibliographystyle{apacite}\\n\\\\bibliography{refs}\\n\\\\section*{About the Authors}\\n\\\\textbf{Ana Paula Chaves} is a Ph.D. Candidate at the Northern Arizona University and a Faculty at the Federal University of Technology - Paraná, Campus Campo Mourão, Brazil. She researches social aspects of human-chatbot interactions as well as technologies for tourism. More information at http://anachaves.pro.br\\n\\\\textbf{Marco Aurelio Gerosa} is an Associate Professor at the Northern Arizona University. He researches Software Engineering and CSCW, having served on the program committee of important conferences, such as FSE, CSCW, SANER, and MSR. He is an ACM and IEEE senior member. For more information, visit http://www.marcoagerosa.com\\n\\\\end{document}\",\n \"id\": \"1430d8e4-a5c1-4f7d-ae08-3a9fcec59bb4\",\n \"level\": \"section\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"48e1448c-68e1-4eb0-9088-8c4cfe65bb83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"How should my chatbot interact? A survey on social characteristics in human-chatbot interaction design\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":87,"string":"87"},"all_cites":{"kind":"list like","value":[6262,6263,7476,6264,6265,6266,2047,650,6267,8990,6268,6269],"string":"[\n 6262,\n 6263,\n 7476,\n 6264,\n 6265,\n 6266,\n 2047,\n 650,\n 6267,\n 8990,\n 6268,\n 6269\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.487598140996128,"string":"1.487598"}}},{"rowIdx":54,"cells":{"authors":{"kind":"list like","value":["Petr Chunaev"],"string":"[\n \"Petr Chunaev\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Community detection in node-attributed social networks: a~survey"},"year":{"kind":"string","value":"2019"},"date":{"kind":"timestamp","value":"2019-12-20T13:35:32Z","string":"2019-12-20T13:35:32Z"},"category":{"kind":"string","value":"cs.SI"},"abstract":{"kind":"string","value":" \t\tCommunity detection is a fundamental problem in social network analysis consisting, roughly speaking, in unsupervised dividing social actors (modelled as nodes in a social graph) with certain social connections (modelled as edges in the social graph) into densely knitted and highly related groups with each group well separated from the others. Classical approaches for community detection usually deal only with the structure of the network and ignore features of the nodes (traditionally called node attributes), although the majority of real-world social networks provide additional actors' information such as age, gender, interests, etc. It is believed that the attributes may clarify and enrich the knowledge about the actors and give sense to the detected communities. This belief has motivated the progress in developing community detection methods that use both the structure and the attributes of the network (modelled already via a node-attributed graph) to yield more informative and qualitative community detection results. \t\tDuring the last decade many such methods based on different ideas and techniques have appeared. Although there exist partial overviews of them, a recent survey is a necessity as the growing number of the methods may cause repetitions in methodology and uncertainty in practice. \t\tIn this paper we aim at describing and clarifying the overall situation in the field of community detection in node-attributed social networks. Namely, we perform an exhaustive search of known methods and propose a classification of them based on when and how the structure and the attributes are fused. We not only give a description of each class but also provide general technical ideas behind each method in the class. Furthermore, we pay attention to available information which methods outperform others and which datasets and quality measures are used for their performance evaluation. Basing on the information collected, we make conclusions on the current state of the field and disclose several problems that seem important to be resolved in future. \t"},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"]],"subsections":["fb35c7bc-cc42-4de0-92ed-61876f9f332b","f91b6565-cd82-4d66-b50d-4e9374bb4beb","f7c5bcf3-a055-4452-990c-b315c5512443","124afd9d-8240-4dd5-9e90-68d0d18e9a07","c0718d2b-4a41-48ee-939a-4ab8664cabe6","1c6b434c-2289-4e6e-9627-ef1149463588","2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","a4712ae6-1690-4727-8bbb-b44187674ee8","ec3f6cdd-43b4-49d2-ab2f-b07a25b101ed","7362ce42-549a-4724-9b4c-b963b4574449"],"title":"root"},{"cite_extract_rate":0.5,"cites":[8481,1710],"content":"Community detection is a fundamental problem in social network analysis consisting, roughly speaking, in unsupervised dividing social actors into densely knitted and highly related groups with each group well separated from the others. One class of classical community detection methods mainly deal only with the {\\it structure} of social networks (i.e. with connections between social actors) and ignore actors' features. There exist a variety of such structure-aware methods that have shown their efficiency in multiple applications (see ). However, the majority of real-world social networks provide more information about social actors than just connections between them. Indeed, it is rather common that certain actors' {\\it attributes} such as age, gender, interests, etc., are available. When it is so, the social network is called {\\it node-attributed} (recall that the actors are represented via nodes). According to , attributes form the second dimension, besides the structural one, in social network representation. There is another class of classical community detection methods (being opposite to the structure-aware ones, in a sense) that use only node attributes to detect communities and completely ignore connections between social actors. A representative of the attributes-aware methods is well-known $k$-means clustering algorithm taking attribute vectors as an input. Clearly, methods that deal only with structure or only with attributes do not use all the information available in a node-attributed social network. Naturally, this issue can be overcome if a method would somehow jointly use structure and attributes while detecting communities. Developing of such methods became a novel field in social network analysis . The field is moreover promising as the joint usage is believed to clarify and enrich the knowledge about social actors and to describe the powers that form their communities . \nDuring the last decade numerous methods based on different ideas and techniques have appeared in the field. Although there exist some partial overviews of them, especially in Related Works sections of published papers and in the survey published in 2015, a recent summary of the subject is a necessity as the growing number of the methods may cause repetitions in methodology and uncertainty in practice. \nIn this survey, we aim at describing and clarifying the overall situation in the field. Namely, we perform an exhaustive search of existing community detection methods for node-attributed social networks. What is more, we propose a classification of them based on when and how they use and fuse network structure and attributes. We not only give a description of each class but also provide general technical ideas behind each method in the class. Furthermore, we pay attention to available information which methods outperform others and which datasets and quality measures are used for their performance evaluation. Basing on the information collected, we make conclusions on the current state of the field and disclose several problems that seem important to be resolved in future. \nTo be more precise, let us describe the content of the survey. In Section~\\ref{section2}, we first provide the reader with the notation used in the survey and state the problem of community detection in node-attributed social networks. We further briefly discuss the traditional argumentation in support of such a community detection and the effect of fusing network structure and attributes. In\nSection~\\ref{section3}, we give information about the related survey works and explain how the search of relevant literature was organized in our case. We also indicate which methods are included in the survey and which are not. Additionally, we explain why references throughout the survey are made in a certain way. Section~\\ref{section4} introduces the classification that we propose for community detection methods under consideration. In Section~\\ref{section5}, we discuss the most popular datasets and quality measures for evaluation of community detection results. This section is also helpful for simplifying exposition in the forthcoming sections. Sections~\\ref{section6}--\\ref{section8} contain descriptions of the classes of methods and their representatives. In Section~\\ref{section9}, we analyze the overall situation in the field basing on the information from Sections~\\ref{section6}--\\ref{section8}. Among other things, we disclose several methodological problems that are important to resolve in future studies, in our opinion. Our conclusions on the topic are summarized in Section~\\ref{section10}.","id":"fb35c7bc-cc42-4de0-92ed-61876f9f332b","level":"section","origin_cites_number":4,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section2}","id":"f91b6565-cd82-4d66-b50d-4e9374bb4beb","level":"section","origin_cites_number":0,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes"]],"subsections":["796f6d71-50d7-49c9-9b0b-67ff77b34e9c","c2f50602-97d7-4993-94e6-9400d76e89e5"],"title":"Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes"},{"cite_extract_rate":0,"cites":[],"content":"We represent a node-attributed social network as triple ({\\it node-attributed graph}) $G=(\\mathcal{V},\\mathcal{E},\\mathcal{A})$, where $\\mathcal{V}=\\{v_i\\}$ is the set of nodes (vertices) representing social actors, $\\mathcal{E}= \\{e_{ij}\\}$ the set of edges representing connections between the actors ($e_{ij}$ stands for the edge between nodes $v_i$ and $v_j$), and $\\mathcal{A}$ the set of attribute vectors $A(v_i)=\\{a_{k}(v_i)\\}$ associated with nodes in $\\mathcal{V}$ and containing information about actors' features. Furthermore, $|\\mathcal{V}|=n$, $|\\mathcal{E}|=m$ and the dimension of the attribute vectors is $d$. The domain of $a_k$, i.e. the set of possible values of $k$th element of attribute vectors $a_k(v_i)$, is denoted by $dom(a_k)$. In these terms, $k$th attribute of node $v_i$ is referred to as $a_k(v_i)$. The notation introduced above is summarized in Figure~\\ref{fig:notation}. Note that pairs $(\\mathcal{V},\\mathcal{E})$ and $(\\mathcal{V},\\mathcal{A})$ are correspondingly called the {\\it structure} (or {\\it topology}) and the {\\it attributes} (or {\\it semantics}) of node-attributed graph $G$.\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{notation.pdf}\n\t\\caption{Notation related to the triple (node-attributed graph) $G=(\\mathcal{V},\\mathcal{E},\\mathcal{A})$, where $\\mathcal{V}=\\{v_i\\}$, $\\mathcal{E}=\\{e_{ij}\\}$ and $\\mathcal{A}=\\{A(v_i)\\}$.}\n\t\\label{fig:notation}\n\\end{figure}\nBy {\\it community detection\\footnote{It is also called ``community discovery'' or ``clusterization''.} in node-attributed graph} $G=(\\mathcal{V},\\mathcal{E},\\mathcal{A})$ we mean {\\it unsupervised} partitioning the set of nodes $\\mathcal{V}$ into $N$ subsets ({\\it communities} or {\\it clusters}) $C_k\\subset \\mathcal{V}$, with $C=\\{C_k\\}_{k=1}^N$ called a {\\it partition}, such that\n$\\mathcal{V}=\\bigcup_{k=1}^N C_k$ and a certain balance between the following two properties is achieved:\n\\begin{itemize}\n\t\\item[(a)] {\\it structural closeness}, i.e. nodes within a community are structurally close to each other, while nodes in different communities are not;\n\t\\item[(b)] {\\it attribute homogeneity}, i.e. nodes within a community have homogeneous attributes, while nodes in different communities do not.\n\\end{itemize}\nSince one can meet variations of the above-mentioned definitions in the relevant literature, it is worth giving several comments on them. First, the number of communities $N$ can be either known in advance or determined during the community detection process automatically. Second, the communities $C_k$ may be defined to be disjoint or overlapping. Third, the property $\\mathcal{V}=\\bigcup_{k=1}^N C_k$ is sometimes omitted if the resulting partition is not required to include all the nodes from $\\mathcal{V}$. Fourth, \nthe notion of structural closeness and attribute homogeneity may seem vague at the moment but hopefully become more evident after Subsection~\\ref{subsection-str-closeness-attr-homo} and Section~\\ref{section4} where reasons and particular measures for them are discussed. Fifth, the definitions given are for the case of nodes and edges each of one type. It is of course possible that social actors and connections between them can be of different types in a social network and thus one should take this heterogeneity into account. This situation is however closer to the notion of multi-layer networks that are out of scope of the present survey (see also Section~\\ref{section3}).","id":"796f6d71-50d7-49c9-9b0b-67ff77b34e9c","level":"subsection","origin_cites_number":0,"parent_id":"f91b6565-cd82-4d66-b50d-4e9374bb4beb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes"],["subsection","Necessary notation and the community detection problem statement"]],"subsections":[],"title":"Necessary notation and the community detection problem statement"},{"cite_extract_rate":0.36363636363636304,"cites":[8483,1712,1710,8482,1711,8484,217,9116],"content":"\\label{subsection-str-closeness-attr-homo}\nThe structural closeness requirement is based on the recent concepts of a (structural) community in a social network. For example, communities are thought in as subsets of nodes with dense connections within the subsets and sparse in between. In its turn, adopts the intuition that nodes within the same community should be better connected than they would be by chance. A possible measure for that is famous Newman's Modularity that has become an influential tool for structure-based community detection in social networks . Multiple Modularity modifications and other measures have been also proposed to assess structural closeness . In fact, the precise meaning of structural closeness in each community detection method is determined by the measure chosen.\nThe attribute homogeneity requirement is based on the social science founding (see e.g. ) that actors' features can reflect and affect community structure in social networks. The well-known principle of homophily in social networks states that like-minded social actors have a higher likelihood to be connected . Thus community detection process taking into account attribute homogeneity may provide results of better quality . Oppositely to the situation with structural closeness measures, the attribute homogeneity is usually measured by Entropy that quantifies the degree of disorder of attribute vectors in $(\\mathcal{V},\\mathcal{A})$ within the communities detected.\nLet us now discuss different points of view on the effect of fusing structure and attributes.\nFrom one side, multiple experiments, e.g. in and many other papers cited in this survey, suggest that the structure and the attributes of a node-attributed social network often provide complementary information that improves community detection quality. For example, attributes may compensate the structural sparseness of a real-world social network , while structural information may be helpful in resolving the problem of missing or noisy attributes . What is more, it is observed in that structure-only or attributes-only community detection is often not as effective as when both sources of information are used. From the other side, some experiments (see e.g. ) suggest that this is not always true, and network structure and attributes may be orthogonal and contradictory thus leading to ambiguous community detection results. Moreover, relations between these sources of information may be highly non-linear and challenging to analyze . \nBesides the above-mentioned points, our general impression is that there is no widely accepted opinion on the effect of fusing structure and attributes and how this fusion can influence community detection quality. Let us illustrate this with an example.\nSuppose that the structure of a certain node-attributed social network is ideally correlated with a half of attributes and is wholly orthogonal to another half. For simplicity, let the dimension of attribute vectors be small so that there is no sense to fight against the curse of dimensionality. Now we follow the popular suggestion that the mismatch between structure and attributes negatively affects community detection quality and that the existence of structure-attributes correlation offers ``a unique opportunity to improve the learning performance of various graph mining tasks'' . The choice is clear then: we need to use the structure and only the ideally correlated attributes for community detection. It turns out however that we are going to use two sources of information that mostly duplicate each other. Why should we expect that this improves the quality of detected communities? From our side, we would presume that the structure and the chosen half of attributes (considered separately or jointly) would yield very similar or even the same communities, with all the ensuing consequences for assessing their quality. Shouldn't we use just one of the sources then? Furthermore, it is not clear to us why the other half of attributes should be omitted. Generally speaking, they may contain valuable information for community detection and thus omitting them because of the lack of correlation with the structure is rather questionable.\nIn any case, a focused theoretical study of when the fusion of structure and attributes is worthy and when not for community detection (ideally, in terms of subclasses of node-attributed social networks) seems to be an extremely important thing that would allow to remove the above-mentioned contradictions.","id":"c2f50602-97d7-4993-94e6-9400d76e89e5","level":"subsection","origin_cites_number":22,"parent_id":"f91b6565-cd82-4d66-b50d-4e9374bb4beb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes"],["subsection","Structural closeness and attribute homogeneity. The effect of fusing structure and attributes"]],"subsections":[],"title":"Structural closeness and attribute homogeneity. The effect of fusing structure and attributes"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section3}","id":"f7c5bcf3-a055-4452-990c-b315c5512443","level":"section","origin_cites_number":0,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Related works and processing the relevant literature"]],"subsections":["e751245c-7c36-43ca-8918-65a42b01d47c","b12abaa9-ebec-4b58-8134-5afb1ff9a9f1","f0793b92-de74-4aae-8b54-7257a1189cf5","e38efd05-8a77-49a3-ba2a-eff593b2c129","29a40dbb-fbd7-48bc-96e3-b7914c7e2c83","32b01039-79ed-4c11-816c-f79bd32ba71d"],"title":"Related works and processing the relevant literature"},{"cite_extract_rate":0.833333333333333,"cites":[8481,1713,1714,8485,1710],"content":"There is a variety of surveys and comparative studies considering community detection in social networks without attributes, in particular, . At the same time, the survey \t seems to be the only one on community detection in node-attributed social networks. Obviously, since it was published in 2015, many new methods adapting different fusion techniques have appeared in the field. Furthermore, a big amount of the methods that had been proposed before 2015 are not covered by , in particular, some based on objective function modification, non-negative matrix factorization, probabilistic models, clustering ensembles, etc. In a sense, the technique-based classification of methods in is also sometimes confusing. For example, CODICIL , a method based on assigning attribute-aware weights on graph edges, is not included in \\cite[Section 3.2. Weight modification according to node attributes]{Bothorel2015}, but to \\cite[Section 3.7. Other methods]{Bothorel2015}. Although is a well-written and highly cited paper, a recent survey of community detection methods for node-attributed social networks is clearly required.\nBesides , almost every paper on the topic contains a Related Works section. It typically has a short survey of preceding methods and an attempt to classify them. We observed that many authors are just partly aware of the relevant literature and this sometimes leads to repetitions in approaches. Furthermore, multiple classifications (usually technique-based) are mainly not full and even contradictory.","id":"e751245c-7c36-43ca-8918-65a42b01d47c","level":"subsection","origin_cites_number":6,"parent_id":"f7c5bcf3-a055-4452-990c-b315c5512443","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Related works and processing the relevant literature"],["subsection","Related works"]],"subsections":[],"title":"Related works"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsection 3.2}\nAt the beginning, we started the search of relevant literature using regular and scientific search engines making the queries like ``community detection'' or ``clustering'' or\n``community discovery'' in ``node-attributed social networks'' or ``node-attributed graphs''. Within the search process it became evident that other queries also lead to the relevant literature. In particular, ``clustering an attribute-aware graph'', ``community detection in networks with node attributes'', ``description-oriented community detection'', ``semantic clustering of social networks'', ``structure and attributes community detection'', ``joint cluster analysis of attribute and relationship data'', ``community discovery integrating links and tags'', ``attributed graph partitioning'', ``node attribute-enhanced community detection'', ``community detection based on structure and content'', etc. It can be also observed that node-attributed networks and graphs are also sometimes called ``augmented networks'', ``graphs with feature vectors'', ``feature-rich networks'' and ``multi-attributed graphs''.\nThis variety of terms suggests that there is still no established terminology in the field and emphasizes the significance of our survey, where we try to use consistent terminology.\nAfter the above-mentioned exhaustive search, we learned the references in the found papers. Among other things, it brought us to ideologically close papers devoted, for example, to ``attributed information networks'', ``annotated document networks'', ``multi-layer networks'' and ``subspace-based clustering''. We stopped further search when we could not find any new relevant references. Since this happened in the middle of 2019, the survey covers the found papers that had been published in journals or conference proceedings before this date.","id":"b12abaa9-ebec-4b58-8134-5afb1ff9a9f1","level":"subsection","origin_cites_number":0,"parent_id":"f7c5bcf3-a055-4452-990c-b315c5512443","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Related works and processing the relevant literature"],["subsection","Relevant literature search process"]],"subsections":[],"title":"Relevant literature search process"},{"cite_extract_rate":0,"cites":[],"content":"It turns out that several methods for community detection in node-attributed social networks can be proposed in one paper. Therefore, a regular reference of the form [ReferenceNumber] may be not informative enough. From the other hand, authors usually provide their methods with short names like SA-Cluster, CODICIL or CESNA\\footnote{If this is not the case, we allowed ourselves to invent our own names suggesting the class the method belongs to in our classification.}. Some of the names are rather familiar to researches in the field. Thus it seems reasonable to make a reference of the form MethodName [ReferenceNumber] and so we do in what follows. However, not all the methods that are mentioned in the survey are included in our classification and thus discussed in a more detailed manner (as some are just out of scope). To distinguish the cases, we write names of the methods included in our classification in {\\bf bold}, e.g. {\\bf SCMAG} , {\\bf UNCut} and {\\bf DCM} . Such a format means that the reader can find short descriptions of the methods {\\bf SCMAG} , {\\bf UNCut} and {\\bf DCM} in our survey. References like DB-CSC , FocusCO and ACM mean that the corresponding methods are not included in our classification. In this case the reader is recommended to go directly to the papers , and to get additional information. \nA similar scheme is applied to names of the node-attributed network datasets discussed in Section~\\ref{section5} and used in further classification. The reader is assumed to have in mind that if a dataset name is written in {\\bf bold}, then its description can be found in Tables~\\ref{tab_dataset1}, \\ref{tab_dataset2} or \\ref{tab_dataset3}. Note also that various versions of the datasets from Tables~\\ref{tab_dataset1}, \\ref{tab_dataset2} or \\ref{tab_dataset3} are in fact used in different papers. To show that a dataset is somehow different from the description in Tables~\\ref{tab_dataset1}, \\ref{tab_dataset2} or \\ref{tab_dataset3}, we mark it by *. For example, a DBLP dataset with the number of nodes and edges different from {\\bf DBLP10K} and {\\bf DBLP84K} in Table~\\ref{tab_dataset2} is denoted by {\\bf DBLP*} in Sections~\\ref{section6}--\\ref{section8}.","id":"f0793b92-de74-4aae-8b54-7257a1189cf5","level":"subsection","origin_cites_number":6,"parent_id":"f7c5bcf3-a055-4452-990c-b315c5512443","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Related works and processing the relevant literature"],["subsection","The format of references to methods and datasets"]],"subsections":[],"title":"The format of references to methods and datasets"},{"cite_extract_rate":0,"cites":[],"content":"In the survey, we do not consider community detection methods for multi-layer networks, where different types of vertices and edges may present at different layers . Nevertheless, we mention some of these methods from time to time in corresponding remarks. It is though important to note that node-attributed networks of different nature may be clearly considered as a particular case of multi-layer ones. In the majority of papers covered by the present survey, this connection is however rarely commented. \nLet us also emphasize that multi-layer networks (graphs) require special analysis taking into account the heterogeneity of vertices and edges on different layers. A~separate survey and an extensive comparable study of such methods is an independent and useful task (see partial overviews e.g. in ).","id":"e38efd05-8a77-49a3-ba2a-eff593b2c129","level":"subsection","origin_cites_number":3,"parent_id":"f7c5bcf3-a055-4452-990c-b315c5512443","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Related works and processing the relevant literature"],["subsection","Note on multi-layer network clustering"]],"subsections":[],"title":"Note on multi-layer network clustering"},{"cite_extract_rate":0,"cites":[],"content":"Following the above-mentioned definition of community detection in node-attributed social networks, we mainly consider in the survey the methods that can use the full attribute space and find communities covering the whole network. However, there is a big class of special methods that explore subspaces of attributes and/or deal with subgraphs of an initial graph, e.g. GAMer , DB-CSC , SSCG , FocusCO and ACM .\nThe main idea behind the subspace-based (also known as projection-based) clustering methods is that not all available semantic information (attributes) is relevant to obtain good-quality communities . For this reason, one has to somehow choose the appropriate attribute subspace to avoid the so-called {\\it curse of dimensionality} (see \\cite[Section 3.2]{Bothorel2015}) and reveal significant communities that would not be detected if all available attributes were considered.\nTo be precise, some of the methods that we discuss below partly use this idea, e.g. {\\bf WCru} (cf. the definition of a {\\it point of view} in the papers), {\\bf DVil} , {\\bf SCMAG} , {\\bf UNCut} , {\\bf DCM} , etc., but still can work with the full attribute space. In any case, a separate survey on subspace-based methods for community detection in node-attributed social networks would be a very valuable complement to the current one.","id":"29a40dbb-fbd7-48bc-96e3-b7914c7e2c83","level":"subsection","origin_cites_number":14,"parent_id":"f7c5bcf3-a055-4452-990c-b315c5512443","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Related works and processing the relevant literature"],["subsection","Note on subspace-based clustering"]],"subsections":[],"title":"Note on subspace-based clustering"},{"cite_extract_rate":0.30000000000000004,"cites":[1715,1714,1712],"content":"Clearly, community detection tools for node-attributed social networks are suitable for networks of different nature. That is why, besides obvious applications in marketing (recommender systems, targeted advertisements and user profiling) , the tools are used for search engine\noptimization and spam detection , in counter-terrorist activities and disclosing fraudulent schemes . They are also applied to analysis of protein-protein interactions, genes and epidemics .\nAnother possible application is document network clustering. Note that such a clustering is historically preceding to community detection in node-attributed social networks and is rich methodologically on its own . One should take into account however that social communities although have similar formal description with document clusters have inner and more complicated forces to be formed and act. What is more, it has been shown that methods for community detection in node-attributed social networks outperform preceding methods for document network clustering in many cases, see . For this reason, we do not consider document network clustering methods in this survey.","id":"32b01039-79ed-4c11-816c-f79bd32ba71d","level":"subsection","origin_cites_number":10,"parent_id":"f7c5bcf3-a055-4452-990c-b315c5512443","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Related works and processing the relevant literature"],["subsection","Note on community detection in node-attributed networks of different type and its applications"]],"subsections":[],"title":"Note on community detection in node-attributed networks of different type and its applications"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section4}\nIn previous works, the classification of methods for community detection in node-attributed social networks was done mostly with respect to the techniques used (e.g. distance-based or random walk-based). We partly follow this principle but at a lower level. At the upper level we group the methods by when structure and attributes are fused with respect to the community detection process, see Figure~\\ref{fig:pics_clafficication}. Namely, we distinguish the classes of \n\t\\begin{itemize}\n\t\t\\item {\\it early fusion methods} that fuse structure and attributes before the community detection process,\n\t\t\\item {\\it simultaneous fusion methods} that fuse structure and attributes simultaneously with the community detection process,\n\t\t\\item {\\it late fusion methods} that first partition structure and attributes separately and further fuse the partitions obtained.\n\t\\end{itemize}\n\t\\begin{figure}[b]\n\\begin{minipage}[t]{0.325\\textwidth}\n \\includegraphics[width=\\textwidth]{early.pdf}\n \\centering \n {\\bf Early fusion methods}\\\\ (Section~\\ref{section6})\n \\vspace{0.3cm}\n \\hrule\n \\vspace{0.3cm}\n Weight-based (Section~\\ref{class-ef-weight}) \n \\vspace{0.3cm}\n Distance-based (Section~\\ref{class-ef-distance})\n \\vspace{0.3cm}\n Node-augmented graph-based (Section~\\ref{class-ef-node-augmented-graph}) \n \\vspace{0.3cm}\n Embedding-based (Section~\\ref{class-ef-embedding})\n \\vspace{0.3cm}\n Pattern mining-based (Section~\\ref{class-ef-patterns})\n\\end{minipage}\n\\begin{minipage}[t]{0.325\\textwidth}\n \\includegraphics[width=\\textwidth]{sim.pdf}\n \\centering \n {\\bf Simultaneous fusion methods} (Section~\\ref{section7})\n \\vspace{0.3cm}\n \\hrule\n \\vspace{0.3cm}\n Methods modifying objective functions of classical clustering algorithms (Section~\\ref{class-sf-modifying}) \n \\vspace{0.3cm}\n Metaheuristic-based (Section~\\ref{class-sf-mataheuristic})\n \\vspace{0.3cm}\n Methods based on non-negative matrix factorization and matrix compression (Section~\\ref{class-sf-nnmf}) \n \\vspace{0.3cm}\n Pattern mining-based (Section~\\ref{class-sf-patterns}) \n \\vspace{0.3cm}\n Probabilistic model-based (Section~\\ref{class-sf-probabilistic-models}) \n \\vspace{0.3cm}\n Dynamical system-based and agent-based (Section~\\ref{class-sf-dynamics})\n\\end{minipage}\n\\begin{minipage}[t]{0.325\\textwidth}\n \\includegraphics[width=\\textwidth]{scheme-late.pdf}\n \\centering\n {\\bf Late fusion methods}\n (Section~\\ref{section8})\n \\vspace{0.3cm}\n \\hrule\n \\vspace{0.3cm}\n Consensus-based (Section~\\ref{class-lf-consensus})\n \\vspace{0.3cm}\n Switch-based (Section~\\ref{class-lf-switch})\n\\end{minipage}\n \\caption{The proposed classification and the survey structure guide.}\n \\label{fig:pics_clafficication}\n\\end{figure}\nSuch a classification allows an interested researcher/data scientist to estimate the labour costs of software implementation of the method chosen for use in practice.\nIndeed, early fusion methods require just a preprocessing (fusion) procedure converting information about structure and attributes into a format that is often suitable for classical community detection algorithms. For example, weight-based early fusion methods convert attribute vectors into the graph form and further merge the structure- and attributes-aware graphs into a weighted graph that can be processed by graph clustering algorithms with existing implementations. Implementation of late fusion methods is also rather simple. Namely, at the first step they detects communities by classical graph and vector clustering algorithms applied to structure and attributes separately. At the second step, the communities obtained are somehow merged by an existing (or at least easy-implemented) algorithm. Oppositely to the early and late fusion methods, simultaneous fusion ones require more programmer work as usually assume either a completely new implementation or essential modifications to existing ones. The situation is aggravated by the fact that their source codes are rarely available.\nAs seen from Figure~\\ref{fig:pics_clafficication}, we also divide the methods within each class into technique-used subclasses. Let us emphasize that the priority in this division is the fusion technique. For example, by ``weight-based methods'' we mean those which form a weighted graph while fusing structure and attributes. The majority of such methods further use graph clustering algorithms for community detection (and this is reasonable). However, some may still transform the graph into the distance matrix form and then use distance-based clustering algorithms. Such methods are still called ``weight-based''. One more example: ``distance-based methods'' are called in this way as directly produce a distance matrix while fusing structure and attributes, independently on how this matrix is further processed.\nWhat is more, we provide short descriptions of methods in each class and subclass in tables in Sections~\\ref{section6}--\\ref{section8}. In particular, we briefly describe the community detection algorithm and its input used in the method, and the type of communities obtained (overlapping or not). Furthermore, we mention which datasets and quality measured are used by method's authors for evaluation of community detection results. In addition, for each method we provide a list of other methods for {\\it community detection in node-attributed networks} (using both structure and attributes) that the method under consideration is compared with. Note that the list may be empty sometimes. This is so, for example, if the method under consideration is compared only with classical community detection methods that deal either with structure or attributes and do not fuse them to detect communities.","id":"124afd9d-8240-4dd5-9e90-68d0d18e9a07","level":"section","origin_cites_number":0,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Our classification of community detection methods for~node-attributed~ social~networks"]],"subsections":[],"title":"Our classification of community detection methods for~node-attributed~ social~networks"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section5}","id":"c0718d2b-4a41-48ee-939a-4ab8664cabe6","level":"section","origin_cites_number":0,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Most used node-attributed network datasets and quality measures for community detection evaluation"]],"subsections":["36da5b81-971c-4135-9405-fdc43b1406db","01bdf590-8667-4b9e-b2aa-932d3480f52b"],"title":"Most used node-attributed network datasets and quality measures for community detection evaluation"},{"cite_extract_rate":0.25,"cites":[1714,8486,1716],"content":"The title of the survey suggests that we are focused on community detection in node-attributed {\\it social} networks. However, the methods that are included in our classification, generally speaking, may be applied to node-attributed networks of different nature. As we have noticed, authors of the methods implicitly share this point of view and freely use various node-attributed network datasets to evaluate community detection quality.\nIn this subsection (Tables~\\ref{tab_dataset1}, \\ref{tab_dataset2} or \\ref{tab_dataset3}) we collect and briefly describe the datasets that are popular in the field\\footnote{An interested reader can find more node-attributed network datasets at \\href{http://www-personal.umich.edu/~mejn/netdata/}{Mark Newman page}, \\href{https://hpi.de/naumann/projects/repeatability/datasets/}{HPI Information Systems Group}, \\href{https://linqs.soe.ucsc.edu/data}{LINQS Statistical Relational Learning Group}, \\href{http://snap.stanford.edu/data/}{Stanford Large Network Dataset Collection}, \\href{http://dp.univr.it/~laudanna/LCTST/downloads/index.html}{University of Verona Laboratory of Cell Trafficking and Signal Transduction}, \\href{https://perso.liris.cnrs.fr/marc.plantevit/doku/doku.php?id=data_sets}{Marc Plantevit page}, \\href{https://toreopsahl.com/datasets/}{Tore Opsahl page}, \\href{https://sites.google.com/site/ucinetsoftware/datasets}{UCINET networks}, \\href{http://networkrepository.com}{Interactive Scientific Network Data Repository}, \\href{https://aminer.org/citation}{Citation Network Dataset}.\n}. Recall that the dataset names written in {\\bf bold} in Sections~\\ref{section6}--\\ref{section8} refer to the tables in this subsection. \nIt can be observed that Tables~\\ref{tab_dataset1}, \\ref{tab_dataset2} or \\ref{tab_dataset3} contain datasets based on social network data (e.g. Facebook, LastFM and Twitter) and document or citation network data (e.g. DBLP, Wiki and Patents). \nFor convenience, we distinguish the datasets by size. Namely, by {\\it small}, {\\it medium} and {\\it large} we mean network datasets with $< 10^3$, $10^3\\ldots 10^5$ and $>10^5$ nodes, correspondingly.\n\\begin{table}\n\t\\caption{Most popular small size datasets.}\n\t\\label{tab_dataset1} \n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.2cm}|p{12.9cm}|p{0.8cm}|} \n\t\t\t\\hline \n\t\t\tDataset & Description & Source\\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Political Books} & All books in this dataset were\n\t\t\tabout U.S. politics published during the 2004 presidential election\n\t\t\tand sold by Amazon.com. Edges between books means\n\t\t\ttwo books are always bought together by customers. Each\n\t\t\tbook has only one attribute termed as political persuasion, with\n\t\t\tthree values: 1) conservative; 2) liberal; and 3) neutrality & \\href{http://www-personal.umich.edu/~mejn/netdata/}{Link} \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf WebKB} & A classified network of 877 webpages (nodes) and 1608 hyperlinks (edges) gathered from four different universities Web sites (Cornell, Texas, Washington, and Wisconsin). Each web page is associated with a binary vector, whose elements take the value $1$ if the corresponding word from the vocabulary is present in that webpage, and $0$ otherwise. The vocabulary consists of 1703 unique words. Nodes are classified into five classes: course, faculty, student, project, or staff. & \\href{https://linqs.soe.ucsc.edu/data}{Link} \\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Twitter} & A collection of several tweet networks: 1) Politics-UK dataset is collected from Twitter accounts of 419 Members of Parliament in the United Kingdom in 2012. Each user has 3614-dimensional attributes, including a list of words repeated more than 500 times in their tweets. The accounts are assigned to five\n\t\t\tdisjoint communities according to their political affiliation.\n\t\t\t2) Politics-IE dataset is collected from 348 Irish\n\t\t\tpoliticians and political organizations, each user has 1047-\n\t\t\tdimensional attributes. The users are distributed into seven\n\t\t\tcommunities. 3) Football dataset contains 248 English\n\t\t\tPremier League football players active on Twitter which are\n\t\t\tassigned to 20 disjoint communities, each corresponding to\n\t\t\ta Premier League club. 4) Olympics dataset contains users\n\t\t\tof 464 athletes and organizations involved in the London\n\t\t\t2012 Summer Olympics. The users are grouped into 28\n\t\t\tdisjoint communities, corresponding to different Olympic\n\t\t\tsports. & \\href{http://mlg.ucd.ie/aggregation/}{Link 1} \\newline \\href{http://mlg.ucd.ie/networks/}{Link 2}\\newline \\\\ \n\t\t\t\\hline \n\t\t{\\bf Lazega} & A corporate law partnership in a Northeastern US corporate law firm; possible attributes: (1: partner; 2: associate), office (1: Boston; 2: Hartford; 3: Providence); 71 nodes and 575 edges & \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Research} & A research team of employees in a manufacturing company; possible attributes: location (1: Paris; 2: Frankfurt; 3: Warsaw; 4: Geneva), tenure (1: 1--12 months; 2: 13--36 months; 3: 37--60 months; 4: 61+ months); 77 nodes and 2228 edges & \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Consult} & Network showing relationships between employees in a consulting company; possible attributes: organisational level (1: Research Assistant; 2: Junior Consultant; 3: Senior Consultant; 4: Managing Consultant; 5: Partner), gender (1: male; 2: female); 46 nodes and 879 edges & \\\\ \n\t\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{table} \n\\begin{table}\n\t\\caption{Most popular medium size datasets.}\n\t\\label{tab_dataset2} \n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.2cm}|p{12.9cm}|p{0.8cm}|}\n\t\t\t\\hline \n\t\t\tDataset & Description & Source \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Political Blogs} & A non-classified network of 1,490 webblogs (nodes) on US politics with\n\t\t\t19,090 hyperlinks (edges) between the webblogs. Each node has an attribute describing its political leaning as either liberal or\n\t\t\tconservative (represented by $0$ and $1$). & \\href{http://www-personal.umich.edu/~mejn/netdata/}{Link}\\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf DBLP10K} & A non-classified co-author network extracted from DBLP Bibliography (four research areas of database, data mining, information retrieval and artificial intelligence) with 10,000 authors (nodes) and their co-author relationships (edges). Each author is associated with two relevant categorical attributes: prolific and primary topic. For attribute ``prolific'', authors with $\\ge 20$ papers are labelled as highly prolific; authors with $>10$ and $<20$ papers are labelled as prolific and authors with $\\le 10$ papers are labelled as low prolific. Node-attribute values for ``primary topic'' (100 research topics) are obtained via topic modelling. Each extracted topic consists of a probability distribution of keywords which are most representative of the topic. & \\href{https://github.com/Issamfalih/ANCL/tree/master/data}{Link}\\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf DBLP84K} & A larger non-classified co-author network extracted from DBLP Bibliography (15 research areas of database, data mining, information retrieval, artificial intelligence, machine learning, computer vision, networking, multimedia, computer systems, simulation, theory, architecture, natural language processing, human-computer interaction, and programming language) with 84,170 authors (nodes) and their co-author relationships (edges). Each author is associated with two relevant categorical attributes: prolific and primary topic, defined in a similar way as in DBLP10. & \\href{https://github.com/Issamfalih/ANCL/tree/master/data}{Link}\\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Cora} & A classified network of machine learning papers with 2,708 papers (nodes) and 5,429 citations (edges). Each node is attributed with a $1433$-dimension binary vector indicating the absence/presence of words from the dictionary of words collected from the corpus of papers. The papers are classified into 7 subcategories: case-based reasoning,genetic algorithms,\n\t\t\tneural networks, probabilistic methods, reinforcement learning,\n\t\t\trule learning and theory.\t & \\href{https://linqs.soe.ucsc.edu/data}{Link 1}\\newline \\href{https://github.com/albertyang33/TADW}{Link 2}\\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf CiteSeer} & A classified citation network in the field of machine learning with 3,312 papers (nodes) and 4,732 citations (edges). Each node is attributed with a binary vector indicating the absence/presence of the corresponding words from the dictionary of the 3,703 words collected from the corpus of papers. Papers are classified into 6 classes. & \\href{https://linqs.soe.ucsc.edu/data}{Link 1}\\newline \\href{https://github.com/albertyang33/TADW}{Link 2}\\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Sinanet} & A classified microblog user relationship network extracted from the sina-microblog website (http://www.weibo.com) with 3,490 users (nodes) and 30,282 relationships (edges). Each node is attributed with 10-dimensional numerical attributes describing the interests of the user. & \\href{https://github.com/smileyan448/Sinanet}{Link}\\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf PubMed Diabetes} & A classified citation networks extracted from the PubMed database pertaining to diabetes. It contains 19,717 publications (nodes) and 44,338 citations (edges). Each node is attributed by a TF-IDF weighted word vector from a dictionary that consists of 500 unique words. & \\href{https://linqs.soe.ucsc.edu/data}{Link} \\\\ \n\t\t\t\\hline\n\t\t\t{\\bf Facebook100} & A non-classified Facebook users network with 6,386 users (nodes) and 435,324 friendships (edges). The network is gathered from Facebook users of 100 colleges and universities (e.g. Caltech, Princeton, Georgetown and UNC Chapel Hill) in September 2005. Each user has the following attributes: ID, a student/faculty status flag, gender, major, second major/minor (if applicable), dormitory(house), year and high school. & \\href{https://archive.org/download/oxford-2005-facebook-matrix}{Link}\\newline \\\\ \n\t\t\t\\hline\n\t\t\t{\\bf ego-Facebook} & Dataset consists of 'circles' ('friends lists') from Facebook with 4039 nodes and 88234 edges. Facebook data was collected from survey participants using a Facebook app. The dataset includes node features (profiles), circles, and ego networks. & \\href{https://snap.stanford.edu/data/ego-Facebook.html}{Link}\\newline \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf LastFM} & A network gathered from the online music system Last.fm with 1,892 users (nodes) and 12,717 friendships on Last.fm (edges). Each node has 11,946-dimensional attributes, including a list of most listened music artists, and tag assignments.\n\t\t\t& \\href{http://ir.ii.uam.es/hetrec2011/datasets.html}{Link} \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Delicious} & A network of 1,861 nodes, 7,664 edges and 1,350 attributes. This is a publicly available dataset from the HetRec 2011 workshop that has been obtained from the Delicious social bookmarking system. Its users are connected in a social network generated from Delicious mutual fan relations. Each user has bookmarks, tag assignments, that is, [user, tag, bookmark] tuples, and contact relations within the social network. The tag assignments were\n\t\t\ttransformed to attribute data by taking all tags that a user ever assigned to any bookmark and assigning those to the user. & \\href{http://ir.ii.uam.es/hetrec2011/}{Link} \\\\ \n\t\t\t\\hline \n\t\t\t{\\bf Wiki} & A network with nodes as web pages. The link among different nodes is the hyperlink in the web page. 2,405 nodes, 12,761 edges, 4,973 attributes, 17 labels & \\href{https://github.com/albertyang33/TADW}{Link} \\\\ \n\t\t\\hline \n\t\t{\\bf ego-Twitter} & This dataset consists of 'circles' (or 'lists') from Twitter. Twitter data was crawled from public sources. The dataset includes node features (profiles), circles, and ego networks. Nodes\t81306,\n\t\tEdges 1768149 & \\href{https://snap.stanford.edu/data/ego-Twitter.html}{Link}\\newline \\\\\n\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{table} \n\\begin{table}\n\t\\caption{Most popular large size datasets.}\n\t\\label{tab_dataset3} \n\t\t{\\tiny\n\\begin{tabular}{|p{1.2cm}|p{12.9cm}|p{0.8cm}|}\n\t\\hline \nDataset & Description & Source \\\\ \n\t\\hline \n{\\bf Flickr} & A network with 100,267 nodes, 3,781,947 edges and 16,215 attributes collected from the internal database of the popular Flickr\nphoto sharing platform. The social network is defined by the contact relation of\nFlickr. Two vertices are connected with an undirected edge if at least one undirected edge exists between them. Each user has a list of tags associated that he/she used at least five times. Tags are limited to those used by at least 50 users. Users are limited to those having a vocabulary of more than 100 and less than 5,000 tags. & \\href{https://snap.stanford.edu/data/web-flickr.html}{Link}\\newline \\\\ \n\\hline \n{\\bf Patents} & A patent citation network with vertices representing patents and edges depicting the citations between. A subgraph containing all the patents from the year 1988 to 1999. Each patent has six attributes, grant year, number of claims, technological category, technological subcategory, assignee type, and main patent class. There are 1,174,908 vertices and 4,967,216 edges in the network. & \\href{http://www.nber.org/patents/}{Link 1}\\newline \\href{https://snap.stanford.edu/data/cit-Patents.html}{Link 2} \\\\ \n\\hline\n{\\bf ego-G+} & This dataset consists of 'circles' from Google+. Google+ data was collected from users who had manually shared their circles using the 'share circle' feature. The dataset includes node features (profiles), circles, and ego networks. Nodes 107,614, Edges 13,673,453. Each node has four features: job\ntitle, current place, university, and workplace. A user-pair(edge) is\ncompared using knowledge graphs based on, Category: Occupations,\nCategory:Companies by country and industry, Category: Countries,\nCategory:Universities and colleges by country. & \\href{https://snap.stanford.edu/data/ego-Gplus.html}{Link}\\newline \\\\\n\\hline \n\\end{tabular}\n}\n\\end{table}","id":"36da5b81-971c-4135-9405-fdc43b1406db","level":"subsection","origin_cites_number":12,"parent_id":"c0718d2b-4a41-48ee-939a-4ab8664cabe6","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Most used node-attributed network datasets and quality measures for community detection evaluation"],["subsection","Datasets"]],"subsections":[],"title":"Datasets"},{"cite_extract_rate":1,"cites":[1711],"content":"Given a set of detected communities (overlapping or not), one needs to evaluate their quality. There are two possible options for this depending on the network dataset under consideration. If the network dataset has no ground truth, one can use various measures of structural closeness and attribute homogeneity. According to our observations, the most popular quality measures in this case are Modularity and Density for the former and Entropy for the latter. Many others such as Conductance, Permanence, Intra- and Inter-Cluster Densities, etc., are also possible. If there is ground truth, it is traditional to compare the detected communities with the ground truth ones. This can be done, for instance, with the following popular measures: Accuracy, Normalized Mutual Information (denoted below by NMI), Adjusted Rand Index or Rand Index (denoted below by ARI and RI, correspondingly) and $F$-measure. We will discuss both the approaches further in Section~\\ref{section9}. \nDue to space limitations, we refer the reader to the comprehensive survey and to \\cite[Sections 2.2 and 4]{Bothorel2015}, where all the above-mentioned quality measures and many others are precisely defined and discussed in detail.","id":"01bdf590-8667-4b9e-b2aa-932d3480f52b","level":"subsection","origin_cites_number":1,"parent_id":"c0718d2b-4a41-48ee-939a-4ab8664cabe6","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Most used node-attributed network datasets and quality measures for community detection evaluation"],["subsection","Community detection quality measures"]],"subsections":[],"title":"Community detection quality measures"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section6}\nAs we have already mentioned, these methods fuse structure and attributes before the community detection process so that the data obtained are suitable for classical community detection algorithms (thus one can use their existing software implementations).\nBefore we proceed, we introduce additional notation applied only to the weight-based and distance-based early fusion methods. The fact is that existing network structure may be saved or modified depending on the heuristics used in a method, therefore we distinguish\n\t\\begin{itemize}\n\t\t\\item {\\it fixed topology methods} that use initial network structure without modifying it with respect to attributes,\n\t\t\\item {\\it non-fixed topology methods} that modify initial network structure with respect to attributes, in particular, add/erase edges and/or vertices.\n\t\\end{itemize}\nAs far as we know, there is no study on which approach is preferable. How each one influences community detection results is yet to be established.\n\\begin{figure}[b]\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{weight.pdf}\n\t\\caption{A typical scheme of a weight-based method (the attributive weights here are the values of normalized matching coefficient for the attribute vectors).}\n\t\\label{fig:weight}\n\\end{figure}","id":"1c6b434c-2289-4e6e-9627-ef1149463588","level":"section","origin_cites_number":0,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Early fusion methods"]],"subsections":["4ebad2e3-f246-4134-89c0-5d1b73a7710d","0bd4586d-c3e0-4e3a-817b-be22385f142c","9080cec6-6c41-47f4-8a10-ec818aed0224","bcc6d0ce-88ee-4a08-8574-4e5664967352","99f1e588-a145-4e2a-aa14-a5084080874e"],"title":"Early fusion methods"},{"cite_extract_rate":0.203389830508474,"cites":[7062,1714,1712,1710,8487,8488,8484],"content":"\\label{class-ef-weight}\nThese methods (see Tables \\ref{tab_weights_alpha0} and \\ref{tab_weights}) convert attributes $(\\mathcal{V},\\mathcal{A})$ into a weighted {\\it attributive} graph and further somehow merge it with structural graph $(\\mathcal{V},\\mathcal{E})$.\nThe result is a weighted graph $G_W$ that is a substitution for node-attributed graph $G$, see Figure~\\ref{fig:weight}. Edge weights of $G_W$ are usually assigned as follows:\n\\begin{equation}\n\\label{main_weight}\nW_\\alpha(v_i,v_j)=\\alpha w_S(v_i,v_j)+(1-\\alpha) w_A (v_i,v_j),\\qquad \\alpha\\in[0,1],\\qquad v_i,v_j\\in \\mathcal{V},\n\\end{equation}\nwhere $w_S$ and $w_A$ are chosen {\\it structural} and {\\it attributive} similarity functions, respectively. The hyperparameter $\\alpha$ controls the balance between structure and attributes. Clearly, if $\\alpha=1$ in (\\ref{main_weight}), one obtains weights based only on structure; if $\\alpha=0$, then they are based only on attributes. As for $w_S$ and $w_A$, $w_S(v_i,v_j)$ usually reflects existing connections in $(\\mathcal{V},\\mathcal{E})$ (e.g. $w_S(v_i,v_j)=1$, if $e_{ij}\\in \\mathcal{E}$, and $w_S(v_i,v_j)=0$, otherwise), while $w_A(v_i,v_j)$ may be Cosine Similarity or Matching Coefficient values for vectors $A(v_i)$ and $A(v_j)$.\nOnce the weighted graph $G_W$ is constructed, one can use classical graph clustering algorithms on it such as Weighted Louvain . Sometimes $G_W$ is instead converted to a certain distance matrix that is further used for detecting communities via distance-based clustering algorithms such as $k$-means or $k$-medoids.\nIt is worth mentioning that the fixed topology methods in this subclass assume that the weights (\\ref{main_weight}) are assigned on the same set of edges $\\mathcal{E}$ as in the initial node-attributed graph $G$, while the non-fixed topology ones assign weights on all (or on the most part of) possible edges between nodes in $\\mathcal{V}$.\nNow let us say some words how the hyperparameter $\\alpha$ can be chosen. A very popular approach in fixed topology methods is assuming $\\alpha=0$ in (\\ref{main_weight}), see Table~\\ref{tab_weights_alpha0}. This actually means that weights in $G_W$ are based only on attributive similarity. Clearly, this may lead to dominance of attributes in the resulting fusion and disappearance of structural connections between nodes with dissimilar attributes. Varying $\\alpha$ over the segment $[0,1]$ in (\\ref{main_weight}), used by the methods in Table~\\ref{tab_weights}, seems more adequate for controlling the impact of both the components. However, tuning $\\alpha$ in this case is usually performed {\\it manually}. In fact, we are unaware of any general non-manual approaches for tuning $\\alpha$ in (\\ref{main_weight}), although the need for such approaches has been repeatedly emphasized .\nNote that the choice of similarity functions $w_S$ and $w_A$ is usually determined by preferences of the authors of a particular method. The systematic study of how such a choice influences the community detection results is yet to be done.\n\\begin{table}\n\t\\caption{Weight-based methods with $\\alpha=0$ in (\\ref{main_weight}).}\n\t\\label{tab_weights_alpha0} \n\t\\begin{center}\n\t\t{\\tiny\n\t\t\t\\begin{tabular}{|p{1.5cm}|p{4.0cm}|p{1.5cm}|p{1.0cm}|p{1.1cm}|p{1cm}|p{1.3cm}|p{1.5cm}|}\n\t\t\t\t\\hline \n\t\t\t\tMethod & Community detection method used and its {\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Topology & Datasets used & Compared with \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WNev}~ & {\\bf Weighted graph}\\newline MinCut \\newline MajorClust \\newline Spectral & No/No & Small & Accuracy & Fixed & Synthetic & --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WSte1}~ & {\\bf Weighted graph}\\newline Threshold & No/No & Large & Modularity & Fixed & Phone Network & --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WSte2}~ & {\\bf Similarity matrix (via Weighted graph and random walks)}\\newline Hierarchical clustering & No/No & Large & Modularity & Fixed & Phone Network & --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WCom1} & {\\bf Weighted graph}\\newline Weighted Louvain & Yes/No & Small & Accuracy & Fixed & {\\bf DBLP*} & {\\bf WCom2} \\newline {\\bf DCom} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WCom2} & {\\bf Distance matrix (via weighted graph)}\\newline Hierarchical agglomerative clustering & Yes/No & Small & Accuracy & Fixed & {\\bf DBLP*} & {\\bf WCom1} \\newline {\\bf DCom} \\\\\n\t\t\t\t\\hline\n\t\t\t\t{\\bf AA-Cluster} & {\\bf Node embeddings (via weighted graph)}\\newline $k$-medoids & Yes/No & Small\\newline Medium\\newline Large & Density\\newline Entropy & Fixed & {\\bf Political Blogs}\\newline {\\bf DBLP*}\\newline {\\bf Patents*} \\newline Synthetic & {\\bf SA-Cluster}~\\newline {\\bf BAGC} \\newline {\\bf CPIP} \n\t\t\t\t\\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf PWMA-MILP} & {\\bf Weighted graph}\\newline Linear programming MILP~ & No/No & Small & RI \\newline NMI & Fixed & {\\bf WebKB} & --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf KDComm} & {\\bf Weighted graph} \\newline Iterative Weighted Louvain & No/No & Small\\newline Medium \\newline Large & $F$-measure\\newline Jaccard measure \\newline Rank Entropy measure & Fixed & \t{\\bf ego-G+} \\newline {\\bf Twitter*} \\newline {\\bf DBLP*} \\newline Reddit \\href{https://archive.org/details/2015_reddit_comments_corpus}{link} & {\\bf CPIP} \\newline {\\bf JCDC} \\newline {\\bf UNCut} \\newline {\\bf SI} \\\\\n\t\t\t\t\\hline \n\t\t\t\\end{tabular}\n\t\t}\n\t\\end{center}\n\\end{table}\n\\begin{table}\n\t\\caption{Weight-based methods with $\\alpha\\in[0,1]$ in (\\ref{main_weight}).}\n\t\\label{tab_weights} \n\t\\begin{center}\n\t\t{\\tiny\n\t\t\t\\begin{tabular}{|p{1.25cm}|p{1.5cm}|p{2.0cm}|p{1.5cm}|p{0.8cm}|p{1.2cm}|p{1.2cm}|p{1.7cm}|p{1.7cm}|}\n\t\t\t\t\\hline \n\t\t\t\tMethod & $\\alpha$ in (\\ref{main_weight}) & Community detection method used and its {\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Topology & Datasets used & Compared with \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WWan} & $[0,1]$ in theory\\newline $\\tfrac{1}{2}$ in experiments & {\\bf Edge similarity matrix (via weighted graph)}\\newline EdgeCluster \\newline($k$-means variant) & Yes & Small & NMI \\newline Micro-F1 \\newline Macro-F1 & Non-fixed: removing edges & Synthetic \\newline BlogCatalog\\newline {\\bf Delicious$^*$} & Non-overlapping co-clustering \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf SAC2} & $[0,1]$ & {\\bf $k$NN (unweighted) graph (via weighted graph)}\\newline (Unweighted) Louvain & No/ No & Small\\newline Medium & Density\\newline Entropy & Non-fixed: removing edges & {\\bf Political Blogs} \\newline {\\bf Facebook100} \\newline {\\bf DBLP10K} & {\\bf SAC1} \\newline {\\bf WSte2} \\newline Fast greedy for weighted graph \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WCru}~\n\t\t\t\t& $[0,1]$ in theory \\newline Not specified in experiments & {\\bf Weighted graph}\\newline Weighted Louvain~ & No & Medium & Modularity\\newline Intracluster distance & Fixed & {\\bf Twitter$^*$} & ---\\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf CODICIL} & $[0,1]$ in theory\\newline\n\t\t\t\t$1/2$ in some experiments & {\\bf Weighted graph}\\newline Metis \\newline Markov Clustering~ & No & Small\\newline Medium\\newline Large & $F$-measure & Non-fixed: adding and removing edges & {\\bf CiteSeer*}\\newline {\\bf Flickr*}\\newline Wikipedia* & {\\bf Inc-Cluster} \\newline {\\bf PCL-DC} \\newline Link-PLSA-LDA \n\t\t\t\t\\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf WMen} & Not specified & {\\bf Weighted graph/Distance matrix for the weighted graph}\\newline SLPA \\newline Weighted Louvain \\newline K-medoids & Yes-No/\\newline Yes-No & Small\\newline Medium & NMI \\newline $F$-measure \\newline Accuracy & Fixed & {\\bf Lazega} \\newline {\\bf Research} \\newline {\\bf Consult} \\newline\n\t\t\t\tLFR benchmark & {\\bf CODICIL}~\\newline {\\bf SA-Cluster}~ \\\\ \n\t\t\t\t\\hline \n\t\t\t\t{\\bf PLCA-MILP} & $[0,1]$ & {\\bf Weighted graph}\\newline Linear programming MILP~ & No/No & Small & RI \\newline NMI & Non-fixed: adding and removing edges & {\\bf WebKB} & {\\bf SCD} \\newline {\\bf ASCD} \\newline {\\bf SCI} \\newline PCL-DC \\newline Block-LDA \\\\\n\t\t\t\t\\hline\n\t\t\t\t{\\bf kNN-enhance} & May be thought as $\\alpha=1/2$, $k$NN by attributes & {\\bf Distance matrix (of an edge-augmented graph)} \\newline $k$NN\\newline $k$-means & No/No & Medium & Accuracy \\newline NMI \\newline F-Measure \\newline Modularity\\newline Entropy & Non-fixed: adding edges & {\\bf Cora} \\newline {\\bf Citeseer}\\newline {\\bf Sinanet}\\newline {\\bf PubMed Diabetes}\\newline {\\bf DBLP*}& {\\bf PCL-DC} \\newline PPL-DC \\newline {\\bf PPSB-DC} \\newline {\\bf CESNA} \\newline {\\bf cohsMix} \\newline {\\bf BAGC} \\newline {\\bf GBAGC} ) \\newline {\\bf SA-Custer} \\newline {\\bf Inc-Cluster} \\newline {\\bf CODICIL} \\newline GLFM \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf IGC-CSM} (\\href{https://github.com/WNawaz/CSM}{source}) & $[0,1]$ in theory\\newline $1/2$ in comparison experiments & {\\bf Distance matrix for the weighted graph} \\newline $k$-Medoids & Yes/ No & Medium & Density\\newline Entropy & Fixed & {\\bf Political Blogs}\\newline {\\bf DBLP10K} & {\\bf SA-Cluster} \\newline {\\bf SA-Cluster-Opt} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf AGPFC} & $[0,1]$ in theory, manually tuned in experiments & Fuzzy equivalent matrix \\newline $\\lambda$-cut set method & No/Yes & Small\\newline Medium & Density\\newline Entropy & Fixed & {\\bf Political Blogs}\\newline {\\bf CiteSeer} \\newline {\\bf Cora} \\newline {\\bf WebKB} & {\\bf SA-Cluster} \\newline {\\bf BAGC} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf NMLPA} & $1/2$ & {\\bf Weighted graph}\\newline A multi-label propagation algorithm & Yes/ Yes & Medium & $F1$-score \\newline Jaccard Similarity & Fixed & {\\bf ego-Facebook}\\newline {\\bf Flickr*} \\newline {\\bf ego-Twitter} & {\\bf CESNA} \\newline {\\bf SCI} \\newline {\\bf CDE} \\\\\n\t\t\t\t\\hline \n\t\t\t\\end{tabular}\n\t\t}\n\t\\end{center}\n\\end{table}\n\\begin{remark} The weight-based strategy has been applied to more general networks than considered in the present survey. For example, and use a similar scheme to detect communities in {\\it multi-layer} networks. Another example is\nSANS that works with {\\it directed} node-attributed graphs. \nEdge weighting similar to (\\ref{main_weight}) is also applied in FocusCO\t. Although it is not a purely unsupervised community detection method (it requires user's preferences on some of the attributes), it can simultaneously extract local clusters and detect outliers in a node-attributed social network.\n\\end{remark}\n\\begin{remark}\nLet us mention that there exist community detection methods similar ideologically (attributes $\\to$ edge weights $\\to$ weighted graph node embeddings\\footnote{Low-dimensional continuous vector representations of graph nodes. See also Subsection~\\ref{class-ef-embedding}.} $\\to$ $k$-means) but preceding to recently proposed {\\bf AA-Cluster} . Namely, GraphEncoder~ and GraRep~ also first convert a node-attributed graph into a weighted one according to (\\ref{main_weight}) with $\\alpha=0$ and then find corresponding node embeddings. The embeddings are further fed to $k$-means algorithm to detect communities. However, in opposite to , and mostly focus on node embedding techniques for a weighted graph than on the community detection task. \n\\end{remark}","id":"4ebad2e3-f246-4134-89c0-5d1b73a7710d","level":"subsection","origin_cites_number":59,"parent_id":"1c6b434c-2289-4e6e-9627-ef1149463588","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Early fusion methods"],["subsection","Weight-based methods"]],"subsections":[],"title":"Weight-based methods"},{"cite_extract_rate":0.1,"cites":[1717,1714],"content":"\\label{class-ef-distance}\n\\begin{table}[b]\n\\caption {Distance-based methods.}\n\\label{tab:distance} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.0cm}|p{1.3cm}|p{1.8cm}|p{1.5cm}|p{1.2cm}|p{1.5cm}|p{1.1cm}|p{1.5cm}|p{1.8cm}|}\n\t\t\t\\hline \n\t\t\tMethod & $\\alpha$ in (\\ref{distance_main}) & Community detection method used and its {\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Topology & Datasets used & Compared with \\\\\n\t\t\t\\hline \n\t\t\t{\\bf DCom} & $[0,1]$ & {\\bf Distance matrix}\\newline Hierarchical agglomerative clustering & Yes/No & Small & Accuracy & Non-fixed: added edges & {\\bf DBLP*} & {\\bf WCom1} \\newline {\\bf WCom2} \\\\\n\t\t\t\\hline\n\t\t\t{\\bf DVil} & $[0,1]$ & {\\bf Distance matrix} \\newline Self-organizing maps & No/No & Small\\newline Medium & NMI & Non-fixed: added edges & Synthetic\\newline \\href{http://graphcomp.univ-tlse2.fr/}{Medieval Notarial Deeds} & --- \\\\\n\t\t\t\\hline \n\t\t\t{\\bf SToC} & Maybe thought depending on the values of $d_S$ and $d_A$ & {\\bf Distance matrix}\\newline $\\tau$-close clustering & No/No & Medium\\newline Large & Modularity\\newline Within-Cluster Sum of Squares & Non-fixed: added edges & {\\bf DBLP10K}\\newline DIRECTORS*\\newline DIRECTORS-gcc* & {\\bf Inc-Cluster} \\newline {\\bf GBAGC} \\\\\n\t\t\t\\hline \n\t\t\t{\\bf @NetGA} & $\\alpha\\in[0,1]$ in general\\newline $\\alpha=1/2$ in experiments & {\\bf Distance matrix}\\newline Genetic algorithm & No/No & Medium & NMI & Non-fixed: added edges & Synthetic & {\\bf SA-Cluster} \\newline CSPA \\newline {\\bf Selection} \\\\\n\t\t\t\\hline \n\t\t\t{\\bf ANCA} & Maybe thought as $\\alpha=1/2$ & {\\bf Distance matrix}\\newline $k$-means & Yes/No & Medium & Adjusted Rand Index\\newline NMI \\newline Density \\newline Modularity \\newline Conductance \\newline Entropy & Fixed & Synthetic \\newline {\\bf DBLP10K} \\newline \n\t\t\tEnron email corpus \\newline & {\\bf SA-Cluster} \\newline {\\bf SAC1-SAC2} \\newline {\\bf IGC-CSM} \\newline {\\bf WSte1} \\newline {\\bf ILouvain} \\\\\n\t\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}\nMethods discussed in the previous subsection aim at representing structure and attributes in a unified graph form suitable for further graph clustering. In opposite, distance-based methods intentionally abandon graph representation in favor of the representation via a distance matrix that contains information about structure and attributes. The distance matrix is usually obtained by fusing the components by a structure- and attributes-aware distance function, see Figure~\\ref{fig:distance}. The matrix can be further fed to distance-based clustering algorithms such as $k$-means and $k$-medoids. Note that the resulting clusters may contain disconnected portions of initial graph $G$ as the graph structure is removed at the fusion step \\cite[Section 3.3]{Akbas2017}.\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{distance.pdf}\n\t\\caption{A typical scheme of a distance-based method.}\n\t\\label{fig:distance}\n\\end{figure}\nThe distance function fusing structure and attributes is often of the form\n\\begin{equation}\n\\label{distance_main}\nD_{\\alpha}(v_i,v_j)=\\alpha d_S(v_i,v_j)+(1-\\alpha)d_A(v_i,v_j),\\qquad \\alpha\\in[0,1],\\qquad v_i,v_j\\in \\mathcal{V},\n\\end{equation}\nwhere $d_S$ and $d_A$ are {\\it structural} and {\\it attributive} distance functions, correspondingly. As in the case of (\\ref{main_weight}), $\\alpha$ in (\\ref{distance_main}) controls the balance between the components; how to choose it ``properly'' seems to be an open problem, too. As for the distance functions, it is common to define $d_S(v_i,v_j)$ as the shortest path length between $v_i$ and $v_j$. Possible options for $d_S(v_i,v_j)$ are the Jaccard or Minkowski distances between vectors $A(v_i)$ and $A(v_j)$.\nShort descriptions for known distance-based methods are given in Table~\\ref{tab:distance}. Note that {\\bf ANCA} and {\\bf SToC} employ a bit different fusion than in (\\ref{distance_main}) but nevertheless still deal with certain structure- and attribute-aware distances.\n\\begin{remark}\nThere exist distance-based methods for {\\it multi-layer} networks. For example, CLAMP is an method for clustering networks with heterogeneous attributes and multiple types of edges that uses a distance function similar to (\\ref{distance_main}), in a sense.\n\\end{remark}\n\\begin{table}\n\\caption {Node-augmented graph distance-based methods.} \\label{tab:sa} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.8cm}|p{1.7cm}|p{2.8cm}|p{1.2cm}|p{1cm}|p{1.0cm}|p{1.4cm}|p{2.2cm}|}\n\t\t\t\\hline \n\t\t\tMethod & Graph augmentation & Community detection method used and its {\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\n\t\t\t\\hline \n\t\t\t{\\bf SA-Cluster} \\newline {\\bf Inc-Cluster} \\newline {\\bf SA-Cluster-Opt} & New nodes and edges & {\\bf Distance matrix (via neighbourhood random walks) }\\newline Modified $k$-medoids & Yes/No & Small\\newline Medium & Density\\newline Entropy & {\\bf Political Blogs}\\newline {\\bf DBLP10K}\\newline {\\bf DBLP84K} & W-Cluster (based on (\\ref{distance_main}))\\newline {\\bf SA-Cluster} \\newline {\\bf Inc-Cluster} \\newline {\\bf SA-Cluster-Opt} \\\\\n\t\t\t\\hline \n\t\t\t{\\bf SCMAG} & New nodes and edges & {\\bf Distance matrix (via neighbourhood random walks) }\\newline Subspace clustering algorithm based on ENCLUS & No/Yes & Medium & Density\\newline Entropy & IMDB\\newline Arnetminer bibliography$^*$ & {\\bf SA-Custer} \\newline GAMer \\\\\t\t\t\n\t\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}","id":"0bd4586d-c3e0-4e3a-817b-be22385f142c","level":"subsection","origin_cites_number":20,"parent_id":"1c6b434c-2289-4e6e-9627-ef1149463588","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Early fusion methods"],["subsection","Distance-based methods"]],"subsections":[],"title":"Distance-based methods"},{"cite_extract_rate":0,"cites":[],"content":"\\label{class-ef-node-augmented-graph}\n\\begin{figure}[b]\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{auggraph.pdf}\n\t\\caption{A typical scheme of a node-augmented graph-based method.}\n\t\\label{fig:node-augmented}\n\\end{figure}\nMethods from this subsection (see Table~\\ref{tab:sa}) transform initial node-attributed graph $G$ into another node-augmented graph $\\tilde{G}$, with nodes from $\\mathcal{V}$ and new {\\it attributive} nodes representing attributes, see Figure~\\ref{fig:node-augmented}.\nTo be more precise, suppose that $dom(a_k)=\\{s_l\\}_{l=1}^{L_k}$, i.e. the domain of $k$th element (attribute) of attribute vector $A$ contains $L_k$ possible values. Then one should create $L_k$ new attribute nodes $\\tilde{v}_{k,l}$ corresponding to $l$th value of $k$th attribute. Such a procedure is performed for $k=1,\\ldots,d$, where $d=\\dim A$. The set $\\mathcal{V}_A:=\\{\\tilde{v}_{k,l}\\}$, where $k=1,\\ldots,d$ and $l=1,\\ldots,L_k$, is then the set of attributive nodes. An edge between structural node $v_i\\in \\mathcal{V}$ and attributive node $\\tilde{v}_{k,l}$ in $\\tilde{G}$ exists if $a_k(v_i)=s_l$. Community detection is further performed in $\\tilde{G}$. The methods in Table~\\ref{tab:sa} propose to apply random walks to obtain a certain distance matrix for $\\tilde{G}$ and further use it in a distance-based clustering algorithm.\nNote that the above-mentioned augmentation is not applicable to continuous attributes. What is more, $\\tilde{G}$ contains much more nodes end edges than $G$ (especially if $d$ and $L_k$, $k=1,\\ldots,d$, are large) and this makes the methods from this subclass rather computationally expensive.","id":"9080cec6-6c41-47f4-8a10-ec818aed0224","level":"subsection","origin_cites_number":1,"parent_id":"1c6b434c-2289-4e6e-9627-ef1149463588","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Early fusion methods"],["subsection","Node-augmented graph-based methods"]],"subsections":[],"title":"Node-augmented graph-based methods"},{"cite_extract_rate":0.35714285714285704,"cites":[1712,242,217,8489,215,282,1719,1718,229,1010],"content":"\\label{class-ef-embedding}\n\\begin{figure}\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{emb.pdf}\n\t\\caption{A possible scheme of an embedding-based method (the attribute vectors are here concatenated with the node embeddings and then fed to $k$-means).}\n\t\\label{fig:embeddings}\n\\end{figure}\nAs is well-known, a graph as a traditional representation of a network brings several difficulties to network analysis. As mentioned in , graph algorithms suffer from high computational complexity, low parallelisability and inapplicability of machine learning methods. Novel embedding techniques aim at tackling this by learning {\\it node embeddings}, i.e. low-dimensional continuous vector representations for network nodes so that main network information is efficiently encoded . Roughly speaking, node embedding techniques allow to convert a graph with $n$ nodes into a set of $n$ vectors.\nIn the context of node-attributed social networks, the objective of node embedding techniques is efficient encoding both structure and attributes . We are not going to provide general details on the techniques as this has been already done in the surveys . Let us just mention that having node embeddings (i.e. vectors) at hand allows one to use classical distance-based clustering algorithms such as $k$-means to detect communities, see Figure~\\ref{fig:embeddings}.\nIt turns out that there exists a rich bibliography on node embedding techniques for attributes networks of different type . However, not all of them have been applied to the community detection task (the classification task is typically considered). Taking this into account, we confine ourselves in this survey only to the embedding-based (early fusion) methods that have been used for community detection in node-attributed social or document networks and compared with other community detection methods. The results are presented in Table~\\ref{tab:emb}.\n\\begin{table}[b]\n\\caption {Embedding-based methods.}\n\\label{tab:emb} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1cm}|p{1.5cm}|p{2.5cm}|p{1.4cm}|p{1cm}|p{1.9cm}|p{1.5cm}|p{2.3cm}|}\n\t\t\t\\hline \n\t\t\tMethod & Embedding technique & Community detection method used and its {\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\n\t\t\t\\hline \n\t\t\t{\\bf PLANE} & A generative model and EM~ & {\\bf Node embeddings}\\newline $k$-means & Yes/No & Small\\newline Medium & Accuracy & {\\bf Cora*} & Relational Topic Model~+Topic Distributions Embedding~\\\\\n\t\t\t\\hline \n\t\t\t{\\bf DANE} & Autoencoder & {\\bf Node Embeddings}\\newline $k$-means & Yes/No & Medium & Accuracy & {\\bf Cora}\\newline {\\bf Citeseer}\\newline {\\bf PubMed Diabetes}\\newline {\\bf Wiki} & Embeddings obtained via TADW \\newline\n\t\t\tLANE \\newline GAE \\newline VGAE \\newline GraphSAGE \\\\\n\t\t\t\\hline \n\t\t\t{\\bf CDE} & Structure embedding matrix & {\\bf Structure embedding matrix and attribute matrix}\\newline Non-negative matrix factorization & Yes/(Yes/No) & Small\\newline Medium & Accuracy \\newline NMI \\newline Jaccard similarity \\newline F1-score & {\\bf Cora}\\newline {\\bf Citeseer}\\newline {\\bf WebKB}\\newline {\\bf Flickr*} \\newline Philosophers \\newline {\\bf ego-Facebook} & {\\bf PCL-DC} \\newline {\\bf Circles} \\newline {\\bf CESNA} \\newline \\newline {\\bf SCI} \\\\\n\t\t\t\\hline \n\t\t\t{\\bf MGAE} & Autoencoder & {\\bf Node embeddings} \\newline Spectral clustering & Yes/No & Medium & Accuracy\\newline NMI \\newline $F$-score \\newline Precision \\newline Recall \\newline Average Entropy \\newline Adjusted Rand Index & {\\bf Cora}\\newline {\\bf CiteSeer} \\newline {\\bf Wiki} & {\\bf Circles} \\newline RTM \\newline RMSC \\newline Embeddings obtained via TADW \\newline VGAE \\\\\n\t\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}\n\\begin{remark}\nVarious embedding techniques applicable for community detection in {\\it multi-layer} networks are considered e.g. in .\n\\end{remark}\n\\begin{table}[b]\n\\caption{Pattern mining-based (early fusion) methods.}\n \\label{tab:pattern} \n\\begin{center}\n\t{\\tiny\n\\begin{tabular}{|p{1.1cm}|p{0.9cm}|p{2.9cm}|p{2.5cm}|p{2cm}|p{1.5cm}|p{1.2cm}|p{1cm}|}\n\t\t\t\\hline \n\t\t\tMethod & Pattern used & Community detection method used and its {\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\n\t\t\t\\hline \n\t\t\t{\\bf AHMotif} & Motif & {\\bf Structure- and attributes-aware adjacency matrix} \\newline Permanence \\newline Affinity Propagation & Yes/No & Medium & NMI \\newline Accuracy & {\\bf Cora}\\newline {\\bf WebKB} & ---\\\\\n\t\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}","id":"bcc6d0ce-88ee-4a08-8574-4e5664967352","level":"subsection","origin_cites_number":28,"parent_id":"1c6b434c-2289-4e6e-9627-ef1149463588","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Early fusion methods"],["subsection","Embedding-based methods"]],"subsections":[],"title":"Embedding-based methods"},{"cite_extract_rate":0,"cites":[],"content":"\\label{class-ef-patterns}\nRecall that a motif is a pattern of the interconnection occurring in real-world networks at numbers that are significantly higher than those in random\nnetworks . Motifs are considered as building blocks for complex networks . We found just one community detection method for node-attributed social networks using such patterns, namely, {\\bf AHMotif} , see Table~\\ref{tab:pattern}. This method equips structural motifs identified in the network with the so-called homogeneity value based on node attributes involved in the motif. This information is stored in a special adjacency matrix that can be an input to classical community detection algorithms.","id":"99f1e588-a145-4e2a-aa14-a5084080874e","level":"subsection","origin_cites_number":2,"parent_id":"1c6b434c-2289-4e6e-9627-ef1149463588","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Early fusion methods"],["subsection","Pattern mining-based (early fusion) methods"]],"subsections":[],"title":"Pattern mining-based (early fusion) methods"},{"cite_extract_rate":0.228571428571428,"cites":[8483,1714,8487,8488,8482,1720],"content":"\\label{section7}\nRecall that simultaneous fusion methods fuse structure and attributes in a joint process with community detection. For this reason, these methods often require special software implementation, in contrast to early and late fusion methods that partially allow one to use existing implementations of classical community detection algorithms.\n\\begin{table}\n\t\\caption {Methods modifying objective functions of classical clustering algorithms}\n\t\\label{tab_Louvain_NCut} \n\t\\begin{center}\n\t\t{\\tiny\n\t\t\t\\begin{tabular}{|p{1.2cm}|p{2.8cm}|p{1.8cm}|p{1.5cm}|p{1.6cm}|p{2.0cm}|p{2.4cm}|}\n\t\t\t\t\\hline \n\t\t\t\tMethod & Modified algorithm & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf OCru} & Louvain \\newline Added attribute Entropy minimization & No/No & Medium & Modularity\\newline Entropy & {\\bf Facebook100} & --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf SAC1} & Louvain \\newline Added attribute similarity maximization & No/ No & Small\\newline Medium & Density\\newline Entropy & {\\bf Political Blogs}\\newline {\\bf Facebook100} \\newline {\\bf DBLP10K} & {\\bf SAC2} \\newline {\\bf WSte2} \\newline Fast greedy for weighted graph \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf ILouvain} (\\href{http://bit.ly/ILouvain}{source}) & Louvain \\newline Added maximization of attribute-aware Inertia & No/ No & Small\\newline Medium & NMI \\newline Accuracy & DBLP+Microsoft Academic Search$^*$ \\newline Synthetic & ToTeM \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf LAA/LOA} & Louvain \\newline Modularity gain depends on attributes & No/No & Small & Density\\newline Modularity & \\href{https://sites.google.com/site/ucinetsoftware/datasets/covert-networks/londongang}{London gang} \\newline\n\t\t\t\t\\href{https://sites.google.com/site/ucinetsoftware/datasets/covert-networks/italiangangs}{Italy gang} \\newline\n\t\t\t\t\\href{http://networkrepository.com/polbooks.php}{Polbooks}\n\t\t\t\t\\newline\n\t\t\t\t\\href{http://networkrepository.com/adjnoun.php}{Adjnoun} \n\t\t\t\t\\newline\n\t\t\t\tFootball & --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf MAM} (\\href{http://ipd.kit.edu/~muellere/mam/}{source}) & Louvain-type algorithm with attribute-aware Modularity+Outlier detection & No/No & Small\\newline Medium \\newline Large & F1-score \\newline Attribute-aware Modularity & Synthetic \\newline Disney \\newline DFB \\newline ARXIV \\newline\n\t\t\t\tIMDB \\newline \n\t\t\t\t{\\bf DBLP*} \\newline\n\t\t\t\t{\\bf Patents*} \\newline\n\t\t\t\tAmazon & CODA \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf UNCut} & Normalized Cut\\newline Added attributes-aware Unimodality Compactness & Yes/No & Small\\newline Medium & NMI \\newline ARI \\newline & Disney \\newline\n\t\t\t\tDFB \\newline\n\t\t\t\tARXIV \\newline\n\t\t\t\t{\\bf Political Blogs} \\newline\n\t\t\t\t4area \\newline\n\t\t\t\t{\\bf Patents} & {\\bf SA-cluster} \\newline SSCG \\newline {\\bf NNM} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf NetScan} & An approximation algorithm for the connected $k$-Center optimization problem (structure and attributes involved) & Yes/Yes & Small\\newline Medium & Accuracy & Professors* \\newline Synthetic \\newline {\\bf DBLP*}\\newline BioGRID+Spellman\n\t\t\t\t& --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf JointClust} & An approximation algorithm for the Connected X Clusters problem (structure and attributes involved) & No/No & Medium & Accuracy & {\\bf DBLP*} \\newline {\\bf CiteSeer*} \\newline Corel stock photo collection & --- \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf SS-Cluster} & $k$-Medoid with structure- and attributes-aware objective functions & Yes/No & Medium & Density \\newline Entropy & {\\bf Political Blogs}\\newline {\\bf DBLP10K} & {\\bf SA-cluster} \\newline W-cluster \\newline $k$SNAP \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf Adapt-SA} & Weighted $k$-means for $d$-dimensional representations of structure and attributes & Yes/No & Medium & Accuracy \\newline \n\t\t\t\tNMI \\newline F-measure \\newline Modularity\\newline\n\t\t\t\tEntropy & Synthetic \\newline {\\bf WebKB}\\newline {\\bf Cora}\\newline {\\bf Political Blogs} \\newline {\\bf CiteSeer} \\newline {\\bf DBLP10K} & {\\bf CODICIL} \\newline {\\bf SA-Cluster} \\newline {\\bf Inc-Cluster} \\newline {\\bf PPSB-DC} \\newline {\\bf PCL-DC} \\newline {\\bf BAGC} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf kNAS} & $kNN$ with added Semantic Similarity Score & Yes/Yes & Medium & Density\\newline Tanimoto Coefficient & {\\bf DBLP*} \\newline {\\bf Facebook*} \\newline {\\bf Twitter*} & {\\bf SA-Cluster-Opt} \\newline {\\bf CODICIL} \\newline NISE \\\\\n\t\t\t\t\\hline \n\t\t\t\\end{tabular}\n\t\t}\n\t\\end{center}\n\\end{table}","id":"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","level":"section","origin_cites_number":35,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Simultaneous fusion methods"]],"subsections":["1af2c98c-c4a1-4d9f-917a-330efea7a2a3","7b216870-6a3a-40ca-874f-da14c2289f12","c986582c-70df-43b2-b7f2-9e2e83ab57f5","f8f23311-f720-42e4-94c4-ed1c64c919d4","45b099b6-e934-4199-b2c6-b9c78b20f692","6a22c128-a708-4842-b8d3-074bec4c46a4"],"title":"Simultaneous fusion methods"},{"cite_extract_rate":0.33333333333333304,"cites":[8487],"content":"\\label{class-sf-modifying}\nTable~\\ref{tab_Louvain_NCut} contains\\footnote{The authors of {\\bf ILouvain} claim that they compare their method with ToTeM , ``another community detection method designed for attributed graphs''. However, it seems that there is an inaccuracy with it as we could not find in any method called ToTeM.} short descriptions of simultaneous fusion methods that modify objective functions of well-known clustering algorithms such as Louvain, Normalized Cut, $k$-means, $k$-medoids and $kNN$. Their main idea is to adapt a classical method (that works, for example, only with network structure originally) for using both structure and attributes in the optimization process. For example, if one wants to modify Louvain whose original objective function is structure-aware Modularity, one can include an attributes-aware objective function, say, Entropy in the optimization process. Then Modularity is maximized and Entropy is minimized simultaneously in an iterative process similar to that of Louvain .","id":"1af2c98c-c4a1-4d9f-917a-330efea7a2a3","level":"subsection","origin_cites_number":3,"parent_id":"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Simultaneous fusion methods"],["subsection","Methods modifying objective functions of classical clustering algorithms"]],"subsections":[],"title":"Methods modifying objective functions of classical clustering algorithms"},{"cite_extract_rate":0.25,"cites":[7062,1714,1712],"content":"\\label{class-sf-mataheuristic}\nMethods in this subclass are rather similar ideologically to those in Subsection~\\ref{class-sf-modifying}. However, instead of modifying objective functions and iterative processes of well-known community detection algorithms, they directly apply metaheuristic algorithms (in particular, evolutionary algorithms and tabu search) to find a node-attributed network partition that provides optimal values for certain measures of structural closeness and attribute homogeneity. More precisely, they use metaheuristics for optimization of a combination of structure- and attributes-aware objective functions, for example, Modularity and Attributes Similarity. Short descriptions of the methods from this subclass are given in Table~\\ref{tab_heur}. \n\\begin{table}\n\t\\caption {Metaheuristic-based methods.}\n\t\\label{tab_heur} \n\t\\begin{center}\n\t\t{\\tiny\n\t\t\t\\begin{tabular}{|p{1.2cm}|p{3.7cm}|p{1.5cm}|p{1.0cm}|p{1.4cm}|p{2.5cm}|p{2.2cm}|}\n\t\t\t\t\\hline \n\t\t\t\tMethod & Metaheuristic optimization algorithm & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf MOEA-SA} & Multiobjective evolutionary algorithm (Modularity and Attribute Similarity are maximized) & No/No & Small \\newline Medium & Density\\newline Entropy & {\\bf Political Books}\\newline {\\bf Political Blogs} \\newline {\\bf Facebook100}\\newline {\\bf ego-Facebook} & {\\bf SAC1-SAC2} \\newline {\\bf SA-Cluster} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf MOGA-@Net} & Multiobjective genetic algorithm (optimizing Modularity, Community score, Conductance, attribute similarity) & No/No & Small\\newline Medium & NMI \\newline Cumulative NMI \\newline Density \\newline Entropy & Synthetic \\newline {\\bf Cora}\\newline {\\bf Citeseer} \\newline {\\bf Political books} \\newline {\\bf Political Blogs} \\newline {\\bf ego-Facebook} & {\\bf SA-cluster} \\newline {\\bf BAGC} \\newline {\\bf OCru} \\newline {\\bf Selection} \\newline HGPA-CSPA \\\\\n\t\t\t\t\\hline\n\t\t\t\t{\\bf JCDC} & Tabu search and gradient ascent for a certain structure- and attributes-aware objective function & Yes/No & Small\\newline Medium & NMI & Synthetic \\newline World trade \\newline {\\bf Lazega} & CASC \\newline {\\bf CESNA} \\newline {\\bf BAGS} \\\\\n\t\t\t\t\\hline \n\t\t\t\\end{tabular}\n\t\t}\n\t\\end{center}\n\\end{table}","id":"7b216870-6a3a-40ca-874f-da14c2289f12","level":"subsection","origin_cites_number":12,"parent_id":"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Simultaneous fusion methods"],["subsection","Metaheuristic-based methods"]],"subsections":[],"title":"Metaheuristic-based methods"},{"cite_extract_rate":0.1,"cites":[1712,1721,8489],"content":"\\label{class-sf-nnmf}\nNon-negative matrix factorization (NNMF) is a matrix technique that consists in approximating a non-negative matrix with high rank by a product of non-negative matrices with lower ranks so that the approximation error by means of the Frobenius norm\\footnote{A matrix norm defined as the square root of the sum of the absolute squares of matrix elements.} $F$ is minimal. As is well known, NNMF is able to find clusters in the input data . \nTo be applied to community detection in node-attributed social networks, NNMF requires a proper adaptation to fuse both structure and attributes. Different versions of such an adaptation have been proposed, see Table~\\ref{tab:NMF}. \nTo be more formal in describing the corresponding NNMF-based methods, let us introduce additional notation. Let ${\\bf S}_{n\\times n}$ denote the adjacency matrix for the network structure (as before, $n$ is the number of nodes), ${\\bf A}_{n\\times d}$ the node attribute matrix for the network attributes ($d$ is the dimension of attribute vector $A$), $N$ the number of required clusters (it is an input in NNMF-based methods), ${\\bf U}_{n\\times N}$ the cluster membership matrix whose elements indicate the association of nodes with communities, and finally ${\\bf V}_{d\\times N}$ denotes the cluster membership matrix whose elements indicate the association of the attributes with the communities. In these terms, the aim of NNMF-based methods is to use known matrices $\\bf S$, $\\bf A$ and the number of clusters $N$ in order to determine the unknown matrices $\\bf U$ and $\\bf V$ in an iterative optimization procedure. For example, {\\bf SCI} models structural closeness as $\\min_{{\\bf U}\\ge 0}\\|{\\bf S}-{\\bf U}{\\bf U}^T\\|^2_F$ and attribute homogeneity as \n$\\min_{{\\bf V}\\ge 0}\\|{\\bf U}-{\\bf A}{\\bf V}\\|^2_F$. It is also proposed to select the most relevant attributes for each community by adding an $l_1$ norm sparsity term to each column of matrix\n$\\bf V$. As a result, one obtains the following optimization problem: \n$$\n\\min_{{\\bf U}\\ge 0,\\,{\\bf V}\\ge 0}\\left(\\alpha_1\\|{\\bf S}-{\\bf U}{\\bf U}^T\\|^2_F+\\|{\\bf U}-{\\bf A}{\\bf V}\\|^2_F+\\alpha_2 \\sum\\nolimits_j \\|{\\bf V}(\\cdot,j)\\|^2_1\\right),\n$$\nwhere $\\alpha_1>0$ controls the impact of structure and $\\alpha_1\\ge 0$ the sparsity penalty. This problem is further approximately solved in an iterative process according to Majorization-Minimization framework .\n\\begin{table}\n\\caption {Non-negative matrix factorization-based and matrix compression-based methods.} \\label{tab:NMF} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.6cm}|p{1.8cm}|p{1.8cm}|p{1cm}|p{2cm}|p{2.5cm}|p{2.8cm}|}\n\t\t\t\\hline \n\t\t\tAlgorithm & Factorization/ compression type & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\n\t\t\t\\hline \n\t\t\t{\\bf NPei} & 3-factor NNMF & Yes/Yes & Small\\newline Medium & Purity & {\\bf Twitter} \\newline DBLP* & Relational Topic Model \\\\\n\t\t\t\\hline \n\t\t\t{\\bf 3NCD} & $2$-factor NNMF & Yes/Yes & Medium\\newline Large & F1-score \\newline Jaccard\n\t\t\tsimilarity & {\\bf ego-Facebook}\\newline {\\bf ego-Twitter}\\newline {\\bf ego-G+} & {\\bf CESNA} \\\\\n\t\t\t\\hline \n\t\t\t{\\bf SCI} & 2-factor NNMF & Yes/Yes & Medium & ACC \\newline NMI \\newline GNMI \\newline $F$-measure \\newline Jaccard similarity & {\\bf Citeseer}\\newline {\\bf Cora}\\newline {\\bf WebKB}\\newline {\\bf LastFM} & {\\bf PCL-DC} \\newline {\\bf CESNA} \\newline {\\bf DCM} \\\\\n\t\t\t\\hline \n\t\t\t{\\bf JWNMF} & 2-factor NNMF & Yes/Yes & Small\\newline Medium & Modularity\\newline Entropy\\newline NMI & Amazon Fail \\href{https://www.ipd.kit.edu/mitarbeiter/muellere/consub/}{dataset} \n\t\t\t\\newline Disney \\href{http://www.perozzi.net/projects/focused-clustering/}{dataset} \\newline Enron \\href{https://www.ipd.kit.edu/mitarbeiter/muellere/consub/}{dataset} \\newline DBLP-4AREA \\href{http://www.perozzi.net/projects/focused-clustering/}{dataset}\\newline {\\bf WebKB} \\newline {\\bf Citeseer} \\newline {\\bf Cora} & {\\bf BAGC} \\newline {\\bf PICS} \\newline SANS \\\\\n\t\t\t\\hline \n\t\t\t{\\bf SCD} & 2- and 3-factor NNMF & Yes/Yes-No & Small\\newline Medium & Accuracy \\newline NMI & {\\bf Twitter}\\newline {\\bf WebKB} & {\\bf SCI} \\\\\n\t\t\t\\hline\n\t\t\t{\\bf ASCD} & 2-factor NNMF &Yes/Yes-No & Small\\newline Medium & ACC \\newline NMI \\newline $F$-measure \\newline Jaccard similarity & {\\bf LastFM} \\newline {\\bf WebKB}\\newline {\\bf Cora} \\newline {\\bf Citeseer} \\newline {\\bf ego-Twitter*} \\newline {\\bf ego-Facebook*} & Block-LDA \\newline {\\bf PCL-DC} \\newline {\\bf SCI} \\newline {\\bf CESNA} \\newline {\\bf Circles} \\\\\n\t\t\t\\hline\n\t\t\t{\\bf CFOND} & $2$- and $3$-factor NMF & Yes/(Yes/No) & Medium & Accuracy \\newline NMI & {\\bf Cora}\\newline {\\bf CiteSeer}\\newline {\\bf PubMed}\\newline {\\bf Attack}\\newline Synthetic\n\t\t\t& GNMF \\newline DRCC \\newline \n\t\t\tLP-NMTF \\newline iTopicModel \\\\\n\t\t\t\\hline \n\t\t\t{\\bf MVCNMF} & $2$-factor NMF & Yes/Yes & Small\\newline Medium & Density\\newline Entropy & {\\bf Political\n\t\t\t\tBlogs}\\newline {\\bf CiteSeer}\\newline {\\bf Cora}\\newline {\\bf WebKB} \\newline ICDM (DBLP*) & FCAN \\newline SACTL \\newline {\\bf kNAS} \\\\\n\t\t\t\\hline\n\t\t\t{\\bf PICS} (\\href{http://www.andrew.cmu.edu/user/lakoglu/pubs.html#pics}{source}) & Matrix compression (finding rectangular blocks) & No/No & Small\\newline Medium & Anecdotal and visual study & Youtube \\newline {\\bf Twitter}* \\newline Phonecall \\newline Device \\newline {\\bf Political Books} (\\href{http://www-personal.umich.edu/~mejn/netdata/}{link}) \\newline {\\bf Political Blogs} & ---\\\\\n\t\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}\n\\begin{remark}\nAn NNMF-based community detection method for {\\it multi-layer} networks is given in .\n\\end{remark}\nNow we briefly describe {\\bf PICS} , the method adopting matrix compression\\footnote{The aim of compression methods is to find a\nshorter form of describing the same information content.} for community detection in node-attributed social networks. {\\bf PICS} uses lossless compression principles from to simultaneously compress the network adjacency matrix $\\bf S$ and the attribute matrix $\\bf A$. As a result of the compression, certain homogeneous rectangular blocks in the matrices can be determined. Groups of the nodes corresponding to the blocks are considered as communities. One should be aware however that nodes within communities found by {\\bf PICS} may not be densely connected due to the definition of a community in .","id":"c986582c-70df-43b2-b7f2-9e2e83ab57f5","level":"subsection","origin_cites_number":30,"parent_id":"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Simultaneous fusion methods"],["subsection","Non-negative matrix factorization-based and matrix compression-based methods"]],"subsections":[],"title":"Non-negative matrix factorization-based and matrix compression-based methods"},{"cite_extract_rate":0.22222222222222202,"cites":[1714,1710],"content":"\\label{class-sf-patterns}\nPattern mining in node-attributed social networks focuses on extraction of patterns, e.g. subsets of specific attributes or connections\\footnote{An example of a pattern is a maximal clique . Recall that a clique is a subset of nodes in an undirected graph such that every two nodes are adjacent. A clique is called maximal if there is no other clique that contains it.}, in network structure and attributes . Among other things, this helps to make sense of a network and to understand how it was formed. In the context of community detection, the extracted patterns are used as building blocks for communities. \nThere are many papers devoted to pattern mining in social networks but the majority of them do not deal with the task of community detection. The ones relevant to the topic of this survey are presented in Table~\\ref{tab:clique}. It is worth mentioning here that it is common for pattern mining-based methods to detect communities not in the whole network but in its part only (e.g. ). \n\\begin{table}\n\\caption {Pattern mining-based (simultaneous fusion) methods.}\n\\label{tab:clique} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.4cm}|p{1.2cm}|p{2.5cm}|p{2cm}|p{2.5cm}|p{2.0cm}|p{1.8cm}|}\n\t\t\t\\hline \n\t\t\tMethod & Patterns used & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\n\t\t\t\\hline \n\t\t\t{\\bf DCM} (\\href{http://www.patternsthatmatter.org/software.php}{source}) & Semantic patterns ({\\it queries}) & Yes/Yes & Small\\newline Medium & Community score\\newline Conductance \\newline Intra-cluster density \\newline Modularity & {\\bf Delicious}\\newline {\\bf LastFM}\\newline {\\bf Flickr} & --- \\\\\n\t\t\t\\hline \n\t\t\t{\\bf COMODO} & Semantic patterns & Yes/Yes & Small\\newline Medium & Description complexity\\newline Community size & BibSonomy \\newline {\\bf Delicious} \\newline {\\bf LastFM} & {\\bf DCM} \\\\\n\t\t\t\\hline \n\t\t\t{\\bf ACDC} & Maximal cliques & Yes/Yes & Medium & Density & {\\bf Political Blogs} & {\\bf SA-Cluster} \\newline {\\bf SAC1-SAC2} \\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}\n\\begin{remark}\nABACUS detects communities by extracting patterns in {\\it multi-layer} networks.\n\\end{remark}","id":"f8f23311-f720-42e4-94c4-ed1c64c919d4","level":"subsection","origin_cites_number":9,"parent_id":"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Simultaneous fusion methods"],["subsection","Pattern mining-based (simultaneous fusion) methods"]],"subsections":[],"title":"Pattern mining-based (simultaneous fusion) methods"},{"cite_extract_rate":0.137931034482758,"cites":[1712,1721,1714,7062],"content":"\\label{class-sf-probabilistic-models}\n\\begin{table}[b]\n\t\\caption{Probabilistic model-based methods.} \\label{tab:statistical} \n\t\\begin{center}\n\t\t{\\tiny\t\\begin{tabular}{|p{1.2cm}|p{2.5cm}|p{1.5cm}|p{1.3cm}|p{1.5cm}|p{3cm}|p{2.5cm}|}\n\t\t\t\t\\hline \n\t\t\t\tMethod & Model features & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf PCL-DC} & Conditional Link Model\\newline Discriminative Content model & Yes/No & Medium & NMI \\newline Pairwise $F$-measure \\newline Modularity\\newline Normalized cut & {\\bf Cora}\\newline {\\bf Siteseer} & PHITS-PLSA \\newline LDA-Link-Word \\newline Link-Content-Factorization \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf CohsMix} & MixNet model & Yes/No & Small & Rand Index & Synthetic \\newline Exalead.com search engine dataset & Multiple view learning \\newline Hidden Markov Random Field \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf BAGC} \\newline {\\bf GBAGC} & a Bayesian treatment on distribution parameters & Yes/No & Medium & Modularity\\newline Entropy & {\\bf Political Blogs}\\newline {\\bf DBLP10K}\\newline {\\bf DBLP84K} & {\\bf Inc-Cluster} \\newline {\\bf PICS} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf VEM-BAGC} & Based on {\\bf BAGC} & Yes/No & Medium & Modularity\\newline Entropy & {\\bf Political Blogs}\\newline Synthetic networks & {\\bf BAGC} \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf PPSB-DC} & Popularity-productivity stochastic block model and discriminative content model & Yes/No & Medium & normalized mutual information (NMI)\\newline Pairwise\n\t\t\t\tF measure (PWF)\\newline Accuracy & {\\bf Cora} \\newline {\\bf CiteSeer} \\newline {\\bf WebKB} & {\\bf PCL-DC} \\newline PPL-DC \\\\\n\t\t\t\t\\hline \n\t\t\t{\\bf CESNA} & A probabilistic generative model assuming communities generate network structure and attributes & No/Yes & Medium\\newline Large & Evaluation & {\\bf ego-Facebook}\\newline {\\bf ego-G+} \\newline {\\bf ego-Twitter}\\newline \n\t\t\t{\\bf Wikipedia* (philosophers)}\\newline {\\bf Flickr} & {\\bf CODICIL} \\newline {\\bf Circles} \\newline Block-LDA \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf Circles} & A generative model for friendships in social circles & Yes/Yes & Medium\\newline Large & Balanced Error Rate & {\\bf ego-Facebook}\\newline {\\bf ego-G+} \\newline {\\bf ego-Twitter} & Block-LDA \\newline Adapted Low-Rank Embedding \\\\\n\t\t\t\t\\hline \n\t\t\t\t{\\bf SI} & A modified version of a stochastic block model & Yes/No & Small\\newline Medium & Normalized\n\t\t\t\tmutual information (NMI) & Synthetic \\newline High school friendship network \\newline Food web of marine species in the Weddell Sea \\newline Harvard Facebook friendship network \\newline Malaria HVR 5 and 6 gene\n\t\t\t\trecombination network & --- \\\\\n\t\t\t\t\\hline\n\t\t\t\t{\\bf NEMBP} & A generative model with learning method using a nested EM algorithm with belief propagation & Yes/(Yes/No) & Small \\newline Medium & Accuracy \\newline NMI \\newline GNMI \\newline F-score \\newline Jaccard & {\\bf WebKB} \\newline {\\bf ego-Twitter*}\\newline\n\t\t\t\t{\\bf ego-Facebook*}\\newline {\\bf CiteSeer} \\newline {\\bf Cora}\\newline {\\bf Wikipedia*}\\newline {\\bf Pubmed} & Block-LDA \\newline {\\bf PCL-DC} \\newline {\\bf CESNA} \\newline {\\bf DCM} \\newline {\\bf SCI} \\\\\n\t\t\t\t\\hline\n\t\t\t\t{\\bf NBAGC-FABAGC} & A nonparametric and asymptotic Bayesian model selection method based on {\\bf BAGC} & No/No & Medium & NMI \\newline Modularity \\newline Entropy & Synthetic \\newline {\\bf Political Blogs} \\newline {\\bf DBLP10K} \\newline {\\bf DBLP84K} & {\\bf PICS} \\\\\n\t\t\t\t\\hline\n\t\t\t\\end{tabular}\n\t\t}\n\t\\end{center}\n\\end{table}\nMethods from this subclass probabilistically infer the distribution of community memberships for nodes in a node-attributed social network under the assumption that network structure and attributes are generated according to chosen parametric distributions. Generative and discriminative models are mainly used for the inferring. It is worth mentioning that it is though a non-trivial task to ``properly'' choose a priori distributions for structure and attributes .\nShort descriptions of the methods from this subclass are given in Table~\\ref{tab:statistical}. Pay attention that this table does not contain any method preceding to and this requires the following additional comments. According to , several probabilistic model-based clustering methods for node-attributed networks had been proposed before , for example, in . However, they focus on node-attributed {\\it document} networks which are out of scope of the present survey. That is why they are non included in Table~\\ref{tab:statistical}.\n\\begin{remark}\nTUCM proposes a generative Bayesian model for detecting communities in {\\it multi-layer} networks where different types of interactions between social actors are possible. \n\\end{remark}","id":"45b099b6-e934-4199-b2c6-b9c78b20f692","level":"subsection","origin_cites_number":29,"parent_id":"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Simultaneous fusion methods"],["subsection","Probabilistic model-based methods"]],"subsections":[],"title":"Probabilistic model-based methods"},{"cite_extract_rate":0.22222222222222202,"cites":[1714,1712],"content":"\\label{class-sf-dynamics}\nMethods from this subclass (see Table~\\ref{tab:agent}) treat a node-attributed social network as a dynamic system and assume that its community structure is a consequence of certain interactions among nodes (of course, the attributes are thought to affect the interactions). Some of the methods assume that the interactions occur in an information propagation process, i.e. while information is sent to or received from every node. Others comprehend each node as an autonomous agent and develop a multi-agent system to detect communities. Note that these approaches are rather recent and consider community detection in node-attributed social networks from a new perspective. Furthermore, they seem to be efficient for large networks as can be easily parallelized.\n\\begin{table}\n\\caption{Dynamical system-based and agent-based methods.}\n\\label{tab:agent} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.2cm}|p{3.2cm}|p{1.7cm}|p{1.5cm}|p{2cm}|p{2cm}|p{1.8cm}|}\n\t\t\t\\hline \n\t\t\tMethod & Description & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\n\t\t\t\\hline \n\t\t\t{\\bf CPIP}-{\\bf CPRW} & Content (information) propagation models: a linear approximate model\n\t\t\tof influence propagation ({\\bf CPIP}) and content propagation\n\t\t\twith the random walk principle ({\\bf CPRW}) & Yes/Yes & Medium & F-score\\newline Jaccard Similarity\\newline NMI & {\\bf CiteSeer}\\newline {\\bf Cora}\\newline {\\bf ego-Facebook}\\newline {\\bf PubMed Diabetes} & Adamic Adar \\newline {\\bf PCL-DC} \\newline {\\bf Circles} \\newline {\\bf CODICIL} \\newline {\\bf CESNA} \\\\\n\t\t\t\\hline \n{\\bf CAMAS} & Each node with attributes as an autonomous agent with influence in a cluster-aware multiagent system & No/Yes & Medium\\newline Large & Coverage Rate\\newline Normalized Tightness \\newline Normalized Homogeneity\\newline F1-Score \\newline Jaccard\\newline Adjusted Rand Index & Synthetic \\newline {\\bf ego-Facebook}\\newline {\\bf ego-Twitter*} \\newline {\\bf ego-G+} & {\\bf CESNA} \\newline EDCAR \\\\\n\\hline \n{\\bf SLA} & A dynamic cluster formation game played by all nodes and clusters in a discrete-time dynamical system & Yes/No & Medium \\newline Large & Density \\newline Entropy \\newline F1-score \\newline & {\\bf Delicious}\\newline {\\bf LastFM} \\newline {\\bf ego-Facebook}\\newline {\\bf ego-Twitter*} \\newline {\\bf ego-G+} & {\\bf CESNA} \\newline EDCAR \\\\\n\\hline \n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}\n\\begin{figure}[b]\n\t\\centering\n\t\\includegraphics[width=0.9\\linewidth]{late.pdf}\n\t\\caption{A typical scheme of a late fusion method.}\n\t\\label{fig:late-fusion}\n\\end{figure}","id":"6a22c128-a708-4842-b8d3-074bec4c46a4","level":"subsection","origin_cites_number":9,"parent_id":"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Simultaneous fusion methods"],["subsection","Dynamical system-based and agent-based methods"]],"subsections":[],"title":"Dynamical system-based and agent-based methods"},{"cite_extract_rate":0.5,"cites":[8487],"content":"\\label{section8}\nRecall that late fusion methods intend to fuse structural and attributive information after the community detection process. More precisely, community detection is first separately performed for structure (e.g. by Louvain ) and attributes (e.g. by $k$-means ). After that, the partitions obtained are fused somehow in order to get the resulting structure- and attributes-aware partition, see Figure~\\ref{fig:late-fusion}.\nNote that late fusion methods usually allow a researcher/data scientist to use existing implementations of classical community detection and consensus clustering algorithms to get the required partition.","id":"a4712ae6-1690-4727-8bbb-b44187674ee8","level":"section","origin_cites_number":2,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Late fusion methods"]],"subsections":["a2976c7a-9d26-4a9e-9cb2-40daafc174e6","edd7b709-6935-46b9-9d57-cbc247202cf8"],"title":"Late fusion methods"},{"cite_extract_rate":0.35714285714285704,"cites":[8484,1712,1722,1715],"content":"\\label{class-lf-consensus}\nGiven a set (also known as an ensemble) of partitions, the general goal of consensus clustering algorithms is to find a single consolidated partition that aggregates information in the ensemble . A recent survey on such methods can be found e.g. in . The idea behind consensus clustering is clearly appropriate for community detection in node-attributed social networks if one has an ensemble of partitions obtained separately (or maybe even jointly) for structure and attributes. Table~\\ref{tab:late} contains short descriptions of the methods applying the idea.\n\\begin{table}[t]\n\\caption {Consensus-based methods.} \\label{tab:late} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.2cm}|p{3.0cm}|p{1.9cm}|p{1.5cm}|p{1.5cm}|p{2.3cm}|p{2cm}|}\n\t\t\t\\hline \n\t\t\tMethod & Fusing the partitions & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\n\t\t\t\\hline \n\t\t\t{\\bf LCru} & Row-manipulation in the contingency matrix for the partitions & No/No & Small\\newline Medium & ARI \\newline Density\\newline Entropy & Facebook$^*$\\newline {\\bf DBLP10K} & ---\\\\\n\t\t\t\\hline \n\t\t\t{\\bf Multiplex} & Multiplex representation scheme (attributes and structure are clustered separately as layers and then combined via consensus ) & No/Yes & Medium\\newline Large & F1-score & Synthetic \\newline {\\bf ego-Twitter} \\newline {\\bf ego-Facebook} \\newline {\\bf ego-G+} & {\\bf CESNA} \\newline {\\bf 3NCD} \\\\\n\t\t\t\\hline\n\t\t\t{\\bf WCMFA} & Association matrix with weighting based on structure- and attributes-aware similarity & Depends on the partitions & Small & Rand index \\newline ARI \\newline NMI & {\\bf Consult}~\\newline \t\t\n\t\t\tLondon Gang \\newline\n\t\t\tMontreal Gang & {\\bf WMen} \\\\\n\t\t\t\\hline \t\t\n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}\n\\begin{remark}\nGeneral-purpose consensus clustering algorithms for {\\it multi-layer} networks\nare considered in .\n\\end{remark}","id":"a2976c7a-9d26-4a9e-9cb2-40daafc174e6","level":"subsection","origin_cites_number":14,"parent_id":"a4712ae6-1690-4727-8bbb-b44187674ee8","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Late fusion methods"],["subsection","Consensus-based methods"]],"subsections":[],"title":"Consensus-based methods"},{"cite_extract_rate":0.16666666666666602,"cites":[8484],"content":"\\label{class-lf-switch}\nThe only method included in this subclass (see Table~\\ref{tab:late2}) also deals with partitions obtained separately for structure and attributes but chooses a more ``preferable'' one instead of finding consensus. Namely, {\\bf Selection} switches from a structure-based to an attributes-based partition when the former one is {\\it ambiguous}. This refers to the case when the so-called {\\it estimated mixing parameter} $\\mu$ for the structure-based partition is less then a certain experimental value $\\mu_{\\lim}$ associated with a significant drop in clustering quality on synthetic networks . An interested reader can find the precise definitions of $\\mu$ and $\\mu_{\\lim}$ in .\n\\begin{table}[t]\n\\caption {Switch-based methods.} \\label{tab:late2} \n\\begin{center}\n\t{\\tiny\n\t\t\\begin{tabular}{|p{1.2cm}|p{3.0cm}|p{1.9cm}|p{1.5cm}|p{1.5cm}|p{2.3cm}|p{2cm}|}\n\t\t\t\\hline \n\t\t\tMethod & Fusing the partitions & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\n\t\t\t\\hline \n\t\t\t{\\bf Selection} & Switching between the partitions & Depends on the partitions & Medium & NMI \\newline Modularity & Synthetic LFR benchmark \\newline {\\bf DBLP84K} & {\\bf BAGC} \\newline {\\bf OCru} \\newline {\\bf SA-Cluster} \\newline HGPA-CSPA \\\\\n\t\t\t\\hline \n\t\t\\end{tabular}\n\t}\n\\end{center}\n\\end{table}","id":"edd7b709-6935-46b9-9d57-cbc247202cf8","level":"subsection","origin_cites_number":6,"parent_id":"a4712ae6-1690-4727-8bbb-b44187674ee8","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Late fusion methods"],["subsection","Switch-based methods"]],"subsections":[],"title":"Switch-based methods"},{"cite_extract_rate":0.347826086956521,"cites":[7062,1712,1714,1723,1724,1721],"content":"\\label{section9}\nThe information collected in Sections~\\ref{section6}--\\ref{section8} allows us to analyze the overall situation in the field. Particularly, we would like to determine which methods are {\\it state-of-the-art}\\footnote{According to Cambridge Dictionary, state-of-the-art means ``the best and most modern of its type''.}. This is probably what a researcher/data scientist facing the community detection problem in node-attributed social networks expects from our survey.\nWe start with observing the directed graph based on the information in Sections~\\ref{section6}--\\ref{section8} and showing method-method comparisons, see Figure~\\ref{fig:graph1}. It requires several comments though. First, there are methods in Sections~\\ref{section6}--\\ref{section8} that are not compared with others for community detection in node-attributed networks or compared with a few. For this reason, we include in the graph only nodes (representing methods) whose degree is at least two. This means that there are at least two comparison experiments with each method presented in Figure~\\ref{fig:graph1}. Note that 46 methods are shown in the graph of 75 classified in the survey. \nSecond, the directed edges in the graph show the existing method-method comparisons. For example, the directed edge from node {\\bf CESNA} to node {\\bf CODICIL} indicates that the authors of {\\bf CESNA} compared their method with {\\bf CODICIL} and showed that {\\bf CESNA} outperforms {\\bf CODICIL} in some sense (community detection quality, computational efficiency, etc.). This is applied to all edges in the graph.\n\\begin{figure}[b]\n\t\\label{fig:graph1}\n \\includegraphics[scale=0.95]{Graph1new.pdf}\n\t\\caption{The directed graph of existing method-method comparisons. Nodes (shown only those with degree $\\ge 2$) represent the methods classified in the present survey. The most influential methods (nodes with highest PageRank) are filled green so that the darker green means the higher PageRank.}\n\\end{figure}\nWhat is more, we used PageRank to detect the most important or, better to say, {\\it most influential methods} in the field. Nodes with the highest PageRank values are filled green in Figure~\\ref{fig:graph1} so that the darker green means the higher PageRank. It turns out that the most influential ones are (in the order they discussed in Sections~\\ref{section6}--\\ref{section8}):\n\\begin{itemize}\n\\item weight-based {\\bf SAC2}~ and {\\bf CODICIL}~ (Subsection~\\ref{class-ef-weight}),\n\\item node-augmented graph-based {\\bf SA-Cluster} , {\\bf Inc-Cluster} and {\\bf SA-Cluster-Opt}~ (Subsection~\\ref{class-ef-node-augmented-graph}),\n\\item {\\bf SAC1}~ modifying the Louvain objective function (Subsection~\\ref{class-sf-modifying}),\n\\item NNMF-based {\\bf SCI} and matrix compression-based {\\bf PICS} (Subsection~\\ref{class-sf-nnmf}),\n\\item pattern mining-based (simultaneous fusion) {\\bf DCM} (Subsection~\\ref{class-sf-patterns}),\n\\item probabilistic model-based {\\bf PCL-DC} , {\\bf BAGC} , {\\bf GBAGC} , {\\bf CESNA} and {\\bf Circles} (Subsection~\\ref{class-sf-probabilistic-models}).\n\\end{itemize}\nIn our opinion, these methods may be though to be chosen by researchers' community as those determining further developments in the field. Thus we encourage a newcomer in the field to get familiar with them first to see the main ideas and techniques existing for community detection in node-attributed social networks.\nAt the same time, we would not consider the most influential methods as state-of-the-art as other methods are shown to outperform them in some sense.\nWhat else the graph in Figure~\\ref{fig:graph1} can tell us about? We emphasize that it does not contain 29 methods of those discussed in Sections~\\ref{section6}--\\ref{section8} and furthermore it is rather sparse and even disconnected. These points lead to the conclusion that the comparison study of the methods in the field is far from being complete. We would even strengthen the conclusion made by saying that a researcher/data scientist cannot be sure that the method chosen for practical use is preferable to other existing ones, even if there is an edge in Figure~\\ref{fig:graph1}. The problem is that the comparison experiments (represented via the edges) are made by different means, e.g. different quality measures, datasets, hyperparameter tuning strategies, etc. Let us discuss it below in more detail.\nSuppose for a second that two methods show the same community detection quality for a number of datasets and measures, and their hyperparameters are tuned to provide the best possible results. Then we should think how much time/space each method uses for it. Particularly, we may think of method's computational complexity in terms of the number of vertices $n$, the number of edges $m$ and the attribute dimension $d$ in a node-attributed graph. Such estimates exist for certain methods, particularly for some of the most influential ones. Examples are {\\bf CODICIL}~ with $O(n^2\\log n)$, {\\bf SA-Cluster} with $O(n^3)$ and {\\bf CESNA} with $O(m)$. However, we could not find such estimates for the majority of methods discussed in Sections~\\ref{section6}--\\ref{section8} as authors often omit such estimation. This makes the overall comparison of methods in terms of computational complexity impossible.\nNow let us discuss the hyperparameter tuning problem. It turns out that some authors tune hyperparameters in their methods manually, some just do not consider the problem at all. Another issue is the lack of a general understanding how to determine ``equal impact'' of structure and attributes on the community detection results. For example, in the weight-based early fusion methods (Subsection~\\ref{class-ef-weight}) some authors choose $\\alpha=1/2$ in experiments hoping that this provides the equal impact. However, if ones takes into account the different nature of structural and attributive information and the disbalance between associated statistical features, this choice seems questionable in a general situation.\nFurthermore, authors use different datasets (of various size and nature, see Section~\\ref{section5}) and quality measures to test their methods so that a unified comparison of experimental results in different papers cannot be carried out\\footnote{A more general observation is that several comparison experiments clearly do not provide generality of conclusions. This however seems to be a hot topic among supporters of scientific research from one side and of empirical one from another side.}. What is more, datasets and software implementations used in comparison experiments are rarely provided by the authors (especially of out-of-date papers) thus making reproducing their results time-consuming or even impossible. Note how few links to source codes are indicated in Sections~\\ref{section6}--\\ref{section8}.\nLet us now discuss the methodology of using quality measures in comparison experiments. There exist two main strategies (see Section~\\ref{section5}). Namely, community detection quality can be evaluated (a) by heuristic measures of structural closeness and attributes homogeneity (e.g. Modularity and Entropy) when the dataset under consideration has no ``ground truth'' and (b) by measures estimating the agreement between the detected communities and the ground truth ones (e.g. NMI or ARI).\nLet us first discuss Option (a) in terms of a typical weight-based early fusion method from Subsection~\\ref{class-ef-weight}. For a given node-attributed graph $G$, suppose that we first convert its attributes $(\\mathcal{V},\\mathcal{A})$ into attributive graph $G_A$ and further fuse it with the structure $(\\mathcal{V},\\mathcal{E})$. This results in weighted graph $G_W$ that is thought to contain information about both the structure and the attributes in a unified form. After that, we find communities in $G_W$ that provide, say, the maximum of Modularity of $G_W$ (e.g. by Weighted Louvain). At the evaluation step we further calculate Modularity of $(\\mathcal{V},\\mathcal{E})$ and Entropy of $(\\mathcal{V},\\mathcal{A})$ basing on the partition found. This seems reasonable as far as you do not look at such a methodology critically. Indeed, note that we deal with one quality measure within the optimization process (Modularity of $G_W$) but evaluate community detection performance by other measures (Modularity of $(\\mathcal{V},\\mathcal{E})$ and Entropy of $(\\mathcal{V},\\mathcal{A})$). Generally speaking, how can we be sure that optimization of one objective function provides optimal values of other objective functions, if there is no mathematically established connection between them? Of course, this may be simply called a heuristic but it looks more a logical gap, from our point of view. Anyways, we are unaware of any explicit explanation of such a methodology. What is more, one should carefully study how the change of data representation within a method (the change of vectors $\\mathcal{A}$ for edge weights in $G_A$ in the above-mentioned example) affects community detection results.\nTake into account that a similar discussion is suitable for the majority of methods in Sections~\\ref{section6}--\\ref{section8} that use quality measures for estimating structural closeness and attributes homogeneity of the detected communities. The exception is the methods that directly optimize the quality measures {\\it within the optimization process}. Examples are the simultaneous fusion methods {\\bf OCru} , \n{\\bf SAC1} ,\n{\\bf ILouvain} , \n{\\bf UNCut} , \n{\\bf MOEA-SA} ,\n{\\bf MOGA-@Net} and \n{\\bf JCDC} .\nJust by construction, they aim at providing optimal values of the quality measures. One should however take into account the precision of the optimization method applied. \nNow we turn our attention to the methodology of evaluating community detection quality by measures that estimate the agreement between the detected communities and the ground truth ones (Option (b)). In our opinion, such a methodology makes sense for synthetic node-attributed social networks as the way how the communities are formed is known in advance. As for real-world networks, it seems somewhat questionable as ground truth communities may reflect only one point of view on network communities (among many possible). Therefore, expecting that a community detection method would share this point of view seems a bit unsuitable. There are several works on this issue, e.g. , discussing connections between ground truth, attributes and quality measures in detail, and therefore we refer the interested reader to them.\nRecall that we started this discussion trying to determine the state-of-the-art methods for community detection in node-attributed social networks. Unfortunately, the above-mentioned facts do not give us a chance to do this. Of course, we could simply list the most recent methods in the field (and then it is enough to check just the time of publication), but this certainly does not meet the requirements imposed on state-of-the-art methods.","id":"ec3f6cdd-43b4-49d2-ab2f-b07a25b101ed","level":"section","origin_cites_number":23,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Analysis of the overall situation in the field"]],"subsections":[],"title":"Analysis of the overall situation in the field"},{"cite_extract_rate":0,"cites":[],"content":"\\label{section10}\nIt is shown in the survey that there exist a large amount of methods for community detection in node-attribute social networks based on different ideas. In particular, we gave short descriptions of 75 most relevant methods and mentioned much more those partly related to the topic (Sections~\\ref{section6}--\\ref{section8}).\nWe also proposed to divide the methods into the three classes --- early fusion, simultaneous fusion and late fusion ones (Section~\\ref{section4}). This classification is based on the moment when network structure and attributes are fused within a method and allows a researcher/data scientist to estimate the ease of method's software implementation. Namely, we concluded that early and late fusion methods can be easier implemented in comparison with simultaneous fusion ones as the former two usually can be combined of several classical community detection algorithms with existing implementations. At a lower lever, we also divided the methods into subclasses of fusion techniques used (Section~\\ref{section4}). This allows one to estimate the methodological variety in the field.\nWithin the classification, we also focused on the experimental part so that one can see which datasets and measures are used for quality evaluation of each method from Sections~\\ref{section5}--\\ref{section8}.\nThe analysis of all the information collected brought us to the unfortunate conclusion that it is impossible now to determine state-of-the-art methods for community detection in node-attributed social networks (Section~\\ref{section9}). This is a result of the presence of the following general problems in the field that we disclosed in the survey:\n \\begin{itemize}\n \\item the terminology in the field is rather unstable (Subsection~\\ref{subsection 3.2});\n \\item there is no generally accepted opinion on the effect of fusing structure and attributes, in particular, on when the fusion is helpful and when not in terms of subclasses of node-attributed social networks (Subsection~\\ref{subsection-str-closeness-attr-homo}); \n\\item there is no unified methodology for experimental comparison of methods that would include estimation of computational complexity, use of a unified set of datasets and quality measures for evaluation, and justified hyperparameter tuning procedures (Section~\\ref{section9});\n\\item there is no general understanding what is the ``equal impact'' of structure and attributes on community detection results (Section~\\ref{section9}),\n \\item as a rule, there is no mathematically established connection between computational processes within a community detection method and the quality measured used for its evaluation (Section~\\ref{section9}).\n \\end{itemize}\nSummarizing, we concluded that the comparison study allowing one to determine the most preferable (in any sense) methods in the field is far from being complete.\nNevertheless, community detection methods dealing both with network structure and attributes remain a powerful tool in social network analysis and can yield useful insights about a node-attributed social network to a researcher/data scientist. Furthermore, they have wide applications even beyond social networks (Section~\\ref{section3}). With respect to these, we believe that the formulation of existing problems in the field done in this survey is the first step in finding solutions to them.\n\\section*{Competing Interests Statement}\nThere are no competing interests in publication of this paper.\n\\section*{Acknowledgements}\nThe author is very grateful to Klavdiya Bochenina and Timofey Gradov for numerous conversations on the topic and especially to the Anonymous Referee for useful comments that helped to improve the quality of exposition in the survey essentially.\nThis research is financially supported by Russian Science Foundation, Agreement 17-71-30029 with co-financing of Bank Saint Petersburg, Russia.\n{\\small\n\\bibliographystyle{acm}\n\\bibliography{sample-base}\n}\n\\end{document}","id":"7362ce42-549a-4724-9b4c-b963b4574449","level":"section","origin_cites_number":0,"parent_id":"4b2a7071-f315-4436-8aea-91d8d6db7afb","prefix_titles":[["title","Community detection in node-attributed social networks: a~survey"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ]\n ],\n \"subsections\": [\n \"fb35c7bc-cc42-4de0-92ed-61876f9f332b\",\n \"f91b6565-cd82-4d66-b50d-4e9374bb4beb\",\n \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"124afd9d-8240-4dd5-9e90-68d0d18e9a07\",\n \"c0718d2b-4a41-48ee-939a-4ab8664cabe6\",\n \"1c6b434c-2289-4e6e-9627-ef1149463588\",\n \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"a4712ae6-1690-4727-8bbb-b44187674ee8\",\n \"ec3f6cdd-43b4-49d2-ab2f-b07a25b101ed\",\n \"7362ce42-549a-4724-9b4c-b963b4574449\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8481,\n 1710\n ],\n \"content\": \"Community detection is a fundamental problem in social network analysis consisting, roughly speaking, in unsupervised dividing social actors into densely knitted and highly related groups with each group well separated from the others. One class of classical community detection methods mainly deal only with the {\\\\it structure} of social networks (i.e. with connections between social actors) and ignore actors' features. There exist a variety of such structure-aware methods that have shown their efficiency in multiple applications (see ). However, the majority of real-world social networks provide more information about social actors than just connections between them. Indeed, it is rather common that certain actors' {\\\\it attributes} such as age, gender, interests, etc., are available. When it is so, the social network is called {\\\\it node-attributed} (recall that the actors are represented via nodes). According to , attributes form the second dimension, besides the structural one, in social network representation. There is another class of classical community detection methods (being opposite to the structure-aware ones, in a sense) that use only node attributes to detect communities and completely ignore connections between social actors. A representative of the attributes-aware methods is well-known $k$-means clustering algorithm taking attribute vectors as an input. Clearly, methods that deal only with structure or only with attributes do not use all the information available in a node-attributed social network. Naturally, this issue can be overcome if a method would somehow jointly use structure and attributes while detecting communities. Developing of such methods became a novel field in social network analysis . The field is moreover promising as the joint usage is believed to clarify and enrich the knowledge about social actors and to describe the powers that form their communities . \\nDuring the last decade numerous methods based on different ideas and techniques have appeared in the field. Although there exist some partial overviews of them, especially in Related Works sections of published papers and in the survey published in 2015, a recent summary of the subject is a necessity as the growing number of the methods may cause repetitions in methodology and uncertainty in practice. \\nIn this survey, we aim at describing and clarifying the overall situation in the field. Namely, we perform an exhaustive search of existing community detection methods for node-attributed social networks. What is more, we propose a classification of them based on when and how they use and fuse network structure and attributes. We not only give a description of each class but also provide general technical ideas behind each method in the class. Furthermore, we pay attention to available information which methods outperform others and which datasets and quality measures are used for their performance evaluation. Basing on the information collected, we make conclusions on the current state of the field and disclose several problems that seem important to be resolved in future. \\nTo be more precise, let us describe the content of the survey. In Section~\\\\ref{section2}, we first provide the reader with the notation used in the survey and state the problem of community detection in node-attributed social networks. We further briefly discuss the traditional argumentation in support of such a community detection and the effect of fusing network structure and attributes. In\\nSection~\\\\ref{section3}, we give information about the related survey works and explain how the search of relevant literature was organized in our case. We also indicate which methods are included in the survey and which are not. Additionally, we explain why references throughout the survey are made in a certain way. Section~\\\\ref{section4} introduces the classification that we propose for community detection methods under consideration. In Section~\\\\ref{section5}, we discuss the most popular datasets and quality measures for evaluation of community detection results. This section is also helpful for simplifying exposition in the forthcoming sections. Sections~\\\\ref{section6}--\\\\ref{section8} contain descriptions of the classes of methods and their representatives. In Section~\\\\ref{section9}, we analyze the overall situation in the field basing on the information from Sections~\\\\ref{section6}--\\\\ref{section8}. Among other things, we disclose several methodological problems that are important to resolve in future studies, in our opinion. Our conclusions on the topic are summarized in Section~\\\\ref{section10}.\",\n \"id\": \"fb35c7bc-cc42-4de0-92ed-61876f9f332b\",\n \"level\": \"section\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section2}\",\n \"id\": \"f91b6565-cd82-4d66-b50d-4e9374bb4beb\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes\"\n ]\n ],\n \"subsections\": [\n \"796f6d71-50d7-49c9-9b0b-67ff77b34e9c\",\n \"c2f50602-97d7-4993-94e6-9400d76e89e5\"\n ],\n \"title\": \"Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"We represent a node-attributed social network as triple ({\\\\it node-attributed graph}) $G=(\\\\mathcal{V},\\\\mathcal{E},\\\\mathcal{A})$, where $\\\\mathcal{V}=\\\\{v_i\\\\}$ is the set of nodes (vertices) representing social actors, $\\\\mathcal{E}= \\\\{e_{ij}\\\\}$ the set of edges representing connections between the actors ($e_{ij}$ stands for the edge between nodes $v_i$ and $v_j$), and $\\\\mathcal{A}$ the set of attribute vectors $A(v_i)=\\\\{a_{k}(v_i)\\\\}$ associated with nodes in $\\\\mathcal{V}$ and containing information about actors' features. Furthermore, $|\\\\mathcal{V}|=n$, $|\\\\mathcal{E}|=m$ and the dimension of the attribute vectors is $d$. The domain of $a_k$, i.e. the set of possible values of $k$th element of attribute vectors $a_k(v_i)$, is denoted by $dom(a_k)$. In these terms, $k$th attribute of node $v_i$ is referred to as $a_k(v_i)$. The notation introduced above is summarized in Figure~\\\\ref{fig:notation}. Note that pairs $(\\\\mathcal{V},\\\\mathcal{E})$ and $(\\\\mathcal{V},\\\\mathcal{A})$ are correspondingly called the {\\\\it structure} (or {\\\\it topology}) and the {\\\\it attributes} (or {\\\\it semantics}) of node-attributed graph $G$.\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{notation.pdf}\\n\\t\\\\caption{Notation related to the triple (node-attributed graph) $G=(\\\\mathcal{V},\\\\mathcal{E},\\\\mathcal{A})$, where $\\\\mathcal{V}=\\\\{v_i\\\\}$, $\\\\mathcal{E}=\\\\{e_{ij}\\\\}$ and $\\\\mathcal{A}=\\\\{A(v_i)\\\\}$.}\\n\\t\\\\label{fig:notation}\\n\\\\end{figure}\\nBy {\\\\it community detection\\\\footnote{It is also called ``community discovery'' or ``clusterization''.} in node-attributed graph} $G=(\\\\mathcal{V},\\\\mathcal{E},\\\\mathcal{A})$ we mean {\\\\it unsupervised} partitioning the set of nodes $\\\\mathcal{V}$ into $N$ subsets ({\\\\it communities} or {\\\\it clusters}) $C_k\\\\subset \\\\mathcal{V}$, with $C=\\\\{C_k\\\\}_{k=1}^N$ called a {\\\\it partition}, such that\\n$\\\\mathcal{V}=\\\\bigcup_{k=1}^N C_k$ and a certain balance between the following two properties is achieved:\\n\\\\begin{itemize}\\n\\t\\\\item[(a)] {\\\\it structural closeness}, i.e. nodes within a community are structurally close to each other, while nodes in different communities are not;\\n\\t\\\\item[(b)] {\\\\it attribute homogeneity}, i.e. nodes within a community have homogeneous attributes, while nodes in different communities do not.\\n\\\\end{itemize}\\nSince one can meet variations of the above-mentioned definitions in the relevant literature, it is worth giving several comments on them. First, the number of communities $N$ can be either known in advance or determined during the community detection process automatically. Second, the communities $C_k$ may be defined to be disjoint or overlapping. Third, the property $\\\\mathcal{V}=\\\\bigcup_{k=1}^N C_k$ is sometimes omitted if the resulting partition is not required to include all the nodes from $\\\\mathcal{V}$. Fourth, \\nthe notion of structural closeness and attribute homogeneity may seem vague at the moment but hopefully become more evident after Subsection~\\\\ref{subsection-str-closeness-attr-homo} and Section~\\\\ref{section4} where reasons and particular measures for them are discussed. Fifth, the definitions given are for the case of nodes and edges each of one type. It is of course possible that social actors and connections between them can be of different types in a social network and thus one should take this heterogeneity into account. This situation is however closer to the notion of multi-layer networks that are out of scope of the present survey (see also Section~\\\\ref{section3}).\",\n \"id\": \"796f6d71-50d7-49c9-9b0b-67ff77b34e9c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"f91b6565-cd82-4d66-b50d-4e9374bb4beb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes\"\n ],\n [\n \"subsection\",\n \"Necessary notation and the community detection problem statement\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Necessary notation and the community detection problem statement\"\n },\n {\n \"cite_extract_rate\": 0.36363636363636304,\n \"cites\": [\n 8483,\n 1712,\n 1710,\n 8482,\n 1711,\n 8484,\n 217,\n 9116\n ],\n \"content\": \"\\\\label{subsection-str-closeness-attr-homo}\\nThe structural closeness requirement is based on the recent concepts of a (structural) community in a social network. For example, communities are thought in as subsets of nodes with dense connections within the subsets and sparse in between. In its turn, adopts the intuition that nodes within the same community should be better connected than they would be by chance. A possible measure for that is famous Newman's Modularity that has become an influential tool for structure-based community detection in social networks . Multiple Modularity modifications and other measures have been also proposed to assess structural closeness . In fact, the precise meaning of structural closeness in each community detection method is determined by the measure chosen.\\nThe attribute homogeneity requirement is based on the social science founding (see e.g. ) that actors' features can reflect and affect community structure in social networks. The well-known principle of homophily in social networks states that like-minded social actors have a higher likelihood to be connected . Thus community detection process taking into account attribute homogeneity may provide results of better quality . Oppositely to the situation with structural closeness measures, the attribute homogeneity is usually measured by Entropy that quantifies the degree of disorder of attribute vectors in $(\\\\mathcal{V},\\\\mathcal{A})$ within the communities detected.\\nLet us now discuss different points of view on the effect of fusing structure and attributes.\\nFrom one side, multiple experiments, e.g. in and many other papers cited in this survey, suggest that the structure and the attributes of a node-attributed social network often provide complementary information that improves community detection quality. For example, attributes may compensate the structural sparseness of a real-world social network , while structural information may be helpful in resolving the problem of missing or noisy attributes . What is more, it is observed in that structure-only or attributes-only community detection is often not as effective as when both sources of information are used. From the other side, some experiments (see e.g. ) suggest that this is not always true, and network structure and attributes may be orthogonal and contradictory thus leading to ambiguous community detection results. Moreover, relations between these sources of information may be highly non-linear and challenging to analyze . \\nBesides the above-mentioned points, our general impression is that there is no widely accepted opinion on the effect of fusing structure and attributes and how this fusion can influence community detection quality. Let us illustrate this with an example.\\nSuppose that the structure of a certain node-attributed social network is ideally correlated with a half of attributes and is wholly orthogonal to another half. For simplicity, let the dimension of attribute vectors be small so that there is no sense to fight against the curse of dimensionality. Now we follow the popular suggestion that the mismatch between structure and attributes negatively affects community detection quality and that the existence of structure-attributes correlation offers ``a unique opportunity to improve the learning performance of various graph mining tasks'' . The choice is clear then: we need to use the structure and only the ideally correlated attributes for community detection. It turns out however that we are going to use two sources of information that mostly duplicate each other. Why should we expect that this improves the quality of detected communities? From our side, we would presume that the structure and the chosen half of attributes (considered separately or jointly) would yield very similar or even the same communities, with all the ensuing consequences for assessing their quality. Shouldn't we use just one of the sources then? Furthermore, it is not clear to us why the other half of attributes should be omitted. Generally speaking, they may contain valuable information for community detection and thus omitting them because of the lack of correlation with the structure is rather questionable.\\nIn any case, a focused theoretical study of when the fusion of structure and attributes is worthy and when not for community detection (ideally, in terms of subclasses of node-attributed social networks) seems to be an extremely important thing that would allow to remove the above-mentioned contradictions.\",\n \"id\": \"c2f50602-97d7-4993-94e6-9400d76e89e5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"f91b6565-cd82-4d66-b50d-4e9374bb4beb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Community detection problem for node-attributed social networks and the~effect of fusing network structure and attributes\"\n ],\n [\n \"subsection\",\n \"Structural closeness and attribute homogeneity. The effect of fusing structure and attributes\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Structural closeness and attribute homogeneity. The effect of fusing structure and attributes\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section3}\",\n \"id\": \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Related works and processing the relevant literature\"\n ]\n ],\n \"subsections\": [\n \"e751245c-7c36-43ca-8918-65a42b01d47c\",\n \"b12abaa9-ebec-4b58-8134-5afb1ff9a9f1\",\n \"f0793b92-de74-4aae-8b54-7257a1189cf5\",\n \"e38efd05-8a77-49a3-ba2a-eff593b2c129\",\n \"29a40dbb-fbd7-48bc-96e3-b7914c7e2c83\",\n \"32b01039-79ed-4c11-816c-f79bd32ba71d\"\n ],\n \"title\": \"Related works and processing the relevant literature\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 8481,\n 1713,\n 1714,\n 8485,\n 1710\n ],\n \"content\": \"There is a variety of surveys and comparative studies considering community detection in social networks without attributes, in particular, . At the same time, the survey \\t seems to be the only one on community detection in node-attributed social networks. Obviously, since it was published in 2015, many new methods adapting different fusion techniques have appeared in the field. Furthermore, a big amount of the methods that had been proposed before 2015 are not covered by , in particular, some based on objective function modification, non-negative matrix factorization, probabilistic models, clustering ensembles, etc. In a sense, the technique-based classification of methods in is also sometimes confusing. For example, CODICIL , a method based on assigning attribute-aware weights on graph edges, is not included in \\\\cite[Section 3.2. Weight modification according to node attributes]{Bothorel2015}, but to \\\\cite[Section 3.7. Other methods]{Bothorel2015}. Although is a well-written and highly cited paper, a recent survey of community detection methods for node-attributed social networks is clearly required.\\nBesides , almost every paper on the topic contains a Related Works section. It typically has a short survey of preceding methods and an attempt to classify them. We observed that many authors are just partly aware of the relevant literature and this sometimes leads to repetitions in approaches. Furthermore, multiple classifications (usually technique-based) are mainly not full and even contradictory.\",\n \"id\": \"e751245c-7c36-43ca-8918-65a42b01d47c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Related works and processing the relevant literature\"\n ],\n [\n \"subsection\",\n \"Related works\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Related works\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsection 3.2}\\nAt the beginning, we started the search of relevant literature using regular and scientific search engines making the queries like ``community detection'' or ``clustering'' or\\n``community discovery'' in ``node-attributed social networks'' or ``node-attributed graphs''. Within the search process it became evident that other queries also lead to the relevant literature. In particular, ``clustering an attribute-aware graph'', ``community detection in networks with node attributes'', ``description-oriented community detection'', ``semantic clustering of social networks'', ``structure and attributes community detection'', ``joint cluster analysis of attribute and relationship data'', ``community discovery integrating links and tags'', ``attributed graph partitioning'', ``node attribute-enhanced community detection'', ``community detection based on structure and content'', etc. It can be also observed that node-attributed networks and graphs are also sometimes called ``augmented networks'', ``graphs with feature vectors'', ``feature-rich networks'' and ``multi-attributed graphs''.\\nThis variety of terms suggests that there is still no established terminology in the field and emphasizes the significance of our survey, where we try to use consistent terminology.\\nAfter the above-mentioned exhaustive search, we learned the references in the found papers. Among other things, it brought us to ideologically close papers devoted, for example, to ``attributed information networks'', ``annotated document networks'', ``multi-layer networks'' and ``subspace-based clustering''. We stopped further search when we could not find any new relevant references. Since this happened in the middle of 2019, the survey covers the found papers that had been published in journals or conference proceedings before this date.\",\n \"id\": \"b12abaa9-ebec-4b58-8134-5afb1ff9a9f1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Related works and processing the relevant literature\"\n ],\n [\n \"subsection\",\n \"Relevant literature search process\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Relevant literature search process\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"It turns out that several methods for community detection in node-attributed social networks can be proposed in one paper. Therefore, a regular reference of the form [ReferenceNumber] may be not informative enough. From the other hand, authors usually provide their methods with short names like SA-Cluster, CODICIL or CESNA\\\\footnote{If this is not the case, we allowed ourselves to invent our own names suggesting the class the method belongs to in our classification.}. Some of the names are rather familiar to researches in the field. Thus it seems reasonable to make a reference of the form MethodName [ReferenceNumber] and so we do in what follows. However, not all the methods that are mentioned in the survey are included in our classification and thus discussed in a more detailed manner (as some are just out of scope). To distinguish the cases, we write names of the methods included in our classification in {\\\\bf bold}, e.g. {\\\\bf SCMAG} , {\\\\bf UNCut} and {\\\\bf DCM} . Such a format means that the reader can find short descriptions of the methods {\\\\bf SCMAG} , {\\\\bf UNCut} and {\\\\bf DCM} in our survey. References like DB-CSC , FocusCO and ACM mean that the corresponding methods are not included in our classification. In this case the reader is recommended to go directly to the papers , and to get additional information. \\nA similar scheme is applied to names of the node-attributed network datasets discussed in Section~\\\\ref{section5} and used in further classification. The reader is assumed to have in mind that if a dataset name is written in {\\\\bf bold}, then its description can be found in Tables~\\\\ref{tab_dataset1}, \\\\ref{tab_dataset2} or \\\\ref{tab_dataset3}. Note also that various versions of the datasets from Tables~\\\\ref{tab_dataset1}, \\\\ref{tab_dataset2} or \\\\ref{tab_dataset3} are in fact used in different papers. To show that a dataset is somehow different from the description in Tables~\\\\ref{tab_dataset1}, \\\\ref{tab_dataset2} or \\\\ref{tab_dataset3}, we mark it by *. For example, a DBLP dataset with the number of nodes and edges different from {\\\\bf DBLP10K} and {\\\\bf DBLP84K} in Table~\\\\ref{tab_dataset2} is denoted by {\\\\bf DBLP*} in Sections~\\\\ref{section6}--\\\\ref{section8}.\",\n \"id\": \"f0793b92-de74-4aae-8b54-7257a1189cf5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Related works and processing the relevant literature\"\n ],\n [\n \"subsection\",\n \"The format of references to methods and datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"The format of references to methods and datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In the survey, we do not consider community detection methods for multi-layer networks, where different types of vertices and edges may present at different layers . Nevertheless, we mention some of these methods from time to time in corresponding remarks. It is though important to note that node-attributed networks of different nature may be clearly considered as a particular case of multi-layer ones. In the majority of papers covered by the present survey, this connection is however rarely commented. \\nLet us also emphasize that multi-layer networks (graphs) require special analysis taking into account the heterogeneity of vertices and edges on different layers. A~separate survey and an extensive comparable study of such methods is an independent and useful task (see partial overviews e.g. in ).\",\n \"id\": \"e38efd05-8a77-49a3-ba2a-eff593b2c129\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Related works and processing the relevant literature\"\n ],\n [\n \"subsection\",\n \"Note on multi-layer network clustering\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Note on multi-layer network clustering\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Following the above-mentioned definition of community detection in node-attributed social networks, we mainly consider in the survey the methods that can use the full attribute space and find communities covering the whole network. However, there is a big class of special methods that explore subspaces of attributes and/or deal with subgraphs of an initial graph, e.g. GAMer , DB-CSC , SSCG , FocusCO and ACM .\\nThe main idea behind the subspace-based (also known as projection-based) clustering methods is that not all available semantic information (attributes) is relevant to obtain good-quality communities . For this reason, one has to somehow choose the appropriate attribute subspace to avoid the so-called {\\\\it curse of dimensionality} (see \\\\cite[Section 3.2]{Bothorel2015}) and reveal significant communities that would not be detected if all available attributes were considered.\\nTo be precise, some of the methods that we discuss below partly use this idea, e.g. {\\\\bf WCru} (cf. the definition of a {\\\\it point of view} in the papers), {\\\\bf DVil} , {\\\\bf SCMAG} , {\\\\bf UNCut} , {\\\\bf DCM} , etc., but still can work with the full attribute space. In any case, a separate survey on subspace-based methods for community detection in node-attributed social networks would be a very valuable complement to the current one.\",\n \"id\": \"29a40dbb-fbd7-48bc-96e3-b7914c7e2c83\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Related works and processing the relevant literature\"\n ],\n [\n \"subsection\",\n \"Note on subspace-based clustering\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Note on subspace-based clustering\"\n },\n {\n \"cite_extract_rate\": 0.30000000000000004,\n \"cites\": [\n 1715,\n 1714,\n 1712\n ],\n \"content\": \"Clearly, community detection tools for node-attributed social networks are suitable for networks of different nature. That is why, besides obvious applications in marketing (recommender systems, targeted advertisements and user profiling) , the tools are used for search engine\\noptimization and spam detection , in counter-terrorist activities and disclosing fraudulent schemes . They are also applied to analysis of protein-protein interactions, genes and epidemics .\\nAnother possible application is document network clustering. Note that such a clustering is historically preceding to community detection in node-attributed social networks and is rich methodologically on its own . One should take into account however that social communities although have similar formal description with document clusters have inner and more complicated forces to be formed and act. What is more, it has been shown that methods for community detection in node-attributed social networks outperform preceding methods for document network clustering in many cases, see . For this reason, we do not consider document network clustering methods in this survey.\",\n \"id\": \"32b01039-79ed-4c11-816c-f79bd32ba71d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"f7c5bcf3-a055-4452-990c-b315c5512443\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Related works and processing the relevant literature\"\n ],\n [\n \"subsection\",\n \"Note on community detection in node-attributed networks of different type and its applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Note on community detection in node-attributed networks of different type and its applications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section4}\\nIn previous works, the classification of methods for community detection in node-attributed social networks was done mostly with respect to the techniques used (e.g. distance-based or random walk-based). We partly follow this principle but at a lower level. At the upper level we group the methods by when structure and attributes are fused with respect to the community detection process, see Figure~\\\\ref{fig:pics_clafficication}. Namely, we distinguish the classes of \\n\\t\\\\begin{itemize}\\n\\t\\t\\\\item {\\\\it early fusion methods} that fuse structure and attributes before the community detection process,\\n\\t\\t\\\\item {\\\\it simultaneous fusion methods} that fuse structure and attributes simultaneously with the community detection process,\\n\\t\\t\\\\item {\\\\it late fusion methods} that first partition structure and attributes separately and further fuse the partitions obtained.\\n\\t\\\\end{itemize}\\n\\t\\\\begin{figure}[b]\\n\\\\begin{minipage}[t]{0.325\\\\textwidth}\\n \\\\includegraphics[width=\\\\textwidth]{early.pdf}\\n \\\\centering \\n {\\\\bf Early fusion methods}\\\\\\\\ (Section~\\\\ref{section6})\\n \\\\vspace{0.3cm}\\n \\\\hrule\\n \\\\vspace{0.3cm}\\n Weight-based (Section~\\\\ref{class-ef-weight}) \\n \\\\vspace{0.3cm}\\n Distance-based (Section~\\\\ref{class-ef-distance})\\n \\\\vspace{0.3cm}\\n Node-augmented graph-based (Section~\\\\ref{class-ef-node-augmented-graph}) \\n \\\\vspace{0.3cm}\\n Embedding-based (Section~\\\\ref{class-ef-embedding})\\n \\\\vspace{0.3cm}\\n Pattern mining-based (Section~\\\\ref{class-ef-patterns})\\n\\\\end{minipage}\\n\\\\begin{minipage}[t]{0.325\\\\textwidth}\\n \\\\includegraphics[width=\\\\textwidth]{sim.pdf}\\n \\\\centering \\n {\\\\bf Simultaneous fusion methods} (Section~\\\\ref{section7})\\n \\\\vspace{0.3cm}\\n \\\\hrule\\n \\\\vspace{0.3cm}\\n Methods modifying objective functions of classical clustering algorithms (Section~\\\\ref{class-sf-modifying}) \\n \\\\vspace{0.3cm}\\n Metaheuristic-based (Section~\\\\ref{class-sf-mataheuristic})\\n \\\\vspace{0.3cm}\\n Methods based on non-negative matrix factorization and matrix compression (Section~\\\\ref{class-sf-nnmf}) \\n \\\\vspace{0.3cm}\\n Pattern mining-based (Section~\\\\ref{class-sf-patterns}) \\n \\\\vspace{0.3cm}\\n Probabilistic model-based (Section~\\\\ref{class-sf-probabilistic-models}) \\n \\\\vspace{0.3cm}\\n Dynamical system-based and agent-based (Section~\\\\ref{class-sf-dynamics})\\n\\\\end{minipage}\\n\\\\begin{minipage}[t]{0.325\\\\textwidth}\\n \\\\includegraphics[width=\\\\textwidth]{scheme-late.pdf}\\n \\\\centering\\n {\\\\bf Late fusion methods}\\n (Section~\\\\ref{section8})\\n \\\\vspace{0.3cm}\\n \\\\hrule\\n \\\\vspace{0.3cm}\\n Consensus-based (Section~\\\\ref{class-lf-consensus})\\n \\\\vspace{0.3cm}\\n Switch-based (Section~\\\\ref{class-lf-switch})\\n\\\\end{minipage}\\n \\\\caption{The proposed classification and the survey structure guide.}\\n \\\\label{fig:pics_clafficication}\\n\\\\end{figure}\\nSuch a classification allows an interested researcher/data scientist to estimate the labour costs of software implementation of the method chosen for use in practice.\\nIndeed, early fusion methods require just a preprocessing (fusion) procedure converting information about structure and attributes into a format that is often suitable for classical community detection algorithms. For example, weight-based early fusion methods convert attribute vectors into the graph form and further merge the structure- and attributes-aware graphs into a weighted graph that can be processed by graph clustering algorithms with existing implementations. Implementation of late fusion methods is also rather simple. Namely, at the first step they detects communities by classical graph and vector clustering algorithms applied to structure and attributes separately. At the second step, the communities obtained are somehow merged by an existing (or at least easy-implemented) algorithm. Oppositely to the early and late fusion methods, simultaneous fusion ones require more programmer work as usually assume either a completely new implementation or essential modifications to existing ones. The situation is aggravated by the fact that their source codes are rarely available.\\nAs seen from Figure~\\\\ref{fig:pics_clafficication}, we also divide the methods within each class into technique-used subclasses. Let us emphasize that the priority in this division is the fusion technique. For example, by ``weight-based methods'' we mean those which form a weighted graph while fusing structure and attributes. The majority of such methods further use graph clustering algorithms for community detection (and this is reasonable). However, some may still transform the graph into the distance matrix form and then use distance-based clustering algorithms. Such methods are still called ``weight-based''. One more example: ``distance-based methods'' are called in this way as directly produce a distance matrix while fusing structure and attributes, independently on how this matrix is further processed.\\nWhat is more, we provide short descriptions of methods in each class and subclass in tables in Sections~\\\\ref{section6}--\\\\ref{section8}. In particular, we briefly describe the community detection algorithm and its input used in the method, and the type of communities obtained (overlapping or not). Furthermore, we mention which datasets and quality measured are used by method's authors for evaluation of community detection results. In addition, for each method we provide a list of other methods for {\\\\it community detection in node-attributed networks} (using both structure and attributes) that the method under consideration is compared with. Note that the list may be empty sometimes. This is so, for example, if the method under consideration is compared only with classical community detection methods that deal either with structure or attributes and do not fuse them to detect communities.\",\n \"id\": \"124afd9d-8240-4dd5-9e90-68d0d18e9a07\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Our classification of community detection methods for~node-attributed~ social~networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Our classification of community detection methods for~node-attributed~ social~networks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section5}\",\n \"id\": \"c0718d2b-4a41-48ee-939a-4ab8664cabe6\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Most used node-attributed network datasets and quality measures for community detection evaluation\"\n ]\n ],\n \"subsections\": [\n \"36da5b81-971c-4135-9405-fdc43b1406db\",\n \"01bdf590-8667-4b9e-b2aa-932d3480f52b\"\n ],\n \"title\": \"Most used node-attributed network datasets and quality measures for community detection evaluation\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 1714,\n 8486,\n 1716\n ],\n \"content\": \"The title of the survey suggests that we are focused on community detection in node-attributed {\\\\it social} networks. However, the methods that are included in our classification, generally speaking, may be applied to node-attributed networks of different nature. As we have noticed, authors of the methods implicitly share this point of view and freely use various node-attributed network datasets to evaluate community detection quality.\\nIn this subsection (Tables~\\\\ref{tab_dataset1}, \\\\ref{tab_dataset2} or \\\\ref{tab_dataset3}) we collect and briefly describe the datasets that are popular in the field\\\\footnote{An interested reader can find more node-attributed network datasets at \\\\href{http://www-personal.umich.edu/~mejn/netdata/}{Mark Newman page}, \\\\href{https://hpi.de/naumann/projects/repeatability/datasets/}{HPI Information Systems Group}, \\\\href{https://linqs.soe.ucsc.edu/data}{LINQS Statistical Relational Learning Group}, \\\\href{http://snap.stanford.edu/data/}{Stanford Large Network Dataset Collection}, \\\\href{http://dp.univr.it/~laudanna/LCTST/downloads/index.html}{University of Verona Laboratory of Cell Trafficking and Signal Transduction}, \\\\href{https://perso.liris.cnrs.fr/marc.plantevit/doku/doku.php?id=data_sets}{Marc Plantevit page}, \\\\href{https://toreopsahl.com/datasets/}{Tore Opsahl page}, \\\\href{https://sites.google.com/site/ucinetsoftware/datasets}{UCINET networks}, \\\\href{http://networkrepository.com}{Interactive Scientific Network Data Repository}, \\\\href{https://aminer.org/citation}{Citation Network Dataset}.\\n}. Recall that the dataset names written in {\\\\bf bold} in Sections~\\\\ref{section6}--\\\\ref{section8} refer to the tables in this subsection. \\nIt can be observed that Tables~\\\\ref{tab_dataset1}, \\\\ref{tab_dataset2} or \\\\ref{tab_dataset3} contain datasets based on social network data (e.g. Facebook, LastFM and Twitter) and document or citation network data (e.g. DBLP, Wiki and Patents). \\nFor convenience, we distinguish the datasets by size. Namely, by {\\\\it small}, {\\\\it medium} and {\\\\it large} we mean network datasets with $< 10^3$, $10^3\\\\ldots 10^5$ and $>10^5$ nodes, correspondingly.\\n\\\\begin{table}\\n\\t\\\\caption{Most popular small size datasets.}\\n\\t\\\\label{tab_dataset1} \\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.2cm}|p{12.9cm}|p{0.8cm}|} \\n\\t\\t\\t\\\\hline \\n\\t\\t\\tDataset & Description & Source\\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Political Books} & All books in this dataset were\\n\\t\\t\\tabout U.S. politics published during the 2004 presidential election\\n\\t\\t\\tand sold by Amazon.com. Edges between books means\\n\\t\\t\\ttwo books are always bought together by customers. Each\\n\\t\\t\\tbook has only one attribute termed as political persuasion, with\\n\\t\\t\\tthree values: 1) conservative; 2) liberal; and 3) neutrality & \\\\href{http://www-personal.umich.edu/~mejn/netdata/}{Link} \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf WebKB} & A classified network of 877 webpages (nodes) and 1608 hyperlinks (edges) gathered from four different universities Web sites (Cornell, Texas, Washington, and Wisconsin). Each web page is associated with a binary vector, whose elements take the value $1$ if the corresponding word from the vocabulary is present in that webpage, and $0$ otherwise. The vocabulary consists of 1703 unique words. Nodes are classified into five classes: course, faculty, student, project, or staff. & \\\\href{https://linqs.soe.ucsc.edu/data}{Link} \\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Twitter} & A collection of several tweet networks: 1) Politics-UK dataset is collected from Twitter accounts of 419 Members of Parliament in the United Kingdom in 2012. Each user has 3614-dimensional attributes, including a list of words repeated more than 500 times in their tweets. The accounts are assigned to five\\n\\t\\t\\tdisjoint communities according to their political affiliation.\\n\\t\\t\\t2) Politics-IE dataset is collected from 348 Irish\\n\\t\\t\\tpoliticians and political organizations, each user has 1047-\\n\\t\\t\\tdimensional attributes. The users are distributed into seven\\n\\t\\t\\tcommunities. 3) Football dataset contains 248 English\\n\\t\\t\\tPremier League football players active on Twitter which are\\n\\t\\t\\tassigned to 20 disjoint communities, each corresponding to\\n\\t\\t\\ta Premier League club. 4) Olympics dataset contains users\\n\\t\\t\\tof 464 athletes and organizations involved in the London\\n\\t\\t\\t2012 Summer Olympics. The users are grouped into 28\\n\\t\\t\\tdisjoint communities, corresponding to different Olympic\\n\\t\\t\\tsports. & \\\\href{http://mlg.ucd.ie/aggregation/}{Link 1} \\\\newline \\\\href{http://mlg.ucd.ie/networks/}{Link 2}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t{\\\\bf Lazega} & A corporate law partnership in a Northeastern US corporate law firm; possible attributes: (1: partner; 2: associate), office (1: Boston; 2: Hartford; 3: Providence); 71 nodes and 575 edges & \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Research} & A research team of employees in a manufacturing company; possible attributes: location (1: Paris; 2: Frankfurt; 3: Warsaw; 4: Geneva), tenure (1: 1--12 months; 2: 13--36 months; 3: 37--60 months; 4: 61+ months); 77 nodes and 2228 edges & \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Consult} & Network showing relationships between employees in a consulting company; possible attributes: organisational level (1: Research Assistant; 2: Junior Consultant; 3: Senior Consultant; 4: Managing Consultant; 5: Partner), gender (1: male; 2: female); 46 nodes and 879 edges & \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{table} \\n\\\\begin{table}\\n\\t\\\\caption{Most popular medium size datasets.}\\n\\t\\\\label{tab_dataset2} \\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.2cm}|p{12.9cm}|p{0.8cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tDataset & Description & Source \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Political Blogs} & A non-classified network of 1,490 webblogs (nodes) on US politics with\\n\\t\\t\\t19,090 hyperlinks (edges) between the webblogs. Each node has an attribute describing its political leaning as either liberal or\\n\\t\\t\\tconservative (represented by $0$ and $1$). & \\\\href{http://www-personal.umich.edu/~mejn/netdata/}{Link}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf DBLP10K} & A non-classified co-author network extracted from DBLP Bibliography (four research areas of database, data mining, information retrieval and artificial intelligence) with 10,000 authors (nodes) and their co-author relationships (edges). Each author is associated with two relevant categorical attributes: prolific and primary topic. For attribute ``prolific'', authors with $\\\\ge 20$ papers are labelled as highly prolific; authors with $>10$ and $<20$ papers are labelled as prolific and authors with $\\\\le 10$ papers are labelled as low prolific. Node-attribute values for ``primary topic'' (100 research topics) are obtained via topic modelling. Each extracted topic consists of a probability distribution of keywords which are most representative of the topic. & \\\\href{https://github.com/Issamfalih/ANCL/tree/master/data}{Link}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf DBLP84K} & A larger non-classified co-author network extracted from DBLP Bibliography (15 research areas of database, data mining, information retrieval, artificial intelligence, machine learning, computer vision, networking, multimedia, computer systems, simulation, theory, architecture, natural language processing, human-computer interaction, and programming language) with 84,170 authors (nodes) and their co-author relationships (edges). Each author is associated with two relevant categorical attributes: prolific and primary topic, defined in a similar way as in DBLP10. & \\\\href{https://github.com/Issamfalih/ANCL/tree/master/data}{Link}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Cora} & A classified network of machine learning papers with 2,708 papers (nodes) and 5,429 citations (edges). Each node is attributed with a $1433$-dimension binary vector indicating the absence/presence of words from the dictionary of words collected from the corpus of papers. The papers are classified into 7 subcategories: case-based reasoning,genetic algorithms,\\n\\t\\t\\tneural networks, probabilistic methods, reinforcement learning,\\n\\t\\t\\trule learning and theory.\\t & \\\\href{https://linqs.soe.ucsc.edu/data}{Link 1}\\\\newline \\\\href{https://github.com/albertyang33/TADW}{Link 2}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf CiteSeer} & A classified citation network in the field of machine learning with 3,312 papers (nodes) and 4,732 citations (edges). Each node is attributed with a binary vector indicating the absence/presence of the corresponding words from the dictionary of the 3,703 words collected from the corpus of papers. Papers are classified into 6 classes. & \\\\href{https://linqs.soe.ucsc.edu/data}{Link 1}\\\\newline \\\\href{https://github.com/albertyang33/TADW}{Link 2}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Sinanet} & A classified microblog user relationship network extracted from the sina-microblog website (http://www.weibo.com) with 3,490 users (nodes) and 30,282 relationships (edges). Each node is attributed with 10-dimensional numerical attributes describing the interests of the user. & \\\\href{https://github.com/smileyan448/Sinanet}{Link}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf PubMed Diabetes} & A classified citation networks extracted from the PubMed database pertaining to diabetes. It contains 19,717 publications (nodes) and 44,338 citations (edges). Each node is attributed by a TF-IDF weighted word vector from a dictionary that consists of 500 unique words. & \\\\href{https://linqs.soe.ucsc.edu/data}{Link} \\\\\\\\ \\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{\\\\bf Facebook100} & A non-classified Facebook users network with 6,386 users (nodes) and 435,324 friendships (edges). The network is gathered from Facebook users of 100 colleges and universities (e.g. Caltech, Princeton, Georgetown and UNC Chapel Hill) in September 2005. Each user has the following attributes: ID, a student/faculty status flag, gender, major, second major/minor (if applicable), dormitory(house), year and high school. & \\\\href{https://archive.org/download/oxford-2005-facebook-matrix}{Link}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{\\\\bf ego-Facebook} & Dataset consists of 'circles' ('friends lists') from Facebook with 4039 nodes and 88234 edges. Facebook data was collected from survey participants using a Facebook app. The dataset includes node features (profiles), circles, and ego networks. & \\\\href{https://snap.stanford.edu/data/ego-Facebook.html}{Link}\\\\newline \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf LastFM} & A network gathered from the online music system Last.fm with 1,892 users (nodes) and 12,717 friendships on Last.fm (edges). Each node has 11,946-dimensional attributes, including a list of most listened music artists, and tag assignments.\\n\\t\\t\\t& \\\\href{http://ir.ii.uam.es/hetrec2011/datasets.html}{Link} \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Delicious} & A network of 1,861 nodes, 7,664 edges and 1,350 attributes. This is a publicly available dataset from the HetRec 2011 workshop that has been obtained from the Delicious social bookmarking system. Its users are connected in a social network generated from Delicious mutual fan relations. Each user has bookmarks, tag assignments, that is, [user, tag, bookmark] tuples, and contact relations within the social network. The tag assignments were\\n\\t\\t\\ttransformed to attribute data by taking all tags that a user ever assigned to any bookmark and assigning those to the user. & \\\\href{http://ir.ii.uam.es/hetrec2011/}{Link} \\\\\\\\ \\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Wiki} & A network with nodes as web pages. The link among different nodes is the hyperlink in the web page. 2,405 nodes, 12,761 edges, 4,973 attributes, 17 labels & \\\\href{https://github.com/albertyang33/TADW}{Link} \\\\\\\\ \\n\\t\\t\\\\hline \\n\\t\\t{\\\\bf ego-Twitter} & This dataset consists of 'circles' (or 'lists') from Twitter. Twitter data was crawled from public sources. The dataset includes node features (profiles), circles, and ego networks. Nodes\\t81306,\\n\\t\\tEdges 1768149 & \\\\href{https://snap.stanford.edu/data/ego-Twitter.html}{Link}\\\\newline \\\\\\\\\\n\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{table} \\n\\\\begin{table}\\n\\t\\\\caption{Most popular large size datasets.}\\n\\t\\\\label{tab_dataset3} \\n\\t\\t{\\\\tiny\\n\\\\begin{tabular}{|p{1.2cm}|p{12.9cm}|p{0.8cm}|}\\n\\t\\\\hline \\nDataset & Description & Source \\\\\\\\ \\n\\t\\\\hline \\n{\\\\bf Flickr} & A network with 100,267 nodes, 3,781,947 edges and 16,215 attributes collected from the internal database of the popular Flickr\\nphoto sharing platform. The social network is defined by the contact relation of\\nFlickr. Two vertices are connected with an undirected edge if at least one undirected edge exists between them. Each user has a list of tags associated that he/she used at least five times. Tags are limited to those used by at least 50 users. Users are limited to those having a vocabulary of more than 100 and less than 5,000 tags. & \\\\href{https://snap.stanford.edu/data/web-flickr.html}{Link}\\\\newline \\\\\\\\ \\n\\\\hline \\n{\\\\bf Patents} & A patent citation network with vertices representing patents and edges depicting the citations between. A subgraph containing all the patents from the year 1988 to 1999. Each patent has six attributes, grant year, number of claims, technological category, technological subcategory, assignee type, and main patent class. There are 1,174,908 vertices and 4,967,216 edges in the network. & \\\\href{http://www.nber.org/patents/}{Link 1}\\\\newline \\\\href{https://snap.stanford.edu/data/cit-Patents.html}{Link 2} \\\\\\\\ \\n\\\\hline\\n{\\\\bf ego-G+} & This dataset consists of 'circles' from Google+. Google+ data was collected from users who had manually shared their circles using the 'share circle' feature. The dataset includes node features (profiles), circles, and ego networks. Nodes 107,614, Edges 13,673,453. Each node has four features: job\\ntitle, current place, university, and workplace. A user-pair(edge) is\\ncompared using knowledge graphs based on, Category: Occupations,\\nCategory:Companies by country and industry, Category: Countries,\\nCategory:Universities and colleges by country. & \\\\href{https://snap.stanford.edu/data/ego-Gplus.html}{Link}\\\\newline \\\\\\\\\\n\\\\hline \\n\\\\end{tabular}\\n}\\n\\\\end{table}\",\n \"id\": \"36da5b81-971c-4135-9405-fdc43b1406db\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"c0718d2b-4a41-48ee-939a-4ab8664cabe6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Most used node-attributed network datasets and quality measures for community detection evaluation\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1711\n ],\n \"content\": \"Given a set of detected communities (overlapping or not), one needs to evaluate their quality. There are two possible options for this depending on the network dataset under consideration. If the network dataset has no ground truth, one can use various measures of structural closeness and attribute homogeneity. According to our observations, the most popular quality measures in this case are Modularity and Density for the former and Entropy for the latter. Many others such as Conductance, Permanence, Intra- and Inter-Cluster Densities, etc., are also possible. If there is ground truth, it is traditional to compare the detected communities with the ground truth ones. This can be done, for instance, with the following popular measures: Accuracy, Normalized Mutual Information (denoted below by NMI), Adjusted Rand Index or Rand Index (denoted below by ARI and RI, correspondingly) and $F$-measure. We will discuss both the approaches further in Section~\\\\ref{section9}. \\nDue to space limitations, we refer the reader to the comprehensive survey and to \\\\cite[Sections 2.2 and 4]{Bothorel2015}, where all the above-mentioned quality measures and many others are precisely defined and discussed in detail.\",\n \"id\": \"01bdf590-8667-4b9e-b2aa-932d3480f52b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"c0718d2b-4a41-48ee-939a-4ab8664cabe6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Most used node-attributed network datasets and quality measures for community detection evaluation\"\n ],\n [\n \"subsection\",\n \"Community detection quality measures\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Community detection quality measures\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section6}\\nAs we have already mentioned, these methods fuse structure and attributes before the community detection process so that the data obtained are suitable for classical community detection algorithms (thus one can use their existing software implementations).\\nBefore we proceed, we introduce additional notation applied only to the weight-based and distance-based early fusion methods. The fact is that existing network structure may be saved or modified depending on the heuristics used in a method, therefore we distinguish\\n\\t\\\\begin{itemize}\\n\\t\\t\\\\item {\\\\it fixed topology methods} that use initial network structure without modifying it with respect to attributes,\\n\\t\\t\\\\item {\\\\it non-fixed topology methods} that modify initial network structure with respect to attributes, in particular, add/erase edges and/or vertices.\\n\\t\\\\end{itemize}\\nAs far as we know, there is no study on which approach is preferable. How each one influences community detection results is yet to be established.\\n\\\\begin{figure}[b]\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{weight.pdf}\\n\\t\\\\caption{A typical scheme of a weight-based method (the attributive weights here are the values of normalized matching coefficient for the attribute vectors).}\\n\\t\\\\label{fig:weight}\\n\\\\end{figure}\",\n \"id\": \"1c6b434c-2289-4e6e-9627-ef1149463588\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Early fusion methods\"\n ]\n ],\n \"subsections\": [\n \"4ebad2e3-f246-4134-89c0-5d1b73a7710d\",\n \"0bd4586d-c3e0-4e3a-817b-be22385f142c\",\n \"9080cec6-6c41-47f4-8a10-ec818aed0224\",\n \"bcc6d0ce-88ee-4a08-8574-4e5664967352\",\n \"99f1e588-a145-4e2a-aa14-a5084080874e\"\n ],\n \"title\": \"Early fusion methods\"\n },\n {\n \"cite_extract_rate\": 0.203389830508474,\n \"cites\": [\n 7062,\n 1714,\n 1712,\n 1710,\n 8487,\n 8488,\n 8484\n ],\n \"content\": \"\\\\label{class-ef-weight}\\nThese methods (see Tables \\\\ref{tab_weights_alpha0} and \\\\ref{tab_weights}) convert attributes $(\\\\mathcal{V},\\\\mathcal{A})$ into a weighted {\\\\it attributive} graph and further somehow merge it with structural graph $(\\\\mathcal{V},\\\\mathcal{E})$.\\nThe result is a weighted graph $G_W$ that is a substitution for node-attributed graph $G$, see Figure~\\\\ref{fig:weight}. Edge weights of $G_W$ are usually assigned as follows:\\n\\\\begin{equation}\\n\\\\label{main_weight}\\nW_\\\\alpha(v_i,v_j)=\\\\alpha w_S(v_i,v_j)+(1-\\\\alpha) w_A (v_i,v_j),\\\\qquad \\\\alpha\\\\in[0,1],\\\\qquad v_i,v_j\\\\in \\\\mathcal{V},\\n\\\\end{equation}\\nwhere $w_S$ and $w_A$ are chosen {\\\\it structural} and {\\\\it attributive} similarity functions, respectively. The hyperparameter $\\\\alpha$ controls the balance between structure and attributes. Clearly, if $\\\\alpha=1$ in (\\\\ref{main_weight}), one obtains weights based only on structure; if $\\\\alpha=0$, then they are based only on attributes. As for $w_S$ and $w_A$, $w_S(v_i,v_j)$ usually reflects existing connections in $(\\\\mathcal{V},\\\\mathcal{E})$ (e.g. $w_S(v_i,v_j)=1$, if $e_{ij}\\\\in \\\\mathcal{E}$, and $w_S(v_i,v_j)=0$, otherwise), while $w_A(v_i,v_j)$ may be Cosine Similarity or Matching Coefficient values for vectors $A(v_i)$ and $A(v_j)$.\\nOnce the weighted graph $G_W$ is constructed, one can use classical graph clustering algorithms on it such as Weighted Louvain . Sometimes $G_W$ is instead converted to a certain distance matrix that is further used for detecting communities via distance-based clustering algorithms such as $k$-means or $k$-medoids.\\nIt is worth mentioning that the fixed topology methods in this subclass assume that the weights (\\\\ref{main_weight}) are assigned on the same set of edges $\\\\mathcal{E}$ as in the initial node-attributed graph $G$, while the non-fixed topology ones assign weights on all (or on the most part of) possible edges between nodes in $\\\\mathcal{V}$.\\nNow let us say some words how the hyperparameter $\\\\alpha$ can be chosen. A very popular approach in fixed topology methods is assuming $\\\\alpha=0$ in (\\\\ref{main_weight}), see Table~\\\\ref{tab_weights_alpha0}. This actually means that weights in $G_W$ are based only on attributive similarity. Clearly, this may lead to dominance of attributes in the resulting fusion and disappearance of structural connections between nodes with dissimilar attributes. Varying $\\\\alpha$ over the segment $[0,1]$ in (\\\\ref{main_weight}), used by the methods in Table~\\\\ref{tab_weights}, seems more adequate for controlling the impact of both the components. However, tuning $\\\\alpha$ in this case is usually performed {\\\\it manually}. In fact, we are unaware of any general non-manual approaches for tuning $\\\\alpha$ in (\\\\ref{main_weight}), although the need for such approaches has been repeatedly emphasized .\\nNote that the choice of similarity functions $w_S$ and $w_A$ is usually determined by preferences of the authors of a particular method. The systematic study of how such a choice influences the community detection results is yet to be done.\\n\\\\begin{table}\\n\\t\\\\caption{Weight-based methods with $\\\\alpha=0$ in (\\\\ref{main_weight}).}\\n\\t\\\\label{tab_weights_alpha0} \\n\\t\\\\begin{center}\\n\\t\\t{\\\\tiny\\n\\t\\t\\t\\\\begin{tabular}{|p{1.5cm}|p{4.0cm}|p{1.5cm}|p{1.0cm}|p{1.1cm}|p{1cm}|p{1.3cm}|p{1.5cm}|}\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\tMethod & Community detection method used and its {\\\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Topology & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WNev}~ & {\\\\bf Weighted graph}\\\\newline MinCut \\\\newline MajorClust \\\\newline Spectral & No/No & Small & Accuracy & Fixed & Synthetic & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WSte1}~ & {\\\\bf Weighted graph}\\\\newline Threshold & No/No & Large & Modularity & Fixed & Phone Network & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WSte2}~ & {\\\\bf Similarity matrix (via Weighted graph and random walks)}\\\\newline Hierarchical clustering & No/No & Large & Modularity & Fixed & Phone Network & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WCom1} & {\\\\bf Weighted graph}\\\\newline Weighted Louvain & Yes/No & Small & Accuracy & Fixed & {\\\\bf DBLP*} & {\\\\bf WCom2} \\\\newline {\\\\bf DCom} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WCom2} & {\\\\bf Distance matrix (via weighted graph)}\\\\newline Hierarchical agglomerative clustering & Yes/No & Small & Accuracy & Fixed & {\\\\bf DBLP*} & {\\\\bf WCom1} \\\\newline {\\\\bf DCom} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t{\\\\bf AA-Cluster} & {\\\\bf Node embeddings (via weighted graph)}\\\\newline $k$-medoids & Yes/No & Small\\\\newline Medium\\\\newline Large & Density\\\\newline Entropy & Fixed & {\\\\bf Political Blogs}\\\\newline {\\\\bf DBLP*}\\\\newline {\\\\bf Patents*} \\\\newline Synthetic & {\\\\bf SA-Cluster}~\\\\newline {\\\\bf BAGC} \\\\newline {\\\\bf CPIP} \\n\\t\\t\\t\\t\\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf PWMA-MILP} & {\\\\bf Weighted graph}\\\\newline Linear programming MILP~ & No/No & Small & RI \\\\newline NMI & Fixed & {\\\\bf WebKB} & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf KDComm} & {\\\\bf Weighted graph} \\\\newline Iterative Weighted Louvain & No/No & Small\\\\newline Medium \\\\newline Large & $F$-measure\\\\newline Jaccard measure \\\\newline Rank Entropy measure & Fixed & \\t{\\\\bf ego-G+} \\\\newline {\\\\bf Twitter*} \\\\newline {\\\\bf DBLP*} \\\\newline Reddit \\\\href{https://archive.org/details/2015_reddit_comments_corpus}{link} & {\\\\bf CPIP} \\\\newline {\\\\bf JCDC} \\\\newline {\\\\bf UNCut} \\\\newline {\\\\bf SI} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\\\end{tabular}\\n\\t\\t}\\n\\t\\\\end{center}\\n\\\\end{table}\\n\\\\begin{table}\\n\\t\\\\caption{Weight-based methods with $\\\\alpha\\\\in[0,1]$ in (\\\\ref{main_weight}).}\\n\\t\\\\label{tab_weights} \\n\\t\\\\begin{center}\\n\\t\\t{\\\\tiny\\n\\t\\t\\t\\\\begin{tabular}{|p{1.25cm}|p{1.5cm}|p{2.0cm}|p{1.5cm}|p{0.8cm}|p{1.2cm}|p{1.2cm}|p{1.7cm}|p{1.7cm}|}\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\tMethod & $\\\\alpha$ in (\\\\ref{main_weight}) & Community detection method used and its {\\\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Topology & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WWan} & $[0,1]$ in theory\\\\newline $\\\\tfrac{1}{2}$ in experiments & {\\\\bf Edge similarity matrix (via weighted graph)}\\\\newline EdgeCluster \\\\newline($k$-means variant) & Yes & Small & NMI \\\\newline Micro-F1 \\\\newline Macro-F1 & Non-fixed: removing edges & Synthetic \\\\newline BlogCatalog\\\\newline {\\\\bf Delicious$^*$} & Non-overlapping co-clustering \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf SAC2} & $[0,1]$ & {\\\\bf $k$NN (unweighted) graph (via weighted graph)}\\\\newline (Unweighted) Louvain & No/ No & Small\\\\newline Medium & Density\\\\newline Entropy & Non-fixed: removing edges & {\\\\bf Political Blogs} \\\\newline {\\\\bf Facebook100} \\\\newline {\\\\bf DBLP10K} & {\\\\bf SAC1} \\\\newline {\\\\bf WSte2} \\\\newline Fast greedy for weighted graph \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WCru}~\\n\\t\\t\\t\\t& $[0,1]$ in theory \\\\newline Not specified in experiments & {\\\\bf Weighted graph}\\\\newline Weighted Louvain~ & No & Medium & Modularity\\\\newline Intracluster distance & Fixed & {\\\\bf Twitter$^*$} & ---\\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf CODICIL} & $[0,1]$ in theory\\\\newline\\n\\t\\t\\t\\t$1/2$ in some experiments & {\\\\bf Weighted graph}\\\\newline Metis \\\\newline Markov Clustering~ & No & Small\\\\newline Medium\\\\newline Large & $F$-measure & Non-fixed: adding and removing edges & {\\\\bf CiteSeer*}\\\\newline {\\\\bf Flickr*}\\\\newline Wikipedia* & {\\\\bf Inc-Cluster} \\\\newline {\\\\bf PCL-DC} \\\\newline Link-PLSA-LDA \\n\\t\\t\\t\\t\\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf WMen} & Not specified & {\\\\bf Weighted graph/Distance matrix for the weighted graph}\\\\newline SLPA \\\\newline Weighted Louvain \\\\newline K-medoids & Yes-No/\\\\newline Yes-No & Small\\\\newline Medium & NMI \\\\newline $F$-measure \\\\newline Accuracy & Fixed & {\\\\bf Lazega} \\\\newline {\\\\bf Research} \\\\newline {\\\\bf Consult} \\\\newline\\n\\t\\t\\t\\tLFR benchmark & {\\\\bf CODICIL}~\\\\newline {\\\\bf SA-Cluster}~ \\\\\\\\ \\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf PLCA-MILP} & $[0,1]$ & {\\\\bf Weighted graph}\\\\newline Linear programming MILP~ & No/No & Small & RI \\\\newline NMI & Non-fixed: adding and removing edges & {\\\\bf WebKB} & {\\\\bf SCD} \\\\newline {\\\\bf ASCD} \\\\newline {\\\\bf SCI} \\\\newline PCL-DC \\\\newline Block-LDA \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t{\\\\bf kNN-enhance} & May be thought as $\\\\alpha=1/2$, $k$NN by attributes & {\\\\bf Distance matrix (of an edge-augmented graph)} \\\\newline $k$NN\\\\newline $k$-means & No/No & Medium & Accuracy \\\\newline NMI \\\\newline F-Measure \\\\newline Modularity\\\\newline Entropy & Non-fixed: adding edges & {\\\\bf Cora} \\\\newline {\\\\bf Citeseer}\\\\newline {\\\\bf Sinanet}\\\\newline {\\\\bf PubMed Diabetes}\\\\newline {\\\\bf DBLP*}& {\\\\bf PCL-DC} \\\\newline PPL-DC \\\\newline {\\\\bf PPSB-DC} \\\\newline {\\\\bf CESNA} \\\\newline {\\\\bf cohsMix} \\\\newline {\\\\bf BAGC} \\\\newline {\\\\bf GBAGC} ) \\\\newline {\\\\bf SA-Custer} \\\\newline {\\\\bf Inc-Cluster} \\\\newline {\\\\bf CODICIL} \\\\newline GLFM \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf IGC-CSM} (\\\\href{https://github.com/WNawaz/CSM}{source}) & $[0,1]$ in theory\\\\newline $1/2$ in comparison experiments & {\\\\bf Distance matrix for the weighted graph} \\\\newline $k$-Medoids & Yes/ No & Medium & Density\\\\newline Entropy & Fixed & {\\\\bf Political Blogs}\\\\newline {\\\\bf DBLP10K} & {\\\\bf SA-Cluster} \\\\newline {\\\\bf SA-Cluster-Opt} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf AGPFC} & $[0,1]$ in theory, manually tuned in experiments & Fuzzy equivalent matrix \\\\newline $\\\\lambda$-cut set method & No/Yes & Small\\\\newline Medium & Density\\\\newline Entropy & Fixed & {\\\\bf Political Blogs}\\\\newline {\\\\bf CiteSeer} \\\\newline {\\\\bf Cora} \\\\newline {\\\\bf WebKB} & {\\\\bf SA-Cluster} \\\\newline {\\\\bf BAGC} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf NMLPA} & $1/2$ & {\\\\bf Weighted graph}\\\\newline A multi-label propagation algorithm & Yes/ Yes & Medium & $F1$-score \\\\newline Jaccard Similarity & Fixed & {\\\\bf ego-Facebook}\\\\newline {\\\\bf Flickr*} \\\\newline {\\\\bf ego-Twitter} & {\\\\bf CESNA} \\\\newline {\\\\bf SCI} \\\\newline {\\\\bf CDE} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\\\end{tabular}\\n\\t\\t}\\n\\t\\\\end{center}\\n\\\\end{table}\\n\\\\begin{remark} The weight-based strategy has been applied to more general networks than considered in the present survey. For example, and use a similar scheme to detect communities in {\\\\it multi-layer} networks. Another example is\\nSANS that works with {\\\\it directed} node-attributed graphs. \\nEdge weighting similar to (\\\\ref{main_weight}) is also applied in FocusCO\\t. Although it is not a purely unsupervised community detection method (it requires user's preferences on some of the attributes), it can simultaneously extract local clusters and detect outliers in a node-attributed social network.\\n\\\\end{remark}\\n\\\\begin{remark}\\nLet us mention that there exist community detection methods similar ideologically (attributes $\\\\to$ edge weights $\\\\to$ weighted graph node embeddings\\\\footnote{Low-dimensional continuous vector representations of graph nodes. See also Subsection~\\\\ref{class-ef-embedding}.} $\\\\to$ $k$-means) but preceding to recently proposed {\\\\bf AA-Cluster} . Namely, GraphEncoder~ and GraRep~ also first convert a node-attributed graph into a weighted one according to (\\\\ref{main_weight}) with $\\\\alpha=0$ and then find corresponding node embeddings. The embeddings are further fed to $k$-means algorithm to detect communities. However, in opposite to , and mostly focus on node embedding techniques for a weighted graph than on the community detection task. \\n\\\\end{remark}\",\n \"id\": \"4ebad2e3-f246-4134-89c0-5d1b73a7710d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 59,\n \"parent_id\": \"1c6b434c-2289-4e6e-9627-ef1149463588\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Early fusion methods\"\n ],\n [\n \"subsection\",\n \"Weight-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Weight-based methods\"\n },\n {\n \"cite_extract_rate\": 0.1,\n \"cites\": [\n 1717,\n 1714\n ],\n \"content\": \"\\\\label{class-ef-distance}\\n\\\\begin{table}[b]\\n\\\\caption {Distance-based methods.}\\n\\\\label{tab:distance} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.0cm}|p{1.3cm}|p{1.8cm}|p{1.5cm}|p{1.2cm}|p{1.5cm}|p{1.1cm}|p{1.5cm}|p{1.8cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & $\\\\alpha$ in (\\\\ref{distance_main}) & Community detection method used and its {\\\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Topology & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf DCom} & $[0,1]$ & {\\\\bf Distance matrix}\\\\newline Hierarchical agglomerative clustering & Yes/No & Small & Accuracy & Non-fixed: added edges & {\\\\bf DBLP*} & {\\\\bf WCom1} \\\\newline {\\\\bf WCom2} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{\\\\bf DVil} & $[0,1]$ & {\\\\bf Distance matrix} \\\\newline Self-organizing maps & No/No & Small\\\\newline Medium & NMI & Non-fixed: added edges & Synthetic\\\\newline \\\\href{http://graphcomp.univ-tlse2.fr/}{Medieval Notarial Deeds} & --- \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf SToC} & Maybe thought depending on the values of $d_S$ and $d_A$ & {\\\\bf Distance matrix}\\\\newline $\\\\tau$-close clustering & No/No & Medium\\\\newline Large & Modularity\\\\newline Within-Cluster Sum of Squares & Non-fixed: added edges & {\\\\bf DBLP10K}\\\\newline DIRECTORS*\\\\newline DIRECTORS-gcc* & {\\\\bf Inc-Cluster} \\\\newline {\\\\bf GBAGC} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf @NetGA} & $\\\\alpha\\\\in[0,1]$ in general\\\\newline $\\\\alpha=1/2$ in experiments & {\\\\bf Distance matrix}\\\\newline Genetic algorithm & No/No & Medium & NMI & Non-fixed: added edges & Synthetic & {\\\\bf SA-Cluster} \\\\newline CSPA \\\\newline {\\\\bf Selection} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf ANCA} & Maybe thought as $\\\\alpha=1/2$ & {\\\\bf Distance matrix}\\\\newline $k$-means & Yes/No & Medium & Adjusted Rand Index\\\\newline NMI \\\\newline Density \\\\newline Modularity \\\\newline Conductance \\\\newline Entropy & Fixed & Synthetic \\\\newline {\\\\bf DBLP10K} \\\\newline \\n\\t\\t\\tEnron email corpus \\\\newline & {\\\\bf SA-Cluster} \\\\newline {\\\\bf SAC1-SAC2} \\\\newline {\\\\bf IGC-CSM} \\\\newline {\\\\bf WSte1} \\\\newline {\\\\bf ILouvain} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\\nMethods discussed in the previous subsection aim at representing structure and attributes in a unified graph form suitable for further graph clustering. In opposite, distance-based methods intentionally abandon graph representation in favor of the representation via a distance matrix that contains information about structure and attributes. The distance matrix is usually obtained by fusing the components by a structure- and attributes-aware distance function, see Figure~\\\\ref{fig:distance}. The matrix can be further fed to distance-based clustering algorithms such as $k$-means and $k$-medoids. Note that the resulting clusters may contain disconnected portions of initial graph $G$ as the graph structure is removed at the fusion step \\\\cite[Section 3.3]{Akbas2017}.\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{distance.pdf}\\n\\t\\\\caption{A typical scheme of a distance-based method.}\\n\\t\\\\label{fig:distance}\\n\\\\end{figure}\\nThe distance function fusing structure and attributes is often of the form\\n\\\\begin{equation}\\n\\\\label{distance_main}\\nD_{\\\\alpha}(v_i,v_j)=\\\\alpha d_S(v_i,v_j)+(1-\\\\alpha)d_A(v_i,v_j),\\\\qquad \\\\alpha\\\\in[0,1],\\\\qquad v_i,v_j\\\\in \\\\mathcal{V},\\n\\\\end{equation}\\nwhere $d_S$ and $d_A$ are {\\\\it structural} and {\\\\it attributive} distance functions, correspondingly. As in the case of (\\\\ref{main_weight}), $\\\\alpha$ in (\\\\ref{distance_main}) controls the balance between the components; how to choose it ``properly'' seems to be an open problem, too. As for the distance functions, it is common to define $d_S(v_i,v_j)$ as the shortest path length between $v_i$ and $v_j$. Possible options for $d_S(v_i,v_j)$ are the Jaccard or Minkowski distances between vectors $A(v_i)$ and $A(v_j)$.\\nShort descriptions for known distance-based methods are given in Table~\\\\ref{tab:distance}. Note that {\\\\bf ANCA} and {\\\\bf SToC} employ a bit different fusion than in (\\\\ref{distance_main}) but nevertheless still deal with certain structure- and attribute-aware distances.\\n\\\\begin{remark}\\nThere exist distance-based methods for {\\\\it multi-layer} networks. For example, CLAMP is an method for clustering networks with heterogeneous attributes and multiple types of edges that uses a distance function similar to (\\\\ref{distance_main}), in a sense.\\n\\\\end{remark}\\n\\\\begin{table}\\n\\\\caption {Node-augmented graph distance-based methods.} \\\\label{tab:sa} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.8cm}|p{1.7cm}|p{2.8cm}|p{1.2cm}|p{1cm}|p{1.0cm}|p{1.4cm}|p{2.2cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & Graph augmentation & Community detection method used and its {\\\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf SA-Cluster} \\\\newline {\\\\bf Inc-Cluster} \\\\newline {\\\\bf SA-Cluster-Opt} & New nodes and edges & {\\\\bf Distance matrix (via neighbourhood random walks) }\\\\newline Modified $k$-medoids & Yes/No & Small\\\\newline Medium & Density\\\\newline Entropy & {\\\\bf Political Blogs}\\\\newline {\\\\bf DBLP10K}\\\\newline {\\\\bf DBLP84K} & W-Cluster (based on (\\\\ref{distance_main}))\\\\newline {\\\\bf SA-Cluster} \\\\newline {\\\\bf Inc-Cluster} \\\\newline {\\\\bf SA-Cluster-Opt} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf SCMAG} & New nodes and edges & {\\\\bf Distance matrix (via neighbourhood random walks) }\\\\newline Subspace clustering algorithm based on ENCLUS & No/Yes & Medium & Density\\\\newline Entropy & IMDB\\\\newline Arnetminer bibliography$^*$ & {\\\\bf SA-Custer} \\\\newline GAMer \\\\\\\\\\t\\t\\t\\n\\t\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\",\n \"id\": \"0bd4586d-c3e0-4e3a-817b-be22385f142c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"1c6b434c-2289-4e6e-9627-ef1149463588\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Early fusion methods\"\n ],\n [\n \"subsection\",\n \"Distance-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Distance-based methods\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{class-ef-node-augmented-graph}\\n\\\\begin{figure}[b]\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{auggraph.pdf}\\n\\t\\\\caption{A typical scheme of a node-augmented graph-based method.}\\n\\t\\\\label{fig:node-augmented}\\n\\\\end{figure}\\nMethods from this subsection (see Table~\\\\ref{tab:sa}) transform initial node-attributed graph $G$ into another node-augmented graph $\\\\tilde{G}$, with nodes from $\\\\mathcal{V}$ and new {\\\\it attributive} nodes representing attributes, see Figure~\\\\ref{fig:node-augmented}.\\nTo be more precise, suppose that $dom(a_k)=\\\\{s_l\\\\}_{l=1}^{L_k}$, i.e. the domain of $k$th element (attribute) of attribute vector $A$ contains $L_k$ possible values. Then one should create $L_k$ new attribute nodes $\\\\tilde{v}_{k,l}$ corresponding to $l$th value of $k$th attribute. Such a procedure is performed for $k=1,\\\\ldots,d$, where $d=\\\\dim A$. The set $\\\\mathcal{V}_A:=\\\\{\\\\tilde{v}_{k,l}\\\\}$, where $k=1,\\\\ldots,d$ and $l=1,\\\\ldots,L_k$, is then the set of attributive nodes. An edge between structural node $v_i\\\\in \\\\mathcal{V}$ and attributive node $\\\\tilde{v}_{k,l}$ in $\\\\tilde{G}$ exists if $a_k(v_i)=s_l$. Community detection is further performed in $\\\\tilde{G}$. The methods in Table~\\\\ref{tab:sa} propose to apply random walks to obtain a certain distance matrix for $\\\\tilde{G}$ and further use it in a distance-based clustering algorithm.\\nNote that the above-mentioned augmentation is not applicable to continuous attributes. What is more, $\\\\tilde{G}$ contains much more nodes end edges than $G$ (especially if $d$ and $L_k$, $k=1,\\\\ldots,d$, are large) and this makes the methods from this subclass rather computationally expensive.\",\n \"id\": \"9080cec6-6c41-47f4-8a10-ec818aed0224\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"1c6b434c-2289-4e6e-9627-ef1149463588\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Early fusion methods\"\n ],\n [\n \"subsection\",\n \"Node-augmented graph-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Node-augmented graph-based methods\"\n },\n {\n \"cite_extract_rate\": 0.35714285714285704,\n \"cites\": [\n 1712,\n 242,\n 217,\n 8489,\n 215,\n 282,\n 1719,\n 1718,\n 229,\n 1010\n ],\n \"content\": \"\\\\label{class-ef-embedding}\\n\\\\begin{figure}\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{emb.pdf}\\n\\t\\\\caption{A possible scheme of an embedding-based method (the attribute vectors are here concatenated with the node embeddings and then fed to $k$-means).}\\n\\t\\\\label{fig:embeddings}\\n\\\\end{figure}\\nAs is well-known, a graph as a traditional representation of a network brings several difficulties to network analysis. As mentioned in , graph algorithms suffer from high computational complexity, low parallelisability and inapplicability of machine learning methods. Novel embedding techniques aim at tackling this by learning {\\\\it node embeddings}, i.e. low-dimensional continuous vector representations for network nodes so that main network information is efficiently encoded . Roughly speaking, node embedding techniques allow to convert a graph with $n$ nodes into a set of $n$ vectors.\\nIn the context of node-attributed social networks, the objective of node embedding techniques is efficient encoding both structure and attributes . We are not going to provide general details on the techniques as this has been already done in the surveys . Let us just mention that having node embeddings (i.e. vectors) at hand allows one to use classical distance-based clustering algorithms such as $k$-means to detect communities, see Figure~\\\\ref{fig:embeddings}.\\nIt turns out that there exists a rich bibliography on node embedding techniques for attributes networks of different type . However, not all of them have been applied to the community detection task (the classification task is typically considered). Taking this into account, we confine ourselves in this survey only to the embedding-based (early fusion) methods that have been used for community detection in node-attributed social or document networks and compared with other community detection methods. The results are presented in Table~\\\\ref{tab:emb}.\\n\\\\begin{table}[b]\\n\\\\caption {Embedding-based methods.}\\n\\\\label{tab:emb} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1cm}|p{1.5cm}|p{2.5cm}|p{1.4cm}|p{1cm}|p{1.9cm}|p{1.5cm}|p{2.3cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & Embedding technique & Community detection method used and its {\\\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf PLANE} & A generative model and EM~ & {\\\\bf Node embeddings}\\\\newline $k$-means & Yes/No & Small\\\\newline Medium & Accuracy & {\\\\bf Cora*} & Relational Topic Model~+Topic Distributions Embedding~\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf DANE} & Autoencoder & {\\\\bf Node Embeddings}\\\\newline $k$-means & Yes/No & Medium & Accuracy & {\\\\bf Cora}\\\\newline {\\\\bf Citeseer}\\\\newline {\\\\bf PubMed Diabetes}\\\\newline {\\\\bf Wiki} & Embeddings obtained via TADW \\\\newline\\n\\t\\t\\tLANE \\\\newline GAE \\\\newline VGAE \\\\newline GraphSAGE \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf CDE} & Structure embedding matrix & {\\\\bf Structure embedding matrix and attribute matrix}\\\\newline Non-negative matrix factorization & Yes/(Yes/No) & Small\\\\newline Medium & Accuracy \\\\newline NMI \\\\newline Jaccard similarity \\\\newline F1-score & {\\\\bf Cora}\\\\newline {\\\\bf Citeseer}\\\\newline {\\\\bf WebKB}\\\\newline {\\\\bf Flickr*} \\\\newline Philosophers \\\\newline {\\\\bf ego-Facebook} & {\\\\bf PCL-DC} \\\\newline {\\\\bf Circles} \\\\newline {\\\\bf CESNA} \\\\newline \\\\newline {\\\\bf SCI} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf MGAE} & Autoencoder & {\\\\bf Node embeddings} \\\\newline Spectral clustering & Yes/No & Medium & Accuracy\\\\newline NMI \\\\newline $F$-score \\\\newline Precision \\\\newline Recall \\\\newline Average Entropy \\\\newline Adjusted Rand Index & {\\\\bf Cora}\\\\newline {\\\\bf CiteSeer} \\\\newline {\\\\bf Wiki} & {\\\\bf Circles} \\\\newline RTM \\\\newline RMSC \\\\newline Embeddings obtained via TADW \\\\newline VGAE \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\\n\\\\begin{remark}\\nVarious embedding techniques applicable for community detection in {\\\\it multi-layer} networks are considered e.g. in .\\n\\\\end{remark}\\n\\\\begin{table}[b]\\n\\\\caption{Pattern mining-based (early fusion) methods.}\\n \\\\label{tab:pattern} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\\\begin{tabular}{|p{1.1cm}|p{0.9cm}|p{2.9cm}|p{2.5cm}|p{2cm}|p{1.5cm}|p{1.2cm}|p{1cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & Pattern used & Community detection method used and its {\\\\bf input} & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf AHMotif} & Motif & {\\\\bf Structure- and attributes-aware adjacency matrix} \\\\newline Permanence \\\\newline Affinity Propagation & Yes/No & Medium & NMI \\\\newline Accuracy & {\\\\bf Cora}\\\\newline {\\\\bf WebKB} & ---\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\",\n \"id\": \"bcc6d0ce-88ee-4a08-8574-4e5664967352\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 28,\n \"parent_id\": \"1c6b434c-2289-4e6e-9627-ef1149463588\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Early fusion methods\"\n ],\n [\n \"subsection\",\n \"Embedding-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Embedding-based methods\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{class-ef-patterns}\\nRecall that a motif is a pattern of the interconnection occurring in real-world networks at numbers that are significantly higher than those in random\\nnetworks . Motifs are considered as building blocks for complex networks . We found just one community detection method for node-attributed social networks using such patterns, namely, {\\\\bf AHMotif} , see Table~\\\\ref{tab:pattern}. This method equips structural motifs identified in the network with the so-called homogeneity value based on node attributes involved in the motif. This information is stored in a special adjacency matrix that can be an input to classical community detection algorithms.\",\n \"id\": \"99f1e588-a145-4e2a-aa14-a5084080874e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"1c6b434c-2289-4e6e-9627-ef1149463588\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Early fusion methods\"\n ],\n [\n \"subsection\",\n \"Pattern mining-based (early fusion) methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pattern mining-based (early fusion) methods\"\n },\n {\n \"cite_extract_rate\": 0.228571428571428,\n \"cites\": [\n 8483,\n 1714,\n 8487,\n 8488,\n 8482,\n 1720\n ],\n \"content\": \"\\\\label{section7}\\nRecall that simultaneous fusion methods fuse structure and attributes in a joint process with community detection. For this reason, these methods often require special software implementation, in contrast to early and late fusion methods that partially allow one to use existing implementations of classical community detection algorithms.\\n\\\\begin{table}\\n\\t\\\\caption {Methods modifying objective functions of classical clustering algorithms}\\n\\t\\\\label{tab_Louvain_NCut} \\n\\t\\\\begin{center}\\n\\t\\t{\\\\tiny\\n\\t\\t\\t\\\\begin{tabular}{|p{1.2cm}|p{2.8cm}|p{1.8cm}|p{1.5cm}|p{1.6cm}|p{2.0cm}|p{2.4cm}|}\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\tMethod & Modified algorithm & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf OCru} & Louvain \\\\newline Added attribute Entropy minimization & No/No & Medium & Modularity\\\\newline Entropy & {\\\\bf Facebook100} & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf SAC1} & Louvain \\\\newline Added attribute similarity maximization & No/ No & Small\\\\newline Medium & Density\\\\newline Entropy & {\\\\bf Political Blogs}\\\\newline {\\\\bf Facebook100} \\\\newline {\\\\bf DBLP10K} & {\\\\bf SAC2} \\\\newline {\\\\bf WSte2} \\\\newline Fast greedy for weighted graph \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf ILouvain} (\\\\href{http://bit.ly/ILouvain}{source}) & Louvain \\\\newline Added maximization of attribute-aware Inertia & No/ No & Small\\\\newline Medium & NMI \\\\newline Accuracy & DBLP+Microsoft Academic Search$^*$ \\\\newline Synthetic & ToTeM \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf LAA/LOA} & Louvain \\\\newline Modularity gain depends on attributes & No/No & Small & Density\\\\newline Modularity & \\\\href{https://sites.google.com/site/ucinetsoftware/datasets/covert-networks/londongang}{London gang} \\\\newline\\n\\t\\t\\t\\t\\\\href{https://sites.google.com/site/ucinetsoftware/datasets/covert-networks/italiangangs}{Italy gang} \\\\newline\\n\\t\\t\\t\\t\\\\href{http://networkrepository.com/polbooks.php}{Polbooks}\\n\\t\\t\\t\\t\\\\newline\\n\\t\\t\\t\\t\\\\href{http://networkrepository.com/adjnoun.php}{Adjnoun} \\n\\t\\t\\t\\t\\\\newline\\n\\t\\t\\t\\tFootball & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf MAM} (\\\\href{http://ipd.kit.edu/~muellere/mam/}{source}) & Louvain-type algorithm with attribute-aware Modularity+Outlier detection & No/No & Small\\\\newline Medium \\\\newline Large & F1-score \\\\newline Attribute-aware Modularity & Synthetic \\\\newline Disney \\\\newline DFB \\\\newline ARXIV \\\\newline\\n\\t\\t\\t\\tIMDB \\\\newline \\n\\t\\t\\t\\t{\\\\bf DBLP*} \\\\newline\\n\\t\\t\\t\\t{\\\\bf Patents*} \\\\newline\\n\\t\\t\\t\\tAmazon & CODA \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf UNCut} & Normalized Cut\\\\newline Added attributes-aware Unimodality Compactness & Yes/No & Small\\\\newline Medium & NMI \\\\newline ARI \\\\newline & Disney \\\\newline\\n\\t\\t\\t\\tDFB \\\\newline\\n\\t\\t\\t\\tARXIV \\\\newline\\n\\t\\t\\t\\t{\\\\bf Political Blogs} \\\\newline\\n\\t\\t\\t\\t4area \\\\newline\\n\\t\\t\\t\\t{\\\\bf Patents} & {\\\\bf SA-cluster} \\\\newline SSCG \\\\newline {\\\\bf NNM} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf NetScan} & An approximation algorithm for the connected $k$-Center optimization problem (structure and attributes involved) & Yes/Yes & Small\\\\newline Medium & Accuracy & Professors* \\\\newline Synthetic \\\\newline {\\\\bf DBLP*}\\\\newline BioGRID+Spellman\\n\\t\\t\\t\\t& --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf JointClust} & An approximation algorithm for the Connected X Clusters problem (structure and attributes involved) & No/No & Medium & Accuracy & {\\\\bf DBLP*} \\\\newline {\\\\bf CiteSeer*} \\\\newline Corel stock photo collection & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf SS-Cluster} & $k$-Medoid with structure- and attributes-aware objective functions & Yes/No & Medium & Density \\\\newline Entropy & {\\\\bf Political Blogs}\\\\newline {\\\\bf DBLP10K} & {\\\\bf SA-cluster} \\\\newline W-cluster \\\\newline $k$SNAP \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf Adapt-SA} & Weighted $k$-means for $d$-dimensional representations of structure and attributes & Yes/No & Medium & Accuracy \\\\newline \\n\\t\\t\\t\\tNMI \\\\newline F-measure \\\\newline Modularity\\\\newline\\n\\t\\t\\t\\tEntropy & Synthetic \\\\newline {\\\\bf WebKB}\\\\newline {\\\\bf Cora}\\\\newline {\\\\bf Political Blogs} \\\\newline {\\\\bf CiteSeer} \\\\newline {\\\\bf DBLP10K} & {\\\\bf CODICIL} \\\\newline {\\\\bf SA-Cluster} \\\\newline {\\\\bf Inc-Cluster} \\\\newline {\\\\bf PPSB-DC} \\\\newline {\\\\bf PCL-DC} \\\\newline {\\\\bf BAGC} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf kNAS} & $kNN$ with added Semantic Similarity Score & Yes/Yes & Medium & Density\\\\newline Tanimoto Coefficient & {\\\\bf DBLP*} \\\\newline {\\\\bf Facebook*} \\\\newline {\\\\bf Twitter*} & {\\\\bf SA-Cluster-Opt} \\\\newline {\\\\bf CODICIL} \\\\newline NISE \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\\\end{tabular}\\n\\t\\t}\\n\\t\\\\end{center}\\n\\\\end{table}\",\n \"id\": \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"level\": \"section\",\n \"origin_cites_number\": 35,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Simultaneous fusion methods\"\n ]\n ],\n \"subsections\": [\n \"1af2c98c-c4a1-4d9f-917a-330efea7a2a3\",\n \"7b216870-6a3a-40ca-874f-da14c2289f12\",\n \"c986582c-70df-43b2-b7f2-9e2e83ab57f5\",\n \"f8f23311-f720-42e4-94c4-ed1c64c919d4\",\n \"45b099b6-e934-4199-b2c6-b9c78b20f692\",\n \"6a22c128-a708-4842-b8d3-074bec4c46a4\"\n ],\n \"title\": \"Simultaneous fusion methods\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 8487\n ],\n \"content\": \"\\\\label{class-sf-modifying}\\nTable~\\\\ref{tab_Louvain_NCut} contains\\\\footnote{The authors of {\\\\bf ILouvain} claim that they compare their method with ToTeM , ``another community detection method designed for attributed graphs''. However, it seems that there is an inaccuracy with it as we could not find in any method called ToTeM.} short descriptions of simultaneous fusion methods that modify objective functions of well-known clustering algorithms such as Louvain, Normalized Cut, $k$-means, $k$-medoids and $kNN$. Their main idea is to adapt a classical method (that works, for example, only with network structure originally) for using both structure and attributes in the optimization process. For example, if one wants to modify Louvain whose original objective function is structure-aware Modularity, one can include an attributes-aware objective function, say, Entropy in the optimization process. Then Modularity is maximized and Entropy is minimized simultaneously in an iterative process similar to that of Louvain .\",\n \"id\": \"1af2c98c-c4a1-4d9f-917a-330efea7a2a3\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Simultaneous fusion methods\"\n ],\n [\n \"subsection\",\n \"Methods modifying objective functions of classical clustering algorithms\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Methods modifying objective functions of classical clustering algorithms\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 7062,\n 1714,\n 1712\n ],\n \"content\": \"\\\\label{class-sf-mataheuristic}\\nMethods in this subclass are rather similar ideologically to those in Subsection~\\\\ref{class-sf-modifying}. However, instead of modifying objective functions and iterative processes of well-known community detection algorithms, they directly apply metaheuristic algorithms (in particular, evolutionary algorithms and tabu search) to find a node-attributed network partition that provides optimal values for certain measures of structural closeness and attribute homogeneity. More precisely, they use metaheuristics for optimization of a combination of structure- and attributes-aware objective functions, for example, Modularity and Attributes Similarity. Short descriptions of the methods from this subclass are given in Table~\\\\ref{tab_heur}. \\n\\\\begin{table}\\n\\t\\\\caption {Metaheuristic-based methods.}\\n\\t\\\\label{tab_heur} \\n\\t\\\\begin{center}\\n\\t\\t{\\\\tiny\\n\\t\\t\\t\\\\begin{tabular}{|p{1.2cm}|p{3.7cm}|p{1.5cm}|p{1.0cm}|p{1.4cm}|p{2.5cm}|p{2.2cm}|}\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\tMethod & Metaheuristic optimization algorithm & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf MOEA-SA} & Multiobjective evolutionary algorithm (Modularity and Attribute Similarity are maximized) & No/No & Small \\\\newline Medium & Density\\\\newline Entropy & {\\\\bf Political Books}\\\\newline {\\\\bf Political Blogs} \\\\newline {\\\\bf Facebook100}\\\\newline {\\\\bf ego-Facebook} & {\\\\bf SAC1-SAC2} \\\\newline {\\\\bf SA-Cluster} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf MOGA-@Net} & Multiobjective genetic algorithm (optimizing Modularity, Community score, Conductance, attribute similarity) & No/No & Small\\\\newline Medium & NMI \\\\newline Cumulative NMI \\\\newline Density \\\\newline Entropy & Synthetic \\\\newline {\\\\bf Cora}\\\\newline {\\\\bf Citeseer} \\\\newline {\\\\bf Political books} \\\\newline {\\\\bf Political Blogs} \\\\newline {\\\\bf ego-Facebook} & {\\\\bf SA-cluster} \\\\newline {\\\\bf BAGC} \\\\newline {\\\\bf OCru} \\\\newline {\\\\bf Selection} \\\\newline HGPA-CSPA \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t{\\\\bf JCDC} & Tabu search and gradient ascent for a certain structure- and attributes-aware objective function & Yes/No & Small\\\\newline Medium & NMI & Synthetic \\\\newline World trade \\\\newline {\\\\bf Lazega} & CASC \\\\newline {\\\\bf CESNA} \\\\newline {\\\\bf BAGS} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\\\end{tabular}\\n\\t\\t}\\n\\t\\\\end{center}\\n\\\\end{table}\",\n \"id\": \"7b216870-6a3a-40ca-874f-da14c2289f12\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Simultaneous fusion methods\"\n ],\n [\n \"subsection\",\n \"Metaheuristic-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Metaheuristic-based methods\"\n },\n {\n \"cite_extract_rate\": 0.1,\n \"cites\": [\n 1712,\n 1721,\n 8489\n ],\n \"content\": \"\\\\label{class-sf-nnmf}\\nNon-negative matrix factorization (NNMF) is a matrix technique that consists in approximating a non-negative matrix with high rank by a product of non-negative matrices with lower ranks so that the approximation error by means of the Frobenius norm\\\\footnote{A matrix norm defined as the square root of the sum of the absolute squares of matrix elements.} $F$ is minimal. As is well known, NNMF is able to find clusters in the input data . \\nTo be applied to community detection in node-attributed social networks, NNMF requires a proper adaptation to fuse both structure and attributes. Different versions of such an adaptation have been proposed, see Table~\\\\ref{tab:NMF}. \\nTo be more formal in describing the corresponding NNMF-based methods, let us introduce additional notation. Let ${\\\\bf S}_{n\\\\times n}$ denote the adjacency matrix for the network structure (as before, $n$ is the number of nodes), ${\\\\bf A}_{n\\\\times d}$ the node attribute matrix for the network attributes ($d$ is the dimension of attribute vector $A$), $N$ the number of required clusters (it is an input in NNMF-based methods), ${\\\\bf U}_{n\\\\times N}$ the cluster membership matrix whose elements indicate the association of nodes with communities, and finally ${\\\\bf V}_{d\\\\times N}$ denotes the cluster membership matrix whose elements indicate the association of the attributes with the communities. In these terms, the aim of NNMF-based methods is to use known matrices $\\\\bf S$, $\\\\bf A$ and the number of clusters $N$ in order to determine the unknown matrices $\\\\bf U$ and $\\\\bf V$ in an iterative optimization procedure. For example, {\\\\bf SCI} models structural closeness as $\\\\min_{{\\\\bf U}\\\\ge 0}\\\\|{\\\\bf S}-{\\\\bf U}{\\\\bf U}^T\\\\|^2_F$ and attribute homogeneity as \\n$\\\\min_{{\\\\bf V}\\\\ge 0}\\\\|{\\\\bf U}-{\\\\bf A}{\\\\bf V}\\\\|^2_F$. It is also proposed to select the most relevant attributes for each community by adding an $l_1$ norm sparsity term to each column of matrix\\n$\\\\bf V$. As a result, one obtains the following optimization problem: \\n$$\\n\\\\min_{{\\\\bf U}\\\\ge 0,\\\\,{\\\\bf V}\\\\ge 0}\\\\left(\\\\alpha_1\\\\|{\\\\bf S}-{\\\\bf U}{\\\\bf U}^T\\\\|^2_F+\\\\|{\\\\bf U}-{\\\\bf A}{\\\\bf V}\\\\|^2_F+\\\\alpha_2 \\\\sum\\\\nolimits_j \\\\|{\\\\bf V}(\\\\cdot,j)\\\\|^2_1\\\\right),\\n$$\\nwhere $\\\\alpha_1>0$ controls the impact of structure and $\\\\alpha_1\\\\ge 0$ the sparsity penalty. This problem is further approximately solved in an iterative process according to Majorization-Minimization framework .\\n\\\\begin{table}\\n\\\\caption {Non-negative matrix factorization-based and matrix compression-based methods.} \\\\label{tab:NMF} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.6cm}|p{1.8cm}|p{1.8cm}|p{1cm}|p{2cm}|p{2.5cm}|p{2.8cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tAlgorithm & Factorization/ compression type & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf NPei} & 3-factor NNMF & Yes/Yes & Small\\\\newline Medium & Purity & {\\\\bf Twitter} \\\\newline DBLP* & Relational Topic Model \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf 3NCD} & $2$-factor NNMF & Yes/Yes & Medium\\\\newline Large & F1-score \\\\newline Jaccard\\n\\t\\t\\tsimilarity & {\\\\bf ego-Facebook}\\\\newline {\\\\bf ego-Twitter}\\\\newline {\\\\bf ego-G+} & {\\\\bf CESNA} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf SCI} & 2-factor NNMF & Yes/Yes & Medium & ACC \\\\newline NMI \\\\newline GNMI \\\\newline $F$-measure \\\\newline Jaccard similarity & {\\\\bf Citeseer}\\\\newline {\\\\bf Cora}\\\\newline {\\\\bf WebKB}\\\\newline {\\\\bf LastFM} & {\\\\bf PCL-DC} \\\\newline {\\\\bf CESNA} \\\\newline {\\\\bf DCM} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf JWNMF} & 2-factor NNMF & Yes/Yes & Small\\\\newline Medium & Modularity\\\\newline Entropy\\\\newline NMI & Amazon Fail \\\\href{https://www.ipd.kit.edu/mitarbeiter/muellere/consub/}{dataset} \\n\\t\\t\\t\\\\newline Disney \\\\href{http://www.perozzi.net/projects/focused-clustering/}{dataset} \\\\newline Enron \\\\href{https://www.ipd.kit.edu/mitarbeiter/muellere/consub/}{dataset} \\\\newline DBLP-4AREA \\\\href{http://www.perozzi.net/projects/focused-clustering/}{dataset}\\\\newline {\\\\bf WebKB} \\\\newline {\\\\bf Citeseer} \\\\newline {\\\\bf Cora} & {\\\\bf BAGC} \\\\newline {\\\\bf PICS} \\\\newline SANS \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf SCD} & 2- and 3-factor NNMF & Yes/Yes-No & Small\\\\newline Medium & Accuracy \\\\newline NMI & {\\\\bf Twitter}\\\\newline {\\\\bf WebKB} & {\\\\bf SCI} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{\\\\bf ASCD} & 2-factor NNMF &Yes/Yes-No & Small\\\\newline Medium & ACC \\\\newline NMI \\\\newline $F$-measure \\\\newline Jaccard similarity & {\\\\bf LastFM} \\\\newline {\\\\bf WebKB}\\\\newline {\\\\bf Cora} \\\\newline {\\\\bf Citeseer} \\\\newline {\\\\bf ego-Twitter*} \\\\newline {\\\\bf ego-Facebook*} & Block-LDA \\\\newline {\\\\bf PCL-DC} \\\\newline {\\\\bf SCI} \\\\newline {\\\\bf CESNA} \\\\newline {\\\\bf Circles} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{\\\\bf CFOND} & $2$- and $3$-factor NMF & Yes/(Yes/No) & Medium & Accuracy \\\\newline NMI & {\\\\bf Cora}\\\\newline {\\\\bf CiteSeer}\\\\newline {\\\\bf PubMed}\\\\newline {\\\\bf Attack}\\\\newline Synthetic\\n\\t\\t\\t& GNMF \\\\newline DRCC \\\\newline \\n\\t\\t\\tLP-NMTF \\\\newline iTopicModel \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf MVCNMF} & $2$-factor NMF & Yes/Yes & Small\\\\newline Medium & Density\\\\newline Entropy & {\\\\bf Political\\n\\t\\t\\t\\tBlogs}\\\\newline {\\\\bf CiteSeer}\\\\newline {\\\\bf Cora}\\\\newline {\\\\bf WebKB} \\\\newline ICDM (DBLP*) & FCAN \\\\newline SACTL \\\\newline {\\\\bf kNAS} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{\\\\bf PICS} (\\\\href{http://www.andrew.cmu.edu/user/lakoglu/pubs.html#pics}{source}) & Matrix compression (finding rectangular blocks) & No/No & Small\\\\newline Medium & Anecdotal and visual study & Youtube \\\\newline {\\\\bf Twitter}* \\\\newline Phonecall \\\\newline Device \\\\newline {\\\\bf Political Books} (\\\\href{http://www-personal.umich.edu/~mejn/netdata/}{link}) \\\\newline {\\\\bf Political Blogs} & ---\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\\n\\\\begin{remark}\\nAn NNMF-based community detection method for {\\\\it multi-layer} networks is given in .\\n\\\\end{remark}\\nNow we briefly describe {\\\\bf PICS} , the method adopting matrix compression\\\\footnote{The aim of compression methods is to find a\\nshorter form of describing the same information content.} for community detection in node-attributed social networks. {\\\\bf PICS} uses lossless compression principles from to simultaneously compress the network adjacency matrix $\\\\bf S$ and the attribute matrix $\\\\bf A$. As a result of the compression, certain homogeneous rectangular blocks in the matrices can be determined. Groups of the nodes corresponding to the blocks are considered as communities. One should be aware however that nodes within communities found by {\\\\bf PICS} may not be densely connected due to the definition of a community in .\",\n \"id\": \"c986582c-70df-43b2-b7f2-9e2e83ab57f5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 30,\n \"parent_id\": \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Simultaneous fusion methods\"\n ],\n [\n \"subsection\",\n \"Non-negative matrix factorization-based and matrix compression-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Non-negative matrix factorization-based and matrix compression-based methods\"\n },\n {\n \"cite_extract_rate\": 0.22222222222222202,\n \"cites\": [\n 1714,\n 1710\n ],\n \"content\": \"\\\\label{class-sf-patterns}\\nPattern mining in node-attributed social networks focuses on extraction of patterns, e.g. subsets of specific attributes or connections\\\\footnote{An example of a pattern is a maximal clique . Recall that a clique is a subset of nodes in an undirected graph such that every two nodes are adjacent. A clique is called maximal if there is no other clique that contains it.}, in network structure and attributes . Among other things, this helps to make sense of a network and to understand how it was formed. In the context of community detection, the extracted patterns are used as building blocks for communities. \\nThere are many papers devoted to pattern mining in social networks but the majority of them do not deal with the task of community detection. The ones relevant to the topic of this survey are presented in Table~\\\\ref{tab:clique}. It is worth mentioning here that it is common for pattern mining-based methods to detect communities not in the whole network but in its part only (e.g. ). \\n\\\\begin{table}\\n\\\\caption {Pattern mining-based (simultaneous fusion) methods.}\\n\\\\label{tab:clique} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.4cm}|p{1.2cm}|p{2.5cm}|p{2cm}|p{2.5cm}|p{2.0cm}|p{1.8cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & Patterns used & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf DCM} (\\\\href{http://www.patternsthatmatter.org/software.php}{source}) & Semantic patterns ({\\\\it queries}) & Yes/Yes & Small\\\\newline Medium & Community score\\\\newline Conductance \\\\newline Intra-cluster density \\\\newline Modularity & {\\\\bf Delicious}\\\\newline {\\\\bf LastFM}\\\\newline {\\\\bf Flickr} & --- \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf COMODO} & Semantic patterns & Yes/Yes & Small\\\\newline Medium & Description complexity\\\\newline Community size & BibSonomy \\\\newline {\\\\bf Delicious} \\\\newline {\\\\bf LastFM} & {\\\\bf DCM} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf ACDC} & Maximal cliques & Yes/Yes & Medium & Density & {\\\\bf Political Blogs} & {\\\\bf SA-Cluster} \\\\newline {\\\\bf SAC1-SAC2} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\\n\\\\begin{remark}\\nABACUS detects communities by extracting patterns in {\\\\it multi-layer} networks.\\n\\\\end{remark}\",\n \"id\": \"f8f23311-f720-42e4-94c4-ed1c64c919d4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Simultaneous fusion methods\"\n ],\n [\n \"subsection\",\n \"Pattern mining-based (simultaneous fusion) methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pattern mining-based (simultaneous fusion) methods\"\n },\n {\n \"cite_extract_rate\": 0.137931034482758,\n \"cites\": [\n 1712,\n 1721,\n 1714,\n 7062\n ],\n \"content\": \"\\\\label{class-sf-probabilistic-models}\\n\\\\begin{table}[b]\\n\\t\\\\caption{Probabilistic model-based methods.} \\\\label{tab:statistical} \\n\\t\\\\begin{center}\\n\\t\\t{\\\\tiny\\t\\\\begin{tabular}{|p{1.2cm}|p{2.5cm}|p{1.5cm}|p{1.3cm}|p{1.5cm}|p{3cm}|p{2.5cm}|}\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\tMethod & Model features & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf PCL-DC} & Conditional Link Model\\\\newline Discriminative Content model & Yes/No & Medium & NMI \\\\newline Pairwise $F$-measure \\\\newline Modularity\\\\newline Normalized cut & {\\\\bf Cora}\\\\newline {\\\\bf Siteseer} & PHITS-PLSA \\\\newline LDA-Link-Word \\\\newline Link-Content-Factorization \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf CohsMix} & MixNet model & Yes/No & Small & Rand Index & Synthetic \\\\newline Exalead.com search engine dataset & Multiple view learning \\\\newline Hidden Markov Random Field \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf BAGC} \\\\newline {\\\\bf GBAGC} & a Bayesian treatment on distribution parameters & Yes/No & Medium & Modularity\\\\newline Entropy & {\\\\bf Political Blogs}\\\\newline {\\\\bf DBLP10K}\\\\newline {\\\\bf DBLP84K} & {\\\\bf Inc-Cluster} \\\\newline {\\\\bf PICS} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf VEM-BAGC} & Based on {\\\\bf BAGC} & Yes/No & Medium & Modularity\\\\newline Entropy & {\\\\bf Political Blogs}\\\\newline Synthetic networks & {\\\\bf BAGC} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf PPSB-DC} & Popularity-productivity stochastic block model and discriminative content model & Yes/No & Medium & normalized mutual information (NMI)\\\\newline Pairwise\\n\\t\\t\\t\\tF measure (PWF)\\\\newline Accuracy & {\\\\bf Cora} \\\\newline {\\\\bf CiteSeer} \\\\newline {\\\\bf WebKB} & {\\\\bf PCL-DC} \\\\newline PPL-DC \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf CESNA} & A probabilistic generative model assuming communities generate network structure and attributes & No/Yes & Medium\\\\newline Large & Evaluation & {\\\\bf ego-Facebook}\\\\newline {\\\\bf ego-G+} \\\\newline {\\\\bf ego-Twitter}\\\\newline \\n\\t\\t\\t{\\\\bf Wikipedia* (philosophers)}\\\\newline {\\\\bf Flickr} & {\\\\bf CODICIL} \\\\newline {\\\\bf Circles} \\\\newline Block-LDA \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf Circles} & A generative model for friendships in social circles & Yes/Yes & Medium\\\\newline Large & Balanced Error Rate & {\\\\bf ego-Facebook}\\\\newline {\\\\bf ego-G+} \\\\newline {\\\\bf ego-Twitter} & Block-LDA \\\\newline Adapted Low-Rank Embedding \\\\\\\\\\n\\t\\t\\t\\t\\\\hline \\n\\t\\t\\t\\t{\\\\bf SI} & A modified version of a stochastic block model & Yes/No & Small\\\\newline Medium & Normalized\\n\\t\\t\\t\\tmutual information (NMI) & Synthetic \\\\newline High school friendship network \\\\newline Food web of marine species in the Weddell Sea \\\\newline Harvard Facebook friendship network \\\\newline Malaria HVR 5 and 6 gene\\n\\t\\t\\t\\trecombination network & --- \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t{\\\\bf NEMBP} & A generative model with learning method using a nested EM algorithm with belief propagation & Yes/(Yes/No) & Small \\\\newline Medium & Accuracy \\\\newline NMI \\\\newline GNMI \\\\newline F-score \\\\newline Jaccard & {\\\\bf WebKB} \\\\newline {\\\\bf ego-Twitter*}\\\\newline\\n\\t\\t\\t\\t{\\\\bf ego-Facebook*}\\\\newline {\\\\bf CiteSeer} \\\\newline {\\\\bf Cora}\\\\newline {\\\\bf Wikipedia*}\\\\newline {\\\\bf Pubmed} & Block-LDA \\\\newline {\\\\bf PCL-DC} \\\\newline {\\\\bf CESNA} \\\\newline {\\\\bf DCM} \\\\newline {\\\\bf SCI} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\t{\\\\bf NBAGC-FABAGC} & A nonparametric and asymptotic Bayesian model selection method based on {\\\\bf BAGC} & No/No & Medium & NMI \\\\newline Modularity \\\\newline Entropy & Synthetic \\\\newline {\\\\bf Political Blogs} \\\\newline {\\\\bf DBLP10K} \\\\newline {\\\\bf DBLP84K} & {\\\\bf PICS} \\\\\\\\\\n\\t\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\end{tabular}\\n\\t\\t}\\n\\t\\\\end{center}\\n\\\\end{table}\\nMethods from this subclass probabilistically infer the distribution of community memberships for nodes in a node-attributed social network under the assumption that network structure and attributes are generated according to chosen parametric distributions. Generative and discriminative models are mainly used for the inferring. It is worth mentioning that it is though a non-trivial task to ``properly'' choose a priori distributions for structure and attributes .\\nShort descriptions of the methods from this subclass are given in Table~\\\\ref{tab:statistical}. Pay attention that this table does not contain any method preceding to and this requires the following additional comments. According to , several probabilistic model-based clustering methods for node-attributed networks had been proposed before , for example, in . However, they focus on node-attributed {\\\\it document} networks which are out of scope of the present survey. That is why they are non included in Table~\\\\ref{tab:statistical}.\\n\\\\begin{remark}\\nTUCM proposes a generative Bayesian model for detecting communities in {\\\\it multi-layer} networks where different types of interactions between social actors are possible. \\n\\\\end{remark}\",\n \"id\": \"45b099b6-e934-4199-b2c6-b9c78b20f692\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 29,\n \"parent_id\": \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Simultaneous fusion methods\"\n ],\n [\n \"subsection\",\n \"Probabilistic model-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Probabilistic model-based methods\"\n },\n {\n \"cite_extract_rate\": 0.22222222222222202,\n \"cites\": [\n 1714,\n 1712\n ],\n \"content\": \"\\\\label{class-sf-dynamics}\\nMethods from this subclass (see Table~\\\\ref{tab:agent}) treat a node-attributed social network as a dynamic system and assume that its community structure is a consequence of certain interactions among nodes (of course, the attributes are thought to affect the interactions). Some of the methods assume that the interactions occur in an information propagation process, i.e. while information is sent to or received from every node. Others comprehend each node as an autonomous agent and develop a multi-agent system to detect communities. Note that these approaches are rather recent and consider community detection in node-attributed social networks from a new perspective. Furthermore, they seem to be efficient for large networks as can be easily parallelized.\\n\\\\begin{table}\\n\\\\caption{Dynamical system-based and agent-based methods.}\\n\\\\label{tab:agent} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.2cm}|p{3.2cm}|p{1.7cm}|p{1.5cm}|p{2cm}|p{2cm}|p{1.8cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & Description & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf CPIP}-{\\\\bf CPRW} & Content (information) propagation models: a linear approximate model\\n\\t\\t\\tof influence propagation ({\\\\bf CPIP}) and content propagation\\n\\t\\t\\twith the random walk principle ({\\\\bf CPRW}) & Yes/Yes & Medium & F-score\\\\newline Jaccard Similarity\\\\newline NMI & {\\\\bf CiteSeer}\\\\newline {\\\\bf Cora}\\\\newline {\\\\bf ego-Facebook}\\\\newline {\\\\bf PubMed Diabetes} & Adamic Adar \\\\newline {\\\\bf PCL-DC} \\\\newline {\\\\bf Circles} \\\\newline {\\\\bf CODICIL} \\\\newline {\\\\bf CESNA} \\\\\\\\\\n\\t\\t\\t\\\\hline \\n{\\\\bf CAMAS} & Each node with attributes as an autonomous agent with influence in a cluster-aware multiagent system & No/Yes & Medium\\\\newline Large & Coverage Rate\\\\newline Normalized Tightness \\\\newline Normalized Homogeneity\\\\newline F1-Score \\\\newline Jaccard\\\\newline Adjusted Rand Index & Synthetic \\\\newline {\\\\bf ego-Facebook}\\\\newline {\\\\bf ego-Twitter*} \\\\newline {\\\\bf ego-G+} & {\\\\bf CESNA} \\\\newline EDCAR \\\\\\\\\\n\\\\hline \\n{\\\\bf SLA} & A dynamic cluster formation game played by all nodes and clusters in a discrete-time dynamical system & Yes/No & Medium \\\\newline Large & Density \\\\newline Entropy \\\\newline F1-score \\\\newline & {\\\\bf Delicious}\\\\newline {\\\\bf LastFM} \\\\newline {\\\\bf ego-Facebook}\\\\newline {\\\\bf ego-Twitter*} \\\\newline {\\\\bf ego-G+} & {\\\\bf CESNA} \\\\newline EDCAR \\\\\\\\\\n\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\\n\\\\begin{figure}[b]\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=0.9\\\\linewidth]{late.pdf}\\n\\t\\\\caption{A typical scheme of a late fusion method.}\\n\\t\\\\label{fig:late-fusion}\\n\\\\end{figure}\",\n \"id\": \"6a22c128-a708-4842-b8d3-074bec4c46a4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"2eeee52e-f7f4-4129-a5ef-03fe424ccf1f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Simultaneous fusion methods\"\n ],\n [\n \"subsection\",\n \"Dynamical system-based and agent-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamical system-based and agent-based methods\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8487\n ],\n \"content\": \"\\\\label{section8}\\nRecall that late fusion methods intend to fuse structural and attributive information after the community detection process. More precisely, community detection is first separately performed for structure (e.g. by Louvain ) and attributes (e.g. by $k$-means ). After that, the partitions obtained are fused somehow in order to get the resulting structure- and attributes-aware partition, see Figure~\\\\ref{fig:late-fusion}.\\nNote that late fusion methods usually allow a researcher/data scientist to use existing implementations of classical community detection and consensus clustering algorithms to get the required partition.\",\n \"id\": \"a4712ae6-1690-4727-8bbb-b44187674ee8\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Late fusion methods\"\n ]\n ],\n \"subsections\": [\n \"a2976c7a-9d26-4a9e-9cb2-40daafc174e6\",\n \"edd7b709-6935-46b9-9d57-cbc247202cf8\"\n ],\n \"title\": \"Late fusion methods\"\n },\n {\n \"cite_extract_rate\": 0.35714285714285704,\n \"cites\": [\n 8484,\n 1712,\n 1722,\n 1715\n ],\n \"content\": \"\\\\label{class-lf-consensus}\\nGiven a set (also known as an ensemble) of partitions, the general goal of consensus clustering algorithms is to find a single consolidated partition that aggregates information in the ensemble . A recent survey on such methods can be found e.g. in . The idea behind consensus clustering is clearly appropriate for community detection in node-attributed social networks if one has an ensemble of partitions obtained separately (or maybe even jointly) for structure and attributes. Table~\\\\ref{tab:late} contains short descriptions of the methods applying the idea.\\n\\\\begin{table}[t]\\n\\\\caption {Consensus-based methods.} \\\\label{tab:late} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.2cm}|p{3.0cm}|p{1.9cm}|p{1.5cm}|p{1.5cm}|p{2.3cm}|p{2cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & Fusing the partitions & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf LCru} & Row-manipulation in the contingency matrix for the partitions & No/No & Small\\\\newline Medium & ARI \\\\newline Density\\\\newline Entropy & Facebook$^*$\\\\newline {\\\\bf DBLP10K} & ---\\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Multiplex} & Multiplex representation scheme (attributes and structure are clustered separately as layers and then combined via consensus ) & No/Yes & Medium\\\\newline Large & F1-score & Synthetic \\\\newline {\\\\bf ego-Twitter} \\\\newline {\\\\bf ego-Facebook} \\\\newline {\\\\bf ego-G+} & {\\\\bf CESNA} \\\\newline {\\\\bf 3NCD} \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t{\\\\bf WCMFA} & Association matrix with weighting based on structure- and attributes-aware similarity & Depends on the partitions & Small & Rand index \\\\newline ARI \\\\newline NMI & {\\\\bf Consult}~\\\\newline \\t\\t\\n\\t\\t\\tLondon Gang \\\\newline\\n\\t\\t\\tMontreal Gang & {\\\\bf WMen} \\\\\\\\\\n\\t\\t\\t\\\\hline \\t\\t\\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\\n\\\\begin{remark}\\nGeneral-purpose consensus clustering algorithms for {\\\\it multi-layer} networks\\nare considered in .\\n\\\\end{remark}\",\n \"id\": \"a2976c7a-9d26-4a9e-9cb2-40daafc174e6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"a4712ae6-1690-4727-8bbb-b44187674ee8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Late fusion methods\"\n ],\n [\n \"subsection\",\n \"Consensus-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Consensus-based methods\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 8484\n ],\n \"content\": \"\\\\label{class-lf-switch}\\nThe only method included in this subclass (see Table~\\\\ref{tab:late2}) also deals with partitions obtained separately for structure and attributes but chooses a more ``preferable'' one instead of finding consensus. Namely, {\\\\bf Selection} switches from a structure-based to an attributes-based partition when the former one is {\\\\it ambiguous}. This refers to the case when the so-called {\\\\it estimated mixing parameter} $\\\\mu$ for the structure-based partition is less then a certain experimental value $\\\\mu_{\\\\lim}$ associated with a significant drop in clustering quality on synthetic networks . An interested reader can find the precise definitions of $\\\\mu$ and $\\\\mu_{\\\\lim}$ in .\\n\\\\begin{table}[t]\\n\\\\caption {Switch-based methods.} \\\\label{tab:late2} \\n\\\\begin{center}\\n\\t{\\\\tiny\\n\\t\\t\\\\begin{tabular}{|p{1.2cm}|p{3.0cm}|p{1.9cm}|p{1.5cm}|p{1.5cm}|p{2.3cm}|p{2cm}|}\\n\\t\\t\\t\\\\hline \\n\\t\\t\\tMethod & Fusing the partitions & Require the number of clusters/ Clusters overlap & Size of datasets used for evaluation & Quality measures & Datasets used & Compared with \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\t{\\\\bf Selection} & Switching between the partitions & Depends on the partitions & Medium & NMI \\\\newline Modularity & Synthetic LFR benchmark \\\\newline {\\\\bf DBLP84K} & {\\\\bf BAGC} \\\\newline {\\\\bf OCru} \\\\newline {\\\\bf SA-Cluster} \\\\newline HGPA-CSPA \\\\\\\\\\n\\t\\t\\t\\\\hline \\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\\\end{center}\\n\\\\end{table}\",\n \"id\": \"edd7b709-6935-46b9-9d57-cbc247202cf8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"a4712ae6-1690-4727-8bbb-b44187674ee8\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Late fusion methods\"\n ],\n [\n \"subsection\",\n \"Switch-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Switch-based methods\"\n },\n {\n \"cite_extract_rate\": 0.347826086956521,\n \"cites\": [\n 7062,\n 1712,\n 1714,\n 1723,\n 1724,\n 1721\n ],\n \"content\": \"\\\\label{section9}\\nThe information collected in Sections~\\\\ref{section6}--\\\\ref{section8} allows us to analyze the overall situation in the field. Particularly, we would like to determine which methods are {\\\\it state-of-the-art}\\\\footnote{According to Cambridge Dictionary, state-of-the-art means ``the best and most modern of its type''.}. This is probably what a researcher/data scientist facing the community detection problem in node-attributed social networks expects from our survey.\\nWe start with observing the directed graph based on the information in Sections~\\\\ref{section6}--\\\\ref{section8} and showing method-method comparisons, see Figure~\\\\ref{fig:graph1}. It requires several comments though. First, there are methods in Sections~\\\\ref{section6}--\\\\ref{section8} that are not compared with others for community detection in node-attributed networks or compared with a few. For this reason, we include in the graph only nodes (representing methods) whose degree is at least two. This means that there are at least two comparison experiments with each method presented in Figure~\\\\ref{fig:graph1}. Note that 46 methods are shown in the graph of 75 classified in the survey. \\nSecond, the directed edges in the graph show the existing method-method comparisons. For example, the directed edge from node {\\\\bf CESNA} to node {\\\\bf CODICIL} indicates that the authors of {\\\\bf CESNA} compared their method with {\\\\bf CODICIL} and showed that {\\\\bf CESNA} outperforms {\\\\bf CODICIL} in some sense (community detection quality, computational efficiency, etc.). This is applied to all edges in the graph.\\n\\\\begin{figure}[b]\\n\\t\\\\label{fig:graph1}\\n \\\\includegraphics[scale=0.95]{Graph1new.pdf}\\n\\t\\\\caption{The directed graph of existing method-method comparisons. Nodes (shown only those with degree $\\\\ge 2$) represent the methods classified in the present survey. The most influential methods (nodes with highest PageRank) are filled green so that the darker green means the higher PageRank.}\\n\\\\end{figure}\\nWhat is more, we used PageRank to detect the most important or, better to say, {\\\\it most influential methods} in the field. Nodes with the highest PageRank values are filled green in Figure~\\\\ref{fig:graph1} so that the darker green means the higher PageRank. It turns out that the most influential ones are (in the order they discussed in Sections~\\\\ref{section6}--\\\\ref{section8}):\\n\\\\begin{itemize}\\n\\\\item weight-based {\\\\bf SAC2}~ and {\\\\bf CODICIL}~ (Subsection~\\\\ref{class-ef-weight}),\\n\\\\item node-augmented graph-based {\\\\bf SA-Cluster} , {\\\\bf Inc-Cluster} and {\\\\bf SA-Cluster-Opt}~ (Subsection~\\\\ref{class-ef-node-augmented-graph}),\\n\\\\item {\\\\bf SAC1}~ modifying the Louvain objective function (Subsection~\\\\ref{class-sf-modifying}),\\n\\\\item NNMF-based {\\\\bf SCI} and matrix compression-based {\\\\bf PICS} (Subsection~\\\\ref{class-sf-nnmf}),\\n\\\\item pattern mining-based (simultaneous fusion) {\\\\bf DCM} (Subsection~\\\\ref{class-sf-patterns}),\\n\\\\item probabilistic model-based {\\\\bf PCL-DC} , {\\\\bf BAGC} , {\\\\bf GBAGC} , {\\\\bf CESNA} and {\\\\bf Circles} (Subsection~\\\\ref{class-sf-probabilistic-models}).\\n\\\\end{itemize}\\nIn our opinion, these methods may be though to be chosen by researchers' community as those determining further developments in the field. Thus we encourage a newcomer in the field to get familiar with them first to see the main ideas and techniques existing for community detection in node-attributed social networks.\\nAt the same time, we would not consider the most influential methods as state-of-the-art as other methods are shown to outperform them in some sense.\\nWhat else the graph in Figure~\\\\ref{fig:graph1} can tell us about? We emphasize that it does not contain 29 methods of those discussed in Sections~\\\\ref{section6}--\\\\ref{section8} and furthermore it is rather sparse and even disconnected. These points lead to the conclusion that the comparison study of the methods in the field is far from being complete. We would even strengthen the conclusion made by saying that a researcher/data scientist cannot be sure that the method chosen for practical use is preferable to other existing ones, even if there is an edge in Figure~\\\\ref{fig:graph1}. The problem is that the comparison experiments (represented via the edges) are made by different means, e.g. different quality measures, datasets, hyperparameter tuning strategies, etc. Let us discuss it below in more detail.\\nSuppose for a second that two methods show the same community detection quality for a number of datasets and measures, and their hyperparameters are tuned to provide the best possible results. Then we should think how much time/space each method uses for it. Particularly, we may think of method's computational complexity in terms of the number of vertices $n$, the number of edges $m$ and the attribute dimension $d$ in a node-attributed graph. Such estimates exist for certain methods, particularly for some of the most influential ones. Examples are {\\\\bf CODICIL}~ with $O(n^2\\\\log n)$, {\\\\bf SA-Cluster} with $O(n^3)$ and {\\\\bf CESNA} with $O(m)$. However, we could not find such estimates for the majority of methods discussed in Sections~\\\\ref{section6}--\\\\ref{section8} as authors often omit such estimation. This makes the overall comparison of methods in terms of computational complexity impossible.\\nNow let us discuss the hyperparameter tuning problem. It turns out that some authors tune hyperparameters in their methods manually, some just do not consider the problem at all. Another issue is the lack of a general understanding how to determine ``equal impact'' of structure and attributes on the community detection results. For example, in the weight-based early fusion methods (Subsection~\\\\ref{class-ef-weight}) some authors choose $\\\\alpha=1/2$ in experiments hoping that this provides the equal impact. However, if ones takes into account the different nature of structural and attributive information and the disbalance between associated statistical features, this choice seems questionable in a general situation.\\nFurthermore, authors use different datasets (of various size and nature, see Section~\\\\ref{section5}) and quality measures to test their methods so that a unified comparison of experimental results in different papers cannot be carried out\\\\footnote{A more general observation is that several comparison experiments clearly do not provide generality of conclusions. This however seems to be a hot topic among supporters of scientific research from one side and of empirical one from another side.}. What is more, datasets and software implementations used in comparison experiments are rarely provided by the authors (especially of out-of-date papers) thus making reproducing their results time-consuming or even impossible. Note how few links to source codes are indicated in Sections~\\\\ref{section6}--\\\\ref{section8}.\\nLet us now discuss the methodology of using quality measures in comparison experiments. There exist two main strategies (see Section~\\\\ref{section5}). Namely, community detection quality can be evaluated (a) by heuristic measures of structural closeness and attributes homogeneity (e.g. Modularity and Entropy) when the dataset under consideration has no ``ground truth'' and (b) by measures estimating the agreement between the detected communities and the ground truth ones (e.g. NMI or ARI).\\nLet us first discuss Option (a) in terms of a typical weight-based early fusion method from Subsection~\\\\ref{class-ef-weight}. For a given node-attributed graph $G$, suppose that we first convert its attributes $(\\\\mathcal{V},\\\\mathcal{A})$ into attributive graph $G_A$ and further fuse it with the structure $(\\\\mathcal{V},\\\\mathcal{E})$. This results in weighted graph $G_W$ that is thought to contain information about both the structure and the attributes in a unified form. After that, we find communities in $G_W$ that provide, say, the maximum of Modularity of $G_W$ (e.g. by Weighted Louvain). At the evaluation step we further calculate Modularity of $(\\\\mathcal{V},\\\\mathcal{E})$ and Entropy of $(\\\\mathcal{V},\\\\mathcal{A})$ basing on the partition found. This seems reasonable as far as you do not look at such a methodology critically. Indeed, note that we deal with one quality measure within the optimization process (Modularity of $G_W$) but evaluate community detection performance by other measures (Modularity of $(\\\\mathcal{V},\\\\mathcal{E})$ and Entropy of $(\\\\mathcal{V},\\\\mathcal{A})$). Generally speaking, how can we be sure that optimization of one objective function provides optimal values of other objective functions, if there is no mathematically established connection between them? Of course, this may be simply called a heuristic but it looks more a logical gap, from our point of view. Anyways, we are unaware of any explicit explanation of such a methodology. What is more, one should carefully study how the change of data representation within a method (the change of vectors $\\\\mathcal{A}$ for edge weights in $G_A$ in the above-mentioned example) affects community detection results.\\nTake into account that a similar discussion is suitable for the majority of methods in Sections~\\\\ref{section6}--\\\\ref{section8} that use quality measures for estimating structural closeness and attributes homogeneity of the detected communities. The exception is the methods that directly optimize the quality measures {\\\\it within the optimization process}. Examples are the simultaneous fusion methods {\\\\bf OCru} , \\n{\\\\bf SAC1} ,\\n{\\\\bf ILouvain} , \\n{\\\\bf UNCut} , \\n{\\\\bf MOEA-SA} ,\\n{\\\\bf MOGA-@Net} and \\n{\\\\bf JCDC} .\\nJust by construction, they aim at providing optimal values of the quality measures. One should however take into account the precision of the optimization method applied. \\nNow we turn our attention to the methodology of evaluating community detection quality by measures that estimate the agreement between the detected communities and the ground truth ones (Option (b)). In our opinion, such a methodology makes sense for synthetic node-attributed social networks as the way how the communities are formed is known in advance. As for real-world networks, it seems somewhat questionable as ground truth communities may reflect only one point of view on network communities (among many possible). Therefore, expecting that a community detection method would share this point of view seems a bit unsuitable. There are several works on this issue, e.g. , discussing connections between ground truth, attributes and quality measures in detail, and therefore we refer the interested reader to them.\\nRecall that we started this discussion trying to determine the state-of-the-art methods for community detection in node-attributed social networks. Unfortunately, the above-mentioned facts do not give us a chance to do this. Of course, we could simply list the most recent methods in the field (and then it is enough to check just the time of publication), but this certainly does not meet the requirements imposed on state-of-the-art methods.\",\n \"id\": \"ec3f6cdd-43b4-49d2-ab2f-b07a25b101ed\",\n \"level\": \"section\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Analysis of the overall situation in the field\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Analysis of the overall situation in the field\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{section10}\\nIt is shown in the survey that there exist a large amount of methods for community detection in node-attribute social networks based on different ideas. In particular, we gave short descriptions of 75 most relevant methods and mentioned much more those partly related to the topic (Sections~\\\\ref{section6}--\\\\ref{section8}).\\nWe also proposed to divide the methods into the three classes --- early fusion, simultaneous fusion and late fusion ones (Section~\\\\ref{section4}). This classification is based on the moment when network structure and attributes are fused within a method and allows a researcher/data scientist to estimate the ease of method's software implementation. Namely, we concluded that early and late fusion methods can be easier implemented in comparison with simultaneous fusion ones as the former two usually can be combined of several classical community detection algorithms with existing implementations. At a lower lever, we also divided the methods into subclasses of fusion techniques used (Section~\\\\ref{section4}). This allows one to estimate the methodological variety in the field.\\nWithin the classification, we also focused on the experimental part so that one can see which datasets and measures are used for quality evaluation of each method from Sections~\\\\ref{section5}--\\\\ref{section8}.\\nThe analysis of all the information collected brought us to the unfortunate conclusion that it is impossible now to determine state-of-the-art methods for community detection in node-attributed social networks (Section~\\\\ref{section9}). This is a result of the presence of the following general problems in the field that we disclosed in the survey:\\n \\\\begin{itemize}\\n \\\\item the terminology in the field is rather unstable (Subsection~\\\\ref{subsection 3.2});\\n \\\\item there is no generally accepted opinion on the effect of fusing structure and attributes, in particular, on when the fusion is helpful and when not in terms of subclasses of node-attributed social networks (Subsection~\\\\ref{subsection-str-closeness-attr-homo}); \\n\\\\item there is no unified methodology for experimental comparison of methods that would include estimation of computational complexity, use of a unified set of datasets and quality measures for evaluation, and justified hyperparameter tuning procedures (Section~\\\\ref{section9});\\n\\\\item there is no general understanding what is the ``equal impact'' of structure and attributes on community detection results (Section~\\\\ref{section9}),\\n \\\\item as a rule, there is no mathematically established connection between computational processes within a community detection method and the quality measured used for its evaluation (Section~\\\\ref{section9}).\\n \\\\end{itemize}\\nSummarizing, we concluded that the comparison study allowing one to determine the most preferable (in any sense) methods in the field is far from being complete.\\nNevertheless, community detection methods dealing both with network structure and attributes remain a powerful tool in social network analysis and can yield useful insights about a node-attributed social network to a researcher/data scientist. Furthermore, they have wide applications even beyond social networks (Section~\\\\ref{section3}). With respect to these, we believe that the formulation of existing problems in the field done in this survey is the first step in finding solutions to them.\\n\\\\section*{Competing Interests Statement}\\nThere are no competing interests in publication of this paper.\\n\\\\section*{Acknowledgements}\\nThe author is very grateful to Klavdiya Bochenina and Timofey Gradov for numerous conversations on the topic and especially to the Anonymous Referee for useful comments that helped to improve the quality of exposition in the survey essentially.\\nThis research is financially supported by Russian Science Foundation, Agreement 17-71-30029 with co-financing of Bank Saint Petersburg, Russia.\\n{\\\\small\\n\\\\bibliographystyle{acm}\\n\\\\bibliography{sample-base}\\n}\\n\\\\end{document}\",\n \"id\": \"7362ce42-549a-4724-9b4c-b963b4574449\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4b2a7071-f315-4436-8aea-91d8d6db7afb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Community detection in node-attributed social networks: a~survey\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":88,"string":"88"},"all_cites":{"kind":"list like","value":[8481,1710,8483,1712,8482,1711,8484,217,9116,1713,1714,8485,1715,8486,1716,7062,8487,8488,1717,242,8489,215,282,1719,1718,229,1010,1720,1721,1722,1723,1724],"string":"[\n 8481,\n 1710,\n 8483,\n 1712,\n 8482,\n 1711,\n 8484,\n 217,\n 9116,\n 1713,\n 1714,\n 8485,\n 1715,\n 8486,\n 1716,\n 7062,\n 8487,\n 8488,\n 1717,\n 242,\n 8489,\n 215,\n 282,\n 1719,\n 1718,\n 229,\n 1010,\n 1720,\n 1721,\n 1722,\n 1723,\n 1724\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.13895468275926,"string":"1.138955"}}},{"rowIdx":55,"cells":{"authors":{"kind":"list like","value":["Xiaocong Chen","Lina Yao","Julian McAuley","Guanglin Zhou","Xianzhi Wang"],"string":"[\n \"Xiaocong Chen\",\n \"Lina Yao\",\n \"Julian McAuley\",\n \"Guanglin Zhou\",\n \"Xianzhi Wang\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-09-08T10:44:55Z","string":"2021-09-08T10:44:55Z"},"category":{"kind":"string","value":"cs.IR"},"abstract":{"kind":"string","value":" In light of the emergence of deep reinforcement learning (DRL) in recommender systems research and several fruitful results in recent years, this survey aims to provide a timely and comprehensive overview of the recent trends of deep reinforcement learning in recommender systems. We start with the motivation of applying DRL in recommender systems. Then, we provide a taxonomy of current DRL-based recommender systems and a summary of existing methods. We discuss emerging topics and open issues, and provide our perspective on advancing the domain. This survey serves as introductory material for readers from academia and industry into the topic and identifies notable opportunities for further research. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4bb82714-5571-4179-9dac-693fb550f058","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"]],"subsections":["f16a1a1e-25a7-4e14-aaa7-69bb7d4fc8b0","a6f31a50-0b05-46c3-ac49-a2a06aea3f6f","696bfdf7-6191-445e-82a8-ae4df2378853","3dd345fb-ce99-4727-b174-a3f36f9f711a","62c3e9c3-0318-4d9a-947c-bf1d76096072","db61769c-00fa-49cf-aca1-794fe6209774","0a382833-431c-4392-982a-66ab0eba2961"],"title":"root"},{"cite_extract_rate":0.4,"cites":[8983,5194],"content":"Recent years have seen significant progress in recommendation techniques, from traditional recommendation techniques, e.g., collaborative filtering, content-based recommendation and matrix factorization~, to deep learning based techniques.\nIn particular, deep learning show strong advantages in solving complex tasks and dealing with complex data, due to its capability to capture non-linear user-item relationships and deal with various types of data sources such as images and text.\nIt has thus been increasingly used in recommender systems.\nDeep learning-based recommender systems have limitations in capturing interest dynamics~ due to distribution shift, i.e., the training phase is based on an existing dataset which may not reflect real user preferences that undergo rapid change.\nIn contrast, deep reinforcement learning (DRL) aims to train an agent that can learn from interaction trajectories provided by the environment by combining the power of deep learning and reinforcement learning. Since an agent in DRL can actively learn from users' real-time feedback to infer dynamic user preferences,\nDRL is especially suitable for learning from interactions, such as human-robot collaboration; it has also driven significant advances in a range of interactive applications ranging from video games, Alpha Go to autonomous driving~.\nIn light of the significance and recent progresses in DRL for recommender sytsems, we aim to timely summaize and comment on DRL-based recommendation systems in this survey.\nA\nrecent survey on reinforcement learning based recommender systems provides a general review about reinforcement learning in recommender systems without a sophsiticated investigation of the growing area of deep reinforcement learning. Our survey distinguishes itself in providing a systematic and comprehensive overview of existing methods in DRL-based recommender systems, along with a discussion of emerging topics, open issues, and future directions. This survey introduces researchers, practitioners and educators into this topic\nand fostering an understanding of the key techniques in the area.\nThe main contributions of this survey include the following:\n\\begin{itemize}\n \\item We provide an up-to-date comprehensive review of deep reinforcement learning in recommender systems, with state of the art techniques and pointers to core references. To the best of our knowledge, this the first comprehensive survey in deep reinforcement learning based recommender systems.\n \\item We present a taxonomy of the literature of deep reinforcement learning in recommender systems. Along with the outlined taxonomy and literature overview, we discuss the benefits, drawbacks and give suggestions for future research directions. \n \\item We shed light on emerging topics and open issues for DRL-based recommender systems. We also point out future directions that could be crucial for advancing DRL-based recommender systems.\n\\end{itemize}\nThe remainder of this survey is organized as follows: Section 2 provides an overview of recommender systems, DRL and their integration. Section 3 provides a literature review with a taxonomy and classification mechanism. Section 4 reviews emerging topics, and Section 5 points out open questions. Finally, Section 6 provides a few promising future directions for further advances in this domain.\n\\begin{figure}\n \\centering\n \\includegraphics[width=\\linewidth]{route.pdf}\n \\caption{Taxonomy of Deep Reinforcement Learning based Recommender Systems in this survey}\n \\label{fig:overview}\n\\end{figure}","id":"f16a1a1e-25a7-4e14-aaa7-69bb7d4fc8b0","level":"section","origin_cites_number":5,"parent_id":"4bb82714-5571-4179-9dac-693fb550f058","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"In this section,we introduce key concepts related to dynamic recommender systems (RS) and deep reinforcement learning (DRL), and motivate the introduction of DRL to dynamic recommender systems.","id":"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f","level":"section","origin_cites_number":0,"parent_id":"4bb82714-5571-4179-9dac-693fb550f058","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Background"]],"subsections":["a7e80129-286c-44ec-84f0-48ae3a07002b","e05ae3e8-b24d-431b-b181-250a9bf1fb50","467d3d60-cfac-40c2-8a3f-6dbe98149287"],"title":"Background"},{"cite_extract_rate":0,"cites":[],"content":"Recommender systems require coping with \\emph{dynamic} environments by estimating rapidly changing users' preferences and proactively recommending items to users.\nLet $\\mathcal{U}$ be a set of users of cardinality $|\\mathcal{U}|$ and $\\mathcal{I}$ be a set of items of cardinality $|\\mathcal{I}|$.\nFor each user $u\\in\\mathcal{U}$, we observe a sequence of user actions $\\mathbb{X}^u = [x_1^u, x_2^u, \\cdots , x_{T_u}^u]$ with item $x_t^u \\in \\mathcal{I}$, i.e., each event in a user sequence comes from the item set. We refer to a user making a decision as an interaction with an item. Suppose the feedback (e.g., ratings or clicking behavior) provided by users is $\\mathcal{F}$, then a dynamic recommender system maintains the corresponding recommendation policy $\\pi^u_t$, which will be updated systematically based on the feedback $f^u_i \\in \\mathcal{F}$ received during the interaction for item $i \\in \\mathcal{I}$ at the timestamp $t$. \nThe marriage of deep learning and reinforcement learning has fueled breakthroughs in recommender systems. \nDRL-based RS consists of a pipeline with three building blocks: environment construction, state representation and recommendation policy learning. Environment construction aims to build an environment based on a set of users' historical behaviors. State representation is provided by the environment containing certain user information including historical behavior, demographic data (etc.). Recommendation policy learning is the key component to understand and predict users' future behavior. \nDL-based RS receives user feedback (e.g., ratings or clicks) to reflect users' interests and update the recommender, while DRL-based RS receives the reward provided by the environment to update the policy. The reward provided by the environment is a pre-defined function containing several factors. The detailed process of DL based RS and DRL-based RS mapping can be found in~\\Cref{fig:catnew}.","id":"a7e80129-286c-44ec-84f0-48ae3a07002b","level":"subsection","origin_cites_number":0,"parent_id":"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Background"],["subsection","Why Deep Reinforcement Learning for Recommendation?"]],"subsections":[],"title":"Why Deep Reinforcement Learning for Recommendation?"},{"cite_extract_rate":0.14285714285714202,"cites":[8105],"content":"The typical defining feature of DRL is to use the deep learning to approximate reinforcement learning's value function and solve high-dimensional Markov Decision Processes (MDPs)~. Formally, a MDP can be represented as a tuple ($\\mathcal{S},\\mathcal{A},\\mathcal{P},\\mathcal{R},\\gamma$). The agent chooses an action $a_t\\in\\mathcal{A}$ according to the policy $\\pi_t(s_t)$ at state $s_t\\in\\mathcal{S}$. The environment receives the action and produces a reward $r_{t+1}\\in\\mathcal{R}$ and transfers the reward into the next state $s_{t+1}$ according to the transition probability $P(s_{t+1}|s_t,a_t)\\in\\mathcal{P}$. The transition probability $\\mathcal{P}$ is unknown beforehand in DRL.\nSuch a process continues until the agent reaches the terminal state or exceeds a pre-defined maximum time step. The overall objective is to maximize the expected discounted cumulative reward,\n\\begin{align}\n \\mathbb{E}_\\pi [r_t] = \\mathbb{E}_\\pi \\big[\\sum_{0}^\\infty \\gamma^k r_{t+k}\\big] \n\\end{align}\nwhere $\\gamma \\in [0,1]$ is the discount factor that balances the future reward and the immediate reward.\n\\begin{wrapfigure}{r}{0.45\\textwidth}\n \\begin{center}\n \\includegraphics[width=0.45\\textwidth]{tax.pdf}\n \\end{center}\n \\caption{Taxonomy of Deep Reinforcement Learning in Recommender Systems}\n \\label{fig:tax}\n\\end{wrapfigure}\nDeep reinforcement learning can be divided into two categories: \\emph{model-based} and \\emph{model-free} methods (a detailed taxonomy can be found in \\Cref{fig:tax}). The major {\\emph difference} between the two is whether the agent can learn a model of the environment. Model-based methods aim to estimate the transition function and reward function, while model-free methods aim to estimate the value function or policy from experience. In model-based methods, the agent accesses the environment and plans ahead while model-free methods gain sample efficiency from using models which are more extensively developed and tested than model-based methods in recent literature~. \nDeep reinforcement learning approaches are divided into three streams: \\emph{value-based}, \\emph{policy-based} and \\emph{hybrid} methods. In value-based methods, the agent updates the value function to learn a policy; policy-based methods learn the policy directly; and hybrid methods combine value-based and policy-based methods called \\emph{actor-critic} methods. Actor-critic contains two different networks where an actor network uses a policy-based method and the critic uses a value-based method to evaluate the policy learned by the agent.\n\\begin{table}[ht]\n \\centering\n \\caption{Notations}\n \\begin{tabular}{cc|ccc}\n \\hline\n Notations & Name & Notations & Name & Notes \\\\\\hline\n $Q(\\cdot)$ & Q-Value Function & $s$ & State & users' preference \\\\\n $V(\\cdot)$ & Value Function & $a$ & Action & Recommended item(s) \\\\\n $\\gamma$ & Discount Factor & $\\pi$, $\\mu(\\cdot)$ & Policy & Recommendation policy \\\\\n $\\mathbb{E}$ & Expected Value & $r(\\cdot,\\cdot)$ & Reward & users' click behavior \\\\\n $\\theta$ & Model Parameter & $\\alpha$ & constant $\\in [0,1]$ & - \\\\\n $p(\\cdot)$ & Transition Probability & $\\tau$ & Sampled Trajectory & A tuple $(s_t,a_t,s'_t,r_t)$ \\\\\\hline\n \\end{tabular}\n \\label{tab:my_label}\n\\end{table}\nDeep reinforcement learning can be divided into \\emph{on-policy} and \\emph{off-policy} methods.\nIn off-policy methods,\nthe behavior policy $\\pi_b$ is used for exploration while the target policy $\\pi$ is used for decision-making.\nFor on-policy methods, the behavior policy is the same as the target policy. \n\\textbf{Q-learning}~ is an off-policy value-based learning scheme for finding a greedy target policy:\n\\begin{align}\n \\pi(s) = \\argmax_a Q_\\pi (s,a)\n\\end{align}\nwhere $Q_u (s,a)$ denotes the $Q$-value and is used in a small discrete action space. For a deterministic policy, the $Q$ value can be calculated as follows\n\\begin{align}\n Q (s_t,a_t) = \\mathbb{E}_{\\tau \\sim \\pi}[r(s_t,a_t) + \\gamma Q(s'_t, a'_t)].\n\\end{align}\nDeep Q learning (DQN)~ uses deep learning to approximate a non-liner Q function parameterized by $\\theta_q$: $Q_{\\theta_q} (s,a)$. DQN designs a network $Q_{\\theta_q}$ that is asynchronously updated by minimizing the MSE:\n\\begin{align}\n \\mathcal{L}(\\theta_q) = \\mathbb{E}_{\\tau \\sim \\pi}\\Big[Q_{\\theta_q}(s_t,a_t)-(r(s_t,a_t) + \\gamma Q_{\\theta_q}(s'_{t},a'_{t}))\\Big]^2 \\label{dqnloss}\n\\end{align}\nwhere $\\tau$ is the sampled trajectory containing $(s,a,s',r(s,a))$. In particular, $s'_t$ and $a'_t$ come from the behavior policy $\\pi_b$ while $s,a$ comes from the target policy $\\pi$.\nIt is worth mentioning that the value function $V_\\pi(s)$ represents the expected return. $V_\\pi(s)$ is used to evaluate the goodness of the state while $Q_\\pi(s_t,a_t)$ is used to evaluate the action. $V_\\pi(s)$ can be defined as\n\\begin{align}\n V_\\pi(s) = \\mathbb{E}_{\\tau\\sim \\pi}\\bigg[\\sum_{t=0}^T\\gamma^t r(s,a)|s_0=s\\bigg].\n\\end{align}\n$V_\\pi(\\cdot)$ and $Q_\\pi(\\cdot)$ have the following relationship:\n\\begin{align}\n V_\\pi(s) = \\mathbb{E}_{a\\sim\\pi}[Q_\\pi(s,a)].\n\\end{align}\nThe value function is updated using the following rule with the Temporal Difference (TD) method,\n\\begin{align}\n V_\\pi(s_t) \\leftarrow V_\\pi(s_t) + \\alpha[\\underbrace{r(s'_t,a'_t) + \\gamma V_\\pi(s'_t) - V_\\pi(s_t)}_{\\text{TD-error}}]\n\\end{align}\nwhere $\\alpha$ is a constant.\n\\textbf{Policy gradient}~ is an on-policy policy-based method which can handle high-dimensional or continuous actions which cannot be easily handled by Q-learning. Policy gradient\naims to find the parameter $\\theta$ of $\\pi_{\\theta}$ to maximize the accumulated reward. To this end, it maximizes the expected return from the start state:\n\\begin{align}\n J(\\pi_\\theta) = \\mathbb{E}_{\\tau \\sim \\pi_{\\theta}}[r(\\tau)] = \\int\\pi_{\\theta}(\\tau) r(\\tau)d\\tau\n\\end{align}\nwhere $\\pi_{\\theta}(\\tau)$ is the probability of the occurrence of $\\tau$. Policy gradient learns the parameter $\\theta$ by the gradient $\\nabla_\\theta J(\\pi_\\theta)$ as defined below:\n\\begin{align}\n \\nabla_\\theta J(\\pi_\\theta) = \\int\\pi_{\\theta}(\\tau) r(\\tau)d\\tau &= \\int\\pi_{\\theta}(\\tau) \\nabla_{\\theta}\\log \\pi_{\\theta}(\\tau)r(\\tau)d\\tau \\notag \\\\\n & =\\mathbb{E}_{\\tau \\sim d_{\\pi_\\theta}}[\\sum_{t=1}^Tr(s_t,a_t)\\sum_{t=1}^T\\nabla_{\\theta} \\log \\pi_{\\theta}(s_t,a_t)].\n\\end{align}\nThe above derivations contain the following substitution,\n\\begin{align}\n \\pi_{\\theta}(\\tau) = p(s_1)\\prod_{t=1}^T \\pi_{\\theta}(s_t,a_t)p(s_{t+1}|s_t,a_t)\n\\end{align}\nwhere $p(\\cdot)$ are independent from the policy parameter $\\theta$, which is omitted during the derivation.\nMonte-Carlo sampling has been used by previous policy gradient algorithm (e.g,. REINFORCE) for $\\tau\\sim d_{\\pi_{\\theta}}$.\n\\textbf{Actor-critic networks} combine the advantages from Q-learning and policy gradient. They can be either on-policy~ or off-policy~. An actor-critic network consists of two components: i) \\textit{an actor}, which optimizes the policy $\\pi_\\theta$ under the guidance of $\\nabla_\\theta J(\\pi_{\\theta})$; and ii) \\textit{a critic}, which evaluates the learned policy $\\pi_\\theta$ by using $Q_{\\theta_q} (s,a)$. The overall gradient is represented as follows:\n\\begin{align}\n \\mathbb{E}_{s \\sim d_{\\pi_\\theta}}[Q_{\\theta_q}(s,a)\\nabla_{\\theta} \\log \\pi_{\\theta}(s,a)].\n\\end{align}\nWhen dealing with off-policy learning, the value function for $\\pi_\\theta (a|s)$ can be further determined by deterministic policy gradient (DPG) as shown below:\n\\begin{align}\n \\mathbb{E}_{s \\sim d_{\\pi_\\theta}}[\\nabla_a Q_{\\theta_q}(s,a)|_{a=\\pi_\\theta(s)}\\nabla_{\\theta} \\pi_{\\theta}(s,a)].\n\\end{align}\nWhile traditional policy gradient calculates the integral for both the state space $\\mathcal{S}$ and the action space $\\mathcal{A}$, DPG only requires computing the integral to the state space $\\mathcal{S}$.\nGiven a state $s\\in\\mathcal{S}$, there will be only one corresponding action $a\\in\\mathcal{A}:\\mu_\\theta(s) = a$ using DPG. Specifically, deep Deterministic Policy Gradients (DDPG) is an algorithm that combines techniques from DQN and DPG. DDPG contains four different neural networks: Q Network $Q$, policy network, target Q network $Q^{\\mathit{tar}}$, and target policy network. It uses the target network for both the Q Network $Q$ and policy network $\\mu$ to ensure stability during training. Assume $\\theta_q, \\theta_{\\pi}, \\theta_{q'}$ and $\\theta_{\\pi'}$ are parameters of the above networks; then DDPG soft-updates the parameters for the target network~:\n\\begin{align}\n \\text{Actor: }\\theta_{\\pi'} \\leftarrow \\alpha \\theta_\\pi + (1-\\alpha)\\theta_{\\pi'} \\text{ Critic: }\\theta_{q'} \\leftarrow \\alpha \\theta_q + (1-\\alpha)\\theta_{q'} \n\\end{align}","id":"e05ae3e8-b24d-431b-b181-250a9bf1fb50","level":"subsection","origin_cites_number":7,"parent_id":"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Background"],["subsection","Preliminaries of Deep Reinforcement Learning"]],"subsections":[],"title":"Preliminaries of Deep Reinforcement Learning"},{"cite_extract_rate":0,"cites":[],"content":"DRL is normally formulated as a Markov Decision Process (MDP).\nGiven a set of users $\\mathcal{U} = \\{u, u_1, u_2, u_3, ...\\}$, a set of items $\\mathcal{I} = \\{i, i_1, i_2, i_3, ...\\}$, the system first recommends item $i$ to user $u$ and then gets feedback $f_i^u$.\nThe system aims to incorporate the feedback to improve future recommendations and needs to determine an optimal policy $\\pi^*$ regarding which item to recommend to the user to achieve positive feedback.\nThe MDP modelling of the problem treats the user as the environment and the system as the agent. The key components of the MDP in DRL-based RS include the following:\n\\begin{itemize}\n \\item State $\\mathcal{S}$: A state $S_t\\in\\mathcal{S}$ is determined by both users' information and the recent $l$ items in which the user was interested before time $t$.\n \\item Action $\\mathcal{A}$: An action $a_t \\in\\mathcal{A}$ represents users' dynamic preference at time $t$ as predicted by the agent. $\\mathcal{A}$ represents the whole set of (potentially millions of) candidate items.\n \\item Transition Probability $\\mathcal{P}$: The transition probability $p(s_{t+1}|s_t,a_t)$ is defined as the probability of state transition from $s_t$ to $s_{t+1}$ when action $a_t$ is executed by the recommendation agent. In a recommender system, the transition probability refers to users' behavior probability. $\\mathcal{P}$ is only used in model-based methods.\n \\item Reward $\\mathcal{R}$: Once the agent chooses a suitable action $a_t$ based on the current state $S_t$ at time $t$, the user will receive the item recommended by the agent. Users' feedback on the recommended item accounts for the reward $r(S_t,a_t)$. The feedback is used to improve the policy $\\pi$ learned by the recommendation agent.\n \\item Discount Factor $\\gamma$: The discount factor $\\gamma \\in [0,1]$ is used to balance between future and immediate rewards---the agent focuses only on the immediate reward when $\\gamma=0$ and takes into account all the (immediate and future) rewards otherwise.\n\\end{itemize}\nThe DRL-based recommendation problem can be defined by using MDP as follows.\n\\textit{Given the historical MDP, i.e., $(\\mathcal{S},\\mathcal{A},\\mathcal{P}, \\mathcal{R},\\gamma)$, the goal is to find a set of recommendation polices ($\\{\\pi\\}: \\mathcal{S}\\to \\mathcal{A}$) that maximizes the cumulative reward during interaction with users.}\n\\begin{dfn*}\nGiven an environment that contains all items $\\mathcal{I}$, when user $u\\in\\mathcal{U}$ interacts with the system, an initial state $s$ is sampled from the environment which contains a list of candidate items and users' historical data. The DRL agent needs to work out a recommendation policy $\\pi$ based on the state $s$ and produces the corresponding recommended item list $a$.\nThe user will provide feedback on the list which is normally represented as click or not click. The DRL agent will then utilize the feedback to improve the recommendation policy and move to the next interaction episode.\n\\end{dfn*}\n\\begin{figure}[ht]\n \\centering\n \\begin{subfigure}[b]{0.68\\textwidth}\n \\centering\n \\includegraphics[width=\\textwidth]{dlrs.pdf}\n \\caption{Deep learning based recommender systems}\n \\end{subfigure}\n \\begin{subfigure}[b]{0.65\\textwidth}\n \\centering\n \\includegraphics[width=\\textwidth]{drlrs.pdf}\n \\caption{Deep reinforcement learning based recommender systems}\n \\end{subfigure}\n \\caption{Difference between deep learning based RS and DRL-based RS. Deep learning based RSs may only update the recommendation policy during the training stage. They often require re-training, which is computationally inefficient, when users' interests change significantly. DRL-based RS will update the recommendation policy time over time as new rewards are received.}\n \\label{fig:catnew}\n\\end{figure}","id":"467d3d60-cfac-40c2-8a3f-6dbe98149287","level":"subsection","origin_cites_number":0,"parent_id":"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Background"],["subsection","DRL meets RS: Problem Formulation"]],"subsections":[],"title":"DRL meets RS: Problem Formulation"},{"cite_extract_rate":0,"cites":[],"content":"DRL-based RS has some unique challenges such as state construction, reward estimation and environment simulation. We categorize the existing work of DRL-based recommendation into model-based and model-free methods (the taxonomy is shown in~\\Cref{fig:tax}).","id":"696bfdf7-6191-445e-82a8-ae4df2378853","level":"section","origin_cites_number":0,"parent_id":"4bb82714-5571-4179-9dac-693fb550f058","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"]],"subsections":["4ebfecfc-b15f-4ce1-abac-7b121665be4a","74b47d11-d3b5-407c-b802-84dc050162e4","c30b91c8-48bc-4d92-ba66-c418e39b9327"],"title":"Deep Reinforcement Learning in Recommender Systems"},{"cite_extract_rate":0.42857142857142805,"cites":[5189,5180,5195],"content":"Model-based methods assume an expected reward or action available for the next step to help the agent update the policy.\n\\begin{table}[ht]\n\\caption{List of publications in model-based DRL-based RS}\n\\begin{tabular}{ccc}\n\\hline\nMethod & Work \\\\\\hline\nValue-based & \\\\\nPolicy-based & \\\\\nHybrid & \n\\\\\\hline\n\\end{tabular}\n\\end{table}","id":"4ebfecfc-b15f-4ce1-abac-7b121665be4a","level":"subsection","origin_cites_number":7,"parent_id":"696bfdf7-6191-445e-82a8-ae4df2378853","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-based Deep Reinforcement Learning based Methods"]],"subsections":["8ceb9826-a858-407a-9b78-7a83d0ea8871"],"title":"Model-based Deep Reinforcement Learning based Methods"},{"cite_extract_rate":0.33333333333333304,"cites":[7217],"content":"IRecGAN~ is a model-based method that adopts generative adversarial training to improve the robustness of policy learning. It can reduce the cost of interaction for RS by using offline data instead of the simulated environment.\nIRecGAN employs a generative adversarial network~ to generate user data based on the offline dataset.\nIt trains a recommendation agent using a policy gradient-based DRL method called REINFORCE. The agent aims to learn a policy based on the following gradient,\n\\begin{align}\n \\mathbb{E}_{\\tau\\sim\\{g,data\\}}\\big[\\sum_{t=0}^T \\sum_{t'=t}^T\\gamma^{t'-t} q_D(\\tau_{0:t}^n)r_t \\nabla_{\\theta_a}(c_t \\in \\pi_{\\theta_a}(s_t)\\big],q_D(\\tau_{0:t}^n)= \\frac{1}{N}\\sum_{n=1}^N D(\\tau_{0:T}^n),\\tau_{0:T}^n\\in MC^{\\mathcal{U}}(N)\n\\end{align}\nwhere the $\\mathit{MC}^{\\mathcal{U}}(N)$ represents the sampled $N$ sequences from the interaction between $\\mathcal{U}$ and the agent using the Monte-Carlo tree search algorithm, $D$ is the discriminator, $T$ is the length of $\\tau$, $g$ represents the offline data, and \\textit{data} represents the generated data. \n propose NRSS for personalized music recommendation. NRSS uses wireless sensing data to learn users' current preferences. NRSS considers three different types of feedback: score, option, and wireless sensing data. Because multiple factors are considered as the reward, NRSS designs a reward model which consists of users' preference reward $r_p$ and a novel transition reward $r_{\\textit{trans}}$ which are parameterized by $\\theta_{r_p}$ and $\\theta_{r_{\\textit{trans}}}$. The goal for NRSS is to find the optimal parameters $\\theta_{r_p}$ and $\\theta_{r_{\\textit{trans}}}$ by using the Monte-Carlo tree search thus improving recommendation performance.\nHowever, wireless sensing feedback lacks generalization ability as it is only available for certain tasks or scenarios, making it hard to determine dynamic user interest.","id":"8ceb9826-a858-407a-9b78-7a83d0ea8871","level":"paragraph","origin_cites_number":3,"parent_id":"4ebfecfc-b15f-4ce1-abac-7b121665be4a","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-based Deep Reinforcement Learning based Methods"],["paragraph","Policy-based methods"]],"subsections":["bb6b42ee-6de2-432f-ab2a-fc4b096f3199","68abbde2-1fcc-48e8-a5f8-eccafeb14c6f","326848cf-b30c-48e3-8c0b-494e313de16b"],"title":"Policy-based methods"},{"cite_extract_rate":0.5,"cites":[5195,5180],"content":"Prior to Q-learning, value iteration is a more traditional value-based reinforcement learning algorithm that focuses on the iteration of the value function. Gradient Value Iteration (GVI)~ is proposed to improve the traditional value iteration algorithm by utilizing the transition probability and a multi-agent setting to predict chronological author collaborations. It introduces a new parameter named `status' to reflect the amount of knowledge that the agent needs to learn from this state. The policy is updated only when the distance between the new status and the old status is lower than a pre-defined threshold. However, value iteration requires the transition probability, which is hard to obtain in most cases. Hence, Q-learning and its variants are widely used in DRL-based RS. Cascading DQN (CDQN) with a generative user model~ is proposed to deal with the environment with unknown reward and environment dynamics. The generative user model adopts GANs to generate a user model based on an offline dataset. Different from previous work, it will generate the reward function for each user to explain the users' behavior. The user model can be written as,\n\\begin{align}\n \\argmax_{\\phi\\in\\triangle^{k-1}} \\mathbb{E}_{\\phi}[r(s_t,a_t)]-R(\\phi)/\\eta\n\\end{align}\nwhere $\\triangle^{k-1}$ is the probability simplex, $R(\\phi)$ is the regularization term for exploration and $\\eta$ is a constant.\nPseudo Dyna-Q (PDQ)~ points out that Monte-Carlo tree search may lead to an extremely large action space and an unbounded importance weight of training samples. Hence, a world model is proposed to reduce the instability of convergence and high computation cost for interacting with users by imitating the offline dataset. With the world model, the agent will interact with the learned world model instead of the environment to improve the sample efficiency and convergence stability. The world model learning process introduced in PDQ can be described as finding the parameter $\\theta_M$,\n\\begin{align}\n \\argmin_{\\theta_M}\\mathbb{E}_{\\xi\\in P_{\\xi}^\\pi}[\\sum_{t}^{T-1}\\gamma^t\\prod_{j=0}^t\\frac{\\pi(s_j,a_j)}{\\pi_b(s_j,a_j)}\\Delta_t(\\theta_M)]\n\\end{align}\nwhere $\\xi$ is generated by the logged policy $\\pi_b$, $\\prod_{j=0}^t\\frac{\\pi(s_j,a_j)}{\\pi_b(s_j,a_j)}$ is the ratio used for importance sampling and $\\Delta$ is the difference between the reward in the world model and real reward. \nFurthermore, GoalRec~ designs a disentangled universal value function to be integrated with the world model to help the agent deal with different recommendation tasks. The universal value function is defined as\n\\begin{align}\n V_{\\pi} (s)= \\mathbb{E}[\\sum_{t=0}^\\infty r(s_t,a_t)\\prod_{k=0}^t\\gamma s_k|s_0=s].\n\\end{align}\nMoreover, GoalRec introduces a new variable goal $g\\in G$ used to represent users' future trajectory and measurement $m\\in M$. $m$ is an indicator that reflects users' response to the given future trajectory based on historical behaviors. Based on that, the optimal action will be selected based on\n\\begin{align}\n a^* = \\max_a U(M(s,a),g)\n\\end{align}\nwith a customized liner function $U(\\cdot)$.\n\\begin{figure}\n \\centering\n \\includegraphics[width=\\linewidth]{GoalRec_wb.pdf}\n \\caption{Left is the general structure of model-free methods. Right is the structure for GoalRec which is a model-based method. A sample trajectory is used to demonstrate the difference between them~. }\n \\label{fig:worldmodel}\n\\end{figure}","id":"bb6b42ee-6de2-432f-ab2a-fc4b096f3199","level":"paragraph","origin_cites_number":4,"parent_id":"8ceb9826-a858-407a-9b78-7a83d0ea8871","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-based Deep Reinforcement Learning based Methods"],["paragraph","Policy-based methods"],["paragraph","Value-based methods"]],"subsections":[],"title":"Value-based methods"},{"cite_extract_rate":1,"cites":[5189],"content":"Hybrid method can be recognized as a midpoint between value-based and policy gradient-based methods. DeepChain~ uses the multi-agent setting to relieve the sub-optimality problem. The sub-optimality problem is caused by the \\textit{one for all} setting that optimizes one policy for all users. Hence, DeepChain designs a multi-agent setting that adopts several agents to learn consecutive scenarios and jointly optimizes multiple recommendation policies. The main training algorithm used is DDPG. To this end, users' actions can be formulated in a model-based form as follows:\n\\begin{align}\n \\sum_{m,d}[p_m^s (s_t,a_t)\\gamma Q_{\\theta}(s'_{t},\\pi_m(s'_{t})) + p_m^c(s_t,a_t)(r_t+\\gamma Q_{\\theta}(s'_{t},\\pi_d(s'_{t}))) + p_m^l (s_t,a_t)r_t]1_m\n\\end{align}\nwhere $m$ represents the number of actor networks, $c,l,s$ represent the three different scenarios, $1_m$ is used to control the activation of two actors and $(m,d)\\in\\{(1,2),(2,1)\\}$.","id":"68abbde2-1fcc-48e8-a5f8-eccafeb14c6f","level":"paragraph","origin_cites_number":1,"parent_id":"8ceb9826-a858-407a-9b78-7a83d0ea8871","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-based Deep Reinforcement Learning based Methods"],["paragraph","Policy-based methods"],["paragraph","Hybrid methods"]],"subsections":[],"title":"Hybrid methods"},{"cite_extract_rate":0.5,"cites":[5189,5180,5195],"content":"Model-based methods aim to learn a model or representation to represent the whole environment so that the agent can plan ahead and receive better sample efficiency.\nThe drawback of such a method is that the ground-truth representation of the environment is unavailable in recommendation scenarios as it dynamically changes, leading to a biased representation. Moreover, model-based methods use the transition probability function $\\mathcal{P}$ to estimate the optimal policy. As mentioned, the transition probability function is normally equivalent to users' behavior probability which is hard to determine in a recommender system. Hence, existing works~ approximate $\\mathcal{P}$ using a neural network or embedding it into the world model. design a probability network to estimate $\\mathcal{P}$ while uses a GAN to generate user behavior where $\\mathcal{P}$ is embedded in the latent space. Different from them,~ relies on the world model to predict users' next behavior and feed it into the policy learning process.\nThe challenges of model-based DRL are not widely used in RS and can be summarized into the following facets:\n\\begin{itemize}\n\\item $\\mathcal{P}$ is hard to determine in real-world recommender systems. \n\\item If approximation is used to estimate $\\mathcal{P}$, the overall model complexity will substantially increase as it requires approximating two different functions $\\mathcal{P}$ and the recommendation policy $\\pi$ by using a large amount of user behavior data. \n\\item World model-based methods require periodic re-training to ensure the model can reflect user interests in time which increases the computation cost. \n\\end{itemize}\n\\begin{figure}\n \\centering\n \\begin{subfigure}{0.48\\linewidth}\n \\includegraphics[width=\\linewidth]{DQN_wb.pdf}\n \\caption{DQN}\n \\end{subfigure}\n \\begin{subfigure}{0.48\\linewidth}\n \\includegraphics[width=\\linewidth]{DDPG_wb.pdf}\n \\caption{DDPG}\n \\end{subfigure}\n \\caption{The typical structure of DQN and DDPG}\n \\label{fig:ddpg_fig}\n\\end{figure}","id":"326848cf-b30c-48e3-8c0b-494e313de16b","level":"paragraph","origin_cites_number":6,"parent_id":"8ceb9826-a858-407a-9b78-7a83d0ea8871","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-based Deep Reinforcement Learning based Methods"],["paragraph","Policy-based methods"],["paragraph","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":0.43181818181818105,"cites":[5190,3604,8884,5165,6153,3616,1389,5184,5191,5177,5169,5199,6154,5181,6152,5172,5194,5170,5185],"content":"Compared with model-based methods, model-free methods are relatively well-studied. Different from model-based methods, $\\mathcal{P}$ is unknown and not required in model-free methods. Model-based methods enable the agent to learn from previous experiences. In this subsection, we categorize model-free based DRL in RS into three parts: value-based, policy-based and hybrid methods. \n\\begin{table}[ht]\n\\caption{List of reviewed publications in Model-free DRL-based RS}\n\\begin{tabular}{ccc}\n\\hline\nTasks & Note & Work \\\\\\hline\n\\multirow{4}{*}{Value-based} & Vanilla DQN and its extensions & \\\\ \\\n& DQN with state/action space optimization & \\\\\n& DQN with graph/image input & \\\\\n& DQN for joint learning & \\\\\\hline\n\\multirow{3}{*}{Policy-based} & Vanilla REINFORCE & \\\\\n& REINFORCE uses graph structure/input & \\\\\n& Non-REINFORCE based & \\\\ \\hline\n\\multirow{2}{*}{Hybrid} & Vanilla DDPG & \\\\\n& with Knowledge Graph & \\\\\n\\hline\n\\end{tabular}\n\\end{table}","id":"74b47d11-d3b5-407c-b802-84dc050162e4","level":"subsection","origin_cites_number":44,"parent_id":"696bfdf7-6191-445e-82a8-ae4df2378853","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-free deep reinforcement learning based methods"]],"subsections":["6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe"],"title":"Model-free deep reinforcement learning based methods"},{"cite_extract_rate":0.41176470588235203,"cites":[5181,5191,5170,6155,5185,5199,1408],"content":"As mentioned, Deep Q-learning and its variants are typical value-based DRL methods widely used in DRL-based RS. DRN~ is the first work utilizing Deep Q-Networks (DQN) in RS. It adopts Double DQN (DDQN)~ to build a user profile and designs an activeness score to reflect how frequently a user returns after one recommendation plus users' action (click or not) as the reward. DRN provides a new approach to integrating DRL into RS when dealing with a dynamic environment. The key objective function can be found as follows,\n\\begin{align}\n \\mathbb{E}[r_{t+1} + \\gamma Q_{\\theta_t'}(s_{t+1}, \\argmax_{a'} Q_{\\theta_t}(s_t,a'))]\n\\end{align}\nwhere $a'$ is the action that gives the maximum future reward according to $\\theta_t$, $\\theta_t$ and $\\theta_t'$ are different parameters for two different DQNs. \n points out that negative feedback will also affect recommendation performance which DRN does not consider. Moreover, positive feedback is sparse due to the large number of candidate items in RS. Only using positive feedback would lead to convergence problems. Hence, DEERS is proposed to consider both positive and negative feedback simultaneously by using DQN. Gated Recurrent Units (GRU) are employed to capture users' preferences for both a positive state $s^+$ and negative state $s^-$ and the final objective function can be computed as:\n\\begin{align}\n \\mathbb{E}[r_{t+1} + \\gamma \\max_{a_{t+1}}Q_{\\theta_q}(s^+_{t+1},s^-_{t+1},a_{t+1})|s^+_t,s^-_t,a_t].\n\\end{align}\n introduces attention mechanisms into the DQN to leverage social influence among users. To be specific, a social impact representation $U_v$ is introduced into the state representation. Matrix factorization is adopted to determine \nsimilarity among\nusers and hence present the social influence. Social attention is introduced to distill the final state representation.\nIn addition, a few studies focus on user profiling to improve recommendation performance~. claims that user feedback contains useful information in the previous feedback even when the user does not like the recommended items. Some existing studies focus on final feedback which ignore the importance from earlier steps to later ones. Hence, user-specific DQN (UQDN) is proposed to consider multi-step feedback from users. It employs Matrix Factorization to generate user-specific latent state spaces. The newly defined objective function with the user-specific latent state space can be represented as\n\\begin{align}\n \\mathbb{E}[r_{t+1} + \\gamma \\max_{a_{t+1}}\\overline{Q}_{\\theta_q}(s_{t+1},a_{t+1}) + \\overline{\\textbf{b}}_u - Q_{\\theta_q} (s_{t+1},a_{t+1})]\n\\end{align}\nwhere $\\overline{\\textbf{b}}_u$ is the learned user latent representation.\n also points out that most studies do not consider users' long-term engagement in the state representation as they focus on the immediate reward. FeedRec is proposed that combines both instant feedback and delayed feedback into the model to represent the long-term reward and optimize the long-term engagement by using DQN. To be specific, time-LSTM is employed to track users' hierarchical behavior over time to represent the delayed feedback which contains three different operations: $h_{\\mathit{skip}},h_{\\mathit{choose}},h_{\\mathit{order}}$. The state space is the concatenation of those operations and users' latent representation. Differently, focuses on the user privacy issue in recommender systems. Deep user profile perturbation (DUPP) is proposed to add perturbation into the user profile by using DQN during the recommendation process. Specifically, DUPP adds a perturbation vector into users' clicked items as well as the state space, which contains users' previous behavior.\nDistinct from previous studies which focus on optimizing user profiles or state spaces, some studies aim to optimize the action space formed by interactions with items. In the situation of basket recommendation, the user is suggested multiple items as a bundle, which is called a recommendation slate. It leads to combinatorially large action spaces making it intractable for DQN based recommendation models.\nSlateQ~ is proposed to decompose slate Q-value to estimate a long-term value for individual items, and it is represented as,\n\\begin{align}\n Q_{\\theta_q}(s_t,a_t) = \\sum_{i \\in a_t}p(i|s_t,a_t)\\overline{Q}_{\\theta_q}(s_t,i)\n\\end{align}\nwhere $\\overline{Q}_{\\theta}(s,i)$ is the decomposed Q-value for item $i$. The decomposed Q-value will be updated by the following rule which is similar to traditional DQN,\n\\begin{align}\n \\overline{Q}_{\\theta_q}(s_t,i) \\leftarrow \\alpha \\bigg(r_t +\\gamma \\sum_{j\\in a_{t+1}}p(j|s_{t+1},a_{t+1}) \\overline{Q}_{\\theta_q}(s_{t+1},j)\\bigg) + (1-\\alpha)\\overline{Q}_{\\theta_q}(s_t,i).\n\\end{align}\nDifferent from other mode-free methods, Slate-Q assumes that the transition probability $p(i|s_t,a_t)$ is known.\nVanilla DQN methods may not have sufficient knowledge to handle complex data such as images and graphs. firstly models users' click behavior as an embedding matrix in the latent space to include the skip behaviors of sequence patterns for sequential recommendation. Based on that, propose DRCGR, which adopts CNN and GAN into DQN to help the agent to better understand high-dimensional data, e.g., a matrix. Two different convolution kernels are used to capture users' positive feedback. In the meantime, DRCGR uses GANs to learn a negative feedback representation to improve robustness. \nAnother typical data format is the graph, which is widely used in RS, including knowledge graphs.\n propose GCQN which adopts Graph Convolutional Networks (GCN)~ into the DQN which constructs the state and action space as a graph-aware representation. Differently, GCQN introduces the attention aggregator: $\\sum_{w\\in \\mathcal{N}(i)}\\alpha_{iu}e_u$ which demonstrates better performance than the mean-aggregator and pooling-aggregator. For item $i$, the graph-aware representation can be represented as,\n\\begin{align}\n \\sigma\\bigg(W_{fc}[e_i\\oplus\\sum_{w\\in \\mathcal{N}(i)}\\alpha_{iu}e_u + b_{fc}]\\bigg)\n\\end{align}\nwhere $W_{fc},b_{fc}$ are the parameters for the fully-connected layer, $e_u$ is the embedding for user $u$ and $\\mathcal{N}(i)$ is the set of one-hot neighbours of item $i$ in graph $G(i)$. propose KGQR uses a similar strategy to transform the information into a knowledge graph which is fed into the GCN to generate the state representation. Notably, KGQR presents a different state representation generation method. For given node $i$, the neighbourhood representation with a $k$-hop neighborhood aggregator can be represented as,\n\\begin{align}\n e_i^k = \\sigma\\bigg(W_{k}\\frac{1}{|\\mathcal{N}(i)|}\\sum_{t\\in\\mathcal{N}(i)} e_{t}^{k-1} + B_k e_i^{k-1}\\bigg)\n\\end{align}\nwhere $\\mathcal{N}(i)$ is the set of neighboring nodes, $W_k, B_k$ are the parameter of the aggregator. Those neighbourhood representations will be fed into a GRU and the state representation will be generated. Another application domain for using graph data is job recommendation which requires considering multiple factors jointly such as salary, job description, job location etc. SRDQN~ constructs a probability graph to represent a candidate's skill set and employs a multiple-task DQN structure to process these different factors concurrently. \nThere are some studies targeting recommendation and advertising simultaneously in e-commerce environments~. mentions when deploying RS into real-world platforms such as e-commerce scenarios, the expectation is to improve the profit of the system.\nA new metric, Gross Merchandise Volume (GMV), is proposed to measure the profitability of the RS to provide a new view about evaluating RS in advertising. Different from GMV, separates recommendation and advertising as two different tasks and proposes the Rec/Ads\nMixed display (RAM) framework. RAM designs two agents: a recommendation agent and an advertising agent, where each agent employs a CDQN to conduct the corresponding task. find that advertising and recommendation may harm each other and formulate a rec/ads trade-off. Their proposed solution, DEARS, contains two RNNs. Two RNNs are employed to capture user preferences toward recommendations and ads separately. Based on that, DQN is employed to take those two outputs as the input to construct the state and output the advertising.","id":"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe","level":"paragraph","origin_cites_number":17,"parent_id":"74b47d11-d3b5-407c-b802-84dc050162e4","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-free deep reinforcement learning based methods"],["paragraph","Value based methods"]],"subsections":["01225554-85b2-42e4-a661-722c3b5972d7","cb915b6e-26a1-44b9-be2e-47ffa839e199","47d0fb4b-4dca-4006-ba90-90530959ef38"],"title":"Value based methods"},{"cite_extract_rate":0.4,"cites":[5190,5165,6153,5184,5177,1911,6152,6156],"content":"Policy-based DRL can be divided into two parts which are Constrained Policy Optimization (CPO) ~ and policy gradient.\n uses CPO to identify the contradiction between text feedback and historical preferences. It provides a solution for using DRL in the situation where users' feedback is entirely different from previous feedback in RS. \nPolicy gradient-based methods are the other stream in policy-based DRL methods for RS. These methods aims to optimize the policy $\\pi$ directly instead of estimating the Q-value like DQN. A well-known and widely used policy gradient method in RS is REINFORCE which uses the following rule for policy $\\pi_{\\theta_\\pi}$,\n\\begin{align}\n \\theta \\leftarrow \\theta + \\alpha \\mathbb{E}_{\\tau\\sim d_{\\pi_{\\theta_\\pi}}}\\bigg[\\sum_{t=1}^Tr(s_t^i,a_t^i)\\sum_{t=1}^T\\nabla_{\\theta_\\pi} \\log \\pi_{\\theta_\\pi}(s_t^i,a_t^i)\\bigg]\n\\end{align}\nwhere $i$ is sampled trajectories from $\\pi_{\\theta}(a_t|s_t)$. propose Policy Gradient for Contextual Recommendation (PGCR), which adopts REINFORCE and considers contextual information. PGCR assumes that the policy follows the multinoulli distribution, in which case the transition probability can be estimated easily through sampling from previously seen context.\n incorporate CNNs and attention mechanisms in REINFORCE for explainable recommendation. Specifically, this work designs a coupled agent structure where one agent generates the explanation and the other makes recommendations based on the generated explanation.\n increases the scalability of REINFORCE to ensure it can deal with the extremely large action space under recommendation scenarios. To be specific, it introduces a policy correction gradient estimator into REINFORCE to reduce the variance of each gradient by doing importance sampling. The new update rule\nbecomes\n\\begin{align}\n \\theta_\\pi \\leftarrow \\theta_\\pi + \\alpha \\sum_{\\tau\\sim\\beta}\\bigg[\\sum_{t=1}^T\\frac{\\pi_{\\theta_\\pi}(s_t,a_t)}{\\pi_\\beta(s_t,a_t)}r(s_t^i,a_t^i)\\sum_{t=1}^T\\nabla_{\\theta_\\pi} \\log \\pi_{\\theta_\\pi}(s_t^i,a_t^i)\\bigg]\n\\end{align}\nwhere $\\pi_\\beta$ is the behavior policy trained by state-action pairs without the long-term reward and $\\pi_{\\theta}$ is trained based on the long-term reward only. It is worth mentioning that the vanilla REINFORCE algorithm is on-policy, and importance sampling will make REINFORCE behave like an off-policy method with the following gradient format,\n\\begin{align}\n \\mathbb{E}_{\\tau \\sim d_{\\pi_\\theta}}\\bigg[\\prod_{t'=1}^t\\frac{\\pi_{\\theta}(s_t,a_t)}{\\pi_{\\theta_s}(s_{t'},a_{t'})}\\sum_{t'=t}^Tr(s_t,a_t)\\sum_{t=1}^T\\nabla_{\\theta} \\log \\pi_{\\theta}(s_t,a_t)\\bigg]\n\\end{align}\nwhere $\\pi_{\\theta_s}$ is the sample policy parameter.\n also finds that the REINFORCE method suffers from a high variance gradient problem and Pairwise Policy Gradient (PPG) is proposed. Different from policy correction, PPG uses Monte Carlo sampling to sample two different actions $a,b$ and compare the gradient to update $\\theta$,\n\\begin{align}\n \\mathbb{E}_{\\tau\\sim d_{\\pi_{\\theta_\\pi}}}\\bigg(\\sum_a\\sum_b(r(s,a)-r(s,b))\\sum_{t=1}^T(\\nabla_{\\theta_\\pi}\\log_{\\pi_{\\theta_\\pi}}(s_t,a_t)-\\nabla_{\\theta_\\pi}\\log_{\\pi_{\\theta_\\pi}}(s_t,b_t))\\bigg).\n\\end{align}\n extends the policy correction gradient estimator into a two-stage setting which are $p(s_t,a^p)$ and $q(s_t,a|a^p)$ and the policy can be written as\n\\begin{align}\n \\sum_{a^p}p(s_t,a^p)q(s_t,a|a^p).\n\\end{align}\nIn addition, weight capping and self-normalized importance sampling are used to further reduce the variance.\nMoreover, a large state space and action space will cause sample inefficiency problems as REINFORCE relies on the current sampled trajectories $\\tau$. finds that the auxiliary loss can help improve the sample efficiency\n~. Specifically, a linear projection is applied to the state $s_t$, the output is combined with action $a_t$ to calculate the auxiliary loss and appended into the final overall objective function for optimization.\nAnother prototype of vanilla policy gradient in DRL-based RS is the policy network. designs a policy network to extract features and represent the relevant feedback that can help the agent make a decision. Similar to DQN, this work uses a neural network to approximate the Q-value and the policy directly without theoretical analysis.\n extend the policy network by introducing spatio-temporal feature fusion to help the agent understand complex features. Specifically, it considers both the current number and the future number of vacant taxis on the route to recommend routes for taxis. introduces multi-modal data as new features to conduct vision-language recommendation by using historical data to train REINFORCE. ResNet and attention are used to encode vision and text information, respectively. Moreover, two rewards are introduced with a customized ratio $\\lambda$ to balance vision and text information. \nKnowledge Graphs (KG) are widely used in RS to enrich side information, provide explainability and improve recommendation performance. Similar to DQN, vanilla REINFORCE cannot properly handle graph-like data. propose a method named Knowledge-guided Reinforcement Learning (KERL), which integrates knowledge graphs into the REINFORCE algorithm. To be specific, KERL adopts TransE~ to transfer the knowledge graph into a graph embedding and utilizes a multilayer perceptron (MLP) to predict future knowledge of user preferences. The state representation can be written as\n\\begin{align}\n h_t \\oplus \\mathit{TransE}(\\mathcal{G}) \\oplus \\mathit{MLP}(\\mathit{TransE}(\\mathcal{G}))\n\\end{align}\nwhere $h_t$ is the hidden representation from the GRU for sequential behavior and $\\mathcal{G}$ is the knowledge graph.\nDifferent from KERL, propose Policy-Guided Path Reasoning (PGPR), which formulates the whole environment as a knowledge graph. The agent is trained to find the policy to find good items conditioned on the starting user in the KG by using REINFORCE. PGPR uses the tuple $(u,e_t,h_t)$ to represent the state instead of the graph embedding where $e_t$ is the entity the agent has reached at $t$ for user $u$ and $h_t$ is the previous action before $t$.\nThe action in PGPR is defined as the prediction of all outgoing edges for $e_t$ based on $h_t$. \n propose a knowledge graph policy network (KGPolicy) which puts the KG into the policy network and adopts REINFORCE to optimize it. In addition, KGPolicy uses negative sampling instead of stochastic sampling to overcome the false negative issue---sampled items behave differently during training and inference. Similar to GCQN, attention is also employed to establish the representation for its neighbors. \nDue to the on-policy nature of REINFORCE, it is difficult to apply it to large-scale RS as the convergence speed will be a key issue. To relieve this, propose TPGR, which designs a tree-structured policy gradient method to handle the large discrete action space hierarchically. TPGR uses balanced hierarchical clustering to construct a clustering tree. Specifically, it splits a large-scale data into several levels and maintains multiple policy networks for each level to conduct the recommendation. The results are integrated at the final stage.\nAs mentioned, policy gradient can be further extended to deterministic policy gradient (DPG) . propose Deterministic Policy Gradient with Full Backup\nEstimation (DPG-FBE) to complete a sub-task of recommendation. DPG-FBE considers a search session MDP (SSMDP) that contains a limited number of samples, where the stochastic policy gradient method like REINFORCE cannot work well.","id":"01225554-85b2-42e4-a661-722c3b5972d7","level":"paragraph","origin_cites_number":20,"parent_id":"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-free deep reinforcement learning based methods"],["paragraph","Value based methods"],["paragraph","Policy-based methods"]],"subsections":[],"title":"Policy-based methods"},{"cite_extract_rate":0.6153846153846151,"cites":[5172,6154,3604,8884,5194,5170,3616,1389],"content":"The most common model-free hybrid method used would be the actor-critic algorithm where the critic network uses the DQN and the actor uses the policy gradient. The common algorithm used to train actor-critic is DDPG with the following objective function,\n\\begin{align}\n \\mathbb{E}[r_t+\\gamma Q_{\\theta_q'}(s_{t+1},\\mu_{\\theta_\\pi'}(s_{t+1})) - Q_{\\theta_q}(s_t,a_t)]\n\\end{align}\nwhere $\\theta_q,\\theta_q'$ is the parameter for Q-learning at time $t,t+1$ while $\\theta_\\pi'$ is the parameter for deterministic policy gradient at time $t+1$.\n propose LIRD, which uses the vanilla actor-critic framework to conduct list-wise recommendations. In order to demonstrate the effectiveness of LIRD, a pre-trained user simulator is used to evaluate the effectiveness of LIRD where the transition probability is approximated using the cosine similarity for a given state-action pair $s_t,a_t$.\n further extend LIRD into page-wise recommendation and proposed DeepPage. Similar to other previous work, GRU is employed to process the sequential pattern. Moreover, similar to DRCGR, DeepPage formulates the state as a page, then CNNs are employed to capture features and fed to the critic network. The final state representation is the concatenation of the sequential pattern and the page features. Additionally, there are a few studies focusing on different scenarios such as top-aware recommendation~, treatment recommendation~, allocating impressions~ etc. introduces a supervised learning module (SLC) as the indicator to identify the difference between the current policy and historical preferences. SLC will conduct the ranking process to ensure the recommendation policy will not be affected by the positional bias -- the item appearing on top receives more clicks. Similarly, also integrates the supervised learning paradigm into DRL but in a different way. An expert action $\\hat{a}_t$ is provided when the critic evaluates the policy and the update rule is slightly different than normal DQN,\n\\begin{align}\n \\theta_q \\leftarrow \\theta_q + \\alpha \\sum_{t}[ Q_{\\theta_q}(s_t,\\hat{a}_t)-r_t-\\gamma Q_{\\theta_{q'}}(s_t,\\mu_{\\theta_{\\pi'}}(s_t))]\\nabla_{\\theta_q}Q_{\\theta_q}(s_t,a_t).\n\\end{align}\nHowever, such a method is not universal as the acquisition of expert action is difficult and depends on the application domain.\nSimilar to policy gradient and DQN, Knowledge Graphs (KG) are also used in actor-critic-based methods. propose KGRL to incorporate the substantial information of knowledge graphs to help the critic to better evaluate the generated policy. A knowledge graph is embedded into the critic network. Different from previous studies which use the KG as the environment or state representation, KGRL uses KG as a component in the critic, which can guide the actor to find a better recommendation policy by measuring the proximity from the optimal path. Specifically, a graph convolutional network is used to weight the graph and Dijkstra's algorithm is employed to find the optimal path for finally identifying the corresponding Q-value. claim that human's demonstration could improve path searching and propose ADAC. ADAC also searches for the optimal path in the KG but further adopts adversarial imitation learning and uses expert paths to facilitate the search process. propose MA-RDPG, which extends the standard actor-critic algorithm to deal with multiple scenarios by utilizing a multi-actor reinforcement learning setting. Specifically, two different actor-networks are initialized while only one critic network will make the final decision. Those two actor networks can communicate with each other to share information and approximate the global state. find that there are multiple factors can affect the selection of electric charging station. Hence, it uses a similar idea to recommend the electric vehicle charging station by considering current supply, future supply, and future demand. figure out that the communication mechanism in MA-RDPG will harm actors as they are dealing with independent modules, and there is no intersection. Hence, extend MA-RDPG into multi-agent settings which contain multiple pairs of actors and critics and remove the communication mechanism to ensure independence. \nDifferent from , use `soft' actor-critic (SAC)~, which introduces a maximum entropy term $\\mathcal{H}(\\pi(s_t,\\phi_t))$ to actor-critic to improve exploration and stability with the stochastic policy $\\pi(s_t,\\phi_t)$. \nSimilar to the multi-agent idea, use a hierarchical setting to help the agent learn multiple goals by setting multiple actors and critics. In comparison, hierarchical RL uses multiple actor-critic networks for the same task. It splits a recommendation task into two sub-tasks: discovering long-term behavior and capturing short-term behavior. The final recommendation policy is the combination of the optimal policies for the two sub-tasks. Similarly, use the hierarchical setting for integrated recommendation by using different sourced data. The objective is to work out the sub-polices for each source hierarchically and form the final recommendation policy afterward.","id":"cb915b6e-26a1-44b9-be2e-47ffa839e199","level":"paragraph","origin_cites_number":13,"parent_id":"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-free deep reinforcement learning based methods"],["paragraph","Value based methods"],["paragraph","Hybrid methods"]],"subsections":[],"title":"Hybrid methods"},{"cite_extract_rate":0.5,"cites":[6157],"content":"In RS, model-free methods are generally more flexible than model-based methods as they do not require knowing the transition probability. We summarize the advantages and disadvantages of the three kinds of methods described under the model-free category. DQN is the first DRL method used in RS, which is suitable for small discrete action spaces. The problems with DQN in RS are: \n\\begin{itemize}\n \\item RS normally contains large and high-dimensional action spaces.\n \\item The reward function is hard to determine which will lead to inaccurate value function approximation.\n\\end{itemize} \nSpecifically, the high dimensional action space in context of recommender systems is recognized as a major drawback of DQN~. The reason lies in the large number of the candidate items. Hence, DQN, as one of the most popular schemes, is not the best choice for RS in many situations. Moreover, some unique factors need to be considered when designing the reward function for RS such as social inference. It introduces extra parameters to the Q-network and hinders the convergence.\nPolicy gradient does not require the reward function to estimate the value function. Instead, it estimates the policy directly. However, policy gradient is designed for continuous action spaces. More importantly, it will introduce high variance in the gradient. Actor-critic algorithms combine the advantages of DQN and policy gradient. Nonetheless, actor-critic will map the large discrete action space into a small continuous action space to ensure it is differentiable, which may cause potential information loss. Actor-critic uses DDPG and thus inherits disadvantages from DQN and DPG, including difficulty in determining the reward function and poor exploration ability.","id":"47d0fb4b-4dca-4006-ba90-90530959ef38","level":"paragraph","origin_cites_number":2,"parent_id":"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Model-free deep reinforcement learning based methods"],["paragraph","Value based methods"],["paragraph","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":0.5,"cites":[6158,5192,5201,8984,5204,3610],"content":"\\label{sec:component}\nThere are a few studies that use DRL in RS for goals other than improving recommendation performance or proposing new application domains. We split the literature based on the following components: environment, state representation, and reward function. Exisitng studies usually focus on optimizing one single component in the DRL setting (as illustrated in~\\Cref{fig:overview}).\n\\begin{table}[h]\n \\centering\n \\caption{List of publications reviewed in this section}\n \\begin{tabular}{c|c}\n \\hline\n Component & Work \\\\\\hline\n Environment & \\\\\n State & \\\\\n Reward & \\\\\n \\hline\n \\end{tabular}\n \\label{tab:component}\n\\end{table}\n\\\\","id":"c30b91c8-48bc-4d92-ba66-c418e39b9327","level":"subsection","origin_cites_number":12,"parent_id":"696bfdf7-6191-445e-82a8-ae4df2378853","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Component Optimization in Deep Reinforcement Learning based RS"]],"subsections":["3de2362c-6535-43b9-860b-ebc2cbce5821","45229725-ce80-46d4-8644-8011bc19bde5","0c8d5ce0-24e2-4a9b-aa04-14e44c18baf4"],"title":"Component Optimization in Deep Reinforcement Learning based RS"},{"cite_extract_rate":0.6666666666666661,"cites":[6158,5201,8984,1346,5204,3610],"content":"~\\\\\nMany environments are available for evaluating deep reinforcement learning. Two popular ones are OpenAI gym-based environment and MuJoCo\\footnote{http://mujoco.org/}.\nUnfortunately, there is no standardized simulation platform or benchmark specific to reinforcement learning based recommender systems. Existing work on DRL in RS is usually evaluated through offline datasets or via deployment in real applications. The drawback for evaluating offline datasets include:\n\\begin{itemize}\n \\item Different studies use different environment construction methods which leads to unfair comparison. For instance, some studies use the KG as the environment while some studies assume the environment is gym-like or design a simulator for specific tasks.\n \\item With offline datasets, users' dynamic interests, and environment dynamics are hard to maintain. Deploying the method into a real application is difficult for academic research as it takes time and costs money. Hence, a standardized simulation environment is a desirable solution.\n\\end{itemize}\n\\begin{figure}\n \\centering\n \\includegraphics[width=\\linewidth]{cofounder.pdf}\n \\caption{Left is the traditional MDP transition; Right is the POMDP which considers the environmental confounders such as social influence~.}\n \\label{fig:cofounder}\n\\end{figure}\nThere are several different studies that provide standardized gym\\footnote{https://gym.openai.com/}-based simulation platforms for DRL-based RS research in different scenarios. RecSim~ is a configurable platform that supports sequential interaction between the system and users. RecSim contains three different tasks: interest evolution, interest exploration and long-term satisfaction. RecoGym~ provides an environment for recommendation and advertising. In addition, RecoGym also provides simulation of online experiments such as A/B-tests. However, RecSim and RecoGym are designed for bandit behavior which means users' interests will not change over time. \nVirtualTB~ is proposed to relieve such problems. VirtualTB employs imitation learning to learn a user model to interact with the agent. GANs are employed to generate users' interests. Similar to VirtualTB, Recsimu~ uses a GAN to tackle the complex item distribution.\nIn addition, PyRecGym~ accommodates standard benchmark datasets into a gym-based environment. MARS-Gym~ provides a benchmark framework for marketplace recommendation.~ suggests that existing simulation environments are biased because of biased logged data. Two common biases are discussed: popularity bias and positivity bias. To reduce the effect from those biases, SOFA introduces an Intermediate Bias Mitigation Step for debiasing purposes.\nOne work discusses environment reconstruction by considering confounders.~ claims that users' interests may be affected by social networks which may introduce extra bias to the state and affect the decision-making process. A multi-agent setting is introduced to treat the environment as an agent which can partially relieve the hidden confounder effect. Specifically, a deconfounded environment reconstruction method DEMER is proposed. Different from previously mentioned methods, DEMER assumes the environment is partially observed and models the whole recommendation task as a Partially Observed MDP (POMDP). Different from an MDP, a POMDP contains one more component observation $o\\in\\mathcal{O}$ and the action $a_t$ is derived based on the observation $o_t$ instead of the state $s_t$ by $a_t = \\pi_a(o_t)$. DEMER assumes there is a confounder policy $\\pi_h$ for observation $o_h$ which is composed by $a_t$ and $o_t$: $a_h = \\pi_h(a_t,o_t)$. Moreover, another observation $o_b$ is introduced to observe the transition as well. $\\pi_b$ is the corresponding policy and $a_b = \\pi_b(o_b) = \\pi_b(o_t,a_t,a_h)$. DEMER uses generative adversarial imitation learning (GAIL) to imitate the policy $A,B$.\nGiven trajectory $\\{o_t,o_h,o_b\\}$ for different policies $A$ and $B$, the objective function is defined as\n\\begin{align}\n & (\\pi_a,\\pi_b,\\pi_h) = \\argmin_{(\\pi_a,\\pi_b,\\pi_h)}\\mathbb{E}_{s\\sim \\tau }(L(s,a_t,a_b)) \\notag \\\\\n & \\text{where }L(s,a_t,a_b) = \\mathbb{E}_{(\\pi_a,\\pi_b,\\pi_h)}[\\log D(s,a_t,a_b)]-\\lambda \\sum_{\\pi\\in\\{\\pi_a,\\pi_b,\\pi_h\\}}H(\\pi)\n\\end{align}\nwhere $L(\\cdot)$ is the loss function, $D(\\cdot)$ is a discriminator and $H(\\pi)$ is introduced in GAIL.\n\\\\","id":"3de2362c-6535-43b9-860b-ebc2cbce5821","level":"subsubsection","origin_cites_number":9,"parent_id":"c30b91c8-48bc-4d92-ba66-c418e39b9327","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Component Optimization in Deep Reinforcement Learning based RS"],["subsubsection","Environment Simulation and Reconstruction"]],"subsections":[],"title":"Environment Simulation and Reconstruction"},{"cite_extract_rate":0.33333333333333304,"cites":[5192],"content":"~\\\\\nState representation is another component in DRL-based RS which exists in both model-based and model-free methods. find that the state representation in model-free methods would affect recommendation performance.\nExisting studies usually directly use the embedding as the state representation. propose a supervised learning method to generate a better state representation by utilizing an attention mechanism and a pooling operation as shown in~\\Cref{fig:staterep}.\nSuch a representation method requires training a representation network when training the main policy network, which increases the model complexity.\n\\begin{figure}[!h]\n \\centering\n \\includegraphics[width=0.5\\linewidth]{staterep_wb.pdf}\n \\caption{State representation used in works~. $h_t$ is the output of an attention layer that takes the representation of users' history at time $t$ as input and $g(\\cdot)$ is the pooling operation. }\n \\label{fig:staterep}\n\\end{figure}\n\\\\","id":"45229725-ce80-46d4-8644-8011bc19bde5","level":"subsubsection","origin_cites_number":3,"parent_id":"c30b91c8-48bc-4d92-ba66-c418e39b9327","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Component Optimization in Deep Reinforcement Learning based RS"],["subsubsection","State Representation"]],"subsections":[],"title":"State Representation"},{"cite_extract_rate":0,"cites":[],"content":"~\\\\\nThe reward function is crucial for methods involving DQN.\nA robust reward function can significantly improve training efficiency and performance. find that the DQN may not receive the correct reward value when entering the absorbing state. That is, when the absorbing state is reached, the agent will receive zero reward and harm policy learning. The reason behind this is that when designing the environment zero reward is implicitly assigned to the absorbing state as it is hard to determine the reward value in such a state. propose a robust DQN method, which can stabilize the reward value when facing the absorbing state. The new reward formula can improve the robustness, which is defined as follows:\n\\begin{align}\n r = \\begin{cases} \n r_t & \\text{if } s_{t+1} \\text{ is an absorbing state} \\\\\n r_t + \\gamma Q_{\\theta'}(s_{t+1},a_{t+1})& \\text{otherwise}.\n \\end{cases}\n\\end{align}\nThe major difference is that $r_t$ is assigned to the absorbing state to ensure the agent can continue learning.\nOne remaining problem in current DRL-based RS is the reward sparsity, i.e,. the large state and action spaces make the reward sparsity problem more serious. One of the possible solution would be a better designed reward by using the reward shaping~.","id":"0c8d5ce0-24e2-4a9b-aa04-14e44c18baf4","level":"subsubsection","origin_cites_number":3,"parent_id":"c30b91c8-48bc-4d92-ba66-c418e39b9327","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Deep Reinforcement Learning in Recommender Systems"],["subsection","Component Optimization in Deep Reinforcement Learning based RS"],["subsubsection","Robustness of Reward Functions"]],"subsections":[],"title":"Robustness of Reward Functions"},{"cite_extract_rate":0,"cites":[],"content":"While existing studies have established a solid foundation for DRL-based RS research, this section\noutlines several promising emerging research directions.","id":"3dd345fb-ce99-4727-b174-a3f36f9f711a","level":"section","origin_cites_number":0,"parent_id":"4bb82714-5571-4179-9dac-693fb550f058","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Emerging Topics"]],"subsections":["fd29fd11-361a-4846-b9fe-151e3edf8271","ca4b40c6-413e-4583-9e93-7c0668b3a545","04addbb4-b221-4c4b-aaeb-b0d4e6b0d4fb","19a7db65-832f-4d59-9f47-c445aa549fa1"],"title":"Emerging Topics"},{"cite_extract_rate":0.30769230769230704,"cites":[8884,5185,3588,3616],"content":"Recommender systems are monolithic systems containing \ntasks such as searching, ranking, recommendation, advertising, personalization, and diverse stakeholders such as users and items. Most existing methods are based on single agent. \nMulti-Agent Reinforcement Learning (MARL) is a subfield of reinforcement learning and it is capable of learning multiple policies and strategies. \nWhile a single-agent reinforcement learning framework can only handle a single task, many studies consider the multi-task situation in RS and employ multi-agent DRL (MADRL) or hierarchical DRL (HDRL).\nHDRL~ is proposed to handle complex tasks by splitting such tasks into several small components and asks the agent to determine sub-policies. HDRL belongs to a single-agent reinforcement learning framework such that the agent contains a meta-controller and several controllers. The meta-controller splits the task, and the controllers learn the value and reward functions for designated tasks to get a series of sub-policies. There are a few studies already utilizing HDRL in RS.\n target integrated recommendation to capture user preferences on both heterogeneous items and recommendation channels. Specifically, the meta-controller is used for item recommendation, and controllers aim to find the personalized channel according to user channel-level preferences. uses HDRL for course recommendation in MOOCs, which contains two different tasks: profile reviser and recommendation. The meta-controller aims to make course recommendations by using the revised profile pruned by the controllers.\nDifferent from HDRL, MADRL~ introduces multiple agents to handle the sub-tasks. uses the MADRL for twitter mention recommendation where three agents are initialized. The three agents need to generate different representations for the following tasks: query text, historical text from authors and historical text from candidate users. Once the representations are finalized, the model will conduct the recommendation based on the concatenation of representations. and provide two different views of the communication mechanism in MADRL and demonstrate that agents could work collaboratively or individually. designs a MADRL framework for two tasks where two agents are designed to conduct advertising and recommendation respectively. uses MADRL for collaborative recommendation where each agent is responsible for a single user. MADRL is adopted to help the recommender consider both collaboration and potential competition between users. designs a charging recommender system for intelligent electric vehicles by using decentralized agents to handle sub-tasks and a centralized critic to make the final decision.\nHierarchical multi-agent RL (HMARL)~ proves that MARL and HRL can be combined.\nRecently, introduces HMADRL into the continuous action space, which provides a direction for RS. uses HMARL for multi-goal recommendation where the meta-controller considers users' long-term preferences and controllers focus on short-term click behavior. While the meta-controller and controllers in HDRL deal with sub-tasks that belong to a single task, HMARL focuses on multi-task or multi-goal learning where the meta-controller and controllers belong to different agents and deal with different tasks or goals.\nHMADRL would be a suitable solution for future research work in DRL-based RS where HDRL can be used to split a complex task into several sub-tasks such as users' long-term interests and short-term click behavior, and MADRL can jointly consider multiple factors such as advertising .","id":"fd29fd11-361a-4846-b9fe-151e3edf8271","level":"subsection","origin_cites_number":13,"parent_id":"3dd345fb-ce99-4727-b174-a3f36f9f711a","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Emerging Topics"],["subsection","Multi-Agent and Hierarchical Deep Reinforcement Learning-based RS"]],"subsections":[],"title":"Multi-Agent and Hierarchical Deep Reinforcement Learning-based RS"},{"cite_extract_rate":0.7142857142857141,"cites":[6160,3592,6159,6161,5199],"content":"As\nmentioned,\nthe reward function plays a critical role in DRL-based recommender systems. In many existing works, reward functions are manually designed. The common method uses users' click behavior to represent the reward and to reflect users' interests. However, such a setting can not represent users' long-term goals~ as clicking or not only depicts part of the feedback information from users. It requires significant effort to design a reward function, due to the large number of factors that can affect users' decision, such as social engagement or bad product reviews, which may adversely affect recommendation performance. It is difficult to include all potential factors into the reward function because not every factor can be represented properly. A few works~ show that manually designed reward functions can be omitted by employing inverse reinforcement learning (IRL)~ or generative adversarial imitation learning (GAIL)~. Such inverse DRL-based methods require using expert demonstration as the ground truth. However, expert demonstration is often hard to obtain for recommendation scenarios.\nThose two studies conduct experiments in an offline dataset-based simulation environment that can access expert demonstration. In contrast, use IRL as the main algorithm to train the agent while use both demonstration and reward to train the agent. also employ GAIL to improve recommendation performance. In this work, GAIL is used to learn the reasoning path inside the KG to provide side information to help the agent learn the policy. Although IRL achieves some progress in RS, the lack of demonstration is a key shortcoming that impedes adoption in RS. One possibility is to use the IRL method in casual reasoning to help improve interpretability~ thus boosting recommendation performance. Alternately, IRL may be suitable for learning users' long-term and static behavior to support the reward function.","id":"ca4b40c6-413e-4583-9e93-7c0668b3a545","level":"subsection","origin_cites_number":7,"parent_id":"3dd345fb-ce99-4727-b174-a3f36f9f711a","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Emerging Topics"],["subsection","Inverse Deep Reinforcement Learning for RS"]],"subsections":[],"title":"Inverse Deep Reinforcement Learning for RS"},{"cite_extract_rate":0.555555555555555,"cites":[1444,6160,180,6162,5194],"content":"Graph data and KG are widely used in RS.\nGraph modeling enables an RS to leverage interactions between users and the recommender for reasoning or improving interpretability. According to existing studies about deep learning-based RS~, embedding is a technique used to get the representation for the input data. Graph embedding is a common solution to handle graph-like data. GCN is a type of graph embedding method which\nare broadly used in RS to process graph data. propose a variant of GCN to learn the embedding for KG. Specifically, they propose knowledge graph convolutional networks (KGCN) to capture the high-order structural proximity among entities in a knowledge graph. \nIn DRL-based RS, graph data are handled similarly---s underthe transformed into an embedding and fed to the agent. uses a traditional graph embedding method TransE~ to generate the state representation for DRL-based RS. There are several studies that use GCN in DRL for recommendations under different settings. propose a graph convolutional RL (DGN) method which integrates the GCN into the Q-learning framework for general RL problems by replacing the state encoding layer with the GCN layer. extend this method into the deep learning field and apply it to recommender systems. To be specific, multiple GCN layers are employed to process the sub-graphs for a given item $i$. employs KG inside the actor-critic algorithm to help the agent learn the policy. Specifically, the critic network contains a GCN layer to give weight to the graph and conduct searches in the graph to find an optimal path and hence guide the optimization of policy learning. However, such a method is relatively computationally expensive as it requires jointly training the GCN and the actor-critic network. adopts a Graph Attention Network (GAT)~ into the actor-critic network to conduct recommendation. In addition, the GAT is used as an encoder to obtain a state representation.\nA common way of using GCN or its variants in DRL-based RS is the state encoder. \nThe related challenge is the difficulty for the environment to provide a graph-like input to the GCN.","id":"04addbb4-b221-4c4b-aaeb-b0d4e6b0d4fb","level":"subsection","origin_cites_number":9,"parent_id":"3dd345fb-ce99-4727-b174-a3f36f9f711a","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Emerging Topics"],["subsection","Graph Neural Networks for Boosting DRL-based RS"]],"subsections":[],"title":"Graph Neural Networks for Boosting DRL-based RS"},{"cite_extract_rate":0.6666666666666661,"cites":[4711,6163,6164,5188],"content":"Self-supervised learning (SSL) is a technique in which the model is trained by itself without external label information. SSL-DRL is receiving growing interest in robotics \n~. shows that SSL can be used to learn the policy when doing navigation by providing real-world experience. demonstrates that SSL-DRL can be used to help the agent learn synergies between two similar policies, thus empowering the agent to conduct two different tasks. \nRecent advances in SSL RL show that SSL can also provide interpretability for RL, which is promising for interpretable RS research~. shows that SSL based RL can highlight the task-relevant information to guide the agent's behavior.\nMoreover, shows that SSL can be used to provide negative feedback for DRL-based RS to improve recommendation performance. To be specific, a self-supervised loss function is appended into the normal DRL loss function,\n\\begin{align}\n -\\sum_{i=1}^nY_i\\log\\bigg(\\frac{e^{y_i}}{\\sum_{i'=1}^ne^{y_{i'}}}\\bigg) + L_{\\mathit{DRL}}\n\\end{align}\nwhere $Y_i$ is an indicator function to show users interact with the item $i$ or not. $L_{\\mathit{DRL}}$ could vary, if the DQN is adopted, \\Cref{dqnloss} should be used.\nSSL demonstrates promising performance in visual representation in recent years, which would be a possible solution to generate the state representation as there are a few DRL-based RS studies that adopt CNNs to process image-like data and transform it into a state~. Furthermore, as an unsupervised learning approach, SSL would provide a new direction about defining the reward function by learning common patterns between different states as well as multi-task learning.","id":"19a7db65-832f-4d59-9f47-c445aa549fa1","level":"subsection","origin_cites_number":6,"parent_id":"3dd345fb-ce99-4727-b174-a3f36f9f711a","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Emerging Topics"],["subsection","Self-Supervised DRL-based RS"]],"subsections":[],"title":"Self-Supervised DRL-based RS"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we outline several open questions and challenges that exist in DRL-based RS research. We believe these issues could be critical for the future development of DRL-based RS.","id":"62c3e9c3-0318-4d9a-947c-bf1d76096072","level":"section","origin_cites_number":0,"parent_id":"4bb82714-5571-4179-9dac-693fb550f058","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Open Questions"]],"subsections":["e24a9f91-4259-41ce-8cfe-67de33facae4","f3bc9826-577b-48f2-9fd1-1cbd7502868e","8fb9ed60-6e10-4fb7-a67b-6f49f9ef3798","bb9fa85c-840e-4ea9-928c-30b0b38ddbc5","317e1a16-83a6-4e39-b6b1-02817b7555bf","03cb7efb-1098-4511-ad2a-26dcf319c6b5"],"title":"Open Questions"},{"cite_extract_rate":0.6666666666666661,"cites":[3586,8985],"content":"Sample inefficiency is a well-known challenge in model-free DRL methods. Model-free DRL requires a significant number of samples as there is no guarantee that the received state is useful. Normally, after a substantial number of episodes, the agent may start learning as the agent finally receives a useful state and reward signal. A common solution is the experience replay technique, which only works in off-policy methods. \nExperience replay still suffers the sample inefficiency problem~ as not every past experience is worth replaying. propose selected experience replay (SER) that only stores valuable experience into the replay buffer and thus improves sample efficiency. while traditional DRL environments only contain several\\footnote{For example, the number of actions in MuJoCo is less than one hounded.} candidate items, in DRL-based RS, the agent must deal with a significantly larger action space as RS may contain lots of candidate items. Existing DRL-based RS studies on traditional experience replay methods often demonstrate slow converge speed. design a user model to improve the sample efficiency through auxiliary learning. Specifically, they apply the auxiliary loss with the state representation, and the model distinguishes low-activity users and asks the agent to update the recommendation policy based on high-activity users more frequently.\nOn the other hand, model-based methods are more sample efficient. However, they introduce extra complexity as the agent is required to learn the environment model as well as the policy. Due to the extremely large action space and possibly large state space (depending on users' contextual information) in RS, approximating the environment model and policy simultaneously\nbecomes challenging.","id":"e24a9f91-4259-41ce-8cfe-67de33facae4","level":"subsection","origin_cites_number":3,"parent_id":"62c3e9c3-0318-4d9a-947c-bf1d76096072","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Open Questions"],["subsection","Sample Efficiency"]],"subsections":[],"title":"Sample Efficiency"},{"cite_extract_rate":0.388888888888888,"cites":[5190,6154,5180,5195,5177,5165,5199],"content":"The exploration and exploitation dilemma is a fundamental and challenging problem in reinforcement learning research and receives lots of attention in DRL.\nThis dilemma describes a trade-off between obtaining new knowledge and the need to use that knowledge to improve performance. Many DQN-based methods focus on exploration before the replay buffer is full and exploitation afterward. Consequently, it requires an extremely large replay buffer to allow all possibilities in recommendation can be stored.\nDRN employs Dueling Bandit Gradient Descent (DBGD)~ to encourage exploration while ~ introduces a regularization or entropy term into the objective function to do so.~ uses the sheer size of the action space to ensure sufficient exploration.~ uses a separate KG or elaborated graph exploration operation to conduct exploration.~ employs Boltzmann exploration to get the benefit of exploratory data without negatively impacting user experience. \nIn addition, $\\epsilon$-greedy is the most common technique used to encourage exploration ~. Remaining studies rely on a simulator to conduct exploration. However, it may suffer from noise and over-fitting~ because of the gap between simulation and real online application. For most DRL-based methods such as vanilla DQN, policy gradient, or actor-critic-based methods, $\\epsilon$-greedy would be a good choice for exploration. In addition, injecting noise into the action space would also be helpful for those actor-critic-based methods~. For methods involving KGs, $\\epsilon$-greedy may help, but the elaborated graph exploration methods may receive better performance.","id":"f3bc9826-577b-48f2-9fd1-1cbd7502868e","level":"subsection","origin_cites_number":18,"parent_id":"62c3e9c3-0318-4d9a-947c-bf1d76096072","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Open Questions"],["subsection","Exploration and Exploitation"]],"subsections":[],"title":"Exploration and Exploitation"},{"cite_extract_rate":1,"cites":[7022,3199,8106],"content":"Existing work generally trains DRL algorithms in simulation environments or offline datasets.\nDeploying DRL algorithms into real applications is challenging due to the gap between simulation and real-world applications.\nSimulation environments do not contain domain knowledge or social impact.\nThey can not cover the domain knowledge and task-specific engineering in the real-world recommendation.\nHow to bridge the gap between simulation and real applications is a challenging topic. Sim2real~ is a transfer learning approach that transfers DRL policies from simulation environments to reality. Sim2real uses domain adaption techniques to help agents transfer learned policy. Specifically, it adopts GANs to conduct adaption by generating different samples.\nRL-CycleGAN~ is a sim2real method for vision-based tasks. It uses CycleGAN~ to conduct pixel-level domain adaption. Specifically, it maintains cycle consistency during GAN training and encourages the adapted image to retain certain attributes of the input image.\nIn DRL-based RS, sim2real would be a possible solution for generalizing the learned policy from simulation environments to reality. However, sim2real is a new technique still under exploration. It shows an adequate capability in simple tasks and requires more effort to handle the complex task such as recommendation. We believe it is a workable solution for generalizing from simulation to reality.","id":"8fb9ed60-6e10-4fb7-a67b-6f49f9ef3798","level":"subsection","origin_cites_number":3,"parent_id":"62c3e9c3-0318-4d9a-947c-bf1d76096072","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Open Questions"],["subsection","Generalizing from Simulation to Real-World Recommendation"]],"subsections":[],"title":"Generalizing from Simulation to Real-World Recommendation"},{"cite_extract_rate":0.5,"cites":[6165,5165,3891],"content":" observe that user behavior data are not experimental but observational, which leads to problems of bias and unfairness.\\\\\nThere are two reasons why bias is so common. First, the inherent characteristic of user behavior data is not experimental but observational. In other words, data that are fed into recommender systems are subject to selection bias . For instance, users in a video recommendation system tend to watch, rate, and comment on those movies that they are interested in. Second, a distribution discrepancy exists, which means the distributions of users and items in the recommender system are not even. Recommender systems may suffer from 'popularity bias', where popular items are recommended far more frequently than the others. However, the ignored products in the “long tail” can be equally critical for businesses as they are the ones less likely to be discovered.\n denote the unfairness as that the system systematically and unfairly discriminates against certain individuals or groups of individuals in favor of others. \nA large number of studies explore dynamic recommendation systems by utilizing the agent mechanism in reinforcement learning (RL), considering the information seeking and decision-making as sequential interactions. How to evaluate a policy efficiently is a big challenge for RL-based recommenders. Online A/B tests are not only expensive and time-consuming but also sometimes hurt the user experience. Off-policy evaluation is an alternative strategy that historical user behavior data are used to evaluate the policy. However, user behavior data are biased, as mentioned before, which causes a gap between the policy of RL-based RS and the optimal policy.\nTo eliminate the effects of bias and unfairness, use the inverse of the probability of historical policy to weight the policy gradients. introduce a debiasing step that corrects the biases presented in the logged data before it is used to simulate user behavior. propose to build a customer simulator that is designed to simulate the environment and handle the selection bias of logged data.","id":"bb9fa85c-840e-4ea9-928c-30b0b38ddbc5","level":"subsection","origin_cites_number":6,"parent_id":"62c3e9c3-0318-4d9a-947c-bf1d76096072","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Open Questions"],["subsection","Bias (Unfairness)"]],"subsections":[],"title":"Bias (Unfairness)"},{"cite_extract_rate":0.4,"cites":[5190,8881],"content":"Although deep learning-based models can generally improve the performance of recommender systems, they are not easily interpretable. As a result, it becomes an important task to make recommender results explainable, along with providing high-quality recommendations.\nHigh explainability in recommender systems not only helps end-users understand the items recommended but also enables system designers to check the internal mechanisms of recommender systems. review different information sources and various types of models that can facilitate explainable recommendation. Attention mechanisms and knowledge graph techniques currently play an important role in realizing explainability in RS.\nAttention models have great advantages in both enhancing predictive performance and having greater explainability~. introduce a reinforcement learning framework incorporated with an attention model for explainable recommendation. Firstly, it achieves model-agnosticism by separating the recommendation model from the explanation generator. Secondly, the agents that are instantiated by attention-based neural networks can generate sentence-level explanations. \nKnowledge graphs contain rich information about users and items, which can help to generate intuitive and more tailored explanations for the recommendation system . Recent work has achieved greater explainability by using reinforcement and knowledge graph reasoning. The algorithm from learns to find a path that navigates from users to items of interest by interacting with the knowledge graph environment. extract imperfect path demonstrations with minimum labeling effort and propose an adversarial actor-critic model for demonstration-guided path-finding. Moreover, it achieves better recommendation accuracy and explainability by reinforcement learning and knowledge graph reasoning.","id":"317e1a16-83a6-4e39-b6b1-02817b7555bf","level":"subsection","origin_cites_number":5,"parent_id":"62c3e9c3-0318-4d9a-947c-bf1d76096072","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Open Questions"],["subsection","Explainability"]],"subsections":[],"title":"Explainability"},{"cite_extract_rate":0.5,"cites":[6167,1192,6166],"content":"Adversarial samples demonstrate that deep learning-based methods are vulnerable. Hence,\nrobustness becomes an open question for both RS and DRL. Specifically, adversarial attack and defense in RS have received a lot of attention in recent years~ as security is crucial in RS. Moreover, DRL policies are vulnerable to adversarial perturbations to agent's observations~. provide an adversarial attack method for perturbing the observations, thus affecting the learned policy. Hence, improving the robustness is the common interest for DRL and RS, which would be a critical problem for DRL-based RS. provide an adversarial attack detection method for DRL-based RS which uses the GRU to encode the action space into a low-dimensional space and design decoders to detect the potential attack. However, it only considers Fast Gradient Sign Method (FGSM)-based attacks and strategically-timed attacks~. Thus, it lacks the capability to detect other types of attack. Moreover, it only provides the detection method while the defence is still an opening question.\nWe believe zero-shot learning techniques would be a good direction for training a universal adversarial attack detector. For defence, it is still an open question for DRL-based RS, though recent advances in adversarial defence in DRL may provide some insights~.","id":"03cb7efb-1098-4511-ad2a-26dcf319c6b5","level":"subsection","origin_cites_number":6,"parent_id":"62c3e9c3-0318-4d9a-947c-bf1d76096072","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Open Questions"],["subsection","Robustness on Adversarial Samples and Attacks"]],"subsections":[],"title":"Robustness on Adversarial Samples and Attacks"},{"cite_extract_rate":0,"cites":[],"content":"In this section, we provide a few potential future directions of DRL-based RS. Benefiting from recent advances in DRL research, we believe those topics can boost the progress of DRL-based RS research.","id":"db61769c-00fa-49cf-aca1-794fe6209774","level":"section","origin_cites_number":0,"parent_id":"4bb82714-5571-4179-9dac-693fb550f058","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Future Directions"]],"subsections":["e714f9af-5ef5-4cc0-b0fc-93a677fe49fd","69b9e3f2-89df-425e-ae0b-359e87390941","18638c1b-d943-46f7-8255-6db369c76c4b"],"title":"Future Directions"},{"cite_extract_rate":0.6666666666666661,"cites":[6168,4573,6169,4428,3610,4416],"content":"Causality is a generic relationship between a cause and effect. Moreover, inferring causal effects is a fundamental problem in many applications like computational advertising, search engines, and recommender systems~. \nRecently, some researchers have connected reinforcement learning with learning causality to improve the effects for solving sequential decision-making problems.\nBesides, Learning agents in reinforcement learning frameworks face a more complicated environment where a large number of heterogeneous data are integrated.\nFrom our point of view, causal relationships would be capable of improving the recommendation results by introducing the directionality of cause and the effect. The users' previous choices have impact on the subsequent actions. This can be cast as an interventional data generating the dynamics of recommender systems. By viewing a policy in RL as an intervention, we can detect unobserved confounders in RL and choose a policy on the expected reward to better estimate the causal effect . Some studies improve RL models with causal knowledge as side information. Another line of work uses causal inference methods to achieve unbiased reward prediction~. \n propose a Causal Inference Q-network which introduces observational inference into DRL by applying extra noise and uncertain inventions to improve resilience. Specifically, in this work, noise and uncertainty are added into the state space during the training state, and the agent is required to learn a causal inference model by considering the perturbation.\n give the first demonstration that model-free reinforcement learning can be used for causal reasoning. They explore meta-reinforcement learning to solve the problem of causal reasoning. The agents trained by a recurrent network able to make causal inferences from observational data and output counterfactual predictions. \n bridge RL and causality by data-fusion for reinforcement learners. Specifically, online agents combine observations, experiments and counterfactual data to learn about the environment, even if unobserved confounders exist. Similarly, make the model-based RL agents work in a causal way to explore the environment under the Partially-Observable Markov Decision Process (POMDP) setting. They consider interventional data and observational data jointly and interprete model-based reinforcement learning as a causal inference problem. In this way, they bridge the gap between RL and causality by relating common concepts in RL and causality. \nRegarding explainability in RL, propose to explain the behavior of agents in reinforcement learning with the help of causal science. The authors encode causal relationships and learn a structural causal model in RL, which is used to generate explanations based on counterfactual analysis. With counterfactual exploration, this work is able to generate two contrastive explanations for `why' and `why not' questions.\nIt is so important to search for a Directed Acyclic Graph (DAG) in causal discovery. Considering traditional methods rely on local heuristics and predefined score functions, propose to use reinforcement learning to search DAG for causal discovery. They use observational data as an input, RL agents as a search strategy and output the causal graph generated from an encoder-decoder NN model.","id":"e714f9af-5ef5-4cc0-b0fc-93a677fe49fd","level":"subsection","origin_cites_number":9,"parent_id":"db61769c-00fa-49cf-aca1-794fe6209774","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Future Directions"],["subsection","Causal and Counterfactual Inference"]],"subsections":[],"title":"Causal and Counterfactual Inference"},{"cite_extract_rate":1,"cites":[7753,6170],"content":"Recommender systems often need to deal with multiple scenarios such as joint recommendation and adverting, offline DRL and meta DRL provide a promising direction for achieving multiple scenarios at the same time.\nOffline DRL is a new paradigm of DRL that can be combined with existing methods such as self-supervised learning and transfer learning to move toward real-world settings. \nOffline DRL~ (also known as batch DRL) is designed for tasks which contain huge amounts of data.\nGiven a large dataset that contains past interactions, offline DRL uses the dataset for training across many epochs but does not interact with the environment.\nOffline DRL provides a solution that can be generalized to new scenarios as it was trained by a large sized dataset.\nSuch generalization ability is critical to RSs, which may need to deal with multiple scenarios or multiple customers.\nWhile offline DRL could provide a new direction for DRL-based RS, it still faces a few problems regarding handling the distributional shifts between existing datasets and real-world interactions. \nMeta DRL~ is defined as \nmeta learning in the filed of DRL. Meta DRL is another approach to help agents to generalize to new tasks or environments. Different from offline DRL, meta DRL contains a memory unit which is formed by the recurrent neural network to memorize the common knowledge for different tasks. Different from offline DRL, meta DRL does not require a large amount of data to train.","id":"69b9e3f2-89df-425e-ae0b-359e87390941","level":"subsection","origin_cites_number":2,"parent_id":"db61769c-00fa-49cf-aca1-794fe6209774","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Future Directions"],["subsection","Offline DRL and Meta DRL"]],"subsections":[],"title":"Offline DRL and Meta DRL"},{"cite_extract_rate":0.8,"cites":[6171,2219,1391,1389],"content":"An actor-critic method uses the traditional policy gradient method, which suffers from the high variance problem due to the gap between behavior policy (i.e., the policy that is being used by an agent for action select) and target policy (i.e., the policy that an agent is trying to learn).\nA method commonly used to \nrelieve\nthe high variance problem is Advantage Actor-critic (A2C). \nDifferent from traditional actor-critic methods, A2C uses an advantage function to replace the Q-function inside the critic network. The advantage function $A(s_t)$ is defined as the expected value of the TD-error. \nThe new objective function for policy gradient can be written as,\n\\begin{align}\n \\mathbb{E}_{\\tau \\sim d_{\\pi_\\theta}}[\\sum_{t=1}^T\\underbrace{(Q(s_t,a_t) - V(s_t))}_{A(s_t)}\\sum_{t=1}^T\\nabla_{\\theta} \\log \\pi_{\\theta}(s_t,a_t)].\n\\end{align}\nHowever, A2C still uses DDPG as the main training algorithm, which may suffer function approximation errors when estimating the Q value.\nTwin-Delayed DDPG (TD3)~ is designed to improve the function approximation problem in DDPG which uses clipped double Q-learning to update the critic. The gradient update can be expressed as,\n\\begin{align}\n \\mathbb{E}_{\\tau \\sim d_{\\pi_\\theta}}[\\sum_{t=1}^Tr(s_t,a_t) + \\gamma \\min(Q_1(s_t,a_t+\\epsilon), Q_2(s_t,a_t+\\epsilon))\\sum_{t=1}^T\\nabla_{\\theta} \\log \\pi_{\\theta}(s_t,a_t)].\n\\end{align}\nwhere $\\epsilon\\sim\\textit{clip}(\\mathcal{N}(0,\\sigma,-c,c))$, $\\sigma$ is the standard deviation and $c$ is a constant for clipping. \nAnother two ways to improve actor-critic methods are Trust Region Policy Optimization (TRPO)~ and Proximal Policy Optimization (PPO)~, which focus on modification of the advantage function. TRPO aims to limit the step size for each gradient to ensure it will not change too much. The core idea is to add a constraint to the advantage function,\n\\begin{align}\n \\frac{\\pi(a|s)}{\\pi_{old}(a|s)}A(s),\n\\end{align}\nwhere the KL divergence will be used to measure the distance between the current policy and the old policy is small enough. PPO has the same goal as TRPO which is to try to find the biggest possible improvement step on a policy using the current data. PPO is a simplified version of TRPO which introduces the clip operation,\n\\begin{align}\n \\min\\bigg(\\frac{\\pi(a|s)}{\\pi_{old}(a|s)}A(s),\\text{clip}\n \\bigg(\\frac{\\pi(a|s)}{\\pi_{old}(a|s)}A(s), 1-\\epsilon, 1+\\epsilon\\bigg)A(s)\\bigg).\n\\end{align}\nSoft Actor-Critic (SAC)~ is another promising variant of the actor-critic algorithm and is widely used in DRL research. SAC uses the entropy term to encourage the agent to explore, which could be a possible direction to solve the exploration and exploitation dilemma. Moreover, SAC assigns an equal probability to actions that are equally attractive to the agent to capture those near-optimal policies.\nAn example of related work uses SAC to improve the stability of the training process in RS.","id":"18638c1b-d943-46f7-8255-6db369c76c4b","level":"subsection","origin_cites_number":5,"parent_id":"db61769c-00fa-49cf-aca1-794fe6209774","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Future Directions"],["subsection","Further Developments in Actor-Critic Methods"]],"subsections":[],"title":"Further Developments in Actor-Critic Methods"},{"cite_extract_rate":0,"cites":[],"content":"In this survey, we provide a comprehensive overview the use of deep reinforcement learning in recommender systems. We introduce a classification scheme for existing studies and discuss them by category. We also provide an overview of such existing emerging topics and point out a few promising directions. \nWe hope this survey can provide a systematic understanding of the key concepts in DRL-based RS and valuable insights for future research. \n\\bibliographystyle{ACM-Reference-Format}\n\\bibliography{sample-base}\n\\end{document}\n\\endinput","id":"0a382833-431c-4392-982a-66ab0eba2961","level":"section","origin_cites_number":0,"parent_id":"4bb82714-5571-4179-9dac-693fb550f058","prefix_titles":[["title","A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ]\n ],\n \"subsections\": [\n \"f16a1a1e-25a7-4e14-aaa7-69bb7d4fc8b0\",\n \"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f\",\n \"696bfdf7-6191-445e-82a8-ae4df2378853\",\n \"3dd345fb-ce99-4727-b174-a3f36f9f711a\",\n \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"db61769c-00fa-49cf-aca1-794fe6209774\",\n \"0a382833-431c-4392-982a-66ab0eba2961\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 8983,\n 5194\n ],\n \"content\": \"Recent years have seen significant progress in recommendation techniques, from traditional recommendation techniques, e.g., collaborative filtering, content-based recommendation and matrix factorization~, to deep learning based techniques.\\nIn particular, deep learning show strong advantages in solving complex tasks and dealing with complex data, due to its capability to capture non-linear user-item relationships and deal with various types of data sources such as images and text.\\nIt has thus been increasingly used in recommender systems.\\nDeep learning-based recommender systems have limitations in capturing interest dynamics~ due to distribution shift, i.e., the training phase is based on an existing dataset which may not reflect real user preferences that undergo rapid change.\\nIn contrast, deep reinforcement learning (DRL) aims to train an agent that can learn from interaction trajectories provided by the environment by combining the power of deep learning and reinforcement learning. Since an agent in DRL can actively learn from users' real-time feedback to infer dynamic user preferences,\\nDRL is especially suitable for learning from interactions, such as human-robot collaboration; it has also driven significant advances in a range of interactive applications ranging from video games, Alpha Go to autonomous driving~.\\nIn light of the significance and recent progresses in DRL for recommender sytsems, we aim to timely summaize and comment on DRL-based recommendation systems in this survey.\\nA\\nrecent survey on reinforcement learning based recommender systems provides a general review about reinforcement learning in recommender systems without a sophsiticated investigation of the growing area of deep reinforcement learning. Our survey distinguishes itself in providing a systematic and comprehensive overview of existing methods in DRL-based recommender systems, along with a discussion of emerging topics, open issues, and future directions. This survey introduces researchers, practitioners and educators into this topic\\nand fostering an understanding of the key techniques in the area.\\nThe main contributions of this survey include the following:\\n\\\\begin{itemize}\\n \\\\item We provide an up-to-date comprehensive review of deep reinforcement learning in recommender systems, with state of the art techniques and pointers to core references. To the best of our knowledge, this the first comprehensive survey in deep reinforcement learning based recommender systems.\\n \\\\item We present a taxonomy of the literature of deep reinforcement learning in recommender systems. Along with the outlined taxonomy and literature overview, we discuss the benefits, drawbacks and give suggestions for future research directions. \\n \\\\item We shed light on emerging topics and open issues for DRL-based recommender systems. We also point out future directions that could be crucial for advancing DRL-based recommender systems.\\n\\\\end{itemize}\\nThe remainder of this survey is organized as follows: Section 2 provides an overview of recommender systems, DRL and their integration. Section 3 provides a literature review with a taxonomy and classification mechanism. Section 4 reviews emerging topics, and Section 5 points out open questions. Finally, Section 6 provides a few promising future directions for further advances in this domain.\\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=\\\\linewidth]{route.pdf}\\n \\\\caption{Taxonomy of Deep Reinforcement Learning based Recommender Systems in this survey}\\n \\\\label{fig:overview}\\n\\\\end{figure}\",\n \"id\": \"f16a1a1e-25a7-4e14-aaa7-69bb7d4fc8b0\",\n \"level\": \"section\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section,we introduce key concepts related to dynamic recommender systems (RS) and deep reinforcement learning (DRL), and motivate the introduction of DRL to dynamic recommender systems.\",\n \"id\": \"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Background\"\n ]\n ],\n \"subsections\": [\n \"a7e80129-286c-44ec-84f0-48ae3a07002b\",\n \"e05ae3e8-b24d-431b-b181-250a9bf1fb50\",\n \"467d3d60-cfac-40c2-8a3f-6dbe98149287\"\n ],\n \"title\": \"Background\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Recommender systems require coping with \\\\emph{dynamic} environments by estimating rapidly changing users' preferences and proactively recommending items to users.\\nLet $\\\\mathcal{U}$ be a set of users of cardinality $|\\\\mathcal{U}|$ and $\\\\mathcal{I}$ be a set of items of cardinality $|\\\\mathcal{I}|$.\\nFor each user $u\\\\in\\\\mathcal{U}$, we observe a sequence of user actions $\\\\mathbb{X}^u = [x_1^u, x_2^u, \\\\cdots , x_{T_u}^u]$ with item $x_t^u \\\\in \\\\mathcal{I}$, i.e., each event in a user sequence comes from the item set. We refer to a user making a decision as an interaction with an item. Suppose the feedback (e.g., ratings or clicking behavior) provided by users is $\\\\mathcal{F}$, then a dynamic recommender system maintains the corresponding recommendation policy $\\\\pi^u_t$, which will be updated systematically based on the feedback $f^u_i \\\\in \\\\mathcal{F}$ received during the interaction for item $i \\\\in \\\\mathcal{I}$ at the timestamp $t$. \\nThe marriage of deep learning and reinforcement learning has fueled breakthroughs in recommender systems. \\nDRL-based RS consists of a pipeline with three building blocks: environment construction, state representation and recommendation policy learning. Environment construction aims to build an environment based on a set of users' historical behaviors. State representation is provided by the environment containing certain user information including historical behavior, demographic data (etc.). Recommendation policy learning is the key component to understand and predict users' future behavior. \\nDL-based RS receives user feedback (e.g., ratings or clicks) to reflect users' interests and update the recommender, while DRL-based RS receives the reward provided by the environment to update the policy. The reward provided by the environment is a pre-defined function containing several factors. The detailed process of DL based RS and DRL-based RS mapping can be found in~\\\\Cref{fig:catnew}.\",\n \"id\": \"a7e80129-286c-44ec-84f0-48ae3a07002b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Why Deep Reinforcement Learning for Recommendation?\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Why Deep Reinforcement Learning for Recommendation?\"\n },\n {\n \"cite_extract_rate\": 0.14285714285714202,\n \"cites\": [\n 8105\n ],\n \"content\": \"The typical defining feature of DRL is to use the deep learning to approximate reinforcement learning's value function and solve high-dimensional Markov Decision Processes (MDPs)~. Formally, a MDP can be represented as a tuple ($\\\\mathcal{S},\\\\mathcal{A},\\\\mathcal{P},\\\\mathcal{R},\\\\gamma$). The agent chooses an action $a_t\\\\in\\\\mathcal{A}$ according to the policy $\\\\pi_t(s_t)$ at state $s_t\\\\in\\\\mathcal{S}$. The environment receives the action and produces a reward $r_{t+1}\\\\in\\\\mathcal{R}$ and transfers the reward into the next state $s_{t+1}$ according to the transition probability $P(s_{t+1}|s_t,a_t)\\\\in\\\\mathcal{P}$. The transition probability $\\\\mathcal{P}$ is unknown beforehand in DRL.\\nSuch a process continues until the agent reaches the terminal state or exceeds a pre-defined maximum time step. The overall objective is to maximize the expected discounted cumulative reward,\\n\\\\begin{align}\\n \\\\mathbb{E}_\\\\pi [r_t] = \\\\mathbb{E}_\\\\pi \\\\big[\\\\sum_{0}^\\\\infty \\\\gamma^k r_{t+k}\\\\big] \\n\\\\end{align}\\nwhere $\\\\gamma \\\\in [0,1]$ is the discount factor that balances the future reward and the immediate reward.\\n\\\\begin{wrapfigure}{r}{0.45\\\\textwidth}\\n \\\\begin{center}\\n \\\\includegraphics[width=0.45\\\\textwidth]{tax.pdf}\\n \\\\end{center}\\n \\\\caption{Taxonomy of Deep Reinforcement Learning in Recommender Systems}\\n \\\\label{fig:tax}\\n\\\\end{wrapfigure}\\nDeep reinforcement learning can be divided into two categories: \\\\emph{model-based} and \\\\emph{model-free} methods (a detailed taxonomy can be found in \\\\Cref{fig:tax}). The major {\\\\emph difference} between the two is whether the agent can learn a model of the environment. Model-based methods aim to estimate the transition function and reward function, while model-free methods aim to estimate the value function or policy from experience. In model-based methods, the agent accesses the environment and plans ahead while model-free methods gain sample efficiency from using models which are more extensively developed and tested than model-based methods in recent literature~. \\nDeep reinforcement learning approaches are divided into three streams: \\\\emph{value-based}, \\\\emph{policy-based} and \\\\emph{hybrid} methods. In value-based methods, the agent updates the value function to learn a policy; policy-based methods learn the policy directly; and hybrid methods combine value-based and policy-based methods called \\\\emph{actor-critic} methods. Actor-critic contains two different networks where an actor network uses a policy-based method and the critic uses a value-based method to evaluate the policy learned by the agent.\\n\\\\begin{table}[ht]\\n \\\\centering\\n \\\\caption{Notations}\\n \\\\begin{tabular}{cc|ccc}\\n \\\\hline\\n Notations & Name & Notations & Name & Notes \\\\\\\\\\\\hline\\n $Q(\\\\cdot)$ & Q-Value Function & $s$ & State & users' preference \\\\\\\\\\n $V(\\\\cdot)$ & Value Function & $a$ & Action & Recommended item(s) \\\\\\\\\\n $\\\\gamma$ & Discount Factor & $\\\\pi$, $\\\\mu(\\\\cdot)$ & Policy & Recommendation policy \\\\\\\\\\n $\\\\mathbb{E}$ & Expected Value & $r(\\\\cdot,\\\\cdot)$ & Reward & users' click behavior \\\\\\\\\\n $\\\\theta$ & Model Parameter & $\\\\alpha$ & constant $\\\\in [0,1]$ & - \\\\\\\\\\n $p(\\\\cdot)$ & Transition Probability & $\\\\tau$ & Sampled Trajectory & A tuple $(s_t,a_t,s'_t,r_t)$ \\\\\\\\\\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:my_label}\\n\\\\end{table}\\nDeep reinforcement learning can be divided into \\\\emph{on-policy} and \\\\emph{off-policy} methods.\\nIn off-policy methods,\\nthe behavior policy $\\\\pi_b$ is used for exploration while the target policy $\\\\pi$ is used for decision-making.\\nFor on-policy methods, the behavior policy is the same as the target policy. \\n\\\\textbf{Q-learning}~ is an off-policy value-based learning scheme for finding a greedy target policy:\\n\\\\begin{align}\\n \\\\pi(s) = \\\\argmax_a Q_\\\\pi (s,a)\\n\\\\end{align}\\nwhere $Q_u (s,a)$ denotes the $Q$-value and is used in a small discrete action space. For a deterministic policy, the $Q$ value can be calculated as follows\\n\\\\begin{align}\\n Q (s_t,a_t) = \\\\mathbb{E}_{\\\\tau \\\\sim \\\\pi}[r(s_t,a_t) + \\\\gamma Q(s'_t, a'_t)].\\n\\\\end{align}\\nDeep Q learning (DQN)~ uses deep learning to approximate a non-liner Q function parameterized by $\\\\theta_q$: $Q_{\\\\theta_q} (s,a)$. DQN designs a network $Q_{\\\\theta_q}$ that is asynchronously updated by minimizing the MSE:\\n\\\\begin{align}\\n \\\\mathcal{L}(\\\\theta_q) = \\\\mathbb{E}_{\\\\tau \\\\sim \\\\pi}\\\\Big[Q_{\\\\theta_q}(s_t,a_t)-(r(s_t,a_t) + \\\\gamma Q_{\\\\theta_q}(s'_{t},a'_{t}))\\\\Big]^2 \\\\label{dqnloss}\\n\\\\end{align}\\nwhere $\\\\tau$ is the sampled trajectory containing $(s,a,s',r(s,a))$. In particular, $s'_t$ and $a'_t$ come from the behavior policy $\\\\pi_b$ while $s,a$ comes from the target policy $\\\\pi$.\\nIt is worth mentioning that the value function $V_\\\\pi(s)$ represents the expected return. $V_\\\\pi(s)$ is used to evaluate the goodness of the state while $Q_\\\\pi(s_t,a_t)$ is used to evaluate the action. $V_\\\\pi(s)$ can be defined as\\n\\\\begin{align}\\n V_\\\\pi(s) = \\\\mathbb{E}_{\\\\tau\\\\sim \\\\pi}\\\\bigg[\\\\sum_{t=0}^T\\\\gamma^t r(s,a)|s_0=s\\\\bigg].\\n\\\\end{align}\\n$V_\\\\pi(\\\\cdot)$ and $Q_\\\\pi(\\\\cdot)$ have the following relationship:\\n\\\\begin{align}\\n V_\\\\pi(s) = \\\\mathbb{E}_{a\\\\sim\\\\pi}[Q_\\\\pi(s,a)].\\n\\\\end{align}\\nThe value function is updated using the following rule with the Temporal Difference (TD) method,\\n\\\\begin{align}\\n V_\\\\pi(s_t) \\\\leftarrow V_\\\\pi(s_t) + \\\\alpha[\\\\underbrace{r(s'_t,a'_t) + \\\\gamma V_\\\\pi(s'_t) - V_\\\\pi(s_t)}_{\\\\text{TD-error}}]\\n\\\\end{align}\\nwhere $\\\\alpha$ is a constant.\\n\\\\textbf{Policy gradient}~ is an on-policy policy-based method which can handle high-dimensional or continuous actions which cannot be easily handled by Q-learning. Policy gradient\\naims to find the parameter $\\\\theta$ of $\\\\pi_{\\\\theta}$ to maximize the accumulated reward. To this end, it maximizes the expected return from the start state:\\n\\\\begin{align}\\n J(\\\\pi_\\\\theta) = \\\\mathbb{E}_{\\\\tau \\\\sim \\\\pi_{\\\\theta}}[r(\\\\tau)] = \\\\int\\\\pi_{\\\\theta}(\\\\tau) r(\\\\tau)d\\\\tau\\n\\\\end{align}\\nwhere $\\\\pi_{\\\\theta}(\\\\tau)$ is the probability of the occurrence of $\\\\tau$. Policy gradient learns the parameter $\\\\theta$ by the gradient $\\\\nabla_\\\\theta J(\\\\pi_\\\\theta)$ as defined below:\\n\\\\begin{align}\\n \\\\nabla_\\\\theta J(\\\\pi_\\\\theta) = \\\\int\\\\pi_{\\\\theta}(\\\\tau) r(\\\\tau)d\\\\tau &= \\\\int\\\\pi_{\\\\theta}(\\\\tau) \\\\nabla_{\\\\theta}\\\\log \\\\pi_{\\\\theta}(\\\\tau)r(\\\\tau)d\\\\tau \\\\notag \\\\\\\\\\n & =\\\\mathbb{E}_{\\\\tau \\\\sim d_{\\\\pi_\\\\theta}}[\\\\sum_{t=1}^Tr(s_t,a_t)\\\\sum_{t=1}^T\\\\nabla_{\\\\theta} \\\\log \\\\pi_{\\\\theta}(s_t,a_t)].\\n\\\\end{align}\\nThe above derivations contain the following substitution,\\n\\\\begin{align}\\n \\\\pi_{\\\\theta}(\\\\tau) = p(s_1)\\\\prod_{t=1}^T \\\\pi_{\\\\theta}(s_t,a_t)p(s_{t+1}|s_t,a_t)\\n\\\\end{align}\\nwhere $p(\\\\cdot)$ are independent from the policy parameter $\\\\theta$, which is omitted during the derivation.\\nMonte-Carlo sampling has been used by previous policy gradient algorithm (e.g,. REINFORCE) for $\\\\tau\\\\sim d_{\\\\pi_{\\\\theta}}$.\\n\\\\textbf{Actor-critic networks} combine the advantages from Q-learning and policy gradient. They can be either on-policy~ or off-policy~. An actor-critic network consists of two components: i) \\\\textit{an actor}, which optimizes the policy $\\\\pi_\\\\theta$ under the guidance of $\\\\nabla_\\\\theta J(\\\\pi_{\\\\theta})$; and ii) \\\\textit{a critic}, which evaluates the learned policy $\\\\pi_\\\\theta$ by using $Q_{\\\\theta_q} (s,a)$. The overall gradient is represented as follows:\\n\\\\begin{align}\\n \\\\mathbb{E}_{s \\\\sim d_{\\\\pi_\\\\theta}}[Q_{\\\\theta_q}(s,a)\\\\nabla_{\\\\theta} \\\\log \\\\pi_{\\\\theta}(s,a)].\\n\\\\end{align}\\nWhen dealing with off-policy learning, the value function for $\\\\pi_\\\\theta (a|s)$ can be further determined by deterministic policy gradient (DPG) as shown below:\\n\\\\begin{align}\\n \\\\mathbb{E}_{s \\\\sim d_{\\\\pi_\\\\theta}}[\\\\nabla_a Q_{\\\\theta_q}(s,a)|_{a=\\\\pi_\\\\theta(s)}\\\\nabla_{\\\\theta} \\\\pi_{\\\\theta}(s,a)].\\n\\\\end{align}\\nWhile traditional policy gradient calculates the integral for both the state space $\\\\mathcal{S}$ and the action space $\\\\mathcal{A}$, DPG only requires computing the integral to the state space $\\\\mathcal{S}$.\\nGiven a state $s\\\\in\\\\mathcal{S}$, there will be only one corresponding action $a\\\\in\\\\mathcal{A}:\\\\mu_\\\\theta(s) = a$ using DPG. Specifically, deep Deterministic Policy Gradients (DDPG) is an algorithm that combines techniques from DQN and DPG. DDPG contains four different neural networks: Q Network $Q$, policy network, target Q network $Q^{\\\\mathit{tar}}$, and target policy network. It uses the target network for both the Q Network $Q$ and policy network $\\\\mu$ to ensure stability during training. Assume $\\\\theta_q, \\\\theta_{\\\\pi}, \\\\theta_{q'}$ and $\\\\theta_{\\\\pi'}$ are parameters of the above networks; then DDPG soft-updates the parameters for the target network~:\\n\\\\begin{align}\\n \\\\text{Actor: }\\\\theta_{\\\\pi'} \\\\leftarrow \\\\alpha \\\\theta_\\\\pi + (1-\\\\alpha)\\\\theta_{\\\\pi'} \\\\text{ Critic: }\\\\theta_{q'} \\\\leftarrow \\\\alpha \\\\theta_q + (1-\\\\alpha)\\\\theta_{q'} \\n\\\\end{align}\",\n \"id\": \"e05ae3e8-b24d-431b-b181-250a9bf1fb50\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"Preliminaries of Deep Reinforcement Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Preliminaries of Deep Reinforcement Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"DRL is normally formulated as a Markov Decision Process (MDP).\\nGiven a set of users $\\\\mathcal{U} = \\\\{u, u_1, u_2, u_3, ...\\\\}$, a set of items $\\\\mathcal{I} = \\\\{i, i_1, i_2, i_3, ...\\\\}$, the system first recommends item $i$ to user $u$ and then gets feedback $f_i^u$.\\nThe system aims to incorporate the feedback to improve future recommendations and needs to determine an optimal policy $\\\\pi^*$ regarding which item to recommend to the user to achieve positive feedback.\\nThe MDP modelling of the problem treats the user as the environment and the system as the agent. The key components of the MDP in DRL-based RS include the following:\\n\\\\begin{itemize}\\n \\\\item State $\\\\mathcal{S}$: A state $S_t\\\\in\\\\mathcal{S}$ is determined by both users' information and the recent $l$ items in which the user was interested before time $t$.\\n \\\\item Action $\\\\mathcal{A}$: An action $a_t \\\\in\\\\mathcal{A}$ represents users' dynamic preference at time $t$ as predicted by the agent. $\\\\mathcal{A}$ represents the whole set of (potentially millions of) candidate items.\\n \\\\item Transition Probability $\\\\mathcal{P}$: The transition probability $p(s_{t+1}|s_t,a_t)$ is defined as the probability of state transition from $s_t$ to $s_{t+1}$ when action $a_t$ is executed by the recommendation agent. In a recommender system, the transition probability refers to users' behavior probability. $\\\\mathcal{P}$ is only used in model-based methods.\\n \\\\item Reward $\\\\mathcal{R}$: Once the agent chooses a suitable action $a_t$ based on the current state $S_t$ at time $t$, the user will receive the item recommended by the agent. Users' feedback on the recommended item accounts for the reward $r(S_t,a_t)$. The feedback is used to improve the policy $\\\\pi$ learned by the recommendation agent.\\n \\\\item Discount Factor $\\\\gamma$: The discount factor $\\\\gamma \\\\in [0,1]$ is used to balance between future and immediate rewards---the agent focuses only on the immediate reward when $\\\\gamma=0$ and takes into account all the (immediate and future) rewards otherwise.\\n\\\\end{itemize}\\nThe DRL-based recommendation problem can be defined by using MDP as follows.\\n\\\\textit{Given the historical MDP, i.e., $(\\\\mathcal{S},\\\\mathcal{A},\\\\mathcal{P}, \\\\mathcal{R},\\\\gamma)$, the goal is to find a set of recommendation polices ($\\\\{\\\\pi\\\\}: \\\\mathcal{S}\\\\to \\\\mathcal{A}$) that maximizes the cumulative reward during interaction with users.}\\n\\\\begin{dfn*}\\nGiven an environment that contains all items $\\\\mathcal{I}$, when user $u\\\\in\\\\mathcal{U}$ interacts with the system, an initial state $s$ is sampled from the environment which contains a list of candidate items and users' historical data. The DRL agent needs to work out a recommendation policy $\\\\pi$ based on the state $s$ and produces the corresponding recommended item list $a$.\\nThe user will provide feedback on the list which is normally represented as click or not click. The DRL agent will then utilize the feedback to improve the recommendation policy and move to the next interaction episode.\\n\\\\end{dfn*}\\n\\\\begin{figure}[ht]\\n \\\\centering\\n \\\\begin{subfigure}[b]{0.68\\\\textwidth}\\n \\\\centering\\n \\\\includegraphics[width=\\\\textwidth]{dlrs.pdf}\\n \\\\caption{Deep learning based recommender systems}\\n \\\\end{subfigure}\\n \\\\begin{subfigure}[b]{0.65\\\\textwidth}\\n \\\\centering\\n \\\\includegraphics[width=\\\\textwidth]{drlrs.pdf}\\n \\\\caption{Deep reinforcement learning based recommender systems}\\n \\\\end{subfigure}\\n \\\\caption{Difference between deep learning based RS and DRL-based RS. Deep learning based RSs may only update the recommendation policy during the training stage. They often require re-training, which is computationally inefficient, when users' interests change significantly. DRL-based RS will update the recommendation policy time over time as new rewards are received.}\\n \\\\label{fig:catnew}\\n\\\\end{figure}\",\n \"id\": \"467d3d60-cfac-40c2-8a3f-6dbe98149287\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a6f31a50-0b05-46c3-ac49-a2a06aea3f6f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Background\"\n ],\n [\n \"subsection\",\n \"DRL meets RS: Problem Formulation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"DRL meets RS: Problem Formulation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"DRL-based RS has some unique challenges such as state construction, reward estimation and environment simulation. We categorize the existing work of DRL-based recommendation into model-based and model-free methods (the taxonomy is shown in~\\\\Cref{fig:tax}).\",\n \"id\": \"696bfdf7-6191-445e-82a8-ae4df2378853\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ]\n ],\n \"subsections\": [\n \"4ebfecfc-b15f-4ce1-abac-7b121665be4a\",\n \"74b47d11-d3b5-407c-b802-84dc050162e4\",\n \"c30b91c8-48bc-4d92-ba66-c418e39b9327\"\n ],\n \"title\": \"Deep Reinforcement Learning in Recommender Systems\"\n },\n {\n \"cite_extract_rate\": 0.42857142857142805,\n \"cites\": [\n 5189,\n 5180,\n 5195\n ],\n \"content\": \"Model-based methods assume an expected reward or action available for the next step to help the agent update the policy.\\n\\\\begin{table}[ht]\\n\\\\caption{List of publications in model-based DRL-based RS}\\n\\\\begin{tabular}{ccc}\\n\\\\hline\\nMethod & Work \\\\\\\\\\\\hline\\nValue-based & \\\\\\\\\\nPolicy-based & \\\\\\\\\\nHybrid & \\n\\\\\\\\\\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"4ebfecfc-b15f-4ce1-abac-7b121665be4a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"696bfdf7-6191-445e-82a8-ae4df2378853\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-based Deep Reinforcement Learning based Methods\"\n ]\n ],\n \"subsections\": [\n \"8ceb9826-a858-407a-9b78-7a83d0ea8871\"\n ],\n \"title\": \"Model-based Deep Reinforcement Learning based Methods\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 7217\n ],\n \"content\": \"IRecGAN~ is a model-based method that adopts generative adversarial training to improve the robustness of policy learning. It can reduce the cost of interaction for RS by using offline data instead of the simulated environment.\\nIRecGAN employs a generative adversarial network~ to generate user data based on the offline dataset.\\nIt trains a recommendation agent using a policy gradient-based DRL method called REINFORCE. The agent aims to learn a policy based on the following gradient,\\n\\\\begin{align}\\n \\\\mathbb{E}_{\\\\tau\\\\sim\\\\{g,data\\\\}}\\\\big[\\\\sum_{t=0}^T \\\\sum_{t'=t}^T\\\\gamma^{t'-t} q_D(\\\\tau_{0:t}^n)r_t \\\\nabla_{\\\\theta_a}(c_t \\\\in \\\\pi_{\\\\theta_a}(s_t)\\\\big],q_D(\\\\tau_{0:t}^n)= \\\\frac{1}{N}\\\\sum_{n=1}^N D(\\\\tau_{0:T}^n),\\\\tau_{0:T}^n\\\\in MC^{\\\\mathcal{U}}(N)\\n\\\\end{align}\\nwhere the $\\\\mathit{MC}^{\\\\mathcal{U}}(N)$ represents the sampled $N$ sequences from the interaction between $\\\\mathcal{U}$ and the agent using the Monte-Carlo tree search algorithm, $D$ is the discriminator, $T$ is the length of $\\\\tau$, $g$ represents the offline data, and \\\\textit{data} represents the generated data. \\n propose NRSS for personalized music recommendation. NRSS uses wireless sensing data to learn users' current preferences. NRSS considers three different types of feedback: score, option, and wireless sensing data. Because multiple factors are considered as the reward, NRSS designs a reward model which consists of users' preference reward $r_p$ and a novel transition reward $r_{\\\\textit{trans}}$ which are parameterized by $\\\\theta_{r_p}$ and $\\\\theta_{r_{\\\\textit{trans}}}$. The goal for NRSS is to find the optimal parameters $\\\\theta_{r_p}$ and $\\\\theta_{r_{\\\\textit{trans}}}$ by using the Monte-Carlo tree search thus improving recommendation performance.\\nHowever, wireless sensing feedback lacks generalization ability as it is only available for certain tasks or scenarios, making it hard to determine dynamic user interest.\",\n \"id\": \"8ceb9826-a858-407a-9b78-7a83d0ea8871\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"4ebfecfc-b15f-4ce1-abac-7b121665be4a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-based Deep Reinforcement Learning based Methods\"\n ],\n [\n \"paragraph\",\n \"Policy-based methods\"\n ]\n ],\n \"subsections\": [\n \"bb6b42ee-6de2-432f-ab2a-fc4b096f3199\",\n \"68abbde2-1fcc-48e8-a5f8-eccafeb14c6f\",\n \"326848cf-b30c-48e3-8c0b-494e313de16b\"\n ],\n \"title\": \"Policy-based methods\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5195,\n 5180\n ],\n \"content\": \"Prior to Q-learning, value iteration is a more traditional value-based reinforcement learning algorithm that focuses on the iteration of the value function. Gradient Value Iteration (GVI)~ is proposed to improve the traditional value iteration algorithm by utilizing the transition probability and a multi-agent setting to predict chronological author collaborations. It introduces a new parameter named `status' to reflect the amount of knowledge that the agent needs to learn from this state. The policy is updated only when the distance between the new status and the old status is lower than a pre-defined threshold. However, value iteration requires the transition probability, which is hard to obtain in most cases. Hence, Q-learning and its variants are widely used in DRL-based RS. Cascading DQN (CDQN) with a generative user model~ is proposed to deal with the environment with unknown reward and environment dynamics. The generative user model adopts GANs to generate a user model based on an offline dataset. Different from previous work, it will generate the reward function for each user to explain the users' behavior. The user model can be written as,\\n\\\\begin{align}\\n \\\\argmax_{\\\\phi\\\\in\\\\triangle^{k-1}} \\\\mathbb{E}_{\\\\phi}[r(s_t,a_t)]-R(\\\\phi)/\\\\eta\\n\\\\end{align}\\nwhere $\\\\triangle^{k-1}$ is the probability simplex, $R(\\\\phi)$ is the regularization term for exploration and $\\\\eta$ is a constant.\\nPseudo Dyna-Q (PDQ)~ points out that Monte-Carlo tree search may lead to an extremely large action space and an unbounded importance weight of training samples. Hence, a world model is proposed to reduce the instability of convergence and high computation cost for interacting with users by imitating the offline dataset. With the world model, the agent will interact with the learned world model instead of the environment to improve the sample efficiency and convergence stability. The world model learning process introduced in PDQ can be described as finding the parameter $\\\\theta_M$,\\n\\\\begin{align}\\n \\\\argmin_{\\\\theta_M}\\\\mathbb{E}_{\\\\xi\\\\in P_{\\\\xi}^\\\\pi}[\\\\sum_{t}^{T-1}\\\\gamma^t\\\\prod_{j=0}^t\\\\frac{\\\\pi(s_j,a_j)}{\\\\pi_b(s_j,a_j)}\\\\Delta_t(\\\\theta_M)]\\n\\\\end{align}\\nwhere $\\\\xi$ is generated by the logged policy $\\\\pi_b$, $\\\\prod_{j=0}^t\\\\frac{\\\\pi(s_j,a_j)}{\\\\pi_b(s_j,a_j)}$ is the ratio used for importance sampling and $\\\\Delta$ is the difference between the reward in the world model and real reward. \\nFurthermore, GoalRec~ designs a disentangled universal value function to be integrated with the world model to help the agent deal with different recommendation tasks. The universal value function is defined as\\n\\\\begin{align}\\n V_{\\\\pi} (s)= \\\\mathbb{E}[\\\\sum_{t=0}^\\\\infty r(s_t,a_t)\\\\prod_{k=0}^t\\\\gamma s_k|s_0=s].\\n\\\\end{align}\\nMoreover, GoalRec introduces a new variable goal $g\\\\in G$ used to represent users' future trajectory and measurement $m\\\\in M$. $m$ is an indicator that reflects users' response to the given future trajectory based on historical behaviors. Based on that, the optimal action will be selected based on\\n\\\\begin{align}\\n a^* = \\\\max_a U(M(s,a),g)\\n\\\\end{align}\\nwith a customized liner function $U(\\\\cdot)$.\\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=\\\\linewidth]{GoalRec_wb.pdf}\\n \\\\caption{Left is the general structure of model-free methods. Right is the structure for GoalRec which is a model-based method. A sample trajectory is used to demonstrate the difference between them~. }\\n \\\\label{fig:worldmodel}\\n\\\\end{figure}\",\n \"id\": \"bb6b42ee-6de2-432f-ab2a-fc4b096f3199\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"8ceb9826-a858-407a-9b78-7a83d0ea8871\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-based Deep Reinforcement Learning based Methods\"\n ],\n [\n \"paragraph\",\n \"Policy-based methods\"\n ],\n [\n \"paragraph\",\n \"Value-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Value-based methods\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 5189\n ],\n \"content\": \"Hybrid method can be recognized as a midpoint between value-based and policy gradient-based methods. DeepChain~ uses the multi-agent setting to relieve the sub-optimality problem. The sub-optimality problem is caused by the \\\\textit{one for all} setting that optimizes one policy for all users. Hence, DeepChain designs a multi-agent setting that adopts several agents to learn consecutive scenarios and jointly optimizes multiple recommendation policies. The main training algorithm used is DDPG. To this end, users' actions can be formulated in a model-based form as follows:\\n\\\\begin{align}\\n \\\\sum_{m,d}[p_m^s (s_t,a_t)\\\\gamma Q_{\\\\theta}(s'_{t},\\\\pi_m(s'_{t})) + p_m^c(s_t,a_t)(r_t+\\\\gamma Q_{\\\\theta}(s'_{t},\\\\pi_d(s'_{t}))) + p_m^l (s_t,a_t)r_t]1_m\\n\\\\end{align}\\nwhere $m$ represents the number of actor networks, $c,l,s$ represent the three different scenarios, $1_m$ is used to control the activation of two actors and $(m,d)\\\\in\\\\{(1,2),(2,1)\\\\}$.\",\n \"id\": \"68abbde2-1fcc-48e8-a5f8-eccafeb14c6f\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"8ceb9826-a858-407a-9b78-7a83d0ea8871\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-based Deep Reinforcement Learning based Methods\"\n ],\n [\n \"paragraph\",\n \"Policy-based methods\"\n ],\n [\n \"paragraph\",\n \"Hybrid methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hybrid methods\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5189,\n 5180,\n 5195\n ],\n \"content\": \"Model-based methods aim to learn a model or representation to represent the whole environment so that the agent can plan ahead and receive better sample efficiency.\\nThe drawback of such a method is that the ground-truth representation of the environment is unavailable in recommendation scenarios as it dynamically changes, leading to a biased representation. Moreover, model-based methods use the transition probability function $\\\\mathcal{P}$ to estimate the optimal policy. As mentioned, the transition probability function is normally equivalent to users' behavior probability which is hard to determine in a recommender system. Hence, existing works~ approximate $\\\\mathcal{P}$ using a neural network or embedding it into the world model. design a probability network to estimate $\\\\mathcal{P}$ while uses a GAN to generate user behavior where $\\\\mathcal{P}$ is embedded in the latent space. Different from them,~ relies on the world model to predict users' next behavior and feed it into the policy learning process.\\nThe challenges of model-based DRL are not widely used in RS and can be summarized into the following facets:\\n\\\\begin{itemize}\\n\\\\item $\\\\mathcal{P}$ is hard to determine in real-world recommender systems. \\n\\\\item If approximation is used to estimate $\\\\mathcal{P}$, the overall model complexity will substantially increase as it requires approximating two different functions $\\\\mathcal{P}$ and the recommendation policy $\\\\pi$ by using a large amount of user behavior data. \\n\\\\item World model-based methods require periodic re-training to ensure the model can reflect user interests in time which increases the computation cost. \\n\\\\end{itemize}\\n\\\\begin{figure}\\n \\\\centering\\n \\\\begin{subfigure}{0.48\\\\linewidth}\\n \\\\includegraphics[width=\\\\linewidth]{DQN_wb.pdf}\\n \\\\caption{DQN}\\n \\\\end{subfigure}\\n \\\\begin{subfigure}{0.48\\\\linewidth}\\n \\\\includegraphics[width=\\\\linewidth]{DDPG_wb.pdf}\\n \\\\caption{DDPG}\\n \\\\end{subfigure}\\n \\\\caption{The typical structure of DQN and DDPG}\\n \\\\label{fig:ddpg_fig}\\n\\\\end{figure}\",\n \"id\": \"326848cf-b30c-48e3-8c0b-494e313de16b\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"8ceb9826-a858-407a-9b78-7a83d0ea8871\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-based Deep Reinforcement Learning based Methods\"\n ],\n [\n \"paragraph\",\n \"Policy-based methods\"\n ],\n [\n \"paragraph\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 0.43181818181818105,\n \"cites\": [\n 5190,\n 3604,\n 8884,\n 5165,\n 6153,\n 3616,\n 1389,\n 5184,\n 5191,\n 5177,\n 5169,\n 5199,\n 6154,\n 5181,\n 6152,\n 5172,\n 5194,\n 5170,\n 5185\n ],\n \"content\": \"Compared with model-based methods, model-free methods are relatively well-studied. Different from model-based methods, $\\\\mathcal{P}$ is unknown and not required in model-free methods. Model-based methods enable the agent to learn from previous experiences. In this subsection, we categorize model-free based DRL in RS into three parts: value-based, policy-based and hybrid methods. \\n\\\\begin{table}[ht]\\n\\\\caption{List of reviewed publications in Model-free DRL-based RS}\\n\\\\begin{tabular}{ccc}\\n\\\\hline\\nTasks & Note & Work \\\\\\\\\\\\hline\\n\\\\multirow{4}{*}{Value-based} & Vanilla DQN and its extensions & \\\\\\\\ \\\\\\n& DQN with state/action space optimization & \\\\\\\\\\n& DQN with graph/image input & \\\\\\\\\\n& DQN for joint learning & \\\\\\\\\\\\hline\\n\\\\multirow{3}{*}{Policy-based} & Vanilla REINFORCE & \\\\\\\\\\n& REINFORCE uses graph structure/input & \\\\\\\\\\n& Non-REINFORCE based & \\\\\\\\ \\\\hline\\n\\\\multirow{2}{*}{Hybrid} & Vanilla DDPG & \\\\\\\\\\n& with Knowledge Graph & \\\\\\\\\\n\\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"74b47d11-d3b5-407c-b802-84dc050162e4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 44,\n \"parent_id\": \"696bfdf7-6191-445e-82a8-ae4df2378853\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-free deep reinforcement learning based methods\"\n ]\n ],\n \"subsections\": [\n \"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe\"\n ],\n \"title\": \"Model-free deep reinforcement learning based methods\"\n },\n {\n \"cite_extract_rate\": 0.41176470588235203,\n \"cites\": [\n 5181,\n 5191,\n 5170,\n 6155,\n 5185,\n 5199,\n 1408\n ],\n \"content\": \"As mentioned, Deep Q-learning and its variants are typical value-based DRL methods widely used in DRL-based RS. DRN~ is the first work utilizing Deep Q-Networks (DQN) in RS. It adopts Double DQN (DDQN)~ to build a user profile and designs an activeness score to reflect how frequently a user returns after one recommendation plus users' action (click or not) as the reward. DRN provides a new approach to integrating DRL into RS when dealing with a dynamic environment. The key objective function can be found as follows,\\n\\\\begin{align}\\n \\\\mathbb{E}[r_{t+1} + \\\\gamma Q_{\\\\theta_t'}(s_{t+1}, \\\\argmax_{a'} Q_{\\\\theta_t}(s_t,a'))]\\n\\\\end{align}\\nwhere $a'$ is the action that gives the maximum future reward according to $\\\\theta_t$, $\\\\theta_t$ and $\\\\theta_t'$ are different parameters for two different DQNs. \\n points out that negative feedback will also affect recommendation performance which DRN does not consider. Moreover, positive feedback is sparse due to the large number of candidate items in RS. Only using positive feedback would lead to convergence problems. Hence, DEERS is proposed to consider both positive and negative feedback simultaneously by using DQN. Gated Recurrent Units (GRU) are employed to capture users' preferences for both a positive state $s^+$ and negative state $s^-$ and the final objective function can be computed as:\\n\\\\begin{align}\\n \\\\mathbb{E}[r_{t+1} + \\\\gamma \\\\max_{a_{t+1}}Q_{\\\\theta_q}(s^+_{t+1},s^-_{t+1},a_{t+1})|s^+_t,s^-_t,a_t].\\n\\\\end{align}\\n introduces attention mechanisms into the DQN to leverage social influence among users. To be specific, a social impact representation $U_v$ is introduced into the state representation. Matrix factorization is adopted to determine \\nsimilarity among\\nusers and hence present the social influence. Social attention is introduced to distill the final state representation.\\nIn addition, a few studies focus on user profiling to improve recommendation performance~. claims that user feedback contains useful information in the previous feedback even when the user does not like the recommended items. Some existing studies focus on final feedback which ignore the importance from earlier steps to later ones. Hence, user-specific DQN (UQDN) is proposed to consider multi-step feedback from users. It employs Matrix Factorization to generate user-specific latent state spaces. The newly defined objective function with the user-specific latent state space can be represented as\\n\\\\begin{align}\\n \\\\mathbb{E}[r_{t+1} + \\\\gamma \\\\max_{a_{t+1}}\\\\overline{Q}_{\\\\theta_q}(s_{t+1},a_{t+1}) + \\\\overline{\\\\textbf{b}}_u - Q_{\\\\theta_q} (s_{t+1},a_{t+1})]\\n\\\\end{align}\\nwhere $\\\\overline{\\\\textbf{b}}_u$ is the learned user latent representation.\\n also points out that most studies do not consider users' long-term engagement in the state representation as they focus on the immediate reward. FeedRec is proposed that combines both instant feedback and delayed feedback into the model to represent the long-term reward and optimize the long-term engagement by using DQN. To be specific, time-LSTM is employed to track users' hierarchical behavior over time to represent the delayed feedback which contains three different operations: $h_{\\\\mathit{skip}},h_{\\\\mathit{choose}},h_{\\\\mathit{order}}$. The state space is the concatenation of those operations and users' latent representation. Differently, focuses on the user privacy issue in recommender systems. Deep user profile perturbation (DUPP) is proposed to add perturbation into the user profile by using DQN during the recommendation process. Specifically, DUPP adds a perturbation vector into users' clicked items as well as the state space, which contains users' previous behavior.\\nDistinct from previous studies which focus on optimizing user profiles or state spaces, some studies aim to optimize the action space formed by interactions with items. In the situation of basket recommendation, the user is suggested multiple items as a bundle, which is called a recommendation slate. It leads to combinatorially large action spaces making it intractable for DQN based recommendation models.\\nSlateQ~ is proposed to decompose slate Q-value to estimate a long-term value for individual items, and it is represented as,\\n\\\\begin{align}\\n Q_{\\\\theta_q}(s_t,a_t) = \\\\sum_{i \\\\in a_t}p(i|s_t,a_t)\\\\overline{Q}_{\\\\theta_q}(s_t,i)\\n\\\\end{align}\\nwhere $\\\\overline{Q}_{\\\\theta}(s,i)$ is the decomposed Q-value for item $i$. The decomposed Q-value will be updated by the following rule which is similar to traditional DQN,\\n\\\\begin{align}\\n \\\\overline{Q}_{\\\\theta_q}(s_t,i) \\\\leftarrow \\\\alpha \\\\bigg(r_t +\\\\gamma \\\\sum_{j\\\\in a_{t+1}}p(j|s_{t+1},a_{t+1}) \\\\overline{Q}_{\\\\theta_q}(s_{t+1},j)\\\\bigg) + (1-\\\\alpha)\\\\overline{Q}_{\\\\theta_q}(s_t,i).\\n\\\\end{align}\\nDifferent from other mode-free methods, Slate-Q assumes that the transition probability $p(i|s_t,a_t)$ is known.\\nVanilla DQN methods may not have sufficient knowledge to handle complex data such as images and graphs. firstly models users' click behavior as an embedding matrix in the latent space to include the skip behaviors of sequence patterns for sequential recommendation. Based on that, propose DRCGR, which adopts CNN and GAN into DQN to help the agent to better understand high-dimensional data, e.g., a matrix. Two different convolution kernels are used to capture users' positive feedback. In the meantime, DRCGR uses GANs to learn a negative feedback representation to improve robustness. \\nAnother typical data format is the graph, which is widely used in RS, including knowledge graphs.\\n propose GCQN which adopts Graph Convolutional Networks (GCN)~ into the DQN which constructs the state and action space as a graph-aware representation. Differently, GCQN introduces the attention aggregator: $\\\\sum_{w\\\\in \\\\mathcal{N}(i)}\\\\alpha_{iu}e_u$ which demonstrates better performance than the mean-aggregator and pooling-aggregator. For item $i$, the graph-aware representation can be represented as,\\n\\\\begin{align}\\n \\\\sigma\\\\bigg(W_{fc}[e_i\\\\oplus\\\\sum_{w\\\\in \\\\mathcal{N}(i)}\\\\alpha_{iu}e_u + b_{fc}]\\\\bigg)\\n\\\\end{align}\\nwhere $W_{fc},b_{fc}$ are the parameters for the fully-connected layer, $e_u$ is the embedding for user $u$ and $\\\\mathcal{N}(i)$ is the set of one-hot neighbours of item $i$ in graph $G(i)$. propose KGQR uses a similar strategy to transform the information into a knowledge graph which is fed into the GCN to generate the state representation. Notably, KGQR presents a different state representation generation method. For given node $i$, the neighbourhood representation with a $k$-hop neighborhood aggregator can be represented as,\\n\\\\begin{align}\\n e_i^k = \\\\sigma\\\\bigg(W_{k}\\\\frac{1}{|\\\\mathcal{N}(i)|}\\\\sum_{t\\\\in\\\\mathcal{N}(i)} e_{t}^{k-1} + B_k e_i^{k-1}\\\\bigg)\\n\\\\end{align}\\nwhere $\\\\mathcal{N}(i)$ is the set of neighboring nodes, $W_k, B_k$ are the parameter of the aggregator. Those neighbourhood representations will be fed into a GRU and the state representation will be generated. Another application domain for using graph data is job recommendation which requires considering multiple factors jointly such as salary, job description, job location etc. SRDQN~ constructs a probability graph to represent a candidate's skill set and employs a multiple-task DQN structure to process these different factors concurrently. \\nThere are some studies targeting recommendation and advertising simultaneously in e-commerce environments~. mentions when deploying RS into real-world platforms such as e-commerce scenarios, the expectation is to improve the profit of the system.\\nA new metric, Gross Merchandise Volume (GMV), is proposed to measure the profitability of the RS to provide a new view about evaluating RS in advertising. Different from GMV, separates recommendation and advertising as two different tasks and proposes the Rec/Ads\\nMixed display (RAM) framework. RAM designs two agents: a recommendation agent and an advertising agent, where each agent employs a CDQN to conduct the corresponding task. find that advertising and recommendation may harm each other and formulate a rec/ads trade-off. Their proposed solution, DEARS, contains two RNNs. Two RNNs are employed to capture user preferences toward recommendations and ads separately. Based on that, DQN is employed to take those two outputs as the input to construct the state and output the advertising.\",\n \"id\": \"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"74b47d11-d3b5-407c-b802-84dc050162e4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-free deep reinforcement learning based methods\"\n ],\n [\n \"paragraph\",\n \"Value based methods\"\n ]\n ],\n \"subsections\": [\n \"01225554-85b2-42e4-a661-722c3b5972d7\",\n \"cb915b6e-26a1-44b9-be2e-47ffa839e199\",\n \"47d0fb4b-4dca-4006-ba90-90530959ef38\"\n ],\n \"title\": \"Value based methods\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 5190,\n 5165,\n 6153,\n 5184,\n 5177,\n 1911,\n 6152,\n 6156\n ],\n \"content\": \"Policy-based DRL can be divided into two parts which are Constrained Policy Optimization (CPO) ~ and policy gradient.\\n uses CPO to identify the contradiction between text feedback and historical preferences. It provides a solution for using DRL in the situation where users' feedback is entirely different from previous feedback in RS. \\nPolicy gradient-based methods are the other stream in policy-based DRL methods for RS. These methods aims to optimize the policy $\\\\pi$ directly instead of estimating the Q-value like DQN. A well-known and widely used policy gradient method in RS is REINFORCE which uses the following rule for policy $\\\\pi_{\\\\theta_\\\\pi}$,\\n\\\\begin{align}\\n \\\\theta \\\\leftarrow \\\\theta + \\\\alpha \\\\mathbb{E}_{\\\\tau\\\\sim d_{\\\\pi_{\\\\theta_\\\\pi}}}\\\\bigg[\\\\sum_{t=1}^Tr(s_t^i,a_t^i)\\\\sum_{t=1}^T\\\\nabla_{\\\\theta_\\\\pi} \\\\log \\\\pi_{\\\\theta_\\\\pi}(s_t^i,a_t^i)\\\\bigg]\\n\\\\end{align}\\nwhere $i$ is sampled trajectories from $\\\\pi_{\\\\theta}(a_t|s_t)$. propose Policy Gradient for Contextual Recommendation (PGCR), which adopts REINFORCE and considers contextual information. PGCR assumes that the policy follows the multinoulli distribution, in which case the transition probability can be estimated easily through sampling from previously seen context.\\n incorporate CNNs and attention mechanisms in REINFORCE for explainable recommendation. Specifically, this work designs a coupled agent structure where one agent generates the explanation and the other makes recommendations based on the generated explanation.\\n increases the scalability of REINFORCE to ensure it can deal with the extremely large action space under recommendation scenarios. To be specific, it introduces a policy correction gradient estimator into REINFORCE to reduce the variance of each gradient by doing importance sampling. The new update rule\\nbecomes\\n\\\\begin{align}\\n \\\\theta_\\\\pi \\\\leftarrow \\\\theta_\\\\pi + \\\\alpha \\\\sum_{\\\\tau\\\\sim\\\\beta}\\\\bigg[\\\\sum_{t=1}^T\\\\frac{\\\\pi_{\\\\theta_\\\\pi}(s_t,a_t)}{\\\\pi_\\\\beta(s_t,a_t)}r(s_t^i,a_t^i)\\\\sum_{t=1}^T\\\\nabla_{\\\\theta_\\\\pi} \\\\log \\\\pi_{\\\\theta_\\\\pi}(s_t^i,a_t^i)\\\\bigg]\\n\\\\end{align}\\nwhere $\\\\pi_\\\\beta$ is the behavior policy trained by state-action pairs without the long-term reward and $\\\\pi_{\\\\theta}$ is trained based on the long-term reward only. It is worth mentioning that the vanilla REINFORCE algorithm is on-policy, and importance sampling will make REINFORCE behave like an off-policy method with the following gradient format,\\n\\\\begin{align}\\n \\\\mathbb{E}_{\\\\tau \\\\sim d_{\\\\pi_\\\\theta}}\\\\bigg[\\\\prod_{t'=1}^t\\\\frac{\\\\pi_{\\\\theta}(s_t,a_t)}{\\\\pi_{\\\\theta_s}(s_{t'},a_{t'})}\\\\sum_{t'=t}^Tr(s_t,a_t)\\\\sum_{t=1}^T\\\\nabla_{\\\\theta} \\\\log \\\\pi_{\\\\theta}(s_t,a_t)\\\\bigg]\\n\\\\end{align}\\nwhere $\\\\pi_{\\\\theta_s}$ is the sample policy parameter.\\n also finds that the REINFORCE method suffers from a high variance gradient problem and Pairwise Policy Gradient (PPG) is proposed. Different from policy correction, PPG uses Monte Carlo sampling to sample two different actions $a,b$ and compare the gradient to update $\\\\theta$,\\n\\\\begin{align}\\n \\\\mathbb{E}_{\\\\tau\\\\sim d_{\\\\pi_{\\\\theta_\\\\pi}}}\\\\bigg(\\\\sum_a\\\\sum_b(r(s,a)-r(s,b))\\\\sum_{t=1}^T(\\\\nabla_{\\\\theta_\\\\pi}\\\\log_{\\\\pi_{\\\\theta_\\\\pi}}(s_t,a_t)-\\\\nabla_{\\\\theta_\\\\pi}\\\\log_{\\\\pi_{\\\\theta_\\\\pi}}(s_t,b_t))\\\\bigg).\\n\\\\end{align}\\n extends the policy correction gradient estimator into a two-stage setting which are $p(s_t,a^p)$ and $q(s_t,a|a^p)$ and the policy can be written as\\n\\\\begin{align}\\n \\\\sum_{a^p}p(s_t,a^p)q(s_t,a|a^p).\\n\\\\end{align}\\nIn addition, weight capping and self-normalized importance sampling are used to further reduce the variance.\\nMoreover, a large state space and action space will cause sample inefficiency problems as REINFORCE relies on the current sampled trajectories $\\\\tau$. finds that the auxiliary loss can help improve the sample efficiency\\n~. Specifically, a linear projection is applied to the state $s_t$, the output is combined with action $a_t$ to calculate the auxiliary loss and appended into the final overall objective function for optimization.\\nAnother prototype of vanilla policy gradient in DRL-based RS is the policy network. designs a policy network to extract features and represent the relevant feedback that can help the agent make a decision. Similar to DQN, this work uses a neural network to approximate the Q-value and the policy directly without theoretical analysis.\\n extend the policy network by introducing spatio-temporal feature fusion to help the agent understand complex features. Specifically, it considers both the current number and the future number of vacant taxis on the route to recommend routes for taxis. introduces multi-modal data as new features to conduct vision-language recommendation by using historical data to train REINFORCE. ResNet and attention are used to encode vision and text information, respectively. Moreover, two rewards are introduced with a customized ratio $\\\\lambda$ to balance vision and text information. \\nKnowledge Graphs (KG) are widely used in RS to enrich side information, provide explainability and improve recommendation performance. Similar to DQN, vanilla REINFORCE cannot properly handle graph-like data. propose a method named Knowledge-guided Reinforcement Learning (KERL), which integrates knowledge graphs into the REINFORCE algorithm. To be specific, KERL adopts TransE~ to transfer the knowledge graph into a graph embedding and utilizes a multilayer perceptron (MLP) to predict future knowledge of user preferences. The state representation can be written as\\n\\\\begin{align}\\n h_t \\\\oplus \\\\mathit{TransE}(\\\\mathcal{G}) \\\\oplus \\\\mathit{MLP}(\\\\mathit{TransE}(\\\\mathcal{G}))\\n\\\\end{align}\\nwhere $h_t$ is the hidden representation from the GRU for sequential behavior and $\\\\mathcal{G}$ is the knowledge graph.\\nDifferent from KERL, propose Policy-Guided Path Reasoning (PGPR), which formulates the whole environment as a knowledge graph. The agent is trained to find the policy to find good items conditioned on the starting user in the KG by using REINFORCE. PGPR uses the tuple $(u,e_t,h_t)$ to represent the state instead of the graph embedding where $e_t$ is the entity the agent has reached at $t$ for user $u$ and $h_t$ is the previous action before $t$.\\nThe action in PGPR is defined as the prediction of all outgoing edges for $e_t$ based on $h_t$. \\n propose a knowledge graph policy network (KGPolicy) which puts the KG into the policy network and adopts REINFORCE to optimize it. In addition, KGPolicy uses negative sampling instead of stochastic sampling to overcome the false negative issue---sampled items behave differently during training and inference. Similar to GCQN, attention is also employed to establish the representation for its neighbors. \\nDue to the on-policy nature of REINFORCE, it is difficult to apply it to large-scale RS as the convergence speed will be a key issue. To relieve this, propose TPGR, which designs a tree-structured policy gradient method to handle the large discrete action space hierarchically. TPGR uses balanced hierarchical clustering to construct a clustering tree. Specifically, it splits a large-scale data into several levels and maintains multiple policy networks for each level to conduct the recommendation. The results are integrated at the final stage.\\nAs mentioned, policy gradient can be further extended to deterministic policy gradient (DPG) . propose Deterministic Policy Gradient with Full Backup\\nEstimation (DPG-FBE) to complete a sub-task of recommendation. DPG-FBE considers a search session MDP (SSMDP) that contains a limited number of samples, where the stochastic policy gradient method like REINFORCE cannot work well.\",\n \"id\": \"01225554-85b2-42e4-a661-722c3b5972d7\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-free deep reinforcement learning based methods\"\n ],\n [\n \"paragraph\",\n \"Value based methods\"\n ],\n [\n \"paragraph\",\n \"Policy-based methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Policy-based methods\"\n },\n {\n \"cite_extract_rate\": 0.6153846153846151,\n \"cites\": [\n 5172,\n 6154,\n 3604,\n 8884,\n 5194,\n 5170,\n 3616,\n 1389\n ],\n \"content\": \"The most common model-free hybrid method used would be the actor-critic algorithm where the critic network uses the DQN and the actor uses the policy gradient. The common algorithm used to train actor-critic is DDPG with the following objective function,\\n\\\\begin{align}\\n \\\\mathbb{E}[r_t+\\\\gamma Q_{\\\\theta_q'}(s_{t+1},\\\\mu_{\\\\theta_\\\\pi'}(s_{t+1})) - Q_{\\\\theta_q}(s_t,a_t)]\\n\\\\end{align}\\nwhere $\\\\theta_q,\\\\theta_q'$ is the parameter for Q-learning at time $t,t+1$ while $\\\\theta_\\\\pi'$ is the parameter for deterministic policy gradient at time $t+1$.\\n propose LIRD, which uses the vanilla actor-critic framework to conduct list-wise recommendations. In order to demonstrate the effectiveness of LIRD, a pre-trained user simulator is used to evaluate the effectiveness of LIRD where the transition probability is approximated using the cosine similarity for a given state-action pair $s_t,a_t$.\\n further extend LIRD into page-wise recommendation and proposed DeepPage. Similar to other previous work, GRU is employed to process the sequential pattern. Moreover, similar to DRCGR, DeepPage formulates the state as a page, then CNNs are employed to capture features and fed to the critic network. The final state representation is the concatenation of the sequential pattern and the page features. Additionally, there are a few studies focusing on different scenarios such as top-aware recommendation~, treatment recommendation~, allocating impressions~ etc. introduces a supervised learning module (SLC) as the indicator to identify the difference between the current policy and historical preferences. SLC will conduct the ranking process to ensure the recommendation policy will not be affected by the positional bias -- the item appearing on top receives more clicks. Similarly, also integrates the supervised learning paradigm into DRL but in a different way. An expert action $\\\\hat{a}_t$ is provided when the critic evaluates the policy and the update rule is slightly different than normal DQN,\\n\\\\begin{align}\\n \\\\theta_q \\\\leftarrow \\\\theta_q + \\\\alpha \\\\sum_{t}[ Q_{\\\\theta_q}(s_t,\\\\hat{a}_t)-r_t-\\\\gamma Q_{\\\\theta_{q'}}(s_t,\\\\mu_{\\\\theta_{\\\\pi'}}(s_t))]\\\\nabla_{\\\\theta_q}Q_{\\\\theta_q}(s_t,a_t).\\n\\\\end{align}\\nHowever, such a method is not universal as the acquisition of expert action is difficult and depends on the application domain.\\nSimilar to policy gradient and DQN, Knowledge Graphs (KG) are also used in actor-critic-based methods. propose KGRL to incorporate the substantial information of knowledge graphs to help the critic to better evaluate the generated policy. A knowledge graph is embedded into the critic network. Different from previous studies which use the KG as the environment or state representation, KGRL uses KG as a component in the critic, which can guide the actor to find a better recommendation policy by measuring the proximity from the optimal path. Specifically, a graph convolutional network is used to weight the graph and Dijkstra's algorithm is employed to find the optimal path for finally identifying the corresponding Q-value. claim that human's demonstration could improve path searching and propose ADAC. ADAC also searches for the optimal path in the KG but further adopts adversarial imitation learning and uses expert paths to facilitate the search process. propose MA-RDPG, which extends the standard actor-critic algorithm to deal with multiple scenarios by utilizing a multi-actor reinforcement learning setting. Specifically, two different actor-networks are initialized while only one critic network will make the final decision. Those two actor networks can communicate with each other to share information and approximate the global state. find that there are multiple factors can affect the selection of electric charging station. Hence, it uses a similar idea to recommend the electric vehicle charging station by considering current supply, future supply, and future demand. figure out that the communication mechanism in MA-RDPG will harm actors as they are dealing with independent modules, and there is no intersection. Hence, extend MA-RDPG into multi-agent settings which contain multiple pairs of actors and critics and remove the communication mechanism to ensure independence. \\nDifferent from , use `soft' actor-critic (SAC)~, which introduces a maximum entropy term $\\\\mathcal{H}(\\\\pi(s_t,\\\\phi_t))$ to actor-critic to improve exploration and stability with the stochastic policy $\\\\pi(s_t,\\\\phi_t)$. \\nSimilar to the multi-agent idea, use a hierarchical setting to help the agent learn multiple goals by setting multiple actors and critics. In comparison, hierarchical RL uses multiple actor-critic networks for the same task. It splits a recommendation task into two sub-tasks: discovering long-term behavior and capturing short-term behavior. The final recommendation policy is the combination of the optimal policies for the two sub-tasks. Similarly, use the hierarchical setting for integrated recommendation by using different sourced data. The objective is to work out the sub-polices for each source hierarchically and form the final recommendation policy afterward.\",\n \"id\": \"cb915b6e-26a1-44b9-be2e-47ffa839e199\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-free deep reinforcement learning based methods\"\n ],\n [\n \"paragraph\",\n \"Value based methods\"\n ],\n [\n \"paragraph\",\n \"Hybrid methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hybrid methods\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6157\n ],\n \"content\": \"In RS, model-free methods are generally more flexible than model-based methods as they do not require knowing the transition probability. We summarize the advantages and disadvantages of the three kinds of methods described under the model-free category. DQN is the first DRL method used in RS, which is suitable for small discrete action spaces. The problems with DQN in RS are: \\n\\\\begin{itemize}\\n \\\\item RS normally contains large and high-dimensional action spaces.\\n \\\\item The reward function is hard to determine which will lead to inaccurate value function approximation.\\n\\\\end{itemize} \\nSpecifically, the high dimensional action space in context of recommender systems is recognized as a major drawback of DQN~. The reason lies in the large number of the candidate items. Hence, DQN, as one of the most popular schemes, is not the best choice for RS in many situations. Moreover, some unique factors need to be considered when designing the reward function for RS such as social inference. It introduces extra parameters to the Q-network and hinders the convergence.\\nPolicy gradient does not require the reward function to estimate the value function. Instead, it estimates the policy directly. However, policy gradient is designed for continuous action spaces. More importantly, it will introduce high variance in the gradient. Actor-critic algorithms combine the advantages of DQN and policy gradient. Nonetheless, actor-critic will map the large discrete action space into a small continuous action space to ensure it is differentiable, which may cause potential information loss. Actor-critic uses DDPG and thus inherits disadvantages from DQN and DPG, including difficulty in determining the reward function and poor exploration ability.\",\n \"id\": \"47d0fb4b-4dca-4006-ba90-90530959ef38\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"6c84bc7f-b9cc-44a6-b8d6-5c59f102ecfe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Model-free deep reinforcement learning based methods\"\n ],\n [\n \"paragraph\",\n \"Value based methods\"\n ],\n [\n \"paragraph\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6158,\n 5192,\n 5201,\n 8984,\n 5204,\n 3610\n ],\n \"content\": \"\\\\label{sec:component}\\nThere are a few studies that use DRL in RS for goals other than improving recommendation performance or proposing new application domains. We split the literature based on the following components: environment, state representation, and reward function. Exisitng studies usually focus on optimizing one single component in the DRL setting (as illustrated in~\\\\Cref{fig:overview}).\\n\\\\begin{table}[h]\\n \\\\centering\\n \\\\caption{List of publications reviewed in this section}\\n \\\\begin{tabular}{c|c}\\n \\\\hline\\n Component & Work \\\\\\\\\\\\hline\\n Environment & \\\\\\\\\\n State & \\\\\\\\\\n Reward & \\\\\\\\\\n \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:component}\\n\\\\end{table}\\n\\\\\\\\\",\n \"id\": \"c30b91c8-48bc-4d92-ba66-c418e39b9327\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"696bfdf7-6191-445e-82a8-ae4df2378853\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Component Optimization in Deep Reinforcement Learning based RS\"\n ]\n ],\n \"subsections\": [\n \"3de2362c-6535-43b9-860b-ebc2cbce5821\",\n \"45229725-ce80-46d4-8644-8011bc19bde5\",\n \"0c8d5ce0-24e2-4a9b-aa04-14e44c18baf4\"\n ],\n \"title\": \"Component Optimization in Deep Reinforcement Learning based RS\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 6158,\n 5201,\n 8984,\n 1346,\n 5204,\n 3610\n ],\n \"content\": \"~\\\\\\\\\\nMany environments are available for evaluating deep reinforcement learning. Two popular ones are OpenAI gym-based environment and MuJoCo\\\\footnote{http://mujoco.org/}.\\nUnfortunately, there is no standardized simulation platform or benchmark specific to reinforcement learning based recommender systems. Existing work on DRL in RS is usually evaluated through offline datasets or via deployment in real applications. The drawback for evaluating offline datasets include:\\n\\\\begin{itemize}\\n \\\\item Different studies use different environment construction methods which leads to unfair comparison. For instance, some studies use the KG as the environment while some studies assume the environment is gym-like or design a simulator for specific tasks.\\n \\\\item With offline datasets, users' dynamic interests, and environment dynamics are hard to maintain. Deploying the method into a real application is difficult for academic research as it takes time and costs money. Hence, a standardized simulation environment is a desirable solution.\\n\\\\end{itemize}\\n\\\\begin{figure}\\n \\\\centering\\n \\\\includegraphics[width=\\\\linewidth]{cofounder.pdf}\\n \\\\caption{Left is the traditional MDP transition; Right is the POMDP which considers the environmental confounders such as social influence~.}\\n \\\\label{fig:cofounder}\\n\\\\end{figure}\\nThere are several different studies that provide standardized gym\\\\footnote{https://gym.openai.com/}-based simulation platforms for DRL-based RS research in different scenarios. RecSim~ is a configurable platform that supports sequential interaction between the system and users. RecSim contains three different tasks: interest evolution, interest exploration and long-term satisfaction. RecoGym~ provides an environment for recommendation and advertising. In addition, RecoGym also provides simulation of online experiments such as A/B-tests. However, RecSim and RecoGym are designed for bandit behavior which means users' interests will not change over time. \\nVirtualTB~ is proposed to relieve such problems. VirtualTB employs imitation learning to learn a user model to interact with the agent. GANs are employed to generate users' interests. Similar to VirtualTB, Recsimu~ uses a GAN to tackle the complex item distribution.\\nIn addition, PyRecGym~ accommodates standard benchmark datasets into a gym-based environment. MARS-Gym~ provides a benchmark framework for marketplace recommendation.~ suggests that existing simulation environments are biased because of biased logged data. Two common biases are discussed: popularity bias and positivity bias. To reduce the effect from those biases, SOFA introduces an Intermediate Bias Mitigation Step for debiasing purposes.\\nOne work discusses environment reconstruction by considering confounders.~ claims that users' interests may be affected by social networks which may introduce extra bias to the state and affect the decision-making process. A multi-agent setting is introduced to treat the environment as an agent which can partially relieve the hidden confounder effect. Specifically, a deconfounded environment reconstruction method DEMER is proposed. Different from previously mentioned methods, DEMER assumes the environment is partially observed and models the whole recommendation task as a Partially Observed MDP (POMDP). Different from an MDP, a POMDP contains one more component observation $o\\\\in\\\\mathcal{O}$ and the action $a_t$ is derived based on the observation $o_t$ instead of the state $s_t$ by $a_t = \\\\pi_a(o_t)$. DEMER assumes there is a confounder policy $\\\\pi_h$ for observation $o_h$ which is composed by $a_t$ and $o_t$: $a_h = \\\\pi_h(a_t,o_t)$. Moreover, another observation $o_b$ is introduced to observe the transition as well. $\\\\pi_b$ is the corresponding policy and $a_b = \\\\pi_b(o_b) = \\\\pi_b(o_t,a_t,a_h)$. DEMER uses generative adversarial imitation learning (GAIL) to imitate the policy $A,B$.\\nGiven trajectory $\\\\{o_t,o_h,o_b\\\\}$ for different policies $A$ and $B$, the objective function is defined as\\n\\\\begin{align}\\n & (\\\\pi_a,\\\\pi_b,\\\\pi_h) = \\\\argmin_{(\\\\pi_a,\\\\pi_b,\\\\pi_h)}\\\\mathbb{E}_{s\\\\sim \\\\tau }(L(s,a_t,a_b)) \\\\notag \\\\\\\\\\n & \\\\text{where }L(s,a_t,a_b) = \\\\mathbb{E}_{(\\\\pi_a,\\\\pi_b,\\\\pi_h)}[\\\\log D(s,a_t,a_b)]-\\\\lambda \\\\sum_{\\\\pi\\\\in\\\\{\\\\pi_a,\\\\pi_b,\\\\pi_h\\\\}}H(\\\\pi)\\n\\\\end{align}\\nwhere $L(\\\\cdot)$ is the loss function, $D(\\\\cdot)$ is a discriminator and $H(\\\\pi)$ is introduced in GAIL.\\n\\\\\\\\\",\n \"id\": \"3de2362c-6535-43b9-860b-ebc2cbce5821\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"c30b91c8-48bc-4d92-ba66-c418e39b9327\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Component Optimization in Deep Reinforcement Learning based RS\"\n ],\n [\n \"subsubsection\",\n \"Environment Simulation and Reconstruction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Environment Simulation and Reconstruction\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 5192\n ],\n \"content\": \"~\\\\\\\\\\nState representation is another component in DRL-based RS which exists in both model-based and model-free methods. find that the state representation in model-free methods would affect recommendation performance.\\nExisting studies usually directly use the embedding as the state representation. propose a supervised learning method to generate a better state representation by utilizing an attention mechanism and a pooling operation as shown in~\\\\Cref{fig:staterep}.\\nSuch a representation method requires training a representation network when training the main policy network, which increases the model complexity.\\n\\\\begin{figure}[!h]\\n \\\\centering\\n \\\\includegraphics[width=0.5\\\\linewidth]{staterep_wb.pdf}\\n \\\\caption{State representation used in works~. $h_t$ is the output of an attention layer that takes the representation of users' history at time $t$ as input and $g(\\\\cdot)$ is the pooling operation. }\\n \\\\label{fig:staterep}\\n\\\\end{figure}\\n\\\\\\\\\",\n \"id\": \"45229725-ce80-46d4-8644-8011bc19bde5\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"c30b91c8-48bc-4d92-ba66-c418e39b9327\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Component Optimization in Deep Reinforcement Learning based RS\"\n ],\n [\n \"subsubsection\",\n \"State Representation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"State Representation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"~\\\\\\\\\\nThe reward function is crucial for methods involving DQN.\\nA robust reward function can significantly improve training efficiency and performance. find that the DQN may not receive the correct reward value when entering the absorbing state. That is, when the absorbing state is reached, the agent will receive zero reward and harm policy learning. The reason behind this is that when designing the environment zero reward is implicitly assigned to the absorbing state as it is hard to determine the reward value in such a state. propose a robust DQN method, which can stabilize the reward value when facing the absorbing state. The new reward formula can improve the robustness, which is defined as follows:\\n\\\\begin{align}\\n r = \\\\begin{cases} \\n r_t & \\\\text{if } s_{t+1} \\\\text{ is an absorbing state} \\\\\\\\\\n r_t + \\\\gamma Q_{\\\\theta'}(s_{t+1},a_{t+1})& \\\\text{otherwise}.\\n \\\\end{cases}\\n\\\\end{align}\\nThe major difference is that $r_t$ is assigned to the absorbing state to ensure the agent can continue learning.\\nOne remaining problem in current DRL-based RS is the reward sparsity, i.e,. the large state and action spaces make the reward sparsity problem more serious. One of the possible solution would be a better designed reward by using the reward shaping~.\",\n \"id\": \"0c8d5ce0-24e2-4a9b-aa04-14e44c18baf4\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"c30b91c8-48bc-4d92-ba66-c418e39b9327\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Deep Reinforcement Learning in Recommender Systems\"\n ],\n [\n \"subsection\",\n \"Component Optimization in Deep Reinforcement Learning based RS\"\n ],\n [\n \"subsubsection\",\n \"Robustness of Reward Functions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Robustness of Reward Functions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"While existing studies have established a solid foundation for DRL-based RS research, this section\\noutlines several promising emerging research directions.\",\n \"id\": \"3dd345fb-ce99-4727-b174-a3f36f9f711a\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Emerging Topics\"\n ]\n ],\n \"subsections\": [\n \"fd29fd11-361a-4846-b9fe-151e3edf8271\",\n \"ca4b40c6-413e-4583-9e93-7c0668b3a545\",\n \"04addbb4-b221-4c4b-aaeb-b0d4e6b0d4fb\",\n \"19a7db65-832f-4d59-9f47-c445aa549fa1\"\n ],\n \"title\": \"Emerging Topics\"\n },\n {\n \"cite_extract_rate\": 0.30769230769230704,\n \"cites\": [\n 8884,\n 5185,\n 3588,\n 3616\n ],\n \"content\": \"Recommender systems are monolithic systems containing \\ntasks such as searching, ranking, recommendation, advertising, personalization, and diverse stakeholders such as users and items. Most existing methods are based on single agent. \\nMulti-Agent Reinforcement Learning (MARL) is a subfield of reinforcement learning and it is capable of learning multiple policies and strategies. \\nWhile a single-agent reinforcement learning framework can only handle a single task, many studies consider the multi-task situation in RS and employ multi-agent DRL (MADRL) or hierarchical DRL (HDRL).\\nHDRL~ is proposed to handle complex tasks by splitting such tasks into several small components and asks the agent to determine sub-policies. HDRL belongs to a single-agent reinforcement learning framework such that the agent contains a meta-controller and several controllers. The meta-controller splits the task, and the controllers learn the value and reward functions for designated tasks to get a series of sub-policies. There are a few studies already utilizing HDRL in RS.\\n target integrated recommendation to capture user preferences on both heterogeneous items and recommendation channels. Specifically, the meta-controller is used for item recommendation, and controllers aim to find the personalized channel according to user channel-level preferences. uses HDRL for course recommendation in MOOCs, which contains two different tasks: profile reviser and recommendation. The meta-controller aims to make course recommendations by using the revised profile pruned by the controllers.\\nDifferent from HDRL, MADRL~ introduces multiple agents to handle the sub-tasks. uses the MADRL for twitter mention recommendation where three agents are initialized. The three agents need to generate different representations for the following tasks: query text, historical text from authors and historical text from candidate users. Once the representations are finalized, the model will conduct the recommendation based on the concatenation of representations. and provide two different views of the communication mechanism in MADRL and demonstrate that agents could work collaboratively or individually. designs a MADRL framework for two tasks where two agents are designed to conduct advertising and recommendation respectively. uses MADRL for collaborative recommendation where each agent is responsible for a single user. MADRL is adopted to help the recommender consider both collaboration and potential competition between users. designs a charging recommender system for intelligent electric vehicles by using decentralized agents to handle sub-tasks and a centralized critic to make the final decision.\\nHierarchical multi-agent RL (HMARL)~ proves that MARL and HRL can be combined.\\nRecently, introduces HMADRL into the continuous action space, which provides a direction for RS. uses HMARL for multi-goal recommendation where the meta-controller considers users' long-term preferences and controllers focus on short-term click behavior. While the meta-controller and controllers in HDRL deal with sub-tasks that belong to a single task, HMARL focuses on multi-task or multi-goal learning where the meta-controller and controllers belong to different agents and deal with different tasks or goals.\\nHMADRL would be a suitable solution for future research work in DRL-based RS where HDRL can be used to split a complex task into several sub-tasks such as users' long-term interests and short-term click behavior, and MADRL can jointly consider multiple factors such as advertising .\",\n \"id\": \"fd29fd11-361a-4846-b9fe-151e3edf8271\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"3dd345fb-ce99-4727-b174-a3f36f9f711a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Emerging Topics\"\n ],\n [\n \"subsection\",\n \"Multi-Agent and Hierarchical Deep Reinforcement Learning-based RS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-Agent and Hierarchical Deep Reinforcement Learning-based RS\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 6160,\n 3592,\n 6159,\n 6161,\n 5199\n ],\n \"content\": \"As\\nmentioned,\\nthe reward function plays a critical role in DRL-based recommender systems. In many existing works, reward functions are manually designed. The common method uses users' click behavior to represent the reward and to reflect users' interests. However, such a setting can not represent users' long-term goals~ as clicking or not only depicts part of the feedback information from users. It requires significant effort to design a reward function, due to the large number of factors that can affect users' decision, such as social engagement or bad product reviews, which may adversely affect recommendation performance. It is difficult to include all potential factors into the reward function because not every factor can be represented properly. A few works~ show that manually designed reward functions can be omitted by employing inverse reinforcement learning (IRL)~ or generative adversarial imitation learning (GAIL)~. Such inverse DRL-based methods require using expert demonstration as the ground truth. However, expert demonstration is often hard to obtain for recommendation scenarios.\\nThose two studies conduct experiments in an offline dataset-based simulation environment that can access expert demonstration. In contrast, use IRL as the main algorithm to train the agent while use both demonstration and reward to train the agent. also employ GAIL to improve recommendation performance. In this work, GAIL is used to learn the reasoning path inside the KG to provide side information to help the agent learn the policy. Although IRL achieves some progress in RS, the lack of demonstration is a key shortcoming that impedes adoption in RS. One possibility is to use the IRL method in casual reasoning to help improve interpretability~ thus boosting recommendation performance. Alternately, IRL may be suitable for learning users' long-term and static behavior to support the reward function.\",\n \"id\": \"ca4b40c6-413e-4583-9e93-7c0668b3a545\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"3dd345fb-ce99-4727-b174-a3f36f9f711a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Emerging Topics\"\n ],\n [\n \"subsection\",\n \"Inverse Deep Reinforcement Learning for RS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Inverse Deep Reinforcement Learning for RS\"\n },\n {\n \"cite_extract_rate\": 0.555555555555555,\n \"cites\": [\n 1444,\n 6160,\n 180,\n 6162,\n 5194\n ],\n \"content\": \"Graph data and KG are widely used in RS.\\nGraph modeling enables an RS to leverage interactions between users and the recommender for reasoning or improving interpretability. According to existing studies about deep learning-based RS~, embedding is a technique used to get the representation for the input data. Graph embedding is a common solution to handle graph-like data. GCN is a type of graph embedding method which\\nare broadly used in RS to process graph data. propose a variant of GCN to learn the embedding for KG. Specifically, they propose knowledge graph convolutional networks (KGCN) to capture the high-order structural proximity among entities in a knowledge graph. \\nIn DRL-based RS, graph data are handled similarly---s underthe transformed into an embedding and fed to the agent. uses a traditional graph embedding method TransE~ to generate the state representation for DRL-based RS. There are several studies that use GCN in DRL for recommendations under different settings. propose a graph convolutional RL (DGN) method which integrates the GCN into the Q-learning framework for general RL problems by replacing the state encoding layer with the GCN layer. extend this method into the deep learning field and apply it to recommender systems. To be specific, multiple GCN layers are employed to process the sub-graphs for a given item $i$. employs KG inside the actor-critic algorithm to help the agent learn the policy. Specifically, the critic network contains a GCN layer to give weight to the graph and conduct searches in the graph to find an optimal path and hence guide the optimization of policy learning. However, such a method is relatively computationally expensive as it requires jointly training the GCN and the actor-critic network. adopts a Graph Attention Network (GAT)~ into the actor-critic network to conduct recommendation. In addition, the GAT is used as an encoder to obtain a state representation.\\nA common way of using GCN or its variants in DRL-based RS is the state encoder. \\nThe related challenge is the difficulty for the environment to provide a graph-like input to the GCN.\",\n \"id\": \"04addbb4-b221-4c4b-aaeb-b0d4e6b0d4fb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"3dd345fb-ce99-4727-b174-a3f36f9f711a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Emerging Topics\"\n ],\n [\n \"subsection\",\n \"Graph Neural Networks for Boosting DRL-based RS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Graph Neural Networks for Boosting DRL-based RS\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4711,\n 6163,\n 6164,\n 5188\n ],\n \"content\": \"Self-supervised learning (SSL) is a technique in which the model is trained by itself without external label information. SSL-DRL is receiving growing interest in robotics \\n~. shows that SSL can be used to learn the policy when doing navigation by providing real-world experience. demonstrates that SSL-DRL can be used to help the agent learn synergies between two similar policies, thus empowering the agent to conduct two different tasks. \\nRecent advances in SSL RL show that SSL can also provide interpretability for RL, which is promising for interpretable RS research~. shows that SSL based RL can highlight the task-relevant information to guide the agent's behavior.\\nMoreover, shows that SSL can be used to provide negative feedback for DRL-based RS to improve recommendation performance. To be specific, a self-supervised loss function is appended into the normal DRL loss function,\\n\\\\begin{align}\\n -\\\\sum_{i=1}^nY_i\\\\log\\\\bigg(\\\\frac{e^{y_i}}{\\\\sum_{i'=1}^ne^{y_{i'}}}\\\\bigg) + L_{\\\\mathit{DRL}}\\n\\\\end{align}\\nwhere $Y_i$ is an indicator function to show users interact with the item $i$ or not. $L_{\\\\mathit{DRL}}$ could vary, if the DQN is adopted, \\\\Cref{dqnloss} should be used.\\nSSL demonstrates promising performance in visual representation in recent years, which would be a possible solution to generate the state representation as there are a few DRL-based RS studies that adopt CNNs to process image-like data and transform it into a state~. Furthermore, as an unsupervised learning approach, SSL would provide a new direction about defining the reward function by learning common patterns between different states as well as multi-task learning.\",\n \"id\": \"19a7db65-832f-4d59-9f47-c445aa549fa1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"3dd345fb-ce99-4727-b174-a3f36f9f711a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Emerging Topics\"\n ],\n [\n \"subsection\",\n \"Self-Supervised DRL-based RS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Self-Supervised DRL-based RS\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we outline several open questions and challenges that exist in DRL-based RS research. We believe these issues could be critical for the future development of DRL-based RS.\",\n \"id\": \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Open Questions\"\n ]\n ],\n \"subsections\": [\n \"e24a9f91-4259-41ce-8cfe-67de33facae4\",\n \"f3bc9826-577b-48f2-9fd1-1cbd7502868e\",\n \"8fb9ed60-6e10-4fb7-a67b-6f49f9ef3798\",\n \"bb9fa85c-840e-4ea9-928c-30b0b38ddbc5\",\n \"317e1a16-83a6-4e39-b6b1-02817b7555bf\",\n \"03cb7efb-1098-4511-ad2a-26dcf319c6b5\"\n ],\n \"title\": \"Open Questions\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 3586,\n 8985\n ],\n \"content\": \"Sample inefficiency is a well-known challenge in model-free DRL methods. Model-free DRL requires a significant number of samples as there is no guarantee that the received state is useful. Normally, after a substantial number of episodes, the agent may start learning as the agent finally receives a useful state and reward signal. A common solution is the experience replay technique, which only works in off-policy methods. \\nExperience replay still suffers the sample inefficiency problem~ as not every past experience is worth replaying. propose selected experience replay (SER) that only stores valuable experience into the replay buffer and thus improves sample efficiency. while traditional DRL environments only contain several\\\\footnote{For example, the number of actions in MuJoCo is less than one hounded.} candidate items, in DRL-based RS, the agent must deal with a significantly larger action space as RS may contain lots of candidate items. Existing DRL-based RS studies on traditional experience replay methods often demonstrate slow converge speed. design a user model to improve the sample efficiency through auxiliary learning. Specifically, they apply the auxiliary loss with the state representation, and the model distinguishes low-activity users and asks the agent to update the recommendation policy based on high-activity users more frequently.\\nOn the other hand, model-based methods are more sample efficient. However, they introduce extra complexity as the agent is required to learn the environment model as well as the policy. Due to the extremely large action space and possibly large state space (depending on users' contextual information) in RS, approximating the environment model and policy simultaneously\\nbecomes challenging.\",\n \"id\": \"e24a9f91-4259-41ce-8cfe-67de33facae4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Open Questions\"\n ],\n [\n \"subsection\",\n \"Sample Efficiency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Sample Efficiency\"\n },\n {\n \"cite_extract_rate\": 0.388888888888888,\n \"cites\": [\n 5190,\n 6154,\n 5180,\n 5195,\n 5177,\n 5165,\n 5199\n ],\n \"content\": \"The exploration and exploitation dilemma is a fundamental and challenging problem in reinforcement learning research and receives lots of attention in DRL.\\nThis dilemma describes a trade-off between obtaining new knowledge and the need to use that knowledge to improve performance. Many DQN-based methods focus on exploration before the replay buffer is full and exploitation afterward. Consequently, it requires an extremely large replay buffer to allow all possibilities in recommendation can be stored.\\nDRN employs Dueling Bandit Gradient Descent (DBGD)~ to encourage exploration while ~ introduces a regularization or entropy term into the objective function to do so.~ uses the sheer size of the action space to ensure sufficient exploration.~ uses a separate KG or elaborated graph exploration operation to conduct exploration.~ employs Boltzmann exploration to get the benefit of exploratory data without negatively impacting user experience. \\nIn addition, $\\\\epsilon$-greedy is the most common technique used to encourage exploration ~. Remaining studies rely on a simulator to conduct exploration. However, it may suffer from noise and over-fitting~ because of the gap between simulation and real online application. For most DRL-based methods such as vanilla DQN, policy gradient, or actor-critic-based methods, $\\\\epsilon$-greedy would be a good choice for exploration. In addition, injecting noise into the action space would also be helpful for those actor-critic-based methods~. For methods involving KGs, $\\\\epsilon$-greedy may help, but the elaborated graph exploration methods may receive better performance.\",\n \"id\": \"f3bc9826-577b-48f2-9fd1-1cbd7502868e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Open Questions\"\n ],\n [\n \"subsection\",\n \"Exploration and Exploitation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Exploration and Exploitation\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7022,\n 3199,\n 8106\n ],\n \"content\": \"Existing work generally trains DRL algorithms in simulation environments or offline datasets.\\nDeploying DRL algorithms into real applications is challenging due to the gap between simulation and real-world applications.\\nSimulation environments do not contain domain knowledge or social impact.\\nThey can not cover the domain knowledge and task-specific engineering in the real-world recommendation.\\nHow to bridge the gap between simulation and real applications is a challenging topic. Sim2real~ is a transfer learning approach that transfers DRL policies from simulation environments to reality. Sim2real uses domain adaption techniques to help agents transfer learned policy. Specifically, it adopts GANs to conduct adaption by generating different samples.\\nRL-CycleGAN~ is a sim2real method for vision-based tasks. It uses CycleGAN~ to conduct pixel-level domain adaption. Specifically, it maintains cycle consistency during GAN training and encourages the adapted image to retain certain attributes of the input image.\\nIn DRL-based RS, sim2real would be a possible solution for generalizing the learned policy from simulation environments to reality. However, sim2real is a new technique still under exploration. It shows an adequate capability in simple tasks and requires more effort to handle the complex task such as recommendation. We believe it is a workable solution for generalizing from simulation to reality.\",\n \"id\": \"8fb9ed60-6e10-4fb7-a67b-6f49f9ef3798\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Open Questions\"\n ],\n [\n \"subsection\",\n \"Generalizing from Simulation to Real-World Recommendation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Generalizing from Simulation to Real-World Recommendation\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6165,\n 5165,\n 3891\n ],\n \"content\": \" observe that user behavior data are not experimental but observational, which leads to problems of bias and unfairness.\\\\\\\\\\nThere are two reasons why bias is so common. First, the inherent characteristic of user behavior data is not experimental but observational. In other words, data that are fed into recommender systems are subject to selection bias . For instance, users in a video recommendation system tend to watch, rate, and comment on those movies that they are interested in. Second, a distribution discrepancy exists, which means the distributions of users and items in the recommender system are not even. Recommender systems may suffer from 'popularity bias', where popular items are recommended far more frequently than the others. However, the ignored products in the “long tail” can be equally critical for businesses as they are the ones less likely to be discovered.\\n denote the unfairness as that the system systematically and unfairly discriminates against certain individuals or groups of individuals in favor of others. \\nA large number of studies explore dynamic recommendation systems by utilizing the agent mechanism in reinforcement learning (RL), considering the information seeking and decision-making as sequential interactions. How to evaluate a policy efficiently is a big challenge for RL-based recommenders. Online A/B tests are not only expensive and time-consuming but also sometimes hurt the user experience. Off-policy evaluation is an alternative strategy that historical user behavior data are used to evaluate the policy. However, user behavior data are biased, as mentioned before, which causes a gap between the policy of RL-based RS and the optimal policy.\\nTo eliminate the effects of bias and unfairness, use the inverse of the probability of historical policy to weight the policy gradients. introduce a debiasing step that corrects the biases presented in the logged data before it is used to simulate user behavior. propose to build a customer simulator that is designed to simulate the environment and handle the selection bias of logged data.\",\n \"id\": \"bb9fa85c-840e-4ea9-928c-30b0b38ddbc5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Open Questions\"\n ],\n [\n \"subsection\",\n \"Bias (Unfairness)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Bias (Unfairness)\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 5190,\n 8881\n ],\n \"content\": \"Although deep learning-based models can generally improve the performance of recommender systems, they are not easily interpretable. As a result, it becomes an important task to make recommender results explainable, along with providing high-quality recommendations.\\nHigh explainability in recommender systems not only helps end-users understand the items recommended but also enables system designers to check the internal mechanisms of recommender systems. review different information sources and various types of models that can facilitate explainable recommendation. Attention mechanisms and knowledge graph techniques currently play an important role in realizing explainability in RS.\\nAttention models have great advantages in both enhancing predictive performance and having greater explainability~. introduce a reinforcement learning framework incorporated with an attention model for explainable recommendation. Firstly, it achieves model-agnosticism by separating the recommendation model from the explanation generator. Secondly, the agents that are instantiated by attention-based neural networks can generate sentence-level explanations. \\nKnowledge graphs contain rich information about users and items, which can help to generate intuitive and more tailored explanations for the recommendation system . Recent work has achieved greater explainability by using reinforcement and knowledge graph reasoning. The algorithm from learns to find a path that navigates from users to items of interest by interacting with the knowledge graph environment. extract imperfect path demonstrations with minimum labeling effort and propose an adversarial actor-critic model for demonstration-guided path-finding. Moreover, it achieves better recommendation accuracy and explainability by reinforcement learning and knowledge graph reasoning.\",\n \"id\": \"317e1a16-83a6-4e39-b6b1-02817b7555bf\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Open Questions\"\n ],\n [\n \"subsection\",\n \"Explainability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Explainability\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6167,\n 1192,\n 6166\n ],\n \"content\": \"Adversarial samples demonstrate that deep learning-based methods are vulnerable. Hence,\\nrobustness becomes an open question for both RS and DRL. Specifically, adversarial attack and defense in RS have received a lot of attention in recent years~ as security is crucial in RS. Moreover, DRL policies are vulnerable to adversarial perturbations to agent's observations~. provide an adversarial attack method for perturbing the observations, thus affecting the learned policy. Hence, improving the robustness is the common interest for DRL and RS, which would be a critical problem for DRL-based RS. provide an adversarial attack detection method for DRL-based RS which uses the GRU to encode the action space into a low-dimensional space and design decoders to detect the potential attack. However, it only considers Fast Gradient Sign Method (FGSM)-based attacks and strategically-timed attacks~. Thus, it lacks the capability to detect other types of attack. Moreover, it only provides the detection method while the defence is still an opening question.\\nWe believe zero-shot learning techniques would be a good direction for training a universal adversarial attack detector. For defence, it is still an open question for DRL-based RS, though recent advances in adversarial defence in DRL may provide some insights~.\",\n \"id\": \"03cb7efb-1098-4511-ad2a-26dcf319c6b5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"62c3e9c3-0318-4d9a-947c-bf1d76096072\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Open Questions\"\n ],\n [\n \"subsection\",\n \"Robustness on Adversarial Samples and Attacks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Robustness on Adversarial Samples and Attacks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this section, we provide a few potential future directions of DRL-based RS. Benefiting from recent advances in DRL research, we believe those topics can boost the progress of DRL-based RS research.\",\n \"id\": \"db61769c-00fa-49cf-aca1-794fe6209774\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ]\n ],\n \"subsections\": [\n \"e714f9af-5ef5-4cc0-b0fc-93a677fe49fd\",\n \"69b9e3f2-89df-425e-ae0b-359e87390941\",\n \"18638c1b-d943-46f7-8255-6db369c76c4b\"\n ],\n \"title\": \"Future Directions\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 6168,\n 4573,\n 6169,\n 4428,\n 3610,\n 4416\n ],\n \"content\": \"Causality is a generic relationship between a cause and effect. Moreover, inferring causal effects is a fundamental problem in many applications like computational advertising, search engines, and recommender systems~. \\nRecently, some researchers have connected reinforcement learning with learning causality to improve the effects for solving sequential decision-making problems.\\nBesides, Learning agents in reinforcement learning frameworks face a more complicated environment where a large number of heterogeneous data are integrated.\\nFrom our point of view, causal relationships would be capable of improving the recommendation results by introducing the directionality of cause and the effect. The users' previous choices have impact on the subsequent actions. This can be cast as an interventional data generating the dynamics of recommender systems. By viewing a policy in RL as an intervention, we can detect unobserved confounders in RL and choose a policy on the expected reward to better estimate the causal effect . Some studies improve RL models with causal knowledge as side information. Another line of work uses causal inference methods to achieve unbiased reward prediction~. \\n propose a Causal Inference Q-network which introduces observational inference into DRL by applying extra noise and uncertain inventions to improve resilience. Specifically, in this work, noise and uncertainty are added into the state space during the training state, and the agent is required to learn a causal inference model by considering the perturbation.\\n give the first demonstration that model-free reinforcement learning can be used for causal reasoning. They explore meta-reinforcement learning to solve the problem of causal reasoning. The agents trained by a recurrent network able to make causal inferences from observational data and output counterfactual predictions. \\n bridge RL and causality by data-fusion for reinforcement learners. Specifically, online agents combine observations, experiments and counterfactual data to learn about the environment, even if unobserved confounders exist. Similarly, make the model-based RL agents work in a causal way to explore the environment under the Partially-Observable Markov Decision Process (POMDP) setting. They consider interventional data and observational data jointly and interprete model-based reinforcement learning as a causal inference problem. In this way, they bridge the gap between RL and causality by relating common concepts in RL and causality. \\nRegarding explainability in RL, propose to explain the behavior of agents in reinforcement learning with the help of causal science. The authors encode causal relationships and learn a structural causal model in RL, which is used to generate explanations based on counterfactual analysis. With counterfactual exploration, this work is able to generate two contrastive explanations for `why' and `why not' questions.\\nIt is so important to search for a Directed Acyclic Graph (DAG) in causal discovery. Considering traditional methods rely on local heuristics and predefined score functions, propose to use reinforcement learning to search DAG for causal discovery. They use observational data as an input, RL agents as a search strategy and output the causal graph generated from an encoder-decoder NN model.\",\n \"id\": \"e714f9af-5ef5-4cc0-b0fc-93a677fe49fd\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"db61769c-00fa-49cf-aca1-794fe6209774\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Causal and Counterfactual Inference\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Causal and Counterfactual Inference\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7753,\n 6170\n ],\n \"content\": \"Recommender systems often need to deal with multiple scenarios such as joint recommendation and adverting, offline DRL and meta DRL provide a promising direction for achieving multiple scenarios at the same time.\\nOffline DRL is a new paradigm of DRL that can be combined with existing methods such as self-supervised learning and transfer learning to move toward real-world settings. \\nOffline DRL~ (also known as batch DRL) is designed for tasks which contain huge amounts of data.\\nGiven a large dataset that contains past interactions, offline DRL uses the dataset for training across many epochs but does not interact with the environment.\\nOffline DRL provides a solution that can be generalized to new scenarios as it was trained by a large sized dataset.\\nSuch generalization ability is critical to RSs, which may need to deal with multiple scenarios or multiple customers.\\nWhile offline DRL could provide a new direction for DRL-based RS, it still faces a few problems regarding handling the distributional shifts between existing datasets and real-world interactions. \\nMeta DRL~ is defined as \\nmeta learning in the filed of DRL. Meta DRL is another approach to help agents to generalize to new tasks or environments. Different from offline DRL, meta DRL contains a memory unit which is formed by the recurrent neural network to memorize the common knowledge for different tasks. Different from offline DRL, meta DRL does not require a large amount of data to train.\",\n \"id\": \"69b9e3f2-89df-425e-ae0b-359e87390941\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"db61769c-00fa-49cf-aca1-794fe6209774\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Offline DRL and Meta DRL\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Offline DRL and Meta DRL\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 6171,\n 2219,\n 1391,\n 1389\n ],\n \"content\": \"An actor-critic method uses the traditional policy gradient method, which suffers from the high variance problem due to the gap between behavior policy (i.e., the policy that is being used by an agent for action select) and target policy (i.e., the policy that an agent is trying to learn).\\nA method commonly used to \\nrelieve\\nthe high variance problem is Advantage Actor-critic (A2C). \\nDifferent from traditional actor-critic methods, A2C uses an advantage function to replace the Q-function inside the critic network. The advantage function $A(s_t)$ is defined as the expected value of the TD-error. \\nThe new objective function for policy gradient can be written as,\\n\\\\begin{align}\\n \\\\mathbb{E}_{\\\\tau \\\\sim d_{\\\\pi_\\\\theta}}[\\\\sum_{t=1}^T\\\\underbrace{(Q(s_t,a_t) - V(s_t))}_{A(s_t)}\\\\sum_{t=1}^T\\\\nabla_{\\\\theta} \\\\log \\\\pi_{\\\\theta}(s_t,a_t)].\\n\\\\end{align}\\nHowever, A2C still uses DDPG as the main training algorithm, which may suffer function approximation errors when estimating the Q value.\\nTwin-Delayed DDPG (TD3)~ is designed to improve the function approximation problem in DDPG which uses clipped double Q-learning to update the critic. The gradient update can be expressed as,\\n\\\\begin{align}\\n \\\\mathbb{E}_{\\\\tau \\\\sim d_{\\\\pi_\\\\theta}}[\\\\sum_{t=1}^Tr(s_t,a_t) + \\\\gamma \\\\min(Q_1(s_t,a_t+\\\\epsilon), Q_2(s_t,a_t+\\\\epsilon))\\\\sum_{t=1}^T\\\\nabla_{\\\\theta} \\\\log \\\\pi_{\\\\theta}(s_t,a_t)].\\n\\\\end{align}\\nwhere $\\\\epsilon\\\\sim\\\\textit{clip}(\\\\mathcal{N}(0,\\\\sigma,-c,c))$, $\\\\sigma$ is the standard deviation and $c$ is a constant for clipping. \\nAnother two ways to improve actor-critic methods are Trust Region Policy Optimization (TRPO)~ and Proximal Policy Optimization (PPO)~, which focus on modification of the advantage function. TRPO aims to limit the step size for each gradient to ensure it will not change too much. The core idea is to add a constraint to the advantage function,\\n\\\\begin{align}\\n \\\\frac{\\\\pi(a|s)}{\\\\pi_{old}(a|s)}A(s),\\n\\\\end{align}\\nwhere the KL divergence will be used to measure the distance between the current policy and the old policy is small enough. PPO has the same goal as TRPO which is to try to find the biggest possible improvement step on a policy using the current data. PPO is a simplified version of TRPO which introduces the clip operation,\\n\\\\begin{align}\\n \\\\min\\\\bigg(\\\\frac{\\\\pi(a|s)}{\\\\pi_{old}(a|s)}A(s),\\\\text{clip}\\n \\\\bigg(\\\\frac{\\\\pi(a|s)}{\\\\pi_{old}(a|s)}A(s), 1-\\\\epsilon, 1+\\\\epsilon\\\\bigg)A(s)\\\\bigg).\\n\\\\end{align}\\nSoft Actor-Critic (SAC)~ is another promising variant of the actor-critic algorithm and is widely used in DRL research. SAC uses the entropy term to encourage the agent to explore, which could be a possible direction to solve the exploration and exploitation dilemma. Moreover, SAC assigns an equal probability to actions that are equally attractive to the agent to capture those near-optimal policies.\\nAn example of related work uses SAC to improve the stability of the training process in RS.\",\n \"id\": \"18638c1b-d943-46f7-8255-6db369c76c4b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"db61769c-00fa-49cf-aca1-794fe6209774\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Further Developments in Actor-Critic Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Further Developments in Actor-Critic Methods\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In this survey, we provide a comprehensive overview the use of deep reinforcement learning in recommender systems. We introduce a classification scheme for existing studies and discuss them by category. We also provide an overview of such existing emerging topics and point out a few promising directions. \\nWe hope this survey can provide a systematic understanding of the key concepts in DRL-based RS and valuable insights for future research. \\n\\\\bibliographystyle{ACM-Reference-Format}\\n\\\\bibliography{sample-base}\\n\\\\end{document}\\n\\\\endinput\",\n \"id\": \"0a382833-431c-4392-982a-66ab0eba2961\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4bb82714-5571-4179-9dac-693fb550f058\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Deep Reinforcement Learning in Recommender Systems: A Systematic Review and Future Directions\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":89,"string":"89"},"all_cites":{"kind":"list like","value":[8983,5194,8105,5189,5180,5195,7217,5190,3604,8884,5165,6153,3616,1389,5184,5191,5177,5169,5199,6154,5181,6152,5172,5170,5185,6155,1408,1911,6156,6157,6158,5192,5201,8984,5204,3610,1346,3588,6160,3592,6159,6161,1444,180,6162,4711,6163,6164,5188,3586,8985,7022,3199,8106,6165,3891,8881,6167,1192,6166,6168,4573,6169,4428,4416,7753,6170,6171,2219,1391],"string":"[\n 8983,\n 5194,\n 8105,\n 5189,\n 5180,\n 5195,\n 7217,\n 5190,\n 3604,\n 8884,\n 5165,\n 6153,\n 3616,\n 1389,\n 5184,\n 5191,\n 5177,\n 5169,\n 5199,\n 6154,\n 5181,\n 6152,\n 5172,\n 5170,\n 5185,\n 6155,\n 1408,\n 1911,\n 6156,\n 6157,\n 6158,\n 5192,\n 5201,\n 8984,\n 5204,\n 3610,\n 1346,\n 3588,\n 6160,\n 3592,\n 6159,\n 6161,\n 1444,\n 180,\n 6162,\n 4711,\n 6163,\n 6164,\n 5188,\n 3586,\n 8985,\n 7022,\n 3199,\n 8106,\n 6165,\n 3891,\n 8881,\n 6167,\n 1192,\n 6166,\n 6168,\n 4573,\n 6169,\n 4428,\n 4416,\n 7753,\n 6170,\n 6171,\n 2219,\n 1391\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.8819139710299241,"string":"0.881914"}}},{"rowIdx":56,"cells":{"authors":{"kind":"list like","value":["Hwanjun Song","Minseok Kim","Dongmin Park","Yooju Shin","Jae-Gil Lee"],"string":"[\n \"Hwanjun Song\",\n \"Minseok Kim\",\n \"Dongmin Park\",\n \"Yooju Shin\",\n \"Jae-Gil Lee\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Learning from Noisy Labels with Deep Neural Networks: A Survey"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-07-16T09:23:13Z","string":"2020-07-16T09:23:13Z"},"category":{"kind":"string","value":"cs.LG"},"abstract":{"kind":"string","value":" Deep learning has achieved remarkable success in numerous domains with help from large amounts of big data. However, the quality of data labels is a concern because of the lack of high-quality labels in many real-world scenarios. As noisy labels severely degrade the generalization performance of deep neural networks, learning from noisy labels\\,(robust training) is becoming an important task in modern deep learning applications. In this survey, we first describe the problem of learning with label noise from a supervised learning perspective. Next, we provide a comprehensive review of 62 state-of-the-art robust training methods, all of which are categorized into five groups according to their methodological difference, followed by a systematic comparison of six properties used to evaluate their superiority. Subsequently, we perform an in-depth analysis of noise rate estimation and summarize the typically used evaluation methodology, including public noisy datasets and evaluation metrics. Finally, we present several promising research directions that can serve as a guideline for future studies. All the contents will be available at \\href{https://github.com/songhwanjun/Awesome-Noisy-Labels}{\\color{blue!50!brown} https://github.com/songhwanjun/Awesome-Noisy-Labels}. \\looseness=-1 "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"]],"subsections":["8a473653-5cf6-4f3e-941b-06c5873beaf5","345d7d1c-66a9-4d03-80b2-cee975c6db53","60bd8c81-37fb-4934-ab18-bb9a24e80724","3b9c1099-a2f2-417d-a753-e663d7ce62fa","8d7ed0bf-5e07-45ed-a147-190ba1a209fb","0788ae52-30c4-4156-89f9-6be2e443db83","01d4e82b-d227-4b85-bfa5-6db6af2b743d","0fd993d8-9dac-43e0-a25d-3f85b6def415"],"title":"root"},{"cite_extract_rate":0.392857142857142,"cites":[7,206,7210,4238,7770,71,3630,8734,4116,4115,7769],"content":"\\label{sec:introduction}\n\\IEEEPARstart{W}ith the recent emergence of large-scale datasets, deep neural networks (DNNs) have exhibited impressive performance in numerous machine learning tasks, such as computer vision , information retrieval , and language processing . Their success is dependent on the availability of massive but carefully labeled data, which are expensive and time-consuming to obtain. Some non-expert sources, such as Amazon's Mechanical Turk and the surrounding text of collected data, have been widely used to mitigate the high labeling cost; however, the use of these source often results in unreliable labels . \nIn addition, data labels can be extremely complex even for experienced domain experts ; \nthey can also be adversarially manipulated by a label-flipping attack . Such unreliable labels are called \\emph{noisy labels} because they may be \\emph{corrupted} from ground-truth labels. The ratio of corrupted labels in real-world datasets is reported to range from $8.0\\%$ to $38.5\\%$ . \\looseness=-1\nIn the presence of noisy labels, training DNNs is known to be susceptible to noisy labels because of the significant number of model parameters that render DNNs overfit to even corrupted labels with the capability of learning any complex function . Zhang et al. demonstrated that DNNs can easily fit an entire training dataset with any ratio of corrupted labels, which eventually resulted in poor generalizability on a test dataset. Unfortunately, popular regularization techniques, such as data augmentation , weight decay , dropout , and batch normalization {have been applied extensively, but they do \\emph{not} completely overcome the overfitting issue by themselves.} As shown in Figure \\ref{fig:convergence_analysis}, the gap in test accuracy between models trained on clean and noisy data remains significant even though all of the aforementioned regularization techniques are activated. Additionally, the accuracy drop with label noise is considered to be more harmful than with other noises, such as input noise . Hence, achieving a good generalization capability in the presence of noisy labels is a key challenge.\n\\begin{figure}[t!]\n\\begin{center}\n\\includegraphics[width=8.8cm]{figures/introduction/convergence.pdf}\n\\end{center}\n\\vspace*{-0.55cm}\n\\caption{Convergence curves of training and test accuracy when training WideResNet-16-8 using a standard training method on the CIFAR-100 dataset with the symmetric noise of $40\\%$: \\enquote{Noisy w/o. Reg.} and \\enquote{Noisy w. Reg.} are the models trained on noisy data without and with regularization, respectively, and \\enquote{Clean w. Reg.} is the model trained on clean data with regularization.}\n\\label{fig:convergence_analysis}\n\\vspace*{-0.4cm}\n\\end{figure}\nSeveral studies have been conducted to investigate supervised learning under noisy labels. Beyond conventional machine learning techniques , deep learning techniques have recently gained significant attention in the machine learning community. In this survey, we present the advances in recent deep learning techniques for overcoming noisy labels. We surveyed {recent studies by recursively tracking relevant bibliographies in papers published at premier research conferences, such as CVPR, ICCV, NeurIPS, ICML, and ICLR. Although we attempted to comprehensively include all recent studies at the time of submission, some of them may not be included because of the quadratic increase in deep learning papers. The studies included were grouped into \\emph{five} categories, as shown in Figure \\ref{fig:categorization} (see Section \\ref{sec:methodology} for details).\n\\begin{figure*}[t!]\n\\begin{center}\n\\includegraphics[width=16.5cm]{figures/introduction/catorization_revise.pdf}\n\\end{center}\n\\vspace*{-0.55cm}\n\\caption{Categorization of recent deep learning methods for overcomming noisy labels.}\n\\label{fig:categorization}\n\\vspace*{-0.4cm}\n\\end{figure*}","id":"8a473653-5cf6-4f3e-941b-06c5873beaf5","level":"section","origin_cites_number":28,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Introduction"]],"subsections":["74284b44-a1b2-476d-b3b5-489a51838419","e5555e5b-acf2-4c85-9ee1-4bf47dcf2ae1"],"title":"Introduction"},{"cite_extract_rate":0.25,"cites":[7771],"content":"\\label{sec:related_surveys}\nFr{\\'e}nay and Verleysen discussed the potential negative consequence of learning from noisy labels and provided a comprehensive survey on noise-robust classification methods, focusing on conventional supervised approaches such as na\\\"ive Bayes and support vector machines. Furthermore, their survey included the definitions and sources of label noise as well as the taxonomy of label noise. Zhang et al. discussed another aspect of label noise in crowdsourced data annotated by non-experts and provided a thorough review of expectation-maximization (EM) algorithms that were proposed to improve the quality of crowdsourced labels. Meanwhile, Nigam et al. provided a brief introduction to deep learning algorithms that were proposed to manage noisy labels; however, the scope of these algorithms was limited to only two categories, i.e., the loss function and sample selection in Figure \\ref{fig:categorization}. Recently, Han et al. summarized the essential components of robust learning with noisy labels, but their categorization is totally different from ours in philosophy; {we mainly focus on systematic methodological difference, whereas they rather focused on more general views, such as input data, objective functions, and optimization policies. Furthermore, this survey is the first to present a comprehensive methodological comparison of existing robust training approaches (see Tables \\ref{table:all_comparision} and \\ref{table:direction_comparison}}).","id":"74284b44-a1b2-476d-b3b5-489a51838419","level":"subsection","origin_cites_number":4,"parent_id":"8a473653-5cf6-4f3e-941b-06c5873beaf5","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Introduction"],["subsection","Related Surveys"]],"subsections":[],"title":"Related Surveys"},{"cite_extract_rate":0.5714285714285711,"cites":[8696,4118,313,4117],"content":"\\label{sec:survey_scope}\n{\nRobust training with DNNs becomes critical to guarantee the reliability of machine learning algorithms. In addition to label noise, two types of flawed training data have been actively studied by different communities . \\emph{Adversarial learning} is designed for small, worst-case perturbations of the inputs, so-called adversarial examples, which are maliciously constructed to deceive an already trained model into making errors . Meanwhile, \\emph{data imputation} primarily deals with missing inputs in training data, where missing values are estimated from the observed ones . Adversarial learning and data imputation are closely related to robust learning, but handling \\emph{feature} noise is beyond the scope of this survey---i.e., learning from noisy \\emph{labels}. \n}","id":"e5555e5b-acf2-4c85-9ee1-4bf47dcf2ae1","level":"subsection","origin_cites_number":7,"parent_id":"8a473653-5cf6-4f3e-941b-06c5873beaf5","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Introduction"],["subsection","Survey Scope"]],"subsections":[],"title":"Survey Scope"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:preliminaries}\nIn this section, the problem statement for supervised learning with noisy labels is provided along with the taxonomy of label noise. Managing noisy labels is a long-standing issue; therefore, we review the basic conventional approaches and {theoretical foundations underlying robust deep learning}. Table \\ref{table:notation} summarizes the notation frequently used in this study. \\looseness=-1\n{\n\\newcolumntype{L}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}m{#1}}\n\\newcolumntype{X}[1]{>{\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}p{#1}}\n\\begin{table}[t!]\n\\caption{Summary of the notation.}\n\\vspace*{-0.3cm}\n\\begin{center}\n\\begin{tabular}{|L{1.7cm} |X{6.2cm}|}\\hline\n\\!\\!\\textbf{Notation} & \\hspace*{2.25cm} \\textbf{Description} \\\\\\hline \\hline \n$\\mathcal{X}$ & the data feature space \\\\\\hline\n$\\mathcal{Y}$, $\\tilde{\\mathcal{Y}}$ & the true and noisy label space \\\\\\hline\n$\\mathcal{D}$, $\\tilde{\\mathcal{D}}$ & the clean and noisy training data \\\\\\hline\n$P_{\\mathcal{D}}$, $P_{\\tilde{\\mathcal{D}}}$ & the joint distributions of clean and noisy data \\\\\\hline\n$\\mathcal{B}_{t}$ & a set of mini-batch examples at time $t$ \\\\\\hline\n$\\Theta_{t}$ & the parameter of a deep neural network at time $t$ \\\\\\hline\n$f(\\,\\cdot\\,;\\Theta_{t})$ & a deep neural network parameterized by $\\Theta_{t}$ \\\\\\hline\n$\\ell$ & a specific loss function \\\\\\hline\n$\\mathcal{R}$ & an empirical risk \\\\\\hline\n$\\mathbb{E}_{\\mathcal{D}}$ & an expectation over $\\mathcal{D}$ \\\\\\hline\n$x$, $x_i$ & a data example of $\\mathcal{X}$ \\\\\\hline\n$y$, $y_i$ & a true label of $\\mathcal{Y}$ \\\\\\hline\n$\\tilde{y}$, $\\tilde{y}_i$ & a noisy label of $\\tilde{\\mathcal{Y}}$ \\\\\\hline\n$\\eta$ & a specific learning rate \\\\\\hline\n$\\tau$ & a true noise rate \\\\\\hline\n$b$ & the number of mini-batch examples in $\\mathcal{B}_{t}$ \\\\\\hline\n$c$ & the number of classes \\\\\\hline\n{T}, $\\hat{\\text{T}}$ & the true and estimated noise transition matrix \\\\\\hline\n\\end{tabular}\n\\end{center}\n\\label{table:notation}\n\\vspace*{-0.55cm}\n\\end{table}\n}\n\\begin{figure*}[t!]\n\\begin{center}\n\\includegraphics[width=15.5cm]{figures/introduction/tree_revise.pdf}\n\\end{center}\n\\vspace*{-0.45cm}\n\\caption{A high level research overview of robust deep learning for noisy labels. The research directions that are actively contributed by the machine learning community are categorized into five groups in blue italic.}\n\\label{fig:tree_categorization}\n\\vspace*{-0.4cm}\n\\end{figure*}","id":"345d7d1c-66a9-4d03-80b2-cee975c6db53","level":"section","origin_cites_number":0,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"]],"subsections":["160721f0-f1c5-43ce-955b-8257be13eaa0","53f5fc74-56fd-44a0-ad89-6075838d51dd","587ccbe7-447b-402d-9bec-671027007ca9","f4b49791-795c-49a1-9ee2-043cc9fa43e4","7dbbb1a5-6e79-42e1-a3b6-1ed894715ef6"],"title":"Preliminaries"},{"cite_extract_rate":0,"cites":[],"content":"\\emph{Classification} is a representative supervised learning task for learning a function that maps an input feature to a label . In this paper, we consider a $c$-class classification problem using a DNN with a softmax output layer. Let $\\mathcal{X} \\subset \\mathbb{R}^{d}$ be the feature space and $\\mathcal{Y} = \\{0,1\\}^{c}$ be the ground-truth label space in a \\emph{one-hot} manner. In a typical classification problem, we are provided with a training dataset $\\mathcal{D}=\\{(x_i, y_i)\\}_{i=1}^{N}$ obtained from an unknown joint distribution $P_{\\mathcal{D}}$ over $\\mathcal{X} \\times \\mathcal{Y}$, where each $(x_i, y_i)$ is \\emph{independent and identically distributed}. The goal of the task is to learn the mapping function $f(\\,\\cdot\\,;\\Theta): \\mathcal{X} \\rightarrow [0, 1]^{c}$ of the DNN parameterized by $\\Theta$ such that the parameter $\\Theta$ minimizes the empirical risk $\\mathcal{R}_{\\mathcal{D}}(f)$,\n\\begin{equation}\n\\label{eq:empirical_risk}\n\\!\\!\\mathcal{R}_{\\mathcal{D}}(f) = \\mathbb{E}_{\\mathcal{D}}[\\ell\\big(f(x;\\Theta), y\\big)] = \\frac{1}{|\\mathcal{D}|}\\!\\sum_{(x,y)\\in\\mathcal{D}}\\!\\!\\!\\!\\ell\\big(f(x;\\Theta), y\\big),\n\\end{equation}\nwhere $\\ell$ is a certain loss function.\nAs data labels are corrupted in various real-world scenarios, we aim to train the DNN from noisy labels. Specifically, we are provided with a noisy training dataset $\\tilde{\\mathcal{D}}=\\{(x_i, \\tilde{y}_i)\\}_{i=1}^{N}$ obtained from a noisy joint distribution $P_{\\tilde{\\mathcal{D}}}$ over $\\mathcal{X}\\times\\tilde{\\mathcal{Y}}$, where $\\tilde{y}$ is a \\emph{noisy} label which may not be true. Hence, following the standard training procedure, a mini-batch $\\mathcal{B}_{t}=\\{(x_i, \\tilde{y}_i)\\}_{i=1}^{b}$ comprising $b$ examples is obtained randomly from the noisy training dataset $\\tilde{\\mathcal{D}}$ at time $t$. Subsequently, the DNN parameter $\\Theta_t$ at time $t$ is updated along the descent direction of the empirical risk on mini-batch $\\mathcal{B}_t$,\n\\begin{equation}\n\\label{eq:corrupted_update}\n\\Theta_{t+1} = \\Theta_{t} - \\eta\\nabla\\Big(\\frac{1}{|\\mathcal{B}_{t}|} \\!\\sum_{(x,\\tilde{y}) \\in \\mathcal{B}_{t}} \\!\\!\\!\\!\\ell\\big(f(x;\\Theta_{t}), \\tilde{y}\\big)\\Big),\n\\end{equation}\nwhere $\\eta$ is a learning rate specified. \\looseness=-1\nHere, the risk minimization process is no longer \\emph{noise-tolerant} because of the loss computed by the noisy labels. DNNs can easily memorize corrupted labels and correspondingly degenerate their generalizations on unseen data . Hence, mitigating the adverse effects of noisy labels is essential to enable noise-tolerant training for deep learning. \n\\vspace*{-0.1cm}","id":"160721f0-f1c5-43ce-955b-8257be13eaa0","level":"subsection","origin_cites_number":4,"parent_id":"345d7d1c-66a9-4d03-80b2-cee975c6db53","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"],["subsection","Supervised Learning with Noisy Labels"]],"subsections":[],"title":"Supervised Learning with Noisy Labels"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:taxonomy}\n{This section presents the types of label noise that have been adopted to design robust training algorithms.} Even if data labels are corrupted from ground-truth labels without \\emph{any} prior assumption, in essence, the corruption probability is affected by the dependency between \\emph{data features} or \\emph{class labels}. A detailed analysis of the taxonomy of label noise was provided by Fr{\\'e}nay and Verleysen . {Most existing algorithms dealt with instance-independent noise, but instance-dependent noise has not yet been extensively investigated owing to its complex modeling.}\n\\smallskip","id":"53f5fc74-56fd-44a0-ad89-6075838d51dd","level":"subsection","origin_cites_number":1,"parent_id":"345d7d1c-66a9-4d03-80b2-cee975c6db53","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"],["subsection","Taxonomy of Label Noise"]],"subsections":["5a2e6abe-21d0-4935-ab73-f9bf0caf9ca6","9e5fdfd7-ba0e-466c-96b0-5d8ad47de6e9"],"title":"Taxonomy of Label Noise"},{"cite_extract_rate":0.5,"cites":[3630],"content":"}\nA typical approach for modeling label noise assumes that the corruption process is conditionally \\emph{independent} of data features when the true label is given . That is, the true label is corrupted by a \\emph{noise transition matrix} $\\text{T} \\in [0, 1]^{c\\times c}$, where $ \\text{T}_{ij} \\coloneqq p(\\tilde{y}=j|y=i)$ is the probability of the true label $i$ being flipped into a corrupted label $j$. \nIn this approach, the noise is called a \\emph{symmetric}\\,(or \\emph{uniform}) noise with a noise rate $\\tau \\in [0,1]$ if $\\forall_{i=j} \\text{T}_{ij} \\!=\\! 1-\\tau \\wedge \\forall_{i \\neq j} \\text{T}_{ij} = \\frac{\\tau}{c-1}$, where a true label is flipped into other labels with equal probability. In contrast to symmetric noise, the noise is called an \\emph{asymmetric}\\,(or \\emph{label-dependent}) noise if $\\forall_{i=j} \\text{T}_{ij} \\!=\\! 1-\\tau \\wedge \\exists_{i \\neq j, i\\neq k, j\\neq k} \\text{T}_{ij} > \\text{T}_{ik} $, where a true label is more likely to be mislabeled into a particular label. For example, a \\enquote{dog} is more likely to be confused with a \\enquote{cat} than with a \\enquote{fish.} In a stricter case when $\\forall_{i=j} \\text{T}_{ij} \\!=\\! 1-\\tau \\wedge \\exists_{i \\neq j} \\text{T}_{ij} = \\tau$, the noise is called a \\emph{pair noise}, where a true label is flipped into only a certain label. \n\\smallskip","id":"5a2e6abe-21d0-4935-ab73-f9bf0caf9ca6","level":"subsubsection","origin_cites_number":2,"parent_id":"53f5fc74-56fd-44a0-ad89-6075838d51dd","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"],["subsection","Taxonomy of Label Noise"],["subsubsection","Instance-independent Label Noise"]],"subsections":[],"title":"Instance-independent Label Noise"},{"cite_extract_rate":0,"cites":[],"content":"}\nFor more realistic noise modeling, the corruption probability is assumed to be \\emph{dependent} on both the data features and class labels . Accordingly, the corruption probability is defined as $\\rho_{ij}(x)\\!= \\!p(\\tilde{y}\\!=\\!j|y\\!=\\!i, x)$. \nUnlike the aforementioned noises, the data feature of an example $x$ also affects the chance of $x$ being mislabeled. \n\\vspace*{-0.1cm}","id":"9e5fdfd7-ba0e-466c-96b0-5d8ad47de6e9","level":"subsubsection","origin_cites_number":2,"parent_id":"53f5fc74-56fd-44a0-ad89-6075838d51dd","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"],["subsection","Taxonomy of Label Noise"],["subsubsection","Instance-dependent Label Noise"]],"subsections":[],"title":"Instance-dependent Label Noise"},{"cite_extract_rate":0.344827586206896,"cites":[3454,7771,4121,8735,4120,4122,3630,4119,4115,8736],"content":"\\label{sec:non_deep_learning}\nFor decades, numerous methods have been proposed to manage noisy labels using conventional machine learning techniques. These methods can be categorized into \\emph{four} groups , as follows:\n\\begin{itemize}[leftmargin=9pt]\n\\item \\textbf{Data Cleaning:} Training data are cleaned by excluding examples whose labels are likely to be corrupted. Bagging and boosting are used to filter out false-labeled examples to remove examples with higher weights because false-labeled examples tend to exhibit much higher weights than true-labeled examples . In addition, various methods, such as $k$-nearest neighbor, outlier detection, and anomaly detection, have been widely exploited to exclude false-labeled examples from noisy training data . Nevertheless, this family of methods suffers from over-cleaning issue that overly removes even the true-labeled examples. \n\\vspace*{0.12cm}\n\\item \\textbf{Surrogate Loss:} Motivated by the noise-tolerance of the 0-1 loss function , many researchers have attempted to resolve its inherent limitations, such as computational hardness and non-convexity that render gradient methods unusable. Hence, several convex surrogate loss functions, which approximate the 0-1 loss function, have been proposed to train a specified classifier under the binary classification setting . However, these loss functions cannot support the multi-class classification task.\n\\vspace*{0.12cm}\n\\item \\textbf{Probabilistic Method:} Under the assumption that the distribution of features is helpful in solving the problem of learning from noisy labels , the confidence of each label is estimated by clustering and then used for a weighted training scheme . This confidence is also used to convert hard labels into soft labels to reflect the uncertainty of labels . In addition to these clustering approaches, several Bayesian methods have been proposed for graphical models such that they can benefit from using any type of prior information in the learning process . However, this family of methods may exacerbate the overfitting issue owing to the increased number of model parameters. \n\\vspace*{0.12cm}\n\\item \\textbf{Model-based Method:} As conventional models, such as the SVM and decision tree, are not robust to noisy labels, significant effort has been expended to improve the robustness of them. To develop a robust SVM model, misclassified examples during learning are penalized in the objective . In addition, several decision tree models are extended using new split criteria to solve the overfitting issue when the training data are not fully reliable . However, it is infeasible to apply their design principles to deep learning.\n\\end{itemize}\n\\smallskip\n{\nMeanwhile, deep learning is more susceptible to label noises than traditional machine learning owing to its high expressive power, as proven by many researchers . \nThere has been significant effort to understand why noisy labels negatively affect the performance of DNNs . This theoretical understanding has led to the algorithmic design which achieves higher robustness than non-deep learning methods. A detailed analysis of theoretical understanding for robust deep learning was provided by Han et al. . \n}\n\\begin{comment}\n\\vspace*{-0.2cm}","id":"587ccbe7-447b-402d-9bec-671027007ca9","level":"subsection","origin_cites_number":29,"parent_id":"345d7d1c-66a9-4d03-80b2-cee975c6db53","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"],["subsection","Non-deep Learning Approaches"]],"subsections":[],"title":"Non-deep Learning Approaches"},{"cite_extract_rate":0.647058823529411,"cites":[7771,3340,4126,3630,4253,8735,4120,4124,4125,4115,4123],"content":"\\label{sec:theoretical}\n{\nDeep learning is more susceptible to label noises than traditional machine learning owing to its high expressive power, as proven by many researchers . \nThere has been significant effort to understand why noisy labels negatively affect the performance of DNNs . This theoretical understanding has led to the architectural or algorithmic design which is more robust than the non-deep learning methods. A detailed analysis of theoretical understanding for robust deep learning was provided by Han et al. ; \\emph{three} high-level perspectives have been widely leveraged to design robust approaches, as follows:\n\\vspace*{0.05cm}\n\\newcommand{\\norm}[1]{\\left\\lVert#1\\right\\rVert}\n\\begin{itemize}[leftmargin=9pt]\n\\item \\textbf{Memorization Effect}: The \\emph{memorization nature} of DNNs was explored theoretically in recent literature . Assuming clusterable data where the clusters are located on the unit Euclidean ball, Li et al. proved the distance from the initial weight ${W}_{0}$ to the weight ${W}_t$ after $t$ iterations,\n\\begin{equation}\n\\label{eq:memorization_effect_foundation}\n\\norm{{W}_t - {W}_0}_{F} \\lesssim \\big( \\sqrt{K} + (K^{2}\\epsilon_{0}/\\norm{{C}}^{2})t \\big),\n\\end{equation}\nwhere $\\norm{\\cdot}_{F}$ is the Frobenius norm, $K$ is the number of clusters, and ${C}$ is the set of cluster centers reaching all input examples within their $\\epsilon_0$ neighborhood.\nEq.\\ \\eqref{eq:memorization_effect_foundation} demonstrates that the weights of DNNs start to stray far from the initial weights when overfitting to corrupted labels, while they are still in the vicinity of the initial weights at the beginning of training . In the empirical studies , this result is also known as the \\emph{memorization effect} that DNNs tend to first learn simple and generalized patterns and then gradually overfit to all the noisy patterns. Thus, to achieve better generalization, early stopping and favoring small-loss training examples are commonly employed to design robust training methods.\n\\end{itemize}\n}\n\\end{comment}\n\\vspace*{-0.2cm}","id":"f4b49791-795c-49a1-9ee2-043cc9fa43e4","level":"subsection","origin_cites_number":17,"parent_id":"345d7d1c-66a9-4d03-80b2-cee975c6db53","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"],["subsection","Theoretical Foundations"]],"subsections":[],"title":"Theoretical Foundations"},{"cite_extract_rate":0.25,"cites":[4127],"content":"\\label{sec:regression}\n{\nIn addition to classification, regression is another main topic of supervised machine learning, which aims to model the relationship between a number of features and a continuous target variable. Unlike the classification task with a \\emph{discrete} label space, the regression task considers the continuous variable as its target label , and thus it learns the mapping function $f(~\\cdot~; \\Theta): \\mathcal{X} \\rightarrow \\mathcal{Y}$, where $\\mathcal{Y} \\in \\mathbb{R}$ is a \\emph{continuous} label space. \nGiven the input feature $x$ and its ground-truth label $y$, two types of label noise are considered in the regression task. An \\emph{additive noise} is formulated by $\\tilde{y} := y + \\epsilon$ where $\\epsilon$ is drawn from a random distribution independent from the input feature; an \\emph{instance-dependent noise} is formulated by $\\tilde{y} := \\rho(x)$ where $\\rho: \\mathcal{X} \\rightarrow \\mathcal{Y}$ is a noise function dependent on the input feature. \nAlthough regression predicts continuous values, regression and classification share the same concept of learning the mapping function from the input feature $x$ to the output label $y$. Thus, many robust approaches for classification are easily extended to the regression problem with simple modification . Thus, in this survey, we focus on the classification setting for which most robust methods are defined.\n}\n\\vspace*{-0.1cm}","id":"7dbbb1a5-6e79-42e1-a3b6-1ed894715ef6","level":"subsection","origin_cites_number":4,"parent_id":"345d7d1c-66a9-4d03-80b2-cee975c6db53","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Preliminaries"],["subsection","Regression with Noisy Labels"]],"subsections":[],"title":"Regression with Noisy Labels"},{"cite_extract_rate":0.777777777777777,"cites":[3340,4129,4143,7133,4137,7773,3345,4134,4132,4140,4141,7775,4126,3342,4142,7772,8737,4130,4138,7191,7162,4135,4144,4145,7774,4139,4125,892,4128,4253,4131,8630,4133,2277,4136],"content":"\\label{sec:methodology}\n\\vspace*{-0.0cm}\nAccording to our comprehensive survey, the robustness of deep learning can be enhanced in numerous approaches . Figure \\ref{fig:tree_categorization} shows an overview of recent research directions conducted by the machine learning community. {All of them\\,(i.e., \\textsection \\ref{sec:robust_architecture}~--~\\textsection \\ref{sec:sample_selection}) focused on making a supervised learning process more robust to label noise: \\looseness=-1\n\\begin{itemize}[leftmargin=9pt]\n\\item\n(\\textsection \\ref{sec:robust_architecture}) Robust architecture: adding a noise adaptation layer at the top of an underlying DNN to learn label transition process or developing a dedicated architecture to reliably support more diverse types of label noise;\n\\vspace*{0.12cm}\n\\item (\\textsection \\ref{sec:robust_regularization}) Robust regularization: enforcing a DNN to overfit less to false-labeled examples explicitly or implicitly;\n\\vspace*{0.12cm}\n\\item (\\textsection \\ref{sec:robust_loss_function}) Robust loss function: improving the loss function;\n\\vspace*{0.12cm}\n\\item (\\textsection \\ref{sec:loss_adjustment}) Loss adjustment: adjusting the loss value according to the confidence of a given loss (or label) by loss correction, loss reweighting, or label refurbishment; \n\\vspace*{0.12cm}\n\\item (\\textsection \\ref{sec:sample_selection}) Sample selection: identifying true-labeled examples from noisy training data via multi-network or multi-round learning. \n\\end{itemize}\nOverall, we categorize all recent deep learning methods into \\emph{five} groups corresponding to popular research directions, as shown in Figure \\ref{fig:tree_categorization}. In \\textsection \\ref{sec:loss_adjustment}, meta learning is also discussed because it finds the optimal hyperparameters for loss reweighting. In \\textsection \\ref{sec:sample_selection}, we discuss the recent efforts for combining sample selection with other orthogonal directions or semi-supervised learning toward the state-of-the-art performance.}\nFigure \\ref{fig:categorization} illustrates the categorization of robust training methods using these five groups. \n\\begin{comment}\n{\n\\newcolumntype{L}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}m{#1}}\n\\newcolumntype{X}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}p{#1}}\n\\begin{table}[t!]\n\\caption{Summary of existing deep learning methods according to the seven categories in Figure \\ref{fig:categorization}.}\n\\vspace*{-0.3cm}\n\\begin{center}\n\\begin{tabular}{L{1.75cm} |X{6.23cm}}\\toprule\n\\!\\!\\textbf{Category} & \\textbf{Deep Learning Method} \\\\\\midrule \nRobust Loss Function & \\makecell[l]{\\!\\!\\emph{Robust MAE}\\,\\!, \\emph{Generalized Cross Entropy}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{Symmetric Cross Entropy}\\,\\!, \\emph{Curriculum Loss}\\,\\!}\\!\\!\\! \\\\\\hline\nRobust Architecture & \\makecell[l]{ \\!\\!\\emph{Webly Learning}\\,\\!, \\emph{Noise Model}\\,\\!, \\emph{Dropout Noise} \\!\\!\\!\\!\\!\\\\ \\!\\!\\emph{Model}\\,\\!, \\emph{S--model}\\,\\!, \\emph{C--model}\\,\\!, \\emph{NLNN}\\,\\!, \\!\\!\\!\\!\\!\\\\ \\!\\!\\emph{Probabilistic Noise Model}\\,\\!, \\emph{Masking}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{Contrastive-additive Noise Network}\\,\\!}\\!\\!\\! \\\\\\hline\nRobust Regularization & \\makecell[l]{ \\!\\!\\!\\! \\emph{Adversarial Training}\\,\\!, \\emph{Label Smoothing}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{Mixup}\\,\\!, \\emph{Bilevel Learning}\\,\\!, \\emph{Annotator}\\!\\!\\!\\\\ \\!\\!\\!\\! \\emph{Confusion}\\,\\!, \\emph{Pre-training}\\,\\!\\!\\!} \\!\\!\\!\\\\\\hline\n\\!\\!Loss Adjustment\\!\\!\\! & \\makecell[l]{ \\!\\!\\emph{Backward Correction}\\,\\!}, \\emph{Forward Correction}\\,\\!}, \\!\\!\\!\\\\ \\!\\!\\emph{Gold}\\! \\emph{Loss}\\! \\emph{Correction}\\,\\!\\!, \\emph{Importance}\\! \\emph{Reweighting}\\,\\!\\!, \\!\\!\\!\\!\\!\\!\\\\ \\!\\!\\emph{Active Bias}\\! , \\emph{Bootstrapping}\\! , \\emph{Dynamic} \\!\\!\\\\ \\!\\!\\!\\! \\emph{Bootstrapping}\\,\\!, \\emph{D2L}\\,\\!, \\emph{SELFIE}\\,\\! }\\!\\!\\!\\\\\\hline\n\\!\\!\\!Sample Selection\\!\\!\\! & \\makecell[l]{\\!\\!\\emph{Decouple}\\,\\!, \\emph{MentorNet}\\,\\!, \\emph{Co-teaching}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{Co-teaching+}\\,\\!, \\emph{Iterative Detection}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{ITLM}\\,\\!, \\emph{INCV}\\,\\!, \\emph{SELFIE}\\,\\!, \\emph{SELF}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{Curriculum Loss}\\,\\!}\\!\\!\\!\\\\\\hline\n\\!\\!\\!Meta Learning\\!\\!\\! & \\makecell[l]{ \\!\\!\\emph{Meta-Regressor}\\,\\!, \\emph{Knowledge Distillation}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{L2LWS}\\,\\!, \\emph{CWS}\\,\\!, \\emph{Automatic Reweighting}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{MLNT}\\,\\!, \\emph{Meta-weight-net}\\,\\!} \\!\\!\\!\\\\\\hline\n\\!\\!Semi-supervised\\!\\! \\!Learning\\! & \\makecell[l]{ \\!\\!\\emph{Label Aggregation}\\,\\!, \\emph{Two-Stage Framework}\\,\\!,\\!\\!\\!\\\\ \\!\\!\\emph{SELF} , \\emph{DivideMix}\\,\\!}\\!\\!\\! \\\\\\bottomrule\n\\end{tabular}\n\\end{center}\n\\label{table:summary_methods}\n\\vspace*{-0.5cm}\n\\end{table}\n}\n\\end{comment}","id":"60bd8c81-37fb-4934-ab18-bb9a24e80724","level":"section","origin_cites_number":45,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"]],"subsections":["f16bcc10-a9fa-4749-b206-8aa388990240","25e15182-1991-4f1a-83d9-017fe1a40d14","59ab6098-f130-49dc-b87f-123eaa6be482","3d40345d-b6d6-4b98-943c-ec552f10d7fb","54d73157-e47b-4943-9b30-b518e1806946"],"title":"Deep Learning Approaches"},{"cite_extract_rate":0.6666666666666661,"cites":[4138,4133,4146,7773,2277,4136],"content":"\\label{sec:robust_architecture}\nIn numerous studies, architectural changes have been made to model the noise transition matrix of a noisy dataset . These changes include adding a noise adaptation layer at the top of the softmax layer and designing a new dedicated architecture. The resulting architectures yield improved generalization through the modification of the DNN output based on the estimated label transition probability.\n\\vspace*{0.15cm}","id":"f16bcc10-a9fa-4749-b206-8aa388990240","level":"subsection","origin_cites_number":9,"parent_id":"60bd8c81-37fb-4934-ab18-bb9a24e80724","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Robust Architecture"]],"subsections":["439a17f4-972c-4739-b63c-f2fd25ca8426","81044d2e-5255-4dbd-8274-7c1d26bb1d92"],"title":"Robust Architecture"},{"cite_extract_rate":0.5,"cites":[7776,2277,4136],"content":"}\n\\label{sec:adaptation_layer}\n{\nFrom the view of training data, the noise process is modeled by discovering the underlying label transition pattern (i.e., the {noise transition matrix} T). Given an example $x$, the noisy class posterior probability for an example $x$ is expressed by\n\\begin{equation}\n\\label{eq:label_transition_matrix}\n\\begin{gathered}\n\\!\\!\\!\\!\\!p(\\tilde{y}=j|x) \\!=\\!\\sum_{i=1}^{c}p(\\tilde{y}=j, y=i|x)\\! = \\!\\sum_{i=1}^{c}\\text{T}_{ij}p(y=i|x),\\\\\n\\text{where} ~~ \\text{T}_{ij} = p(\\tilde{y}=j|y=i,x).\n\\end{gathered}\n\\end{equation} \nIn light of this, the noise adaptation layer is intended to mimic the label transition behavior in learning a DNN. Let $p(y|x;\\Theta)$ be the output of the base DNN with a softmax output layer.\nThen, following Eq.\\,\\eqref{eq:label_transition_matrix}, the probability of an example $x$ being predicted as its noisy label $\\tilde{y}$ is parameterized by }\n\\begin{equation}\n\\label{eq:noise_adaption_process}\n\\begin{split}\n\\!\\!\\!p(\\tilde{y}=j|x;\\Theta, \\mathcal{W})&= \\sum_{i=1}^{c}p(\\tilde{y}=j, y\\!=\\!i|x; \\Theta, \\mathcal{W})\\\\\n&= \\sum_{i=1}^{c}\\underbrace{p(\\tilde{y}=j|y\\!=\\!i;\\mathcal{W})}_{\\text{Noise Adaptation Layer}}\\underbrace{p(y\\!=\\!i|x;\\Theta)}_{\\text{Base Model}}.\n\\end{split}\n\\end{equation}\nHere, the noisy label $\\tilde{y}$ is assumed to be \\emph{conditionally independent} of the input $x$ in general. \nAccordingly, as shown in Figure \\ref{fig:noise_adaption_process}, the noisy adaptation layer is added at the top of the base DNN to model the noise transition matrix parameterized by $\\mathcal{W}$. This layer should be removed when test data is to be predicted. \\looseness=-1\n\\begin{figure}[t!]\n\\begin{center}\n\\includegraphics[width=8.5cm]{figures/methodology/noise_adaption_layer.pdf}\n\\end{center}\n\\vspace*{-0.4cm}\n\\caption{Noise modeling process using the noise adaptation layer.}\n\\label{fig:noise_adaption_process}\n\\vspace*{-0.3cm}\n\\end{figure}\n\\smallskip\n\\noindent \\underline{Technical Detail}: \\emph{Webly learning} first trains the base DNN only for easy examples retrieved by search engines; subsequently, the confusion matrix for all training examples is used as the initial weight $\\mathcal{W}$ of the noise adaptation layer. It fine-tunes the entire model in an end-to-end manner for hard training examples. In contrast, the \\emph{noise model} initializes $\\mathcal{W}$ to an identity matrix and adds a regularizer to force $\\mathcal{W}$ to diffuse during DNN training. The \\emph{dropout noise model} applies dropout regularization to the adaptation layer, whose output is normalized by the softmax function to implicitly diffuse $\\mathcal{W}$. The \\emph{s-model} is similar to the \\emph{dropout noise model} but dropout is not applied. The \\emph{c-model} is an extension of the s-model that models the instance-dependent noise, which is more realistic than the symmetric and asymmetric noises. Meanwhile, \\emph{NLNN} adopts the EM algorithm to iterate the E-step to estimate the noise transition matrix and the M-step to back-propagate the DNN. \\looseness=-1\n\\smallskip\n\\noindent {\\underline{Remark}: A common drawback of this family is their inability to identify false-labeled examples, treating all the examples equally. Thus, the estimation error for the transition matrix is generally large when only noisy training data is used or when the noise rate is high .}\nMeanwhile, for the EM-based method, becoming stuck in local optima is inevitable, and high computational costs are incurred .\n\\vspace*{0.15cm}","id":"439a17f4-972c-4739-b63c-f2fd25ca8426","level":"subsubsection","origin_cites_number":6,"parent_id":"f16bcc10-a9fa-4749-b206-8aa388990240","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Robust Architecture"],["subsubsection","Noise Adaptation Layer"]],"subsections":[],"title":"Noise Adaptation Layer"},{"cite_extract_rate":0.6000000000000001,"cites":[4138,7773,8738],"content":"} \n{Beyond the label-dependent label noise, several studies have been conducted to support more complex noise, leading to the design of dedicated architectures .} They typically aimed at increasing the reliability of estimating the label transition probability to handle more complex and realistic label noise. \n\\smallskip\n\\noindent \\underline{Technical Detail}: \\emph{Probabilistic noise modeling} manages two independent networks, each of which is specialized to predict the noise type and label transition probability.\nBecause an EM-based approach with random initialization is impractical for training the entire network, both networks are trained with massive noisy labeled data after the pre-training step with a small amount of clean data.\nMeanwhile, \\emph{masking} is a human-assisted approach to convey the human cognition of invalid label transitions.\nConsidering that noisy labels are mainly from the interaction between humans and tasks, the invalid transition investigated by humans was leveraged to constrain the noise modeling process. Owing to the difficulty in specifying the explicit constraint, a variant of generative adversarial networks\\,(GANs) was employed in this study. Recently, the \\emph{contrastive-additive noise network} was proposed to adjust incorrectly estimated label transition probabilities by introducing a new concept of quality embedding, which models the trustworthiness of noisy labels. {\\emph{RoG} builds a simple yet robust generative classifier on top of any discriminative DNN pre-trained on noisy data.}\n\\smallskip\n\\noindent \\underline{Remark}: Compared with the noise adaptation layer, this family of methods significantly improves the robustness to more diverse types of label noise, but it cannot be easily extended to other architectures in general.","id":"81044d2e-5255-4dbd-8274-7c1d26bb1d92","level":"subsubsection","origin_cites_number":5,"parent_id":"f16bcc10-a9fa-4749-b206-8aa388990240","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Robust Architecture"],["subsubsection","Dedicated Architecture"]],"subsections":[],"title":"Dedicated Architecture"},{"cite_extract_rate":0.4,"cites":[4142,71],"content":"\\label{sec:robust_regularization}\nRegularization methods have been widely studied to improve the generalizability of a learned model in the machine learning community . \nBy avoiding overfitting in model training, the robustness to label noise improves with widely-used regularization techniques such as \\emph{data augmentation} , \\emph{weight decay} , \\emph{dropout} , and \\emph{batch normalization} . {These canonical regularization methods operate well on moderately noisy data, but they alone do \\emph{not sufficiently} improve the test accuracy}; poor generalization could be obtained when the noise is heavy . Thus, more advanced regularization techniques have been recently proposed, which further improved robustness to label noise when used along with the canonical methods. The main advantage of this family is its \\emph{flexibility} in collaborating with other directions because it only requires simple modifications. \n\\smallskip","id":"25e15182-1991-4f1a-83d9-017fe1a40d14","level":"subsection","origin_cites_number":5,"parent_id":"60bd8c81-37fb-4934-ab18-bb9a24e80724","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Robust Regularization"]],"subsections":["84774a8d-19b4-4ef8-ba90-419f51709a46","50da4bbd-f1a0-49f9-b327-8d43650a94c6"],"title":"Robust Regularization"},{"cite_extract_rate":0.6666666666666661,"cites":[4147,4131,7133,4142],"content":"}\n{\nThe regularization can be an {explicit} form that modifies the expected training loss, e.g., weight decay and dropout. \n}\n\\smallskip\n\\noindent \\underline{Technical Detail}:\n\\emph{Bilevel learning} uses a clean validation dataset to regularize the overfitting of a model by introducing a bilevel optimization approach, which differs from the conventional one in that its regularization constraint is also an optimization problem. Overfitting is controlled by adjusting the weights on each mini-batch and selecting their values such that they minimize the error on the validation dataset. Meanwhile, \\emph{annotator confusion} assumes the existence of multiple annotators and introduces a regularized EM-based approach to model the label transition probability; its regularizer enables the estimated transition probability to converge to the true confusion matrix of the annotators. \nIn contrast, \\emph{pre-training} empirically proves that fine-tuning on a pre-trained model provides a significant improvement in robustness compared with models trained from scratch; the universal representations of pre-training prevent the model parameters from being updated in the wrong direction by noisy labels. { \\emph{PHuber} proposes a composite loss-based gradient clipping, which is a variation of standard gradient clipping for label noise robustness. \\emph{Robust early-learning} classifies critical parameters and non-critical parameters for fitting clean and noise labels, respectively. Then, it penalizes only the non-critical ones with a different update rule.} \n{\\emph{ODLN} leverages open-set auxiliary data and prevents the overfitting to noisy labels by assigning random labels to the open-set examples, which are uniformly sampled from the label set.}\n\\smallskip\n\\noindent {\\underline{Remark}: The explicit regularization often introduces sensitive model-dependent hyperparameters or requires deeper architectures to compensate for the reduced capacity, yet it can lead to significant performance gain if they are optimally tuned. \\looseness=-1}\n\\smallskip","id":"84774a8d-19b4-4ef8-ba90-419f51709a46","level":"subsubsection","origin_cites_number":6,"parent_id":"25e15182-1991-4f1a-83d9-017fe1a40d14","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Robust Regularization"],["subsubsection","Explicit Regularization"]],"subsections":[],"title":"Explicit Regularization"},{"cite_extract_rate":1,"cites":[7191,4141,4148,892],"content":"}\nThe regularization can also be an {implicit} form that gives the effect of stochasticity, e.g., data augmentation and mini-batch stochastic gradient descent.\n\\smallskip\n\\noindent \\underline{Technical Detail}: \\emph{Adversarial training} enhances the noise tolerance by encouraging the DNN to correctly classify both original inputs and hostilely perturbed ones. {\\emph{Label smoothing} estimates the marginalized effect of label noise during training, thereby reducing overfitting by preventing the DNN from assigning a full probability to noisy training examples. Instead of the one-hot label, the noisy label is mixed with a uniform mixture over all possible labels,\n\\begin{equation}\n\\begin{gathered}\n\\bar{y} = \\big\\langle \\bar{y}(1), \\bar{y}(2), \\dots, \\bar{y}(c) \\big\\rangle,\\\\ \n\\text{where} ~ \\bar{y}(i) = (1-\\alpha) \\cdot [\\tilde{y} = i] + \\alpha / c ~\\text{and}~ \\alpha \\in [0, 1].\n\\end{gathered}\n\\end{equation}\nHere, $[\\cdot]$ is the Iverson bracket and $\\alpha$ is the smoothing degree.} In contrast, \\emph{mixup} regularizes the DNN to favor simple linear behaviors in between training examples. \nFirst, the mini-batch is constructed using virtual training examples, each of which is formed by the linear interpolation of two noisy training examples $(x_i, \\tilde{y}_i)$ and $(x_j, \\tilde{y}_j)$ obtained at random from noisy training data $\\tilde{\\mathcal{D}}$,\n\\begin{equation}\n\\label{eq:mixup_construction}\n{x}_{mix} = \\lambda x_i + (1-\\lambda) x_j~~ \\text{and} ~~{y}_{mix} = \\lambda \\tilde{y}_i + (1-\\lambda) \\tilde{y}_j,\n\\end{equation}\nwhere $\\lambda \\in [0,1]$ is the balance parameter between two examples. Thus, \\emph{mixup} extends the training distribution by updating the DNN for the constructed mini-batch. \n\\smallskip\n\\noindent {\\underline{Remark}: The implicit regularization improves the generalization capability of the DNN without reducing the representational capacity. It also does not introduce sensitive model-dependent hyperparameters because it is applied to the training data. However, the extended feature or label space slows down the convergence of training.}","id":"50da4bbd-f1a0-49f9-b327-8d43650a94c6","level":"subsubsection","origin_cites_number":4,"parent_id":"25e15182-1991-4f1a-83d9-017fe1a40d14","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Robust Regularization"],["subsubsection","Implicit Regularization"]],"subsections":[],"title":"Implicit Regularization"},{"cite_extract_rate":0.6000000000000001,"cites":[4130,3454,4132,4151,4150,4128,4149,8736,4134],"content":"\\label{sec:robust_loss_function}\n{\nIt was proven that a learned DNN with a \\emph{suitably modified} loss function $\\ell^{\\prime}$ for noisy data $\\tilde{\\mathcal{D}}$ can approach the Bayes optimal classifier $f^{*}$, which achieves the optimal Bayes risk $\\mathcal{R}^{*} = \\mathcal{R}_{\\mathcal{D}}(f^{*})$ for clean data $\\mathcal{D}$. Let $\\hat{f} = \\text{argmin}_{f \\in \\mathcal{F}} \\hat{\\mathcal{R}}_{\\ell^{\\prime}, \\tilde{\\mathcal{D}}}(f)$ be the learned classifier with the modified loss $\\ell^{\\prime}$ for the noisy data, where $\\hat{\\mathcal{R}}_{\\ell^{\\prime}, \\tilde{\\mathcal{D}}}(f) = \\mathbb{E}_{\\tilde{\\mathcal{D}}}[\\ell(f(x;\\Theta), \\tilde{y})]$. If $\\ell$ is $L$-Lipschitz and classification-calibrated , with probability at least $1\\!-\\!\\delta$, there exists a non-decreasing function $\\zeta_{\\ell}$ with $\\zeta_{\\ell}(0)=0$ such that \\looseness=-1\n\\vspace*{-0.6cm}\n\\begin{equation}\n\\begin{split}\n\\mathcal{R}_{\\mathcal{D}}(\\hat{f}) - \\mathcal{R}^{*} \\leq &~\\overbrace{\\zeta_{\\ell} \\Big( \\text{min}_{f\\in \\mathcal{F}}\\mathcal{R}_{\\ell,\\mathcal{D}}(f) - \\text{min}_{f}\\mathcal{R}_{\\ell, \\mathcal{D}}(f)}^{\\text{Approximation and Estimation Errors}}\\\\\n&~~+ 4L_{p}\\text{RC}(\\mathcal{F}) + 2\\sqrt{{\\text{log}(1/\\delta)}\\big/{2|\\mathcal{D}|}} \\Big),\\!\\!\\!\\!\\!\n\\end{split}\n\\end{equation}\n$L_{p}$ is the Lipschitz constant of $\\ell^{\\prime}$ and RC is the Rademacher complexity of the hypothesis class $\\mathcal{F}$. Then, by the universal approximation theorem , the Bayes optimal classifier $f^*$ is guaranteed to be in the hypothesis class $\\mathcal{F}$ with DNNs. \nBased on this theoretical foundation, researchers have attempted to design robust loss functions such that they achieve a small risk for unseen clean data even when noisy labels exist in the training data . \n}\n\\smallskip\n\\noindent \\underline{Technical Detail}: \nInitially, Manwani and Sastry theoretically proved a sufficient condition for the loss function such that risk minimization with that function becomes noise-tolerant for binary classification. Subsequently, the sufficient condition was extended for multi-class classification using deep learning . Specifically, a loss function is defined to be \\emph{noise-tolerant} for a $c$-class classification under \\emph{symmetric} noise if the function satisfies the noise rate $\\tau<\\frac{c-1}{c}$ and\n\\begin{equation}\n\\label{eq:symmetric_function}\n\\sum_{j=1}^{c}\\ell\\big(f(x;\\Theta),y=j\\big)=C, ~\\forall x\\in\\mathcal{X}, ~\\forall f,\n\\end{equation}\nwhere $C$ is a constant. This condition guarantees that the classifier trained on noisy data has the same misclassification probability as that trained on noise-free data under the specified assumption. {An extension for \\emph{multi-label} classification was provided by Kumar et al. .} Moreover, if $\\mathcal{R}_{\\mathcal{D}}(f^{*})=0$, then the function is also noise-tolerant under an \\emph{asymmetric} noise, where $f^{*}$ is a global risk minimizer of $\\mathcal{R}_{\\mathcal{D}}$. \\looseness=-1\nFor the classification task, the categorical cross entropy\\,(CCE) loss is the most widely used loss function owing to its fast convergence and high generalization capability. However, in the presence of noisy labels, the \\emph{robust MAE} showed that the mean absolute error\\,(MAE) loss achieves better generalization than the CCE loss because only the MAE loss satisfies the aforementioned condition. A limitation of the MAE loss is that its generalization performance degrades significantly when complicated data are involved. Hence, the \\emph{generalized cross entropy}\\,(GCE) was proposed to achieve the advantages of both MAE and CCE losses; the GCE loss is a more general class of noise-robust loss that encompasses both of them. \n{\nAmid et al. extended the GCE loss by introducing two temperatures based on the Tsallis divergence. \\emph{Bi-tempered loss} introduces a proper unbiased generalization of the CE loss based on the Bregman divergence.}\nIn addition, inspired by the symmetricity of the Kullback-Leibler divergence, the symmetric cross entropy\\,(SCE) was proposed by combining a noise tolerance term, namely reverse cross entropy loss, with the standard CCE loss. \nMeanwhile, the \\emph{curriculum loss}\\,(CL) is a surrogate loss of the 0-1 loss function; it provides a tight upper bound and can easily be extended to multi-class classification. {The \\emph{active passive loss}\\,(APL) is a combination of two types of robust loss functions, an active loss that maximizes the probability of belonging to the given class and a passive loss that minimizes the probability of belonging to other classes.}\n\\smallskip\n\\noindent \\underline{Remark}: The robustness of these methods is theoretically supported well. However, they perform well only in simple cases, when learning is easy or the number of classes is small . Moreover, the modification of the loss function increases the training time for convergence .\n\\vspace*{-0.2cm}","id":"59ab6098-f130-49dc-b87f-123eaa6be482","level":"subsection","origin_cites_number":15,"parent_id":"60bd8c81-37fb-4934-ab18-bb9a24e80724","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Robust Loss Function"]],"subsections":[],"title":"Robust Loss Function"},{"cite_extract_rate":0.6666666666666661,"cites":[7162,3340,7775,4135,4145,4139],"content":"\\label{sec:loss_adjustment}\n\\vspace*{-0.0cm}\nLoss adjustment is effective for reducing the negative impact of noisy labels by adjusting the loss of all training examples before updating the DNN . The methods associated with it can be categorized into three groups depending on their adjustment philosophy: \\emph{1)} {\\emph{loss correction}} that estimates the noise transition matrix to correct the forward or backward loss, \\emph{2)} {\\emph{loss reweighting}} that imposes different importance to each example for a weighted training scheme, \\emph{3)} {\\emph{label refurbishment}} that adjusts the loss using the refurbished label obtained from a convex combination of noisy and predicted labels, and \\emph{4)} {\\emph{meta learning}} that automatically infers the optimal rule for loss adjustment. {Unlike the robust loss function newly designed for robustness, this family of methods aims to make the traditional optimization process robust to label noise. Hence, in the middle of training, the {update rule} is adjusted such that the negative impact of label noise is minimized.}\nIn general, loss adjustment allows for a \\emph{full exploration} of the training data by adjusting the loss of every example. However, the error incurred by \\emph{false} correction is accumulated, especially when the number of classes or the number of mislabeled examples is large . \n\\vspace*{0.15cm}","id":"3d40345d-b6d6-4b98-943c-ec552f10d7fb","level":"subsection","origin_cites_number":9,"parent_id":"60bd8c81-37fb-4934-ab18-bb9a24e80724","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Loss Adjustment"]],"subsections":["26fbf349-d8eb-41b0-9964-18115fa4704c","67aed536-256b-4b94-954e-fe79fe9aa42b","6322f675-ca49-4162-8e1c-216bde9693b8","5cc6d843-bb17-416b-bbf4-7e089941b605"],"title":"Loss Adjustment"},{"cite_extract_rate":0.6666666666666661,"cites":[7777,4145,4153,4152],"content":"}\n\\label{sec:loss_correction}\nSimilar to the noise adaptation layer presented in Section \\ref{sec:robust_architecture}, this approach modifies the loss of each example by multiplying the estimated label transition probability by the output of a specified DNN. The main difference is that the learning of the transition probability is decoupled from that of the model. \n\\smallskip\n\\noindent \\underline{Technical Detail}:\n\\emph{Backward correction} initially approximates the noise transition matrix using the softmax output of the DNN trained without loss correction. Subsequently, it retrains the DNN while correcting the original loss based on the estimated matrix. The corrected loss of a example $(x,\\tilde{y})$ is computed by a linear combination of its loss values for observable labels, whose coefficient is the inverse transition matrix $\\text{T}^{-1}$ to the observable label $y\\in\\{1,\\dots, c\\}$, given its target label $\\tilde{y}$.\nTherefore, the backward correction $\\cev{\\ell}$ is performed by multiplying the inverse transition matrix to the prediction for all the observable labels, \n{\n\\begin{equation}\n\\small\n\\label{eq:backward_correction}\n\\begin{split}\n\\cev{\\ell}\\big(f(x;&\\Theta),\\tilde{y}\\big) = \\hat{\\text{T}}^{-1}\\Big\\langle\\ell\\big(f(x;\\Theta), 1\\big), \\dots, \\ell\\big(f(x;\\Theta), c\\big)\\Big\\rangle^{\\!\\top}\\!\\!,\n\\end{split}\n\\end{equation}\n}\n\\noindent where $\\hat{\\text{T}}$ is the estimated noise transition matrix. \nConversely, \\emph{forward correction} uses a linear combination of a DNN's softmax outputs before applying the loss function. Hence, the forward correction $\\vec{\\ell}$ is performed by multiplying the estimated transition probability with the softmax outputs during the forward propagation step,\n{\n\\begin{equation}\n\\small\n\\label{eq:forward_correction}\n\\begin{split}\n\\vec{\\ell}\\big(f(x;\\Theta),\\tilde{y}\\big)& = \\ell\\Big(\\Big\\langle\\hat{p}(\\tilde{y}|1),\\dots,\\hat{p}(\\tilde{y}|c)\\Big\\rangle f(x;\\Theta)^{\\top},\\tilde{y}\\Big)\\\\\n&=\\ell\\big(\\hat{\\text{T}}^{\\top}f(x;\\Theta)^{\\top},\\tilde{y}\\big).\n\\end{split}\n\\end{equation}\n}\n\\vspace*{-0.2cm}\nFurthermore, \\emph{gold loss correction} assumes the availability of clean validation data or anchor points for loss correction. Thus, a more accurate transition matrix is obtained by using them as additional information, which further improves the robustness of the loss correction. {Recently, \\emph{T-Revision} provides a solution that can infer the transition matrix without anchor points, and \\emph{Dual T} factorizes the matrix into the product of two easy-to-estimate matrices to avoid directly estimating the noisy class posterior.}\n{\nBeyond the instance-independent noise assumption, Zhang et al. introduced the instance-confidence embedding to model instance-dependent noise in estimating the transition matrix. On the other hand, Yang et al. proposed to use the Bayes optimal transition matrix estimated from the distilled examples for the instance-dependent noise transition matrix.}\n\\smallskip\n\\noindent {\\underline{Remark}: The robustness of these approaches is highly dependent on how precisely the transition matrix is estimated. To acquire such a transition matrix, they require prior knowledge in general, such as anchor points or clean validation data.}\n\\vspace*{0.15cm}","id":"26fbf349-d8eb-41b0-9964-18115fa4704c","level":"subsubsection","origin_cites_number":6,"parent_id":"3d40345d-b6d6-4b98-943c-ec552f10d7fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Loss Adjustment"],["subsubsection","Loss Correction"]],"subsections":[],"title":"Loss Correction"},{"cite_extract_rate":0.5,"cites":[3453,7775],"content":"} Inspired by the concept of importance reweighting , loss reweighting aims to assign smaller weights to the examples with false labels and greater weights to those with true labels. Accordingly, the reweighted loss on the mini-batch $\\mathcal{B}_t$ is used to update the DNN, \n{\n\\begin{equation}\n\\label{eq:loss_reweighting}\n\\Theta_{t+1} = \\Theta_{t} - \\eta\\nabla \\Big( \\frac{1}{|\\mathcal{B}_{t}|}\\!\\sum_{(x,\\tilde{y})\\in\\mathcal{B}_t}\\!\\!\\!\\!\\overbrace{w(x,\\tilde{y})\\ell\\big(f(x;\\Theta_t), \\tilde{y}\\big)}^{\\text{Reweighted Loss}}\\Big),\n\\end{equation}}\nwhere $w(x,\\tilde{y})$ is the weight of an example $x$ with its noisy label $\\tilde{y}$. Hence, the examples with smaller weights do not significantly affect the DNN learning.\n\\smallskip\n\\noindent \\underline{Technical Detail}:\nIn \\emph{importance reweighting} , the ratio of two joint data distributions {$w(x,\\tilde{y})=P_{\\mathcal{D}}(x,\\tilde{y})/P_{\\tilde{\\mathcal{D}}}(x,\\tilde{y})$} determines the contribution of the loss of each noisy example. An approximate solution to estimate the ratio was developed because the two distributions are difficult to determine from noisy data. Meanwhile, \\emph{active bias} emphasizes uncertain examples with inconsistent label predictions by assigning their prediction variances as the weights for training. \n{\n\\emph{DualGraph} employs graph neural networks and reweights the examples according to the structural relations among labels, eliminating the abnormal noise examples.}\n\\smallskip\n\\noindent {\\underline{Remark}: These approaches need to manually pre-specify the weighting function as well as there additional hyper-parameters, which is fairly hard to be applied in practice due to the significant variation of appropriate weighting schemes that rely on the noise type and training data.}\n\\vspace*{0.15cm}","id":"67aed536-256b-4b94-954e-fe79fe9aa42b","level":"subsubsection","origin_cites_number":4,"parent_id":"3d40345d-b6d6-4b98-943c-ec552f10d7fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Loss Adjustment"],["subsubsection","Loss Reweighting"]],"subsections":[],"title":"Loss Reweighting"},{"cite_extract_rate":0.75,"cites":[7162,4135,4156,4154,4155,4139],"content":"} Refurbishing a noisy label $\\tilde{y}$ effectively prevents overfitting to false labels. Let $\\hat{y}$ be the current prediction of the DNN $f(x;\\Theta)$. Therefore, the refurbished label $y^{refurb}$ can be obtained by a convex combination of the noisy label $\\tilde{y}$ and the DNN prediction $\\hat{y}$,\n\\begin{equation}\n\\label{eq:label_correction}\ny^{refurb} = \\alpha \\tilde{y} + (1-\\alpha) \\hat{y},\n\\end{equation}\nwhere $\\alpha \\in [0,1]$ is the label confidence of $\\tilde{y}$. To mitigate the damage of incorrect labeling, this approach backpropagates the loss for the refurbished label instead of the noisy one, thereby yielding substantial robustness to noisy labels. \n\\smallskip\n\\noindent \\underline{Technical Detail}:\n\\emph{Bootstrapping} is the first method that proposes the concept of label refurbishment to update the target label of training examples. It develops a more coherent network that improves its ability to evaluate the consistency of noisy labels, with the label confidence $\\alpha$ obtained via cross-validation.\n\\emph{Dynamic bootstrapping} dynamically adjusts the confidence $\\alpha$ of individual training examples. The confidence $\\alpha$ is obtained by fitting a two-component and one-dimensional beta mixture model to the loss distribution of all training examples. {\\emph{Self-adaptive training} applies the exponential moving average to alleviate the instability issue of using instantaneous prediction of the current DNN, \n\\begin{equation}\n\\!\\!y_{t+1}^{refurb} = \\alpha y_t^{refurb} + (1-\\alpha)\\hat{y}, ~\\text{where}~ y_0^{refurb} = \\tilde{y}\\!\\!\n\\end{equation}\n}\n\\emph{D2L} trains a DNN using a dimensionality-driven learning strategy to avoid overfitting to false labels. A simple measure called \\emph{local intrinsic dimensionality} is adopted to evaluate the confidence $\\alpha$ in considering that the overfitting is exacerbated by dimensional expansion. Hence, refurbished labels are generated to prevent the dimensionality of the representation subspace from expanding at a later stage of training. \nRecently, \\emph{SELFIE} introduces a novel concept of \\emph{refurbishable examples} that can be corrected with high precision. The key idea is to consider the example with consistent label predictions as refurbishable because such consistent predictions correspond to its true label with a high probability owing to the learner's perceptual consistency. Accordingly, the labels of only refurbishable examples are corrected to minimize the number of falsely corrected cases. {Similarly, \\emph{AdaCorr} selectively refurbishes the label of noisy examples, but a theoretical error-bound is provided. Alternatively, \\emph{SEAL} averages the softmax output of a DNN on each example over the whole training process, then re-trains the DNN using the averaged soft labels.}\n\\smallskip\n\\noindent {\\underline{Remark}: Differently from loss correction and reweighting, all the noisy labels are explicitly replaced with other expected clean labels (or their combination). If there are not many confusing classes in data, these methods work well by refurbishing the noisy labels with high precision. In the opposite case, the DNN could overfit to wrongly refurbished labels.}\n\\vspace*{0.15cm}","id":"6322f675-ca49-4162-8e1c-216bde9693b8","level":"subsubsection","origin_cites_number":8,"parent_id":"3d40345d-b6d6-4b98-943c-ec552f10d7fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Loss Adjustment"],["subsubsection","Label Refurbishment"]],"subsections":[],"title":"Label Refurbishment"},{"cite_extract_rate":0.8888888888888881,"cites":[4157,4158,4128,3345,1695,7774,7772,8737],"content":"}\n{In recent years, meta learning becomes an important topic in the machine learning community and is applied to improve noise robustness . The key concept is \\emph{learning to learn} that performs learning at a level higher than conventional learning, thus achieving data-agnostic and noise type-agnostic rules for better practical use. It is similar to loss reweighting and label refurbishment, but the adjustment is automated in a meta learning manner.\n\\smallskip\n\\noindent \\underline{Technical Detail}:\nFor the loss reweighting in Eq.~\\eqref{eq:loss_reweighting}, the goal is to learn the weight function $w(x,\\tilde{y})$. Specifically, \\emph{L2LWS} and \\emph{CWS} are unified neural architectures composed of a target DNN and a meta-DNN. The meta-DNN is trained on a small clean validation dataset; it then provides guidance to evaluate the weight score for the target DNN. Here, part of the two DNNs are shared and jointly trained to benefit from each other. \\emph{Automatic reweighting} is a meta learning algorithm that learns the weights of training examples based on their gradient directions. It includes a small clean validation dataset into the training dataset and reweights the backward loss of the mini-batch examples such that the updated gradient minimizes the loss of this validation dataset. \\emph{Meta-weight-net} parameterizes the weighting function as a multi-layer perceptron network with only one hidden layer. A meta-objective is defined to update its parameters such that they minimize the empirical risk of a small clean dataset. At each iteration, the parameter of the target network is guided by the weight function updated via the meta-objective. \n{Likewise, \\emph{data coefficients}\\,(i.e., exemplar weights and true labels) are estimated by meta-optimization with a small clean set, which is only $0.2$\\% of the entire training set, while refurbishing the examples probably mislabeled.}\nFor the label refurbishment in Eq.~\\eqref{eq:label_correction}, \\emph{knowledge distillation} adopts the technique of transferring knowledge from one expert model to a target model. The prediction from the expert DNN trained on small clean validation data is used instead of the prediction $\\hat{y}$ from the target DNN. \n{\\emph{MLC} updates the target model with corrected labels provided by a meta model trained on clean validation data. The two models are trained concurrently via a bi-level optimization.} \n\\smallskip\n\\noindent \\underline{Remark}:\nBy learning the update rule via meta learning, the trained network easily adapts to various types of data and label noise. Nevertheless, unbiased clean validation data is essential to minimize the auxiliary objective, although it may not be available in real-world data.}\n\\vspace*{-0.1cm}","id":"5cc6d843-bb17-416b-bbf4-7e089941b605","level":"subsubsection","origin_cites_number":9,"parent_id":"3d40345d-b6d6-4b98-943c-ec552f10d7fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Loss Adjustment"],["subsubsection","Meta Learning"]],"subsections":[],"title":"Meta Learning"},{"cite_extract_rate":0.631578947368421,"cites":[3340,4137,7771,4140,4126,3342,4125,4124,4123,4253,8630,4115],"content":"\\label{sec:sample_selection}\nTo avoid any false corrections, many recent studies have adopted sample selection that involves selecting true-labeled examples from a noisy training dataset. In this case, the update equation in Eq.\\ \\eqref{eq:corrupted_update} is modified to render a DNN more robust for noisy labels. Let $\\mathcal{C}_t \\subseteq \\mathcal{B}_t$ be the identified \\emph{clean} examples at time $t$. Then, the DNN is updated only for the selected clean examples $\\mathcal{C}_t$, \n\\begin{equation}\n\\label{eq:clean_update}\n\\Theta_{t+1} = \\Theta_{t} - \\eta\\nabla\\Big(\\frac{1}{|\\mathcal{C}_t|} \\!\\sum_{(x,\\tilde{y}) \\in \\mathcal{C}_t} \\!\\!\\!\\!\\ell\\big(f(x;\\Theta_{t}), \\tilde{y}\\big)\\Big),\n\\end{equation}\nwhere the rest mini-batch examples, which are likely to be false-labeled, are excluded to pursue robust learning. \n\\newcommand{\\norm}[1]{\\left\\lVert#1\\right\\rVert}\n{The {memorization nature} of DNNs has been explored theoretically and empirically to identify clean examples from noisy training data .\nSpecifically, assuming clusterable data where the clusters are located on the unit Euclidean ball, Li et al. proved the distance from the initial weight ${W}_{0}$ to the weight ${W}_t$ after $t$ iterations,\n\\begin{equation}\n\\label{eq:memorization_effect_foundation}\n\\norm{{W}_t - {W}_0}_{F} \\lesssim \\big( \\sqrt{K} + (K^{2}\\epsilon_{0}/\\norm{{C}}^{2})t \\big),\n\\end{equation}\nwhere $\\norm{\\cdot}_{F}$ is the Frobenius norm, $K$ is the number of clusters, and ${C}$ is the set of cluster centers reaching all input examples within their $\\epsilon_0$ neighborhood.\nEq.\\ \\eqref{eq:memorization_effect_foundation} demonstrates that the weights of DNNs start to stray far from the initial weights when overfitting to corrupted labels, while they are still in the vicinity of the initial weights at an early stage of training . In the empirical studies , the \\emph{memorization effect} is also observed since DNNs tend to first learn simple and generalized patterns and then gradually overfit to all noisy patterns. As such, favoring small-loss training examples as the clean ones are commonly employed to design robust training methods .\nLearning with sample selection is well motivated and works well in general, but this approach suffers from accumulated error caused by incorrect selection, especially when there are many ambiguous classes in training data.} \nHence, recent approaches often leverage multiple DNNs to cooperate with one another or run multiple training rounds .\nMoreover, to benefit from even false-labeled examples, loss correction or semi-supervised learning have been recently combined with the sample selection strategy .\n\\begin{figure}\n\\vspace*{+0.08cm}\n\\begin{center}\n\\includegraphics[width=8.7cm]{figures/methodology/loss_distribution.pdf}\n\\end{center}\n\\vspace*{-0.2cm}\n\\hspace*{0.80cm} {\\small (a) Symmetric Noise $40\\%$.} \\hspace*{0.525cm} {\\small (b) Asymmetric Noise $40\\%$.}\n\\vspace*{-0.15cm}\n\\caption{Loss distribution of training examples at the training accuracy of $50\\%$ on noisy CIFAR-100. (This figure is adapted from Song et al. .)}\n\\label{fig:loss_distribution}\n\\vspace*{-0.4cm}\n\\end{figure}\n\\smallskip\\smallskip","id":"54d73157-e47b-4943-9b30-b518e1806946","level":"subsection","origin_cites_number":19,"parent_id":"60bd8c81-37fb-4934-ab18-bb9a24e80724","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Sample Selection"]],"subsections":["a00439ee-902b-4911-93a8-2198ddea3d64","3f9409b8-29c2-4f75-91a3-9614418f8bb4","3b9e7951-da54-4090-b74a-113ced54d832"],"title":"Sample Selection"},{"cite_extract_rate":0.8,"cites":[3340,4140,4137,4159],"content":"}\n{\nCollaborative learning and co-training are widely used for the multi-network training. Consequently, the sample selection process is guided by the mentor network in the case of collaborative learning or the peer network in the case of co-training.\n}\n\\smallskip\n\\noindent \\underline{Technical Detail}:\nInitially, \\emph{Decouple} proposes the decoupling of when to update from how to update. Hence, two DNNs are maintained simultaneously and updated only the examples selected based on a disagreement between the two DNNs.\nNext, due to the memorization effect of DNNs, many researchers have adopted another selection criterion, called a \\emph{small-loss} trick, which treats a certain number of small-loss training examples as true-labeled examples; many true-labeled examples tend to exhibit smaller losses than false-labeled examples, as illustrated in Figure \\ref{fig:loss_distribution}(a). In \\emph{MentorNet} , a pre-trained mentor network guides the training of a student network in a collaborative learning manner. Based on the small-loss trick, the mentor network provides the student network with examples whose labels are likely to be correct. \\emph{Co-teaching} and \\emph{Co-teaching+} also maintain two DNNs, but each DNN selects a certain number of small-loss examples and feeds them to its peer DNN for further training. \\emph{Co-teaching+} further employs the disagreement strategy of \\emph{Decouple} compared with \\emph{Co-teaching}. {In contrast, \\emph{JoCoR} reduces the diversity of two networks via co-regularization, making predictions of the two networks closer.}\n\\smallskip\n\\noindent {\\underline{Remark}: The co-training methods help reduce the confirmation bias , which is a hazard of favoring the examples selected at the beginning of training, while the increase in the number of learnable parameters makes their learning pipeline inefficient. In addition, the small-loss trick does not work well when the loss distribution of true-labeled and false-labeled examples largely overlap, as in the asymmetric noise in Figure \\ref{fig:loss_distribution}(b).}\n\\smallskip\\smallskip","id":"a00439ee-902b-4911-93a8-2198ddea3d64","level":"subsubsection","origin_cites_number":5,"parent_id":"54d73157-e47b-4943-9b30-b518e1806946","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Sample Selection"],["subsubsection","Multi-network Learning"]],"subsections":[],"title":"Multi-network Learning"},{"cite_extract_rate":0.75,"cites":[4126,3343,8630,4125,7163,4123],"content":"}\nWithout maintaining additional DNNs, multi-round learning iteratively refines the selected set of clean examples by repeating the training round. Thus, the selected set keeps improved as the number of rounds increases. \n\\smallskip\n\\noindent \\underline{Technical Detail}:\n\\emph{ITLM} iteratively minimizes the trimmed loss by alternating between selecting true-labeled examples at the current moment and retraining the DNN using them. At each training round, only a fraction of small-loss examples obtained in the current round are used to retrain the DNN in the next round. \\emph{INCV} randomly divides noisy training data and then employs cross-validation to classify true-labeled examples while removing large-loss examples at each training round. Here, \\emph{Co-teaching} is adopted to train the DNN on the identified examples in the final round of training. {Similarly, \\emph{O2U-Net} repeats the whole training process with the cyclical learning rate until enough loss statistics of every examples are gathered. Next, the DNN is re-trained from scratch only for the clean data where false-labeled examples have been detected and removed based on statistics.} \\looseness=-1\nA number of variations have been proposed to achieve high performance using iterative refinement only in a single training round.\nBeyond the small-loss trick, \\emph{iterative detection} detects false-labeled examples by employing the local outlier factor algorithm . With a Siamese network, it gradually pulls away false-labeled examples from true-labeled samples in the deep feature space. \\emph{MORPH} introduces the concept of memorized examples which is used to iteratively expand an initial safe set into a maximal safe set via self-transitional learning. \\emph{TopoFilter} utilizes the spatial topological pattern of learned representations to detect true-labeled examples, not relying on the prediction of the noisy classifier.\n{\n\\emph{NGC} iteratively constructs the nearest neighbor graph using latent representations and performs geometry-based sample selection by aggregating information from neighborhoods. Soft pesudo-labels are assigned to the examples not selected.\n}\n\\smallskip\n\\noindent {\\underline{Remark}: The selected clean set keeps expanded and purified with iterative refinement, mainly through multi-round learning. As a side effect, the computational cost for training increases linearly for the number of training rounds.}\n\\smallskip","id":"3f9409b8-29c2-4f75-91a3-9614418f8bb4","level":"subsubsection","origin_cites_number":8,"parent_id":"54d73157-e47b-4943-9b30-b518e1806946","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Sample Selection"],["subsubsection","Multi-round Learning"]],"subsections":[],"title":"Multi-round Learning"},{"cite_extract_rate":0.7741935483870961,"cites":[3340,3343,7133,4137,7773,3345,8738,4159,4134,4132,4140,4141,7775,7778,4126,3342,4156,4142,7772,4150,8737,4155,4138,4130,7191,4151,4147,4135,7162,4157,4145,7774,4139,4125,4123,892,4128,4152,4253,4158,7777,4131,8630,7163,4133,4154,2277,4136],"content":"} {An inherent limitation of sample selection is to discard all the \\emph{unselected} training examples, thus resulting in a \\emph{partial} exploration of training data. To exploit all the noisy examples, researchers have attempted to combine sample selection with other orthogonal ideas.\n\\begin{figure}[t!]\n\\begin{center}\n\\includegraphics[width=8.45cm]{figures/methodology/semi_supervised.pdf}\n\\end{center}\n\\vspace*{-0.1cm}\n\\hspace*{0.2cm} \\small{(a) Noisy Data.} \\hspace*{0.4cm} \\small{(b) Transformed Data.} \\hspace*{0.8cm} \\small{(c) SSL.}\n\\vspace*{-0.2cm}\n\\caption{{Procedures for semi-supervised learning under label noise.}}\n\\label{fig:noise_semi_supervised}\n\\vspace*{-0.45cm}\n\\end{figure}\n\\smallskip\n\\noindent \\underline{Technical Detail}:\nThe most prominent method in this direction is combining a specific sample selection strategy with a specific semi-supervised learning model. As illustrated in Figure \\ref{fig:noise_semi_supervised}, selected examples are treated as labeled clean data, whereas the remaining examples are treated as unlabeled. Subsequently, semi-supervised learning is performed using the transformed data. \\emph{SELF} is combined with a semi-supervised learning approach to progressively filter out false-labeled examples from noisy data. By maintaining the running average model called the {mean-teacher} as the backbone, it obtains the self-ensemble predictions of all training examples and then progressively removes examples whose ensemble predictions do not agree with their annotated labels. This method further leverages unsupervised loss from the examples not included in the selected clean set. \\emph{DivideMix} uses two-component and one-dimensional Gaussian mixture models to transform noisy data into labeled (clean) and unlabeled (noisy) sets. Then, it applies a semi-supervised technique \\emph{MixMatch} . Recently, \\emph{RoCL} employs two-phase learning strategies: supervised training on selected clean examples and then semi-supervised learning on relabeled noisy examples with self-supervision. For selection and relabeling, it computes the exponential moving average of the loss over training iterations. \n}\n\\clearpage\n{\n\\newcolumntype{L}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}m{#1}}\n\\newcolumntype{X}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}p{#1}}\n\\newcolumntype{Y}[1]{>{\\let\\newline\\\\\\arraybackslash\\hspace{1pt}}m{#1}}\n\\begin{table*}[h]\n\\vspace*{-3.9cm}\n\\caption{{Comparison of proposed robust deep learning methods with respect to the following six properties: (P1)\\,Flexibility, (P2)\\,No Pre-training, (P3)\\,Full Exploration, (P4)\\,No Supervision, (P5)\\,Heavy Noise, and (P6)\\,Complex Noise.}}\n\\vspace*{-0.2cm}\n\\begin{tabular}{L{0.7cm}|L{2.5cm}|Y{4.7cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|Y{2.9cm}}\n\\toprule\n\\multicolumn{2}{c|}{\\textbf{Category}} & \\hspace*{2.0cm}\\textbf{Method} & \\textbf{P1} & \\textbf{P2} & \\textbf{P3} & \\textbf{P4} & \\textbf{P5} & \\textbf{P6} & \\,\\,\\,\\,\\,\\,\\,\\,\\,\\textbf{Implementation}\\!\\! \\\\ \\hline\n\\multirow{9}{*}{\\vspace*{-0.35cm}\\hspace*{0.13cm}\\!\\!\\!\\!\\!\\!\\rotatebox[origin=c]{90}{\\makecell[l]{Robust Architecture\\\\\\hspace*{0.66cm}(\\textsection \\ref{sec:robust_architecture})}}\\!\\!\\!\\!\\!} \n& \\multirow{6}{*}{\\vspace*{-0.2cm}\\makecell[l]{Noisy Adaptation\\\\\\hspace*{0.67cm}Layer}} \n& \\emph{Webly Learning} & \\cellcolor{gray!10}$\\bigtriangleup$ & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(Caffe)\\protect\\footnotemark[1] \\\\\\cline{3-10} \n& & \\emph{Noise Model} & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Unofficial\\,(Keras)\\protect\\footnotemark[2] \\\\\\cline{3-10}\n& & \\emph{Dropout Noise Model} & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(MATLAB)\\protect\\footnotemark[3] \\\\\\cline{3-10}\n& & \\emph{S-model} & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(Keras)\\protect\\footnotemark[4]\\\\\\cline{3-10}\n& & \\emph{C-model} & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & Official\\,(Keras)\\protect\\footnotemark[4] \\\\\\cline{3-10}\n& & \\emph{NLNN} & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Unofficial\\,(Chainer)\\protect\\footnotemark[5] \\\\\\cmidrule{2-10}\n& \\multirow{4}{*}{\\makecell[l]{\\hspace*{0.12cm}Dedicated\\\\Architecture}} \n& \\emph{Probablistic Noise Model} & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & $ \\cellcolor{gray!10}\\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & Official\\,(Caffe)\\protect\\footnotemark[6] \\\\\\cline{3-10}\n& & \\emph{Masking} & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & $ \\cellcolor{gray!10}\\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & Official\\,(TensorFlow)\\protect\\footnotemark[7] \\\\\\cline{3-10}\n& & \\emph{Contrastive-Additive Noise Network} & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark &\\cellcolor{blue!10}\\cmark & $ \\cellcolor{gray!10}\\bigtriangleup$ &\\cellcolor{blue!10}\\cmark & N/A \\\\\\cline{3-10}\n& & {\\emph{RoG}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark &\\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark &$ \\cellcolor{gray!10}\\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[8] \\\\\\midrule\n\\multirow{9}{*}{\\vspace*{-0.25cm}\\hspace*{0.13cm}\\!\\!\\!\\!\\!\\!\\rotatebox[origin=c]{90}{\\makecell[l]{Robust Regularization\\\\\\hspace*{0.8cm}(\\textsection \\ref{sec:robust_regularization})}}\\!\\!\\!\\!\\!} \n& \\multirow{6}{*}{\\makecell[l]{\\hspace*{0.35cm}Explicit\\\\Regularization}} \n& \\emph{Bilevel Learning} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[9] \\\\\\cline{3-10}\n& & \\emph{Annotator Confusion} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[10] \\\\\\cline{3-10}\n& & \\emph{Pre-training} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[11] \\\\\\cline{3-10}\n& & {\\emph{PHuber}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(PyTorch)\\protect\\footnotemark[12] \\\\\\cline{3-10}\n& & {\\emph{Robust Early-learning}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[13] \\\\\\cline{3-10}\n& & {\\emph{ODLN}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[14] \\\\\\cmidrule{2-10}\n& \\multirow{3}{*}{\\makecell[l]{\\hspace*{0.35cm}Implicit\\\\Regularization}} \n& \\emph{Adversarial Training} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(PyTorch)\\protect\\footnotemark[15] \\\\\\cline{3-10}\n& & \\emph{Label Smoothing} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(PyTorch)\\protect\\footnotemark[16] \\\\\\cline{3-10}\n& & \\emph{Mixup} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[17] \\\\\\midrule\n\\multicolumn{2}{c|}{\\multirow{7}{*}{{\\makecell[l]{Robust Loss Function\\\\\\hspace*{0.75cm}(\\textsection \\ref{sec:robust_loss_function})}}}} \n& \\emph{Robust MAE} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & N/A \\\\\\cline{3-10}\n\\multicolumn{2}{c|}{} & \\emph{Generalized Cross Entropy} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Unofficial\\,(PyTorch)\\protect\\footnotemark[18] \\\\\\cline{3-10}\n\\multicolumn{2}{c|}{} & \\emph{Symmetric Cross Entropy} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(Keras)\\protect\\footnotemark[19] \\\\\\cline{3-10}\n\\multicolumn{2}{c|}{} & {\\emph{Bi-tempered Loss}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$\\bigtriangleup$ & \\cellcolor{gray!10}$\\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[20] \\\\\\cline{3-10}\n\\multicolumn{2}{c|}{} & \\emph{Curriculum Learning} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$\\bigtriangleup$ & N/A \\\\\\cline{3-10}\n\\multicolumn{2}{c|}{} & {\\emph{Active Passive Loss}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$\\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[21] \\\\\\midrule\n\\multirow{21}{*}{\\vspace*{-0.7cm}\\hspace*{0.13cm}\\!\\!\\!\\!\\!\\!\\rotatebox[origin=c]{90}{\\makecell[l]{Loss Adjustment\\\\\\hspace*{0.5cm}(\\textsection \\ref{sec:loss_adjustment})}}\\!\\!\\!\\!\\!} \n& \\multirow{5}{*}{\\makecell[l]{Loss Correction}} \n& \\emph{Backward Correction} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(Keras)\\protect\\footnotemark[22] \\\\\\cline{3-10}\n& & \\emph{Forward Correction} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(Keras)\\protect\\footnotemark[22] \\\\\\cline{3-10}\n& & \\emph{Gold Loss Correction} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(PyTorch)\\protect\\footnotemark[23] \\\\\\cline{3-10}\n& & {\\emph{T-revision}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & Official\\,(PyTorch)\\protect\\footnotemark[24] \\\\\\cline{3-10}\n& & {\\emph{Dual T}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{red!10}\\xmark & N/A \\\\\\cmidrule{2-10}\n& \\multirow{3}{*}{\\makecell[l]{Loss Reweigting}} \n& \\emph{Importance Reweighting} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(PyTorch)\\protect\\footnotemark[25] \\\\\\cline{3-10}\n& & \\emph{Active Bias} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(TensorFlow)\\protect\\footnotemark[26]\\!\\! \\\\\\cline{3-10}\n& & {\\emph{DualGraph}} & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & N/A \\\\\\cmidrule{2-10}\n& \\multirow{6}{*}{\\makecell[l]{Label Refurbishment}} \n& \\emph{Bootstrapping} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(Keras)\\protect\\footnotemark[27] \\\\\\cline{3-10}\n& & \\emph{Dynamic Bootstrapping} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[28] \\\\\\cline{3-10}\n& & {\\emph{Self-adaptive Training}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[29] \\\\\\cline{3-10}\n& & \\emph{D2L} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(Keras)\\protect\\footnotemark[30] \\\\\\cline{3-10}\n& & {\\emph{AdaCorr}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[31] \\\\\\cline{3-10}\n& & {\\emph{SEAL}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & Official\\,(PyTorch)\\protect\\footnotemark[32] \\\\\\cmidrule{2-10}\n& \\multirow{7}{*}{\\makecell[l]{Meta Learning}} \n& \\emph{L2LWS} & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(TensorFlow)\\protect\\footnotemark[33]\\!\\! \\\\\\cline{3-10}\n& & \\emph{CWS} & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & N/A \\\\\\cline{3-10}\n& & \\emph{Automatic Reweighting} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[34] \\\\\\cline{3-10}\n& & \\emph{Meta-weight-net} & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[35] \\\\\\cline{3-10}\n& & {\\emph{Data Coefficients}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[36] \\\\\\cline{3-10}\n& & \\emph{Knowledge Distillation} & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & N/A \\\\\\cline{3-10}\n& & {\\emph{MLC}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & Official\\,(PyTorch)\\protect\\footnotemark[37] \\\\\\midrule\n\\multirow{13}{*}{\\vspace*{-0.45cm}\\hspace*{0.13cm}\\!\\!\\!\\!\\!\\!\\rotatebox[origin=c]{90}{\\makecell[l]{Sample Selection\\\\\\hspace*{0.5cm}(\\textsection \\ref{sec:sample_selection})}}\\!\\!\\!\\!\\!} \n& \\multirow{4}{*}{\\makecell[l]{Multi-Network\\\\\\hspace*{0.33cm}Learning}} \n& \\emph{Decouple} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[38] \\\\\\cline{3-10}\n& & \\emph{MentorNet} & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[39] \\\\\\cline{3-10}\n& & \\emph{Co-teaching} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[40] \\\\\\cline{3-10}\n& & \\emph{Co-teaching+} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[41] \\\\\\cline{3-10}\n& & {\\emph{JoCoR}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[42] \\\\\\cmidrule{2-10}\n& \\multirow{6}{*}{\\makecell[l]{Multi-Round\\\\\\hspace*{0.23cm}Learning}} \n& \\emph{ITLM} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(GluonCV)\\protect\\footnotemark[43] \\\\\\cline{3-10}\n& & \\emph{INCV} & \\cellcolor{blue!10}\\cmark &\\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(Keras)\\protect\\footnotemark[44] \\\\\\cline{3-10}\n& & {\\emph{O2U-Net}} & \\cellcolor{blue!10}\\cmark &\\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Unofficial\\,(PyTorch)\\protect\\footnotemark[45] \\\\\\cline{3-10}\n& & \\emph{Iterative Detection} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(Keras)\\protect\\footnotemark[46] \\\\\\cline{3-10}\n& & {\\emph{MORPH}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & N/A \\\\\\cline{3-10}\n& & {\\emph{TopoFilter}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[47] \\\\\\cline{3-10}\n& & {\\emph{NGC}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & N/A\\\\\\cmidrule{2-10}\n& \\multirow{3}{*}{\\vspace*{-0.24cm}\\makecell[l]{Hybrid Approach}} \n& \\emph{SELFIE} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(TensorFlow)\\protect\\footnotemark[48] \\\\\\cline{3-10}\n& & \\emph{SELF} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & N/A \\\\\\cline{3-10}\n& & \\emph{DivideMix} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & Official\\,(PyTorch)\\protect\\footnotemark[49] \\\\\\cline{3-10}\n& & {\\emph{RoCL}} & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & N/A \\\\\\bottomrule\n\\end{tabular}\n\\label{table:all_comparision}\n\\vspace*{-4cm}\n\\end{table*}\n}\n\\clearpage\n{\n\\footnotetext[1]{\\url{https://github.com/endernewton/webly-supervised}}\n\\footnotetext[2]{\\url{https://github.com/delchiaro/training-cnn-noisy-labels-keras}}\n\\footnotetext[3]{\\url{https://github.com/ijindal/Noisy_Dropout_regularization}}\n\\footnotetext[4]{\\url{https://github.com/udibr/noisy_labels}}\n\\footnotetext[5]{\\url{https://github.com/Ryo-Ito/Noisy-Labels-Neural-Network}}\n\\footnotetext[6]{\\url{https://github.com/Cysu/noisy_label}}\n\\footnotetext[7]{\\url{https://github.com/bhanML/Masking}}\n\\footnotetext[8]{\\url{https://github.com/pokaxpoka/RoGNoisyLabel}}\n\\footnotetext[9]{\\url{https://github.com/sjenni/DeepBilevel}}\n\\footnotetext[10]{\\url{https://rt416.github.io/pdf/trace_codes.pdf}}\n\\footnotetext[11]{\\url{github.com/hendrycks/pre-training}}\n\\footnotetext[12]{\\url{https://github.com/dmizr/phuber}}\n\\footnotetext[13]{\\url{https://github.com/xiaoboxia/CDR}}\n\\footnotetext[14]{\\url{https://github.com/hongxin001/ODNL?ref=pythonrepo.com}}\n\\footnotetext[15]{\\url{https://https://github.com/sarathknv/adversarial-examples-pytorch}}\n\\footnotetext[16]{\\url{https://github.com/CoinCheung/pytorch-loss}}\n\\footnotetext[17]{\\url{https://github.com/facebookresearch/mixup-cifar10}}\n\\footnotetext[18]{\\url{https://github.com/AlanChou/Truncated-Loss}}\n\\footnotetext[19]{\\href{https://github.com/YisenWang/symmetric\\_cross\\_entropy\\_for\\_noisy_label}{https://github.com/YisenWang/symmetric\\_cross\\_entropy}}\n\\footnotetext[20]{\\url{https://github.com/google/bi-tempered-loss}}\n\\footnotetext[21]{\\url{https://github.com/HanxunH/Active-Passive-Losses}}\n\\footnotetext[22]{\\url{https://github.com/giorgiop/loss-correction}}\n\\footnotetext[23]{\\url{https://github.com/mmazeika/glc}}\n\\footnotetext[24]{\\url{https://github.com/xiaoboxia/T-Revision}}\n\\footnotetext[25]{\\href{https://github.com/xiaoboxia/Classification-with-noisy-labels-by-importance-reweighting}{https://github.com/xiaoboxia/Classification-with-noisy-labels}}\n\\newcolumntype{L}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}m{#1}}\n\\newcolumntype{X}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}p{#1}}\n\\begin{table*}[!htb]\n\\centering\n\\vspace*{-0.1cm}\n\\caption{Comparison of robust deep learning categories for overcoming noisy labels.}\n\\vspace*{-0.2cm}\n\\begin{tabular}{L{2.2cm} | L{2.85cm}|X{1.5cm}|X{1.55cm}|X{1.8cm}|X{1.65cm}|X{1.5cm}|X{1.6cm}}\n\\toprule\n\\multicolumn{2}{>{}c|}{\\multirow{2}{*}{\\textbf{Category}}} & \\textbf{P1} & \\textbf{P2} & \\textbf{P3} & \\textbf{P4} & \\textbf{P5} & \\textbf{P6} \\\\\n\\multicolumn{2}{>{}c|}{} & \\!\\!\\!{Flexibility}\\!\\!\\! & \\!\\!\\!{No Pre-train}\\!\\!\\! & \\!\\!\\!{Full Exploration}\\!\\!\\! & \\!\\!\\!{No Supervision}\\!\\!\\! & \\!\\!\\!{Heavy Noise}\\!\\!\\! & \\!\\!\\!{Complex Noise}\\!\\!\\! \\\\\\midrule\n\\multirow{2}{*}\n{\\makecell[l]{\\!\\!\\!Robust Architecture\\!\\!\\! \\\\\\hspace*{0.5cm}(\\textsection \\ref{sec:robust_architecture})}}\n& \\!\\!Noise Adaptation Layer & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark \\\\\n\\cline{2-8}\n& \\!\\!Dedicated Architecture & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{blue!10}\\cmark\\\\ \\midrule\n\\multirow{2}{*}\n{\\makecell[l]{\\!\\!\\!\\!\\!\\!Robust Regularization\\!\\!\\!\\!\\!\\! \\\\\\hspace*{0.5cm}(\\textsection \\ref{sec:robust_regularization})}}\n& \\!\\!Implicit Regularization & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ \\\\\n\\cline{2-8}\n& \\!\\!Explicit Regularization & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ \\\\ \\midrule\n\\multicolumn{2}{>{}c|}{\\hspace*{0cm}Robust Loss Function (\\textsection \\ref{sec:robust_loss_function}) }& \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark \\\\\\midrule\n\\multirow{2}{*}\n{\\vspace*{-0.58cm}\\!\\makecell[l]{Loss Adjustment \\\\\\hspace*{0.52cm}(\\textsection \\ref{sec:loss_adjustment})}}\n& \\!\\!Loss Correction &\\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark &\\cellcolor{red!10}\\xmark \\\\\n\\cline{2-8}\n& \\!\\!Loss Reweighting & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$\\\\\n\\cline{2-8}\n& \\!\\!Label Refurbishment & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$\\\\\\cline{2-8}\n& \\!\\!Meta Learning & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{gray!10}$ \\bigtriangleup$ & \\cellcolor{gray!10}$ \\bigtriangleup$ \\\\ \\midrule\n{\\vspace*{-0.55cm}\\hspace*{0.075cm}\\makecell[l]{Sample Selection\\\\\\hspace*{0.55cm}(\\textsection \\ref{sec:sample_selection})}}\n& \\!\\!Multi-Network Learning &\\cellcolor{blue!10}\\cmark &\\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$ \\\\\n\\cline{2-8}\n& \\!\\!Multi-Round Learning & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{red!10}\\xmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$\\\\\\cline{2-8}\n& \\!\\!Hybrid Approach & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{blue!10}\\cmark & \\cellcolor{gray!10}$ \\bigtriangleup$\\\\\\bottomrule\n\\end{tabular}\n\\label{table:direction_comparison}\n\\vspace*{-0.4cm}\n\\end{table*}\n}\nMeanwhile, \\emph{SELFIE} is a hybrid approach of sample selection and loss correction. The loss of refurbishable examples is corrected (i.e., loss correction) and then used together with that of small-loss examples (i.e., sample selection). Consequently, more training examples are considered for updating the DNN.\nThe \\emph{curriculum loss\\,(CL)} is combined with the robust loss function approach and used to extract the true-labeled examples from noisy data.\n\\smallskip\n\\noindent \\underline{Remark}:\nNoise robustness is significantly improved by combining with other techniques. However, the hyperparameters introduced by these techniques render a DNN more susceptible to changes in data and noise types, and an increase in computational cost is inevitable","id":"3b9e7951-da54-4090-b74a-113ced54d832","level":"subsubsection","origin_cites_number":62,"parent_id":"54d73157-e47b-4943-9b30-b518e1806946","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Deep Learning Approaches"],["subsection","Sample Selection"],["subsubsection","Hybrid Approach"]],"subsections":[],"title":"Hybrid Approach"},{"cite_extract_rate":0.5,"cites":[3340],"content":"\\label{sec:comparison}\nIn this section, we compare the {$62$} deep learning methods for overcoming noisy labels introduced in Section \\ref{sec:methodology} with respect to the following \\emph{six} properties. When selecting the properties, we refer to the properties that are typically used to compare the performance of robust deep learning methods . To the best of our knowledge, this survey is the first to provide a systematic comparison of robust training methods. \nThis comprehensive comparison will provide useful insights that can enlighten new future directions.\n\\vspace*{0.1cm}\n\\begin{itemize}[leftmargin=9pt]\n\\item \\textbf{(P1)\\,Flexibility:} With the rapid evolution of deep learning research, a number of new network architectures are constantly emerging and becoming available. Hence, the ability to support any type of architecture is important. \\enquote{Flexibility} ensures that the proposed method can quickly adapt to the state-of-the-art architecture.\n\\vspace*{0.1cm}\n\\item \\textbf{(P2)\\,No Pre-training:} A typical approach to improve noise robustness is to use a pre-trained network; however, this incurs an additional computational cost to the learning process. \\enquote{No Pre-training} ensures that the proposed method can be trained from scratch without any pre-training.\n\\vspace*{0.1cm}\n\\item \\textbf{(P3)\\,Full Exploration:} Excluding unreliable examples from the update is an effective method for robust deep learning; however, it eliminates hard but useful training examples as well. \\enquote{Full Exploration} ensures that the proposed methods can use \\emph{all} training examples without severe overfitting to false-labeled examples by adjusting their training losses or applying semi-supervised learning.\n\\vspace*{0.1cm}\n\\item \\textbf{(P4)\\,No Supervision:} Learning with supervision, such as a clean validation set or a known noise rate, is often impractical because they are difficult to obtain. Hence, such supervision had better be avoided to increase practicality in real-world scenarios. \\enquote{No Supervision} ensures that the proposed methods can be trained without any supervision.\n\\vspace*{0.1cm}\n\\item \\textbf{(P5)\\,Heavy Noise:} In real-world noisy data, the noise rate can vary from light to heavy. Hence, learning methods should achieve consistent noise robustness with respect to the noise rate. \\enquote{Heavy Noise} ensures that the proposed methods can combat even the heavy noise.\n\\end{itemize}\n{\n\\footnotetext[26]{\\url{https://github.com/songhwanjun/ActiveBias}}\n\\footnotetext[27]{\\url{https://github.com/dr-darryl-wright/Noisy-Labels-with-Bootstrapping}}\n\\footnotetext[28]{\\url{https://github.com/PaulAlbert31/LabelNoiseCorrection}}\n\\footnotetext[29]{\\url{https://github.com/LayneH/self-adaptive-training}}\n\\footnotetext[30]{\\url{https://github.com/xingjunm/dimensionality-driven-learning}}\n\\footnotetext[31]{\\url{https://github.com/pingqingsheng/LRT}}\n\\footnotetext[32]{\\url{https://github.com/chenpf1025/IDN}}\n\\footnotetext[33]{\\url{https://github.com/krayush07/learn-by-weak-supervision}}\n\\footnotetext[34]{\\url{https://github.com/uber-research/learning-to-reweight-examples}}\n\\footnotetext[35]{\\url{https://github.com/xjtushujun/meta-weight-net}}\n\\footnotetext[36]{\\url{https://github.com/google-research/google-research/tree/master/ieg}}\n\\footnotetext[37]{\\url{https://aka.ms/MLC}}\n\\footnotetext[38]{\\url{https://github.com/emalach/UpdateByDisagreement}}\n\\footnotetext[39]{\\url{https://github.com/google/mentornet}}\n\\footnotetext[40]{\\url{https://github.com/bhanML/Co-teaching}}\n\\footnotetext[41]{\\url{https://github.com/bhanML/coteaching_plus}}\n\\footnotetext[42]{\\url{https://github.com/hongxin001/JoCoR}}\n\\footnotetext[43]{\\url{https://github.com/yanyao-shen/ITLM-simplecode}}\n\\footnotetext[44]{\\url{https://github.com/chenpf1025/noisy_label_understanding_utilizing}}\n\\footnotetext[45]{\\url{https://github.com/hjimce/O2U-Net}}\n\\footnotetext[46]{\\url{https://github.com/YisenWang/Iterative_learning}}\n\\footnotetext[47]{\\url{https://github.com/pxiangwu/TopoFilter}}\n\\footnotetext[48]{\\url{https://github.com/kaist-dmlab/SELFIE}}\n\\footnotetext[49]{\\url{https://github.com/LiJunnan1992/DivideMix}}\n}\n\\begin{itemize}[leftmargin=9pt]\n\\item \\textbf{(P6)\\,Complex Noise:} The type of label noise significantly affects the performance of a learning method. To manage real-world noisy data, diverse types of label noise should be considered when designing a robust training method. \\enquote{Complex Noise} ensures that the proposed method can combat even the complex label noise.\n\\end{itemize}\n\\vspace*{0.05cm}\nTable \\ref{table:all_comparision} shows a comparison of all robust deep learning methods, which are grouped according to the most appropriate category. In the first row, the aforementioned six properties are labeled as P1--P6, and the availability of open-source implementation is added in the last column. For each property, we assign \\enquote{\\cmark} if it is completely supported, \\enquote{\\xmark} if it is not supported, and \\enquote{$\\bigtriangleup$} if it is supported but not completely. \nMore specifically, \\enquote{$\\bigtriangleup$} is assigned to P1 if the method can be flexible but requires additional effort, to P5 if the method can combat only moderate label noise, {and to P6 if the method does not make a strict assumption about the noise type but without explicitly modeling instance-dependent noise. Thus, for P6, the method marked with \\enquote{\\xmark} only deals with the instance-independent noise, while the method marked with \\enquote{\\cmark} deals with both instance-independent and -dependent noises.} The remaining properties\\,(i.e., P2, P3, and P4) are only assigned \\enquote{\\cmark} or \\enquote{\\xmark}. Regarding the implementation, we assign \\enquote{N/A} if a publicly available source code is not available.\nNo existing method supports all the properties. Each method achieves noise robustness by supporting a different combination of the properties. The supported properties are similar among the methods of the same (sub-)category because those methods share the same methodological philosophy; however, they differ significantly depending on the (sub-)category. Therefore, we investigate the properties generally supported in each (sub-)category and summarize them in Table \\ref{table:direction_comparison}. Here, the property of a (sub-)category is marked as the majority of the belonging methods. If no clear trend is observed among those methods, then the property is marked \\enquote{$\\bigtriangleup$}.","id":"3b9c1099-a2f2-417d-a753-e663d7ce62fa","level":"section","origin_cites_number":2,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Methodological Comparison"]],"subsections":[],"title":"Methodological Comparison"},{"cite_extract_rate":0.9,"cites":[4130,3340,4126,4135,7777,3453,4123,4160,4152],"content":"\\label{sec:noise_rate}\n{\nThe estimation of a noise rate is an imperative part of utilizing robust methods for better practical use, especially with the approaches belonging to the loss adjustment and sample selection. The estimated noise rate is widely used to reweight examples for a robust classifier or to determine how many examples should be selected as clean ones . However, detailed analysis has yet to be performed properly, though many robust approaches highly rely on the accuracy of noise rate estimation. The noise rate can be estimated by exploiting the inferred noise transition matrix , the Gaussian mixture model , or the cross-validation .\n}\n\\begin{comment}\n1. Yao, Y., Liu, T., Han, B., Gong, M., Deng, J., Niu, G., & Sugiyama, M.\nDual T: Reducing estimation error for transition matrix in label-noise learning. Advances in Neural Information Processing Systems 33, pp. 7260-7271, 2020. \n-> noise transition matrix\n2. Xia, X., Liu, T., Wang, N., Han, B., Gong, C., Niu, G., & Sugiyama, M.\nAre anchor points really indispensable in label-noise learning? Advances in Neural Information Processing Systems 32, pp. 6835-6846, 2019.\n-> noise transition matrix\n3. Li, X., Liu, T., Han, B., Niu, G., & Sugiyama, M.\nProvably end-to-end label-noise learning without anchor points.\n-> noise transition matrix\nProceedings of 38th International Conference on Machine Learning (ICML2021), pp. 6403-6413, 2021.\n4. Patrini, Giorgio et al. “Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach.” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017): 2233-2241.\n-> noise transition matrix\n5. Liu, Tongliang, and Dacheng Tao. \"Classification with noisy labels by importance reweighting.\" IEEE Transactions on pattern analysis and machine intelligence 38.3 (2015): 447-461.\n6. Scott, Clayton, Gilles Blanchard, and Gregory Handy. \"Classification with asymmetric label noise: Consistency and maximal denoising.\" Conference on learning theory. PMLR, 2013.\n\\end{comment}\n\\vspace*{-0.15cm}","id":"8d7ed0bf-5e07-45ed-a147-190ba1a209fb","level":"section","origin_cites_number":10,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Noise Rate Estimation"]],"subsections":["841668d4-4892-46d7-ae08-7e210e7c2add","39b0234a-04ab-44e2-98d0-fa20ed4e0f74","8516d834-e901-490a-aa9d-864c2f267139","685f1250-3a8b-48b6-b510-042be23d64f6"],"title":"Noise Rate Estimation"},{"cite_extract_rate":0.5,"cites":[7777,4238,4145,4161,652,3453,7769,4160,4152],"content":"{\nThe noise transition matrix has been used to build a statistically consistent robust classifier because it represents the class posterior probabilities for noisy and clean data, as in Eq.\\,\\eqref{eq:label_transition_matrix}. The first method to estimate the noise rate is exploiting this noise transition matrix, which can be inferred or trained accurately by using perfectly clean examples, i.e., \\emph{anchor points} ; an example $x$ with its label $i$ is defined as an anchor point if $p(y=i|x)=1$ and $p(y=k|x)=0$ for $k \\neq i$. Thus, let $\\mathcal{A}_i$ be the set of anchor points with label $i$, then the element of the noise transition matrix $T_{ij}$ is estimated by \\looseness=-1\n\\begin{equation}\n\\begin{split}\n\\hat{T}_{ij} &= \\frac{1}{|\\mathcal{A}_{i}|}\\sum_{x\\in\\mathcal{A}_{i}} \\sum_{k=1}^{c} p(\\tilde{y}=j|y=k)p(y=k|x) \\\\\n&= \\frac{1}{|\\mathcal{A}_{i}|}\\sum_{x\\in\\mathcal{A}_{i}} p(\\tilde{y}=j|x; \\Theta), \n\\end{split}\n\\end{equation}\nwhere $p(\\tilde{y}=j|x; \\Theta)$ is the noisy class posterior probability of the classifier trained on noisy training data for the anchor point $x$ (see the detailed proof in ). Next, based on the inferred noise transition matrix, the noise rate of a balanced training data is obtained by averaging the label transition probabilities between classes,\n\\begin{equation}\n\\hat{\\tau} = \\frac{1}{c} \\sum_{i=1}^{c} \\sum_{j\\neq i}^{c} p(\\tilde{y}=j | {y}=i) = \\frac{1}{c} \\sum_{i=1}^{c} \\sum_{j\\neq i}^{c} \\hat{T}_{ij}.\n\\end{equation}\nHowever, since the anchor points are typically unknown in real-world data, they are identified from noisy training data by either theoretical derivations or heuristics . \nIn addition, there have been recent efforts to learn the noise transition matrix without anchor points. \\emph{T-Revision} initializes a transition matrix by exploiting the examples with high noisy class posterior probabilities and then refines the matrix by adding a slack variable. \\emph{Dual-T} introduces an intermediate class that factorizes the transition matrix into two easy-to-estimate matrices for better accuracy. \\emph{VolMinNet} realizes an end-to-end framework and relaxes the need for anchor points under the sufficiently scattered assumption.\n}\n{\n\\newcolumntype{L}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}m{#1}}\n\\newcolumntype{X}[1]{>{\\centering\\let\\newline\\\\\\arraybackslash\\hspace{0pt}}p{#1}}\n\\newcolumntype{Y}[1]{>{\\let\\newline\\\\\\arraybackslash\\hspace{1pt}}m{#1}}\n\\begin{table*}[t!]\n\\small\n\\caption{Summary of publicly available datasets used for studying label noise.}\n\\vspace*{-0.7cm}\n\\begin{center}\n\\begin{tabular}{L{2.0cm} | Y{3.4cm} |X{1.9cm}| X{1.9cm} |X{1.9cm} |X{1.9cm} |Y{2.1cm}}\\toprule \n\\multicolumn{2}{c|}{\\textbf{Dataset}} & \\textbf{\\# Training} & \\textbf{\\# Validation} & \\textbf{\\# Testing} & \\textbf{\\# Classes} & \\textbf{Noise Rate\\,(\\%)} \\\\\\hline\n\\multirow{7}{*}{\\hspace*{0.13cm}\\!\\!\\!\\!\\!\\!{\\makecell[l]{Clean Data}}\\!\\!\\!\\!\\!} \n& {MNIST} \\footnote[50] & {60K}& N/A & {10K} & $10$ & $\\approx0.0$ \\\\\\cline{2-7}\n& {Fashion-MNIST} \\footnote[51]\\!\\!\\!\\!\\! & {60K}& N/A & {10K} & $10$ & $\\approx0.0$ \\\\\\cline{2-7}\n& {CIFAR-10} \\footnote[52] & {50K}& N/A & {10K} & $10$ & $\\approx0.0$ \\\\\\cline{2-7}\n& {CIFAR-100} \\footnote[52] & {50K}& N/A & {10K} & $100$ & $\\approx0.0$ \\\\\\cline{2-7}\n& {SVHN} \\footnote[53] & {73K}& N/A & {26K} & $10$ & $\\approx0.0$ \\\\\\cline{2-7}\n& {Tiny-ImageNet} \\footnote[55] & {100K}& {10K} & {10K} & $200$ & $\\approx0.0$ \\\\\\cline{2-7}\n& {ImageNet} \\footnote[54] & {1.3M}& {50K} & {50K} & $1000$ & $\\approx0.0$ \\\\\\midrule\n\\multirow{6}{*}{\\hspace*{0.13cm}\\!\\!\\!\\!\\!\\!{\\makecell[l]{Real-world\\\\\\hspace*{-0.02cm}Noisy Data}}\\!\\!\\!\\!\\!} \n& {ANIMAL-10N} \\footnote[56] & {50K}& N/A & {5K} & $10$ & $\\approx8.0$ \\\\\\cline{2-7}\n& {CIFAR-10N }\\footnote[57] & {50K}& N/A & {10K} & $10$ & $\\approx9.0/18.0/40.2$\\!\\!\\!\\!\\!\\!\\! \\\\\\cline{2-7}\n& {CIFAR-100N }\\footnote[57] & {50K}& N/A & {10K} & $100$ & $\\approx 25.6/40.2$ \\\\\\cline{2-7}\n& {Food-101N} \\footnote[58] & {310K}& {5K} & {25K} & $101$ & $\\approx18.4$ \\\\\\cline{2-7}\n& {Clothing1M} \\footnote[59] & {1M}& {14K} & {10K} & $14$ & $\\approx38.5$ \\\\\\cline{2-7}\n& {WebVision} \\footnote[60] & {2.4M}& {50K} & {50K} & $1000$ & $\\approx20.0$ \\\\\\bottomrule\n\\end{tabular}\n\\end{center}\n\\label{table:summary_dataset}\n\\vspace*{-0.6cm}\n\\end{table*}\n}\n\\vspace*{-0.15cm}","id":"841668d4-4892-46d7-ae08-7e210e7c2add","level":"subsection","origin_cites_number":18,"parent_id":"8d7ed0bf-5e07-45ed-a147-190ba1a209fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Noise Rate Estimation"],["subsection","Noise Transition Matrix"]],"subsections":[],"title":"Noise Transition Matrix"},{"cite_extract_rate":1,"cites":[4135],"content":"\\label{sec:gmm}\n\\begin{figure}[t!]\n\\begin{center}\n\\vspace*{0.1cm}\n\\includegraphics[width=8.8cm]{figures/evaluation/gmm_analysis.pdf}\n\\end{center}\n\\vspace*{-0.15cm}\n\\hspace*{1.3cm} \\small{(a) Symmetric Noise.} \\hspace*{1.2cm} \\small{(b) Asymmetric Noise.} \n\\vspace*{-0.15cm}\n\\caption{Training loss distributions of true-labeled and false-labeled examples using the ground-truth label and the GMM on CIFAR-100 data with two synthetic noises of $40\\%$.}\n\\label{fig:loss_gmm}\n\\vspace*{-0.4cm}\n\\end{figure}\nThe second method is exploiting a one-dimensional and two-component GMM to model the loss distribution of true-labeled and false-labeled examples . As shown in Figure \\ref{fig:loss_gmm}, since the loss distribution tends to be \\emph{bi-modal}, the two Gaussian components are fitted to the training loss by using the EM algorithm; the probability of an example being a false-labeled one is obtained through its posterior probability. Hence, the noise rate is estimated at each epoch $t$ by computing the expectation of the posterior probability for all training examples, \\looseness=-1\n\\begin{equation}\n\\label{eq:naive_gmm}\n\\begin{gathered}\n\\hat{\\tau} = \\mathbb{E}_{(x,\\tilde{y})\\in\\tilde{\\mathcal{D}}}\\Big[\\,p\\big(g\\,|\\,\\ell\\big(f(x;\\Theta_{t}), \\tilde{y}\\big)\\big)\\,\\Big],\\\\ \n\\end{gathered}\n\\end{equation}\nwhere $g$ is the Gaussian component with a larger loss. However, Pleiss et al. recently pointed out that the training loss becomes less separable by the GMM as the training progresses, and thus proposed the \\emph{area under the loss}\\,(AUL) curve, which is the sum of the example's training losses obtained from all previous training epochs. Even after the loss signal decays in later epochs, the distributions remain separable. Therefore, the noise rate is finally estimated by\n\\begin{equation}\n\\label{eq:aul_gmm}\n\\begin{gathered}\n\\hat{\\tau} = \\mathbb{E}_{(x,\\tilde{y})\\in\\tilde{\\mathcal{D}}}\\Big[\\,p\\big(g\\,|\\,{\\rm AUL}_{t}(x,\\tilde{y})\\big)\\,\\Big],\\\\ \n\\text{where}~ \\text{AUL}_{t}(x,\\tilde{y})= \\sum_{i=1}^{t}\\ell\\big(f(x;\\Theta_t),\\tilde{y}\\big).\n\\end{gathered}\n\\end{equation}\n\\vspace*{-0.4cm}","id":"39b0234a-04ab-44e2-98d0-fa20ed4e0f74","level":"subsection","origin_cites_number":1,"parent_id":"8d7ed0bf-5e07-45ed-a147-190ba1a209fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Noise Rate Estimation"],["subsection","Gaussian Mixture Model\\,(GMM)"]],"subsections":[],"title":"Gaussian Mixture Model\\,(GMM)"},{"cite_extract_rate":0.75,"cites":[3340,4140,4126],"content":"\\label{sec:cross_val}\nThe third method is estimating the noise rate by applying cross validation, which typically requires clean validation data . However, such clean validation data is hard to acquire in real-world applications. Thus, Chen et al. leveraged two randomly divided noisy training datasets for cross validation. Under the assumption that the two datasets share exactly the same noise transition matrix, the noise rate quantifies the test accuracy of DNNs that are respectively trained and tested on the two divided sets,\n\\begin{equation}\n\\label{eq:cross_val}\n\\!\\!\\!{\\rm Test\\,Accuracy} \\!= \\!\\!\\!\\\n\\begin{cases}\n(1 - \\hat{\\tau})^{2} + \\hat{\\tau}^{2} / (c - 1) \\!\\!&\\!\\! \\text{if symmetric}\\!\\!\\!\\!\\\\\n(1 - \\hat{\\tau})^{2} + \\hat{\\tau}^{2} \\!\\!& \\!\\!\\text{if asymmetric}.\\!\\!\\!\\!\n\\end{cases}\n\\end{equation}\nTherefore, the noise rate is estimated from the test accuracy obtained by cross validation.\n\\begin{comment}","id":"8516d834-e901-490a-aa9d-864c2f267139","level":"subsection","origin_cites_number":4,"parent_id":"8d7ed0bf-5e07-45ed-a147-190ba1a209fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Noise Rate Estimation"],["subsection","Cross Validation"]],"subsections":[],"title":"Cross Validation"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:nr_comp}\nTo compare the estimation performance of using the GMM and cross validation, we trained a WideResNet-16-8 for three benchmark datasets with varying noise rates. The results are plotted in Figure \\ref{fig:noise_rate_symmetric}. Generally, both methods performed well on the easy dataset\\,(i.e., CIFAR-10), but their performance worsened as the training difficulty increased from CIFAR-10 to Tiny-ImageNet because the true-labeled but hard examples are not clearly distinguishable from the false-labeled ones. Nevertheless, the GMM method showed considerably better performance than the cross validation method even in the two difficult datasets, CIFAR-100 and Tiny-ImageNet. \nOverall, this empirical analysis will be helpful for practitioners or researchers who design robust algorithms for noisy labels.\n\\begin{figure}[!t]\n\\vspace*{0.0cm}\n\\begin{minipage}[t]{1.0\\linewidth}\n\\begin{center}\n\\includegraphics[width=8.8cm]{figures/evaluation/noise_rate_symmetric.pdf}\n\\end{center}\n\\vspace*{-0.15cm}\n\\hspace*{2.1cm} \\small{(a) GMM.} \\hspace*{2.0cm} \\small{(b) Cross Validation.}\n\\vspace*{-0.2cm}\n\\caption{Noise rate estimation using the Gaussian mixture model and cross validation when training a WideResNet-16-8 on three datasets with symmetric noise.}\n\\label{fig:noise_rate_symmetric}\n\\end{minipage}\n\\vspace*{-0.35cm}\n\\end{figure}\n\\end{comment}\n\\vspace*{-0.15cm}","id":"685f1250-3a8b-48b6-b510-042be23d64f6","level":"subsection","origin_cites_number":0,"parent_id":"8d7ed0bf-5e07-45ed-a147-190ba1a209fb","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Noise Rate Estimation"],["subsection","Comparison of Noise Rate Estimation"]],"subsections":[],"title":"Comparison of Noise Rate Estimation"},{"cite_extract_rate":0,"cites":[],"content":"This section describes the typically used experimental design for comparing robust training methods in the presence of label noise. We introduce publicly available image datasets and then describe widely-used evaluation metrics.\n\\vspace*{-0.3cm}","id":"0788ae52-30c4-4156-89f9-6be2e443db83","level":"section","origin_cites_number":0,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Experimental Design"]],"subsections":["b05d418b-8b51-4e39-b585-58d8b1bcd7ff","6b4a1fff-c3c1-412a-b6f0-eeaa7d3721a4"],"title":"Experimental Design"},{"cite_extract_rate":0,"cites":[],"content":"To validate the robustness of the proposed algorithms, an image classification task was widely conducted on numerous image benchmark datasets. Table \\ref{table:summary_dataset} summarizes popularly-used public benchmark datasets, which are classified into two categories: \\emph{1)} a \\enquote{clean dataset} that consists of mostly true-labeled examples annotated by human experts and \\emph{2)} a \\enquote{real-world noisy dataset} that comprises real-world noisy examples with varying numbers of false labels.\n\\vspace*{0.15cm}","id":"b05d418b-8b51-4e39-b585-58d8b1bcd7ff","level":"subsection","origin_cites_number":0,"parent_id":"0788ae52-30c4-4156-89f9-6be2e443db83","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Experimental Design"],["subsection","Publicly Available Datasets"]],"subsections":["0c8efb50-e5a7-4707-a3b7-dcff2ec802f5","2e98f74b-5745-4f37-aea0-711f64e9e6c5"],"title":"Publicly Available Datasets"},{"cite_extract_rate":0.33333333333333304,"cites":[4253,8630,652],"content":"According to the literature , \\emph{seven} clean datasets are widely used: MNIST\\footnote[50]{\\url{http://yann.lecun.com/exdb/mnist}}, classification of handwritten digits ; Fashion-MNIST\\footnote[51]{\\url{https://github.com/zalandoresearch/fashion-mnist}}, classification of various clothing ; CIFAR-10\\footnote[52]{\\url{https://www.cs.toronto.edu/~kriz/cifar.html}} and CIFAR-100\\footnotemark[52], classification of a subset of $80$ million categorical images ; SVHN\\footnote[53]{\\url{http://ufldl.stanford.edu/housenumbers}}, classification of house numbers in Google Street view images ; ImageNet\\footnote[54]{\\url{http://www.image-net.org}} and Tiny-ImageNet\\footnote[55]{\\url{https://www.kaggle.com/c/tiny-imagenet}}, image database organized according to the WordNet hierarchy and its small subset . Because the labels in these datasets are almost all true-labeled, their labels in the training data should be artificially corrupted for the evaluation of synthetic noises, namely \\emph{symmetric} noise and \\emph{asymmetric} noise. \n\\vspace*{0.15cm}","id":"0c8efb50-e5a7-4707-a3b7-dcff2ec802f5","level":"subsubsection","origin_cites_number":9,"parent_id":"b05d418b-8b51-4e39-b585-58d8b1bcd7ff","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Experimental Design"],["subsection","Publicly Available Datasets"],["subsubsection","Clean Datasets"]],"subsections":[],"title":"Clean Datasets"},{"cite_extract_rate":0.5,"cites":[4238,4161,7769],"content":"Unlike the clean datasets, real-world noisy datasets inherently contain many mislabeled examples annotated by non-experts. According to the literature , \\emph{six} real-world noisy datasets are widely used: ANIMAL-10N\\footnote[56]{\\url{https://dm.kaist.ac.kr/datasets/animal-10n}}, real-world noisy data of human-labeled online images for 10 confusing animals ;\n{CIFAR-10N\\footnote[57]{\\url{http://noisylabels.com/}} and CIFAR-100N\\footnotemark[57], variations of CIFAR-10 and CIFAR-100 with human-annotated real-world noisy labels collected from Amazon’s Mechanical Turk . They provide human labels with different noise rates, as shown in Table \\ref{table:summary_dataset};}\nFood-101N\\footnote[58]{\\url{https://kuanghuei.github.io/Food-101N}}, real-world noisy data of crawled food images annotated by their search keywords in the Food-101 taxonomy ; Clothing1M\\footnote[59]{\\url{https://www.floydhub.com/lukasmyth/datasets/clothing1m}}, real-world noisy data of large-scale crawled clothing images from several online shopping websites ; WebVision\\footnote[60]{\\url{https://data.vision.ee.ethz.ch/cvl/webvision/download.html}}, real-world noisy data of large-scale web images crawled from Flickr and Google Images search . To support sophisticated evaluation, most real-world noisy datasets contain their own clean validation set and provide the estimated noise rate of their training set. \\looseness=-1","id":"2e98f74b-5745-4f37-aea0-711f64e9e6c5","level":"subsubsection","origin_cites_number":6,"parent_id":"b05d418b-8b51-4e39-b585-58d8b1bcd7ff","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Experimental Design"],["subsection","Publicly Available Datasets"],["subsubsection","Real-world Noisy Datasets"]],"subsections":[],"title":"Real-world Noisy Datasets"},{"cite_extract_rate":0.6000000000000001,"cites":[3340,4126,3630],"content":"A typical metric to assess the robustness of a particular method is the prediction accuracy for unbiased and clean examples that are not used in training. The prediction accuracy degrades significantly if the DNN overfits to false-labeled examples . Hence, \\emph{test accuracy} has generally been adopted for evaluation . For a test set $\\mathcal{T}=\\{(x_i,y_i)\\}_{i=1}^{|\\mathcal{T}|}$, let $\\hat{y}_i$ be the predicted label of the $i$-th example in $\\mathcal{T}$. Subsequently, the test accuracy is formalized by\n\\begin{equation}\n\\label{eq:test_accuracy}\n\\text{Test Accuracy} = \\frac{|\\{(x_i,y_i)\\in\\mathcal{T}:\\hat{y}_i=y_i\\}|}{|\\mathcal{T}|}.\n\\end{equation}\nIf the test data are not available, \\emph{validation accuracy} can be used by replacing $\\mathcal{T}$ in Eq.\\,\\eqref{eq:test_accuracy} with validation data $\\mathcal{V}=\\{(x_i,y_i)\\}_{i=1}^{|\\mathcal{V}|}$ as an alternative,\n\\begin{equation}\n\\label{eq:validation_accuracy}\n\\text{Validation Accuracy} = \\frac{|\\{(x_i,y_i)\\in\\mathcal{V}:\\hat{y}_i=y_i\\}|}{|\\mathcal{V}|}.\n\\end{equation}\nFurthermore, if the specified method belongs to the \\enquote{sample selection} category, \\emph{label precision} and \\emph{label recall} can be used as the metrics,\n\\noindent\n\\begin{equation}\n\\label{eq:label_precision}\n\\begin{gathered}\n\\text{Label Precision} = \\frac{|\\{(x_i,\\tilde{y}_i)\\in\\mathcal{S}_t: \\tilde{y}_i = y_i\\}|}{|\\mathcal{S}_t|},\\\\\n\\text{Label Recall} = \\frac{|\\{(x_i,\\tilde{y}_i)\\in\\mathcal{S}_t: \\tilde{y}_i = y_i\\}|}{|\\{(x_i,\\tilde{y}_i)\\in\\mathcal{B}_t: \\tilde{y}_i = y_i\\}|},\n\\end{gathered}\n\\end{equation}\nwhere $\\mathcal{S}_t$ is the set of selected clean examples in a mini-batch $\\mathcal{B}_t$. The two metrics are performance indicators for the examples selected from the mini-batch as true-labeled ones . \\looseness=-1\nMeanwhile, if the specified method belongs to the \\enquote{label refurbishment} category, \\emph{correction error} can be used as an indicator of how many examples are incorrectly refurbished,\n\\begin{equation}\n\\label{eq:correction_error}\n\\!\\text{Correction Error} = \\frac{|\\{x_i\\!\\in\\!\\mathcal{R}: \\text{argmax}(y_{i}^{refurb}) \\neq y_i\\}|}{|\\mathcal{R}|},\n\\end{equation}\nwhere $\\mathcal{R}$ is the set of examples whose labels are refurbished by Eq.\\,(\\ref{eq:label_correction}) and $y_i^{refurb}$ is the refurbished label of the $i$-th examples in $\\mathcal{R}$. \n\\vspace*{-0.12cm}","id":"6b4a1fff-c3c1-412a-b6f0-eeaa7d3721a4","level":"subsection","origin_cites_number":5,"parent_id":"0788ae52-30c4-4156-89f9-6be2e443db83","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Experimental Design"],["subsection","Evaluation Metrics"]],"subsections":[],"title":"Evaluation Metrics"},{"cite_extract_rate":0,"cites":[],"content":"With recent efforts in the machine learning community, the robustness of DNNs becomes evolving in several directions. Thus, the existing approaches covered in our survey face a variety of future challenges. This section provides discussion for future research that can facilitate and envision the development of deep learning in the label noise area. \\looseness=-1\n\\vspace*{-0.22cm}","id":"01d4e82b-d227-4b85-bfa5-6db6af2b743d","level":"section","origin_cites_number":0,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Future Research Directions"]],"subsections":["4ad866ea-f321-4ab2-be85-14135292873f","e53153d2-a39d-4573-9711-598781d00529","cfbd0ac1-5575-43b8-9a59-26d13838ff8e","23870233-7f5b-4ef1-9d38-e6f3adb0fa48","bf2c4a41-a623-4acf-b905-fe345054727a","a23d6d53-b5ad-4828-8bb9-730c9fff2012"],"title":"Future Research Directions"},{"cite_extract_rate":0.6666666666666661,"cites":[7777,4120,4124,4155,4136,4152],"content":"}\n{\nExisting theoretical and empirical studies for \\emph{robust loss function} and \\emph{loss correction} are largely built upon the instance-independent noise assumption that the label noise is independent of input features . However, this assumption may not be a good approximation of the real-world label noise. In particular, Chen et al. conducted a theoretical hypothesis testing\\footnote[61]{In Clothing1M, the result showed that the instance-independent noise happens with probability lower than $10^{-21250}$, which is statistically impossible.} using a popular real-world dataset, Clothing1M, and proved that its label noise is statistically different from the instance-independent noise. This testing confirms that the label noise should depend on the instance. \\looseness=-1\nConversely, most methods for the other direction (especially, \\emph{sample selection}) work well even under the instance-dependent label noise in general since they do not rely on the assumption. Nevertheless, Song et al. pointed out that their performance could considerably worsen in the instance-dependent\\,(or real-world) noise compared to symmetric noise due to the confusion between true-labeled and false-labeled examples. The loss distribution of true-labeled examples heavily overlaps that of false-labeled samples in the asymmetric noise, which is similar to the real-world noise, in Figure \\ref{fig:loss_distribution}(b). Thus, identifying clean examples becomes more challenging when dealing with the instance-dependent label noise.\nBeyond the instance-independent label noise, there have been a few recent studies for the instance-dependent label noise. Mostly, they only focus on a binary classification task or a restricted small-scale machine learning model such as logistic regression . Therefore, learning with the instance-dependent label noise is an important topic that deserves more research attention.} \n\\vspace*{-0.32cm}","id":"4ad866ea-f321-4ab2-be85-14135292873f","level":"subsection","origin_cites_number":9,"parent_id":"01d4e82b-d227-4b85-bfa5-6db6af2b743d","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Future Research Directions"],["subsection","{Instance-dependent Label Noise"]],"subsections":[],"title":"{Instance-dependent Label Noise"},{"cite_extract_rate":0.25,"cites":[4162],"content":"}\n{\nMost of the existing methods are applicable only for a \\emph{single-label} multi-class classification problem, where each data example is assumed to have only one true label. However, in the case of \\emph{multi-label} learning, each data example can be associated with a set of multiple true class labels. In music categorization, each music can belong to multiple categories . In semantic scene classification, each scene may belong to multiple scene classes . Thus, contrary to the single-label setup, the multi-label classifier aims to predict a set of target objects simultaneously. \nIn this setup, a multi-label dataset of millions of examples are reported to contain over $26.6\\%$ false-positive labels as well as a significant number of omitted labels . \nEven worse, the difference in occurrence between classes makes this problem more challenging; some minor class labels occur less in training data than other major class labels. Considering such aspects that can arise in multi-label classification, the simple extension of existing methods may not learn the proper correlations among multiple labels. Therefore, learning from noisy labels with multi-label data is another important topic for future research. We refer the readers to a recent study that discusses the evaluation of multi-label classifiers trained with noisy labels.\n}\n\\vspace*{-0.12cm}","id":"e53153d2-a39d-4573-9711-598781d00529","level":"subsection","origin_cites_number":4,"parent_id":"01d4e82b-d227-4b85-bfa5-6db6af2b743d","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Future Research Directions"],["subsection","{Multi-label Data with Label Noise"]],"subsections":[],"title":"{Multi-label Data with Label Noise"},{"cite_extract_rate":0.5,"cites":[7774],"content":"}\n{\nThe \\emph{class imbalance} in training data is commonly observed, where a few classes account for most of the data. Especially when working with large data in many real-world applications, this problem becomes more severe and is often associated with the problem of noisy labels .\nNevertheless, to ease the label noise problem, it is commonly assumed that training examples are equally distributed over all class labels in the training data. This assumption is quite strong when collecting large-scale data, and thus we need to consider a more realistic scenario in which the two problems coexist. \\looseness=-1\nMost of the existing robust methods may not work well with the class imbalance, especially when they rely on the learning dynamics of DNNs, e.g., the small-loss trick or memorization effect. Under the existence of the class imbalance, the training model converges to major classes faster than minor classes such that most examples in the major class exhibit small losses\\,(i.e., early memorization). That is, there is a risk of discarding most examples in the minor class. Furthermore, in terms of example importance, high-loss examples are commonly favored for the class imbalance problem , while small-loss examples are favored for the label noise problem. This conceptual contradiction hinders the applicability of the existing methods that neglect the class imbalance. Therefore, these two problems should be considered simultaneously to deal with more general situations.}\n\\vspace*{-0.12cm}","id":"cfbd0ac1-5575-43b8-9a59-26d13838ff8e","level":"subsection","origin_cites_number":2,"parent_id":"01d4e82b-d227-4b85-bfa5-6db6af2b743d","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Future Research Directions"],["subsection","{Class Imbalance Data with Label Noise"]],"subsections":[],"title":"{Class Imbalance Data with Label Noise"},{"cite_extract_rate":0.833333333333333,"cites":[7771,4163,3899,8740,8739],"content":"}\n{\nMachine learning classifiers can perpetuate and amplify the existing systemic injustices in society . Hence, fairness is becoming another important topic. Traditionally, robust training and fair training have been studied by separate communities; robust training with noisy labels has mostly focused on combating label noise without regarding data bias , whereas fair training has focused on dealing with data bias, not necessarily noise . However, noisy labels and data bias, in fact, coexist in real-world data. Satisfying both robustness and fairness is more realistic but challenging because the bias in data is pertinent to label noise. \nIn general, many fairness criteria are group-based, where a target metric is equalized or enforced over subpopulations in the data, also known as \\emph{protected groups} such as race or gender . Accordingly, the goal of fair training is building a model that satisfies such fairness criteria for the \\emph{true} protected groups. However, if the \\emph {noisy} protection group is involved, such fairness criteria cannot be directly applied. Recently, mostly after 2020, a few pioneering studies have emerged to consider both robustness and fairness objectives at the same time under the binary classification setting . Therefore, more research attention is needed for the convergence of robust training and fair training.\n}\n\\vspace*{-0.12cm}","id":"23870233-7f5b-4ef1-9d38-e6f3adb0fa48","level":"subsection","origin_cites_number":6,"parent_id":"01d4e82b-d227-4b85-bfa5-6db6af2b743d","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Future Research Directions"],["subsection","{Robust and Fair Training"]],"subsections":[],"title":"{Robust and Fair Training"},{"cite_extract_rate":0.4,"cites":[9127,4164],"content":"}\n{\nThere has been a lot of research on the robustness of deep learning under input perturbation, mainly in the field of adversarial training where the {input feature} is maliciously perturbed to distort the output of the DNN . Although learning with noisy labels and learning with noisy inputs have been regarded as separate research fields, their goals are similar in that they learn noise-robust representations from noisy data. Based on this common point of view, a few recent studies have investigated the interaction of adversarial training with noisy labels .\nInterestingly, it was turned out that adversarial training makes DNNs robust to label noise . Based on this finding, Damodaran et al. proposed a new regularization term, called Wasserstein adversarial regularization, to address the problem of learning with noisy labels. Zhu et al. proposed to use the number of projected gradient descent steps as a new criterion for sample selection such that clean examples are filtered out from noisy data. These approaches are regarded as a new perspective on label noise compared to traditional work. Therefore, understanding the connection between input perturbation and label noise could be another future topic for better representation learning toward robustness. \n}\n\\vspace*{-0.12cm}","id":"bf2c4a41-a623-4acf-b905-fe345054727a","level":"subsection","origin_cites_number":5,"parent_id":"01d4e82b-d227-4b85-bfa5-6db6af2b743d","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Future Research Directions"],["subsection","{Connection with Input Perturbation"]],"subsections":[],"title":"{Connection with Input Perturbation"},{"cite_extract_rate":0,"cites":[],"content":"}\n{\n{The efficiency of the learning pipeline is another important aspect to design deep learning approaches.} However, for robust deep learning, most studies have neglected the efficiency of the algorithm because their main goal is to improve the robustness to label noise. For example, maintaining multiple DNNs or training a DNN in multiple rounds is frequently used, but these approaches significantly degrade the efficiency of the learning pipeline. On ther other hand, the need for more efficient algorithms is increasing owing to the rapid increase in the amount of available data . \nAccording to our literature survey, most work did not even report the efficiency\\,(or time complexity) of their approaches. However, it is evident that saving the training time is helpful under the restricted budget for computation.\nTherefore, enhancing the efficiency will significantly increase the usability of robust deep learning in the big data era.\n}","id":"a23d6d53-b5ad-4828-8bb9-730c9fff2012","level":"subsection","origin_cites_number":1,"parent_id":"01d4e82b-d227-4b85-bfa5-6db6af2b743d","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Future Research Directions"],["subsection","{Efficient Learning Pipeline"]],"subsections":[],"title":"{Efficient Learning Pipeline"},{"cite_extract_rate":0,"cites":[],"content":"DNNs easily overfit to false labels owing to their high capacity in totally memorizing all noisy training samples. This overfitting issue still remains even with various conventional regularization techniques, such as dropout and batch normalization, thereby significantly decreasing their generalization performance. Even worse, in real-world applications, the difficulty in labeling renders the overfitting issue more severe. Therefore, learning from noisy labels has recently become one of the most active research topics. \nIn this survey, we presented a comprehensive understanding of modern deep learning methods to address the negative consequences of learning from noisy labels. All the methods were grouped into five categories according to their underlying strategies and described along with their methodological weaknesses. Furthermore, a systematic comparison was conducted using six popular properties used for evaluation in the recent literature. According to the comparison results, there is no ideal method that supports all the required properties; the supported properties varied depending on the category to which each method belonged. Several experimental guidelines were also discussed, including {noise rate estimation,} publicly available datasets, and evaluation metrics. Finally, we provided insights and directions for future research in this domain. \n\\section*{Acknowledgements}\nThis work was supported by Institute of Information \\& Communications Technology Planning \\& Evaluation\\,(IITP) grant funded by the Korea government\\,(MSIT) (No. 2020-0-00862, DB4DL: High-Usability and Performance In-Memory Distributed DBMS for Deep Learning).\n\\end{document}","id":"0fd993d8-9dac-43e0-a25d-3f85b6def415","level":"section","origin_cites_number":0,"parent_id":"4c23e240-6b38-42a9-a3ad-77c4f677661a","prefix_titles":[["title","Learning from Noisy Labels with Deep Neural Networks: A Survey"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ]\n ],\n \"subsections\": [\n \"8a473653-5cf6-4f3e-941b-06c5873beaf5\",\n \"345d7d1c-66a9-4d03-80b2-cee975c6db53\",\n \"60bd8c81-37fb-4934-ab18-bb9a24e80724\",\n \"3b9c1099-a2f2-417d-a753-e663d7ce62fa\",\n \"8d7ed0bf-5e07-45ed-a147-190ba1a209fb\",\n \"0788ae52-30c4-4156-89f9-6be2e443db83\",\n \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"0fd993d8-9dac-43e0-a25d-3f85b6def415\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.392857142857142,\n \"cites\": [\n 7,\n 206,\n 7210,\n 4238,\n 7770,\n 71,\n 3630,\n 8734,\n 4116,\n 4115,\n 7769\n ],\n \"content\": \"\\\\label{sec:introduction}\\n\\\\IEEEPARstart{W}ith the recent emergence of large-scale datasets, deep neural networks (DNNs) have exhibited impressive performance in numerous machine learning tasks, such as computer vision , information retrieval , and language processing . Their success is dependent on the availability of massive but carefully labeled data, which are expensive and time-consuming to obtain. Some non-expert sources, such as Amazon's Mechanical Turk and the surrounding text of collected data, have been widely used to mitigate the high labeling cost; however, the use of these source often results in unreliable labels . \\nIn addition, data labels can be extremely complex even for experienced domain experts ; \\nthey can also be adversarially manipulated by a label-flipping attack . Such unreliable labels are called \\\\emph{noisy labels} because they may be \\\\emph{corrupted} from ground-truth labels. The ratio of corrupted labels in real-world datasets is reported to range from $8.0\\\\%$ to $38.5\\\\%$ . \\\\looseness=-1\\nIn the presence of noisy labels, training DNNs is known to be susceptible to noisy labels because of the significant number of model parameters that render DNNs overfit to even corrupted labels with the capability of learning any complex function . Zhang et al. demonstrated that DNNs can easily fit an entire training dataset with any ratio of corrupted labels, which eventually resulted in poor generalizability on a test dataset. Unfortunately, popular regularization techniques, such as data augmentation , weight decay , dropout , and batch normalization {have been applied extensively, but they do \\\\emph{not} completely overcome the overfitting issue by themselves.} As shown in Figure \\\\ref{fig:convergence_analysis}, the gap in test accuracy between models trained on clean and noisy data remains significant even though all of the aforementioned regularization techniques are activated. Additionally, the accuracy drop with label noise is considered to be more harmful than with other noises, such as input noise . Hence, achieving a good generalization capability in the presence of noisy labels is a key challenge.\\n\\\\begin{figure}[t!]\\n\\\\begin{center}\\n\\\\includegraphics[width=8.8cm]{figures/introduction/convergence.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.55cm}\\n\\\\caption{Convergence curves of training and test accuracy when training WideResNet-16-8 using a standard training method on the CIFAR-100 dataset with the symmetric noise of $40\\\\%$: \\\\enquote{Noisy w/o. Reg.} and \\\\enquote{Noisy w. Reg.} are the models trained on noisy data without and with regularization, respectively, and \\\\enquote{Clean w. Reg.} is the model trained on clean data with regularization.}\\n\\\\label{fig:convergence_analysis}\\n\\\\vspace*{-0.4cm}\\n\\\\end{figure}\\nSeveral studies have been conducted to investigate supervised learning under noisy labels. Beyond conventional machine learning techniques , deep learning techniques have recently gained significant attention in the machine learning community. In this survey, we present the advances in recent deep learning techniques for overcoming noisy labels. We surveyed {recent studies by recursively tracking relevant bibliographies in papers published at premier research conferences, such as CVPR, ICCV, NeurIPS, ICML, and ICLR. Although we attempted to comprehensively include all recent studies at the time of submission, some of them may not be included because of the quadratic increase in deep learning papers. The studies included were grouped into \\\\emph{five} categories, as shown in Figure \\\\ref{fig:categorization} (see Section \\\\ref{sec:methodology} for details).\\n\\\\begin{figure*}[t!]\\n\\\\begin{center}\\n\\\\includegraphics[width=16.5cm]{figures/introduction/catorization_revise.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.55cm}\\n\\\\caption{Categorization of recent deep learning methods for overcomming noisy labels.}\\n\\\\label{fig:categorization}\\n\\\\vspace*{-0.4cm}\\n\\\\end{figure*}\",\n \"id\": \"8a473653-5cf6-4f3e-941b-06c5873beaf5\",\n \"level\": \"section\",\n \"origin_cites_number\": 28,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [\n \"74284b44-a1b2-476d-b3b5-489a51838419\",\n \"e5555e5b-acf2-4c85-9ee1-4bf47dcf2ae1\"\n ],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 7771\n ],\n \"content\": \"\\\\label{sec:related_surveys}\\nFr{\\\\'e}nay and Verleysen discussed the potential negative consequence of learning from noisy labels and provided a comprehensive survey on noise-robust classification methods, focusing on conventional supervised approaches such as na\\\\\\\"ive Bayes and support vector machines. Furthermore, their survey included the definitions and sources of label noise as well as the taxonomy of label noise. Zhang et al. discussed another aspect of label noise in crowdsourced data annotated by non-experts and provided a thorough review of expectation-maximization (EM) algorithms that were proposed to improve the quality of crowdsourced labels. Meanwhile, Nigam et al. provided a brief introduction to deep learning algorithms that were proposed to manage noisy labels; however, the scope of these algorithms was limited to only two categories, i.e., the loss function and sample selection in Figure \\\\ref{fig:categorization}. Recently, Han et al. summarized the essential components of robust learning with noisy labels, but their categorization is totally different from ours in philosophy; {we mainly focus on systematic methodological difference, whereas they rather focused on more general views, such as input data, objective functions, and optimization policies. Furthermore, this survey is the first to present a comprehensive methodological comparison of existing robust training approaches (see Tables \\\\ref{table:all_comparision} and \\\\ref{table:direction_comparison}}).\",\n \"id\": \"74284b44-a1b2-476d-b3b5-489a51838419\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"8a473653-5cf6-4f3e-941b-06c5873beaf5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Related Surveys\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Related Surveys\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 8696,\n 4118,\n 313,\n 4117\n ],\n \"content\": \"\\\\label{sec:survey_scope}\\n{\\nRobust training with DNNs becomes critical to guarantee the reliability of machine learning algorithms. In addition to label noise, two types of flawed training data have been actively studied by different communities . \\\\emph{Adversarial learning} is designed for small, worst-case perturbations of the inputs, so-called adversarial examples, which are maliciously constructed to deceive an already trained model into making errors . Meanwhile, \\\\emph{data imputation} primarily deals with missing inputs in training data, where missing values are estimated from the observed ones . Adversarial learning and data imputation are closely related to robust learning, but handling \\\\emph{feature} noise is beyond the scope of this survey---i.e., learning from noisy \\\\emph{labels}. \\n}\",\n \"id\": \"e5555e5b-acf2-4c85-9ee1-4bf47dcf2ae1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"8a473653-5cf6-4f3e-941b-06c5873beaf5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Survey Scope\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Survey Scope\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:preliminaries}\\nIn this section, the problem statement for supervised learning with noisy labels is provided along with the taxonomy of label noise. Managing noisy labels is a long-standing issue; therefore, we review the basic conventional approaches and {theoretical foundations underlying robust deep learning}. Table \\\\ref{table:notation} summarizes the notation frequently used in this study. \\\\looseness=-1\\n{\\n\\\\newcolumntype{L}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}m{#1}}\\n\\\\newcolumntype{X}[1]{>{\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}p{#1}}\\n\\\\begin{table}[t!]\\n\\\\caption{Summary of the notation.}\\n\\\\vspace*{-0.3cm}\\n\\\\begin{center}\\n\\\\begin{tabular}{|L{1.7cm} |X{6.2cm}|}\\\\hline\\n\\\\!\\\\!\\\\textbf{Notation} & \\\\hspace*{2.25cm} \\\\textbf{Description} \\\\\\\\\\\\hline \\\\hline \\n$\\\\mathcal{X}$ & the data feature space \\\\\\\\\\\\hline\\n$\\\\mathcal{Y}$, $\\\\tilde{\\\\mathcal{Y}}$ & the true and noisy label space \\\\\\\\\\\\hline\\n$\\\\mathcal{D}$, $\\\\tilde{\\\\mathcal{D}}$ & the clean and noisy training data \\\\\\\\\\\\hline\\n$P_{\\\\mathcal{D}}$, $P_{\\\\tilde{\\\\mathcal{D}}}$ & the joint distributions of clean and noisy data \\\\\\\\\\\\hline\\n$\\\\mathcal{B}_{t}$ & a set of mini-batch examples at time $t$ \\\\\\\\\\\\hline\\n$\\\\Theta_{t}$ & the parameter of a deep neural network at time $t$ \\\\\\\\\\\\hline\\n$f(\\\\,\\\\cdot\\\\,;\\\\Theta_{t})$ & a deep neural network parameterized by $\\\\Theta_{t}$ \\\\\\\\\\\\hline\\n$\\\\ell$ & a specific loss function \\\\\\\\\\\\hline\\n$\\\\mathcal{R}$ & an empirical risk \\\\\\\\\\\\hline\\n$\\\\mathbb{E}_{\\\\mathcal{D}}$ & an expectation over $\\\\mathcal{D}$ \\\\\\\\\\\\hline\\n$x$, $x_i$ & a data example of $\\\\mathcal{X}$ \\\\\\\\\\\\hline\\n$y$, $y_i$ & a true label of $\\\\mathcal{Y}$ \\\\\\\\\\\\hline\\n$\\\\tilde{y}$, $\\\\tilde{y}_i$ & a noisy label of $\\\\tilde{\\\\mathcal{Y}}$ \\\\\\\\\\\\hline\\n$\\\\eta$ & a specific learning rate \\\\\\\\\\\\hline\\n$\\\\tau$ & a true noise rate \\\\\\\\\\\\hline\\n$b$ & the number of mini-batch examples in $\\\\mathcal{B}_{t}$ \\\\\\\\\\\\hline\\n$c$ & the number of classes \\\\\\\\\\\\hline\\n{T}, $\\\\hat{\\\\text{T}}$ & the true and estimated noise transition matrix \\\\\\\\\\\\hline\\n\\\\end{tabular}\\n\\\\end{center}\\n\\\\label{table:notation}\\n\\\\vspace*{-0.55cm}\\n\\\\end{table}\\n}\\n\\\\begin{figure*}[t!]\\n\\\\begin{center}\\n\\\\includegraphics[width=15.5cm]{figures/introduction/tree_revise.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.45cm}\\n\\\\caption{A high level research overview of robust deep learning for noisy labels. The research directions that are actively contributed by the machine learning community are categorized into five groups in blue italic.}\\n\\\\label{fig:tree_categorization}\\n\\\\vspace*{-0.4cm}\\n\\\\end{figure*}\",\n \"id\": \"345d7d1c-66a9-4d03-80b2-cee975c6db53\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ]\n ],\n \"subsections\": [\n \"160721f0-f1c5-43ce-955b-8257be13eaa0\",\n \"53f5fc74-56fd-44a0-ad89-6075838d51dd\",\n \"587ccbe7-447b-402d-9bec-671027007ca9\",\n \"f4b49791-795c-49a1-9ee2-043cc9fa43e4\",\n \"7dbbb1a5-6e79-42e1-a3b6-1ed894715ef6\"\n ],\n \"title\": \"Preliminaries\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\emph{Classification} is a representative supervised learning task for learning a function that maps an input feature to a label . In this paper, we consider a $c$-class classification problem using a DNN with a softmax output layer. Let $\\\\mathcal{X} \\\\subset \\\\mathbb{R}^{d}$ be the feature space and $\\\\mathcal{Y} = \\\\{0,1\\\\}^{c}$ be the ground-truth label space in a \\\\emph{one-hot} manner. In a typical classification problem, we are provided with a training dataset $\\\\mathcal{D}=\\\\{(x_i, y_i)\\\\}_{i=1}^{N}$ obtained from an unknown joint distribution $P_{\\\\mathcal{D}}$ over $\\\\mathcal{X} \\\\times \\\\mathcal{Y}$, where each $(x_i, y_i)$ is \\\\emph{independent and identically distributed}. The goal of the task is to learn the mapping function $f(\\\\,\\\\cdot\\\\,;\\\\Theta): \\\\mathcal{X} \\\\rightarrow [0, 1]^{c}$ of the DNN parameterized by $\\\\Theta$ such that the parameter $\\\\Theta$ minimizes the empirical risk $\\\\mathcal{R}_{\\\\mathcal{D}}(f)$,\\n\\\\begin{equation}\\n\\\\label{eq:empirical_risk}\\n\\\\!\\\\!\\\\mathcal{R}_{\\\\mathcal{D}}(f) = \\\\mathbb{E}_{\\\\mathcal{D}}[\\\\ell\\\\big(f(x;\\\\Theta), y\\\\big)] = \\\\frac{1}{|\\\\mathcal{D}|}\\\\!\\\\sum_{(x,y)\\\\in\\\\mathcal{D}}\\\\!\\\\!\\\\!\\\\!\\\\ell\\\\big(f(x;\\\\Theta), y\\\\big),\\n\\\\end{equation}\\nwhere $\\\\ell$ is a certain loss function.\\nAs data labels are corrupted in various real-world scenarios, we aim to train the DNN from noisy labels. Specifically, we are provided with a noisy training dataset $\\\\tilde{\\\\mathcal{D}}=\\\\{(x_i, \\\\tilde{y}_i)\\\\}_{i=1}^{N}$ obtained from a noisy joint distribution $P_{\\\\tilde{\\\\mathcal{D}}}$ over $\\\\mathcal{X}\\\\times\\\\tilde{\\\\mathcal{Y}}$, where $\\\\tilde{y}$ is a \\\\emph{noisy} label which may not be true. Hence, following the standard training procedure, a mini-batch $\\\\mathcal{B}_{t}=\\\\{(x_i, \\\\tilde{y}_i)\\\\}_{i=1}^{b}$ comprising $b$ examples is obtained randomly from the noisy training dataset $\\\\tilde{\\\\mathcal{D}}$ at time $t$. Subsequently, the DNN parameter $\\\\Theta_t$ at time $t$ is updated along the descent direction of the empirical risk on mini-batch $\\\\mathcal{B}_t$,\\n\\\\begin{equation}\\n\\\\label{eq:corrupted_update}\\n\\\\Theta_{t+1} = \\\\Theta_{t} - \\\\eta\\\\nabla\\\\Big(\\\\frac{1}{|\\\\mathcal{B}_{t}|} \\\\!\\\\sum_{(x,\\\\tilde{y}) \\\\in \\\\mathcal{B}_{t}} \\\\!\\\\!\\\\!\\\\!\\\\ell\\\\big(f(x;\\\\Theta_{t}), \\\\tilde{y}\\\\big)\\\\Big),\\n\\\\end{equation}\\nwhere $\\\\eta$ is a learning rate specified. \\\\looseness=-1\\nHere, the risk minimization process is no longer \\\\emph{noise-tolerant} because of the loss computed by the noisy labels. DNNs can easily memorize corrupted labels and correspondingly degenerate their generalizations on unseen data . Hence, mitigating the adverse effects of noisy labels is essential to enable noise-tolerant training for deep learning. \\n\\\\vspace*{-0.1cm}\",\n \"id\": \"160721f0-f1c5-43ce-955b-8257be13eaa0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"345d7d1c-66a9-4d03-80b2-cee975c6db53\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ],\n [\n \"subsection\",\n \"Supervised Learning with Noisy Labels\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Supervised Learning with Noisy Labels\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:taxonomy}\\n{This section presents the types of label noise that have been adopted to design robust training algorithms.} Even if data labels are corrupted from ground-truth labels without \\\\emph{any} prior assumption, in essence, the corruption probability is affected by the dependency between \\\\emph{data features} or \\\\emph{class labels}. A detailed analysis of the taxonomy of label noise was provided by Fr{\\\\'e}nay and Verleysen . {Most existing algorithms dealt with instance-independent noise, but instance-dependent noise has not yet been extensively investigated owing to its complex modeling.}\\n\\\\smallskip\",\n \"id\": \"53f5fc74-56fd-44a0-ad89-6075838d51dd\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"345d7d1c-66a9-4d03-80b2-cee975c6db53\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ],\n [\n \"subsection\",\n \"Taxonomy of Label Noise\"\n ]\n ],\n \"subsections\": [\n \"5a2e6abe-21d0-4935-ab73-f9bf0caf9ca6\",\n \"9e5fdfd7-ba0e-466c-96b0-5d8ad47de6e9\"\n ],\n \"title\": \"Taxonomy of Label Noise\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 3630\n ],\n \"content\": \"}\\nA typical approach for modeling label noise assumes that the corruption process is conditionally \\\\emph{independent} of data features when the true label is given . That is, the true label is corrupted by a \\\\emph{noise transition matrix} $\\\\text{T} \\\\in [0, 1]^{c\\\\times c}$, where $ \\\\text{T}_{ij} \\\\coloneqq p(\\\\tilde{y}=j|y=i)$ is the probability of the true label $i$ being flipped into a corrupted label $j$. \\nIn this approach, the noise is called a \\\\emph{symmetric}\\\\,(or \\\\emph{uniform}) noise with a noise rate $\\\\tau \\\\in [0,1]$ if $\\\\forall_{i=j} \\\\text{T}_{ij} \\\\!=\\\\! 1-\\\\tau \\\\wedge \\\\forall_{i \\\\neq j} \\\\text{T}_{ij} = \\\\frac{\\\\tau}{c-1}$, where a true label is flipped into other labels with equal probability. In contrast to symmetric noise, the noise is called an \\\\emph{asymmetric}\\\\,(or \\\\emph{label-dependent}) noise if $\\\\forall_{i=j} \\\\text{T}_{ij} \\\\!=\\\\! 1-\\\\tau \\\\wedge \\\\exists_{i \\\\neq j, i\\\\neq k, j\\\\neq k} \\\\text{T}_{ij} > \\\\text{T}_{ik} $, where a true label is more likely to be mislabeled into a particular label. For example, a \\\\enquote{dog} is more likely to be confused with a \\\\enquote{cat} than with a \\\\enquote{fish.} In a stricter case when $\\\\forall_{i=j} \\\\text{T}_{ij} \\\\!=\\\\! 1-\\\\tau \\\\wedge \\\\exists_{i \\\\neq j} \\\\text{T}_{ij} = \\\\tau$, the noise is called a \\\\emph{pair noise}, where a true label is flipped into only a certain label. \\n\\\\smallskip\",\n \"id\": \"5a2e6abe-21d0-4935-ab73-f9bf0caf9ca6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"53f5fc74-56fd-44a0-ad89-6075838d51dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ],\n [\n \"subsection\",\n \"Taxonomy of Label Noise\"\n ],\n [\n \"subsubsection\",\n \"Instance-independent Label Noise\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Instance-independent Label Noise\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\nFor more realistic noise modeling, the corruption probability is assumed to be \\\\emph{dependent} on both the data features and class labels . Accordingly, the corruption probability is defined as $\\\\rho_{ij}(x)\\\\!= \\\\!p(\\\\tilde{y}\\\\!=\\\\!j|y\\\\!=\\\\!i, x)$. \\nUnlike the aforementioned noises, the data feature of an example $x$ also affects the chance of $x$ being mislabeled. \\n\\\\vspace*{-0.1cm}\",\n \"id\": \"9e5fdfd7-ba0e-466c-96b0-5d8ad47de6e9\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"53f5fc74-56fd-44a0-ad89-6075838d51dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ],\n [\n \"subsection\",\n \"Taxonomy of Label Noise\"\n ],\n [\n \"subsubsection\",\n \"Instance-dependent Label Noise\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Instance-dependent Label Noise\"\n },\n {\n \"cite_extract_rate\": 0.344827586206896,\n \"cites\": [\n 3454,\n 7771,\n 4121,\n 8735,\n 4120,\n 4122,\n 3630,\n 4119,\n 4115,\n 8736\n ],\n \"content\": \"\\\\label{sec:non_deep_learning}\\nFor decades, numerous methods have been proposed to manage noisy labels using conventional machine learning techniques. These methods can be categorized into \\\\emph{four} groups , as follows:\\n\\\\begin{itemize}[leftmargin=9pt]\\n\\\\item \\\\textbf{Data Cleaning:} Training data are cleaned by excluding examples whose labels are likely to be corrupted. Bagging and boosting are used to filter out false-labeled examples to remove examples with higher weights because false-labeled examples tend to exhibit much higher weights than true-labeled examples . In addition, various methods, such as $k$-nearest neighbor, outlier detection, and anomaly detection, have been widely exploited to exclude false-labeled examples from noisy training data . Nevertheless, this family of methods suffers from over-cleaning issue that overly removes even the true-labeled examples. \\n\\\\vspace*{0.12cm}\\n\\\\item \\\\textbf{Surrogate Loss:} Motivated by the noise-tolerance of the 0-1 loss function , many researchers have attempted to resolve its inherent limitations, such as computational hardness and non-convexity that render gradient methods unusable. Hence, several convex surrogate loss functions, which approximate the 0-1 loss function, have been proposed to train a specified classifier under the binary classification setting . However, these loss functions cannot support the multi-class classification task.\\n\\\\vspace*{0.12cm}\\n\\\\item \\\\textbf{Probabilistic Method:} Under the assumption that the distribution of features is helpful in solving the problem of learning from noisy labels , the confidence of each label is estimated by clustering and then used for a weighted training scheme . This confidence is also used to convert hard labels into soft labels to reflect the uncertainty of labels . In addition to these clustering approaches, several Bayesian methods have been proposed for graphical models such that they can benefit from using any type of prior information in the learning process . However, this family of methods may exacerbate the overfitting issue owing to the increased number of model parameters. \\n\\\\vspace*{0.12cm}\\n\\\\item \\\\textbf{Model-based Method:} As conventional models, such as the SVM and decision tree, are not robust to noisy labels, significant effort has been expended to improve the robustness of them. To develop a robust SVM model, misclassified examples during learning are penalized in the objective . In addition, several decision tree models are extended using new split criteria to solve the overfitting issue when the training data are not fully reliable . However, it is infeasible to apply their design principles to deep learning.\\n\\\\end{itemize}\\n\\\\smallskip\\n{\\nMeanwhile, deep learning is more susceptible to label noises than traditional machine learning owing to its high expressive power, as proven by many researchers . \\nThere has been significant effort to understand why noisy labels negatively affect the performance of DNNs . This theoretical understanding has led to the algorithmic design which achieves higher robustness than non-deep learning methods. A detailed analysis of theoretical understanding for robust deep learning was provided by Han et al. . \\n}\\n\\\\begin{comment}\\n\\\\vspace*{-0.2cm}\",\n \"id\": \"587ccbe7-447b-402d-9bec-671027007ca9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 29,\n \"parent_id\": \"345d7d1c-66a9-4d03-80b2-cee975c6db53\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ],\n [\n \"subsection\",\n \"Non-deep Learning Approaches\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Non-deep Learning Approaches\"\n },\n {\n \"cite_extract_rate\": 0.647058823529411,\n \"cites\": [\n 7771,\n 3340,\n 4126,\n 3630,\n 4253,\n 8735,\n 4120,\n 4124,\n 4125,\n 4115,\n 4123\n ],\n \"content\": \"\\\\label{sec:theoretical}\\n{\\nDeep learning is more susceptible to label noises than traditional machine learning owing to its high expressive power, as proven by many researchers . \\nThere has been significant effort to understand why noisy labels negatively affect the performance of DNNs . This theoretical understanding has led to the architectural or algorithmic design which is more robust than the non-deep learning methods. A detailed analysis of theoretical understanding for robust deep learning was provided by Han et al. ; \\\\emph{three} high-level perspectives have been widely leveraged to design robust approaches, as follows:\\n\\\\vspace*{0.05cm}\\n\\\\newcommand{\\\\norm}[1]{\\\\left\\\\lVert#1\\\\right\\\\rVert}\\n\\\\begin{itemize}[leftmargin=9pt]\\n\\\\item \\\\textbf{Memorization Effect}: The \\\\emph{memorization nature} of DNNs was explored theoretically in recent literature . Assuming clusterable data where the clusters are located on the unit Euclidean ball, Li et al. proved the distance from the initial weight ${W}_{0}$ to the weight ${W}_t$ after $t$ iterations,\\n\\\\begin{equation}\\n\\\\label{eq:memorization_effect_foundation}\\n\\\\norm{{W}_t - {W}_0}_{F} \\\\lesssim \\\\big( \\\\sqrt{K} + (K^{2}\\\\epsilon_{0}/\\\\norm{{C}}^{2})t \\\\big),\\n\\\\end{equation}\\nwhere $\\\\norm{\\\\cdot}_{F}$ is the Frobenius norm, $K$ is the number of clusters, and ${C}$ is the set of cluster centers reaching all input examples within their $\\\\epsilon_0$ neighborhood.\\nEq.\\\\ \\\\eqref{eq:memorization_effect_foundation} demonstrates that the weights of DNNs start to stray far from the initial weights when overfitting to corrupted labels, while they are still in the vicinity of the initial weights at the beginning of training . In the empirical studies , this result is also known as the \\\\emph{memorization effect} that DNNs tend to first learn simple and generalized patterns and then gradually overfit to all the noisy patterns. Thus, to achieve better generalization, early stopping and favoring small-loss training examples are commonly employed to design robust training methods.\\n\\\\end{itemize}\\n}\\n\\\\end{comment}\\n\\\\vspace*{-0.2cm}\",\n \"id\": \"f4b49791-795c-49a1-9ee2-043cc9fa43e4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"345d7d1c-66a9-4d03-80b2-cee975c6db53\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ],\n [\n \"subsection\",\n \"Theoretical Foundations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Theoretical Foundations\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 4127\n ],\n \"content\": \"\\\\label{sec:regression}\\n{\\nIn addition to classification, regression is another main topic of supervised machine learning, which aims to model the relationship between a number of features and a continuous target variable. Unlike the classification task with a \\\\emph{discrete} label space, the regression task considers the continuous variable as its target label , and thus it learns the mapping function $f(~\\\\cdot~; \\\\Theta): \\\\mathcal{X} \\\\rightarrow \\\\mathcal{Y}$, where $\\\\mathcal{Y} \\\\in \\\\mathbb{R}$ is a \\\\emph{continuous} label space. \\nGiven the input feature $x$ and its ground-truth label $y$, two types of label noise are considered in the regression task. An \\\\emph{additive noise} is formulated by $\\\\tilde{y} := y + \\\\epsilon$ where $\\\\epsilon$ is drawn from a random distribution independent from the input feature; an \\\\emph{instance-dependent noise} is formulated by $\\\\tilde{y} := \\\\rho(x)$ where $\\\\rho: \\\\mathcal{X} \\\\rightarrow \\\\mathcal{Y}$ is a noise function dependent on the input feature. \\nAlthough regression predicts continuous values, regression and classification share the same concept of learning the mapping function from the input feature $x$ to the output label $y$. Thus, many robust approaches for classification are easily extended to the regression problem with simple modification . Thus, in this survey, we focus on the classification setting for which most robust methods are defined.\\n}\\n\\\\vspace*{-0.1cm}\",\n \"id\": \"7dbbb1a5-6e79-42e1-a3b6-1ed894715ef6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"345d7d1c-66a9-4d03-80b2-cee975c6db53\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ],\n [\n \"subsection\",\n \"Regression with Noisy Labels\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Regression with Noisy Labels\"\n },\n {\n \"cite_extract_rate\": 0.777777777777777,\n \"cites\": [\n 3340,\n 4129,\n 4143,\n 7133,\n 4137,\n 7773,\n 3345,\n 4134,\n 4132,\n 4140,\n 4141,\n 7775,\n 4126,\n 3342,\n 4142,\n 7772,\n 8737,\n 4130,\n 4138,\n 7191,\n 7162,\n 4135,\n 4144,\n 4145,\n 7774,\n 4139,\n 4125,\n 892,\n 4128,\n 4253,\n 4131,\n 8630,\n 4133,\n 2277,\n 4136\n ],\n \"content\": \"\\\\label{sec:methodology}\\n\\\\vspace*{-0.0cm}\\nAccording to our comprehensive survey, the robustness of deep learning can be enhanced in numerous approaches . Figure \\\\ref{fig:tree_categorization} shows an overview of recent research directions conducted by the machine learning community. {All of them\\\\,(i.e., \\\\textsection \\\\ref{sec:robust_architecture}~--~\\\\textsection \\\\ref{sec:sample_selection}) focused on making a supervised learning process more robust to label noise: \\\\looseness=-1\\n\\\\begin{itemize}[leftmargin=9pt]\\n\\\\item\\n(\\\\textsection \\\\ref{sec:robust_architecture}) Robust architecture: adding a noise adaptation layer at the top of an underlying DNN to learn label transition process or developing a dedicated architecture to reliably support more diverse types of label noise;\\n\\\\vspace*{0.12cm}\\n\\\\item (\\\\textsection \\\\ref{sec:robust_regularization}) Robust regularization: enforcing a DNN to overfit less to false-labeled examples explicitly or implicitly;\\n\\\\vspace*{0.12cm}\\n\\\\item (\\\\textsection \\\\ref{sec:robust_loss_function}) Robust loss function: improving the loss function;\\n\\\\vspace*{0.12cm}\\n\\\\item (\\\\textsection \\\\ref{sec:loss_adjustment}) Loss adjustment: adjusting the loss value according to the confidence of a given loss (or label) by loss correction, loss reweighting, or label refurbishment; \\n\\\\vspace*{0.12cm}\\n\\\\item (\\\\textsection \\\\ref{sec:sample_selection}) Sample selection: identifying true-labeled examples from noisy training data via multi-network or multi-round learning. \\n\\\\end{itemize}\\nOverall, we categorize all recent deep learning methods into \\\\emph{five} groups corresponding to popular research directions, as shown in Figure \\\\ref{fig:tree_categorization}. In \\\\textsection \\\\ref{sec:loss_adjustment}, meta learning is also discussed because it finds the optimal hyperparameters for loss reweighting. In \\\\textsection \\\\ref{sec:sample_selection}, we discuss the recent efforts for combining sample selection with other orthogonal directions or semi-supervised learning toward the state-of-the-art performance.}\\nFigure \\\\ref{fig:categorization} illustrates the categorization of robust training methods using these five groups. \\n\\\\begin{comment}\\n{\\n\\\\newcolumntype{L}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}m{#1}}\\n\\\\newcolumntype{X}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}p{#1}}\\n\\\\begin{table}[t!]\\n\\\\caption{Summary of existing deep learning methods according to the seven categories in Figure \\\\ref{fig:categorization}.}\\n\\\\vspace*{-0.3cm}\\n\\\\begin{center}\\n\\\\begin{tabular}{L{1.75cm} |X{6.23cm}}\\\\toprule\\n\\\\!\\\\!\\\\textbf{Category} & \\\\textbf{Deep Learning Method} \\\\\\\\\\\\midrule \\nRobust Loss Function & \\\\makecell[l]{\\\\!\\\\!\\\\emph{Robust MAE}\\\\,\\\\!, \\\\emph{Generalized Cross Entropy}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Symmetric Cross Entropy}\\\\,\\\\!, \\\\emph{Curriculum Loss}\\\\,\\\\!}\\\\!\\\\!\\\\! \\\\\\\\\\\\hline\\nRobust Architecture & \\\\makecell[l]{ \\\\!\\\\!\\\\emph{Webly Learning}\\\\,\\\\!, \\\\emph{Noise Model}\\\\,\\\\!, \\\\emph{Dropout Noise} \\\\!\\\\!\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Model}\\\\,\\\\!, \\\\emph{S--model}\\\\,\\\\!, \\\\emph{C--model}\\\\,\\\\!, \\\\emph{NLNN}\\\\,\\\\!, \\\\!\\\\!\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Probabilistic Noise Model}\\\\,\\\\!, \\\\emph{Masking}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Contrastive-additive Noise Network}\\\\,\\\\!}\\\\!\\\\!\\\\! \\\\\\\\\\\\hline\\nRobust Regularization & \\\\makecell[l]{ \\\\!\\\\!\\\\!\\\\! \\\\emph{Adversarial Training}\\\\,\\\\!, \\\\emph{Label Smoothing}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Mixup}\\\\,\\\\!, \\\\emph{Bilevel Learning}\\\\,\\\\!, \\\\emph{Annotator}\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\!\\\\! \\\\emph{Confusion}\\\\,\\\\!, \\\\emph{Pre-training}\\\\,\\\\!\\\\!\\\\!} \\\\!\\\\!\\\\!\\\\\\\\\\\\hline\\n\\\\!\\\\!Loss Adjustment\\\\!\\\\!\\\\! & \\\\makecell[l]{ \\\\!\\\\!\\\\emph{Backward Correction}\\\\,\\\\!}, \\\\emph{Forward Correction}\\\\,\\\\!}, \\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Gold}\\\\! \\\\emph{Loss}\\\\! \\\\emph{Correction}\\\\,\\\\!\\\\!, \\\\emph{Importance}\\\\! \\\\emph{Reweighting}\\\\,\\\\!\\\\!, \\\\!\\\\!\\\\!\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Active Bias}\\\\! , \\\\emph{Bootstrapping}\\\\! , \\\\emph{Dynamic} \\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\!\\\\! \\\\emph{Bootstrapping}\\\\,\\\\!, \\\\emph{D2L}\\\\,\\\\!, \\\\emph{SELFIE}\\\\,\\\\! }\\\\!\\\\!\\\\!\\\\\\\\\\\\hline\\n\\\\!\\\\!\\\\!Sample Selection\\\\!\\\\!\\\\! & \\\\makecell[l]{\\\\!\\\\!\\\\emph{Decouple}\\\\,\\\\!, \\\\emph{MentorNet}\\\\,\\\\!, \\\\emph{Co-teaching}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Co-teaching+}\\\\,\\\\!, \\\\emph{Iterative Detection}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{ITLM}\\\\,\\\\!, \\\\emph{INCV}\\\\,\\\\!, \\\\emph{SELFIE}\\\\,\\\\!, \\\\emph{SELF}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{Curriculum Loss}\\\\,\\\\!}\\\\!\\\\!\\\\!\\\\\\\\\\\\hline\\n\\\\!\\\\!\\\\!Meta Learning\\\\!\\\\!\\\\! & \\\\makecell[l]{ \\\\!\\\\!\\\\emph{Meta-Regressor}\\\\,\\\\!, \\\\emph{Knowledge Distillation}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{L2LWS}\\\\,\\\\!, \\\\emph{CWS}\\\\,\\\\!, \\\\emph{Automatic Reweighting}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{MLNT}\\\\,\\\\!, \\\\emph{Meta-weight-net}\\\\,\\\\!} \\\\!\\\\!\\\\!\\\\\\\\\\\\hline\\n\\\\!\\\\!Semi-supervised\\\\!\\\\! \\\\!Learning\\\\! & \\\\makecell[l]{ \\\\!\\\\!\\\\emph{Label Aggregation}\\\\,\\\\!, \\\\emph{Two-Stage Framework}\\\\,\\\\!,\\\\!\\\\!\\\\!\\\\\\\\ \\\\!\\\\!\\\\emph{SELF} , \\\\emph{DivideMix}\\\\,\\\\!}\\\\!\\\\!\\\\! \\\\\\\\\\\\bottomrule\\n\\\\end{tabular}\\n\\\\end{center}\\n\\\\label{table:summary_methods}\\n\\\\vspace*{-0.5cm}\\n\\\\end{table}\\n}\\n\\\\end{comment}\",\n \"id\": \"60bd8c81-37fb-4934-ab18-bb9a24e80724\",\n \"level\": \"section\",\n \"origin_cites_number\": 45,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ]\n ],\n \"subsections\": [\n \"f16bcc10-a9fa-4749-b206-8aa388990240\",\n \"25e15182-1991-4f1a-83d9-017fe1a40d14\",\n \"59ab6098-f130-49dc-b87f-123eaa6be482\",\n \"3d40345d-b6d6-4b98-943c-ec552f10d7fb\",\n \"54d73157-e47b-4943-9b30-b518e1806946\"\n ],\n \"title\": \"Deep Learning Approaches\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4138,\n 4133,\n 4146,\n 7773,\n 2277,\n 4136\n ],\n \"content\": \"\\\\label{sec:robust_architecture}\\nIn numerous studies, architectural changes have been made to model the noise transition matrix of a noisy dataset . These changes include adding a noise adaptation layer at the top of the softmax layer and designing a new dedicated architecture. The resulting architectures yield improved generalization through the modification of the DNN output based on the estimated label transition probability.\\n\\\\vspace*{0.15cm}\",\n \"id\": \"f16bcc10-a9fa-4749-b206-8aa388990240\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"60bd8c81-37fb-4934-ab18-bb9a24e80724\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Robust Architecture\"\n ]\n ],\n \"subsections\": [\n \"439a17f4-972c-4739-b63c-f2fd25ca8426\",\n \"81044d2e-5255-4dbd-8274-7c1d26bb1d92\"\n ],\n \"title\": \"Robust Architecture\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7776,\n 2277,\n 4136\n ],\n \"content\": \"}\\n\\\\label{sec:adaptation_layer}\\n{\\nFrom the view of training data, the noise process is modeled by discovering the underlying label transition pattern (i.e., the {noise transition matrix} T). Given an example $x$, the noisy class posterior probability for an example $x$ is expressed by\\n\\\\begin{equation}\\n\\\\label{eq:label_transition_matrix}\\n\\\\begin{gathered}\\n\\\\!\\\\!\\\\!\\\\!\\\\!p(\\\\tilde{y}=j|x) \\\\!=\\\\!\\\\sum_{i=1}^{c}p(\\\\tilde{y}=j, y=i|x)\\\\! = \\\\!\\\\sum_{i=1}^{c}\\\\text{T}_{ij}p(y=i|x),\\\\\\\\\\n\\\\text{where} ~~ \\\\text{T}_{ij} = p(\\\\tilde{y}=j|y=i,x).\\n\\\\end{gathered}\\n\\\\end{equation} \\nIn light of this, the noise adaptation layer is intended to mimic the label transition behavior in learning a DNN. Let $p(y|x;\\\\Theta)$ be the output of the base DNN with a softmax output layer.\\nThen, following Eq.\\\\,\\\\eqref{eq:label_transition_matrix}, the probability of an example $x$ being predicted as its noisy label $\\\\tilde{y}$ is parameterized by }\\n\\\\begin{equation}\\n\\\\label{eq:noise_adaption_process}\\n\\\\begin{split}\\n\\\\!\\\\!\\\\!p(\\\\tilde{y}=j|x;\\\\Theta, \\\\mathcal{W})&= \\\\sum_{i=1}^{c}p(\\\\tilde{y}=j, y\\\\!=\\\\!i|x; \\\\Theta, \\\\mathcal{W})\\\\\\\\\\n&= \\\\sum_{i=1}^{c}\\\\underbrace{p(\\\\tilde{y}=j|y\\\\!=\\\\!i;\\\\mathcal{W})}_{\\\\text{Noise Adaptation Layer}}\\\\underbrace{p(y\\\\!=\\\\!i|x;\\\\Theta)}_{\\\\text{Base Model}}.\\n\\\\end{split}\\n\\\\end{equation}\\nHere, the noisy label $\\\\tilde{y}$ is assumed to be \\\\emph{conditionally independent} of the input $x$ in general. \\nAccordingly, as shown in Figure \\\\ref{fig:noise_adaption_process}, the noisy adaptation layer is added at the top of the base DNN to model the noise transition matrix parameterized by $\\\\mathcal{W}$. This layer should be removed when test data is to be predicted. \\\\looseness=-1\\n\\\\begin{figure}[t!]\\n\\\\begin{center}\\n\\\\includegraphics[width=8.5cm]{figures/methodology/noise_adaption_layer.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.4cm}\\n\\\\caption{Noise modeling process using the noise adaptation layer.}\\n\\\\label{fig:noise_adaption_process}\\n\\\\vspace*{-0.3cm}\\n\\\\end{figure}\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}: \\\\emph{Webly learning} first trains the base DNN only for easy examples retrieved by search engines; subsequently, the confusion matrix for all training examples is used as the initial weight $\\\\mathcal{W}$ of the noise adaptation layer. It fine-tunes the entire model in an end-to-end manner for hard training examples. In contrast, the \\\\emph{noise model} initializes $\\\\mathcal{W}$ to an identity matrix and adds a regularizer to force $\\\\mathcal{W}$ to diffuse during DNN training. The \\\\emph{dropout noise model} applies dropout regularization to the adaptation layer, whose output is normalized by the softmax function to implicitly diffuse $\\\\mathcal{W}$. The \\\\emph{s-model} is similar to the \\\\emph{dropout noise model} but dropout is not applied. The \\\\emph{c-model} is an extension of the s-model that models the instance-dependent noise, which is more realistic than the symmetric and asymmetric noises. Meanwhile, \\\\emph{NLNN} adopts the EM algorithm to iterate the E-step to estimate the noise transition matrix and the M-step to back-propagate the DNN. \\\\looseness=-1\\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: A common drawback of this family is their inability to identify false-labeled examples, treating all the examples equally. Thus, the estimation error for the transition matrix is generally large when only noisy training data is used or when the noise rate is high .}\\nMeanwhile, for the EM-based method, becoming stuck in local optima is inevitable, and high computational costs are incurred .\\n\\\\vspace*{0.15cm}\",\n \"id\": \"439a17f4-972c-4739-b63c-f2fd25ca8426\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"f16bcc10-a9fa-4749-b206-8aa388990240\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Robust Architecture\"\n ],\n [\n \"subsubsection\",\n \"Noise Adaptation Layer\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Noise Adaptation Layer\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 4138,\n 7773,\n 8738\n ],\n \"content\": \"} \\n{Beyond the label-dependent label noise, several studies have been conducted to support more complex noise, leading to the design of dedicated architectures .} They typically aimed at increasing the reliability of estimating the label transition probability to handle more complex and realistic label noise. \\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}: \\\\emph{Probabilistic noise modeling} manages two independent networks, each of which is specialized to predict the noise type and label transition probability.\\nBecause an EM-based approach with random initialization is impractical for training the entire network, both networks are trained with massive noisy labeled data after the pre-training step with a small amount of clean data.\\nMeanwhile, \\\\emph{masking} is a human-assisted approach to convey the human cognition of invalid label transitions.\\nConsidering that noisy labels are mainly from the interaction between humans and tasks, the invalid transition investigated by humans was leveraged to constrain the noise modeling process. Owing to the difficulty in specifying the explicit constraint, a variant of generative adversarial networks\\\\,(GANs) was employed in this study. Recently, the \\\\emph{contrastive-additive noise network} was proposed to adjust incorrectly estimated label transition probabilities by introducing a new concept of quality embedding, which models the trustworthiness of noisy labels. {\\\\emph{RoG} builds a simple yet robust generative classifier on top of any discriminative DNN pre-trained on noisy data.}\\n\\\\smallskip\\n\\\\noindent \\\\underline{Remark}: Compared with the noise adaptation layer, this family of methods significantly improves the robustness to more diverse types of label noise, but it cannot be easily extended to other architectures in general.\",\n \"id\": \"81044d2e-5255-4dbd-8274-7c1d26bb1d92\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"f16bcc10-a9fa-4749-b206-8aa388990240\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Robust Architecture\"\n ],\n [\n \"subsubsection\",\n \"Dedicated Architecture\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dedicated Architecture\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 4142,\n 71\n ],\n \"content\": \"\\\\label{sec:robust_regularization}\\nRegularization methods have been widely studied to improve the generalizability of a learned model in the machine learning community . \\nBy avoiding overfitting in model training, the robustness to label noise improves with widely-used regularization techniques such as \\\\emph{data augmentation} , \\\\emph{weight decay} , \\\\emph{dropout} , and \\\\emph{batch normalization} . {These canonical regularization methods operate well on moderately noisy data, but they alone do \\\\emph{not sufficiently} improve the test accuracy}; poor generalization could be obtained when the noise is heavy . Thus, more advanced regularization techniques have been recently proposed, which further improved robustness to label noise when used along with the canonical methods. The main advantage of this family is its \\\\emph{flexibility} in collaborating with other directions because it only requires simple modifications. \\n\\\\smallskip\",\n \"id\": \"25e15182-1991-4f1a-83d9-017fe1a40d14\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"60bd8c81-37fb-4934-ab18-bb9a24e80724\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Robust Regularization\"\n ]\n ],\n \"subsections\": [\n \"84774a8d-19b4-4ef8-ba90-419f51709a46\",\n \"50da4bbd-f1a0-49f9-b327-8d43650a94c6\"\n ],\n \"title\": \"Robust Regularization\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4147,\n 4131,\n 7133,\n 4142\n ],\n \"content\": \"}\\n{\\nThe regularization can be an {explicit} form that modifies the expected training loss, e.g., weight decay and dropout. \\n}\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\n\\\\emph{Bilevel learning} uses a clean validation dataset to regularize the overfitting of a model by introducing a bilevel optimization approach, which differs from the conventional one in that its regularization constraint is also an optimization problem. Overfitting is controlled by adjusting the weights on each mini-batch and selecting their values such that they minimize the error on the validation dataset. Meanwhile, \\\\emph{annotator confusion} assumes the existence of multiple annotators and introduces a regularized EM-based approach to model the label transition probability; its regularizer enables the estimated transition probability to converge to the true confusion matrix of the annotators. \\nIn contrast, \\\\emph{pre-training} empirically proves that fine-tuning on a pre-trained model provides a significant improvement in robustness compared with models trained from scratch; the universal representations of pre-training prevent the model parameters from being updated in the wrong direction by noisy labels. { \\\\emph{PHuber} proposes a composite loss-based gradient clipping, which is a variation of standard gradient clipping for label noise robustness. \\\\emph{Robust early-learning} classifies critical parameters and non-critical parameters for fitting clean and noise labels, respectively. Then, it penalizes only the non-critical ones with a different update rule.} \\n{\\\\emph{ODLN} leverages open-set auxiliary data and prevents the overfitting to noisy labels by assigning random labels to the open-set examples, which are uniformly sampled from the label set.}\\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: The explicit regularization often introduces sensitive model-dependent hyperparameters or requires deeper architectures to compensate for the reduced capacity, yet it can lead to significant performance gain if they are optimally tuned. \\\\looseness=-1}\\n\\\\smallskip\",\n \"id\": \"84774a8d-19b4-4ef8-ba90-419f51709a46\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"25e15182-1991-4f1a-83d9-017fe1a40d14\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Robust Regularization\"\n ],\n [\n \"subsubsection\",\n \"Explicit Regularization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Explicit Regularization\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7191,\n 4141,\n 4148,\n 892\n ],\n \"content\": \"}\\nThe regularization can also be an {implicit} form that gives the effect of stochasticity, e.g., data augmentation and mini-batch stochastic gradient descent.\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}: \\\\emph{Adversarial training} enhances the noise tolerance by encouraging the DNN to correctly classify both original inputs and hostilely perturbed ones. {\\\\emph{Label smoothing} estimates the marginalized effect of label noise during training, thereby reducing overfitting by preventing the DNN from assigning a full probability to noisy training examples. Instead of the one-hot label, the noisy label is mixed with a uniform mixture over all possible labels,\\n\\\\begin{equation}\\n\\\\begin{gathered}\\n\\\\bar{y} = \\\\big\\\\langle \\\\bar{y}(1), \\\\bar{y}(2), \\\\dots, \\\\bar{y}(c) \\\\big\\\\rangle,\\\\\\\\ \\n\\\\text{where} ~ \\\\bar{y}(i) = (1-\\\\alpha) \\\\cdot [\\\\tilde{y} = i] + \\\\alpha / c ~\\\\text{and}~ \\\\alpha \\\\in [0, 1].\\n\\\\end{gathered}\\n\\\\end{equation}\\nHere, $[\\\\cdot]$ is the Iverson bracket and $\\\\alpha$ is the smoothing degree.} In contrast, \\\\emph{mixup} regularizes the DNN to favor simple linear behaviors in between training examples. \\nFirst, the mini-batch is constructed using virtual training examples, each of which is formed by the linear interpolation of two noisy training examples $(x_i, \\\\tilde{y}_i)$ and $(x_j, \\\\tilde{y}_j)$ obtained at random from noisy training data $\\\\tilde{\\\\mathcal{D}}$,\\n\\\\begin{equation}\\n\\\\label{eq:mixup_construction}\\n{x}_{mix} = \\\\lambda x_i + (1-\\\\lambda) x_j~~ \\\\text{and} ~~{y}_{mix} = \\\\lambda \\\\tilde{y}_i + (1-\\\\lambda) \\\\tilde{y}_j,\\n\\\\end{equation}\\nwhere $\\\\lambda \\\\in [0,1]$ is the balance parameter between two examples. Thus, \\\\emph{mixup} extends the training distribution by updating the DNN for the constructed mini-batch. \\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: The implicit regularization improves the generalization capability of the DNN without reducing the representational capacity. It also does not introduce sensitive model-dependent hyperparameters because it is applied to the training data. However, the extended feature or label space slows down the convergence of training.}\",\n \"id\": \"50da4bbd-f1a0-49f9-b327-8d43650a94c6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"25e15182-1991-4f1a-83d9-017fe1a40d14\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Robust Regularization\"\n ],\n [\n \"subsubsection\",\n \"Implicit Regularization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Implicit Regularization\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 4130,\n 3454,\n 4132,\n 4151,\n 4150,\n 4128,\n 4149,\n 8736,\n 4134\n ],\n \"content\": \"\\\\label{sec:robust_loss_function}\\n{\\nIt was proven that a learned DNN with a \\\\emph{suitably modified} loss function $\\\\ell^{\\\\prime}$ for noisy data $\\\\tilde{\\\\mathcal{D}}$ can approach the Bayes optimal classifier $f^{*}$, which achieves the optimal Bayes risk $\\\\mathcal{R}^{*} = \\\\mathcal{R}_{\\\\mathcal{D}}(f^{*})$ for clean data $\\\\mathcal{D}$. Let $\\\\hat{f} = \\\\text{argmin}_{f \\\\in \\\\mathcal{F}} \\\\hat{\\\\mathcal{R}}_{\\\\ell^{\\\\prime}, \\\\tilde{\\\\mathcal{D}}}(f)$ be the learned classifier with the modified loss $\\\\ell^{\\\\prime}$ for the noisy data, where $\\\\hat{\\\\mathcal{R}}_{\\\\ell^{\\\\prime}, \\\\tilde{\\\\mathcal{D}}}(f) = \\\\mathbb{E}_{\\\\tilde{\\\\mathcal{D}}}[\\\\ell(f(x;\\\\Theta), \\\\tilde{y})]$. If $\\\\ell$ is $L$-Lipschitz and classification-calibrated , with probability at least $1\\\\!-\\\\!\\\\delta$, there exists a non-decreasing function $\\\\zeta_{\\\\ell}$ with $\\\\zeta_{\\\\ell}(0)=0$ such that \\\\looseness=-1\\n\\\\vspace*{-0.6cm}\\n\\\\begin{equation}\\n\\\\begin{split}\\n\\\\mathcal{R}_{\\\\mathcal{D}}(\\\\hat{f}) - \\\\mathcal{R}^{*} \\\\leq &~\\\\overbrace{\\\\zeta_{\\\\ell} \\\\Big( \\\\text{min}_{f\\\\in \\\\mathcal{F}}\\\\mathcal{R}_{\\\\ell,\\\\mathcal{D}}(f) - \\\\text{min}_{f}\\\\mathcal{R}_{\\\\ell, \\\\mathcal{D}}(f)}^{\\\\text{Approximation and Estimation Errors}}\\\\\\\\\\n&~~+ 4L_{p}\\\\text{RC}(\\\\mathcal{F}) + 2\\\\sqrt{{\\\\text{log}(1/\\\\delta)}\\\\big/{2|\\\\mathcal{D}|}} \\\\Big),\\\\!\\\\!\\\\!\\\\!\\\\!\\n\\\\end{split}\\n\\\\end{equation}\\n$L_{p}$ is the Lipschitz constant of $\\\\ell^{\\\\prime}$ and RC is the Rademacher complexity of the hypothesis class $\\\\mathcal{F}$. Then, by the universal approximation theorem , the Bayes optimal classifier $f^*$ is guaranteed to be in the hypothesis class $\\\\mathcal{F}$ with DNNs. \\nBased on this theoretical foundation, researchers have attempted to design robust loss functions such that they achieve a small risk for unseen clean data even when noisy labels exist in the training data . \\n}\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}: \\nInitially, Manwani and Sastry theoretically proved a sufficient condition for the loss function such that risk minimization with that function becomes noise-tolerant for binary classification. Subsequently, the sufficient condition was extended for multi-class classification using deep learning . Specifically, a loss function is defined to be \\\\emph{noise-tolerant} for a $c$-class classification under \\\\emph{symmetric} noise if the function satisfies the noise rate $\\\\tau<\\\\frac{c-1}{c}$ and\\n\\\\begin{equation}\\n\\\\label{eq:symmetric_function}\\n\\\\sum_{j=1}^{c}\\\\ell\\\\big(f(x;\\\\Theta),y=j\\\\big)=C, ~\\\\forall x\\\\in\\\\mathcal{X}, ~\\\\forall f,\\n\\\\end{equation}\\nwhere $C$ is a constant. This condition guarantees that the classifier trained on noisy data has the same misclassification probability as that trained on noise-free data under the specified assumption. {An extension for \\\\emph{multi-label} classification was provided by Kumar et al. .} Moreover, if $\\\\mathcal{R}_{\\\\mathcal{D}}(f^{*})=0$, then the function is also noise-tolerant under an \\\\emph{asymmetric} noise, where $f^{*}$ is a global risk minimizer of $\\\\mathcal{R}_{\\\\mathcal{D}}$. \\\\looseness=-1\\nFor the classification task, the categorical cross entropy\\\\,(CCE) loss is the most widely used loss function owing to its fast convergence and high generalization capability. However, in the presence of noisy labels, the \\\\emph{robust MAE} showed that the mean absolute error\\\\,(MAE) loss achieves better generalization than the CCE loss because only the MAE loss satisfies the aforementioned condition. A limitation of the MAE loss is that its generalization performance degrades significantly when complicated data are involved. Hence, the \\\\emph{generalized cross entropy}\\\\,(GCE) was proposed to achieve the advantages of both MAE and CCE losses; the GCE loss is a more general class of noise-robust loss that encompasses both of them. \\n{\\nAmid et al. extended the GCE loss by introducing two temperatures based on the Tsallis divergence. \\\\emph{Bi-tempered loss} introduces a proper unbiased generalization of the CE loss based on the Bregman divergence.}\\nIn addition, inspired by the symmetricity of the Kullback-Leibler divergence, the symmetric cross entropy\\\\,(SCE) was proposed by combining a noise tolerance term, namely reverse cross entropy loss, with the standard CCE loss. \\nMeanwhile, the \\\\emph{curriculum loss}\\\\,(CL) is a surrogate loss of the 0-1 loss function; it provides a tight upper bound and can easily be extended to multi-class classification. {The \\\\emph{active passive loss}\\\\,(APL) is a combination of two types of robust loss functions, an active loss that maximizes the probability of belonging to the given class and a passive loss that minimizes the probability of belonging to other classes.}\\n\\\\smallskip\\n\\\\noindent \\\\underline{Remark}: The robustness of these methods is theoretically supported well. However, they perform well only in simple cases, when learning is easy or the number of classes is small . Moreover, the modification of the loss function increases the training time for convergence .\\n\\\\vspace*{-0.2cm}\",\n \"id\": \"59ab6098-f130-49dc-b87f-123eaa6be482\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"60bd8c81-37fb-4934-ab18-bb9a24e80724\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Robust Loss Function\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Robust Loss Function\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 7162,\n 3340,\n 7775,\n 4135,\n 4145,\n 4139\n ],\n \"content\": \"\\\\label{sec:loss_adjustment}\\n\\\\vspace*{-0.0cm}\\nLoss adjustment is effective for reducing the negative impact of noisy labels by adjusting the loss of all training examples before updating the DNN . The methods associated with it can be categorized into three groups depending on their adjustment philosophy: \\\\emph{1)} {\\\\emph{loss correction}} that estimates the noise transition matrix to correct the forward or backward loss, \\\\emph{2)} {\\\\emph{loss reweighting}} that imposes different importance to each example for a weighted training scheme, \\\\emph{3)} {\\\\emph{label refurbishment}} that adjusts the loss using the refurbished label obtained from a convex combination of noisy and predicted labels, and \\\\emph{4)} {\\\\emph{meta learning}} that automatically infers the optimal rule for loss adjustment. {Unlike the robust loss function newly designed for robustness, this family of methods aims to make the traditional optimization process robust to label noise. Hence, in the middle of training, the {update rule} is adjusted such that the negative impact of label noise is minimized.}\\nIn general, loss adjustment allows for a \\\\emph{full exploration} of the training data by adjusting the loss of every example. However, the error incurred by \\\\emph{false} correction is accumulated, especially when the number of classes or the number of mislabeled examples is large . \\n\\\\vspace*{0.15cm}\",\n \"id\": \"3d40345d-b6d6-4b98-943c-ec552f10d7fb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"60bd8c81-37fb-4934-ab18-bb9a24e80724\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Loss Adjustment\"\n ]\n ],\n \"subsections\": [\n \"26fbf349-d8eb-41b0-9964-18115fa4704c\",\n \"67aed536-256b-4b94-954e-fe79fe9aa42b\",\n \"6322f675-ca49-4162-8e1c-216bde9693b8\",\n \"5cc6d843-bb17-416b-bbf4-7e089941b605\"\n ],\n \"title\": \"Loss Adjustment\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 7777,\n 4145,\n 4153,\n 4152\n ],\n \"content\": \"}\\n\\\\label{sec:loss_correction}\\nSimilar to the noise adaptation layer presented in Section \\\\ref{sec:robust_architecture}, this approach modifies the loss of each example by multiplying the estimated label transition probability by the output of a specified DNN. The main difference is that the learning of the transition probability is decoupled from that of the model. \\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\n\\\\emph{Backward correction} initially approximates the noise transition matrix using the softmax output of the DNN trained without loss correction. Subsequently, it retrains the DNN while correcting the original loss based on the estimated matrix. The corrected loss of a example $(x,\\\\tilde{y})$ is computed by a linear combination of its loss values for observable labels, whose coefficient is the inverse transition matrix $\\\\text{T}^{-1}$ to the observable label $y\\\\in\\\\{1,\\\\dots, c\\\\}$, given its target label $\\\\tilde{y}$.\\nTherefore, the backward correction $\\\\cev{\\\\ell}$ is performed by multiplying the inverse transition matrix to the prediction for all the observable labels, \\n{\\n\\\\begin{equation}\\n\\\\small\\n\\\\label{eq:backward_correction}\\n\\\\begin{split}\\n\\\\cev{\\\\ell}\\\\big(f(x;&\\\\Theta),\\\\tilde{y}\\\\big) = \\\\hat{\\\\text{T}}^{-1}\\\\Big\\\\langle\\\\ell\\\\big(f(x;\\\\Theta), 1\\\\big), \\\\dots, \\\\ell\\\\big(f(x;\\\\Theta), c\\\\big)\\\\Big\\\\rangle^{\\\\!\\\\top}\\\\!\\\\!,\\n\\\\end{split}\\n\\\\end{equation}\\n}\\n\\\\noindent where $\\\\hat{\\\\text{T}}$ is the estimated noise transition matrix. \\nConversely, \\\\emph{forward correction} uses a linear combination of a DNN's softmax outputs before applying the loss function. Hence, the forward correction $\\\\vec{\\\\ell}$ is performed by multiplying the estimated transition probability with the softmax outputs during the forward propagation step,\\n{\\n\\\\begin{equation}\\n\\\\small\\n\\\\label{eq:forward_correction}\\n\\\\begin{split}\\n\\\\vec{\\\\ell}\\\\big(f(x;\\\\Theta),\\\\tilde{y}\\\\big)& = \\\\ell\\\\Big(\\\\Big\\\\langle\\\\hat{p}(\\\\tilde{y}|1),\\\\dots,\\\\hat{p}(\\\\tilde{y}|c)\\\\Big\\\\rangle f(x;\\\\Theta)^{\\\\top},\\\\tilde{y}\\\\Big)\\\\\\\\\\n&=\\\\ell\\\\big(\\\\hat{\\\\text{T}}^{\\\\top}f(x;\\\\Theta)^{\\\\top},\\\\tilde{y}\\\\big).\\n\\\\end{split}\\n\\\\end{equation}\\n}\\n\\\\vspace*{-0.2cm}\\nFurthermore, \\\\emph{gold loss correction} assumes the availability of clean validation data or anchor points for loss correction. Thus, a more accurate transition matrix is obtained by using them as additional information, which further improves the robustness of the loss correction. {Recently, \\\\emph{T-Revision} provides a solution that can infer the transition matrix without anchor points, and \\\\emph{Dual T} factorizes the matrix into the product of two easy-to-estimate matrices to avoid directly estimating the noisy class posterior.}\\n{\\nBeyond the instance-independent noise assumption, Zhang et al. introduced the instance-confidence embedding to model instance-dependent noise in estimating the transition matrix. On the other hand, Yang et al. proposed to use the Bayes optimal transition matrix estimated from the distilled examples for the instance-dependent noise transition matrix.}\\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: The robustness of these approaches is highly dependent on how precisely the transition matrix is estimated. To acquire such a transition matrix, they require prior knowledge in general, such as anchor points or clean validation data.}\\n\\\\vspace*{0.15cm}\",\n \"id\": \"26fbf349-d8eb-41b0-9964-18115fa4704c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"3d40345d-b6d6-4b98-943c-ec552f10d7fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Loss Adjustment\"\n ],\n [\n \"subsubsection\",\n \"Loss Correction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Loss Correction\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 3453,\n 7775\n ],\n \"content\": \"} Inspired by the concept of importance reweighting , loss reweighting aims to assign smaller weights to the examples with false labels and greater weights to those with true labels. Accordingly, the reweighted loss on the mini-batch $\\\\mathcal{B}_t$ is used to update the DNN, \\n{\\n\\\\begin{equation}\\n\\\\label{eq:loss_reweighting}\\n\\\\Theta_{t+1} = \\\\Theta_{t} - \\\\eta\\\\nabla \\\\Big( \\\\frac{1}{|\\\\mathcal{B}_{t}|}\\\\!\\\\sum_{(x,\\\\tilde{y})\\\\in\\\\mathcal{B}_t}\\\\!\\\\!\\\\!\\\\!\\\\overbrace{w(x,\\\\tilde{y})\\\\ell\\\\big(f(x;\\\\Theta_t), \\\\tilde{y}\\\\big)}^{\\\\text{Reweighted Loss}}\\\\Big),\\n\\\\end{equation}}\\nwhere $w(x,\\\\tilde{y})$ is the weight of an example $x$ with its noisy label $\\\\tilde{y}$. Hence, the examples with smaller weights do not significantly affect the DNN learning.\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\nIn \\\\emph{importance reweighting} , the ratio of two joint data distributions {$w(x,\\\\tilde{y})=P_{\\\\mathcal{D}}(x,\\\\tilde{y})/P_{\\\\tilde{\\\\mathcal{D}}}(x,\\\\tilde{y})$} determines the contribution of the loss of each noisy example. An approximate solution to estimate the ratio was developed because the two distributions are difficult to determine from noisy data. Meanwhile, \\\\emph{active bias} emphasizes uncertain examples with inconsistent label predictions by assigning their prediction variances as the weights for training. \\n{\\n\\\\emph{DualGraph} employs graph neural networks and reweights the examples according to the structural relations among labels, eliminating the abnormal noise examples.}\\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: These approaches need to manually pre-specify the weighting function as well as there additional hyper-parameters, which is fairly hard to be applied in practice due to the significant variation of appropriate weighting schemes that rely on the noise type and training data.}\\n\\\\vspace*{0.15cm}\",\n \"id\": \"67aed536-256b-4b94-954e-fe79fe9aa42b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"3d40345d-b6d6-4b98-943c-ec552f10d7fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Loss Adjustment\"\n ],\n [\n \"subsubsection\",\n \"Loss Reweighting\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Loss Reweighting\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 7162,\n 4135,\n 4156,\n 4154,\n 4155,\n 4139\n ],\n \"content\": \"} Refurbishing a noisy label $\\\\tilde{y}$ effectively prevents overfitting to false labels. Let $\\\\hat{y}$ be the current prediction of the DNN $f(x;\\\\Theta)$. Therefore, the refurbished label $y^{refurb}$ can be obtained by a convex combination of the noisy label $\\\\tilde{y}$ and the DNN prediction $\\\\hat{y}$,\\n\\\\begin{equation}\\n\\\\label{eq:label_correction}\\ny^{refurb} = \\\\alpha \\\\tilde{y} + (1-\\\\alpha) \\\\hat{y},\\n\\\\end{equation}\\nwhere $\\\\alpha \\\\in [0,1]$ is the label confidence of $\\\\tilde{y}$. To mitigate the damage of incorrect labeling, this approach backpropagates the loss for the refurbished label instead of the noisy one, thereby yielding substantial robustness to noisy labels. \\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\n\\\\emph{Bootstrapping} is the first method that proposes the concept of label refurbishment to update the target label of training examples. It develops a more coherent network that improves its ability to evaluate the consistency of noisy labels, with the label confidence $\\\\alpha$ obtained via cross-validation.\\n\\\\emph{Dynamic bootstrapping} dynamically adjusts the confidence $\\\\alpha$ of individual training examples. The confidence $\\\\alpha$ is obtained by fitting a two-component and one-dimensional beta mixture model to the loss distribution of all training examples. {\\\\emph{Self-adaptive training} applies the exponential moving average to alleviate the instability issue of using instantaneous prediction of the current DNN, \\n\\\\begin{equation}\\n\\\\!\\\\!y_{t+1}^{refurb} = \\\\alpha y_t^{refurb} + (1-\\\\alpha)\\\\hat{y}, ~\\\\text{where}~ y_0^{refurb} = \\\\tilde{y}\\\\!\\\\!\\n\\\\end{equation}\\n}\\n\\\\emph{D2L} trains a DNN using a dimensionality-driven learning strategy to avoid overfitting to false labels. A simple measure called \\\\emph{local intrinsic dimensionality} is adopted to evaluate the confidence $\\\\alpha$ in considering that the overfitting is exacerbated by dimensional expansion. Hence, refurbished labels are generated to prevent the dimensionality of the representation subspace from expanding at a later stage of training. \\nRecently, \\\\emph{SELFIE} introduces a novel concept of \\\\emph{refurbishable examples} that can be corrected with high precision. The key idea is to consider the example with consistent label predictions as refurbishable because such consistent predictions correspond to its true label with a high probability owing to the learner's perceptual consistency. Accordingly, the labels of only refurbishable examples are corrected to minimize the number of falsely corrected cases. {Similarly, \\\\emph{AdaCorr} selectively refurbishes the label of noisy examples, but a theoretical error-bound is provided. Alternatively, \\\\emph{SEAL} averages the softmax output of a DNN on each example over the whole training process, then re-trains the DNN using the averaged soft labels.}\\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: Differently from loss correction and reweighting, all the noisy labels are explicitly replaced with other expected clean labels (or their combination). If there are not many confusing classes in data, these methods work well by refurbishing the noisy labels with high precision. In the opposite case, the DNN could overfit to wrongly refurbished labels.}\\n\\\\vspace*{0.15cm}\",\n \"id\": \"6322f675-ca49-4162-8e1c-216bde9693b8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"3d40345d-b6d6-4b98-943c-ec552f10d7fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Loss Adjustment\"\n ],\n [\n \"subsubsection\",\n \"Label Refurbishment\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Label Refurbishment\"\n },\n {\n \"cite_extract_rate\": 0.8888888888888881,\n \"cites\": [\n 4157,\n 4158,\n 4128,\n 3345,\n 1695,\n 7774,\n 7772,\n 8737\n ],\n \"content\": \"}\\n{In recent years, meta learning becomes an important topic in the machine learning community and is applied to improve noise robustness . The key concept is \\\\emph{learning to learn} that performs learning at a level higher than conventional learning, thus achieving data-agnostic and noise type-agnostic rules for better practical use. It is similar to loss reweighting and label refurbishment, but the adjustment is automated in a meta learning manner.\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\nFor the loss reweighting in Eq.~\\\\eqref{eq:loss_reweighting}, the goal is to learn the weight function $w(x,\\\\tilde{y})$. Specifically, \\\\emph{L2LWS} and \\\\emph{CWS} are unified neural architectures composed of a target DNN and a meta-DNN. The meta-DNN is trained on a small clean validation dataset; it then provides guidance to evaluate the weight score for the target DNN. Here, part of the two DNNs are shared and jointly trained to benefit from each other. \\\\emph{Automatic reweighting} is a meta learning algorithm that learns the weights of training examples based on their gradient directions. It includes a small clean validation dataset into the training dataset and reweights the backward loss of the mini-batch examples such that the updated gradient minimizes the loss of this validation dataset. \\\\emph{Meta-weight-net} parameterizes the weighting function as a multi-layer perceptron network with only one hidden layer. A meta-objective is defined to update its parameters such that they minimize the empirical risk of a small clean dataset. At each iteration, the parameter of the target network is guided by the weight function updated via the meta-objective. \\n{Likewise, \\\\emph{data coefficients}\\\\,(i.e., exemplar weights and true labels) are estimated by meta-optimization with a small clean set, which is only $0.2$\\\\% of the entire training set, while refurbishing the examples probably mislabeled.}\\nFor the label refurbishment in Eq.~\\\\eqref{eq:label_correction}, \\\\emph{knowledge distillation} adopts the technique of transferring knowledge from one expert model to a target model. The prediction from the expert DNN trained on small clean validation data is used instead of the prediction $\\\\hat{y}$ from the target DNN. \\n{\\\\emph{MLC} updates the target model with corrected labels provided by a meta model trained on clean validation data. The two models are trained concurrently via a bi-level optimization.} \\n\\\\smallskip\\n\\\\noindent \\\\underline{Remark}:\\nBy learning the update rule via meta learning, the trained network easily adapts to various types of data and label noise. Nevertheless, unbiased clean validation data is essential to minimize the auxiliary objective, although it may not be available in real-world data.}\\n\\\\vspace*{-0.1cm}\",\n \"id\": \"5cc6d843-bb17-416b-bbf4-7e089941b605\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"3d40345d-b6d6-4b98-943c-ec552f10d7fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Loss Adjustment\"\n ],\n [\n \"subsubsection\",\n \"Meta Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Meta Learning\"\n },\n {\n \"cite_extract_rate\": 0.631578947368421,\n \"cites\": [\n 3340,\n 4137,\n 7771,\n 4140,\n 4126,\n 3342,\n 4125,\n 4124,\n 4123,\n 4253,\n 8630,\n 4115\n ],\n \"content\": \"\\\\label{sec:sample_selection}\\nTo avoid any false corrections, many recent studies have adopted sample selection that involves selecting true-labeled examples from a noisy training dataset. In this case, the update equation in Eq.\\\\ \\\\eqref{eq:corrupted_update} is modified to render a DNN more robust for noisy labels. Let $\\\\mathcal{C}_t \\\\subseteq \\\\mathcal{B}_t$ be the identified \\\\emph{clean} examples at time $t$. Then, the DNN is updated only for the selected clean examples $\\\\mathcal{C}_t$, \\n\\\\begin{equation}\\n\\\\label{eq:clean_update}\\n\\\\Theta_{t+1} = \\\\Theta_{t} - \\\\eta\\\\nabla\\\\Big(\\\\frac{1}{|\\\\mathcal{C}_t|} \\\\!\\\\sum_{(x,\\\\tilde{y}) \\\\in \\\\mathcal{C}_t} \\\\!\\\\!\\\\!\\\\!\\\\ell\\\\big(f(x;\\\\Theta_{t}), \\\\tilde{y}\\\\big)\\\\Big),\\n\\\\end{equation}\\nwhere the rest mini-batch examples, which are likely to be false-labeled, are excluded to pursue robust learning. \\n\\\\newcommand{\\\\norm}[1]{\\\\left\\\\lVert#1\\\\right\\\\rVert}\\n{The {memorization nature} of DNNs has been explored theoretically and empirically to identify clean examples from noisy training data .\\nSpecifically, assuming clusterable data where the clusters are located on the unit Euclidean ball, Li et al. proved the distance from the initial weight ${W}_{0}$ to the weight ${W}_t$ after $t$ iterations,\\n\\\\begin{equation}\\n\\\\label{eq:memorization_effect_foundation}\\n\\\\norm{{W}_t - {W}_0}_{F} \\\\lesssim \\\\big( \\\\sqrt{K} + (K^{2}\\\\epsilon_{0}/\\\\norm{{C}}^{2})t \\\\big),\\n\\\\end{equation}\\nwhere $\\\\norm{\\\\cdot}_{F}$ is the Frobenius norm, $K$ is the number of clusters, and ${C}$ is the set of cluster centers reaching all input examples within their $\\\\epsilon_0$ neighborhood.\\nEq.\\\\ \\\\eqref{eq:memorization_effect_foundation} demonstrates that the weights of DNNs start to stray far from the initial weights when overfitting to corrupted labels, while they are still in the vicinity of the initial weights at an early stage of training . In the empirical studies , the \\\\emph{memorization effect} is also observed since DNNs tend to first learn simple and generalized patterns and then gradually overfit to all noisy patterns. As such, favoring small-loss training examples as the clean ones are commonly employed to design robust training methods .\\nLearning with sample selection is well motivated and works well in general, but this approach suffers from accumulated error caused by incorrect selection, especially when there are many ambiguous classes in training data.} \\nHence, recent approaches often leverage multiple DNNs to cooperate with one another or run multiple training rounds .\\nMoreover, to benefit from even false-labeled examples, loss correction or semi-supervised learning have been recently combined with the sample selection strategy .\\n\\\\begin{figure}\\n\\\\vspace*{+0.08cm}\\n\\\\begin{center}\\n\\\\includegraphics[width=8.7cm]{figures/methodology/loss_distribution.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.2cm}\\n\\\\hspace*{0.80cm} {\\\\small (a) Symmetric Noise $40\\\\%$.} \\\\hspace*{0.525cm} {\\\\small (b) Asymmetric Noise $40\\\\%$.}\\n\\\\vspace*{-0.15cm}\\n\\\\caption{Loss distribution of training examples at the training accuracy of $50\\\\%$ on noisy CIFAR-100. (This figure is adapted from Song et al. .)}\\n\\\\label{fig:loss_distribution}\\n\\\\vspace*{-0.4cm}\\n\\\\end{figure}\\n\\\\smallskip\\\\smallskip\",\n \"id\": \"54d73157-e47b-4943-9b30-b518e1806946\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"60bd8c81-37fb-4934-ab18-bb9a24e80724\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Sample Selection\"\n ]\n ],\n \"subsections\": [\n \"a00439ee-902b-4911-93a8-2198ddea3d64\",\n \"3f9409b8-29c2-4f75-91a3-9614418f8bb4\",\n \"3b9e7951-da54-4090-b74a-113ced54d832\"\n ],\n \"title\": \"Sample Selection\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 3340,\n 4140,\n 4137,\n 4159\n ],\n \"content\": \"}\\n{\\nCollaborative learning and co-training are widely used for the multi-network training. Consequently, the sample selection process is guided by the mentor network in the case of collaborative learning or the peer network in the case of co-training.\\n}\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\nInitially, \\\\emph{Decouple} proposes the decoupling of when to update from how to update. Hence, two DNNs are maintained simultaneously and updated only the examples selected based on a disagreement between the two DNNs.\\nNext, due to the memorization effect of DNNs, many researchers have adopted another selection criterion, called a \\\\emph{small-loss} trick, which treats a certain number of small-loss training examples as true-labeled examples; many true-labeled examples tend to exhibit smaller losses than false-labeled examples, as illustrated in Figure \\\\ref{fig:loss_distribution}(a). In \\\\emph{MentorNet} , a pre-trained mentor network guides the training of a student network in a collaborative learning manner. Based on the small-loss trick, the mentor network provides the student network with examples whose labels are likely to be correct. \\\\emph{Co-teaching} and \\\\emph{Co-teaching+} also maintain two DNNs, but each DNN selects a certain number of small-loss examples and feeds them to its peer DNN for further training. \\\\emph{Co-teaching+} further employs the disagreement strategy of \\\\emph{Decouple} compared with \\\\emph{Co-teaching}. {In contrast, \\\\emph{JoCoR} reduces the diversity of two networks via co-regularization, making predictions of the two networks closer.}\\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: The co-training methods help reduce the confirmation bias , which is a hazard of favoring the examples selected at the beginning of training, while the increase in the number of learnable parameters makes their learning pipeline inefficient. In addition, the small-loss trick does not work well when the loss distribution of true-labeled and false-labeled examples largely overlap, as in the asymmetric noise in Figure \\\\ref{fig:loss_distribution}(b).}\\n\\\\smallskip\\\\smallskip\",\n \"id\": \"a00439ee-902b-4911-93a8-2198ddea3d64\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"54d73157-e47b-4943-9b30-b518e1806946\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Sample Selection\"\n ],\n [\n \"subsubsection\",\n \"Multi-network Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-network Learning\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 4126,\n 3343,\n 8630,\n 4125,\n 7163,\n 4123\n ],\n \"content\": \"}\\nWithout maintaining additional DNNs, multi-round learning iteratively refines the selected set of clean examples by repeating the training round. Thus, the selected set keeps improved as the number of rounds increases. \\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\n\\\\emph{ITLM} iteratively minimizes the trimmed loss by alternating between selecting true-labeled examples at the current moment and retraining the DNN using them. At each training round, only a fraction of small-loss examples obtained in the current round are used to retrain the DNN in the next round. \\\\emph{INCV} randomly divides noisy training data and then employs cross-validation to classify true-labeled examples while removing large-loss examples at each training round. Here, \\\\emph{Co-teaching} is adopted to train the DNN on the identified examples in the final round of training. {Similarly, \\\\emph{O2U-Net} repeats the whole training process with the cyclical learning rate until enough loss statistics of every examples are gathered. Next, the DNN is re-trained from scratch only for the clean data where false-labeled examples have been detected and removed based on statistics.} \\\\looseness=-1\\nA number of variations have been proposed to achieve high performance using iterative refinement only in a single training round.\\nBeyond the small-loss trick, \\\\emph{iterative detection} detects false-labeled examples by employing the local outlier factor algorithm . With a Siamese network, it gradually pulls away false-labeled examples from true-labeled samples in the deep feature space. \\\\emph{MORPH} introduces the concept of memorized examples which is used to iteratively expand an initial safe set into a maximal safe set via self-transitional learning. \\\\emph{TopoFilter} utilizes the spatial topological pattern of learned representations to detect true-labeled examples, not relying on the prediction of the noisy classifier.\\n{\\n\\\\emph{NGC} iteratively constructs the nearest neighbor graph using latent representations and performs geometry-based sample selection by aggregating information from neighborhoods. Soft pesudo-labels are assigned to the examples not selected.\\n}\\n\\\\smallskip\\n\\\\noindent {\\\\underline{Remark}: The selected clean set keeps expanded and purified with iterative refinement, mainly through multi-round learning. As a side effect, the computational cost for training increases linearly for the number of training rounds.}\\n\\\\smallskip\",\n \"id\": \"3f9409b8-29c2-4f75-91a3-9614418f8bb4\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"54d73157-e47b-4943-9b30-b518e1806946\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Sample Selection\"\n ],\n [\n \"subsubsection\",\n \"Multi-round Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-round Learning\"\n },\n {\n \"cite_extract_rate\": 0.7741935483870961,\n \"cites\": [\n 3340,\n 3343,\n 7133,\n 4137,\n 7773,\n 3345,\n 8738,\n 4159,\n 4134,\n 4132,\n 4140,\n 4141,\n 7775,\n 7778,\n 4126,\n 3342,\n 4156,\n 4142,\n 7772,\n 4150,\n 8737,\n 4155,\n 4138,\n 4130,\n 7191,\n 4151,\n 4147,\n 4135,\n 7162,\n 4157,\n 4145,\n 7774,\n 4139,\n 4125,\n 4123,\n 892,\n 4128,\n 4152,\n 4253,\n 4158,\n 7777,\n 4131,\n 8630,\n 7163,\n 4133,\n 4154,\n 2277,\n 4136\n ],\n \"content\": \"} {An inherent limitation of sample selection is to discard all the \\\\emph{unselected} training examples, thus resulting in a \\\\emph{partial} exploration of training data. To exploit all the noisy examples, researchers have attempted to combine sample selection with other orthogonal ideas.\\n\\\\begin{figure}[t!]\\n\\\\begin{center}\\n\\\\includegraphics[width=8.45cm]{figures/methodology/semi_supervised.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.1cm}\\n\\\\hspace*{0.2cm} \\\\small{(a) Noisy Data.} \\\\hspace*{0.4cm} \\\\small{(b) Transformed Data.} \\\\hspace*{0.8cm} \\\\small{(c) SSL.}\\n\\\\vspace*{-0.2cm}\\n\\\\caption{{Procedures for semi-supervised learning under label noise.}}\\n\\\\label{fig:noise_semi_supervised}\\n\\\\vspace*{-0.45cm}\\n\\\\end{figure}\\n\\\\smallskip\\n\\\\noindent \\\\underline{Technical Detail}:\\nThe most prominent method in this direction is combining a specific sample selection strategy with a specific semi-supervised learning model. As illustrated in Figure \\\\ref{fig:noise_semi_supervised}, selected examples are treated as labeled clean data, whereas the remaining examples are treated as unlabeled. Subsequently, semi-supervised learning is performed using the transformed data. \\\\emph{SELF} is combined with a semi-supervised learning approach to progressively filter out false-labeled examples from noisy data. By maintaining the running average model called the {mean-teacher} as the backbone, it obtains the self-ensemble predictions of all training examples and then progressively removes examples whose ensemble predictions do not agree with their annotated labels. This method further leverages unsupervised loss from the examples not included in the selected clean set. \\\\emph{DivideMix} uses two-component and one-dimensional Gaussian mixture models to transform noisy data into labeled (clean) and unlabeled (noisy) sets. Then, it applies a semi-supervised technique \\\\emph{MixMatch} . Recently, \\\\emph{RoCL} employs two-phase learning strategies: supervised training on selected clean examples and then semi-supervised learning on relabeled noisy examples with self-supervision. For selection and relabeling, it computes the exponential moving average of the loss over training iterations. \\n}\\n\\\\clearpage\\n{\\n\\\\newcolumntype{L}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}m{#1}}\\n\\\\newcolumntype{X}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}p{#1}}\\n\\\\newcolumntype{Y}[1]{>{\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{1pt}}m{#1}}\\n\\\\begin{table*}[h]\\n\\\\vspace*{-3.9cm}\\n\\\\caption{{Comparison of proposed robust deep learning methods with respect to the following six properties: (P1)\\\\,Flexibility, (P2)\\\\,No Pre-training, (P3)\\\\,Full Exploration, (P4)\\\\,No Supervision, (P5)\\\\,Heavy Noise, and (P6)\\\\,Complex Noise.}}\\n\\\\vspace*{-0.2cm}\\n\\\\begin{tabular}{L{0.7cm}|L{2.5cm}|Y{4.7cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|X{0.5cm}|Y{2.9cm}}\\n\\\\toprule\\n\\\\multicolumn{2}{c|}{\\\\textbf{Category}} & \\\\hspace*{2.0cm}\\\\textbf{Method} & \\\\textbf{P1} & \\\\textbf{P2} & \\\\textbf{P3} & \\\\textbf{P4} & \\\\textbf{P5} & \\\\textbf{P6} & \\\\,\\\\,\\\\,\\\\,\\\\,\\\\,\\\\,\\\\,\\\\,\\\\textbf{Implementation}\\\\!\\\\! \\\\\\\\ \\\\hline\\n\\\\multirow{9}{*}{\\\\vspace*{-0.35cm}\\\\hspace*{0.13cm}\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!\\\\rotatebox[origin=c]{90}{\\\\makecell[l]{Robust Architecture\\\\\\\\\\\\hspace*{0.66cm}(\\\\textsection \\\\ref{sec:robust_architecture})}}\\\\!\\\\!\\\\!\\\\!\\\\!} \\n& \\\\multirow{6}{*}{\\\\vspace*{-0.2cm}\\\\makecell[l]{Noisy Adaptation\\\\\\\\\\\\hspace*{0.67cm}Layer}} \\n& \\\\emph{Webly Learning} & \\\\cellcolor{gray!10}$\\\\bigtriangleup$ & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(Caffe)\\\\protect\\\\footnotemark[1] \\\\\\\\\\\\cline{3-10} \\n& & \\\\emph{Noise Model} & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Unofficial\\\\,(Keras)\\\\protect\\\\footnotemark[2] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Dropout Noise Model} & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(MATLAB)\\\\protect\\\\footnotemark[3] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{S-model} & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(Keras)\\\\protect\\\\footnotemark[4]\\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{C-model} & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & Official\\\\,(Keras)\\\\protect\\\\footnotemark[4] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{NLNN} & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Unofficial\\\\,(Chainer)\\\\protect\\\\footnotemark[5] \\\\\\\\\\\\cmidrule{2-10}\\n& \\\\multirow{4}{*}{\\\\makecell[l]{\\\\hspace*{0.12cm}Dedicated\\\\\\\\Architecture}} \\n& \\\\emph{Probablistic Noise Model} & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & $ \\\\cellcolor{gray!10}\\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & Official\\\\,(Caffe)\\\\protect\\\\footnotemark[6] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Masking} & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & $ \\\\cellcolor{gray!10}\\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[7] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Contrastive-Additive Noise Network} & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark &\\\\cellcolor{blue!10}\\\\cmark & $ \\\\cellcolor{gray!10}\\\\bigtriangleup$ &\\\\cellcolor{blue!10}\\\\cmark & N/A \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{RoG}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark &\\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark &$ \\\\cellcolor{gray!10}\\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[8] \\\\\\\\\\\\midrule\\n\\\\multirow{9}{*}{\\\\vspace*{-0.25cm}\\\\hspace*{0.13cm}\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!\\\\rotatebox[origin=c]{90}{\\\\makecell[l]{Robust Regularization\\\\\\\\\\\\hspace*{0.8cm}(\\\\textsection \\\\ref{sec:robust_regularization})}}\\\\!\\\\!\\\\!\\\\!\\\\!} \\n& \\\\multirow{6}{*}{\\\\makecell[l]{\\\\hspace*{0.35cm}Explicit\\\\\\\\Regularization}} \\n& \\\\emph{Bilevel Learning} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[9] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Annotator Confusion} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[10] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Pre-training} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[11] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{PHuber}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(PyTorch)\\\\protect\\\\footnotemark[12] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{Robust Early-learning}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[13] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{ODLN}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[14] \\\\\\\\\\\\cmidrule{2-10}\\n& \\\\multirow{3}{*}{\\\\makecell[l]{\\\\hspace*{0.35cm}Implicit\\\\\\\\Regularization}} \\n& \\\\emph{Adversarial Training} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(PyTorch)\\\\protect\\\\footnotemark[15] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Label Smoothing} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(PyTorch)\\\\protect\\\\footnotemark[16] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Mixup} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[17] \\\\\\\\\\\\midrule\\n\\\\multicolumn{2}{c|}{\\\\multirow{7}{*}{{\\\\makecell[l]{Robust Loss Function\\\\\\\\\\\\hspace*{0.75cm}(\\\\textsection \\\\ref{sec:robust_loss_function})}}}} \\n& \\\\emph{Robust MAE} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & N/A \\\\\\\\\\\\cline{3-10}\\n\\\\multicolumn{2}{c|}{} & \\\\emph{Generalized Cross Entropy} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Unofficial\\\\,(PyTorch)\\\\protect\\\\footnotemark[18] \\\\\\\\\\\\cline{3-10}\\n\\\\multicolumn{2}{c|}{} & \\\\emph{Symmetric Cross Entropy} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(Keras)\\\\protect\\\\footnotemark[19] \\\\\\\\\\\\cline{3-10}\\n\\\\multicolumn{2}{c|}{} & {\\\\emph{Bi-tempered Loss}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$\\\\bigtriangleup$ & \\\\cellcolor{gray!10}$\\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[20] \\\\\\\\\\\\cline{3-10}\\n\\\\multicolumn{2}{c|}{} & \\\\emph{Curriculum Learning} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$\\\\bigtriangleup$ & N/A \\\\\\\\\\\\cline{3-10}\\n\\\\multicolumn{2}{c|}{} & {\\\\emph{Active Passive Loss}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$\\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[21] \\\\\\\\\\\\midrule\\n\\\\multirow{21}{*}{\\\\vspace*{-0.7cm}\\\\hspace*{0.13cm}\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!\\\\rotatebox[origin=c]{90}{\\\\makecell[l]{Loss Adjustment\\\\\\\\\\\\hspace*{0.5cm}(\\\\textsection \\\\ref{sec:loss_adjustment})}}\\\\!\\\\!\\\\!\\\\!\\\\!} \\n& \\\\multirow{5}{*}{\\\\makecell[l]{Loss Correction}} \\n& \\\\emph{Backward Correction} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(Keras)\\\\protect\\\\footnotemark[22] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Forward Correction} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(Keras)\\\\protect\\\\footnotemark[22] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Gold Loss Correction} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[23] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{T-revision}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[24] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{Dual T}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{red!10}\\\\xmark & N/A \\\\\\\\\\\\cmidrule{2-10}\\n& \\\\multirow{3}{*}{\\\\makecell[l]{Loss Reweigting}} \\n& \\\\emph{Importance Reweighting} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(PyTorch)\\\\protect\\\\footnotemark[25] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Active Bias} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(TensorFlow)\\\\protect\\\\footnotemark[26]\\\\!\\\\! \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{DualGraph}} & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & N/A \\\\\\\\\\\\cmidrule{2-10}\\n& \\\\multirow{6}{*}{\\\\makecell[l]{Label Refurbishment}} \\n& \\\\emph{Bootstrapping} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(Keras)\\\\protect\\\\footnotemark[27] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Dynamic Bootstrapping} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[28] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{Self-adaptive Training}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[29] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{D2L} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(Keras)\\\\protect\\\\footnotemark[30] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{AdaCorr}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[31] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{SEAL}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[32] \\\\\\\\\\\\cmidrule{2-10}\\n& \\\\multirow{7}{*}{\\\\makecell[l]{Meta Learning}} \\n& \\\\emph{L2LWS} & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(TensorFlow)\\\\protect\\\\footnotemark[33]\\\\!\\\\! \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{CWS} & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & N/A \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Automatic Reweighting} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[34] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Meta-weight-net} & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[35] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{Data Coefficients}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[36] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Knowledge Distillation} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & N/A \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{MLC}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[37] \\\\\\\\\\\\midrule\\n\\\\multirow{13}{*}{\\\\vspace*{-0.45cm}\\\\hspace*{0.13cm}\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!\\\\rotatebox[origin=c]{90}{\\\\makecell[l]{Sample Selection\\\\\\\\\\\\hspace*{0.5cm}(\\\\textsection \\\\ref{sec:sample_selection})}}\\\\!\\\\!\\\\!\\\\!\\\\!} \\n& \\\\multirow{4}{*}{\\\\makecell[l]{Multi-Network\\\\\\\\\\\\hspace*{0.33cm}Learning}} \\n& \\\\emph{Decouple} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[38] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{MentorNet} & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[39] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Co-teaching} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[40] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Co-teaching+} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[41] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{JoCoR}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[42] \\\\\\\\\\\\cmidrule{2-10}\\n& \\\\multirow{6}{*}{\\\\makecell[l]{Multi-Round\\\\\\\\\\\\hspace*{0.23cm}Learning}} \\n& \\\\emph{ITLM} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(GluonCV)\\\\protect\\\\footnotemark[43] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{INCV} & \\\\cellcolor{blue!10}\\\\cmark &\\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(Keras)\\\\protect\\\\footnotemark[44] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{O2U-Net}} & \\\\cellcolor{blue!10}\\\\cmark &\\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Unofficial\\\\,(PyTorch)\\\\protect\\\\footnotemark[45] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{Iterative Detection} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(Keras)\\\\protect\\\\footnotemark[46] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{MORPH}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & N/A \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{TopoFilter}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[47] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{NGC}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & N/A\\\\\\\\\\\\cmidrule{2-10}\\n& \\\\multirow{3}{*}{\\\\vspace*{-0.24cm}\\\\makecell[l]{Hybrid Approach}} \\n& \\\\emph{SELFIE} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(TensorFlow)\\\\protect\\\\footnotemark[48] \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{SELF} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & N/A \\\\\\\\\\\\cline{3-10}\\n& & \\\\emph{DivideMix} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & Official\\\\,(PyTorch)\\\\protect\\\\footnotemark[49] \\\\\\\\\\\\cline{3-10}\\n& & {\\\\emph{RoCL}} & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & N/A \\\\\\\\\\\\bottomrule\\n\\\\end{tabular}\\n\\\\label{table:all_comparision}\\n\\\\vspace*{-4cm}\\n\\\\end{table*}\\n}\\n\\\\clearpage\\n{\\n\\\\footnotetext[1]{\\\\url{https://github.com/endernewton/webly-supervised}}\\n\\\\footnotetext[2]{\\\\url{https://github.com/delchiaro/training-cnn-noisy-labels-keras}}\\n\\\\footnotetext[3]{\\\\url{https://github.com/ijindal/Noisy_Dropout_regularization}}\\n\\\\footnotetext[4]{\\\\url{https://github.com/udibr/noisy_labels}}\\n\\\\footnotetext[5]{\\\\url{https://github.com/Ryo-Ito/Noisy-Labels-Neural-Network}}\\n\\\\footnotetext[6]{\\\\url{https://github.com/Cysu/noisy_label}}\\n\\\\footnotetext[7]{\\\\url{https://github.com/bhanML/Masking}}\\n\\\\footnotetext[8]{\\\\url{https://github.com/pokaxpoka/RoGNoisyLabel}}\\n\\\\footnotetext[9]{\\\\url{https://github.com/sjenni/DeepBilevel}}\\n\\\\footnotetext[10]{\\\\url{https://rt416.github.io/pdf/trace_codes.pdf}}\\n\\\\footnotetext[11]{\\\\url{github.com/hendrycks/pre-training}}\\n\\\\footnotetext[12]{\\\\url{https://github.com/dmizr/phuber}}\\n\\\\footnotetext[13]{\\\\url{https://github.com/xiaoboxia/CDR}}\\n\\\\footnotetext[14]{\\\\url{https://github.com/hongxin001/ODNL?ref=pythonrepo.com}}\\n\\\\footnotetext[15]{\\\\url{https://https://github.com/sarathknv/adversarial-examples-pytorch}}\\n\\\\footnotetext[16]{\\\\url{https://github.com/CoinCheung/pytorch-loss}}\\n\\\\footnotetext[17]{\\\\url{https://github.com/facebookresearch/mixup-cifar10}}\\n\\\\footnotetext[18]{\\\\url{https://github.com/AlanChou/Truncated-Loss}}\\n\\\\footnotetext[19]{\\\\href{https://github.com/YisenWang/symmetric\\\\_cross\\\\_entropy\\\\_for\\\\_noisy_label}{https://github.com/YisenWang/symmetric\\\\_cross\\\\_entropy}}\\n\\\\footnotetext[20]{\\\\url{https://github.com/google/bi-tempered-loss}}\\n\\\\footnotetext[21]{\\\\url{https://github.com/HanxunH/Active-Passive-Losses}}\\n\\\\footnotetext[22]{\\\\url{https://github.com/giorgiop/loss-correction}}\\n\\\\footnotetext[23]{\\\\url{https://github.com/mmazeika/glc}}\\n\\\\footnotetext[24]{\\\\url{https://github.com/xiaoboxia/T-Revision}}\\n\\\\footnotetext[25]{\\\\href{https://github.com/xiaoboxia/Classification-with-noisy-labels-by-importance-reweighting}{https://github.com/xiaoboxia/Classification-with-noisy-labels}}\\n\\\\newcolumntype{L}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}m{#1}}\\n\\\\newcolumntype{X}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}p{#1}}\\n\\\\begin{table*}[!htb]\\n\\\\centering\\n\\\\vspace*{-0.1cm}\\n\\\\caption{Comparison of robust deep learning categories for overcoming noisy labels.}\\n\\\\vspace*{-0.2cm}\\n\\\\begin{tabular}{L{2.2cm} | L{2.85cm}|X{1.5cm}|X{1.55cm}|X{1.8cm}|X{1.65cm}|X{1.5cm}|X{1.6cm}}\\n\\\\toprule\\n\\\\multicolumn{2}{>{}c|}{\\\\multirow{2}{*}{\\\\textbf{Category}}} & \\\\textbf{P1} & \\\\textbf{P2} & \\\\textbf{P3} & \\\\textbf{P4} & \\\\textbf{P5} & \\\\textbf{P6} \\\\\\\\\\n\\\\multicolumn{2}{>{}c|}{} & \\\\!\\\\!\\\\!{Flexibility}\\\\!\\\\!\\\\! & \\\\!\\\\!\\\\!{No Pre-train}\\\\!\\\\!\\\\! & \\\\!\\\\!\\\\!{Full Exploration}\\\\!\\\\!\\\\! & \\\\!\\\\!\\\\!{No Supervision}\\\\!\\\\!\\\\! & \\\\!\\\\!\\\\!{Heavy Noise}\\\\!\\\\!\\\\! & \\\\!\\\\!\\\\!{Complex Noise}\\\\!\\\\!\\\\! \\\\\\\\\\\\midrule\\n\\\\multirow{2}{*}\\n{\\\\makecell[l]{\\\\!\\\\!\\\\!Robust Architecture\\\\!\\\\!\\\\! \\\\\\\\\\\\hspace*{0.5cm}(\\\\textsection \\\\ref{sec:robust_architecture})}}\\n& \\\\!\\\\!Noise Adaptation Layer & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark \\\\\\\\\\n\\\\cline{2-8}\\n& \\\\!\\\\!Dedicated Architecture & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{blue!10}\\\\cmark\\\\\\\\ \\\\midrule\\n\\\\multirow{2}{*}\\n{\\\\makecell[l]{\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!Robust Regularization\\\\!\\\\!\\\\!\\\\!\\\\!\\\\! \\\\\\\\\\\\hspace*{0.5cm}(\\\\textsection \\\\ref{sec:robust_regularization})}}\\n& \\\\!\\\\!Implicit Regularization & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ \\\\\\\\\\n\\\\cline{2-8}\\n& \\\\!\\\\!Explicit Regularization & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ \\\\\\\\ \\\\midrule\\n\\\\multicolumn{2}{>{}c|}{\\\\hspace*{0cm}Robust Loss Function (\\\\textsection \\\\ref{sec:robust_loss_function}) }& \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark \\\\\\\\\\\\midrule\\n\\\\multirow{2}{*}\\n{\\\\vspace*{-0.58cm}\\\\!\\\\makecell[l]{Loss Adjustment \\\\\\\\\\\\hspace*{0.52cm}(\\\\textsection \\\\ref{sec:loss_adjustment})}}\\n& \\\\!\\\\!Loss Correction &\\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark &\\\\cellcolor{red!10}\\\\xmark \\\\\\\\\\n\\\\cline{2-8}\\n& \\\\!\\\\!Loss Reweighting & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$\\\\\\\\\\n\\\\cline{2-8}\\n& \\\\!\\\\!Label Refurbishment & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$\\\\\\\\\\\\cline{2-8}\\n& \\\\!\\\\!Meta Learning & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ \\\\\\\\ \\\\midrule\\n{\\\\vspace*{-0.55cm}\\\\hspace*{0.075cm}\\\\makecell[l]{Sample Selection\\\\\\\\\\\\hspace*{0.55cm}(\\\\textsection \\\\ref{sec:sample_selection})}}\\n& \\\\!\\\\!Multi-Network Learning &\\\\cellcolor{blue!10}\\\\cmark &\\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$ \\\\\\\\\\n\\\\cline{2-8}\\n& \\\\!\\\\!Multi-Round Learning & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{red!10}\\\\xmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$\\\\\\\\\\\\cline{2-8}\\n& \\\\!\\\\!Hybrid Approach & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{blue!10}\\\\cmark & \\\\cellcolor{gray!10}$ \\\\bigtriangleup$\\\\\\\\\\\\bottomrule\\n\\\\end{tabular}\\n\\\\label{table:direction_comparison}\\n\\\\vspace*{-0.4cm}\\n\\\\end{table*}\\n}\\nMeanwhile, \\\\emph{SELFIE} is a hybrid approach of sample selection and loss correction. The loss of refurbishable examples is corrected (i.e., loss correction) and then used together with that of small-loss examples (i.e., sample selection). Consequently, more training examples are considered for updating the DNN.\\nThe \\\\emph{curriculum loss\\\\,(CL)} is combined with the robust loss function approach and used to extract the true-labeled examples from noisy data.\\n\\\\smallskip\\n\\\\noindent \\\\underline{Remark}:\\nNoise robustness is significantly improved by combining with other techniques. However, the hyperparameters introduced by these techniques render a DNN more susceptible to changes in data and noise types, and an increase in computational cost is inevitable\",\n \"id\": \"3b9e7951-da54-4090-b74a-113ced54d832\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 62,\n \"parent_id\": \"54d73157-e47b-4943-9b30-b518e1806946\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Deep Learning Approaches\"\n ],\n [\n \"subsection\",\n \"Sample Selection\"\n ],\n [\n \"subsubsection\",\n \"Hybrid Approach\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hybrid Approach\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 3340\n ],\n \"content\": \"\\\\label{sec:comparison}\\nIn this section, we compare the {$62$} deep learning methods for overcoming noisy labels introduced in Section \\\\ref{sec:methodology} with respect to the following \\\\emph{six} properties. When selecting the properties, we refer to the properties that are typically used to compare the performance of robust deep learning methods . To the best of our knowledge, this survey is the first to provide a systematic comparison of robust training methods. \\nThis comprehensive comparison will provide useful insights that can enlighten new future directions.\\n\\\\vspace*{0.1cm}\\n\\\\begin{itemize}[leftmargin=9pt]\\n\\\\item \\\\textbf{(P1)\\\\,Flexibility:} With the rapid evolution of deep learning research, a number of new network architectures are constantly emerging and becoming available. Hence, the ability to support any type of architecture is important. \\\\enquote{Flexibility} ensures that the proposed method can quickly adapt to the state-of-the-art architecture.\\n\\\\vspace*{0.1cm}\\n\\\\item \\\\textbf{(P2)\\\\,No Pre-training:} A typical approach to improve noise robustness is to use a pre-trained network; however, this incurs an additional computational cost to the learning process. \\\\enquote{No Pre-training} ensures that the proposed method can be trained from scratch without any pre-training.\\n\\\\vspace*{0.1cm}\\n\\\\item \\\\textbf{(P3)\\\\,Full Exploration:} Excluding unreliable examples from the update is an effective method for robust deep learning; however, it eliminates hard but useful training examples as well. \\\\enquote{Full Exploration} ensures that the proposed methods can use \\\\emph{all} training examples without severe overfitting to false-labeled examples by adjusting their training losses or applying semi-supervised learning.\\n\\\\vspace*{0.1cm}\\n\\\\item \\\\textbf{(P4)\\\\,No Supervision:} Learning with supervision, such as a clean validation set or a known noise rate, is often impractical because they are difficult to obtain. Hence, such supervision had better be avoided to increase practicality in real-world scenarios. \\\\enquote{No Supervision} ensures that the proposed methods can be trained without any supervision.\\n\\\\vspace*{0.1cm}\\n\\\\item \\\\textbf{(P5)\\\\,Heavy Noise:} In real-world noisy data, the noise rate can vary from light to heavy. Hence, learning methods should achieve consistent noise robustness with respect to the noise rate. \\\\enquote{Heavy Noise} ensures that the proposed methods can combat even the heavy noise.\\n\\\\end{itemize}\\n{\\n\\\\footnotetext[26]{\\\\url{https://github.com/songhwanjun/ActiveBias}}\\n\\\\footnotetext[27]{\\\\url{https://github.com/dr-darryl-wright/Noisy-Labels-with-Bootstrapping}}\\n\\\\footnotetext[28]{\\\\url{https://github.com/PaulAlbert31/LabelNoiseCorrection}}\\n\\\\footnotetext[29]{\\\\url{https://github.com/LayneH/self-adaptive-training}}\\n\\\\footnotetext[30]{\\\\url{https://github.com/xingjunm/dimensionality-driven-learning}}\\n\\\\footnotetext[31]{\\\\url{https://github.com/pingqingsheng/LRT}}\\n\\\\footnotetext[32]{\\\\url{https://github.com/chenpf1025/IDN}}\\n\\\\footnotetext[33]{\\\\url{https://github.com/krayush07/learn-by-weak-supervision}}\\n\\\\footnotetext[34]{\\\\url{https://github.com/uber-research/learning-to-reweight-examples}}\\n\\\\footnotetext[35]{\\\\url{https://github.com/xjtushujun/meta-weight-net}}\\n\\\\footnotetext[36]{\\\\url{https://github.com/google-research/google-research/tree/master/ieg}}\\n\\\\footnotetext[37]{\\\\url{https://aka.ms/MLC}}\\n\\\\footnotetext[38]{\\\\url{https://github.com/emalach/UpdateByDisagreement}}\\n\\\\footnotetext[39]{\\\\url{https://github.com/google/mentornet}}\\n\\\\footnotetext[40]{\\\\url{https://github.com/bhanML/Co-teaching}}\\n\\\\footnotetext[41]{\\\\url{https://github.com/bhanML/coteaching_plus}}\\n\\\\footnotetext[42]{\\\\url{https://github.com/hongxin001/JoCoR}}\\n\\\\footnotetext[43]{\\\\url{https://github.com/yanyao-shen/ITLM-simplecode}}\\n\\\\footnotetext[44]{\\\\url{https://github.com/chenpf1025/noisy_label_understanding_utilizing}}\\n\\\\footnotetext[45]{\\\\url{https://github.com/hjimce/O2U-Net}}\\n\\\\footnotetext[46]{\\\\url{https://github.com/YisenWang/Iterative_learning}}\\n\\\\footnotetext[47]{\\\\url{https://github.com/pxiangwu/TopoFilter}}\\n\\\\footnotetext[48]{\\\\url{https://github.com/kaist-dmlab/SELFIE}}\\n\\\\footnotetext[49]{\\\\url{https://github.com/LiJunnan1992/DivideMix}}\\n}\\n\\\\begin{itemize}[leftmargin=9pt]\\n\\\\item \\\\textbf{(P6)\\\\,Complex Noise:} The type of label noise significantly affects the performance of a learning method. To manage real-world noisy data, diverse types of label noise should be considered when designing a robust training method. \\\\enquote{Complex Noise} ensures that the proposed method can combat even the complex label noise.\\n\\\\end{itemize}\\n\\\\vspace*{0.05cm}\\nTable \\\\ref{table:all_comparision} shows a comparison of all robust deep learning methods, which are grouped according to the most appropriate category. In the first row, the aforementioned six properties are labeled as P1--P6, and the availability of open-source implementation is added in the last column. For each property, we assign \\\\enquote{\\\\cmark} if it is completely supported, \\\\enquote{\\\\xmark} if it is not supported, and \\\\enquote{$\\\\bigtriangleup$} if it is supported but not completely. \\nMore specifically, \\\\enquote{$\\\\bigtriangleup$} is assigned to P1 if the method can be flexible but requires additional effort, to P5 if the method can combat only moderate label noise, {and to P6 if the method does not make a strict assumption about the noise type but without explicitly modeling instance-dependent noise. Thus, for P6, the method marked with \\\\enquote{\\\\xmark} only deals with the instance-independent noise, while the method marked with \\\\enquote{\\\\cmark} deals with both instance-independent and -dependent noises.} The remaining properties\\\\,(i.e., P2, P3, and P4) are only assigned \\\\enquote{\\\\cmark} or \\\\enquote{\\\\xmark}. Regarding the implementation, we assign \\\\enquote{N/A} if a publicly available source code is not available.\\nNo existing method supports all the properties. Each method achieves noise robustness by supporting a different combination of the properties. The supported properties are similar among the methods of the same (sub-)category because those methods share the same methodological philosophy; however, they differ significantly depending on the (sub-)category. Therefore, we investigate the properties generally supported in each (sub-)category and summarize them in Table \\\\ref{table:direction_comparison}. Here, the property of a (sub-)category is marked as the majority of the belonging methods. If no clear trend is observed among those methods, then the property is marked \\\\enquote{$\\\\bigtriangleup$}.\",\n \"id\": \"3b9c1099-a2f2-417d-a753-e663d7ce62fa\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Methodological Comparison\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Methodological Comparison\"\n },\n {\n \"cite_extract_rate\": 0.9,\n \"cites\": [\n 4130,\n 3340,\n 4126,\n 4135,\n 7777,\n 3453,\n 4123,\n 4160,\n 4152\n ],\n \"content\": \"\\\\label{sec:noise_rate}\\n{\\nThe estimation of a noise rate is an imperative part of utilizing robust methods for better practical use, especially with the approaches belonging to the loss adjustment and sample selection. The estimated noise rate is widely used to reweight examples for a robust classifier or to determine how many examples should be selected as clean ones . However, detailed analysis has yet to be performed properly, though many robust approaches highly rely on the accuracy of noise rate estimation. The noise rate can be estimated by exploiting the inferred noise transition matrix , the Gaussian mixture model , or the cross-validation .\\n}\\n\\\\begin{comment}\\n1. Yao, Y., Liu, T., Han, B., Gong, M., Deng, J., Niu, G., & Sugiyama, M.\\nDual T: Reducing estimation error for transition matrix in label-noise learning. Advances in Neural Information Processing Systems 33, pp. 7260-7271, 2020. \\n-> noise transition matrix\\n2. Xia, X., Liu, T., Wang, N., Han, B., Gong, C., Niu, G., & Sugiyama, M.\\nAre anchor points really indispensable in label-noise learning? Advances in Neural Information Processing Systems 32, pp. 6835-6846, 2019.\\n-> noise transition matrix\\n3. Li, X., Liu, T., Han, B., Niu, G., & Sugiyama, M.\\nProvably end-to-end label-noise learning without anchor points.\\n-> noise transition matrix\\nProceedings of 38th International Conference on Machine Learning (ICML2021), pp. 6403-6413, 2021.\\n4. Patrini, Giorgio et al. “Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach.” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017): 2233-2241.\\n-> noise transition matrix\\n5. Liu, Tongliang, and Dacheng Tao. \\\"Classification with noisy labels by importance reweighting.\\\" IEEE Transactions on pattern analysis and machine intelligence 38.3 (2015): 447-461.\\n6. Scott, Clayton, Gilles Blanchard, and Gregory Handy. \\\"Classification with asymmetric label noise: Consistency and maximal denoising.\\\" Conference on learning theory. PMLR, 2013.\\n\\\\end{comment}\\n\\\\vspace*{-0.15cm}\",\n \"id\": \"8d7ed0bf-5e07-45ed-a147-190ba1a209fb\",\n \"level\": \"section\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Noise Rate Estimation\"\n ]\n ],\n \"subsections\": [\n \"841668d4-4892-46d7-ae08-7e210e7c2add\",\n \"39b0234a-04ab-44e2-98d0-fa20ed4e0f74\",\n \"8516d834-e901-490a-aa9d-864c2f267139\",\n \"685f1250-3a8b-48b6-b510-042be23d64f6\"\n ],\n \"title\": \"Noise Rate Estimation\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7777,\n 4238,\n 4145,\n 4161,\n 652,\n 3453,\n 7769,\n 4160,\n 4152\n ],\n \"content\": \"{\\nThe noise transition matrix has been used to build a statistically consistent robust classifier because it represents the class posterior probabilities for noisy and clean data, as in Eq.\\\\,\\\\eqref{eq:label_transition_matrix}. The first method to estimate the noise rate is exploiting this noise transition matrix, which can be inferred or trained accurately by using perfectly clean examples, i.e., \\\\emph{anchor points} ; an example $x$ with its label $i$ is defined as an anchor point if $p(y=i|x)=1$ and $p(y=k|x)=0$ for $k \\\\neq i$. Thus, let $\\\\mathcal{A}_i$ be the set of anchor points with label $i$, then the element of the noise transition matrix $T_{ij}$ is estimated by \\\\looseness=-1\\n\\\\begin{equation}\\n\\\\begin{split}\\n\\\\hat{T}_{ij} &= \\\\frac{1}{|\\\\mathcal{A}_{i}|}\\\\sum_{x\\\\in\\\\mathcal{A}_{i}} \\\\sum_{k=1}^{c} p(\\\\tilde{y}=j|y=k)p(y=k|x) \\\\\\\\\\n&= \\\\frac{1}{|\\\\mathcal{A}_{i}|}\\\\sum_{x\\\\in\\\\mathcal{A}_{i}} p(\\\\tilde{y}=j|x; \\\\Theta), \\n\\\\end{split}\\n\\\\end{equation}\\nwhere $p(\\\\tilde{y}=j|x; \\\\Theta)$ is the noisy class posterior probability of the classifier trained on noisy training data for the anchor point $x$ (see the detailed proof in ). Next, based on the inferred noise transition matrix, the noise rate of a balanced training data is obtained by averaging the label transition probabilities between classes,\\n\\\\begin{equation}\\n\\\\hat{\\\\tau} = \\\\frac{1}{c} \\\\sum_{i=1}^{c} \\\\sum_{j\\\\neq i}^{c} p(\\\\tilde{y}=j | {y}=i) = \\\\frac{1}{c} \\\\sum_{i=1}^{c} \\\\sum_{j\\\\neq i}^{c} \\\\hat{T}_{ij}.\\n\\\\end{equation}\\nHowever, since the anchor points are typically unknown in real-world data, they are identified from noisy training data by either theoretical derivations or heuristics . \\nIn addition, there have been recent efforts to learn the noise transition matrix without anchor points. \\\\emph{T-Revision} initializes a transition matrix by exploiting the examples with high noisy class posterior probabilities and then refines the matrix by adding a slack variable. \\\\emph{Dual-T} introduces an intermediate class that factorizes the transition matrix into two easy-to-estimate matrices for better accuracy. \\\\emph{VolMinNet} realizes an end-to-end framework and relaxes the need for anchor points under the sufficiently scattered assumption.\\n}\\n{\\n\\\\newcolumntype{L}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}m{#1}}\\n\\\\newcolumntype{X}[1]{>{\\\\centering\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{0pt}}p{#1}}\\n\\\\newcolumntype{Y}[1]{>{\\\\let\\\\newline\\\\\\\\\\\\arraybackslash\\\\hspace{1pt}}m{#1}}\\n\\\\begin{table*}[t!]\\n\\\\small\\n\\\\caption{Summary of publicly available datasets used for studying label noise.}\\n\\\\vspace*{-0.7cm}\\n\\\\begin{center}\\n\\\\begin{tabular}{L{2.0cm} | Y{3.4cm} |X{1.9cm}| X{1.9cm} |X{1.9cm} |X{1.9cm} |Y{2.1cm}}\\\\toprule \\n\\\\multicolumn{2}{c|}{\\\\textbf{Dataset}} & \\\\textbf{\\\\# Training} & \\\\textbf{\\\\# Validation} & \\\\textbf{\\\\# Testing} & \\\\textbf{\\\\# Classes} & \\\\textbf{Noise Rate\\\\,(\\\\%)} \\\\\\\\\\\\hline\\n\\\\multirow{7}{*}{\\\\hspace*{0.13cm}\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!{\\\\makecell[l]{Clean Data}}\\\\!\\\\!\\\\!\\\\!\\\\!} \\n& {MNIST} \\\\footnote[50] & {60K}& N/A & {10K} & $10$ & $\\\\approx0.0$ \\\\\\\\\\\\cline{2-7}\\n& {Fashion-MNIST} \\\\footnote[51]\\\\!\\\\!\\\\!\\\\!\\\\! & {60K}& N/A & {10K} & $10$ & $\\\\approx0.0$ \\\\\\\\\\\\cline{2-7}\\n& {CIFAR-10} \\\\footnote[52] & {50K}& N/A & {10K} & $10$ & $\\\\approx0.0$ \\\\\\\\\\\\cline{2-7}\\n& {CIFAR-100} \\\\footnote[52] & {50K}& N/A & {10K} & $100$ & $\\\\approx0.0$ \\\\\\\\\\\\cline{2-7}\\n& {SVHN} \\\\footnote[53] & {73K}& N/A & {26K} & $10$ & $\\\\approx0.0$ \\\\\\\\\\\\cline{2-7}\\n& {Tiny-ImageNet} \\\\footnote[55] & {100K}& {10K} & {10K} & $200$ & $\\\\approx0.0$ \\\\\\\\\\\\cline{2-7}\\n& {ImageNet} \\\\footnote[54] & {1.3M}& {50K} & {50K} & $1000$ & $\\\\approx0.0$ \\\\\\\\\\\\midrule\\n\\\\multirow{6}{*}{\\\\hspace*{0.13cm}\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!{\\\\makecell[l]{Real-world\\\\\\\\\\\\hspace*{-0.02cm}Noisy Data}}\\\\!\\\\!\\\\!\\\\!\\\\!} \\n& {ANIMAL-10N} \\\\footnote[56] & {50K}& N/A & {5K} & $10$ & $\\\\approx8.0$ \\\\\\\\\\\\cline{2-7}\\n& {CIFAR-10N }\\\\footnote[57] & {50K}& N/A & {10K} & $10$ & $\\\\approx9.0/18.0/40.2$\\\\!\\\\!\\\\!\\\\!\\\\!\\\\!\\\\! \\\\\\\\\\\\cline{2-7}\\n& {CIFAR-100N }\\\\footnote[57] & {50K}& N/A & {10K} & $100$ & $\\\\approx 25.6/40.2$ \\\\\\\\\\\\cline{2-7}\\n& {Food-101N} \\\\footnote[58] & {310K}& {5K} & {25K} & $101$ & $\\\\approx18.4$ \\\\\\\\\\\\cline{2-7}\\n& {Clothing1M} \\\\footnote[59] & {1M}& {14K} & {10K} & $14$ & $\\\\approx38.5$ \\\\\\\\\\\\cline{2-7}\\n& {WebVision} \\\\footnote[60] & {2.4M}& {50K} & {50K} & $1000$ & $\\\\approx20.0$ \\\\\\\\\\\\bottomrule\\n\\\\end{tabular}\\n\\\\end{center}\\n\\\\label{table:summary_dataset}\\n\\\\vspace*{-0.6cm}\\n\\\\end{table*}\\n}\\n\\\\vspace*{-0.15cm}\",\n \"id\": \"841668d4-4892-46d7-ae08-7e210e7c2add\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"8d7ed0bf-5e07-45ed-a147-190ba1a209fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Noise Rate Estimation\"\n ],\n [\n \"subsection\",\n \"Noise Transition Matrix\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Noise Transition Matrix\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4135\n ],\n \"content\": \"\\\\label{sec:gmm}\\n\\\\begin{figure}[t!]\\n\\\\begin{center}\\n\\\\vspace*{0.1cm}\\n\\\\includegraphics[width=8.8cm]{figures/evaluation/gmm_analysis.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.15cm}\\n\\\\hspace*{1.3cm} \\\\small{(a) Symmetric Noise.} \\\\hspace*{1.2cm} \\\\small{(b) Asymmetric Noise.} \\n\\\\vspace*{-0.15cm}\\n\\\\caption{Training loss distributions of true-labeled and false-labeled examples using the ground-truth label and the GMM on CIFAR-100 data with two synthetic noises of $40\\\\%$.}\\n\\\\label{fig:loss_gmm}\\n\\\\vspace*{-0.4cm}\\n\\\\end{figure}\\nThe second method is exploiting a one-dimensional and two-component GMM to model the loss distribution of true-labeled and false-labeled examples . As shown in Figure \\\\ref{fig:loss_gmm}, since the loss distribution tends to be \\\\emph{bi-modal}, the two Gaussian components are fitted to the training loss by using the EM algorithm; the probability of an example being a false-labeled one is obtained through its posterior probability. Hence, the noise rate is estimated at each epoch $t$ by computing the expectation of the posterior probability for all training examples, \\\\looseness=-1\\n\\\\begin{equation}\\n\\\\label{eq:naive_gmm}\\n\\\\begin{gathered}\\n\\\\hat{\\\\tau} = \\\\mathbb{E}_{(x,\\\\tilde{y})\\\\in\\\\tilde{\\\\mathcal{D}}}\\\\Big[\\\\,p\\\\big(g\\\\,|\\\\,\\\\ell\\\\big(f(x;\\\\Theta_{t}), \\\\tilde{y}\\\\big)\\\\big)\\\\,\\\\Big],\\\\\\\\ \\n\\\\end{gathered}\\n\\\\end{equation}\\nwhere $g$ is the Gaussian component with a larger loss. However, Pleiss et al. recently pointed out that the training loss becomes less separable by the GMM as the training progresses, and thus proposed the \\\\emph{area under the loss}\\\\,(AUL) curve, which is the sum of the example's training losses obtained from all previous training epochs. Even after the loss signal decays in later epochs, the distributions remain separable. Therefore, the noise rate is finally estimated by\\n\\\\begin{equation}\\n\\\\label{eq:aul_gmm}\\n\\\\begin{gathered}\\n\\\\hat{\\\\tau} = \\\\mathbb{E}_{(x,\\\\tilde{y})\\\\in\\\\tilde{\\\\mathcal{D}}}\\\\Big[\\\\,p\\\\big(g\\\\,|\\\\,{\\\\rm AUL}_{t}(x,\\\\tilde{y})\\\\big)\\\\,\\\\Big],\\\\\\\\ \\n\\\\text{where}~ \\\\text{AUL}_{t}(x,\\\\tilde{y})= \\\\sum_{i=1}^{t}\\\\ell\\\\big(f(x;\\\\Theta_t),\\\\tilde{y}\\\\big).\\n\\\\end{gathered}\\n\\\\end{equation}\\n\\\\vspace*{-0.4cm}\",\n \"id\": \"39b0234a-04ab-44e2-98d0-fa20ed4e0f74\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"8d7ed0bf-5e07-45ed-a147-190ba1a209fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Noise Rate Estimation\"\n ],\n [\n \"subsection\",\n \"Gaussian Mixture Model\\\\,(GMM)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Gaussian Mixture Model\\\\,(GMM)\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 3340,\n 4140,\n 4126\n ],\n \"content\": \"\\\\label{sec:cross_val}\\nThe third method is estimating the noise rate by applying cross validation, which typically requires clean validation data . However, such clean validation data is hard to acquire in real-world applications. Thus, Chen et al. leveraged two randomly divided noisy training datasets for cross validation. Under the assumption that the two datasets share exactly the same noise transition matrix, the noise rate quantifies the test accuracy of DNNs that are respectively trained and tested on the two divided sets,\\n\\\\begin{equation}\\n\\\\label{eq:cross_val}\\n\\\\!\\\\!\\\\!{\\\\rm Test\\\\,Accuracy} \\\\!= \\\\!\\\\!\\\\!\\\\\\n\\\\begin{cases}\\n(1 - \\\\hat{\\\\tau})^{2} + \\\\hat{\\\\tau}^{2} / (c - 1) \\\\!\\\\!&\\\\!\\\\! \\\\text{if symmetric}\\\\!\\\\!\\\\!\\\\!\\\\\\\\\\n(1 - \\\\hat{\\\\tau})^{2} + \\\\hat{\\\\tau}^{2} \\\\!\\\\!& \\\\!\\\\!\\\\text{if asymmetric}.\\\\!\\\\!\\\\!\\\\!\\n\\\\end{cases}\\n\\\\end{equation}\\nTherefore, the noise rate is estimated from the test accuracy obtained by cross validation.\\n\\\\begin{comment}\",\n \"id\": \"8516d834-e901-490a-aa9d-864c2f267139\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"8d7ed0bf-5e07-45ed-a147-190ba1a209fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Noise Rate Estimation\"\n ],\n [\n \"subsection\",\n \"Cross Validation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Cross Validation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:nr_comp}\\nTo compare the estimation performance of using the GMM and cross validation, we trained a WideResNet-16-8 for three benchmark datasets with varying noise rates. The results are plotted in Figure \\\\ref{fig:noise_rate_symmetric}. Generally, both methods performed well on the easy dataset\\\\,(i.e., CIFAR-10), but their performance worsened as the training difficulty increased from CIFAR-10 to Tiny-ImageNet because the true-labeled but hard examples are not clearly distinguishable from the false-labeled ones. Nevertheless, the GMM method showed considerably better performance than the cross validation method even in the two difficult datasets, CIFAR-100 and Tiny-ImageNet. \\nOverall, this empirical analysis will be helpful for practitioners or researchers who design robust algorithms for noisy labels.\\n\\\\begin{figure}[!t]\\n\\\\vspace*{0.0cm}\\n\\\\begin{minipage}[t]{1.0\\\\linewidth}\\n\\\\begin{center}\\n\\\\includegraphics[width=8.8cm]{figures/evaluation/noise_rate_symmetric.pdf}\\n\\\\end{center}\\n\\\\vspace*{-0.15cm}\\n\\\\hspace*{2.1cm} \\\\small{(a) GMM.} \\\\hspace*{2.0cm} \\\\small{(b) Cross Validation.}\\n\\\\vspace*{-0.2cm}\\n\\\\caption{Noise rate estimation using the Gaussian mixture model and cross validation when training a WideResNet-16-8 on three datasets with symmetric noise.}\\n\\\\label{fig:noise_rate_symmetric}\\n\\\\end{minipage}\\n\\\\vspace*{-0.35cm}\\n\\\\end{figure}\\n\\\\end{comment}\\n\\\\vspace*{-0.15cm}\",\n \"id\": \"685f1250-3a8b-48b6-b510-042be23d64f6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"8d7ed0bf-5e07-45ed-a147-190ba1a209fb\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Noise Rate Estimation\"\n ],\n [\n \"subsection\",\n \"Comparison of Noise Rate Estimation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Comparison of Noise Rate Estimation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"This section describes the typically used experimental design for comparing robust training methods in the presence of label noise. We introduce publicly available image datasets and then describe widely-used evaluation metrics.\\n\\\\vspace*{-0.3cm}\",\n \"id\": \"0788ae52-30c4-4156-89f9-6be2e443db83\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Experimental Design\"\n ]\n ],\n \"subsections\": [\n \"b05d418b-8b51-4e39-b585-58d8b1bcd7ff\",\n \"6b4a1fff-c3c1-412a-b6f0-eeaa7d3721a4\"\n ],\n \"title\": \"Experimental Design\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"To validate the robustness of the proposed algorithms, an image classification task was widely conducted on numerous image benchmark datasets. Table \\\\ref{table:summary_dataset} summarizes popularly-used public benchmark datasets, which are classified into two categories: \\\\emph{1)} a \\\\enquote{clean dataset} that consists of mostly true-labeled examples annotated by human experts and \\\\emph{2)} a \\\\enquote{real-world noisy dataset} that comprises real-world noisy examples with varying numbers of false labels.\\n\\\\vspace*{0.15cm}\",\n \"id\": \"b05d418b-8b51-4e39-b585-58d8b1bcd7ff\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"0788ae52-30c4-4156-89f9-6be2e443db83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Experimental Design\"\n ],\n [\n \"subsection\",\n \"Publicly Available Datasets\"\n ]\n ],\n \"subsections\": [\n \"0c8efb50-e5a7-4707-a3b7-dcff2ec802f5\",\n \"2e98f74b-5745-4f37-aea0-711f64e9e6c5\"\n ],\n \"title\": \"Publicly Available Datasets\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 4253,\n 8630,\n 652\n ],\n \"content\": \"According to the literature , \\\\emph{seven} clean datasets are widely used: MNIST\\\\footnote[50]{\\\\url{http://yann.lecun.com/exdb/mnist}}, classification of handwritten digits ; Fashion-MNIST\\\\footnote[51]{\\\\url{https://github.com/zalandoresearch/fashion-mnist}}, classification of various clothing ; CIFAR-10\\\\footnote[52]{\\\\url{https://www.cs.toronto.edu/~kriz/cifar.html}} and CIFAR-100\\\\footnotemark[52], classification of a subset of $80$ million categorical images ; SVHN\\\\footnote[53]{\\\\url{http://ufldl.stanford.edu/housenumbers}}, classification of house numbers in Google Street view images ; ImageNet\\\\footnote[54]{\\\\url{http://www.image-net.org}} and Tiny-ImageNet\\\\footnote[55]{\\\\url{https://www.kaggle.com/c/tiny-imagenet}}, image database organized according to the WordNet hierarchy and its small subset . Because the labels in these datasets are almost all true-labeled, their labels in the training data should be artificially corrupted for the evaluation of synthetic noises, namely \\\\emph{symmetric} noise and \\\\emph{asymmetric} noise. \\n\\\\vspace*{0.15cm}\",\n \"id\": \"0c8efb50-e5a7-4707-a3b7-dcff2ec802f5\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"b05d418b-8b51-4e39-b585-58d8b1bcd7ff\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Experimental Design\"\n ],\n [\n \"subsection\",\n \"Publicly Available Datasets\"\n ],\n [\n \"subsubsection\",\n \"Clean Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Clean Datasets\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 4238,\n 4161,\n 7769\n ],\n \"content\": \"Unlike the clean datasets, real-world noisy datasets inherently contain many mislabeled examples annotated by non-experts. According to the literature , \\\\emph{six} real-world noisy datasets are widely used: ANIMAL-10N\\\\footnote[56]{\\\\url{https://dm.kaist.ac.kr/datasets/animal-10n}}, real-world noisy data of human-labeled online images for 10 confusing animals ;\\n{CIFAR-10N\\\\footnote[57]{\\\\url{http://noisylabels.com/}} and CIFAR-100N\\\\footnotemark[57], variations of CIFAR-10 and CIFAR-100 with human-annotated real-world noisy labels collected from Amazon’s Mechanical Turk . They provide human labels with different noise rates, as shown in Table \\\\ref{table:summary_dataset};}\\nFood-101N\\\\footnote[58]{\\\\url{https://kuanghuei.github.io/Food-101N}}, real-world noisy data of crawled food images annotated by their search keywords in the Food-101 taxonomy ; Clothing1M\\\\footnote[59]{\\\\url{https://www.floydhub.com/lukasmyth/datasets/clothing1m}}, real-world noisy data of large-scale crawled clothing images from several online shopping websites ; WebVision\\\\footnote[60]{\\\\url{https://data.vision.ee.ethz.ch/cvl/webvision/download.html}}, real-world noisy data of large-scale web images crawled from Flickr and Google Images search . To support sophisticated evaluation, most real-world noisy datasets contain their own clean validation set and provide the estimated noise rate of their training set. \\\\looseness=-1\",\n \"id\": \"2e98f74b-5745-4f37-aea0-711f64e9e6c5\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"b05d418b-8b51-4e39-b585-58d8b1bcd7ff\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Experimental Design\"\n ],\n [\n \"subsection\",\n \"Publicly Available Datasets\"\n ],\n [\n \"subsubsection\",\n \"Real-world Noisy Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Real-world Noisy Datasets\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 3340,\n 4126,\n 3630\n ],\n \"content\": \"A typical metric to assess the robustness of a particular method is the prediction accuracy for unbiased and clean examples that are not used in training. The prediction accuracy degrades significantly if the DNN overfits to false-labeled examples . Hence, \\\\emph{test accuracy} has generally been adopted for evaluation . For a test set $\\\\mathcal{T}=\\\\{(x_i,y_i)\\\\}_{i=1}^{|\\\\mathcal{T}|}$, let $\\\\hat{y}_i$ be the predicted label of the $i$-th example in $\\\\mathcal{T}$. Subsequently, the test accuracy is formalized by\\n\\\\begin{equation}\\n\\\\label{eq:test_accuracy}\\n\\\\text{Test Accuracy} = \\\\frac{|\\\\{(x_i,y_i)\\\\in\\\\mathcal{T}:\\\\hat{y}_i=y_i\\\\}|}{|\\\\mathcal{T}|}.\\n\\\\end{equation}\\nIf the test data are not available, \\\\emph{validation accuracy} can be used by replacing $\\\\mathcal{T}$ in Eq.\\\\,\\\\eqref{eq:test_accuracy} with validation data $\\\\mathcal{V}=\\\\{(x_i,y_i)\\\\}_{i=1}^{|\\\\mathcal{V}|}$ as an alternative,\\n\\\\begin{equation}\\n\\\\label{eq:validation_accuracy}\\n\\\\text{Validation Accuracy} = \\\\frac{|\\\\{(x_i,y_i)\\\\in\\\\mathcal{V}:\\\\hat{y}_i=y_i\\\\}|}{|\\\\mathcal{V}|}.\\n\\\\end{equation}\\nFurthermore, if the specified method belongs to the \\\\enquote{sample selection} category, \\\\emph{label precision} and \\\\emph{label recall} can be used as the metrics,\\n\\\\noindent\\n\\\\begin{equation}\\n\\\\label{eq:label_precision}\\n\\\\begin{gathered}\\n\\\\text{Label Precision} = \\\\frac{|\\\\{(x_i,\\\\tilde{y}_i)\\\\in\\\\mathcal{S}_t: \\\\tilde{y}_i = y_i\\\\}|}{|\\\\mathcal{S}_t|},\\\\\\\\\\n\\\\text{Label Recall} = \\\\frac{|\\\\{(x_i,\\\\tilde{y}_i)\\\\in\\\\mathcal{S}_t: \\\\tilde{y}_i = y_i\\\\}|}{|\\\\{(x_i,\\\\tilde{y}_i)\\\\in\\\\mathcal{B}_t: \\\\tilde{y}_i = y_i\\\\}|},\\n\\\\end{gathered}\\n\\\\end{equation}\\nwhere $\\\\mathcal{S}_t$ is the set of selected clean examples in a mini-batch $\\\\mathcal{B}_t$. The two metrics are performance indicators for the examples selected from the mini-batch as true-labeled ones . \\\\looseness=-1\\nMeanwhile, if the specified method belongs to the \\\\enquote{label refurbishment} category, \\\\emph{correction error} can be used as an indicator of how many examples are incorrectly refurbished,\\n\\\\begin{equation}\\n\\\\label{eq:correction_error}\\n\\\\!\\\\text{Correction Error} = \\\\frac{|\\\\{x_i\\\\!\\\\in\\\\!\\\\mathcal{R}: \\\\text{argmax}(y_{i}^{refurb}) \\\\neq y_i\\\\}|}{|\\\\mathcal{R}|},\\n\\\\end{equation}\\nwhere $\\\\mathcal{R}$ is the set of examples whose labels are refurbished by Eq.\\\\,(\\\\ref{eq:label_correction}) and $y_i^{refurb}$ is the refurbished label of the $i$-th examples in $\\\\mathcal{R}$. \\n\\\\vspace*{-0.12cm}\",\n \"id\": \"6b4a1fff-c3c1-412a-b6f0-eeaa7d3721a4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"0788ae52-30c4-4156-89f9-6be2e443db83\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Experimental Design\"\n ],\n [\n \"subsection\",\n \"Evaluation Metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Evaluation Metrics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"With recent efforts in the machine learning community, the robustness of DNNs becomes evolving in several directions. Thus, the existing approaches covered in our survey face a variety of future challenges. This section provides discussion for future research that can facilitate and envision the development of deep learning in the label noise area. \\\\looseness=-1\\n\\\\vspace*{-0.22cm}\",\n \"id\": \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ]\n ],\n \"subsections\": [\n \"4ad866ea-f321-4ab2-be85-14135292873f\",\n \"e53153d2-a39d-4573-9711-598781d00529\",\n \"cfbd0ac1-5575-43b8-9a59-26d13838ff8e\",\n \"23870233-7f5b-4ef1-9d38-e6f3adb0fa48\",\n \"bf2c4a41-a623-4acf-b905-fe345054727a\",\n \"a23d6d53-b5ad-4828-8bb9-730c9fff2012\"\n ],\n \"title\": \"Future Research Directions\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 7777,\n 4120,\n 4124,\n 4155,\n 4136,\n 4152\n ],\n \"content\": \"}\\n{\\nExisting theoretical and empirical studies for \\\\emph{robust loss function} and \\\\emph{loss correction} are largely built upon the instance-independent noise assumption that the label noise is independent of input features . However, this assumption may not be a good approximation of the real-world label noise. In particular, Chen et al. conducted a theoretical hypothesis testing\\\\footnote[61]{In Clothing1M, the result showed that the instance-independent noise happens with probability lower than $10^{-21250}$, which is statistically impossible.} using a popular real-world dataset, Clothing1M, and proved that its label noise is statistically different from the instance-independent noise. This testing confirms that the label noise should depend on the instance. \\\\looseness=-1\\nConversely, most methods for the other direction (especially, \\\\emph{sample selection}) work well even under the instance-dependent label noise in general since they do not rely on the assumption. Nevertheless, Song et al. pointed out that their performance could considerably worsen in the instance-dependent\\\\,(or real-world) noise compared to symmetric noise due to the confusion between true-labeled and false-labeled examples. The loss distribution of true-labeled examples heavily overlaps that of false-labeled samples in the asymmetric noise, which is similar to the real-world noise, in Figure \\\\ref{fig:loss_distribution}(b). Thus, identifying clean examples becomes more challenging when dealing with the instance-dependent label noise.\\nBeyond the instance-independent label noise, there have been a few recent studies for the instance-dependent label noise. Mostly, they only focus on a binary classification task or a restricted small-scale machine learning model such as logistic regression . Therefore, learning with the instance-dependent label noise is an important topic that deserves more research attention.} \\n\\\\vspace*{-0.32cm}\",\n \"id\": \"4ad866ea-f321-4ab2-be85-14135292873f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"{Instance-dependent Label Noise\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Instance-dependent Label Noise\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 4162\n ],\n \"content\": \"}\\n{\\nMost of the existing methods are applicable only for a \\\\emph{single-label} multi-class classification problem, where each data example is assumed to have only one true label. However, in the case of \\\\emph{multi-label} learning, each data example can be associated with a set of multiple true class labels. In music categorization, each music can belong to multiple categories . In semantic scene classification, each scene may belong to multiple scene classes . Thus, contrary to the single-label setup, the multi-label classifier aims to predict a set of target objects simultaneously. \\nIn this setup, a multi-label dataset of millions of examples are reported to contain over $26.6\\\\%$ false-positive labels as well as a significant number of omitted labels . \\nEven worse, the difference in occurrence between classes makes this problem more challenging; some minor class labels occur less in training data than other major class labels. Considering such aspects that can arise in multi-label classification, the simple extension of existing methods may not learn the proper correlations among multiple labels. Therefore, learning from noisy labels with multi-label data is another important topic for future research. We refer the readers to a recent study that discusses the evaluation of multi-label classifiers trained with noisy labels.\\n}\\n\\\\vspace*{-0.12cm}\",\n \"id\": \"e53153d2-a39d-4573-9711-598781d00529\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"{Multi-label Data with Label Noise\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Multi-label Data with Label Noise\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7774\n ],\n \"content\": \"}\\n{\\nThe \\\\emph{class imbalance} in training data is commonly observed, where a few classes account for most of the data. Especially when working with large data in many real-world applications, this problem becomes more severe and is often associated with the problem of noisy labels .\\nNevertheless, to ease the label noise problem, it is commonly assumed that training examples are equally distributed over all class labels in the training data. This assumption is quite strong when collecting large-scale data, and thus we need to consider a more realistic scenario in which the two problems coexist. \\\\looseness=-1\\nMost of the existing robust methods may not work well with the class imbalance, especially when they rely on the learning dynamics of DNNs, e.g., the small-loss trick or memorization effect. Under the existence of the class imbalance, the training model converges to major classes faster than minor classes such that most examples in the major class exhibit small losses\\\\,(i.e., early memorization). That is, there is a risk of discarding most examples in the minor class. Furthermore, in terms of example importance, high-loss examples are commonly favored for the class imbalance problem , while small-loss examples are favored for the label noise problem. This conceptual contradiction hinders the applicability of the existing methods that neglect the class imbalance. Therefore, these two problems should be considered simultaneously to deal with more general situations.}\\n\\\\vspace*{-0.12cm}\",\n \"id\": \"cfbd0ac1-5575-43b8-9a59-26d13838ff8e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"{Class Imbalance Data with Label Noise\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Class Imbalance Data with Label Noise\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 7771,\n 4163,\n 3899,\n 8740,\n 8739\n ],\n \"content\": \"}\\n{\\nMachine learning classifiers can perpetuate and amplify the existing systemic injustices in society . Hence, fairness is becoming another important topic. Traditionally, robust training and fair training have been studied by separate communities; robust training with noisy labels has mostly focused on combating label noise without regarding data bias , whereas fair training has focused on dealing with data bias, not necessarily noise . However, noisy labels and data bias, in fact, coexist in real-world data. Satisfying both robustness and fairness is more realistic but challenging because the bias in data is pertinent to label noise. \\nIn general, many fairness criteria are group-based, where a target metric is equalized or enforced over subpopulations in the data, also known as \\\\emph{protected groups} such as race or gender . Accordingly, the goal of fair training is building a model that satisfies such fairness criteria for the \\\\emph{true} protected groups. However, if the \\\\emph {noisy} protection group is involved, such fairness criteria cannot be directly applied. Recently, mostly after 2020, a few pioneering studies have emerged to consider both robustness and fairness objectives at the same time under the binary classification setting . Therefore, more research attention is needed for the convergence of robust training and fair training.\\n}\\n\\\\vspace*{-0.12cm}\",\n \"id\": \"23870233-7f5b-4ef1-9d38-e6f3adb0fa48\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"{Robust and Fair Training\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Robust and Fair Training\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 9127,\n 4164\n ],\n \"content\": \"}\\n{\\nThere has been a lot of research on the robustness of deep learning under input perturbation, mainly in the field of adversarial training where the {input feature} is maliciously perturbed to distort the output of the DNN . Although learning with noisy labels and learning with noisy inputs have been regarded as separate research fields, their goals are similar in that they learn noise-robust representations from noisy data. Based on this common point of view, a few recent studies have investigated the interaction of adversarial training with noisy labels .\\nInterestingly, it was turned out that adversarial training makes DNNs robust to label noise . Based on this finding, Damodaran et al. proposed a new regularization term, called Wasserstein adversarial regularization, to address the problem of learning with noisy labels. Zhu et al. proposed to use the number of projected gradient descent steps as a new criterion for sample selection such that clean examples are filtered out from noisy data. These approaches are regarded as a new perspective on label noise compared to traditional work. Therefore, understanding the connection between input perturbation and label noise could be another future topic for better representation learning toward robustness. \\n}\\n\\\\vspace*{-0.12cm}\",\n \"id\": \"bf2c4a41-a623-4acf-b905-fe345054727a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"{Connection with Input Perturbation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Connection with Input Perturbation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\n{\\n{The efficiency of the learning pipeline is another important aspect to design deep learning approaches.} However, for robust deep learning, most studies have neglected the efficiency of the algorithm because their main goal is to improve the robustness to label noise. For example, maintaining multiple DNNs or training a DNN in multiple rounds is frequently used, but these approaches significantly degrade the efficiency of the learning pipeline. On ther other hand, the need for more efficient algorithms is increasing owing to the rapid increase in the amount of available data . \\nAccording to our literature survey, most work did not even report the efficiency\\\\,(or time complexity) of their approaches. However, it is evident that saving the training time is helpful under the restricted budget for computation.\\nTherefore, enhancing the efficiency will significantly increase the usability of robust deep learning in the big data era.\\n}\",\n \"id\": \"a23d6d53-b5ad-4828-8bb9-730c9fff2012\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"01d4e82b-d227-4b85-bfa5-6db6af2b743d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"{Efficient Learning Pipeline\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"{Efficient Learning Pipeline\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"DNNs easily overfit to false labels owing to their high capacity in totally memorizing all noisy training samples. This overfitting issue still remains even with various conventional regularization techniques, such as dropout and batch normalization, thereby significantly decreasing their generalization performance. Even worse, in real-world applications, the difficulty in labeling renders the overfitting issue more severe. Therefore, learning from noisy labels has recently become one of the most active research topics. \\nIn this survey, we presented a comprehensive understanding of modern deep learning methods to address the negative consequences of learning from noisy labels. All the methods were grouped into five categories according to their underlying strategies and described along with their methodological weaknesses. Furthermore, a systematic comparison was conducted using six popular properties used for evaluation in the recent literature. According to the comparison results, there is no ideal method that supports all the required properties; the supported properties varied depending on the category to which each method belonged. Several experimental guidelines were also discussed, including {noise rate estimation,} publicly available datasets, and evaluation metrics. Finally, we provided insights and directions for future research in this domain. \\n\\\\section*{Acknowledgements}\\nThis work was supported by Institute of Information \\\\& Communications Technology Planning \\\\& Evaluation\\\\,(IITP) grant funded by the Korea government\\\\,(MSIT) (No. 2020-0-00862, DB4DL: High-Usability and Performance In-Memory Distributed DBMS for Deep Learning).\\n\\\\end{document}\",\n \"id\": \"0fd993d8-9dac-43e0-a25d-3f85b6def415\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4c23e240-6b38-42a9-a3ad-77c4f677661a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Learning from Noisy Labels with Deep Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":90,"string":"90"},"all_cites":{"kind":"list like","value":[7,206,7210,4238,7770,71,3630,8734,4116,4115,7769,7771,8696,4118,313,4117,3454,4121,8735,4120,4122,4119,8736,3340,4126,4253,4124,4125,4123,4127,4129,4143,7133,4137,7773,3345,4134,4132,4140,4141,7775,3342,4142,7772,8737,4130,4138,7191,7162,4135,4144,4145,7774,4139,892,4128,4131,8630,4133,2277,4136,4146,7776,8738,4147,4148,4151,4150,4149,7777,4153,4152,3453,4156,4154,4155,4157,4158,1695,4159,3343,7163,7778,4160,4161,652,4162,4163,3899,8740,8739,9127,4164],"string":"[\n 7,\n 206,\n 7210,\n 4238,\n 7770,\n 71,\n 3630,\n 8734,\n 4116,\n 4115,\n 7769,\n 7771,\n 8696,\n 4118,\n 313,\n 4117,\n 3454,\n 4121,\n 8735,\n 4120,\n 4122,\n 4119,\n 8736,\n 3340,\n 4126,\n 4253,\n 4124,\n 4125,\n 4123,\n 4127,\n 4129,\n 4143,\n 7133,\n 4137,\n 7773,\n 3345,\n 4134,\n 4132,\n 4140,\n 4141,\n 7775,\n 3342,\n 4142,\n 7772,\n 8737,\n 4130,\n 4138,\n 7191,\n 7162,\n 4135,\n 4144,\n 4145,\n 7774,\n 4139,\n 892,\n 4128,\n 4131,\n 8630,\n 4133,\n 2277,\n 4136,\n 4146,\n 7776,\n 8738,\n 4147,\n 4148,\n 4151,\n 4150,\n 4149,\n 7777,\n 4153,\n 4152,\n 3453,\n 4156,\n 4154,\n 4155,\n 4157,\n 4158,\n 1695,\n 4159,\n 3343,\n 7163,\n 7778,\n 4160,\n 4161,\n 652,\n 4162,\n 4163,\n 3899,\n 8740,\n 8739,\n 9127,\n 4164\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.9143923895532031,"string":"0.914392"}}},{"rowIdx":57,"cells":{"authors":{"kind":"list like","value":["Tiago D. Perez","Samuel Pagliarini"],"string":"[\n \"Tiago D. Perez\",\n \"Samuel Pagliarini\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-06-08T14:24:49Z","string":"2020-06-08T14:24:49Z"},"category":{"kind":"string","value":"cs.CR"},"abstract":{"kind":"string","value":" In today's integrated circuit (IC) ecosystem, owning a foundry is not economically viable, and therefore most IC design houses are now working under a fabless business model. In order to overcome security concerns associated with the outsorcing of IC fabrication, the Split Manufacturing technique was proposed. In Split Manufacturing, the Front End of Line (FEOL) layers (transistors and lower metal layers) are fabricated at an untrusted high-end foundry, while the Back End of Line (BEOL) layers (higher metal layers) are manufactured at a trusted low-end foundry. This approach hides the BEOL connections from the untrusted foundry, thus preventing overproduction and piracy threats. However, many works demonstrate that BEOL connections can be derived by exploiting layout characteristics that are introduced by heuristics employed in typical floorplanning, placement, and routing algorithms. Since straightforward Split Manufacturing may not afford a desirable security level, many authors propose defense techniques to be used along with Split Manufacturing. In our survey, we present a detailed overview of the technique, the many types of attacks towards Split Manufacturing, as well as possible defense techniques described in the literature. For the attacks, we present a concise discussion on the different threat models and assumptions, while for the defenses we classify the studies into three categories: proximity perturbation, wire lifting, and layout obfuscation. The main outcome of our survey is to highlight the discrepancy between many studies -- some claim netlists can be reconstructed with near perfect precision, while others claim marginal success in retrieving BEOL connections. Finally, we also discuss future trends and challenges inherent to Split Manufacturing, including the fundamental difficulty of evaluating the efficiency of the technique. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4df5dfa6-5506-4133-aae0-08641ac96086","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"]],"subsections":["bcde8dcf-d020-417f-80e7-74fb7e142a07","c3cea202-e3a7-4e28-a205-e6374a147781","63582e79-19ab-44ed-8225-2fdf07c8ddcd","0381ac67-f861-41a5-94dd-d57b1d10cc30","5c7f1b9c-d130-400e-ae7c-f7e824d126c3","347259ba-3c12-43d7-b6a0-2402a507ae62"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:introduction}\nCounterfeiting and intellectual property (IP) infringement are growing problems in several industries, including the electronics sector. In Europe, for instance, seizures of counterfeit electronics products increased by almost 30\\% when comparing the 2014-2016 to the 2011-2013 period . Legitimate electronics companies reported about \\$100 billion in sales losses every year because of counterfeiting .\n \\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\linewidth]{counter_taxiderm.pdf}\n{Taxonomy of counterfeit electronics (adapted from ). \\label{fig:cf_tax}}\nAs electronic systems are being increasingly deployed in critical infrastructure, counterfeit and maliciously modified integrated circuits (ICs) have become a major concern. The globalized nature of the IC supply chain contributes to the problem as we lack the means to assess the trustworthiness of the design and fabrication of ICs. It is conceivable -- if not likely -- that a fault in a low-quality counterfeit IC (or even a maliciously modified IC) will effectively disrupt critical infrastructure with grave consequences. Therefore, hardware security has gained more attention in the past decades, emerging as a relevant research topic.\nAs the IC supply chain has become more globalized, ensuring the integrity and trustworthiness of ICs becomes more challenging . When a modern IC is conceived, the probability that all involved parties are trusted is, in practice, close to zero. The process of conceiving an IC can be broken down into three major steps: design, manufacturing, and validation. \\textit{Designing} an IC involves arranging blocks and their interconnections. Some blocks are in-house developed, while some are third-party IPs. Finally, a layout is generated by instantiating libraries that might also be in-house developed or provided by third parties. The resulting layout is then sent to a foundry for \\textit{manufacturing}. The process of \\textit{validation} requires test for physical defects as well as verification of packaged parts for correct functionality. Both test and packaging facilities may be untrusted, as these efforts are often offshored. Thus, in order to produce an IC, sensitive information almost inevitably is exposed to untrusted parties. Today's reality is that ICs are vulnerable to many hardware-based threats, including insertion of hardware trojans, IP piracy, IC overbuilding, reverse engineering, side-channel attacks, and counterfeiting. These threats are discussed in details in . \nHardware trojans, in particular, are malicious modifications to an IC, where attackers insert circuitry (or modify the existing logic) for their own malicious purposes. This type of attack is (typically) mounted during manufacturing, as the foundry holds the entire layout and can easily identify critical locations for trojan insertion. Third-party IPs can also contain trojans/backdoors that may contain hidden functionalities, and which can be used to access restricted parts of the design and/or expose data that would otherwise be unknown to the adversary.\nIP piracy and IC overbuilding are, essentially, illegal ownership claims of different degrees. As said before, during designing an IC, third-party and in-house developed IPs are utilized. The untrusted foundry (or a rogue employee of it) can copy one of those IPs without the owner's authorization. Similarly, malicious foundries can manufacture a surplus of ICs (overbuilding) without the owner's knowledge, and sell these parts in the grey market.\nReversing engineering of ICs has been extensively demonstrated in the specialized literature . An attacker can identify the technology node and underlying components (memory, analog, and standard cells), from which a gate-level netlist can be extracted and even a high-level abstraction can be inferred . Reverse engineering can be effortlessly executed during manufacturing, as the foundry holds the entire layout and most likely promptly recognizes some of the IP as well. After fabrication, -- when ICs are already packaged and deployed -- reverse engineering is more laborious but can still be executed by a knowledgeable adversary.\nAccording to , counterfeit components are classified into seven distinct categories, as illustrated in Figure~\\ref{fig:cf_tax}. Recycled, remarked, out-of-spec/defective, and forged documentation are intrinsic after-market problems, where products are offered by parties other than the original component manufacturer or authorized vendors. These cases are highlighted in red. On the other hand, overproducing, cloning, and tampering are problems faced during the designing and/or fabrication of ICs. These cases are highlighted in blue. For this reason, in this paper, we will focus on these threats. It is important to realize that these threats could be avoided if a trusted fabrication scheme was in place. However, the escalating cost and complexity of semiconductor manufacturing on advanced technologies made owning an advanced foundry unfeasible for design companies, which now have the tendency to adopt the fabless business model . Consequentially, outsourcing of the manufacturing exposes their entire layout to untrusted foundries, leaving their designs vulnerable to malicious attacks.\nWhile many \\textit{ad hoc} techniques have been proposed to individually combat these threats, very few solutions directly address the lack of trust in the fabrication process. Split Manufacturing stands out from other techniques as it promotes a hybrid solution between trusted and untrusted fabrication. The technique was first pitched to DARPA circa 2006 in a white paper authored by Carnegie Mellon and Stanford universities. Later, it was picked by IARPA which then launched the Trusted IC program that successfully stewarded much of the research in the area and led to this survey. \nIn Split Manufacturing, the key concept is to \\textit{split} the circuit in two distinct parts before manufacturing, one containing the transistors and some routing wires, and the other containing the remaining routing wires. These parts are then fabricated in different foundries. The anatomy of an IC is illustrated in Figure~\\ref{fig:ic_anatomy}, containing two set of layers, the bottom layer where the transistors are built, called Front end of the Line (FEOL), and the top layer where the metal layers are built for routing purposes, called Back end of Line (BEOL). The metal layers are referred as MX, where X is the level of the layer. M1 is the lowest layer at the bottom of the stack. Connections between metals are made by vias referred as VX, following the same naming scheme for metals. In Split Manufacturing, the FEOL is first manufactured in a high-end foundry, and later the BEOL is stacked on top of it by a second (and possibly low-end) foundry. This process requires electrical, mechanical, and/or optical alignment techniques to ensure the connections between them. Additionally, FEOL and BEOL technologies have to be compliant with each other regarding the rules for metal/via dimensions where the split is to be performed. The split can be performed at the lowest metal layer (M1) or at higher layers, for which trade-offs are established between attained security and overheads. If the BEOL and FEOL technologies are vastly different from one another, Split Manufacturing may incur heavy overheads.\nIn this work, the focus is on the Split Manufacturing technique. As described above, Split Manufacturing can tackle threats that occur during the fabrication. It avoids overproduction, reverse engineering (to some extent) and unwanted modifications, limiting the capability of attackers. In Section II, we provide a background and more in-depth explanation of the technique. We address security threats in Section \\ref{sec:bck_gnd}, demonstrating the potential vulnerabilities found in split circuits and describing the state-of-the-art attacks proposed until the present day. In Section \\ref{sec:defenses}, the security of split circuits is discussed, showing how it can be improved using enhancement techniques. Future trends and lessons learnt are discussed in Section \\ref{sec:trends}. Finally, our conclusions are presented in Section \\ref{sec:conclusion}.\n \\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=.9\\linewidth]{cross_section.pdf}\n{Anatomy of an integrated circuit (adapted from ). \\label{fig:ic_anatomy}}","id":"bcde8dcf-d020-417f-80e7-74fb7e142a07","level":"section","origin_cites_number":10,"parent_id":"4df5dfa6-5506-4133-aae0-08641ac96086","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:bck_gnd}\nAs mentioned before, in order to have access to advanced technologies, many design companies have to outsource their IC manufacturing to untrusted high-end foundries. Protecting their designs against threats that may occur during manufacturing is a concern. Designs can be protected by applying the Split Manufacturing technique, thus combating all threats highlighted in blue in Figure~\\ref{fig:cf_tax}.\nSplit Manufacturing protects a design by hiding sensitive data from the untrusted foundry. This is achieved by splitting the IC into two parts before manufacturing, a horizontal cut that breaks the circuit into one part containing the transistors and some (local) routing wires, and another containing only routing wires. These parts are termed FEOL and BEOL. \nAs the FEOL and the BEOL of an IC are built sequentially, first FEOL and then BEOL, this characteristic enables the Split Manufacturing technique. Since the FEOL contains the transistors and possibly a few of the lowest ultra-thin metal layers -- the most complex parts of an CMOS process --, it is logical to seek to use a high-end foundry for its manufacturing, even if said foundry is not trusted. Completing the IC can then be done in a trusted low-end foundry, where the BEOL is stacked on top of the FEOL. Split Manufacturing was successfully demonstrated in , where designs were manufactured with \\textasciitilde 0\\% of faults, and are reported to present a performance overhead of about 5\\%. Therefore, the technique is, in principle, feasible and available for design companies to make use of, such that they can utilize advanced foundries without fully exposing their layouts during manufacturing.\nHowever, there are many caveats to Split Manufacturing. The technique can be successfully applied only if the technologies used to build the FEOL and BEOL are ``compatible''. In theory, a layout can be split at any layer if the chosen layer presents a good interface between FEOL and BEOL. However, since advanced technologies utilize the dual-damascene fabrication process, the layout can only be split on metal layers . Thus, the FEOL cannot terminate in a via layer. The dual-damascene process is characterized by patterning the vias and trenches in such a way that the metal deposition fills both at the same time, i.e., via-metal pairs (e.g., V1 and M2) must mandatorily be built by the same foundry. \nTwo technologies are said to be compatible with each other if there is a way for a BEOL via to land on the FEOL uppermost layer while respecting all design rule checks (DRCs) of both technologies. DRCs are used to guarantee the manufacturability and functionality of an IC, and are defined with respected to the characteristics of the materials utilized and to tolerance ranges of the manufacturing processes (e.g., polishing, patterning, and deposition). These rules encompass minimum enclosure, width, spacing, and density checks. Modern technologies have several options for via shapes -- as long as one via shape is valid, the technologies are compatible for Split Manufacturing purposes. However, in practice, to keep the overhead of the technique under control, an array of via shapes must be feasible, thus providing the physical synthesis with a rich selection for both power and signal routing. \nAccording to , compatibility between two technologies can be generalized by enclosure rules as in Eq.~\\ref{eq:drc_1}, where MW.U.x is the minimum width of Mx on untrusted foundry, VW.T.x is the minimum width of Vx on trusted foundry and EN.T.x is the minimum enclosure on trusted foundry. These rules are illustrated in Figure \\ref{fig:drc_mdim}, where the left side of the image portrays a cross section view and the right side shows the top view. According to Figure \\ref{fig:drc_mdim}, the minimum enclosure width, Mx.EX.Vx, must be compatible between the two foundries. In modern technologies, Equation~\\ref{eq:drc_1} is no longer sufficient as it does not capture the intricate rules for vias and line endings (enclosure from 1 side, 2 sides, 3 sides, T-shaped/hammerheads, etc.).\n \\begin{equation} \\label{eq:drc_1}\n MW.U.x \\geq VW.T.x + (2 EN.T.x)\n \\end{equation}\n\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\linewidth]{metal_view.pdf}\n{Compatibility rules between FEOL and BEOL (adapted from ). \\label{fig:drc_mdim}}\nSplit Manufacturing also presents challenges on the design flow front, which is illustrated in Figure~\\ref{fig:split_sm_flow}. An in-house team designs the circuit, from RTL to layout. Most likely, the layout contains IPs obtained from third parties. Depending on the metal layer where the layout is to be split, it may affect existing IP. Logic and memory IP may use higher metal layers -- memories typically require 4 to 5 metal layers, while standard cells typically require 2 metal layers --, limiting where the split can be done. Standard cells and memories have to be re-designed if they use metal layers that will be split, a grave challenge that may render Split Manufacturing significantly harder to execute.\nStill referring to Figure~\\ref{fig:split_sm_flow}, the FEOL and BEOL are generated using a hybrid process design kit (PDK), and then later split to be manufactured. After splitting the layout correctly, the FEOL is first manufactured in a high-end foundry, and later the BEOL is stacked on top of it by a second (and possibly low-end) foundry. \n\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\linewidth]{sm_flow_dia.pdf}\n{Split Manufacturing Design Flow. \\label{fig:split_sm_flow}}\nEven by splitting the layout, it is often argued that the FEOL exposes enough information to be exploited. Attacks towards the FEOL can effectively retrieve the BEOL connections by making educated guesses. The efficiency of the guessing process is inherently linked to the threat model assumed, which determines the information the attacker possesses to begin with. The literature describes two distinct threat models:\n \\begin{itemize}\n \\item \\textbf{Threat model I:} an attacker located at the untrusted foundry holds the FEOL layout and wants to retrieve the BEOL connections.\n \\item \\textbf{Threat model II:} an attacker located at the untrusted foundry holds the entire gate-level netlist that is assumed to be provided by a malicious observer. The attacker here still holds the FEOL layout and wants to retrieve the BEOL connections. .\n \\end{itemize}\nIt is important to emphasize that the second threat model completely nullifies the security introduced by Split Manufacturing. Possessing the gate-level netlist makes reverse engineering the layout a trivial task, as if the attacker held the entire layout, not only the FEOL. Assuming the attacker has knowledge about the netlist challenges the integrity of the design company itself. If a rogue element exists inside the design company, other representations of the design could be equally stolen, such as the register-transfer level (RTL) code of the design, or even the entire layout (including the BEOL). It could be argued that this vulnerability is so severe that Split Manufacturing virtually stops making sense. For this reason, threat model I is the focus in this work. However, as our goal is to present a comprehensive survey, related works that use threat model II will be covered as well. \nAssuming threat model I, an attacker already knowing all the layers that make up the FEOL, is now interested in retrieving the BEOL connections to recreate the full design (or as close as possible). The commonly used assumption is that attackers are powerful and work within the untrusted foundry in some capacity. Thus, the attackers have deep understanding about the technology. Extracting the (still incomplete) gate-level netlist from a layout is, therefore, a trivial task for the attacker.\nMany approaches to retrieve the BEOL connectivity have been proposed, several of which are termed \\textit{proximity attacks} . Since EDA tools focus on optimizing power, performance, and area (PPA), the solution found by a placement algorithm (that uses heuristics internally) tends to place connected cells close to one another as this will, in turn, reduce area, wirelength, and delay. Therefore, finding the correct missing connections between FEOL and BEOL can be done by assessing input and output pins that are in proximity (thus, the name proximity attack). The more input and output pins to connect, the higher is the probability to make a wrong connectivity guess. Thus, a higher level of security is achieved by splitting the circuit at the lowest metal layer possible.\nAs a promising technique to enhance the security of ICs in this era of fabless chip design, Split Manufacturing still faces some enormous challenges:\n \\begin{itemize}\n \\item[] \\textit{Logistical challenge:} Split Manufacturing is not presently incorporated into the IC supply chain. Finding foundries with compliant technologies that are willing to work with each other is not trivial.\n \\item[] \\textit{Technological challenge:} even within compliant technologies, non-negligible overheads can be introduced if they are vastly different\\footnote{For a thorough discussion and silicon results on BEOL-related overheads, please refer to .}. In the worst-case scenario, it can make routing impossible. Thus, this fact narrows down the technology choices available and feasibility of certain layers as candidates for splitting.\n \\item[] \\textit{Security challenge:} the attained security of straightforward Split Manufacturing is still under debate. Attacks towards the FEOL can be effective, where the hidden connections can be retrieved.\n \\end{itemize}\nFor the purpose of this survey, we categorized related works in the literature as attacks and defenses. In \\textit{attacks}, authors proposed new attack models or modifications of existing attacks in order to improve their effectiveness. In \\textit{defenses}, authors proposed new techniques to use together with Split Manufacturing in order to improve its security level.","id":"c3cea202-e3a7-4e28-a205-e6374a147781","level":"section","origin_cites_number":10,"parent_id":"4df5dfa6-5506-4133-aae0-08641ac96086","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Split Manufacturing: Background"]],"subsections":[],"title":"Split Manufacturing: Background"},{"cite_extract_rate":0.130434782608695,"cites":[7800,7798,7799],"content":"\\label{sec:sec_thrts}\nThe Split Manufacturing technique was developed to protect ICs against threats related to manufacturing in potentially untrusted foundries. In practical terms, to split the layout means to hide some connections from the untrusted foundry. The security provided by Split Manufacturing is based on the fact that the attacker in the FEOL foundry cannot infer the missing BEOL connections. This assumption, however, was challenged by several works where authors proposed attack approaches that can potentially retrieve the missing connections with varying degrees of success. In the text that follows, we present works that proposed Split Manufacturing attacks. These attacks are discussed in chronological order as compiled in Table~\\ref{tab:atcks_desc}. For each studied attack, we reported the related threat model, attack type, novelty, and benchmark circuits used in experiments. Additionally, we reported the largest and average size of the circuits utilized in each work (measured in number of gates). The circuit size is an important metric when analyzing the Split Manufacturing technique because its effectiveness is often proportional to circuit size. In Table~\\ref{tab:atcks_effect}, we compiled results for when the smallest and largest studied circuits are under attack. Also, we reported the circuit size and, if defined, the split layer that the author selected to split the circuit at.\nThe first reported attack is by Jeyavijayan \\etal~and is described in . In this work, the authors assumed that naive Split Manufacturing (i.e., splitting a layout without care for the connections) is inherently insecure. They introduced the concept of proximity attacks that exploits ``vulnerabilities'' introduced by EDA design tools. Since EDA tools focus on optimizing for PPA, the solution found by a placement algorithm tends to place logically connected cells close to one another so they become physically connected during routing. Therefore, the distance between output-input pairs can be used as a metric to recover the missing BEOL connections.\n\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.90\\linewidth]{partitions.pdf}\n{Example of a partitioned circuit. \\label{fig:cir_part}}\nDesigns are commonly partitioned during physical implementation, i.e., separated into small logical blocks with few connections between them. That way, the designer has total control of the floorplaning regarding the placement of blocks. This approach also allows for blocks to be implemented separately and later integrated, creating a sense of parallelism in the design flow which can reduce the overall time required for executing physical synthesis. Consider as an example the circuit illustrated in Figure~\\ref{fig:cir_part} which contains two partitions, A and B, each with 3 gates. Partition A has 3 input pins and one output pin, while partition B has three input pins and two output pins. The partitions have only one connection between them, connecting the output pin of gate G2 with one of the inputs of gate G3. Consider a target output pin from partition A $P_{x,A,out}$ and its corresponding candidate input pin from partition B $P_{x,B,in}$. During placement, the EDA tool will attempt to place the pin $P_{x,A,out}$ as close as possible to $P_{x,B,in}$, possibly closer than any other pin in partition B. Using this insight, an attacker may recover the missing connections in the FEOL layout, performing then a proximity attack. The authors argued that their proposed attack flow is successful due to being able to leverage the following ``hints'' provided by the EDA tools:\n \\begin{itemize}\n \\item[] \\textit{Hint 1 - Input-Output Relationship}: partition input pins are connected either to another partition output pin or to an input port of the IC (i.e., input to input connections are excluded from the search space).\n \\item[] \\textit{Hint 2 - Unique Inputs per Partition}: input-output pins between partitions are connected by only one net. If a single partition output pin feeds more than one input pin, the fan-in and fan-out nodes are usually placed within the partitions (i.e., one-to-many connections are ruled out from the search space). \n \\item[] \\textit{Hint 3 - Combinational Loops}: in general, only very specific structures are allowed to utilize combinational loops (e.g., ring oscillators). These structures are very easy to identify. In most cases, random logic does not contain combinational loops (i.e., connections that would lead to combinational loops can be excluded from the search space).\n \\end{itemize}\nAn attacker can correctly connect a target pin to a candidate pin by identifying the closest pin from a list of possible candidates. The list of possible candidates is created by observing the hints mentioned above. A possible candidate pin is an unassigned output pin of another partition and an unassigned input port of the design. Then, a minimum distance metric is used to connect the pins based on the previously discussed heuristic behavior of EDA tools.\nIn Algorithm \\ref{alg:prox_jevs}, we described the proximity attack detailed in . The input to the algorithm is the FEOL layout, from which the information about unassigned input-output ports can be derived. The algorithm does not describe the specifics of how to derive a netlist from a layout. However, the complexity of this task is rather straightforward. It is assumed that the attacker possesses information about both the PDK and the standard cell library. In many cases, the untrusted foundry is the actual provider of both\\footnote{For the PDK, it is very natural that it is created by the untrusted foundry itself. For standard cell libraries, the cells might be designed by the foundry or by a third-party licensed by the foundry. In either case, the effort to revert a layout to a netlist remains trivial.}. From there, a layout in GDSII or OASIS format can be easily reverted to a netlist by any custom design EDA tool. \n\\begin{algorithm}[h]\n\\DontPrintSemicolon\n \\KwInput{FEOL layers}\n \\KwOutput{Netlist with BEOL connections}\n Reverse engineer FEOL layers and obtain partitions;\n \\While{Unassigned partition pins or ports exist}\n {\n Select arbitrary unassigned pin/port as a targetPin;\n \tListOfCandPins = BuildCandPinsList(targetPin);\n \tSelect candPin from ListOfCandPins that is closest to targetPin;\n \tConnect targetPin and candPin;\n \tUpdate netlist;\n }\n \\textbf{Return:} netlist\n \\textbf{BuildCandPinsList}(targetPin)\n \\KwInput{targetPin $P_{X,i,in}$}\n \\KwOutput{CandPins for targetPin}\n CandPins = Unassigned output pins of other partitions + unassigned input ports of the design;\n \\For{\\textbf{each} $Pin_J \\in$ CandPins}\n {\n \\If{CombinationalLoop(targetPin, $Pin_J$)}\n {\n CandPins -= $Pin_J$;\n }\n }\n \\textbf{Return:} CandPins\n\\caption{Proximity attack}\n\\label{alg:prox_jevs}\n\\end{algorithm}\nFrom the gate-level netlist, the algorithm chooses an arbitrary \\textit{TargetPin} from the unassigned partition input pins and output ports, creates a list of possible \\textit{CandidatePins}, and then connects the \\textit{TargetPin} to the closest pin in this list. After each connection, the netlist is updated. This procedure is repeated until all unassigned ports are connected. When the procedure is over, the attacker obtains the possible missing BEOL connections. If all guesses were correct, the original design has been recovered and Split Manufacturing has been defeated.\n \\begin{table*}[htb]\n \\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Threat Models, Attacks, and Metrics.}\n \\begin{tabular}{ccclp{4.0cm}lp{1.7cm}p{1.2cm}}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Work} & \\textbf{Year} & \\parbox{1cm}{\\textbf{Threat} \\\\\\textbf{model}} & \\textbf{Attack type } & \\textbf{Novelty} & \\textbf{Benchmark suite(s)} & \\parbox{2cm}{\\textbf{Largest circuit} \\\\\\textbf{size (gates)}} & \\parbox{2cm}{\\textbf{Avg. circuit} \\\\\\textbf{size (gates)} } \\vspace{1pt} \\\\ \n \\hline\n & 2013 & I & Proximity & Attack Based on Proximity & ISCAS\\textquotesingle85 & 3.51k & 1288\\\\\n & 2016 & II & Proximity & \\parbox{4.5cm}{Placement and routing proximity \\\\ used in conjunction} & ISPD\\textquotesingle11 & 1.29M & 951k \\\\\n & 2018 & I & Proximity & \\parbox{4.5cm}{Network-Flow-Based with Design \\\\ Based Hints} & ISCAS\\textquotesingle85 \\& ITC\\textquotesingle99 & 190.21k & 9856\\\\\n & 2018 & I & Proximity & \\parbox{4.5cm}{Proximity Attack Based on Machine \\\\ Learning} & ISPD\\textquotesingle11 & 1.29M & 951k\\\\\n & 2019 & I & SAT & SAT Attack without Proximity Information & ISCAS\\textquotesingle85 \\& ITC\\textquotesingle99 & 190.21k & 9856\\\\\n & 2019 & I & SAT & \\parbox{4.5cm}{SAT attack dynamically adjusted \\\\based on proximity information} & ISCAS\\textquotesingle85 \\& ITC\\textquotesingle99 & 190.21K & 9856\\\\\n \\hline \\hline\n \\end{tabular}\n \\label{tab:atcks_desc}\n\\end{table*}\nAlgorithm \\ref{alg:prox_jevs} was originally applied to the ISCAS\\textquotesingle85 suite of benchmark circuits. These circuits were originally selected and published to help in comparing automatic test pattern generation (ATPG) tools. Due to the small size of these circuits, they may not be the best option to assess the effectiveness of Split Manufacturing. The difficulty to retrieve the BEOL connections is directly proportional to the size of the circuit. The authors reported an average effectiveness of 96\\% of Correct Connection Rate (CCR) across all the benchmarks considered. For the c17 circuit (smallest circuit in the ISCAS\\textquotesingle85 suite with only 6 gates), all the connections were retrieved correctly, thus demonstrating that the algorithm is capable of retrieving the missing BEOL connections. In Table~\\ref{tab:atcks_effect}, we highlight the best and worst results in terms of CCR.\nJeyavijayan \\etal were the first to question the security of straightforward Split Manufacturing. Their proximity attack showed promising results, even if the considered benchmark circuits were rather small in size. This was the starting point for other studies proposing different attacks to Split Manufacturing in an attempt to retrieve the missing BEOL connections. Improvements over the original proximity attack, as well as other attacks, are compiled in Table~\\ref{tab:atcks_desc}. \nThe effectiveness of the proximity attack utilizing distance of unassigned pins alone as metric to find missing BEOL connections was questioned by Maga\\~na \\etal . The authors proposed to utilize both placement and routing information in augmented proximity attacks. For their results, large-sized circuits from the ISPD-2011 routability-driven placement contest were used. These benchmarks are better representatives of modern circuits as they contain 9 metal layers and up to two million nets in a design. Thus, in an attempt to increase the success rate of the attack for large-sized circuits, they proposed routing-based proximity in conjunction with placement-centric proximity attacks.\n \\begin{table*}[htb]\n \\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Benchmark Size and Comparison of Attack Results.}\n \\begin{tabular}{cllcllc}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Work} & \\textbf{Benchmark} & \\textbf{Attack} & \\textbf{Split Layer} & \\textbf{Size (In Gate Count)} & \\textbf{Metric} & \\textbf{Result} \\vspace{5pt} \\\\\n \\hline\n & c17 & Proximity & Not Defined & 6 & CCR(\\%) & 100 \\\\\n & c7552 & Proximity & Not Defined & 3513 & CCR(\\%) & 94 \\\\\n & Superblue 1 & Placement Proximity & M2 & 847k & \\% Match in List & 12.84 \\\\\n & Superblue 1 & Placement Proximity & M2 & 847k & CCR(\\%) & 5.479 \\\\\n & Superblue 1 & Routing Proximity & M2 & 847k & \\% Match in List & 71.08 \\\\\n & Superblue 1 & Routing Proximity & M2 & 847k & CCR(\\%) & 0.651 \\\\\n & Superblue 1 & Overlap (P\\&R) Proximity & M2 & 847k & \\% Match in List & 13.05 \\\\\n & Superblue 1 & Overlap (P\\&R) Proximity & M2 & 847k & CCR(\\%) & 3.977 \\\\\n & Superblue 1 & Crouting Proximity & M2 & 847k & \\% Match in List & 82.08 \\\\\n & Superblue 1 & Crouting Proximity & M2 & 847k & CCR(\\%) & 0.651 \\\\\n & c7552 & Network-flow Based Proximity & Not Defined & 3513 & CCR(\\%) & 93 \\\\\n & c7552 & Proximity & Not Defined & 3513 & CCR(\\%) & 42 \\\\\n & B18 & Network-flow Based Proximity & Not Defined & 94249 & CCR(\\%) & 17 \\\\\n & B18 & Proximity & Not Defined & 94249 & CCR(\\%) & < 1 \\\\\n & Superblue 1 & Proximity & M6 & 847k & \\% Match in list & 33.40 \\\\\n & Superblue 1 & Proximity & M6 & 847k & CCR(\\%) & 0.76 \\\\\n & Superblue 1 & ML & M6 & 847k & \\% Match in list & 83.12 \\\\\n & Superblue 1 & ML & M6 & 847k & CCR(\\%) & 1.91 \\\\\n & Superblue 1 & ML-imp & M6 & 847k & \\% Match in list & 74.65 \\\\\n & Superblue 1 & ML-imp & M6 & 847k & CCR(\\%) & 2.11 \\\\\n & Superblue 1 & ML-imp & M4 & 847k & \\% Match in list & 75.45 \\\\\n & Superblue 1 & ML-imp & M4 & 847k & CCR(\\%) & 2.58 \\\\\n & c7552 & SAT Attack & Not Defined & 3513 & Logical Equivalence(\\%) & 100 \\\\\n & B18 & SAT Attack & Not Defined & 94249 & Logical Equivalence(\\%) & 100 \\\\\n & c7552 & Improved SAT Attack & Not Defined & 3513 & Logical Equivalence(\\%) & 100 \\\\\n & B18 & Improved SAT Attack & Not Defined & 94249 & Logical Equivalence(\\%) & 100 \\\\\n \\hline \\hline\n \\end{tabular}\n \\label{tab:atcks_effect}\n\\end{table*}\nA key difference present in is that this work utilizes a different threat model (model II), claiming that the untrusted foundry possesses information about the entire place \\& routed netlist (as well as the FEOL layout). This assumption is hard to reason if the attacker's intent was to overproduce the IC or pirate the IP. For these goals, clearly, this assumption is unnecessary. The attacker himself can, if he indeed possesses the netlist, perform his own physical synthesis and generate his own layout. The interest in reverse engineering the BEOL connections of the original design diminishes. Nevertheless, we report on the strategies employed by the authors of since they build on the approach proposed by .\n\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.90\\linewidth]{sa_onlycir.pdf}\n{Representation of the routing for the first 3 metal layers of a simple circuit (adapted from ). \\label{fig:circ_abs}}\nRegarding the attacks, the authors of proposed four different techniques to identify a small search neighborhood for each pin. The goal is to create a neighborhood that is small enough to make further pruning feasible, and therefore increase the likelihood of including the matching pins. The techniques are called \\textit{placement proximity}, \\textit{routing proximity}, \\textit{crouting proximity} and \\textit{overlap of placement and routing proximity}, and are described in the text that follows. The circuit illustrated in Figure~\\ref{fig:circ_abs} is the example (before the split) that will guide the discussion on these four techniques.\n\\textit{Placement proximity} exploits the placement information of cells. Each split wire is taken from the pin location of the corresponding standard cell that is connected to it. A search neighborhood is defined as a square region centered around the corresponding pin with an area equal to the average areas of the bounding boxes (BB) in a typical design. The authors argued that it can also be measured based on BBs of the non-split wires in the design under attack, under the assumption that the number of wires that remain in the FEOL is also very large in practice. Let us consider the circuit illustrated in Figure~\\ref{fig:circ_abs} as an example. If the split is done at M2, the search area defined using the placement proximity would contain three gates as illustrated in Figure~\\ref{fig:prox_sa} (a). Thus, using the placement proximity search area, the most likely connections are illustrated by the green squares (candidate pins) and by the red squares (target pins). Note that the layer at which the layout is split does not affect the search area defined by placement proximity.\n\\textit{Routing proximity} exploits the routing information. First, for each split wire, pins are identified as the point where the wire is actually cut at the split layer, i.e., the via location. Next, a square area centered around those pins is defined. The size of the square area is defined based on the average BBs of the pins on that layer in the design. This procedure for identification results in different neighborhood sizes according to the split layer location, i.e., the search radius adapts to the routing resources of each layer. A search area defined using routing proximity is illustrated in Figure~\\ref{fig:prox_sa} (b), highlighted in gray and containing a set of routing wires and its respective pins. \n\\textit{Crouting proximity} takes into account routing congestion by exploiting the union of placement and routing proximity. The search area for each pin is defined in such way that the ratio of number of pins to the search area is equal across all the pins in the split layer. Thus, if a pin is located at a high routing congestion area, the search area will be expanded until the pin density in the new search area reaches a target value or the search area grows to four times its starting value. The starting value is set according to the split layer, set as the average of numbers of pins which fall within a BB. A search area defined using crouting proximity is illustrated in Figure~\\ref{fig:prox_sa} (c). \nThe last strategy proposed by also combines placement and routing information. It is referred to as \\textit{Overlap of placement and routing proximities}. The concept here is to include a subset of pins identified by the placement proximity list which have their corresponding pins included in the routing proximity list. According to the author, intuitively, the overlap then identifies a subset of pins which may be more likely to point towards the direction of the matching pin. A search area defined using the overlap of placement and routing proximities is illustrated in Figure~\\ref{fig:prox_sa} (d). \n \\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\linewidth]{prox_sa_new.pdf}\n{Multiple strategies for pin/connectivity search areas according to . \\label{fig:prox_sa}}\nMaga\\~na et al. assessed each strategy using the benchmark circuit \\textit{superblue1}. Different split layers were also considered. In Table~\\ref{tab:atcks_effect}, we compiled the results for split layer M2. By comparing the results, it becomes clear that no strategy was able to recover 100\\% of the missing BEOL connections. The best result was only 5.479\\% of CCR. This is in heavy contrast with the findings of . However, as we previously noted, the circuit sizes differ by orders of magnitude.\nAccording to the authors of , proximity alone is in no way sufficient to reverse engineer the FEOL. However, proximity attacks have merit as they can be used to narrow down the list of candidates to a significantly smaller size. Using crouting proximity, in 82.08\\% of the cases, the search area defined contained the matched pin in the list of candidates. The authors also present results for a circuit split at M8. We opt not to show these results in Table~\\ref{tab:atcks_effect}. Using the circuit \\textit{superblue1} as an example, the number of unassigned pins when the circuit is split at M8 is only 1.2\\% of the pins when split at M2. Therefore, the small number of unassigned pins to be connected overshadows the large circuit used for their experiments. It must also be emphasized that splitting a circuit in such higher layers is rather impractical since M8 tends to be a very thick metal reserved for power distribution in typical 10-metal stacks. There is very little value in hiding a power distribution network from an adversary that wants to pirate an IP. Once again, we opt not to show this result in our comparisons.\nA network-flow based attack model towards flattened designs was proposed by Wang et. al . The authors argued that the proximity attack originally proposed by utilizes hints that can be used only by hierarchical designs, and that modern designs are often flattened \\footnote{We highlight that best practices in circuit design have changed over the years. Hierarchical design was heavily utilized for many years, but it lost favor due to the difficulty in performing reasonable timing budgeting between the many blocks of a system. Thus, flattened designs are often used to facilitate timing closure.}. Based on the original proximity attack, they proposed a proximity attack utilizing five hints: physical proximity, acyclic combinational logic circuit, load capacitance constraint, directionality of dangling wires, and timing constraint. Note that the first two hints are already described by and . The three novel hints are described below:\n \\begin{itemize}\n \\item[] \\textit{Load Capacitance Constraint}: gates can drive a limited load to honor slew constraints. Typically, maximum load capacitance is constrained and has a maximum value defined by the PDK and/or the standard cell characterization boundaries. Hence, an attacker will consider only connections that will not violate load capacitance constraints.\n \\item[] \\textit{Directionality of Dangling Wires}: routing engines tend to route wires from a source to a sink node along the direction of the sink node. Therefore, the directionality of remaining dangling wires at lower metal layers may indicate the direction of their destination cell with a high degree of certainty\\footnote{Metals usually have preferred directions that alternate along the stack (i.e., if M1 is vertical, then M2 is horizontal). Therefore, this hint becomes more effective if the attacker can observe more than one routing layer of the FEOL}. An attacker can disregard connections in the other directions.\n \\item[] \\textit{Timing Constraint}: connections that create timing paths that violate timing constraints can be excluded. An attacker, through an educated guess of the clock period, can determine a conservative timing constraint and exclude any connections that would lead to slower paths. \n \\end{itemize}\nThe network-flow based attacked framework proposed by Yang \\etal~considers two hints proposed by plus the aforementioned hints to create a directed graph $G = (V,E)$, where $V$ is a set of vertices and $E$ is a set of edges. The set ($V$) is composed by the set of vertices corresponding to the output pins $V_o$, and a set corresponding to the input pins ($V_i$), the source vertex ($S$), and the target vertex ($T$). The set $E$ consists of $E_{So}$, edges from $S$ to every output pin vertex, $E_{oi}$, edges from output pin vertices to input pin vertices, and $E_{iT}$, which includes edges from every input vertex to the target vertex. An example of this kind of representation is shown in Figure~\\ref{fig:network_flow}, where (a) is the circuit with missing connections and (b) is the network-flow representation. The detailed problem formulation is omitted from this work. To find the connections, a min-cost network-flow problem is solved, where the decision variables are the flow $x_{i,j}$ going through each edge $(i,j) \\in E$ . The authors utilized the Edmons-karp algorithm to solve this problem. Complexity of the algorithm alone is given by $O(VE^2)$, however, in the worst case it is required to run the algorithm $V$ times; thus, the run-time of the complete network attack is given by $O(V^2E^2)$ in the worst case.\n\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.90\\linewidth]{network_flow.pdf}\n{(a) Circuit with missing connections. (b) Network-flow model for inferring the missing connections. (Adapted from ). \\label{fig:network_flow}}\nThe network-flow approach was applied to ISCAS\\textquotesingle85 and ITC\\textquotesingle99 benchmark circuits. The ITC\\textquotesingle99 circuits were proposed as an evolution of the ISCAS\\textquotesingle85 set, since the Test community acknowledged that newer and larger circuits were already in demand. For comparison, the authors applied both the original proximity attack and the network-flow attack to flattened designs. As shown in Table~\\ref{tab:atcks_effect}, their network-flow proximity attack outperformed the original attack in terms of CCR. However, despite the evident improvement, the attack could only retrieve 17\\% of the missing BEOL connections for a medium sized circuit (b18 from the ITC\\textquotesingle99 suite).\nA Machine Learning (ML) framework was used by Zhang \\etal in an attempt to improve the attack proposed in . The same setup as previously discussed was utilized. However, more layout features were incorporated in their ML formulation, including placement, routing, cell sizes, and cell pin types. \nA high-level overview of their modeling framework is shown in Figure~\\ref{fig:ml_framework} (a). First, they create a challenge instance from the entire layout and only FEOL view. Next, for each virtual pin (point where a net is broken on the split layer), layout information is collected, including placement, routing, cell areas, and cell pin as illustrated in Figure~\\ref{fig:ml_framework} (b). Using this information, samples are generated which are fed into the ML training process. Each sample carries information of a pair of virtual pins which may or may not be matched. Classifiers then are built by the ML framework using training samples. After training and building the regression model, cross validation is used for evaluation which ensures validation of the model is done on data samples which were not used for training. Their framework faces scaling issues when applied to lower split metal layers. An improved ML framework is then proposed as well, denoted by ML-imp, to solve the scaling issues.\nFor their experiments, Zhang \\etal utilize the ISPD\\textquotesingle11 benchmark suite. They compare results from their previous work with their ML and ML-imp frameworks. However, they do not show results for lower split metal layers (e.g., M2). Instead, results are provided for M8, M6, and M4 splits. As pointed out before, utilizing higher layers for the split effectively shrinks the otherwise large circuits used in their experiments. A drastic reduction of unassigned pins is expected for such higher layers, as higher metal layers are used often for power routing, not for signal routing. Results for the \\textit{superblue1} circuit are shown in Table~\\ref{tab:atcks_effect}. Regarding recovering missing BEOL connections, ML and ML-imp could only retrieve around 2\\%, therefore not showing a huge improvement over their previous work. However, search list area accuracy showed significantly better results when compared to their prior work. A caveat worth mentioning is that the proposed machine learning framework needs the entire layout during its modeling phase. This characteristic may, in an extreme case, nullify the applicability for an attacker that only holds the FEOL layout and cannot produce training samples from other sources.\n \\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\linewidth]{ml_features.pdf}\n{(a) Machine Learning Modeling by Zhang et al. . (b) Few Exmples of Layout Features. \\label{fig:ml_framework}}\nAttacks using proximity information as a metric are not the only solution to recover missing BEOL connections. An effective methodology to apply a Boolean satisfiability based strategy is proposed by Chen \\etal . The authors claim that their attack methodology does not need (or depend on) any proximity information, or even any other insights into the nature of EDA tools utilized in the design process. The key insight in their work is to model the interconnect network as key-controlled multiplexers (MUX). Initially, all combinations of signal connections between the FEOL partitions are allowed, as illustrated in Figure~\\ref{fig:mux_network}. First, a MUX network is created in order to connect all missing paths in the circuit. This MUX network leads to potential cyclic paths, thus, there is a possibility to generate many combinational loops during the attack process corresponding to incorrect key guesses. Therefore, constraints on the key values are generated in order to avoid activating the cyclic paths. The attack can be summarized in 4 steps: \\textit{identification of all cyclic paths}, \\textit{generation of cycle constraints}, \\textit{cycle constraints optimizations}, and finally, \\textit{SAT attack}. The authors utilize a SAT solver-based attack method derived from CycSat . The SAT attack algorithm has as input the FEOL circuit with MUX network and a packaged IC that serves as an oracle. The algorithm outputs keys to be used in the MUXes such that correct BEOL connections are made. \nIn reality, presents a different interpretation of Threat model I since the attacker is assumed to possess a functional IC. This IC would then have to be available in the open market for the attacker to be able to purchase it. This characteristic severely narrows down the applicability of this SAT attack. For instance, ICs designed for space or military use will not be freely available, thus an oracle may not be known to the attacker.\nExperimental results presented by utilize ISCAS\\textquotesingle85 and ITC\\textquotesingle99 benchmark circuits. It has been shown that their attack could recover a logically correct netlist for all the studied circuits. However, there is a small clarification to be made that relates to what is a logically correct circuit. In Table~\\ref{tab:atcks_effect}, two of those results are shown. For seven of the studied benchmarks (c1908, c2670, c5315, c7552, b14, b15, b17), the connections recovered are identical to the BEOL connections. For the remaining benchmarks, the recovered connections are not identical but logically equivalent to the original circuit. In practice, the logically equivalent circuit may present performance deviations from the original design. Matching the performance of the original design can be done by re-executing place and route using the logically equivalent gate-level netlist. Depending on the attack goal, it is possible that the attacker had already planned to re-execute the physical synthesis flow again (say, to resell the IP in a different form or shape). An attack that guarantees 100\\% of logic equivalence of the recovered netlist is powerful enough, allowing attackers to copy and modify split layouts.\nIn order to increase the efficiency and capacity of the SAT attack proposed in , the authors proposed two improvements in . First, the size of the key-controlled interconnect network that models the possible BEOL connections is reduced. Second, after the MUX network is inserted into the FEOL circuit, the number of combinational cycles it induces in the design for incorrect key guesses should also be reduced. Proximity information is then exploited to achieve the proposed improvements. The improved SAT attack method which exploits proximity information showed significant reduction in the attack time and increase in the capacity. Same as in , the circuits tested were 100\\% recovered, as shown in Table~\\ref{tab:atcks_effect}. \n\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\linewidth]{mux_network.pdf}\n{MUX network for a bipartitioned FEOL circuit (Adapted from ). \\label{fig:mux_network}}\n\\begin{table*}[htb]\n\\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Split Manufacturing Defenses.}\n \\begin{tabular}{cp{1cm}ccp{2.2cm}p{2cm}l}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Work} & \\textbf{Year} & \\textbf{Threat Model} & \\textbf{Category} & \\textbf{Defense} & \\textbf{Metrics} & \\textbf{Defense Overheads Presented} \\vspace{5pt} \\\\\n \\hline \n & 2013 & I & Proximity Perturbation &Pin Swapping & Hamming Distance & \\multicolumn{1}{c}{-*} \\\\\n & 2013 & II & Wire Lifting &Wire Lifting & k-Distance & Power, Area, Delay and Wire-Length \\\\\n & 2014 & I & Layout Obfuscation & Layout Obfuscation for SRAMs and Analog IPs & - & Performance, Power and Area \\\\\n & 2014 & I & Layout Obfuscation & Obfuscation Techniques & Neighbor Connectedness and Entropy & Performance and Area \\\\\n & 2015 & I & Layout Obfuscation & Automatic Obfuscation Cell Layout & Neighbor Connectedness and Entropy & Performance, Power and Area \\\\\n & 2015 & I & Layout Obfuscation & Obfuscated Built-in Self-Authentication & Obfuscation Connection & Number of Nets \\\\\n & 2016 & I & Wire Lifting & Artificial Blockage Insertion & Number of Pins & \\multicolumn{1}{c}{-*} \\\\\n & 2016 & I & Wire Lifting & Net Partition, Cell Hidden and Pin Shaken & - & \\multicolumn{1}{c}{-*} \\\\\n & 2017 & I & Proximity Perturbation & Routing Perturbation & Hamming Distance & Performance and Wire-Length \\\\\n & 2017 & I & Wire Lifting & Secure Routing Perturbation for Manufacturability & Hamming Distance & Performance and Wire-Length \\\\\n & 2017 & I & Proximity Perturbation & placement-centric Techniques & CCR & Performance, Power and Area \\\\\n & 2017 & II & Proximity Perturbation & Gate Swapping and Wire Lifting & Effective Mapped Set Ratio and Average Mapped Set Pruning Ratio & Wire-Length \\\\\n & 2018 & I & Wire Lifting & Concerted Wire Lifting & Hamming Distance & Performance, Power and Area \\\\\n & 2018 & I & Proximity Perturbation & Secure Driven Placement Perturbation & Hamming Distance & Power and Wire-Length\\\\\n & 2018 & I & Proximity Perturbation & placement and routing perturbation & Hamming Distance & Performance, Power and Area \\\\\n & 2019 & I & Layout Obfuscation & Isomorphic replacement for Cell Obfuscation & Isomorphic Entropy & \\multicolumn{1}{c}{-*} \\\\\n & 2019 & II & Layout Obfuscation & Dummy Cell and Wire Insertion & k-security & Area and Wire-Length \\\\\n \\hline \\hline\n \\multicolumn{7}{l}{\\cellcolor{white}* Authors do not present any discussion regarding overhead.} \n \\end{tabular}\n \\label{tab:sm_defenses}\n\\end{table*}","id":"63582e79-19ab-44ed-8225-2fdf07c8ddcd","level":"section","origin_cites_number":23,"parent_id":"4df5dfa6-5506-4133-aae0-08641ac96086","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Attacks on Split Manufacturing"]],"subsections":[],"title":"Attacks on Split Manufacturing"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:defenses}\nAttacks toward Split Manufacturing showed promising results, as described in the previous section. A malicious attacker has the real potential to recover the missing BEOL connections. If the missing connections are successfully recovered, the security introduced by applying the technique is nullified. Therefore, straightforward Split Manufacturing is questioned by several works. Several authors proposed defense techniques that augment the technique, i.e., techniques that, when used together with Split Manufacturing, do increase the achieved security against attacks. In Table~\\ref{tab:sm_defenses}, we compile a comprehensive list of defense techniques found in the literature. Each defense technique utilizes a different metric and Threat model, depending upon the type of attack they are trying to overcome. Since many of the studied defense techniques often introduce heavy PPA overheads, Table~\\ref{tab:sm_defenses} also shows if the studied work assessed overheads and which ones were addressed. \nIn the text that follows, the many defense techniques are divided into categories, namely Proximity Perturbation (i.e., change the location of cells or pins), Wire Lifting (i.e., move routing wires to upper layers), and Layout Obfuscation (i.e., hide the circuit structure). We present the categories in this exact order. For some techniques, it is worth mentioning that overlaps do exist and that techniques could be categorized differently. Thus, this categorization is our interpretation of the state of the art and may not be definitive. Furthermore, the boundaries between categories are not strict. For example, a technique may perform a layout modification that promotes proximity perturbation and leads to (indirect) wire lifting. In Tables~\\ref{tab:def_res} and \\ref{tab:def_res2}, we compile the results for the Proximity Perturbation and Wire Lifting categories, respectively. To demonstrate the effectiveness of each defense technique, we compile the results for when the attack is done with and without the defense. The results showed in Tables~\\ref{tab:def_res} and \\ref{tab:def_res2} are for the smallest and largest circuits addressed in each studied work. Additionally, we show the PPA overhead introduced and, if specified, the split layer.","id":"0381ac67-f861-41a5-94dd-d57b1d10cc30","level":"section","origin_cites_number":0,"parent_id":"4df5dfa6-5506-4133-aae0-08641ac96086","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Split Manufacturing Defenses"]],"subsections":["ee25dbd5-cfaf-41a3-8225-ebde1d228b40","90e5175d-afc8-4844-8dfa-513e3fcc4fbd","4032829b-2e93-4b12-822b-7b2a6e653dbc"],"title":"Split Manufacturing Defenses"},{"cite_extract_rate":0.28571428571428503,"cites":[7798,7799],"content":"Attacks toward split circuits are generally based on leveraging proximity information. The first category of defenses, Proximity Perturbation, addresses this hint left by the EDA tools. The goal of the techniques within this category is to promote changes in the circuit such that the proximity information between the FEOL pins is less evident. Therefore, the success rate of the proximity attacks is decreased.\nIn , the authors proposed pin swapping to overcome proximity attacks. Rearranging the partition pins can alter their distance in such a way that the attacker is deceived. As an example, if the pins $P_{G3,B,in}$ and $P_{G6,A,in}$ (Figure~\\ref{fig:cir_part}) are swapped, the proximity attack will incorrectly guess the connection between $P_{G2,A,out}$ and $P_{G3,B,in}$. Thus, a sufficient number of pins have to be swapped in order to create a netlist that is significantly different from the original netlist (based on some sort of metric for similarity). In , Hamming distance is proposed as a way to quantify the difference between the outputs of the original netlist and the modified netlist. Assuming the outputs of a circuit are arranged as a string of bits, Hamming distance is defined as the number of bits that change when two instances of this string are compared to one another. The authors argued that the optimum netlist is created when the Hamming distance is 50\\%. Therefore, inducing the maximal ambiguity for a potential attacker. Since the best rearrangement for N pins of partitions might take $N!$ computations (rather computationally expensive), pair-wise swapping of pins is considered in . Pair-wise swapping of pins results in $O(N^2)$ computations. \nThe modified netlist is created based on a series of rules. Similarly, to the proximity attack, a list of candidates pins to be swapped is created before the actual swap is applied. Since not every pin can be swapped, a candidate pin to be swapped should:\n \\begin{itemize}\n \\item be an output pin of the partition where the target pin resides\n \\item not be connected to the partition where the candidate pin resides\n \\item not form a combinational loop\n \\end{itemize}\nUsing the above constraints, a candidate pin is selected. The target pin also needs to be chosen carefully. In , IC testing principles and hints from the original proximity attack are used to choose the target pin. The swapping procedure is described in Algorithm \\ref{alg:pin_swap}, where $TestMetric$ is a metric based on IC testing principles, such as stuck-at fault models which are still utilized in Test today. More details can be obtained from . The proposed defense technique is validated using ISCAS\\textquotesingle85 circuits and the original proximity attack. For the smallest circuit, c17, it takes only one swap to achieve a Hamming Distance of 50\\%. For the largest studied circuit, c7552, it takes 49 swaps. These results are summarized in Table~\\ref{tab:def_res}.\nAs demonstrated in , rearranging the partition pins can thwart proximity attacks. However, according to Chen \\etal , pin swapping at partition level has limited efficacy. They demonstrated that an attacker holding the FEOL layout as well as the nestlist can insert hardware trojans even when the defense approach of is applied. It must be highlighted that assumes threat model II, which we have previously argued that has the potential to nullify the vast majority of defenses towards split circuits. Thus, they proposed a defense to counter the threat from hardware trojans. Their defense incorporates the global wire-length information, with the goal to hide the gates from their candidate locations, and as result decreasing the effective mapped set ratio (EMSR). The EMSR metric is an attempt to quantify the ratio of real gates location of a given mapping during a simulated annealing-based attack. This defense consists of two steps, first a greedy gate swapping defense , and second, a measurement of the security elevation in terms of EMSR. The technique is evaluated using ISCAS\\textquotesingle85 benchmarks circuits and the EMSR metric to quantify the defense effectiveness. The results are shown in Table~\\ref{tab:def_res}.\nFollowing the same principle of increasing the Hamming Distance, Wang \\etal proposed a routing perturbation based defense. The optimum Hamming distance is sought to be achieved by layer elevation, routing detours, and wire decoys, while test principles are used to drive the perturbation choices. Layer elevation is essentially a wire lifting technique: without changing the choice of split metal layer, wires are forced to route using higher metal layers, thus being lifted from the FEOL to the BEOL. Intentional routing detours are a way to increase the distance between disconnected pins of the FEOL. If done properly, disconnected pins will not be the closest to each other, deceiving the proximity attack. In some cases, routing detours will increase the distance between disconnected pins, however, they still remain the closest to each other. In this scenario, wire decoys can be drawn near disconnected pins, in such a way that decoys are now the closest and will instead be picked as the ideal candidate pin. \nThe perturbations proposed in can incur heavy overheads, and, for this reason, wires to be perturbed are chosen by utilizing IC test principles. In , fault observability, as defined in SCOAP , is used as a surrogate metric for this task. The technique is evaluated using ISCAS\\textquotesingle85 and ITC\\textquotesingle99 benchmark circuits. For all studied circuits, the Hamming distance increased by an average of 27\\% at a cost of only 2.9\\% wire length overhead (WLO), on average. The results for the largest and smallest studied circuits are shown in Table~\\ref{tab:def_res}.\n \\begin{algorithm}[htb]\n\\DontPrintSemicolon\n \\KwInput{Partitions}\n \\KwOutput{List of target and swapping pins}\n $ListofTargetPins$ = $\\emptyset$;\\\\\n $ListofSwappingPins$ = $\\emptyset$;\\\\\n $ListofUntouchedPins$ = All partition pins and I/O ports;\\\\\n \\While{Untouched output partitions pins or input ports exist}\n {\n \\For{$UntouchedPin$}\n {\n $SwappingPins$ = \\\\\n BuildSwappingPinsList($UntouchedPin$);\n \\For{$SwappingPin$ $\\in$ $SwappingPins$}\n {\n Compute \\\\\n $TestMetric$($UntouchedPin,\n SwappingPin$);\n }\n }\n Find the $TargetPin$ and $SwappingPin$ with the Highest $TestMetric$ from its SwappingPins;\\\\\n $ListofTargetPins += TargetPins$;\\\\\n $ListofSwappingPins += SwappingPins$;\\\\\n $ListofUntouchedPins -= TargetPins$;\\\\\n $LisofUntouchedPins -= SwappingPin$; \\\\\n Swap TargetPin and SwappingPin; \\\\\n Update netlist;\n }\n \\textbf{Return:} ListofTargetPins and ListofswappingPins;\n \\textbf{BuildSwappingPinList}($TargetPin$); \\\\\n \\KwInput{$TargetPin P_{x,i,out}$}\n \\KwOutput{$SwappingPins$ for $TagetPin$}\n \\For{$Pin_J \\in SwappingPins$}\n {\n \\If{$CombinationalLoop(TargetPin,Pin_J)$}\n {\n $SwappingPins -= Pin_J$;\n }\n }\n \\textbf{Return:} $SwappingPins$;\n\\caption{ Fault analysis-based swapping of pins to thwart proximity attack (adapted from ).}\n\\label{alg:pin_swap}\n\\end{algorithm}\nSengupta \\etal take a different direction from other works. They utilized an information-theoretic metric to increase the resilience of a layout against proximity attacks. As demonstrated in , mutual information (MI) can be used to quantify the amount of information revealed by the connectivity distance between cells. Mutual information is calculated by taking into account the cells connectivity $D$, if they are connected or not, and their Manhattan distance $X$, described by Eq. \\ref{eq:mutual_info}, where $H[\\cdot]$ is the entropy. The Manhattan distance of two cells is defined as the sum of horizontal and vertical distances between them. Entropy is a measure of disorder of a system. Therefore, in this work, entropy is utilized as a measure of disorder in the FEOL layer. The distribution of the variable $X$ for a given layout is determined pair-wise for all gates, allowing a straightforward computation of $I(X;D)$. Thus, layouts with the lowest mutual information, i.e., the correlation between cell connectivity and their distance is low, are more resilient against proximity attacks.\n \\begin{equation}\\label{eq:mutual_info}\n MI = I(X;D) = H[X]-H[X/D]\n \\end{equation}\nIn order to minimize the information ``leaked'' from mutual information, applies cell placement randomization and three other techniques: g-color, g-type1, and g-type2. Randomizing the cell placement can achieve the desired low mutual information; however, the PPA overhead incurred is excessive. Minimizing mutual information without excessive PPA overhead can be achieved by the other techniques. From a graph representation of the circuit, graph coloring can be used to hide connectivity information, where gates of the same color must not be connected. Thus, the resulting colored netlist is then partitioned by clustering all cells of same color together. During cell placement, the cells with the same color will be confined within their respective clusters. According to , these constraints naturally mitigate the information leakage to a great extent. The g-color technique utilizes only the graph coloring as described above. The other two, g-type1 and g-type2, consider the type of the gate when creating clusters. The g-type1 approach clusters gates only by their functionality, while g-type2 utilizes functionality and the number of inputs for clustering. The authors assessed their techniques utilizing ISCAS\\textquotesingle85 and MNCN benchmark suites. Results for the smallest and largest circuits are shown in Table~\\ref{tab:def_res}. \n\\begin{table*}[htb]\n\\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Results for Defense Techniques based on Proximity Perturbation.}\n \\begin{tabular}{p{.8cm}p{1cm}p{1.2cm}p{2.5cm}lp{2.2cm}p{.5cm}p{1.5cm}p{1.5cm}}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Work} & \\textbf{Attack Type} & \\textbf{Benchmark} & \\textbf{Defense Technique} & \\textbf{Defense Metric} & \\textbf{Defense Overhead} & \\textbf{Split Layer} & \\textbf{Result without Defense} & \\textbf{Result with Defense} \\vspace{5pt} \\\\\n \\hline\n & Proximity & c17 & - & Hamming Distance & 1 Swap for 50\\% HD & -* & 100\\% CCR & 78\\% CCR\\\\\n & Proximity & c7552 & - & Hamming Distance & 49 Swaps for 50\\% HD & -* & 94\\% CCR & 91\\% CCR\\\\\n & Proximity & c432 & Modifed Greedy Gate Swapping & EMSR & 75\\% of WLO & -* & 90\\% EMSR & 25\\% EMSR \\\\\n & Proximity & c432 & Modifed Greedy Gate Swapping & EMSR & 300\\% of WLO & -* & 78\\% EMSR & 10\\% EMSR \\\\\n & Proximity & c432 & - & Hamming Distance & 3.1\\% WLO for 46.1\\% HD & -* & 92.4\\% CCR & 78.8\\% CCR \\\\\n & Proximity & c432 & - & Hamming Distance & 4.1\\% WLO for 31.7\\% HD & -* & 62.8\\% CCR & 37.9\\% CCR \\\\\n & Proximity & c432 & Random & Mutual Information & < 10\\% PPA & M1 & 17\\% CCR & < 1\\% CCR \\\\\n & Proximity & c432 & g-color & Mutual Information & < 10\\% PPA & M1 & 17\\% CCR & 2\\% CCR \\\\\n & Proximity & c432 & g-type1 & Mutual Information & < 10\\% PPA & M1 & 17\\% CCR & 6\\% CCR \\\\\n & Proximity & c432 & g-type2 & Mutual Information & < 10\\% PPA & M1 & 17\\% CCR & 4.5\\% CCR \\\\\n & Proximity & c7552 & Random & Mutual Information & < 10\\% PPA & M1 & 13\\% CCR & < 1\\% CCR \\\\\n & Proximity & c7552 & g-color & Mutual Information & < 10\\% PPA & M1 & 13\\% CCR & 2\\% CCR \\\\\n & Proximity & c7552 & g-type1 & Mutual Information & < 10\\% PPA & M1 & 13\\% CCR & 4\\% CCR \\\\\n & Proximity & c7552 & g-type2 & Mutual Information & < 10\\% PPA & M1 & 13\\% CCR & 3\\% CCR \\\\\n & SAT & c432 & BEOL+Physical & Perturbation & 4.5\\% WLO & -* & 58\\% CCR & 56\\% CCR \\\\\n & SAT & c432 & Logic+Physical & Perturbation & 5.57\\% WLO & -* & 58\\% CCR & 58\\% CCR \\\\\n & SAT & c432 & Logic+Logic & WLD & 1.68\\% WLO & -* & 58\\% CCR & 52\\% CCR \\\\\n & SAT & b18 & BEOL+Physical & Perturbation & 8.06\\% WLO & -* & 15\\% CCR & 14\\% CCR \\\\\n & SAT & b18 & Logic+Physical & Perturbation & 1.70\\% WLO & -* & 15\\% CCR & 17\\% CCR** \\\\\n & SAT & b18 & Logic+Logic & WLD & 0.61\\% WLO & -* & 15\\% CCR & 16\\% CCR** \\\\\n & Proximity & c432 & Netlist Randomization & Hamming Distance & < 10\\% PPA overall & -* & 92.4\\% CCR & 0\\% CCR \\\\\n & Proximity & c7552 & Netlist Randomization & Hamming Distance & < 10\\% PPA overall & -* & 94.4\\% CCR & 0\\% CCR \\\\\n \\hline\n \\hline\n \\multicolumn{9}{l}{\\cellcolor{white}* Split layer not specified by the authors.} \\\\\n \\multicolumn{9}{l}{\\cellcolor{white}** These results are counter-intuitive, the applied defense degrades the metric.} \n \\end{tabular}\n \\label{tab:def_res}\n\\end{table*} \nSimilar to the pin swapping technique proposed in , Wang \\etal proposed a placement-based defense with the same objective of deceiving a proximity attack by perturbing proximity information. Differently from pin swapping, their placement-based defense considers the incurred wire-length overhead as a metric. This technique is based on changing gate locations such that the proximity hint is no longer effective. Their algorithm consists of two phases, one to select which gates to be perturbed and a second phase where the selected gates are (re)placed. Gate selection is done by extracting a set of trees using two techniques, BEOL-driven and logic-ware extraction. The first approach selects all gate trees that contain any metal wires in the FEOL, i.e., connections that are not hidden from the attacker. The second approach considers the wire-length impact and the gate tree impact on the overall security. After extracting the set of trees, the placement perturbation is done in one of two ways: physical-driven or logic-driven. For each extracted tree, the physical-driven perturbation changes the location of gates using a Pareto optimization approach. Also, each solution is evaluated by its wire-length overhead and a perturbation metric that discerns the placement difference from the original layout. According to , geometric-based difference alone may be insufficient to enhance the split circuit security. Thus, a logic-driven perturbation is performed with a weighted logical difference (WLD) metric, which encourages perturbation solutions with large logical difference from its neighbors. The authors assessed their techniques combining the gate selection and perturbation as BEOL+Physical, Logic+Physical and Logic+Logic, using ISCAS\\textquotesingle85 and ITC\\textquotesingle99 circuit benchmarks. Results for the smallest and largest circuits considered are shown in Table~\\ref{tab:def_res}.\nA considerably different approach is proposed by Patnaik \\etal , whereas netlist modifications are promoted (instead of placement/routing modifications during physical synthesis). The goal is to modify the netlist of a design in order to insert (partial) randomization. According to , this approach helps to retain the misleading modifications throughout any regular design flow, thereby obtaining more resilient FEOL layouts where the netlist changes are later ``corrected'' in the BEOL. This methodology is implemented as an extension to commercial EDA tools with custom in-house scripts. The process goes as follows: first, the netlist is randomized. Second, the modified netlist is place and routed. Lastly, the true functionality is restored by re-routing in the BEOL. For the netlist randomization, pairs of drivers and their sinks are randomly selected and swapped. This is done in such way to avoid combinational loops that may be introduced by swapping. The modified netlist then is place and routed, utilizing a `do not touch'\\footnote{This terminology is used in IC design to mean that a specific cell or family of cells should not be optimized, i.e., not to be touched.} setting for the swapped drivers/sinks to avoid logic restructuring/removal of the related nets. Finally, the true connectivity is restored in the BEOL with the help of correction cells that resemble switch boxes. The technique is evaluated using ISCAS\\textquotesingle85 circuits, and the results for the largest and smallest circuit are shown in Table~\\ref{tab:def_res}.","id":"ee25dbd5-cfaf-41a3-8225-ebde1d228b40","level":"subsection","origin_cites_number":7,"parent_id":"0381ac67-f861-41a5-94dd-d57b1d10cc30","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Split Manufacturing Defenses"],["subsection","Proximity Perturbation"]],"subsections":[],"title":"Proximity Perturbation"},{"cite_extract_rate":0.07692307692307601,"cites":[7800],"content":"Hiding routing information from untrusted foundries is the main objective of the Split Manufacturing technique. Since attacks mainly rely on hints left by EDA tools to recover the missing BEOL connections, the amount of hidden information is related to the circuit performance -- splitting the circuits at low metal layers increases the security level. Following the same idea, wire lifting proposes `lifting' wires from the FEOL layer to the BEOL. That is, changing the routing to split metal layers has the potential to increase the security level. \n \\begin{table*}[htb]\n \\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Results for Defense Techniques based on Wire Lifting.}\n \\begin{tabular}{cp{1cm}p{1.5cm}p{2.5cm}llp{1cm}p{1.5cm}p{1.5cm}}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Work} & \\textbf{Attack Type} & \\textbf{Benchmark} & \\textbf{Defense Technique} & \\textbf{Defense Metric} & \\textbf{Defense Overhead} & \\textbf{Split Layer} & \\textbf{Result without Defense} & \\textbf{Result with Defense} \\vspace{5pt}\\\\\n \\hline\n & SAT & c432 & Wire Lifting & \\textit{k-security} & 477\\% of WLO & -* & k=1 & k=48 \\\\\n & Proximity & Superblue 1 & Routing Blockage Insertion & $E[LS]$ & Not Presented & M4 & 1.51 & 1.77 \\\\\n & Proximity & Superblue 1 & Routing Blockage Insertion & $FOM$ & Not Presented & M4 & 1222.8 & 1433 \\\\\n & Proximity & c432 & Concerted Lifting & Hamming Distance & 7.7\\% of Area & Average** & 23.4 & 45.9 \\\\\n & Proximity & c432 & Concerted Lifting & CCR & 13.2\\% of Power & Average** & 92.4 & 0 \\\\\n & Proximity & c7552 & Concerted Lifting & Hamming Distance & 16.7\\% of Area & Average** & 1.6 & 25.7 \\\\\n & Proximity & c7552 & Concerted Lifting & CCR & 9.3\\% of Power & Average** & 97.8 & 0 \\\\\n & Proximity & c2670 & CMP-Friendly & Hamming Distance & 3.4\\% of WLO & -* & 14.5\\% & 20.4\\% \\\\\n & Proximity & c2670 & CMP-Friendly & CCR(\\%) & 3.4\\% of WLO & -* & 48.1\\% & 33.4\\% \\\\\n & Proximity & b18 & CMP-Friendly & Hamming Distance & 0.4\\% of WLO & -* & 21.6\\% & 27.6\\% \\\\\n & Proximity & b18 & CMP-Friendly & CCR(\\%) & 0.4\\% of WLO & -* & 12.1\\% & 10.7\\% \\\\\n & Proximity & c2670 & SADP-Compliant & Hamming Distance & 7.49\\% of WLO & -* & 14.5\\% & 24.4\\% \\\\\n & Proximity & c2670 & SADP-Compliant & CCR(\\%) & 7.49\\% of WLO & -* & 48.1\\% & 6.4\\% \\\\\n & Proximity & b18 & SADP-Compliant & Hamming Distance & 4.64\\% of WLO & -* & 21.6\\% & 29.6\\% \\\\\n & Proximity & b18 & SADP-Compliant & CCR(\\%) & 4.64\\% of WLO & -* & 12.1\\% & 2.7\\% \\\\ \\hline\n & Proximity & s526 & Net Partitioning & CCR(\\%) & Not Presented & -* & 40\\%*** & 0\\%*** \\\\\n & Proximity & s526 & Net Partitioning \\& Cell Hiding & CCR(\\%) & Not Presented & -* & 40\\%*** & 0\\%*** \\\\\n & Proximity & s526 & Net Partitioning \\& Cell Hiding \\& Pin Shaking & CCR(\\%) & Not Presented & -* & 40\\%*** & 0\\%*** \\\\\n & Proximity & s9234.1 & Net Partitioning & CCR(\\%) & Not Presented & -* & 30\\%*** & 4\\%*** \\\\\n & Proximity & s9234.1 & Net Partitioning \\& Cell Hiding & CCR(\\%) & Not Presented & -* & 30\\%*** & 1.5\\%*** \\\\\n & Proximity & s9234.1 & Net Partitioning \\& Cell Hiding \\& Pin Shaking & CCR(\\%) & Not Presented & -* & 30\\%***& 1.5\\%*** \\\\\n \\hline \n \\hline\n \\multicolumn{9}{l}{\\cellcolor{white}* Split layer not specified by the authors.} \\\\\n \\multicolumn{9}{l}{\\cellcolor{white}** Results are given as an average between M3, M4, and M5.} \\\\\n \\multicolumn{9}{l}{\\cellcolor{white}*** These results cannot be directly compared with previous ones as the transistor technology is vastly different.} \n \\end{tabular}\n \\label{tab:def_res2}\n\\end{table*} \nWire lifting was first presented by Imerson \\etal where Split Manufacturing is considered as a 3D IC implementation . For the sake of argument, we will continue to refer to this technique as Split Manufacturing, even if the notion of untrusted FEOL vs. trusted BEOL is shifted. This type of 3D implementation consists of two or more independently manufactured ICs, where each IC represents a tier that is vertically integrated on top of each other. Connections between the tiers are done using vertical metal pillars, referred to as through-silicon vias (TSVs). In , a 3D implementation consisting of two tiers is used for their experiments. The bottom tier containing the transistors and some routing wires (akin to the FEOL), and the top tier, containing only routing wires (akin to the BEOL). Regarding the manufacturing of these 3D ICs, the bottom tier is built in a high-end untrusted foundry, and the top tier is built in an also untrusted foundry (not necessarily high-end, however).\nIn , threat model II is used, i.e., the adversary is assumed to possess the entire netlist. The problem is formulated as the attacker being the FEOL foundry, which in turn also possesses the so called `unlifted netlist' extracted from the FEOL layout. By utilizing a graph to represent the circuits as previously described, the attacker seeks a bijective mapping of gates of the unlifted netlist to gates in the complete netlist. According to , if the attacker can distinguish any gate between the two netlists, the split circuit does not provide any security. A security notion is provided by the authors, based on existing multiple mapping between gates in the unlifted and complete netlists. Referred to as \\textit{k-security}, this metric qualifies that gates across the design are indistinguishable from at least $k-1$ other gates. Thus, a defender wants to lift wires in a way to guarantee the higher $k-security$ possible. Two procedures are proposed to achieve this goal, one utilizing a greedy heuristic targeted at small circuits (due to scalability issues), and another procedure that utilizes partitioning to solve those issues. For their experimental study, they have utilized the ISCAS\\textquotesingle85 benchmark suite and a DES crypto circuit with approximated 35000 gates. The results are shown in Table~\\ref{tab:def_res2}, where $k=1$ is the original circuit and $k=48$ is achieved when all the wires are lifted. It is worth to mention that; besides the notion of the security metric, their defense technique was not validated using an actual proximity attack towards the modified netlist.\nAn artificial routing blockage\\footnote{This terminology is used in IC design to mean that a specific area should be avoided by the EDA tool for a specific task. A blockage can be for placement and/or for routing.} insertion that promotes wire lifting is proposed by Maga\\~na \\etal . The goal of this technique is to deceive proximity attacks by wire lifting. As discussed before, the objective of commercial EDA tools is to guarantee the best PPA possible. During the routing stage, lower metals are preferred for signal routing, promoting better PPA. Thus, routing blockages can be inserted at the split layer, forcing signals to be routed above the split layer. The result is an artificial wire lifting done during the routing stage. \nApplying this type of procedure must be done considering the design routability and overhead introduced, as well as top level floorplan decisions for the power grid, clock distribution, and resources for busses. Larger designs are generally difficult to be routed -- simply reducing the number of routing layers can make the design unroutable. In , a procedure is proposed to insert routing blockages ensuring the design routability is kept. After a first routing stage, the design is divided into small rectangular non-overlapping windows. The routing congestion is then analyzed in each window at the split layer for the blockage insertion. If the area has capacity for more routing, a routing blockage is inserted, otherwise the original routing is kept. Utilizing ISPD\\textquotesingle11 circuits, the technique is evaluated using the proximity attack proposed by , and its effectiveness is measured using two metrics, $E[LS]$ and $FOM$. The $E[LS]$ metric reports the candidate list size, being an average over different search areas. The $FOM$ metric is a figure of merit obtained from the ratio of candidates list size divided by the search area, when averaged over all the search areas at the split layer. According to , a higher value of $FOM$ means it is more challenging for an attack to be mounted because of the density of candidates (over the same search area). The results for the \\textit{Superblue 1} circuit are shown in Table~\\ref{tab:def_res2}.\nDesign for Manufacturability (DFM) has become an extremely important aspect of IC design for many years now. Manufacturing an IC is a sensitive process that involves many critical steps. Hence, a layout is required to be compliant to several rules to ensure its manufacturability. A layout is said to be manufacturable if there are no DRC violations. However, for a design to also achieve high yield, the layout must also pass strict DFM checks. The most common checks are related to wire and via density over predetermined region sizes. Until now, defense techniques discussed were mainly concerned about security and PPA overheads. Feng \\etal argued that previous works have largely neglected manufacturability concerns. Therefore, they proposed two wire-lifting techniques that address two important DFM-related techniques: Chemical Mechanical Planarization (CMP) and Self-Aligned Double Patterning (SADP) . The first technique, CMP-friendly routing defense is divided into layer elevation, wire selection, and re-routing. Layer elevation selects wires for lifting according to following principles :\n \\begin{itemize}\n \\item The wire has a significant logic difference from its neighboring wires. As such, an incorrect connection in attacking this wire may lead to more signal differences.\n \\item The wire has large observability such that an erroneous guess by the adversary can easily affect the circuit primary output signals.\n \\item The wire segment is originally at a wire-dense region. The wire density of this region would be reduced by the layer elevation and makes the corresponding FEOL layer have more uniform wire density.\n \\item The BEOL region where the wire is elevated to has low wire density so that the density of the corresponding BEOL layer is more uniform.\n \\end{itemize}\nPrinciples 1 and 2 have the goal to increase security in the same way as described in . After the wire lifting step, a set of wires is selected for re-routing. The selection has two purposes, CMP-friendliness and security improvement. For CMP-friendliness, wires located in dense regions are selected to be re-rerouted in sparse areas. For the security improvement, decoys are inserted if the routing detour passes through a sparse area. A suspicious attacker may realize that the detour is a defense measure. After selecting the set of wires to be re-routed, wires are re-routed one at a time. According to , their routing approach considers wire density, while the routing perturbation proposed by can be solely focused on security, and may not be CMP-friendly. Utilizing a graph representation, their re-routing method is based on the Dijkstra's shortest path algorithm where the density of wires is used as a metric. \nWith a few exceptions, the SADP-compliant routing defense follows the same approach as described above. During wire lifting, the density is not considered. Wire re-routing is actually wire extension of FEOL wires as in . This wire extension of FEOL wires inevitably leads to re-routing of connected BEOL wires. According to , solving SADP violations by wire extension can also increase security, as its increase the distance between vias. The wire extension for simultaneous SADP-compliance and security is realized using Integer Linear Programming. In their experiments, ISCAS\\textquotesingle85 and ISPD\\textquotesingle11 are used to evaluate their techniques. Each technique, CMP-friendly and SADP-compliant routing, is evaluated separately. The results for the smallest and largest circuits are shown in Table~\\ref{tab:def_res2}.\nWire lifting approaches, in general, are not cost-free. As shown in the discussed results, wire-lifting based defenses introduce a considerable PPA overhead. An approach to establish a cost-security trade-off is proposed by Paitinak \\etal , i.e., PPA margins for a given security budget. In , a concerted wire-lifting method is proposed. The authors claim to enable higher degrees of security while being cost-effective. For their method, custom elevating cells are used for executing the wire-lifting. Elevating cells connect gates or short wires directly to the first layer above the split layer. Their wire-lifting method utilizes three strategies: lifting high-fanout nets, controlling the distance for open pin pairs, and obfuscation of short nets. High-fanout nets are chosen to be lifted for two reasons: (a) a wrong connection made by the attacker propagates the error to multiple locations, and, (b) introduces multiple open pin pairs. As the attack to overcome is the proximity one, controlling the distance between open pin pairs is necessary, which is achieved at will simply by controlling the placement of the elevating cells. According to , short nets may be easy for an attacker to identify and localize (from assessing driving strengths). Short wires are obfuscated by inserting an elevating cell with two pins close to each other, one being the true connection and the other a dummy connection. Finally, wires are lifted according to those strategies until a given PPA budget is reached. For their experimental study, ISCAS\\textquotesingle85 and ISPD\\textquotesingle11 circuits are utilized. However, results for attacks are presented only for ISCAS\\textquotesingle85 circuits. For ISPD\\textquotesingle11, only the PPA impact result introduced by their technique is presented. Once again, we present the results for the smallest and largest of the studied circuits in Table~\\ref{tab:def_res2}.\nWhile the majority of studies reported in our survey make use of conventional transistors (bulk CMOS technologies with either planar or FinFET transistors), Yang \\etal proposed a design methodology to secure Split Manufacturing for Vertical Slit Field Effect Transistor (VeSFET)-based integrated circuits. VeSFET is a twin-gate device with a horizontal channel and four metal pillars implementing vertical terminals . While a detailed explanation on VeSFETs is beyond the scope of this work, we do highlight the differences between VesFETs and conventional transistors. In contrast with conventional transistors, a VeSFET can be accessed by both top and bottom pillars, allowing two-side routing and offering a friendly monolithic 3D integration . While we have so far considered ICs that have two distinct layers, the FEOL and BEOL, a VeSFET-based IC has tiers of the layer containing the transistors. Connections between tiers can be made directly, same as TSV by the pillars, or by a layer containing connections between tiers. A 2D VeSFET design contains only one tier and both top and bottom connections, whereas a 3D design contains two or more tiers. In summary, the notion of tier is pushed down to the transistor level in this device topology, thus making it an interesting platform for Split Manufacturing.\nThe method proposed by assumes that both foundries are untrusted and have the same capability (i.e., same technology). For 2D designs, the first foundry manufactures the tier with the top connections, comprising most of the connections. Then, the rest of the bottom connections, comprising of the critical minority connections, are completed by the second foundry. For 3D IC designs, they proposed special types of standard cells, referred as cell A and B. Cell A has two tiers that are visible and manufactured by the first foundry, as well as inter-tier connections. Cell B has only the top tier visible and manufactured by the first foundry, the low tier is completed by the second foundry, without inter-tier connections. Thus, transistors can be hidden from the first foundry as a security feature. Vulnerabilities claimed by for both 2D and 3D methods are described in Table~\\ref{tab:vesfet_vul}. Practices of reverse engineering and IC overbuilding are claimed to be impossible because the first foundry controls the number of wafers available to the second foundry.\nIncreasing the security of both 2D and 3D VesFET designs is achieved by net partitioning, and exclusively for 3D designs, by transistor hiding and pin shacking. Net partitioning is performed similarly to the wire lifting techniques described above, where nets are chosen to be routed in the bottom connection layer, thus, hiding those from the first foundry. Their selection method is done by selecting nets from sequential logic. First, all the high-fanout nets are selected to be partitioned. Next, the remaining nets are selected by a search area, where two approaches are used, distance-first search and high-fanout first search. In distance-first method, a pin in a predefined search window connecting to an un-partitioned net is selected when it has the minimum distance to the currently processed pin pair. The FO-first search method selects the pin connecting to a net having the highest FO in the searching window. Transistor hiding in 3D designs is done by utilizing cells similar to the cell B. Cells connected only by partitioned nets are candidates for hiding. After selecting the candidates, availability of unused transistors that are accessible to the second foundry in the lower tiers of the nearby cells is checked. If the available transistor count is sufficient, then the cell is hidden. The empty space created could provide clues for the first foundry about the security technique. Pin shaking is then applied to obfuscate the empty spaces. Some nearby cells are moved to this area to obfuscate the layout for any distance-based proximity attackers. In , 10 MCNC LGSynth\\textquotesingle91 benchmark circuits are used to evaluate the effectiveness of their methodology. The best and worst results are shown in Table~\\ref{tab:def_res2}. It is worth to mention that, even though the VesFET implementation mimics the layered structure of Split Manufacturing, the results cannot be compared side by side in a fair manner.\n \\begin{table}[htb]\n \\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Vulnerabilities of Split Manufactured VesFET-based Designs Described by .}\n \\begin{tabular}{lll}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Threats} & \\textbf{1st Foundry} & \\textbf{2nd Foundry} \\vspace{5pt} \\\\\n \\hline\n \\parbox{1cm}{Design\\\\ Reconstruction} & \\parbox{2.5cm}{2D IC: Very Difficult \\\\ 3D IC: Impossible} & \\parbox{2cm}{Impossible due to a very limited\\\\ information} \\\\\n Trojan Insertion & \\parbox{2.5cm}{Possible, but will be\\\\ detected} & \\parbox{2cm}{No control of devices} \\\\\n \\parbox{1cm}{Reverse\\\\ Engineering} & Meaningless & Impossible\\\\\n IC Overbuilding & Meaningless & Impossible \\\\\n \\hline\n \\hline\n \\end{tabular}\n \\label{tab:vesfet_vul}\n \\end{table}","id":"90e5175d-afc8-4844-8dfa-513e3fcc4fbd","level":"subsection","origin_cites_number":13,"parent_id":"0381ac67-f861-41a5-94dd-d57b1d10cc30","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Split Manufacturing Defenses"],["subsection","Wire Lifting"]],"subsections":[],"title":"Wire Lifting"},{"cite_extract_rate":0,"cites":[],"content":"The main goal of Split Manufacturing -- to hide sensitive information from untrusted foundries -- is compromised once we start to consider more regular structures such as memory. Even without knowing where all the routing goes to, an attacker can easily identify regular structures just by looking at the FEOL layout, possibly leading to easier attacks. Mitigating attacks towards regular structures could be done by obfuscating those structures in such a way that they become indistinguishable. In this section, we discuss works that propose layout obfuscation techniques to be used in a Split Manufacturing context.\nDuring the development of a modern IC, third-party IPs are sought to close a technological gap or to minimize time-to-market. IPs are typically categorized as soft and hard IPs: soft IPs typically come in code form, giving the customer flexibility to modify the IP such that it meets a given specification during synthesis. Therefore, soft IPs do not present a direct challenge for a Split Manufacturing design flow. Perhaps, and on a very specific scenario, a given IP can facilitate a proximity attack because it promotes certain library cells over others (i.e., it leads to a biased composition).\nOn the other hand, hard IPs are completely designed by the vendor and are technology dependent. In some instances, the vendor only provides an abstract of the IP; the customer then has to rely on the foundry to replace the abstract by the actual layout. Thus, splitting a hard IP is not trivial. Additional information is needed to be provided by the vendor, which is not guaranteed to be provided, making the IP completely incompatible with Split Manufacturing. Even when the customer holds the entirety of the IP layout, differences between the FEOL foundry and the BEOL foundry could make the IP no longer compliant and therefore virtually useless. Furthermore, defense techniques cannot be applied due to the lack of information or lack of feasibility. Hard IPs, such as embedded memories and specialized analog circuits, have been heavily optimized for maximum compactness, performance and yield. In today's IP market, there is still little concern with security in general, so it is not conceivable that any vendors will start to offer split IP any time soon.\nThe security of hard IPs in a Split Manufacturing context was first analyzed by Vaidyanathan \\etal . A recognition attack flow was proposed for this purpose. An attacker holding the FEOL layer starts his attack by isolating a target embedded memory or analog hard IP. Since the targeted hard IP has a high probability of being constructed by compilation of leaf-cells, layout pattern recognition software can be used for trivial leaf-cell identification. After recognizing all the leaf-cells, the attacker attempts to infer the missing BEOL connections. Using proximity hints together with the knowledge about the regularized structure, the connections have a high likelihood to be guessed correctly. Demonstrated in , embedded memories, such as SRAM, are susceptible to the proposed recognition attack. Defending against recognition attacks can be achieved by means of obfuscation. According to , SRAM IPs can be obfuscated by the following methods:\n \\begin{itemize}\n \\item Randomization of periphery cells, thus avoiding predictable connections. \n \\item Minimization of regularized topologies used for peripheral circuits such as pre-decoders, word line decoders, sense amplifiers, etc.\n \\item Adding non-standard application-specific functions to improve obfuscation and performance.\n \\end{itemize}\nA synthesis framework is proposed by to obfuscate SRAM IPs. Referred as application-specific SRAM, the methodology synthesizes SRAMs using augmented bitcell arrays and standard cell logic IP (instead of using leaf-cells). Such synthesis, when compared with conventional SRAM compilation, accomplishes all the three obfuscation goals described above while still providing similar performance. \nAnalog hard IPs are also vulnerable to recognition attacks. In contrast with embedded memories (that are often compiled), analog hard IPs are mostly hand designed to cater for a challenging specification or interface. Even when such degree of customization is employed, the majority of the design is done utilizing leaf-cells (e.g., current mirrors, matched arrays, etc.). Thus, disclosing important information that could be used as leverage for recognition attacks. In , two methods are proposed to defend analog hard IPs against such attacks:\n \\begin{itemize}\n \\item Obfuscation of analog leaf-cells.\n \\item Use of diverse topologies and architectures that enable obfuscation and efficiency. \n \\end{itemize}\nNext, let us discuss the techniques utilized in order to achieve the goals listed above. First, adding camouflaging dummy transistors in empty spaces can turn leaf-cells indistinguishable. Second, regularizing transistor widths, which allows transistor with different channel lengths to abut each other, thereby obscuring boundaries across different sized transistors. Third, utilizing the same idea behind wire-lifting, routing blockages can be inserted between transistors below the split layer. Such routing scheme would make it difficult to infer the missing BEOL connections, virtually in the same way as it does for a standard-cell based design. \nTo demonstrate the feasibility and efficacy of their proposed approaches, the authors of designed and fabricated test chips in 130nm technology. For comparison, the same designs were Split Manufactured and conventionally manufactured. Split Manufacturing used Global Foundries Singapore as the untrusted foundry and IBM Burlington as the trusted foundry. Conventional manufacturing was entirely done in Global Foundries Singapore. The first reported design is a smart SRAM that targets an imaging application. Two implementations of a parallel 2x2 access 1Kb SRAM were demonstrated. For conventional manufacturing, the SRAMs were traditionally implemented, and for Split Manufacturing, the SRAMs were implemented using their smart synthesis approach. For their measurements, 10 chips were used to demonstrate the feasibility regarding PPA. Area reported for the split manufactured samples was 75\\% of the conventional approach, and, while the power consumption was 88\\%. Performance was the same between conventional and split manufactured, i.e., both could work with the same clock frequency. The PPA advantage of Split Manufacturing reported in is not from the manufacturing itself. This advantage is from their smart memory synthesis approach, that was not applied on the conventional manufacturing samples.\nThe second demonstrated design is a DAC with statistical element selection. The test chip contains a high resolution 15-bit current steering DAC. Only a description of the results is presented, where the authors claim there are tiny measurements differences between the performance of the conventional and the split manufactured, emphasizing that the differences are within measurement noise.\nAn attacker trying to reverse engineer a split IC will try to recover the maximum number of connections as possible, while minimizing the Time To Evaluate (TTE), i.e., the amount of time needed to reverse engineering the IC. For Jagasivamani \\etal , the goal of a designer seeking to secure his design is to create an IC with a high TTE while being cost-effective regarding design effort and PPA overheads. If the TTE is high enough, an adversary would be discouraged from reverse engineering the IC. To achieve this goal, proposed obfuscation methods that do not require any modifications to standard cells nor the implementation of any specialized cell. \nFour techniques are proposed by for layout obfuscation, (1) limited standard-cell library, (2) smart-dummy cell insertion, (3) isomorphic cells and (4) non-optimal cell placement. Along with the techniques, a set of metrics is presented to help assess the obfuscation level of a design. \\textit{Neighbor connectedness}, a measure of how interconnected cells are to their respective neighbors, i.e., how much proximity information is exposed to the attacker. For a specific cell, this metric is computed as how many connections that cell has for a given radius around it. \\textit{Standard-cell composition bias}, a metric that addresses the effort required for composition analysis of a design. The bias signature could leave information of the function of the cell. Thus, this metric measure how skewed a design is according to a specific bias cell. In , they utilized three types of bias cells for this analysis: XOR-type, flip-flop type, and adder type of cells. \\textit{Cell-level obfuscation}, a metric that measures the percentage of standard-cells that have been obfuscated. \\textit{Entropy}, which is similar to the concept of mutual information previously discussed. \nTechnique (1) aims to achieve obfuscation by reducing the use of complex cells and instead favor only simple cells to compose the design (i.e., to prefer single stage cells over complex multi-stage cells). Removing specialized complex cells could obfuscate functional signatures due to the larger granularity that is employed to construct the cell. However, since the functionality of complex gates will have to be reconstructed through basic cells, a heavy PPA overhead is likely to occur when applying this technique. Technique (2) aims to obfuscate composition analysis by adding dummy cells in such a way that a neutral bias composition is achieved. Dummy cells are inserted as spare cells \\footnote{Spare cells are extra logic usually inserted during physical synthesis. These cells are used when an engineering change order (ECO) is required, such that small tweaks to the circuit logic can be performed with minimal changes to placement and routing.}, focusing solely on obfuscating the composition analysis. Technique (3) obfuscates the layout by regularizing the shapes of the cells in a library. All layouts of logic standard cells are made FEOL-identical such that the overall circuit layout appears to be a sea of gates. The functionality of the cells is defined later by the BEOL connections. Thus, the true functionally of the cell is hidden at the BEOL, making cell-level identification harder. Technique (4) employs the same strategy from placement perturbation discussed before. \nFor their experimental study, the authors of made use of a multiplier block with a high number of adder cells and a crypto-like circuit. Experiments were separated into limited library and smart dummy insertion. Results are shown as a percentage relative to the baseline circuit, i.e., without any protection approach applied. Neighbor connectedness (\\%) for a radius $\\leq 25nm$ decreased substantially for both test cases and circuits (for more information see ). Overhead results are shown in Table~\\ref{tab:over_javs}, where the figures presented are normalized with respect to the baseline circuit.\n \\begin{table}[htb]\n \\centering\n \\caption{Impact on Performance from the Defense Approaches of .}\n \\begin{tabular}{cccc}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Benchmark} & \\textbf{Metric} & \\textbf{Limited Library} & S\\textbf{mart Dummy} \\vspace{5pt}\\\\\n \\hline\n \\multirow{2}{*}{ mult24} &\\cellcolor[gray]{0.9} Area & \\cellcolor[gray]{0.9} 94.9\\% & \\cellcolor[gray]{0.9} 72.6\\% \\\\\n & Timing Slack & -64.8\\% & 3.4\\% \\\\\n \\multirow{2}{*}{a5/1} & \\cellcolor[gray]{0.9} Area & \\cellcolor[gray]{0.9} 69.8\\% & \\cellcolor[gray]{0.9} 69.4\\% \\\\\n & Timing Slack & -27.2\\% & -1.2\\% \\\\ \n \\hline \\hline\n \\end{tabular}\n \\label{tab:over_javs}\n \\end{table}\nUtilizing exactly the same concepts and metrics described in , Otero \\etal proposed a ``trusted split-foundry toolflow'' based on cellTK . The concept of the cellTK-based flow is to have on-demand cell generation from a transistor-level netlist. This is heavily in contrast with a traditional ASIC flow that relies on a predefined (and thoroughly characterized) cell library. Leveraging cellTK, proposed an extension referred as split-cellTK. This extension can generate multiple physical layouts for each unique cell without modifying the circuit topology, which is then used to implement obfuscation strategies. Two strategies are proposed, referred as Uniform and Random. The Uniform strategy tries to standardize the size and spacing between cells by inserting dummy transistors to equalize the number of nMOS and pMOS devices and, after the cell placement, dummy cells are inserted in empty gaps. A Random strategy is also proposed in order to reduce the overheads introduced by the Uniform strategy. Instead of deliberately standardizing the size and spacing between cells, a specific number of empty spaces is chosen for these tasks. A more in-depth explanation about their strategies is beyond the scope of this work because they are closely related to cellTK itself. However, their goals and evaluations are the same as in . For their experimental study, they utilized the island-style asynchronous FPGA developed in . A test chip was Split Manufactured in 65nm and the design was synthesized with a cellTK-based approach, i.e., without any defense strategy. Their defense strategies were evaluated only by simulation. The performance results for the baseline, Uniform, and Random strategies, when applied to obfuscate an adder, are shown in Table~\\ref{tab:over_celltk}. Trustworthiness results are given in terms of neighbor connectedness and, for all implementations discussed, neighbor connectedness results were significantly smaller than the results reported in .\n \\begin{table}[htb]\n \\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Performance Results Reported by to Obfuscate an Adder Circuit.}\n \\begin{tabular}{lllll}\n \\hline \\hline \\\\ [-1.5ex]\n \\parbox{1cm}{\\textbf{Technique}} & \\parbox{0.6cm}{\\textbf{Area}\\\\ ($\\mathbf{\\mu m^2}$)} & \\parbox{0.6cm}{\\textbf{Power (mW)}} & \\parbox{0.6cm}{\\textbf{Energy (pJ)}} & \\parbox{0.7cm}{\\textbf{Perf. (MHz)}} \\vspace{5pt} \\\\\n \\hline\n Baseline & 462 & 0.146 & 0.257 & 568 \\\\\n Uniform & 717 & 0.149 & 0.307 & 486 \\\\\n Random & 760 & 0.164 & 0.303 & 542 \\\\\n \\hline\n \\hline\n \\end{tabular}\n \\label{tab:over_celltk}\n \\end{table}\nIn the context of obfuscation, but also generally for Split Manufacturing, a higher level of security is achieved when the chosen layer to perform the split is the lowest possible. Xiao \\etal pointed out that splitting at lower metal layers could increase the cost to manufacture the IC; it is argued that the FEOL-BEOL integration process must be more `precise' for correct alignment. Thus, a closer technology match between the trusted and untrusted foundries is required. As we previously argued, if the goal is to make use of the best silicon available from an untrusted foundry, the implication is that the trusted foundry cannot provide a legacy technology, but perhaps a mature yet still relevant technology is sufficient. In , a methodology for obfuscating the layout is proposed for split at M3 or higher, meanwhile keeping the cost as low as possible and at the same time providing a high level of security. Their strategy is similar to the insertion of dummy cells; however, functional cells are inserted instead. Referred as obfuscated built-in self-authentication (OBISA) cells, the inserted functional cells are connected together to form a circuit. As the circuit is connected to the original circuit it is trying to protect, they claim this fact makes it extremely difficult for an attacker to separate the OBISA design from the original design. The idea behind OBISA is to obfuscate the layout by hindering neighbor connectedness analysis and standard-cell composition bias analysis while also perturbing the proximity between gates. As illustrated in Figure~\\ref{fig:obisa}, two additional functional cells, O1 and O2, were placed between gates G2 and G3, and, G3 and G4, respectively. The insertion of these additional functional cells could deceive proximity attacks, assuming that the EDA tool would place the gates between OBISA cells farther apart than in the original circuit.\nThe proposed OBISA circuitry has two operating modes: functional and authentication. During functional mode, the OBISA circuitry stays idle and incoming signals and clock are gated/blocked. Thus, the original circuit is not affected by OBISA operating as it should. As the name suggests, when in authentication mode, OBISA is used to verify the trustworthiness of the manufactured IC (in the field). The specifics of the authentication are beyond the scope of this work and will not be discussed. The insertion of OBISA cells follows a similar strategy of dummy cell insertion as discussed in . The connections of the inserted cells are done in a way to promote the testability of the OBISA circuit and increase the obfuscation strength. Their approaches were evaluated using benchmark circuits from OpenCores. Results for the smallest and largest circuits are shown in Table~\\ref{tab:obisa_res}. \n \\begin{table}[htb]\n \\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Implementation Results from .}\n \\begin{tabular}{lllll}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Benchmark} & \\parbox{0.8cm}{\\textbf{Gate count}} & \\parbox{1.5cm}{\\textbf{OBISA cell count}} & \\textbf{Total nets} & \\textbf{Nets} $\\mathbf{\\geq M4}$ \\vspace{5pt}\\\\\n \\hline\n DES3 & 1559 & 158 & 1799 & 127 \\\\\n DES\\_perf & 49517 & 2090 & 49951 & 1343 \\\\\n \\hline \\hline\n \\end{tabular}\n \\label{tab:obisa_res}\n \\end{table}\n \\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\linewidth]{obisa_cir.pdf}\n{Circuit representation with OBISA cells (square cells) inserted (adapted from ). \\label{fig:obisa}}\nAnother study using look-alike cells is reported by Masoud \\etal where the goal remains to make the attacker unable to distinguish cells and their inputs/outputs, thus mitigating attacks to some degree. In this study, two types of search algorithms are proposed to replace cells for isomorphic cells. In contrast with where all cells are replaced, in , only cells with high impact on the security are replaced. Thus, the overhead introduced by cell replacement can be controlled (i.e., a trade-off is established). The proposed algorithms are based on `gate normalization', whereby truth tables of cells are analyzed in order to balance the occurrence of 0s and 1s (e.g., XOR and XNOR gates are normalized by definition). An analysis is made by replacing existing gates by XORs and comparing the deviation from the original circuit. If the deviation is larger than a given deviation threshold, the gates are effectively replaced. \nA novel layout obfuscation framework is proposed by Li \\etal which builds on the wire lifting concept of . According to the authors, wire lifting alone is not enough to secure a design. If an attacker can tell the functionality of a specific gate that had its wires lifted, the security is already compromised. To address this problem, a framework that considers dummy cells and wire insertion simultaneously with wire lifting is proposed. As in , threat model II was used. The proposed framework makes use of mixed-integer linear programming (MILP) formulation for the FEOL layer generation and a Lagragian relaxation (LR) algorithm to improve scalability. The generation of the new FEOL layout considers three operations: wire-lifting, dummy cell insertion, and dummy wire insertion. Dummy wire insertion is done only on dummy cells; thus, the original functionality of the circuit is guaranteed to remain and floating pins are avoided. Utilizing a graph representation, they re-formulate the security metric to accommodate dummy cell and dummy wire insertion. Since the original graph isomorphic relationship is lost when new nodes are inserted, a new approach has to be used to formalize the relationship between the original and the new FEOL; this concept is denoted as k-isomorphism and the associated security analysis is denoted as \\textit{k-security}. In their experimental study, TrustHub trojan insertion methods are used to select the nodes for protection. They used ISCAS\\textquotesingle85 benchmark circuits together with functional units (shifter, alu, and div) from the OpenSPARC T1 processor . Comparison between MILP and LR algorithms are done for several \\textit{k-security} levels, and the results are given in terms of area overhead (AO) and wire-length overhead (WLO). The results for a few of the security levels are shown in Table~\\ref{tab:li_milplr}. \n \\begin{table}[htb]\n \\rowcolors{2}{gray!25}{white}\n \\centering\n \\caption{Comparison Between MILP and LR Algorithms for the c4232 circuit .}\n \\begin{tabular}{clll}\n \\hline \\hline \\\\ [-1.5ex]\n \\textbf{Security Level } & \\textbf{Algorithm} & \\textbf{AO(\\%)} & \\textbf{WLO(\\%)} \\vspace{5pt}\\\\\n \\hline\n 15 & MILP & 18& 180\\\\\n 20 & MILP & 41 & 220\\\\\n 25 & MILP & 58 & 295\\\\\n 15 & LR & 18& 200\\\\\n 20 & LR & 40 & 230\\\\\n 25 & LR & 60 & 305\\\\\n \\hline \\hline\n \\end{tabular}\n \\label{tab:li_milplr}\n \\end{table}","id":"4032829b-2e93-4b12-822b-7b2a6e653dbc","level":"subsection","origin_cites_number":12,"parent_id":"0381ac67-f861-41a5-94dd-d57b1d10cc30","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Split Manufacturing Defenses"],["subsection","Layout Obfuscation"]],"subsections":[],"title":"Layout Obfuscation"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:trends}\nDespite our effort to present the results of the many studied papers in the most fair way possible, it is clear that the hardware security community lacks a \\textit{unified benchmark suite} and/or a \\textit{common criteria} for assessing results. Often, researchers make use of benchmark suites that are popular in the Test community but have no real applicability in security. For instance, the ISCAS\\textquotesingle85 suite has no crypto cores in it, which are the bread and butter of the research in the area. Furthermore, we believe the community would largely benefit from using circuits that better represent IC design practices of this decade where IPs often have millions of gates and ICs have billions of transistors.\nWhile the lack of a common criteria is an issue for the academic community, the lack of an industry-supported path for Split Manufacturing is even more troubling. Today, more than ever, foundries compete for the title of `best silicon' and rarely engage in cross-foundry cooperation. Efforts of the past, such as the now defunct Common Platform of IBM, Samsung and GF, could have been a catalyzer for the adoption of Split Manufacturing. Without such collaboration, it is hard to foresee a future where the technique will gain traction again. Furthermore, the study of DFM-related implications of the technique is really cumbered by the fact that we cannot measure yield from massively produced Split Manufactured chips. \nWe have discussed in details how many attacks leverage heuristics and hints left behind by the EDA tools. Many of these hints are very logical and can be appreciated, even graphically, as we illustrated in Figure~\\ref{fig:prox_sa}. It is entirely possible that machine learning approaches can detect subtle biases in the tools that are not easy to appreciate graphically. There is no consolidated knowledge of what these biases are and to which extent machine learning is effective in detecting them. This avenue of research is certainly interesting and we believe it will be the target of many papers in the near future.\n \\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\linewidth]{pie_chart.pdf}\n{Techniques validated in silicon among presented works. \\label{fig:pizza}}\nIt is also worth discussing the attack models that have been proposed so far. We have previously highlighted how strong Threat model II is, but the fact of the matter is that defining the threat model is a complicated exercise in which we seek to establish what are the capabilities of the attacker. By definition, formalizing the capabilities of an attacker requires understanding his motivations, technical proficiency, and availability of resources. If the attacker is underestimated, useless defense strategies can be devised and assumed to be effective. If the attacker is overestimated, convoluted defense strategies might be employed, leading to unnecessary PPA overheads. This is a challenge for Split Manufacturing and many other techniques that promote obfuscation.\nAnother topic that has led to no consensus is whether an attacker can make use of a partially recovered netlist. For instance, let us assume a design that instantiates the same block multiple times. If one of the blocks is correctly recovered, perhaps a cursory inspection of the structure will allow the attacker to recover all other instances of the same block. The same line of thinking can be applied to datapaths and some cryptographic structures that are regular in nature. In a sense, an analysis of the functionality of the recovered netlist could be combined with existing attacks for further improvement of correctly guessed connections.\nWe note that many of the works studied in this survey have not actually demonstrated their approach in silicon. This fact is summarized in Figure~\\ref{fig:pizza}. As a community effort, we should strive to validate our approaches in silicon as often as possible. However, as discussed before, finding two foundries willing to diverge from their established practices could be next to impossible. This is likely the main reason that such small percentage of the works herein reported have validated their techniques in silicon.","id":"5c7f1b9c-d130-400e-ae7c-f7e824d126c3","level":"section","origin_cites_number":0,"parent_id":"4df5dfa6-5506-4133-aae0-08641ac96086","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Future trends and challenges"]],"subsections":[],"title":"Future trends and challenges"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclusion} \nOur findings showed a big disparity on how the Split Manufacturing technique is approached among the surveyed studies. A variety of benchmark suites and metrics were used for evaluation, making direct comparisons between studies very difficult -- and, in some cases, impossible. In spite of that, we were able to classify the studies, clearly demonstrating the many interpretations of the technique, its attacks, and defenses. Our belief is that this survey assesses the most significant studies about Split Manufacturing as we focused on papers that appear on highly-regarded venues. Results gathered from the surveyed studies were compiled such that main features, metrics, and performance results are available. Regarding the results themselves, these are presented in such manner to illustrate the present state of the technique. Therefore, this work can be very helpful for future researchers to contextualize their own techniques for augmenting Split Manufacturing.\nOverall, the security of Split Manufacturing is still under debate. Some studies conclude that the technique is indeed secure, and others that it is not. However, these conclusions are reached for different scenarios, i.e., using different benchmark circuits and set of metrics. Creating a unified benchmark suite suitable for Split Manufacturing evaluation, along with a unified set of metrics to quantify/qualify its performance, could facilitate the discussion about its security. In addition, increasing the number of demonstrations in silicon could also help with evaluation and adoption issues related to Split Manufacturing.\n\\bibliographystyle{ieeetr}\n\\bibliography{split_refs}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{perez_t_d.jpg}}]{Tiago D. Perez} received the M.S. degree in electric engineering from the University of Campinas, S\\~ao Paulo, Brazil, in 2019. He is currently pursuing a Ph.D. degree at Tallinn University of Technology (TalTech), Tallinn, Estonia. \nFrom 2014 to 2019, he was a Digital Designer Engineer with Eldorado Research Institute, S\\~ao Paulo, Brazil. His fields of work include digital signal processing, telecommunication systems and IC implementation. His current research interests include the study of hardware security from the point of view of digital circuit design and IC implementation.\n\\end{IEEEbiography}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{pagliarini_s.png}}]{Samuel Pagliarini}\n(M'14) received the PhD degree from Telecom ParisTech, Paris, France, in 2013. \nHe has held research positions with the University of Bristol, Bristol, UK, and with Carnegie Mellon University, Pittsburgh, PA, USA. He is currently a Professor of Hardware Security with Tallinn University of Technology (TalTech) in Tallinn, Estonia where he leads the Centre for Hardware Security. His current research interests include many facets of digital circuit design, with a focus on circuit reliability, dependability, and hardware trustworthiness.\n\\end{IEEEbiography}\n\\EOD\n\\end{document}","id":"347259ba-3c12-43d7-b6a0-2402a507ae62","level":"section","origin_cites_number":0,"parent_id":"4df5dfa6-5506-4133-aae0-08641ac96086","prefix_titles":[["title","A Survey on Split Manufacturing: Attacks, Defenses, and Challenges"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4df5dfa6-5506-4133-aae0-08641ac96086\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ]\n ],\n \"subsections\": [\n \"bcde8dcf-d020-417f-80e7-74fb7e142a07\",\n \"c3cea202-e3a7-4e28-a205-e6374a147781\",\n \"63582e79-19ab-44ed-8225-2fdf07c8ddcd\",\n \"0381ac67-f861-41a5-94dd-d57b1d10cc30\",\n \"5c7f1b9c-d130-400e-ae7c-f7e824d126c3\",\n \"347259ba-3c12-43d7-b6a0-2402a507ae62\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:introduction}\\nCounterfeiting and intellectual property (IP) infringement are growing problems in several industries, including the electronics sector. In Europe, for instance, seizures of counterfeit electronics products increased by almost 30\\\\% when comparing the 2014-2016 to the 2011-2013 period . Legitimate electronics companies reported about \\\\$100 billion in sales losses every year because of counterfeiting .\\n \\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\\\linewidth]{counter_taxiderm.pdf}\\n{Taxonomy of counterfeit electronics (adapted from ). \\\\label{fig:cf_tax}}\\nAs electronic systems are being increasingly deployed in critical infrastructure, counterfeit and maliciously modified integrated circuits (ICs) have become a major concern. The globalized nature of the IC supply chain contributes to the problem as we lack the means to assess the trustworthiness of the design and fabrication of ICs. It is conceivable -- if not likely -- that a fault in a low-quality counterfeit IC (or even a maliciously modified IC) will effectively disrupt critical infrastructure with grave consequences. Therefore, hardware security has gained more attention in the past decades, emerging as a relevant research topic.\\nAs the IC supply chain has become more globalized, ensuring the integrity and trustworthiness of ICs becomes more challenging . When a modern IC is conceived, the probability that all involved parties are trusted is, in practice, close to zero. The process of conceiving an IC can be broken down into three major steps: design, manufacturing, and validation. \\\\textit{Designing} an IC involves arranging blocks and their interconnections. Some blocks are in-house developed, while some are third-party IPs. Finally, a layout is generated by instantiating libraries that might also be in-house developed or provided by third parties. The resulting layout is then sent to a foundry for \\\\textit{manufacturing}. The process of \\\\textit{validation} requires test for physical defects as well as verification of packaged parts for correct functionality. Both test and packaging facilities may be untrusted, as these efforts are often offshored. Thus, in order to produce an IC, sensitive information almost inevitably is exposed to untrusted parties. Today's reality is that ICs are vulnerable to many hardware-based threats, including insertion of hardware trojans, IP piracy, IC overbuilding, reverse engineering, side-channel attacks, and counterfeiting. These threats are discussed in details in . \\nHardware trojans, in particular, are malicious modifications to an IC, where attackers insert circuitry (or modify the existing logic) for their own malicious purposes. This type of attack is (typically) mounted during manufacturing, as the foundry holds the entire layout and can easily identify critical locations for trojan insertion. Third-party IPs can also contain trojans/backdoors that may contain hidden functionalities, and which can be used to access restricted parts of the design and/or expose data that would otherwise be unknown to the adversary.\\nIP piracy and IC overbuilding are, essentially, illegal ownership claims of different degrees. As said before, during designing an IC, third-party and in-house developed IPs are utilized. The untrusted foundry (or a rogue employee of it) can copy one of those IPs without the owner's authorization. Similarly, malicious foundries can manufacture a surplus of ICs (overbuilding) without the owner's knowledge, and sell these parts in the grey market.\\nReversing engineering of ICs has been extensively demonstrated in the specialized literature . An attacker can identify the technology node and underlying components (memory, analog, and standard cells), from which a gate-level netlist can be extracted and even a high-level abstraction can be inferred . Reverse engineering can be effortlessly executed during manufacturing, as the foundry holds the entire layout and most likely promptly recognizes some of the IP as well. After fabrication, -- when ICs are already packaged and deployed -- reverse engineering is more laborious but can still be executed by a knowledgeable adversary.\\nAccording to , counterfeit components are classified into seven distinct categories, as illustrated in Figure~\\\\ref{fig:cf_tax}. Recycled, remarked, out-of-spec/defective, and forged documentation are intrinsic after-market problems, where products are offered by parties other than the original component manufacturer or authorized vendors. These cases are highlighted in red. On the other hand, overproducing, cloning, and tampering are problems faced during the designing and/or fabrication of ICs. These cases are highlighted in blue. For this reason, in this paper, we will focus on these threats. It is important to realize that these threats could be avoided if a trusted fabrication scheme was in place. However, the escalating cost and complexity of semiconductor manufacturing on advanced technologies made owning an advanced foundry unfeasible for design companies, which now have the tendency to adopt the fabless business model . Consequentially, outsourcing of the manufacturing exposes their entire layout to untrusted foundries, leaving their designs vulnerable to malicious attacks.\\nWhile many \\\\textit{ad hoc} techniques have been proposed to individually combat these threats, very few solutions directly address the lack of trust in the fabrication process. Split Manufacturing stands out from other techniques as it promotes a hybrid solution between trusted and untrusted fabrication. The technique was first pitched to DARPA circa 2006 in a white paper authored by Carnegie Mellon and Stanford universities. Later, it was picked by IARPA which then launched the Trusted IC program that successfully stewarded much of the research in the area and led to this survey. \\nIn Split Manufacturing, the key concept is to \\\\textit{split} the circuit in two distinct parts before manufacturing, one containing the transistors and some routing wires, and the other containing the remaining routing wires. These parts are then fabricated in different foundries. The anatomy of an IC is illustrated in Figure~\\\\ref{fig:ic_anatomy}, containing two set of layers, the bottom layer where the transistors are built, called Front end of the Line (FEOL), and the top layer where the metal layers are built for routing purposes, called Back end of Line (BEOL). The metal layers are referred as MX, where X is the level of the layer. M1 is the lowest layer at the bottom of the stack. Connections between metals are made by vias referred as VX, following the same naming scheme for metals. In Split Manufacturing, the FEOL is first manufactured in a high-end foundry, and later the BEOL is stacked on top of it by a second (and possibly low-end) foundry. This process requires electrical, mechanical, and/or optical alignment techniques to ensure the connections between them. Additionally, FEOL and BEOL technologies have to be compliant with each other regarding the rules for metal/via dimensions where the split is to be performed. The split can be performed at the lowest metal layer (M1) or at higher layers, for which trade-offs are established between attained security and overheads. If the BEOL and FEOL technologies are vastly different from one another, Split Manufacturing may incur heavy overheads.\\nIn this work, the focus is on the Split Manufacturing technique. As described above, Split Manufacturing can tackle threats that occur during the fabrication. It avoids overproduction, reverse engineering (to some extent) and unwanted modifications, limiting the capability of attackers. In Section II, we provide a background and more in-depth explanation of the technique. We address security threats in Section \\\\ref{sec:bck_gnd}, demonstrating the potential vulnerabilities found in split circuits and describing the state-of-the-art attacks proposed until the present day. In Section \\\\ref{sec:defenses}, the security of split circuits is discussed, showing how it can be improved using enhancement techniques. Future trends and lessons learnt are discussed in Section \\\\ref{sec:trends}. Finally, our conclusions are presented in Section \\\\ref{sec:conclusion}.\\n \\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=.9\\\\linewidth]{cross_section.pdf}\\n{Anatomy of an integrated circuit (adapted from ). \\\\label{fig:ic_anatomy}}\",\n \"id\": \"bcde8dcf-d020-417f-80e7-74fb7e142a07\",\n \"level\": \"section\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"4df5dfa6-5506-4133-aae0-08641ac96086\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:bck_gnd}\\nAs mentioned before, in order to have access to advanced technologies, many design companies have to outsource their IC manufacturing to untrusted high-end foundries. Protecting their designs against threats that may occur during manufacturing is a concern. Designs can be protected by applying the Split Manufacturing technique, thus combating all threats highlighted in blue in Figure~\\\\ref{fig:cf_tax}.\\nSplit Manufacturing protects a design by hiding sensitive data from the untrusted foundry. This is achieved by splitting the IC into two parts before manufacturing, a horizontal cut that breaks the circuit into one part containing the transistors and some (local) routing wires, and another containing only routing wires. These parts are termed FEOL and BEOL. \\nAs the FEOL and the BEOL of an IC are built sequentially, first FEOL and then BEOL, this characteristic enables the Split Manufacturing technique. Since the FEOL contains the transistors and possibly a few of the lowest ultra-thin metal layers -- the most complex parts of an CMOS process --, it is logical to seek to use a high-end foundry for its manufacturing, even if said foundry is not trusted. Completing the IC can then be done in a trusted low-end foundry, where the BEOL is stacked on top of the FEOL. Split Manufacturing was successfully demonstrated in , where designs were manufactured with \\\\textasciitilde 0\\\\% of faults, and are reported to present a performance overhead of about 5\\\\%. Therefore, the technique is, in principle, feasible and available for design companies to make use of, such that they can utilize advanced foundries without fully exposing their layouts during manufacturing.\\nHowever, there are many caveats to Split Manufacturing. The technique can be successfully applied only if the technologies used to build the FEOL and BEOL are ``compatible''. In theory, a layout can be split at any layer if the chosen layer presents a good interface between FEOL and BEOL. However, since advanced technologies utilize the dual-damascene fabrication process, the layout can only be split on metal layers . Thus, the FEOL cannot terminate in a via layer. The dual-damascene process is characterized by patterning the vias and trenches in such a way that the metal deposition fills both at the same time, i.e., via-metal pairs (e.g., V1 and M2) must mandatorily be built by the same foundry. \\nTwo technologies are said to be compatible with each other if there is a way for a BEOL via to land on the FEOL uppermost layer while respecting all design rule checks (DRCs) of both technologies. DRCs are used to guarantee the manufacturability and functionality of an IC, and are defined with respected to the characteristics of the materials utilized and to tolerance ranges of the manufacturing processes (e.g., polishing, patterning, and deposition). These rules encompass minimum enclosure, width, spacing, and density checks. Modern technologies have several options for via shapes -- as long as one via shape is valid, the technologies are compatible for Split Manufacturing purposes. However, in practice, to keep the overhead of the technique under control, an array of via shapes must be feasible, thus providing the physical synthesis with a rich selection for both power and signal routing. \\nAccording to , compatibility between two technologies can be generalized by enclosure rules as in Eq.~\\\\ref{eq:drc_1}, where MW.U.x is the minimum width of Mx on untrusted foundry, VW.T.x is the minimum width of Vx on trusted foundry and EN.T.x is the minimum enclosure on trusted foundry. These rules are illustrated in Figure \\\\ref{fig:drc_mdim}, where the left side of the image portrays a cross section view and the right side shows the top view. According to Figure \\\\ref{fig:drc_mdim}, the minimum enclosure width, Mx.EX.Vx, must be compatible between the two foundries. In modern technologies, Equation~\\\\ref{eq:drc_1} is no longer sufficient as it does not capture the intricate rules for vias and line endings (enclosure from 1 side, 2 sides, 3 sides, T-shaped/hammerheads, etc.).\\n \\\\begin{equation} \\\\label{eq:drc_1}\\n MW.U.x \\\\geq VW.T.x + (2 EN.T.x)\\n \\\\end{equation}\\n\\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\\\linewidth]{metal_view.pdf}\\n{Compatibility rules between FEOL and BEOL (adapted from ). \\\\label{fig:drc_mdim}}\\nSplit Manufacturing also presents challenges on the design flow front, which is illustrated in Figure~\\\\ref{fig:split_sm_flow}. An in-house team designs the circuit, from RTL to layout. Most likely, the layout contains IPs obtained from third parties. Depending on the metal layer where the layout is to be split, it may affect existing IP. Logic and memory IP may use higher metal layers -- memories typically require 4 to 5 metal layers, while standard cells typically require 2 metal layers --, limiting where the split can be done. Standard cells and memories have to be re-designed if they use metal layers that will be split, a grave challenge that may render Split Manufacturing significantly harder to execute.\\nStill referring to Figure~\\\\ref{fig:split_sm_flow}, the FEOL and BEOL are generated using a hybrid process design kit (PDK), and then later split to be manufactured. After splitting the layout correctly, the FEOL is first manufactured in a high-end foundry, and later the BEOL is stacked on top of it by a second (and possibly low-end) foundry. \\n\\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\\\linewidth]{sm_flow_dia.pdf}\\n{Split Manufacturing Design Flow. \\\\label{fig:split_sm_flow}}\\nEven by splitting the layout, it is often argued that the FEOL exposes enough information to be exploited. Attacks towards the FEOL can effectively retrieve the BEOL connections by making educated guesses. The efficiency of the guessing process is inherently linked to the threat model assumed, which determines the information the attacker possesses to begin with. The literature describes two distinct threat models:\\n \\\\begin{itemize}\\n \\\\item \\\\textbf{Threat model I:} an attacker located at the untrusted foundry holds the FEOL layout and wants to retrieve the BEOL connections.\\n \\\\item \\\\textbf{Threat model II:} an attacker located at the untrusted foundry holds the entire gate-level netlist that is assumed to be provided by a malicious observer. The attacker here still holds the FEOL layout and wants to retrieve the BEOL connections. .\\n \\\\end{itemize}\\nIt is important to emphasize that the second threat model completely nullifies the security introduced by Split Manufacturing. Possessing the gate-level netlist makes reverse engineering the layout a trivial task, as if the attacker held the entire layout, not only the FEOL. Assuming the attacker has knowledge about the netlist challenges the integrity of the design company itself. If a rogue element exists inside the design company, other representations of the design could be equally stolen, such as the register-transfer level (RTL) code of the design, or even the entire layout (including the BEOL). It could be argued that this vulnerability is so severe that Split Manufacturing virtually stops making sense. For this reason, threat model I is the focus in this work. However, as our goal is to present a comprehensive survey, related works that use threat model II will be covered as well. \\nAssuming threat model I, an attacker already knowing all the layers that make up the FEOL, is now interested in retrieving the BEOL connections to recreate the full design (or as close as possible). The commonly used assumption is that attackers are powerful and work within the untrusted foundry in some capacity. Thus, the attackers have deep understanding about the technology. Extracting the (still incomplete) gate-level netlist from a layout is, therefore, a trivial task for the attacker.\\nMany approaches to retrieve the BEOL connectivity have been proposed, several of which are termed \\\\textit{proximity attacks} . Since EDA tools focus on optimizing power, performance, and area (PPA), the solution found by a placement algorithm (that uses heuristics internally) tends to place connected cells close to one another as this will, in turn, reduce area, wirelength, and delay. Therefore, finding the correct missing connections between FEOL and BEOL can be done by assessing input and output pins that are in proximity (thus, the name proximity attack). The more input and output pins to connect, the higher is the probability to make a wrong connectivity guess. Thus, a higher level of security is achieved by splitting the circuit at the lowest metal layer possible.\\nAs a promising technique to enhance the security of ICs in this era of fabless chip design, Split Manufacturing still faces some enormous challenges:\\n \\\\begin{itemize}\\n \\\\item[] \\\\textit{Logistical challenge:} Split Manufacturing is not presently incorporated into the IC supply chain. Finding foundries with compliant technologies that are willing to work with each other is not trivial.\\n \\\\item[] \\\\textit{Technological challenge:} even within compliant technologies, non-negligible overheads can be introduced if they are vastly different\\\\footnote{For a thorough discussion and silicon results on BEOL-related overheads, please refer to .}. In the worst-case scenario, it can make routing impossible. Thus, this fact narrows down the technology choices available and feasibility of certain layers as candidates for splitting.\\n \\\\item[] \\\\textit{Security challenge:} the attained security of straightforward Split Manufacturing is still under debate. Attacks towards the FEOL can be effective, where the hidden connections can be retrieved.\\n \\\\end{itemize}\\nFor the purpose of this survey, we categorized related works in the literature as attacks and defenses. In \\\\textit{attacks}, authors proposed new attack models or modifications of existing attacks in order to improve their effectiveness. In \\\\textit{defenses}, authors proposed new techniques to use together with Split Manufacturing in order to improve its security level.\",\n \"id\": \"c3cea202-e3a7-4e28-a205-e6374a147781\",\n \"level\": \"section\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"4df5dfa6-5506-4133-aae0-08641ac96086\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Split Manufacturing: Background\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Split Manufacturing: Background\"\n },\n {\n \"cite_extract_rate\": 0.130434782608695,\n \"cites\": [\n 7800,\n 7798,\n 7799\n ],\n \"content\": \"\\\\label{sec:sec_thrts}\\nThe Split Manufacturing technique was developed to protect ICs against threats related to manufacturing in potentially untrusted foundries. In practical terms, to split the layout means to hide some connections from the untrusted foundry. The security provided by Split Manufacturing is based on the fact that the attacker in the FEOL foundry cannot infer the missing BEOL connections. This assumption, however, was challenged by several works where authors proposed attack approaches that can potentially retrieve the missing connections with varying degrees of success. In the text that follows, we present works that proposed Split Manufacturing attacks. These attacks are discussed in chronological order as compiled in Table~\\\\ref{tab:atcks_desc}. For each studied attack, we reported the related threat model, attack type, novelty, and benchmark circuits used in experiments. Additionally, we reported the largest and average size of the circuits utilized in each work (measured in number of gates). The circuit size is an important metric when analyzing the Split Manufacturing technique because its effectiveness is often proportional to circuit size. In Table~\\\\ref{tab:atcks_effect}, we compiled results for when the smallest and largest studied circuits are under attack. Also, we reported the circuit size and, if defined, the split layer that the author selected to split the circuit at.\\nThe first reported attack is by Jeyavijayan \\\\etal~and is described in . In this work, the authors assumed that naive Split Manufacturing (i.e., splitting a layout without care for the connections) is inherently insecure. They introduced the concept of proximity attacks that exploits ``vulnerabilities'' introduced by EDA design tools. Since EDA tools focus on optimizing for PPA, the solution found by a placement algorithm tends to place logically connected cells close to one another so they become physically connected during routing. Therefore, the distance between output-input pairs can be used as a metric to recover the missing BEOL connections.\\n\\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.90\\\\linewidth]{partitions.pdf}\\n{Example of a partitioned circuit. \\\\label{fig:cir_part}}\\nDesigns are commonly partitioned during physical implementation, i.e., separated into small logical blocks with few connections between them. That way, the designer has total control of the floorplaning regarding the placement of blocks. This approach also allows for blocks to be implemented separately and later integrated, creating a sense of parallelism in the design flow which can reduce the overall time required for executing physical synthesis. Consider as an example the circuit illustrated in Figure~\\\\ref{fig:cir_part} which contains two partitions, A and B, each with 3 gates. Partition A has 3 input pins and one output pin, while partition B has three input pins and two output pins. The partitions have only one connection between them, connecting the output pin of gate G2 with one of the inputs of gate G3. Consider a target output pin from partition A $P_{x,A,out}$ and its corresponding candidate input pin from partition B $P_{x,B,in}$. During placement, the EDA tool will attempt to place the pin $P_{x,A,out}$ as close as possible to $P_{x,B,in}$, possibly closer than any other pin in partition B. Using this insight, an attacker may recover the missing connections in the FEOL layout, performing then a proximity attack. The authors argued that their proposed attack flow is successful due to being able to leverage the following ``hints'' provided by the EDA tools:\\n \\\\begin{itemize}\\n \\\\item[] \\\\textit{Hint 1 - Input-Output Relationship}: partition input pins are connected either to another partition output pin or to an input port of the IC (i.e., input to input connections are excluded from the search space).\\n \\\\item[] \\\\textit{Hint 2 - Unique Inputs per Partition}: input-output pins between partitions are connected by only one net. If a single partition output pin feeds more than one input pin, the fan-in and fan-out nodes are usually placed within the partitions (i.e., one-to-many connections are ruled out from the search space). \\n \\\\item[] \\\\textit{Hint 3 - Combinational Loops}: in general, only very specific structures are allowed to utilize combinational loops (e.g., ring oscillators). These structures are very easy to identify. In most cases, random logic does not contain combinational loops (i.e., connections that would lead to combinational loops can be excluded from the search space).\\n \\\\end{itemize}\\nAn attacker can correctly connect a target pin to a candidate pin by identifying the closest pin from a list of possible candidates. The list of possible candidates is created by observing the hints mentioned above. A possible candidate pin is an unassigned output pin of another partition and an unassigned input port of the design. Then, a minimum distance metric is used to connect the pins based on the previously discussed heuristic behavior of EDA tools.\\nIn Algorithm \\\\ref{alg:prox_jevs}, we described the proximity attack detailed in . The input to the algorithm is the FEOL layout, from which the information about unassigned input-output ports can be derived. The algorithm does not describe the specifics of how to derive a netlist from a layout. However, the complexity of this task is rather straightforward. It is assumed that the attacker possesses information about both the PDK and the standard cell library. In many cases, the untrusted foundry is the actual provider of both\\\\footnote{For the PDK, it is very natural that it is created by the untrusted foundry itself. For standard cell libraries, the cells might be designed by the foundry or by a third-party licensed by the foundry. In either case, the effort to revert a layout to a netlist remains trivial.}. From there, a layout in GDSII or OASIS format can be easily reverted to a netlist by any custom design EDA tool. \\n\\\\begin{algorithm}[h]\\n\\\\DontPrintSemicolon\\n \\\\KwInput{FEOL layers}\\n \\\\KwOutput{Netlist with BEOL connections}\\n Reverse engineer FEOL layers and obtain partitions;\\n \\\\While{Unassigned partition pins or ports exist}\\n {\\n Select arbitrary unassigned pin/port as a targetPin;\\n \\tListOfCandPins = BuildCandPinsList(targetPin);\\n \\tSelect candPin from ListOfCandPins that is closest to targetPin;\\n \\tConnect targetPin and candPin;\\n \\tUpdate netlist;\\n }\\n \\\\textbf{Return:} netlist\\n \\\\textbf{BuildCandPinsList}(targetPin)\\n \\\\KwInput{targetPin $P_{X,i,in}$}\\n \\\\KwOutput{CandPins for targetPin}\\n CandPins = Unassigned output pins of other partitions + unassigned input ports of the design;\\n \\\\For{\\\\textbf{each} $Pin_J \\\\in$ CandPins}\\n {\\n \\\\If{CombinationalLoop(targetPin, $Pin_J$)}\\n {\\n CandPins -= $Pin_J$;\\n }\\n }\\n \\\\textbf{Return:} CandPins\\n\\\\caption{Proximity attack}\\n\\\\label{alg:prox_jevs}\\n\\\\end{algorithm}\\nFrom the gate-level netlist, the algorithm chooses an arbitrary \\\\textit{TargetPin} from the unassigned partition input pins and output ports, creates a list of possible \\\\textit{CandidatePins}, and then connects the \\\\textit{TargetPin} to the closest pin in this list. After each connection, the netlist is updated. This procedure is repeated until all unassigned ports are connected. When the procedure is over, the attacker obtains the possible missing BEOL connections. If all guesses were correct, the original design has been recovered and Split Manufacturing has been defeated.\\n \\\\begin{table*}[htb]\\n \\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Threat Models, Attacks, and Metrics.}\\n \\\\begin{tabular}{ccclp{4.0cm}lp{1.7cm}p{1.2cm}}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Work} & \\\\textbf{Year} & \\\\parbox{1cm}{\\\\textbf{Threat} \\\\\\\\\\\\textbf{model}} & \\\\textbf{Attack type } & \\\\textbf{Novelty} & \\\\textbf{Benchmark suite(s)} & \\\\parbox{2cm}{\\\\textbf{Largest circuit} \\\\\\\\\\\\textbf{size (gates)}} & \\\\parbox{2cm}{\\\\textbf{Avg. circuit} \\\\\\\\\\\\textbf{size (gates)} } \\\\vspace{1pt} \\\\\\\\ \\n \\\\hline\\n & 2013 & I & Proximity & Attack Based on Proximity & ISCAS\\\\textquotesingle85 & 3.51k & 1288\\\\\\\\\\n & 2016 & II & Proximity & \\\\parbox{4.5cm}{Placement and routing proximity \\\\\\\\ used in conjunction} & ISPD\\\\textquotesingle11 & 1.29M & 951k \\\\\\\\\\n & 2018 & I & Proximity & \\\\parbox{4.5cm}{Network-Flow-Based with Design \\\\\\\\ Based Hints} & ISCAS\\\\textquotesingle85 \\\\& ITC\\\\textquotesingle99 & 190.21k & 9856\\\\\\\\\\n & 2018 & I & Proximity & \\\\parbox{4.5cm}{Proximity Attack Based on Machine \\\\\\\\ Learning} & ISPD\\\\textquotesingle11 & 1.29M & 951k\\\\\\\\\\n & 2019 & I & SAT & SAT Attack without Proximity Information & ISCAS\\\\textquotesingle85 \\\\& ITC\\\\textquotesingle99 & 190.21k & 9856\\\\\\\\\\n & 2019 & I & SAT & \\\\parbox{4.5cm}{SAT attack dynamically adjusted \\\\\\\\based on proximity information} & ISCAS\\\\textquotesingle85 \\\\& ITC\\\\textquotesingle99 & 190.21K & 9856\\\\\\\\\\n \\\\hline \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:atcks_desc}\\n\\\\end{table*}\\nAlgorithm \\\\ref{alg:prox_jevs} was originally applied to the ISCAS\\\\textquotesingle85 suite of benchmark circuits. These circuits were originally selected and published to help in comparing automatic test pattern generation (ATPG) tools. Due to the small size of these circuits, they may not be the best option to assess the effectiveness of Split Manufacturing. The difficulty to retrieve the BEOL connections is directly proportional to the size of the circuit. The authors reported an average effectiveness of 96\\\\% of Correct Connection Rate (CCR) across all the benchmarks considered. For the c17 circuit (smallest circuit in the ISCAS\\\\textquotesingle85 suite with only 6 gates), all the connections were retrieved correctly, thus demonstrating that the algorithm is capable of retrieving the missing BEOL connections. In Table~\\\\ref{tab:atcks_effect}, we highlight the best and worst results in terms of CCR.\\nJeyavijayan \\\\etal were the first to question the security of straightforward Split Manufacturing. Their proximity attack showed promising results, even if the considered benchmark circuits were rather small in size. This was the starting point for other studies proposing different attacks to Split Manufacturing in an attempt to retrieve the missing BEOL connections. Improvements over the original proximity attack, as well as other attacks, are compiled in Table~\\\\ref{tab:atcks_desc}. \\nThe effectiveness of the proximity attack utilizing distance of unassigned pins alone as metric to find missing BEOL connections was questioned by Maga\\\\~na \\\\etal . The authors proposed to utilize both placement and routing information in augmented proximity attacks. For their results, large-sized circuits from the ISPD-2011 routability-driven placement contest were used. These benchmarks are better representatives of modern circuits as they contain 9 metal layers and up to two million nets in a design. Thus, in an attempt to increase the success rate of the attack for large-sized circuits, they proposed routing-based proximity in conjunction with placement-centric proximity attacks.\\n \\\\begin{table*}[htb]\\n \\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Benchmark Size and Comparison of Attack Results.}\\n \\\\begin{tabular}{cllcllc}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Work} & \\\\textbf{Benchmark} & \\\\textbf{Attack} & \\\\textbf{Split Layer} & \\\\textbf{Size (In Gate Count)} & \\\\textbf{Metric} & \\\\textbf{Result} \\\\vspace{5pt} \\\\\\\\\\n \\\\hline\\n & c17 & Proximity & Not Defined & 6 & CCR(\\\\%) & 100 \\\\\\\\\\n & c7552 & Proximity & Not Defined & 3513 & CCR(\\\\%) & 94 \\\\\\\\\\n & Superblue 1 & Placement Proximity & M2 & 847k & \\\\% Match in List & 12.84 \\\\\\\\\\n & Superblue 1 & Placement Proximity & M2 & 847k & CCR(\\\\%) & 5.479 \\\\\\\\\\n & Superblue 1 & Routing Proximity & M2 & 847k & \\\\% Match in List & 71.08 \\\\\\\\\\n & Superblue 1 & Routing Proximity & M2 & 847k & CCR(\\\\%) & 0.651 \\\\\\\\\\n & Superblue 1 & Overlap (P\\\\&R) Proximity & M2 & 847k & \\\\% Match in List & 13.05 \\\\\\\\\\n & Superblue 1 & Overlap (P\\\\&R) Proximity & M2 & 847k & CCR(\\\\%) & 3.977 \\\\\\\\\\n & Superblue 1 & Crouting Proximity & M2 & 847k & \\\\% Match in List & 82.08 \\\\\\\\\\n & Superblue 1 & Crouting Proximity & M2 & 847k & CCR(\\\\%) & 0.651 \\\\\\\\\\n & c7552 & Network-flow Based Proximity & Not Defined & 3513 & CCR(\\\\%) & 93 \\\\\\\\\\n & c7552 & Proximity & Not Defined & 3513 & CCR(\\\\%) & 42 \\\\\\\\\\n & B18 & Network-flow Based Proximity & Not Defined & 94249 & CCR(\\\\%) & 17 \\\\\\\\\\n & B18 & Proximity & Not Defined & 94249 & CCR(\\\\%) & < 1 \\\\\\\\\\n & Superblue 1 & Proximity & M6 & 847k & \\\\% Match in list & 33.40 \\\\\\\\\\n & Superblue 1 & Proximity & M6 & 847k & CCR(\\\\%) & 0.76 \\\\\\\\\\n & Superblue 1 & ML & M6 & 847k & \\\\% Match in list & 83.12 \\\\\\\\\\n & Superblue 1 & ML & M6 & 847k & CCR(\\\\%) & 1.91 \\\\\\\\\\n & Superblue 1 & ML-imp & M6 & 847k & \\\\% Match in list & 74.65 \\\\\\\\\\n & Superblue 1 & ML-imp & M6 & 847k & CCR(\\\\%) & 2.11 \\\\\\\\\\n & Superblue 1 & ML-imp & M4 & 847k & \\\\% Match in list & 75.45 \\\\\\\\\\n & Superblue 1 & ML-imp & M4 & 847k & CCR(\\\\%) & 2.58 \\\\\\\\\\n & c7552 & SAT Attack & Not Defined & 3513 & Logical Equivalence(\\\\%) & 100 \\\\\\\\\\n & B18 & SAT Attack & Not Defined & 94249 & Logical Equivalence(\\\\%) & 100 \\\\\\\\\\n & c7552 & Improved SAT Attack & Not Defined & 3513 & Logical Equivalence(\\\\%) & 100 \\\\\\\\\\n & B18 & Improved SAT Attack & Not Defined & 94249 & Logical Equivalence(\\\\%) & 100 \\\\\\\\\\n \\\\hline \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:atcks_effect}\\n\\\\end{table*}\\nA key difference present in is that this work utilizes a different threat model (model II), claiming that the untrusted foundry possesses information about the entire place \\\\& routed netlist (as well as the FEOL layout). This assumption is hard to reason if the attacker's intent was to overproduce the IC or pirate the IP. For these goals, clearly, this assumption is unnecessary. The attacker himself can, if he indeed possesses the netlist, perform his own physical synthesis and generate his own layout. The interest in reverse engineering the BEOL connections of the original design diminishes. Nevertheless, we report on the strategies employed by the authors of since they build on the approach proposed by .\\n\\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.90\\\\linewidth]{sa_onlycir.pdf}\\n{Representation of the routing for the first 3 metal layers of a simple circuit (adapted from ). \\\\label{fig:circ_abs}}\\nRegarding the attacks, the authors of proposed four different techniques to identify a small search neighborhood for each pin. The goal is to create a neighborhood that is small enough to make further pruning feasible, and therefore increase the likelihood of including the matching pins. The techniques are called \\\\textit{placement proximity}, \\\\textit{routing proximity}, \\\\textit{crouting proximity} and \\\\textit{overlap of placement and routing proximity}, and are described in the text that follows. The circuit illustrated in Figure~\\\\ref{fig:circ_abs} is the example (before the split) that will guide the discussion on these four techniques.\\n\\\\textit{Placement proximity} exploits the placement information of cells. Each split wire is taken from the pin location of the corresponding standard cell that is connected to it. A search neighborhood is defined as a square region centered around the corresponding pin with an area equal to the average areas of the bounding boxes (BB) in a typical design. The authors argued that it can also be measured based on BBs of the non-split wires in the design under attack, under the assumption that the number of wires that remain in the FEOL is also very large in practice. Let us consider the circuit illustrated in Figure~\\\\ref{fig:circ_abs} as an example. If the split is done at M2, the search area defined using the placement proximity would contain three gates as illustrated in Figure~\\\\ref{fig:prox_sa} (a). Thus, using the placement proximity search area, the most likely connections are illustrated by the green squares (candidate pins) and by the red squares (target pins). Note that the layer at which the layout is split does not affect the search area defined by placement proximity.\\n\\\\textit{Routing proximity} exploits the routing information. First, for each split wire, pins are identified as the point where the wire is actually cut at the split layer, i.e., the via location. Next, a square area centered around those pins is defined. The size of the square area is defined based on the average BBs of the pins on that layer in the design. This procedure for identification results in different neighborhood sizes according to the split layer location, i.e., the search radius adapts to the routing resources of each layer. A search area defined using routing proximity is illustrated in Figure~\\\\ref{fig:prox_sa} (b), highlighted in gray and containing a set of routing wires and its respective pins. \\n\\\\textit{Crouting proximity} takes into account routing congestion by exploiting the union of placement and routing proximity. The search area for each pin is defined in such way that the ratio of number of pins to the search area is equal across all the pins in the split layer. Thus, if a pin is located at a high routing congestion area, the search area will be expanded until the pin density in the new search area reaches a target value or the search area grows to four times its starting value. The starting value is set according to the split layer, set as the average of numbers of pins which fall within a BB. A search area defined using crouting proximity is illustrated in Figure~\\\\ref{fig:prox_sa} (c). \\nThe last strategy proposed by also combines placement and routing information. It is referred to as \\\\textit{Overlap of placement and routing proximities}. The concept here is to include a subset of pins identified by the placement proximity list which have their corresponding pins included in the routing proximity list. According to the author, intuitively, the overlap then identifies a subset of pins which may be more likely to point towards the direction of the matching pin. A search area defined using the overlap of placement and routing proximities is illustrated in Figure~\\\\ref{fig:prox_sa} (d). \\n \\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\\\linewidth]{prox_sa_new.pdf}\\n{Multiple strategies for pin/connectivity search areas according to . \\\\label{fig:prox_sa}}\\nMaga\\\\~na et al. assessed each strategy using the benchmark circuit \\\\textit{superblue1}. Different split layers were also considered. In Table~\\\\ref{tab:atcks_effect}, we compiled the results for split layer M2. By comparing the results, it becomes clear that no strategy was able to recover 100\\\\% of the missing BEOL connections. The best result was only 5.479\\\\% of CCR. This is in heavy contrast with the findings of . However, as we previously noted, the circuit sizes differ by orders of magnitude.\\nAccording to the authors of , proximity alone is in no way sufficient to reverse engineer the FEOL. However, proximity attacks have merit as they can be used to narrow down the list of candidates to a significantly smaller size. Using crouting proximity, in 82.08\\\\% of the cases, the search area defined contained the matched pin in the list of candidates. The authors also present results for a circuit split at M8. We opt not to show these results in Table~\\\\ref{tab:atcks_effect}. Using the circuit \\\\textit{superblue1} as an example, the number of unassigned pins when the circuit is split at M8 is only 1.2\\\\% of the pins when split at M2. Therefore, the small number of unassigned pins to be connected overshadows the large circuit used for their experiments. It must also be emphasized that splitting a circuit in such higher layers is rather impractical since M8 tends to be a very thick metal reserved for power distribution in typical 10-metal stacks. There is very little value in hiding a power distribution network from an adversary that wants to pirate an IP. Once again, we opt not to show this result in our comparisons.\\nA network-flow based attack model towards flattened designs was proposed by Wang et. al . The authors argued that the proximity attack originally proposed by utilizes hints that can be used only by hierarchical designs, and that modern designs are often flattened \\\\footnote{We highlight that best practices in circuit design have changed over the years. Hierarchical design was heavily utilized for many years, but it lost favor due to the difficulty in performing reasonable timing budgeting between the many blocks of a system. Thus, flattened designs are often used to facilitate timing closure.}. Based on the original proximity attack, they proposed a proximity attack utilizing five hints: physical proximity, acyclic combinational logic circuit, load capacitance constraint, directionality of dangling wires, and timing constraint. Note that the first two hints are already described by and . The three novel hints are described below:\\n \\\\begin{itemize}\\n \\\\item[] \\\\textit{Load Capacitance Constraint}: gates can drive a limited load to honor slew constraints. Typically, maximum load capacitance is constrained and has a maximum value defined by the PDK and/or the standard cell characterization boundaries. Hence, an attacker will consider only connections that will not violate load capacitance constraints.\\n \\\\item[] \\\\textit{Directionality of Dangling Wires}: routing engines tend to route wires from a source to a sink node along the direction of the sink node. Therefore, the directionality of remaining dangling wires at lower metal layers may indicate the direction of their destination cell with a high degree of certainty\\\\footnote{Metals usually have preferred directions that alternate along the stack (i.e., if M1 is vertical, then M2 is horizontal). Therefore, this hint becomes more effective if the attacker can observe more than one routing layer of the FEOL}. An attacker can disregard connections in the other directions.\\n \\\\item[] \\\\textit{Timing Constraint}: connections that create timing paths that violate timing constraints can be excluded. An attacker, through an educated guess of the clock period, can determine a conservative timing constraint and exclude any connections that would lead to slower paths. \\n \\\\end{itemize}\\nThe network-flow based attacked framework proposed by Yang \\\\etal~considers two hints proposed by plus the aforementioned hints to create a directed graph $G = (V,E)$, where $V$ is a set of vertices and $E$ is a set of edges. The set ($V$) is composed by the set of vertices corresponding to the output pins $V_o$, and a set corresponding to the input pins ($V_i$), the source vertex ($S$), and the target vertex ($T$). The set $E$ consists of $E_{So}$, edges from $S$ to every output pin vertex, $E_{oi}$, edges from output pin vertices to input pin vertices, and $E_{iT}$, which includes edges from every input vertex to the target vertex. An example of this kind of representation is shown in Figure~\\\\ref{fig:network_flow}, where (a) is the circuit with missing connections and (b) is the network-flow representation. The detailed problem formulation is omitted from this work. To find the connections, a min-cost network-flow problem is solved, where the decision variables are the flow $x_{i,j}$ going through each edge $(i,j) \\\\in E$ . The authors utilized the Edmons-karp algorithm to solve this problem. Complexity of the algorithm alone is given by $O(VE^2)$, however, in the worst case it is required to run the algorithm $V$ times; thus, the run-time of the complete network attack is given by $O(V^2E^2)$ in the worst case.\\n\\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.90\\\\linewidth]{network_flow.pdf}\\n{(a) Circuit with missing connections. (b) Network-flow model for inferring the missing connections. (Adapted from ). \\\\label{fig:network_flow}}\\nThe network-flow approach was applied to ISCAS\\\\textquotesingle85 and ITC\\\\textquotesingle99 benchmark circuits. The ITC\\\\textquotesingle99 circuits were proposed as an evolution of the ISCAS\\\\textquotesingle85 set, since the Test community acknowledged that newer and larger circuits were already in demand. For comparison, the authors applied both the original proximity attack and the network-flow attack to flattened designs. As shown in Table~\\\\ref{tab:atcks_effect}, their network-flow proximity attack outperformed the original attack in terms of CCR. However, despite the evident improvement, the attack could only retrieve 17\\\\% of the missing BEOL connections for a medium sized circuit (b18 from the ITC\\\\textquotesingle99 suite).\\nA Machine Learning (ML) framework was used by Zhang \\\\etal in an attempt to improve the attack proposed in . The same setup as previously discussed was utilized. However, more layout features were incorporated in their ML formulation, including placement, routing, cell sizes, and cell pin types. \\nA high-level overview of their modeling framework is shown in Figure~\\\\ref{fig:ml_framework} (a). First, they create a challenge instance from the entire layout and only FEOL view. Next, for each virtual pin (point where a net is broken on the split layer), layout information is collected, including placement, routing, cell areas, and cell pin as illustrated in Figure~\\\\ref{fig:ml_framework} (b). Using this information, samples are generated which are fed into the ML training process. Each sample carries information of a pair of virtual pins which may or may not be matched. Classifiers then are built by the ML framework using training samples. After training and building the regression model, cross validation is used for evaluation which ensures validation of the model is done on data samples which were not used for training. Their framework faces scaling issues when applied to lower split metal layers. An improved ML framework is then proposed as well, denoted by ML-imp, to solve the scaling issues.\\nFor their experiments, Zhang \\\\etal utilize the ISPD\\\\textquotesingle11 benchmark suite. They compare results from their previous work with their ML and ML-imp frameworks. However, they do not show results for lower split metal layers (e.g., M2). Instead, results are provided for M8, M6, and M4 splits. As pointed out before, utilizing higher layers for the split effectively shrinks the otherwise large circuits used in their experiments. A drastic reduction of unassigned pins is expected for such higher layers, as higher metal layers are used often for power routing, not for signal routing. Results for the \\\\textit{superblue1} circuit are shown in Table~\\\\ref{tab:atcks_effect}. Regarding recovering missing BEOL connections, ML and ML-imp could only retrieve around 2\\\\%, therefore not showing a huge improvement over their previous work. However, search list area accuracy showed significantly better results when compared to their prior work. A caveat worth mentioning is that the proposed machine learning framework needs the entire layout during its modeling phase. This characteristic may, in an extreme case, nullify the applicability for an attacker that only holds the FEOL layout and cannot produce training samples from other sources.\\n \\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=1\\\\linewidth]{ml_features.pdf}\\n{(a) Machine Learning Modeling by Zhang et al. . (b) Few Exmples of Layout Features. \\\\label{fig:ml_framework}}\\nAttacks using proximity information as a metric are not the only solution to recover missing BEOL connections. An effective methodology to apply a Boolean satisfiability based strategy is proposed by Chen \\\\etal . The authors claim that their attack methodology does not need (or depend on) any proximity information, or even any other insights into the nature of EDA tools utilized in the design process. The key insight in their work is to model the interconnect network as key-controlled multiplexers (MUX). Initially, all combinations of signal connections between the FEOL partitions are allowed, as illustrated in Figure~\\\\ref{fig:mux_network}. First, a MUX network is created in order to connect all missing paths in the circuit. This MUX network leads to potential cyclic paths, thus, there is a possibility to generate many combinational loops during the attack process corresponding to incorrect key guesses. Therefore, constraints on the key values are generated in order to avoid activating the cyclic paths. The attack can be summarized in 4 steps: \\\\textit{identification of all cyclic paths}, \\\\textit{generation of cycle constraints}, \\\\textit{cycle constraints optimizations}, and finally, \\\\textit{SAT attack}. The authors utilize a SAT solver-based attack method derived from CycSat . The SAT attack algorithm has as input the FEOL circuit with MUX network and a packaged IC that serves as an oracle. The algorithm outputs keys to be used in the MUXes such that correct BEOL connections are made. \\nIn reality, presents a different interpretation of Threat model I since the attacker is assumed to possess a functional IC. This IC would then have to be available in the open market for the attacker to be able to purchase it. This characteristic severely narrows down the applicability of this SAT attack. For instance, ICs designed for space or military use will not be freely available, thus an oracle may not be known to the attacker.\\nExperimental results presented by utilize ISCAS\\\\textquotesingle85 and ITC\\\\textquotesingle99 benchmark circuits. It has been shown that their attack could recover a logically correct netlist for all the studied circuits. However, there is a small clarification to be made that relates to what is a logically correct circuit. In Table~\\\\ref{tab:atcks_effect}, two of those results are shown. For seven of the studied benchmarks (c1908, c2670, c5315, c7552, b14, b15, b17), the connections recovered are identical to the BEOL connections. For the remaining benchmarks, the recovered connections are not identical but logically equivalent to the original circuit. In practice, the logically equivalent circuit may present performance deviations from the original design. Matching the performance of the original design can be done by re-executing place and route using the logically equivalent gate-level netlist. Depending on the attack goal, it is possible that the attacker had already planned to re-execute the physical synthesis flow again (say, to resell the IP in a different form or shape). An attack that guarantees 100\\\\% of logic equivalence of the recovered netlist is powerful enough, allowing attackers to copy and modify split layouts.\\nIn order to increase the efficiency and capacity of the SAT attack proposed in , the authors proposed two improvements in . First, the size of the key-controlled interconnect network that models the possible BEOL connections is reduced. Second, after the MUX network is inserted into the FEOL circuit, the number of combinational cycles it induces in the design for incorrect key guesses should also be reduced. Proximity information is then exploited to achieve the proposed improvements. The improved SAT attack method which exploits proximity information showed significant reduction in the attack time and increase in the capacity. Same as in , the circuits tested were 100\\\\% recovered, as shown in Table~\\\\ref{tab:atcks_effect}. \\n\\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\\\linewidth]{mux_network.pdf}\\n{MUX network for a bipartitioned FEOL circuit (Adapted from ). \\\\label{fig:mux_network}}\\n\\\\begin{table*}[htb]\\n\\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Split Manufacturing Defenses.}\\n \\\\begin{tabular}{cp{1cm}ccp{2.2cm}p{2cm}l}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Work} & \\\\textbf{Year} & \\\\textbf{Threat Model} & \\\\textbf{Category} & \\\\textbf{Defense} & \\\\textbf{Metrics} & \\\\textbf{Defense Overheads Presented} \\\\vspace{5pt} \\\\\\\\\\n \\\\hline \\n & 2013 & I & Proximity Perturbation &Pin Swapping & Hamming Distance & \\\\multicolumn{1}{c}{-*} \\\\\\\\\\n & 2013 & II & Wire Lifting &Wire Lifting & k-Distance & Power, Area, Delay and Wire-Length \\\\\\\\\\n & 2014 & I & Layout Obfuscation & Layout Obfuscation for SRAMs and Analog IPs & - & Performance, Power and Area \\\\\\\\\\n & 2014 & I & Layout Obfuscation & Obfuscation Techniques & Neighbor Connectedness and Entropy & Performance and Area \\\\\\\\\\n & 2015 & I & Layout Obfuscation & Automatic Obfuscation Cell Layout & Neighbor Connectedness and Entropy & Performance, Power and Area \\\\\\\\\\n & 2015 & I & Layout Obfuscation & Obfuscated Built-in Self-Authentication & Obfuscation Connection & Number of Nets \\\\\\\\\\n & 2016 & I & Wire Lifting & Artificial Blockage Insertion & Number of Pins & \\\\multicolumn{1}{c}{-*} \\\\\\\\\\n & 2016 & I & Wire Lifting & Net Partition, Cell Hidden and Pin Shaken & - & \\\\multicolumn{1}{c}{-*} \\\\\\\\\\n & 2017 & I & Proximity Perturbation & Routing Perturbation & Hamming Distance & Performance and Wire-Length \\\\\\\\\\n & 2017 & I & Wire Lifting & Secure Routing Perturbation for Manufacturability & Hamming Distance & Performance and Wire-Length \\\\\\\\\\n & 2017 & I & Proximity Perturbation & placement-centric Techniques & CCR & Performance, Power and Area \\\\\\\\\\n & 2017 & II & Proximity Perturbation & Gate Swapping and Wire Lifting & Effective Mapped Set Ratio and Average Mapped Set Pruning Ratio & Wire-Length \\\\\\\\\\n & 2018 & I & Wire Lifting & Concerted Wire Lifting & Hamming Distance & Performance, Power and Area \\\\\\\\\\n & 2018 & I & Proximity Perturbation & Secure Driven Placement Perturbation & Hamming Distance & Power and Wire-Length\\\\\\\\\\n & 2018 & I & Proximity Perturbation & placement and routing perturbation & Hamming Distance & Performance, Power and Area \\\\\\\\\\n & 2019 & I & Layout Obfuscation & Isomorphic replacement for Cell Obfuscation & Isomorphic Entropy & \\\\multicolumn{1}{c}{-*} \\\\\\\\\\n & 2019 & II & Layout Obfuscation & Dummy Cell and Wire Insertion & k-security & Area and Wire-Length \\\\\\\\\\n \\\\hline \\\\hline\\n \\\\multicolumn{7}{l}{\\\\cellcolor{white}* Authors do not present any discussion regarding overhead.} \\n \\\\end{tabular}\\n \\\\label{tab:sm_defenses}\\n\\\\end{table*}\",\n \"id\": \"63582e79-19ab-44ed-8225-2fdf07c8ddcd\",\n \"level\": \"section\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"4df5dfa6-5506-4133-aae0-08641ac96086\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Attacks on Split Manufacturing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks on Split Manufacturing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:defenses}\\nAttacks toward Split Manufacturing showed promising results, as described in the previous section. A malicious attacker has the real potential to recover the missing BEOL connections. If the missing connections are successfully recovered, the security introduced by applying the technique is nullified. Therefore, straightforward Split Manufacturing is questioned by several works. Several authors proposed defense techniques that augment the technique, i.e., techniques that, when used together with Split Manufacturing, do increase the achieved security against attacks. In Table~\\\\ref{tab:sm_defenses}, we compile a comprehensive list of defense techniques found in the literature. Each defense technique utilizes a different metric and Threat model, depending upon the type of attack they are trying to overcome. Since many of the studied defense techniques often introduce heavy PPA overheads, Table~\\\\ref{tab:sm_defenses} also shows if the studied work assessed overheads and which ones were addressed. \\nIn the text that follows, the many defense techniques are divided into categories, namely Proximity Perturbation (i.e., change the location of cells or pins), Wire Lifting (i.e., move routing wires to upper layers), and Layout Obfuscation (i.e., hide the circuit structure). We present the categories in this exact order. For some techniques, it is worth mentioning that overlaps do exist and that techniques could be categorized differently. Thus, this categorization is our interpretation of the state of the art and may not be definitive. Furthermore, the boundaries between categories are not strict. For example, a technique may perform a layout modification that promotes proximity perturbation and leads to (indirect) wire lifting. In Tables~\\\\ref{tab:def_res} and \\\\ref{tab:def_res2}, we compile the results for the Proximity Perturbation and Wire Lifting categories, respectively. To demonstrate the effectiveness of each defense technique, we compile the results for when the attack is done with and without the defense. The results showed in Tables~\\\\ref{tab:def_res} and \\\\ref{tab:def_res2} are for the smallest and largest circuits addressed in each studied work. Additionally, we show the PPA overhead introduced and, if specified, the split layer.\",\n \"id\": \"0381ac67-f861-41a5-94dd-d57b1d10cc30\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4df5dfa6-5506-4133-aae0-08641ac96086\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Split Manufacturing Defenses\"\n ]\n ],\n \"subsections\": [\n \"ee25dbd5-cfaf-41a3-8225-ebde1d228b40\",\n \"90e5175d-afc8-4844-8dfa-513e3fcc4fbd\",\n \"4032829b-2e93-4b12-822b-7b2a6e653dbc\"\n ],\n \"title\": \"Split Manufacturing Defenses\"\n },\n {\n \"cite_extract_rate\": 0.28571428571428503,\n \"cites\": [\n 7798,\n 7799\n ],\n \"content\": \"Attacks toward split circuits are generally based on leveraging proximity information. The first category of defenses, Proximity Perturbation, addresses this hint left by the EDA tools. The goal of the techniques within this category is to promote changes in the circuit such that the proximity information between the FEOL pins is less evident. Therefore, the success rate of the proximity attacks is decreased.\\nIn , the authors proposed pin swapping to overcome proximity attacks. Rearranging the partition pins can alter their distance in such a way that the attacker is deceived. As an example, if the pins $P_{G3,B,in}$ and $P_{G6,A,in}$ (Figure~\\\\ref{fig:cir_part}) are swapped, the proximity attack will incorrectly guess the connection between $P_{G2,A,out}$ and $P_{G3,B,in}$. Thus, a sufficient number of pins have to be swapped in order to create a netlist that is significantly different from the original netlist (based on some sort of metric for similarity). In , Hamming distance is proposed as a way to quantify the difference between the outputs of the original netlist and the modified netlist. Assuming the outputs of a circuit are arranged as a string of bits, Hamming distance is defined as the number of bits that change when two instances of this string are compared to one another. The authors argued that the optimum netlist is created when the Hamming distance is 50\\\\%. Therefore, inducing the maximal ambiguity for a potential attacker. Since the best rearrangement for N pins of partitions might take $N!$ computations (rather computationally expensive), pair-wise swapping of pins is considered in . Pair-wise swapping of pins results in $O(N^2)$ computations. \\nThe modified netlist is created based on a series of rules. Similarly, to the proximity attack, a list of candidates pins to be swapped is created before the actual swap is applied. Since not every pin can be swapped, a candidate pin to be swapped should:\\n \\\\begin{itemize}\\n \\\\item be an output pin of the partition where the target pin resides\\n \\\\item not be connected to the partition where the candidate pin resides\\n \\\\item not form a combinational loop\\n \\\\end{itemize}\\nUsing the above constraints, a candidate pin is selected. The target pin also needs to be chosen carefully. In , IC testing principles and hints from the original proximity attack are used to choose the target pin. The swapping procedure is described in Algorithm \\\\ref{alg:pin_swap}, where $TestMetric$ is a metric based on IC testing principles, such as stuck-at fault models which are still utilized in Test today. More details can be obtained from . The proposed defense technique is validated using ISCAS\\\\textquotesingle85 circuits and the original proximity attack. For the smallest circuit, c17, it takes only one swap to achieve a Hamming Distance of 50\\\\%. For the largest studied circuit, c7552, it takes 49 swaps. These results are summarized in Table~\\\\ref{tab:def_res}.\\nAs demonstrated in , rearranging the partition pins can thwart proximity attacks. However, according to Chen \\\\etal , pin swapping at partition level has limited efficacy. They demonstrated that an attacker holding the FEOL layout as well as the nestlist can insert hardware trojans even when the defense approach of is applied. It must be highlighted that assumes threat model II, which we have previously argued that has the potential to nullify the vast majority of defenses towards split circuits. Thus, they proposed a defense to counter the threat from hardware trojans. Their defense incorporates the global wire-length information, with the goal to hide the gates from their candidate locations, and as result decreasing the effective mapped set ratio (EMSR). The EMSR metric is an attempt to quantify the ratio of real gates location of a given mapping during a simulated annealing-based attack. This defense consists of two steps, first a greedy gate swapping defense , and second, a measurement of the security elevation in terms of EMSR. The technique is evaluated using ISCAS\\\\textquotesingle85 benchmarks circuits and the EMSR metric to quantify the defense effectiveness. The results are shown in Table~\\\\ref{tab:def_res}.\\nFollowing the same principle of increasing the Hamming Distance, Wang \\\\etal proposed a routing perturbation based defense. The optimum Hamming distance is sought to be achieved by layer elevation, routing detours, and wire decoys, while test principles are used to drive the perturbation choices. Layer elevation is essentially a wire lifting technique: without changing the choice of split metal layer, wires are forced to route using higher metal layers, thus being lifted from the FEOL to the BEOL. Intentional routing detours are a way to increase the distance between disconnected pins of the FEOL. If done properly, disconnected pins will not be the closest to each other, deceiving the proximity attack. In some cases, routing detours will increase the distance between disconnected pins, however, they still remain the closest to each other. In this scenario, wire decoys can be drawn near disconnected pins, in such a way that decoys are now the closest and will instead be picked as the ideal candidate pin. \\nThe perturbations proposed in can incur heavy overheads, and, for this reason, wires to be perturbed are chosen by utilizing IC test principles. In , fault observability, as defined in SCOAP , is used as a surrogate metric for this task. The technique is evaluated using ISCAS\\\\textquotesingle85 and ITC\\\\textquotesingle99 benchmark circuits. For all studied circuits, the Hamming distance increased by an average of 27\\\\% at a cost of only 2.9\\\\% wire length overhead (WLO), on average. The results for the largest and smallest studied circuits are shown in Table~\\\\ref{tab:def_res}.\\n \\\\begin{algorithm}[htb]\\n\\\\DontPrintSemicolon\\n \\\\KwInput{Partitions}\\n \\\\KwOutput{List of target and swapping pins}\\n $ListofTargetPins$ = $\\\\emptyset$;\\\\\\\\\\n $ListofSwappingPins$ = $\\\\emptyset$;\\\\\\\\\\n $ListofUntouchedPins$ = All partition pins and I/O ports;\\\\\\\\\\n \\\\While{Untouched output partitions pins or input ports exist}\\n {\\n \\\\For{$UntouchedPin$}\\n {\\n $SwappingPins$ = \\\\\\\\\\n BuildSwappingPinsList($UntouchedPin$);\\n \\\\For{$SwappingPin$ $\\\\in$ $SwappingPins$}\\n {\\n Compute \\\\\\\\\\n $TestMetric$($UntouchedPin,\\n SwappingPin$);\\n }\\n }\\n Find the $TargetPin$ and $SwappingPin$ with the Highest $TestMetric$ from its SwappingPins;\\\\\\\\\\n $ListofTargetPins += TargetPins$;\\\\\\\\\\n $ListofSwappingPins += SwappingPins$;\\\\\\\\\\n $ListofUntouchedPins -= TargetPins$;\\\\\\\\\\n $LisofUntouchedPins -= SwappingPin$; \\\\\\\\\\n Swap TargetPin and SwappingPin; \\\\\\\\\\n Update netlist;\\n }\\n \\\\textbf{Return:} ListofTargetPins and ListofswappingPins;\\n \\\\textbf{BuildSwappingPinList}($TargetPin$); \\\\\\\\\\n \\\\KwInput{$TargetPin P_{x,i,out}$}\\n \\\\KwOutput{$SwappingPins$ for $TagetPin$}\\n \\\\For{$Pin_J \\\\in SwappingPins$}\\n {\\n \\\\If{$CombinationalLoop(TargetPin,Pin_J)$}\\n {\\n $SwappingPins -= Pin_J$;\\n }\\n }\\n \\\\textbf{Return:} $SwappingPins$;\\n\\\\caption{ Fault analysis-based swapping of pins to thwart proximity attack (adapted from ).}\\n\\\\label{alg:pin_swap}\\n\\\\end{algorithm}\\nSengupta \\\\etal take a different direction from other works. They utilized an information-theoretic metric to increase the resilience of a layout against proximity attacks. As demonstrated in , mutual information (MI) can be used to quantify the amount of information revealed by the connectivity distance between cells. Mutual information is calculated by taking into account the cells connectivity $D$, if they are connected or not, and their Manhattan distance $X$, described by Eq. \\\\ref{eq:mutual_info}, where $H[\\\\cdot]$ is the entropy. The Manhattan distance of two cells is defined as the sum of horizontal and vertical distances between them. Entropy is a measure of disorder of a system. Therefore, in this work, entropy is utilized as a measure of disorder in the FEOL layer. The distribution of the variable $X$ for a given layout is determined pair-wise for all gates, allowing a straightforward computation of $I(X;D)$. Thus, layouts with the lowest mutual information, i.e., the correlation between cell connectivity and their distance is low, are more resilient against proximity attacks.\\n \\\\begin{equation}\\\\label{eq:mutual_info}\\n MI = I(X;D) = H[X]-H[X/D]\\n \\\\end{equation}\\nIn order to minimize the information ``leaked'' from mutual information, applies cell placement randomization and three other techniques: g-color, g-type1, and g-type2. Randomizing the cell placement can achieve the desired low mutual information; however, the PPA overhead incurred is excessive. Minimizing mutual information without excessive PPA overhead can be achieved by the other techniques. From a graph representation of the circuit, graph coloring can be used to hide connectivity information, where gates of the same color must not be connected. Thus, the resulting colored netlist is then partitioned by clustering all cells of same color together. During cell placement, the cells with the same color will be confined within their respective clusters. According to , these constraints naturally mitigate the information leakage to a great extent. The g-color technique utilizes only the graph coloring as described above. The other two, g-type1 and g-type2, consider the type of the gate when creating clusters. The g-type1 approach clusters gates only by their functionality, while g-type2 utilizes functionality and the number of inputs for clustering. The authors assessed their techniques utilizing ISCAS\\\\textquotesingle85 and MNCN benchmark suites. Results for the smallest and largest circuits are shown in Table~\\\\ref{tab:def_res}. \\n\\\\begin{table*}[htb]\\n\\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Results for Defense Techniques based on Proximity Perturbation.}\\n \\\\begin{tabular}{p{.8cm}p{1cm}p{1.2cm}p{2.5cm}lp{2.2cm}p{.5cm}p{1.5cm}p{1.5cm}}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Work} & \\\\textbf{Attack Type} & \\\\textbf{Benchmark} & \\\\textbf{Defense Technique} & \\\\textbf{Defense Metric} & \\\\textbf{Defense Overhead} & \\\\textbf{Split Layer} & \\\\textbf{Result without Defense} & \\\\textbf{Result with Defense} \\\\vspace{5pt} \\\\\\\\\\n \\\\hline\\n & Proximity & c17 & - & Hamming Distance & 1 Swap for 50\\\\% HD & -* & 100\\\\% CCR & 78\\\\% CCR\\\\\\\\\\n & Proximity & c7552 & - & Hamming Distance & 49 Swaps for 50\\\\% HD & -* & 94\\\\% CCR & 91\\\\% CCR\\\\\\\\\\n & Proximity & c432 & Modifed Greedy Gate Swapping & EMSR & 75\\\\% of WLO & -* & 90\\\\% EMSR & 25\\\\% EMSR \\\\\\\\\\n & Proximity & c432 & Modifed Greedy Gate Swapping & EMSR & 300\\\\% of WLO & -* & 78\\\\% EMSR & 10\\\\% EMSR \\\\\\\\\\n & Proximity & c432 & - & Hamming Distance & 3.1\\\\% WLO for 46.1\\\\% HD & -* & 92.4\\\\% CCR & 78.8\\\\% CCR \\\\\\\\\\n & Proximity & c432 & - & Hamming Distance & 4.1\\\\% WLO for 31.7\\\\% HD & -* & 62.8\\\\% CCR & 37.9\\\\% CCR \\\\\\\\\\n & Proximity & c432 & Random & Mutual Information & < 10\\\\% PPA & M1 & 17\\\\% CCR & < 1\\\\% CCR \\\\\\\\\\n & Proximity & c432 & g-color & Mutual Information & < 10\\\\% PPA & M1 & 17\\\\% CCR & 2\\\\% CCR \\\\\\\\\\n & Proximity & c432 & g-type1 & Mutual Information & < 10\\\\% PPA & M1 & 17\\\\% CCR & 6\\\\% CCR \\\\\\\\\\n & Proximity & c432 & g-type2 & Mutual Information & < 10\\\\% PPA & M1 & 17\\\\% CCR & 4.5\\\\% CCR \\\\\\\\\\n & Proximity & c7552 & Random & Mutual Information & < 10\\\\% PPA & M1 & 13\\\\% CCR & < 1\\\\% CCR \\\\\\\\\\n & Proximity & c7552 & g-color & Mutual Information & < 10\\\\% PPA & M1 & 13\\\\% CCR & 2\\\\% CCR \\\\\\\\\\n & Proximity & c7552 & g-type1 & Mutual Information & < 10\\\\% PPA & M1 & 13\\\\% CCR & 4\\\\% CCR \\\\\\\\\\n & Proximity & c7552 & g-type2 & Mutual Information & < 10\\\\% PPA & M1 & 13\\\\% CCR & 3\\\\% CCR \\\\\\\\\\n & SAT & c432 & BEOL+Physical & Perturbation & 4.5\\\\% WLO & -* & 58\\\\% CCR & 56\\\\% CCR \\\\\\\\\\n & SAT & c432 & Logic+Physical & Perturbation & 5.57\\\\% WLO & -* & 58\\\\% CCR & 58\\\\% CCR \\\\\\\\\\n & SAT & c432 & Logic+Logic & WLD & 1.68\\\\% WLO & -* & 58\\\\% CCR & 52\\\\% CCR \\\\\\\\\\n & SAT & b18 & BEOL+Physical & Perturbation & 8.06\\\\% WLO & -* & 15\\\\% CCR & 14\\\\% CCR \\\\\\\\\\n & SAT & b18 & Logic+Physical & Perturbation & 1.70\\\\% WLO & -* & 15\\\\% CCR & 17\\\\% CCR** \\\\\\\\\\n & SAT & b18 & Logic+Logic & WLD & 0.61\\\\% WLO & -* & 15\\\\% CCR & 16\\\\% CCR** \\\\\\\\\\n & Proximity & c432 & Netlist Randomization & Hamming Distance & < 10\\\\% PPA overall & -* & 92.4\\\\% CCR & 0\\\\% CCR \\\\\\\\\\n & Proximity & c7552 & Netlist Randomization & Hamming Distance & < 10\\\\% PPA overall & -* & 94.4\\\\% CCR & 0\\\\% CCR \\\\\\\\\\n \\\\hline\\n \\\\hline\\n \\\\multicolumn{9}{l}{\\\\cellcolor{white}* Split layer not specified by the authors.} \\\\\\\\\\n \\\\multicolumn{9}{l}{\\\\cellcolor{white}** These results are counter-intuitive, the applied defense degrades the metric.} \\n \\\\end{tabular}\\n \\\\label{tab:def_res}\\n\\\\end{table*} \\nSimilar to the pin swapping technique proposed in , Wang \\\\etal proposed a placement-based defense with the same objective of deceiving a proximity attack by perturbing proximity information. Differently from pin swapping, their placement-based defense considers the incurred wire-length overhead as a metric. This technique is based on changing gate locations such that the proximity hint is no longer effective. Their algorithm consists of two phases, one to select which gates to be perturbed and a second phase where the selected gates are (re)placed. Gate selection is done by extracting a set of trees using two techniques, BEOL-driven and logic-ware extraction. The first approach selects all gate trees that contain any metal wires in the FEOL, i.e., connections that are not hidden from the attacker. The second approach considers the wire-length impact and the gate tree impact on the overall security. After extracting the set of trees, the placement perturbation is done in one of two ways: physical-driven or logic-driven. For each extracted tree, the physical-driven perturbation changes the location of gates using a Pareto optimization approach. Also, each solution is evaluated by its wire-length overhead and a perturbation metric that discerns the placement difference from the original layout. According to , geometric-based difference alone may be insufficient to enhance the split circuit security. Thus, a logic-driven perturbation is performed with a weighted logical difference (WLD) metric, which encourages perturbation solutions with large logical difference from its neighbors. The authors assessed their techniques combining the gate selection and perturbation as BEOL+Physical, Logic+Physical and Logic+Logic, using ISCAS\\\\textquotesingle85 and ITC\\\\textquotesingle99 circuit benchmarks. Results for the smallest and largest circuits considered are shown in Table~\\\\ref{tab:def_res}.\\nA considerably different approach is proposed by Patnaik \\\\etal , whereas netlist modifications are promoted (instead of placement/routing modifications during physical synthesis). The goal is to modify the netlist of a design in order to insert (partial) randomization. According to , this approach helps to retain the misleading modifications throughout any regular design flow, thereby obtaining more resilient FEOL layouts where the netlist changes are later ``corrected'' in the BEOL. This methodology is implemented as an extension to commercial EDA tools with custom in-house scripts. The process goes as follows: first, the netlist is randomized. Second, the modified netlist is place and routed. Lastly, the true functionality is restored by re-routing in the BEOL. For the netlist randomization, pairs of drivers and their sinks are randomly selected and swapped. This is done in such way to avoid combinational loops that may be introduced by swapping. The modified netlist then is place and routed, utilizing a `do not touch'\\\\footnote{This terminology is used in IC design to mean that a specific cell or family of cells should not be optimized, i.e., not to be touched.} setting for the swapped drivers/sinks to avoid logic restructuring/removal of the related nets. Finally, the true connectivity is restored in the BEOL with the help of correction cells that resemble switch boxes. The technique is evaluated using ISCAS\\\\textquotesingle85 circuits, and the results for the largest and smallest circuit are shown in Table~\\\\ref{tab:def_res}.\",\n \"id\": \"ee25dbd5-cfaf-41a3-8225-ebde1d228b40\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"0381ac67-f861-41a5-94dd-d57b1d10cc30\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Split Manufacturing Defenses\"\n ],\n [\n \"subsection\",\n \"Proximity Perturbation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Proximity Perturbation\"\n },\n {\n \"cite_extract_rate\": 0.07692307692307601,\n \"cites\": [\n 7800\n ],\n \"content\": \"Hiding routing information from untrusted foundries is the main objective of the Split Manufacturing technique. Since attacks mainly rely on hints left by EDA tools to recover the missing BEOL connections, the amount of hidden information is related to the circuit performance -- splitting the circuits at low metal layers increases the security level. Following the same idea, wire lifting proposes `lifting' wires from the FEOL layer to the BEOL. That is, changing the routing to split metal layers has the potential to increase the security level. \\n \\\\begin{table*}[htb]\\n \\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Results for Defense Techniques based on Wire Lifting.}\\n \\\\begin{tabular}{cp{1cm}p{1.5cm}p{2.5cm}llp{1cm}p{1.5cm}p{1.5cm}}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Work} & \\\\textbf{Attack Type} & \\\\textbf{Benchmark} & \\\\textbf{Defense Technique} & \\\\textbf{Defense Metric} & \\\\textbf{Defense Overhead} & \\\\textbf{Split Layer} & \\\\textbf{Result without Defense} & \\\\textbf{Result with Defense} \\\\vspace{5pt}\\\\\\\\\\n \\\\hline\\n & SAT & c432 & Wire Lifting & \\\\textit{k-security} & 477\\\\% of WLO & -* & k=1 & k=48 \\\\\\\\\\n & Proximity & Superblue 1 & Routing Blockage Insertion & $E[LS]$ & Not Presented & M4 & 1.51 & 1.77 \\\\\\\\\\n & Proximity & Superblue 1 & Routing Blockage Insertion & $FOM$ & Not Presented & M4 & 1222.8 & 1433 \\\\\\\\\\n & Proximity & c432 & Concerted Lifting & Hamming Distance & 7.7\\\\% of Area & Average** & 23.4 & 45.9 \\\\\\\\\\n & Proximity & c432 & Concerted Lifting & CCR & 13.2\\\\% of Power & Average** & 92.4 & 0 \\\\\\\\\\n & Proximity & c7552 & Concerted Lifting & Hamming Distance & 16.7\\\\% of Area & Average** & 1.6 & 25.7 \\\\\\\\\\n & Proximity & c7552 & Concerted Lifting & CCR & 9.3\\\\% of Power & Average** & 97.8 & 0 \\\\\\\\\\n & Proximity & c2670 & CMP-Friendly & Hamming Distance & 3.4\\\\% of WLO & -* & 14.5\\\\% & 20.4\\\\% \\\\\\\\\\n & Proximity & c2670 & CMP-Friendly & CCR(\\\\%) & 3.4\\\\% of WLO & -* & 48.1\\\\% & 33.4\\\\% \\\\\\\\\\n & Proximity & b18 & CMP-Friendly & Hamming Distance & 0.4\\\\% of WLO & -* & 21.6\\\\% & 27.6\\\\% \\\\\\\\\\n & Proximity & b18 & CMP-Friendly & CCR(\\\\%) & 0.4\\\\% of WLO & -* & 12.1\\\\% & 10.7\\\\% \\\\\\\\\\n & Proximity & c2670 & SADP-Compliant & Hamming Distance & 7.49\\\\% of WLO & -* & 14.5\\\\% & 24.4\\\\% \\\\\\\\\\n & Proximity & c2670 & SADP-Compliant & CCR(\\\\%) & 7.49\\\\% of WLO & -* & 48.1\\\\% & 6.4\\\\% \\\\\\\\\\n & Proximity & b18 & SADP-Compliant & Hamming Distance & 4.64\\\\% of WLO & -* & 21.6\\\\% & 29.6\\\\% \\\\\\\\\\n & Proximity & b18 & SADP-Compliant & CCR(\\\\%) & 4.64\\\\% of WLO & -* & 12.1\\\\% & 2.7\\\\% \\\\\\\\ \\\\hline\\n & Proximity & s526 & Net Partitioning & CCR(\\\\%) & Not Presented & -* & 40\\\\%*** & 0\\\\%*** \\\\\\\\\\n & Proximity & s526 & Net Partitioning \\\\& Cell Hiding & CCR(\\\\%) & Not Presented & -* & 40\\\\%*** & 0\\\\%*** \\\\\\\\\\n & Proximity & s526 & Net Partitioning \\\\& Cell Hiding \\\\& Pin Shaking & CCR(\\\\%) & Not Presented & -* & 40\\\\%*** & 0\\\\%*** \\\\\\\\\\n & Proximity & s9234.1 & Net Partitioning & CCR(\\\\%) & Not Presented & -* & 30\\\\%*** & 4\\\\%*** \\\\\\\\\\n & Proximity & s9234.1 & Net Partitioning \\\\& Cell Hiding & CCR(\\\\%) & Not Presented & -* & 30\\\\%*** & 1.5\\\\%*** \\\\\\\\\\n & Proximity & s9234.1 & Net Partitioning \\\\& Cell Hiding \\\\& Pin Shaking & CCR(\\\\%) & Not Presented & -* & 30\\\\%***& 1.5\\\\%*** \\\\\\\\\\n \\\\hline \\n \\\\hline\\n \\\\multicolumn{9}{l}{\\\\cellcolor{white}* Split layer not specified by the authors.} \\\\\\\\\\n \\\\multicolumn{9}{l}{\\\\cellcolor{white}** Results are given as an average between M3, M4, and M5.} \\\\\\\\\\n \\\\multicolumn{9}{l}{\\\\cellcolor{white}*** These results cannot be directly compared with previous ones as the transistor technology is vastly different.} \\n \\\\end{tabular}\\n \\\\label{tab:def_res2}\\n\\\\end{table*} \\nWire lifting was first presented by Imerson \\\\etal where Split Manufacturing is considered as a 3D IC implementation . For the sake of argument, we will continue to refer to this technique as Split Manufacturing, even if the notion of untrusted FEOL vs. trusted BEOL is shifted. This type of 3D implementation consists of two or more independently manufactured ICs, where each IC represents a tier that is vertically integrated on top of each other. Connections between the tiers are done using vertical metal pillars, referred to as through-silicon vias (TSVs). In , a 3D implementation consisting of two tiers is used for their experiments. The bottom tier containing the transistors and some routing wires (akin to the FEOL), and the top tier, containing only routing wires (akin to the BEOL). Regarding the manufacturing of these 3D ICs, the bottom tier is built in a high-end untrusted foundry, and the top tier is built in an also untrusted foundry (not necessarily high-end, however).\\nIn , threat model II is used, i.e., the adversary is assumed to possess the entire netlist. The problem is formulated as the attacker being the FEOL foundry, which in turn also possesses the so called `unlifted netlist' extracted from the FEOL layout. By utilizing a graph to represent the circuits as previously described, the attacker seeks a bijective mapping of gates of the unlifted netlist to gates in the complete netlist. According to , if the attacker can distinguish any gate between the two netlists, the split circuit does not provide any security. A security notion is provided by the authors, based on existing multiple mapping between gates in the unlifted and complete netlists. Referred to as \\\\textit{k-security}, this metric qualifies that gates across the design are indistinguishable from at least $k-1$ other gates. Thus, a defender wants to lift wires in a way to guarantee the higher $k-security$ possible. Two procedures are proposed to achieve this goal, one utilizing a greedy heuristic targeted at small circuits (due to scalability issues), and another procedure that utilizes partitioning to solve those issues. For their experimental study, they have utilized the ISCAS\\\\textquotesingle85 benchmark suite and a DES crypto circuit with approximated 35000 gates. The results are shown in Table~\\\\ref{tab:def_res2}, where $k=1$ is the original circuit and $k=48$ is achieved when all the wires are lifted. It is worth to mention that; besides the notion of the security metric, their defense technique was not validated using an actual proximity attack towards the modified netlist.\\nAn artificial routing blockage\\\\footnote{This terminology is used in IC design to mean that a specific area should be avoided by the EDA tool for a specific task. A blockage can be for placement and/or for routing.} insertion that promotes wire lifting is proposed by Maga\\\\~na \\\\etal . The goal of this technique is to deceive proximity attacks by wire lifting. As discussed before, the objective of commercial EDA tools is to guarantee the best PPA possible. During the routing stage, lower metals are preferred for signal routing, promoting better PPA. Thus, routing blockages can be inserted at the split layer, forcing signals to be routed above the split layer. The result is an artificial wire lifting done during the routing stage. \\nApplying this type of procedure must be done considering the design routability and overhead introduced, as well as top level floorplan decisions for the power grid, clock distribution, and resources for busses. Larger designs are generally difficult to be routed -- simply reducing the number of routing layers can make the design unroutable. In , a procedure is proposed to insert routing blockages ensuring the design routability is kept. After a first routing stage, the design is divided into small rectangular non-overlapping windows. The routing congestion is then analyzed in each window at the split layer for the blockage insertion. If the area has capacity for more routing, a routing blockage is inserted, otherwise the original routing is kept. Utilizing ISPD\\\\textquotesingle11 circuits, the technique is evaluated using the proximity attack proposed by , and its effectiveness is measured using two metrics, $E[LS]$ and $FOM$. The $E[LS]$ metric reports the candidate list size, being an average over different search areas. The $FOM$ metric is a figure of merit obtained from the ratio of candidates list size divided by the search area, when averaged over all the search areas at the split layer. According to , a higher value of $FOM$ means it is more challenging for an attack to be mounted because of the density of candidates (over the same search area). The results for the \\\\textit{Superblue 1} circuit are shown in Table~\\\\ref{tab:def_res2}.\\nDesign for Manufacturability (DFM) has become an extremely important aspect of IC design for many years now. Manufacturing an IC is a sensitive process that involves many critical steps. Hence, a layout is required to be compliant to several rules to ensure its manufacturability. A layout is said to be manufacturable if there are no DRC violations. However, for a design to also achieve high yield, the layout must also pass strict DFM checks. The most common checks are related to wire and via density over predetermined region sizes. Until now, defense techniques discussed were mainly concerned about security and PPA overheads. Feng \\\\etal argued that previous works have largely neglected manufacturability concerns. Therefore, they proposed two wire-lifting techniques that address two important DFM-related techniques: Chemical Mechanical Planarization (CMP) and Self-Aligned Double Patterning (SADP) . The first technique, CMP-friendly routing defense is divided into layer elevation, wire selection, and re-routing. Layer elevation selects wires for lifting according to following principles :\\n \\\\begin{itemize}\\n \\\\item The wire has a significant logic difference from its neighboring wires. As such, an incorrect connection in attacking this wire may lead to more signal differences.\\n \\\\item The wire has large observability such that an erroneous guess by the adversary can easily affect the circuit primary output signals.\\n \\\\item The wire segment is originally at a wire-dense region. The wire density of this region would be reduced by the layer elevation and makes the corresponding FEOL layer have more uniform wire density.\\n \\\\item The BEOL region where the wire is elevated to has low wire density so that the density of the corresponding BEOL layer is more uniform.\\n \\\\end{itemize}\\nPrinciples 1 and 2 have the goal to increase security in the same way as described in . After the wire lifting step, a set of wires is selected for re-routing. The selection has two purposes, CMP-friendliness and security improvement. For CMP-friendliness, wires located in dense regions are selected to be re-rerouted in sparse areas. For the security improvement, decoys are inserted if the routing detour passes through a sparse area. A suspicious attacker may realize that the detour is a defense measure. After selecting the set of wires to be re-routed, wires are re-routed one at a time. According to , their routing approach considers wire density, while the routing perturbation proposed by can be solely focused on security, and may not be CMP-friendly. Utilizing a graph representation, their re-routing method is based on the Dijkstra's shortest path algorithm where the density of wires is used as a metric. \\nWith a few exceptions, the SADP-compliant routing defense follows the same approach as described above. During wire lifting, the density is not considered. Wire re-routing is actually wire extension of FEOL wires as in . This wire extension of FEOL wires inevitably leads to re-routing of connected BEOL wires. According to , solving SADP violations by wire extension can also increase security, as its increase the distance between vias. The wire extension for simultaneous SADP-compliance and security is realized using Integer Linear Programming. In their experiments, ISCAS\\\\textquotesingle85 and ISPD\\\\textquotesingle11 are used to evaluate their techniques. Each technique, CMP-friendly and SADP-compliant routing, is evaluated separately. The results for the smallest and largest circuits are shown in Table~\\\\ref{tab:def_res2}.\\nWire lifting approaches, in general, are not cost-free. As shown in the discussed results, wire-lifting based defenses introduce a considerable PPA overhead. An approach to establish a cost-security trade-off is proposed by Paitinak \\\\etal , i.e., PPA margins for a given security budget. In , a concerted wire-lifting method is proposed. The authors claim to enable higher degrees of security while being cost-effective. For their method, custom elevating cells are used for executing the wire-lifting. Elevating cells connect gates or short wires directly to the first layer above the split layer. Their wire-lifting method utilizes three strategies: lifting high-fanout nets, controlling the distance for open pin pairs, and obfuscation of short nets. High-fanout nets are chosen to be lifted for two reasons: (a) a wrong connection made by the attacker propagates the error to multiple locations, and, (b) introduces multiple open pin pairs. As the attack to overcome is the proximity one, controlling the distance between open pin pairs is necessary, which is achieved at will simply by controlling the placement of the elevating cells. According to , short nets may be easy for an attacker to identify and localize (from assessing driving strengths). Short wires are obfuscated by inserting an elevating cell with two pins close to each other, one being the true connection and the other a dummy connection. Finally, wires are lifted according to those strategies until a given PPA budget is reached. For their experimental study, ISCAS\\\\textquotesingle85 and ISPD\\\\textquotesingle11 circuits are utilized. However, results for attacks are presented only for ISCAS\\\\textquotesingle85 circuits. For ISPD\\\\textquotesingle11, only the PPA impact result introduced by their technique is presented. Once again, we present the results for the smallest and largest of the studied circuits in Table~\\\\ref{tab:def_res2}.\\nWhile the majority of studies reported in our survey make use of conventional transistors (bulk CMOS technologies with either planar or FinFET transistors), Yang \\\\etal proposed a design methodology to secure Split Manufacturing for Vertical Slit Field Effect Transistor (VeSFET)-based integrated circuits. VeSFET is a twin-gate device with a horizontal channel and four metal pillars implementing vertical terminals . While a detailed explanation on VeSFETs is beyond the scope of this work, we do highlight the differences between VesFETs and conventional transistors. In contrast with conventional transistors, a VeSFET can be accessed by both top and bottom pillars, allowing two-side routing and offering a friendly monolithic 3D integration . While we have so far considered ICs that have two distinct layers, the FEOL and BEOL, a VeSFET-based IC has tiers of the layer containing the transistors. Connections between tiers can be made directly, same as TSV by the pillars, or by a layer containing connections between tiers. A 2D VeSFET design contains only one tier and both top and bottom connections, whereas a 3D design contains two or more tiers. In summary, the notion of tier is pushed down to the transistor level in this device topology, thus making it an interesting platform for Split Manufacturing.\\nThe method proposed by assumes that both foundries are untrusted and have the same capability (i.e., same technology). For 2D designs, the first foundry manufactures the tier with the top connections, comprising most of the connections. Then, the rest of the bottom connections, comprising of the critical minority connections, are completed by the second foundry. For 3D IC designs, they proposed special types of standard cells, referred as cell A and B. Cell A has two tiers that are visible and manufactured by the first foundry, as well as inter-tier connections. Cell B has only the top tier visible and manufactured by the first foundry, the low tier is completed by the second foundry, without inter-tier connections. Thus, transistors can be hidden from the first foundry as a security feature. Vulnerabilities claimed by for both 2D and 3D methods are described in Table~\\\\ref{tab:vesfet_vul}. Practices of reverse engineering and IC overbuilding are claimed to be impossible because the first foundry controls the number of wafers available to the second foundry.\\nIncreasing the security of both 2D and 3D VesFET designs is achieved by net partitioning, and exclusively for 3D designs, by transistor hiding and pin shacking. Net partitioning is performed similarly to the wire lifting techniques described above, where nets are chosen to be routed in the bottom connection layer, thus, hiding those from the first foundry. Their selection method is done by selecting nets from sequential logic. First, all the high-fanout nets are selected to be partitioned. Next, the remaining nets are selected by a search area, where two approaches are used, distance-first search and high-fanout first search. In distance-first method, a pin in a predefined search window connecting to an un-partitioned net is selected when it has the minimum distance to the currently processed pin pair. The FO-first search method selects the pin connecting to a net having the highest FO in the searching window. Transistor hiding in 3D designs is done by utilizing cells similar to the cell B. Cells connected only by partitioned nets are candidates for hiding. After selecting the candidates, availability of unused transistors that are accessible to the second foundry in the lower tiers of the nearby cells is checked. If the available transistor count is sufficient, then the cell is hidden. The empty space created could provide clues for the first foundry about the security technique. Pin shaking is then applied to obfuscate the empty spaces. Some nearby cells are moved to this area to obfuscate the layout for any distance-based proximity attackers. In , 10 MCNC LGSynth\\\\textquotesingle91 benchmark circuits are used to evaluate the effectiveness of their methodology. The best and worst results are shown in Table~\\\\ref{tab:def_res2}. It is worth to mention that, even though the VesFET implementation mimics the layered structure of Split Manufacturing, the results cannot be compared side by side in a fair manner.\\n \\\\begin{table}[htb]\\n \\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Vulnerabilities of Split Manufactured VesFET-based Designs Described by .}\\n \\\\begin{tabular}{lll}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Threats} & \\\\textbf{1st Foundry} & \\\\textbf{2nd Foundry} \\\\vspace{5pt} \\\\\\\\\\n \\\\hline\\n \\\\parbox{1cm}{Design\\\\\\\\ Reconstruction} & \\\\parbox{2.5cm}{2D IC: Very Difficult \\\\\\\\ 3D IC: Impossible} & \\\\parbox{2cm}{Impossible due to a very limited\\\\\\\\ information} \\\\\\\\\\n Trojan Insertion & \\\\parbox{2.5cm}{Possible, but will be\\\\\\\\ detected} & \\\\parbox{2cm}{No control of devices} \\\\\\\\\\n \\\\parbox{1cm}{Reverse\\\\\\\\ Engineering} & Meaningless & Impossible\\\\\\\\\\n IC Overbuilding & Meaningless & Impossible \\\\\\\\\\n \\\\hline\\n \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:vesfet_vul}\\n \\\\end{table}\",\n \"id\": \"90e5175d-afc8-4844-8dfa-513e3fcc4fbd\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"0381ac67-f861-41a5-94dd-d57b1d10cc30\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Split Manufacturing Defenses\"\n ],\n [\n \"subsection\",\n \"Wire Lifting\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Wire Lifting\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The main goal of Split Manufacturing -- to hide sensitive information from untrusted foundries -- is compromised once we start to consider more regular structures such as memory. Even without knowing where all the routing goes to, an attacker can easily identify regular structures just by looking at the FEOL layout, possibly leading to easier attacks. Mitigating attacks towards regular structures could be done by obfuscating those structures in such a way that they become indistinguishable. In this section, we discuss works that propose layout obfuscation techniques to be used in a Split Manufacturing context.\\nDuring the development of a modern IC, third-party IPs are sought to close a technological gap or to minimize time-to-market. IPs are typically categorized as soft and hard IPs: soft IPs typically come in code form, giving the customer flexibility to modify the IP such that it meets a given specification during synthesis. Therefore, soft IPs do not present a direct challenge for a Split Manufacturing design flow. Perhaps, and on a very specific scenario, a given IP can facilitate a proximity attack because it promotes certain library cells over others (i.e., it leads to a biased composition).\\nOn the other hand, hard IPs are completely designed by the vendor and are technology dependent. In some instances, the vendor only provides an abstract of the IP; the customer then has to rely on the foundry to replace the abstract by the actual layout. Thus, splitting a hard IP is not trivial. Additional information is needed to be provided by the vendor, which is not guaranteed to be provided, making the IP completely incompatible with Split Manufacturing. Even when the customer holds the entirety of the IP layout, differences between the FEOL foundry and the BEOL foundry could make the IP no longer compliant and therefore virtually useless. Furthermore, defense techniques cannot be applied due to the lack of information or lack of feasibility. Hard IPs, such as embedded memories and specialized analog circuits, have been heavily optimized for maximum compactness, performance and yield. In today's IP market, there is still little concern with security in general, so it is not conceivable that any vendors will start to offer split IP any time soon.\\nThe security of hard IPs in a Split Manufacturing context was first analyzed by Vaidyanathan \\\\etal . A recognition attack flow was proposed for this purpose. An attacker holding the FEOL layer starts his attack by isolating a target embedded memory or analog hard IP. Since the targeted hard IP has a high probability of being constructed by compilation of leaf-cells, layout pattern recognition software can be used for trivial leaf-cell identification. After recognizing all the leaf-cells, the attacker attempts to infer the missing BEOL connections. Using proximity hints together with the knowledge about the regularized structure, the connections have a high likelihood to be guessed correctly. Demonstrated in , embedded memories, such as SRAM, are susceptible to the proposed recognition attack. Defending against recognition attacks can be achieved by means of obfuscation. According to , SRAM IPs can be obfuscated by the following methods:\\n \\\\begin{itemize}\\n \\\\item Randomization of periphery cells, thus avoiding predictable connections. \\n \\\\item Minimization of regularized topologies used for peripheral circuits such as pre-decoders, word line decoders, sense amplifiers, etc.\\n \\\\item Adding non-standard application-specific functions to improve obfuscation and performance.\\n \\\\end{itemize}\\nA synthesis framework is proposed by to obfuscate SRAM IPs. Referred as application-specific SRAM, the methodology synthesizes SRAMs using augmented bitcell arrays and standard cell logic IP (instead of using leaf-cells). Such synthesis, when compared with conventional SRAM compilation, accomplishes all the three obfuscation goals described above while still providing similar performance. \\nAnalog hard IPs are also vulnerable to recognition attacks. In contrast with embedded memories (that are often compiled), analog hard IPs are mostly hand designed to cater for a challenging specification or interface. Even when such degree of customization is employed, the majority of the design is done utilizing leaf-cells (e.g., current mirrors, matched arrays, etc.). Thus, disclosing important information that could be used as leverage for recognition attacks. In , two methods are proposed to defend analog hard IPs against such attacks:\\n \\\\begin{itemize}\\n \\\\item Obfuscation of analog leaf-cells.\\n \\\\item Use of diverse topologies and architectures that enable obfuscation and efficiency. \\n \\\\end{itemize}\\nNext, let us discuss the techniques utilized in order to achieve the goals listed above. First, adding camouflaging dummy transistors in empty spaces can turn leaf-cells indistinguishable. Second, regularizing transistor widths, which allows transistor with different channel lengths to abut each other, thereby obscuring boundaries across different sized transistors. Third, utilizing the same idea behind wire-lifting, routing blockages can be inserted between transistors below the split layer. Such routing scheme would make it difficult to infer the missing BEOL connections, virtually in the same way as it does for a standard-cell based design. \\nTo demonstrate the feasibility and efficacy of their proposed approaches, the authors of designed and fabricated test chips in 130nm technology. For comparison, the same designs were Split Manufactured and conventionally manufactured. Split Manufacturing used Global Foundries Singapore as the untrusted foundry and IBM Burlington as the trusted foundry. Conventional manufacturing was entirely done in Global Foundries Singapore. The first reported design is a smart SRAM that targets an imaging application. Two implementations of a parallel 2x2 access 1Kb SRAM were demonstrated. For conventional manufacturing, the SRAMs were traditionally implemented, and for Split Manufacturing, the SRAMs were implemented using their smart synthesis approach. For their measurements, 10 chips were used to demonstrate the feasibility regarding PPA. Area reported for the split manufactured samples was 75\\\\% of the conventional approach, and, while the power consumption was 88\\\\%. Performance was the same between conventional and split manufactured, i.e., both could work with the same clock frequency. The PPA advantage of Split Manufacturing reported in is not from the manufacturing itself. This advantage is from their smart memory synthesis approach, that was not applied on the conventional manufacturing samples.\\nThe second demonstrated design is a DAC with statistical element selection. The test chip contains a high resolution 15-bit current steering DAC. Only a description of the results is presented, where the authors claim there are tiny measurements differences between the performance of the conventional and the split manufactured, emphasizing that the differences are within measurement noise.\\nAn attacker trying to reverse engineer a split IC will try to recover the maximum number of connections as possible, while minimizing the Time To Evaluate (TTE), i.e., the amount of time needed to reverse engineering the IC. For Jagasivamani \\\\etal , the goal of a designer seeking to secure his design is to create an IC with a high TTE while being cost-effective regarding design effort and PPA overheads. If the TTE is high enough, an adversary would be discouraged from reverse engineering the IC. To achieve this goal, proposed obfuscation methods that do not require any modifications to standard cells nor the implementation of any specialized cell. \\nFour techniques are proposed by for layout obfuscation, (1) limited standard-cell library, (2) smart-dummy cell insertion, (3) isomorphic cells and (4) non-optimal cell placement. Along with the techniques, a set of metrics is presented to help assess the obfuscation level of a design. \\\\textit{Neighbor connectedness}, a measure of how interconnected cells are to their respective neighbors, i.e., how much proximity information is exposed to the attacker. For a specific cell, this metric is computed as how many connections that cell has for a given radius around it. \\\\textit{Standard-cell composition bias}, a metric that addresses the effort required for composition analysis of a design. The bias signature could leave information of the function of the cell. Thus, this metric measure how skewed a design is according to a specific bias cell. In , they utilized three types of bias cells for this analysis: XOR-type, flip-flop type, and adder type of cells. \\\\textit{Cell-level obfuscation}, a metric that measures the percentage of standard-cells that have been obfuscated. \\\\textit{Entropy}, which is similar to the concept of mutual information previously discussed. \\nTechnique (1) aims to achieve obfuscation by reducing the use of complex cells and instead favor only simple cells to compose the design (i.e., to prefer single stage cells over complex multi-stage cells). Removing specialized complex cells could obfuscate functional signatures due to the larger granularity that is employed to construct the cell. However, since the functionality of complex gates will have to be reconstructed through basic cells, a heavy PPA overhead is likely to occur when applying this technique. Technique (2) aims to obfuscate composition analysis by adding dummy cells in such a way that a neutral bias composition is achieved. Dummy cells are inserted as spare cells \\\\footnote{Spare cells are extra logic usually inserted during physical synthesis. These cells are used when an engineering change order (ECO) is required, such that small tweaks to the circuit logic can be performed with minimal changes to placement and routing.}, focusing solely on obfuscating the composition analysis. Technique (3) obfuscates the layout by regularizing the shapes of the cells in a library. All layouts of logic standard cells are made FEOL-identical such that the overall circuit layout appears to be a sea of gates. The functionality of the cells is defined later by the BEOL connections. Thus, the true functionally of the cell is hidden at the BEOL, making cell-level identification harder. Technique (4) employs the same strategy from placement perturbation discussed before. \\nFor their experimental study, the authors of made use of a multiplier block with a high number of adder cells and a crypto-like circuit. Experiments were separated into limited library and smart dummy insertion. Results are shown as a percentage relative to the baseline circuit, i.e., without any protection approach applied. Neighbor connectedness (\\\\%) for a radius $\\\\leq 25nm$ decreased substantially for both test cases and circuits (for more information see ). Overhead results are shown in Table~\\\\ref{tab:over_javs}, where the figures presented are normalized with respect to the baseline circuit.\\n \\\\begin{table}[htb]\\n \\\\centering\\n \\\\caption{Impact on Performance from the Defense Approaches of .}\\n \\\\begin{tabular}{cccc}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Benchmark} & \\\\textbf{Metric} & \\\\textbf{Limited Library} & S\\\\textbf{mart Dummy} \\\\vspace{5pt}\\\\\\\\\\n \\\\hline\\n \\\\multirow{2}{*}{ mult24} &\\\\cellcolor[gray]{0.9} Area & \\\\cellcolor[gray]{0.9} 94.9\\\\% & \\\\cellcolor[gray]{0.9} 72.6\\\\% \\\\\\\\\\n & Timing Slack & -64.8\\\\% & 3.4\\\\% \\\\\\\\\\n \\\\multirow{2}{*}{a5/1} & \\\\cellcolor[gray]{0.9} Area & \\\\cellcolor[gray]{0.9} 69.8\\\\% & \\\\cellcolor[gray]{0.9} 69.4\\\\% \\\\\\\\\\n & Timing Slack & -27.2\\\\% & -1.2\\\\% \\\\\\\\ \\n \\\\hline \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:over_javs}\\n \\\\end{table}\\nUtilizing exactly the same concepts and metrics described in , Otero \\\\etal proposed a ``trusted split-foundry toolflow'' based on cellTK . The concept of the cellTK-based flow is to have on-demand cell generation from a transistor-level netlist. This is heavily in contrast with a traditional ASIC flow that relies on a predefined (and thoroughly characterized) cell library. Leveraging cellTK, proposed an extension referred as split-cellTK. This extension can generate multiple physical layouts for each unique cell without modifying the circuit topology, which is then used to implement obfuscation strategies. Two strategies are proposed, referred as Uniform and Random. The Uniform strategy tries to standardize the size and spacing between cells by inserting dummy transistors to equalize the number of nMOS and pMOS devices and, after the cell placement, dummy cells are inserted in empty gaps. A Random strategy is also proposed in order to reduce the overheads introduced by the Uniform strategy. Instead of deliberately standardizing the size and spacing between cells, a specific number of empty spaces is chosen for these tasks. A more in-depth explanation about their strategies is beyond the scope of this work because they are closely related to cellTK itself. However, their goals and evaluations are the same as in . For their experimental study, they utilized the island-style asynchronous FPGA developed in . A test chip was Split Manufactured in 65nm and the design was synthesized with a cellTK-based approach, i.e., without any defense strategy. Their defense strategies were evaluated only by simulation. The performance results for the baseline, Uniform, and Random strategies, when applied to obfuscate an adder, are shown in Table~\\\\ref{tab:over_celltk}. Trustworthiness results are given in terms of neighbor connectedness and, for all implementations discussed, neighbor connectedness results were significantly smaller than the results reported in .\\n \\\\begin{table}[htb]\\n \\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Performance Results Reported by to Obfuscate an Adder Circuit.}\\n \\\\begin{tabular}{lllll}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\parbox{1cm}{\\\\textbf{Technique}} & \\\\parbox{0.6cm}{\\\\textbf{Area}\\\\\\\\ ($\\\\mathbf{\\\\mu m^2}$)} & \\\\parbox{0.6cm}{\\\\textbf{Power (mW)}} & \\\\parbox{0.6cm}{\\\\textbf{Energy (pJ)}} & \\\\parbox{0.7cm}{\\\\textbf{Perf. (MHz)}} \\\\vspace{5pt} \\\\\\\\\\n \\\\hline\\n Baseline & 462 & 0.146 & 0.257 & 568 \\\\\\\\\\n Uniform & 717 & 0.149 & 0.307 & 486 \\\\\\\\\\n Random & 760 & 0.164 & 0.303 & 542 \\\\\\\\\\n \\\\hline\\n \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:over_celltk}\\n \\\\end{table}\\nIn the context of obfuscation, but also generally for Split Manufacturing, a higher level of security is achieved when the chosen layer to perform the split is the lowest possible. Xiao \\\\etal pointed out that splitting at lower metal layers could increase the cost to manufacture the IC; it is argued that the FEOL-BEOL integration process must be more `precise' for correct alignment. Thus, a closer technology match between the trusted and untrusted foundries is required. As we previously argued, if the goal is to make use of the best silicon available from an untrusted foundry, the implication is that the trusted foundry cannot provide a legacy technology, but perhaps a mature yet still relevant technology is sufficient. In , a methodology for obfuscating the layout is proposed for split at M3 or higher, meanwhile keeping the cost as low as possible and at the same time providing a high level of security. Their strategy is similar to the insertion of dummy cells; however, functional cells are inserted instead. Referred as obfuscated built-in self-authentication (OBISA) cells, the inserted functional cells are connected together to form a circuit. As the circuit is connected to the original circuit it is trying to protect, they claim this fact makes it extremely difficult for an attacker to separate the OBISA design from the original design. The idea behind OBISA is to obfuscate the layout by hindering neighbor connectedness analysis and standard-cell composition bias analysis while also perturbing the proximity between gates. As illustrated in Figure~\\\\ref{fig:obisa}, two additional functional cells, O1 and O2, were placed between gates G2 and G3, and, G3 and G4, respectively. The insertion of these additional functional cells could deceive proximity attacks, assuming that the EDA tool would place the gates between OBISA cells farther apart than in the original circuit.\\nThe proposed OBISA circuitry has two operating modes: functional and authentication. During functional mode, the OBISA circuitry stays idle and incoming signals and clock are gated/blocked. Thus, the original circuit is not affected by OBISA operating as it should. As the name suggests, when in authentication mode, OBISA is used to verify the trustworthiness of the manufactured IC (in the field). The specifics of the authentication are beyond the scope of this work and will not be discussed. The insertion of OBISA cells follows a similar strategy of dummy cell insertion as discussed in . The connections of the inserted cells are done in a way to promote the testability of the OBISA circuit and increase the obfuscation strength. Their approaches were evaluated using benchmark circuits from OpenCores. Results for the smallest and largest circuits are shown in Table~\\\\ref{tab:obisa_res}. \\n \\\\begin{table}[htb]\\n \\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Implementation Results from .}\\n \\\\begin{tabular}{lllll}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Benchmark} & \\\\parbox{0.8cm}{\\\\textbf{Gate count}} & \\\\parbox{1.5cm}{\\\\textbf{OBISA cell count}} & \\\\textbf{Total nets} & \\\\textbf{Nets} $\\\\mathbf{\\\\geq M4}$ \\\\vspace{5pt}\\\\\\\\\\n \\\\hline\\n DES3 & 1559 & 158 & 1799 & 127 \\\\\\\\\\n DES\\\\_perf & 49517 & 2090 & 49951 & 1343 \\\\\\\\\\n \\\\hline \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:obisa_res}\\n \\\\end{table}\\n \\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\\\linewidth]{obisa_cir.pdf}\\n{Circuit representation with OBISA cells (square cells) inserted (adapted from ). \\\\label{fig:obisa}}\\nAnother study using look-alike cells is reported by Masoud \\\\etal where the goal remains to make the attacker unable to distinguish cells and their inputs/outputs, thus mitigating attacks to some degree. In this study, two types of search algorithms are proposed to replace cells for isomorphic cells. In contrast with where all cells are replaced, in , only cells with high impact on the security are replaced. Thus, the overhead introduced by cell replacement can be controlled (i.e., a trade-off is established). The proposed algorithms are based on `gate normalization', whereby truth tables of cells are analyzed in order to balance the occurrence of 0s and 1s (e.g., XOR and XNOR gates are normalized by definition). An analysis is made by replacing existing gates by XORs and comparing the deviation from the original circuit. If the deviation is larger than a given deviation threshold, the gates are effectively replaced. \\nA novel layout obfuscation framework is proposed by Li \\\\etal which builds on the wire lifting concept of . According to the authors, wire lifting alone is not enough to secure a design. If an attacker can tell the functionality of a specific gate that had its wires lifted, the security is already compromised. To address this problem, a framework that considers dummy cells and wire insertion simultaneously with wire lifting is proposed. As in , threat model II was used. The proposed framework makes use of mixed-integer linear programming (MILP) formulation for the FEOL layer generation and a Lagragian relaxation (LR) algorithm to improve scalability. The generation of the new FEOL layout considers three operations: wire-lifting, dummy cell insertion, and dummy wire insertion. Dummy wire insertion is done only on dummy cells; thus, the original functionality of the circuit is guaranteed to remain and floating pins are avoided. Utilizing a graph representation, they re-formulate the security metric to accommodate dummy cell and dummy wire insertion. Since the original graph isomorphic relationship is lost when new nodes are inserted, a new approach has to be used to formalize the relationship between the original and the new FEOL; this concept is denoted as k-isomorphism and the associated security analysis is denoted as \\\\textit{k-security}. In their experimental study, TrustHub trojan insertion methods are used to select the nodes for protection. They used ISCAS\\\\textquotesingle85 benchmark circuits together with functional units (shifter, alu, and div) from the OpenSPARC T1 processor . Comparison between MILP and LR algorithms are done for several \\\\textit{k-security} levels, and the results are given in terms of area overhead (AO) and wire-length overhead (WLO). The results for a few of the security levels are shown in Table~\\\\ref{tab:li_milplr}. \\n \\\\begin{table}[htb]\\n \\\\rowcolors{2}{gray!25}{white}\\n \\\\centering\\n \\\\caption{Comparison Between MILP and LR Algorithms for the c4232 circuit .}\\n \\\\begin{tabular}{clll}\\n \\\\hline \\\\hline \\\\\\\\ [-1.5ex]\\n \\\\textbf{Security Level } & \\\\textbf{Algorithm} & \\\\textbf{AO(\\\\%)} & \\\\textbf{WLO(\\\\%)} \\\\vspace{5pt}\\\\\\\\\\n \\\\hline\\n 15 & MILP & 18& 180\\\\\\\\\\n 20 & MILP & 41 & 220\\\\\\\\\\n 25 & MILP & 58 & 295\\\\\\\\\\n 15 & LR & 18& 200\\\\\\\\\\n 20 & LR & 40 & 230\\\\\\\\\\n 25 & LR & 60 & 305\\\\\\\\\\n \\\\hline \\\\hline\\n \\\\end{tabular}\\n \\\\label{tab:li_milplr}\\n \\\\end{table}\",\n \"id\": \"4032829b-2e93-4b12-822b-7b2a6e653dbc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"0381ac67-f861-41a5-94dd-d57b1d10cc30\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Split Manufacturing Defenses\"\n ],\n [\n \"subsection\",\n \"Layout Obfuscation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Layout Obfuscation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:trends}\\nDespite our effort to present the results of the many studied papers in the most fair way possible, it is clear that the hardware security community lacks a \\\\textit{unified benchmark suite} and/or a \\\\textit{common criteria} for assessing results. Often, researchers make use of benchmark suites that are popular in the Test community but have no real applicability in security. For instance, the ISCAS\\\\textquotesingle85 suite has no crypto cores in it, which are the bread and butter of the research in the area. Furthermore, we believe the community would largely benefit from using circuits that better represent IC design practices of this decade where IPs often have millions of gates and ICs have billions of transistors.\\nWhile the lack of a common criteria is an issue for the academic community, the lack of an industry-supported path for Split Manufacturing is even more troubling. Today, more than ever, foundries compete for the title of `best silicon' and rarely engage in cross-foundry cooperation. Efforts of the past, such as the now defunct Common Platform of IBM, Samsung and GF, could have been a catalyzer for the adoption of Split Manufacturing. Without such collaboration, it is hard to foresee a future where the technique will gain traction again. Furthermore, the study of DFM-related implications of the technique is really cumbered by the fact that we cannot measure yield from massively produced Split Manufactured chips. \\nWe have discussed in details how many attacks leverage heuristics and hints left behind by the EDA tools. Many of these hints are very logical and can be appreciated, even graphically, as we illustrated in Figure~\\\\ref{fig:prox_sa}. It is entirely possible that machine learning approaches can detect subtle biases in the tools that are not easy to appreciate graphically. There is no consolidated knowledge of what these biases are and to which extent machine learning is effective in detecting them. This avenue of research is certainly interesting and we believe it will be the target of many papers in the near future.\\n \\\\Figure[t!](topskip=0pt, botskip=0pt, midskip=0pt)[width=0.95\\\\linewidth]{pie_chart.pdf}\\n{Techniques validated in silicon among presented works. \\\\label{fig:pizza}}\\nIt is also worth discussing the attack models that have been proposed so far. We have previously highlighted how strong Threat model II is, but the fact of the matter is that defining the threat model is a complicated exercise in which we seek to establish what are the capabilities of the attacker. By definition, formalizing the capabilities of an attacker requires understanding his motivations, technical proficiency, and availability of resources. If the attacker is underestimated, useless defense strategies can be devised and assumed to be effective. If the attacker is overestimated, convoluted defense strategies might be employed, leading to unnecessary PPA overheads. This is a challenge for Split Manufacturing and many other techniques that promote obfuscation.\\nAnother topic that has led to no consensus is whether an attacker can make use of a partially recovered netlist. For instance, let us assume a design that instantiates the same block multiple times. If one of the blocks is correctly recovered, perhaps a cursory inspection of the structure will allow the attacker to recover all other instances of the same block. The same line of thinking can be applied to datapaths and some cryptographic structures that are regular in nature. In a sense, an analysis of the functionality of the recovered netlist could be combined with existing attacks for further improvement of correctly guessed connections.\\nWe note that many of the works studied in this survey have not actually demonstrated their approach in silicon. This fact is summarized in Figure~\\\\ref{fig:pizza}. As a community effort, we should strive to validate our approaches in silicon as often as possible. However, as discussed before, finding two foundries willing to diverge from their established practices could be next to impossible. This is likely the main reason that such small percentage of the works herein reported have validated their techniques in silicon.\",\n \"id\": \"5c7f1b9c-d130-400e-ae7c-f7e824d126c3\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4df5dfa6-5506-4133-aae0-08641ac96086\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Future trends and challenges\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Future trends and challenges\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclusion} \\nOur findings showed a big disparity on how the Split Manufacturing technique is approached among the surveyed studies. A variety of benchmark suites and metrics were used for evaluation, making direct comparisons between studies very difficult -- and, in some cases, impossible. In spite of that, we were able to classify the studies, clearly demonstrating the many interpretations of the technique, its attacks, and defenses. Our belief is that this survey assesses the most significant studies about Split Manufacturing as we focused on papers that appear on highly-regarded venues. Results gathered from the surveyed studies were compiled such that main features, metrics, and performance results are available. Regarding the results themselves, these are presented in such manner to illustrate the present state of the technique. Therefore, this work can be very helpful for future researchers to contextualize their own techniques for augmenting Split Manufacturing.\\nOverall, the security of Split Manufacturing is still under debate. Some studies conclude that the technique is indeed secure, and others that it is not. However, these conclusions are reached for different scenarios, i.e., using different benchmark circuits and set of metrics. Creating a unified benchmark suite suitable for Split Manufacturing evaluation, along with a unified set of metrics to quantify/qualify its performance, could facilitate the discussion about its security. In addition, increasing the number of demonstrations in silicon could also help with evaluation and adoption issues related to Split Manufacturing.\\n\\\\bibliographystyle{ieeetr}\\n\\\\bibliography{split_refs}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{perez_t_d.jpg}}]{Tiago D. Perez} received the M.S. degree in electric engineering from the University of Campinas, S\\\\~ao Paulo, Brazil, in 2019. He is currently pursuing a Ph.D. degree at Tallinn University of Technology (TalTech), Tallinn, Estonia. \\nFrom 2014 to 2019, he was a Digital Designer Engineer with Eldorado Research Institute, S\\\\~ao Paulo, Brazil. His fields of work include digital signal processing, telecommunication systems and IC implementation. His current research interests include the study of hardware security from the point of view of digital circuit design and IC implementation.\\n\\\\end{IEEEbiography}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{pagliarini_s.png}}]{Samuel Pagliarini}\\n(M'14) received the PhD degree from Telecom ParisTech, Paris, France, in 2013. \\nHe has held research positions with the University of Bristol, Bristol, UK, and with Carnegie Mellon University, Pittsburgh, PA, USA. He is currently a Professor of Hardware Security with Tallinn University of Technology (TalTech) in Tallinn, Estonia where he leads the Centre for Hardware Security. His current research interests include many facets of digital circuit design, with a focus on circuit reliability, dependability, and hardware trustworthiness.\\n\\\\end{IEEEbiography}\\n\\\\EOD\\n\\\\end{document}\",\n \"id\": \"347259ba-3c12-43d7-b6a0-2402a507ae62\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4df5dfa6-5506-4133-aae0-08641ac96086\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey on Split Manufacturing: Attacks, Defenses, and Challenges\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":91,"string":"91"},"all_cites":{"kind":"list like","value":[7800,7798,7799],"string":"[\n 7800,\n 7798,\n 7799\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.35959075103987004,"string":"0.359591"}}},{"rowIdx":58,"cells":{"authors":{"kind":"list like","value":["Christopher Schröder","Andreas Niekler"],"string":"[\n \"Christopher Schröder\",\n \"Andreas Niekler\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey of Active Learning for Text Classification using Deep Neural Networks"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-08-17T12:53:20Z","string":"2020-08-17T12:53:20Z"},"category":{"kind":"string","value":"cs.CL"},"abstract":{"kind":"string","value":" Natural language processing (NLP) and neural networks (NNs) have both undergone significant changes in recent years. For active learning (AL) purposes, NNs are, however, less commonly used -- despite their current popularity. By using the superior text classification performance of NNs for AL, we can either increase a model's performance using the same amount of data or reduce the data and therefore the required annotation efforts while keeping the same performance. We review AL for text classification using deep neural networks (DNNs) and elaborate on two main causes which used to hinder the adoption: (a) the inability of NNs to provide reliable uncertainty estimates, on which the most commonly used query strategies rely, and (b) the challenge of training DNNs on small data. To investigate the former, we construct a taxonomy of query strategies, which distinguishes between data-based, model-based, and prediction-based instance selection, and investigate the prevalence of these classes in recent research. Moreover, we review recent NN-based advances in NLP like word embeddings or language models in the context of (D)NNs, survey the current state-of-the-art at the intersection of AL, text classification, and DNNs and relate recent advances in NLP to AL. Finally, we analyze recent work in AL for text classification, connect the respective query strategies to the taxonomy, and outline commonalities and shortcomings. As a result, we highlight gaps in current research and present open research questions. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"]],"subsections":["f9d1ed5f-ebd8-41c8-b85b-c14ee4a51e33","10ab6316-68f8-4bea-9503-5f9388a60998","3e095066-9c33-4519-a7ec-ec4550b8cb2d","909d9662-92b7-4a13-b530-6c9553233a2e","847fa582-b0b5-403a-9242-ea4d17738bff","3ba159ca-0eee-4f69-aedd-91724bce7349","ca5cf519-fe65-47a3-9e14-707edbf02174"],"title":"root"},{"cite_extract_rate":0.32,"cites":[7764,1159,7165,1096,4022,8385,1684,11],"content":"\\label{sec:introduction}\nData is the fuel of machine learning applications and therefore has been steadily increasing in value.\nIn many settings an abundant amount of unlabeled data is produced,\nbut in order to use such data in supervised machine learning, \none has no choice but to provide labels.\nThis usually entails a manual labeling process, which is often non-trivial and can even require a domain expert, \ne.g., in patent classification , or clinical text classification .\nMoreover, this is time-consuming and rapidly increases monetary costs, thereby quickly rendering this approach infeasible.\nEven if an expert is available, it is often impossible to label each datum due to the vast size of modern datasets.\nThis especially impedes the field of Natural Language Processing (NLP), \nin which both the dataset and the amount of text within each document can be huge,\nresulting in unbearable amounts of annotation efforts for human experts.\nActive Learning (AL) aims to reduce the amount of data annotated by the human expert.\n\\revised{It is an iterative cyclic process between an {\\em oracle} (usually the human annotator)\nand an {\\em active learner}.}\nIn contrast to passive learning, in which the data is simply fed to the algorithm, \nthe active learner chooses which samples are to be labeled next. \nThe labeling itself, however, is done by a human expert, the so-called human in the loop.\nHaving received new labels, the active learner trains a new model and the process starts from the beginning.\n\\revised{Using the term active learner, we refer to the composition of a {\\em model}, a {\\em query strategy}, and a {\\em stopping criterion}.}\nIn this work the model is w.l.o.g. a text classification model, \nthe query strategy decides which instances should be labeled next, and the stopping criterion defines when to stop the AL loop.\nAccording to \\textcite{settles2010active} there are three main scenarios for AL:\n(1) Pool-based, in which the learner has access to the closed set of unlabeled instances, called the pool;\n(2) stream-based, where the learner receives one instance at a time and has the options to keep it, or to discard;\n(3) membership query synthesis, \\revised{in which the learner creates new artificial instances to be labeled}.\nIf the pool-based scenario operates not on a single instance, but on a batch of instances, this is called {\\em batch-mode} AL .\nThroughout this work we assume a pool-based batch-mode scenario because in a text classification setting the dataset is usually a closed set, and the batch-wise operation reduces the number of retraining operations, which cause waiting periods for the user.\\\\\n\\pindent \\revised{The underlying idea of AL is that few representative instances can be used as surrogate for the full dataset.}\nNot only does a smaller subset of the data reduce the computational costs, \nbut also it has been shown that AL can even increase the quality of the resulting model compared to learning on the full dataset .\nAs a consequence, AL has been used in many NLP tasks, e.g. text classification , named entity recognition , \nor machine translation and is still an active area of research.\nIn recent years, deep learning (DL) approaches have dominated most NLP tasks' state-of-the-art results.\nThis can be attributed to advances in neural networks (NNs), above all Convolutional Neural Networks (CNN; ) and (Bidirectional-)Long Short-Term Memory (LSTM; ),\nwhich were eventually adopted into the NLP domain,\nand to the advances of using word embeddings and contextualized word embeddings .\nBoth NN architectures and text representations have raised the state-of-the-art results in the field of text classification considerably (e.g., ).\nIf these improvements were transferrable to AL, this would result in a huge increase in efficiency.\nFor the AL practitioner, this either means achieving the same performance using fewer samples, or having an increase in performance using the same amount of data.\nAnother favorable development is that transfer learning, especially the paradigm of fine-tuning pre-trained language models (LMs), has become popular in NLP.\nIn the context of AL this helps especially in the small data scenario, in which a pre-trained model can be leveraged to train a model by fine-tuning using only little data, which would otherwise be infeasible.\nFinally, by operating on sub-word units LMs also handle out-of-vocabulary tokens, which is an advantage over many traditional methods.\nResulting from these advances, existing AL surveys have become both incomplete in some parts and outdated in others:\nThey lack comparison against the current state of the art models, do not provide results for more recent large-scale datasets, \nand most importantly, they are lacking the aforementioned advances in NNs and text representations.\nSurprisingly, despite the current popularity of NNs, there is only little research about NN-based active learning in the context of NLP, and even less thereof in the context of text classification (see Section \\ref{subsect:nn_al} and Section \\ref{subsect:active_tc} for a detailed summary).\nWe suspect this is due to the following reasons: \n(1) Many DL models are known to require large amounts of data ,\nwhich is in strong contrast to AL aiming at requiring as little data as possible\n(2) there is a whole AL scenario based on artificial data generation, which unfortunately is a lot more challenging for text in contrast to for example images, for which data augmentation is commonly used in classification tasks ;\n(3) NNs are lacking uncertainty information regarding their predictions \\revised{(as explained in Section \\ref{subsect:nn_al})}, which complicates the use of a whole prominent class of query strategies.\n\\hiddennotes{\n \\begin{itemize}\n \\item No clear evaluation protocol / huge differences in the evaluation approaches\n \\end{itemize}\n}\nThis survey aims at summarizing the existing approaches of (D)NN-based AL for text classification.\nOur main contributions are as follows:\n\\begin{enumerate}\n \\item We provide a taxonomy of query strategies and classify strategies relevant for AL for text classification.\n \\item We survey existing work at the intersection of AL, text classification, and (D)NNs.\n \\item Recent advances in text classification are summarized and related to the AL process. It is then investigated, if and to what degree they have been adopted for AL.\n \\item The experimental setup of previous research is collectively analyzed regarding datasets, models, and query strategies in order to identify recent trends, commonalities, and shortcomings in the experiments.\n \\item \\revised{We identify research gaps and outline future research directions.}\n\\end{enumerate}\n\\noindent Thereby we provide a comprehensive survey of recent advances in NN-based active text classification. \nHaving reviewed these recent advances, we illuminate areas that either need re-evaluation, or have not yet been evaluated in a more recent context.\nAs a final result, we develop research questions outlining the scope of future research.","id":"f9d1ed5f-ebd8-41c8-b85b-c14ee4a51e33","level":"section","origin_cites_number":25,"parent_id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0.33333333333333304,"cites":[1159],"content":"\\textcite{settles2010active} provides a general active learning survey, \nsummarizing the prevalent AL scenario types and query strategies.\nThey present variations of the basic AL setup like \n variable labeling costs or alternative query types, \nand most notably, they discuss empirical and theoretical research investigating the effectiveness of AL: \nThey mention research suggesting that AL is effective in practice \nand has increasingly gained adoption in real world applications. However, it is pointed out that empirical research also reported cases in which AL performed worse than passive learning and that the theoretical analysis of AL is incomplete.\nFinally, relations to related research areas are illustrated, thereby connecting AL among others to reinforcement learning and semi-supervised learning.\n\\hiddennotes{\n settles2009/settles2010active \n \\begin{itemize}\n \\item scenarios: membership query synthesis, stream-based, pool-based\n \\item query strategies\n \\begin{itemize}\n \\item keine Klassifikation vorgenommen\n \\end{itemize}\n \\item most important (at that time): analysis\n \\begin{itemize}\n \\item does it work?\n \\item both empirical and theoretical evidence\n \\end{itemize}\n \\item TODO: problem setting variants, practical considerations (only in 2010 version)\n \\item related areas: semi-supervised, reinforcement, submodular optimization, equivalence query learning, model parroting and compression\n \\end{itemize} \n}\nThe survey of \\textcite{fu2013survey} is focused around a thorough analysis of uncertainty-based query strategies, which are categorized into a taxonomy.\nThis taxonomy differentiates at the topmost level between the uncertainty of i.i.d. instances and instance correlation. The latter is a superset of the former \nand intends to reduce redundancy among instances by considering feature, label, and structure correlation when querying.\nMoreover, they perform an algorithmic analysis for each query strategy \nand order the strategies by their respective time complexity, highlighting the increased complexity for correlation-based strategies.\n\\hiddennotes{\n \\begin{itemize}\n \\item survey on instance selection\n \\end{itemize} \n}\nAnother general survey covering a wide range of topics was conducted by \\textcite{aggarwal2014active}. \nThey provide a flat categorization of query strategies, which is quite different from the taxonomy of \\textcite{fu2013survey} and divides them into the following three categories: (1) ``heterogenity-based'', which sample instances by their prediction uncertainty or dissimilarity compared to existing labeled instances, (2) ``performance-based''', which select instances based on a predicted change of the model loss, and (3) ``representativeness-based'', which select data points to reflect a larger set in terms of their properties, usually achieved by the means of distribution density .\nSimilarly to , they present and discuss many non-standard variations of the active learning scenario.\n\\revised{\n An NLP-focused active learning survey was performed by \\textcite{olsson2009literature}.\n This work's main contribution is a survey of disagreement-based query strategies, which use the disagreement among multiple classifiers to select instances.\n Moreover, Olsson reviews practical considerations, e.g., selecting an initial seed set,\n deciding between stream-based and pool-based scenario, and deciding when to terminate the learning process.\n Although some NN-based applications are mentioned, none of the above surveys covers NN-based AL in depth.\n Besides, none is recent enough to cover NN-architectures, which have only recently been adapted successfully to text classification problems like e.g., KimCNN .\n The same holds true for recent advances in NLP such as word embeddings, contextualized language models (explained in Section \\ref{subsect:tc_recent_adv}), or resulting advances in text classification (discussed in Section \\ref{subsect:nn_text_classification} and Section \\ref{subsect:active_tc}).\n We intend to fill these gaps in the remainder of this survey.\n}\n\\hiddennotes{\n \\begin{itemize}\n \\item NLP-centric Active Learning Survey\n \\item describes multiple active learning modes: Query by uncertainty, query by committee, redundant views\n \\item extensively reviews strategies to measure disagrement\n \\end{itemize}\n}","id":"10ab6316-68f8-4bea-9503-5f9388a60998","level":"section","origin_cites_number":3,"parent_id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Related Work"]],"subsections":[],"title":"Related Work"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sect:active_learning}\nThe goal of AL is to create a model using as few labeled instances as possible,\ni.e. minimizing the interactions between the oracle and the active learner.\nThe AL process (illustrated in Figure \\ref{fig:al_process}) is as follows: The oracle requests unlabeled instances from the active learner ({\\em query}, see Figure \\ref{fig:al_process}: step 1), \nwhich are then selected by the active learner (based on the selected query strategy) and passed to the oracle (see Figure \\ref{fig:al_process}: step 2).\nSubsequently, these instances are labeled by the oracle and returned to the active learner ({\\em update}, see Figure \\ref{fig:al_process}: step 3).\nAfter each update step the active learner's model is retrained, which makes this operation at least as expensive as a training of the underlying model.\nThis process is repeated until a stopping criterion is met (e.g., a maximum number of iterations or a minimum threshold of change in classification accuracy).\n\\begin{figure}[!h]\n \\centering \n \\includegraphics{figures/al_process.pdf}\n \\caption{An overview of the AL process: Model, query strategy, and (optionally a) stopping criterion are the key components of an active learner.\n The main loop is as follows:\n First the oracle queries the active learner, which returns a fixed amount of unlabeled instances.\n Then, for all selected unlabeled unstances are assigned labels by the oracle.\n This process is repeated until the oracle stops, or a predefined stopping criterion is met.}\n \\label{fig:al_process}\n\\end{figure}\n\\noindent The most important component for AL is the query strategy.\nIn the introduction we claimed that a large fraction of query strategies are uncertainty-based.\nTo analyze this we provide a taxonomy of query strategies in the following section and highlight the parts in which uncertainty is involved.\nFor a general and more detailed introduction on AL refer to the surveys of \\textcite{settles2010active} and \\textcite{aggarwal2014active}.","id":"3e095066-9c33-4519-a7ec-ec4550b8cb2d","level":"section","origin_cites_number":0,"parent_id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"]],"subsections":["3effa5dc-9905-4ec7-8c3d-464cfd4bd1d9","aff7fdcf-a8ec-4ce1-9895-2dad247f6097"],"title":"Active Learning"},{"cite_extract_rate":0.33333333333333304,"cites":[1042,4024,4023,11,8716,8717],"content":"\\label{subsec:query_strategies}\nIn Figure \\ref{fig:query_strategies} we classify the most common AL query strategies based on a strategy's {\\em input information}, which denotes the numeric value(s) a strategy operates on.\nIn our taxonomy the input information can be either random or one of data, model, and prediction.\n\\phrasing{These categories are ordered by increasing complexity and are not mutually exclusive.}\n\\phrasing{Obviously, the model is a function of the data, as well as the prediction is a function of model and data, and moreover, in many cases a strategy use multiple of these criteria.}\nIn such cases we assign the query strategy to the most specific category (i.e. prediction-based precedes model-based, which in turn precedes data-based).\n\\begin{figure}[!h]\n {\n \\scriptsize\n \\centering\n \\begin{forest} for tree={\n grow'=0, draw\n },\n forked edges,\n [\\scshape query strategies, no edge, root\n [\\scshape random, onode_dashed\n [,phantom]\n ]\n [\\scshape data-based, onode\n [\\scshape data uncertainty, onode\n [discriminative\\\\ , tier=leaf, tnode]\n ]\n [\\scshape representativeness, onode\n [\\scshape clustering, onode\n [flat\\\\ , tier=leaf, tnode]\n [hierarchical\\\\ , tier=leaf, tnode]\n ]\n [\\scshape set construction, onode\n [core-set\\\\ , tier=leaf, tnode]\n ]\n ]\n ]\n [\\scshape model-based, onode\n [\\scshape model uncertainty, onode\n [UNC-IE\\\\ , tier=leaf, tnode]\n ]\n [\\scshape expected parameter\\\\change, onode\n [expected gradient length\\\\ , tier=leaf, tnode]\n [expected weight change\\\\ , tier=leaf, tnode]\n ]\n [\\scshape adversarial, tnode\n [DFAL\\\\ , tier=leaf, tnode]\n ] \n ]\n [\\scshape prediction-based, name=prediction_based, onode\n [\\scshape prediction\\newline uncertainty, name=prediction_uncertainty, tnode\n [\\scshape probabilistic, onode\n [uncertainty sampling , tier=leaf,tnode]\n ]\n [\\scshape margin-based, onode\n [version space\\\\ , tier=leaf,tnode]\n [closest to hyperplane\\\\ , tier=leaf,tnode]\n ]\n [\\scshape entropy, name=entropy, onode\n [BALD\\\\ , tier=leaf,tnode]\n ]\n ]\n [\\scshape discriminative, onode\n [DAL\\\\ , tier=leaf , tnode]\n ] \n [\\scshape expected prediction change\\\\, onode\n [expected error reduction\\\\ , name=lastnode, tier=leaf, tnode]\n ]\n [\\scshape disagreement, name=disagreement, onode_dashed\n [,phantom, tier=leaf]\n ]\n ] \n ]\n \\draw[zlevel,-] (disagreement.east) -| (entropy.south);\n \\draw[zlevel,-] (entropy.north) |- (prediction_uncertainty.east);\n \\path (current bounding box.south) coordinate (s1);\n \\draw[decorate,decoration={brace,amplitude=1em,mirror}] ([yshift=-3pt]prediction_based.south west|-s1) -- node[below=1em] {class} ([yshift=-3pt,xshift=-1pt]prediction_based.south east|-s1);\n \\draw[decorate,decoration={brace,amplitude=1em,mirror}] ([yshift=-3pt,xshift=1pt]prediction_based.south east|-s1) -- node[below=1em] {subclass(es)} ([yshift=-3pt,xshift=-1pt]lastnode.south west|-s1);\n \\draw[decorate,decoration={brace,amplitude=1em,mirror}] ([yshift=-3pt,xshift=1pt]lastnode.south west|-s1) -- node[below=1em] {example} ([yshift=-3pt,xshift=-1pt]lastnode.south east|-s1);\n \\end{forest}\n \\caption{\\revised{A taxonomy of query strategies for AL.\n The key distinction is at the first level, where the query strategies are categorized by their access to different kinds of input information. \n From the second to the penultimate level we form coherent subclasses,\n and the final level shows examples for the respective class.\n This taxonomy is not exhaustive due to the abundance of existing query strategies, and it is biased towards query strategies in NLP.}\n }\n \\label{fig:query_strategies}\n }\n\\end{figure}\n\\hiddennotes{\n \\begin{itemize}\n \\item Table in appendix: For each strategy here: Applicability: B / MC / ML\\\\\n \\end{itemize}\n}","id":"3effa5dc-9905-4ec7-8c3d-464cfd4bd1d9","level":"subsection","origin_cites_number":18,"parent_id":"3e095066-9c33-4519-a7ec-ec4550b8cb2d","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Query Strategies"]],"subsections":["78f9fdb7-e132-4e81-b968-042f74647827"],"title":"Query Strategies"},{"cite_extract_rate":0.6000000000000001,"cites":[1042,4024,8716],"content":"Randomness has traditionally been used as a baseline for many tasks.\nIn this case, random sampling selects instances at random and is a strong baseline for AL instance selection .\n\\revised{It often performs competitive to more sophisticated strategies, especially when the labeled pool has grown larger .}","id":"78f9fdb7-e132-4e81-b968-042f74647827","level":"paragraph","origin_cites_number":5,"parent_id":"3effa5dc-9905-4ec7-8c3d-464cfd4bd1d9","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Query Strategies"],["paragraph","Random"]],"subsections":["734bf4e0-5809-4e56-9af4-79cebf71ecdb","d53e7115-2a38-4cfc-bd3f-72de86cf317d","c472f10c-b421-401e-aea9-953609e5c153","ff2fa7ea-f6cb-4eae-aca9-41c04d989749"],"title":"Random"},{"cite_extract_rate":0,"cites":[],"content":"Data-based strategies have the lowest level of knowledge, i.e. they only operate on the raw input data and optionally the labels of the labeled pool. \nWe categorize them further into (1) strategies relying on data-uncertainty, which may use information about the data distribution, label distribution, and label correlation,\nand (2) representativeness, \\revised{which tries to geometrically compress a set of points, by using fewer representative instances to represent the properties of the entirety}.","id":"734bf4e0-5809-4e56-9af4-79cebf71ecdb","level":"paragraph","origin_cites_number":0,"parent_id":"78f9fdb7-e132-4e81-b968-042f74647827","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Query Strategies"],["paragraph","Random"],["paragraph","Data-based"]],"subsections":[],"title":"Data-based"},{"cite_extract_rate":0,"cites":[],"content":"The class of model-based strategies has knowledge about both the data and the model.\nThese strategies query instances based on measure provided by the model given an instance. \nAn example for this would be a measure of confidence for the model's explanation of the given instance ,\nfor example, how reliable the model rates encountered features.\nThis can also be an expected quantity, for example in terms of the gradient's magnitude . \nWhile predictions from the model can still be obtained, we impose the restriction that the target metric must be an (observed or expected) quantity of the model, excluding the final prediction.\n\\revised{Model-based uncertainty is a noteworthy subclass here, which operates using the uncertainty of a model's weights .}\n\\revised{\\textcite{sharma2017evidence-based} describe a similar class, in which the uncertainty stems from not finding enough evidence in the training data, i.e. failing to separate classes at training time.}\n\\revised{They refer to this kind of uncertainty as {\\em insufficient evidence uncertainty}.}","id":"d53e7115-2a38-4cfc-bd3f-72de86cf317d","level":"paragraph","origin_cites_number":2,"parent_id":"78f9fdb7-e132-4e81-b968-042f74647827","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Query Strategies"],["paragraph","Random"],["paragraph","Model-based"]],"subsections":[],"title":"Model-based"},{"cite_extract_rate":1,"cites":[8716],"content":"Prediction-based strategies select instances by scoring their prediction output.\nThe most prominent members of this class are prediction-uncertainty-based and disagreement-based approaches.\n\\textcite{sharma2017evidence-based} denote prediction-based uncertainty by {\\em conflicting-evidence uncertainty}, which they, contrary to this work, count as another form of model-based uncertainty.\nThere is sometimes only a thin line between the concepts of model-based und prediction-based uncertainty.\n\\revised{Roughly speaking, prediction-based uncertainty corresponds in a classification setting to inter-class uncertainty, as opposed to model-based uncertainty, which corresponds to intra-class uncertainty.}\nIn literature, uncertainty sampling usually refers to prediction-based uncertainty, unless otherwise specified.","id":"c472f10c-b421-401e-aea9-953609e5c153","level":"paragraph","origin_cites_number":1,"parent_id":"78f9fdb7-e132-4e81-b968-042f74647827","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Query Strategies"],["paragraph","Random"],["paragraph","Prediction-based"]],"subsections":[],"title":"Prediction-based"},{"cite_extract_rate":0,"cites":[],"content":"\\revised{When a query strategy combines the output of multiple other strategies, this is called an {\\em ensemble}.}\n\\revised{We only classify the concept of ensemble strategies within the taxonomy (see disagreement-based subclass in Figure \\ref{fig:query_strategies}) without going into detail due to several reasons:}\n(1) Ensembles are again composed of primitive query strategies, which can be classified using our taxonomy.\n(2) Ensembles can be hybrids, i.e. they can be a mixture of different classes of query strategies.\nMoreover, the output of an ensemble is usually a function of the disagreement among the single classifiers, which is already covered in previous surveys of \\textcite{olsson2009literature} and \\textcite{fu2013survey}.\\\\\n\\noindent We are not the first to provide a classification of query strategies: \\textcite{aggarwal2014active} provide an alternative classification, which divides the query strategies into heterogenity-based models, \nperformance-based models, and representativeness-based models.\nHeterogenity-based models try to sample diverse data points, w.r.t the current labeled pool. This class includes among others uncertainty sampling and ensembles, i.e. no distinction is made between ensembles and single-model strategies. \nPerformance-based models aim to sample data targeting an increase of the models performance, for example a reduction in the model's error. \nThis intersects with our model-based class, however, it lacks strategies which focus on a change of parameters (e.g., expected gradient length ) as opposed to changes in a metric.\nLastly, representativeness-based strategies sample instances so that the distribution of the subsample is as similar as possible to the training set.\nAlthough similar to our data-based class, they always assume the existence of a model, which is not the case for data-based strategies.\\\\\n\\pindent \\textcite{fu2013survey} separate query strategies into \nuncertainty-based and diversity-based classes. \nUncertainty-based strategies assume the i.i.d. distribution of instances; they compute a separate score for each instance, which is the basis for the instance selection.\nDiversity-based strategies are a superset thereof and additionally consider correlation amongst instances.\nThereby they characterize uncertainty and correlation as critical components for query strategies.\nThis classification successfully distinguishes query strategies by considering exclusively uncertainty and correlation.\nHowever, it is less transparent in terms of the input information, \nwhich our taxonomy highlights.\nNevertheless, correlation is a factor orthogonal to our taxonomy and can be added as an additional criterion.\n\\\\\n\\pindent After creating our taxonomy, we discovered a recent categorization of uncertainty in deep learning , which distinguishes between data-, model-, and predict{\\em{}ive}-{\\em uncertainty}, similar to the taxonomy's first level (data-, model-, prediction-based query strategies).\nAlthough this classification comes naturally from the data's degree of processing, we emphasize that we are not the first to come up with this abstraction.\nBy using the input information as decisive criterion, this taxonomy provides an information-oriented view on query strategies.\nIt highlights in which parts and how uncertainty has been involed in existing query strategies. Uncertainty in terms of NNs is, however, is known to be challenging as described in Section \\ref{subsect:nn_al}. Moreover, we use the taxonomy to categorize recent work in AL for text classification in Section \\ref{subsec:experiments}.","id":"ff2fa7ea-f6cb-4eae-aca9-41c04d989749","level":"paragraph","origin_cites_number":2,"parent_id":"78f9fdb7-e132-4e81-b968-042f74647827","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Query Strategies"],["paragraph","Random"],["paragraph","Ensembles"]],"subsections":[],"title":"Ensembles"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsect:nn_al}\nIn this section we investigate the question, why neural networks are not more prevalent in AL applications. \nThis can be attributed to two central topics: Uncertainty estimation in NNs, and the contrast of NNs requiring between big data and AL dealing with small data.\nWe examine these issues from a NN perspective, alleviating the NLP focus.","id":"aff7fdcf-a8ec-4ce1-9895-2dad247f6097","level":"subsection","origin_cites_number":0,"parent_id":"3e095066-9c33-4519-a7ec-ec4550b8cb2d","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Neural-Network-Based Active Learning"]],"subsections":["fcf92379-9f52-4b46-8476-f3cf95e27e6c"],"title":"Neural-Network-Based Active Learning"},{"cite_extract_rate":0.18181818181818102,"cites":[1042,3986],"content":"\\hiddennotes{\n \\begin{itemize}\n \\item Überleitung fehlt\n \\item Argumentationsstruktur unklar?\n \\begin{itemize}\n \\item Classic -> RNN -> CNN -> LSTM -> GANs (-> Meta?) \n \\end{itemize}\n \\end{itemize}\n}\nEarly research in NN-based AL can be divided into uncertainty-based ,\nand ensemble-based strategies.\nThe former often use prediction entropy as measure of uncertainty,\nwhile the latter utilize the disagreement among the single classifiers.\n\\textcite{settles2008multiple-instance} proposed the expected gradient length (EGL) query strategy, which selects instances by the expected change in the model's weights. \\textcite{zhang2017active} were first to use a CNN for AL.\nThey proposed a variant of the expected gradient length strategy , in which they select instances that are expected to result in the largest change in embedding space, thereby training highly discriminative representations.\n\\textcite{sener2017active} observed uncertainty-based query strategies not to be effective for CNN-based batch-mode AL, and proposed core-set selection, which samples a small subset to represent the full dataset. \\textcite{ash2019deep} proposed BADGE, a query strategy for DNNs, which uses k-means++ seeding on the gradients of the final layer, in order to query by uncertainty and diversity.\nFinally, Generative Adversarial Networks (GANs; ) have also been applied successfully for AL tasks: \n\\textcite{zhu2017generative} use GANs for query synthesis of images within an active learner using an SVM model.\nThe instances are synthesized so that they would be classified with high uncertainty. \nThe authors report this approach to outperform random sampling, pool-based uncertainty sampling using an SVM , and in some cases passive learning, while having the weakness to generate too similar instances.\nThe approach itself is neither pure NN-based, nor does it belong to the pool-based scenario, however, it is the first reported use of GANs for AL.\n\\textcite{ducoffe2018adversarial} use adversarial attacks to find instances that cross the decision boundary with the aim to increase the model robustness.\nThey train two CNN architectures and report results superior to the core-set strategy on image classification tasks.\nIt is obvious that GANs inherently belong to the membership query synthesis scenario.\n\\revised{Therefore their performance correlates with the quality of artificial data synthesis, i.e. they are usually not that effective for NLP tasks. This has already been recognized and first improvements towards a better text generation have been made }.","id":"fcf92379-9f52-4b46-8476-f3cf95e27e6c","level":"paragraph","origin_cites_number":11,"parent_id":"aff7fdcf-a8ec-4ce1-9895-2dad247f6097","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Neural-Network-Based Active Learning"],["paragraph","Previous Work"]],"subsections":["72864fc8-a21e-4cfd-8e06-6087c4b614bf","17995907-7b29-4bbf-84e0-f3a0b81b8627"],"title":"Previous Work"},{"cite_extract_rate":0.33333333333333304,"cites":[4025,759,1042,3288],"content":"One of the earliest and in many variations adopted class of strategies is uncertainty sampling .\n\\label{sect:nn_uncertainty}\nUnfortunately, this widely-used concept is not straightforward to apply for NNs, as they do not provide an inherent indicator of uncertainty.\nIn the past, this has been tackled among others by ensembling , or by learning error estimates .\nMore recent approaches furthermore use Bayesian extensions , \nobtain uncertainty estimations using dropout ,\nor use probabilistic NNs to estimate predictive uncertainty .\nHowever, ensemble and Bayesian approaches quickly become infeasible on larger datasets,\nand NN architectures are generally known to be overconfident in their predictions .\nConsequently, uncertainty in NNs is only insufficiently solved and therefore still remains a highly relevant research area.\n\\hiddennotes{\n To circumvent this, various approaches have been suggested:\n (1) Bayes by Backprop uses Bayesian variational inference in order to learn approximated distributions over the weights of a NN.\n (2) extend dropout to represent uncertainties \n by interpreting dropout-enhanced NNs as a Bayesian approximation to deep Gaussian Processes.\n However, this has been reported to not scale well for larger datasets .\n (3) Traditionally this has been tackled by ensembling , i.e. producing $k$-many different models and outputting a prediction by interpreting their disagreement (see \\textcite{olsson2009literature}).\n The disadvantage of this approach is the obvious additional cost in runtime and model size.\n \\textcite{sener2017active} report uncertainty-based approaches to be not effective for CNNs, and \\textcite{ducoffe2018adversarial} observed worse performance compared to random sampling.\\\\\n}","id":"72864fc8-a21e-4cfd-8e06-6087c4b614bf","level":"paragraph","origin_cites_number":12,"parent_id":"fcf92379-9f52-4b46-8476-f3cf95e27e6c","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Neural-Network-Based Active Learning"],["paragraph","Previous Work"],["paragraph","Uncertainty in Neural Networks"]],"subsections":[],"title":"Uncertainty in Neural Networks"},{"cite_extract_rate":0.25,"cites":[7764,1096],"content":"DNNs are known to excel in particularly at large-scale datasets,\nbut often having large amounts of data available is a strict requirement to perform well at all (e.g., ).\nAL on the other hand tries to minimize the labeled data.\nThe small labeled datasets can be a problem for DNNs, since they are known to overfit on small datasets (e.g., ), \nwhich results in bad generalization performance on the test set.\nMoreover, DNNs often offer little advantage over shallow models when they are trained using small datasets , thereby lacking justification for their higher computational costs.\nOn the other hand we clearly cannot require AL to label more data, since this would defeat its purpose.\nTherefore there has been research on dealing with (D)NNs using small datasets, however, it is only a scarce amount, especially in relation to the large volume of NN literature in general.\nHandling small datasets is mostly circumvented by using pre-training or other transfer learning approaches .\nFinally, the search for optimal hyperparameters is often neglected and instead the hyperparameters of related work are used, which are optimized for large datasets, if at all.","id":"17995907-7b29-4bbf-84e0-f3a0b81b8627","level":"paragraph","origin_cites_number":8,"parent_id":"fcf92379-9f52-4b46-8476-f3cf95e27e6c","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning"],["subsection","Neural-Network-Based Active Learning"],["paragraph","Previous Work"],["paragraph","Contrasting Paradigms"]],"subsections":[],"title":"Contrasting Paradigms"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sect:text_classification}\nIn Sections \\ref{subsect:tc_recent_adv} and \\ref{subsect:active_tc} we first summarize recent methods in text classification and NNs.\nWe elaborate on each method's importance in the context of AL, and analyze its adoption by recent research where applicable.\nFor insufficiently adopted methods, we present how they could advance AL for text classification.\nMost importantly, we present an overview of recent experiments in AL for text classification and analyze commonalities and shortcomings.","id":"909d9662-92b7-4a13-b530-6c9553233a2e","level":"section","origin_cites_number":0,"parent_id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning for Text Classification"]],"subsections":["54497d60-bdb0-4315-828f-3d8f53e9c84c","5247a74b-cf2b-493b-a901-1dbfdda01e22","c22085b6-c5bc-4141-87dc-c353d7a13732"],"title":"Active Learning for Text Classification"},{"cite_extract_rate":0,"cites":[],"content":"\\label{subsect:tc_recent_adv}","id":"54497d60-bdb0-4315-828f-3d8f53e9c84c","level":"subsection","origin_cites_number":0,"parent_id":"909d9662-92b7-4a13-b530-6c9553233a2e","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning for Text Classification"],["subsection","Recent Advances in Text Classification"]],"subsections":["ad081844-efd8-44ae-a153-552a2838fa22"],"title":"Recent Advances in Text Classification"},{"cite_extract_rate":0.833333333333333,"cites":[7265,7165,856,1684,7370,11,826,1557,8385,673],"content":"Traditional methods use the bag-of-words (BoW) representation, which are sparse and high-dimensional.\nHowever, with the introduction of word embeddings like word2vec , \nGloVe , or fastText ,\nword embeddings have replaced BoW representations in many cases.\nThis is due to several reasons:\n(1) They represent semantic relations in vectors space and avoid the problem of mismatching features as for example due to synonymy;\n(2) incorporating word embeddings resulted in superior performance for many downstream tasks ;\n(3) unlike bag-of-words, word vectors are dense, low-dimensional representations, \nwhich makes them applicable to a wider range of algorithms --\nespecially in the context of NNs which favor fixed-size inputs.\nVarious approaches have been presented in order to obtain similar fixed size representations for word sequences, i.e. sentences, paragraphs or documents .\\\\\n\\pindent Word embeddings are representations, which provide exactly one vector per word and in consequence one meaning as well.\nThis makes them also unaware of the current word's context and therefore makes them unable to detect and handle ambiguities.\nUnlike word embeddings, language models (LMs) compute the word vector using the word and the surrounding context .\nThis results in a contextualized representation, which inherits the advantages of word embeddings, and at the same time allows for context-specific representation (in contrast to static embeddings) . \nELMo was the first LM to gain wide adoption and surpassed state of the art models on several NLP tasks .\nShortly thereafter, BERT was introduced and provided bidirectional pre-training-based language modelling.\nThe process to create a BERT-based model consists of a pre-training and a fine-tuning step as opposed to ELMO's direct feature-based approach in which contextualized vectors are obtained from the pre-trained model and used directly as features .\nBy masking, i.e. randomly removing a fraction of tokens during training, the training was adapted to predict the masked words.\nThis made the bidrectional training possible, which would otherwise be obstructed because a word could \"see itself\" when computing its probability of occurrence given a context .\nFollowing this, XLNet introduced a similar approach of pre-training and fine-tuning using an autoregressive language model, however, it overcame BERT's limitation as it does not rely on masking data during pre-training , and moreover, successfully manages to integrate the recent TransformerXL architecture .\nSince then, a variety of LMs have been published, which further optimize the pre-training of previous LM architectures (e.g., RoBERTa and ELECTRA ), or distill the knowledge into a smaller model (e.g., DistilBERT ).\nSimilarly to word embeddings, there are approaches to use LMs in order to obtain sentence representations from LMs .\\\\\n\\pindent All mentioned representations offer a richer expressiveness than traditional BoW representations and therefore are well-suited for active learning purposes.","id":"ad081844-efd8-44ae-a153-552a2838fa22","level":"paragraph","origin_cites_number":12,"parent_id":"54497d60-bdb0-4315-828f-3d8f53e9c84c","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning for Text Classification"],["subsection","Recent Advances in Text Classification"],["paragraph","Representations"]],"subsections":["e093196b-af7c-4543-b2db-908cbfcfea1d"],"title":"Representations"},{"cite_extract_rate":0.5,"cites":[7595,673,11,4026],"content":"\\label{subsect:nn_text_classification}\nA well-known CNN architecture presented by \\textcite{kim2014convolutional} (KimCNN) operates on pre-trained word vectors and achieved state of the art results at the time using only a simple but elegant architecture.\nThe investigated CNN setups did not require much hyperparameter tuning and confirmed the effectiveness of dropout as a regularizer for CNN-based text classification.\\\\\n\\pindent The word embeddings of fastText differ from other word embeddings in the sense that the approach is (1) supervised and (2) specifically designed for text classification.\nBeing a shallow neural network, it is still very efficient, while still obtaining performances comparable to deep learning approaches at that time.\\\\\n\\pindent\\textcite{howard2018universal} developed Universal Language Model Fine-tuning (ULMFiT), a LM transfer learning method using the AWD-LSTM architecture ,\nwhich outperformed the state of the art on several text classification datasets when trained on only $100$ labeled examples, and thereby achieved results significantly superior to more sophisticated architectures of previous work.\nContext-specific LMs like BERT and XLNet yield a context-dependent vector for each token, thereby strongly improving NN-based text classification .\nState of the art in NN-based text classification is LM-based fine-tuning with XLNet, which has a slight edge over BERT in terms of test error rate .\nULMFiT follows closely thereafter, and KimCNN is still a strong contender.\nNotably, ULMFiT, BERT and XLNet all perform {\\em transfer learning}, which aims to transfer knowledge from one model to another ,\nthereby massively reducing the required amounts of data.","id":"e093196b-af7c-4543-b2db-908cbfcfea1d","level":"paragraph","origin_cites_number":8,"parent_id":"ad081844-efd8-44ae-a153-552a2838fa22","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning for Text Classification"],["subsection","Recent Advances in Text Classification"],["paragraph","Representations"],["paragraph","Neural-Network-Based Text Classification"]],"subsections":[],"title":"Neural-Network-Based Text Classification"},{"cite_extract_rate":0.33333333333333304,"cites":[7765,4023,4022,8716,11],"content":"\\label{subsect:active_tc} \nTraditional AL for text classification heavily relied on query strategies based on prediction-uncertainty and ensembling . \nCommon model choices included support vector machines (SVMs; ), naive bayes , logistic regression and neural networks .\nTo the best of our knowledge, no previous survey covered traditional AL for text classification, however, ensembling-based AL for NLP has been covered in depth by \\textcite{olsson2009literature}.\\\\\n\\pindent Regarding modern NN-based AL for text classification, the relevant models are primarily CNN- and LSTM-based deep architectures: \\textcite{zhang2017active} claim to be the first to consider AL for text classification using DNNs.\nThey use CNNs and contribute a query strategy, which selects the instances \\phrasing{based on the expected change of the word embeddings and the model's uncertainty given the instance,} thereby learning discriminative embeddings for text classification.\n\\textcite{an2018deep} evaluated SVM, LSTM and gated recurrent unit (GRU; \\textcite{cho2014properties}) models, and reported that the latter two significantly outperformed the SVM baseline on the Chinese news dataset ThucNews.\n\\textcite{lu2020investigating} investigated the performance of different text representations in a pool-based AL scenario.\nThey compared frequency-based text representations, word embeddings and transformer-based representations used as input features for a SVM-based AL and different query strategies,\nin which transformer-based representations yielded consistently higher scores.\n\\textcite{prabhu2019sampling} investigate sampling bias and apply active text classification on the large scale text corpora of \\textcite{zhang2016character-level}.\nThey demonstrate FastText.zip with (entropy-based) uncertainty sampling to be a strong baseline, which is competitive compared to recent approaches in active text classification.\nMoreover, they use this strategy to obtain a surrogate dataset (comprising from 5\\% to 40\\% of the total data) on which a LSTM-based LM is trained using ULMFiT , reaching accuracy levels close to a training on the full dataset.\nUnlike past publications, they report this uncertainty-based strategy to be effective, robust, and at the same time computationally cheap.\nThis is the most relevant work in terms of \\revised{the intersection between text classification, NNs and DL}.\n\\hiddennotes{\n \\begin{itemize}\n \\item Previous Research nach Query Strategy Class / Embeddings untersuchen\n \\item Reinforcement Learning using NNs\n \\end{itemize}\n}\n\\begin{table}[h!]\n \\begin{center}\n \\begin{tabular}{l P{3cm} P{2.5cm} P{4.5cm}}\n \\hline\n \\textbf{Publication} & \\textbf{Datasets} & \\textbf{Model(s)} & \\textbf{Query Strategy Class(es)}\\\\\n \\hline\n \\mbox{} & 20N, R21, RV2, SPM & NB, SVM, kNN & 1. Prediction uncertainty (LC)\\newline 2. Prediction uncertainty (CTH)\\newline 3. Prediction uncertainty (disagreement) \\\\\n \\mbox{} & CR, MR, SJ, MRL, MUR, DR & CNN & 1. Model uncertainty (EGL)\\newline 2. Prediction Uncertainty (entropy)\\\\\n \\mbox{} & RMA & SVM & 1. Prediction uncertainty (CTH)\\newline 2. Prediction uncertainty (disagreement)\\\\\n \\mbox{} & TQA, MR & SVM, CNN, BiLSTM & Prediction uncertainty (disagreement)\\\\\n \\mbox{} & MR, SJ, TQA, CR & SVM, CNN, BiLSTM & 1. Prediction uncertainty (entropy)\\newline 2. Prediction uncertainty (disagreement)\\\\\n \\mbox{} & SGN, DBP, YHA, YRP, YRF, AGN, ARP, ARF & FTZ, ULMFiT & Prediction uncertainty (entropy)\\\\\n \\mbox{} & MRL, MDS, BAG, G13, ACR, SJ, AGN, DBP & SVM & 1. Prediction uncertainty (CTH) \\newline 2. Prediction uncertainty (disagreement) \\newline 3. Data-based (EGAL) \\newline 4. Data-based (density)\\\\\n \\hline \n \\end{tabular}\n \\captionsetup{aboveskip=10pt}\n \\caption{An overview of recent work on AL for text classification.\n We referred to the datasets using short keys, which can be looked up in Table \\ref{tab:datasets} in the Appendix. Models: Naive Bayes (NB), Support Vector Machine (SVM), k-Nearest Neighbours (kNN), Convolutional Neural Network (CNN), [Bidirectional] Long Short-Term Memory ([Bi]LSTM), FastText.zip (FTZ), Univeral Language Model Fine-Tuning (ULMFiT). Query strategies: Least confidence (LC), Closest-to-hyperplane (CTH), expected gradient length (EGL).\n Random selection baselines were omitted.}\n \\label{table:recent_work}\n\\end{center}\n\\end{table}","id":"5247a74b-cf2b-493b-a901-1dbfdda01e22","level":"subsection","origin_cites_number":15,"parent_id":"909d9662-92b7-4a13-b530-6c9553233a2e","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning for Text Classification"],["subsection","Text Classification for Active Learning"]],"subsections":[],"title":"Text Classification for Active Learning"},{"cite_extract_rate":0.5,"cites":[11,1096],"content":"\\label{subsec:experiments}\nTable \\ref{table:recent_work} shows the most recent AL for text classification experiments, all of them more recent than the surveys of \\textcite{settles2010active} and \\textcite{olsson2009literature}. \nFor each publication we list the utilized datasets, models, and classes of query strategies (with respect to the taxonomy in Section \\ref{subsec:query_strategies}).\nWe present this table in order to get insights about the recently preferred classification models and query strategy classes.\\\\\n\\pindent We can draw multiple conclusions from Table \\ref{table:recent_work}:\nIt is obvious that a significant majority of these query strategies belong to the class of prediction-based query strategies, more specifically to the prediction-uncertainty and disagreement-based sub-classes. \nIn addition to that, we can identify several shortcomings:\nFirst, in many experiments two or more standard datasets are evaluated, but very often there is little to no intersection between the experiments in terms of their datasets. \nAs a result we lose comparability against previous research.\nFor recent research, this can seen in Table \\ref{table:recent_work}, where the only larger intersections is between the works of \\textcite{zhang2017active} and \\textcite{lowell2019practical}.\n\\textcite{siddhant2018deep} provide at least some comparability against \\textcite{zhang2017active} and \\textcite{lowell2019practical} through one dataset each.\nAdditionally, RMA is a subset of R21 , which are used by \\textcite{bloodgood2018support} and \\textcite{hu2016active}, so they might be comparable to some degree.\n are the only ones to evaluate on the more recent large-scale text classification datasets , and although these datasets are more realistic in terms of their size, the authors omitted the classic datasets, so it is difficult to relate their contributions to previous work.\nMoreover, as a result of this, we do not know if and to what degree past experiments generalize to DNNs .\n\\noindent Finally, it is not clear if recent (D)NNs benefit from the same query strategies, i.e. past findings may not apply to modern NN architectures:\n\\textcite{prabhu2019sampling} identified contradicting statements in recent literature about the effectiveness of using prediction uncertainty in combination with NNs.\nThey achieved competitive results using a FastText.zip (FTZ) model and a prediction uncertainty query strategy, which proved to be very effective while requiring only a small amount of data, despite all reported weaknesses concering NNs and uncertainty estimates.","id":"c22085b6-c5bc-4141-87dc-c353d7a13732","level":"subsection","origin_cites_number":4,"parent_id":"909d9662-92b7-4a13-b530-6c9553233a2e","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Active Learning for Text Classification"],["subsection","Commonalities and Limitations of Previous Experiments"]],"subsections":[],"title":"Commonalities and Limitations of Previous Experiments"},{"cite_extract_rate":0,"cites":[],"content":"","id":"847fa582-b0b5-403a-9242-ea4d17738bff","level":"section","origin_cites_number":0,"parent_id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Open Research Questions"]],"subsections":["b759602b-ca2d-4e2f-9b70-cf51e332ff01"],"title":"Open Research Questions"},{"cite_extract_rate":0,"cites":[],"content":"In Section \\ref{sect:active_learning} it was illustrated that uncertainty-based strategies \nhave been used successfully in combination with non-NN models, and in Section \\ref{subsec:experiments} it was shown that they also account for the largest fraction of query strategies in recent NN-based AL. \nUnfortunately, uncertainty in NNs is still challenging due to inaccurate uncertainty estimates, or limited scalability (as described in Section \\ref{sect:nn_uncertainty}).","id":"b759602b-ca2d-4e2f-9b70-cf51e332ff01","level":"paragraph","origin_cites_number":0,"parent_id":"847fa582-b0b5-403a-9242-ea4d17738bff","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Open Research Questions"],["paragraph","Uncertainty Estimates in Neural Networks"]],"subsections":["00aa4769-dd36-4d2f-ba39-18878b0e6be3"],"title":"Uncertainty Estimates in Neural Networks"},{"cite_extract_rate":0.6666666666666661,"cites":[11,4023],"content":"As outlined in Section \\ref{subsect:tc_recent_adv}, the use of text representations in NLP has shifted from bag-of-words to static and contextualized word embeddings.\nThese representations evidentially provide many advantages like disambiguation capabilities, non-sparse vectors, and an increase in performance for many tasks.\nAlthough there have been some applications , there is no AL-specific systematic evaluation to compare word embeddings and LMs using NNs.\nMoreover, they are currently only scarcely used,\nwhich hints at either a slow adoption, or some non-investigated practical issues.","id":"00aa4769-dd36-4d2f-ba39-18878b0e6be3","level":"paragraph","origin_cites_number":3,"parent_id":"b759602b-ca2d-4e2f-9b70-cf51e332ff01","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Open Research Questions"],["paragraph","Uncertainty Estimates in Neural Networks"],["paragraph","Representations"]],"subsections":["6e1cfbc6-56a2-4243-bbb5-f779c6fa69b3","fd5f0c0a-5428-46a3-b6af-76f50107da5a","1b466353-341c-4aec-900d-656fddb7ff9e"],"title":"Representations"},{"cite_extract_rate":0,"cites":[],"content":"DL approaches are usually applied in the context of large datasets.\nAL, however, necessarily intends to keep the (labeled) dataset as small as possible.\nIn Section \\ref{sect:active_learning} we outlined why small datasets can be challenging for DNNs, and as a direct consequence as well for DNN-based AL.\nUsing pre-trained language models, this problem is alleviated to some degree because fine-tuning allows training models using considerably smaller datasets.\nNonetheless, it is to be investigated how little data is still necessary to successfully fine-tune a model.","id":"6e1cfbc6-56a2-4243-bbb5-f779c6fa69b3","level":"paragraph","origin_cites_number":0,"parent_id":"00aa4769-dd36-4d2f-ba39-18878b0e6be3","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Open Research Questions"],["paragraph","Uncertainty Estimates in Neural Networks"],["paragraph","Representations"],["paragraph","Small Data DNNs"]],"subsections":[],"title":"Small Data DNNs"},{"cite_extract_rate":1,"cites":[11],"content":"In Section \\ref{subsec:experiments} we provided an overview of the most common AL strategies for text classification.\nUnfortunately, the combinations of datasets used in the experiments are often completely disjoint, e.g. \\textcite{siddhant2018deep}, \\textcite{lowell2019practical}, and \\textcite{prabhu2019sampling}.\nAs a consequence, comparability is decreased or even lost, especially between more recent and past work.\nComparibility is, however, crucial to verify if past insights regarding shallow NN-based AL still apply in context of DNN-based AL .","id":"fd5f0c0a-5428-46a3-b6af-76f50107da5a","level":"paragraph","origin_cites_number":1,"parent_id":"00aa4769-dd36-4d2f-ba39-18878b0e6be3","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Open Research Questions"],["paragraph","Uncertainty Estimates in Neural Networks"],["paragraph","Representations"],["paragraph","Comparable Evaluations"]],"subsections":[],"title":"Comparable Evaluations"},{"cite_extract_rate":0.6666666666666661,"cites":[8718,4027],"content":"There is an abundance of query strategies to choose from, which we have (non-exhaustively) categorized in Section \\ref{subsec:query_strategies}.\nThis introduces the problem of choosing the optimal strategy.\nThe right choice depends on many factors like data, model, or task, and can even vary between different iterations during the AL process.\nAs a result, {\\em learning to learn} (or {\\em meta-learning}) has become popular and can be used to learn the optimal selection , or even learn query strategies as a whole .","id":"1b466353-341c-4aec-900d-656fddb7ff9e","level":"paragraph","origin_cites_number":3,"parent_id":"00aa4769-dd36-4d2f-ba39-18878b0e6be3","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Open Research Questions"],["paragraph","Uncertainty Estimates in Neural Networks"],["paragraph","Representations"],["paragraph","Learning to Learn"]],"subsections":[],"title":"Learning to Learn"},{"cite_extract_rate":0.5,"cites":[4023],"content":"In this survey, we investigated (D)NN-based AL for text classification and inspected factors obstructing its adoption.\nWe created a taxonomy, distinguishing query strategies by their reliance on data-based, model-based, and prediction-based input information.\nWe analyzed query strategies used in AL for text classification and categorized them into the respective taxonomy classes.\nWe presented the intersection between AL, text classification and DNNs, which is to the best of our knowledge the first survey of this topic.\nFurthermore, we reviewed (D)NN-based AL, identified current challenges and state of the art, and pointed out that it is both underresearched and often lacks comparability.\nIn addition to that, we presented relevant recent advances in NLP, related them to AL, and showed gaps and limitations for their application.\nOne of our main findings is that uncertainty-based query strategies are still the most widely used class, regardless of whether the analysis is restricted to NNs.\nLM-based representations offer finer-grained context-specific representations while also handling out-of-vocabulary words. \nMoreover, we find fine-tuning-based transfer learning alleviates the small data problem to some degree but lacks adoption.\nMost important DNNs are known for their strong performance on many tasks and first adoptions in AL have shown promising results .\nAll these gains would be highly desirable for AL.\nTherefore improving the adoption of DNNs in AL is crucial, especially since the expected increases in performance could be either used to improve the classification results while using the same amount of data or to increase the efficiency of the labeling process by reducing the data and therefore the labeling efforts.\nBased on these findings we identify research directions for future work in order to further advance (D)NN-based AL.\n\\section*{Acknowledgements}\nWe thank Gerhard Heyer for his valuable feedback on the manuscript, Lydia Müller for fruitful discussions about the taxonomy and advice thereon, and Janos Borst for sharing his thoughts on recent advances in language models.\nThis research was partially funded by the Development Bank of Saxony\n(SAB) under project number 100335729.\n\\printbibliography\n\\appendix","id":"3ba159ca-0eee-4f69-aedd-91724bce7349","level":"section","origin_cites_number":2,"parent_id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Conclusions"]],"subsections":[],"title":"Conclusions"},{"cite_extract_rate":0,"cites":[],"content":"","id":"ca5cf519-fe65-47a3-9e14-707edbf02174","level":"section","origin_cites_number":0,"parent_id":"4e01ebc7-074c-4060-b09d-956a012ff8ea","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Appendix"]],"subsections":["01e0ed63-a97a-4079-8877-7a6544e5b88a"],"title":"Appendix"},{"cite_extract_rate":0.17647058823529402,"cites":[1096,4028,8420],"content":"The following table provides additional information about the datasets which were referred to in Section \\ref{subsect:active_tc}. \n{\n\\setlength{\\LTcapwidth}{\\textwidth}\n\\begin{longtable}{l p{3.8cm} l p{4.1cm} r r}\n \\hline\n \\textbf{Id} & \\textbf{Name} & \\textbf{Type} & \\textbf{Publication} & \\textbf{\\#Train} & \\textbf{\\#Test} \\\\\n \\hline\n TQA & TREC QA & MC & & 5,500 & 500 \\\\\n CR & Customer Reviews & MC & & \\textsuperscript{*}{315} & - \\\\\n ACR & Additional Customer\\newline Reviews & MC & & \\textsuperscript{*}{325} & - \\\\\n MDS & Multi-Domain Sentiment & B & & \\textsuperscript{**}{8,000} & - \\\\\n BAG & Blog Author Gender & B & & 3,100 & - \\\\\n G13 & Guardian 2013 & MC & & 6,520 & -\\\\\n MR & Movie Reviews & B & & 10,662 & - \\\\\n MRL & Movie Reviews Long & B & & 2,000 & - \\\\\n MUR & Music Review & B & & 2,000 & - \\\\\n DR & Doctor Reviews & MC & & 58,110 & - \\\\\n SJ & Subjectivity & B & & 10,000 & -\\\\\n 20N & 20newsgroups & MC & & \\textsuperscript{***}18,846 & - \\\\\n R21 & Reuters-21578 & ML & & 21578 & - \\\\\n RMA & Reuters ModApté & ML & & 9,603 & 3,299\\\\\n RV2 & RCV1-V2 & ML & & 23,149 & 781,265 \\\\\n SPM & Spam & B & & 1,000 & - \\\\\n AGN & AG News & MC & \\newline & 120,000 & 7,600 \\\\\n SGN & Sogou News & MC & & 450,000 & 60,000\\\\\n DBP & DBPedia & MC & & 560,000 & 70,000\\\\\n YRP & Yelp Review Polarity & B & & 560,000 & 38,000\\\\\n YRF & Yelp Review Full & MC & & 650,000 & 50,000\\\\\n YAH & Yahoo! Answers & MC & & 1,400,000 & 60,000\\\\\n ARP & Amazon Review Polarity & B & & 3,600,000 & 40,000\\\\\n ARF & Amazon Review Full & MC & & 3,000,000 & 650,000\\\\\n \\hline\n \\caption{A collection of widely-used text classification datasets.\n The column \"Type\" denotes the classification setting (B = binary, MC = multi-class, ML = multi-class multi-label). \n The columns \"\\#Train\" and \"\\#Test\" show the size of the train and of the test set.\n In the case that no predefined splits were available \"\\#Train\" represents the full dataset's size. \n Each dataset was assigned a short id (first column), which we use in the paper for reference.\\\\\n \\newline\n (*): documents, (**) labels reduced to positive/negative, (***) 20news-bydate with duplicates removed}\n \\label{tab:datasets}\n\\end{longtable}\n}\n\\end{document}","id":"01e0ed63-a97a-4079-8877-7a6544e5b88a","level":"subsection","origin_cites_number":17,"parent_id":"ca5cf519-fe65-47a3-9e14-707edbf02174","prefix_titles":[["title","A Survey of Active Learning for Text Classification using Deep Neural Networks"],["section","Appendix"],["subsection","Datasets"]],"subsections":[],"title":"Datasets"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ]\n ],\n \"subsections\": [\n \"f9d1ed5f-ebd8-41c8-b85b-c14ee4a51e33\",\n \"10ab6316-68f8-4bea-9503-5f9388a60998\",\n \"3e095066-9c33-4519-a7ec-ec4550b8cb2d\",\n \"909d9662-92b7-4a13-b530-6c9553233a2e\",\n \"847fa582-b0b5-403a-9242-ea4d17738bff\",\n \"3ba159ca-0eee-4f69-aedd-91724bce7349\",\n \"ca5cf519-fe65-47a3-9e14-707edbf02174\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.32,\n \"cites\": [\n 7764,\n 1159,\n 7165,\n 1096,\n 4022,\n 8385,\n 1684,\n 11\n ],\n \"content\": \"\\\\label{sec:introduction}\\nData is the fuel of machine learning applications and therefore has been steadily increasing in value.\\nIn many settings an abundant amount of unlabeled data is produced,\\nbut in order to use such data in supervised machine learning, \\none has no choice but to provide labels.\\nThis usually entails a manual labeling process, which is often non-trivial and can even require a domain expert, \\ne.g., in patent classification , or clinical text classification .\\nMoreover, this is time-consuming and rapidly increases monetary costs, thereby quickly rendering this approach infeasible.\\nEven if an expert is available, it is often impossible to label each datum due to the vast size of modern datasets.\\nThis especially impedes the field of Natural Language Processing (NLP), \\nin which both the dataset and the amount of text within each document can be huge,\\nresulting in unbearable amounts of annotation efforts for human experts.\\nActive Learning (AL) aims to reduce the amount of data annotated by the human expert.\\n\\\\revised{It is an iterative cyclic process between an {\\\\em oracle} (usually the human annotator)\\nand an {\\\\em active learner}.}\\nIn contrast to passive learning, in which the data is simply fed to the algorithm, \\nthe active learner chooses which samples are to be labeled next. \\nThe labeling itself, however, is done by a human expert, the so-called human in the loop.\\nHaving received new labels, the active learner trains a new model and the process starts from the beginning.\\n\\\\revised{Using the term active learner, we refer to the composition of a {\\\\em model}, a {\\\\em query strategy}, and a {\\\\em stopping criterion}.}\\nIn this work the model is w.l.o.g. a text classification model, \\nthe query strategy decides which instances should be labeled next, and the stopping criterion defines when to stop the AL loop.\\nAccording to \\\\textcite{settles2010active} there are three main scenarios for AL:\\n(1) Pool-based, in which the learner has access to the closed set of unlabeled instances, called the pool;\\n(2) stream-based, where the learner receives one instance at a time and has the options to keep it, or to discard;\\n(3) membership query synthesis, \\\\revised{in which the learner creates new artificial instances to be labeled}.\\nIf the pool-based scenario operates not on a single instance, but on a batch of instances, this is called {\\\\em batch-mode} AL .\\nThroughout this work we assume a pool-based batch-mode scenario because in a text classification setting the dataset is usually a closed set, and the batch-wise operation reduces the number of retraining operations, which cause waiting periods for the user.\\\\\\\\\\n\\\\pindent \\\\revised{The underlying idea of AL is that few representative instances can be used as surrogate for the full dataset.}\\nNot only does a smaller subset of the data reduce the computational costs, \\nbut also it has been shown that AL can even increase the quality of the resulting model compared to learning on the full dataset .\\nAs a consequence, AL has been used in many NLP tasks, e.g. text classification , named entity recognition , \\nor machine translation and is still an active area of research.\\nIn recent years, deep learning (DL) approaches have dominated most NLP tasks' state-of-the-art results.\\nThis can be attributed to advances in neural networks (NNs), above all Convolutional Neural Networks (CNN; ) and (Bidirectional-)Long Short-Term Memory (LSTM; ),\\nwhich were eventually adopted into the NLP domain,\\nand to the advances of using word embeddings and contextualized word embeddings .\\nBoth NN architectures and text representations have raised the state-of-the-art results in the field of text classification considerably (e.g., ).\\nIf these improvements were transferrable to AL, this would result in a huge increase in efficiency.\\nFor the AL practitioner, this either means achieving the same performance using fewer samples, or having an increase in performance using the same amount of data.\\nAnother favorable development is that transfer learning, especially the paradigm of fine-tuning pre-trained language models (LMs), has become popular in NLP.\\nIn the context of AL this helps especially in the small data scenario, in which a pre-trained model can be leveraged to train a model by fine-tuning using only little data, which would otherwise be infeasible.\\nFinally, by operating on sub-word units LMs also handle out-of-vocabulary tokens, which is an advantage over many traditional methods.\\nResulting from these advances, existing AL surveys have become both incomplete in some parts and outdated in others:\\nThey lack comparison against the current state of the art models, do not provide results for more recent large-scale datasets, \\nand most importantly, they are lacking the aforementioned advances in NNs and text representations.\\nSurprisingly, despite the current popularity of NNs, there is only little research about NN-based active learning in the context of NLP, and even less thereof in the context of text classification (see Section \\\\ref{subsect:nn_al} and Section \\\\ref{subsect:active_tc} for a detailed summary).\\nWe suspect this is due to the following reasons: \\n(1) Many DL models are known to require large amounts of data ,\\nwhich is in strong contrast to AL aiming at requiring as little data as possible\\n(2) there is a whole AL scenario based on artificial data generation, which unfortunately is a lot more challenging for text in contrast to for example images, for which data augmentation is commonly used in classification tasks ;\\n(3) NNs are lacking uncertainty information regarding their predictions \\\\revised{(as explained in Section \\\\ref{subsect:nn_al})}, which complicates the use of a whole prominent class of query strategies.\\n\\\\hiddennotes{\\n \\\\begin{itemize}\\n \\\\item No clear evaluation protocol / huge differences in the evaluation approaches\\n \\\\end{itemize}\\n}\\nThis survey aims at summarizing the existing approaches of (D)NN-based AL for text classification.\\nOur main contributions are as follows:\\n\\\\begin{enumerate}\\n \\\\item We provide a taxonomy of query strategies and classify strategies relevant for AL for text classification.\\n \\\\item We survey existing work at the intersection of AL, text classification, and (D)NNs.\\n \\\\item Recent advances in text classification are summarized and related to the AL process. It is then investigated, if and to what degree they have been adopted for AL.\\n \\\\item The experimental setup of previous research is collectively analyzed regarding datasets, models, and query strategies in order to identify recent trends, commonalities, and shortcomings in the experiments.\\n \\\\item \\\\revised{We identify research gaps and outline future research directions.}\\n\\\\end{enumerate}\\n\\\\noindent Thereby we provide a comprehensive survey of recent advances in NN-based active text classification. \\nHaving reviewed these recent advances, we illuminate areas that either need re-evaluation, or have not yet been evaluated in a more recent context.\\nAs a final result, we develop research questions outlining the scope of future research.\",\n \"id\": \"f9d1ed5f-ebd8-41c8-b85b-c14ee4a51e33\",\n \"level\": \"section\",\n \"origin_cites_number\": 25,\n \"parent_id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 1159\n ],\n \"content\": \"\\\\textcite{settles2010active} provides a general active learning survey, \\nsummarizing the prevalent AL scenario types and query strategies.\\nThey present variations of the basic AL setup like \\n variable labeling costs or alternative query types, \\nand most notably, they discuss empirical and theoretical research investigating the effectiveness of AL: \\nThey mention research suggesting that AL is effective in practice \\nand has increasingly gained adoption in real world applications. However, it is pointed out that empirical research also reported cases in which AL performed worse than passive learning and that the theoretical analysis of AL is incomplete.\\nFinally, relations to related research areas are illustrated, thereby connecting AL among others to reinforcement learning and semi-supervised learning.\\n\\\\hiddennotes{\\n settles2009/settles2010active \\n \\\\begin{itemize}\\n \\\\item scenarios: membership query synthesis, stream-based, pool-based\\n \\\\item query strategies\\n \\\\begin{itemize}\\n \\\\item keine Klassifikation vorgenommen\\n \\\\end{itemize}\\n \\\\item most important (at that time): analysis\\n \\\\begin{itemize}\\n \\\\item does it work?\\n \\\\item both empirical and theoretical evidence\\n \\\\end{itemize}\\n \\\\item TODO: problem setting variants, practical considerations (only in 2010 version)\\n \\\\item related areas: semi-supervised, reinforcement, submodular optimization, equivalence query learning, model parroting and compression\\n \\\\end{itemize} \\n}\\nThe survey of \\\\textcite{fu2013survey} is focused around a thorough analysis of uncertainty-based query strategies, which are categorized into a taxonomy.\\nThis taxonomy differentiates at the topmost level between the uncertainty of i.i.d. instances and instance correlation. The latter is a superset of the former \\nand intends to reduce redundancy among instances by considering feature, label, and structure correlation when querying.\\nMoreover, they perform an algorithmic analysis for each query strategy \\nand order the strategies by their respective time complexity, highlighting the increased complexity for correlation-based strategies.\\n\\\\hiddennotes{\\n \\\\begin{itemize}\\n \\\\item survey on instance selection\\n \\\\end{itemize} \\n}\\nAnother general survey covering a wide range of topics was conducted by \\\\textcite{aggarwal2014active}. \\nThey provide a flat categorization of query strategies, which is quite different from the taxonomy of \\\\textcite{fu2013survey} and divides them into the following three categories: (1) ``heterogenity-based'', which sample instances by their prediction uncertainty or dissimilarity compared to existing labeled instances, (2) ``performance-based''', which select instances based on a predicted change of the model loss, and (3) ``representativeness-based'', which select data points to reflect a larger set in terms of their properties, usually achieved by the means of distribution density .\\nSimilarly to , they present and discuss many non-standard variations of the active learning scenario.\\n\\\\revised{\\n An NLP-focused active learning survey was performed by \\\\textcite{olsson2009literature}.\\n This work's main contribution is a survey of disagreement-based query strategies, which use the disagreement among multiple classifiers to select instances.\\n Moreover, Olsson reviews practical considerations, e.g., selecting an initial seed set,\\n deciding between stream-based and pool-based scenario, and deciding when to terminate the learning process.\\n Although some NN-based applications are mentioned, none of the above surveys covers NN-based AL in depth.\\n Besides, none is recent enough to cover NN-architectures, which have only recently been adapted successfully to text classification problems like e.g., KimCNN .\\n The same holds true for recent advances in NLP such as word embeddings, contextualized language models (explained in Section \\\\ref{subsect:tc_recent_adv}), or resulting advances in text classification (discussed in Section \\\\ref{subsect:nn_text_classification} and Section \\\\ref{subsect:active_tc}).\\n We intend to fill these gaps in the remainder of this survey.\\n}\\n\\\\hiddennotes{\\n \\\\begin{itemize}\\n \\\\item NLP-centric Active Learning Survey\\n \\\\item describes multiple active learning modes: Query by uncertainty, query by committee, redundant views\\n \\\\item extensively reviews strategies to measure disagrement\\n \\\\end{itemize}\\n}\",\n \"id\": \"10ab6316-68f8-4bea-9503-5f9388a60998\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Related Work\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Related Work\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sect:active_learning}\\nThe goal of AL is to create a model using as few labeled instances as possible,\\ni.e. minimizing the interactions between the oracle and the active learner.\\nThe AL process (illustrated in Figure \\\\ref{fig:al_process}) is as follows: The oracle requests unlabeled instances from the active learner ({\\\\em query}, see Figure \\\\ref{fig:al_process}: step 1), \\nwhich are then selected by the active learner (based on the selected query strategy) and passed to the oracle (see Figure \\\\ref{fig:al_process}: step 2).\\nSubsequently, these instances are labeled by the oracle and returned to the active learner ({\\\\em update}, see Figure \\\\ref{fig:al_process}: step 3).\\nAfter each update step the active learner's model is retrained, which makes this operation at least as expensive as a training of the underlying model.\\nThis process is repeated until a stopping criterion is met (e.g., a maximum number of iterations or a minimum threshold of change in classification accuracy).\\n\\\\begin{figure}[!h]\\n \\\\centering \\n \\\\includegraphics{figures/al_process.pdf}\\n \\\\caption{An overview of the AL process: Model, query strategy, and (optionally a) stopping criterion are the key components of an active learner.\\n The main loop is as follows:\\n First the oracle queries the active learner, which returns a fixed amount of unlabeled instances.\\n Then, for all selected unlabeled unstances are assigned labels by the oracle.\\n This process is repeated until the oracle stops, or a predefined stopping criterion is met.}\\n \\\\label{fig:al_process}\\n\\\\end{figure}\\n\\\\noindent The most important component for AL is the query strategy.\\nIn the introduction we claimed that a large fraction of query strategies are uncertainty-based.\\nTo analyze this we provide a taxonomy of query strategies in the following section and highlight the parts in which uncertainty is involved.\\nFor a general and more detailed introduction on AL refer to the surveys of \\\\textcite{settles2010active} and \\\\textcite{aggarwal2014active}.\",\n \"id\": \"3e095066-9c33-4519-a7ec-ec4550b8cb2d\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ]\n ],\n \"subsections\": [\n \"3effa5dc-9905-4ec7-8c3d-464cfd4bd1d9\",\n \"aff7fdcf-a8ec-4ce1-9895-2dad247f6097\"\n ],\n \"title\": \"Active Learning\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 1042,\n 4024,\n 4023,\n 11,\n 8716,\n 8717\n ],\n \"content\": \"\\\\label{subsec:query_strategies}\\nIn Figure \\\\ref{fig:query_strategies} we classify the most common AL query strategies based on a strategy's {\\\\em input information}, which denotes the numeric value(s) a strategy operates on.\\nIn our taxonomy the input information can be either random or one of data, model, and prediction.\\n\\\\phrasing{These categories are ordered by increasing complexity and are not mutually exclusive.}\\n\\\\phrasing{Obviously, the model is a function of the data, as well as the prediction is a function of model and data, and moreover, in many cases a strategy use multiple of these criteria.}\\nIn such cases we assign the query strategy to the most specific category (i.e. prediction-based precedes model-based, which in turn precedes data-based).\\n\\\\begin{figure}[!h]\\n {\\n \\\\scriptsize\\n \\\\centering\\n \\\\begin{forest} for tree={\\n grow'=0, draw\\n },\\n forked edges,\\n [\\\\scshape query strategies, no edge, root\\n [\\\\scshape random, onode_dashed\\n [,phantom]\\n ]\\n [\\\\scshape data-based, onode\\n [\\\\scshape data uncertainty, onode\\n [discriminative\\\\\\\\ , tier=leaf, tnode]\\n ]\\n [\\\\scshape representativeness, onode\\n [\\\\scshape clustering, onode\\n [flat\\\\\\\\ , tier=leaf, tnode]\\n [hierarchical\\\\\\\\ , tier=leaf, tnode]\\n ]\\n [\\\\scshape set construction, onode\\n [core-set\\\\\\\\ , tier=leaf, tnode]\\n ]\\n ]\\n ]\\n [\\\\scshape model-based, onode\\n [\\\\scshape model uncertainty, onode\\n [UNC-IE\\\\\\\\ , tier=leaf, tnode]\\n ]\\n [\\\\scshape expected parameter\\\\\\\\change, onode\\n [expected gradient length\\\\\\\\ , tier=leaf, tnode]\\n [expected weight change\\\\\\\\ , tier=leaf, tnode]\\n ]\\n [\\\\scshape adversarial, tnode\\n [DFAL\\\\\\\\ , tier=leaf, tnode]\\n ] \\n ]\\n [\\\\scshape prediction-based, name=prediction_based, onode\\n [\\\\scshape prediction\\\\newline uncertainty, name=prediction_uncertainty, tnode\\n [\\\\scshape probabilistic, onode\\n [uncertainty sampling , tier=leaf,tnode]\\n ]\\n [\\\\scshape margin-based, onode\\n [version space\\\\\\\\ , tier=leaf,tnode]\\n [closest to hyperplane\\\\\\\\ , tier=leaf,tnode]\\n ]\\n [\\\\scshape entropy, name=entropy, onode\\n [BALD\\\\\\\\ , tier=leaf,tnode]\\n ]\\n ]\\n [\\\\scshape discriminative, onode\\n [DAL\\\\\\\\ , tier=leaf , tnode]\\n ] \\n [\\\\scshape expected prediction change\\\\\\\\, onode\\n [expected error reduction\\\\\\\\ , name=lastnode, tier=leaf, tnode]\\n ]\\n [\\\\scshape disagreement, name=disagreement, onode_dashed\\n [,phantom, tier=leaf]\\n ]\\n ] \\n ]\\n \\\\draw[zlevel,-] (disagreement.east) -| (entropy.south);\\n \\\\draw[zlevel,-] (entropy.north) |- (prediction_uncertainty.east);\\n \\\\path (current bounding box.south) coordinate (s1);\\n \\\\draw[decorate,decoration={brace,amplitude=1em,mirror}] ([yshift=-3pt]prediction_based.south west|-s1) -- node[below=1em] {class} ([yshift=-3pt,xshift=-1pt]prediction_based.south east|-s1);\\n \\\\draw[decorate,decoration={brace,amplitude=1em,mirror}] ([yshift=-3pt,xshift=1pt]prediction_based.south east|-s1) -- node[below=1em] {subclass(es)} ([yshift=-3pt,xshift=-1pt]lastnode.south west|-s1);\\n \\\\draw[decorate,decoration={brace,amplitude=1em,mirror}] ([yshift=-3pt,xshift=1pt]lastnode.south west|-s1) -- node[below=1em] {example} ([yshift=-3pt,xshift=-1pt]lastnode.south east|-s1);\\n \\\\end{forest}\\n \\\\caption{\\\\revised{A taxonomy of query strategies for AL.\\n The key distinction is at the first level, where the query strategies are categorized by their access to different kinds of input information. \\n From the second to the penultimate level we form coherent subclasses,\\n and the final level shows examples for the respective class.\\n This taxonomy is not exhaustive due to the abundance of existing query strategies, and it is biased towards query strategies in NLP.}\\n }\\n \\\\label{fig:query_strategies}\\n }\\n\\\\end{figure}\\n\\\\hiddennotes{\\n \\\\begin{itemize}\\n \\\\item Table in appendix: For each strategy here: Applicability: B / MC / ML\\\\\\\\\\n \\\\end{itemize}\\n}\",\n \"id\": \"3effa5dc-9905-4ec7-8c3d-464cfd4bd1d9\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"3e095066-9c33-4519-a7ec-ec4550b8cb2d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Query Strategies\"\n ]\n ],\n \"subsections\": [\n \"78f9fdb7-e132-4e81-b968-042f74647827\"\n ],\n \"title\": \"Query Strategies\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 1042,\n 4024,\n 8716\n ],\n \"content\": \"Randomness has traditionally been used as a baseline for many tasks.\\nIn this case, random sampling selects instances at random and is a strong baseline for AL instance selection .\\n\\\\revised{It often performs competitive to more sophisticated strategies, especially when the labeled pool has grown larger .}\",\n \"id\": \"78f9fdb7-e132-4e81-b968-042f74647827\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"3effa5dc-9905-4ec7-8c3d-464cfd4bd1d9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Query Strategies\"\n ],\n [\n \"paragraph\",\n \"Random\"\n ]\n ],\n \"subsections\": [\n \"734bf4e0-5809-4e56-9af4-79cebf71ecdb\",\n \"d53e7115-2a38-4cfc-bd3f-72de86cf317d\",\n \"c472f10c-b421-401e-aea9-953609e5c153\",\n \"ff2fa7ea-f6cb-4eae-aca9-41c04d989749\"\n ],\n \"title\": \"Random\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Data-based strategies have the lowest level of knowledge, i.e. they only operate on the raw input data and optionally the labels of the labeled pool. \\nWe categorize them further into (1) strategies relying on data-uncertainty, which may use information about the data distribution, label distribution, and label correlation,\\nand (2) representativeness, \\\\revised{which tries to geometrically compress a set of points, by using fewer representative instances to represent the properties of the entirety}.\",\n \"id\": \"734bf4e0-5809-4e56-9af4-79cebf71ecdb\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"78f9fdb7-e132-4e81-b968-042f74647827\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Query Strategies\"\n ],\n [\n \"paragraph\",\n \"Random\"\n ],\n [\n \"paragraph\",\n \"Data-based\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data-based\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The class of model-based strategies has knowledge about both the data and the model.\\nThese strategies query instances based on measure provided by the model given an instance. \\nAn example for this would be a measure of confidence for the model's explanation of the given instance ,\\nfor example, how reliable the model rates encountered features.\\nThis can also be an expected quantity, for example in terms of the gradient's magnitude . \\nWhile predictions from the model can still be obtained, we impose the restriction that the target metric must be an (observed or expected) quantity of the model, excluding the final prediction.\\n\\\\revised{Model-based uncertainty is a noteworthy subclass here, which operates using the uncertainty of a model's weights .}\\n\\\\revised{\\\\textcite{sharma2017evidence-based} describe a similar class, in which the uncertainty stems from not finding enough evidence in the training data, i.e. failing to separate classes at training time.}\\n\\\\revised{They refer to this kind of uncertainty as {\\\\em insufficient evidence uncertainty}.}\",\n \"id\": \"d53e7115-2a38-4cfc-bd3f-72de86cf317d\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"78f9fdb7-e132-4e81-b968-042f74647827\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Query Strategies\"\n ],\n [\n \"paragraph\",\n \"Random\"\n ],\n [\n \"paragraph\",\n \"Model-based\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Model-based\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 8716\n ],\n \"content\": \"Prediction-based strategies select instances by scoring their prediction output.\\nThe most prominent members of this class are prediction-uncertainty-based and disagreement-based approaches.\\n\\\\textcite{sharma2017evidence-based} denote prediction-based uncertainty by {\\\\em conflicting-evidence uncertainty}, which they, contrary to this work, count as another form of model-based uncertainty.\\nThere is sometimes only a thin line between the concepts of model-based und prediction-based uncertainty.\\n\\\\revised{Roughly speaking, prediction-based uncertainty corresponds in a classification setting to inter-class uncertainty, as opposed to model-based uncertainty, which corresponds to intra-class uncertainty.}\\nIn literature, uncertainty sampling usually refers to prediction-based uncertainty, unless otherwise specified.\",\n \"id\": \"c472f10c-b421-401e-aea9-953609e5c153\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"78f9fdb7-e132-4e81-b968-042f74647827\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Query Strategies\"\n ],\n [\n \"paragraph\",\n \"Random\"\n ],\n [\n \"paragraph\",\n \"Prediction-based\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Prediction-based\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\revised{When a query strategy combines the output of multiple other strategies, this is called an {\\\\em ensemble}.}\\n\\\\revised{We only classify the concept of ensemble strategies within the taxonomy (see disagreement-based subclass in Figure \\\\ref{fig:query_strategies}) without going into detail due to several reasons:}\\n(1) Ensembles are again composed of primitive query strategies, which can be classified using our taxonomy.\\n(2) Ensembles can be hybrids, i.e. they can be a mixture of different classes of query strategies.\\nMoreover, the output of an ensemble is usually a function of the disagreement among the single classifiers, which is already covered in previous surveys of \\\\textcite{olsson2009literature} and \\\\textcite{fu2013survey}.\\\\\\\\\\n\\\\noindent We are not the first to provide a classification of query strategies: \\\\textcite{aggarwal2014active} provide an alternative classification, which divides the query strategies into heterogenity-based models, \\nperformance-based models, and representativeness-based models.\\nHeterogenity-based models try to sample diverse data points, w.r.t the current labeled pool. This class includes among others uncertainty sampling and ensembles, i.e. no distinction is made between ensembles and single-model strategies. \\nPerformance-based models aim to sample data targeting an increase of the models performance, for example a reduction in the model's error. \\nThis intersects with our model-based class, however, it lacks strategies which focus on a change of parameters (e.g., expected gradient length ) as opposed to changes in a metric.\\nLastly, representativeness-based strategies sample instances so that the distribution of the subsample is as similar as possible to the training set.\\nAlthough similar to our data-based class, they always assume the existence of a model, which is not the case for data-based strategies.\\\\\\\\\\n\\\\pindent \\\\textcite{fu2013survey} separate query strategies into \\nuncertainty-based and diversity-based classes. \\nUncertainty-based strategies assume the i.i.d. distribution of instances; they compute a separate score for each instance, which is the basis for the instance selection.\\nDiversity-based strategies are a superset thereof and additionally consider correlation amongst instances.\\nThereby they characterize uncertainty and correlation as critical components for query strategies.\\nThis classification successfully distinguishes query strategies by considering exclusively uncertainty and correlation.\\nHowever, it is less transparent in terms of the input information, \\nwhich our taxonomy highlights.\\nNevertheless, correlation is a factor orthogonal to our taxonomy and can be added as an additional criterion.\\n\\\\\\\\\\n\\\\pindent After creating our taxonomy, we discovered a recent categorization of uncertainty in deep learning , which distinguishes between data-, model-, and predict{\\\\em{}ive}-{\\\\em uncertainty}, similar to the taxonomy's first level (data-, model-, prediction-based query strategies).\\nAlthough this classification comes naturally from the data's degree of processing, we emphasize that we are not the first to come up with this abstraction.\\nBy using the input information as decisive criterion, this taxonomy provides an information-oriented view on query strategies.\\nIt highlights in which parts and how uncertainty has been involed in existing query strategies. Uncertainty in terms of NNs is, however, is known to be challenging as described in Section \\\\ref{subsect:nn_al}. Moreover, we use the taxonomy to categorize recent work in AL for text classification in Section \\\\ref{subsec:experiments}.\",\n \"id\": \"ff2fa7ea-f6cb-4eae-aca9-41c04d989749\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"78f9fdb7-e132-4e81-b968-042f74647827\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Query Strategies\"\n ],\n [\n \"paragraph\",\n \"Random\"\n ],\n [\n \"paragraph\",\n \"Ensembles\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Ensembles\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsect:nn_al}\\nIn this section we investigate the question, why neural networks are not more prevalent in AL applications. \\nThis can be attributed to two central topics: Uncertainty estimation in NNs, and the contrast of NNs requiring between big data and AL dealing with small data.\\nWe examine these issues from a NN perspective, alleviating the NLP focus.\",\n \"id\": \"aff7fdcf-a8ec-4ce1-9895-2dad247f6097\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"3e095066-9c33-4519-a7ec-ec4550b8cb2d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Neural-Network-Based Active Learning\"\n ]\n ],\n \"subsections\": [\n \"fcf92379-9f52-4b46-8476-f3cf95e27e6c\"\n ],\n \"title\": \"Neural-Network-Based Active Learning\"\n },\n {\n \"cite_extract_rate\": 0.18181818181818102,\n \"cites\": [\n 1042,\n 3986\n ],\n \"content\": \"\\\\hiddennotes{\\n \\\\begin{itemize}\\n \\\\item Überleitung fehlt\\n \\\\item Argumentationsstruktur unklar?\\n \\\\begin{itemize}\\n \\\\item Classic -> RNN -> CNN -> LSTM -> GANs (-> Meta?) \\n \\\\end{itemize}\\n \\\\end{itemize}\\n}\\nEarly research in NN-based AL can be divided into uncertainty-based ,\\nand ensemble-based strategies.\\nThe former often use prediction entropy as measure of uncertainty,\\nwhile the latter utilize the disagreement among the single classifiers.\\n\\\\textcite{settles2008multiple-instance} proposed the expected gradient length (EGL) query strategy, which selects instances by the expected change in the model's weights. \\\\textcite{zhang2017active} were first to use a CNN for AL.\\nThey proposed a variant of the expected gradient length strategy , in which they select instances that are expected to result in the largest change in embedding space, thereby training highly discriminative representations.\\n\\\\textcite{sener2017active} observed uncertainty-based query strategies not to be effective for CNN-based batch-mode AL, and proposed core-set selection, which samples a small subset to represent the full dataset. \\\\textcite{ash2019deep} proposed BADGE, a query strategy for DNNs, which uses k-means++ seeding on the gradients of the final layer, in order to query by uncertainty and diversity.\\nFinally, Generative Adversarial Networks (GANs; ) have also been applied successfully for AL tasks: \\n\\\\textcite{zhu2017generative} use GANs for query synthesis of images within an active learner using an SVM model.\\nThe instances are synthesized so that they would be classified with high uncertainty. \\nThe authors report this approach to outperform random sampling, pool-based uncertainty sampling using an SVM , and in some cases passive learning, while having the weakness to generate too similar instances.\\nThe approach itself is neither pure NN-based, nor does it belong to the pool-based scenario, however, it is the first reported use of GANs for AL.\\n\\\\textcite{ducoffe2018adversarial} use adversarial attacks to find instances that cross the decision boundary with the aim to increase the model robustness.\\nThey train two CNN architectures and report results superior to the core-set strategy on image classification tasks.\\nIt is obvious that GANs inherently belong to the membership query synthesis scenario.\\n\\\\revised{Therefore their performance correlates with the quality of artificial data synthesis, i.e. they are usually not that effective for NLP tasks. This has already been recognized and first improvements towards a better text generation have been made }.\",\n \"id\": \"fcf92379-9f52-4b46-8476-f3cf95e27e6c\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"aff7fdcf-a8ec-4ce1-9895-2dad247f6097\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Neural-Network-Based Active Learning\"\n ],\n [\n \"paragraph\",\n \"Previous Work\"\n ]\n ],\n \"subsections\": [\n \"72864fc8-a21e-4cfd-8e06-6087c4b614bf\",\n \"17995907-7b29-4bbf-84e0-f3a0b81b8627\"\n ],\n \"title\": \"Previous Work\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 4025,\n 759,\n 1042,\n 3288\n ],\n \"content\": \"One of the earliest and in many variations adopted class of strategies is uncertainty sampling .\\n\\\\label{sect:nn_uncertainty}\\nUnfortunately, this widely-used concept is not straightforward to apply for NNs, as they do not provide an inherent indicator of uncertainty.\\nIn the past, this has been tackled among others by ensembling , or by learning error estimates .\\nMore recent approaches furthermore use Bayesian extensions , \\nobtain uncertainty estimations using dropout ,\\nor use probabilistic NNs to estimate predictive uncertainty .\\nHowever, ensemble and Bayesian approaches quickly become infeasible on larger datasets,\\nand NN architectures are generally known to be overconfident in their predictions .\\nConsequently, uncertainty in NNs is only insufficiently solved and therefore still remains a highly relevant research area.\\n\\\\hiddennotes{\\n To circumvent this, various approaches have been suggested:\\n (1) Bayes by Backprop uses Bayesian variational inference in order to learn approximated distributions over the weights of a NN.\\n (2) extend dropout to represent uncertainties \\n by interpreting dropout-enhanced NNs as a Bayesian approximation to deep Gaussian Processes.\\n However, this has been reported to not scale well for larger datasets .\\n (3) Traditionally this has been tackled by ensembling , i.e. producing $k$-many different models and outputting a prediction by interpreting their disagreement (see \\\\textcite{olsson2009literature}).\\n The disadvantage of this approach is the obvious additional cost in runtime and model size.\\n \\\\textcite{sener2017active} report uncertainty-based approaches to be not effective for CNNs, and \\\\textcite{ducoffe2018adversarial} observed worse performance compared to random sampling.\\\\\\\\\\n}\",\n \"id\": \"72864fc8-a21e-4cfd-8e06-6087c4b614bf\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"fcf92379-9f52-4b46-8476-f3cf95e27e6c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Neural-Network-Based Active Learning\"\n ],\n [\n \"paragraph\",\n \"Previous Work\"\n ],\n [\n \"paragraph\",\n \"Uncertainty in Neural Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Uncertainty in Neural Networks\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 7764,\n 1096\n ],\n \"content\": \"DNNs are known to excel in particularly at large-scale datasets,\\nbut often having large amounts of data available is a strict requirement to perform well at all (e.g., ).\\nAL on the other hand tries to minimize the labeled data.\\nThe small labeled datasets can be a problem for DNNs, since they are known to overfit on small datasets (e.g., ), \\nwhich results in bad generalization performance on the test set.\\nMoreover, DNNs often offer little advantage over shallow models when they are trained using small datasets , thereby lacking justification for their higher computational costs.\\nOn the other hand we clearly cannot require AL to label more data, since this would defeat its purpose.\\nTherefore there has been research on dealing with (D)NNs using small datasets, however, it is only a scarce amount, especially in relation to the large volume of NN literature in general.\\nHandling small datasets is mostly circumvented by using pre-training or other transfer learning approaches .\\nFinally, the search for optimal hyperparameters is often neglected and instead the hyperparameters of related work are used, which are optimized for large datasets, if at all.\",\n \"id\": \"17995907-7b29-4bbf-84e0-f3a0b81b8627\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"fcf92379-9f52-4b46-8476-f3cf95e27e6c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning\"\n ],\n [\n \"subsection\",\n \"Neural-Network-Based Active Learning\"\n ],\n [\n \"paragraph\",\n \"Previous Work\"\n ],\n [\n \"paragraph\",\n \"Contrasting Paradigms\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Contrasting Paradigms\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sect:text_classification}\\nIn Sections \\\\ref{subsect:tc_recent_adv} and \\\\ref{subsect:active_tc} we first summarize recent methods in text classification and NNs.\\nWe elaborate on each method's importance in the context of AL, and analyze its adoption by recent research where applicable.\\nFor insufficiently adopted methods, we present how they could advance AL for text classification.\\nMost importantly, we present an overview of recent experiments in AL for text classification and analyze commonalities and shortcomings.\",\n \"id\": \"909d9662-92b7-4a13-b530-6c9553233a2e\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning for Text Classification\"\n ]\n ],\n \"subsections\": [\n \"54497d60-bdb0-4315-828f-3d8f53e9c84c\",\n \"5247a74b-cf2b-493b-a901-1dbfdda01e22\",\n \"c22085b6-c5bc-4141-87dc-c353d7a13732\"\n ],\n \"title\": \"Active Learning for Text Classification\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{subsect:tc_recent_adv}\",\n \"id\": \"54497d60-bdb0-4315-828f-3d8f53e9c84c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"909d9662-92b7-4a13-b530-6c9553233a2e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning for Text Classification\"\n ],\n [\n \"subsection\",\n \"Recent Advances in Text Classification\"\n ]\n ],\n \"subsections\": [\n \"ad081844-efd8-44ae-a153-552a2838fa22\"\n ],\n \"title\": \"Recent Advances in Text Classification\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 7265,\n 7165,\n 856,\n 1684,\n 7370,\n 11,\n 826,\n 1557,\n 8385,\n 673\n ],\n \"content\": \"Traditional methods use the bag-of-words (BoW) representation, which are sparse and high-dimensional.\\nHowever, with the introduction of word embeddings like word2vec , \\nGloVe , or fastText ,\\nword embeddings have replaced BoW representations in many cases.\\nThis is due to several reasons:\\n(1) They represent semantic relations in vectors space and avoid the problem of mismatching features as for example due to synonymy;\\n(2) incorporating word embeddings resulted in superior performance for many downstream tasks ;\\n(3) unlike bag-of-words, word vectors are dense, low-dimensional representations, \\nwhich makes them applicable to a wider range of algorithms --\\nespecially in the context of NNs which favor fixed-size inputs.\\nVarious approaches have been presented in order to obtain similar fixed size representations for word sequences, i.e. sentences, paragraphs or documents .\\\\\\\\\\n\\\\pindent Word embeddings are representations, which provide exactly one vector per word and in consequence one meaning as well.\\nThis makes them also unaware of the current word's context and therefore makes them unable to detect and handle ambiguities.\\nUnlike word embeddings, language models (LMs) compute the word vector using the word and the surrounding context .\\nThis results in a contextualized representation, which inherits the advantages of word embeddings, and at the same time allows for context-specific representation (in contrast to static embeddings) . \\nELMo was the first LM to gain wide adoption and surpassed state of the art models on several NLP tasks .\\nShortly thereafter, BERT was introduced and provided bidirectional pre-training-based language modelling.\\nThe process to create a BERT-based model consists of a pre-training and a fine-tuning step as opposed to ELMO's direct feature-based approach in which contextualized vectors are obtained from the pre-trained model and used directly as features .\\nBy masking, i.e. randomly removing a fraction of tokens during training, the training was adapted to predict the masked words.\\nThis made the bidrectional training possible, which would otherwise be obstructed because a word could \\\"see itself\\\" when computing its probability of occurrence given a context .\\nFollowing this, XLNet introduced a similar approach of pre-training and fine-tuning using an autoregressive language model, however, it overcame BERT's limitation as it does not rely on masking data during pre-training , and moreover, successfully manages to integrate the recent TransformerXL architecture .\\nSince then, a variety of LMs have been published, which further optimize the pre-training of previous LM architectures (e.g., RoBERTa and ELECTRA ), or distill the knowledge into a smaller model (e.g., DistilBERT ).\\nSimilarly to word embeddings, there are approaches to use LMs in order to obtain sentence representations from LMs .\\\\\\\\\\n\\\\pindent All mentioned representations offer a richer expressiveness than traditional BoW representations and therefore are well-suited for active learning purposes.\",\n \"id\": \"ad081844-efd8-44ae-a153-552a2838fa22\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"54497d60-bdb0-4315-828f-3d8f53e9c84c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning for Text Classification\"\n ],\n [\n \"subsection\",\n \"Recent Advances in Text Classification\"\n ],\n [\n \"paragraph\",\n \"Representations\"\n ]\n ],\n \"subsections\": [\n \"e093196b-af7c-4543-b2db-908cbfcfea1d\"\n ],\n \"title\": \"Representations\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7595,\n 673,\n 11,\n 4026\n ],\n \"content\": \"\\\\label{subsect:nn_text_classification}\\nA well-known CNN architecture presented by \\\\textcite{kim2014convolutional} (KimCNN) operates on pre-trained word vectors and achieved state of the art results at the time using only a simple but elegant architecture.\\nThe investigated CNN setups did not require much hyperparameter tuning and confirmed the effectiveness of dropout as a regularizer for CNN-based text classification.\\\\\\\\\\n\\\\pindent The word embeddings of fastText differ from other word embeddings in the sense that the approach is (1) supervised and (2) specifically designed for text classification.\\nBeing a shallow neural network, it is still very efficient, while still obtaining performances comparable to deep learning approaches at that time.\\\\\\\\\\n\\\\pindent\\\\textcite{howard2018universal} developed Universal Language Model Fine-tuning (ULMFiT), a LM transfer learning method using the AWD-LSTM architecture ,\\nwhich outperformed the state of the art on several text classification datasets when trained on only $100$ labeled examples, and thereby achieved results significantly superior to more sophisticated architectures of previous work.\\nContext-specific LMs like BERT and XLNet yield a context-dependent vector for each token, thereby strongly improving NN-based text classification .\\nState of the art in NN-based text classification is LM-based fine-tuning with XLNet, which has a slight edge over BERT in terms of test error rate .\\nULMFiT follows closely thereafter, and KimCNN is still a strong contender.\\nNotably, ULMFiT, BERT and XLNet all perform {\\\\em transfer learning}, which aims to transfer knowledge from one model to another ,\\nthereby massively reducing the required amounts of data.\",\n \"id\": \"e093196b-af7c-4543-b2db-908cbfcfea1d\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"ad081844-efd8-44ae-a153-552a2838fa22\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning for Text Classification\"\n ],\n [\n \"subsection\",\n \"Recent Advances in Text Classification\"\n ],\n [\n \"paragraph\",\n \"Representations\"\n ],\n [\n \"paragraph\",\n \"Neural-Network-Based Text Classification\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Neural-Network-Based Text Classification\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 7765,\n 4023,\n 4022,\n 8716,\n 11\n ],\n \"content\": \"\\\\label{subsect:active_tc} \\nTraditional AL for text classification heavily relied on query strategies based on prediction-uncertainty and ensembling . \\nCommon model choices included support vector machines (SVMs; ), naive bayes , logistic regression and neural networks .\\nTo the best of our knowledge, no previous survey covered traditional AL for text classification, however, ensembling-based AL for NLP has been covered in depth by \\\\textcite{olsson2009literature}.\\\\\\\\\\n\\\\pindent Regarding modern NN-based AL for text classification, the relevant models are primarily CNN- and LSTM-based deep architectures: \\\\textcite{zhang2017active} claim to be the first to consider AL for text classification using DNNs.\\nThey use CNNs and contribute a query strategy, which selects the instances \\\\phrasing{based on the expected change of the word embeddings and the model's uncertainty given the instance,} thereby learning discriminative embeddings for text classification.\\n\\\\textcite{an2018deep} evaluated SVM, LSTM and gated recurrent unit (GRU; \\\\textcite{cho2014properties}) models, and reported that the latter two significantly outperformed the SVM baseline on the Chinese news dataset ThucNews.\\n\\\\textcite{lu2020investigating} investigated the performance of different text representations in a pool-based AL scenario.\\nThey compared frequency-based text representations, word embeddings and transformer-based representations used as input features for a SVM-based AL and different query strategies,\\nin which transformer-based representations yielded consistently higher scores.\\n\\\\textcite{prabhu2019sampling} investigate sampling bias and apply active text classification on the large scale text corpora of \\\\textcite{zhang2016character-level}.\\nThey demonstrate FastText.zip with (entropy-based) uncertainty sampling to be a strong baseline, which is competitive compared to recent approaches in active text classification.\\nMoreover, they use this strategy to obtain a surrogate dataset (comprising from 5\\\\% to 40\\\\% of the total data) on which a LSTM-based LM is trained using ULMFiT , reaching accuracy levels close to a training on the full dataset.\\nUnlike past publications, they report this uncertainty-based strategy to be effective, robust, and at the same time computationally cheap.\\nThis is the most relevant work in terms of \\\\revised{the intersection between text classification, NNs and DL}.\\n\\\\hiddennotes{\\n \\\\begin{itemize}\\n \\\\item Previous Research nach Query Strategy Class / Embeddings untersuchen\\n \\\\item Reinforcement Learning using NNs\\n \\\\end{itemize}\\n}\\n\\\\begin{table}[h!]\\n \\\\begin{center}\\n \\\\begin{tabular}{l P{3cm} P{2.5cm} P{4.5cm}}\\n \\\\hline\\n \\\\textbf{Publication} & \\\\textbf{Datasets} & \\\\textbf{Model(s)} & \\\\textbf{Query Strategy Class(es)}\\\\\\\\\\n \\\\hline\\n \\\\mbox{} & 20N, R21, RV2, SPM & NB, SVM, kNN & 1. Prediction uncertainty (LC)\\\\newline 2. Prediction uncertainty (CTH)\\\\newline 3. Prediction uncertainty (disagreement) \\\\\\\\\\n \\\\mbox{} & CR, MR, SJ, MRL, MUR, DR & CNN & 1. Model uncertainty (EGL)\\\\newline 2. Prediction Uncertainty (entropy)\\\\\\\\\\n \\\\mbox{} & RMA & SVM & 1. Prediction uncertainty (CTH)\\\\newline 2. Prediction uncertainty (disagreement)\\\\\\\\\\n \\\\mbox{} & TQA, MR & SVM, CNN, BiLSTM & Prediction uncertainty (disagreement)\\\\\\\\\\n \\\\mbox{} & MR, SJ, TQA, CR & SVM, CNN, BiLSTM & 1. Prediction uncertainty (entropy)\\\\newline 2. Prediction uncertainty (disagreement)\\\\\\\\\\n \\\\mbox{} & SGN, DBP, YHA, YRP, YRF, AGN, ARP, ARF & FTZ, ULMFiT & Prediction uncertainty (entropy)\\\\\\\\\\n \\\\mbox{} & MRL, MDS, BAG, G13, ACR, SJ, AGN, DBP & SVM & 1. Prediction uncertainty (CTH) \\\\newline 2. Prediction uncertainty (disagreement) \\\\newline 3. Data-based (EGAL) \\\\newline 4. Data-based (density)\\\\\\\\\\n \\\\hline \\n \\\\end{tabular}\\n \\\\captionsetup{aboveskip=10pt}\\n \\\\caption{An overview of recent work on AL for text classification.\\n We referred to the datasets using short keys, which can be looked up in Table \\\\ref{tab:datasets} in the Appendix. Models: Naive Bayes (NB), Support Vector Machine (SVM), k-Nearest Neighbours (kNN), Convolutional Neural Network (CNN), [Bidirectional] Long Short-Term Memory ([Bi]LSTM), FastText.zip (FTZ), Univeral Language Model Fine-Tuning (ULMFiT). Query strategies: Least confidence (LC), Closest-to-hyperplane (CTH), expected gradient length (EGL).\\n Random selection baselines were omitted.}\\n \\\\label{table:recent_work}\\n\\\\end{center}\\n\\\\end{table}\",\n \"id\": \"5247a74b-cf2b-493b-a901-1dbfdda01e22\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"909d9662-92b7-4a13-b530-6c9553233a2e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning for Text Classification\"\n ],\n [\n \"subsection\",\n \"Text Classification for Active Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Text Classification for Active Learning\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 11,\n 1096\n ],\n \"content\": \"\\\\label{subsec:experiments}\\nTable \\\\ref{table:recent_work} shows the most recent AL for text classification experiments, all of them more recent than the surveys of \\\\textcite{settles2010active} and \\\\textcite{olsson2009literature}. \\nFor each publication we list the utilized datasets, models, and classes of query strategies (with respect to the taxonomy in Section \\\\ref{subsec:query_strategies}).\\nWe present this table in order to get insights about the recently preferred classification models and query strategy classes.\\\\\\\\\\n\\\\pindent We can draw multiple conclusions from Table \\\\ref{table:recent_work}:\\nIt is obvious that a significant majority of these query strategies belong to the class of prediction-based query strategies, more specifically to the prediction-uncertainty and disagreement-based sub-classes. \\nIn addition to that, we can identify several shortcomings:\\nFirst, in many experiments two or more standard datasets are evaluated, but very often there is little to no intersection between the experiments in terms of their datasets. \\nAs a result we lose comparability against previous research.\\nFor recent research, this can seen in Table \\\\ref{table:recent_work}, where the only larger intersections is between the works of \\\\textcite{zhang2017active} and \\\\textcite{lowell2019practical}.\\n\\\\textcite{siddhant2018deep} provide at least some comparability against \\\\textcite{zhang2017active} and \\\\textcite{lowell2019practical} through one dataset each.\\nAdditionally, RMA is a subset of R21 , which are used by \\\\textcite{bloodgood2018support} and \\\\textcite{hu2016active}, so they might be comparable to some degree.\\n are the only ones to evaluate on the more recent large-scale text classification datasets , and although these datasets are more realistic in terms of their size, the authors omitted the classic datasets, so it is difficult to relate their contributions to previous work.\\nMoreover, as a result of this, we do not know if and to what degree past experiments generalize to DNNs .\\n\\\\noindent Finally, it is not clear if recent (D)NNs benefit from the same query strategies, i.e. past findings may not apply to modern NN architectures:\\n\\\\textcite{prabhu2019sampling} identified contradicting statements in recent literature about the effectiveness of using prediction uncertainty in combination with NNs.\\nThey achieved competitive results using a FastText.zip (FTZ) model and a prediction uncertainty query strategy, which proved to be very effective while requiring only a small amount of data, despite all reported weaknesses concering NNs and uncertainty estimates.\",\n \"id\": \"c22085b6-c5bc-4141-87dc-c353d7a13732\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"909d9662-92b7-4a13-b530-6c9553233a2e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Active Learning for Text Classification\"\n ],\n [\n \"subsection\",\n \"Commonalities and Limitations of Previous Experiments\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Commonalities and Limitations of Previous Experiments\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"847fa582-b0b5-403a-9242-ea4d17738bff\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Open Research Questions\"\n ]\n ],\n \"subsections\": [\n \"b759602b-ca2d-4e2f-9b70-cf51e332ff01\"\n ],\n \"title\": \"Open Research Questions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In Section \\\\ref{sect:active_learning} it was illustrated that uncertainty-based strategies \\nhave been used successfully in combination with non-NN models, and in Section \\\\ref{subsec:experiments} it was shown that they also account for the largest fraction of query strategies in recent NN-based AL. \\nUnfortunately, uncertainty in NNs is still challenging due to inaccurate uncertainty estimates, or limited scalability (as described in Section \\\\ref{sect:nn_uncertainty}).\",\n \"id\": \"b759602b-ca2d-4e2f-9b70-cf51e332ff01\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"847fa582-b0b5-403a-9242-ea4d17738bff\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Open Research Questions\"\n ],\n [\n \"paragraph\",\n \"Uncertainty Estimates in Neural Networks\"\n ]\n ],\n \"subsections\": [\n \"00aa4769-dd36-4d2f-ba39-18878b0e6be3\"\n ],\n \"title\": \"Uncertainty Estimates in Neural Networks\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 11,\n 4023\n ],\n \"content\": \"As outlined in Section \\\\ref{subsect:tc_recent_adv}, the use of text representations in NLP has shifted from bag-of-words to static and contextualized word embeddings.\\nThese representations evidentially provide many advantages like disambiguation capabilities, non-sparse vectors, and an increase in performance for many tasks.\\nAlthough there have been some applications , there is no AL-specific systematic evaluation to compare word embeddings and LMs using NNs.\\nMoreover, they are currently only scarcely used,\\nwhich hints at either a slow adoption, or some non-investigated practical issues.\",\n \"id\": \"00aa4769-dd36-4d2f-ba39-18878b0e6be3\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"b759602b-ca2d-4e2f-9b70-cf51e332ff01\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Open Research Questions\"\n ],\n [\n \"paragraph\",\n \"Uncertainty Estimates in Neural Networks\"\n ],\n [\n \"paragraph\",\n \"Representations\"\n ]\n ],\n \"subsections\": [\n \"6e1cfbc6-56a2-4243-bbb5-f779c6fa69b3\",\n \"fd5f0c0a-5428-46a3-b6af-76f50107da5a\",\n \"1b466353-341c-4aec-900d-656fddb7ff9e\"\n ],\n \"title\": \"Representations\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"DL approaches are usually applied in the context of large datasets.\\nAL, however, necessarily intends to keep the (labeled) dataset as small as possible.\\nIn Section \\\\ref{sect:active_learning} we outlined why small datasets can be challenging for DNNs, and as a direct consequence as well for DNN-based AL.\\nUsing pre-trained language models, this problem is alleviated to some degree because fine-tuning allows training models using considerably smaller datasets.\\nNonetheless, it is to be investigated how little data is still necessary to successfully fine-tune a model.\",\n \"id\": \"6e1cfbc6-56a2-4243-bbb5-f779c6fa69b3\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"00aa4769-dd36-4d2f-ba39-18878b0e6be3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Open Research Questions\"\n ],\n [\n \"paragraph\",\n \"Uncertainty Estimates in Neural Networks\"\n ],\n [\n \"paragraph\",\n \"Representations\"\n ],\n [\n \"paragraph\",\n \"Small Data DNNs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Small Data DNNs\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 11\n ],\n \"content\": \"In Section \\\\ref{subsec:experiments} we provided an overview of the most common AL strategies for text classification.\\nUnfortunately, the combinations of datasets used in the experiments are often completely disjoint, e.g. \\\\textcite{siddhant2018deep}, \\\\textcite{lowell2019practical}, and \\\\textcite{prabhu2019sampling}.\\nAs a consequence, comparability is decreased or even lost, especially between more recent and past work.\\nComparibility is, however, crucial to verify if past insights regarding shallow NN-based AL still apply in context of DNN-based AL .\",\n \"id\": \"fd5f0c0a-5428-46a3-b6af-76f50107da5a\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"00aa4769-dd36-4d2f-ba39-18878b0e6be3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Open Research Questions\"\n ],\n [\n \"paragraph\",\n \"Uncertainty Estimates in Neural Networks\"\n ],\n [\n \"paragraph\",\n \"Representations\"\n ],\n [\n \"paragraph\",\n \"Comparable Evaluations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Comparable Evaluations\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 8718,\n 4027\n ],\n \"content\": \"There is an abundance of query strategies to choose from, which we have (non-exhaustively) categorized in Section \\\\ref{subsec:query_strategies}.\\nThis introduces the problem of choosing the optimal strategy.\\nThe right choice depends on many factors like data, model, or task, and can even vary between different iterations during the AL process.\\nAs a result, {\\\\em learning to learn} (or {\\\\em meta-learning}) has become popular and can be used to learn the optimal selection , or even learn query strategies as a whole .\",\n \"id\": \"1b466353-341c-4aec-900d-656fddb7ff9e\",\n \"level\": \"paragraph\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"00aa4769-dd36-4d2f-ba39-18878b0e6be3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Open Research Questions\"\n ],\n [\n \"paragraph\",\n \"Uncertainty Estimates in Neural Networks\"\n ],\n [\n \"paragraph\",\n \"Representations\"\n ],\n [\n \"paragraph\",\n \"Learning to Learn\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Learning to Learn\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 4023\n ],\n \"content\": \"In this survey, we investigated (D)NN-based AL for text classification and inspected factors obstructing its adoption.\\nWe created a taxonomy, distinguishing query strategies by their reliance on data-based, model-based, and prediction-based input information.\\nWe analyzed query strategies used in AL for text classification and categorized them into the respective taxonomy classes.\\nWe presented the intersection between AL, text classification and DNNs, which is to the best of our knowledge the first survey of this topic.\\nFurthermore, we reviewed (D)NN-based AL, identified current challenges and state of the art, and pointed out that it is both underresearched and often lacks comparability.\\nIn addition to that, we presented relevant recent advances in NLP, related them to AL, and showed gaps and limitations for their application.\\nOne of our main findings is that uncertainty-based query strategies are still the most widely used class, regardless of whether the analysis is restricted to NNs.\\nLM-based representations offer finer-grained context-specific representations while also handling out-of-vocabulary words. \\nMoreover, we find fine-tuning-based transfer learning alleviates the small data problem to some degree but lacks adoption.\\nMost important DNNs are known for their strong performance on many tasks and first adoptions in AL have shown promising results .\\nAll these gains would be highly desirable for AL.\\nTherefore improving the adoption of DNNs in AL is crucial, especially since the expected increases in performance could be either used to improve the classification results while using the same amount of data or to increase the efficiency of the labeling process by reducing the data and therefore the labeling efforts.\\nBased on these findings we identify research directions for future work in order to further advance (D)NN-based AL.\\n\\\\section*{Acknowledgements}\\nWe thank Gerhard Heyer for his valuable feedback on the manuscript, Lydia Müller for fruitful discussions about the taxonomy and advice thereon, and Janos Borst for sharing his thoughts on recent advances in language models.\\nThis research was partially funded by the Development Bank of Saxony\\n(SAB) under project number 100335729.\\n\\\\printbibliography\\n\\\\appendix\",\n \"id\": \"3ba159ca-0eee-4f69-aedd-91724bce7349\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"ca5cf519-fe65-47a3-9e14-707edbf02174\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e01ebc7-074c-4060-b09d-956a012ff8ea\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Appendix\"\n ]\n ],\n \"subsections\": [\n \"01e0ed63-a97a-4079-8877-7a6544e5b88a\"\n ],\n \"title\": \"Appendix\"\n },\n {\n \"cite_extract_rate\": 0.17647058823529402,\n \"cites\": [\n 1096,\n 4028,\n 8420\n ],\n \"content\": \"The following table provides additional information about the datasets which were referred to in Section \\\\ref{subsect:active_tc}. \\n{\\n\\\\setlength{\\\\LTcapwidth}{\\\\textwidth}\\n\\\\begin{longtable}{l p{3.8cm} l p{4.1cm} r r}\\n \\\\hline\\n \\\\textbf{Id} & \\\\textbf{Name} & \\\\textbf{Type} & \\\\textbf{Publication} & \\\\textbf{\\\\#Train} & \\\\textbf{\\\\#Test} \\\\\\\\\\n \\\\hline\\n TQA & TREC QA & MC & & 5,500 & 500 \\\\\\\\\\n CR & Customer Reviews & MC & & \\\\textsuperscript{*}{315} & - \\\\\\\\\\n ACR & Additional Customer\\\\newline Reviews & MC & & \\\\textsuperscript{*}{325} & - \\\\\\\\\\n MDS & Multi-Domain Sentiment & B & & \\\\textsuperscript{**}{8,000} & - \\\\\\\\\\n BAG & Blog Author Gender & B & & 3,100 & - \\\\\\\\\\n G13 & Guardian 2013 & MC & & 6,520 & -\\\\\\\\\\n MR & Movie Reviews & B & & 10,662 & - \\\\\\\\\\n MRL & Movie Reviews Long & B & & 2,000 & - \\\\\\\\\\n MUR & Music Review & B & & 2,000 & - \\\\\\\\\\n DR & Doctor Reviews & MC & & 58,110 & - \\\\\\\\\\n SJ & Subjectivity & B & & 10,000 & -\\\\\\\\\\n 20N & 20newsgroups & MC & & \\\\textsuperscript{***}18,846 & - \\\\\\\\\\n R21 & Reuters-21578 & ML & & 21578 & - \\\\\\\\\\n RMA & Reuters ModApté & ML & & 9,603 & 3,299\\\\\\\\\\n RV2 & RCV1-V2 & ML & & 23,149 & 781,265 \\\\\\\\\\n SPM & Spam & B & & 1,000 & - \\\\\\\\\\n AGN & AG News & MC & \\\\newline & 120,000 & 7,600 \\\\\\\\\\n SGN & Sogou News & MC & & 450,000 & 60,000\\\\\\\\\\n DBP & DBPedia & MC & & 560,000 & 70,000\\\\\\\\\\n YRP & Yelp Review Polarity & B & & 560,000 & 38,000\\\\\\\\\\n YRF & Yelp Review Full & MC & & 650,000 & 50,000\\\\\\\\\\n YAH & Yahoo! Answers & MC & & 1,400,000 & 60,000\\\\\\\\\\n ARP & Amazon Review Polarity & B & & 3,600,000 & 40,000\\\\\\\\\\n ARF & Amazon Review Full & MC & & 3,000,000 & 650,000\\\\\\\\\\n \\\\hline\\n \\\\caption{A collection of widely-used text classification datasets.\\n The column \\\"Type\\\" denotes the classification setting (B = binary, MC = multi-class, ML = multi-class multi-label). \\n The columns \\\"\\\\#Train\\\" and \\\"\\\\#Test\\\" show the size of the train and of the test set.\\n In the case that no predefined splits were available \\\"\\\\#Train\\\" represents the full dataset's size. \\n Each dataset was assigned a short id (first column), which we use in the paper for reference.\\\\\\\\\\n \\\\newline\\n (*): documents, (**) labels reduced to positive/negative, (***) 20news-bydate with duplicates removed}\\n \\\\label{tab:datasets}\\n\\\\end{longtable}\\n}\\n\\\\end{document}\",\n \"id\": \"01e0ed63-a97a-4079-8877-7a6544e5b88a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"ca5cf519-fe65-47a3-9e14-707edbf02174\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Active Learning for Text Classification using Deep Neural Networks\"\n ],\n [\n \"section\",\n \"Appendix\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":7,"string":"7"},"all_cites":{"kind":"list like","value":[7764,1159,7165,1096,4022,8385,1684,11,1042,4024,4023,8716,8717,3986,4025,759,3288,7265,856,7370,826,1557,673,7595,4026,7765,8718,4027,4028,8420],"string":"[\n 7764,\n 1159,\n 7165,\n 1096,\n 4022,\n 8385,\n 1684,\n 11,\n 1042,\n 4024,\n 4023,\n 8716,\n 8717,\n 3986,\n 4025,\n 759,\n 3288,\n 7265,\n 856,\n 7370,\n 826,\n 1557,\n 673,\n 7595,\n 4026,\n 7765,\n 8718,\n 4027,\n 4028,\n 8420\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.728934256609426,"string":"1.728934"}}},{"rowIdx":59,"cells":{"authors":{"kind":"list like","value":["Shuanghong Shen","Qi Liu","Zhenya Huang","Yonghe Zheng","Minghao Yin","Minjuan Wang","Enhong Chen"],"string":"[\n \"Shuanghong Shen\",\n \"Qi Liu\",\n \"Zhenya Huang\",\n \"Yonghe Zheng\",\n \"Minghao Yin\",\n \"Minjuan Wang\",\n \"Enhong Chen\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey of Knowledge Tracing: Models, Variants, and Applications"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-05-06T13:05:55Z","string":"2021-05-06T13:05:55Z"},"category":{"kind":"string","value":"cs.CY"},"abstract":{"kind":"string","value":" Modern online education has the capacity to provide intelligent educational services by automatically analyzing substantial amounts of student behavioral data. Knowledge Tracing (KT) is one of the fundamental tasks for student behavioral data analysis, aiming to monitor students' evolving knowledge state during their problem-solving process. In recent years, a substantial number of studies have concentrated on this rapidly growing field, significantly contributing to its advancements. In this survey, we will conduct a thorough investigation of these progressions. Firstly, we present three types of fundamental KT models with distinct technical routes. Subsequently, we review extensive variants of the fundamental KT models that consider more stringent learning assumptions. Moreover, the development of KT cannot be separated from its applications, thereby we present typical KT applications in various scenarios. To facilitate the work of researchers and practitioners in this field, we have developed two open-source algorithm libraries: EduData that enables the download and preprocessing of KT-related datasets, and EduKTM that provides an extensible and unified implementation of existing mainstream KT models. Finally, we discuss potential directions for future research in this rapidly growing field. We hope that the current survey will assist both researchers and practitioners in fostering the development of KT, thereby benefiting a broader range of students. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"]],"subsections":["a96dd0ad-2f86-4f16-b5a9-882602bf6a06","6987e516-0520-40ba-bdfa-9ecf587cfc88","bd30f2af-cabd-4e04-ab90-5536a2f60dde","a705e159-25ef-44ad-8166-1cef83a7c265","5fc15152-8ae0-4a7f-83a2-d64f287123fe","34caa8b6-1d19-4064-837a-9dcb1549a45d","18821b92-3ccc-4057-9352-24e5c0409dd6","8a28f5df-8a26-4509-8bdf-0a8c518eb23c"],"title":"root"},{"cite_extract_rate":0.21739130434782603,"cites":[5371,5372,5370,5374,5373],"content":"\\IEEEPARstart{W}{ith} the proliferation of the Internet and mobile communication technology, online education has become increasingly popular and is now developing at an unprecedented scale .\n{This innovative style of learning provides a degree of flexibility that conventional education cannot match, which enables teaching and learning to occur at any time and any place.}\nMeanwhile, online learning systems (e.g., Coursera, ASSISTment) have proven to be even more effective than traditional learning styles, since they can offer more intelligent educational services, such as recommending individualized learning resources to students . \nTo provide these intelligent services, online learning systems continuously record a massive amount of available data about student-system interactions {(e.g., responding to exercises)}, which can be further mined to assess their knowledge levels, learning preferences, and other attributes. Specifically, Knowledge Tracing (KT) is one of the most fundamental and critical tasks for analyzing students' learning behavior data, which aims to explore the recorded student-system interactions to monitor their evolving knowledge states .\n\\begin{figure*}[t]\n\t\\vspace{-0.5cm}\n\t\\centerline{\\includegraphics[width=\\textwidth]{Figures/learning_case.pdf}}\n\t\\vspace{-0.2cm}\n\t\\caption{A simple schematic diagram of knowledge tracing. Different knowledge concepts are represented in different colors, while exercises are also depicted in the color relevant to the knowledge concepts. During the learning process, different kinds of side information are also recorded. The evolving process of the knowledge state is assessed by KT models and illustrated by the radar maps.}\n\t\\label{kt}\n\t\\vspace{-0.5cm}\n\\end{figure*}\nFig. \\ref{kt} presents a simple schematic diagram of knowledge tracing. During the learning process, online learning systems continuously record students' learning behavioral data, including exercises and their related knowledge concepts (e.g., \\emph{equality, inequality, plane vector}, \\emph{probability}, represented in various colors), and students' answers (i.e., correct or incorrect responses). A substantial amount of supplementary information is also simultaneously recorded, including response time, opportunity count, and tutor intervention, which provides a more comprehensive reflection of students' learning process. \nBased on the collected learning data, researchers are striving to maintain an estimate of students' evolving knowledge states. {For illustration, we give a case in Fig. \\ref{kt}, where the student's prior knowledge is quantified as 0.2, 0.4, 0.4, and 0.5 across four distinct knowledge concepts. The radar map serves as a visual representation of the student's knowledge mastery, which progressively expands as the student continues to acquire new knowledge in learning.} After a period of learning, the student's knowledge states reach 0.9, 0.8, 0.8, and 0.7 respectively, suggesting good knowledge growth. In the aforementioned learning process, KT models aim to monitor changes in students' knowledge states. Once we understand students' knowledge states, the learning system can customize more suitable learning schemes for different students, thereby enabling the teaching of students in accordance with their proficiency. It also allows students to better comprehend their learning process and gradually focus on improving their skills with poorly mastered concepts .\nKnowledge tracing has been studied for decades, with the first studies tracing back to the late 1970s. These initial works primarily focused on confirming the effectiveness of mastery learning . To the best of our knowledge, ~ were the first to introduce the concept of knowledge tracing, employing Bayesian networks to model the student learning process, which they referred to as Bayesian Knowledge Tracing. Since then, the significance of KT has been recognized by a broader spectrum of researchers, and increasing attention has been directed towards KT-related research. Many logistic models have been applied to KT, including Learning Factor Analysis and Performance Factor Analysis . In recent years, deep learning has greatly enhanced research into the KT task, largely due to its capacity to extract and represent features and discover intricate structure. For instance, Deep Knowledge Tracing introduced recurrent neural networks (RNNs) into the KT task and was found to significantly outperform previous methods . Following this, various methods have been introduced that employ various types of neural networks to the KT task, considering various characteristics of the learning sequence . Moreover, due to the requirements of practical applications, many variants of KT models have been continuously developed, and KT has already been broadly applied in numerous educational scenarios. \nWhile novel KT models continue to emerge, there remains a lack of comprehensive surveys exploring this young research field, particularly {regarding its} numerous variants and applications. To this end, {the current survey aims to systematically review the development of KT}. As depicted in Figure \\ref{tax}, we initially categorize existing KT models from a technical perspective, which is consistent with the majority of existing surveys . This categorization splits them into three categories: (1) Bayesian models, (2) logistic models, and (3) deep learning models. In each category, we further organize specific KT methods according to their various techniques. \nSubsequently, we introduce extensive variants of these fundamental KT models, which consider more stringent assumptions about more complete learning process in different learning phases. In addition, we present several typical applications of KT in real learning scenarios. Due to the complexity of different KT models, we have open sourced two algorithm libraries to better aid researchers and practitioners in implementing KT models and facilitate community development in this domain. These libraries, EduData\\footnote{https://github.com/bigdata-ustc/EduData} and EduKTM\\footnote{https://github.com/bigdata-ustc/EduKTM}, include most existing KT-related datasets, extensible, and unified implementations of existing KT models, and relevant resources. Finally, we discuss potential future research directions. In summary, this paper presents an extensive survey of KT that can serve as a comprehensive guide for both researchers and practitioners. \nThe remainder of this survey is structured as follows. Section \\ref{sec:overview} presents an overview of the KT task and we discuss the differences between this and previous surveys. Section \\ref{sec:models} provides a review of the three categories of fundamental KT models. \nSection \\ref{sec:variants} describes the variants of fundamental KT models. \nSection \\ref{sec:applications} introduces the extensive applications of KT in different scenarios. Section \\ref{sec:dataset} gives the summary of existing datasets for evaluating KT models and details of the algorithm libraries we have released. Section \\ref{sec:future} discusses some potential future research directions. Finally, section \\ref{sec:conclution} summarizes the paper.\n\\begin{figure*}[t]\n\t\\vspace{-0.2cm}\n\t\\centerline{\\includegraphics[width=\\textwidth]{Figures/taxonomy.pdf}}\n\t\\vspace{-0.2cm}\n\t\\caption{{An overview of knowledge tracing models.} }\n\t\\label{tax}\n\t\\vspace{-0.2cm}\n\\end{figure*}","id":"a96dd0ad-2f86-4f16-b5a9-882602bf6a06","level":"section","origin_cites_number":23,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:overview}","id":"6987e516-0520-40ba-bdfa-9ecf587cfc88","level":"section","origin_cites_number":0,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Overview"]],"subsections":["0c71baa7-2bd1-4dbb-a83e-a14394274744","c68cad9f-5ccb-45ac-9535-78433e591dfe","1f61ba5e-8077-4921-924c-62e9bddd3afc"],"title":"Overview"},{"cite_extract_rate":0.5,"cites":[5376,5375,38],"content":"In an online learning system, supposing there exists a set of students $\\mathbb{S}$ and a set of exercises $\\mathbb{E}$. {Each exercise is related to specific Knowledge Concepts (KCs). Generally, the name given to the knowledge related to exercises differs across online learning platforms. For instance, it is named \\emph{skill} in ASSISTments . To promote better understanding, we refer to these uniformly as knowledge concepts throughout this paper, and denote the set of all KCs as $\\mathbb{KC}$. Moreover, $M$ and $K$ are respectively used to represent the total number of exercises and KCs. Students are asked to answer different exercises in order to achieve mastery of the related knowledge.}\nTherefore, the learning sequence of a student can be formulated as $\\bm{X} = \\{([e_1, k_{e_1}], a_1, r_1), ([e_2, k_{e_2}], a_2, r_2), ..., ([e_t, k_{e_t}], a_t, r_t), ..., \\\\([e_N, k_{e_N}], a_N, r_N)\\}$, where the tuple $([e_t, k_{e_t}], a_t, r_t)$ represents the learning interaction at the $t-$th time step, $e_t$ represents the exercise, $k_{e_t}$ represents the exercise's related KCs, $a_t$ represents the correctness label (i.e., with 1 for correct and 0 for incorrect answers), $r_t$ stands for the side information recorded in this learning interaction, and $N$ is the length of the learning sequence. The research problem of knowledge tracing can thus be defined as follows:\n\\textbf{Given sequences of learning interactions in online learning systems, knowledge tracing aims to monitor students' evolving knowledge states during the learning process and predict their performance on future exercises. The measured knowledge states can be further applied to individualize students' learning schemes in order to maximize their learning efficiency.}\nSome recent works directly regarded the KT task as student performance prediction, without considering students' knowledge states . We agree that predicting student performance is of great significance, as it is now {the best way} to evaluate the quality of the knowledge state traced by KT models. However, we have to point out that KT focuses more on students' knowledge states, especially their interpretability and rationality, which is related to the students' acceptance of the conclusions given based on the KT model .","id":"0c71baa7-2bd1-4dbb-a83e-a14394274744","level":"subsection","origin_cites_number":6,"parent_id":"6987e516-0520-40ba-bdfa-9ecf587cfc88","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Overview"],["subsection","Problem Definition"]],"subsections":[],"title":"Problem Definition"},{"cite_extract_rate":0,"cites":[],"content":"{ As shown in Fig. \\ref{tax}, we categorize and summarize the existing KT models according to their technical differences. In more detail, the proposed taxonomy splits existing KT methods into three categories: (1) Bayesian models, which are implemented through probability model, (2) logistic models, which are implemented through logistic functions, and (3) deep learning models, which are implemented through neural networks. Specifically, we divide the deep learning models into four sub-categories according to four various neural networks, i.e., deep knowledge tracing based on recurrent neural networks, memory-aware knowledge tracing based on memory networks, attentive knowledge tracing based on self-attention mechanism, and graph-based knowledge tracing based on graph neural networks.\nIn addition to these fundamental KT models, we also introduce a large number of their variants, which respectively consider a more complete learning process in distinct learning phases: (1) modeling individualization before learning, (2) incorporating engagement during learning, (3) considering forgetting after learning, and (4) utilizing side information across learning. Moreover, we also summarize the extensive applications of KT in different educational scenarios, including learning resources recommendation, adaptive learning and broader applications beyond student learning.}","id":"c68cad9f-5ccb-45ac-9535-78433e591dfe","level":"subsection","origin_cites_number":0,"parent_id":"6987e516-0520-40ba-bdfa-9ecf587cfc88","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Overview"],["subsection","Categorization"]],"subsections":[],"title":"Categorization"},{"cite_extract_rate":0.41176470588235203,"cites":[5372,5377,5374,5370,8911,5371,1206],"content":"Given the increasing importance of KT, several recent surveys have also examined this area. These include works by , . Here, we will briefly discuss the key distinctions between these studies to highlight the necessity and significance of this survey. \nExisting surveys have either focused on specific categories of KT models or comprehensively reviewed all available KT models. {For example, provided an overview of KT in terms of Bayesian models and logistic models, compared and discussed deep learning based KT models, while presented to pay more attention to hybird models in KT.} conducted a bibliometric analysis to examine the evolution of KT research from 1992 to 2021. also presented a comprehensive survey for the KT literature, including a broad range of methods starting from the early attempts to the recent state-of-the-art techniques utilizing deep learning. summarized KT methods in the context of student performance modeling problems. \nHowever, current surveys are somewhat limited in their scope, they only provide a detailed introduction to various KT methods and comparisons between them. Given the complexity of online learning systems and the significant importance of KT research in practical applications, this survey places a greater emphasis on the variants and applications of KT models, rather than solely introducing and comparing different KT methods. Moreover, considering that datasets are collected from different systems with various setting, subjects, learning stages, and scales, \nwe do not report and compare the performance of KT models on the student performance prediction task across various datasets in this survey. have also empirically verified that no single KT model was always the best, a specific better model must consider multiple student features and the learning context.\nInstead, we have open sourced two algorithm libraries which include the majority of existing KT-related datasets and unified implementations of existing KT models. Consequently, researchers and practitioners can freely select appropriate KT models based on their specific requirements in various application scenarios. \n\\begin{table*}[t]\n\t\\vspace{-0.6cm}\n\t\\centering\n\t\\renewcommand\\arraystretch{1.0}\n\t\\caption{A summary of different types of fundamental knowledge tracing models.}\n\t\\vspace{-0.3cm}\n\t\\resizebox{\\textwidth}{!}{\n\t\t\\begin{tabular}{|c|c|c|c|c|}\n\t\t\t\\hline\n\t\t\t\\multicolumn{1}{|c|}{Category} & Typical approach & Technique & KC relationship & Knowledge state \\bigstrut\\\\\n\t\t\t\\hline\n\t\t\t\\multicolumn{1}{|c|}{\\multirow{2}[5]{*}{Bayesian models}} & Bayesian knowledge tracing & Bayesian networks & independent & \\multirow{2}[5]{*}{\\shortstack{unobservable node \\\\ in HMM}} \\bigstrut\\\\\n\t\t\t\\cline{2-4} & dynamic Bayesian knowledge tracing & dynamic Bayesian networks & pre-defined & \\multicolumn{1}{c|}{} \\bigstrut\\\\\n\t\t\t\\hline\n\t\t\t\\multicolumn{1}{|c|}{\\multirow{3}[8]{*}{Logistic models}} & learning factor analysis & \\multirow{2}[4]{*}{logistic regression} & \\multirow{3}[8]{*}{independent} & \\multirow{3}[6]{*}{\\shortstack{the output of \\\\ logistic regression function}} \\bigstrut\\\\\n\t\t\t\\cline{2-2} & performance factor analysis & \\multicolumn{1}{c|}{} & \\multicolumn{1}{c|}{} & \\multicolumn{1}{c|}{} \\bigstrut\\\\\n\t\t\t\\cline{2-3} & knowledge tracing machines & factorization machines & \\multicolumn{1}{c|}{} & \\multicolumn{1}{c|}{} \\bigstrut\\\\\n\t\t\t\\hline\n\t\t\t\\multicolumn{1}{|c|}{\\multirow{4}[6]{*}{\\shortstack{Deep learning \\\\ models}}} & deep knowledge tracing & RNN/LSTM & discover automatically & the hidden state \\bigstrut\\\\\n\t\t\t\\cline{2-5} & memory-aware knowledge tracing & memory networks & correlation weights & \\emph{value} matrix \\bigstrut\\\\\n\t\t\t\\cline{2-5} & attentive knowledge tracing & self-attention mechanism & attention weights & attentive historical knowledge state \\\\\n\t\t\t\\cline{2-5} & graph-based knowledge tracing & graph neural networks & edges in graph & aggregate in the graph \\bigstrut\\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t}\n\t\\vspace{-0.6cm}\n\t\\label{tab1}\n\\end{table*}","id":"1f61ba5e-8077-4921-924c-62e9bddd3afc","level":"subsection","origin_cites_number":17,"parent_id":"6987e516-0520-40ba-bdfa-9ecf587cfc88","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Overview"],["subsection","Differences between this and previous surveys"]],"subsections":[],"title":"Differences between this and previous surveys"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:models}\nIn this section, as shown in Table \\ref{tab1}, we will present the fundamental KT models, based on our taxonomic framework. {Specifically, we will introduce these models in accordance with their development timeline. Subsequently, we will provide a summary of these fundamental KT models.}","id":"bd30f2af-cabd-4e04-ab90-5536a2f60dde","level":"section","origin_cites_number":0,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"]],"subsections":["22232da7-baf6-4ee6-a966-d59421dce543","737cf102-cd65-4a28-91e5-d7967b691341","0b0df85b-c041-47ba-8171-c36927847515","2737aa4f-b2f6-4977-83f3-36d33949ab77"],"title":"Fundamental Knowledge Tracing Models"},{"cite_extract_rate":0,"cites":[],"content":"Bayesian models assume that the learning process adheres to a Markov process. This process allows for the estimation of students' latent knowledge states based on their observed performance. In the subsequent section, we will present two Bayesian models in our taxonomy framework: the Bayesian Knowledge Tracing (BKT) and the Dynamic Bayesian Knowledge Tracing (DBKT).","id":"22232da7-baf6-4ee6-a966-d59421dce543","level":"subsection","origin_cites_number":0,"parent_id":"bd30f2af-cabd-4e04-ab90-5536a2f60dde","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Bayesian Models"]],"subsections":["9274e535-9216-4905-8c97-6e2b15501b96","2d449595-4bbc-4a8e-8c76-39195e6d0bf8"],"title":"Bayesian Models"},{"cite_extract_rate":0,"cites":[],"content":"{BKT's structure is illustrated in Fig. \\ref{fbkt}; here, the unshaded nodes represent unobservable latent knowledge states, while the shaded nodes represent the observable answers of the student.}\nBKT is a unique instance of the Hidden Markov Model (HMM). There are two types of parameters in HMM: transition probabilities and emission probabilities. In BKT, the transition probabilities are defined by two learning parameters: (1) $P(T)$, the probability of transition from the unlearned state to the learned state; (2) $P(F)$, the probability of forgetting previously mastered knowledge. Moreover, the emission probabilities are determined by two performance parameters: (1) $P(G)$ - the probability that a student will guess correctly, despite non-mastery; (2) $P(S)$ - the probability that a student will make a mistake, despite mastery. Additionally, $P(L_0)$ represents the initial probability of mastery. BKT operates within a two-state student modeling framework: knowledge is either learned or unlearned, and there is no forgetting once a student has mastered the knowledge. Based on the observations of the student's learning interactions, the following equation is utilized to estimate the knowledge state and the probability of correct answers: \n\\begin{equation}\n\t\\small\n\t\\begin{aligned}\n\t\tP(L_n) &= P(L_{n}|Answer) + (1 - P(L_{n}|Answer))P(T), \\\\\n\t\tP(C_{n+1}) &= P(L_n)(1 - P(S)) + (1 - P(L_n))P(G),\n\t\\end{aligned}\n\\end{equation}\nwhere $P(L_n)$ is the probability that a KC is mastered at the $n$-th learning interaction, $P(C_{n+1})$ is the probability of correct answers at the next learning interaction. $P(L_n)$ is the sum of two probabilities: (1) the probability that the KC is already mastered; (2) the probability that the knowledge state will convert to the mastered state. The posterior probability $P(L_{n}|Answer)$ is estimated as follows:\n\\begin{equation}\n\t\\scriptsize\n\t\\begin{aligned}\n\t\tP(L_{n}|correct)&=\n\t\t\\frac{P(L_{n-1})(1 - P(S))}{P(L_{n-1})(1 - P(S)) + (1 - P(L_{n-1}))P(G)}, \\\\\n\t\tP(L_{n}|incorrect)&=\n\t\t\\frac{P(L_{n-1})P(S)}{P(L_{n-1})P(S) + (1 - P(L_{n-1}))(1 - P(G))}.\n\t\\end{aligned}\n\\end{equation}\n\\begin{figure}[t]\n\t\\vspace{-0.0cm}\n\t\\centerline{\\includegraphics[width=0.9\\columnwidth]{Figures/BKT_model.pdf}}\n\t\\vspace{-0.2cm}\n\t\\caption{The topology of Bayesian Knowledge Tracing . $K$ are the unobserved knowledge nodes, $A$ are the observed performance (answer) nodes, $P(L_0)$ represents the initial probability, $P(T)$ is the transition probability, $P(G)$ is the guessing probability and $P(S)$ is the slipping probability.}\n\t\\label{fbkt}\n\t\\vspace{-0.3cm}\n\\end{figure}","id":"9274e535-9216-4905-8c97-6e2b15501b96","level":"subsubsection","origin_cites_number":1,"parent_id":"22232da7-baf6-4ee6-a966-d59421dce543","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Bayesian Models"],["subsubsection","Bayesian Knowledge Tracing"]],"subsections":[],"title":"Bayesian Knowledge Tracing"},{"cite_extract_rate":0,"cites":[],"content":"\\label{dynamic}\nBKT independently models the parameters of each KC, employing a specific model for each KC. However, KCs are not completely independent, but rather hierarchical and closely related . Dynamic Bayesian networks are capable of jointly representing multiple skills within a single model, potentially enhancing the representational power of BKT. Consequently, propose dynamic Bayesian knowledge tracing (DBKT) to represent the hierarchies and relationships within KCs using dynamic Bayesian networks. This approach considers different KCs jointly within a single model. \nIn DBKT, a student's knowledge mastery is represented by binary latent variables, which are estimated based on their learning interactions. Unlike BKT, DBKT takes into account the dependencies between various KCs. For example, if $KC_1$ and $KC_2$ are prerequisites for mastering $KC_3$, students' mastery of $KC_3$ depends on their mastery of $KC_1$ and $KC_2$.\nLet $H$ denote the unobserved variables, i.e., lack of student answers and binary mastery variables. Assuming the student correctly answers an exercise associated with $KC_1$ at time step $t_1$, i.e., $a_{1,1} = 1$. The observed variables are then $a_m = a_{1,1}$ and the unobserved variables are $h_m = \\{KC_{1,1}, KC_{2,1}, KC_{3,1}, a_{2,1}, a_{3,1}\\}$. The objective of DBKT is to find the parameters $\\theta$ that maximize the joint probability $p(a_m, h_m|\\theta)$. The log-likelihood can alternatively be formulated using a log-linear model, as follows: \n\\begin{equation}\n\tL(\\bm{w}) = \\sum_{m}ln(\\sum_{h_m}exp(\\bm{w}^T\\varPhi(a_m, h_m) - ln(Z))),\n\t\\vspace{-0.2cm}\n\\end{equation}\nwhere $\\varPhi: A \\times H \\rightarrow \\mathbb{R}^{F}$ denotes a mapping from the observed space $A$ and the latent space $H$ to an $F$-dimensional feature vector. $Z$ is a normalizing constant, $\\bm{w}$ denotes the weights.","id":"2d449595-4bbc-4a8e-8c76-39195e6d0bf8","level":"subsubsection","origin_cites_number":2,"parent_id":"22232da7-baf6-4ee6-a966-d59421dce543","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Bayesian Models"],["subsubsection","Dynamic Bayesian Knowledge Tracing"]],"subsections":[],"title":"Dynamic Bayesian Knowledge Tracing"},{"cite_extract_rate":0,"cites":[],"content":"{Logistic models represent the probability of students correctly answering exercises as a logistic function of the student and KC parameters.} They first use different factors in students' learning interactions to compute an estimation of the student and KC parameters, then utilize a logistic function to transform this estimation into the prediction of the probability of mastery .\nIn the subsequent section, we will introduce three types of logistic models: (1) Learning Factor Analysis (LFA), (2) Performance Factor Analysis (PFA) and (3) Knowledge Tracing Machines (KTM).","id":"737cf102-cd65-4a28-91e5-d7967b691341","level":"subsection","origin_cites_number":1,"parent_id":"bd30f2af-cabd-4e04-ab90-5536a2f60dde","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Logistic Models"]],"subsections":["5d298d54-78eb-4db4-91fe-1e91f13c0cdc","d181c1c4-a973-41d5-b53b-df201574371b","609eb20d-a953-453f-b163-f5e296a7c090"],"title":"Logistic Models"},{"cite_extract_rate":0,"cites":[],"content":"The LFA model considers the following learning factors:\n\\begin{itemize}\n\t\\item{Initial knowledge state}: parameter $\\alpha$ estimates the initial knowledge state of each student;\n\t\\item{Easiness of KCs}: parameter $\\beta$ captures the easiness of different KCs;\n\t\\item{Learning rate of KCs}: parameter $\\gamma$ denotes the learning rate of KCs.\n\\end{itemize}\nThe standard LFA model takes the following form:\n\\vspace{-0.2cm}\n\\begin{equation}\n\tp(\\theta) = \\sigma(\\sum_{i \\in N}\\alpha_iS_i\t+ \\sum_{j \\in KCs}(\\beta_j + \\gamma_jT_j)K_j), \n\t\\vspace{-0.2cm}\n\\end{equation}\nwhere $\\sigma$ is the sigmoid function, $S_i$ is the covariate for the student $i$, $T_j$ represents the covariate for the number of interactions on KC $j$, $K_j$ is the covariate for KC $j$, $p(\\theta)$ is the estimation of the probability of a correct answer.","id":"5d298d54-78eb-4db4-91fe-1e91f13c0cdc","level":"subsubsection","origin_cites_number":1,"parent_id":"737cf102-cd65-4a28-91e5-d7967b691341","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Logistic Models"],["subsubsection","Learning Factor Analysis"]],"subsections":[],"title":"Learning Factor Analysis"},{"cite_extract_rate":0.5,"cites":[5377],"content":"The PFA model can be seen as an extension of the LFA model that is especially sensitive to the student performance. In contrast to the LFA model, PFA considers the following different factors:\n\\begin{itemize}\n\t\\item{Previous failures}: parameter $f$ is the prior failures for the KC of the student;\n\t\\item{Previous successes}: parameter $s$ represents the prior successes for the KC of the student;\n\t\\item{Easiness of KCs}: parameter $\\beta$ means the easiness of different KCs, which is the same as in the LFA model.\n\\end{itemize}\nThe standard PFA model takes the following form:\n\\vspace{-0.15cm}\n\\begin{equation}\n\tp(\\theta) = \\sigma(\\sum_{j \\in KCs}(\\beta_j + \\mu_js_{ij} + \\nu_jf_{ij})),\n\t\\label{pfa}\n\t\\vspace{-0.15cm}\n\\end{equation}\nwhere $\\mu$ and $\\nu$ are the coefficients for $s$ and $f$, which denote the learning rates\nfor successes and failures.\n\\begin{figure}[t]\n\t\\vspace{-0.2cm}\n\t\\centerline{\\includegraphics[width= 0.9\\columnwidth]{Figures/KTM.pdf}}\n\t\\vspace{-0.2cm}\n\t\\caption{Example of activation of a knowledge tracing machine . $V$ refers to the matrix of embeddings, $w$ refers to the vector of biases, $x$ is the encoding vector of the learning interaction.}\n\t\\label{fktm}\n\t\\vspace{-0.6cm}\n\\end{figure}","id":"d181c1c4-a973-41d5-b53b-df201574371b","level":"subsubsection","origin_cites_number":2,"parent_id":"737cf102-cd65-4a28-91e5-d7967b691341","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Logistic Models"],["subsubsection","Performance Factor Analysis"]],"subsections":[],"title":"Performance Factor Analysis"},{"cite_extract_rate":0.25,"cites":[5377],"content":"The KTM model, developed by Vie et al. , employs factorization machines (FMs) to generalize logistic models to higher dimensions. FMs were initially introduced as a general predictor capable of working with any real-valued feature vector, enabling the model to represent all interactions between variables using factorized parameters . FMs provide a means of encoding side information about exercises or students into the model. Figure \\ref{fktm} illustrates the example of KTM, which models the knowledge mastery of the student based on a sparse set of weights for all features involved in the event. Let $L$ be the number of features; here, the features can be related to students, exercises, KCs, or any other side information. The learning interaction is encoded by a sparse vector $\\bm{l}$ of length $L$. When feature $i$ is involved in the interaction, $l_i > 0$. The probability $p(\\theta)$ of the correct answer is determined by the following equations: \n\\vspace{-0.3cm}\n\\begin{equation}\n\tp(\\theta) = \\sigma(\\mu + \\sum_{i = 1}^{L}w_il_i + \\sum_{1 \\leq i < j \\leq L}l_il_j\\langle\\bm{v_i}, \\bm{v_j}\\rangle ),\n\t\\vspace{-0.2cm}\n\\end{equation}\nwhere $\\mu$ is the global bias, the feature $i$ is modeled by the bias $w_i \\in \\bm{R}$ and the embedding $\\bm{v}_i \\in \\bm{R}^d$ ($d$ is the dimension). Note that only features with $l_i > 0$ will have impacts on the predictions.","id":"609eb20d-a953-453f-b163-f5e296a7c090","level":"subsubsection","origin_cites_number":4,"parent_id":"737cf102-cd65-4a28-91e5-d7967b691341","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Logistic Models"],["subsubsection","Knowledge Tracing Machines"]],"subsections":[],"title":"Knowledge Tracing Machines"},{"cite_extract_rate":1,"cites":[5378,5371],"content":"\\label{deep}\nThe cognitive process can be influenced by various factors at both the macro and micro levels. It is difficult for Bayesian models or logistic models to adequately capture a cognitive process of high complexity .\nDeep learning, with its potent ability to achieve non-linearity and feature extraction, is well-suited for modeling complex learning processes, particularly when a significant amount of learning interaction data is available . {In recent years, numerous research works have been proposed on deep learning KT models, we will introduce deep learning models from four sub-categories: (1) deep knowledge tracing, (2) memory-aware knowledge tracing, (3) attentive knowledge tracing, and (4) graph-based knowledge tracing.}","id":"0b0df85b-c041-47ba-8171-c36927847515","level":"subsection","origin_cites_number":2,"parent_id":"bd30f2af-cabd-4e04-ab90-5536a2f60dde","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Deep Learning Models"]],"subsections":["6bda0552-a0c8-4448-bccb-783dc9dbe658","a1cbba76-ad36-4969-8878-16538f12140f","b60bbb0c-acf3-4bdd-ace6-fb43306d7dda","9d414c8f-554a-43a7-afe7-209c3c4d813e"],"title":"Deep Learning Models"},{"cite_extract_rate":0.5,"cites":[5378,5379,5371],"content":"Deep Knowledge Tracing (DKT) is the pioneering approach that introduces deep learning {to complete the KT task. DKT employs Recurrent Neural Networks (RNNs) to process the input sequence of learning interactions over time,} maintaining a hidden state that implicitly contains information about the history of all past elements of the sequence. This hidden state evolves based on both the previous knowledge state and the present input learning interaction . DKT provides a high-dimensional and continuous representation of the knowledge state, enabling it to more effectively model the complex learning process. Typically, RNNs' variant, the Long Short-Term Memory (LSTM) networks , are more frequently used in the implementation of DKT, which is further strengthened by considering forgetting. \nFig. \\ref{fdkt} illustrates the process of deep knowledge tracing. In DKT, exercises are represented by their contained KCs. For datasets with different numbers of KCs, DKT applies two different methods to convert students' learning interactions $\\bm{X} = \\{(e_1, a_1), (e_2, a_2), ..., (e_t, a_t), ..., (e_N, a_N)\\}$ into a sequence of fixed-length input vectors. More specifically, for datasets with a small number $K$ of unique KCs, $\\bm{x}_t \\in \\{0,1\\}^{2K}$ is set as a one-hot embedding, where $ \\bm{x}_t^k = 1 $ if the answer $a_t$ of the exercise with KC $k$ was correct or $\\bm{x}_t^{k + K} = 1$ if the answer was incorrect. For datasets with a large number of unique KCs, one-hot embeddings are considered too sparse. Therefore, DKT assigns each input vector $\\bm{x}_t$ to a corresponding random vector, and then uses the embedded learning sequence as the input of RNNs. A linear mapping and activation function are then applied to the output hidden states to obtain the knowledge state of students:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\begin{aligned}\\label{dkt}\n\t\t&\\bm{h}_t = tanh(\\bm{W}_{hs}\\bm{x}_t + \\bm{W}_{hh}\\bm{h}_{t-1} + \\bm{b}_h), \\\\\n\t\t&\\bm{y}_t = \\sigma (\\bm{W}_{yh}\\bm{h}_{t} + \\bm{b}_y),\n\t\\end{aligned}\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $tanh$ is the activation function, $\\bm{W}_{hs}$ is the input weights, $\\bm{W}_{hh}$ is the recurrent weights, $\\bm{W}_{yh}$ is the readout weights, and $\\bm{b}_h$ and $\\bm{b}_y$ are the bias terms.\n\\begin{figure}[t]\n\t\\vspace{-0.3cm}\n\t\\centerline{\\includegraphics[width=0.9\\columnwidth]{Figures/DKT_model.pdf}}\n\t\\vspace{-0.3cm}\n\t\\caption{The architecture of DKT .}\n\t\\label{fdkt}\n\t\\vspace{-0.6cm}\n\\end{figure}\nDespite demonstrating superior performance compared to Bayesian and logistic models, DKT has several inherent shortcomings. For instance, the lack of interpretability is a significant drawback. It is challenging to understand how the hidden states represent students' knowledge states, and the model cannot explicitly determine a student's knowledge mastery from the hidden state . Additionally, identified two unreasonable phenomena in DKT that contravene common sense. These are: (1) the inability to reconstruct observed input, and (2) inconsistent predicted knowledge states across time-steps. However, despite these shortcomings, DKT remains a promising KT model .","id":"6bda0552-a0c8-4448-bccb-783dc9dbe658","level":"subsubsection","origin_cites_number":6,"parent_id":"0b0df85b-c041-47ba-8171-c36927847515","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Deep Learning Models"],["subsubsection","Deep Knowledge Tracing"]],"subsections":[],"title":"Deep Knowledge Tracing"},{"cite_extract_rate":0.6666666666666661,"cites":[5380,1206],"content":"To enhance the interpretability of DKT, memory-aware knowledge tracing introduces an external memory module, as proposed by . This module is designed to {store the and update the corresponding knowledge mastery of the student.} The most representative example is Dynamic Key-Value Memory Networks (DKVMN) for knowledge tracing, as proposed by . DKVMN highlights students' specific knowledge states on various knowledge categories. It initializes a static matrix, referred to as a $key$ matrix to store latent KCs and a dynamic matrix, called a $value$ matrix to store and update the mastery of corresponding KCs through read and write operations over time.\nAs shown in Fig. \\ref{fdkvmn}, an embedding matrix is first defined to obtain the embedding vector $k_t$ of the exercises. A correlation weight $\\bm{w}_t$ is then obtained by taking the inner product between the exercise embedding $k_t$ and the $key$ vectors $M^k$, followed by the softmax activation:\n\\vspace{-0.2cm}\n\\begin{equation}\n\t\\bm{w}_t = Softmax(k_tM^k),\n\t\\vspace{-0.2cm}\n\\end{equation}\nwhere the correlation weight $\\bm{w}_t$ represents the correlation between the exercises and all latent KCs.\nIn the read operation, DKVMN predicts student performance based on the student's knowledge mastery. Specifically, DKVMN reads students’ mastery of the exercise $\\bm{r}_t$ with reference to the weighted sum of all memory vectors in the $value$ matrix using the correlation weight. The read content and the input exercise embeddings are then concatenated together and passed to a fully connected layer to yield a summary vector $\\bm{f}_t$, which contains both the student's knowledge mastery and the prior difficulty of the exercise. Furthermore, the student's performance can be predicted by applying another fully connected layer with a sigmoid activation function to the summary vector:\n\\vspace{-0.2cm}\n\\begin{equation}\n\t\\begin{aligned}\n\t\t&\\bm{r}_t = \\sum_{i=1}^{N}w_t(i)M_t^v(i), \\\\\n\t\t&\\bm{f}_t = tanh(\\bm{W}_f[\\bm{r}_t, k_t] + \\bm{b}_f),\\\\\n\t\t&p_t = \\sigma(\\bm{W}_p\\bm{f}_t + \\bm{b}_p),\n\t\\end{aligned}\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $\\bm{W}_f$ and $\\bm{W}_p$ are the weights, $\\bm{b}_f$ and $\\bm{b}_p$ are bias terms.\nIn the write operation, after an exercise has been answered, DKVMN updates students' knowledge mastery (i.e., the $value$ matrix) based on their performance. Specifically, the learning interaction $(e_t, a_t)$ is first embedded with an embedding matrix $\\bm{B}$ to obtain the student's knowledge growth $\\bm{v}_t$. Then DKVMN calculates an erase vector $\\bm{erase}_t$ from $\\bm{v}_t$ and decides to erase the previous memory with reference to both the erase vector and the correlation weight $\\bm{w}_t$.\nFollowing erasure, the new memory vectors are updated by the new knowledge state and the add vector $\\bm{add}_t$, which forms an $erase$-followed-by-$add$ mechanism that allows forgetting and strengthening knowledge mastery in the learning process:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\begin{aligned}\n\t\t&\\bm{erase}_t = \\sigma(\\bm{W}_e\\bm{v}_t + \\bm{b}_e), \\\\\n\t\t&\\widetilde{M}_t^v(i) = M_{t-1}^v(i)[1 - w_t(i)\\bm{erase}_t],\\\\\n\t\t&\\bm{add}_t = tanh(\\bm{W}_d\\bm{v}_t + \\bm{b}_d),\\\\\n\t\t&M_{t}^v(i) = \\widetilde{M}_t^v(i) + w_t(i)\\bm{add}_t,\\\\\n\t\\end{aligned}\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $\\bm{W}_e$ and $\\bm{W}_d$ are the weights, $\\bm{b}_e$ and $\\bm{b}_d$ are bias terms.\n point out that DKVMN failed to capture long-term dependencies in the learning process. Therefore, they propose a Sequential Key-Value Memory Network (SKVMN) to combine the strengths of DKT's recurrent modelling capacity and DKVMN's memory capacity. In SKVMN, a modified LSTM called \\emph{Hop-LSTM} is used to hop across LSTM cells according to the relevance of the latent KCs, which directly captures the long-term dependencies. During the writing process, SKVMN allows for the calculation of the knowledge growth of a new exercise, taking into consideration the current knowledge state, thereby yielding more reasonable results. \n\\begin{figure}[t]\n\t\\centerline{\\includegraphics[width=0.9\\columnwidth]{Figures/DKVMN.pdf}}\n\t\\vspace{-0.3cm}\n\t\\caption{The architecture of DKVMN .}\n\t\\label{fdkvmn}\n\t\\vspace{-0.6cm}\n\\end{figure}","id":"a1cbba76-ad36-4969-8878-16538f12140f","level":"subsubsection","origin_cites_number":3,"parent_id":"0b0df85b-c041-47ba-8171-c36927847515","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Deep Learning Models"],["subsubsection","Memory-aware Knowledge Tracing"]],"subsections":[],"title":"Memory-aware Knowledge Tracing"},{"cite_extract_rate":0.5,"cites":[679,5376,5374,5375,5370,38],"content":"\\label{attentive}\nIn the development of deep learning, the Transformer is initially proposed for neural machine translation , which abandons recurrence and solely relies on the self-attention mechanism to capture global dependencies within a sequence. The Transformer has been demonstrated to excel in feature extraction and dependency capture, while maintaining high computational efficiency. Some representative pre-training models based on the Transformer, such as BERT and GPT , have obtained state-of-the-art results on various natural language processing tasks. propose a self-attentive model for knowledge tracing (SAKT), which directly apply the Transformer to capture long-term dependencies between students' learning interactions. Furthermore, introduce an adaptive sparse self-attention network to generate missing features and simultaneously produce fine-grained predictions of student performance. employ a multi-head ProbSparse self-attention mechanism to mitigate the time complexity and effectively capture the long-term dependencies in students' learning interactions. \nHowever, the complexity of the KT task often limits the performance of the aforementioned simple Transformer applications. introduce a novel approach named Separated Self-Attentive Neural Knowledge Tracing (SAINT) to enhance self-attentive computation for KT adaptability. Specifically, SAINT employs an encoder-decoder structure, with the exercise and answer embeddings being separately encoded and decoded by self-attention layers. The separation of the input allows SAINT to stack self-attention layers multiple times, thus capturing complex relations in student interactions. Subsequently, introduce the SAINT+ model, which integrates two temporal features into SAINT: namely, the time taken to answer each exercise and the interval time between consecutive learning interactions. Both SAINT and SAINT+ have outperformed the SAKT model on the student performance prediction task. \nAdditionally, observe that SAKT does not surpass DKT and DKVMN in their experiments. Unlike SAINT and SAINT+, they present a context-aware attentive knowledge tracing (AKT) model. This model integrates the self-attention mechanism with psychometric models, creating a more effective system. AKT is composed of four modules: Rasch model-based embeddings, exercise encoder, knowledge encoder, and knowledge retriever. Specifically, the embedding module employs the classic Rasch model in psychometrics to construct embeddings for exercises and KCs:\n\\vspace{-0.2cm}\n\\begin{equation}\n\t\\begin{aligned}\n\t\t\\bm{x}_t = \\bm{c}_{c_t} + \\mu_{e_t} \\cdot \\bm{d}_{c_t}, \\\\\n\t\\end{aligned}\n\t\\vspace{-0.2cm}\n\\end{equation}\nwhere $\\bm{c}_{c_t} \\in \\mathbb{R}^{\\bm{D}}$ is the embedding of the KC of this exercise, $\\bm{d}_{c_t} \\in \\mathbb{R}^{\\bm{D}}$ is a vector that summarizes the variation in exercises with the related KC, and $\\mu_{e_t} \\in \\mathbb{R}^{\\bm{D}}$ is a scalar difficulty parameter that controls the extent to which this exercise deviates from the related KC.\nThe exercise-answer tuple $(e_t, a_t)$ is similarly extended using the scalar difficulty parameter for each pair:\n\\vspace{-0.2cm}\n\\begin{equation}\n\t\\begin{aligned}\n\t\t\\bm{y}_t = \\bm{q}_{(c_t, a_t)} + \\mu_{e_t} \\cdot \\bm{f}_{(c_t, a_t)}, \\\\\n\t\\end{aligned}\n\t\\vspace{-0.2cm}\n\\end{equation}\nwhere $\\bm{q}_{(c_t, a_t)} \\in \\mathbb{R}^{\\bm{D}}$ is the KC-answer embedding, $\\bm{f}_{(c_t, a_t)} \\in \\mathbb{R}^{\\bm{D}}$ is the variation vector. Through the above embedding, exercises labeled as the same KCs are determined to be closely related while retaining important individual characteristics.\nThen, in the exercise encoder, the input is the exercise embeddings $ \\{\\bm{e}_1, . . . , \\bm{e}_t \\} $ and the output is a sequence of context-aware exercise embeddings $\\{\\widetilde{\\bm{e}}_1, . . ., \\widetilde{\\bm{e}}_t\\}$. AKT designs a monotonic attention mechanism to accomplish the above process, where the context-aware embedding of each exercise depends on both itself and the previous exercises, i.e., $\\widetilde{\\bm{e}}_t = f_{enc_1}(\\bm{e}_1, . . . , \\bm{e}_t)$. Similarly, the knowledge encoder takes exercise-answer embeddings $ \\{\\bm{y}_1, . . . , \\bm{y}_t \\} $ as input and\noutputs a sequence of context-aware embeddings of the knowledge acquisitions $\\{\\widetilde{\\bm{y}}_1, . . ., \\widetilde{\\bm{y}}_t\\}$ using the same monotonic attention mechanism; these are also determined by students' answers to both the current exercise and prior exercises, i.e., $\\widetilde{\\bm{y}}_t = f_{enc_1}(\\bm{y}_1, . . . , \\bm{y}_t)$.\nFinally, the knowledge retriever takes the context-aware exercise embedding $\\widetilde{\\bm{e}}_{1:t}$ and exercise-answer pair embeddings $\\widetilde{\\bm{y}}_{1:t}$ as input and outputs a retrieved knowledge state $\\bm{h}_t$ for the current exercise. Since the student’s current knowledge state depends on answering the related exercise, it is also context-aware in AKT.\nThe novel monotonic attention mechanism proposed in AKT is based on the assumption that the learning process is temporal and students' knowledge will decay over time. Therefore, the scaled inner-product attention mechanism utilized in the original Transformer is not suitable for the KT task. AKT uses exponential decay and a context-aware relative distance measure to compute the attention weights. Finally, AKT achieves outstanding performance in predicting students' future answers, as well as demonstrating interpretability due to the combination of the psychometric model.\nIt is important to note that have recently proposed that attentive knowledge tracing models significantly benefit from students' continuous, repeated interactions on the same exercises throughout the learning process. In their experiments, the removal of these repeated interactions in the dataset led to a decline in AKT's performance, bringing it close to that of DKVMN. \n{Moreover, according to the findings of ,} existing attentive attentive KT models primarily trace patterns of a learner's learning activities, rather than their evolving knowledge states. Consequently, they develop the DTransformer model to facilitate stable knowledge state estimation and tracing, rather than solely focusing on next performance prediction. \n\\begin{figure}[t]\n\t\\centerline{\\includegraphics[width=0.9\\columnwidth]{Figures/gkt.pdf}}\n\t\\vspace{-0.3cm}\n\t\\caption{The architecture of graph-based knowledge tracing .}\n\t\\label{fgkt}\n\t\\vspace{-0.3cm}\n\\end{figure}","id":"b60bbb0c-acf3-4bdd-ace6-fb43306d7dda","level":"subsubsection","origin_cites_number":12,"parent_id":"0b0df85b-c041-47ba-8171-c36927847515","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Deep Learning Models"],["subsubsection","Attentive Knowledge Tracing"]],"subsections":[],"title":"Attentive Knowledge Tracing"},{"cite_extract_rate":0.28571428571428503,"cites":[243,553],"content":"\\label{graph}\nGraph neural networks (GNNs), which are designed to handle complex graph-related data, have developed rapidly in recent years . The graph represents a kind of data structure that models a set of objects (nodes) and their relationships (edges). From a data structure perspective, there is a naturally existing graph structure within the KCs. Therefore, incorporating the graph structure of the KCs as additional information should be beneficial to the KT task.\n presented graph-based knowledge tracing (GKT), which conceptualizes the potential graph structure of the KCs as a graph $G = (V, E)$, where nodes $V = \\{v_1, v_2, ..., v_N\\}$ represent the set of KCs and the edges $E \\subseteq V \\times V $ represent relationships of these KCs; moreover, $ \\bm{h}^t = \\{\\bm{h}^t_{i \\in V}\\} $ represents the student's temporal knowledge state after answering the exercise at time $t$.\nThe architecture for graph-based knowledge tracing is presented in Figure \\ref{fgkt}, which is composed of three parts: (1) $aggregate$, (2) $update$ and (3) $predict$.\nIn the $aggregate$ module, GKT aggregates the temporal knowledge state and the embedding for the answered KC $i$ and its neighboring KC $j$:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\bm{h}_k^{'t}=\\left\\{\n\t\\begin{aligned}\n\t\t&[\\bm{h}_k^{t}, a^t\\bm{E}_s] &(k = i), \\\\\n\t\t&[\\bm{h}_k^{t}, \\bm{E}_e(k)] &(k \\neq i),\n\t\\end{aligned}\n\t\\right.\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $a^t$ represents the exercises answered correctly or incorrectly at time step $t$, $\\bm{E}_s$ is the embedding matrix for the learning interactions, $\\bm{E}_e$ is the embedding matrix for the KC, and $k$ represents the $k$-th row of $\\bm{E}_e$.\nIn the $update$ module, GKT updates the temporal knowledge state based on the aggregated features and the knowledge graph structure, as follows:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\begin{aligned}\n\t\t&\\bm{m}_k^{t+1}=\n\t\t\\begin{cases}\n\t\t\t&f_{self}(\\bm{h}_k^{'t}) (k = i), \\\\\n\t\t\t&f_{neighbor}(\\bm{h}_i^{'t}, \\bm{h}_k^{'t}) (k \\neq i),\\\\\n\t\t\\end{cases}\\\\\n\t\t&\\widetilde{\\bm{m}}_k^{t+1} = G_{ea}(\\bm{m}_k^{t+1}), \\\\\n\t\t&\\bm{h}_{k}^{t+1} = G_{gru}(\\widetilde{\\bm{m}}_k^{t+1}, \\bm{h}_{k}^{t}), \\\\\n\t\\end{aligned}\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $f_{self}$ is the multilayer perceptron, $G_{ea}$ is the same $erase$-followed-by-$add$ mechanism used in DKVMN, and $G_{gru}$ is the gated recurrent unit (GRU) gate . Moreover, $f_{neighbor}$ defines the information propagation to neighboring nodes based on the knowledge graph structure.\nIn the $predict$ module, GKT predicts the student's performance at the next time step according to the updated temporal knowledge state:\n\\vspace{-0.1cm}\n\\begin{equation}\n\ty_k^t = \\sigma(\\bm{W}_k\\bm{h}_k^{t+1} + \\bm{b}_k),\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $\\bm{W}_k$ is the weight parameter and $\\bm{b}_k$ is the bias term.\nIn addition to modeling the graph structure in KCs by graph neural networks, propose to model the educational relation and topology in the concept map, which will be intended to act as mathematical constraints for the construction of the KT model.\nRecently, in the attempt to further explore knowledge structure, propose structure-based knowledge tracing (SKT), which aims to capture the multiple relations in knowledge structure to model the influence propagation among concepts. SKT is mainly motivated by an education theory, \\emph{transfer of knowledge} , which claims that students' knowledge states on some relevant KCs will also be changed when they are practicing on a specific KC due to the potential knowledge structure among KCs. Therefore, a student's knowledge state is determined by not only the temporal effect from the exercise sequence, but also the spatial effect from the knowledge structure. To concurrently model the latent spatial effects, SKT presents the synchronization and partial propagation methods to characterize the undirected and directed relations between KCs, respectively. In this way, SKT measures influence propagation in the knowledge structure with both temporal and spatial relations. To get rid of dependence on knowledge structure, propose the Automatical Graph-based Knowledge Tracing (AGKT), which utilizes the automatical\ngraph to measure students’ knowledge states automatically without annotation manual annotations.","id":"9d414c8f-554a-43a7-afe7-209c3c4d813e","level":"subsubsection","origin_cites_number":7,"parent_id":"0b0df85b-c041-47ba-8171-c36927847515","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Deep Learning Models"],["subsubsection","Graph-based Knowledge Tracing"]],"subsections":[],"title":"Graph-based Knowledge Tracing"},{"cite_extract_rate":0.2,"cites":[5381],"content":"{It is crucial to emphasize that, despite deep learning models exhibiting superior performance compared to Bayesian models and logistic models, there remains significant room for improvement in their interpretability and explainability. \nDue to the end-to-end learning strategy, deep learning models are notoriously difficult to interpret. The modeling process itself is also challenging to explain. Specifically, deep learning models are predominantly data-driven and benefit a lot from large-scale student learning data. It is challenging to understand how they calculate a student's knowledge state with no theoretical guidance . \nAll we have are the results generated by these models. Therefore, any errors made by these models will lead students to doubt their reliability. The lack of explainability and interpretability has thus limited their further applicability. }\n{To make the complex KT models interpretable, especially those deep learning models, researchers have attempted various methods.\n presented a post-hoc approach to reveal the interpretability of DKT. Specifically, they employed the layer-wise relevance propagation (LRP) technique to interpret DKT by measuring the relevance between DKT's output and input. Preliminary experimental results suggest that this post-hoc approach could be a promising method for explaining DKT. Besides, explainable AI (xAI) is proposed to make the black-box deep learning models more transparent, thereby promoting its applications . proposed to use the xAI technique to interpret the complex KT models based on deep learning. The interpreting results of DKT have been demonstrated to aid in enhancing the trust of students and teachers. Their findings suggested that it is promising to utilize xAI techniques to interpret the deep learning KT models, thereby assisting users in accepting and applying the suggestions provided by these models. \n}","id":"2737aa4f-b2f6-4977-83f3-36d33949ab77","level":"subsection","origin_cites_number":5,"parent_id":"bd30f2af-cabd-4e04-ab90-5536a2f60dde","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Fundamental Knowledge Tracing Models"],["subsection","Summarization"]],"subsections":[],"title":"Summarization"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:variants}\n{The aforementioned fundamental KT models are} typically based on simplistic assumptions regarding the learning process. Specifically, these models predominantly employ learning interactions such as exercises and responses to estimate students' knowledge states. However, the learning process is not solely represented by exercises and responses, but is influenced by various factors. In summary, the aforementioned fundamental KT models, while straightforward, may have reduced performance in real-world learning scenarios. Consequently, numerous variants have been proposed under more stringent assumptions, reflecting a more comprehensive learning process in real-world scenarios. Accordingly, we classify and review current variants of fundamental KT models into four categories: (1) Modeling individualization before learning, (2) Incorporating engagement \\ during learning, (3) Considering forgetting after learning, and (4) utilizing side information across learning.","id":"a705e159-25ef-44ad-8166-1cef83a7c265","level":"section","origin_cites_number":0,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Variants of Knowledge Tracing Models"]],"subsections":["242f5268-6395-4e4d-b14a-e118bc61d29b","b2eec265-034c-4c04-b806-03ebaf934ee5","2cb054ef-697a-4d85-a2b7-9c0e6282362f","60b4234d-a28f-4026-ab3b-0291b62ceca6"],"title":"Variants of Knowledge Tracing Models"},{"cite_extract_rate":0.14285714285714202,"cites":[5382,166],"content":"Everything and everyone possess unique characteristics. For instance, explored several personalized factors of various tourists to recommend personalized travel packages. Similarly, the concept of individualization in the KT task implies that different students often exhibit different learning characteristics (such as varying learning rates or prior knowledge). Considering the student-specific variability in learning could potentially enhance the KT process, as suggested by . In the subsequent sections, we will introduce various variant KT models that {take into account individualization before learning.}\nThe initial BKT paper has delved into the concept of individualization. Specifically, it uses all students' learning interactions on a specific KC to learn the individual parameter. Similarly, for a specific student, all her learning interactions are utilized to fit her individual learning parameters . Consequently, BKT is able to ascertain different learning and performance parameters for various students and KCs. However, this approach only offers a marginal improvement compared to original BKT. \nSubsequently, propose two simple variants of BKT that respectively individualize students' initial probability of mastery and the probability of transition from the unlearned state to the learned state. {Specifically, a student node is added to individualize the initial probability of mastery for each student. The student node assigns each student with a personalized initial probability of mastery. A conditional probability table is designed to determine the value of the student node. Similarly, if changing the connection of the student node to the subsequent knowledge nodes, the transition probability parameter can also be individualized. In this case, the student node gives individualized transition parameters to each student.}\nMoreover, rather than individualizing only one kind of parameter in BKT, some other variants of BKT opt to individualize all four BKT parameters simultaneously . suggest that when applied in an intelligent tutoring system, the individualized BKT model can yield good improvements to student learning efficacy, reducing by about half the amount of questions required for 20\\% of students to achieve mastery.\nAnother means of modeling individualization is clustering, which considers a wider range of students in different groups . \nBy clustering the students into various groups, we can train different KT models and make predictions on the test data. The number of clusters is then varied according to the student groups and the predicting process is repeated iteratively. Finally, we can obtain a set of different predictions. Furthermore, there are two common methods used to combine these predictions : (1) uniform averaging, which simply averages the predictions; (2) weighted averaging, which combines the models by means of a weighted average.\nTo realize clustering, K-means is a basic clustering algorithm that randomly initializes a set of cluster centroids, which are identified using Euclidean distance. Another popular clustering algorithm is spectral clustering, which represents the data as an undirected graph and analyzes the spectrum of the graph Laplacian obtained from the pairwise similarities of data points. Recently, some novel clustering algorithms have been proposed, including discrete nonnegative spectral clustering and clustering uncertain data .\n propose a model named deep knowledge tracing with dynamic student classification (DKT-DSC), which introduces individualization to DKT by exploiting the idea of clustering. \nAccording to students' previous performance, DKT-DSC assigns students with similar learning ability to the same group . The knowledge states of students in different groups are then traced by different DKT models. Moreover, considering the dynamic property of the learning ability, each student's learning sequence is segmented into multiple time intervals. At the start of each time interval, DKT-DSC will reassess students' learning ability and reassign their groups. \nIn DKT-DSC, the K-means clustering algorithm is utilized to split students with similar ability levels into the same group at each time interval. After learning the centroids of all K clusters, each student is assigned to the nearest cluster. Through dynamic student clustering, DKT-DSC offers an effective approach to realizing individualization in DKT.\n claim that it is significant to consider both individual exercise representation and individual prior knowledge. They propose a fine-grained knowledge tracing model, named FGKT. FGKT obtains the individual exercise representation through the acquisition of knowledge cells (KCs) and exercise distinctions. Subsequently, it assesses the individual prior knowledge by evaluating the relevance between current and historical learning interactions. Finally, the above individual representations will be utilized as the input of LSTM in FGKT to evaluate students' evolving knowledge states. also notice that the individualization of exercises is significant for measuring students' knowledge states. They propose to consider multiple exercise factors, including the difficulty\nand the discrimination, to enhance the performance of DKT.\n propose a convolutional knowledge tracing model (CKT) to implicitly measure student individualization. Specifically, CKT considers two factors that influence students' individualization: individualized learning rates and individualized prior knowledge.\nIndividualized learning rates represent students' differing capacities to absorb knowledge. The sequence of student learning interactions can reflect different learning rates in the sense that students with high learning rates can rapidly master knowledge, while others need to spend more time trying and failing. Therefore, it is reasonable to assess the differences in learning rate by simultaneously processing several continuous learning interactions within a sliding window of convolutional neural networks . Besides, individualized prior knowledge refers to students' prior knowledge, which can be assessed via their historical learning interactions.","id":"242f5268-6395-4e4d-b14a-e118bc61d29b","level":"subsection","origin_cites_number":14,"parent_id":"a705e159-25ef-44ad-8166-1cef83a7c265","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Variants of Knowledge Tracing Models"],["subsection","Modeling Individualization before Learning"]],"subsections":[],"title":"Modeling Individualization before Learning"},{"cite_extract_rate":0,"cites":[],"content":"Student engagement is defined as \\emph{\"the quality of effort students themselves devote to educationally purposeful activities that contribute directly to desired outcomes\"} . This definition highlights a strong connection {between student engagement and the learning process.} Generally, higher engagement leads to enhanced knowledge gains. Consequently, considering student engagement in the learning process could potentially improve knowledge tracing results . In this section, {we will present some variants that integrate student engagement into the KT models. }\nStudent engagement is difficult to be directly measured. In practice, some online learning systems have made use of sensor data to measure student engagement. For example, inexpensive portable electroencephalography (EEG) devices can help to detect a variety of student mental states in learning, which can be seen as reflections of student engagement . propose two methods that combine EEG-measured mental states to improve the performance of BKT. Concretely, the first one inserts a one-dimensional binary EEG measure into BKT, forming the EEG-BKT structure that extends BKT by adding a binary variable node $E$ between the knowledge node and the answer node. The second one, i.e., EEG-LRKT, utilizes logistic regression to combine an $m$-dimensional continuous variable $E$ extracted from the raw EEG signal in BKT.\nHowever, in most cases, it is difficult to collect sensor data on every student. Therefore, propose the knowledge and affect tracing (KAT) to model both knowledge and engagement in parallel. KAT is a sensorless model that does not rely on any sensor data. In this model, both knowledge and engagement are assumed to have direct influences on student performance. KAT considers three kinds of disengagement behaviors: quick guess (the student makes an attempt very quickly), bottom-out hint (all available hints are used) and many attempts (making more than three attempts at an exercise). These three behaviors are grouped as “gaming” behaviors in order to predict students' knowledge and engagement at each learning interaction. \nRather than assuming equal influence of knowledge and engagement on students' knowledge state, one variation on the KAT model defines the connection between knowledge and engagement, and accordingly considers that students' knowledge states will influence their engagement. For example, students are more likely to disengage from knowledge they are not familiar with.\nMoreover, rather than explicitly modeling student engagement, further propose the knowledge tracing with behavior (KTB) model, which has only one latent knowledge node that acts as a combination of both knowledge and engagement. KTB assumes that both engagement and performance are expressions of knowledge. The Bayesian estimation of the knowledge state needs to be inferred by both student engagement and performance.\n propose to add five features in the process of watching videos on MOOCs to the input of DKT. These features reflect student engagement from various aspects, including playback speed, whether or not the video was paused, fast-forwarded or rewound, and whether or not the video was completed. For example, if a student watches a video at a much faster playback speed, it is likely that he/she is impatient and absent-minded. This model incorporates two further features: whether or not the exercise was submitted with an answer selected and whether or not the exercise was a part of an end-of-unit quiz, both of which are considered together. Experimental results indicate that DKT can achieve better performance through incorporating the above binarized engagement covariates.","id":"b2eec265-034c-4c04-b806-03ebaf934ee5","level":"subsection","origin_cites_number":7,"parent_id":"a705e159-25ef-44ad-8166-1cef83a7c265","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Variants of Knowledge Tracing Models"],["subsection","Incorporating Engagement during Learning"]],"subsections":[],"title":"Incorporating Engagement during Learning"},{"cite_extract_rate":0.08333333333333301,"cites":[5378],"content":"In real-world scenarios, while learning, forgetting is inevitable . The \\emph{Ebbinghaus forgetting curve theory} indicates that students' knowledge proficiency will decline due to forgetting . Recently, proposed the concept of 'Knowledge Proficiency Tracing' (KPT), a model that can dynamically capture the changes in students' proficiency levels on knowledge concepts over time. This model effectively tracks these changes in an interpretable manner. Therefore, the assumption that students' knowledge states will remain constant over time is untenable. However, fundamental KT models, such as the BKT, often overlook forgetting. In the following, we will introduce some variants of fundamental KT models that have attempted to consider forgetting after learning for more precise knowledge states. \n discover that BKT consistently overestimates the accuracy of students' answers when a day or more had elapsed since {their previous responses.} The underlying reason is that BKT assumes that student performance will remain the same regardless of how much time has passed. To consider how student performance declines with time, they propose a BKT-Forget model, which hypothesizes that students may forget information they have learned as days go by. In the BKT-Forget model, a time node is added to specify which parameters should be affected by a new day and the new day node is fixed with a prior probability of 0.2. {It also introduced parameters to represent the forgetting rate on a new day and denote the forgetting rate on the same day.} However, although BKT-forget does consider the decline in student performance, it can only model forgetting that occurs over the time scale of days. \nTo model the continuous decay of knowledge as time progresses, incorporate forgetting into BKT based on the assumption that learned knowledge decays exponentially over time . An exponential decay function is thus utilized to update the knowledge mastery level. They further assumed that the chance of forgetting will increase if a student does not practice the knowledge concepts within 30 days.\nMoreover, introduce an approach that counts the number of intervening trials and treats each as an independent opportunity for forgetting to occur.\nRecall the PFA model in Eq.(\\ref{pfa}), in which the probability of students' mastery is estimated using a logistic function: $ p(\\theta) = \\sigma(\\beta + \\mu s + \\nu f)$. The original PFA model ignores the order of answers, in addition to the time between learning interactions. It is therefore difficult to directly incorporate time information into the original PFA model.\n propose PFAE (PFA Elo/Extended), a variant of the PFA model that combines PFA with some aspects of the Elo rating system . The Elo rating system is originally devised for chess rating (estimating players' skills based on match results). In PFAE, $\\theta$ is updated after each learning interaction:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\theta :=\\left\\{\n\t\\begin{aligned}\n\t\t& \\theta + \\mu \\cdot (1 - p(\\theta))& \\mbox{if the answer was correct}, \\\\\n\t\t& \\theta + \\nu \\cdot p(\\theta)& \\mbox{if the answer was wrong}.\n\t\\end{aligned}\n\t\\right.\n\t\\vspace{-0.1cm}\n\\end{equation}\nAs the forgetting behavior of students is closely related to time, in order to consider forgetting, add a time effect function $f$ to $\\theta$, i.e., using $p(\\theta+f(t))$ instead of $p(\\theta)$, where $t$ is the time (in seconds) from the last learning interaction, and $f$ is the time effect function.\nTo represent the complex forgetting behavior, the DKT-forget model introduces forgetting into DKT, which considers three types of side information related to forgetting: (1) the repeated time gap that represents the interval time between the present interaction and the previous interaction with the same KC, (2) the sequence time gap that represents the interval time between the present interaction and the previous interaction, and (3) past trial counts that represent the number of times a student has attempted on the exercise with the same KC. All these three features are discretized at $log_2$ scale.\nThose side information is concatenated as additional information and represented as a multi-hot vector $\\bm{c}_t$, which is integrated with the embedding vector $\\bm{v}_t$ of the learning interaction, as follows:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\bm{v}_t^c = \\theta^{in}(\\bm{v}_t, \\bm{c}_t),\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $\\theta^{in}$ is the input integration function. The integrated input $\\bm{v}_t^c$ and the previous knowledge state $\\bm{h}_{t-1}$ are passed through the RNNs to update $\\bm{h}_t$ in the same way as in Eq.(\\ref{dkt}). The additional information at the next time step $\\bm{c}_{t+1}$ is also integrated with the updated $\\bm{h}_t$:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\bm{h}_t^c = \\theta^{out}(\\bm{h}_t, \\bm{c}_{t+1}),\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $\\theta^{out}$ is the output integration function.\n propose a novel HawkesKT model, which introduces the Hawkes process to adaptively model temporal cross-effects. The Hawkes process performs well at modeling sequential events localized in time, as it controls corresponding temporal trends by the intensity function. The intensity function in HawkesKT is designed to characterize the accumulative effects of previous learning interactions, along with their evolutions over time. In HawkesKT, the temporal cross-effects and the ways in which they evolve between historical learning interactions combine to form a dynamic learning process.","id":"2cb054ef-697a-4d85-a2b7-9c0e6282362f","level":"subsection","origin_cites_number":12,"parent_id":"a705e159-25ef-44ad-8166-1cef83a7c265","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Variants of Knowledge Tracing Models"],["subsection","Considering Forgetting after Learning"]],"subsections":[],"title":"Considering Forgetting after Learning"},{"cite_extract_rate":0.157894736842105,"cites":[1684,5383,5384],"content":"\\label{sec:side information}\nMost KT models primarily {rely on exercises and student responses} to evaluate students' knowledge states. These models have yielded impressive results and have been effectively implemented in online learning systems. Despite this, there are various other types of side information collected across the learning process that could be utilized to enhance {these models.} In this section, we will introduce several variants that aim to leverage this diverse side information across learning. \nIn terms of a student's first response time, a short initial response time could indicate either high proficiency or 'gaming' behavior, while a long initial response time could indicate either careful thinking or lack of concentration. Since the connection between initial response time and knowledge state could be influenced by {complex factors}, propose to discretize the continuous first response time into four categories (i.e., extremely short, short, long, extremely long) {to eliminate unnecessary information and simplify the latent complex possibilities. They then build a one-by-four parameter table, in which each column represents the category of the initial response time of the previous exercise, while the relevant values represent the probability of correct answers}. \nRegarding tutor intervention, propose the Bayesian evaluation and assessment model, which simultaneously assesses students' knowledge states and evaluates the lasting impact of tutor intervention. More specifically, it adds one observable binary intervention node to BKT: \\emph{True} means that the tutor intervention occurs in corresponding interactions while \\emph{False} indicates the opposite. The connection between the intervention node and knowledge node indicates the potential impact of the tutor intervention on students' knowledge states. The intervention node is linked to all four BKT parameters. As a result, there are a total of eight parameters to learn in order to incorporate tutor intervention. One possible way to reduce the number of parameters is choosing to link only the intervention node to the learning rate parameter .\nSimilarly, develop the intervention-Bayesian knowledge tracing (Intervention-BKT) model, {which incorporates two types of interventions into BKT and distinguishes their different effects: \\emph{elicit and tell}}. The relations between the intervention and performance nodes represent the impact of teaching interventions on student performance, while the relations between the intervention and knowledge nodes represent the impact of teaching interventions on students' knowledge states. Therefore, at each learning interaction, while the present knowledge state is conditional on both the previous knowledge state and the current intervention, the student's performance depends on both the present knowledge state and the current intervention.\nRather than considering only one kind of side information, propose a feature-aware student knowledge tracing (FAST) model, which allows for the utilization of all kinds of side information. Traditional BKT uses conditional probability tables for the guessing, slipping, transition and learning probabilities, meaning that the number of features involved in inference grows exponentially. Therefore, as the number of features increases, the time and space complexity of the model also grow exponentially. To deal with this large number of features, FAST uses logistic regression parameters rather than conditional probability tables. The number of features and complexity increase linearly rather than exponentially. \nFor parameter learning, FAST uses the Expectation Maximization with Features algorithm and focuses on only emission features. The E step uses the current parameter estimates $\\lambda$ to infer the probability of the student having mastered the KC at each learning interaction.\nThe parameters $\\lambda$ are now a function of the weight $\\beta$ and the feature vector $\\bm{f}(t)$. $\\bm{f}$ is the feature extraction function, and $\\bm{f}(t)$ is the feature vector constructed from the observations at the relevant time step. The emission probability is represented with a logistic function:\n\\vspace{-0.1cm}\n\\begin{equation}\n\t\\begin{aligned}\n\t\t\\lambda(\\beta)^{y^{'},k^{'}} & = \\frac{1}{1 + exp(-\\beta^T \\cdot \\bm{f}(t))},\n\t\\end{aligned}\n\t\\vspace{-0.1cm}\n\\end{equation}\nwhere $\\beta$ is learned by training a weighted regularized logistic regression using a gradient-based search algorithm.\n propose an extension to DKT that explored the inclusion of additional features. Specifically, it incorporates an auto-encoder network layer to convert the higher-dimensional input data into smaller representative feature vectors, thereby reducing both the resource and time requirement for training. Students' response time, opportunity count, and first action are selected as incorporated side information and all input features are converted into a fixed-length input vector. First, all input features are converted into categorical data and represented as a sparse vector by means of one-hot encodings. These encoded features are concatenated together to construct the higher-dimensional input vector:\n\\vspace{-0.2cm}\n\\begin{equation}\n\t\\begin{aligned}\n\t\tC(e_t,a_t) &= e_t + (max(e) +1)a_t, \\\\\n\t\tv_t &= O(C(e_t,a_t)) \\oplus O(C(t_t,a_t)) \\oplus O(t_t), \\\\\n\t\tv_t^{'} &= tanh(W_vv_t + b_v),\n\t\\end{aligned}\n\t\\vspace{-0.2cm}\n\\end{equation}\nwhere $C$ is the cross feature, $O$ is the one-hot encoder format, $v_t$ represents the resulting input vector of each learning interaction, $e_t$ is the exercise, $a_t$ refers to the answer, $t_t$ is the response time, $W_v$ is the weight parameter and $b_v$ is the bias term. Subsequently, an auto-encoder is introduced to reduce the dimensionality without incurring the loss of too much important information. Finally, the feature vectors extracted by auto-encoder will be the new input of DKT.\n presented another extension to DKT, namely the Exercise-aware Knowledge Tracing (EKT), which utilized the potential value of exercises' text contents. Generally, the text content is of great significance for students to understand the exercises. For example, used text materials to automatically predict their difficulties, utilized the text content to find similar exercises. further proposed a pre-training model called QuesNet for learning the unified representations of heterogeneous exercises. \nTherefore, instead of using one-hot encoding of exercises, EKT automatically learns the semantic representation of each exercise from its text contents. EKT first uses $Word2vec$ to pre-train the embedding vector for each word in the exercise. It then constructs a bidirectional LSTM, which captures the word sequence from both forward and backward directions to learn the semantic word representation. The element-wise max-pooling operation is utilized to merge words’ contextual representations into a global embedding. Finally, EKT can update the student's knowledge state with the aid of the semantic representation of each exercise. \nTo achieve more feasible integration of side information, present a deep knowledge tracing method with decision trees (DKT-DT), which takes advantage of Classification And Regression Trees (CART) to preprocess the heterogeneous input features . Specifically, CART is utilized to automatically partition the feature space and outputs whether or not a student can answer an exercise correctly. {The predicted response and the true response are encoded into a four-bit binary code; for example, the code is $1010$ if the predicted response and the true response are both correct. This binary code is then concatenated with the original one-hot encoding of the exercise as the new input of DKT to train the corresponding model.}\n suggests that the student's language proficiency can serve as supplementary information to improve existing KT models. The student’s language proficiency is extracted by Elo rating score and time window features. Then, the language proficiency information is demonstrated to be effective in promoting several KT models, including DKT, DKVMN, and SAKT. Additionally, the problem of cold start in the KT task is alleviated with the assistance of language proficiency information. \n explore to add side information to the original KT model by auxiliary learning tasks. They specifically introduced two tasks: (1) predicting the KCs of the question, and (2) predicting the individualized prior knowledge. By training with these tasks, KT can enhance its understanding of the intrinsic relationships between questions and KCs, while explicitly capturing student-level variability. \n{When solving programming problems, we can record students' full code submissions, which can be employed to analyze their programming ability. collected student programming data and analyzed their personalized programming preferences. The study commenced with a statistical analysis of student errors, followed by an examination of students' programming structures derived from their code submissions, and finally used BKT to measure students' programming abilities.\n transformed students' code submissions into embedded vectors, and applied them in DKT to model students' fine-graded programming knowledge states.\n noticed that a single programming problem generally involves in multiple KCs, thereby they proposed to learn useful information about the programming problem's multiple requirements from students' code submissions. }","id":"60b4234d-a28f-4026-ab3b-0291b62ceca6","level":"subsection","origin_cites_number":19,"parent_id":"a705e159-25ef-44ad-8166-1cef83a7c265","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Variants of Knowledge Tracing Models"],["subsection","Utilizing Side Information across Learning"]],"subsections":[],"title":"Utilizing Side Information across Learning"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:applications}\nAlthough knowledge tracing is an emerging research area, it has already been applied in a wide variety of scenarios. {In the following, we will first survey the applications of KT models in two typical educational scenarios: learning resources recommendation and adaptive learning. Then, we will discuss broader applications of KT beyond student learning.}","id":"5fc15152-8ae0-4a7f-83a2-d64f287123fe","level":"section","origin_cites_number":0,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Applications"]],"subsections":["260a9040-27f0-4099-a17f-ea39955bc0eb","f6f4178c-41e5-4941-b347-f40654197f09","c5ab0210-1d4b-4488-91f4-cb2e837345b7"],"title":"Applications"},{"cite_extract_rate":0.2,"cites":[1206],"content":"Traditionally, learning resources for each student are selected in one of two ways. The first one requires teachers to manually select suitable resources that match students' knowledge levels. However, this approach requires substantial time and effort, and different teachers may have different preferences. The second one allows students themselves to freely choose resources to learn. However, this may result in students choosing too easy or too difficult materials that will not benefit their learning , leading to low learning efficiency. In recent years, the prevalence of intelligent tutoring systems and the development of KT methods have made it possible to automatically recommend appropriate exercises to each student based on artificially designed algorithms.\nExercises are the most common learning resources in learning. Given the inferred knowledge states, one common strategy is selecting the next exercise that will best advance students' knowledge acquisition. propose two extensions of the original BKT model, which respectively considered exercises' difficulties and students' multiple-attempt behaviors. These two extensions are integrated into a BKT-sequence algorithm to recommend exercises to students based on their knowledge states. Specifically, BKT-sequence first determines the predicted range of scores for each exercise. It then computes an expected score for each exercise that the student should get to achieve mastery, which is dependent on their current knowledge state (for instance, a lower knowledge state will result in higher expected scores). Finally, the algorithm returns the exercise with a predicted score that is closest to that of the expected score. Therefore, as the knowledge state of a particular KC grows, more difficult exercises will be recommended, as harder exercises are associated with a lower predictive score. Experimental results have shown that students using the BKT-sequence algorithm were able to solve more difficult exercises, obtained higher performance and spent more time in the system than students who used the traditional approach. Moreover, students also expressed that the BKT-sequence algorithm was more efficient. expanded the DKVMN model to include the exercise's type and difficulty. This model is then used to assess students' knowledge state and subsequently recommends personalized exercises for each student in Ssmall Private Online Courses (SPOCs). They conducted a randomized controlled trial to show the proposed personalized exercise recommendation could enhance students' learning efficiency.\nIn addition to exercises, there are also some other types of multi-modal learning resources, such as videos and figures. utilizes an adaptation of BKT to improve student performance prediction by incorporating video observation. Experimental verification demonstrates the impact of both using and eschewing video data, as well as the learning rate associated with a particular video. In this way, they further developed a method to help people evaluate the quality of video resources. Concretely, they proposed the Template 1 Video model to incorporate video observations into BKT, which adds video activity as additional independent observation nodes to the BKT model. This model accordingly considers the probability that a given video resource will impart knowledge to a student. Moreover, the transition probability in BKT is conditional only on the presence of either a video or an exercise. Thus, the quality of the video can be determined by its promotion of learning, and this model can be leveraged as a tool to aid in evaluating and recommending video resources.\nWhen recommending learning resources, the primary aim of existing solutions is to choose a simple strategy for assigning non-mastered exercises to students. \nWhile reasonable, it is also too broad to advance learning effectively. accordingly propose three more beneficial and specific objectives: \\emph{review and explore}, \\emph{smoothness of difficulty level} and \\emph{student engagement}. In more detail, \\emph{review and explore} considers both enhancing students' non-mastered concepts with timely reviews and reserving certain opportunities to explore new knowledge; \\emph{smoothness of difficulty level} indicates that the difficulty levels of several continuous exercises should vary within a small range as students gradually learn new knowledge; finally, \\emph{student engagement} considers that to promote students' enthusiasm during learning, the recommended exercises should be in line with their preferences. In order to support online intelligent education with the above three domain-specific objectives,\nthey developed a more reasonable multi-objective deep reinforcement learning (DRE) framework. DRE presented three corresponding novel reward functions to capture and quantify the effects of the above three objectives. This DRE framework is a unified platform designed to optimize multiple learning objectives, where more reasonable objectives also can be incorporated if necessary.\nExperimental results show that DRE can effectively learn from the students' learning records to optimize multiple objectives and adaptively recommend suitable exercises.","id":"260a9040-27f0-4099-a17f-ea39955bc0eb","level":"subsection","origin_cites_number":5,"parent_id":"5fc15152-8ae0-4a7f-83a2-d64f287123fe","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Applications"],["subsection","Learning Resources Recommendation"]],"subsections":[],"title":"Learning Resources Recommendation"},{"cite_extract_rate":0.25,"cites":[5373],"content":"{Adaptive learning, unlike learning resource recommendations, goes beyond the mere provision of resources. It not only concentrates on the selection of appropriate learning materials but also designs effective learning strategies and dynamic learning pathways. These are structured based on both the learning rules and students' evolving knowledge states.\nSpecifically, adaptive learning broadly refers to \\emph{\"a learning process in which the content taught, or the way such content is presented, changes or 'adapts' based on individual student responses, and which dynamically adjusts the level or types of instruction based on individual student abilities or preferences\"} . }\nThe first few attempt made to apply KT to adaptive learning was the ACT Programming Tutor (APT) , where students were asked to write short programs and BKT was utilized to estimate their evolving knowledge state. This tutor can present an individualized sequence of exercises to each student based on their estimated knowledge states until the student has \"mastered\" each rule. \nIn recent years, Massive Open Online Courses (MOOCs) have become an emerging modality of learning, particularly in higher education. adapt BKT on the edX platform. The research object was a 14-week online course that included weekly video lectures and corresponding lecture problems. BKT was applied to enhance students' learning on this course. In order to better adapt BKT to the learning platform, the original BKT was modified in several respects. First, due to the lack of labeled KCs, the problems would be directly seen as the KCs, while the questions would be seen as the exercises belonging to the KC.\nSecond, in order to capture the varying degrees of students' knowledge acquisition at each attempt, the modified model assigned different guess and slip parameters to different attempt counts. Third, to deal with the problem of multiple pathways in the system, which reflected that the impacts on learning may come from various resources, they framed the influence of resources on learning as a credit/blame inference problem.\nGenerally, students' cognitive structures include both students' knowledge level and the knowledge structure of learning items (e.g., \\emph{one-digit addition} is the prerequisite knowledge of \\emph{two-digit addition}). Therefore, adaptive learning should maintain consistency with both students' knowledge level and the latent knowledge structure. Nevertheless, existing methods for adaptive learning often focus separately on either the knowledge levels of students (i.e., with the help of specific KT models) or the knowledge structure of learning items. To fully exploit the cognitive structure for adaptive learning, propose a Cognitive Structure Enhanced framework for adaptive Learning (CSEAL). CSEAL conceptualized adaptive learning as a Markov Decision Process. It first utilized DKT to trace the evolving knowledge states of students at each learning step. Subsequently, the authors designed a navigation algorithm based on the knowledge structure to ensure that the learning paths in adaptive learning were logical and reasonable, which also reduced the search space in the decision process. Finally, CSEAL utilized the actor-critic algorithm to dynamically determine what should be learned next. In this way, CSEAL can sequentially identify the most suitable learning resources for different students.","id":"f6f4178c-41e5-4941-b347-f40654197f09","level":"subsection","origin_cites_number":4,"parent_id":"5fc15152-8ae0-4a7f-83a2-d64f287123fe","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Applications"],["subsection","Adaptive Learning"]],"subsections":[],"title":"Adaptive Learning"},{"cite_extract_rate":0.22222222222222202,"cites":[5385,8912],"content":"{The above two types of applications are most commonly used for KT in student learning. In addition, the KT methods can be expanded to be utilized in any systems that necessitate continuous evaluation of user capabilities or states. We will introduce some broader applications of KT in this section.}\nIn gaming systems, the paradigm of tracing students' knowledge state can also work for player modeling. Here, player modeling, which is the study of computational models of players in games, aims to capture human players' characteristics and cognitive features . For instance, reveal that children engage in cycles of increasingly sophisticated mathematical thinking over the course of playing an online game. present an approach to trace player knowledge in a parallel programming educational game, which is capable of measuring the current players' real-time state across the different skills required to play an educational game based only on in-game player activities. conduct a classroom experiment comparing a commercial game for equation solving, i.e., \\emph{DragonBox}, with a research-based intelligent tutoring system, i.e., \\emph{Lynnette}. The results indicated that students who used \\emph{DragonBox} enjoyed the experience more, while students who used \\emph{Lynnette} performed significantly better on the test. Therefore, it is possible to enable students to learn effectively and happily by designing suitable educational games on the online learning platform.\n{In crowdsourcing, unlabeled data or specific tasks are assigned to various crowd annotators. Understanding the dynamic capabilities of these annotators is crucial in ensuring the reliability of their annotations and promoting the annotation efficiency. developed a framework called KT4Crowd, which utilized KT methods to predict the performance of annotators, which surpassed traditional rating systems. \nBesides, observed that students' participation in crowdsourcing tasks can enhance their learning. KT methods can also better comprehend students' knowledge states, aided by the students' annotated items. \n}\n develop an online citizen science project that employs machine learning techniques to improve the training of new volunteers using authentic tasks featuring uncertain outcomes, such as image classification. Specifically, they employ the BKT model to monitor the knowledge states of volunteers, enabling them to more efficiently complete assigned tasks and contribute meaningfully to the project. \n develop an automated exercise collection approach for teachers, employing the KT model and reinforcement learning. Specifically, the exercise collection need to be well-designed to align with students' abilities. This study first leverages the KT model to forecast students' performance on unseen exercise candidates. Subsequently, the exercise selector is designed based on the KT model's predictions, ensuring that the exercise collection is both approximate and optimized. Similarly, design a reinforcement learning guided\nmethod for exam paper generation, where the DKT model is utilized to measure examinees' knowledge states. \n\\iffalse\n\\begin{table*}[t]\n\t\\vspace{-0.2cm}\n\t\\renewcommand\\arraystretch{1.0}\n\t\\caption{A Summary of Public Datasets for Knowledge Tracing. }\n\t\\vspace{-0.3cm}\n\t\\centering\n\t\\begin{tabular}{l|l}\n\t\t\\hline\n\t\tDatasets & Link\\\\\n\t\t\\hline\n\t\tASSISTments2009 & https://sites.google.com/site/assistmentsdata/home/assistment-2009-2010-data\\\\\n\t\tASSISTments2012 & https://sites.google.com/site/assistmentsdata/home/2012-13-school-data-with-affect\\\\\n\t\tASSISTments2015 & https://sites.google.com/site/assistmentsdata/home/2015-assistments-skill-builder-data \\\\\n\t\tASSISTments2017 & https://sites.google.com/view/assistmentsdatamining/dataset \\\\\n\t\tKDD-Cup 2010 & https://pslcdatashop.web.cmu.edu/KDDCup/rules\\_data\\_format.jsp \\\\\n\t\tStatics 2011 & https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=507 \\\\\n\t\tslepemapy.cz & https://www.fi.muni.cz/adaptivelearning/?a=data\\\\\n\t\tsynthetic & https://github.com/chrispiech/DeepKnowledgeTracing/tree/master/data/synthetic\\\\\n\t\tJunyi & https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=1198 \\\\\n\t\tEdNet & http://ednet-leaderboard.s3-website-ap-northeast-1.amazonaws.com/\\\\\n\t\tpoj & http://poj.org/ \\\\\n\t\t\\hline\n\t\\end{tabular}\n\t\\label{tab3}\n\t\\vspace{-0.6cm}\n\\end{table*}\n\\fi","id":"c5ab0210-1d4b-4488-91f4-cb2e837345b7","level":"subsection","origin_cites_number":9,"parent_id":"5fc15152-8ae0-4a7f-83a2-d64f287123fe","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Applications"],["subsection","Broader Applications"]],"subsections":[],"title":"Broader Applications"},{"cite_extract_rate":0.14285714285714202,"cites":[5386],"content":"\\label{sec:dataset}\nAfter introducing above KT models and variants, to better help researchers and practitioners who want to further conduct related work and promote the application of KT, we have open sourced two algorithm libraries, i.e., EduData that for downloading and preprocessing most existing KT-related datasets, and EduKTM that includes extensible and unified implementations of existing popular KT models. In the following, we will give detailed introduction of these two algorithm libraries.\n\\begin{table*}[t]\n\t\\vspace{-0.6cm}\n\t\\renewcommand\\arraystretch{1.2}\n\t\\caption{{Basic information and statistics of existing datasets available for evaluating KT models. }}\n\t\\vspace{-0.2cm}\n\t\\centering\n\t\\resizebox{\\textwidth}{!}\n\t{\n\t\t\\begin{tabular}{l|c|c|c|c|c|r|r|r|r}\n\t\t\t\\hline\n\t\t\t\\multirow{3}*{Datasets}&\\multirow{3}{*}{Subjects} &\\multirow{3}{*}{Learning Stages} & \\multirow{3}*{\\makecell[c]{Side \\\\ Information}}& \\multicolumn{2}{c|}{Sources}&\\multicolumn{4}{c}{Statistics} \\\\\n\t\t\t\\cline{5-10} \n\t\t\t& & & & \\makecell[c]{Online \\\\ platforms} & \\makecell[c]{Educational \\\\ challenges} & \\makecell[c]{\\# of \\\\ Students} & \\makecell[c]{\\# of \\\\ Exercises} & \\makecell[c]{\\# of \\\\ KCs} & \\makecell[c]{\\# of \\\\ Learning records}\\\\\n\t\t\t\\hline\n\t\t\tASSISTments2009 & mathematics & middle school & Yes & \\Checkmark & & 4,163 & 17,751& 123 & 346,860 \\\\\n\t\t\t\\hline\n\t\t\tASSISTments2012 & mathematics & middle school & Yes & \\Checkmark & & 46,674 & 179,999& 265 &6,123,270 \\\\\n\t\t\t\\hline\n\t\t\tASSISTments2015 & mathematics & middle school & No & \\Checkmark & & 19,917 & /& 100 & 708,631 \\\\\n\t\t\t\\hline\n\t\t\tASSISTments2017& mathematics &\\makecell[c]{from middle school \\\\ to college} & Yes & & \\Checkmark & 1,709 &3,162& 102 & 942,816 \\\\\n\t\t\t\\hline\n\t\t\tJunyi & mathematics & \\makecell[c]{from primary \\\\ to high school} & Yes & \\Checkmark & & 247,606 & 722 & 41 & 25,925,922 \\\\\n\t\t\t\\hline\n\t\t\tEedi2020 & mathematics & \\makecell[c]{from primary \\\\ to high school} & Yes & & \\Checkmark & 118,971 & 27,613 & 388 & 15,867,850 \\\\\n\t\t\t\\hline\n\t\t\tStatics2011 &engineering &university &Yes & \\Checkmark & & 335 & 1,224 & 80 & 361,092 \\\\\n\t\t\t\\hline\n\t\t\tEdNet-KT1 & english & / & Yes & \\Checkmark & & 784,309 & 13,169 & 188 & 95,293,926 \\\\\n\t\t\t\\hline\n\t\t\tEdNet-KT2 & english &/ & Yes & \\Checkmark & & 297,444 & 13,169 & 188 & 56,360,602 \\\\\n\t\t\t\\hline\n\t\t\tEdNet-KT3 & english &/ & Yes & \\Checkmark & & 297,915 & 13,169 & 293 & 89,270,654 \\\\\n\t\t\t\\hline\n\t\t\tEdNet-KT4 & english &/& Yes& \\Checkmark & & 297,915 & 13,169 & 293 & 131,441,538 \\\\\n\t\t\t\\hline\n\t\t\tCodeWorkout & programming & university & Yes & & \\Checkmark & 819 & 50 & 50 & \\textgreater 130,000 \\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t}\n\t\\label{data}\n\t\\vspace{-0.6cm}\n\\end{table*}","id":"34caa8b6-1d19-4064-837a-9dcb1549a45d","level":"section","origin_cites_number":7,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"]],"subsections":["8d0ee175-d87a-4c6c-bbdc-2113cd06121b","66f9b876-576e-4b0f-af2b-91235cb5ec11"],"title":"Datasets and Baselines"},{"cite_extract_rate":0,"cites":[],"content":"As we have mentioned, KT emerges from the development of online education, where a large number of students' learning data is collected for analyzing their learning behaviors and knowledge states. In this section, we mainly introduce existing public datasets available for evaluating KT models. Table \\ref{data} lists all datasets, as well as their basic information and statistics. In our released EduData, we provide the service of downloading and preprocessing all these datasets, which is convenient to help beginners analyze and utilize them quickly.\nIn summary, these datasets are collected in different learning scenarios, so that they exhibit an extremely distinct difference in data scale, subject, and so on, indicating complex applications in practical for KT models.","id":"8d0ee175-d87a-4c6c-bbdc-2113cd06121b","level":"subsection","origin_cites_number":0,"parent_id":"34caa8b6-1d19-4064-837a-9dcb1549a45d","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Datasets"]],"subsections":["3d73fd7b-d881-422a-a516-5cd421232c1c","efcd8f6a-3777-4dd9-8b4a-f9125974da7d","db5459a4-5149-42f0-85da-a9bc2ed105e4","266d224a-ec55-40cb-a11c-e27be3a867bf","fd615d1a-1d1e-4b41-9cdf-9cd4873ba07b","6cb5a8b7-c682-457d-85c6-8ed300213a0a"],"title":"Datasets"},{"cite_extract_rate":0,"cites":[],"content":"ASSISTments , created in 2004, is an online tutoring system in the United States, which provides students with both assessment information and tutoring assistance. While working on ASSISTments, students will be provided with instructional assistance to help them solve the problem in several substeps when they give wrong answers. After obtaining correct answers, they will be given a new one. Meanwhile, the system will learn about the students' knowledge states and predict how they will do in future tests. Up to now, the organizers have released four public avaliable datasets from ASSISTments, which are respectively ASSISTments2009, ASSISTments2012\\footnote{https://sites.google.com/site/assistmentsdata/datasets/2012-13-school-data-with-affect}, ASSISTments2015\\footnote{https://sites.google.com/site/assistmentsdata/datasets/2015-assistments-skill-builder-data}, and ASSISTments2017\\footnote{https://sites.google.com/view/assistmentsdatamining/dataset}. The ASSISTments datasets have built profound impacts in the research community of KT, which can be seen in many related papers.\nMost of these datasets were collected from mathematics in middle school, the details of them are listed as follows.\n\\begin{itemize}[leftmargin=*]\n\t\\item{\\textbf{ASSISTments2009.}}\n\tThe full name of ASSISTments2009 is the ASSISTments 2009-2010 skill builder data set , which was collected form ASSISTments’ skill builder problem sets during the school year from 2009 to 2010. Students were asked to work on exercises with similar knowledge concepts until they can answer correctly for three or more times in a row. This dataset contains many valuable side information, such as \\textit{attempt count} that represents the number of attempts of the student, \\textit{ms first reponse} that represents the time in the milliseconds for the student’s first response, and \\textit{opportunity} that represents the number of opportunities the student has to practice. It worth noting that the original version of ASSISTments2009 has three serious problems that lead to unreliable experimental results : (1) a large number of duplicated records, (2) treating scaffolding problems as the same as main problems, and (3) repeated response sequences with different KCs. The latest version of ASSISTments2009 has fixed these problems.\n\t\\item{\\textbf{ASSISTments2012.}}\n\tASSISTments2012 is collected from the ASSISTments system, in the school year from 2012 to 2013 with affect predictions. Compared with ASSISTments2009, ASSISTments2012 has much more students, exercises, as well as learning records. However, many learning records have missed the related KCs in this dataset. After filtering these records, there are only 29,018 students, 53,091 exercises, and 2,711,813 learning records.\n\tBesides, this dataset contains more side information, such as \\textit{start time}, \\textit{end time}, \\textit{problem type}, and \\textit{average confidence}. \n\tIt worth noting that you can access to the text of the exercises in ASSISTments2012 for more fine-grained research. Specifically, you can email nth@wpi.edu and cc td@wpi.edu, explaining your purpose and promising not to share it with anyone else.\n\t\\item{\\textbf{ASSISTments2015.}}\n\tASSISTments2015 covers the 2015 school years' student response records on ASSISTments. This dataset only contains student learning records on 100 KCs, there is no information about the exercise, as well as any other side information. \n\t\\item{\\textbf{ASSISTments2017.}}\n\tASSISTments2017 is the dataset used in the 2017 ASSISTments Longitudinal Data Mining Competition. This dataset collected data over a decade, which tracked students' intercations at the ASSISTments learning platform from middle school study in 2004-2007 to their high school course-taking, until they graduated from the college. Therefore, the average learning record of students in ASSISTments2017 is much longer than other dataset, which is beyond the length of 1,000. ASSISTments2017 also has rich side information, including but not limited to \\textit{AveKnow} that indicates students' average knowledge level based on BKT, \\textit{timeTaken} that represents the time spent on the current exercise, \\textit{frIsHelpRequest} that represents whether the first response is a help request.\n\\end{itemize}","id":"3d73fd7b-d881-422a-a516-5cd421232c1c","level":"subsubsection","origin_cites_number":2,"parent_id":"8d0ee175-d87a-4c6c-bbdc-2113cd06121b","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Datasets"],["subsubsection","ASSISTments Datasets"]],"subsections":[],"title":"ASSISTments Datasets"},{"cite_extract_rate":0,"cites":[],"content":"The Junyi dataset contains the problem log and exercise-related information on the Junyi Academy\\footnote{http://www.junyiacademy.org/}, a Chinese e-learning website established in 2012 on the basis of the open-source code released by Khan Academy. In contrast to the ASSISTments datasets, Junyi has less exercsies and KCs, but includes an exercise hierarchy labeled by experts, the annotations of exercise relationship are also available. Therefore, many research works that focused on the knowledge structure in KT had utilized this dataset . Junyi provides the prerequisite exercise of a specific exercise in the knowledge map, the topic and area of each exercise, as well as the coordiate position of the knowledge map.","id":"efcd8f6a-3777-4dd9-8b4a-f9125974da7d","level":"subsubsection","origin_cites_number":2,"parent_id":"8d0ee175-d87a-4c6c-bbdc-2113cd06121b","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Datasets"],["subsubsection","Junyi Dataset"]],"subsections":[],"title":"Junyi Dataset"},{"cite_extract_rate":0.5,"cites":[8913],"content":"The Eedi2020 dataset is also released in an achedemic challenge, i.e., NeurIPS 2020 Education Challenge\\footnote{https://eedi.com/projects/neurips-education-challenge}. This dataset contains students’ answers to mathematics questions from Eedi, an online educational platform which millions of students interact with daily around the globe from school year 2018 to 2020. All exercises are multiple-choice problems with 4 possible answer choices, exactly one of which is correct. In Table \\ref{data}, we give the statistics based on the training data in this competition, the total number of learning records in the full dataset exceeds 17 million. It worth noting that Eedi2020 gives students' exact answer choice so that we can also predict students options . Moreover, for the students, Eedi2020 records lots of valuable context information, including the \\textit{Gender}, \\textit{DateOfBirth}, \\textit{PremiumPupil}. For the learning records, Eedi2020 also presents their \\textit{Confidence}, \\textit{GroupId}, \\textit{QuizId}, and \\textit{SchemeOfWorkId}.","id":"db5459a4-5149-42f0-85da-a9bc2ed105e4","level":"subsubsection","origin_cites_number":2,"parent_id":"8d0ee175-d87a-4c6c-bbdc-2113cd06121b","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Datasets"],["subsubsection","Eedi2020 Dataset"]],"subsections":[],"title":"Eedi2020 Dataset"},{"cite_extract_rate":0,"cites":[],"content":"Different from the above datasets that focus on mathematics exercises, the Statics2011 dataset is obtained from a college-level engineering statics course via online educational system developed by Carnegie Mellon University\\footnote{https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=507}. {The problems in college engineering course are quite complex, often comprising numerous independent sub-steps. Consequently, we treat each sub-problem as an exercise and calculate the number of exercises and KCs within this dataset.}","id":"266d224a-ec55-40cb-a11c-e27be3a867bf","level":"subsubsection","origin_cites_number":1,"parent_id":"8d0ee175-d87a-4c6c-bbdc-2113cd06121b","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Datasets"],["subsubsection","Statics2011 Dataset"]],"subsections":[],"title":"Statics2011 Dataset"},{"cite_extract_rate":1,"cites":[5386],"content":"The EdNet dataset is related to the english subject, which is consisted of students' learning records in the multi-platform AI tutoring system Santa in South Korea\\footnote{https://github.com/riiid/ednet}. EdNet collected learning data of students over two years for their preparation of the e TOEIC (Test of English for International Communication) Listening and Reading Test. EdNet is now the largest public dataset in KT field with a total of 131,441,538 learning records from 784,309 students. Besides, it contains various features of students' learning actions, such as the specific learning material they have interacted, how much time they have spent for answering a given exercise. There are four differnet versions of EdNet, respectively named EdNet-KT1, EdNet-KT2, EdNet-KT3, and EdNet-KT4 with different extents. {We note that the students in this dataset may be in different learning states. Therefore we do not give the learning state information in Table \\ref{data}.}\n\\begin{itemize}[leftmargin=*]\n\t\\item{\\textbf{EdNet-KT1.}}\n\tEdNet-KT1 contains students' basic exercise-answering logs. This dataset has 784,309 students, 13,169 exercises, 188 KCs, and a total of 95,293,926 learning records.\n\tExercises in EdNet-KT1 are organized by bundles, i.e., a collection of exercises sharing a common passage, picture or listening material. Therefore, exercises come up in bundles and students have to answer all contained exercises when a bundle is given.\n\t\\item{\\textbf{EdNet-KT2.}}\n\tEdNet-KT2 recorded students' action sequences, which indicated their full learning behaviors. For example, a student who is not confident about the answer may alternately select among several answer choices before submitting. Such learning behaviors can reflect more fine-grained knowledge state of students. EdNet-KT2 contains three kinds of actions: \\textit{enter} when student first receives and views a bundle , \\textit{respond} when the student selects an answer choice to the exercise, and \\textit{submit} when the student submits his final answers to the the given bundle. It is worth noting that EdNet-KT2 is a subset of EdNet-KT1.\n\t\\item{\\textbf{EdNet-KT3.}}\n\tOn the basis of EdNet-KT2, EdNet-KT3 collected more students' learning activities, such as reading explanations or watching lectures. These learning activities have potential impacts on students' knowledge state so that they are valuable to be analyzed.\n\t\\item{\\textbf{EdNet-KT4.}}\n\tIn EdNet-KT4, the very fine details of actions were provided. In particular, the following types of actions are added to EdNet-KT3: erase choice, undo erase choice, play audio, pause audio, play video, pause video, pay, refund, and enroll coupon.\n\\end{itemize}","id":"fd615d1a-1d1e-4b41-9cdf-9cd4873ba07b","level":"subsubsection","origin_cites_number":1,"parent_id":"8d0ee175-d87a-4c6c-bbdc-2113cd06121b","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Datasets"],["subsubsection","EdNet Dataset"]],"subsections":[],"title":"EdNet Dataset"},{"cite_extract_rate":0,"cites":[],"content":"The CodeWorkout dataset is utilized in the 2nd Computer Science Educational Data Mining Challenge (CSEDM)\\footnote{https://sites.google.com/ncsu.edu/csedm-dc-2021/home}. This dataset is collected from a CS1 course in the Spring and Fall 2019 semesters at a public university in United States. It contains the code submissions from students for 50 coding problems, each requiring ~10-26 lines of code. {In this dataset, each exercise has a unique KC, thereby the number of KCs is equal to the number of exercises.} In total, there are 329 and 490 students in the Spring and Fall semesters who completed the course. Each dataset contains more than 65,000 code submissions, the scores of the submissions (\\% of unit tests passes) are also available, as well as the compiler message if the compilation is not successful. The final grades of students are also provided for this dataset.","id":"6cb5a8b7-c682-457d-85c6-8ed300213a0a","level":"subsubsection","origin_cites_number":1,"parent_id":"8d0ee175-d87a-4c6c-bbdc-2113cd06121b","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Datasets"],["subsubsection","CodeWorkout Dataset"]],"subsections":[],"title":"CodeWorkout Dataset"},{"cite_extract_rate":0,"cites":[],"content":"The implementations of existing KT methods are not standardized, which may use different program languages (e.g., python, lua) and different deep learning frameworks (e.g., tensorflow, torch). Furthermore, some works did not well organize the codes systematically (e.g., the missing of running environments and dependencies), which brings difficulties in reproducing the models. To this end, we put forward the algorithm library for KT baselines, named EduKTM, which now has contained the concurrent popular works. EduKTM will be always under development for including the latest KT models, more algorithms and features are going to be added. Besides, we provide detailed guidelines for everyone who is interested in contributing to EduKTM.\nIt worth nothing that we do not provide the performance evaluation of these baselines using the aforementioned benchmark datasets. The reasons are two-folds: (1) For each baseline method, their experimental settings are quite different as they were designed to handle with various learning scenarios. No single KT model can always be the best one under various learning contexts. It is not fair to compare all baselines under a fixed setting.\n(2) Existing evaluation standards primarily concentrate on the performance of student performance prediction task, which cannot directly reflect the effectiveness of various KT models in practical applications.\nTherefore, we decide to open source EduData and EduKTM. Researchers and practitioners can freely compare and select appropriate KT models based on their specific requirements in various application scenarios.","id":"66f9b876-576e-4b0f-af2b-91235cb5ec11","level":"subsection","origin_cites_number":0,"parent_id":"34caa8b6-1d19-4064-837a-9dcb1549a45d","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Datasets and Baselines"],["subsection","Baselines"]],"subsections":[],"title":"Baselines"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:future}\nThis survey has comprehensively reviewed the abundant current developments in the field of knowledge tracing, including fundamental KT models, their variations and typical applications. Despite this, as knowledge tracing is a relatively new but promising research area, there are still a substantial number of problems that require urgent resolution. In this section, we discuss several potential future research directions.","id":"18821b92-3ccc-4057-9352-24e5c0409dd6","level":"section","origin_cites_number":0,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Future Research Directions"]],"subsections":["ee9a478b-7e5c-4c3d-a3b7-01ab255d95f7","203ad4ee-8eb1-4e57-ac27-fd2974c1a08f","1df76c0e-efc5-4366-b813-6204fb5c7b9f","ed0bdca3-3fb6-4028-a4d7-cc6fb028657b","01240c30-0ca2-4a6c-bf1e-ccabebf5cfe7","c7707957-a296-4aea-b19d-29d5e30c34c6"],"title":"Future Research Directions"},{"cite_extract_rate":0.4,"cites":[5374,8914],"content":"The performance of KT models is now evaluated indirectly through the student performance prediction task. The higher the precision of students' responses on future exercises, the better the KT model's performance. Nonetheless, interpretability plays a significant role in education, as students often express more concern about the 'why' than the 'what' of a learning decision . Enhancing the interpretability of KT models is therefore crucial. Some educational theories, such as the Rasch model used in AKT and the transfer of knowledge used in SKT , could be considered for this purpose. noticed the significance of interpretable KT, particularly for deep learning models, as they have numerous parameters that are challenging to explain meaningfully. Thus, they introduced a straightforward and interpretable KT approach, based on the causal relationships within the latent features extracted from students' behavioral data. \n attempted to introduce causal inference for explanatory analysis on KT, and they achieved more stable and explainable knowledge tracing based on the analysis results. \nIt is imperative to further refine existing KT models or to explore additional methods for interpretable KT researches. This will lead to the production of more accurate and interpretable evaluations of students' knowledge states.","id":"ee9a478b-7e5c-4c3d-a3b7-01ab255d95f7","level":"subsection","origin_cites_number":5,"parent_id":"18821b92-3ccc-4057-9352-24e5c0409dd6","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Future Research Directions"],["subsection","Knowledge Tracing with Interpretability"]],"subsections":[],"title":"Knowledge Tracing with Interpretability"},{"cite_extract_rate":0.6000000000000001,"cites":[5388,5387,8915],"content":"The acquisition of high-quality KT models necessarily requires a substantial amount of data to ensure training stability. However, online learning systems often achieve less learning interactions in practical educational scenarios, thereby leading to the data sparsity problem. To address this issue, utilized graph convolutional network to include exercise-KC correlations. \n turned to improve the attention mechanism.\nMore recently, researchers have employed contrastive learning to alleviate the data sparsity problem in KT . For example, presented a contrastive learning framework to enhance KT, which measure the semantical similarity between various learning interactions to effectively learn their representations. They further designed data augmentation methods to enhance the semantics of students' learning interactions. However, while the aforementioned methods alleviate the data sparsity problem in various aspects, there remains a need for further improvement in addressing this issue comprehensively.","id":"203ad4ee-8eb1-4e57-ac27-fd2974c1a08f","level":"subsection","origin_cites_number":5,"parent_id":"18821b92-3ccc-4057-9352-24e5c0409dd6","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Future Research Directions"],["subsection","Knowledge Tracing with Sparse Learning Interactions"]],"subsections":[],"title":"Knowledge Tracing with Sparse Learning Interactions"},{"cite_extract_rate":0,"cites":[],"content":"Most existing KT researches are capable of dealing with objective exercises, assuming student responses are binary. However, they overlook students' open-ended answers, often presented in natural language, on subjective exercises. As per the work by , an open-ended KT approach for computer science education was explored. This method aimed to predict students' precise solutions to programming questions, utilizing language models for assistance. \nRecent advancements in large language models offer promising potential to enhance the ability of KT models to comprehend the critical information regarding students' knowledge states embedded within their open-ended answers on subjective exercises .","id":"1df76c0e-efc5-4366-b813-6204fb5c7b9f","level":"subsection","origin_cites_number":2,"parent_id":"18821b92-3ccc-4057-9352-24e5c0409dd6","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Future Research Directions"],["subsection","Knowledge Tracing with Subjective Exercises"]],"subsections":[],"title":"Knowledge Tracing with Subjective Exercises"},{"cite_extract_rate":0,"cites":[],"content":"Learning records are passive representations of students' knowledge states. In contrast, students' feedback provides us with their proactive understanding of their knowledge states, thereby offering direct and authentic indicators of their learning situation. However, there are few KT models that take advantage of training data related to students' feedback, even though it can play an important role in fixing the KT results . have noted that feedback plays a positive role in learning, which may promote transfer and retention in learning from worked-out examples. Therefore, incorporating students' feedback presents a meaningful avenue that could lead to significant improvements.","id":"ed0bdca3-3fb6-4028-a4d7-cc6fb028657b","level":"subsection","origin_cites_number":2,"parent_id":"18821b92-3ccc-4057-9352-24e5c0409dd6","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Future Research Directions"],["subsection","Knowledge Tracing with Students' Feedback"]],"subsections":[],"title":"Knowledge Tracing with Students' Feedback"},{"cite_extract_rate":0,"cites":[],"content":"Generally, user modeling refers to tools for characterizing users' behaviors (e.g., frequent locations), personal information (e.g., age, gender, and occupation) and latent features (e.g., interests and abilities), which facilitate the provision of targeted services for different users . As a type of latent feature modeling, knowledge tracing diagnoses the proficiency of users (not only individuals, but also groups of individuals, like user teams and companies) on specific skills/concepts. Thus, in addition to education, knowledge tracing can be developed and applied across a wide range of domains for user modeling, including games, sports, and recruitment.","id":"01240c30-0ca2-4a6c-bf1e-ccabebf5cfe7","level":"subsection","origin_cites_number":1,"parent_id":"18821b92-3ccc-4057-9352-24e5c0409dd6","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Future Research Directions"],["subsection","Knowledge Tracing for General User Modeling"]],"subsections":[],"title":"Knowledge Tracing for General User Modeling"},{"cite_extract_rate":0.33333333333333304,"cites":[8461],"content":"{In recent years, the advancement of large language models (LLM), notably ChatGPT, has led to significant impacts and garnered considerable research interest worldwide . The use of LLM in education is promising to revolutionize the current learning pattern . The current study mainly focuses on generating learning materials, improving student-system interaction, and explaining educational contents. However, it remains unclear how LLM can assist in understanding students' knowledge states. Given the powerful capabilities of LLM, it has the potential to enhance the generalization and interpretability of existing KT methods. Specifically, LLM can analyze the content of students' responses and evaluate their quality, which directly reflects students' knowledge states. Besides, LLM can play the role of the instructor, answering questions for students and identifying their strengths and weaknesses. Furthermore, LLM itself can serve as a KT model, which can output reasonable results about students' knowledge states given their previous learning interactions.\nSimultaneously, it is imperative to safeguard the privacy and security of student data when utilizing LLM . \n}","id":"c7707957-a296-4aea-b19d-29d5e30c34c6","level":"subsection","origin_cites_number":3,"parent_id":"18821b92-3ccc-4057-9352-24e5c0409dd6","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Future Research Directions"],["subsection","Knowledge Tracing with Large Language Models"]],"subsections":[],"title":"Knowledge Tracing with Large Language Models"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclution}\nIn this survey, we conducted a thorough review of knowledge tracing. Specifically, we initially conducted a comprehensive review of existing fundamental KT models. Considering the complexity of online learning systems and the significant importance of KT research in practical scenarios, we investigated a broad range of variant KT research. Subsequently, we summarized some typical applications of KT in common educational scenarios. Furthermore, we released two algorithm libraries for KT-related datasets and baselines, thereby facilitating researchers and practitioners in selecting appropriate KT models based on their specific requirements. Finally, we outlined some potential future directions. \nWe hope that this comprehensive survey of knowledge tracing will assist readers in understanding the problem of modeling students' dynamic knowledge states. It serves as a fundamental framework for both researchers and practitioners in future studies, fostering the development of KT. The development of KT will directly benefit millions of students, with its impact potentially extending to a broader audience as online learning continues to evolve. In this context, KT will play an increasingly important role in enabling individuals to adapt to the ever-changing society . Furthermore, the development of more refined KT methods tailored to students in various subjects and age groups will enhance their ability to comprehend students' individual knowledge states. \nHowever, we recognize some limitations of the current survey. \nFirstly, as new KT methods continue to emerge, although we have conducted a thorough investigation, there may be some representative works that we have overlooked. If necessary, we will incorporate these into our proposed framework in the future. \nSecondly, the complexity of KT methods warrants attention, as it continues to grow, particularly in those based on deep learning. The complexity of the model may enhance its accuracy, but it could also compromise its applicability, as users will question the reliability of complex models. Despite the fact that some studies have attempted to reveal the interpretability of complex KT models, for instance, by utilizing the xAI technique, there remains a significant distance to traverse before achieving a completely transparent deep learning KT model. \nWe posit that the application of KT methods in online education presents a promising avenue for research. It significantly enhances the learning experience for students while simultaneously alleviating the burden on teachers. Despite the numerous challenges and obstacles, researchers are be encouraged to overcome them and ensure a reliable and equitable access to students' evolving knowledge state in learning. \n\\section*{Acknowledgment}\nThis research was supported by grants from the National Key Research and Development Program of China (Grant No. 2021YFF0901003), and the National Natural Science Foundation of China (Grant No. 62337001, U20A20229).\n\\bibliographystyle{IEEEtranN}\n\\begin{footnotesize}\n\t\\bibliography{cite}\n\\end{footnotesize}\n\\vspace{-0.6cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/shen.pdf}}]\n\t{Shuanghong Shen} recieved the B.E. degree from Wuhan University, Wuhan, China, in 2018. He is currently working toward the Ph.D. degree in the School of Computer Science and Technology, University of Science and Technology of China. His main research interests include data mining, knowledge discovery, natural language processing and intelligent tutoring systems. He won the first prize in task 2 of the NeurIPS 2020 Education Challenge. He has published papers in referred conference proceedings, such as KDD2023, AAAI2022, SIGIR2022.\n\\end{IEEEbiography}\n\\vspace{-0.6cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/liu.pdf}}]\n\t{Qi Liu} (Member, IEEE) received the Ph.D. degree from University of Science and Technology of China (USTC), Hefei, China, in 2013. He is currently a Professor in the School of Computer Science and Technology at USTC. His general area of research is data mining and intelligent education. He has published prolifically in refereed journals and conference proceedings (e.g., TKDE, TOIS, KDD). He is an Associate Editor of IEEE TBD and Neurocomputing. He was the recipient of KDD’ 18 Best Student Paper Award and ICDM’ 11 Best Research Paper Award. He is a member of the Alibaba DAMO Academy Young Fellow. He was also the recipient of China Outstanding Youth Science Foundation in 2019.\n\\end{IEEEbiography}\n\\vspace{-0.6cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/huang.pdf}}]\n\t{Zhenya Huang} (Member, IEEE) received the B.E. degree from Shandong University, in 2014 and the Ph.D. degree from USTC, in 2020. He is currently an associate researcher of the School of Computer Science and Technology, University of Science and Technology of China (USTC). His main research interests include data mining, knowledge discovery representation learning and intelligent tutoring systems. He has published more than 30 papers in refereed journals and conference proceedings including TKDE, TOIS, KDD, AAAI. He has served regularly in the program committees of a number of conferences, and is reviewers for the leading academic journals.\n\\end{IEEEbiography}\n\\vspace{-0.6cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/zheng.png}}]\n\t{Yonghe Zheng} is the funding dean of Research Institute of Science Education in Beijing Normal University, Professor in Education Faculty of BNU, PhD Supervisor. Prof. Zheng is the Vice president of the 8th Council of China Association of Children's Science Instructors and is the member of Teaching Guidance Committee of the Basic Education, the member of comprehensive group of the compulsory education curriculum revision committee, the science curriculum standard revision group; the General Secretary of the Awarding Scheme Council for outstanding computer teachers in Universities; the Vice director of the academic committee of Chinese society of Natural Science Museum; and former Director General of the Science Policy Bureau of National Natural Science Foundation of China. Prof. Zheng has engaged in project management of basic research and S\\&T policy research for a long time. More than 60 articles have been published. And Zheng is the chief editor of Disciplinary and Interdisciplinary Science Education Research Journal. Prof. Zheng joined in Educational Faculty of Normal University in April, 2018, with specific research interests on science education and educational technology. Prof. Zheng established the Research Institute of Science Education of Beijing Normal University in November of 2019.\n\\end{IEEEbiography}\n\\vspace{-0.6cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/yin.png}}]\n\t{Minghao Yin} (Member, IEEE) received the B.S. and M.S. degrees from Northeast Normal University, Changchun, China, in 2001 and 2004, respectively, and the Ph.D. degree from Jilin University, Changchun, China, in 2008, all in computer science. He is currently a professor, the dean of the School of Information Science and Information Technology at Northeast Normal University, and the director of the Key Laboratory of Intelligent Information Processing in Jilin Province Colleges and Universities. He also contributes as the Associate Editor for the IEEE Transactions on Learning Technologies (TLT). He has published multiple papers in prestigious journals and conferences, including Artificial Intelligence, IEEE Transactions, ACM Transactions, JAIR, AAAI, IJCAI, and WWW. His research interests include heuristic search, data mining, and combinatorial optimization.\n\\end{IEEEbiography}\n\\vspace{-0.6cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/wang.png}}]\n\t{Minjuan Wang} (Member, IEEE) obtained her B.A. in Chinese literature from the Beijing/Peking University in 1995, M.A. in comparative literature from Penn State University in 1997, and Ph.D. in information science and learning technologies from University of Missouri-Columbia in 2001. She is\ta newly appointed Chair Professor of Emerging Technologies and Future Education at The Education University of Hong Kong. She has also been a professor and program chair of Learning design and technology at San Diego State University, California, USA since 2000. In addition, she serves as the Editor-in-Chief for the IEEE Transactions on Learning Technologies (TLT). Her research specialties are multidisciplinary, focusing on STEM education, new and emerging technologies in various educational settings, Metaverse and immersive learning, and\tthe design and implementation of artificial intelligence including AIGC for education and training. She is an internationally recognized scholar and has keynoted about 45 international conferences. She is recognized internationally for her research, publishing and dedicated service to IEEE and other scholarly communities. Dr. Wang is a member of IEEE and the IEEE Education Society. She also co-chairs the Education Society’s technical committee for immersive learning (TC-ILE). \n\\end{IEEEbiography}\n\\vspace{-0.6cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/chen.pdf}}]\n\t{Enhong Chen} (Fellow, IEEE) received the B.S. degree from Anhui University, Hefei, China, the M.S. degree from the Hefei University of Technology, Hefei, China, and the Ph.D. degree in computer science from the University of Science and Technology of China (USTC), Hefei, China, in 1989, 1992 and 1996 respectively. He is currently a Professor and the Executive Dean of School of Data Science, the Director of Anhui Province Key Laboratory of Big Data Analysis and Application. He has published a number of papers on refereed journals and conferences, such as TKDE, TIST, TMC, KDD, ICDM, NIPS and CIKM. His current research interests include data mining and machine learning, social network analysis, and recommender system. Dr. Chen was a recipient of the National Science Fund for Distinguished Young Scholars of China, the Best Application Paper Award on KDD-2008, the Best Student Paper Award on KDD-2018 (Research), and the Best Research Paper Award on ICDM-2011. He is an IEEE Fellow and an ACM Distinguished Scientist.\n\\end{IEEEbiography}\n\\vspace{-0.6cm}\n\\end{document}","id":"8a28f5df-8a26-4509-8bdf-0a8c518eb23c","level":"section","origin_cites_number":1,"parent_id":"4e5af7af-552b-4c7f-939b-54b6b4dadeb9","prefix_titles":[["title","A Survey of Knowledge Tracing: Models, Variants, and Applications"],["section","Discussions and Conclusions"]],"subsections":[],"title":"Discussions and Conclusions"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ]\n ],\n \"subsections\": [\n \"a96dd0ad-2f86-4f16-b5a9-882602bf6a06\",\n \"6987e516-0520-40ba-bdfa-9ecf587cfc88\",\n \"bd30f2af-cabd-4e04-ab90-5536a2f60dde\",\n \"a705e159-25ef-44ad-8166-1cef83a7c265\",\n \"5fc15152-8ae0-4a7f-83a2-d64f287123fe\",\n \"34caa8b6-1d19-4064-837a-9dcb1549a45d\",\n \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"8a28f5df-8a26-4509-8bdf-0a8c518eb23c\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.21739130434782603,\n \"cites\": [\n 5371,\n 5372,\n 5370,\n 5374,\n 5373\n ],\n \"content\": \"\\\\IEEEPARstart{W}{ith} the proliferation of the Internet and mobile communication technology, online education has become increasingly popular and is now developing at an unprecedented scale .\\n{This innovative style of learning provides a degree of flexibility that conventional education cannot match, which enables teaching and learning to occur at any time and any place.}\\nMeanwhile, online learning systems (e.g., Coursera, ASSISTment) have proven to be even more effective than traditional learning styles, since they can offer more intelligent educational services, such as recommending individualized learning resources to students . \\nTo provide these intelligent services, online learning systems continuously record a massive amount of available data about student-system interactions {(e.g., responding to exercises)}, which can be further mined to assess their knowledge levels, learning preferences, and other attributes. Specifically, Knowledge Tracing (KT) is one of the most fundamental and critical tasks for analyzing students' learning behavior data, which aims to explore the recorded student-system interactions to monitor their evolving knowledge states .\\n\\\\begin{figure*}[t]\\n\\t\\\\vspace{-0.5cm}\\n\\t\\\\centerline{\\\\includegraphics[width=\\\\textwidth]{Figures/learning_case.pdf}}\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\caption{A simple schematic diagram of knowledge tracing. Different knowledge concepts are represented in different colors, while exercises are also depicted in the color relevant to the knowledge concepts. During the learning process, different kinds of side information are also recorded. The evolving process of the knowledge state is assessed by KT models and illustrated by the radar maps.}\\n\\t\\\\label{kt}\\n\\t\\\\vspace{-0.5cm}\\n\\\\end{figure*}\\nFig. \\\\ref{kt} presents a simple schematic diagram of knowledge tracing. During the learning process, online learning systems continuously record students' learning behavioral data, including exercises and their related knowledge concepts (e.g., \\\\emph{equality, inequality, plane vector}, \\\\emph{probability}, represented in various colors), and students' answers (i.e., correct or incorrect responses). A substantial amount of supplementary information is also simultaneously recorded, including response time, opportunity count, and tutor intervention, which provides a more comprehensive reflection of students' learning process. \\nBased on the collected learning data, researchers are striving to maintain an estimate of students' evolving knowledge states. {For illustration, we give a case in Fig. \\\\ref{kt}, where the student's prior knowledge is quantified as 0.2, 0.4, 0.4, and 0.5 across four distinct knowledge concepts. The radar map serves as a visual representation of the student's knowledge mastery, which progressively expands as the student continues to acquire new knowledge in learning.} After a period of learning, the student's knowledge states reach 0.9, 0.8, 0.8, and 0.7 respectively, suggesting good knowledge growth. In the aforementioned learning process, KT models aim to monitor changes in students' knowledge states. Once we understand students' knowledge states, the learning system can customize more suitable learning schemes for different students, thereby enabling the teaching of students in accordance with their proficiency. It also allows students to better comprehend their learning process and gradually focus on improving their skills with poorly mastered concepts .\\nKnowledge tracing has been studied for decades, with the first studies tracing back to the late 1970s. These initial works primarily focused on confirming the effectiveness of mastery learning . To the best of our knowledge, ~ were the first to introduce the concept of knowledge tracing, employing Bayesian networks to model the student learning process, which they referred to as Bayesian Knowledge Tracing. Since then, the significance of KT has been recognized by a broader spectrum of researchers, and increasing attention has been directed towards KT-related research. Many logistic models have been applied to KT, including Learning Factor Analysis and Performance Factor Analysis . In recent years, deep learning has greatly enhanced research into the KT task, largely due to its capacity to extract and represent features and discover intricate structure. For instance, Deep Knowledge Tracing introduced recurrent neural networks (RNNs) into the KT task and was found to significantly outperform previous methods . Following this, various methods have been introduced that employ various types of neural networks to the KT task, considering various characteristics of the learning sequence . Moreover, due to the requirements of practical applications, many variants of KT models have been continuously developed, and KT has already been broadly applied in numerous educational scenarios. \\nWhile novel KT models continue to emerge, there remains a lack of comprehensive surveys exploring this young research field, particularly {regarding its} numerous variants and applications. To this end, {the current survey aims to systematically review the development of KT}. As depicted in Figure \\\\ref{tax}, we initially categorize existing KT models from a technical perspective, which is consistent with the majority of existing surveys . This categorization splits them into three categories: (1) Bayesian models, (2) logistic models, and (3) deep learning models. In each category, we further organize specific KT methods according to their various techniques. \\nSubsequently, we introduce extensive variants of these fundamental KT models, which consider more stringent assumptions about more complete learning process in different learning phases. In addition, we present several typical applications of KT in real learning scenarios. Due to the complexity of different KT models, we have open sourced two algorithm libraries to better aid researchers and practitioners in implementing KT models and facilitate community development in this domain. These libraries, EduData\\\\footnote{https://github.com/bigdata-ustc/EduData} and EduKTM\\\\footnote{https://github.com/bigdata-ustc/EduKTM}, include most existing KT-related datasets, extensible, and unified implementations of existing KT models, and relevant resources. Finally, we discuss potential future research directions. In summary, this paper presents an extensive survey of KT that can serve as a comprehensive guide for both researchers and practitioners. \\nThe remainder of this survey is structured as follows. Section \\\\ref{sec:overview} presents an overview of the KT task and we discuss the differences between this and previous surveys. Section \\\\ref{sec:models} provides a review of the three categories of fundamental KT models. \\nSection \\\\ref{sec:variants} describes the variants of fundamental KT models. \\nSection \\\\ref{sec:applications} introduces the extensive applications of KT in different scenarios. Section \\\\ref{sec:dataset} gives the summary of existing datasets for evaluating KT models and details of the algorithm libraries we have released. Section \\\\ref{sec:future} discusses some potential future research directions. Finally, section \\\\ref{sec:conclution} summarizes the paper.\\n\\\\begin{figure*}[t]\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\centerline{\\\\includegraphics[width=\\\\textwidth]{Figures/taxonomy.pdf}}\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\caption{{An overview of knowledge tracing models.} }\\n\\t\\\\label{tax}\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{figure*}\",\n \"id\": \"a96dd0ad-2f86-4f16-b5a9-882602bf6a06\",\n \"level\": \"section\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:overview}\",\n \"id\": \"6987e516-0520-40ba-bdfa-9ecf587cfc88\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Overview\"\n ]\n ],\n \"subsections\": [\n \"0c71baa7-2bd1-4dbb-a83e-a14394274744\",\n \"c68cad9f-5ccb-45ac-9535-78433e591dfe\",\n \"1f61ba5e-8077-4921-924c-62e9bddd3afc\"\n ],\n \"title\": \"Overview\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5376,\n 5375,\n 38\n ],\n \"content\": \"In an online learning system, supposing there exists a set of students $\\\\mathbb{S}$ and a set of exercises $\\\\mathbb{E}$. {Each exercise is related to specific Knowledge Concepts (KCs). Generally, the name given to the knowledge related to exercises differs across online learning platforms. For instance, it is named \\\\emph{skill} in ASSISTments . To promote better understanding, we refer to these uniformly as knowledge concepts throughout this paper, and denote the set of all KCs as $\\\\mathbb{KC}$. Moreover, $M$ and $K$ are respectively used to represent the total number of exercises and KCs. Students are asked to answer different exercises in order to achieve mastery of the related knowledge.}\\nTherefore, the learning sequence of a student can be formulated as $\\\\bm{X} = \\\\{([e_1, k_{e_1}], a_1, r_1), ([e_2, k_{e_2}], a_2, r_2), ..., ([e_t, k_{e_t}], a_t, r_t), ..., \\\\\\\\([e_N, k_{e_N}], a_N, r_N)\\\\}$, where the tuple $([e_t, k_{e_t}], a_t, r_t)$ represents the learning interaction at the $t-$th time step, $e_t$ represents the exercise, $k_{e_t}$ represents the exercise's related KCs, $a_t$ represents the correctness label (i.e., with 1 for correct and 0 for incorrect answers), $r_t$ stands for the side information recorded in this learning interaction, and $N$ is the length of the learning sequence. The research problem of knowledge tracing can thus be defined as follows:\\n\\\\textbf{Given sequences of learning interactions in online learning systems, knowledge tracing aims to monitor students' evolving knowledge states during the learning process and predict their performance on future exercises. The measured knowledge states can be further applied to individualize students' learning schemes in order to maximize their learning efficiency.}\\nSome recent works directly regarded the KT task as student performance prediction, without considering students' knowledge states . We agree that predicting student performance is of great significance, as it is now {the best way} to evaluate the quality of the knowledge state traced by KT models. However, we have to point out that KT focuses more on students' knowledge states, especially their interpretability and rationality, which is related to the students' acceptance of the conclusions given based on the KT model .\",\n \"id\": \"0c71baa7-2bd1-4dbb-a83e-a14394274744\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"6987e516-0520-40ba-bdfa-9ecf587cfc88\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Overview\"\n ],\n [\n \"subsection\",\n \"Problem Definition\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Problem Definition\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"{ As shown in Fig. \\\\ref{tax}, we categorize and summarize the existing KT models according to their technical differences. In more detail, the proposed taxonomy splits existing KT methods into three categories: (1) Bayesian models, which are implemented through probability model, (2) logistic models, which are implemented through logistic functions, and (3) deep learning models, which are implemented through neural networks. Specifically, we divide the deep learning models into four sub-categories according to four various neural networks, i.e., deep knowledge tracing based on recurrent neural networks, memory-aware knowledge tracing based on memory networks, attentive knowledge tracing based on self-attention mechanism, and graph-based knowledge tracing based on graph neural networks.\\nIn addition to these fundamental KT models, we also introduce a large number of their variants, which respectively consider a more complete learning process in distinct learning phases: (1) modeling individualization before learning, (2) incorporating engagement during learning, (3) considering forgetting after learning, and (4) utilizing side information across learning. Moreover, we also summarize the extensive applications of KT in different educational scenarios, including learning resources recommendation, adaptive learning and broader applications beyond student learning.}\",\n \"id\": \"c68cad9f-5ccb-45ac-9535-78433e591dfe\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"6987e516-0520-40ba-bdfa-9ecf587cfc88\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Overview\"\n ],\n [\n \"subsection\",\n \"Categorization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Categorization\"\n },\n {\n \"cite_extract_rate\": 0.41176470588235203,\n \"cites\": [\n 5372,\n 5377,\n 5374,\n 5370,\n 8911,\n 5371,\n 1206\n ],\n \"content\": \"Given the increasing importance of KT, several recent surveys have also examined this area. These include works by , . Here, we will briefly discuss the key distinctions between these studies to highlight the necessity and significance of this survey. \\nExisting surveys have either focused on specific categories of KT models or comprehensively reviewed all available KT models. {For example, provided an overview of KT in terms of Bayesian models and logistic models, compared and discussed deep learning based KT models, while presented to pay more attention to hybird models in KT.} conducted a bibliometric analysis to examine the evolution of KT research from 1992 to 2021. also presented a comprehensive survey for the KT literature, including a broad range of methods starting from the early attempts to the recent state-of-the-art techniques utilizing deep learning. summarized KT methods in the context of student performance modeling problems. \\nHowever, current surveys are somewhat limited in their scope, they only provide a detailed introduction to various KT methods and comparisons between them. Given the complexity of online learning systems and the significant importance of KT research in practical applications, this survey places a greater emphasis on the variants and applications of KT models, rather than solely introducing and comparing different KT methods. Moreover, considering that datasets are collected from different systems with various setting, subjects, learning stages, and scales, \\nwe do not report and compare the performance of KT models on the student performance prediction task across various datasets in this survey. have also empirically verified that no single KT model was always the best, a specific better model must consider multiple student features and the learning context.\\nInstead, we have open sourced two algorithm libraries which include the majority of existing KT-related datasets and unified implementations of existing KT models. Consequently, researchers and practitioners can freely select appropriate KT models based on their specific requirements in various application scenarios. \\n\\\\begin{table*}[t]\\n\\t\\\\vspace{-0.6cm}\\n\\t\\\\centering\\n\\t\\\\renewcommand\\\\arraystretch{1.0}\\n\\t\\\\caption{A summary of different types of fundamental knowledge tracing models.}\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\resizebox{\\\\textwidth}{!}{\\n\\t\\t\\\\begin{tabular}{|c|c|c|c|c|}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multicolumn{1}{|c|}{Category} & Typical approach & Technique & KC relationship & Knowledge state \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multicolumn{1}{|c|}{\\\\multirow{2}[5]{*}{Bayesian models}} & Bayesian knowledge tracing & Bayesian networks & independent & \\\\multirow{2}[5]{*}{\\\\shortstack{unobservable node \\\\\\\\ in HMM}} \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\cline{2-4} & dynamic Bayesian knowledge tracing & dynamic Bayesian networks & pre-defined & \\\\multicolumn{1}{c|}{} \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multicolumn{1}{|c|}{\\\\multirow{3}[8]{*}{Logistic models}} & learning factor analysis & \\\\multirow{2}[4]{*}{logistic regression} & \\\\multirow{3}[8]{*}{independent} & \\\\multirow{3}[6]{*}{\\\\shortstack{the output of \\\\\\\\ logistic regression function}} \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\cline{2-2} & performance factor analysis & \\\\multicolumn{1}{c|}{} & \\\\multicolumn{1}{c|}{} & \\\\multicolumn{1}{c|}{} \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\cline{2-3} & knowledge tracing machines & factorization machines & \\\\multicolumn{1}{c|}{} & \\\\multicolumn{1}{c|}{} \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multicolumn{1}{|c|}{\\\\multirow{4}[6]{*}{\\\\shortstack{Deep learning \\\\\\\\ models}}} & deep knowledge tracing & RNN/LSTM & discover automatically & the hidden state \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\cline{2-5} & memory-aware knowledge tracing & memory networks & correlation weights & \\\\emph{value} matrix \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\cline{2-5} & attentive knowledge tracing & self-attention mechanism & attention weights & attentive historical knowledge state \\\\\\\\\\n\\t\\t\\t\\\\cline{2-5} & graph-based knowledge tracing & graph neural networks & edges in graph & aggregate in the graph \\\\bigstrut\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\t\\\\vspace{-0.6cm}\\n\\t\\\\label{tab1}\\n\\\\end{table*}\",\n \"id\": \"1f61ba5e-8077-4921-924c-62e9bddd3afc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"6987e516-0520-40ba-bdfa-9ecf587cfc88\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Overview\"\n ],\n [\n \"subsection\",\n \"Differences between this and previous surveys\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Differences between this and previous surveys\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:models}\\nIn this section, as shown in Table \\\\ref{tab1}, we will present the fundamental KT models, based on our taxonomic framework. {Specifically, we will introduce these models in accordance with their development timeline. Subsequently, we will provide a summary of these fundamental KT models.}\",\n \"id\": \"bd30f2af-cabd-4e04-ab90-5536a2f60dde\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ]\n ],\n \"subsections\": [\n \"22232da7-baf6-4ee6-a966-d59421dce543\",\n \"737cf102-cd65-4a28-91e5-d7967b691341\",\n \"0b0df85b-c041-47ba-8171-c36927847515\",\n \"2737aa4f-b2f6-4977-83f3-36d33949ab77\"\n ],\n \"title\": \"Fundamental Knowledge Tracing Models\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Bayesian models assume that the learning process adheres to a Markov process. This process allows for the estimation of students' latent knowledge states based on their observed performance. In the subsequent section, we will present two Bayesian models in our taxonomy framework: the Bayesian Knowledge Tracing (BKT) and the Dynamic Bayesian Knowledge Tracing (DBKT).\",\n \"id\": \"22232da7-baf6-4ee6-a966-d59421dce543\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"bd30f2af-cabd-4e04-ab90-5536a2f60dde\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Bayesian Models\"\n ]\n ],\n \"subsections\": [\n \"9274e535-9216-4905-8c97-6e2b15501b96\",\n \"2d449595-4bbc-4a8e-8c76-39195e6d0bf8\"\n ],\n \"title\": \"Bayesian Models\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"{BKT's structure is illustrated in Fig. \\\\ref{fbkt}; here, the unshaded nodes represent unobservable latent knowledge states, while the shaded nodes represent the observable answers of the student.}\\nBKT is a unique instance of the Hidden Markov Model (HMM). There are two types of parameters in HMM: transition probabilities and emission probabilities. In BKT, the transition probabilities are defined by two learning parameters: (1) $P(T)$, the probability of transition from the unlearned state to the learned state; (2) $P(F)$, the probability of forgetting previously mastered knowledge. Moreover, the emission probabilities are determined by two performance parameters: (1) $P(G)$ - the probability that a student will guess correctly, despite non-mastery; (2) $P(S)$ - the probability that a student will make a mistake, despite mastery. Additionally, $P(L_0)$ represents the initial probability of mastery. BKT operates within a two-state student modeling framework: knowledge is either learned or unlearned, and there is no forgetting once a student has mastered the knowledge. Based on the observations of the student's learning interactions, the following equation is utilized to estimate the knowledge state and the probability of correct answers: \\n\\\\begin{equation}\\n\\t\\\\small\\n\\t\\\\begin{aligned}\\n\\t\\tP(L_n) &= P(L_{n}|Answer) + (1 - P(L_{n}|Answer))P(T), \\\\\\\\\\n\\t\\tP(C_{n+1}) &= P(L_n)(1 - P(S)) + (1 - P(L_n))P(G),\\n\\t\\\\end{aligned}\\n\\\\end{equation}\\nwhere $P(L_n)$ is the probability that a KC is mastered at the $n$-th learning interaction, $P(C_{n+1})$ is the probability of correct answers at the next learning interaction. $P(L_n)$ is the sum of two probabilities: (1) the probability that the KC is already mastered; (2) the probability that the knowledge state will convert to the mastered state. The posterior probability $P(L_{n}|Answer)$ is estimated as follows:\\n\\\\begin{equation}\\n\\t\\\\scriptsize\\n\\t\\\\begin{aligned}\\n\\t\\tP(L_{n}|correct)&=\\n\\t\\t\\\\frac{P(L_{n-1})(1 - P(S))}{P(L_{n-1})(1 - P(S)) + (1 - P(L_{n-1}))P(G)}, \\\\\\\\\\n\\t\\tP(L_{n}|incorrect)&=\\n\\t\\t\\\\frac{P(L_{n-1})P(S)}{P(L_{n-1})P(S) + (1 - P(L_{n-1}))(1 - P(G))}.\\n\\t\\\\end{aligned}\\n\\\\end{equation}\\n\\\\begin{figure}[t]\\n\\t\\\\vspace{-0.0cm}\\n\\t\\\\centerline{\\\\includegraphics[width=0.9\\\\columnwidth]{Figures/BKT_model.pdf}}\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\caption{The topology of Bayesian Knowledge Tracing . $K$ are the unobserved knowledge nodes, $A$ are the observed performance (answer) nodes, $P(L_0)$ represents the initial probability, $P(T)$ is the transition probability, $P(G)$ is the guessing probability and $P(S)$ is the slipping probability.}\\n\\t\\\\label{fbkt}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\",\n \"id\": \"9274e535-9216-4905-8c97-6e2b15501b96\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"22232da7-baf6-4ee6-a966-d59421dce543\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Bayesian Models\"\n ],\n [\n \"subsubsection\",\n \"Bayesian Knowledge Tracing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Bayesian Knowledge Tracing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{dynamic}\\nBKT independently models the parameters of each KC, employing a specific model for each KC. However, KCs are not completely independent, but rather hierarchical and closely related . Dynamic Bayesian networks are capable of jointly representing multiple skills within a single model, potentially enhancing the representational power of BKT. Consequently, propose dynamic Bayesian knowledge tracing (DBKT) to represent the hierarchies and relationships within KCs using dynamic Bayesian networks. This approach considers different KCs jointly within a single model. \\nIn DBKT, a student's knowledge mastery is represented by binary latent variables, which are estimated based on their learning interactions. Unlike BKT, DBKT takes into account the dependencies between various KCs. For example, if $KC_1$ and $KC_2$ are prerequisites for mastering $KC_3$, students' mastery of $KC_3$ depends on their mastery of $KC_1$ and $KC_2$.\\nLet $H$ denote the unobserved variables, i.e., lack of student answers and binary mastery variables. Assuming the student correctly answers an exercise associated with $KC_1$ at time step $t_1$, i.e., $a_{1,1} = 1$. The observed variables are then $a_m = a_{1,1}$ and the unobserved variables are $h_m = \\\\{KC_{1,1}, KC_{2,1}, KC_{3,1}, a_{2,1}, a_{3,1}\\\\}$. The objective of DBKT is to find the parameters $\\\\theta$ that maximize the joint probability $p(a_m, h_m|\\\\theta)$. The log-likelihood can alternatively be formulated using a log-linear model, as follows: \\n\\\\begin{equation}\\n\\tL(\\\\bm{w}) = \\\\sum_{m}ln(\\\\sum_{h_m}exp(\\\\bm{w}^T\\\\varPhi(a_m, h_m) - ln(Z))),\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{equation}\\nwhere $\\\\varPhi: A \\\\times H \\\\rightarrow \\\\mathbb{R}^{F}$ denotes a mapping from the observed space $A$ and the latent space $H$ to an $F$-dimensional feature vector. $Z$ is a normalizing constant, $\\\\bm{w}$ denotes the weights.\",\n \"id\": \"2d449595-4bbc-4a8e-8c76-39195e6d0bf8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"22232da7-baf6-4ee6-a966-d59421dce543\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Bayesian Models\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Bayesian Knowledge Tracing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Bayesian Knowledge Tracing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"{Logistic models represent the probability of students correctly answering exercises as a logistic function of the student and KC parameters.} They first use different factors in students' learning interactions to compute an estimation of the student and KC parameters, then utilize a logistic function to transform this estimation into the prediction of the probability of mastery .\\nIn the subsequent section, we will introduce three types of logistic models: (1) Learning Factor Analysis (LFA), (2) Performance Factor Analysis (PFA) and (3) Knowledge Tracing Machines (KTM).\",\n \"id\": \"737cf102-cd65-4a28-91e5-d7967b691341\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"bd30f2af-cabd-4e04-ab90-5536a2f60dde\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Logistic Models\"\n ]\n ],\n \"subsections\": [\n \"5d298d54-78eb-4db4-91fe-1e91f13c0cdc\",\n \"d181c1c4-a973-41d5-b53b-df201574371b\",\n \"609eb20d-a953-453f-b163-f5e296a7c090\"\n ],\n \"title\": \"Logistic Models\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The LFA model considers the following learning factors:\\n\\\\begin{itemize}\\n\\t\\\\item{Initial knowledge state}: parameter $\\\\alpha$ estimates the initial knowledge state of each student;\\n\\t\\\\item{Easiness of KCs}: parameter $\\\\beta$ captures the easiness of different KCs;\\n\\t\\\\item{Learning rate of KCs}: parameter $\\\\gamma$ denotes the learning rate of KCs.\\n\\\\end{itemize}\\nThe standard LFA model takes the following form:\\n\\\\vspace{-0.2cm}\\n\\\\begin{equation}\\n\\tp(\\\\theta) = \\\\sigma(\\\\sum_{i \\\\in N}\\\\alpha_iS_i\\t+ \\\\sum_{j \\\\in KCs}(\\\\beta_j + \\\\gamma_jT_j)K_j), \\n\\t\\\\vspace{-0.2cm}\\n\\\\end{equation}\\nwhere $\\\\sigma$ is the sigmoid function, $S_i$ is the covariate for the student $i$, $T_j$ represents the covariate for the number of interactions on KC $j$, $K_j$ is the covariate for KC $j$, $p(\\\\theta)$ is the estimation of the probability of a correct answer.\",\n \"id\": \"5d298d54-78eb-4db4-91fe-1e91f13c0cdc\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"737cf102-cd65-4a28-91e5-d7967b691341\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Logistic Models\"\n ],\n [\n \"subsubsection\",\n \"Learning Factor Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Learning Factor Analysis\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5377\n ],\n \"content\": \"The PFA model can be seen as an extension of the LFA model that is especially sensitive to the student performance. In contrast to the LFA model, PFA considers the following different factors:\\n\\\\begin{itemize}\\n\\t\\\\item{Previous failures}: parameter $f$ is the prior failures for the KC of the student;\\n\\t\\\\item{Previous successes}: parameter $s$ represents the prior successes for the KC of the student;\\n\\t\\\\item{Easiness of KCs}: parameter $\\\\beta$ means the easiness of different KCs, which is the same as in the LFA model.\\n\\\\end{itemize}\\nThe standard PFA model takes the following form:\\n\\\\vspace{-0.15cm}\\n\\\\begin{equation}\\n\\tp(\\\\theta) = \\\\sigma(\\\\sum_{j \\\\in KCs}(\\\\beta_j + \\\\mu_js_{ij} + \\\\nu_jf_{ij})),\\n\\t\\\\label{pfa}\\n\\t\\\\vspace{-0.15cm}\\n\\\\end{equation}\\nwhere $\\\\mu$ and $\\\\nu$ are the coefficients for $s$ and $f$, which denote the learning rates\\nfor successes and failures.\\n\\\\begin{figure}[t]\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\centerline{\\\\includegraphics[width= 0.9\\\\columnwidth]{Figures/KTM.pdf}}\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\caption{Example of activation of a knowledge tracing machine . $V$ refers to the matrix of embeddings, $w$ refers to the vector of biases, $x$ is the encoding vector of the learning interaction.}\\n\\t\\\\label{fktm}\\n\\t\\\\vspace{-0.6cm}\\n\\\\end{figure}\",\n \"id\": \"d181c1c4-a973-41d5-b53b-df201574371b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"737cf102-cd65-4a28-91e5-d7967b691341\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Logistic Models\"\n ],\n [\n \"subsubsection\",\n \"Performance Factor Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Performance Factor Analysis\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 5377\n ],\n \"content\": \"The KTM model, developed by Vie et al. , employs factorization machines (FMs) to generalize logistic models to higher dimensions. FMs were initially introduced as a general predictor capable of working with any real-valued feature vector, enabling the model to represent all interactions between variables using factorized parameters . FMs provide a means of encoding side information about exercises or students into the model. Figure \\\\ref{fktm} illustrates the example of KTM, which models the knowledge mastery of the student based on a sparse set of weights for all features involved in the event. Let $L$ be the number of features; here, the features can be related to students, exercises, KCs, or any other side information. The learning interaction is encoded by a sparse vector $\\\\bm{l}$ of length $L$. When feature $i$ is involved in the interaction, $l_i > 0$. The probability $p(\\\\theta)$ of the correct answer is determined by the following equations: \\n\\\\vspace{-0.3cm}\\n\\\\begin{equation}\\n\\tp(\\\\theta) = \\\\sigma(\\\\mu + \\\\sum_{i = 1}^{L}w_il_i + \\\\sum_{1 \\\\leq i < j \\\\leq L}l_il_j\\\\langle\\\\bm{v_i}, \\\\bm{v_j}\\\\rangle ),\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{equation}\\nwhere $\\\\mu$ is the global bias, the feature $i$ is modeled by the bias $w_i \\\\in \\\\bm{R}$ and the embedding $\\\\bm{v}_i \\\\in \\\\bm{R}^d$ ($d$ is the dimension). Note that only features with $l_i > 0$ will have impacts on the predictions.\",\n \"id\": \"609eb20d-a953-453f-b163-f5e296a7c090\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"737cf102-cd65-4a28-91e5-d7967b691341\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Logistic Models\"\n ],\n [\n \"subsubsection\",\n \"Knowledge Tracing Machines\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Tracing Machines\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 5378,\n 5371\n ],\n \"content\": \"\\\\label{deep}\\nThe cognitive process can be influenced by various factors at both the macro and micro levels. It is difficult for Bayesian models or logistic models to adequately capture a cognitive process of high complexity .\\nDeep learning, with its potent ability to achieve non-linearity and feature extraction, is well-suited for modeling complex learning processes, particularly when a significant amount of learning interaction data is available . {In recent years, numerous research works have been proposed on deep learning KT models, we will introduce deep learning models from four sub-categories: (1) deep knowledge tracing, (2) memory-aware knowledge tracing, (3) attentive knowledge tracing, and (4) graph-based knowledge tracing.}\",\n \"id\": \"0b0df85b-c041-47ba-8171-c36927847515\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"bd30f2af-cabd-4e04-ab90-5536a2f60dde\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Deep Learning Models\"\n ]\n ],\n \"subsections\": [\n \"6bda0552-a0c8-4448-bccb-783dc9dbe658\",\n \"a1cbba76-ad36-4969-8878-16538f12140f\",\n \"b60bbb0c-acf3-4bdd-ace6-fb43306d7dda\",\n \"9d414c8f-554a-43a7-afe7-209c3c4d813e\"\n ],\n \"title\": \"Deep Learning Models\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 5378,\n 5379,\n 5371\n ],\n \"content\": \"Deep Knowledge Tracing (DKT) is the pioneering approach that introduces deep learning {to complete the KT task. DKT employs Recurrent Neural Networks (RNNs) to process the input sequence of learning interactions over time,} maintaining a hidden state that implicitly contains information about the history of all past elements of the sequence. This hidden state evolves based on both the previous knowledge state and the present input learning interaction . DKT provides a high-dimensional and continuous representation of the knowledge state, enabling it to more effectively model the complex learning process. Typically, RNNs' variant, the Long Short-Term Memory (LSTM) networks , are more frequently used in the implementation of DKT, which is further strengthened by considering forgetting. \\nFig. \\\\ref{fdkt} illustrates the process of deep knowledge tracing. In DKT, exercises are represented by their contained KCs. For datasets with different numbers of KCs, DKT applies two different methods to convert students' learning interactions $\\\\bm{X} = \\\\{(e_1, a_1), (e_2, a_2), ..., (e_t, a_t), ..., (e_N, a_N)\\\\}$ into a sequence of fixed-length input vectors. More specifically, for datasets with a small number $K$ of unique KCs, $\\\\bm{x}_t \\\\in \\\\{0,1\\\\}^{2K}$ is set as a one-hot embedding, where $ \\\\bm{x}_t^k = 1 $ if the answer $a_t$ of the exercise with KC $k$ was correct or $\\\\bm{x}_t^{k + K} = 1$ if the answer was incorrect. For datasets with a large number of unique KCs, one-hot embeddings are considered too sparse. Therefore, DKT assigns each input vector $\\\\bm{x}_t$ to a corresponding random vector, and then uses the embedded learning sequence as the input of RNNs. A linear mapping and activation function are then applied to the output hidden states to obtain the knowledge state of students:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\\\label{dkt}\\n\\t\\t&\\\\bm{h}_t = tanh(\\\\bm{W}_{hs}\\\\bm{x}_t + \\\\bm{W}_{hh}\\\\bm{h}_{t-1} + \\\\bm{b}_h), \\\\\\\\\\n\\t\\t&\\\\bm{y}_t = \\\\sigma (\\\\bm{W}_{yh}\\\\bm{h}_{t} + \\\\bm{b}_y),\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $tanh$ is the activation function, $\\\\bm{W}_{hs}$ is the input weights, $\\\\bm{W}_{hh}$ is the recurrent weights, $\\\\bm{W}_{yh}$ is the readout weights, and $\\\\bm{b}_h$ and $\\\\bm{b}_y$ are the bias terms.\\n\\\\begin{figure}[t]\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\centerline{\\\\includegraphics[width=0.9\\\\columnwidth]{Figures/DKT_model.pdf}}\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\caption{The architecture of DKT .}\\n\\t\\\\label{fdkt}\\n\\t\\\\vspace{-0.6cm}\\n\\\\end{figure}\\nDespite demonstrating superior performance compared to Bayesian and logistic models, DKT has several inherent shortcomings. For instance, the lack of interpretability is a significant drawback. It is challenging to understand how the hidden states represent students' knowledge states, and the model cannot explicitly determine a student's knowledge mastery from the hidden state . Additionally, identified two unreasonable phenomena in DKT that contravene common sense. These are: (1) the inability to reconstruct observed input, and (2) inconsistent predicted knowledge states across time-steps. However, despite these shortcomings, DKT remains a promising KT model .\",\n \"id\": \"6bda0552-a0c8-4448-bccb-783dc9dbe658\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"0b0df85b-c041-47ba-8171-c36927847515\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Deep Learning Models\"\n ],\n [\n \"subsubsection\",\n \"Deep Knowledge Tracing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Knowledge Tracing\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 5380,\n 1206\n ],\n \"content\": \"To enhance the interpretability of DKT, memory-aware knowledge tracing introduces an external memory module, as proposed by . This module is designed to {store the and update the corresponding knowledge mastery of the student.} The most representative example is Dynamic Key-Value Memory Networks (DKVMN) for knowledge tracing, as proposed by . DKVMN highlights students' specific knowledge states on various knowledge categories. It initializes a static matrix, referred to as a $key$ matrix to store latent KCs and a dynamic matrix, called a $value$ matrix to store and update the mastery of corresponding KCs through read and write operations over time.\\nAs shown in Fig. \\\\ref{fdkvmn}, an embedding matrix is first defined to obtain the embedding vector $k_t$ of the exercises. A correlation weight $\\\\bm{w}_t$ is then obtained by taking the inner product between the exercise embedding $k_t$ and the $key$ vectors $M^k$, followed by the softmax activation:\\n\\\\vspace{-0.2cm}\\n\\\\begin{equation}\\n\\t\\\\bm{w}_t = Softmax(k_tM^k),\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{equation}\\nwhere the correlation weight $\\\\bm{w}_t$ represents the correlation between the exercises and all latent KCs.\\nIn the read operation, DKVMN predicts student performance based on the student's knowledge mastery. Specifically, DKVMN reads students’ mastery of the exercise $\\\\bm{r}_t$ with reference to the weighted sum of all memory vectors in the $value$ matrix using the correlation weight. The read content and the input exercise embeddings are then concatenated together and passed to a fully connected layer to yield a summary vector $\\\\bm{f}_t$, which contains both the student's knowledge mastery and the prior difficulty of the exercise. Furthermore, the student's performance can be predicted by applying another fully connected layer with a sigmoid activation function to the summary vector:\\n\\\\vspace{-0.2cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\t&\\\\bm{r}_t = \\\\sum_{i=1}^{N}w_t(i)M_t^v(i), \\\\\\\\\\n\\t\\t&\\\\bm{f}_t = tanh(\\\\bm{W}_f[\\\\bm{r}_t, k_t] + \\\\bm{b}_f),\\\\\\\\\\n\\t\\t&p_t = \\\\sigma(\\\\bm{W}_p\\\\bm{f}_t + \\\\bm{b}_p),\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $\\\\bm{W}_f$ and $\\\\bm{W}_p$ are the weights, $\\\\bm{b}_f$ and $\\\\bm{b}_p$ are bias terms.\\nIn the write operation, after an exercise has been answered, DKVMN updates students' knowledge mastery (i.e., the $value$ matrix) based on their performance. Specifically, the learning interaction $(e_t, a_t)$ is first embedded with an embedding matrix $\\\\bm{B}$ to obtain the student's knowledge growth $\\\\bm{v}_t$. Then DKVMN calculates an erase vector $\\\\bm{erase}_t$ from $\\\\bm{v}_t$ and decides to erase the previous memory with reference to both the erase vector and the correlation weight $\\\\bm{w}_t$.\\nFollowing erasure, the new memory vectors are updated by the new knowledge state and the add vector $\\\\bm{add}_t$, which forms an $erase$-followed-by-$add$ mechanism that allows forgetting and strengthening knowledge mastery in the learning process:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\t&\\\\bm{erase}_t = \\\\sigma(\\\\bm{W}_e\\\\bm{v}_t + \\\\bm{b}_e), \\\\\\\\\\n\\t\\t&\\\\widetilde{M}_t^v(i) = M_{t-1}^v(i)[1 - w_t(i)\\\\bm{erase}_t],\\\\\\\\\\n\\t\\t&\\\\bm{add}_t = tanh(\\\\bm{W}_d\\\\bm{v}_t + \\\\bm{b}_d),\\\\\\\\\\n\\t\\t&M_{t}^v(i) = \\\\widetilde{M}_t^v(i) + w_t(i)\\\\bm{add}_t,\\\\\\\\\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $\\\\bm{W}_e$ and $\\\\bm{W}_d$ are the weights, $\\\\bm{b}_e$ and $\\\\bm{b}_d$ are bias terms.\\n point out that DKVMN failed to capture long-term dependencies in the learning process. Therefore, they propose a Sequential Key-Value Memory Network (SKVMN) to combine the strengths of DKT's recurrent modelling capacity and DKVMN's memory capacity. In SKVMN, a modified LSTM called \\\\emph{Hop-LSTM} is used to hop across LSTM cells according to the relevance of the latent KCs, which directly captures the long-term dependencies. During the writing process, SKVMN allows for the calculation of the knowledge growth of a new exercise, taking into consideration the current knowledge state, thereby yielding more reasonable results. \\n\\\\begin{figure}[t]\\n\\t\\\\centerline{\\\\includegraphics[width=0.9\\\\columnwidth]{Figures/DKVMN.pdf}}\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\caption{The architecture of DKVMN .}\\n\\t\\\\label{fdkvmn}\\n\\t\\\\vspace{-0.6cm}\\n\\\\end{figure}\",\n \"id\": \"a1cbba76-ad36-4969-8878-16538f12140f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"0b0df85b-c041-47ba-8171-c36927847515\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Deep Learning Models\"\n ],\n [\n \"subsubsection\",\n \"Memory-aware Knowledge Tracing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Memory-aware Knowledge Tracing\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 679,\n 5376,\n 5374,\n 5375,\n 5370,\n 38\n ],\n \"content\": \"\\\\label{attentive}\\nIn the development of deep learning, the Transformer is initially proposed for neural machine translation , which abandons recurrence and solely relies on the self-attention mechanism to capture global dependencies within a sequence. The Transformer has been demonstrated to excel in feature extraction and dependency capture, while maintaining high computational efficiency. Some representative pre-training models based on the Transformer, such as BERT and GPT , have obtained state-of-the-art results on various natural language processing tasks. propose a self-attentive model for knowledge tracing (SAKT), which directly apply the Transformer to capture long-term dependencies between students' learning interactions. Furthermore, introduce an adaptive sparse self-attention network to generate missing features and simultaneously produce fine-grained predictions of student performance. employ a multi-head ProbSparse self-attention mechanism to mitigate the time complexity and effectively capture the long-term dependencies in students' learning interactions. \\nHowever, the complexity of the KT task often limits the performance of the aforementioned simple Transformer applications. introduce a novel approach named Separated Self-Attentive Neural Knowledge Tracing (SAINT) to enhance self-attentive computation for KT adaptability. Specifically, SAINT employs an encoder-decoder structure, with the exercise and answer embeddings being separately encoded and decoded by self-attention layers. The separation of the input allows SAINT to stack self-attention layers multiple times, thus capturing complex relations in student interactions. Subsequently, introduce the SAINT+ model, which integrates two temporal features into SAINT: namely, the time taken to answer each exercise and the interval time between consecutive learning interactions. Both SAINT and SAINT+ have outperformed the SAKT model on the student performance prediction task. \\nAdditionally, observe that SAKT does not surpass DKT and DKVMN in their experiments. Unlike SAINT and SAINT+, they present a context-aware attentive knowledge tracing (AKT) model. This model integrates the self-attention mechanism with psychometric models, creating a more effective system. AKT is composed of four modules: Rasch model-based embeddings, exercise encoder, knowledge encoder, and knowledge retriever. Specifically, the embedding module employs the classic Rasch model in psychometrics to construct embeddings for exercises and KCs:\\n\\\\vspace{-0.2cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\t\\\\bm{x}_t = \\\\bm{c}_{c_t} + \\\\mu_{e_t} \\\\cdot \\\\bm{d}_{c_t}, \\\\\\\\\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{equation}\\nwhere $\\\\bm{c}_{c_t} \\\\in \\\\mathbb{R}^{\\\\bm{D}}$ is the embedding of the KC of this exercise, $\\\\bm{d}_{c_t} \\\\in \\\\mathbb{R}^{\\\\bm{D}}$ is a vector that summarizes the variation in exercises with the related KC, and $\\\\mu_{e_t} \\\\in \\\\mathbb{R}^{\\\\bm{D}}$ is a scalar difficulty parameter that controls the extent to which this exercise deviates from the related KC.\\nThe exercise-answer tuple $(e_t, a_t)$ is similarly extended using the scalar difficulty parameter for each pair:\\n\\\\vspace{-0.2cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\t\\\\bm{y}_t = \\\\bm{q}_{(c_t, a_t)} + \\\\mu_{e_t} \\\\cdot \\\\bm{f}_{(c_t, a_t)}, \\\\\\\\\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{equation}\\nwhere $\\\\bm{q}_{(c_t, a_t)} \\\\in \\\\mathbb{R}^{\\\\bm{D}}$ is the KC-answer embedding, $\\\\bm{f}_{(c_t, a_t)} \\\\in \\\\mathbb{R}^{\\\\bm{D}}$ is the variation vector. Through the above embedding, exercises labeled as the same KCs are determined to be closely related while retaining important individual characteristics.\\nThen, in the exercise encoder, the input is the exercise embeddings $ \\\\{\\\\bm{e}_1, . . . , \\\\bm{e}_t \\\\} $ and the output is a sequence of context-aware exercise embeddings $\\\\{\\\\widetilde{\\\\bm{e}}_1, . . ., \\\\widetilde{\\\\bm{e}}_t\\\\}$. AKT designs a monotonic attention mechanism to accomplish the above process, where the context-aware embedding of each exercise depends on both itself and the previous exercises, i.e., $\\\\widetilde{\\\\bm{e}}_t = f_{enc_1}(\\\\bm{e}_1, . . . , \\\\bm{e}_t)$. Similarly, the knowledge encoder takes exercise-answer embeddings $ \\\\{\\\\bm{y}_1, . . . , \\\\bm{y}_t \\\\} $ as input and\\noutputs a sequence of context-aware embeddings of the knowledge acquisitions $\\\\{\\\\widetilde{\\\\bm{y}}_1, . . ., \\\\widetilde{\\\\bm{y}}_t\\\\}$ using the same monotonic attention mechanism; these are also determined by students' answers to both the current exercise and prior exercises, i.e., $\\\\widetilde{\\\\bm{y}}_t = f_{enc_1}(\\\\bm{y}_1, . . . , \\\\bm{y}_t)$.\\nFinally, the knowledge retriever takes the context-aware exercise embedding $\\\\widetilde{\\\\bm{e}}_{1:t}$ and exercise-answer pair embeddings $\\\\widetilde{\\\\bm{y}}_{1:t}$ as input and outputs a retrieved knowledge state $\\\\bm{h}_t$ for the current exercise. Since the student’s current knowledge state depends on answering the related exercise, it is also context-aware in AKT.\\nThe novel monotonic attention mechanism proposed in AKT is based on the assumption that the learning process is temporal and students' knowledge will decay over time. Therefore, the scaled inner-product attention mechanism utilized in the original Transformer is not suitable for the KT task. AKT uses exponential decay and a context-aware relative distance measure to compute the attention weights. Finally, AKT achieves outstanding performance in predicting students' future answers, as well as demonstrating interpretability due to the combination of the psychometric model.\\nIt is important to note that have recently proposed that attentive knowledge tracing models significantly benefit from students' continuous, repeated interactions on the same exercises throughout the learning process. In their experiments, the removal of these repeated interactions in the dataset led to a decline in AKT's performance, bringing it close to that of DKVMN. \\n{Moreover, according to the findings of ,} existing attentive attentive KT models primarily trace patterns of a learner's learning activities, rather than their evolving knowledge states. Consequently, they develop the DTransformer model to facilitate stable knowledge state estimation and tracing, rather than solely focusing on next performance prediction. \\n\\\\begin{figure}[t]\\n\\t\\\\centerline{\\\\includegraphics[width=0.9\\\\columnwidth]{Figures/gkt.pdf}}\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\caption{The architecture of graph-based knowledge tracing .}\\n\\t\\\\label{fgkt}\\n\\t\\\\vspace{-0.3cm}\\n\\\\end{figure}\",\n \"id\": \"b60bbb0c-acf3-4bdd-ace6-fb43306d7dda\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"0b0df85b-c041-47ba-8171-c36927847515\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Deep Learning Models\"\n ],\n [\n \"subsubsection\",\n \"Attentive Knowledge Tracing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attentive Knowledge Tracing\"\n },\n {\n \"cite_extract_rate\": 0.28571428571428503,\n \"cites\": [\n 243,\n 553\n ],\n \"content\": \"\\\\label{graph}\\nGraph neural networks (GNNs), which are designed to handle complex graph-related data, have developed rapidly in recent years . The graph represents a kind of data structure that models a set of objects (nodes) and their relationships (edges). From a data structure perspective, there is a naturally existing graph structure within the KCs. Therefore, incorporating the graph structure of the KCs as additional information should be beneficial to the KT task.\\n presented graph-based knowledge tracing (GKT), which conceptualizes the potential graph structure of the KCs as a graph $G = (V, E)$, where nodes $V = \\\\{v_1, v_2, ..., v_N\\\\}$ represent the set of KCs and the edges $E \\\\subseteq V \\\\times V $ represent relationships of these KCs; moreover, $ \\\\bm{h}^t = \\\\{\\\\bm{h}^t_{i \\\\in V}\\\\} $ represents the student's temporal knowledge state after answering the exercise at time $t$.\\nThe architecture for graph-based knowledge tracing is presented in Figure \\\\ref{fgkt}, which is composed of three parts: (1) $aggregate$, (2) $update$ and (3) $predict$.\\nIn the $aggregate$ module, GKT aggregates the temporal knowledge state and the embedding for the answered KC $i$ and its neighboring KC $j$:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\bm{h}_k^{'t}=\\\\left\\\\{\\n\\t\\\\begin{aligned}\\n\\t\\t&[\\\\bm{h}_k^{t}, a^t\\\\bm{E}_s] &(k = i), \\\\\\\\\\n\\t\\t&[\\\\bm{h}_k^{t}, \\\\bm{E}_e(k)] &(k \\\\neq i),\\n\\t\\\\end{aligned}\\n\\t\\\\right.\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $a^t$ represents the exercises answered correctly or incorrectly at time step $t$, $\\\\bm{E}_s$ is the embedding matrix for the learning interactions, $\\\\bm{E}_e$ is the embedding matrix for the KC, and $k$ represents the $k$-th row of $\\\\bm{E}_e$.\\nIn the $update$ module, GKT updates the temporal knowledge state based on the aggregated features and the knowledge graph structure, as follows:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\t&\\\\bm{m}_k^{t+1}=\\n\\t\\t\\\\begin{cases}\\n\\t\\t\\t&f_{self}(\\\\bm{h}_k^{'t}) (k = i), \\\\\\\\\\n\\t\\t\\t&f_{neighbor}(\\\\bm{h}_i^{'t}, \\\\bm{h}_k^{'t}) (k \\\\neq i),\\\\\\\\\\n\\t\\t\\\\end{cases}\\\\\\\\\\n\\t\\t&\\\\widetilde{\\\\bm{m}}_k^{t+1} = G_{ea}(\\\\bm{m}_k^{t+1}), \\\\\\\\\\n\\t\\t&\\\\bm{h}_{k}^{t+1} = G_{gru}(\\\\widetilde{\\\\bm{m}}_k^{t+1}, \\\\bm{h}_{k}^{t}), \\\\\\\\\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $f_{self}$ is the multilayer perceptron, $G_{ea}$ is the same $erase$-followed-by-$add$ mechanism used in DKVMN, and $G_{gru}$ is the gated recurrent unit (GRU) gate . Moreover, $f_{neighbor}$ defines the information propagation to neighboring nodes based on the knowledge graph structure.\\nIn the $predict$ module, GKT predicts the student's performance at the next time step according to the updated temporal knowledge state:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\ty_k^t = \\\\sigma(\\\\bm{W}_k\\\\bm{h}_k^{t+1} + \\\\bm{b}_k),\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $\\\\bm{W}_k$ is the weight parameter and $\\\\bm{b}_k$ is the bias term.\\nIn addition to modeling the graph structure in KCs by graph neural networks, propose to model the educational relation and topology in the concept map, which will be intended to act as mathematical constraints for the construction of the KT model.\\nRecently, in the attempt to further explore knowledge structure, propose structure-based knowledge tracing (SKT), which aims to capture the multiple relations in knowledge structure to model the influence propagation among concepts. SKT is mainly motivated by an education theory, \\\\emph{transfer of knowledge} , which claims that students' knowledge states on some relevant KCs will also be changed when they are practicing on a specific KC due to the potential knowledge structure among KCs. Therefore, a student's knowledge state is determined by not only the temporal effect from the exercise sequence, but also the spatial effect from the knowledge structure. To concurrently model the latent spatial effects, SKT presents the synchronization and partial propagation methods to characterize the undirected and directed relations between KCs, respectively. In this way, SKT measures influence propagation in the knowledge structure with both temporal and spatial relations. To get rid of dependence on knowledge structure, propose the Automatical Graph-based Knowledge Tracing (AGKT), which utilizes the automatical\\ngraph to measure students’ knowledge states automatically without annotation manual annotations.\",\n \"id\": \"9d414c8f-554a-43a7-afe7-209c3c4d813e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"0b0df85b-c041-47ba-8171-c36927847515\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Deep Learning Models\"\n ],\n [\n \"subsubsection\",\n \"Graph-based Knowledge Tracing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Graph-based Knowledge Tracing\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 5381\n ],\n \"content\": \"{It is crucial to emphasize that, despite deep learning models exhibiting superior performance compared to Bayesian models and logistic models, there remains significant room for improvement in their interpretability and explainability. \\nDue to the end-to-end learning strategy, deep learning models are notoriously difficult to interpret. The modeling process itself is also challenging to explain. Specifically, deep learning models are predominantly data-driven and benefit a lot from large-scale student learning data. It is challenging to understand how they calculate a student's knowledge state with no theoretical guidance . \\nAll we have are the results generated by these models. Therefore, any errors made by these models will lead students to doubt their reliability. The lack of explainability and interpretability has thus limited their further applicability. }\\n{To make the complex KT models interpretable, especially those deep learning models, researchers have attempted various methods.\\n presented a post-hoc approach to reveal the interpretability of DKT. Specifically, they employed the layer-wise relevance propagation (LRP) technique to interpret DKT by measuring the relevance between DKT's output and input. Preliminary experimental results suggest that this post-hoc approach could be a promising method for explaining DKT. Besides, explainable AI (xAI) is proposed to make the black-box deep learning models more transparent, thereby promoting its applications . proposed to use the xAI technique to interpret the complex KT models based on deep learning. The interpreting results of DKT have been demonstrated to aid in enhancing the trust of students and teachers. Their findings suggested that it is promising to utilize xAI techniques to interpret the deep learning KT models, thereby assisting users in accepting and applying the suggestions provided by these models. \\n}\",\n \"id\": \"2737aa4f-b2f6-4977-83f3-36d33949ab77\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"bd30f2af-cabd-4e04-ab90-5536a2f60dde\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Fundamental Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Summarization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summarization\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:variants}\\n{The aforementioned fundamental KT models are} typically based on simplistic assumptions regarding the learning process. Specifically, these models predominantly employ learning interactions such as exercises and responses to estimate students' knowledge states. However, the learning process is not solely represented by exercises and responses, but is influenced by various factors. In summary, the aforementioned fundamental KT models, while straightforward, may have reduced performance in real-world learning scenarios. Consequently, numerous variants have been proposed under more stringent assumptions, reflecting a more comprehensive learning process in real-world scenarios. Accordingly, we classify and review current variants of fundamental KT models into four categories: (1) Modeling individualization before learning, (2) Incorporating engagement \\\\ during learning, (3) Considering forgetting after learning, and (4) utilizing side information across learning.\",\n \"id\": \"a705e159-25ef-44ad-8166-1cef83a7c265\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Variants of Knowledge Tracing Models\"\n ]\n ],\n \"subsections\": [\n \"242f5268-6395-4e4d-b14a-e118bc61d29b\",\n \"b2eec265-034c-4c04-b806-03ebaf934ee5\",\n \"2cb054ef-697a-4d85-a2b7-9c0e6282362f\",\n \"60b4234d-a28f-4026-ab3b-0291b62ceca6\"\n ],\n \"title\": \"Variants of Knowledge Tracing Models\"\n },\n {\n \"cite_extract_rate\": 0.14285714285714202,\n \"cites\": [\n 5382,\n 166\n ],\n \"content\": \"Everything and everyone possess unique characteristics. For instance, explored several personalized factors of various tourists to recommend personalized travel packages. Similarly, the concept of individualization in the KT task implies that different students often exhibit different learning characteristics (such as varying learning rates or prior knowledge). Considering the student-specific variability in learning could potentially enhance the KT process, as suggested by . In the subsequent sections, we will introduce various variant KT models that {take into account individualization before learning.}\\nThe initial BKT paper has delved into the concept of individualization. Specifically, it uses all students' learning interactions on a specific KC to learn the individual parameter. Similarly, for a specific student, all her learning interactions are utilized to fit her individual learning parameters . Consequently, BKT is able to ascertain different learning and performance parameters for various students and KCs. However, this approach only offers a marginal improvement compared to original BKT. \\nSubsequently, propose two simple variants of BKT that respectively individualize students' initial probability of mastery and the probability of transition from the unlearned state to the learned state. {Specifically, a student node is added to individualize the initial probability of mastery for each student. The student node assigns each student with a personalized initial probability of mastery. A conditional probability table is designed to determine the value of the student node. Similarly, if changing the connection of the student node to the subsequent knowledge nodes, the transition probability parameter can also be individualized. In this case, the student node gives individualized transition parameters to each student.}\\nMoreover, rather than individualizing only one kind of parameter in BKT, some other variants of BKT opt to individualize all four BKT parameters simultaneously . suggest that when applied in an intelligent tutoring system, the individualized BKT model can yield good improvements to student learning efficacy, reducing by about half the amount of questions required for 20\\\\% of students to achieve mastery.\\nAnother means of modeling individualization is clustering, which considers a wider range of students in different groups . \\nBy clustering the students into various groups, we can train different KT models and make predictions on the test data. The number of clusters is then varied according to the student groups and the predicting process is repeated iteratively. Finally, we can obtain a set of different predictions. Furthermore, there are two common methods used to combine these predictions : (1) uniform averaging, which simply averages the predictions; (2) weighted averaging, which combines the models by means of a weighted average.\\nTo realize clustering, K-means is a basic clustering algorithm that randomly initializes a set of cluster centroids, which are identified using Euclidean distance. Another popular clustering algorithm is spectral clustering, which represents the data as an undirected graph and analyzes the spectrum of the graph Laplacian obtained from the pairwise similarities of data points. Recently, some novel clustering algorithms have been proposed, including discrete nonnegative spectral clustering and clustering uncertain data .\\n propose a model named deep knowledge tracing with dynamic student classification (DKT-DSC), which introduces individualization to DKT by exploiting the idea of clustering. \\nAccording to students' previous performance, DKT-DSC assigns students with similar learning ability to the same group . The knowledge states of students in different groups are then traced by different DKT models. Moreover, considering the dynamic property of the learning ability, each student's learning sequence is segmented into multiple time intervals. At the start of each time interval, DKT-DSC will reassess students' learning ability and reassign their groups. \\nIn DKT-DSC, the K-means clustering algorithm is utilized to split students with similar ability levels into the same group at each time interval. After learning the centroids of all K clusters, each student is assigned to the nearest cluster. Through dynamic student clustering, DKT-DSC offers an effective approach to realizing individualization in DKT.\\n claim that it is significant to consider both individual exercise representation and individual prior knowledge. They propose a fine-grained knowledge tracing model, named FGKT. FGKT obtains the individual exercise representation through the acquisition of knowledge cells (KCs) and exercise distinctions. Subsequently, it assesses the individual prior knowledge by evaluating the relevance between current and historical learning interactions. Finally, the above individual representations will be utilized as the input of LSTM in FGKT to evaluate students' evolving knowledge states. also notice that the individualization of exercises is significant for measuring students' knowledge states. They propose to consider multiple exercise factors, including the difficulty\\nand the discrimination, to enhance the performance of DKT.\\n propose a convolutional knowledge tracing model (CKT) to implicitly measure student individualization. Specifically, CKT considers two factors that influence students' individualization: individualized learning rates and individualized prior knowledge.\\nIndividualized learning rates represent students' differing capacities to absorb knowledge. The sequence of student learning interactions can reflect different learning rates in the sense that students with high learning rates can rapidly master knowledge, while others need to spend more time trying and failing. Therefore, it is reasonable to assess the differences in learning rate by simultaneously processing several continuous learning interactions within a sliding window of convolutional neural networks . Besides, individualized prior knowledge refers to students' prior knowledge, which can be assessed via their historical learning interactions.\",\n \"id\": \"242f5268-6395-4e4d-b14a-e118bc61d29b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"a705e159-25ef-44ad-8166-1cef83a7c265\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Variants of Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Modeling Individualization before Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Modeling Individualization before Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Student engagement is defined as \\\\emph{\\\"the quality of effort students themselves devote to educationally purposeful activities that contribute directly to desired outcomes\\\"} . This definition highlights a strong connection {between student engagement and the learning process.} Generally, higher engagement leads to enhanced knowledge gains. Consequently, considering student engagement in the learning process could potentially improve knowledge tracing results . In this section, {we will present some variants that integrate student engagement into the KT models. }\\nStudent engagement is difficult to be directly measured. In practice, some online learning systems have made use of sensor data to measure student engagement. For example, inexpensive portable electroencephalography (EEG) devices can help to detect a variety of student mental states in learning, which can be seen as reflections of student engagement . propose two methods that combine EEG-measured mental states to improve the performance of BKT. Concretely, the first one inserts a one-dimensional binary EEG measure into BKT, forming the EEG-BKT structure that extends BKT by adding a binary variable node $E$ between the knowledge node and the answer node. The second one, i.e., EEG-LRKT, utilizes logistic regression to combine an $m$-dimensional continuous variable $E$ extracted from the raw EEG signal in BKT.\\nHowever, in most cases, it is difficult to collect sensor data on every student. Therefore, propose the knowledge and affect tracing (KAT) to model both knowledge and engagement in parallel. KAT is a sensorless model that does not rely on any sensor data. In this model, both knowledge and engagement are assumed to have direct influences on student performance. KAT considers three kinds of disengagement behaviors: quick guess (the student makes an attempt very quickly), bottom-out hint (all available hints are used) and many attempts (making more than three attempts at an exercise). These three behaviors are grouped as “gaming” behaviors in order to predict students' knowledge and engagement at each learning interaction. \\nRather than assuming equal influence of knowledge and engagement on students' knowledge state, one variation on the KAT model defines the connection between knowledge and engagement, and accordingly considers that students' knowledge states will influence their engagement. For example, students are more likely to disengage from knowledge they are not familiar with.\\nMoreover, rather than explicitly modeling student engagement, further propose the knowledge tracing with behavior (KTB) model, which has only one latent knowledge node that acts as a combination of both knowledge and engagement. KTB assumes that both engagement and performance are expressions of knowledge. The Bayesian estimation of the knowledge state needs to be inferred by both student engagement and performance.\\n propose to add five features in the process of watching videos on MOOCs to the input of DKT. These features reflect student engagement from various aspects, including playback speed, whether or not the video was paused, fast-forwarded or rewound, and whether or not the video was completed. For example, if a student watches a video at a much faster playback speed, it is likely that he/she is impatient and absent-minded. This model incorporates two further features: whether or not the exercise was submitted with an answer selected and whether or not the exercise was a part of an end-of-unit quiz, both of which are considered together. Experimental results indicate that DKT can achieve better performance through incorporating the above binarized engagement covariates.\",\n \"id\": \"b2eec265-034c-4c04-b806-03ebaf934ee5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"a705e159-25ef-44ad-8166-1cef83a7c265\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Variants of Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Incorporating Engagement during Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Incorporating Engagement during Learning\"\n },\n {\n \"cite_extract_rate\": 0.08333333333333301,\n \"cites\": [\n 5378\n ],\n \"content\": \"In real-world scenarios, while learning, forgetting is inevitable . The \\\\emph{Ebbinghaus forgetting curve theory} indicates that students' knowledge proficiency will decline due to forgetting . Recently, proposed the concept of 'Knowledge Proficiency Tracing' (KPT), a model that can dynamically capture the changes in students' proficiency levels on knowledge concepts over time. This model effectively tracks these changes in an interpretable manner. Therefore, the assumption that students' knowledge states will remain constant over time is untenable. However, fundamental KT models, such as the BKT, often overlook forgetting. In the following, we will introduce some variants of fundamental KT models that have attempted to consider forgetting after learning for more precise knowledge states. \\n discover that BKT consistently overestimates the accuracy of students' answers when a day or more had elapsed since {their previous responses.} The underlying reason is that BKT assumes that student performance will remain the same regardless of how much time has passed. To consider how student performance declines with time, they propose a BKT-Forget model, which hypothesizes that students may forget information they have learned as days go by. In the BKT-Forget model, a time node is added to specify which parameters should be affected by a new day and the new day node is fixed with a prior probability of 0.2. {It also introduced parameters to represent the forgetting rate on a new day and denote the forgetting rate on the same day.} However, although BKT-forget does consider the decline in student performance, it can only model forgetting that occurs over the time scale of days. \\nTo model the continuous decay of knowledge as time progresses, incorporate forgetting into BKT based on the assumption that learned knowledge decays exponentially over time . An exponential decay function is thus utilized to update the knowledge mastery level. They further assumed that the chance of forgetting will increase if a student does not practice the knowledge concepts within 30 days.\\nMoreover, introduce an approach that counts the number of intervening trials and treats each as an independent opportunity for forgetting to occur.\\nRecall the PFA model in Eq.(\\\\ref{pfa}), in which the probability of students' mastery is estimated using a logistic function: $ p(\\\\theta) = \\\\sigma(\\\\beta + \\\\mu s + \\\\nu f)$. The original PFA model ignores the order of answers, in addition to the time between learning interactions. It is therefore difficult to directly incorporate time information into the original PFA model.\\n propose PFAE (PFA Elo/Extended), a variant of the PFA model that combines PFA with some aspects of the Elo rating system . The Elo rating system is originally devised for chess rating (estimating players' skills based on match results). In PFAE, $\\\\theta$ is updated after each learning interaction:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\theta :=\\\\left\\\\{\\n\\t\\\\begin{aligned}\\n\\t\\t& \\\\theta + \\\\mu \\\\cdot (1 - p(\\\\theta))& \\\\mbox{if the answer was correct}, \\\\\\\\\\n\\t\\t& \\\\theta + \\\\nu \\\\cdot p(\\\\theta)& \\\\mbox{if the answer was wrong}.\\n\\t\\\\end{aligned}\\n\\t\\\\right.\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nAs the forgetting behavior of students is closely related to time, in order to consider forgetting, add a time effect function $f$ to $\\\\theta$, i.e., using $p(\\\\theta+f(t))$ instead of $p(\\\\theta)$, where $t$ is the time (in seconds) from the last learning interaction, and $f$ is the time effect function.\\nTo represent the complex forgetting behavior, the DKT-forget model introduces forgetting into DKT, which considers three types of side information related to forgetting: (1) the repeated time gap that represents the interval time between the present interaction and the previous interaction with the same KC, (2) the sequence time gap that represents the interval time between the present interaction and the previous interaction, and (3) past trial counts that represent the number of times a student has attempted on the exercise with the same KC. All these three features are discretized at $log_2$ scale.\\nThose side information is concatenated as additional information and represented as a multi-hot vector $\\\\bm{c}_t$, which is integrated with the embedding vector $\\\\bm{v}_t$ of the learning interaction, as follows:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\bm{v}_t^c = \\\\theta^{in}(\\\\bm{v}_t, \\\\bm{c}_t),\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $\\\\theta^{in}$ is the input integration function. The integrated input $\\\\bm{v}_t^c$ and the previous knowledge state $\\\\bm{h}_{t-1}$ are passed through the RNNs to update $\\\\bm{h}_t$ in the same way as in Eq.(\\\\ref{dkt}). The additional information at the next time step $\\\\bm{c}_{t+1}$ is also integrated with the updated $\\\\bm{h}_t$:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\bm{h}_t^c = \\\\theta^{out}(\\\\bm{h}_t, \\\\bm{c}_{t+1}),\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $\\\\theta^{out}$ is the output integration function.\\n propose a novel HawkesKT model, which introduces the Hawkes process to adaptively model temporal cross-effects. The Hawkes process performs well at modeling sequential events localized in time, as it controls corresponding temporal trends by the intensity function. The intensity function in HawkesKT is designed to characterize the accumulative effects of previous learning interactions, along with their evolutions over time. In HawkesKT, the temporal cross-effects and the ways in which they evolve between historical learning interactions combine to form a dynamic learning process.\",\n \"id\": \"2cb054ef-697a-4d85-a2b7-9c0e6282362f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"a705e159-25ef-44ad-8166-1cef83a7c265\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Variants of Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Considering Forgetting after Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Considering Forgetting after Learning\"\n },\n {\n \"cite_extract_rate\": 0.157894736842105,\n \"cites\": [\n 1684,\n 5383,\n 5384\n ],\n \"content\": \"\\\\label{sec:side information}\\nMost KT models primarily {rely on exercises and student responses} to evaluate students' knowledge states. These models have yielded impressive results and have been effectively implemented in online learning systems. Despite this, there are various other types of side information collected across the learning process that could be utilized to enhance {these models.} In this section, we will introduce several variants that aim to leverage this diverse side information across learning. \\nIn terms of a student's first response time, a short initial response time could indicate either high proficiency or 'gaming' behavior, while a long initial response time could indicate either careful thinking or lack of concentration. Since the connection between initial response time and knowledge state could be influenced by {complex factors}, propose to discretize the continuous first response time into four categories (i.e., extremely short, short, long, extremely long) {to eliminate unnecessary information and simplify the latent complex possibilities. They then build a one-by-four parameter table, in which each column represents the category of the initial response time of the previous exercise, while the relevant values represent the probability of correct answers}. \\nRegarding tutor intervention, propose the Bayesian evaluation and assessment model, which simultaneously assesses students' knowledge states and evaluates the lasting impact of tutor intervention. More specifically, it adds one observable binary intervention node to BKT: \\\\emph{True} means that the tutor intervention occurs in corresponding interactions while \\\\emph{False} indicates the opposite. The connection between the intervention node and knowledge node indicates the potential impact of the tutor intervention on students' knowledge states. The intervention node is linked to all four BKT parameters. As a result, there are a total of eight parameters to learn in order to incorporate tutor intervention. One possible way to reduce the number of parameters is choosing to link only the intervention node to the learning rate parameter .\\nSimilarly, develop the intervention-Bayesian knowledge tracing (Intervention-BKT) model, {which incorporates two types of interventions into BKT and distinguishes their different effects: \\\\emph{elicit and tell}}. The relations between the intervention and performance nodes represent the impact of teaching interventions on student performance, while the relations between the intervention and knowledge nodes represent the impact of teaching interventions on students' knowledge states. Therefore, at each learning interaction, while the present knowledge state is conditional on both the previous knowledge state and the current intervention, the student's performance depends on both the present knowledge state and the current intervention.\\nRather than considering only one kind of side information, propose a feature-aware student knowledge tracing (FAST) model, which allows for the utilization of all kinds of side information. Traditional BKT uses conditional probability tables for the guessing, slipping, transition and learning probabilities, meaning that the number of features involved in inference grows exponentially. Therefore, as the number of features increases, the time and space complexity of the model also grow exponentially. To deal with this large number of features, FAST uses logistic regression parameters rather than conditional probability tables. The number of features and complexity increase linearly rather than exponentially. \\nFor parameter learning, FAST uses the Expectation Maximization with Features algorithm and focuses on only emission features. The E step uses the current parameter estimates $\\\\lambda$ to infer the probability of the student having mastered the KC at each learning interaction.\\nThe parameters $\\\\lambda$ are now a function of the weight $\\\\beta$ and the feature vector $\\\\bm{f}(t)$. $\\\\bm{f}$ is the feature extraction function, and $\\\\bm{f}(t)$ is the feature vector constructed from the observations at the relevant time step. The emission probability is represented with a logistic function:\\n\\\\vspace{-0.1cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\t\\\\lambda(\\\\beta)^{y^{'},k^{'}} & = \\\\frac{1}{1 + exp(-\\\\beta^T \\\\cdot \\\\bm{f}(t))},\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.1cm}\\n\\\\end{equation}\\nwhere $\\\\beta$ is learned by training a weighted regularized logistic regression using a gradient-based search algorithm.\\n propose an extension to DKT that explored the inclusion of additional features. Specifically, it incorporates an auto-encoder network layer to convert the higher-dimensional input data into smaller representative feature vectors, thereby reducing both the resource and time requirement for training. Students' response time, opportunity count, and first action are selected as incorporated side information and all input features are converted into a fixed-length input vector. First, all input features are converted into categorical data and represented as a sparse vector by means of one-hot encodings. These encoded features are concatenated together to construct the higher-dimensional input vector:\\n\\\\vspace{-0.2cm}\\n\\\\begin{equation}\\n\\t\\\\begin{aligned}\\n\\t\\tC(e_t,a_t) &= e_t + (max(e) +1)a_t, \\\\\\\\\\n\\t\\tv_t &= O(C(e_t,a_t)) \\\\oplus O(C(t_t,a_t)) \\\\oplus O(t_t), \\\\\\\\\\n\\t\\tv_t^{'} &= tanh(W_vv_t + b_v),\\n\\t\\\\end{aligned}\\n\\t\\\\vspace{-0.2cm}\\n\\\\end{equation}\\nwhere $C$ is the cross feature, $O$ is the one-hot encoder format, $v_t$ represents the resulting input vector of each learning interaction, $e_t$ is the exercise, $a_t$ refers to the answer, $t_t$ is the response time, $W_v$ is the weight parameter and $b_v$ is the bias term. Subsequently, an auto-encoder is introduced to reduce the dimensionality without incurring the loss of too much important information. Finally, the feature vectors extracted by auto-encoder will be the new input of DKT.\\n presented another extension to DKT, namely the Exercise-aware Knowledge Tracing (EKT), which utilized the potential value of exercises' text contents. Generally, the text content is of great significance for students to understand the exercises. For example, used text materials to automatically predict their difficulties, utilized the text content to find similar exercises. further proposed a pre-training model called QuesNet for learning the unified representations of heterogeneous exercises. \\nTherefore, instead of using one-hot encoding of exercises, EKT automatically learns the semantic representation of each exercise from its text contents. EKT first uses $Word2vec$ to pre-train the embedding vector for each word in the exercise. It then constructs a bidirectional LSTM, which captures the word sequence from both forward and backward directions to learn the semantic word representation. The element-wise max-pooling operation is utilized to merge words’ contextual representations into a global embedding. Finally, EKT can update the student's knowledge state with the aid of the semantic representation of each exercise. \\nTo achieve more feasible integration of side information, present a deep knowledge tracing method with decision trees (DKT-DT), which takes advantage of Classification And Regression Trees (CART) to preprocess the heterogeneous input features . Specifically, CART is utilized to automatically partition the feature space and outputs whether or not a student can answer an exercise correctly. {The predicted response and the true response are encoded into a four-bit binary code; for example, the code is $1010$ if the predicted response and the true response are both correct. This binary code is then concatenated with the original one-hot encoding of the exercise as the new input of DKT to train the corresponding model.}\\n suggests that the student's language proficiency can serve as supplementary information to improve existing KT models. The student’s language proficiency is extracted by Elo rating score and time window features. Then, the language proficiency information is demonstrated to be effective in promoting several KT models, including DKT, DKVMN, and SAKT. Additionally, the problem of cold start in the KT task is alleviated with the assistance of language proficiency information. \\n explore to add side information to the original KT model by auxiliary learning tasks. They specifically introduced two tasks: (1) predicting the KCs of the question, and (2) predicting the individualized prior knowledge. By training with these tasks, KT can enhance its understanding of the intrinsic relationships between questions and KCs, while explicitly capturing student-level variability. \\n{When solving programming problems, we can record students' full code submissions, which can be employed to analyze their programming ability. collected student programming data and analyzed their personalized programming preferences. The study commenced with a statistical analysis of student errors, followed by an examination of students' programming structures derived from their code submissions, and finally used BKT to measure students' programming abilities.\\n transformed students' code submissions into embedded vectors, and applied them in DKT to model students' fine-graded programming knowledge states.\\n noticed that a single programming problem generally involves in multiple KCs, thereby they proposed to learn useful information about the programming problem's multiple requirements from students' code submissions. }\",\n \"id\": \"60b4234d-a28f-4026-ab3b-0291b62ceca6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"a705e159-25ef-44ad-8166-1cef83a7c265\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Variants of Knowledge Tracing Models\"\n ],\n [\n \"subsection\",\n \"Utilizing Side Information across Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Utilizing Side Information across Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:applications}\\nAlthough knowledge tracing is an emerging research area, it has already been applied in a wide variety of scenarios. {In the following, we will first survey the applications of KT models in two typical educational scenarios: learning resources recommendation and adaptive learning. Then, we will discuss broader applications of KT beyond student learning.}\",\n \"id\": \"5fc15152-8ae0-4a7f-83a2-d64f287123fe\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Applications\"\n ]\n ],\n \"subsections\": [\n \"260a9040-27f0-4099-a17f-ea39955bc0eb\",\n \"f6f4178c-41e5-4941-b347-f40654197f09\",\n \"c5ab0210-1d4b-4488-91f4-cb2e837345b7\"\n ],\n \"title\": \"Applications\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 1206\n ],\n \"content\": \"Traditionally, learning resources for each student are selected in one of two ways. The first one requires teachers to manually select suitable resources that match students' knowledge levels. However, this approach requires substantial time and effort, and different teachers may have different preferences. The second one allows students themselves to freely choose resources to learn. However, this may result in students choosing too easy or too difficult materials that will not benefit their learning , leading to low learning efficiency. In recent years, the prevalence of intelligent tutoring systems and the development of KT methods have made it possible to automatically recommend appropriate exercises to each student based on artificially designed algorithms.\\nExercises are the most common learning resources in learning. Given the inferred knowledge states, one common strategy is selecting the next exercise that will best advance students' knowledge acquisition. propose two extensions of the original BKT model, which respectively considered exercises' difficulties and students' multiple-attempt behaviors. These two extensions are integrated into a BKT-sequence algorithm to recommend exercises to students based on their knowledge states. Specifically, BKT-sequence first determines the predicted range of scores for each exercise. It then computes an expected score for each exercise that the student should get to achieve mastery, which is dependent on their current knowledge state (for instance, a lower knowledge state will result in higher expected scores). Finally, the algorithm returns the exercise with a predicted score that is closest to that of the expected score. Therefore, as the knowledge state of a particular KC grows, more difficult exercises will be recommended, as harder exercises are associated with a lower predictive score. Experimental results have shown that students using the BKT-sequence algorithm were able to solve more difficult exercises, obtained higher performance and spent more time in the system than students who used the traditional approach. Moreover, students also expressed that the BKT-sequence algorithm was more efficient. expanded the DKVMN model to include the exercise's type and difficulty. This model is then used to assess students' knowledge state and subsequently recommends personalized exercises for each student in Ssmall Private Online Courses (SPOCs). They conducted a randomized controlled trial to show the proposed personalized exercise recommendation could enhance students' learning efficiency.\\nIn addition to exercises, there are also some other types of multi-modal learning resources, such as videos and figures. utilizes an adaptation of BKT to improve student performance prediction by incorporating video observation. Experimental verification demonstrates the impact of both using and eschewing video data, as well as the learning rate associated with a particular video. In this way, they further developed a method to help people evaluate the quality of video resources. Concretely, they proposed the Template 1 Video model to incorporate video observations into BKT, which adds video activity as additional independent observation nodes to the BKT model. This model accordingly considers the probability that a given video resource will impart knowledge to a student. Moreover, the transition probability in BKT is conditional only on the presence of either a video or an exercise. Thus, the quality of the video can be determined by its promotion of learning, and this model can be leveraged as a tool to aid in evaluating and recommending video resources.\\nWhen recommending learning resources, the primary aim of existing solutions is to choose a simple strategy for assigning non-mastered exercises to students. \\nWhile reasonable, it is also too broad to advance learning effectively. accordingly propose three more beneficial and specific objectives: \\\\emph{review and explore}, \\\\emph{smoothness of difficulty level} and \\\\emph{student engagement}. In more detail, \\\\emph{review and explore} considers both enhancing students' non-mastered concepts with timely reviews and reserving certain opportunities to explore new knowledge; \\\\emph{smoothness of difficulty level} indicates that the difficulty levels of several continuous exercises should vary within a small range as students gradually learn new knowledge; finally, \\\\emph{student engagement} considers that to promote students' enthusiasm during learning, the recommended exercises should be in line with their preferences. In order to support online intelligent education with the above three domain-specific objectives,\\nthey developed a more reasonable multi-objective deep reinforcement learning (DRE) framework. DRE presented three corresponding novel reward functions to capture and quantify the effects of the above three objectives. This DRE framework is a unified platform designed to optimize multiple learning objectives, where more reasonable objectives also can be incorporated if necessary.\\nExperimental results show that DRE can effectively learn from the students' learning records to optimize multiple objectives and adaptively recommend suitable exercises.\",\n \"id\": \"260a9040-27f0-4099-a17f-ea39955bc0eb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"5fc15152-8ae0-4a7f-83a2-d64f287123fe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Learning Resources Recommendation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Learning Resources Recommendation\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 5373\n ],\n \"content\": \"{Adaptive learning, unlike learning resource recommendations, goes beyond the mere provision of resources. It not only concentrates on the selection of appropriate learning materials but also designs effective learning strategies and dynamic learning pathways. These are structured based on both the learning rules and students' evolving knowledge states.\\nSpecifically, adaptive learning broadly refers to \\\\emph{\\\"a learning process in which the content taught, or the way such content is presented, changes or 'adapts' based on individual student responses, and which dynamically adjusts the level or types of instruction based on individual student abilities or preferences\\\"} . }\\nThe first few attempt made to apply KT to adaptive learning was the ACT Programming Tutor (APT) , where students were asked to write short programs and BKT was utilized to estimate their evolving knowledge state. This tutor can present an individualized sequence of exercises to each student based on their estimated knowledge states until the student has \\\"mastered\\\" each rule. \\nIn recent years, Massive Open Online Courses (MOOCs) have become an emerging modality of learning, particularly in higher education. adapt BKT on the edX platform. The research object was a 14-week online course that included weekly video lectures and corresponding lecture problems. BKT was applied to enhance students' learning on this course. In order to better adapt BKT to the learning platform, the original BKT was modified in several respects. First, due to the lack of labeled KCs, the problems would be directly seen as the KCs, while the questions would be seen as the exercises belonging to the KC.\\nSecond, in order to capture the varying degrees of students' knowledge acquisition at each attempt, the modified model assigned different guess and slip parameters to different attempt counts. Third, to deal with the problem of multiple pathways in the system, which reflected that the impacts on learning may come from various resources, they framed the influence of resources on learning as a credit/blame inference problem.\\nGenerally, students' cognitive structures include both students' knowledge level and the knowledge structure of learning items (e.g., \\\\emph{one-digit addition} is the prerequisite knowledge of \\\\emph{two-digit addition}). Therefore, adaptive learning should maintain consistency with both students' knowledge level and the latent knowledge structure. Nevertheless, existing methods for adaptive learning often focus separately on either the knowledge levels of students (i.e., with the help of specific KT models) or the knowledge structure of learning items. To fully exploit the cognitive structure for adaptive learning, propose a Cognitive Structure Enhanced framework for adaptive Learning (CSEAL). CSEAL conceptualized adaptive learning as a Markov Decision Process. It first utilized DKT to trace the evolving knowledge states of students at each learning step. Subsequently, the authors designed a navigation algorithm based on the knowledge structure to ensure that the learning paths in adaptive learning were logical and reasonable, which also reduced the search space in the decision process. Finally, CSEAL utilized the actor-critic algorithm to dynamically determine what should be learned next. In this way, CSEAL can sequentially identify the most suitable learning resources for different students.\",\n \"id\": \"f6f4178c-41e5-4941-b347-f40654197f09\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"5fc15152-8ae0-4a7f-83a2-d64f287123fe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Adaptive Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adaptive Learning\"\n },\n {\n \"cite_extract_rate\": 0.22222222222222202,\n \"cites\": [\n 5385,\n 8912\n ],\n \"content\": \"{The above two types of applications are most commonly used for KT in student learning. In addition, the KT methods can be expanded to be utilized in any systems that necessitate continuous evaluation of user capabilities or states. We will introduce some broader applications of KT in this section.}\\nIn gaming systems, the paradigm of tracing students' knowledge state can also work for player modeling. Here, player modeling, which is the study of computational models of players in games, aims to capture human players' characteristics and cognitive features . For instance, reveal that children engage in cycles of increasingly sophisticated mathematical thinking over the course of playing an online game. present an approach to trace player knowledge in a parallel programming educational game, which is capable of measuring the current players' real-time state across the different skills required to play an educational game based only on in-game player activities. conduct a classroom experiment comparing a commercial game for equation solving, i.e., \\\\emph{DragonBox}, with a research-based intelligent tutoring system, i.e., \\\\emph{Lynnette}. The results indicated that students who used \\\\emph{DragonBox} enjoyed the experience more, while students who used \\\\emph{Lynnette} performed significantly better on the test. Therefore, it is possible to enable students to learn effectively and happily by designing suitable educational games on the online learning platform.\\n{In crowdsourcing, unlabeled data or specific tasks are assigned to various crowd annotators. Understanding the dynamic capabilities of these annotators is crucial in ensuring the reliability of their annotations and promoting the annotation efficiency. developed a framework called KT4Crowd, which utilized KT methods to predict the performance of annotators, which surpassed traditional rating systems. \\nBesides, observed that students' participation in crowdsourcing tasks can enhance their learning. KT methods can also better comprehend students' knowledge states, aided by the students' annotated items. \\n}\\n develop an online citizen science project that employs machine learning techniques to improve the training of new volunteers using authentic tasks featuring uncertain outcomes, such as image classification. Specifically, they employ the BKT model to monitor the knowledge states of volunteers, enabling them to more efficiently complete assigned tasks and contribute meaningfully to the project. \\n develop an automated exercise collection approach for teachers, employing the KT model and reinforcement learning. Specifically, the exercise collection need to be well-designed to align with students' abilities. This study first leverages the KT model to forecast students' performance on unseen exercise candidates. Subsequently, the exercise selector is designed based on the KT model's predictions, ensuring that the exercise collection is both approximate and optimized. Similarly, design a reinforcement learning guided\\nmethod for exam paper generation, where the DKT model is utilized to measure examinees' knowledge states. \\n\\\\iffalse\\n\\\\begin{table*}[t]\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\renewcommand\\\\arraystretch{1.0}\\n\\t\\\\caption{A Summary of Public Datasets for Knowledge Tracing. }\\n\\t\\\\vspace{-0.3cm}\\n\\t\\\\centering\\n\\t\\\\begin{tabular}{l|l}\\n\\t\\t\\\\hline\\n\\t\\tDatasets & Link\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\tASSISTments2009 & https://sites.google.com/site/assistmentsdata/home/assistment-2009-2010-data\\\\\\\\\\n\\t\\tASSISTments2012 & https://sites.google.com/site/assistmentsdata/home/2012-13-school-data-with-affect\\\\\\\\\\n\\t\\tASSISTments2015 & https://sites.google.com/site/assistmentsdata/home/2015-assistments-skill-builder-data \\\\\\\\\\n\\t\\tASSISTments2017 & https://sites.google.com/view/assistmentsdatamining/dataset \\\\\\\\\\n\\t\\tKDD-Cup 2010 & https://pslcdatashop.web.cmu.edu/KDDCup/rules\\\\_data\\\\_format.jsp \\\\\\\\\\n\\t\\tStatics 2011 & https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=507 \\\\\\\\\\n\\t\\tslepemapy.cz & https://www.fi.muni.cz/adaptivelearning/?a=data\\\\\\\\\\n\\t\\tsynthetic & https://github.com/chrispiech/DeepKnowledgeTracing/tree/master/data/synthetic\\\\\\\\\\n\\t\\tJunyi & https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=1198 \\\\\\\\\\n\\t\\tEdNet & http://ednet-leaderboard.s3-website-ap-northeast-1.amazonaws.com/\\\\\\\\\\n\\t\\tpoj & http://poj.org/ \\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\\\end{tabular}\\n\\t\\\\label{tab3}\\n\\t\\\\vspace{-0.6cm}\\n\\\\end{table*}\\n\\\\fi\",\n \"id\": \"c5ab0210-1d4b-4488-91f4-cb2e837345b7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"5fc15152-8ae0-4a7f-83a2-d64f287123fe\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Applications\"\n ],\n [\n \"subsection\",\n \"Broader Applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Broader Applications\"\n },\n {\n \"cite_extract_rate\": 0.14285714285714202,\n \"cites\": [\n 5386\n ],\n \"content\": \"\\\\label{sec:dataset}\\nAfter introducing above KT models and variants, to better help researchers and practitioners who want to further conduct related work and promote the application of KT, we have open sourced two algorithm libraries, i.e., EduData that for downloading and preprocessing most existing KT-related datasets, and EduKTM that includes extensible and unified implementations of existing popular KT models. In the following, we will give detailed introduction of these two algorithm libraries.\\n\\\\begin{table*}[t]\\n\\t\\\\vspace{-0.6cm}\\n\\t\\\\renewcommand\\\\arraystretch{1.2}\\n\\t\\\\caption{{Basic information and statistics of existing datasets available for evaluating KT models. }}\\n\\t\\\\vspace{-0.2cm}\\n\\t\\\\centering\\n\\t\\\\resizebox{\\\\textwidth}{!}\\n\\t{\\n\\t\\t\\\\begin{tabular}{l|c|c|c|c|c|r|r|r|r}\\n\\t\\t\\t\\\\hline\\n\\t\\t\\t\\\\multirow{3}*{Datasets}&\\\\multirow{3}{*}{Subjects} &\\\\multirow{3}{*}{Learning Stages} & \\\\multirow{3}*{\\\\makecell[c]{Side \\\\\\\\ Information}}& \\\\multicolumn{2}{c|}{Sources}&\\\\multicolumn{4}{c}{Statistics} \\\\\\\\\\n\\t\\t\\t\\\\cline{5-10} \\n\\t\\t\\t& & & & \\\\makecell[c]{Online \\\\\\\\ platforms} & \\\\makecell[c]{Educational \\\\\\\\ challenges} & \\\\makecell[c]{\\\\# of \\\\\\\\ Students} & \\\\makecell[c]{\\\\# of \\\\\\\\ Exercises} & \\\\makecell[c]{\\\\# of \\\\\\\\ KCs} & \\\\makecell[c]{\\\\# of \\\\\\\\ Learning records}\\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tASSISTments2009 & mathematics & middle school & Yes & \\\\Checkmark & & 4,163 & 17,751& 123 & 346,860 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tASSISTments2012 & mathematics & middle school & Yes & \\\\Checkmark & & 46,674 & 179,999& 265 &6,123,270 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tASSISTments2015 & mathematics & middle school & No & \\\\Checkmark & & 19,917 & /& 100 & 708,631 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tASSISTments2017& mathematics &\\\\makecell[c]{from middle school \\\\\\\\ to college} & Yes & & \\\\Checkmark & 1,709 &3,162& 102 & 942,816 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tJunyi & mathematics & \\\\makecell[c]{from primary \\\\\\\\ to high school} & Yes & \\\\Checkmark & & 247,606 & 722 & 41 & 25,925,922 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tEedi2020 & mathematics & \\\\makecell[c]{from primary \\\\\\\\ to high school} & Yes & & \\\\Checkmark & 118,971 & 27,613 & 388 & 15,867,850 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tStatics2011 &engineering &university &Yes & \\\\Checkmark & & 335 & 1,224 & 80 & 361,092 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tEdNet-KT1 & english & / & Yes & \\\\Checkmark & & 784,309 & 13,169 & 188 & 95,293,926 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tEdNet-KT2 & english &/ & Yes & \\\\Checkmark & & 297,444 & 13,169 & 188 & 56,360,602 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tEdNet-KT3 & english &/ & Yes & \\\\Checkmark & & 297,915 & 13,169 & 293 & 89,270,654 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tEdNet-KT4 & english &/& Yes& \\\\Checkmark & & 297,915 & 13,169 & 293 & 131,441,538 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\tCodeWorkout & programming & university & Yes & & \\\\Checkmark & 819 & 50 & 50 & \\\\textgreater 130,000 \\\\\\\\\\n\\t\\t\\t\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t}\\n\\t\\\\label{data}\\n\\t\\\\vspace{-0.6cm}\\n\\\\end{table*}\",\n \"id\": \"34caa8b6-1d19-4064-837a-9dcb1549a45d\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ]\n ],\n \"subsections\": [\n \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"66f9b876-576e-4b0f-af2b-91235cb5ec11\"\n ],\n \"title\": \"Datasets and Baselines\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As we have mentioned, KT emerges from the development of online education, where a large number of students' learning data is collected for analyzing their learning behaviors and knowledge states. In this section, we mainly introduce existing public datasets available for evaluating KT models. Table \\\\ref{data} lists all datasets, as well as their basic information and statistics. In our released EduData, we provide the service of downloading and preprocessing all these datasets, which is convenient to help beginners analyze and utilize them quickly.\\nIn summary, these datasets are collected in different learning scenarios, so that they exhibit an extremely distinct difference in data scale, subject, and so on, indicating complex applications in practical for KT models.\",\n \"id\": \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"34caa8b6-1d19-4064-837a-9dcb1549a45d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [\n \"3d73fd7b-d881-422a-a516-5cd421232c1c\",\n \"efcd8f6a-3777-4dd9-8b4a-f9125974da7d\",\n \"db5459a4-5149-42f0-85da-a9bc2ed105e4\",\n \"266d224a-ec55-40cb-a11c-e27be3a867bf\",\n \"fd615d1a-1d1e-4b41-9cdf-9cd4873ba07b\",\n \"6cb5a8b7-c682-457d-85c6-8ed300213a0a\"\n ],\n \"title\": \"Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"ASSISTments , created in 2004, is an online tutoring system in the United States, which provides students with both assessment information and tutoring assistance. While working on ASSISTments, students will be provided with instructional assistance to help them solve the problem in several substeps when they give wrong answers. After obtaining correct answers, they will be given a new one. Meanwhile, the system will learn about the students' knowledge states and predict how they will do in future tests. Up to now, the organizers have released four public avaliable datasets from ASSISTments, which are respectively ASSISTments2009, ASSISTments2012\\\\footnote{https://sites.google.com/site/assistmentsdata/datasets/2012-13-school-data-with-affect}, ASSISTments2015\\\\footnote{https://sites.google.com/site/assistmentsdata/datasets/2015-assistments-skill-builder-data}, and ASSISTments2017\\\\footnote{https://sites.google.com/view/assistmentsdatamining/dataset}. The ASSISTments datasets have built profound impacts in the research community of KT, which can be seen in many related papers.\\nMost of these datasets were collected from mathematics in middle school, the details of them are listed as follows.\\n\\\\begin{itemize}[leftmargin=*]\\n\\t\\\\item{\\\\textbf{ASSISTments2009.}}\\n\\tThe full name of ASSISTments2009 is the ASSISTments 2009-2010 skill builder data set , which was collected form ASSISTments’ skill builder problem sets during the school year from 2009 to 2010. Students were asked to work on exercises with similar knowledge concepts until they can answer correctly for three or more times in a row. This dataset contains many valuable side information, such as \\\\textit{attempt count} that represents the number of attempts of the student, \\\\textit{ms first reponse} that represents the time in the milliseconds for the student’s first response, and \\\\textit{opportunity} that represents the number of opportunities the student has to practice. It worth noting that the original version of ASSISTments2009 has three serious problems that lead to unreliable experimental results : (1) a large number of duplicated records, (2) treating scaffolding problems as the same as main problems, and (3) repeated response sequences with different KCs. The latest version of ASSISTments2009 has fixed these problems.\\n\\t\\\\item{\\\\textbf{ASSISTments2012.}}\\n\\tASSISTments2012 is collected from the ASSISTments system, in the school year from 2012 to 2013 with affect predictions. Compared with ASSISTments2009, ASSISTments2012 has much more students, exercises, as well as learning records. However, many learning records have missed the related KCs in this dataset. After filtering these records, there are only 29,018 students, 53,091 exercises, and 2,711,813 learning records.\\n\\tBesides, this dataset contains more side information, such as \\\\textit{start time}, \\\\textit{end time}, \\\\textit{problem type}, and \\\\textit{average confidence}. \\n\\tIt worth noting that you can access to the text of the exercises in ASSISTments2012 for more fine-grained research. Specifically, you can email nth@wpi.edu and cc td@wpi.edu, explaining your purpose and promising not to share it with anyone else.\\n\\t\\\\item{\\\\textbf{ASSISTments2015.}}\\n\\tASSISTments2015 covers the 2015 school years' student response records on ASSISTments. This dataset only contains student learning records on 100 KCs, there is no information about the exercise, as well as any other side information. \\n\\t\\\\item{\\\\textbf{ASSISTments2017.}}\\n\\tASSISTments2017 is the dataset used in the 2017 ASSISTments Longitudinal Data Mining Competition. This dataset collected data over a decade, which tracked students' intercations at the ASSISTments learning platform from middle school study in 2004-2007 to their high school course-taking, until they graduated from the college. Therefore, the average learning record of students in ASSISTments2017 is much longer than other dataset, which is beyond the length of 1,000. ASSISTments2017 also has rich side information, including but not limited to \\\\textit{AveKnow} that indicates students' average knowledge level based on BKT, \\\\textit{timeTaken} that represents the time spent on the current exercise, \\\\textit{frIsHelpRequest} that represents whether the first response is a help request.\\n\\\\end{itemize}\",\n \"id\": \"3d73fd7b-d881-422a-a516-5cd421232c1c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"ASSISTments Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ASSISTments Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The Junyi dataset contains the problem log and exercise-related information on the Junyi Academy\\\\footnote{http://www.junyiacademy.org/}, a Chinese e-learning website established in 2012 on the basis of the open-source code released by Khan Academy. In contrast to the ASSISTments datasets, Junyi has less exercsies and KCs, but includes an exercise hierarchy labeled by experts, the annotations of exercise relationship are also available. Therefore, many research works that focused on the knowledge structure in KT had utilized this dataset . Junyi provides the prerequisite exercise of a specific exercise in the knowledge map, the topic and area of each exercise, as well as the coordiate position of the knowledge map.\",\n \"id\": \"efcd8f6a-3777-4dd9-8b4a-f9125974da7d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"Junyi Dataset\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Junyi Dataset\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8913\n ],\n \"content\": \"The Eedi2020 dataset is also released in an achedemic challenge, i.e., NeurIPS 2020 Education Challenge\\\\footnote{https://eedi.com/projects/neurips-education-challenge}. This dataset contains students’ answers to mathematics questions from Eedi, an online educational platform which millions of students interact with daily around the globe from school year 2018 to 2020. All exercises are multiple-choice problems with 4 possible answer choices, exactly one of which is correct. In Table \\\\ref{data}, we give the statistics based on the training data in this competition, the total number of learning records in the full dataset exceeds 17 million. It worth noting that Eedi2020 gives students' exact answer choice so that we can also predict students options . Moreover, for the students, Eedi2020 records lots of valuable context information, including the \\\\textit{Gender}, \\\\textit{DateOfBirth}, \\\\textit{PremiumPupil}. For the learning records, Eedi2020 also presents their \\\\textit{Confidence}, \\\\textit{GroupId}, \\\\textit{QuizId}, and \\\\textit{SchemeOfWorkId}.\",\n \"id\": \"db5459a4-5149-42f0-85da-a9bc2ed105e4\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"Eedi2020 Dataset\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Eedi2020 Dataset\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Different from the above datasets that focus on mathematics exercises, the Statics2011 dataset is obtained from a college-level engineering statics course via online educational system developed by Carnegie Mellon University\\\\footnote{https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=507}. {The problems in college engineering course are quite complex, often comprising numerous independent sub-steps. Consequently, we treat each sub-problem as an exercise and calculate the number of exercises and KCs within this dataset.}\",\n \"id\": \"266d224a-ec55-40cb-a11c-e27be3a867bf\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"Statics2011 Dataset\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Statics2011 Dataset\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 5386\n ],\n \"content\": \"The EdNet dataset is related to the english subject, which is consisted of students' learning records in the multi-platform AI tutoring system Santa in South Korea\\\\footnote{https://github.com/riiid/ednet}. EdNet collected learning data of students over two years for their preparation of the e TOEIC (Test of English for International Communication) Listening and Reading Test. EdNet is now the largest public dataset in KT field with a total of 131,441,538 learning records from 784,309 students. Besides, it contains various features of students' learning actions, such as the specific learning material they have interacted, how much time they have spent for answering a given exercise. There are four differnet versions of EdNet, respectively named EdNet-KT1, EdNet-KT2, EdNet-KT3, and EdNet-KT4 with different extents. {We note that the students in this dataset may be in different learning states. Therefore we do not give the learning state information in Table \\\\ref{data}.}\\n\\\\begin{itemize}[leftmargin=*]\\n\\t\\\\item{\\\\textbf{EdNet-KT1.}}\\n\\tEdNet-KT1 contains students' basic exercise-answering logs. This dataset has 784,309 students, 13,169 exercises, 188 KCs, and a total of 95,293,926 learning records.\\n\\tExercises in EdNet-KT1 are organized by bundles, i.e., a collection of exercises sharing a common passage, picture or listening material. Therefore, exercises come up in bundles and students have to answer all contained exercises when a bundle is given.\\n\\t\\\\item{\\\\textbf{EdNet-KT2.}}\\n\\tEdNet-KT2 recorded students' action sequences, which indicated their full learning behaviors. For example, a student who is not confident about the answer may alternately select among several answer choices before submitting. Such learning behaviors can reflect more fine-grained knowledge state of students. EdNet-KT2 contains three kinds of actions: \\\\textit{enter} when student first receives and views a bundle , \\\\textit{respond} when the student selects an answer choice to the exercise, and \\\\textit{submit} when the student submits his final answers to the the given bundle. It is worth noting that EdNet-KT2 is a subset of EdNet-KT1.\\n\\t\\\\item{\\\\textbf{EdNet-KT3.}}\\n\\tOn the basis of EdNet-KT2, EdNet-KT3 collected more students' learning activities, such as reading explanations or watching lectures. These learning activities have potential impacts on students' knowledge state so that they are valuable to be analyzed.\\n\\t\\\\item{\\\\textbf{EdNet-KT4.}}\\n\\tIn EdNet-KT4, the very fine details of actions were provided. In particular, the following types of actions are added to EdNet-KT3: erase choice, undo erase choice, play audio, pause audio, play video, pause video, pay, refund, and enroll coupon.\\n\\\\end{itemize}\",\n \"id\": \"fd615d1a-1d1e-4b41-9cdf-9cd4873ba07b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"EdNet Dataset\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"EdNet Dataset\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The CodeWorkout dataset is utilized in the 2nd Computer Science Educational Data Mining Challenge (CSEDM)\\\\footnote{https://sites.google.com/ncsu.edu/csedm-dc-2021/home}. This dataset is collected from a CS1 course in the Spring and Fall 2019 semesters at a public university in United States. It contains the code submissions from students for 50 coding problems, each requiring ~10-26 lines of code. {In this dataset, each exercise has a unique KC, thereby the number of KCs is equal to the number of exercises.} In total, there are 329 and 490 students in the Spring and Fall semesters who completed the course. Each dataset contains more than 65,000 code submissions, the scores of the submissions (\\\\% of unit tests passes) are also available, as well as the compiler message if the compilation is not successful. The final grades of students are also provided for this dataset.\",\n \"id\": \"6cb5a8b7-c682-457d-85c6-8ed300213a0a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"8d0ee175-d87a-4c6c-bbdc-2113cd06121b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ],\n [\n \"subsubsection\",\n \"CodeWorkout Dataset\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"CodeWorkout Dataset\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The implementations of existing KT methods are not standardized, which may use different program languages (e.g., python, lua) and different deep learning frameworks (e.g., tensorflow, torch). Furthermore, some works did not well organize the codes systematically (e.g., the missing of running environments and dependencies), which brings difficulties in reproducing the models. To this end, we put forward the algorithm library for KT baselines, named EduKTM, which now has contained the concurrent popular works. EduKTM will be always under development for including the latest KT models, more algorithms and features are going to be added. Besides, we provide detailed guidelines for everyone who is interested in contributing to EduKTM.\\nIt worth nothing that we do not provide the performance evaluation of these baselines using the aforementioned benchmark datasets. The reasons are two-folds: (1) For each baseline method, their experimental settings are quite different as they were designed to handle with various learning scenarios. No single KT model can always be the best one under various learning contexts. It is not fair to compare all baselines under a fixed setting.\\n(2) Existing evaluation standards primarily concentrate on the performance of student performance prediction task, which cannot directly reflect the effectiveness of various KT models in practical applications.\\nTherefore, we decide to open source EduData and EduKTM. Researchers and practitioners can freely compare and select appropriate KT models based on their specific requirements in various application scenarios.\",\n \"id\": \"66f9b876-576e-4b0f-af2b-91235cb5ec11\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"34caa8b6-1d19-4064-837a-9dcb1549a45d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Datasets and Baselines\"\n ],\n [\n \"subsection\",\n \"Baselines\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Baselines\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:future}\\nThis survey has comprehensively reviewed the abundant current developments in the field of knowledge tracing, including fundamental KT models, their variations and typical applications. Despite this, as knowledge tracing is a relatively new but promising research area, there are still a substantial number of problems that require urgent resolution. In this section, we discuss several potential future research directions.\",\n \"id\": \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ]\n ],\n \"subsections\": [\n \"ee9a478b-7e5c-4c3d-a3b7-01ab255d95f7\",\n \"203ad4ee-8eb1-4e57-ac27-fd2974c1a08f\",\n \"1df76c0e-efc5-4366-b813-6204fb5c7b9f\",\n \"ed0bdca3-3fb6-4028-a4d7-cc6fb028657b\",\n \"01240c30-0ca2-4a6c-bf1e-ccabebf5cfe7\",\n \"c7707957-a296-4aea-b19d-29d5e30c34c6\"\n ],\n \"title\": \"Future Research Directions\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 5374,\n 8914\n ],\n \"content\": \"The performance of KT models is now evaluated indirectly through the student performance prediction task. The higher the precision of students' responses on future exercises, the better the KT model's performance. Nonetheless, interpretability plays a significant role in education, as students often express more concern about the 'why' than the 'what' of a learning decision . Enhancing the interpretability of KT models is therefore crucial. Some educational theories, such as the Rasch model used in AKT and the transfer of knowledge used in SKT , could be considered for this purpose. noticed the significance of interpretable KT, particularly for deep learning models, as they have numerous parameters that are challenging to explain meaningfully. Thus, they introduced a straightforward and interpretable KT approach, based on the causal relationships within the latent features extracted from students' behavioral data. \\n attempted to introduce causal inference for explanatory analysis on KT, and they achieved more stable and explainable knowledge tracing based on the analysis results. \\nIt is imperative to further refine existing KT models or to explore additional methods for interpretable KT researches. This will lead to the production of more accurate and interpretable evaluations of students' knowledge states.\",\n \"id\": \"ee9a478b-7e5c-4c3d-a3b7-01ab255d95f7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"Knowledge Tracing with Interpretability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Tracing with Interpretability\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 5388,\n 5387,\n 8915\n ],\n \"content\": \"The acquisition of high-quality KT models necessarily requires a substantial amount of data to ensure training stability. However, online learning systems often achieve less learning interactions in practical educational scenarios, thereby leading to the data sparsity problem. To address this issue, utilized graph convolutional network to include exercise-KC correlations. \\n turned to improve the attention mechanism.\\nMore recently, researchers have employed contrastive learning to alleviate the data sparsity problem in KT . For example, presented a contrastive learning framework to enhance KT, which measure the semantical similarity between various learning interactions to effectively learn their representations. They further designed data augmentation methods to enhance the semantics of students' learning interactions. However, while the aforementioned methods alleviate the data sparsity problem in various aspects, there remains a need for further improvement in addressing this issue comprehensively.\",\n \"id\": \"203ad4ee-8eb1-4e57-ac27-fd2974c1a08f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"Knowledge Tracing with Sparse Learning Interactions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Tracing with Sparse Learning Interactions\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Most existing KT researches are capable of dealing with objective exercises, assuming student responses are binary. However, they overlook students' open-ended answers, often presented in natural language, on subjective exercises. As per the work by , an open-ended KT approach for computer science education was explored. This method aimed to predict students' precise solutions to programming questions, utilizing language models for assistance. \\nRecent advancements in large language models offer promising potential to enhance the ability of KT models to comprehend the critical information regarding students' knowledge states embedded within their open-ended answers on subjective exercises .\",\n \"id\": \"1df76c0e-efc5-4366-b813-6204fb5c7b9f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"Knowledge Tracing with Subjective Exercises\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Tracing with Subjective Exercises\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Learning records are passive representations of students' knowledge states. In contrast, students' feedback provides us with their proactive understanding of their knowledge states, thereby offering direct and authentic indicators of their learning situation. However, there are few KT models that take advantage of training data related to students' feedback, even though it can play an important role in fixing the KT results . have noted that feedback plays a positive role in learning, which may promote transfer and retention in learning from worked-out examples. Therefore, incorporating students' feedback presents a meaningful avenue that could lead to significant improvements.\",\n \"id\": \"ed0bdca3-3fb6-4028-a4d7-cc6fb028657b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"Knowledge Tracing with Students' Feedback\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Tracing with Students' Feedback\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Generally, user modeling refers to tools for characterizing users' behaviors (e.g., frequent locations), personal information (e.g., age, gender, and occupation) and latent features (e.g., interests and abilities), which facilitate the provision of targeted services for different users . As a type of latent feature modeling, knowledge tracing diagnoses the proficiency of users (not only individuals, but also groups of individuals, like user teams and companies) on specific skills/concepts. Thus, in addition to education, knowledge tracing can be developed and applied across a wide range of domains for user modeling, including games, sports, and recruitment.\",\n \"id\": \"01240c30-0ca2-4a6c-bf1e-ccabebf5cfe7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"Knowledge Tracing for General User Modeling\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Tracing for General User Modeling\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 8461\n ],\n \"content\": \"{In recent years, the advancement of large language models (LLM), notably ChatGPT, has led to significant impacts and garnered considerable research interest worldwide . The use of LLM in education is promising to revolutionize the current learning pattern . The current study mainly focuses on generating learning materials, improving student-system interaction, and explaining educational contents. However, it remains unclear how LLM can assist in understanding students' knowledge states. Given the powerful capabilities of LLM, it has the potential to enhance the generalization and interpretability of existing KT methods. Specifically, LLM can analyze the content of students' responses and evaluate their quality, which directly reflects students' knowledge states. Besides, LLM can play the role of the instructor, answering questions for students and identifying their strengths and weaknesses. Furthermore, LLM itself can serve as a KT model, which can output reasonable results about students' knowledge states given their previous learning interactions.\\nSimultaneously, it is imperative to safeguard the privacy and security of student data when utilizing LLM . \\n}\",\n \"id\": \"c7707957-a296-4aea-b19d-29d5e30c34c6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"18821b92-3ccc-4057-9352-24e5c0409dd6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Future Research Directions\"\n ],\n [\n \"subsection\",\n \"Knowledge Tracing with Large Language Models\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Knowledge Tracing with Large Language Models\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclution}\\nIn this survey, we conducted a thorough review of knowledge tracing. Specifically, we initially conducted a comprehensive review of existing fundamental KT models. Considering the complexity of online learning systems and the significant importance of KT research in practical scenarios, we investigated a broad range of variant KT research. Subsequently, we summarized some typical applications of KT in common educational scenarios. Furthermore, we released two algorithm libraries for KT-related datasets and baselines, thereby facilitating researchers and practitioners in selecting appropriate KT models based on their specific requirements. Finally, we outlined some potential future directions. \\nWe hope that this comprehensive survey of knowledge tracing will assist readers in understanding the problem of modeling students' dynamic knowledge states. It serves as a fundamental framework for both researchers and practitioners in future studies, fostering the development of KT. The development of KT will directly benefit millions of students, with its impact potentially extending to a broader audience as online learning continues to evolve. In this context, KT will play an increasingly important role in enabling individuals to adapt to the ever-changing society . Furthermore, the development of more refined KT methods tailored to students in various subjects and age groups will enhance their ability to comprehend students' individual knowledge states. \\nHowever, we recognize some limitations of the current survey. \\nFirstly, as new KT methods continue to emerge, although we have conducted a thorough investigation, there may be some representative works that we have overlooked. If necessary, we will incorporate these into our proposed framework in the future. \\nSecondly, the complexity of KT methods warrants attention, as it continues to grow, particularly in those based on deep learning. The complexity of the model may enhance its accuracy, but it could also compromise its applicability, as users will question the reliability of complex models. Despite the fact that some studies have attempted to reveal the interpretability of complex KT models, for instance, by utilizing the xAI technique, there remains a significant distance to traverse before achieving a completely transparent deep learning KT model. \\nWe posit that the application of KT methods in online education presents a promising avenue for research. It significantly enhances the learning experience for students while simultaneously alleviating the burden on teachers. Despite the numerous challenges and obstacles, researchers are be encouraged to overcome them and ensure a reliable and equitable access to students' evolving knowledge state in learning. \\n\\\\section*{Acknowledgment}\\nThis research was supported by grants from the National Key Research and Development Program of China (Grant No. 2021YFF0901003), and the National Natural Science Foundation of China (Grant No. 62337001, U20A20229).\\n\\\\bibliographystyle{IEEEtranN}\\n\\\\begin{footnotesize}\\n\\t\\\\bibliography{cite}\\n\\\\end{footnotesize}\\n\\\\vspace{-0.6cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/shen.pdf}}]\\n\\t{Shuanghong Shen} recieved the B.E. degree from Wuhan University, Wuhan, China, in 2018. He is currently working toward the Ph.D. degree in the School of Computer Science and Technology, University of Science and Technology of China. His main research interests include data mining, knowledge discovery, natural language processing and intelligent tutoring systems. He won the first prize in task 2 of the NeurIPS 2020 Education Challenge. He has published papers in referred conference proceedings, such as KDD2023, AAAI2022, SIGIR2022.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-0.6cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/liu.pdf}}]\\n\\t{Qi Liu} (Member, IEEE) received the Ph.D. degree from University of Science and Technology of China (USTC), Hefei, China, in 2013. He is currently a Professor in the School of Computer Science and Technology at USTC. His general area of research is data mining and intelligent education. He has published prolifically in refereed journals and conference proceedings (e.g., TKDE, TOIS, KDD). He is an Associate Editor of IEEE TBD and Neurocomputing. He was the recipient of KDD’ 18 Best Student Paper Award and ICDM’ 11 Best Research Paper Award. He is a member of the Alibaba DAMO Academy Young Fellow. He was also the recipient of China Outstanding Youth Science Foundation in 2019.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-0.6cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/huang.pdf}}]\\n\\t{Zhenya Huang} (Member, IEEE) received the B.E. degree from Shandong University, in 2014 and the Ph.D. degree from USTC, in 2020. He is currently an associate researcher of the School of Computer Science and Technology, University of Science and Technology of China (USTC). His main research interests include data mining, knowledge discovery representation learning and intelligent tutoring systems. He has published more than 30 papers in refereed journals and conference proceedings including TKDE, TOIS, KDD, AAAI. He has served regularly in the program committees of a number of conferences, and is reviewers for the leading academic journals.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-0.6cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/zheng.png}}]\\n\\t{Yonghe Zheng} is the funding dean of Research Institute of Science Education in Beijing Normal University, Professor in Education Faculty of BNU, PhD Supervisor. Prof. Zheng is the Vice president of the 8th Council of China Association of Children's Science Instructors and is the member of Teaching Guidance Committee of the Basic Education, the member of comprehensive group of the compulsory education curriculum revision committee, the science curriculum standard revision group; the General Secretary of the Awarding Scheme Council for outstanding computer teachers in Universities; the Vice director of the academic committee of Chinese society of Natural Science Museum; and former Director General of the Science Policy Bureau of National Natural Science Foundation of China. Prof. Zheng has engaged in project management of basic research and S\\\\&T policy research for a long time. More than 60 articles have been published. And Zheng is the chief editor of Disciplinary and Interdisciplinary Science Education Research Journal. Prof. Zheng joined in Educational Faculty of Normal University in April, 2018, with specific research interests on science education and educational technology. Prof. Zheng established the Research Institute of Science Education of Beijing Normal University in November of 2019.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-0.6cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/yin.png}}]\\n\\t{Minghao Yin} (Member, IEEE) received the B.S. and M.S. degrees from Northeast Normal University, Changchun, China, in 2001 and 2004, respectively, and the Ph.D. degree from Jilin University, Changchun, China, in 2008, all in computer science. He is currently a professor, the dean of the School of Information Science and Information Technology at Northeast Normal University, and the director of the Key Laboratory of Intelligent Information Processing in Jilin Province Colleges and Universities. He also contributes as the Associate Editor for the IEEE Transactions on Learning Technologies (TLT). He has published multiple papers in prestigious journals and conferences, including Artificial Intelligence, IEEE Transactions, ACM Transactions, JAIR, AAAI, IJCAI, and WWW. His research interests include heuristic search, data mining, and combinatorial optimization.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-0.6cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/wang.png}}]\\n\\t{Minjuan Wang} (Member, IEEE) obtained her B.A. in Chinese literature from the Beijing/Peking University in 1995, M.A. in comparative literature from Penn State University in 1997, and Ph.D. in information science and learning technologies from University of Missouri-Columbia in 2001. She is\\ta newly appointed Chair Professor of Emerging Technologies and Future Education at The Education University of Hong Kong. She has also been a professor and program chair of Learning design and technology at San Diego State University, California, USA since 2000. In addition, she serves as the Editor-in-Chief for the IEEE Transactions on Learning Technologies (TLT). Her research specialties are multidisciplinary, focusing on STEM education, new and emerging technologies in various educational settings, Metaverse and immersive learning, and\\tthe design and implementation of artificial intelligence including AIGC for education and training. She is an internationally recognized scholar and has keynoted about 45 international conferences. She is recognized internationally for her research, publishing and dedicated service to IEEE and other scholarly communities. Dr. Wang is a member of IEEE and the IEEE Education Society. She also co-chairs the Education Society’s technical committee for immersive learning (TC-ILE). \\n\\\\end{IEEEbiography}\\n\\\\vspace{-0.6cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{Photos/chen.pdf}}]\\n\\t{Enhong Chen} (Fellow, IEEE) received the B.S. degree from Anhui University, Hefei, China, the M.S. degree from the Hefei University of Technology, Hefei, China, and the Ph.D. degree in computer science from the University of Science and Technology of China (USTC), Hefei, China, in 1989, 1992 and 1996 respectively. He is currently a Professor and the Executive Dean of School of Data Science, the Director of Anhui Province Key Laboratory of Big Data Analysis and Application. He has published a number of papers on refereed journals and conferences, such as TKDE, TIST, TMC, KDD, ICDM, NIPS and CIKM. His current research interests include data mining and machine learning, social network analysis, and recommender system. Dr. Chen was a recipient of the National Science Fund for Distinguished Young Scholars of China, the Best Application Paper Award on KDD-2008, the Best Student Paper Award on KDD-2018 (Research), and the Best Research Paper Award on ICDM-2011. He is an IEEE Fellow and an ACM Distinguished Scientist.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-0.6cm}\\n\\\\end{document}\",\n \"id\": \"8a28f5df-8a26-4509-8bdf-0a8c518eb23c\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"4e5af7af-552b-4c7f-939b-54b6b4dadeb9\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Knowledge Tracing: Models, Variants, and Applications\"\n ],\n [\n \"section\",\n \"Discussions and Conclusions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussions and Conclusions\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":92,"string":"92"},"all_cites":{"kind":"list like","value":[5371,5372,5370,5374,5373,5376,5375,38,5377,8911,1206,5378,5379,5380,679,243,553,5381,5382,166,1684,5383,5384,5385,8912,5386,8913,8914,5388,5387,8915,8461],"string":"[\n 5371,\n 5372,\n 5370,\n 5374,\n 5373,\n 5376,\n 5375,\n 38,\n 5377,\n 8911,\n 1206,\n 5378,\n 5379,\n 5380,\n 679,\n 243,\n 553,\n 5381,\n 5382,\n 166,\n 1684,\n 5383,\n 5384,\n 5385,\n 8912,\n 5386,\n 8913,\n 8914,\n 5388,\n 5387,\n 8915,\n 8461\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.291702768660087,"string":"1.291703"}}},{"rowIdx":60,"cells":{"authors":{"kind":"list like","value":["Yizeng Han","Gao Huang","Shiji Song","Le Yang","Honghui Wang","Yulin Wang"],"string":"[\n \"Yizeng Han\",\n \"Gao Huang\",\n \"Shiji Song\",\n \"Le Yang\",\n \"Honghui Wang\",\n \"Yulin Wang\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Dynamic Neural Networks: A Survey"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-02-09T16:02:00Z","string":"2021-02-09T16:02:00Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" Dynamic neural network is an emerging research topic in deep learning. Compared to static models which have fixed computational graphs and parameters at the inference stage, dynamic networks can adapt their structures or parameters to different inputs, leading to notable advantages in terms of accuracy, computational efficiency, adaptiveness, etc. In this survey, we comprehensively review this rapidly developing area by dividing dynamic networks into three main categories: 1) {\\emph{sample-wise}} dynamic models that process each sample with data-dependent architectures or parameters; 2) \\emph{spatial-wise} dynamic networks that conduct adaptive computation with respect to different spatial locations of image data; and 3) \\emph{temporal-wise} dynamic models that perform adaptive inference along the temporal dimension for sequential data such as videos and texts. The important research problems of dynamic networks, e.g., architecture design, decision making scheme, optimization technique and applications, are reviewed systematically. Finally, we discuss the open problems in this field together with interesting future research directions. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Dynamic Neural Networks: A Survey"]],"subsections":["ed4f3055-e5dc-420d-8009-404bb4d82113","58436ad0-55e1-4da8-b7e8-6bf9edce3fb1","d1dfc651-19ab-46ab-9db8-02eef3371ca0","8a8770e6-47f0-428a-8e00-4bb401ff006f","4ffefefb-4651-456b-a07d-75e465a728b0","3cd25c7f-359c-4b70-8c02-f2738746b35c","e9013abd-3d22-4eab-a882-4ba5e86719f2"],"title":"root"},{"cite_extract_rate":0.642857142857142,"cites":[71,96,679,681,687,544,305,682,683,675,38,504,97,686,684,680,677,676,7266,8372,7268,7,678,7267,674,685,306],"content":"\\label{sec:introduction}}\n\\IEEEPARstart{D}{eep} neural networks (DNNs) are playing an important role in various areas including computer vision (CV) and natural language processing (NLP) . \nRecent years have witnessed many successful deep models such as AlexNet , VGG , GoogleNet , ResNet , DenseNet and Transformer . These {architectural innovations} have enabled the training of deeper, more accurate and more efficient models. The recent research on neural architecture search (NAS) further speeds up the process of designing more powerful structures. However, most of the prevalent deep learning models perform inference in a static manner, i.e., both the computational graph and the network parameters are fixed once trained, which may limit their representation power, efficiency and interpretability . \nDynamic networks, as opposed to static ones, can adapt their structures or parameters to the input during inference, and therefore enjoy favorable properties that are absent in static models. In general, dynamic computation in the context of deep learning has the following advantages: \n\\begin{figure*}\n \\centering\n \\vspace{-2ex}\n \\includegraphics[width=0.875\\linewidth]{1_overall.pdf}\n \\vskip -0.175in\n \\centering\\caption{Overview of the survey. {We first review the dynamic networks that perform adaptive computation at three different granularities (i.e. sample-wise, spatial-wise and temporal-wise). Then we summarize the decision making strategy, training technique and applications of dynamic models. Existing open problems in this field together with some future research directions are finally discussed. Best viewed in color.}}\n \\label{framework}\n \\vspace{-3ex}\n\\end{figure*}\n\\textbf{1) Efficiency.} One of the most notable advantages of dynamic networks is that they are able to allocate computations on demand at test time, by selectively activating model components (e.g. layers , channels or sub-networks ) \\emph{conditioned} on the input. Consequently, less computation is spent on canonical samples that are relatively easy to recognize, or on less informative spatial/temporal locations of an input. {In addition to \\emph{computational efficiency}, dynamic models have also shown promising results for improving \\emph{data efficiency} in the scenario of few-shot learning .}\n\\textbf{2) Representation power.} Due to the data-dependent network architecture/parameters, dynamic networks have significantly enlarged parameter space and improved representation power. For example, with a minor increase of computation, model capacity can be boosted by applying feature-conditioned attention weights on an ensemble of convolutional kernels . It is worth noting that the popular soft attention mechanism could also be unified in the framework of dynamic networks, as different channels , spatial areas or temporal locations of features are dynamically re-weighted at test time. \n\\textbf{3) Adaptiveness.} Dynamic models are able to achieve a desired trade-off between accuracy and efficiency for dealing with varying computational budgets on the fly. Therefore, they are more adaptable to different hardware platforms and changing environments, compared to static models with a fixed computational cost.\n\\textbf{4) Compatibility.} Dynamic networks are compatible with most advanced techniques in deep learning, including architecture design , optimization algorithms and data preprocessing , which ensures that they can benefit from the most recent advances in the field to achieve state-of-the-art performance. For example, dynamic networks can inherit architectural innovations in lightweight models , or be designed via NAS approaches . Their efficiency could also be further improved by acceleration methods developed for static models, such as network pruning , weight quantization , knowledge distillation and low-rank approximation .\n\\textbf{5) Generality.} As a substitute for static deep learning techniques, many dynamic models are general approaches that can be applied seamlessly to a wide range of applications, such as image classification , object detection and semantic segmentation . Moreover, the techniques developed in CV tasks are proven to transfer well to language models in NLP tasks , and vice versa.\n\\textbf{6) Interpretability.} We finally note that the research on dynamic networks {may bridge} the gap between the underlying mechanism of deep models and brains, as it is believed that the brains process information in a dynamic way~. It is possible to analyze which components of a dynamic model are activated when processing an input sample, and to observe which parts of the input are accountable for certain predictions . These properties may shed light on interpreting the decision process of DNNs.\n\\begin{table}\n \\vspace{-1ex}\n \\caption{Notations used in this paper.}\n \\label{tab:notations}\n \\vspace{-4ex}\n \\begin{center}\n \\begin{tabular}{c|c}\n \\hline\n \\textbf{Notations} & \\textbf{Descriptions} \\\\\n \\hline\n $\\mathbb{R}^m$ & $m$-dimensional real number domain \\\\\n \\hline\n $a, \\mathbf{a}$ & Scalar, vector/matrix/tensor \\\\\n \\hline\n $\\mathbf{x,y}$ & Input, output feature \\\\\n \\hline\n $\\mathbf{x}^{\\ell}$ & Feature at layer $\\ell$ \\\\\n \\hline\n $\\mathbf{h}_t$ & Hidden state at time step $t$ \\\\\n \\hline\n $\\mathbf{x(p)}$ & Feature at spatial location $\\mathbf{p}$ on $\\mathbf{x}$ \\\\\n \\hline\n $\\bm{\\Theta}$ & Learnable parameter \\\\\n \\hline\n $\\bm{\\hat{\\Theta}}|\\mathbf{x}$ & Dynamic parameter conditioned on $\\mathbf{x}$ \\\\\n \\hline\n $\\mathbf{x}\\star\\mathbf{W}$ & Convolution of feature $\\mathbf{x}$ and weight $\\mathbf{W}$ \\\\\n \\hline\n $\\otimes$ & Channel-wise or element-wise multiplication \\\\\n \\hline\n $\\mathcal{F}(\\cdot,\\bm{\\Theta})$ & Functional Operation parameterized by $\\bm{\\Theta}$ \\\\\n \\hline\n $\\mathcal{F}\\circ\\mathcal{G}$ & Composition of function $\\mathcal{F}$ and $\\mathcal{G}$ \\\\\n \\hline\n \\end{tabular}\n \\end{center}\n \\vskip -0.35in\n\\end{table}\nIn fact, adaptive inference, the key idea underlying dynamic networks, has been studied before the popularity of modern DNNs. The most classical approach is building a model ensemble through a cascaded or parallel structure, and selectively activating the models conditioned on the input. Spiking neural networks (SNNs) also perform data-dependent inference by propagating pulse signals. However, the training strategy for SNNs is quite different from that of popular convolutional neural networks (CNNs), and they are less used in vision tasks. Therefore, we leave out the work related to SNNs in this survey. \n\\begin{figure}\n \\centering\n \\vspace{-1ex}\n \\includegraphics[width=\\linewidth]{2_dynamic_cascade.pdf}\n \\vskip -0.15in\n \\caption{Two early-exiting schemes. The dashed lines and shaded modules are not executed, conditioned on the decisions made by the routers.}\n \\label{cascading_models}\n \\vskip -0.2in\n\\end{figure}\n\\begin{figure*}\n \\centering\n \\vspace{-2ex}\n \\includegraphics[width=\\linewidth]{3_multi_scale_SuperNet.pdf}\n \\vskip -0.175in\n \\caption{{Multi-scale architectures with dynamic inference graphs. The first three models (a, b, c) perform adaptive early exiting with specific architecture designs and exiting policies. Dynamic routing is achieved inside a SuperNet (d) to activate data-dependent inference paths.}}\n \\label{multi_scale}\n \\vspace{-3ex}\n\\end{figure*}\nIn the context of deep learning, dynamic inference with modern deep architectures, has raised many new research questions and has attracted great research interests in the past three years.\nDespite the extensive work on designing various types of dynamic networks, a systematic and comprehensive review on this topic is still lacking. This motivates us to write this survey, to review the recent advances in this rapidly developing area, with the purposes of 1) providing an overview as well as new perspectives for researchers who are interested in this topic; 2) pointing out the close relations of different subareas and reducing the risk of reinventing the wheel; and 3) summarizing the key challenges and possible future research directions.\n{This survey is organized as follows (see \\figurename~\\ref{framework} for an overview). In Sec. \\ref{sec_sample_wise}, we introduce the most common \\emph{sample-wise} dynamic networks which adapt their architectures or parameters conditioned on each input sample. Dynamic models working on a finer granularity, i.e., \\emph{spatially} adaptive and \\emph{temporally} adaptive models, are reviewed in Sec. \\ref{sec_spatially_adaptive} and Sec.\\ref{sec_temporal_adaptive}, respectively\\footnote{{These two categories can also be viewed as sample-wise dynamic networks as they perform adaptive computation within each sample at a finer granularity, and we adopt such a split for narrative convenience.}}. Then we investigate the decision making strategies and the training techniques of dynamic networks in Sec. \\ref{inference_and_train}. The applications of dynamic models are further summarized in Sec. \\ref{sec_tasks}. Finally, we conclude this survey with a discussion on a number of open problems and future research directions in Sec. \\ref{sec:discussion}. For better readability, we list the notations that will be used in this survey in Table \\ref{tab:notations}.}\n\\vspace{-2ex}","id":"ed4f3055-e5dc-420d-8009-404bb4d82113","level":"section","origin_cites_number":42,"parent_id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"}\n\\label{sec_sample_wise}\nAiming at processing different inputs in data-dependent manners, sample-wise dynamic networks are typically designed from two perspectives: 1) adjusting model architectures to allocate appropriate computation based on each sample, and therefore reducing redundant computation for increased efficiency (Sec. \\ref{dynamic_arch}); \n2) adapting network \\emph{parameters} to every input sample with fixed computational graphs, with the goal of boosting the representation power with minimal increase of computational cost (Sec. \\ref{adaptive_params}).\n\\vspace{-2ex}","id":"58436ad0-55e1-4da8-b7e8-6bf9edce3fb1","level":"section","origin_cites_number":0,"parent_id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"]],"subsections":["3befdcee-686f-4399-9d86-c1c11ef02509","fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4"],"title":"{Sample-wise Dynamic Networks"},{"cite_extract_rate":1,"cites":[688,685,687],"content":"\\label{dynamic_arch}\nConsidering different inputs may have diverse computational demands, it is natural to perform inference with dynamic architectures conditioned on each sample. Specifically, one can adjust the network depth (Sec. \\ref{dynamic_depth}), width (Sec. \\ref{dynamic_width}), or perform dynamic routing within a super network (SuperNet) that includes multiple possible paths (Sec. \\ref{SuperNets}). Networks with dynamic architectures not only save redundant computation for canonical (\"easy\") samples, but also preserve their representation power when recognizing non-canonical (\"hard\") samples. Such a property leads to remarkable advantages in efficiency compared to the acceleration techniques for static models , which handle \"easy\" and \"hard\" inputs with identical computation, and fail to reduce intrinsic computational redundancy.\n\\vspace{-1.5ex}","id":"3befdcee-686f-4399-9d86-c1c11ef02509","level":"subsection","origin_cites_number":3,"parent_id":"58436ad0-55e1-4da8-b7e8-6bf9edce3fb1","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Architectures"]],"subsections":["ca8e1af0-84b8-4669-8aed-b0ba1c15924a","9747d6b0-66bd-4364-a0a9-4ece0a43fe6e","ad51ce7b-9d07-4edc-88e3-0822d3ae0397"],"title":"Dynamic Architectures"},{"cite_extract_rate":0.621621621621621,"cites":[698,695,7273,305,693,683,690,691,7269,8373,38,7270,697,97,692,694,696,676,7272,7271,7266,689,7],"content":"\\label{dynamic_depth}\nAs modern DNNs are getting increasingly deep for recognizing more \"hard\" samples, a straightforward solution to reducing redundant computation is performing inference with dynamic depth, which can be realized by 1) \\emph{early exiting}, i.e. allowing \"easy\" samples to be output at shallow exits without executing deeper layers ; or 2) \\emph{layer skipping}, i.e. selectively skipping intermediate layers conditioned on each sample .\n\\begin{figure*}\n \\vspace{-2ex}\n \\centering\n \\includegraphics[width=\\linewidth]{4_skip_layers.pdf}\n \\vskip -0.175in\n \\caption{Dynamic layer skipping. Feature $\\mathbf{x}_4$ in (a) are not calculated conditioned on the halting score, and the gating module in (b) decides whether to execute the block based on the intermediate feature. The policy network in (c) generates the skipping decisions for all layers in the main network.}\n \\label{skipping_layers}\n \\vspace{-3ex}\n\\end{figure*}\n\\noindent\\textbf{1) Early exiting.} The complexity (or \"difficulty\") of input samples varies in most real-world scenarios, and shallow networks are capable of correctly identifying many canonical inputs. Ideally, these samples should be output at certain early exits without executing deeper layers.\nFor an input sample $\\mathbf{x}$, the forward propagation of an $L$-layer deep network $\\mathcal{F}$ could be represented by \n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{y} = \\mathcal{F}^{L}\\circ\\mathcal{F}^{L-1}\\circ\\cdots\\circ\\mathcal{F}^1(\\mathbf{x}),\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\n\\noindent where $\\mathcal{F}^{\\ell}$ denotes the operational function at layer $\\ell,1\\!\\le \\ell\\!\\le L$. In contrast, early exiting allows to terminate the inference procedure at an intermediate layer. For the $i$-th input sample $\\mathbf{x}_i$, the forward propagation can be written as\n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{y}_i = \\mathcal{F}^{\\ell_i}\\circ\\mathcal{F}^{\\ell_i-1}\\circ\\cdots\\circ\\mathcal{F}^1(\\mathbf{x}_i), 1\\!\\le \\ell_i\\!\\le L.\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nNote that $\\ell_i$ is adaptively determined based on $\\mathbf{x}_i$. Extensive architectures have been studied to endow DNNs with such early exiting behaviors, as discussed in the following.\na) \\emph{Cascading of DNNs.} The most intuitive approach to enabling early exiting is cascading multiple models (see \\figurename~\\ref{cascading_models} (a)), and adaptively retrieving the prediction of an early network without activating latter ones. For example, Big/little-Net cascades two CNNs with different depths. After obtaining the \\emph{SoftMax} output of the first model, early exiting is conducted when the score margin between the two largest elements exceeds a threshold. {Moreover, a number of classic CNNs are cascaded in and }. After each model, a decision function is trained to determine whether the obtained feature should be fed to a linear classifier for immediate prediction, or be sent to subsequent models.\nb) \\emph{Intermediate classifiers.}\nThe models in the aforementioned cascading structures are mutually independent. Consequently, \nonce a \"difficult\" sample is decided to be fed to a latter network, a whole inference procedure needs to be executed from scratch without reusing the already learned features. A more compact design is involving intermediate classifiers within one backbone network (see \\figurename~\\ref{cascading_models} (b)), so that early features can be propagated to deep layers if needed. Based on such a multi-exit architecture, adaptive early exiting is typically achieved according to confidence-based criteria or learned functions .\nc) \\emph{Multi-scale architecture with early exits.} Researchers have observed that in chain-structured networks, the multiple classifiers may interfere with each other, which degrades the overall performance. A reasonable interpretation could be that in regular CNNs, the high-resolution features lack the coarse-level information that is essential for classification, leading to unsatisfying results for early exits. Moreover, early classifiers would force the shallow layers to generate \\emph{task-specialized} features, while a part of \\emph{general} information is lost, resulting in degraded performance for deep exits. To tackle this issue, multi-scale dense network (MSDNet) adopts 1) a \\emph{multi-scale} architecture, {which consists of multiple sub-networks for processing feature maps with different resolutions (scales)}, to quickly generate coarse-level features that are suitable for classification; 2) \\emph{dense connections}, to reuse early features and improve the performance of deep classifiers (see \\figurename~\\ref{multi_scale} (a)). Such a specially-designed architecture effectively enhances the overall accuracy of all the classifiers in the network. \n{Based on the multi-scale architecture design, researchers have also studied the exiting policies (see \\figurename~\\ref{multi_scale} (b)) and training schemes of early-exiting dynamic models. More discussion about the inference and training schemes for dynamic models will be presented in Sec. \\ref{inference_and_train}.}\nPrevious methods typically achieve the adaptation of network depths. From the perspective of exploiting spatial redundancy in features, resolution adaptive network (RANet, see \\figurename~\\ref{multi_scale} (c)) first processes each sample with low-resolution features, while high-resolution representations are conditionally utilized based on early predictions.\nAdaptive early exiting is also extended to language models (e.g. BERT ) for improving their efficiency on NLP tasks . In addition, it can be implemented in recurrent neural networks (RNNs) for \\emph{temporally} dynamic inference when processing sequential data such as videos and texts (see Sec. \\ref{sec_temporal_adaptive}).\n\\noindent\\textbf{2) Layer skipping.} The general idea of the aforementioned early-exiting paradigm is skipping the execution of all the deep layers after a certain classifier. More flexibly, the network depth can also be adapted on the fly by strategically skipping the calculation of \\emph{intermediate layers} without placing extra classifiers. Given the $i$-th input sample $\\mathbf{x}_i$, dynamic layer skipping could be generally written as\n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{y}_i = (\\mathds{1}^{L}\\circ\\mathcal{F}^{L})\\circ (\\mathds{1}^{L-1}\\circ\\mathcal{F}^{L-1})\\circ\\cdots\\circ (\\mathds{1}^1\\circ\\mathcal{F}^1)(\\mathbf{x}_i),\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nwhere $\\mathds{1}^{\\ell}$ denotes the indicator function determining the execution of layer $\\mathcal{F}^{\\ell}, 1\\!\\le\\!\\ell\\!\\le\\!L$.\nThis scheme is typically implemented on structures with skip connections (e.g. ResNet ) to guarantee the continuity of forward propagation, and here we summarize three common approaches.\na) \\emph{Halting score} is first proposed in , where an accumulated scalar named halting score adaptively decides whether the hidden state of an RNN will be directly fed to the next time step. The halting scheme is extended to vision tasks by viewing residual blocks within a ResNet stage \\footnote{Here we refer to a stage as a stack of multiple residual blocks with the same feature resolution.} as linear layers within a step of RNN (see \\figurename~\\ref{skipping_layers} (a)).\nRather than skipping the execution of layers with independent parameters, multiple blocks in each ResNet stage could be replaced by one weight-sharing block, leading to a significant reduction of parameters . In every stage, the block is executed for an adaptive number of steps according to the halting score.\nIn addition to RNNs and CNNs, the halting scheme is further implemented on Transformers by and to achieve dynamic network depth on NLP tasks.\nb) {\\emph{Gating function} is also a prevalent option for dynamic layer skipping due to its plug-and-play property.} Take ResNet as an example (see \\figurename~\\ref{skipping_layers} (b)), let $\\mathbf{x}^{\\ell}$ denote the input feature of the $\\ell$-th residual block, gating function $\\mathcal{G}^{\\ell}$ generates a binary value to decide the execution of residual block $\\mathcal{F}^{\\ell}$. This procedure could be represented by\\footnote{For simplicity and without generality, the subscript for sample index will be omitted in the following.}\n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{x}^{\\ell+1} = \\mathcal{G}^{\\ell}(\\mathbf{x}^{\\ell})\\mathcal{F}^{\\ell}(\\mathbf{x}^{\\ell}) + \\mathbf{x}^{\\ell}, \\mathcal{G}^{\\ell}(\\mathbf{x}^{\\ell})\\in\\{0,1\\}.\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nSkipNet and convolutional network with adaptive inference graph (Conv-AIG) are two typical approaches to enabling dynamic layer skipping. Both methods induce lightweight computational overheads to efficiently produce the binary decisions on whether skipping the calculation of a residual block. Specifically, Conv-AIG utilizes two FC layers in each residual block, while the gating function in SkipNet is implemented as an RNN for parameter sharing.\n\\begin{figure*}\n \\centering\n \\vspace{-2ex}\n \\includegraphics[width=0.85\\linewidth]{5_multi_branch.pdf}\n \\vskip -0.175in\n \\caption{MoE with soft weighting (a) and hard gating (b) schemes both adopt an auxiliary module to generate the weights or gates. In the tree structure (c), features (nodes) and transformations (paths) are represented as circles and lines with arrows respectively. Only the full lines are activated.}\n \\label{multi_branch}\n \\vspace{-3ex}\n\\end{figure*}\nRather than skipping layers in classic ResNets, dynamic recursive network iteratively executes one block with shared parameters in each stage. Although the weight-sharing scheme seems similar to the aforementioned IamNN , the skipping policy of differs significantly: gating modules are exploited to decide the recursion depth.\nInstead of either skipping a layer, or executing it thoroughly with a full numerical precision, a line of work studies adaptive \\emph{bit-width} for different layers conditioned on the \\emph{resource budget}. Furthermore, fractional skipping adaptively selects a bit-width for each residual block by a gating function based on \\emph{input features}.\nc) {\\emph{Policy network} can be built to take in an input sample, and directly produces the skipping decisions for all the layers in a backbone network (see \\figurename~\\ref{skipping_layers} (c)).}\n\\vspace{-1.5ex}","id":"ca8e1af0-84b8-4669-8aed-b0ba1c15924a","level":"subsubsection","origin_cites_number":37,"parent_id":"3befdcee-686f-4399-9d86-c1c11ef02509","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Architectures"],["subsubsection","Dynamic Depth"]],"subsections":[],"title":"Dynamic Depth"},{"cite_extract_rate":0.764705882352941,"cites":[706,688,705,707,8375,544,691,8374,710,8373,712,713,709,699,38,711,703,680,696,702,7266,700,701,708,704,687],"content":"\\label{dynamic_width}\nIn addition to dynamic network \\emph{depth} (Sec. \\ref{dynamic_depth}), a finer-grained form of conditional computation is performing inference with dynamic \\emph{width}: although every layer is executed, its multiple units (e.g. neurons, channels or branches) are selectively activated conditioned on the input.\n\\noindent\\textbf{1) {Skipping neurons in fully-connected (FC) layers.}} The computational cost of an FC layer is determined by its input and output dimensions. It is commonly believed that different neuron units are responsible for representing different features, and therefore not all of them need to be activated for every sample. Early studies learn to adaptively control the neuron activations by auxiliary branches or other techniques such as low-rank approximation .\n\\noindent\\textbf{2) {Skipping branches in mixture-of-experts (MoE).}} In Sec. \\ref{dynamic_depth}, adaptive model ensemble is achieved via a \\emph{cascading} way, and later networks are conditionally executed based on early predictions. An alternative approach to improving the model capacity is the MoE structure, which means that multiple network branches are built as experts in \\emph{parallel}. \nThese experts could be selectively executed, and their outputs are fused with data-dependent weights.\nConventional \\emph{soft} MoEs adopt real-valued weights to dynamically rescale the representations obtained from different experts (\\figurename~\\ref{multi_branch} (a)). In this way, all the branches still need to be executed, and thus the computation cannot be reduced at test time. \\emph{Hard} gates with only a fraction of non-zero elements are developed to increase the inference efficiency of the MoE structure (see \\figurename~\\ref{multi_branch} (b)) : let $\\mathcal{G}$ denote a gating module whose output is a $N$-dimensional vector $\\bm{\\alpha}$ controlling the execution of $N$ experts $\\mathcal{F}_1,\\mathcal{F}_2, \\cdots, \\mathcal{F}_N$, the final output can be written as\n\\begin{equation} \\label{eq_moe}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{y} = \\sum\\nolimits_{n=1}^N [\\mathcal{G}(\\mathbf{x})]_n \\mathcal{F}_n (\\mathbf{x}) = \\sum\\nolimits_{n=1}^N \\alpha_n \\mathcal{F}_n (\\mathbf{x}),\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nand the $n$-th expert will not be executed if $\\alpha_n\\! =\\!0$. \nHard MoE has been implemented in diverse network structures. For example, HydraNet replaces the convolutional blocks in a CNN by multiple branches, and selectively execute them conditioned on the input. For another example, dynamic routing network (DRNet) performs a branch selection in each cell structure which is commonly used in NAS . On NLP tasks, sparely gated MoE and switch Transformer embeds hard MoE in a long short-term memory (LSTM) network and a Transformer , respectively. Instead of making choice with \\emph{binary} gates as in , only the branches corresponding to the \\emph{top-K} elements of the real-valued gates are activated in .\n\\noindent\\textbf{{3) Skipping channels in CNNs.}}\nModern CNNs usually have considerable channel redundancy. Based on the common belief that the same feature channel can be of disparate importance for different samples, adaptive width of CNNs could be realized by dynamically activating convolutional channels. Compared to the \\emph{static} pruning methods which remove \"unimportant\" filters permanently, such a \\emph{data-dependent} pruning approach improves the inference efficiency without degrading the model capacity.\na) \\emph{Multi-stage architectures along the channel dimension.} Recall that the early-exiting networks discussed in Sec. \\ref{dynamic_depth} can be viewed as multi-stage models along the \\emph{depth} dimension, where late stages are conditionally executed based on early predictions. One can also build multi-stage architectures along the \\emph{width} (channel) dimension, and progressively execute these stages on demand.\nAlong this direction, an optimal architecture is searched among multiple structures with different widths, and any sample can be output at an early stage when a confident prediction is obtained . Channel gating network (CGNet) first executes a subset of convolutional filters in every layer, and the remaining filters are only activated on strategically selected areas.\nb) \\emph{Dynamic pruning with gating functions.} In the aforementioned progressive activation paradigm, the execution of a later stage is decided based on previous output. As a result, a complete forward propagation is required for every stage, which might be suboptimal for reducing the practical inference latency. Another prevalent solution is to decide the execution of channels at every layer by gating functions. For example, runtime neural pruning (RNP) models the layer-wise pruning as a Markov decision process, and an RNN is used to select specific channel groups at every layer. Moreover, pooling operations followed by FC layers are utilized to generate {\\emph{channel-wise hard attention} (i.e. making discrete decisions on whether to activate each channel)} for each sample . The recent work uses a gate module to decide the width for a whole stage of a ResNet. Different reparameterization and optimizing techniques are required for training these gating functions, which will be reviewed in Sec. \\ref{training}.\nRather than placing plug-in gating modules \\emph{inside} a CNN, GaterNet builds an \\emph{extra} network, which takes in the input sample and directly generates all the channel selection decisions for the backbone CNN. This implementation is similar to BlockDrop that exploits an additional policy network for dynamic layer skipping (Sec. \\ref{dynamic_depth}).\nc) \\emph{Dynamic pruning based on feature activations directly} has also been realized without auxiliary branches and computational overheads , where a regularization item is induced in training to encourage the sparsity of features.\n{On basis of the existing literature on dynamically skipping either network \\emph{layers} or convolutional \\emph{filters} , recent work has realized dynamic inference with respect to network \\emph{depth} and \\emph{width} simultaneously: only if a layer is determined to be executed, its channels will be selectively activated, leading to a more flexible adaptation of computational graphs.}\n\\vspace{-1ex}","id":"9747d6b0-66bd-4364-a0a9-4ece0a43fe6e","level":"subsubsection","origin_cites_number":34,"parent_id":"3befdcee-686f-4399-9d86-c1c11ef02509","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Architectures"],["subsubsection","Dynamic Width"]],"subsections":[],"title":"Dynamic Width"},{"cite_extract_rate":0.684210526315789,"cites":[715,714,8374,8373,719,720,718,696,7266,7274,716,717,306],"content":"\\label{SuperNets}\nThe aforementioned methods mostly adjust the depth (Sec. \\ref{dynamic_depth}) or width (Sec. \\ref{dynamic_width}) of \\emph{classic architectures} by activating their computational units (e.g. layers or channels ) conditioned on the input. {Another line of work develops different forms of SuperNets with various possible inference paths, and performs dynamic routing inside the SuperNets to adapt the computational graph to each sample.}\nTo achieve dynamic routing, there are typically routing nodes that are responsible for allocating features to different paths. For node $s$ at layer $\\ell$, let $\\alpha_{s\\rightarrow j}^{\\ell}$ denote the probability of assigning the reached feature $\\mathbf{x}^{\\ell}_s$ to node $j$ at layer $\\ell+1$, the path $\\mathcal{F}_{s\\rightarrow j}^{\\ell}$ will be activated only when $\\alpha_{s\\rightarrow j}^{\\ell}\\!>\\!0$. The resulting feature reaching node $j$ is represented by\n\\begin{equation}\\label{eq_supernet}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{x}^{\\ell+1}_j = \\sum\\nolimits_{s\\in\\left\\{s: \\alpha_{s\\rightarrow j}^{\\ell}>0\\right\\}} \\alpha_{s\\rightarrow j}^{\\ell}\\mathcal{F}_{s\\rightarrow j}^{\\ell}(\\mathbf{x}^{\\ell}_s).\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nThe probability $\\alpha_{s\\rightarrow j}^{\\ell}$ can be obtained in different manners. \nNote that the dynamic early-exiting networks are a special form of SuperNets, where the routing decisions are only made at intermediate classifiers. The CapsuleNets also perform dynamic routing between capsules, i.e. groups of neurons, to character the relations between (parts of) objects. Here we mainly focus on specific architecture designs of the SuperNets and their routing policies.\n\\noindent\\textbf{1) Path selection in multi-branch structures.} The simplest dynamic routing can be implemented by selectively executing \\emph{one} of multiple candidate modules at each layer , which is equivalent to producing a one-hot probability distribution $\\alpha_{s\\rightarrow \\cdot}^{\\ell}$ in Eq. \\ref{eq_supernet}. The main difference of this approach to hard MoE (\\figurename~\\ref{multi_branch} (b)) is that only one branch is activated without any fusion operations.\n\\noindent\\textbf{2) Neural trees and tree-structured networks.} \nAs decision trees always perform inference along one forward path that is dependent on input properties, combining tree structure with neural networks can naturally enjoy the adaptive inference paradigm and the representation power of DNNs simultaneously. Note that in a tree structure, the outputs of different nodes are routed to \\emph{independent} paths rather than being \\emph{fused} as in MoE structures (compare \\figurename~\\ref{multi_branch} (b), (c)).\na) \\emph{Soft decision tree} (SDT) adopts neural units as routing nodes (blue nodes in \\figurename~\\ref{multi_branch} (c)), which decides the portion that the inputs are assigned to their left/right sub-tree. Each leaf node generates a probability distribution over the output space, and the final prediction is the expectation of the results from all leaf nodes. Although the probability for a sample reaching each leaf node in an SDT is data-dependent, all the paths are still executed, which limits the inference efficiency.\nb) \\emph{Neural trees with deterministic routing policies} are designed to make hard routing decisions during inference, avoiding computation on those unselected paths.\nc) \\emph{Tree-structured DNNs.} Instead of developing decision trees containing neural units, a line of work builds special network architectures to endow them with the routing behavior of decision trees. For instance, a small CNN is first executed to classify each sample into coarse categories, and specific sub-networks are conditionally activated based on the coarse predictions . A subsequent work not only partitions samples to different sub-networks, but also divides and routes the feature channels.\nDifferent to those networks using neural units only in routing nodes , or routing each sample to pre-designed sub-networks , adaptive neural tree (ANT) adopts CNN modules as feature transformers in a hard neural tree (see lines with arrows in \\figurename~\\ref{multi_branch} (c)), and learns the tree structure together with the network parameters simultaneously in the training stage.\n\\noindent\\textbf{3) Others.} \n{Performing dynamic routing within more general SuperNet architectures is also a recent research trend. Representatively, an architecture distribution with partly shared parameters is \\emph{searched} from a SuperNet containing $\\sim \\!\\! 10^{25}$ sub-networks . During inference, every sample is allocated by a controller network to one sub-network with appropriate computational cost. Instead of training a standalone controller network, gating modules are plugged inside the \\emph{hand-designed} SuperNet (see \\figurename~\\ref{multi_scale} (d)) to decide the routing path based on intermediate features . \n\\begin{figure*}\n \\centering\n \\vspace{-2ex}\n \\includegraphics[width=0.9\\linewidth]{6_dynamic_parameter.pdf}\n \\vskip -0.2in\n \\caption{{Three implementations of dynamic parameters: adjusting (a) or generating (b) the backbone parameters based on the input, and (c) dynamically rescaling the features with the attention mechanism.}}\n \\label{dynamic_params}\n \\vskip -0.2in\n\\end{figure*}\n\\vspace{-1.5ex}","id":"ad51ce7b-9d07-4edc-88e3-0822d3ae0397","level":"subsubsection","origin_cites_number":19,"parent_id":"3befdcee-686f-4399-9d86-c1c11ef02509","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Architectures"],["subsubsection","Dynamic Routing"]],"subsections":[],"title":"Dynamic Routing"},{"cite_extract_rate":0,"cites":[],"content":"\\label{adaptive_params}\n\\vspace{-0.25ex}\nAlthough the networks with dynamic \\emph{architectures} in Sec. \\ref{dynamic_arch} can adapt their inference graphs to each sample and achieve an efficient allocation of computation, they usually have special architecture designs, requiring specific training strategies or careful hyper-parameters tuning (Sec. \\ref{sec:discussion}).\nAnother line of work adapts network \\emph{parameters} to different inputs while keeping the architectures fixed, which has been shown effective in improving the representation power of networks with a minor increase of computational cost. Given an input sample $\\mathbf{x}$, \nthe output of a conventional network (module) with static parameters can be written as $\\mathbf{y}\\! =\\!\\mathcal{F}(\\mathbf{x},\\bm{\\Theta})$. In contrast, the output of a model with dynamic parameters could be represented by\n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{y} = \\mathcal{F}(\\mathbf{x},\\bm{\\hat{\\Theta}}|\\mathbf{x}) = \\mathcal{F}(\\mathbf{x},\\mathcal{W}(\\mathbf{x}, \\bm{\\Theta})),\n \\label{eq:dynamic_parameter}\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nwhere $\\mathcal{W}(\\cdot, \\bm{\\Theta})$ is the operation producing input-dependent parameters, and its design has been extensively explored.\nIn general, the parameter adaptation can be achieved from three aspects (see \\figurename~\\ref{dynamic_params}): 1) adjusting the trained parameters based on the input (Sec. \\ref{dynamic_param_adjust}); 2) directly generating the network parameters from the input (Sec. \\ref{weight_predict}); and 3) rescaling the features with soft attention (Sec. \\ref{sec:attention}).\n\\vspace{-1ex}","id":"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4","level":"subsection","origin_cites_number":0,"parent_id":"58436ad0-55e1-4da8-b7e8-6bf9edce3fb1","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Parameters"]],"subsections":["8a3a1403-da1a-48be-b1e9-80d1325e0909","81d9301c-7d1b-4044-b144-7027a7b2165b","0651cec9-e0a5-4c53-aca3-aa4ed8ac7d94"],"title":"Dynamic Parameters"},{"cite_extract_rate":0.6666666666666661,"cites":[675,8376,721,7276,7275,722],"content":"\\label{dynamic_param_adjust}\n\\vspace{-0.5ex}\nA typical approach to parameter adaptation is adjusting the weights based on their input during inference as presented in \\figurename~\\ref{dynamic_params} (a). This implementation usually evokes little computation to obtain the adjustments, e.g., attention weights or sampling offsets .\n\\noindent\\textbf{1) Attention on weights.} {To improve the representation power without noticeably increasing the computation, \nsoft attention can be performed on multiple convolutional kernels, producing an adaptive ensemble of parameters .}\nAssuming that there are $N$ kernels $\\mathbf{W}_n, n\\! =1,2,\\!\\cdots\\!,N$, such a dynamic convolution \ncan be formulated as \n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n\\mathbf{y} = \\mathbf{x} \\star \\mathbf{\\tilde{W}} = \\mathbf{x} \\star (\\sum\\nolimits_{n=1}^N \\alpha_n\\mathbf{W}_n). \n\\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nThis procedure increases the model capacity yet remains high efficiency, as the result obtained through fusing the outputs of $N$ convolutional branches (as in MoE structures, see \\figurename~\\ref{multi_branch} (a)) is equivalent to that produced by performing once convolution with $\\mathbf{\\tilde{W}}$. However, only $\\sim\\!1/N$ times of computation is consumed in the latter approach.\nWeight adjustment could also be achieved by performing soft attention over the \\emph{spatial locations} of convolutional weights . For example, segmentation-aware convolutional network applies locally masked convolution to aggregate information with larger weights from similar pixels, which are more likely to belong to the same object.\nUnlike that requires a sub-network for feature embedding, pixel-adaptive convolution (PAC) adapts the convolutional weights based on the attention mask generated from the input feature at each layer.\n{Instead of adjusting weights conditioned on every sample itself, meta-neighborhoods adapt the network parameters to each input sample based on its similarity to the neighbors stored in a dictionary.}\n\\noindent\\textbf{2) Kernel shape adaptation.} \nApart from adaptively scaling the weight \\emph{values}, parameter adjustment can also be realized to reshape the convolutional kernels and achieve \\emph{dynamic reception of fields}. Towards this direction, \ndeformable convolutions sample feature pixels from adaptive locations when performing convolution on each pixel. Deformable kernels samples weights in the kernel space to adapt the \\emph{effective} reception field (ERF) while leaving the reception field unchanged. Table \\ref{tab:deform_kernels} summarizes the formulations of the above three methods. \n{Due to their irregular memory access and computation pattern, these kernel shape adaptation approaches typically require customized CUDA kernels for the implementation on GPUs. However, recent literature has shown that the practical efficiency of deformable convolution could be effectively improved by co-designing algorithm and hardware based on embedded devices such as FPGAs .}\n\\begin{table*}\n \\vspace{-2ex}\n \\caption{{Kernel shape adaptation by dynamically sampling feature pixels or convolutional weights .}}\n \\vspace{-4ex}\n \\label{tab:deform_kernels}\n \\begin{center}\n \\begin{tabular}{c|c|c|c}\n \\hline\n \\textbf{Method} & \\textbf{Formulation} & \\textbf{Sampled Target} & \\textbf{Dynamic Mask} \\\\\n \\hline\n Regular Convolution & $\\mathbf{y(p)} = \\sum\\nolimits_{k=1}^K \\mathbf{W}(\\mathbf{p}_k) \\mathbf{x}(\\mathbf{p+p}_k)$ & - & - \\\\\n \\hline\n Deformable ConvNet-v1 & $\\mathbf{y(p)} = \\sum\\nolimits_{k=1}^K \\mathbf{W}(\\mathbf{p}_k) \\mathbf{x}(\\mathbf{p+p}_k+\\Delta \\mathbf{p}_k)$ & Feature map & No \\\\\n Deformable ConvNet-v2 & $\\mathbf{y(p)} = \\sum\\nolimits_{k=1}^K \\mathbf{W}(\\mathbf{p}_k) \\mathbf{x}(\\mathbf{p+p}_k+\\Delta \\mathbf{p}_k)\\Delta \\mathbf{m}_k$ & Feature map & Yes \\\\\n Deformable Kernels & $\\mathbf{y(p)} = \\sum\\nolimits_{k=1}^K \\mathbf{W}(\\mathbf{p}_k+\\Delta \\mathbf{p}_k) \\mathbf{x}(\\mathbf{p+p}_k)$ & Conv kernel & No \\\\\n \\hline\n \\end{tabular}\n \\end{center}\n \\vspace{-5ex}\n\\end{table*}\n\\vspace{-1ex}","id":"8a3a1403-da1a-48be-b1e9-80d1325e0909","level":"subsubsection","origin_cites_number":9,"parent_id":"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Parameters"],["subsubsection","Parameter Adjustment"]],"subsections":[],"title":"Parameter Adjustment"},{"cite_extract_rate":0.785714285714285,"cites":[675,726,7277,724,7215,678,727,725,8377,677,723],"content":"\\label{weight_predict}\n\\vspace{-0.25ex}\nCompared to making modifications on model parameters on the fly (Sec. \\ref{dynamic_param_adjust}), weight prediction is more straightforward: it directly generates (a subset of) input-adaptive parameters with an independent model at test time (see \\figurename~\\ref{dynamic_params} (b)). This idea was first suggested in , where both the weight prediction model and the backbone model were feedforward networks. Recent work has further extended the paradigm to modern network architectures and tasks.\n\\noindent\\textbf{1) General architectures.} Dynamic filter networks (DFN) and HyperNetworks are two classic approaches realizing runtime weight prediction for CNNs and RNNs, respectively. Specifically, a filter generation network is built in DFN to produce the filters for a convolutional layer. As for processing sequential data (e.g. a sentence), the weight matrices of the main RNN are predicted by a smaller one at each time step conditioned on the input (e.g. a word) . WeightNet unifies the dynamic schemes of and by predicting the convolutional weights via simple grouped FC layers, achieving competitive results in terms of the accuracy-FLOPs\\footnote{{Floating point operations, which is widely used as a measure of inference efficiency of deep networks.}} and accuracy-parameters trade-offs.\n{Rather than generating standard \\emph{convolutional} weights, LambdaNetworks learns to predict the weights of \\emph{linear} projections based on the contexts of each pixel together with the relative position embeddings, showing advantages in terms of computational cost and memory footprint.}\n\\noindent\\textbf{2) Task-specific information} \nhas also been exploited to predict model parameters on the fly, enabling dynamic networks to generate task-aware feature embeddings. For example, edge attributes are utilized in to generate filters for graph convolution, and camera perspective is incorporated in to generate weights for image convolution. {Such task-aware weight prediction has been shown effective in improving the data efficiency on many tasks, including visual question answering and few-shot learning .}\n\\vspace{-1ex}","id":"81d9301c-7d1b-4044-b144-7027a7b2165b","level":"subsubsection","origin_cites_number":14,"parent_id":"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Parameters"],["subsubsection","Weight Prediction"]],"subsections":[],"title":"Weight Prediction"},{"cite_extract_rate":0.7619047619047611,"cites":[736,730,729,7278,8378,38,734,8379,733,731,7040,728,735,7268,7,732],"content":"\\label{sec:attention}\n\\vspace{-0.25ex}\nThe main goal of either \\emph{adjusting} (Sec. \\ref{dynamic_param_adjust}) or \\emph{predicting} (Sec. \\ref{weight_predict}) model parameters is producing more dynamic and informative features, and therefore enhancing the representation power of deep networks. A more straightforward solution is rescaling the features with input-dependent soft attention (see \\figurename~\\ref{dynamic_params} (c)), which requires minor modifications on computational graphs. Note that for a linear transformation $\\mathcal{F}$, applying attention $\\bm{\\alpha}$ on the output is equivalent to performing computation with re-weighted parameters, i.e.\n\\begin{equation}\\label{dynamic_feature_key}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathcal{F}(\\mathbf{x},\\bm{\\Theta})\\otimes \\bm{\\alpha}= \\mathcal{F}(\\mathbf{x},\\bm{\\Theta}\\otimes\\bm{\\alpha}).\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\n\\noindent\\textbf{1) Channel-wise attention} is one of the most common soft attention mechanisms. Existing work typically follows \nthe form in squeeze-and-excitation network (SENet) :\n\\begin{equation}\\label{se_attention}\n \\setlength{\\abovedisplayskip}{3pt}\n\\mathbf{\\tilde{y}}\\! = \\!\\mathbf{y} \\otimes \\bm{\\alpha}\\! =\\!\\mathbf{y} \\otimes \\mathcal{A}(\\mathbf{y}), \\bm{\\alpha}\\in\\left[0,1\\right]^C.\n\\setlength{\\belowdisplayskip}{3pt}\n\\end{equation} \nIn Eq. \\ref{se_attention}, $\\mathbf{y}\\! =\\!\\mathbf{x} \\star \\mathbf{W}$ is the output feature of a convolutional layer with $C$ channels, and $\\mathcal{A}(\\cdot)$ is a lightweight function composed of pooling and linear layers for producing $\\bm{\\alpha}$.\nTaking the convolution into account, the procedure can also be written as $\\mathbf{\\tilde{y}}\\! = \\!(\\mathbf{x} \\star \\mathbf{W}) \\otimes \\bm{\\alpha}\\! =\\! \\mathbf{x} \\star (\\mathbf{W} \\otimes \\bm{\\alpha})$, from which we can observe that applying attention on features is equivalent to performing convolution with dynamic weights.\nOther implementations for attention modules have also been developed, including using standard deviation to provide more statistics , or replacing FC layers with efficient 1D convolutions .\nThe empirical performance of three computational graphs for soft attention is studied in : 1) $\\mathbf{\\tilde{y}}\\! =\\!\\mathbf{y}\\otimes \\mathcal{A}(\\mathbf{y})$, 2) $\\mathbf{\\tilde{y}}\\! =\\!\\mathbf{y}\\otimes \\mathcal{A}(\\mathbf{x})$ and 3) $\\mathbf{\\tilde{y}}\\! =\\!\\mathbf{y}\\otimes \\mathcal{A}(\\mathrm{Conv}(\\mathbf{x}))$. It is found that the three forms yield different performance in different backbone networks.\n\\noindent\\textbf{2) Spatial-wise attention}. Spatial locations in features could also be dynamically rescaled with attention to improve the representation power of deep models . Instead of using pooling operations to efficiently gather global information as in channel-wise attention, convolutions are often adopted in spatial-wise attention to encode local information. Moreover, these two types of attention modules can be integrated in one framework (see \\figurename~\\ref{dynamic_params} (c)).\n\\noindent\\textbf{3) Dynamic activation functions.} The aforementioned approaches to generating dynamic features usually apply soft attention before static activation functions. A recent line of work has sought to increase the representation power of models with dynamic activation functions . For instance, \nDY-ReLU replaces ReLU $(\\mathbf{y}_c\\! =\\!\\max (\\mathbf{x}_c, 0))$ with the max value among $N$ linear transformations $\\mathbf{y}_c\\! =\\!\\max_{n} \\left\\{a_c^n \\mathbf{x}_c + b_c^n\\right\\}$, where $c$ is the channel index, and $a_c^n, b_c^n$ are linear coefficients calculated from $\\mathbf{x}$. On many vision tasks, these dynamic activation functions can effectively improve the performance of different network architectures with negligible computational overhead.\nTo summarize, soft attention has been exploited in many fields due to its simplicity and effectiveness. Moreover, it can be incorporated with other methods conveniently. E.g., by replacing the weighting scalar $\\alpha_n$ in Eq. \\ref{eq_moe} with channel-wise or spatial-wise attention, the output of multiple branches with independent kernel sizes or feature resolutions are adaptively fused.\nNote that we leave out the detailed discussion on the self attention mechanism, which is widely studied in both NLP and CV fields to re-weight features based on the similarity between queries and keys at different locations (temporal or spatial). \nReaders who are interested in this topic may refer to review studies . In this survey, we mainly focus on the feature re-weighting scheme in the framework of dynamic inference. \n\\vspace{-2ex}","id":"0651cec9-e0a5-4c53-aca3-aa4ed8ac7d94","level":"subsubsection","origin_cites_number":21,"parent_id":"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Sample-wise Dynamic Networks"],["subsection","Dynamic Parameters"],["subsubsection","Dynamic Features"]],"subsections":[],"title":"Dynamic Features"},{"cite_extract_rate":1,"cites":[737,504],"content":"\\vspace{-0.25ex}\n\\label{sec_spatially_adaptive}\nIn visual learning, it has been found that not all locations contribute equally to the final prediction of CNNs , which suggests that \\emph{spatially} dynamic computation has great potential for reducing computational redundancy.\nIn other words, making a correct prediction may only require processing a fraction of pixels or regions with an adaptive amount of computation. Moreover, based on the observations that low-resolution representations are sufficient to yield decent performance for most inputs , the static CNNs that take in all the input with the same resolution may also induce considerable redundancy.\nTo this end, spatial-wise dynamic networks are built to perform adaptive inference with respect to different spatial locations of images.\nAccording to the granularity of dynamic computation, we further categorize the relevant approaches into three levels: {\\emph{pixel level} (Sec. \\ref{pixel_level}), \\emph{region level} (Sec. \\ref{region_level}) and \\emph{resolution level} (Sec. \\ref{dynamic_resolution}).}\n\\vspace{-2ex}","id":"d1dfc651-19ab-46ab-9db8-02eef3371ca0","level":"section","origin_cites_number":2,"parent_id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"]],"subsections":["bb93a738-d7fa-4ffe-811c-d2e5d3062ddd","0d52450e-77a8-4861-a2eb-00ddb1ece863","ab0c50b9-9d18-4c18-976a-34aae739e7ac"],"title":"Spatial-wise Dynamic Networks"},{"cite_extract_rate":0,"cites":[],"content":"\\vspace{-0.25ex}\n\\label{pixel_level}\nCommonly seen spatial-wise dynamic networks perform adaptive computation at the pixel level. Similar to the categorization in Sec. \\ref{sec_sample_wise}, pixel-level dynamic networks are grouped into two types: models with pixel-specific \\emph{dynamic architectures} (Sec. \\ref{pixel_dynamic_arch}) and \\emph{dynamic parameters} (Sec. \\ref{pixel_dynamic_param}).\n\\vspace{-1ex}","id":"bb93a738-d7fa-4ffe-811c-d2e5d3062ddd","level":"subsection","origin_cites_number":0,"parent_id":"d1dfc651-19ab-46ab-9db8-02eef3371ca0","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Pixel-level Dynamic Networks"]],"subsections":["fa853b1f-afd4-40d2-bed1-be3a1d6cb128","530cae77-537a-4dca-80a6-e035f5680e22"],"title":"Pixel-level Dynamic Networks"},{"cite_extract_rate":0.7272727272727271,"cites":[701,738,7279,7280,676,739,7267,7281],"content":"\\label{pixel_dynamic_arch}\n\\vspace{-0.25ex}\nBased on the common belief that foreground pixels are more informative and computational demanding than those in the background, some dynamic networks learn to adjust their architectures for each pixel. Existing literature generally achieves this by 1) \\emph{dynamic sparse convolution}, which only performs convolutions on a subset of sampled pixels; 2) \\emph{additional refinement}, which strategically allocates extra computation (e.g. layers or channels) on certain spatial positions.\n\\begin{figure}\n \\centering\n \\vspace{-1ex}\n \\includegraphics[width=\\linewidth]{8_spatial_pixel_sparse.pdf}\n \\vskip -0.15in\n \\caption{{Dynamic convolution on selected spatial locations. The 1 elements (black) in the spatial mask determine the pixels (green) that require computation in the output feature map.}}\n \\label{fig_spatial_location_specific}\n \\vspace{-3ex}\n\\end{figure}\n\\noindent\\textbf{1) Dynamic sparse convolution.} To reduce the unnecessary computation on less informative locations, convolution can be performed only on strategically sampled pixels. Existing sampling strategies include 1) making use of the intrinsic sparsity of the input ; 2) predicting the positions of zero elements on the output ; and 3) estimating the saliency of pixels . A typical approach is using an extra branch to generate a spatial mask, determining the execution of convolution on each pixel (see \\figurename~\\ref{fig_spatial_location_specific}). {Pixel-wise dynamic depth could also be achieved based on a halting scheme } (see Sec. \\ref{dynamic_depth}). These dynamic convolutions usually neglect the unselected positions, which might degrade the network performance. Interpolation is utilized in to efficiently fill those locations, therefore alleviating the aforementioned disadvantage.\n\\noindent\\textbf{2) Dynamic additional refinement.} Instead of only sampling certain pixels to perform convolutions, another line of work first conducts relatively cheap computation on the whole feature map, and adaptively activate extra modules on selected pixels for further \\emph{refinement}. Representatively, dynamic capacity network generates coarse features with a shallow model, and salient pixels are sampled based on the gradient information. \nFor these salient pixels, extra layers are applied to extract finer features. Similarly, specific positions are additionally processed by a fraction of convolutional filters in . These methods adapt their network architectures in terms of \\emph{depth} or \\emph{width} at the pixel level, achieving a spatially adaptive allocation of computation.\n{The aforementioned dynamic additional refinement approaches are mainly developed for image classification.} On the semantic segmentation task, pixel-wise \\emph{early exiting} (see also Sec. \\ref{dynamic_depth}) is proposed in , where the pixels with high prediction confidence are output without being processed by deeper layers. PointRend shares a similar idea, and applies additional FC layers on selected pixels with low prediction confidence, which are more likely to be on borders of objects. All these researches demonstrate that by exploiting the spatial redundancy in image data, dynamic computation at the pixel level beyond sample level significantly increases the model efficiency.\n\\vspace{-1ex}","id":"fa853b1f-afd4-40d2-bed1-be3a1d6cb128","level":"subsubsection","origin_cites_number":11,"parent_id":"bb93a738-d7fa-4ffe-811c-d2e5d3062ddd","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Pixel-level Dynamic Networks"],["subsubsection","Pixel-wise Dynamic Architectures"]],"subsections":[],"title":"Pixel-wise Dynamic Architectures"},{"cite_extract_rate":0.8,"cites":[7257,740,741,7268,8376,731,722,7282,7276,742,7278,7275],"content":"\\label{pixel_dynamic_param}\n\\vspace{-0.25ex}\nIn contrast to entirely skipping the convolution operation on a subset of pixels, dynamic networks can also apply data-dependent parameters on different pixels for improved representation power or adaptive reception fields.\n\\noindent\\textbf{1) Dynamic weights.} {Similar to the sample-wise dynamic parameter methods (Sec. \\ref{adaptive_params}), pixel-level dynamic weights are achieved by test-time \\emph{adjustment} , \\emph{prediction} or \\emph{dynamic features} . Take weight prediction as an example,} typical approaches generate an $H\\!\\times\\!W\\!\\times\\! k^2$ kernel map to produce spatially dynamic weights ($H,W$ are the spatial size of the output feature and $k$ is the kernel size). Considering the pixels belonging to the same object may share identical weights, dynamic region-aware convolution (DRConv) generates a segmentation mask for an input image, dividing it into $m$ regions, for each of which a weight generation network is responsible for producing a data-dependent kernel.\n\\noindent\\textbf{2) Dynamic reception fields.} Traditional convolution operations usually have a fixed shape and size of kernels (e.g. the commonly used $3\\!\\times\\!3$ 2D convolution). The resulting uniform reception field across all the layers may have limitations for recognizing objects with varying shapes and sizes. {To tackle this, a line of work learns to adapt the reception field for different feature pixels , as discussed in Sec. \\ref{dynamic_param_adjust}.} Instead of adapting the sampling location of features or kernels, adaptive connected network realizes a dynamic trade-off among self transformation (e.g. $1\\!\\times\\!1$ convolution), local inference (e.g. $3\\!\\times\\!3$ convolution) and global inference (e.g. FC layer). The three branches of outputs are fused with data-dependent weighted summation. \nBesides images, the local and global information in non-Euclidean data, such as graphs, could also be adaptively aggregated.\n\\vspace{-1.5ex}","id":"530cae77-537a-4dca-80a6-e035f5680e22","level":"subsubsection","origin_cites_number":15,"parent_id":"bb93a738-d7fa-4ffe-811c-d2e5d3062ddd","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Pixel-level Dynamic Networks"],["subsubsection","Pixel-wise Dynamic Parameters"]],"subsections":[],"title":"Pixel-wise Dynamic Parameters"},{"cite_extract_rate":1,"cites":[7281],"content":"\\label{region_level}\n\\vspace{-0.5ex}\n{Pixel-level dynamic networks mentioned in Sec. \\ref{pixel_level} often require specific implementations for sparse computation, and consequently may face challenges in terms of achieving real acceleration on hardware .} An alternative approach is performing adaptive inference on \\emph{regions/patches} of input images. There mainly exists two lines of work along this direction (see \\figurename~\\ref{fig_spatial_select_regions}): one performs parameterized \\emph{transformations} on a region of feature maps for more accurate prediction (Sec. \\ref{dynamic_transofrm}), and the other learns patch-level \\emph{hard attention}, with the goal of improving the effectiveness and/or efficiency of models (Sec. \\ref{hard_attention_pathces}).\n\\vspace{-1ex}","id":"0d52450e-77a8-4861-a2eb-00ddb1ece863","level":"subsection","origin_cites_number":1,"parent_id":"d1dfc651-19ab-46ab-9db8-02eef3371ca0","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Region-level Dynamic Networks"]],"subsections":["b67eaf7b-932e-433e-a7e5-732f3b779d92","27095343-62ad-48ed-a110-4d5657e3a267"],"title":"Region-level Dynamic Networks"},{"cite_extract_rate":1,"cites":[743,531],"content":"\\label{dynamic_transofrm}\n\\vspace{-0.25ex}\nDynamic transformations (e.g. affine/projective/thin plate spline transformation) can be performed on images to undo certain variations for better generalization ability, or to exaggerate the salient regions for discriminative feature representation. For example, spatial transformer adopts a localization network to generate the transformation parameters, and then applies the parameterized transformation to recover the input from the corresponding variations. \nMoreover, transformations are learned to adaptively zoom-in the salient regions on some tasks where the model performance is sensitive to a small portion of regions. \n\\begin{figure}\n \\centering\n \\vspace{-1ex}\n \\includegraphics[width=\\linewidth]{9_spatial_select_regions.pdf}\n \\vskip -0.15in\n \\caption{{Region-level dynamic inference. The region selection module generates the transformation/localization parameters, and the subsequent network performs inference on the transformed/cropped region.}}\n \\label{fig_spatial_select_regions}\n \\vspace{-4ex}\n\\end{figure}\n\\vspace{-1ex}","id":"b67eaf7b-932e-433e-a7e5-732f3b779d92","level":"subsubsection","origin_cites_number":2,"parent_id":"0d52450e-77a8-4861-a2eb-00ddb1ece863","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Region-level Dynamic Networks"],["subsubsection","Dynamic Transformations"]],"subsections":[],"title":"Dynamic Transformations"},{"cite_extract_rate":0.8571428571428571,"cites":[747,745,737,744,674,746],"content":"\\label{hard_attention_pathces}\n\\vspace{-0.25ex}\nInspired by the fact that informative features may only be contained in certain regions of an image, dynamic networks with hard spatial attention are explored to strategically select patches from the input for improved efficiency.\n\\noindent\\textbf{1) Hard attention with RNNs.} The most typical approach is formulating a classification task as a sequential decision process, {and adopting RNNs to make iterative predictions based on selected patches .} For example, images are classified within a fixed number of steps, and at each step, the classifier RNN only sees a cropped patch, deciding the next attentional location until the last step is reached . An adaptive step number is further achieved by including early stopping in the action space . Glance-and-focus network (GFNet) builds a general framework of region-level adaptive inference by sequentially focusing on a series of selected patches, and is compatible with most existing CNN architectures. The recurrent attention mechanism together with the early exiting paradigm enables both \\emph{spatially} and \\emph{temporally} adaptive inference .\n\\noindent\\textbf{2) Hard attention with other implementations.} Rather than using an RNN to predict the region position that the model should pay attention to, class activation mapping (CAM) is leveraged in to iteratively focus on salient patches. At each iteration, the selection is performed on the previously cropped input, leading to a progressive refinement procedure. A multi-scale CNN is built in , where the sub-network in each scale takes in the cropped patch from the previous scale, and is responsible for simultaneously producing 1) the feature representations for classification and 2) the attention map for the next scale. {Without an iterative manner, the recent differentiable patch selection adopts a differentiable top-K module to select a fixed number of patches in one step.}\n\\vspace{-2ex}","id":"27095343-62ad-48ed-a110-4d5657e3a267","level":"subsubsection","origin_cites_number":7,"parent_id":"0d52450e-77a8-4861-a2eb-00ddb1ece863","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Region-level Dynamic Networks"],["subsubsection","Hard Attention on Selected Patches"]],"subsections":[],"title":"Hard Attention on Selected Patches"},{"cite_extract_rate":1,"cites":[504],"content":"\\label{dynamic_resolution}\n\\vspace{-0.5ex}\nThe researches discussed above typically divide feature maps into different areas (pixel-level or region-level) for adaptive inference.\nOn a coarser granularity, some dynamic networks could treat each image as a whole by processing feature representations with adaptive resolutions. Although it has been observed that a low resolution might be sufficient for recognizing most \"easy\" samples , conventional CNNs mostly process all the inputs with the same resolution, inducing considerable redundancy. Therefore, resolution-level dynamic networks exploit spatial redundancy from the perspective of feature resolution rather than the saliency of different locations. Existing approaches mainly include 1) scaling the inputs with adaptive ratios (Sec. \\ref{adaptive_scaling_ratio}); 2) selectively activating the sub-networks with different resolutions in a multi-scale architecture (Sec. \\ref{dynamic_res_multiscale}).\n\\vspace{-1ex}","id":"ab0c50b9-9d18-4c18-976a-34aae739e7ac","level":"subsection","origin_cites_number":1,"parent_id":"d1dfc651-19ab-46ab-9db8-02eef3371ca0","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Resolution-level Dynamic Networks"]],"subsections":["6a7330af-c2c7-4497-af0d-0ab86ce108a0","16545c48-ec80-4bbd-871e-b151b745999d"],"title":"Resolution-level Dynamic Networks"},{"cite_extract_rate":0.5,"cites":[748],"content":"\\label{adaptive_scaling_ratio}\n\\vspace{-0.25ex}\nDynamic resolution can be achieved by scaling features with adaptive ratios. For example, a small sub-network is first executed to predict a scale distribution of faces on the face detection task, then the input images are adaptively zoomed, so that all the faces fall in a suitable range for recognition . A plug-in module is used by to predict the stride for the first convolution block in each ResNet stage, producing features with dynamic resolution.\n\\vspace{-1ex}","id":"6a7330af-c2c7-4497-af0d-0ab86ce108a0","level":"subsubsection","origin_cites_number":2,"parent_id":"ab0c50b9-9d18-4c18-976a-34aae739e7ac","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Resolution-level Dynamic Networks"],["subsubsection","Adaptive Scaling Ratios"]],"subsections":[],"title":"Adaptive Scaling Ratios"},{"cite_extract_rate":0.5,"cites":[7266],"content":"\\label{dynamic_res_multiscale} \n\\vspace{-0.25ex}\nAn alternative approach to achieving dynamic resolution is building multiple sub-networks in a parallel or cascading way. These sub-networks with different feature resolutions are selectively activated conditioned on the input during inference. For instance, Elastic realizes a \\emph{soft} selection from multiple branches at every layer, where each branch performs a downsample-convolution-upsample procedure with an independent scaling ratio. To practically avoid redundant computation, a \\emph{hard} selection is realized by , which allows each sample to conditionally activate sub-networks that process feature representations with resolution from low to high (see \\figurename~\\ref{multi_scale} (c) in Sec. \\ref{dynamic_depth}).\n\\vspace{-2ex}\n\\begin{figure*}\n \\centering\n \\vspace{-2ex}\n \\includegraphics[width=0.9\\linewidth]{10_temporal_skim.pdf}\n \\vskip -0.15in\n \\caption{{Temporally adaptive inference. The first three approaches dynamically allocate computation in each step by (a) skipping the update, (b) partially updating the state, or (c) conditional computation in a hierarchical structure. The agent in (d) decides where to read in the next step.}}\n \\label{fig_temporal_skim}\n \\vspace{-3ex}\n\\end{figure*}","id":"16545c48-ec80-4bbd-871e-b151b745999d","level":"subsubsection","origin_cites_number":2,"parent_id":"ab0c50b9-9d18-4c18-976a-34aae739e7ac","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Spatial-wise Dynamic Networks"],["subsection","Resolution-level Dynamic Networks"],["subsubsection","Dynamic Resolution in Multi-scale Architectures"]],"subsections":[],"title":"Dynamic Resolution in Multi-scale Architectures"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec_temporal_adaptive}\n\\vspace{-0.25ex}\nApart from the spatial dimension (Sec. \\ref{sec_spatially_adaptive}), adaptive computation could also be performed along the temporal dimension of sequential data, such as texts (Sec. \\ref{temporal_text}) and videos (Sec. \\ref{sec:tempoal_video}). {In general, network efficiency can be improved by dynamically allocating less/no computation to the inputs at unimportant temporal locations.}\n\\vspace{-1.5ex}","id":"8a8770e6-47f0-428a-8e00-4bb401ff006f","level":"section","origin_cites_number":0,"parent_id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"]],"subsections":["d0476612-fbf1-40f0-a9de-cc704c8eb7d7","844590b4-9a40-42f9-8052-47ad9e4a7783"],"title":"Temporal-wise Dynamic Networks"},{"cite_extract_rate":0,"cites":[],"content":"\\vspace{-0.25ex}\n\\label{temporal_text}\nTraditional RNNs mostly follow a static inference paradigm, i.e. input tokens are read sequentially to update a hidden state at each time step, which could be written as\n\\begin{equation}\\label{eq_rnn}\n \\setlength{\\abovedisplayskip}{3pt}\n\\mathbf{h}_t = \\mathcal{F}(\\mathbf{x}_t, \\mathbf{h}_{t-1}), t=1,2,\\cdots, T.\n\\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nSuch a static inference paradigm induces significant redundant computation, as different tokens usually have different contributions to the downstream tasks.\nA type of dynamic RNN is developed for allocating appropriate computational cost at each step. Some learn to \\emph{\"skim\"} unimportant tokens by dynamic update of hidden states (Sec. \\ref{skimming_text}), and others conduct \\emph{adaptive reading} to avoid processing task-irrelevant tokens. Specifically, such adaptive reading can be achieved by \\emph{early exiting} (Sec. \\ref{temporal_early_exit}) or \\emph{dynamic jumping} (Sec. \\ref{text_jumping}).\n\\vspace{-1ex}","id":"d0476612-fbf1-40f0-a9de-cc704c8eb7d7","level":"subsection","origin_cites_number":0,"parent_id":"8a8770e6-47f0-428a-8e00-4bb401ff006f","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"],["subsection","RNN-based Dynamic Text Processing"]],"subsections":["7450af29-777b-40c4-bbdb-5820eb26595d","04644c19-76d1-481e-b8a1-a56033239887","64d0b8b6-25b1-481c-a175-60d486a5210d"],"title":"RNN-based Dynamic Text Processing"},{"cite_extract_rate":0.625,"cites":[683,7284,750,7283,749],"content":"\\vspace{-0.25ex}\n\\label{skimming_text}\nSince not all the tokens are essential for capturing the task-relevant information in a sequence, dynamic RNNs can be built to adaptively update their hidden states at each time step. Less informative tokens will be coarsely \\emph{skimmed}, i.e. the states are updated with cheaper computation.\n\\noindent\\textbf{1) Skipping the update.} For unimportant inputs at certain temporal locations, dynamic models can learn to entirely skip the update of hidden states (see \\figurename~\\ref{fig_temporal_skim} (a)), i.e.\n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{h}_t = \\alpha_t\\mathcal{F}(\\mathbf{x}_t, \\mathbf{h}_{t-1}) + (1-\\alpha_t)\\mathbf{h}_{t-1}, \\alpha_t\\in\\left\\{0,1\\right\\}.\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nFor instance, Skip-RNN updates a controlling signal in every step to determine whether to update or \\emph{copy} the hidden state from the previous step. An extra agent is adopted by Structural-Jump-LSTM to make the skipping decision conditioned on the previous state and the current input. Without training the RNNs and the controllers jointly as in and , a predictor is trained in to estimate whether each input will make a \"significant change\" on the hidden state. The update is identified worthy to be executed only when the predicted change is greater than a threshold.\n\\noindent\\textbf{2) Coarse update. } As directly skipping the update may be too aggressive, dynamic models could also update the hidden states with adaptively allocated operations. In specific, a network can adapt its architecture in every step, i.e. \n\\begin{equation}\n \\setlength{\\abovedisplayskip}{3pt}\n \\mathbf{h}_t = \\mathcal{F}_t(\\mathbf{x}_t, \\mathbf{h}_{t-1}), t=1,2,\\cdots, T,\n \\setlength{\\belowdisplayskip}{3pt}\n\\end{equation}\nwhere $\\mathcal{F}_t$ is determined based on the input $\\mathbf{x}_t$. One implementation is selecting a subset of dimensions of the hidden state to calculate, and copying the remaining from the previous step , as shown in \\figurename~\\ref{fig_temporal_skim} (b). To achieve the partial update, a subset of rows in weight matrices of the RNN is dynamically activated in , while Skim-RNN makes a choice between two independent RNNs.\nWhen the hidden states are generated by a multi-layer network, the update could be interrupted at an intermediate layer based on an accumulated halting score .\nTo summarize, a coarse update can be realized by data-dependent network \\emph{depth} or \\emph{width} . \n{\\noindent\\textbf{3) Selective updates in hierarchical RNNs.} Considering the intrinsic hierarchical structure of texts (e.g. sentence-word-character), researchers have developed hierarchical RNNs to encode the temporal dependencies with different timescales using a dynamic update mechanism . During inference, the RNNs at higher levels will selectively update their states conditioned on the output of low-level ones (see \\figurename~\\ref{fig_temporal_skim} (c)).\nFor example, when a character-level model in detects that the input satisfies certain conditions, it will \\emph{\"flush\"} (reset) its states and feed them to a word-level network. Similar operations have also been realized by a gating module on question answering tasks .}\n\\vspace{-1ex}","id":"7450af29-777b-40c4-bbdb-5820eb26595d","level":"subsubsection","origin_cites_number":8,"parent_id":"d0476612-fbf1-40f0-a9de-cc704c8eb7d7","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"],["subsection","RNN-based Dynamic Text Processing"],["subsubsection","Dynamic Update of Hidden States"]],"subsections":[],"title":"Dynamic Update of Hidden States"},{"cite_extract_rate":0,"cites":[],"content":"\\vspace{-0.25ex}\n\\label{temporal_early_exit}\nDespite that the dynamic RNNs in Sec. \\ref{skimming_text} are able to update their states with data-dependent computational costs at each step, all the tokens still must be read, leading to inefficiency in scenarios where the task-relevant results can be obtained before reading the entire sequence.\nIdeally, an efficient model should adaptively stop reading before the last step $T$ in Eq. \\ref{eq_rnn} is reached, once the captured information is satisfactory to solve the task. For instance, reasoning network (ReasoNet) terminates its reading procedure when sufficient evidence has been found for question answering. \nSimilarly, early stopping is implemented for sentence-level and paragraph-level text classification, respectively. Note that the approaches discussed here focus on making early predictions with respect to the \\emph{temporal} dimension of sequential input, rather than along the \\emph{depth} dimension of networks as in Sec. \\ref{dynamic_depth}.\n\\vspace{-1ex}","id":"04644c19-76d1-481e-b8a1-a56033239887","level":"subsubsection","origin_cites_number":3,"parent_id":"d0476612-fbf1-40f0-a9de-cc704c8eb7d7","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"],["subsection","RNN-based Dynamic Text Processing"],["subsubsection","Temporally Early Exiting in RNNs"]],"subsections":[],"title":"Temporally Early Exiting in RNNs"},{"cite_extract_rate":0.5,"cites":[7283,7285],"content":"\\vspace{-0.25ex}\n\\label{text_jumping}\nAlthough early exiting in Sec. \\ref{temporal_early_exit} can largely reduce redundant computation, all the tokens must still be fed to the model one by one. More aggressively, dynamic RNNs could further learn to decide \\emph{\"where to read\"} by strategically skipping some tokens without reading them, and directly jumping to an arbitrary temporal location (see \\figurename~\\ref{fig_temporal_skim} (d)).\nSuch dynamic jumping, together with early exiting, is realized in and . Specifically, LSTM-Jump implements an auxiliary unit to predict the jumping stride within a defined range, and the reading process ends when the unit outputs zero. The model in first decides whether to stop at each step. If not, it will further choose to re-read the current input, or to skip a flexible number of words. Moreover, structural information is exploited by Structural-Jump-LSTM , which utilizes an agent to decide whether to jump to the next punctuation. Apart from looking ahead, LSTM-Shuttle also allows backward jumping to supplement the missed history information.\n\\vspace{-1.5ex}","id":"64d0b8b6-25b1-481c-a175-60d486a5210d","level":"subsubsection","origin_cites_number":4,"parent_id":"d0476612-fbf1-40f0-a9de-cc704c8eb7d7","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"],["subsection","RNN-based Dynamic Text Processing"],["subsubsection","Jumping in Texts"]],"subsections":[],"title":"Jumping in Texts"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:tempoal_video}\n\\vspace{-0.25ex}\nFor video recognition, where a video could be seen as a sequential input of frames, temporal-wise dynamic networks are designed to allocate adaptive computational resources for different frames. This can generally be achieved by two approaches: 1) dynamically updating the hidden states in each time step of \\emph{recurrent} models (Sec. \\ref{recurrent_video}), and 2) performing adaptive \\emph{pre-sampling} for key frames (Sec. \\ref{frame_sampling}).\n\\vspace{-1ex}","id":"844590b4-9a40-42f9-8052-47ad9e4a7783","level":"subsection","origin_cites_number":0,"parent_id":"8a8770e6-47f0-428a-8e00-4bb401ff006f","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"],["subsection","Temporal-wise Dynamic Video Recognition"]],"subsections":["b2f1acb3-4b7a-41ce-96e3-e8abe77bc9b2","ce2a3a82-4498-49fd-b6ea-af13eecfa56b"],"title":"Temporal-wise Dynamic Video Recognition"},{"cite_extract_rate":0.6000000000000001,"cites":[7286,755,7270,752,751,754,7283,8380,753],"content":"\\label{recurrent_video}\n\\vspace{-0.25ex}\nVideo recognition is often conducted via a recurrent procedure, where the video frames are first encoded by a 2D CNN, and the obtained frame features are fed to an RNN sequentially for updating its hidden state. Similar to the approaches introduced in Sec. \\ref{temporal_text}, RNN-based adaptive video recognition is typically realized by 1) treating unimportant frames with relatively cheap computation (\\emph{\"glimpse\"}) ; 2) \\emph{early exiting} ; and 3) performing dynamic \\emph{jumping} to decide {\"where to see\"} .\n\\noindent\\textbf{1) Dynamic update of hidden states.} To reduce redundant computation at each time step, LiteEval makes a choice between two LSTMs with different computational costs. ActionSpotter decides whether to update the hidden state according to each input frame. {AdaFuse selectively reuses certain feature channels from the previous step to efficiently make use of historical information.} Recent work has also proposed to adaptively decide the numerical precision or modalities when processing the sequential input frames. Such a \\emph{glimpse} procedure (i.e. allocating cheap operations to unimportant frames) is similar to the aforementioned text \\emph{skimming} .\n\\noindent\\textbf{2) Temporally early exiting.} Humans are able to comprehend the contents easily before watching an entire video. Such early stopping is also implemented in dynamic networks to make predictions only based on a portion of video frames . Together with the \\emph{temporal} dimension, the model in further achieves early exiting from the aspect of network \\emph{depth} as discussed in Sec. \\ref{dynamic_depth}.\n\\noindent\\textbf{3) Jumping in videos.} Considering encoding those unimportant frames with a CNN still requires considerable computation, a more efficient solution could be dynamically skipping some frames without watching them. Existing arts typically learn to predict the location that the network should jump to at each time step.\nFurthermore, both early stopping and dynamic jumping are allowed in , where the jumping stride is limited in a discrete range.\nAdaptive frame (AdaFrame) generates a continuous scalar within the range of $[0,1]$ as the relative location. \n\\vspace{-1ex}","id":"b2f1acb3-4b7a-41ce-96e3-e8abe77bc9b2","level":"subsubsection","origin_cites_number":15,"parent_id":"844590b4-9a40-42f9-8052-47ad9e4a7783","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"],["subsection","Temporal-wise Dynamic Video Recognition"],["subsubsection","Video Recognition with Dynamic RNNs"]],"subsections":[],"title":"Video Recognition with Dynamic RNNs"},{"cite_extract_rate":0.642857142857142,"cites":[8381,756,6975,8382,757,7288,7287,686,758],"content":"\\label{frame_sampling}\n\\vspace{-0.25ex}\nRather than processing video frames recurrently as in Sec. \\ref{recurrent_video}, another line of work first performs an adaptive \\emph{pre-sampling} procedure, and then makes prediction by processing the selected subset of key frames or clips.\n\\noindent{\\textbf{1) Temporal attention} is a common technique for networks to focus on salient frames.} For face recognition, neural aggregation network uses \\emph{soft} attention to adaptively aggregate frame features. To improve the inference efficiency, \\emph{hard} attention is realized to remove unimportant frames iteratively with RL for efficient video face verification .\n\\noindent{\\textbf{2) Sampling module} is also a prevalent option for dynamically selecting the key frames/clips in a video.} For example, the frames are first sampled uniformly in , and discrete decisions are made for each selected frame to go forward or backward step by step. As for clip-level sampling, SCSample is designed based on a trained classifier to find the most informative clips for prediction. Moreover, dynamic sampling network (DSN) segments each video into multiple sections, and a sampling module with shared weights across the sections is exploited to sample one clip from each section.\n{Adjusting multiple factors of deep models simultaneously has attracted researches in both static and dynamic networks . For example, together with \\emph{temporal-wise} frame sampling, \\emph{spatially} adaptive computation can be achieved by spatial /temporal resolution adaptation and patch selection . It would be promising to exploit the redundancy in both \\emph{input data} and \\emph{network structure} for further improving the efficiency of deep networks.}\n\\vspace{-2ex}","id":"ce2a3a82-4498-49fd-b6ea-af13eecfa56b","level":"subsubsection","origin_cites_number":14,"parent_id":"844590b4-9a40-42f9-8052-47ad9e4a7783","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Temporal-wise Dynamic Networks"],["subsection","Temporal-wise Dynamic Video Recognition"],["subsubsection","Dynamic Key Frame Sampling"]],"subsections":[],"title":"Dynamic Key Frame Sampling"},{"cite_extract_rate":0.6666666666666661,"cites":[675,7277],"content":"\\label{inference_and_train}\n\\vspace{-0.25ex}\nIn previous sections, we have reviewed three different types of dynamic networks (sample-wise (Sec. \\ref{sec_sample_wise}), spatial-wise (Sec. \\ref{sec_spatially_adaptive}) and temporal-wise (Sec. \\ref{sec_temporal_adaptive})). It can be observed that making data-dependent \\emph{decisions} at the inference stage is essential to achieve high efficiency and effectiveness. Moreover, \\emph{training} dynamic models is usually more challenging than optimizing static networks.\nNote that since parameter adaptation (Sec. \\ref{adaptive_params}) could be conveniently achieved by differentiable operations, models with dynamic parameters can be directly trained by stochastic gradient descent (SGD) without specific techniques. Therefore, in this section we mainly focus on discrete decision making (Sec. \\ref{inference}) and its training strategies (Sec. \\ref{training}), which are absent in most static models.\n\\vspace{-1.5ex}","id":"4ffefefb-4651-456b-a07d-75e465a728b0","level":"section","origin_cites_number":3,"parent_id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"]],"subsections":["8d1ee291-ab1d-4e7f-8c16-431c07d68081","2bd26a27-bb2c-4752-8590-7d1e502a061d"],"title":"Inference and Training"},{"cite_extract_rate":0,"cites":[],"content":"\\label{inference}\n\\vspace{-0.5ex}\nAs described above, dynamic networks are capable of making data-dependent decisions during inference to transform their architectures, parameters, or to select salient spatial/temporal locations in the input. Here we summarize three commonly seen decision making schemes as follows.\n\\vspace{-1ex}","id":"8d1ee291-ab1d-4e7f-8c16-431c07d68081","level":"subsection","origin_cites_number":0,"parent_id":"4ffefefb-4651-456b-a07d-75e465a728b0","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"],["subsection","Decision Making of Dynamic Networks"]],"subsections":["16951c8b-a4dd-48e3-a004-db0aa23dc8c4","2a7772db-4dcb-4d14-b74b-bd6ac1bb74cf","6e023be0-dfce-4cb5-b69e-f84dd0c01501"],"title":"Decision Making of Dynamic Networks"},{"cite_extract_rate":0.636363636363636,"cites":[7266,689,683,759,7269,676,7273],"content":"\\vspace{-0.25ex}\nMany dynamic networks are able to output \"easy\" samples at early exits if a certain confidence-based criterion is satisfied. These methods generally require estimating the confidence of intermediate predictions, which is compared to a predefined threshold for decision making.\nIn classification tasks, the confidence is usually represented by the maximum element of the \\emph{SoftMax} output . Alternative criteria include the entropy and the score margin . On NLP tasks, a \\emph{model patience} is proposed in : when the predictions for one sample stay unchanged after a number of classifiers, the inference procedure stops. \nIn addition, the halting score in could also be viewed as confidence for whether the current feature could be output to the next time step or calculation stage.\nEmpirically, the confidence-based criteria are easy to implement, and generally require no specific training techniques. A trade-off between accuracy and efficiency is controlled by manipulating the thresholds, which are usually tuned on a validation dataset. Note that the \\emph{overconfidence} issue in deep models might affect the effectiveness of such decision paradigm, when the incorrectly classified samples could obtain a high confidence at early exits.\n\\vspace{-1ex}","id":"16951c8b-a4dd-48e3-a004-db0aa23dc8c4","level":"subsubsection","origin_cites_number":11,"parent_id":"8d1ee291-ab1d-4e7f-8c16-431c07d68081","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"],["subsection","Decision Making of Dynamic Networks"],["subsubsection","Confidence-based Criteria"]],"subsections":[],"title":"Confidence-based Criteria"},{"cite_extract_rate":1,"cites":[702,718,691],"content":"\\label{inference_policy}\n\\vspace{-0.25ex}\nIt is a common option to build an additional policy network learning to adapt the network topology based on different samples. \nSpecifically, each input sample is first processed by the policy network, whose output directly determines which parts of the main network should be activated. For example, BlockDrop and GaterNet use a policy network to adaptively decide the \\emph{depth} and \\emph{width} of a backbone network. More generally, dynamic routing in a \\emph{SuperNet} can also be controlled by a policy network .\nOne possible limitation of this scheme is that the architectures and the training process of some policy networks are developed for a specific backbone , and may not be easily adapted to different architectures. \n\\vspace{-1ex}","id":"2a7772db-4dcb-4d14-b74b-bd6ac1bb74cf","level":"subsubsection","origin_cites_number":3,"parent_id":"8d1ee291-ab1d-4e7f-8c16-431c07d68081","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"],["subsection","Decision Making of Dynamic Networks"],["subsubsection","Policy Networks"]],"subsections":[],"title":"Policy Networks"},{"cite_extract_rate":0.7142857142857141,"cites":[7274,708,751,8374,710,8373,696,7281,713,699],"content":"\\label{inference_gating}\n\\vspace{-0.25ex}\nGating function is a general and flexible approach to decision making in dynamic networks. It can be conveniently adopted as a plug-in module at arbitrary locations in any backbone network.\nDuring inference, each module is responsible for controlling the local inference graph of a layer or block. The gating functions take in intermediate features and efficiently produce binary-valued gate vectors to decide: 1) which channels need to be activated \\emph{width}, 2) which layers need to be skipped , 3) which paths should be selected in a SuperNet , or 4) what locations of the input should be allocated computations .\nCompared to the aforementioned decision policies, the gating functions demonstrate notable generality and applicability. However, due to their lack of differentiability, these gating functions usually need specific training techniques, which will be introduced in the following Sec. \\ref{training}.\n\\vspace{-2ex}","id":"6e023be0-dfce-4cb5-b69e-f84dd0c01501","level":"subsubsection","origin_cites_number":14,"parent_id":"8d1ee291-ab1d-4e7f-8c16-431c07d68081","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"],["subsection","Decision Making of Dynamic Networks"],["subsubsection","Gating Functions"]],"subsections":[],"title":"Gating Functions"},{"cite_extract_rate":0.581196581196581,"cites":[779,725,722,7273,683,770,771,776,7257,778,741,7294,7283,7267,674,8381,7290,772,746,7285,150,8382,760,7281,7289,7272,761,747,752,757,7288,8383,764,750,727,743,765,748,762,8380,777,775,769,686,680,767,7295,7291,763,773,744,774,7286,7284,7282,7278,749,7215,7292,726,766,768,676,7293,738,7274,780,7287],"content":"\\vspace{-0.5ex}\n\\label{training}\nBesides architecture design, training is also essential for dynamic networks. Here we summarize the existing training strategies for dynamic models from the perspectives of objectives and optimization.\n\\begin{table*}\n \\scriptsize\n \\vspace{-2ex}\n \\caption{Applications of Dynamic Networks. For the type column, Sa, Sp and Te stand for sample-wise, spatial-wise and temporal-wise respectively.}\n \\label{tab_tasks}\n \\begin{center}\n \\vspace{-6ex}\n \\resizebox{\\linewidth}{!}{\n \\begin{tabular}{cccc}\n \\toprule\n \\textbf{Fields} & \\textbf{Data} & \\textbf{Type} & \\textbf{Subfields \\& references} \\\\\n \\midrule\n & \\multirow{6}*{\\textbf{Image}} & \\multirow{2}*{Sa} & Object detection (face , facial point , pedestrian , general ) \\\\\n & & & Image segmentation , Super resolution , Style transfer , Coarse-to-fine classification \\\\ \n \\cmidrule{3-4}\n & & \\multirow{3}*{Sa \\& Sp} & Image segmentation , Image-to-image \\\\\n & & & translation , Object detection , Semantic image synthesis , \\\\\n \\multirow{1}*{\\textbf{Computer}} & & & Image denoising , Fine-grained classification Eye tracking , Super resolution \\\\\n \\cmidrule{3-4}\n \\multirow{1}*{\\textbf{Vision}}& & Sa \\& Sp \\& Te & General classification , Multi-object classification , Fine-grained classification \\\\\n \\cmidrule{2-4}\n & \\multirow{4}*{\\textbf{Video}} & Sa & Multi-task learning (human action recognition and frame prediction) \\\\\n \\cmidrule{3-4}\n & & \\multirow{2}*{Sa \\& Te} & Classification (action recognition) , Semantic segmentation \\\\\n & & & Video face recognition , Action detection , Action spotting \\\\\n \\cmidrule{3-4}\n & & Sa \\& Sp \\& Te & Classification , Frame interpolation , Super resolution , Video deblurring , Action prediction \\\\\n \\cmidrule{2-4}\n & \\textbf{Point Cloud} & Sa \\& Sp & 3D Shape classification and segmentation, 3D scene segmentation , 3D semantic scene completion \\\\\n \\midrule\n \\multirow{2}*{\\textbf{Natural}} & \\multirow{3}*{\\textbf{Text}} & Sa & Neural language inference, Text classification, Paraphrase similarity matching, and Sentiment analysis \\\\\n \\cmidrule{3-4}\n \\textbf{Language} & & \\multirow{2}*{Sa \\& Te} & Language modeling , Machine translation , Classification , \\\\\n \\textbf{Processing} & & & Sentiment analysis , Question answering \\\\\n \\midrule\n \\textbf{Cross-Field} & \\multicolumn{3}{c}{Image captioning , {Video captioning , Visual question answering , Multi-modal sentiment analysis }} \\\\\n \\midrule\n \\multirow{2}{*}{\\textbf{Others}} & \\multicolumn{3}{c}{{Time series forecasting} , Link prediction , {Recommendation system} }\\\\\n & \\multicolumn{3}{c}{Graph classification , {Document classification} , Stereo confidence estimation }\\\\\n \\bottomrule\n \\end{tabular}}\n \\end{center}\n \\vskip -0.3in\n\\end{table*}\n\\vspace{-1ex}","id":"2bd26a27-bb2c-4752-8590-7d1e502a061d","level":"subsection","origin_cites_number":117,"parent_id":"4ffefefb-4651-456b-a07d-75e465a728b0","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"],["subsection","Training of Dynamic Networks"]],"subsections":["4d7d59fa-e730-411f-b178-1c7c2fc0254e","34c67f80-85f1-4cea-8d35-f3e767fa3802"],"title":"Training of Dynamic Networks"},{"cite_extract_rate":0.769230769230769,"cites":[7266,701,693,7274,708,8374,8373,696,7281,753],"content":"\\label{sec_train_objectives}\n\\vspace{-0.25ex}\n\\noindent\\textbf{1) Training multi-exit networks.} We first notice that early-exiting dynamic networks are generally trained by minimizing a weighted cumulative loss of intermediate classifiers. One challenge for training such models is the joint optimization of multiple classifiers, which may interfere with each other. MSDNet alleviates the problem through its special architecture design. Several improved training techniques are proposed for multi-exit networks, including a gradient equilibrium algorithm to stable the training process, and a bi-directional knowledge transfer approach to boost the collaboration of classifiers. {For temporal-wise early exiting, the training of the policy network in FrameExit is supervised by pseudo labels.}\n\\noindent\\textbf{2) Encouraging sparsity.} Many dynamic networks adapt their inference procedure by conditionally activating their computational units or strategically sampling locations from the input .\nTraining these models without additional constraints would result in superfluous computational redundancy, as a network could tend to activate all the candidate units for minimizing the task-specific loss. \nThe overall objective function for restraining such redundancy are typically written as $\\mathfrak{L}\\! =\\!\\mathfrak{L}_\\mathrm{task}\\!+\\!\\gamma \\mathfrak{L}_\\mathrm{sparse}$, where $\\gamma$ is the hyper-parameter balancing the two items for the trade-off between accuracy and efficiency. In real-world applications, the second item can be designed based on the gate/mask values of candidate units (e.g. channels , layers or spatial locations ). Specifically, one may set a target activation rate or minimizing the $\\mathcal{L}_1$ norm of the gates/masks . It is also practical to directly optimize a resource-aware loss (e.g. FLOPs) , which can be estimated according to the input and output feature dimension for every candidate unit.\n\\noindent{\\textbf{3) Others.} Note that extra loss items are mostly designed for but not limited to improving efficiency. Take as an example, the model progressively focuses on a selected region, and is trained with an additional \\emph{inter-scale pairwise ranking loss} for proposing more discriminative regions. Moreover, knowledge distilling is utilized to boost the co-training of multiple sub-networks in and .}\n\\vspace{-1ex}","id":"4d7d59fa-e730-411f-b178-1c7c2fc0254e","level":"subsubsection","origin_cites_number":13,"parent_id":"2bd26a27-bb2c-4752-8590-7d1e502a061d","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"],["subsection","Training of Dynamic Networks"],["subsubsection","Training Objectives for Efficient Inference"]],"subsections":[],"title":"Training Objectives for Efficient Inference"},{"cite_extract_rate":0.8235294117647051,"cites":[8381,708,702,7279,696,781,691,782,750,8373,8375,674,7281,699],"content":"\\vspace{-0.25ex}\nA variety of dynamic networks contain non-differentiable functions that make discrete decisions to modify their architectures or sampling spatial/temporal locations from the input. These functions can not be trained directly with back-propagation. Therefore, specific techniques are studied to enable the end-to-end training as follows.\n\\noindent\\textbf{1) Gradient estimation} is proposed to approximate the gradients for those non-differentiable functions and enable back-propagation. In , straight-through estimator (STE) is exploited to heuristically copy the gradient with respect to the stochastic output directly as an estimator of the gradient with respect to the \\emph{Sigmoid} argument.\n\\noindent\\textbf{2) Reparameterization} is also a popular technique to optimize the discrete decision functions. For instance, the gating functions controlling the network width or depth can both be trained with \\emph{Gumbel SoftMax} , which is also used for pixel-level dynamic convolution . An alternative technique is \\emph{Improved SemHash} adopted in and to train their hard gating modules.\n{Note that although these reparameterization techniques enable joint optimizing dynamic models together with gating modules in an end-to-end fashion, they usually lead to a longer training process for the decision functions to converge into a stable situation . Moreover, the model performance might be sensitive to some extra hyper-parameters (e.g. temperature in \\emph{Gumbel SoftMax}), which might also increase the training cost for these dynamic networks.}\n\\noindent\\textbf{3) Reinforcement learning (RL)} is widely exploited for training non-differentiable decision functions. In specific, the backbones are trained by standard SGD, while the agents (either policy networks in Sec. \\ref{inference_policy} or gating functions in Sec. \\ref{inference_gating}) are trained with RL to take discrete actions for dynamic inference graphs or spatial/temporal sampling strategies .\n{One challenge for RL-based training is the design of reward functions, which is important to the accuracy-efficiency tradeoff of dynamic models. Commonly seen reward signals are usually constructed to minimize a penalty item of the computational cost . Moreover, the training could be costly due to a multi-stage procedure: a pre-training process may be required for the backbone networks before the optimization of decision or sampling modules, and joint finetuning may be indispensable finally.}\n\\vspace{-2ex}","id":"34c67f80-85f1-4cea-8d35-f3e767fa3802","level":"subsubsection","origin_cites_number":17,"parent_id":"2bd26a27-bb2c-4752-8590-7d1e502a061d","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Inference and Training"],["subsection","Training of Dynamic Networks"],["subsubsection","Optimization of Non-differentiable Functions"]],"subsections":[],"title":"Optimization of Non-differentiable Functions"},{"cite_extract_rate":0.8148148148148141,"cites":[764,744,785,7273,765,770,7269,777,758,38,786,692,768,783,7296,7295,7272,7271,7,757,674,784],"content":"\\label{sec_tasks}\n\\vspace{-0.25ex}\nIn this section, we summarize the applications of dynamic networks. Representative methods are listed in Table \\ref{tab_tasks} based on the input data modality.\nFor image recognition, most dynamic CNNs are designed to conduct \\emph{sample-wise} or \\emph{spatial-wise} adaptive inference on classification tasks, and many inference paradigms can be generalized to other applications. Note that as mentioned in Sec. \\ref{region_level}, the object recognition could be formulated as a sequential decision problem . By allowing early exiting in these approaches, \\emph{temporally} adaptive inference procedure could also be enabled.\nFor text data, reducing its intrinsic temporal redundancy has attracted great research interests, and the inference paradigm of \\emph{temporal-wise} dynamic RNNs (see Sec. \\ref{temporal_text}) is also general enough to process audios . Based on large language models such as Transformer and BERT , adaptive depths are extensively studied to reduce redundant computation in network architectures.\nFor video-related tasks, the three types of dynamic inference can be implemented simultaneously . However, for the networks that do not process videos recurrently, e.g. 3D CNNs , most of them still follow a static inference scheme. Few researches have been committed to building dynamic 3D CNNs , which might be an interesting future research direction.\n{Moreover, dynamic networks (especially the attention mechanism) have also been applied to dynamically fuse the features from different modalities in some multi-modal learning tasks, e.g. RGB-D image segmentation and image/video captioning .}\nFinally, dynamic networks have also been exploited to tackle some fundamental problems in deep learning. For example, multi-exit models can be used to: 1) alleviate the \\emph{over-thinking} issue while reducing the overall computation ; 2) perform \\emph{long-tailed classification} by inducing early exiting in the training stage; and 3) improve the model \\emph{robustness} . For another example, the idea of dynamic routing is implemented for: 1) {reducing the \\emph{training cost} under a multi-task setting } and 2) finding the optimal fine-tuning strategy for per example in \\emph{transfer learning} .\n\\vspace{-2ex}","id":"3cd25c7f-359c-4b70-8c02-f2738746b35c","level":"section","origin_cites_number":27,"parent_id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","Application of Dynamic Networks"]],"subsections":[],"title":"Application of Dynamic Networks"},{"cite_extract_rate":0,"cites":[],"content":"}\n\\label{sec:discussion}\n\\vspace{-0.5ex}\n{Though recent years have witnessed significant progress in the research of dynamic neural networks, there still exist many open problems that are worth exploring.} In this section, we summarize a few challenges together with possible future directions in this field.\n\\vspace{-1ex}","id":"e9013abd-3d22-4eab-a882-4ba5e86719f2","level":"section","origin_cites_number":0,"parent_id":"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Challenges and Future Directions"]],"subsections":["3fc7f7a7-7403-4a95-918e-d6e40acaa928","9458c790-c005-432b-a25a-958ab55678a8","baaa0a8b-88d7-44dc-8c2b-1c3a26c2c457","23446088-fa89-4b21-8a64-51ee553e33cb","a8fb9eb8-777e-40f9-a6cc-da50d402237b","80ebe8d1-1699-4384-9941-1c27455ceb08"],"title":"{Challenges and Future Directions"},{"cite_extract_rate":1,"cites":[7266,706],"content":"\\vspace{-0.5ex}\nDespite the success of dynamic neural networks, relatively few researches has been committed to analyze them from the theoretical perspective. In fact, theories for a deep understanding of current dynamic learning models and further improving them in principled ways are highly valuable. {Notably, it has been proven that a dynamic network with an adaptive width can preserve the representation power of an unsparsified model .} However, there are more theoretical problems that are fundamental for dynamic networks. Here we list several of them as follows.\n\\noindent\\textbf{1) Optimal decision in dynamic networks.}\nAn essential operation in most dynamic networks (especially those designed for improving computational efficiency) is making data-dependent decisions, e.g., determining whether a module should be evaluated or skipped. Existing solutions either use confidence-based criteria, or introduce policy networks and gating functions. Although being effective in practice (as mentioned in Sec. \\ref{inference_and_train}), they may not be optimal and lack theoretical justifications. Take early exiting as an example, the current heuristic methods might face the issues of overconfidence, high sensitivity for threshold setting and poor transferability. As for policy networks or gating modules, runtime decisions can be made based on a learned function.\nHowever, they often introduce extra computations, and usually require a long and unstable training procedure. Therefore, principled approaches with theoretical guarantees for decision function design in dynamic networks is a valuable research topic.\n\\noindent\\textbf{2) Generalization issues.} \nIn a dynamic model, a sub-network might be activated for a set of test samples that are not uniformly sampled from the data distribution, e.g., smaller sub-networks tend to handle ``easy'' samples, while larger sub-networks are used for ``hard'' inputs . This brings a divergence between the training data distribution and that of the inference stage, and thus violates the common \\emph{i.i.d.} assumption in classical machine learning. Therefore, it would be interesting to develop new theories to analyze the generalization properties of dynamic networks under such distribution mismatch. Note that transfer learning also aims to address the issue of distributional shift at test time, but the samples of the target domain are assumed to be accessible in advance. In contrast, for dynamic models, the test distribution is not available until the training process is finished, when the network architecture and parameters are finalized. This poses greater challenges than analyzing the generalization issues in transfer learning.","id":"3fc7f7a7-7403-4a95-918e-d6e40acaa928","level":"subsection","origin_cites_number":2,"parent_id":"e9013abd-3d22-4eab-a882-4ba5e86719f2","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Challenges and Future Directions"],["subsection","Theories for Dynamic Networks"]],"subsections":[],"title":"Theories for Dynamic Networks"},{"cite_extract_rate":0.764705882352941,"cites":[7266,96,700,8372,718,7297,97,787,732,696,685,504,687],"content":"Architecture design has been proven to be essential for deep networks. Existing researches on architectural innovations are mainly proposed for static models , while relatively few are dedicated to developing architectures specially for dynamic networks. It is expected that architectures developed specifically for dynamic networks may further improve their effectiveness and efficiency. For example, the interference among multiple classifiers in an early-exiting network could be mitigated by a carefully designed multi-scale architecture with dense connections . \nPossible research direction include designing dynamic network structures either by hand (as in ), or by leveraging the NAS techniques (as in ). Moreover, considering the popularity of Transformers , {recent work has proposed dynamic vision Transformers with adaptive early exiting or token sparsification .} Developing a dynamic version of this family of models could also be an interesting direction.\n{Note that the research on dynamic networks differs from a seemingly close topic, i.e. model compression . One common goal of them is improving the network efficiency with minimal accuracy drop. However, model compression may focus on reducing the \\emph{size} of deep networks, while dynamic networks pay more attention to the \\emph{computation}, even at the price of slightly \\emph{increasing} model size . Moreover, model compression typically adopts pruning or quantization techniques to produce compact \\emph{static} models, which treat all the inputs in the same way. In contrast, dynamic networks perform data-dependent computation on different inputs, which can effectively reduce the intrinsic redundancy in static models.}\n\\vspace{-1ex}","id":"9458c790-c005-432b-a25a-958ab55678a8","level":"subsection","origin_cites_number":17,"parent_id":"e9013abd-3d22-4eab-a882-4ba5e86719f2","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Challenges and Future Directions"],["subsection","Architecture Design for Dynamic Networks"]],"subsections":[],"title":"Architecture Design for Dynamic Networks"},{"cite_extract_rate":0.833333333333333,"cites":[7281,675,7268,674,676],"content":"\\vspace{-0.5ex}\nMany existing dynamic networks (e.g., most of the sample-wise adaptive networks) are designed specially for classification tasks, and cannot be applied to other vision tasks such as object detection and semantic segmentation. The difficulty arises from the fact that for these tasks there is no simple criterion to assert whether an input image is easy or hard, as it usually contains multiple objects and pixels that have different levels of difficulty. Although many efforts, e.g., spatially adaptive models and soft attention based models , have been made to address this issue, it remains a challenging problem to develop a unified and elegant dynamic network that can serve as an off-the-shelf backbone for a variety of tasks.\n\\vspace{-1.5ex}","id":"baaa0a8b-88d7-44dc-8c2b-1c3a26c2c457","level":"subsection","origin_cites_number":6,"parent_id":"e9013abd-3d22-4eab-a882-4ba5e86719f2","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Challenges and Future Directions"],["subsection","Applicability for More Diverse Tasks"]],"subsections":[],"title":"Applicability for More Diverse Tasks"},{"cite_extract_rate":0.33333333333333304,"cites":[721,788,7281],"content":"\\label{discuss_implement}\n\\vspace{-0.5ex}\nThe current deep learning hardware and libraries are mostly optimized for static models, and they may not be friendly to dynamic networks. Therefore, we usually observe that the practical runtime of dynamic models lags behind the theoretical efficiency. For example, some spatially adaptive networks involve sparse computation, {which is known to be inefficient on modern computing devices due to the memory access bottleneck . A recent line of work focuses on the codesign of algorithm and hardware for accelerating deep models on platforms with more flexibility such as FPGA . Many input-dependent operations, including pixel-level dynamic computation , adaptive channel pruning and early exiting , have also been tailored together with hardware for further improving their practical efficiency. It is an interesting research direction to simultaneously optimize the algorithm, hardware and deep learning libraries to harvest the theoretical efficiency gains of dynamic networks.}\n{In addition, a data-dependent inference procedure, especially for the dynamic \\emph{architectures}, usually requires a model to handle input samples sequentially, which also poses challenge for parallel computation. Although inference with batches has been enabled for early-exiting networks , the conflict between adaptive computational graph and parallel computation still exists for other types of dynamic architectures. This issue is mitigated in the scenario of mobile/edge computing, where the input signal by itself is sequential and the computing hardware is less powerful than high-end platforms. However, designing dynamic networks that are more compatible with existing hardware and software is still a valuable and challenging topic.}\n\\vspace{-1.5ex}","id":"23446088-fa89-4b21-8a64-51ee553e33cb","level":"subsection","origin_cites_number":9,"parent_id":"e9013abd-3d22-4eab-a882-4ba5e86719f2","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Challenges and Future Directions"],["subsection","Gap between Theoretical \\& Practical Efficiency"]],"subsections":[],"title":"Gap between Theoretical \\& Practical Efficiency"},{"cite_extract_rate":0.33333333333333304,"cites":[789],"content":"\\vspace{-0.5ex}\nDynamic models may provide new perspectives for the research on adversarial robustness of deep neural networks. {For example, recent work has leveraged the multi-exit structure to improve the robustness against adversarial attacks. Moreover, traditional attacks are usually aimed at causing \\emph{misclassification}. For dynamic networks, it is possible to launch attacks on \\emph{efficiency} . Specifically, by adjusting the objective function of the adversarial attack, input-adaptive models could be fooled to activate all their intermediate layers or yielding confusing predictions at early exits even for \"easy\" samples. It has also been observed that the commonly used adversarial training is not effective to defend such attacks.} The robustness of dynamic network is an interesting yet understudied topic.\n\\vspace{-1.5ex}","id":"a8fb9eb8-777e-40f9-a6cc-da50d402237b","level":"subsection","origin_cites_number":3,"parent_id":"e9013abd-3d22-4eab-a882-4ba5e86719f2","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Challenges and Future Directions"],["subsection","Robustness Against Adversarial Attack"]],"subsections":[],"title":"Robustness Against Adversarial Attack"},{"cite_extract_rate":0,"cites":[],"content":"\\vspace{-0.5ex}\nDynamic networks inherit the black-box nature of deep learning models, and thus also invite research on interpreting their working mechanism. What is special here is that the adaptive inference paradigm, e.g., spatial/temporal adaptiveness, conforms well with that of the human visual system, and may provide new possibilities for making the model more transparent to humans. In a dynamic network, it is usually convenient to analyze which part of the model is activated for a given input or to locate which part of the input the model mostly relies on in making its prediction. {It is expected that the research on dynamic network will inspire new work on the interpretability of deep learning.}\n\\vspace{-2ex}\n\\appendices\n\\ifCLASSOPTIONcompsoc\n \\section*{Acknowledgments}\n\\else\n \\section*{Acknowledgment}\n\\fi\nThis work is supported in part by the National Science and Technology Major Project of the Ministry of Science and Technology of China under Grants 2018AAA0100701, the National Natural Science Foundation of China under Grants 61906106 and 62022048, the Institute for Guo Qiang of Tsinghua University and Beijing Academy of Artificial Intelligence.\n\\ifCLASSOPTIONcaptionsoff\n \\newpage\n\\fi\n\\small{\n \\bibliography{reference}\n}\n\\end{document}","id":"80ebe8d1-1699-4384-9941-1c27455ceb08","level":"subsection","origin_cites_number":0,"parent_id":"e9013abd-3d22-4eab-a882-4ba5e86719f2","prefix_titles":[["title","Dynamic Neural Networks: A Survey"],["section","{Challenges and Future Directions"],["subsection","Interpretability"]],"subsections":[],"title":"Interpretability"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ]\n ],\n \"subsections\": [\n \"ed4f3055-e5dc-420d-8009-404bb4d82113\",\n \"58436ad0-55e1-4da8-b7e8-6bf9edce3fb1\",\n \"d1dfc651-19ab-46ab-9db8-02eef3371ca0\",\n \"8a8770e6-47f0-428a-8e00-4bb401ff006f\",\n \"4ffefefb-4651-456b-a07d-75e465a728b0\",\n \"3cd25c7f-359c-4b70-8c02-f2738746b35c\",\n \"e9013abd-3d22-4eab-a882-4ba5e86719f2\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.642857142857142,\n \"cites\": [\n 71,\n 96,\n 679,\n 681,\n 687,\n 544,\n 305,\n 682,\n 683,\n 675,\n 38,\n 504,\n 97,\n 686,\n 684,\n 680,\n 677,\n 676,\n 7266,\n 8372,\n 7268,\n 7,\n 678,\n 7267,\n 674,\n 685,\n 306\n ],\n \"content\": \"\\\\label{sec:introduction}}\\n\\\\IEEEPARstart{D}{eep} neural networks (DNNs) are playing an important role in various areas including computer vision (CV) and natural language processing (NLP) . \\nRecent years have witnessed many successful deep models such as AlexNet , VGG , GoogleNet , ResNet , DenseNet and Transformer . These {architectural innovations} have enabled the training of deeper, more accurate and more efficient models. The recent research on neural architecture search (NAS) further speeds up the process of designing more powerful structures. However, most of the prevalent deep learning models perform inference in a static manner, i.e., both the computational graph and the network parameters are fixed once trained, which may limit their representation power, efficiency and interpretability . \\nDynamic networks, as opposed to static ones, can adapt their structures or parameters to the input during inference, and therefore enjoy favorable properties that are absent in static models. In general, dynamic computation in the context of deep learning has the following advantages: \\n\\\\begin{figure*}\\n \\\\centering\\n \\\\vspace{-2ex}\\n \\\\includegraphics[width=0.875\\\\linewidth]{1_overall.pdf}\\n \\\\vskip -0.175in\\n \\\\centering\\\\caption{Overview of the survey. {We first review the dynamic networks that perform adaptive computation at three different granularities (i.e. sample-wise, spatial-wise and temporal-wise). Then we summarize the decision making strategy, training technique and applications of dynamic models. Existing open problems in this field together with some future research directions are finally discussed. Best viewed in color.}}\\n \\\\label{framework}\\n \\\\vspace{-3ex}\\n\\\\end{figure*}\\n\\\\textbf{1) Efficiency.} One of the most notable advantages of dynamic networks is that they are able to allocate computations on demand at test time, by selectively activating model components (e.g. layers , channels or sub-networks ) \\\\emph{conditioned} on the input. Consequently, less computation is spent on canonical samples that are relatively easy to recognize, or on less informative spatial/temporal locations of an input. {In addition to \\\\emph{computational efficiency}, dynamic models have also shown promising results for improving \\\\emph{data efficiency} in the scenario of few-shot learning .}\\n\\\\textbf{2) Representation power.} Due to the data-dependent network architecture/parameters, dynamic networks have significantly enlarged parameter space and improved representation power. For example, with a minor increase of computation, model capacity can be boosted by applying feature-conditioned attention weights on an ensemble of convolutional kernels . It is worth noting that the popular soft attention mechanism could also be unified in the framework of dynamic networks, as different channels , spatial areas or temporal locations of features are dynamically re-weighted at test time. \\n\\\\textbf{3) Adaptiveness.} Dynamic models are able to achieve a desired trade-off between accuracy and efficiency for dealing with varying computational budgets on the fly. Therefore, they are more adaptable to different hardware platforms and changing environments, compared to static models with a fixed computational cost.\\n\\\\textbf{4) Compatibility.} Dynamic networks are compatible with most advanced techniques in deep learning, including architecture design , optimization algorithms and data preprocessing , which ensures that they can benefit from the most recent advances in the field to achieve state-of-the-art performance. For example, dynamic networks can inherit architectural innovations in lightweight models , or be designed via NAS approaches . Their efficiency could also be further improved by acceleration methods developed for static models, such as network pruning , weight quantization , knowledge distillation and low-rank approximation .\\n\\\\textbf{5) Generality.} As a substitute for static deep learning techniques, many dynamic models are general approaches that can be applied seamlessly to a wide range of applications, such as image classification , object detection and semantic segmentation . Moreover, the techniques developed in CV tasks are proven to transfer well to language models in NLP tasks , and vice versa.\\n\\\\textbf{6) Interpretability.} We finally note that the research on dynamic networks {may bridge} the gap between the underlying mechanism of deep models and brains, as it is believed that the brains process information in a dynamic way~. It is possible to analyze which components of a dynamic model are activated when processing an input sample, and to observe which parts of the input are accountable for certain predictions . These properties may shed light on interpreting the decision process of DNNs.\\n\\\\begin{table}\\n \\\\vspace{-1ex}\\n \\\\caption{Notations used in this paper.}\\n \\\\label{tab:notations}\\n \\\\vspace{-4ex}\\n \\\\begin{center}\\n \\\\begin{tabular}{c|c}\\n \\\\hline\\n \\\\textbf{Notations} & \\\\textbf{Descriptions} \\\\\\\\\\n \\\\hline\\n $\\\\mathbb{R}^m$ & $m$-dimensional real number domain \\\\\\\\\\n \\\\hline\\n $a, \\\\mathbf{a}$ & Scalar, vector/matrix/tensor \\\\\\\\\\n \\\\hline\\n $\\\\mathbf{x,y}$ & Input, output feature \\\\\\\\\\n \\\\hline\\n $\\\\mathbf{x}^{\\\\ell}$ & Feature at layer $\\\\ell$ \\\\\\\\\\n \\\\hline\\n $\\\\mathbf{h}_t$ & Hidden state at time step $t$ \\\\\\\\\\n \\\\hline\\n $\\\\mathbf{x(p)}$ & Feature at spatial location $\\\\mathbf{p}$ on $\\\\mathbf{x}$ \\\\\\\\\\n \\\\hline\\n $\\\\bm{\\\\Theta}$ & Learnable parameter \\\\\\\\\\n \\\\hline\\n $\\\\bm{\\\\hat{\\\\Theta}}|\\\\mathbf{x}$ & Dynamic parameter conditioned on $\\\\mathbf{x}$ \\\\\\\\\\n \\\\hline\\n $\\\\mathbf{x}\\\\star\\\\mathbf{W}$ & Convolution of feature $\\\\mathbf{x}$ and weight $\\\\mathbf{W}$ \\\\\\\\\\n \\\\hline\\n $\\\\otimes$ & Channel-wise or element-wise multiplication \\\\\\\\\\n \\\\hline\\n $\\\\mathcal{F}(\\\\cdot,\\\\bm{\\\\Theta})$ & Functional Operation parameterized by $\\\\bm{\\\\Theta}$ \\\\\\\\\\n \\\\hline\\n $\\\\mathcal{F}\\\\circ\\\\mathcal{G}$ & Composition of function $\\\\mathcal{F}$ and $\\\\mathcal{G}$ \\\\\\\\\\n \\\\hline\\n \\\\end{tabular}\\n \\\\end{center}\\n \\\\vskip -0.35in\\n\\\\end{table}\\nIn fact, adaptive inference, the key idea underlying dynamic networks, has been studied before the popularity of modern DNNs. The most classical approach is building a model ensemble through a cascaded or parallel structure, and selectively activating the models conditioned on the input. Spiking neural networks (SNNs) also perform data-dependent inference by propagating pulse signals. However, the training strategy for SNNs is quite different from that of popular convolutional neural networks (CNNs), and they are less used in vision tasks. Therefore, we leave out the work related to SNNs in this survey. \\n\\\\begin{figure}\\n \\\\centering\\n \\\\vspace{-1ex}\\n \\\\includegraphics[width=\\\\linewidth]{2_dynamic_cascade.pdf}\\n \\\\vskip -0.15in\\n \\\\caption{Two early-exiting schemes. The dashed lines and shaded modules are not executed, conditioned on the decisions made by the routers.}\\n \\\\label{cascading_models}\\n \\\\vskip -0.2in\\n\\\\end{figure}\\n\\\\begin{figure*}\\n \\\\centering\\n \\\\vspace{-2ex}\\n \\\\includegraphics[width=\\\\linewidth]{3_multi_scale_SuperNet.pdf}\\n \\\\vskip -0.175in\\n \\\\caption{{Multi-scale architectures with dynamic inference graphs. The first three models (a, b, c) perform adaptive early exiting with specific architecture designs and exiting policies. Dynamic routing is achieved inside a SuperNet (d) to activate data-dependent inference paths.}}\\n \\\\label{multi_scale}\\n \\\\vspace{-3ex}\\n\\\\end{figure*}\\nIn the context of deep learning, dynamic inference with modern deep architectures, has raised many new research questions and has attracted great research interests in the past three years.\\nDespite the extensive work on designing various types of dynamic networks, a systematic and comprehensive review on this topic is still lacking. This motivates us to write this survey, to review the recent advances in this rapidly developing area, with the purposes of 1) providing an overview as well as new perspectives for researchers who are interested in this topic; 2) pointing out the close relations of different subareas and reducing the risk of reinventing the wheel; and 3) summarizing the key challenges and possible future research directions.\\n{This survey is organized as follows (see \\\\figurename~\\\\ref{framework} for an overview). In Sec. \\\\ref{sec_sample_wise}, we introduce the most common \\\\emph{sample-wise} dynamic networks which adapt their architectures or parameters conditioned on each input sample. Dynamic models working on a finer granularity, i.e., \\\\emph{spatially} adaptive and \\\\emph{temporally} adaptive models, are reviewed in Sec. \\\\ref{sec_spatially_adaptive} and Sec.\\\\ref{sec_temporal_adaptive}, respectively\\\\footnote{{These two categories can also be viewed as sample-wise dynamic networks as they perform adaptive computation within each sample at a finer granularity, and we adopt such a split for narrative convenience.}}. Then we investigate the decision making strategies and the training techniques of dynamic networks in Sec. \\\\ref{inference_and_train}. The applications of dynamic models are further summarized in Sec. \\\\ref{sec_tasks}. Finally, we conclude this survey with a discussion on a number of open problems and future research directions in Sec. \\\\ref{sec:discussion}. For better readability, we list the notations that will be used in this survey in Table \\\\ref{tab:notations}.}\\n\\\\vspace{-2ex}\",\n \"id\": \"ed4f3055-e5dc-420d-8009-404bb4d82113\",\n \"level\": \"section\",\n \"origin_cites_number\": 42,\n \"parent_id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\n\\\\label{sec_sample_wise}\\nAiming at processing different inputs in data-dependent manners, sample-wise dynamic networks are typically designed from two perspectives: 1) adjusting model architectures to allocate appropriate computation based on each sample, and therefore reducing redundant computation for increased efficiency (Sec. \\\\ref{dynamic_arch}); \\n2) adapting network \\\\emph{parameters} to every input sample with fixed computational graphs, with the goal of boosting the representation power with minimal increase of computational cost (Sec. \\\\ref{adaptive_params}).\\n\\\\vspace{-2ex}\",\n \"id\": \"58436ad0-55e1-4da8-b7e8-6bf9edce3fb1\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"3befdcee-686f-4399-9d86-c1c11ef02509\",\n \"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4\"\n ],\n \"title\": \"{Sample-wise Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 688,\n 685,\n 687\n ],\n \"content\": \"\\\\label{dynamic_arch}\\nConsidering different inputs may have diverse computational demands, it is natural to perform inference with dynamic architectures conditioned on each sample. Specifically, one can adjust the network depth (Sec. \\\\ref{dynamic_depth}), width (Sec. \\\\ref{dynamic_width}), or perform dynamic routing within a super network (SuperNet) that includes multiple possible paths (Sec. \\\\ref{SuperNets}). Networks with dynamic architectures not only save redundant computation for canonical (\\\"easy\\\") samples, but also preserve their representation power when recognizing non-canonical (\\\"hard\\\") samples. Such a property leads to remarkable advantages in efficiency compared to the acceleration techniques for static models , which handle \\\"easy\\\" and \\\"hard\\\" inputs with identical computation, and fail to reduce intrinsic computational redundancy.\\n\\\\vspace{-1.5ex}\",\n \"id\": \"3befdcee-686f-4399-9d86-c1c11ef02509\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"58436ad0-55e1-4da8-b7e8-6bf9edce3fb1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Architectures\"\n ]\n ],\n \"subsections\": [\n \"ca8e1af0-84b8-4669-8aed-b0ba1c15924a\",\n \"9747d6b0-66bd-4364-a0a9-4ece0a43fe6e\",\n \"ad51ce7b-9d07-4edc-88e3-0822d3ae0397\"\n ],\n \"title\": \"Dynamic Architectures\"\n },\n {\n \"cite_extract_rate\": 0.621621621621621,\n \"cites\": [\n 698,\n 695,\n 7273,\n 305,\n 693,\n 683,\n 690,\n 691,\n 7269,\n 8373,\n 38,\n 7270,\n 697,\n 97,\n 692,\n 694,\n 696,\n 676,\n 7272,\n 7271,\n 7266,\n 689,\n 7\n ],\n \"content\": \"\\\\label{dynamic_depth}\\nAs modern DNNs are getting increasingly deep for recognizing more \\\"hard\\\" samples, a straightforward solution to reducing redundant computation is performing inference with dynamic depth, which can be realized by 1) \\\\emph{early exiting}, i.e. allowing \\\"easy\\\" samples to be output at shallow exits without executing deeper layers ; or 2) \\\\emph{layer skipping}, i.e. selectively skipping intermediate layers conditioned on each sample .\\n\\\\begin{figure*}\\n \\\\vspace{-2ex}\\n \\\\centering\\n \\\\includegraphics[width=\\\\linewidth]{4_skip_layers.pdf}\\n \\\\vskip -0.175in\\n \\\\caption{Dynamic layer skipping. Feature $\\\\mathbf{x}_4$ in (a) are not calculated conditioned on the halting score, and the gating module in (b) decides whether to execute the block based on the intermediate feature. The policy network in (c) generates the skipping decisions for all layers in the main network.}\\n \\\\label{skipping_layers}\\n \\\\vspace{-3ex}\\n\\\\end{figure*}\\n\\\\noindent\\\\textbf{1) Early exiting.} The complexity (or \\\"difficulty\\\") of input samples varies in most real-world scenarios, and shallow networks are capable of correctly identifying many canonical inputs. Ideally, these samples should be output at certain early exits without executing deeper layers.\\nFor an input sample $\\\\mathbf{x}$, the forward propagation of an $L$-layer deep network $\\\\mathcal{F}$ could be represented by \\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{y} = \\\\mathcal{F}^{L}\\\\circ\\\\mathcal{F}^{L-1}\\\\circ\\\\cdots\\\\circ\\\\mathcal{F}^1(\\\\mathbf{x}),\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\n\\\\noindent where $\\\\mathcal{F}^{\\\\ell}$ denotes the operational function at layer $\\\\ell,1\\\\!\\\\le \\\\ell\\\\!\\\\le L$. In contrast, early exiting allows to terminate the inference procedure at an intermediate layer. For the $i$-th input sample $\\\\mathbf{x}_i$, the forward propagation can be written as\\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{y}_i = \\\\mathcal{F}^{\\\\ell_i}\\\\circ\\\\mathcal{F}^{\\\\ell_i-1}\\\\circ\\\\cdots\\\\circ\\\\mathcal{F}^1(\\\\mathbf{x}_i), 1\\\\!\\\\le \\\\ell_i\\\\!\\\\le L.\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nNote that $\\\\ell_i$ is adaptively determined based on $\\\\mathbf{x}_i$. Extensive architectures have been studied to endow DNNs with such early exiting behaviors, as discussed in the following.\\na) \\\\emph{Cascading of DNNs.} The most intuitive approach to enabling early exiting is cascading multiple models (see \\\\figurename~\\\\ref{cascading_models} (a)), and adaptively retrieving the prediction of an early network without activating latter ones. For example, Big/little-Net cascades two CNNs with different depths. After obtaining the \\\\emph{SoftMax} output of the first model, early exiting is conducted when the score margin between the two largest elements exceeds a threshold. {Moreover, a number of classic CNNs are cascaded in and }. After each model, a decision function is trained to determine whether the obtained feature should be fed to a linear classifier for immediate prediction, or be sent to subsequent models.\\nb) \\\\emph{Intermediate classifiers.}\\nThe models in the aforementioned cascading structures are mutually independent. Consequently, \\nonce a \\\"difficult\\\" sample is decided to be fed to a latter network, a whole inference procedure needs to be executed from scratch without reusing the already learned features. A more compact design is involving intermediate classifiers within one backbone network (see \\\\figurename~\\\\ref{cascading_models} (b)), so that early features can be propagated to deep layers if needed. Based on such a multi-exit architecture, adaptive early exiting is typically achieved according to confidence-based criteria or learned functions .\\nc) \\\\emph{Multi-scale architecture with early exits.} Researchers have observed that in chain-structured networks, the multiple classifiers may interfere with each other, which degrades the overall performance. A reasonable interpretation could be that in regular CNNs, the high-resolution features lack the coarse-level information that is essential for classification, leading to unsatisfying results for early exits. Moreover, early classifiers would force the shallow layers to generate \\\\emph{task-specialized} features, while a part of \\\\emph{general} information is lost, resulting in degraded performance for deep exits. To tackle this issue, multi-scale dense network (MSDNet) adopts 1) a \\\\emph{multi-scale} architecture, {which consists of multiple sub-networks for processing feature maps with different resolutions (scales)}, to quickly generate coarse-level features that are suitable for classification; 2) \\\\emph{dense connections}, to reuse early features and improve the performance of deep classifiers (see \\\\figurename~\\\\ref{multi_scale} (a)). Such a specially-designed architecture effectively enhances the overall accuracy of all the classifiers in the network. \\n{Based on the multi-scale architecture design, researchers have also studied the exiting policies (see \\\\figurename~\\\\ref{multi_scale} (b)) and training schemes of early-exiting dynamic models. More discussion about the inference and training schemes for dynamic models will be presented in Sec. \\\\ref{inference_and_train}.}\\nPrevious methods typically achieve the adaptation of network depths. From the perspective of exploiting spatial redundancy in features, resolution adaptive network (RANet, see \\\\figurename~\\\\ref{multi_scale} (c)) first processes each sample with low-resolution features, while high-resolution representations are conditionally utilized based on early predictions.\\nAdaptive early exiting is also extended to language models (e.g. BERT ) for improving their efficiency on NLP tasks . In addition, it can be implemented in recurrent neural networks (RNNs) for \\\\emph{temporally} dynamic inference when processing sequential data such as videos and texts (see Sec. \\\\ref{sec_temporal_adaptive}).\\n\\\\noindent\\\\textbf{2) Layer skipping.} The general idea of the aforementioned early-exiting paradigm is skipping the execution of all the deep layers after a certain classifier. More flexibly, the network depth can also be adapted on the fly by strategically skipping the calculation of \\\\emph{intermediate layers} without placing extra classifiers. Given the $i$-th input sample $\\\\mathbf{x}_i$, dynamic layer skipping could be generally written as\\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{y}_i = (\\\\mathds{1}^{L}\\\\circ\\\\mathcal{F}^{L})\\\\circ (\\\\mathds{1}^{L-1}\\\\circ\\\\mathcal{F}^{L-1})\\\\circ\\\\cdots\\\\circ (\\\\mathds{1}^1\\\\circ\\\\mathcal{F}^1)(\\\\mathbf{x}_i),\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nwhere $\\\\mathds{1}^{\\\\ell}$ denotes the indicator function determining the execution of layer $\\\\mathcal{F}^{\\\\ell}, 1\\\\!\\\\le\\\\!\\\\ell\\\\!\\\\le\\\\!L$.\\nThis scheme is typically implemented on structures with skip connections (e.g. ResNet ) to guarantee the continuity of forward propagation, and here we summarize three common approaches.\\na) \\\\emph{Halting score} is first proposed in , where an accumulated scalar named halting score adaptively decides whether the hidden state of an RNN will be directly fed to the next time step. The halting scheme is extended to vision tasks by viewing residual blocks within a ResNet stage \\\\footnote{Here we refer to a stage as a stack of multiple residual blocks with the same feature resolution.} as linear layers within a step of RNN (see \\\\figurename~\\\\ref{skipping_layers} (a)).\\nRather than skipping the execution of layers with independent parameters, multiple blocks in each ResNet stage could be replaced by one weight-sharing block, leading to a significant reduction of parameters . In every stage, the block is executed for an adaptive number of steps according to the halting score.\\nIn addition to RNNs and CNNs, the halting scheme is further implemented on Transformers by and to achieve dynamic network depth on NLP tasks.\\nb) {\\\\emph{Gating function} is also a prevalent option for dynamic layer skipping due to its plug-and-play property.} Take ResNet as an example (see \\\\figurename~\\\\ref{skipping_layers} (b)), let $\\\\mathbf{x}^{\\\\ell}$ denote the input feature of the $\\\\ell$-th residual block, gating function $\\\\mathcal{G}^{\\\\ell}$ generates a binary value to decide the execution of residual block $\\\\mathcal{F}^{\\\\ell}$. This procedure could be represented by\\\\footnote{For simplicity and without generality, the subscript for sample index will be omitted in the following.}\\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{x}^{\\\\ell+1} = \\\\mathcal{G}^{\\\\ell}(\\\\mathbf{x}^{\\\\ell})\\\\mathcal{F}^{\\\\ell}(\\\\mathbf{x}^{\\\\ell}) + \\\\mathbf{x}^{\\\\ell}, \\\\mathcal{G}^{\\\\ell}(\\\\mathbf{x}^{\\\\ell})\\\\in\\\\{0,1\\\\}.\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nSkipNet and convolutional network with adaptive inference graph (Conv-AIG) are two typical approaches to enabling dynamic layer skipping. Both methods induce lightweight computational overheads to efficiently produce the binary decisions on whether skipping the calculation of a residual block. Specifically, Conv-AIG utilizes two FC layers in each residual block, while the gating function in SkipNet is implemented as an RNN for parameter sharing.\\n\\\\begin{figure*}\\n \\\\centering\\n \\\\vspace{-2ex}\\n \\\\includegraphics[width=0.85\\\\linewidth]{5_multi_branch.pdf}\\n \\\\vskip -0.175in\\n \\\\caption{MoE with soft weighting (a) and hard gating (b) schemes both adopt an auxiliary module to generate the weights or gates. In the tree structure (c), features (nodes) and transformations (paths) are represented as circles and lines with arrows respectively. Only the full lines are activated.}\\n \\\\label{multi_branch}\\n \\\\vspace{-3ex}\\n\\\\end{figure*}\\nRather than skipping layers in classic ResNets, dynamic recursive network iteratively executes one block with shared parameters in each stage. Although the weight-sharing scheme seems similar to the aforementioned IamNN , the skipping policy of differs significantly: gating modules are exploited to decide the recursion depth.\\nInstead of either skipping a layer, or executing it thoroughly with a full numerical precision, a line of work studies adaptive \\\\emph{bit-width} for different layers conditioned on the \\\\emph{resource budget}. Furthermore, fractional skipping adaptively selects a bit-width for each residual block by a gating function based on \\\\emph{input features}.\\nc) {\\\\emph{Policy network} can be built to take in an input sample, and directly produces the skipping decisions for all the layers in a backbone network (see \\\\figurename~\\\\ref{skipping_layers} (c)).}\\n\\\\vspace{-1.5ex}\",\n \"id\": \"ca8e1af0-84b8-4669-8aed-b0ba1c15924a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 37,\n \"parent_id\": \"3befdcee-686f-4399-9d86-c1c11ef02509\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Architectures\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Depth\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Depth\"\n },\n {\n \"cite_extract_rate\": 0.764705882352941,\n \"cites\": [\n 706,\n 688,\n 705,\n 707,\n 8375,\n 544,\n 691,\n 8374,\n 710,\n 8373,\n 712,\n 713,\n 709,\n 699,\n 38,\n 711,\n 703,\n 680,\n 696,\n 702,\n 7266,\n 700,\n 701,\n 708,\n 704,\n 687\n ],\n \"content\": \"\\\\label{dynamic_width}\\nIn addition to dynamic network \\\\emph{depth} (Sec. \\\\ref{dynamic_depth}), a finer-grained form of conditional computation is performing inference with dynamic \\\\emph{width}: although every layer is executed, its multiple units (e.g. neurons, channels or branches) are selectively activated conditioned on the input.\\n\\\\noindent\\\\textbf{1) {Skipping neurons in fully-connected (FC) layers.}} The computational cost of an FC layer is determined by its input and output dimensions. It is commonly believed that different neuron units are responsible for representing different features, and therefore not all of them need to be activated for every sample. Early studies learn to adaptively control the neuron activations by auxiliary branches or other techniques such as low-rank approximation .\\n\\\\noindent\\\\textbf{2) {Skipping branches in mixture-of-experts (MoE).}} In Sec. \\\\ref{dynamic_depth}, adaptive model ensemble is achieved via a \\\\emph{cascading} way, and later networks are conditionally executed based on early predictions. An alternative approach to improving the model capacity is the MoE structure, which means that multiple network branches are built as experts in \\\\emph{parallel}. \\nThese experts could be selectively executed, and their outputs are fused with data-dependent weights.\\nConventional \\\\emph{soft} MoEs adopt real-valued weights to dynamically rescale the representations obtained from different experts (\\\\figurename~\\\\ref{multi_branch} (a)). In this way, all the branches still need to be executed, and thus the computation cannot be reduced at test time. \\\\emph{Hard} gates with only a fraction of non-zero elements are developed to increase the inference efficiency of the MoE structure (see \\\\figurename~\\\\ref{multi_branch} (b)) : let $\\\\mathcal{G}$ denote a gating module whose output is a $N$-dimensional vector $\\\\bm{\\\\alpha}$ controlling the execution of $N$ experts $\\\\mathcal{F}_1,\\\\mathcal{F}_2, \\\\cdots, \\\\mathcal{F}_N$, the final output can be written as\\n\\\\begin{equation} \\\\label{eq_moe}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{y} = \\\\sum\\\\nolimits_{n=1}^N [\\\\mathcal{G}(\\\\mathbf{x})]_n \\\\mathcal{F}_n (\\\\mathbf{x}) = \\\\sum\\\\nolimits_{n=1}^N \\\\alpha_n \\\\mathcal{F}_n (\\\\mathbf{x}),\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nand the $n$-th expert will not be executed if $\\\\alpha_n\\\\! =\\\\!0$. \\nHard MoE has been implemented in diverse network structures. For example, HydraNet replaces the convolutional blocks in a CNN by multiple branches, and selectively execute them conditioned on the input. For another example, dynamic routing network (DRNet) performs a branch selection in each cell structure which is commonly used in NAS . On NLP tasks, sparely gated MoE and switch Transformer embeds hard MoE in a long short-term memory (LSTM) network and a Transformer , respectively. Instead of making choice with \\\\emph{binary} gates as in , only the branches corresponding to the \\\\emph{top-K} elements of the real-valued gates are activated in .\\n\\\\noindent\\\\textbf{{3) Skipping channels in CNNs.}}\\nModern CNNs usually have considerable channel redundancy. Based on the common belief that the same feature channel can be of disparate importance for different samples, adaptive width of CNNs could be realized by dynamically activating convolutional channels. Compared to the \\\\emph{static} pruning methods which remove \\\"unimportant\\\" filters permanently, such a \\\\emph{data-dependent} pruning approach improves the inference efficiency without degrading the model capacity.\\na) \\\\emph{Multi-stage architectures along the channel dimension.} Recall that the early-exiting networks discussed in Sec. \\\\ref{dynamic_depth} can be viewed as multi-stage models along the \\\\emph{depth} dimension, where late stages are conditionally executed based on early predictions. One can also build multi-stage architectures along the \\\\emph{width} (channel) dimension, and progressively execute these stages on demand.\\nAlong this direction, an optimal architecture is searched among multiple structures with different widths, and any sample can be output at an early stage when a confident prediction is obtained . Channel gating network (CGNet) first executes a subset of convolutional filters in every layer, and the remaining filters are only activated on strategically selected areas.\\nb) \\\\emph{Dynamic pruning with gating functions.} In the aforementioned progressive activation paradigm, the execution of a later stage is decided based on previous output. As a result, a complete forward propagation is required for every stage, which might be suboptimal for reducing the practical inference latency. Another prevalent solution is to decide the execution of channels at every layer by gating functions. For example, runtime neural pruning (RNP) models the layer-wise pruning as a Markov decision process, and an RNN is used to select specific channel groups at every layer. Moreover, pooling operations followed by FC layers are utilized to generate {\\\\emph{channel-wise hard attention} (i.e. making discrete decisions on whether to activate each channel)} for each sample . The recent work uses a gate module to decide the width for a whole stage of a ResNet. Different reparameterization and optimizing techniques are required for training these gating functions, which will be reviewed in Sec. \\\\ref{training}.\\nRather than placing plug-in gating modules \\\\emph{inside} a CNN, GaterNet builds an \\\\emph{extra} network, which takes in the input sample and directly generates all the channel selection decisions for the backbone CNN. This implementation is similar to BlockDrop that exploits an additional policy network for dynamic layer skipping (Sec. \\\\ref{dynamic_depth}).\\nc) \\\\emph{Dynamic pruning based on feature activations directly} has also been realized without auxiliary branches and computational overheads , where a regularization item is induced in training to encourage the sparsity of features.\\n{On basis of the existing literature on dynamically skipping either network \\\\emph{layers} or convolutional \\\\emph{filters} , recent work has realized dynamic inference with respect to network \\\\emph{depth} and \\\\emph{width} simultaneously: only if a layer is determined to be executed, its channels will be selectively activated, leading to a more flexible adaptation of computational graphs.}\\n\\\\vspace{-1ex}\",\n \"id\": \"9747d6b0-66bd-4364-a0a9-4ece0a43fe6e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 34,\n \"parent_id\": \"3befdcee-686f-4399-9d86-c1c11ef02509\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Architectures\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Width\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Width\"\n },\n {\n \"cite_extract_rate\": 0.684210526315789,\n \"cites\": [\n 715,\n 714,\n 8374,\n 8373,\n 719,\n 720,\n 718,\n 696,\n 7266,\n 7274,\n 716,\n 717,\n 306\n ],\n \"content\": \"\\\\label{SuperNets}\\nThe aforementioned methods mostly adjust the depth (Sec. \\\\ref{dynamic_depth}) or width (Sec. \\\\ref{dynamic_width}) of \\\\emph{classic architectures} by activating their computational units (e.g. layers or channels ) conditioned on the input. {Another line of work develops different forms of SuperNets with various possible inference paths, and performs dynamic routing inside the SuperNets to adapt the computational graph to each sample.}\\nTo achieve dynamic routing, there are typically routing nodes that are responsible for allocating features to different paths. For node $s$ at layer $\\\\ell$, let $\\\\alpha_{s\\\\rightarrow j}^{\\\\ell}$ denote the probability of assigning the reached feature $\\\\mathbf{x}^{\\\\ell}_s$ to node $j$ at layer $\\\\ell+1$, the path $\\\\mathcal{F}_{s\\\\rightarrow j}^{\\\\ell}$ will be activated only when $\\\\alpha_{s\\\\rightarrow j}^{\\\\ell}\\\\!>\\\\!0$. The resulting feature reaching node $j$ is represented by\\n\\\\begin{equation}\\\\label{eq_supernet}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{x}^{\\\\ell+1}_j = \\\\sum\\\\nolimits_{s\\\\in\\\\left\\\\{s: \\\\alpha_{s\\\\rightarrow j}^{\\\\ell}>0\\\\right\\\\}} \\\\alpha_{s\\\\rightarrow j}^{\\\\ell}\\\\mathcal{F}_{s\\\\rightarrow j}^{\\\\ell}(\\\\mathbf{x}^{\\\\ell}_s).\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nThe probability $\\\\alpha_{s\\\\rightarrow j}^{\\\\ell}$ can be obtained in different manners. \\nNote that the dynamic early-exiting networks are a special form of SuperNets, where the routing decisions are only made at intermediate classifiers. The CapsuleNets also perform dynamic routing between capsules, i.e. groups of neurons, to character the relations between (parts of) objects. Here we mainly focus on specific architecture designs of the SuperNets and their routing policies.\\n\\\\noindent\\\\textbf{1) Path selection in multi-branch structures.} The simplest dynamic routing can be implemented by selectively executing \\\\emph{one} of multiple candidate modules at each layer , which is equivalent to producing a one-hot probability distribution $\\\\alpha_{s\\\\rightarrow \\\\cdot}^{\\\\ell}$ in Eq. \\\\ref{eq_supernet}. The main difference of this approach to hard MoE (\\\\figurename~\\\\ref{multi_branch} (b)) is that only one branch is activated without any fusion operations.\\n\\\\noindent\\\\textbf{2) Neural trees and tree-structured networks.} \\nAs decision trees always perform inference along one forward path that is dependent on input properties, combining tree structure with neural networks can naturally enjoy the adaptive inference paradigm and the representation power of DNNs simultaneously. Note that in a tree structure, the outputs of different nodes are routed to \\\\emph{independent} paths rather than being \\\\emph{fused} as in MoE structures (compare \\\\figurename~\\\\ref{multi_branch} (b), (c)).\\na) \\\\emph{Soft decision tree} (SDT) adopts neural units as routing nodes (blue nodes in \\\\figurename~\\\\ref{multi_branch} (c)), which decides the portion that the inputs are assigned to their left/right sub-tree. Each leaf node generates a probability distribution over the output space, and the final prediction is the expectation of the results from all leaf nodes. Although the probability for a sample reaching each leaf node in an SDT is data-dependent, all the paths are still executed, which limits the inference efficiency.\\nb) \\\\emph{Neural trees with deterministic routing policies} are designed to make hard routing decisions during inference, avoiding computation on those unselected paths.\\nc) \\\\emph{Tree-structured DNNs.} Instead of developing decision trees containing neural units, a line of work builds special network architectures to endow them with the routing behavior of decision trees. For instance, a small CNN is first executed to classify each sample into coarse categories, and specific sub-networks are conditionally activated based on the coarse predictions . A subsequent work not only partitions samples to different sub-networks, but also divides and routes the feature channels.\\nDifferent to those networks using neural units only in routing nodes , or routing each sample to pre-designed sub-networks , adaptive neural tree (ANT) adopts CNN modules as feature transformers in a hard neural tree (see lines with arrows in \\\\figurename~\\\\ref{multi_branch} (c)), and learns the tree structure together with the network parameters simultaneously in the training stage.\\n\\\\noindent\\\\textbf{3) Others.} \\n{Performing dynamic routing within more general SuperNet architectures is also a recent research trend. Representatively, an architecture distribution with partly shared parameters is \\\\emph{searched} from a SuperNet containing $\\\\sim \\\\!\\\\! 10^{25}$ sub-networks . During inference, every sample is allocated by a controller network to one sub-network with appropriate computational cost. Instead of training a standalone controller network, gating modules are plugged inside the \\\\emph{hand-designed} SuperNet (see \\\\figurename~\\\\ref{multi_scale} (d)) to decide the routing path based on intermediate features . \\n\\\\begin{figure*}\\n \\\\centering\\n \\\\vspace{-2ex}\\n \\\\includegraphics[width=0.9\\\\linewidth]{6_dynamic_parameter.pdf}\\n \\\\vskip -0.2in\\n \\\\caption{{Three implementations of dynamic parameters: adjusting (a) or generating (b) the backbone parameters based on the input, and (c) dynamically rescaling the features with the attention mechanism.}}\\n \\\\label{dynamic_params}\\n \\\\vskip -0.2in\\n\\\\end{figure*}\\n\\\\vspace{-1.5ex}\",\n \"id\": \"ad51ce7b-9d07-4edc-88e3-0822d3ae0397\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"3befdcee-686f-4399-9d86-c1c11ef02509\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Architectures\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Routing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Routing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{adaptive_params}\\n\\\\vspace{-0.25ex}\\nAlthough the networks with dynamic \\\\emph{architectures} in Sec. \\\\ref{dynamic_arch} can adapt their inference graphs to each sample and achieve an efficient allocation of computation, they usually have special architecture designs, requiring specific training strategies or careful hyper-parameters tuning (Sec. \\\\ref{sec:discussion}).\\nAnother line of work adapts network \\\\emph{parameters} to different inputs while keeping the architectures fixed, which has been shown effective in improving the representation power of networks with a minor increase of computational cost. Given an input sample $\\\\mathbf{x}$, \\nthe output of a conventional network (module) with static parameters can be written as $\\\\mathbf{y}\\\\! =\\\\!\\\\mathcal{F}(\\\\mathbf{x},\\\\bm{\\\\Theta})$. In contrast, the output of a model with dynamic parameters could be represented by\\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{y} = \\\\mathcal{F}(\\\\mathbf{x},\\\\bm{\\\\hat{\\\\Theta}}|\\\\mathbf{x}) = \\\\mathcal{F}(\\\\mathbf{x},\\\\mathcal{W}(\\\\mathbf{x}, \\\\bm{\\\\Theta})),\\n \\\\label{eq:dynamic_parameter}\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nwhere $\\\\mathcal{W}(\\\\cdot, \\\\bm{\\\\Theta})$ is the operation producing input-dependent parameters, and its design has been extensively explored.\\nIn general, the parameter adaptation can be achieved from three aspects (see \\\\figurename~\\\\ref{dynamic_params}): 1) adjusting the trained parameters based on the input (Sec. \\\\ref{dynamic_param_adjust}); 2) directly generating the network parameters from the input (Sec. \\\\ref{weight_predict}); and 3) rescaling the features with soft attention (Sec. \\\\ref{sec:attention}).\\n\\\\vspace{-1ex}\",\n \"id\": \"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"58436ad0-55e1-4da8-b7e8-6bf9edce3fb1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Parameters\"\n ]\n ],\n \"subsections\": [\n \"8a3a1403-da1a-48be-b1e9-80d1325e0909\",\n \"81d9301c-7d1b-4044-b144-7027a7b2165b\",\n \"0651cec9-e0a5-4c53-aca3-aa4ed8ac7d94\"\n ],\n \"title\": \"Dynamic Parameters\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 675,\n 8376,\n 721,\n 7276,\n 7275,\n 722\n ],\n \"content\": \"\\\\label{dynamic_param_adjust}\\n\\\\vspace{-0.5ex}\\nA typical approach to parameter adaptation is adjusting the weights based on their input during inference as presented in \\\\figurename~\\\\ref{dynamic_params} (a). This implementation usually evokes little computation to obtain the adjustments, e.g., attention weights or sampling offsets .\\n\\\\noindent\\\\textbf{1) Attention on weights.} {To improve the representation power without noticeably increasing the computation, \\nsoft attention can be performed on multiple convolutional kernels, producing an adaptive ensemble of parameters .}\\nAssuming that there are $N$ kernels $\\\\mathbf{W}_n, n\\\\! =1,2,\\\\!\\\\cdots\\\\!,N$, such a dynamic convolution \\ncan be formulated as \\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n\\\\mathbf{y} = \\\\mathbf{x} \\\\star \\\\mathbf{\\\\tilde{W}} = \\\\mathbf{x} \\\\star (\\\\sum\\\\nolimits_{n=1}^N \\\\alpha_n\\\\mathbf{W}_n). \\n\\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nThis procedure increases the model capacity yet remains high efficiency, as the result obtained through fusing the outputs of $N$ convolutional branches (as in MoE structures, see \\\\figurename~\\\\ref{multi_branch} (a)) is equivalent to that produced by performing once convolution with $\\\\mathbf{\\\\tilde{W}}$. However, only $\\\\sim\\\\!1/N$ times of computation is consumed in the latter approach.\\nWeight adjustment could also be achieved by performing soft attention over the \\\\emph{spatial locations} of convolutional weights . For example, segmentation-aware convolutional network applies locally masked convolution to aggregate information with larger weights from similar pixels, which are more likely to belong to the same object.\\nUnlike that requires a sub-network for feature embedding, pixel-adaptive convolution (PAC) adapts the convolutional weights based on the attention mask generated from the input feature at each layer.\\n{Instead of adjusting weights conditioned on every sample itself, meta-neighborhoods adapt the network parameters to each input sample based on its similarity to the neighbors stored in a dictionary.}\\n\\\\noindent\\\\textbf{2) Kernel shape adaptation.} \\nApart from adaptively scaling the weight \\\\emph{values}, parameter adjustment can also be realized to reshape the convolutional kernels and achieve \\\\emph{dynamic reception of fields}. Towards this direction, \\ndeformable convolutions sample feature pixels from adaptive locations when performing convolution on each pixel. Deformable kernels samples weights in the kernel space to adapt the \\\\emph{effective} reception field (ERF) while leaving the reception field unchanged. Table \\\\ref{tab:deform_kernels} summarizes the formulations of the above three methods. \\n{Due to their irregular memory access and computation pattern, these kernel shape adaptation approaches typically require customized CUDA kernels for the implementation on GPUs. However, recent literature has shown that the practical efficiency of deformable convolution could be effectively improved by co-designing algorithm and hardware based on embedded devices such as FPGAs .}\\n\\\\begin{table*}\\n \\\\vspace{-2ex}\\n \\\\caption{{Kernel shape adaptation by dynamically sampling feature pixels or convolutional weights .}}\\n \\\\vspace{-4ex}\\n \\\\label{tab:deform_kernels}\\n \\\\begin{center}\\n \\\\begin{tabular}{c|c|c|c}\\n \\\\hline\\n \\\\textbf{Method} & \\\\textbf{Formulation} & \\\\textbf{Sampled Target} & \\\\textbf{Dynamic Mask} \\\\\\\\\\n \\\\hline\\n Regular Convolution & $\\\\mathbf{y(p)} = \\\\sum\\\\nolimits_{k=1}^K \\\\mathbf{W}(\\\\mathbf{p}_k) \\\\mathbf{x}(\\\\mathbf{p+p}_k)$ & - & - \\\\\\\\\\n \\\\hline\\n Deformable ConvNet-v1 & $\\\\mathbf{y(p)} = \\\\sum\\\\nolimits_{k=1}^K \\\\mathbf{W}(\\\\mathbf{p}_k) \\\\mathbf{x}(\\\\mathbf{p+p}_k+\\\\Delta \\\\mathbf{p}_k)$ & Feature map & No \\\\\\\\\\n Deformable ConvNet-v2 & $\\\\mathbf{y(p)} = \\\\sum\\\\nolimits_{k=1}^K \\\\mathbf{W}(\\\\mathbf{p}_k) \\\\mathbf{x}(\\\\mathbf{p+p}_k+\\\\Delta \\\\mathbf{p}_k)\\\\Delta \\\\mathbf{m}_k$ & Feature map & Yes \\\\\\\\\\n Deformable Kernels & $\\\\mathbf{y(p)} = \\\\sum\\\\nolimits_{k=1}^K \\\\mathbf{W}(\\\\mathbf{p}_k+\\\\Delta \\\\mathbf{p}_k) \\\\mathbf{x}(\\\\mathbf{p+p}_k)$ & Conv kernel & No \\\\\\\\\\n \\\\hline\\n \\\\end{tabular}\\n \\\\end{center}\\n \\\\vspace{-5ex}\\n\\\\end{table*}\\n\\\\vspace{-1ex}\",\n \"id\": \"8a3a1403-da1a-48be-b1e9-80d1325e0909\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Parameters\"\n ],\n [\n \"subsubsection\",\n \"Parameter Adjustment\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Parameter Adjustment\"\n },\n {\n \"cite_extract_rate\": 0.785714285714285,\n \"cites\": [\n 675,\n 726,\n 7277,\n 724,\n 7215,\n 678,\n 727,\n 725,\n 8377,\n 677,\n 723\n ],\n \"content\": \"\\\\label{weight_predict}\\n\\\\vspace{-0.25ex}\\nCompared to making modifications on model parameters on the fly (Sec. \\\\ref{dynamic_param_adjust}), weight prediction is more straightforward: it directly generates (a subset of) input-adaptive parameters with an independent model at test time (see \\\\figurename~\\\\ref{dynamic_params} (b)). This idea was first suggested in , where both the weight prediction model and the backbone model were feedforward networks. Recent work has further extended the paradigm to modern network architectures and tasks.\\n\\\\noindent\\\\textbf{1) General architectures.} Dynamic filter networks (DFN) and HyperNetworks are two classic approaches realizing runtime weight prediction for CNNs and RNNs, respectively. Specifically, a filter generation network is built in DFN to produce the filters for a convolutional layer. As for processing sequential data (e.g. a sentence), the weight matrices of the main RNN are predicted by a smaller one at each time step conditioned on the input (e.g. a word) . WeightNet unifies the dynamic schemes of and by predicting the convolutional weights via simple grouped FC layers, achieving competitive results in terms of the accuracy-FLOPs\\\\footnote{{Floating point operations, which is widely used as a measure of inference efficiency of deep networks.}} and accuracy-parameters trade-offs.\\n{Rather than generating standard \\\\emph{convolutional} weights, LambdaNetworks learns to predict the weights of \\\\emph{linear} projections based on the contexts of each pixel together with the relative position embeddings, showing advantages in terms of computational cost and memory footprint.}\\n\\\\noindent\\\\textbf{2) Task-specific information} \\nhas also been exploited to predict model parameters on the fly, enabling dynamic networks to generate task-aware feature embeddings. For example, edge attributes are utilized in to generate filters for graph convolution, and camera perspective is incorporated in to generate weights for image convolution. {Such task-aware weight prediction has been shown effective in improving the data efficiency on many tasks, including visual question answering and few-shot learning .}\\n\\\\vspace{-1ex}\",\n \"id\": \"81d9301c-7d1b-4044-b144-7027a7b2165b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Parameters\"\n ],\n [\n \"subsubsection\",\n \"Weight Prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Weight Prediction\"\n },\n {\n \"cite_extract_rate\": 0.7619047619047611,\n \"cites\": [\n 736,\n 730,\n 729,\n 7278,\n 8378,\n 38,\n 734,\n 8379,\n 733,\n 731,\n 7040,\n 728,\n 735,\n 7268,\n 7,\n 732\n ],\n \"content\": \"\\\\label{sec:attention}\\n\\\\vspace{-0.25ex}\\nThe main goal of either \\\\emph{adjusting} (Sec. \\\\ref{dynamic_param_adjust}) or \\\\emph{predicting} (Sec. \\\\ref{weight_predict}) model parameters is producing more dynamic and informative features, and therefore enhancing the representation power of deep networks. A more straightforward solution is rescaling the features with input-dependent soft attention (see \\\\figurename~\\\\ref{dynamic_params} (c)), which requires minor modifications on computational graphs. Note that for a linear transformation $\\\\mathcal{F}$, applying attention $\\\\bm{\\\\alpha}$ on the output is equivalent to performing computation with re-weighted parameters, i.e.\\n\\\\begin{equation}\\\\label{dynamic_feature_key}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathcal{F}(\\\\mathbf{x},\\\\bm{\\\\Theta})\\\\otimes \\\\bm{\\\\alpha}= \\\\mathcal{F}(\\\\mathbf{x},\\\\bm{\\\\Theta}\\\\otimes\\\\bm{\\\\alpha}).\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\n\\\\noindent\\\\textbf{1) Channel-wise attention} is one of the most common soft attention mechanisms. Existing work typically follows \\nthe form in squeeze-and-excitation network (SENet) :\\n\\\\begin{equation}\\\\label{se_attention}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n\\\\mathbf{\\\\tilde{y}}\\\\! = \\\\!\\\\mathbf{y} \\\\otimes \\\\bm{\\\\alpha}\\\\! =\\\\!\\\\mathbf{y} \\\\otimes \\\\mathcal{A}(\\\\mathbf{y}), \\\\bm{\\\\alpha}\\\\in\\\\left[0,1\\\\right]^C.\\n\\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation} \\nIn Eq. \\\\ref{se_attention}, $\\\\mathbf{y}\\\\! =\\\\!\\\\mathbf{x} \\\\star \\\\mathbf{W}$ is the output feature of a convolutional layer with $C$ channels, and $\\\\mathcal{A}(\\\\cdot)$ is a lightweight function composed of pooling and linear layers for producing $\\\\bm{\\\\alpha}$.\\nTaking the convolution into account, the procedure can also be written as $\\\\mathbf{\\\\tilde{y}}\\\\! = \\\\!(\\\\mathbf{x} \\\\star \\\\mathbf{W}) \\\\otimes \\\\bm{\\\\alpha}\\\\! =\\\\! \\\\mathbf{x} \\\\star (\\\\mathbf{W} \\\\otimes \\\\bm{\\\\alpha})$, from which we can observe that applying attention on features is equivalent to performing convolution with dynamic weights.\\nOther implementations for attention modules have also been developed, including using standard deviation to provide more statistics , or replacing FC layers with efficient 1D convolutions .\\nThe empirical performance of three computational graphs for soft attention is studied in : 1) $\\\\mathbf{\\\\tilde{y}}\\\\! =\\\\!\\\\mathbf{y}\\\\otimes \\\\mathcal{A}(\\\\mathbf{y})$, 2) $\\\\mathbf{\\\\tilde{y}}\\\\! =\\\\!\\\\mathbf{y}\\\\otimes \\\\mathcal{A}(\\\\mathbf{x})$ and 3) $\\\\mathbf{\\\\tilde{y}}\\\\! =\\\\!\\\\mathbf{y}\\\\otimes \\\\mathcal{A}(\\\\mathrm{Conv}(\\\\mathbf{x}))$. It is found that the three forms yield different performance in different backbone networks.\\n\\\\noindent\\\\textbf{2) Spatial-wise attention}. Spatial locations in features could also be dynamically rescaled with attention to improve the representation power of deep models . Instead of using pooling operations to efficiently gather global information as in channel-wise attention, convolutions are often adopted in spatial-wise attention to encode local information. Moreover, these two types of attention modules can be integrated in one framework (see \\\\figurename~\\\\ref{dynamic_params} (c)).\\n\\\\noindent\\\\textbf{3) Dynamic activation functions.} The aforementioned approaches to generating dynamic features usually apply soft attention before static activation functions. A recent line of work has sought to increase the representation power of models with dynamic activation functions . For instance, \\nDY-ReLU replaces ReLU $(\\\\mathbf{y}_c\\\\! =\\\\!\\\\max (\\\\mathbf{x}_c, 0))$ with the max value among $N$ linear transformations $\\\\mathbf{y}_c\\\\! =\\\\!\\\\max_{n} \\\\left\\\\{a_c^n \\\\mathbf{x}_c + b_c^n\\\\right\\\\}$, where $c$ is the channel index, and $a_c^n, b_c^n$ are linear coefficients calculated from $\\\\mathbf{x}$. On many vision tasks, these dynamic activation functions can effectively improve the performance of different network architectures with negligible computational overhead.\\nTo summarize, soft attention has been exploited in many fields due to its simplicity and effectiveness. Moreover, it can be incorporated with other methods conveniently. E.g., by replacing the weighting scalar $\\\\alpha_n$ in Eq. \\\\ref{eq_moe} with channel-wise or spatial-wise attention, the output of multiple branches with independent kernel sizes or feature resolutions are adaptively fused.\\nNote that we leave out the detailed discussion on the self attention mechanism, which is widely studied in both NLP and CV fields to re-weight features based on the similarity between queries and keys at different locations (temporal or spatial). \\nReaders who are interested in this topic may refer to review studies . In this survey, we mainly focus on the feature re-weighting scheme in the framework of dynamic inference. \\n\\\\vspace{-2ex}\",\n \"id\": \"0651cec9-e0a5-4c53-aca3-aa4ed8ac7d94\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"fdd3c2e8-dc39-4dd8-8a63-1d07741f61b4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Sample-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Dynamic Parameters\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Features\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Features\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 737,\n 504\n ],\n \"content\": \"\\\\vspace{-0.25ex}\\n\\\\label{sec_spatially_adaptive}\\nIn visual learning, it has been found that not all locations contribute equally to the final prediction of CNNs , which suggests that \\\\emph{spatially} dynamic computation has great potential for reducing computational redundancy.\\nIn other words, making a correct prediction may only require processing a fraction of pixels or regions with an adaptive amount of computation. Moreover, based on the observations that low-resolution representations are sufficient to yield decent performance for most inputs , the static CNNs that take in all the input with the same resolution may also induce considerable redundancy.\\nTo this end, spatial-wise dynamic networks are built to perform adaptive inference with respect to different spatial locations of images.\\nAccording to the granularity of dynamic computation, we further categorize the relevant approaches into three levels: {\\\\emph{pixel level} (Sec. \\\\ref{pixel_level}), \\\\emph{region level} (Sec. \\\\ref{region_level}) and \\\\emph{resolution level} (Sec. \\\\ref{dynamic_resolution}).}\\n\\\\vspace{-2ex}\",\n \"id\": \"d1dfc651-19ab-46ab-9db8-02eef3371ca0\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"bb93a738-d7fa-4ffe-811c-d2e5d3062ddd\",\n \"0d52450e-77a8-4861-a2eb-00ddb1ece863\",\n \"ab0c50b9-9d18-4c18-976a-34aae739e7ac\"\n ],\n \"title\": \"Spatial-wise Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\vspace{-0.25ex}\\n\\\\label{pixel_level}\\nCommonly seen spatial-wise dynamic networks perform adaptive computation at the pixel level. Similar to the categorization in Sec. \\\\ref{sec_sample_wise}, pixel-level dynamic networks are grouped into two types: models with pixel-specific \\\\emph{dynamic architectures} (Sec. \\\\ref{pixel_dynamic_arch}) and \\\\emph{dynamic parameters} (Sec. \\\\ref{pixel_dynamic_param}).\\n\\\\vspace{-1ex}\",\n \"id\": \"bb93a738-d7fa-4ffe-811c-d2e5d3062ddd\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"d1dfc651-19ab-46ab-9db8-02eef3371ca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Pixel-level Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"fa853b1f-afd4-40d2-bed1-be3a1d6cb128\",\n \"530cae77-537a-4dca-80a6-e035f5680e22\"\n ],\n \"title\": \"Pixel-level Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0.7272727272727271,\n \"cites\": [\n 701,\n 738,\n 7279,\n 7280,\n 676,\n 739,\n 7267,\n 7281\n ],\n \"content\": \"\\\\label{pixel_dynamic_arch}\\n\\\\vspace{-0.25ex}\\nBased on the common belief that foreground pixels are more informative and computational demanding than those in the background, some dynamic networks learn to adjust their architectures for each pixel. Existing literature generally achieves this by 1) \\\\emph{dynamic sparse convolution}, which only performs convolutions on a subset of sampled pixels; 2) \\\\emph{additional refinement}, which strategically allocates extra computation (e.g. layers or channels) on certain spatial positions.\\n\\\\begin{figure}\\n \\\\centering\\n \\\\vspace{-1ex}\\n \\\\includegraphics[width=\\\\linewidth]{8_spatial_pixel_sparse.pdf}\\n \\\\vskip -0.15in\\n \\\\caption{{Dynamic convolution on selected spatial locations. The 1 elements (black) in the spatial mask determine the pixels (green) that require computation in the output feature map.}}\\n \\\\label{fig_spatial_location_specific}\\n \\\\vspace{-3ex}\\n\\\\end{figure}\\n\\\\noindent\\\\textbf{1) Dynamic sparse convolution.} To reduce the unnecessary computation on less informative locations, convolution can be performed only on strategically sampled pixels. Existing sampling strategies include 1) making use of the intrinsic sparsity of the input ; 2) predicting the positions of zero elements on the output ; and 3) estimating the saliency of pixels . A typical approach is using an extra branch to generate a spatial mask, determining the execution of convolution on each pixel (see \\\\figurename~\\\\ref{fig_spatial_location_specific}). {Pixel-wise dynamic depth could also be achieved based on a halting scheme } (see Sec. \\\\ref{dynamic_depth}). These dynamic convolutions usually neglect the unselected positions, which might degrade the network performance. Interpolation is utilized in to efficiently fill those locations, therefore alleviating the aforementioned disadvantage.\\n\\\\noindent\\\\textbf{2) Dynamic additional refinement.} Instead of only sampling certain pixels to perform convolutions, another line of work first conducts relatively cheap computation on the whole feature map, and adaptively activate extra modules on selected pixels for further \\\\emph{refinement}. Representatively, dynamic capacity network generates coarse features with a shallow model, and salient pixels are sampled based on the gradient information. \\nFor these salient pixels, extra layers are applied to extract finer features. Similarly, specific positions are additionally processed by a fraction of convolutional filters in . These methods adapt their network architectures in terms of \\\\emph{depth} or \\\\emph{width} at the pixel level, achieving a spatially adaptive allocation of computation.\\n{The aforementioned dynamic additional refinement approaches are mainly developed for image classification.} On the semantic segmentation task, pixel-wise \\\\emph{early exiting} (see also Sec. \\\\ref{dynamic_depth}) is proposed in , where the pixels with high prediction confidence are output without being processed by deeper layers. PointRend shares a similar idea, and applies additional FC layers on selected pixels with low prediction confidence, which are more likely to be on borders of objects. All these researches demonstrate that by exploiting the spatial redundancy in image data, dynamic computation at the pixel level beyond sample level significantly increases the model efficiency.\\n\\\\vspace{-1ex}\",\n \"id\": \"fa853b1f-afd4-40d2-bed1-be3a1d6cb128\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"bb93a738-d7fa-4ffe-811c-d2e5d3062ddd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Pixel-level Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Pixel-wise Dynamic Architectures\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pixel-wise Dynamic Architectures\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 7257,\n 740,\n 741,\n 7268,\n 8376,\n 731,\n 722,\n 7282,\n 7276,\n 742,\n 7278,\n 7275\n ],\n \"content\": \"\\\\label{pixel_dynamic_param}\\n\\\\vspace{-0.25ex}\\nIn contrast to entirely skipping the convolution operation on a subset of pixels, dynamic networks can also apply data-dependent parameters on different pixels for improved representation power or adaptive reception fields.\\n\\\\noindent\\\\textbf{1) Dynamic weights.} {Similar to the sample-wise dynamic parameter methods (Sec. \\\\ref{adaptive_params}), pixel-level dynamic weights are achieved by test-time \\\\emph{adjustment} , \\\\emph{prediction} or \\\\emph{dynamic features} . Take weight prediction as an example,} typical approaches generate an $H\\\\!\\\\times\\\\!W\\\\!\\\\times\\\\! k^2$ kernel map to produce spatially dynamic weights ($H,W$ are the spatial size of the output feature and $k$ is the kernel size). Considering the pixels belonging to the same object may share identical weights, dynamic region-aware convolution (DRConv) generates a segmentation mask for an input image, dividing it into $m$ regions, for each of which a weight generation network is responsible for producing a data-dependent kernel.\\n\\\\noindent\\\\textbf{2) Dynamic reception fields.} Traditional convolution operations usually have a fixed shape and size of kernels (e.g. the commonly used $3\\\\!\\\\times\\\\!3$ 2D convolution). The resulting uniform reception field across all the layers may have limitations for recognizing objects with varying shapes and sizes. {To tackle this, a line of work learns to adapt the reception field for different feature pixels , as discussed in Sec. \\\\ref{dynamic_param_adjust}.} Instead of adapting the sampling location of features or kernels, adaptive connected network realizes a dynamic trade-off among self transformation (e.g. $1\\\\!\\\\times\\\\!1$ convolution), local inference (e.g. $3\\\\!\\\\times\\\\!3$ convolution) and global inference (e.g. FC layer). The three branches of outputs are fused with data-dependent weighted summation. \\nBesides images, the local and global information in non-Euclidean data, such as graphs, could also be adaptively aggregated.\\n\\\\vspace{-1.5ex}\",\n \"id\": \"530cae77-537a-4dca-80a6-e035f5680e22\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"bb93a738-d7fa-4ffe-811c-d2e5d3062ddd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Pixel-level Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Pixel-wise Dynamic Parameters\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pixel-wise Dynamic Parameters\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7281\n ],\n \"content\": \"\\\\label{region_level}\\n\\\\vspace{-0.5ex}\\n{Pixel-level dynamic networks mentioned in Sec. \\\\ref{pixel_level} often require specific implementations for sparse computation, and consequently may face challenges in terms of achieving real acceleration on hardware .} An alternative approach is performing adaptive inference on \\\\emph{regions/patches} of input images. There mainly exists two lines of work along this direction (see \\\\figurename~\\\\ref{fig_spatial_select_regions}): one performs parameterized \\\\emph{transformations} on a region of feature maps for more accurate prediction (Sec. \\\\ref{dynamic_transofrm}), and the other learns patch-level \\\\emph{hard attention}, with the goal of improving the effectiveness and/or efficiency of models (Sec. \\\\ref{hard_attention_pathces}).\\n\\\\vspace{-1ex}\",\n \"id\": \"0d52450e-77a8-4861-a2eb-00ddb1ece863\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"d1dfc651-19ab-46ab-9db8-02eef3371ca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Region-level Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"b67eaf7b-932e-433e-a7e5-732f3b779d92\",\n \"27095343-62ad-48ed-a110-4d5657e3a267\"\n ],\n \"title\": \"Region-level Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 743,\n 531\n ],\n \"content\": \"\\\\label{dynamic_transofrm}\\n\\\\vspace{-0.25ex}\\nDynamic transformations (e.g. affine/projective/thin plate spline transformation) can be performed on images to undo certain variations for better generalization ability, or to exaggerate the salient regions for discriminative feature representation. For example, spatial transformer adopts a localization network to generate the transformation parameters, and then applies the parameterized transformation to recover the input from the corresponding variations. \\nMoreover, transformations are learned to adaptively zoom-in the salient regions on some tasks where the model performance is sensitive to a small portion of regions. \\n\\\\begin{figure}\\n \\\\centering\\n \\\\vspace{-1ex}\\n \\\\includegraphics[width=\\\\linewidth]{9_spatial_select_regions.pdf}\\n \\\\vskip -0.15in\\n \\\\caption{{Region-level dynamic inference. The region selection module generates the transformation/localization parameters, and the subsequent network performs inference on the transformed/cropped region.}}\\n \\\\label{fig_spatial_select_regions}\\n \\\\vspace{-4ex}\\n\\\\end{figure}\\n\\\\vspace{-1ex}\",\n \"id\": \"b67eaf7b-932e-433e-a7e5-732f3b779d92\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"0d52450e-77a8-4861-a2eb-00ddb1ece863\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Region-level Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Transformations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Transformations\"\n },\n {\n \"cite_extract_rate\": 0.8571428571428571,\n \"cites\": [\n 747,\n 745,\n 737,\n 744,\n 674,\n 746\n ],\n \"content\": \"\\\\label{hard_attention_pathces}\\n\\\\vspace{-0.25ex}\\nInspired by the fact that informative features may only be contained in certain regions of an image, dynamic networks with hard spatial attention are explored to strategically select patches from the input for improved efficiency.\\n\\\\noindent\\\\textbf{1) Hard attention with RNNs.} The most typical approach is formulating a classification task as a sequential decision process, {and adopting RNNs to make iterative predictions based on selected patches .} For example, images are classified within a fixed number of steps, and at each step, the classifier RNN only sees a cropped patch, deciding the next attentional location until the last step is reached . An adaptive step number is further achieved by including early stopping in the action space . Glance-and-focus network (GFNet) builds a general framework of region-level adaptive inference by sequentially focusing on a series of selected patches, and is compatible with most existing CNN architectures. The recurrent attention mechanism together with the early exiting paradigm enables both \\\\emph{spatially} and \\\\emph{temporally} adaptive inference .\\n\\\\noindent\\\\textbf{2) Hard attention with other implementations.} Rather than using an RNN to predict the region position that the model should pay attention to, class activation mapping (CAM) is leveraged in to iteratively focus on salient patches. At each iteration, the selection is performed on the previously cropped input, leading to a progressive refinement procedure. A multi-scale CNN is built in , where the sub-network in each scale takes in the cropped patch from the previous scale, and is responsible for simultaneously producing 1) the feature representations for classification and 2) the attention map for the next scale. {Without an iterative manner, the recent differentiable patch selection adopts a differentiable top-K module to select a fixed number of patches in one step.}\\n\\\\vspace{-2ex}\",\n \"id\": \"27095343-62ad-48ed-a110-4d5657e3a267\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"0d52450e-77a8-4861-a2eb-00ddb1ece863\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Region-level Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Hard Attention on Selected Patches\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hard Attention on Selected Patches\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 504\n ],\n \"content\": \"\\\\label{dynamic_resolution}\\n\\\\vspace{-0.5ex}\\nThe researches discussed above typically divide feature maps into different areas (pixel-level or region-level) for adaptive inference.\\nOn a coarser granularity, some dynamic networks could treat each image as a whole by processing feature representations with adaptive resolutions. Although it has been observed that a low resolution might be sufficient for recognizing most \\\"easy\\\" samples , conventional CNNs mostly process all the inputs with the same resolution, inducing considerable redundancy. Therefore, resolution-level dynamic networks exploit spatial redundancy from the perspective of feature resolution rather than the saliency of different locations. Existing approaches mainly include 1) scaling the inputs with adaptive ratios (Sec. \\\\ref{adaptive_scaling_ratio}); 2) selectively activating the sub-networks with different resolutions in a multi-scale architecture (Sec. \\\\ref{dynamic_res_multiscale}).\\n\\\\vspace{-1ex}\",\n \"id\": \"ab0c50b9-9d18-4c18-976a-34aae739e7ac\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"d1dfc651-19ab-46ab-9db8-02eef3371ca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Resolution-level Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"6a7330af-c2c7-4497-af0d-0ab86ce108a0\",\n \"16545c48-ec80-4bbd-871e-b151b745999d\"\n ],\n \"title\": \"Resolution-level Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 748\n ],\n \"content\": \"\\\\label{adaptive_scaling_ratio}\\n\\\\vspace{-0.25ex}\\nDynamic resolution can be achieved by scaling features with adaptive ratios. For example, a small sub-network is first executed to predict a scale distribution of faces on the face detection task, then the input images are adaptively zoomed, so that all the faces fall in a suitable range for recognition . A plug-in module is used by to predict the stride for the first convolution block in each ResNet stage, producing features with dynamic resolution.\\n\\\\vspace{-1ex}\",\n \"id\": \"6a7330af-c2c7-4497-af0d-0ab86ce108a0\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"ab0c50b9-9d18-4c18-976a-34aae739e7ac\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Resolution-level Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Adaptive Scaling Ratios\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adaptive Scaling Ratios\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7266\n ],\n \"content\": \"\\\\label{dynamic_res_multiscale} \\n\\\\vspace{-0.25ex}\\nAn alternative approach to achieving dynamic resolution is building multiple sub-networks in a parallel or cascading way. These sub-networks with different feature resolutions are selectively activated conditioned on the input during inference. For instance, Elastic realizes a \\\\emph{soft} selection from multiple branches at every layer, where each branch performs a downsample-convolution-upsample procedure with an independent scaling ratio. To practically avoid redundant computation, a \\\\emph{hard} selection is realized by , which allows each sample to conditionally activate sub-networks that process feature representations with resolution from low to high (see \\\\figurename~\\\\ref{multi_scale} (c) in Sec. \\\\ref{dynamic_depth}).\\n\\\\vspace{-2ex}\\n\\\\begin{figure*}\\n \\\\centering\\n \\\\vspace{-2ex}\\n \\\\includegraphics[width=0.9\\\\linewidth]{10_temporal_skim.pdf}\\n \\\\vskip -0.15in\\n \\\\caption{{Temporally adaptive inference. The first three approaches dynamically allocate computation in each step by (a) skipping the update, (b) partially updating the state, or (c) conditional computation in a hierarchical structure. The agent in (d) decides where to read in the next step.}}\\n \\\\label{fig_temporal_skim}\\n \\\\vspace{-3ex}\\n\\\\end{figure*}\",\n \"id\": \"16545c48-ec80-4bbd-871e-b151b745999d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"ab0c50b9-9d18-4c18-976a-34aae739e7ac\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Spatial-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Resolution-level Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Resolution in Multi-scale Architectures\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Resolution in Multi-scale Architectures\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec_temporal_adaptive}\\n\\\\vspace{-0.25ex}\\nApart from the spatial dimension (Sec. \\\\ref{sec_spatially_adaptive}), adaptive computation could also be performed along the temporal dimension of sequential data, such as texts (Sec. \\\\ref{temporal_text}) and videos (Sec. \\\\ref{sec:tempoal_video}). {In general, network efficiency can be improved by dynamically allocating less/no computation to the inputs at unimportant temporal locations.}\\n\\\\vspace{-1.5ex}\",\n \"id\": \"8a8770e6-47f0-428a-8e00-4bb401ff006f\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"d0476612-fbf1-40f0-a9de-cc704c8eb7d7\",\n \"844590b4-9a40-42f9-8052-47ad9e4a7783\"\n ],\n \"title\": \"Temporal-wise Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\vspace{-0.25ex}\\n\\\\label{temporal_text}\\nTraditional RNNs mostly follow a static inference paradigm, i.e. input tokens are read sequentially to update a hidden state at each time step, which could be written as\\n\\\\begin{equation}\\\\label{eq_rnn}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n\\\\mathbf{h}_t = \\\\mathcal{F}(\\\\mathbf{x}_t, \\\\mathbf{h}_{t-1}), t=1,2,\\\\cdots, T.\\n\\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nSuch a static inference paradigm induces significant redundant computation, as different tokens usually have different contributions to the downstream tasks.\\nA type of dynamic RNN is developed for allocating appropriate computational cost at each step. Some learn to \\\\emph{\\\"skim\\\"} unimportant tokens by dynamic update of hidden states (Sec. \\\\ref{skimming_text}), and others conduct \\\\emph{adaptive reading} to avoid processing task-irrelevant tokens. Specifically, such adaptive reading can be achieved by \\\\emph{early exiting} (Sec. \\\\ref{temporal_early_exit}) or \\\\emph{dynamic jumping} (Sec. \\\\ref{text_jumping}).\\n\\\\vspace{-1ex}\",\n \"id\": \"d0476612-fbf1-40f0-a9de-cc704c8eb7d7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"8a8770e6-47f0-428a-8e00-4bb401ff006f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"RNN-based Dynamic Text Processing\"\n ]\n ],\n \"subsections\": [\n \"7450af29-777b-40c4-bbdb-5820eb26595d\",\n \"04644c19-76d1-481e-b8a1-a56033239887\",\n \"64d0b8b6-25b1-481c-a175-60d486a5210d\"\n ],\n \"title\": \"RNN-based Dynamic Text Processing\"\n },\n {\n \"cite_extract_rate\": 0.625,\n \"cites\": [\n 683,\n 7284,\n 750,\n 7283,\n 749\n ],\n \"content\": \"\\\\vspace{-0.25ex}\\n\\\\label{skimming_text}\\nSince not all the tokens are essential for capturing the task-relevant information in a sequence, dynamic RNNs can be built to adaptively update their hidden states at each time step. Less informative tokens will be coarsely \\\\emph{skimmed}, i.e. the states are updated with cheaper computation.\\n\\\\noindent\\\\textbf{1) Skipping the update.} For unimportant inputs at certain temporal locations, dynamic models can learn to entirely skip the update of hidden states (see \\\\figurename~\\\\ref{fig_temporal_skim} (a)), i.e.\\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{h}_t = \\\\alpha_t\\\\mathcal{F}(\\\\mathbf{x}_t, \\\\mathbf{h}_{t-1}) + (1-\\\\alpha_t)\\\\mathbf{h}_{t-1}, \\\\alpha_t\\\\in\\\\left\\\\{0,1\\\\right\\\\}.\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nFor instance, Skip-RNN updates a controlling signal in every step to determine whether to update or \\\\emph{copy} the hidden state from the previous step. An extra agent is adopted by Structural-Jump-LSTM to make the skipping decision conditioned on the previous state and the current input. Without training the RNNs and the controllers jointly as in and , a predictor is trained in to estimate whether each input will make a \\\"significant change\\\" on the hidden state. The update is identified worthy to be executed only when the predicted change is greater than a threshold.\\n\\\\noindent\\\\textbf{2) Coarse update. } As directly skipping the update may be too aggressive, dynamic models could also update the hidden states with adaptively allocated operations. In specific, a network can adapt its architecture in every step, i.e. \\n\\\\begin{equation}\\n \\\\setlength{\\\\abovedisplayskip}{3pt}\\n \\\\mathbf{h}_t = \\\\mathcal{F}_t(\\\\mathbf{x}_t, \\\\mathbf{h}_{t-1}), t=1,2,\\\\cdots, T,\\n \\\\setlength{\\\\belowdisplayskip}{3pt}\\n\\\\end{equation}\\nwhere $\\\\mathcal{F}_t$ is determined based on the input $\\\\mathbf{x}_t$. One implementation is selecting a subset of dimensions of the hidden state to calculate, and copying the remaining from the previous step , as shown in \\\\figurename~\\\\ref{fig_temporal_skim} (b). To achieve the partial update, a subset of rows in weight matrices of the RNN is dynamically activated in , while Skim-RNN makes a choice between two independent RNNs.\\nWhen the hidden states are generated by a multi-layer network, the update could be interrupted at an intermediate layer based on an accumulated halting score .\\nTo summarize, a coarse update can be realized by data-dependent network \\\\emph{depth} or \\\\emph{width} . \\n{\\\\noindent\\\\textbf{3) Selective updates in hierarchical RNNs.} Considering the intrinsic hierarchical structure of texts (e.g. sentence-word-character), researchers have developed hierarchical RNNs to encode the temporal dependencies with different timescales using a dynamic update mechanism . During inference, the RNNs at higher levels will selectively update their states conditioned on the output of low-level ones (see \\\\figurename~\\\\ref{fig_temporal_skim} (c)).\\nFor example, when a character-level model in detects that the input satisfies certain conditions, it will \\\\emph{\\\"flush\\\"} (reset) its states and feed them to a word-level network. Similar operations have also been realized by a gating module on question answering tasks .}\\n\\\\vspace{-1ex}\",\n \"id\": \"7450af29-777b-40c4-bbdb-5820eb26595d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"d0476612-fbf1-40f0-a9de-cc704c8eb7d7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"RNN-based Dynamic Text Processing\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Update of Hidden States\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Update of Hidden States\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\vspace{-0.25ex}\\n\\\\label{temporal_early_exit}\\nDespite that the dynamic RNNs in Sec. \\\\ref{skimming_text} are able to update their states with data-dependent computational costs at each step, all the tokens still must be read, leading to inefficiency in scenarios where the task-relevant results can be obtained before reading the entire sequence.\\nIdeally, an efficient model should adaptively stop reading before the last step $T$ in Eq. \\\\ref{eq_rnn} is reached, once the captured information is satisfactory to solve the task. For instance, reasoning network (ReasoNet) terminates its reading procedure when sufficient evidence has been found for question answering. \\nSimilarly, early stopping is implemented for sentence-level and paragraph-level text classification, respectively. Note that the approaches discussed here focus on making early predictions with respect to the \\\\emph{temporal} dimension of sequential input, rather than along the \\\\emph{depth} dimension of networks as in Sec. \\\\ref{dynamic_depth}.\\n\\\\vspace{-1ex}\",\n \"id\": \"04644c19-76d1-481e-b8a1-a56033239887\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"d0476612-fbf1-40f0-a9de-cc704c8eb7d7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"RNN-based Dynamic Text Processing\"\n ],\n [\n \"subsubsection\",\n \"Temporally Early Exiting in RNNs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Temporally Early Exiting in RNNs\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 7283,\n 7285\n ],\n \"content\": \"\\\\vspace{-0.25ex}\\n\\\\label{text_jumping}\\nAlthough early exiting in Sec. \\\\ref{temporal_early_exit} can largely reduce redundant computation, all the tokens must still be fed to the model one by one. More aggressively, dynamic RNNs could further learn to decide \\\\emph{\\\"where to read\\\"} by strategically skipping some tokens without reading them, and directly jumping to an arbitrary temporal location (see \\\\figurename~\\\\ref{fig_temporal_skim} (d)).\\nSuch dynamic jumping, together with early exiting, is realized in and . Specifically, LSTM-Jump implements an auxiliary unit to predict the jumping stride within a defined range, and the reading process ends when the unit outputs zero. The model in first decides whether to stop at each step. If not, it will further choose to re-read the current input, or to skip a flexible number of words. Moreover, structural information is exploited by Structural-Jump-LSTM , which utilizes an agent to decide whether to jump to the next punctuation. Apart from looking ahead, LSTM-Shuttle also allows backward jumping to supplement the missed history information.\\n\\\\vspace{-1.5ex}\",\n \"id\": \"64d0b8b6-25b1-481c-a175-60d486a5210d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"d0476612-fbf1-40f0-a9de-cc704c8eb7d7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"RNN-based Dynamic Text Processing\"\n ],\n [\n \"subsubsection\",\n \"Jumping in Texts\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Jumping in Texts\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:tempoal_video}\\n\\\\vspace{-0.25ex}\\nFor video recognition, where a video could be seen as a sequential input of frames, temporal-wise dynamic networks are designed to allocate adaptive computational resources for different frames. This can generally be achieved by two approaches: 1) dynamically updating the hidden states in each time step of \\\\emph{recurrent} models (Sec. \\\\ref{recurrent_video}), and 2) performing adaptive \\\\emph{pre-sampling} for key frames (Sec. \\\\ref{frame_sampling}).\\n\\\\vspace{-1ex}\",\n \"id\": \"844590b4-9a40-42f9-8052-47ad9e4a7783\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"8a8770e6-47f0-428a-8e00-4bb401ff006f\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Temporal-wise Dynamic Video Recognition\"\n ]\n ],\n \"subsections\": [\n \"b2f1acb3-4b7a-41ce-96e3-e8abe77bc9b2\",\n \"ce2a3a82-4498-49fd-b6ea-af13eecfa56b\"\n ],\n \"title\": \"Temporal-wise Dynamic Video Recognition\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 7286,\n 755,\n 7270,\n 752,\n 751,\n 754,\n 7283,\n 8380,\n 753\n ],\n \"content\": \"\\\\label{recurrent_video}\\n\\\\vspace{-0.25ex}\\nVideo recognition is often conducted via a recurrent procedure, where the video frames are first encoded by a 2D CNN, and the obtained frame features are fed to an RNN sequentially for updating its hidden state. Similar to the approaches introduced in Sec. \\\\ref{temporal_text}, RNN-based adaptive video recognition is typically realized by 1) treating unimportant frames with relatively cheap computation (\\\\emph{\\\"glimpse\\\"}) ; 2) \\\\emph{early exiting} ; and 3) performing dynamic \\\\emph{jumping} to decide {\\\"where to see\\\"} .\\n\\\\noindent\\\\textbf{1) Dynamic update of hidden states.} To reduce redundant computation at each time step, LiteEval makes a choice between two LSTMs with different computational costs. ActionSpotter decides whether to update the hidden state according to each input frame. {AdaFuse selectively reuses certain feature channels from the previous step to efficiently make use of historical information.} Recent work has also proposed to adaptively decide the numerical precision or modalities when processing the sequential input frames. Such a \\\\emph{glimpse} procedure (i.e. allocating cheap operations to unimportant frames) is similar to the aforementioned text \\\\emph{skimming} .\\n\\\\noindent\\\\textbf{2) Temporally early exiting.} Humans are able to comprehend the contents easily before watching an entire video. Such early stopping is also implemented in dynamic networks to make predictions only based on a portion of video frames . Together with the \\\\emph{temporal} dimension, the model in further achieves early exiting from the aspect of network \\\\emph{depth} as discussed in Sec. \\\\ref{dynamic_depth}.\\n\\\\noindent\\\\textbf{3) Jumping in videos.} Considering encoding those unimportant frames with a CNN still requires considerable computation, a more efficient solution could be dynamically skipping some frames without watching them. Existing arts typically learn to predict the location that the network should jump to at each time step.\\nFurthermore, both early stopping and dynamic jumping are allowed in , where the jumping stride is limited in a discrete range.\\nAdaptive frame (AdaFrame) generates a continuous scalar within the range of $[0,1]$ as the relative location. \\n\\\\vspace{-1ex}\",\n \"id\": \"b2f1acb3-4b7a-41ce-96e3-e8abe77bc9b2\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"844590b4-9a40-42f9-8052-47ad9e4a7783\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Temporal-wise Dynamic Video Recognition\"\n ],\n [\n \"subsubsection\",\n \"Video Recognition with Dynamic RNNs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Video Recognition with Dynamic RNNs\"\n },\n {\n \"cite_extract_rate\": 0.642857142857142,\n \"cites\": [\n 8381,\n 756,\n 6975,\n 8382,\n 757,\n 7288,\n 7287,\n 686,\n 758\n ],\n \"content\": \"\\\\label{frame_sampling}\\n\\\\vspace{-0.25ex}\\nRather than processing video frames recurrently as in Sec. \\\\ref{recurrent_video}, another line of work first performs an adaptive \\\\emph{pre-sampling} procedure, and then makes prediction by processing the selected subset of key frames or clips.\\n\\\\noindent{\\\\textbf{1) Temporal attention} is a common technique for networks to focus on salient frames.} For face recognition, neural aggregation network uses \\\\emph{soft} attention to adaptively aggregate frame features. To improve the inference efficiency, \\\\emph{hard} attention is realized to remove unimportant frames iteratively with RL for efficient video face verification .\\n\\\\noindent{\\\\textbf{2) Sampling module} is also a prevalent option for dynamically selecting the key frames/clips in a video.} For example, the frames are first sampled uniformly in , and discrete decisions are made for each selected frame to go forward or backward step by step. As for clip-level sampling, SCSample is designed based on a trained classifier to find the most informative clips for prediction. Moreover, dynamic sampling network (DSN) segments each video into multiple sections, and a sampling module with shared weights across the sections is exploited to sample one clip from each section.\\n{Adjusting multiple factors of deep models simultaneously has attracted researches in both static and dynamic networks . For example, together with \\\\emph{temporal-wise} frame sampling, \\\\emph{spatially} adaptive computation can be achieved by spatial /temporal resolution adaptation and patch selection . It would be promising to exploit the redundancy in both \\\\emph{input data} and \\\\emph{network structure} for further improving the efficiency of deep networks.}\\n\\\\vspace{-2ex}\",\n \"id\": \"ce2a3a82-4498-49fd-b6ea-af13eecfa56b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"844590b4-9a40-42f9-8052-47ad9e4a7783\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Temporal-wise Dynamic Networks\"\n ],\n [\n \"subsection\",\n \"Temporal-wise Dynamic Video Recognition\"\n ],\n [\n \"subsubsection\",\n \"Dynamic Key Frame Sampling\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Dynamic Key Frame Sampling\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 675,\n 7277\n ],\n \"content\": \"\\\\label{inference_and_train}\\n\\\\vspace{-0.25ex}\\nIn previous sections, we have reviewed three different types of dynamic networks (sample-wise (Sec. \\\\ref{sec_sample_wise}), spatial-wise (Sec. \\\\ref{sec_spatially_adaptive}) and temporal-wise (Sec. \\\\ref{sec_temporal_adaptive})). It can be observed that making data-dependent \\\\emph{decisions} at the inference stage is essential to achieve high efficiency and effectiveness. Moreover, \\\\emph{training} dynamic models is usually more challenging than optimizing static networks.\\nNote that since parameter adaptation (Sec. \\\\ref{adaptive_params}) could be conveniently achieved by differentiable operations, models with dynamic parameters can be directly trained by stochastic gradient descent (SGD) without specific techniques. Therefore, in this section we mainly focus on discrete decision making (Sec. \\\\ref{inference}) and its training strategies (Sec. \\\\ref{training}), which are absent in most static models.\\n\\\\vspace{-1.5ex}\",\n \"id\": \"4ffefefb-4651-456b-a07d-75e465a728b0\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ]\n ],\n \"subsections\": [\n \"8d1ee291-ab1d-4e7f-8c16-431c07d68081\",\n \"2bd26a27-bb2c-4752-8590-7d1e502a061d\"\n ],\n \"title\": \"Inference and Training\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{inference}\\n\\\\vspace{-0.5ex}\\nAs described above, dynamic networks are capable of making data-dependent decisions during inference to transform their architectures, parameters, or to select salient spatial/temporal locations in the input. Here we summarize three commonly seen decision making schemes as follows.\\n\\\\vspace{-1ex}\",\n \"id\": \"8d1ee291-ab1d-4e7f-8c16-431c07d68081\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4ffefefb-4651-456b-a07d-75e465a728b0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ],\n [\n \"subsection\",\n \"Decision Making of Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"16951c8b-a4dd-48e3-a004-db0aa23dc8c4\",\n \"2a7772db-4dcb-4d14-b74b-bd6ac1bb74cf\",\n \"6e023be0-dfce-4cb5-b69e-f84dd0c01501\"\n ],\n \"title\": \"Decision Making of Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0.636363636363636,\n \"cites\": [\n 7266,\n 689,\n 683,\n 759,\n 7269,\n 676,\n 7273\n ],\n \"content\": \"\\\\vspace{-0.25ex}\\nMany dynamic networks are able to output \\\"easy\\\" samples at early exits if a certain confidence-based criterion is satisfied. These methods generally require estimating the confidence of intermediate predictions, which is compared to a predefined threshold for decision making.\\nIn classification tasks, the confidence is usually represented by the maximum element of the \\\\emph{SoftMax} output . Alternative criteria include the entropy and the score margin . On NLP tasks, a \\\\emph{model patience} is proposed in : when the predictions for one sample stay unchanged after a number of classifiers, the inference procedure stops. \\nIn addition, the halting score in could also be viewed as confidence for whether the current feature could be output to the next time step or calculation stage.\\nEmpirically, the confidence-based criteria are easy to implement, and generally require no specific training techniques. A trade-off between accuracy and efficiency is controlled by manipulating the thresholds, which are usually tuned on a validation dataset. Note that the \\\\emph{overconfidence} issue in deep models might affect the effectiveness of such decision paradigm, when the incorrectly classified samples could obtain a high confidence at early exits.\\n\\\\vspace{-1ex}\",\n \"id\": \"16951c8b-a4dd-48e3-a004-db0aa23dc8c4\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"8d1ee291-ab1d-4e7f-8c16-431c07d68081\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ],\n [\n \"subsection\",\n \"Decision Making of Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Confidence-based Criteria\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Confidence-based Criteria\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 702,\n 718,\n 691\n ],\n \"content\": \"\\\\label{inference_policy}\\n\\\\vspace{-0.25ex}\\nIt is a common option to build an additional policy network learning to adapt the network topology based on different samples. \\nSpecifically, each input sample is first processed by the policy network, whose output directly determines which parts of the main network should be activated. For example, BlockDrop and GaterNet use a policy network to adaptively decide the \\\\emph{depth} and \\\\emph{width} of a backbone network. More generally, dynamic routing in a \\\\emph{SuperNet} can also be controlled by a policy network .\\nOne possible limitation of this scheme is that the architectures and the training process of some policy networks are developed for a specific backbone , and may not be easily adapted to different architectures. \\n\\\\vspace{-1ex}\",\n \"id\": \"2a7772db-4dcb-4d14-b74b-bd6ac1bb74cf\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"8d1ee291-ab1d-4e7f-8c16-431c07d68081\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ],\n [\n \"subsection\",\n \"Decision Making of Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Policy Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Policy Networks\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 7274,\n 708,\n 751,\n 8374,\n 710,\n 8373,\n 696,\n 7281,\n 713,\n 699\n ],\n \"content\": \"\\\\label{inference_gating}\\n\\\\vspace{-0.25ex}\\nGating function is a general and flexible approach to decision making in dynamic networks. It can be conveniently adopted as a plug-in module at arbitrary locations in any backbone network.\\nDuring inference, each module is responsible for controlling the local inference graph of a layer or block. The gating functions take in intermediate features and efficiently produce binary-valued gate vectors to decide: 1) which channels need to be activated \\\\emph{width}, 2) which layers need to be skipped , 3) which paths should be selected in a SuperNet , or 4) what locations of the input should be allocated computations .\\nCompared to the aforementioned decision policies, the gating functions demonstrate notable generality and applicability. However, due to their lack of differentiability, these gating functions usually need specific training techniques, which will be introduced in the following Sec. \\\\ref{training}.\\n\\\\vspace{-2ex}\",\n \"id\": \"6e023be0-dfce-4cb5-b69e-f84dd0c01501\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"8d1ee291-ab1d-4e7f-8c16-431c07d68081\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ],\n [\n \"subsection\",\n \"Decision Making of Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Gating Functions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Gating Functions\"\n },\n {\n \"cite_extract_rate\": 0.581196581196581,\n \"cites\": [\n 779,\n 725,\n 722,\n 7273,\n 683,\n 770,\n 771,\n 776,\n 7257,\n 778,\n 741,\n 7294,\n 7283,\n 7267,\n 674,\n 8381,\n 7290,\n 772,\n 746,\n 7285,\n 150,\n 8382,\n 760,\n 7281,\n 7289,\n 7272,\n 761,\n 747,\n 752,\n 757,\n 7288,\n 8383,\n 764,\n 750,\n 727,\n 743,\n 765,\n 748,\n 762,\n 8380,\n 777,\n 775,\n 769,\n 686,\n 680,\n 767,\n 7295,\n 7291,\n 763,\n 773,\n 744,\n 774,\n 7286,\n 7284,\n 7282,\n 7278,\n 749,\n 7215,\n 7292,\n 726,\n 766,\n 768,\n 676,\n 7293,\n 738,\n 7274,\n 780,\n 7287\n ],\n \"content\": \"\\\\vspace{-0.5ex}\\n\\\\label{training}\\nBesides architecture design, training is also essential for dynamic networks. Here we summarize the existing training strategies for dynamic models from the perspectives of objectives and optimization.\\n\\\\begin{table*}\\n \\\\scriptsize\\n \\\\vspace{-2ex}\\n \\\\caption{Applications of Dynamic Networks. For the type column, Sa, Sp and Te stand for sample-wise, spatial-wise and temporal-wise respectively.}\\n \\\\label{tab_tasks}\\n \\\\begin{center}\\n \\\\vspace{-6ex}\\n \\\\resizebox{\\\\linewidth}{!}{\\n \\\\begin{tabular}{cccc}\\n \\\\toprule\\n \\\\textbf{Fields} & \\\\textbf{Data} & \\\\textbf{Type} & \\\\textbf{Subfields \\\\& references} \\\\\\\\\\n \\\\midrule\\n & \\\\multirow{6}*{\\\\textbf{Image}} & \\\\multirow{2}*{Sa} & Object detection (face , facial point , pedestrian , general ) \\\\\\\\\\n & & & Image segmentation , Super resolution , Style transfer , Coarse-to-fine classification \\\\\\\\ \\n \\\\cmidrule{3-4}\\n & & \\\\multirow{3}*{Sa \\\\& Sp} & Image segmentation , Image-to-image \\\\\\\\\\n & & & translation , Object detection , Semantic image synthesis , \\\\\\\\\\n \\\\multirow{1}*{\\\\textbf{Computer}} & & & Image denoising , Fine-grained classification Eye tracking , Super resolution \\\\\\\\\\n \\\\cmidrule{3-4}\\n \\\\multirow{1}*{\\\\textbf{Vision}}& & Sa \\\\& Sp \\\\& Te & General classification , Multi-object classification , Fine-grained classification \\\\\\\\\\n \\\\cmidrule{2-4}\\n & \\\\multirow{4}*{\\\\textbf{Video}} & Sa & Multi-task learning (human action recognition and frame prediction) \\\\\\\\\\n \\\\cmidrule{3-4}\\n & & \\\\multirow{2}*{Sa \\\\& Te} & Classification (action recognition) , Semantic segmentation \\\\\\\\\\n & & & Video face recognition , Action detection , Action spotting \\\\\\\\\\n \\\\cmidrule{3-4}\\n & & Sa \\\\& Sp \\\\& Te & Classification , Frame interpolation , Super resolution , Video deblurring , Action prediction \\\\\\\\\\n \\\\cmidrule{2-4}\\n & \\\\textbf{Point Cloud} & Sa \\\\& Sp & 3D Shape classification and segmentation, 3D scene segmentation , 3D semantic scene completion \\\\\\\\\\n \\\\midrule\\n \\\\multirow{2}*{\\\\textbf{Natural}} & \\\\multirow{3}*{\\\\textbf{Text}} & Sa & Neural language inference, Text classification, Paraphrase similarity matching, and Sentiment analysis \\\\\\\\\\n \\\\cmidrule{3-4}\\n \\\\textbf{Language} & & \\\\multirow{2}*{Sa \\\\& Te} & Language modeling , Machine translation , Classification , \\\\\\\\\\n \\\\textbf{Processing} & & & Sentiment analysis , Question answering \\\\\\\\\\n \\\\midrule\\n \\\\textbf{Cross-Field} & \\\\multicolumn{3}{c}{Image captioning , {Video captioning , Visual question answering , Multi-modal sentiment analysis }} \\\\\\\\\\n \\\\midrule\\n \\\\multirow{2}{*}{\\\\textbf{Others}} & \\\\multicolumn{3}{c}{{Time series forecasting} , Link prediction , {Recommendation system} }\\\\\\\\\\n & \\\\multicolumn{3}{c}{Graph classification , {Document classification} , Stereo confidence estimation }\\\\\\\\\\n \\\\bottomrule\\n \\\\end{tabular}}\\n \\\\end{center}\\n \\\\vskip -0.3in\\n\\\\end{table*}\\n\\\\vspace{-1ex}\",\n \"id\": \"2bd26a27-bb2c-4752-8590-7d1e502a061d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 117,\n \"parent_id\": \"4ffefefb-4651-456b-a07d-75e465a728b0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ],\n [\n \"subsection\",\n \"Training of Dynamic Networks\"\n ]\n ],\n \"subsections\": [\n \"4d7d59fa-e730-411f-b178-1c7c2fc0254e\",\n \"34c67f80-85f1-4cea-8d35-f3e767fa3802\"\n ],\n \"title\": \"Training of Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0.769230769230769,\n \"cites\": [\n 7266,\n 701,\n 693,\n 7274,\n 708,\n 8374,\n 8373,\n 696,\n 7281,\n 753\n ],\n \"content\": \"\\\\label{sec_train_objectives}\\n\\\\vspace{-0.25ex}\\n\\\\noindent\\\\textbf{1) Training multi-exit networks.} We first notice that early-exiting dynamic networks are generally trained by minimizing a weighted cumulative loss of intermediate classifiers. One challenge for training such models is the joint optimization of multiple classifiers, which may interfere with each other. MSDNet alleviates the problem through its special architecture design. Several improved training techniques are proposed for multi-exit networks, including a gradient equilibrium algorithm to stable the training process, and a bi-directional knowledge transfer approach to boost the collaboration of classifiers. {For temporal-wise early exiting, the training of the policy network in FrameExit is supervised by pseudo labels.}\\n\\\\noindent\\\\textbf{2) Encouraging sparsity.} Many dynamic networks adapt their inference procedure by conditionally activating their computational units or strategically sampling locations from the input .\\nTraining these models without additional constraints would result in superfluous computational redundancy, as a network could tend to activate all the candidate units for minimizing the task-specific loss. \\nThe overall objective function for restraining such redundancy are typically written as $\\\\mathfrak{L}\\\\! =\\\\!\\\\mathfrak{L}_\\\\mathrm{task}\\\\!+\\\\!\\\\gamma \\\\mathfrak{L}_\\\\mathrm{sparse}$, where $\\\\gamma$ is the hyper-parameter balancing the two items for the trade-off between accuracy and efficiency. In real-world applications, the second item can be designed based on the gate/mask values of candidate units (e.g. channels , layers or spatial locations ). Specifically, one may set a target activation rate or minimizing the $\\\\mathcal{L}_1$ norm of the gates/masks . It is also practical to directly optimize a resource-aware loss (e.g. FLOPs) , which can be estimated according to the input and output feature dimension for every candidate unit.\\n\\\\noindent{\\\\textbf{3) Others.} Note that extra loss items are mostly designed for but not limited to improving efficiency. Take as an example, the model progressively focuses on a selected region, and is trained with an additional \\\\emph{inter-scale pairwise ranking loss} for proposing more discriminative regions. Moreover, knowledge distilling is utilized to boost the co-training of multiple sub-networks in and .}\\n\\\\vspace{-1ex}\",\n \"id\": \"4d7d59fa-e730-411f-b178-1c7c2fc0254e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"2bd26a27-bb2c-4752-8590-7d1e502a061d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ],\n [\n \"subsection\",\n \"Training of Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Training Objectives for Efficient Inference\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Training Objectives for Efficient Inference\"\n },\n {\n \"cite_extract_rate\": 0.8235294117647051,\n \"cites\": [\n 8381,\n 708,\n 702,\n 7279,\n 696,\n 781,\n 691,\n 782,\n 750,\n 8373,\n 8375,\n 674,\n 7281,\n 699\n ],\n \"content\": \"\\\\vspace{-0.25ex}\\nA variety of dynamic networks contain non-differentiable functions that make discrete decisions to modify their architectures or sampling spatial/temporal locations from the input. These functions can not be trained directly with back-propagation. Therefore, specific techniques are studied to enable the end-to-end training as follows.\\n\\\\noindent\\\\textbf{1) Gradient estimation} is proposed to approximate the gradients for those non-differentiable functions and enable back-propagation. In , straight-through estimator (STE) is exploited to heuristically copy the gradient with respect to the stochastic output directly as an estimator of the gradient with respect to the \\\\emph{Sigmoid} argument.\\n\\\\noindent\\\\textbf{2) Reparameterization} is also a popular technique to optimize the discrete decision functions. For instance, the gating functions controlling the network width or depth can both be trained with \\\\emph{Gumbel SoftMax} , which is also used for pixel-level dynamic convolution . An alternative technique is \\\\emph{Improved SemHash} adopted in and to train their hard gating modules.\\n{Note that although these reparameterization techniques enable joint optimizing dynamic models together with gating modules in an end-to-end fashion, they usually lead to a longer training process for the decision functions to converge into a stable situation . Moreover, the model performance might be sensitive to some extra hyper-parameters (e.g. temperature in \\\\emph{Gumbel SoftMax}), which might also increase the training cost for these dynamic networks.}\\n\\\\noindent\\\\textbf{3) Reinforcement learning (RL)} is widely exploited for training non-differentiable decision functions. In specific, the backbones are trained by standard SGD, while the agents (either policy networks in Sec. \\\\ref{inference_policy} or gating functions in Sec. \\\\ref{inference_gating}) are trained with RL to take discrete actions for dynamic inference graphs or spatial/temporal sampling strategies .\\n{One challenge for RL-based training is the design of reward functions, which is important to the accuracy-efficiency tradeoff of dynamic models. Commonly seen reward signals are usually constructed to minimize a penalty item of the computational cost . Moreover, the training could be costly due to a multi-stage procedure: a pre-training process may be required for the backbone networks before the optimization of decision or sampling modules, and joint finetuning may be indispensable finally.}\\n\\\\vspace{-2ex}\",\n \"id\": \"34c67f80-85f1-4cea-8d35-f3e767fa3802\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"2bd26a27-bb2c-4752-8590-7d1e502a061d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Inference and Training\"\n ],\n [\n \"subsection\",\n \"Training of Dynamic Networks\"\n ],\n [\n \"subsubsection\",\n \"Optimization of Non-differentiable Functions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Optimization of Non-differentiable Functions\"\n },\n {\n \"cite_extract_rate\": 0.8148148148148141,\n \"cites\": [\n 764,\n 744,\n 785,\n 7273,\n 765,\n 770,\n 7269,\n 777,\n 758,\n 38,\n 786,\n 692,\n 768,\n 783,\n 7296,\n 7295,\n 7272,\n 7271,\n 7,\n 757,\n 674,\n 784\n ],\n \"content\": \"\\\\label{sec_tasks}\\n\\\\vspace{-0.25ex}\\nIn this section, we summarize the applications of dynamic networks. Representative methods are listed in Table \\\\ref{tab_tasks} based on the input data modality.\\nFor image recognition, most dynamic CNNs are designed to conduct \\\\emph{sample-wise} or \\\\emph{spatial-wise} adaptive inference on classification tasks, and many inference paradigms can be generalized to other applications. Note that as mentioned in Sec. \\\\ref{region_level}, the object recognition could be formulated as a sequential decision problem . By allowing early exiting in these approaches, \\\\emph{temporally} adaptive inference procedure could also be enabled.\\nFor text data, reducing its intrinsic temporal redundancy has attracted great research interests, and the inference paradigm of \\\\emph{temporal-wise} dynamic RNNs (see Sec. \\\\ref{temporal_text}) is also general enough to process audios . Based on large language models such as Transformer and BERT , adaptive depths are extensively studied to reduce redundant computation in network architectures.\\nFor video-related tasks, the three types of dynamic inference can be implemented simultaneously . However, for the networks that do not process videos recurrently, e.g. 3D CNNs , most of them still follow a static inference scheme. Few researches have been committed to building dynamic 3D CNNs , which might be an interesting future research direction.\\n{Moreover, dynamic networks (especially the attention mechanism) have also been applied to dynamically fuse the features from different modalities in some multi-modal learning tasks, e.g. RGB-D image segmentation and image/video captioning .}\\nFinally, dynamic networks have also been exploited to tackle some fundamental problems in deep learning. For example, multi-exit models can be used to: 1) alleviate the \\\\emph{over-thinking} issue while reducing the overall computation ; 2) perform \\\\emph{long-tailed classification} by inducing early exiting in the training stage; and 3) improve the model \\\\emph{robustness} . For another example, the idea of dynamic routing is implemented for: 1) {reducing the \\\\emph{training cost} under a multi-task setting } and 2) finding the optimal fine-tuning strategy for per example in \\\\emph{transfer learning} .\\n\\\\vspace{-2ex}\",\n \"id\": \"3cd25c7f-359c-4b70-8c02-f2738746b35c\",\n \"level\": \"section\",\n \"origin_cites_number\": 27,\n \"parent_id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"Application of Dynamic Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Application of Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"}\\n\\\\label{sec:discussion}\\n\\\\vspace{-0.5ex}\\n{Though recent years have witnessed significant progress in the research of dynamic neural networks, there still exist many open problems that are worth exploring.} In this section, we summarize a few challenges together with possible future directions in this field.\\n\\\\vspace{-1ex}\",\n \"id\": \"e9013abd-3d22-4eab-a882-4ba5e86719f2\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4fd906cb-4aa6-49a2-a5b7-ec10dfbd9280\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Challenges and Future Directions\"\n ]\n ],\n \"subsections\": [\n \"3fc7f7a7-7403-4a95-918e-d6e40acaa928\",\n \"9458c790-c005-432b-a25a-958ab55678a8\",\n \"baaa0a8b-88d7-44dc-8c2b-1c3a26c2c457\",\n \"23446088-fa89-4b21-8a64-51ee553e33cb\",\n \"a8fb9eb8-777e-40f9-a6cc-da50d402237b\",\n \"80ebe8d1-1699-4384-9941-1c27455ceb08\"\n ],\n \"title\": \"{Challenges and Future Directions\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7266,\n 706\n ],\n \"content\": \"\\\\vspace{-0.5ex}\\nDespite the success of dynamic neural networks, relatively few researches has been committed to analyze them from the theoretical perspective. In fact, theories for a deep understanding of current dynamic learning models and further improving them in principled ways are highly valuable. {Notably, it has been proven that a dynamic network with an adaptive width can preserve the representation power of an unsparsified model .} However, there are more theoretical problems that are fundamental for dynamic networks. Here we list several of them as follows.\\n\\\\noindent\\\\textbf{1) Optimal decision in dynamic networks.}\\nAn essential operation in most dynamic networks (especially those designed for improving computational efficiency) is making data-dependent decisions, e.g., determining whether a module should be evaluated or skipped. Existing solutions either use confidence-based criteria, or introduce policy networks and gating functions. Although being effective in practice (as mentioned in Sec. \\\\ref{inference_and_train}), they may not be optimal and lack theoretical justifications. Take early exiting as an example, the current heuristic methods might face the issues of overconfidence, high sensitivity for threshold setting and poor transferability. As for policy networks or gating modules, runtime decisions can be made based on a learned function.\\nHowever, they often introduce extra computations, and usually require a long and unstable training procedure. Therefore, principled approaches with theoretical guarantees for decision function design in dynamic networks is a valuable research topic.\\n\\\\noindent\\\\textbf{2) Generalization issues.} \\nIn a dynamic model, a sub-network might be activated for a set of test samples that are not uniformly sampled from the data distribution, e.g., smaller sub-networks tend to handle ``easy'' samples, while larger sub-networks are used for ``hard'' inputs . This brings a divergence between the training data distribution and that of the inference stage, and thus violates the common \\\\emph{i.i.d.} assumption in classical machine learning. Therefore, it would be interesting to develop new theories to analyze the generalization properties of dynamic networks under such distribution mismatch. Note that transfer learning also aims to address the issue of distributional shift at test time, but the samples of the target domain are assumed to be accessible in advance. In contrast, for dynamic models, the test distribution is not available until the training process is finished, when the network architecture and parameters are finalized. This poses greater challenges than analyzing the generalization issues in transfer learning.\",\n \"id\": \"3fc7f7a7-7403-4a95-918e-d6e40acaa928\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"e9013abd-3d22-4eab-a882-4ba5e86719f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Theories for Dynamic Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Theories for Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0.764705882352941,\n \"cites\": [\n 7266,\n 96,\n 700,\n 8372,\n 718,\n 7297,\n 97,\n 787,\n 732,\n 696,\n 685,\n 504,\n 687\n ],\n \"content\": \"Architecture design has been proven to be essential for deep networks. Existing researches on architectural innovations are mainly proposed for static models , while relatively few are dedicated to developing architectures specially for dynamic networks. It is expected that architectures developed specifically for dynamic networks may further improve their effectiveness and efficiency. For example, the interference among multiple classifiers in an early-exiting network could be mitigated by a carefully designed multi-scale architecture with dense connections . \\nPossible research direction include designing dynamic network structures either by hand (as in ), or by leveraging the NAS techniques (as in ). Moreover, considering the popularity of Transformers , {recent work has proposed dynamic vision Transformers with adaptive early exiting or token sparsification .} Developing a dynamic version of this family of models could also be an interesting direction.\\n{Note that the research on dynamic networks differs from a seemingly close topic, i.e. model compression . One common goal of them is improving the network efficiency with minimal accuracy drop. However, model compression may focus on reducing the \\\\emph{size} of deep networks, while dynamic networks pay more attention to the \\\\emph{computation}, even at the price of slightly \\\\emph{increasing} model size . Moreover, model compression typically adopts pruning or quantization techniques to produce compact \\\\emph{static} models, which treat all the inputs in the same way. In contrast, dynamic networks perform data-dependent computation on different inputs, which can effectively reduce the intrinsic redundancy in static models.}\\n\\\\vspace{-1ex}\",\n \"id\": \"9458c790-c005-432b-a25a-958ab55678a8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"e9013abd-3d22-4eab-a882-4ba5e86719f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Architecture Design for Dynamic Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Architecture Design for Dynamic Networks\"\n },\n {\n \"cite_extract_rate\": 0.833333333333333,\n \"cites\": [\n 7281,\n 675,\n 7268,\n 674,\n 676\n ],\n \"content\": \"\\\\vspace{-0.5ex}\\nMany existing dynamic networks (e.g., most of the sample-wise adaptive networks) are designed specially for classification tasks, and cannot be applied to other vision tasks such as object detection and semantic segmentation. The difficulty arises from the fact that for these tasks there is no simple criterion to assert whether an input image is easy or hard, as it usually contains multiple objects and pixels that have different levels of difficulty. Although many efforts, e.g., spatially adaptive models and soft attention based models , have been made to address this issue, it remains a challenging problem to develop a unified and elegant dynamic network that can serve as an off-the-shelf backbone for a variety of tasks.\\n\\\\vspace{-1.5ex}\",\n \"id\": \"baaa0a8b-88d7-44dc-8c2b-1c3a26c2c457\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"e9013abd-3d22-4eab-a882-4ba5e86719f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Applicability for More Diverse Tasks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Applicability for More Diverse Tasks\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 721,\n 788,\n 7281\n ],\n \"content\": \"\\\\label{discuss_implement}\\n\\\\vspace{-0.5ex}\\nThe current deep learning hardware and libraries are mostly optimized for static models, and they may not be friendly to dynamic networks. Therefore, we usually observe that the practical runtime of dynamic models lags behind the theoretical efficiency. For example, some spatially adaptive networks involve sparse computation, {which is known to be inefficient on modern computing devices due to the memory access bottleneck . A recent line of work focuses on the codesign of algorithm and hardware for accelerating deep models on platforms with more flexibility such as FPGA . Many input-dependent operations, including pixel-level dynamic computation , adaptive channel pruning and early exiting , have also been tailored together with hardware for further improving their practical efficiency. It is an interesting research direction to simultaneously optimize the algorithm, hardware and deep learning libraries to harvest the theoretical efficiency gains of dynamic networks.}\\n{In addition, a data-dependent inference procedure, especially for the dynamic \\\\emph{architectures}, usually requires a model to handle input samples sequentially, which also poses challenge for parallel computation. Although inference with batches has been enabled for early-exiting networks , the conflict between adaptive computational graph and parallel computation still exists for other types of dynamic architectures. This issue is mitigated in the scenario of mobile/edge computing, where the input signal by itself is sequential and the computing hardware is less powerful than high-end platforms. However, designing dynamic networks that are more compatible with existing hardware and software is still a valuable and challenging topic.}\\n\\\\vspace{-1.5ex}\",\n \"id\": \"23446088-fa89-4b21-8a64-51ee553e33cb\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"e9013abd-3d22-4eab-a882-4ba5e86719f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Gap between Theoretical \\\\& Practical Efficiency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Gap between Theoretical \\\\& Practical Efficiency\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 789\n ],\n \"content\": \"\\\\vspace{-0.5ex}\\nDynamic models may provide new perspectives for the research on adversarial robustness of deep neural networks. {For example, recent work has leveraged the multi-exit structure to improve the robustness against adversarial attacks. Moreover, traditional attacks are usually aimed at causing \\\\emph{misclassification}. For dynamic networks, it is possible to launch attacks on \\\\emph{efficiency} . Specifically, by adjusting the objective function of the adversarial attack, input-adaptive models could be fooled to activate all their intermediate layers or yielding confusing predictions at early exits even for \\\"easy\\\" samples. It has also been observed that the commonly used adversarial training is not effective to defend such attacks.} The robustness of dynamic network is an interesting yet understudied topic.\\n\\\\vspace{-1.5ex}\",\n \"id\": \"a8fb9eb8-777e-40f9-a6cc-da50d402237b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"e9013abd-3d22-4eab-a882-4ba5e86719f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Robustness Against Adversarial Attack\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Robustness Against Adversarial Attack\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\vspace{-0.5ex}\\nDynamic networks inherit the black-box nature of deep learning models, and thus also invite research on interpreting their working mechanism. What is special here is that the adaptive inference paradigm, e.g., spatial/temporal adaptiveness, conforms well with that of the human visual system, and may provide new possibilities for making the model more transparent to humans. In a dynamic network, it is usually convenient to analyze which part of the model is activated for a given input or to locate which part of the input the model mostly relies on in making its prediction. {It is expected that the research on dynamic network will inspire new work on the interpretability of deep learning.}\\n\\\\vspace{-2ex}\\n\\\\appendices\\n\\\\ifCLASSOPTIONcompsoc\\n \\\\section*{Acknowledgments}\\n\\\\else\\n \\\\section*{Acknowledgment}\\n\\\\fi\\nThis work is supported in part by the National Science and Technology Major Project of the Ministry of Science and Technology of China under Grants 2018AAA0100701, the National Natural Science Foundation of China under Grants 61906106 and 62022048, the Institute for Guo Qiang of Tsinghua University and Beijing Academy of Artificial Intelligence.\\n\\\\ifCLASSOPTIONcaptionsoff\\n \\\\newpage\\n\\\\fi\\n\\\\small{\\n \\\\bibliography{reference}\\n}\\n\\\\end{document}\",\n \"id\": \"80ebe8d1-1699-4384-9941-1c27455ceb08\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"e9013abd-3d22-4eab-a882-4ba5e86719f2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Dynamic Neural Networks: A Survey\"\n ],\n [\n \"section\",\n \"{Challenges and Future Directions\"\n ],\n [\n \"subsection\",\n \"Interpretability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Interpretability\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":93,"string":"93"},"all_cites":{"kind":"list like","value":[71,96,679,681,687,544,305,682,683,675,38,504,97,686,684,680,677,676,7266,8372,7268,7,678,7267,674,685,306,688,698,695,7273,693,690,691,7269,8373,7270,697,692,694,696,7272,7271,689,706,705,707,8375,8374,710,712,713,709,699,711,703,702,700,701,708,704,715,714,719,720,718,7274,716,717,8376,721,7276,7275,722,726,7277,724,7215,727,725,8377,723,736,730,729,7278,8378,734,8379,733,731,7040,728,735,732,737,738,7279,7280,739,7281,7257,740,741,7282,742,743,531,747,745,744,746,748,7284,750,7283,749,7285,7286,755,752,751,754,8380,753,8381,756,6975,8382,757,7288,7287,758,759,779,770,771,776,778,7294,7290,772,150,760,7289,761,8383,764,765,762,777,775,769,767,7295,7291,763,773,774,7292,766,768,7293,780,781,782,785,786,783,7296,784,7297,787,788,789],"string":"[\n 71,\n 96,\n 679,\n 681,\n 687,\n 544,\n 305,\n 682,\n 683,\n 675,\n 38,\n 504,\n 97,\n 686,\n 684,\n 680,\n 677,\n 676,\n 7266,\n 8372,\n 7268,\n 7,\n 678,\n 7267,\n 674,\n 685,\n 306,\n 688,\n 698,\n 695,\n 7273,\n 693,\n 690,\n 691,\n 7269,\n 8373,\n 7270,\n 697,\n 692,\n 694,\n 696,\n 7272,\n 7271,\n 689,\n 706,\n 705,\n 707,\n 8375,\n 8374,\n 710,\n 712,\n 713,\n 709,\n 699,\n 711,\n 703,\n 702,\n 700,\n 701,\n 708,\n 704,\n 715,\n 714,\n 719,\n 720,\n 718,\n 7274,\n 716,\n 717,\n 8376,\n 721,\n 7276,\n 7275,\n 722,\n 726,\n 7277,\n 724,\n 7215,\n 727,\n 725,\n 8377,\n 723,\n 736,\n 730,\n 729,\n 7278,\n 8378,\n 734,\n 8379,\n 733,\n 731,\n 7040,\n 728,\n 735,\n 732,\n 737,\n 738,\n 7279,\n 7280,\n 739,\n 7281,\n 7257,\n 740,\n 741,\n 7282,\n 742,\n 743,\n 531,\n 747,\n 745,\n 744,\n 746,\n 748,\n 7284,\n 750,\n 7283,\n 749,\n 7285,\n 7286,\n 755,\n 752,\n 751,\n 754,\n 8380,\n 753,\n 8381,\n 756,\n 6975,\n 8382,\n 757,\n 7288,\n 7287,\n 758,\n 759,\n 779,\n 770,\n 771,\n 776,\n 778,\n 7294,\n 7290,\n 772,\n 150,\n 760,\n 7289,\n 761,\n 8383,\n 764,\n 765,\n 762,\n 777,\n 775,\n 769,\n 767,\n 7295,\n 7291,\n 763,\n 773,\n 774,\n 7292,\n 766,\n 768,\n 7293,\n 780,\n 781,\n 782,\n 785,\n 786,\n 783,\n 7296,\n 784,\n 7297,\n 787,\n 788,\n 789\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.199999018836062,"string":"1.199999"}}},{"rowIdx":61,"cells":{"authors":{"kind":"list like","value":["Deqiang Li","Qianmu Li","Yanfang Ye","Shouhuai Xu"],"string":"[\n \"Deqiang Li\",\n \"Qianmu Li\",\n \"Yanfang Ye\",\n \"Shouhuai Xu\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Arms Race in Adversarial Malware Detection: A Survey"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-05-24T07:20:42Z","string":"2020-05-24T07:20:42Z"},"category":{"kind":"string","value":"cs.CR"},"abstract":{"kind":"string","value":" Malicious software (malware) is a major cyber threat that has to be tackled with Machine Learning (ML) techniques because millions of new malware examples are injected into cyberspace on a daily basis. However, ML is vulnerable to attacks known as adversarial examples. In this paper, we survey and systematize the field of Adversarial Malware Detection (AMD) through the lens of a unified conceptual framework of assumptions, attacks, defenses, and security properties. This not only leads us to map attacks and defenses to partial order structures, but also allows us to clearly describe the attack-defense arms race in the AMD context. We draw a number of insights, including: knowing the defender's feature set is critical to the success of transfer attacks; the effectiveness of practical evasion attacks largely depends on the attacker's freedom in conducting manipulations in the problem space; knowing the attacker's manipulation set is critical to the defender's success; the effectiveness of adversarial training depends on the defender's capability in identifying the most powerful attack. We also discuss a number of future research directions. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"4fe25043-06f8-4577-bd62-3fe42c7ab899","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"]],"subsections":["74b358c3-1042-4f31-a16f-380254516292","a1283bee-37c1-4371-bed3-ceeb8af64d75","e26ef155-9b9c-4a92-9ffa-e334fbe6ea87","965da080-653e-47f5-949b-de42288cf6a5","82494750-c0aa-44ab-add1-1d1d59481f4a"],"title":"root"},{"cite_extract_rate":0.32692307692307604,"cites":[314,7755,7337,893,3856,3857,937,3853,3854,7754,7153,3855,3861,7154,3860,3859,3858],"content":"\\label{sec:intro}\nMalware (malicious software) is a big cyber threat and has received a due amount of attention. For instance, Kaspersky reports that 21,643,946 unique malicious files were detected in the year 2018, 24,610,126 in 2019, and 33,412,568 in 2020 .\nA popular defense against malware is to use {\\em signature-based} detectors , where a signature is often extracted by malware analysts from known malware examples. This approach has two drawbacks: signatures are tedious to extract and can be evaded by a range of techniques (e.g., encryption, repacking, polymorphism ). This incompetence has motivated the use of Machine Learning (ML) based malware detectors, which can be automated to some degree and can possibly detect \\emph{new} malware examples (via model generalization or knowledge adaptation ). More recently, Deep Learning (DL) has been used for malware detection (see, e.g., ). \nWhile promising, ML-based malware detectors are vulnerable to attacks known as {\\em adversarial examples} . \nThere are two kinds of attacks. One is {\\em evasion attack}, where the attacker perturbs test examples to {\\em adversarial examples} to evade malware detectors . The other is {\\em poisoning attack}, where the attacker manipulates the training dataset for learning malware detectors . These attacks usher in the new field of Adversarial Malware Detection (AMD) . \nThe state-of-the-art in AMD is that there are some specific results scattered in the literature but there is no systematic understanding.\nThis is true despite that there have been attempts at systematizing the related field of Adversarial Machine Learning (AML) , which however cannot be automatically translated to AMD. This is so because malware detection \nhas three {\\em unique} characteristics which are not exhibited by the other application domains (e.g., image or audio processing). (i) There are no common, standard feature definitions because both attackers and defenders can define their own features to represent computer files. As a consequence, attackers can leverage this ``freedom'' in feature definition to craft adversarial examples. (ii) Malware features are often discrete rather than continuous and program files are often highly structured with multiple modalities. This means that arbitrarily perturbing malware files or their feature representations might make the perturbed files no more executable. \nThis also means that the discrete domain makes perturbation a {\\em non-differentiable} and {\\em non-convex} task. \n(iii) Any meaningful perturbation to a malware example or its feature representation must preserve its malicious functionality. For example, the Android Package Kit (APK) requires that the used permissions are publicized in the AndroidManifest.xml, meaning that removing permissions in this manifest file would incur a runtime error. The preceding (ii) and (iii) make both the attacker's and defender's tasks more challenging than their \ncounterparts where small perturbations are not noticeable (e.g., images).\n\\noindent{\\bf Our Contributions}. We propose a conceptual framework for systematizing the AMD field through the lens of {\\em assumptions}, {\\em attacks}, {\\em defenses}, and {\\em security properties}. In specifying these, we seek rigorous definitions whenever possible, while noting that these definitions have been scattered in the literature. Rigorous definitions are important because they can serve as a common reference for future studies. The framework allows us to map the known attacks and defenses into some partial order structures and systematize the AMD attack-defense arms race. \nWe make a number of observations, including: (i) the indiscriminate attack that treats malicious examples as equally important has been extensively investigated, but targeted and availability attacks are much less investigated; (ii) the evasion attack is much more extensively studied than the poisoning attack; (iii) there is no silver-bullet defense against evasion and poisoning attacks; (iv) sanitizing examples is effective against black-box and grey-box attacks, but not white-box attacks; (v) AMD security properties have been evaluated empirically rather than rigorously; (vi) there is no theoretical evidence to support that the effectiveness of defense techniques on the training set can generalize to other adversarial examples.\nWe draw a number of insights, including: (i) knowing defender's feature set is critical to the success of transfer attacks, highlighting the importance of keeping defender's feature set secret (e.g., by randomizing defender's feature set);\n(ii) the effectiveness of practical evasion attacks largely depends on the attacker's degree of freedom in conducting manipulations in the problem space (i.e., a small degree of freedom means harder to succeed); (iii) effective defenses often require the defender to know the attacker's manipulation set, explaining from one perspective why it is hard to design effective defenses;\n(iv) effectiveness of adversarial training depends on the defender's capability in identifying the most powerful attack.\nFinally, we discuss a number of future research directions, which hopefully will inspire and encourage many researchers to explore them.\n\\smallskip\n\\noindent{\\bf Related Work}.\nThe closely related prior work is Maiorca et al. , which surveys previous studies in adversarial malicious PDF document detection. In contrast, we consider the broader context of AMD and propose novel partial orders to accommodate AMD assumptions, attacks, defenses, and properties.\nThere are loosely-related prior studies, which survey prior AML studies (but not focusing on AMD), including .\nFor example, Yuan et al. survey attack methods for generating adversarial examples, while briefly discussing evasion attacks in the AMD context; \nBarreno et al. propose a taxonomy of AML attacks\n(causative vs. exploratory attacks, integrity vs. availability attacks, and targeted vs. indiscriminate attacks); Biggio et al. propose a defense framework for protecting \nSupport Vector Machines (SVMs) from evasion attacks, poisoning attacks and privacy violations; Papernot et al. systematize AML security and privacy with emphasis on demonstrating the trade-off between detection accuracy and robustness.\n\\smallskip\n\\noindent{\\bf Paper Outline}.\nSection \\ref{sec:fw} describes our survey and systematization methodology and framework.\nSection \\ref{sec:systematizstion-AMD-studies} applies our framework to systematize the literature AMD studies.\nSection \\ref{sec:cd} discusses future research directions. \nSection \\ref{sec:conc} concludes the paper.","id":"74b358c3-1042-4f31-a16f-380254516292","level":"section","origin_cites_number":52,"parent_id":"4fe25043-06f8-4577-bd62-3fe42c7ab899","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:fw}\n\\begin{table}[htbp!]\n\t\\centering\\caption{Main notations used in the paper \\label{table:notations}}\n\t\\begin{adjustbox}{width=0.99\\textwidth}\n\t\\begin{tabular}{l|p{.8\\textwidth}}\n\t\t\\hline\n\t\tNotation & Meaning\\\\\\hline\n\t\t$\\mathbb R$ ($\\mathbb R_{+}$) & the set of (positive) real numbers \\\\\n\t\t$\\mathcal{A},\\mathcal{I}$ & attacker and defender (treated as algorithms)\\\\\n\t\t$\\mathbb{P}$ & the probability function \\\\\n\t\t$z, z'\\in\\mathcal{Z}$ & $\\mathcal{Z}$ is example space; $z'$ is obtained by perturbing $z$\\\\\n\t\t$S$ & defender $\\mathcal{I}$'s feature set for representing files \\\\\n\t\t${\\mathbf x}, {\\mathbf x}'\\in \\mathcal X$ & $\\mathcal X=\\mathbb{R}^d$ is $d$-dimensional feature space; ${\\mathbf x},{\\mathbf x}' \\in \\mathcal{X}$ are respectively feature representations of $z,z'\\in \\mathcal{Z}$ \\\\\n\t\t$\\mathcal{Y}$, $y$ & $\\mathcal{Y}$ is the label space of binary classification, $\\mathcal{Y}=\\{+/1,-/0\\}$; $y\\in \\mathcal{Y}$ \\\\\n\t\t$\\mathcal{D}=\\mathcal Z \\times \\mathcal Y$ & the file-label (i.e., example-label) space \\\\\n\t\t$D_{train}\\subset \\mathcal D,n$ & the training set in file-label space; $n=|D_{train}|$ \\\\\n\t\t$D_{test}$ & the test set in file-label space \\\\\n\t\t$D_{poison},D'_{poison}$& $D'_{poison}$ is set of adversarial file-label pairs obtained by perturbing non-adversarial files in $D_{poison}\\subset\\mathcal{D}$ \\\\\n\t\t$\\mathcal{O}(z,z')$ & $\\mathcal{O}(z,z'):\\mathcal{Z}\\times\\mathcal{Z}\\to\\{{\\tt true}, {\\tt false}\\}$ is an oracle telling if two files have the same functionality or not \\\\\n\t\t$\\delta$ & a manipulation for perturbing files {\\em with} preserving their functionalities \\\\\n\t\t$\\mathcal{M}$, $\\mathcal{Z}_\\mathcal{M}\\subseteq \\mathcal{Z}$ & $\\mathcal{M}$ is manipulation set in the problem space;\n\t\t$\\mathcal{Z}_\\mathcal{M}$ is set of adversarial files generated using $\\mathcal{M}$\\\\\n\t\t$\\mathbf{M}$, $\\mathcal{X}_\\mathbf{M}\\subseteq \\mathcal{X}$ & $\\mathbf{M}$ is feature manipulation set;\n\t\t$\\mathcal{X}_\\mathbf{M} $ is set of adversarial feature vectors generated using $\\mathbf{M}$ \\\\\n\t\t$\\Gamma(z,z')$ & $\\Gamma(z,z'):\\mathcal Z \\times \\mathcal{Z} \\to \\mathbb R_+$ measures the degree of manipulation for perturbing $z \\in \\mathcal{Z}$ into $z'\\in \\mathcal{Z}$ \\\\\n\t\t$C(\\mathbf{x},\\mathbf{x}')$& $C(\\mathbf{x},\\mathbf{x}'):\\mathcal X \\times \\mathcal X \\to \\mathbb R_+$ is the function measuring the cost incurred by changing feature vector $\\mathbf{x}$ to $\\mathbf{x}'$ \\\\\n\t\t$\\delta_z \\in \\mathcal{M}$ & $\\delta_z$ is a set of manipulations of $z$ w.r.t. $z'$ \\\\\n\t\t$\\delta_{\\mathbf{x}} \\in \\mathbf{M}$ & $\\delta_{\\mathbf{x}}={\\mathbf x}' - {\\mathbf x}$ is a perturbation vector of $\\mathbf{x}$ w.r.t. $\\mathbf{x}'$\\\\\n\t\t$\\phi:\\mathcal Z \\to \\mathcal{X} $ & feature extraction function; $\\mathbf{x}\\leftarrow\\phi(z)$, $\\mathbf{x}'\\leftarrow\\phi(z')$ \\\\\n\t\t$\\varphi,f$ & $\\varphi:\\mathcal{X} \\to \\mathbb{R}$ is classification function; $f:\\mathcal Z \\to \\mathbb{R}$ is classifier $f=\\varphi(\\phi(\\cdot))$; by abusing notation a little bit, we also use ``$+\\leftarrow f(z)$'' to mean that $f$ predicts $z$ as malicious when\n\t\t$f(z)\\geq \\tau$ for a threshold $\\tau$\\\\\n\t\t$F_{\\theta}:\\mathcal X \\to \\mathbb R$& machine learning algorithm \n\t\twith parameters $\\theta$ \\\\\n\t\t$L:\\mathbb{R}\n\t\t\\times\\mathcal{Y}\\to\\mathbb{R}$ & loss function measuring prediction error of $F_\\theta$ \\\\\n\t\t${\\sf EL},{\\sf WR},{\\sf AT}$ & defense techniques: Ensemble Learning, Weight Regularization, Adversarial Training \\\\\n\t\t${\\sf VL},{\\sf RF},{\\sf IT},{\\sf CD},{\\sf SE}$ & defense techniques: Verifiable Learning, Robust Feature, Input Transformation, Classifier ranDomization, Sanitizing Examples \\\\\n\t\t${\\sf BE},{\\sf OE},{\\sf BP},{\\sf OP}$ & attack tactics: basic and optimal evasion; basic and optimal poisoning\\\\\n\t\t{\\sf GO},{\\sf SF},{\\sf MI} & attack techniques: Gradient-based Optimization, Sensitive Features, MImicry\\\\\n\t\t{\\sf TR},{\\sf HS},{\\sf GM},{\\sf MS} & attack techniques: TRansferability, Heuristic Search, Generative Model, Mixture Strategy \\\\\n\t\t$A_1,\\ldots,A_5$ & the 5 attributes under $\\mathcal{I}$'s control; they are known to $\\mathcal{A}$ at respective degrees\n\t\t$a_1,\\ldots,a_5$ \\\\\n\t\t$A_6,\\ldots,A_9$ & the 4 attributes under $\\mathcal{A}$'s control; they are known to $\\mathcal{I}$ at respective degree $a_6,\\ldots,a_9$ \\\\\n\t\t${\\sf RR},{\\sf CR},{\\sf DR},{\\sf TR}$ & security properties: Representation Robustness, Classification Robustness, Detection Robustness, Training Robustness\\\\\n\t\t\\hline\n\t\\end{tabular}\n\t\\end{adjustbox}\n\\end{table}\n\\noindent{\\bf Terminology, Scope and Notations}. \nIn the AMD context, a defender $\\mathcal{I}$ aims to use ML to detect or classify computer files as benign or malicious; i.e., we focus on {\\em binary classification}. An attacker $\\mathcal{A}$ attempts to make malicious files evade $\\mathcal{I}$'s detection by leveraging {\\em adversarial files} (interchangeably, {\\em adversarial examples}). Adversarial malware examples are often generated by perturbing or manipulating malware examples, explaining why we will use the two terms, perturbation and manipulation, interchangeably. Adversarial attacks can be waged in the training phase of a ML model (a.k.a., poisoning attack) or in the test phase (a.k.a., evasion attack). It is worth mentioning that the {\\em privacy violation} attack is waged in addition to the preceding two attacks because $\\mathcal{A}$ can always probe $\\mathcal{I}$'s detectors.\nA file, benign and malicious alike, is adversarial if it is intentionally crafted to (help malicious files) evade $\\mathcal{I}$'s detection, and non-adversarial otherwise.\nWe focus on $\\mathcal{I}$ using supervised learning to detect malicious files, which may be adversarial or non-adversarial because they co-exist in the real world with no self-identification. This means that we do not consider the large body of malware detection literature that does not cope with AMD, which has been addressed elsewhere (e.g., ). \nTable \\ref{table:notations} summarizes the main notations used in the paper.","id":"a1283bee-37c1-4371-bed3-ceeb8af64d75","level":"section","origin_cites_number":2,"parent_id":"4fe25043-06f8-4577-bd62-3fe42c7ab899","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Survey and Systematization Methodology"]],"subsections":["b7dbff6a-c1f8-49f8-8c03-efdb7ba23041","8e32af6c-587d-43e1-aabc-dc5d8345628d","b05fe1bc-da2e-4a87-8f17-d43f1acc81fc"],"title":"Survey and Systematization Methodology"},{"cite_extract_rate":0.277777777777777,"cites":[1159,8565,318,166,3861],"content":"\\label{sec:ml}\nLet $\\mathcal Z$ be the {\\em example space} of benign/malicious adversarial/non-adversarial files. Let $\\mathcal{Y}=\\{+,-\\}$ or $\\mathcal{Y}=\\{1,0\\}$ be the {\\em label space} of binary classification, where $+$/$1$ ($-$/$0$) means a file is malicious (benign). Let $\\mathcal D =\\mathcal Z \\times \\mathcal Y$ be the file-label (example-label) space. For training and evaluating a classifier in the absence of adversarial files, $\\mathcal I$ is given a set $D\\subset \\mathcal{D}$ of non-adversarial benign/malicious files as well as their {\\em ground-truth} labels. $\\mathcal I$ splits $D$ into three disjoint sets: a training set $D_{train}=\\{(z_i,y_i)\\}_{i=1}^n$, a validation set for model selection, and a test set for evaluation. \nEach file $z_i\\in \\mathcal{Z}$ is characterized by a set $S$ of features and represented by a numerical vector $\\mathbf{x}_i=(x_{i,1},\\ldots,x_{i,d})$ in the $d$-dimensional {\\em feature space} $\\mathcal X=\\mathbb R^d$, which accommodates both continuous and discrete feature representations .\nThe process for obtaining feature representation $\\mathbf{x}_i$ of $z_i\\in \\mathcal{Z}$ is called {\\it feature extraction}, denoted by a function $\\phi:\\mathcal{Z}\\to \\mathcal{X}$ with ${\\bf x}_i\\leftarrow \\phi(S,z_i)$. Because $\\phi$ can be hand-crafted (denoted by $\\phi_c$), automatically learned (denoted by $\\phi_a$), or a hybrid of both , we unify them into $\\phi$ such that $\\phi(S,z)=\\phi_a(\\phi_c(S,z))$; when only manual (automatic) feature extraction is involved, we can set $\\phi_a$ ($\\phi_c$) as the identity map.\nThere are two kinds of features: {\\em static features} are extracted via static analysis (e.g., strings, API calls\n); {\\em dynamic features} are extracted via dynamic analysis (e.g., instructions, registry activities\n).\n\\begin{figure}[!htbp]\n\t\\centering\n\t\\scalebox{0.32}{\n\t\\includegraphics{figures/ml.pdf}\n\t}\n\\caption{Illustration of ML-based malware detector.\n}\n\t\\label{fig:ml}\n\\end{figure}\nAs highlighted in Figure \\ref{fig:ml},\n$\\mathcal{I}$ uses $\\{(z_i,y_i)\\}_{i=1}^n$ to learn a malware detector or classifier $f:\\mathcal{Z}\\to [0,1]$, where $f(z)=\\varphi(\\phi(S,z))$ is composed of feature extraction function $\\phi: \\mathcal{Z}\\to \\mathcal{X}$ and classification function $\\varphi:\\mathcal{X} \\to [0,1]$.\nNote that $f(z)\\in [0,1]$, namely $\\varphi({\\bf x})\\in [0,1]$ with ${\\bf x}\\leftarrow \\phi(S,z)$, can be interpreted as the probability that $z$ is malicious (while noting that calibration may be needed ). For a given threshold $\\tau\\in[0,1]$, we further say (by slightly abusing notations) $z$ is labeled by $f$ as $+$, or $+\\leftarrow f(z)$, if $f(z)\\geq \\tau$, and labeled as $-$ or $-\\leftarrow f(z)$ otherwise. In practice, $f$ is often specified by a learning algorithm $F$ with learnable parameter $\\theta$ (e.g., weights) and a hand-crafted feature extraction $\\phi_c$; then, $\\theta$ is tuned to minimize the empirical risk associated with a loss function $L:[0,1] \\times \\mathcal Y \\to \\mathbb{R}$ measuring the prediction error of $F_\\theta$ (e.g., cross-entropy ), namely\n\\begin{equation}\n\\min \\limits_{\\theta}~\\mathcal{L}(\\theta, D_{train}) = \\min \\limits_{\\theta}~\\frac{1}{n}\\sum_{(z_i, y_i)\\in D_{train}}\\left(L(F_\\theta(\\phi_c(S,z_i)), y_i)\\right). \\label{eq:rsk_emp}\n\\end{equation}\n\\noindent{\\bf Example 1: The Drebin malware detector}. Drebin is an Android malware detector trained from static features . Table \\ref{table:drebin} summarizes Drebin's feature set, which includes 4 subsets \n\\begin{table}[htbp!]\n\t\\centering\\caption{Drebin features \\label{table:drebin}}\n\t\\begin{tabular}{c|p{.32\\textwidth}}\n\t\t\\hline\n\t\t\\multicolumn{2}{c}{Feature set} \\\\\\hline\n\t\t\\multirow{4}{*}{Manifest} \n\t\t& $S_1$~~Hardware components \\\\\n\t\t& $S_2$~~Requested permissions\\\\\n\t\t& $S_3$~~App components \\\\\n\t\t& $S_4$~~Filtered intents \\\\\\hline\n\t\t\\multirow{4}{*}{Dexcode}\n\t\t& $S_5$~~Restricted API calls\\\\\n\t\t& $S_6$~~Used permissions\\\\\n\t\t& $S_7$~~Suspicious API calls\\\\\n\t\t& $S_8$~~Network addresses\\\\\n\t\t\\hline\n\t\\end{tabular}\n\\end{table}\nof features $S_1, S_2, S_3, S_4$ extracted from the AndroidManifest.xml and another 4 subsets of features $S_5, S_6, S_7, S_8$ extracted from the disassembled DEX code files (recalling that DEX code is compiled from a program written in some language and can be understood by the Android Runtime). Specifically, $S_1$ contains features related to the access of an Android package (APK) to smartphone hardware (e.g., camera, touchscreen, or GPS module); $S_2$ contains features related to APK's requested permissions listed in the manifest prior to installation; $S_3$ contains features related to application components (e.g., {\\em activities}, {\\em service}, {\\em receivers}, and {\\em providers}.); $S_4$ contains features related to APK's communications with the other APKs; $S_5$ contains features related to critical system API calls, which cannot run without appropriate permissions or the {\\em root} privilege; $S_6$ contains features corresponding to the used permissions; $S_7$ contains features related to API calls that can access sensitive data or resources in a smartphone; and $S_8$ contains features related to IP addresses, hostnames and URLs found in the disassembled codes.\nThe feature representation is binary, meaning $\\phi=\\phi_c:\\mathcal{Z} \\mapsto \\{0,1\\}^d$ with $|S|=d$ and ${\\bf x}=(x_1,\\ldots,x_d)$, where $x_i=1$ if the corresponding feature is present in the APK $z$ and $x_i=0$ otherwise. A file $z$ in the feature space looks like the following: \n\\[\n\\mathbf x = \\phi(z) \\to\n\\begin{pmatrix}\n\\cdots \\\\ 0\\\\ 1\\\\\n\\cdots \\\\ 1\\\\ 0\\\\\n\\cdots\\\\\n\\end{pmatrix}\n\\begin{array}{ll}\n\\cdots & \\multirow{4}{*}{\\hspace{-1mm}\\bigg \\} $S_2$ }\\\\\n\\texttt{\\small permission::SEND\\_SMS} \\\\\n\\texttt{\\small permission::READ\\_CONTACTS}\\\\\n\\cdots & \\multirow{4}{*}{\\hspace{-1mm}\\bigg \\} $S_5$ }\\\\\n\\texttt{\\small api\\_call::getDeviceID}\\\\\n\\texttt{\\small api\\_call::setWifiEnabled}\\\\\n\\cdots & \\\\\n\\end{array}\n\\]\nDrebin uses a linear Support Vector Machine (SVM) to learn classifiers.\n\\smallskip\n\\noindent{\\bf Example 2: The MalConv malware detector}.\nMalConv is Convolutional Neural Network (CNN)-based Windows Portable Executable (PE) malware detector learned from raw binary programs (i.e., end-to-end detection)\n. \nFigure \\ref{fig:malconv} depicts its architecture.\nThe sequence of binary code is transformed into byte values (between 0 to 255) with the maximum length bounded by $N_{max}$ (e.g., $N_{max}=2^{21}$ bytes or 2MB). Each byte is further mapped into a real-valued vector using the embedding . The CNN layer and pooling layer learn abstract representations. The embedding, CNN and pooling layers belong to feature extraction $\\phi_a$, and the fully-connected and softmax layers belong to the classification operation $\\varphi$.\n\\begin{figure}[!htbp]\n\t\\centering\n\t\\includegraphics[width=.7\\textwidth]{figures/malconv.pdf}\n\t\\caption{MalConv architecture .}\n\t\\label{fig:malconv}\n\\end{figure}","id":"b7dbff6a-c1f8-49f8-8c03-efdb7ba23041","level":"subsection","origin_cites_number":18,"parent_id":"a1283bee-37c1-4371-bed3-ceeb8af64d75","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Survey and Systematization Methodology"],["subsection","Brief Review on ML-based Malware Detection"]],"subsections":[],"title":"Brief Review on ML-based Malware Detection"},{"cite_extract_rate":0,"cites":[],"content":"We systematize AMD studies through the lens of four aspects: (i) the assumptions that are made; (ii) the attack or threat model in terms of attacker ${\\mathcal A}$'s {\\em objective} and ${\\mathcal A}$'s {\\em input}, with the latter including ${\\mathcal A}$'s information about the defender $\\mathcal{I}$ and ${\\mathcal A}$'s own; (iii) the defense in terms of $\\mathcal{I}$'s {\\em objective} and $\\mathcal{I}$'s {\\em input}, with the latter including $\\mathcal{I}$'s information about $\\mathcal{A}$ and $\\mathcal{I}$'s own; (iv) the security properties that are at stake. These four aspects are respectively elaborated below.","id":"8e32af6c-587d-43e1-aabc-dc5d8345628d","level":"subsection","origin_cites_number":0,"parent_id":"a1283bee-37c1-4371-bed3-ceeb8af64d75","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Survey and Systematization Methodology"],["subsection","Framework"]],"subsections":["18cf934c-a897-4b88-8a38-4cd32c20a2a0","5f85312b-af7a-44ed-87ac-7e94a51dd8cc","1e0de86d-a9b2-453b-b977-3d5abab5bd99"],"title":"Framework"},{"cite_extract_rate":0.4375,"cites":[318,3857,3862,937,7312,9099,3854],"content":"\\label{sec:assumption}\nFive assumptions have been made in the AMD literature. Assumption \\ref{assumption:iid} below says that the data samples in $D$ are Independent and Identically Distributed (IID), which is a strong assumption and researchers have started to weaken it . \n\\begin{assumption}[{\\sf IID} assumption; see, e.g., ]\n\\label{assumption:iid}\nComputer files in training data and testing data are independently drawn from the same distribution.\n\\end{assumption}\nAssumption \\ref{assumption:oracle} below is adapted from AML context, where humans can serve as an oracle $\\mathcal{O}$ for determining whether two images are the same . In the AMD context, $\\mathcal{O}$ can be instantiated as (or approximated by) malware analysts or automated tools (e.g., Sandbox ), with the latter often using heuristic rules produced by malware analysts (e.g., YARA ).\n\\begin{assumption}[{\\sf Oracle} assumption; adapted from ]\n\\label{assumption:oracle}\nThere is an oracle $\\mathcal{O}:\\mathcal{Z}\\times\\mathcal{Z}\\to\\{{\\tt true},{\\tt false}\\}$ that tells if two files $z,z'\\in\\mathcal{Z}$ have the same functionality or not; ${\\tt true}\\leftarrow {\\mathcal{O}}(z,z')$ if and only if $z$ and $z'$ have the same functionality. \n\\end{assumption}\nAssumption \\ref{assumption:measurability} below says that there is a way to measure the degree of manipulations by which one file \nis transformed to another.\n\\begin{assumption}[{\\sf Measurability} assumption ]\n\\label{assumption:measurability}\nThere is a function $\\Gamma(z,z'):\\mathcal Z \\times \\mathcal Z \\to \\mathbb R_+$ that measures the degree of manipulations according to which a file $z'\\in \\mathcal{Z}$ can be derived from the file $z \\in \\mathcal{Z}$.\n\\end{assumption}\nSince Assumption \\ref{assumption:measurability} is often difficult to validate, $\\Gamma(z,z')$ may be replaced by a function that quantifies the degree of manipulation that can turn feature representation $\\mathbf{x}$ into $\\mathbf{x}'$, where\n$\\mathbf{x}=\\phi(S,z)$ and $\\mathbf{x}'=\\phi(S,z')$. This leads to:\n\\begin{assumption}[{\\sf Smoothness} assumption ]\n\\label{assumption:smoothness}\n\\ignore{\nThis assumption says that for any $z,z'\\in \\mathcal{Z}$ and $\\delta \\in \\Delta$, it holds that $C(\\phi(z),\\phi(z'))\\approx 0$ when $(z'\\leftarrow \\mathcal{A}(z,\\delta)) \\wedge (\\Gamma(z,z')\\approx 0) \\wedge ({\\tt true}\\leftarrow {\\mathcal{O}(z,z')})$.\n}\nThere is a function $C(\\mathbf{x},\\mathbf{x}'):\\mathcal X \\times \\mathcal X \\to \\mathbb R_+$ such that $C(\\phi(S,z),\\phi(S,z'))\\approx 0$ when $ (\\Gamma(z,z')\\approx 0) \\wedge ({\\tt true}\\leftarrow {\\mathcal{O}(z,z')})$.\n\\end{assumption}\nAssumption \\ref{assumption:inverse} below says that the inverse of feature extraction, $\\phi^{-1}$, is solvable so that a perturbed representation $\\mathbf{x}'$ can be mapped back to a legitimate file. \n\\begin{assumption}[{\\sf Invertibility} assumption ] \\label{assumption:inverse}\nFeature extraction $\\phi$ is invertible, meaning that given $\\mathbf{x}'$, the function $\\phi^{-1}:\\mathcal{X}\\to\\mathcal{Z}$ produces $z'=\\phi^{-1}(\\mathbf{x}')$. \n\\end{assumption}\nRecall that the feature extraction function $\\phi$ may be composed of a hand-crafted $\\phi_c$ and an automated $\\phi_a$, where $\\phi_c$ may be neither differentiable nor invertible . This means $\\mathbf{x}'$ may not be mapped to a legitimate file. Researchers tend to relax the assumption by overlooking the interdependent features , while suffering from the side-effect $\\mathbf{x}'\\neq \\phi(\\phi^{-1}(\\mathbf{x}'))$ .","id":"18cf934c-a897-4b88-8a38-4cd32c20a2a0","level":"subsubsection","origin_cites_number":16,"parent_id":"8e32af6c-587d-43e1-aabc-dc5d8345628d","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Survey and Systematization Methodology"],["subsection","Framework"],["subsubsection","Systematizing Assumptions"]],"subsections":[],"title":"Systematizing Assumptions"},{"cite_extract_rate":0.5111111111111111,"cites":[8695,890,3866,7156,888,3868,1193,3867,3865,166,6172,3864,7155,3862,937,3853,9099,9100,7754,3863,7153,3869],"content":"\\label{sec:sys-attack}\nWe systematize attacks from two perspectives: {\\em attacker's objective} (i.e., what the attacker attempts to accomplish) and {\\em attacker's input} (i.e., what leverages the attacker can use). Whenever possible, we seek rigorous definitions to specify the attacker's input, while noting that these definitions have been scattered in the literature. We believe this specification is important because it can serve as a common reference for future studies. To demonstrate this, we discuss how to apply it to formulate a partial-order structure for comparing attacks.\n\\smallskip\n\\noindent{\\bf Attacker's Objective}.\nThere are three kinds of objectives: (i)\n{\\sf Indiscriminate}, meaning $\\mathcal{A}$ attempts to cause as many false-negatives as possible ; \n(ii) {\\sf Targeted}, meaning $\\mathcal{A}$ attempts to cause specific false-negatives (i.e., making certain malicious files evade the detection );\n(iii) {\\sf Availability}, meaning $\\mathcal{A}$ attempts to frustrate defender $\\mathcal{I}$ by rendering $\\mathcal{I}$'s classifier $f$ unusable (e.g., causing substantially high false-positives ).\n\\begin{table}[!htbp]\n\\caption{Attributes for specifying $\\mathcal{A}$'s and $\\mathcal{I}$'s input.}\n\t\\begin{minipage}{\\columnwidth}\n\t\t\\centering\n\t\t\\begin{tabular}{l|c|c}\n\t\t\\hline\nAttributes & Attacker $\\mathcal{A}$'s input & Defender $\\mathcal{I}$'s input \\\\\\hline\\hline\n\\multicolumn{3}{l}{Attributes under $\\mathcal{I}$'s control but may be known to $\\mathcal{A}$ to some extent} \\\\\\hline\n$A_1$: Training set $D_{train}$ & $a_1\\in [0,1]$ & 1 \\\\\\hline\n$A_2$: Defense technique & $a_2\\in \\{0,1\\}$ & 1 \\\\\\hline\n$A_3$: Feature set $S$ & $a_3\\in [0,1]$ & 1\\\\\\hline\n$A_4$: Learning algorithm $F_\\theta$ & $a_4\\in [0,1]$ & 1 \\\\\\hline\n$A_5$: Response & $a_5\\in \\{0,1\\}$ & 1 \\\\\\hline\\hline\n\\multicolumn{3}{l}{Attributes under $\\mathcal{A}$'s control but may be known to $\\mathcal{I}$ to some extent} \\\\\\hline\n$A_6$: Manipulation set $\\mathcal{M}$ & 1 & $a_6\\in [0,1]$ \\\\\\hline\n$A_7$: Attack tactic & 1 & $a_7\\in \\{0,1\\}$ \\\\\\hline\n$A_8$: Attack technique & 1& $a_8\\in \\{0,1\\}$ \\\\\\hline\n$A_9$: Adversarial examples & 1 & $a_9\\in [0,1]$ \\\\\\hline\n\t\t\\end{tabular}\n\t\\end{minipage}\n\t\\label{tab:attr}\n\\end{table}\n\\smallskip\n\\noindent{\\bf Attacker's Input}.\nTable \\ref{tab:attr} highlights the attributes we define to describe $\\mathcal{A}$'s input, including: five attributes $A_1,\\ldots,A_5$ that are under $\\mathcal{I}$'s control (indicated by 1) but may be known to $\\mathcal{A}$ at some extent $a_1,\\ldots,a_5$, respectively; and four attributes $A_6,\\ldots,A_9$ that are under $\\mathcal{A}$'s control (indicated by 1). These attributes are elaborated below.\n({\\bf i}) $A_1$: it describes $\\mathcal{I}$'s training set $D_{train}$ for learning classifier $f$. \nWe use $a_1\\in [0,1]$ to represent the extent at which $D_{train}$ is known to $\\mathcal{A}$.\nLet $\\hat{D}_{train}$ be the training files that are known to $\\mathcal{A}$. Then, $a_1=|\\hat{D}_{train}\\cap{D}_{train}|/|{D}_{train}|$.\n({\\bf ii}) $A_2$: it describes $\\mathcal{I}$'s techniques, which can be Ensemble Learning (${\\sf EL}$), Weight Regularization (${\\sf WR}$), Adversarial Training (${\\sf AT}$), Verifiable Learning (${\\sf VL}$), Robust Feature ({\\sf RF}), Input Transformation (${\\sf IT}$), Classifier ranDomization (${\\sf CD}$), Sanitizing Examples ({\\sf SE}). Let $A_2\\in\\{${\\sf EL},{\\sf WR},{\\sf AT},{\\sf VL},{\\sf RF},{\\sf IT},{\\sf CD},{\\sf SE}$\\}$ and $a_2\\in \\{0,1\\}$ such that $a_2=0$ means $\\mathcal{A}$ does not know $\\mathcal{I}$'s techniques and $a_2=1$ means $\\mathcal{A}$ knows $\\mathcal{I}$'s technique. The techniques are defined as follows.\nDefinition \\ref{definition:el} says that $\\mathcal{I}$ constructs multiple classifiers and uses them collectively in malware detection.\n\\begin{definition} [ensemble learning or {\\sf EL} ] \\label{definition:el}\nLet $\\mathcal{H}$ be $\\mathcal{I}$'s classifier space.\nGiven $K$ classifiers $\\{f_i\\}_{i=1}^{K}$ where $f_i\\in\\mathcal{H}$ and $f_i:\\mathcal{Z}\\to[0,1]$, let $f_i$ be assigned with weight $\\omega_i$ with $\\sum_{i=1}^K{\\omega_i}=1$ and $\\omega_i\\geq 0$. Then, $f=\\sum_{i=1}^{K}\\omega_i f_i$.\n\\end{definition}\nDefinition \\ref{definition:wr} says that $\\mathcal{I}$ uses regularization (e.g., $\\ell_2$ regularization or {\\em dropout} ) to decrease model's sensitivity to adversarial examples.\n\\begin{definition} [weight regularization or {\\sf WR} ] \\label{definition:wr}\nGiven a regularization item $\\Omega$ (e.g., constraints imposed on the learnable parameters), the empirical risk is\n$\\min \\limits_{\\theta}~\\left[\\mathcal{L}(\\theta, D_{train})+\\Omega(\\theta)\\right]$, where $\\mathcal{L}$ is defined in Eq. \\eqref{eq:rsk_emp}.\n\\end{definition}\nDefinition \\ref{definition:at} says that $\\mathcal{I}$ proactively makes its classifier $f$ perceive some information about adversarial files. That is, $\\mathcal{I}$ augments the training set by incorporating adversarial examples that may be produced by $\\mathcal{I}$, $\\mathcal{A}$, or both.\n\\begin{definition} [adversarial training or {\\sf AT} ] \\label{definition:at}\nLet $D'$ denote a set of adversarial file-label pairs.\nThen, $\\mathcal{I}$ tunes model parameters by minimizing the empirical risk: $\\min \\limits_{\\theta}~\\left[\\mathcal{L}(\\theta, D_{train}) + \\beta \\mathcal{L}(\\theta, D')\\right]$, where $\\beta\\geq 0$ denotes a balance factor.\n\\end{definition}\nDefinition \\ref{definiton:vl} says that $\\mathcal{I}$ intentionally over-estimates the error incurred by $\\mathcal{A}$'s manipulations and then minimizes it.\n\\begin{definition}[verifiable learning or {\\sf VL} ] \\label{definiton:vl}\nGiven $(z,y)\\in D_{train}$ and a manipulation set $\\hat{\\mathcal{M}}$ known by $\\mathcal{I}$, let $z(\\hat{\\mathcal{M}})$ denote the upper and lower boundaries on $\\hat{\\mathcal{M}}$. Then, this defense technique minimizes the following loss function derived from Eq.\\eqref{eq:rsk_emp}:\n$L(F_\\theta(\\phi_c(S,z)), y)+\\beta L(F_\\theta(\\phi_c(S,z(\\hat{\\mathcal{M}}))), y).$\n\\end{definition}\nDefinition \\ref{eq:robust-feature-defense-technique} says that $\\mathcal{I}$ uses a set of features $S^*\\subseteq S$ that can lead to higher detection capability against adversarial example attacks.\n\\begin{definition} [robust feature or {\\sf RF}; adapted from ]\n\\label{eq:robust-feature-defense-technique}\nGiven a training set $D_{train} \\cup D'$ that contains (adversarial) file-label pairs, the set of robust feature set $S^*$ is $$S^*=\\argmin_{\\tilde{S}\\subset S}\\sum_{(z,y)\\in {D_{train}\\cup D'}}L(\\widetilde{F}_\\theta(\\phi_c(\\tilde{S},z)),y),$$ \nwhere $\\widetilde{F}_\\theta$ is $F_\\theta$ or a simplified learning algorithm that is computationally faster than $F_\\theta$ .\n\\end{definition}\nDefinition \\ref{def:it} says that $\\mathcal{I}$ aims to use non-learning methods (e.g., de-obfuscation as shown in Proguard ) to offset $\\mathcal{A}$'s manipulations.\n\\begin{definition} [input transformation or {\\sf IT}, adapted from ] \\label{def:it}\nLet ${\\tt IT}:\\mathcal{Z}\\to\\mathcal{Z}$ denote an input transformation in the file space. Given file $z$ and transformation ${\\tt IT}$, the classifier is $f=\\varphi(\\phi({\\tt IT}(z)))$.\n\\end{definition}\nDefinition \\ref{definition:cd} says that $\\mathcal{I}$ randomly chooses $m$ classifiers and uses their results for prediction. That is, $\\mathcal{I}$ aims to randomize the feature representation used by $f$, the learning algorithm, and/or response to $\\mathcal{A}$'s queries (to prevent $\\mathcal{A}$ from inferring information about $f$). \n\\begin{definition} [classifier randomization or {\\sf CD}; adapted from ] \\label{definition:cd}\nGiven $\\mathcal{I}$'s classifier space $\\mathcal{H}$ and an input file $z$, $\\mathcal{I}$ randomly selects $m$ classifiers from $\\mathcal{H}$ with replacement, say $\\{f_i\\}_{i=1}^m$. Then,\n\t$f=\\frac{1}{m}\\sum_{i=1}^m~f_i(z)$. \n\\end{definition}\nInstead of enhancing malware detectors, Definition \\ref{definition:se} provides an alternative that detects the adversarial examples for further analysis.\n\\begin{definition}[sanitizing examples or ${\\sf SE}$; adapted from ] \\label{definition:se}\n$\\mathcal{I}$ aims to detect adversarial files by using function \n${\\sf flag}:\\mathcal{Z}\\to \\{{\\tt yes},{\\tt no}\\}$ to flag a file as adversarial ({\\tt yes}) or not ({\\tt no}).\n\\end{definition}\n({\\bf iii}) $A_3$: it describes $\\mathcal{I}$'s feature set $S$. We use $a_3\\in [0,1]$ to represent the extent at which $\\mathcal{A}$ knows about $S$. Let $\\hat{S}$\ndenote the features that are known to $\\mathcal{A}$. Then,\n$a_3=|\\hat{S}\\cap S|/|S|$.\n({\\bf iv}) $A_4$: it describes $\\mathcal{I}$'s learning algorithm $F_\\theta$, the set of trainable parameters $\\theta$, and hyperparameters (which are set manually, e.g., $\\beta$ in Definition \\ref{definition:at})\n. We use $a_4\\in [0,1]$ to represent that $\\mathcal{A}$ knows an $a_4$ degree about $A_4$, where $a_4=0$ means $\\mathcal{A}$ knows nothing and $a_4=1$ means $\\mathcal{A}$ knows everything.\n({\\bf v}) $A_5$: it describes $\\mathcal{I}$'s response to $\\mathcal{A}$'s query to $f$ (if applicable), which is relevant because $\\mathcal{A}$ can learn useful information about $f$ by observing $f$'s responses .\nWe define $a_5\\in \\{0,1\\}$ such that $a_5=0$ means there is a limit on the response that can be made by $\\mathcal{A}$ to $f$\n(referred as ${\\sf LQ}$) and $a_5=1$ means there is no limit (referred as ${\\sf FQ}$).\n({\\bf vi}) $A_6$: it describes $\\mathcal{A}$'s manipulation set in the problem space, which describes perturbations for generating adversarial files (adapted from {\\em perturbation set} in the AML literature ):\n\\begin{eqnarray*}\n\\mathcal{M}=\\{\\delta: (z'\\leftarrow \\mathcal{A}(z,\\delta)) \\wedge ({\\tt true}\\leftarrow \\mathcal{O}(z,z')\n) \\wedge (z \\in \\mathcal{Z}) \\wedge (z'\\neq z)\\}.\n\\end{eqnarray*}\n$\\mathcal{M}$ is application-specific. For instance, an Android Package Kit (APK) permits adding codes or renaming class names , a Windows Portable Executable (PE) permits adding codes or changing PE {\\em section} names , and a Portable Document Format (PDF) file permits appending dead-code at its end or add new instructions . This means that a perturbation $\\delta\\in\\mathcal{M}$ can be a tuple specifying an operator (e.g., addition or removal), an object (e.g., a feature used by $\\mathcal{I}$), and other kinds of information (e.g., perturbation location in a file). \nSince it is often infeasible to enumerate the entire manipulation set, $\\mathcal{A}$ may leverage an empirical one $\\widetilde{\\mathcal{M}}$ , which can be defined in the problem or feature space. Manipulations in the problem space must not violate the relevant constraints (e.g., adding APIs in an APK should not cause the request of unauthorized permissions). Manipulations in the feature space facilitate efficient computing via gradient-based methods as long as the inverse feature mapping $\\phi^{-1}$ is available.\nFurthermore, we can use the manipulation set $\\mathcal{M}$ to define a {\\em feature manipulation set} $\\mathbf{M}$:\n\\begin{eqnarray}\n\\mathbf{M}=\\{\\delta_{\\mathbf{x}}=\\mathbf{x}'-\\mathbf{x}: (\\mathbf{x}=\\phi(z)) \\land (\\mathbf{x}'=\\phi(z')) \\land (z'\\leftarrow \\mathcal{A}(z,\\delta)) \\land (\\delta \\in \\mathcal{M}) \\land (z \\in \\mathcal{Z})\n\\}. \\label{eq:manipulations}\n\\end{eqnarray}\nIn order to compute $\\mathbf{M}$ efficiently, one strategy is to estimate a feature-space analog of $\\widetilde{\\mathcal{M}}$, denoted by $\\widetilde{\\mathbf{M}}$ . This however demands resolving the invertibility Assumption \\ref{assumption:inverse}.\n({\\bf vii}) $A_7$: it describes $\\mathcal{A}$'s attack tactics. We consider two tactics: classifier {\\em evasion} and classifier {\\em poisoning}. For the evasion attack, we consider three variants: basic evasion (${\\sf BE}$), optimal evasion 1 (${\\sf OE1}$) and optimal evasion 2 (${\\sf OE2}$).\nFor the poisoning attack, we consider two variants: basic poisoning (${\\sf BP}$) and optimal poisoning (${\\sf OP}$). Correspondingly, we have $A_7\\in\\{{\\sf BE},{\\sf OE1},{\\sf OE 2},{\\sf BP},{\\sf OP}\\}$.\nThese tactics are elaborated below, while noting that they do not explicitly call oracle $\\mathcal{O}$ because definitions of manipulation sets $\\mathcal{M}$ already assure that manipulations preserve functionalities of non-adversarial files.\nAs shown in Definition \\ref{defintion:basic_evs}, the basic evasion attack is that $\\mathcal{A}$ uses a set of perturbations $\\delta_z\\subseteq\\mathcal{M}$ to manipulate a malicious file $z$, which is classified by $\\mathcal{I}$'s classifier $f$ as $+\\leftarrow f(z)$, to an adversarial file $z'$ such that $-\\leftarrow f(z')$.\n\\begin{definition}[basic evasion or ${\\sf BE}$ ] \\label{defintion:basic_evs}\n$\\mathcal{A}$ looks for $\\delta_z\\subseteq\\mathcal{M}$ to achieve the following for $z\\in \\mathcal{Z}$ with $+\\leftarrow f(z)$:\n\\begin{eqnarray*}\n-\\leftarrow f(z') ~\\text{where}~\n(z'\\leftarrow \\mathcal{A}(z,\\delta_z)) \\wedge\n(\\delta_z\\subseteq\\mathcal{M}).\n\\end{eqnarray*}\n\\end{definition}\nAs shown in Definition \\ref{def:opt_evasion1}, the attacker attempts to minimize the degree of perturbations. In other words, this attack tactic is the same as {\\sf BE}, except that $\\mathcal{A}$ attempts to minimize the manipulation when perturbing a non-adversarial file $z\\in \\mathcal{Z}$ into an adversarial file $z'\\in \\mathcal{Z}$.\n\\begin{definition}[optimal evasion 1 or ${\\sf OE1}$; adapted from ] \\label{def:opt_evasion1}\n$\\mathcal{A}$ attempts to achieve the following for $z\\in \\mathcal{Z}$ with $+\\leftarrow f(z)$:\n\\begin{eqnarray*}\n\\min \\limits_{z'}\n\\Gamma(z', z) \n~\\text{s.t.}~(z'\\leftarrow \\mathcal{A}(z,\\delta_z))\\land (\\delta_z\\subseteq\\mathcal{M}) \\land (-\\leftarrow f(z')). \n\\end{eqnarray*}\n\\end{definition}\nAs shown in Definition \\ref{def:opt_evasion2}, the attacker attempts to maximize $\\mathcal{I}$'s loss for waging high-confidence evasion attacks, while noting the small perturbations may be incorporated.\n\\begin{definition}[optimal evasion 2 or ${\\sf OE2}$; adapted from ] \\label{def:opt_evasion2}\n$\\mathcal{A}$ attempts to achieve the following for $z\\in \\mathcal{Z}$ with $+\\leftarrow f(z)$:\n\\begin{eqnarray*}\n\\max\\limits_{z'}\nL(F_\\theta(\\phi_c(S,z')),+)\n~\\text{s.t.}~(z'\\leftarrow \\mathcal{A}(z,\\delta_z))\\land (\\delta_z\\subseteq\\mathcal{M}) \\land (-\\leftarrow f(z')).\n\\end{eqnarray*}\n\\end{definition}\nLet $D'_{poison}\\subset \\mathcal{D}$ be a set of adversarial file-label pairs obtained by manipulating non-adversarial files in $D_{poison}$. Let $D'_{train}\\leftarrow D_{train} \\cup D'_{poison}$ be the contaminated training data for learning a classifier $f'$ with parameters $\\theta'$. As shown in Definition \\ref{def:basic poisoning attack}, the basic poisoning attack is that the attacker aims to make $f'$ mis-classify the files in a dataset $D_{target}$, while accommodating the attacks that $\\mathcal{A}$ manipulates labels of the files in $D_{poison}$ .\n\\begin{definition}[basic poisoning or ${\\sf BP}$ ]\n\\label{def:basic poisoning attack}\nGiven a set $D_{target}$ of files where $+\\leftarrow f(\\dot{z})$ for $\\dot{z}\\in D_{target}$ and a set $D_{poison}$ of non-adversarial files, $\\mathcal{A}$ attempts to perturb files in $D_{poison}$ to adversarial ones\n$D'_{poison}=\n\\{(\\mathcal{A}(z,\\delta_z), \\mathcal{A}(y)): ((z,y)\\in D_{poison})\\land (\\delta_z\\subseteq \\mathcal{M})\\land $ $ (\\mathcal{A}(y)\\in\\{+,-\\}) \\}$ such that classifier\n$f'$ learned from $D'_{train}\\leftarrow D_{train} \\cup D'_{poison}$\nmis-classifies the files in $D_{target}$.\nFormally, the attacker intents to achieve the following for $\\forall~\\dot{z} \\in D_{target}$: \n$-\\leftarrow f'(\\dot{z})$ where $f'$ is learned from $D'_{train}\\leftarrow D_{train}\\cup D'_{poison}$.\n\\end{definition}\nAs shown in Definition \\ref{def:optimal poisoning attack}, the optimal poisoning attack is the same as {\\sf BF}, except that $\\mathcal{A}$ attempts to\nmaximize the loss when using classifier $f'$ with parameter $\\theta'$ to classify\nfiles in $D_{target}$. \nDefinition \\ref{def:optimal poisoning attack} can have multiple variants by considering bounds on $|D'_{poison}|$ or bounds on the degree of perturbations $|\\delta_z|$ . \n\\begin{definition}[optimal poisoning or ${\\sf OP}$ ]\n\\label{def:optimal poisoning attack}\nGiven $D_{poison}$, $\\mathcal{A}$ perturbs $D_{poison}$ into $D'_{poison}$ for achieving:\n\\begin{align*}\n\\max \\limits_{D'_{poison}}~\\mathcal{L}(\\theta', D_{target})\n~\\text{where}~{\\theta'} \\leftarrow \\operatorname*{\\arg\\min}_{\\theta} \\mathcal{L}(\\theta, D_{train} \\cup D'_{poison}).\n\\end{align*}\n\\end{definition}\n{({\\bf viii})} $A_8$: it describes $\\mathcal{A}$'s attack techniques, such as Gradient-based Optimization ({\\sf GO}), Sensitive Features ({\\sf SF}), MImicry ({\\sf MI}), TRansferability ({\\sf TR}), Heuristic Search ({\\sf HS}), Generative Model ({\\sf GM}), and Mixture Strategy ({\\sf MS}). We denote this by $A_8\\in\\{${\\sf GO}, {\\sf SF}, {\\sf MI}, {\\sf TR}, {\\sf HS}, {\\sf GM}, {\\sf MS}$\\}$. Let $\\mathcal{A}$ have a classifier $\\hat{f}$, which consists of a hand-crafted feature extraction $\\hat{\\phi}_c$ and a parameterized model $\\hat{F}_{\\hat\\theta}$. Let $\\mathcal{A}$ also have an objective function $L_\\mathcal{A}:[0,1]\\times\\mathcal{Y}\\to\\mathcal{R}$, which measures $\\hat{f}$'s error or $\\mathcal{A}$' failure in evasion . Note that $\\hat{f}$ and $L_\\mathcal{A}$ can be the same as, or can mimic (by leveraging $\\mathcal{A}$'s knowledge about $\\mathcal{I}$'s attributes $A_1,\\ldots,A_5$), $\\mathcal{I}$'s classifier $f$ and loss function $L$, respectively. \nThe attack technique specified by Definition \\ref{definition:go} is that $\\mathcal{A}$ solves the feature-space optimization problems described in Definitions \\ref{def:opt_evasion1}, \\ref{def:opt_evasion2} and \\ref{def:optimal poisoning attack} by using some gradient-based optimization method and then leverages the invertibility Assumption \\ref{assumption:inverse} to generate adversarial malware examples.\n\\begin{definition}[Gradient-based Optimization or {\\sf GO}, adapted from ] \\label{definition:go}\nLet $\\mathbf{x}\\leftarrow\\hat{\\phi}_c(\\hat{S},z)$ and $\\mathbf{x}'\\leftarrow\\mathbf{x}+\\delta_\\mathbf{x}$. The feature-space optimization problem in Definition \\ref{def:opt_evasion1} can be written as\n\t\\begin{equation}\n\t\t\\min_{\\delta_\\mathbf{x}} C(\\mathbf{x}, \\mathbf{x}+\\delta_\\mathbf{x})~~~\\text{s.t.}~~~ (\\delta_\\mathbf{x}\\in[\\underline{\\mathbf{u}},\\overline{\\mathbf{u}}]) \\land (\\hat{F}_{\\hat{\\theta}}(\\mathbf{x}')<\\tau), \\label{eq:a8:evs1}\n\t\\end{equation}\nwhere $\\underline{\\mathbf{u}}$ and $\\overline{\\mathbf{u}}$ are respectively the lower and upper bounds on $\\mathbf{M}$ (e.g., $\\delta_\\mathbf{x}\\in[-\\mathbf{x}, 1-\\mathbf{x}]$ for binary representation $\\mathbf{x}$). \nThe feature-space optimization problem in Definition \\ref{def:opt_evasion2} can be written as \n \\begin{equation}\n \t\\max_{\\delta_\\mathbf{x}} L_\\mathcal{A}\\left(\\hat{F}_{\\hat{\\theta}}(\\mathbf{x}+\\delta_\\mathbf{x}), +\\right)~~\\text{s.t.}~~(\\delta_\\mathbf{x}\\in[\\underline{\\mathbf{u}},\\overline{\\mathbf{u}}]). \\label{eq:a8:evs2}\n \\end{equation}\nThe feature-space optimization problem specified in Definition \\ref{def:optimal poisoning attack} can be written as \n \\begin{align}\n \t&\\max_{\\delta_\\mathbf{x}\\in[\\underline{\\mathbf{u}},\\overline{\\mathbf{u}}]}\\mathbb{E}_{(\\dot{z},\\dot{y})\\in D_{target}}L_\\mathcal{A}(\\hat{F}_{\\hat{\\theta}'}(\\hat{\\phi}_c(\\hat{S},\\dot{z}),\\dot{y})),~~ \\forall (z,y)\\in D_{poison} \\label{eq:a8:poi}\\\\\n \t~~\\text{where}~~&{\\hat\\theta}'\\leftarrow\\argmin_{\\hat\\theta}\\mathbb{E}_{(z_t,y_t)\\in{\\hat{D}_{train}\\cup\\{(\\hat{\\phi}_c^{-1}(\\hat{\\phi}_c(z)+\\delta_\\mathbf{x}),y')\\}}}L_\\mathcal{A}\\left(\\hat{F}_{\\hat\\theta}(\\hat{\\phi}_c(\\hat{S},z_t),y_t)\\right). \\nonumber\n \\end{align}\n\\end{definition}\nIn order to calculate the gradients of loss function $L_\\mathcal{A}$ with respect to $\\delta_\\mathbf{x}$ in Eqs.\\eqref{eq:a8:evs1} and \\eqref{eq:a8:evs2}, inequality constraints can be handled by appending penalty items to the loss function in question and box-constraints can be coped with by using gradient projection . Since $\\mathbf{x}+\\delta_\\mathbf{x}$ is continuous, the {\\sf GO} attack technique needs to map $\\delta_\\mathbf{x}$ to a discrete perturbation vector in $\\mathbf{M}$, for instance by using the nearest neighbor search .\nThe gradients of loss function $L_\\mathcal{A}$ with respect to $\\delta_\\mathbf{x}$ in Eq. \\eqref{eq:a8:poi} are delicate to deal with. One issue is the indirect relation between $L_\\mathcal{A}$ and $\\delta_\\mathbf{x}$, which can be handled by the chain rule . \nAnother issue is the difficulty that is encountered when computing the partial derivatives $\\partial{\\hat\\theta'}/\\partial\\delta_{\\mathbf{x}}$ . \nFor dealing with this, researchers often relax the underlying constraints (e.g., by supposing that $\\hat{F}_{\\hat\\theta}$ is a linear model). \nThe attack technique specified by Definition \\ref{definition:sf} is that $\\mathcal{A}$ perturbs malware examples by injecting or removing a small number of features to decrease the classification error measured by the loss function ${L}_\\mathcal{A}$ as much as possible.\n\\begin{definition}[Sensitive Features or {\\sf SF}, adapted from ] \\label{definition:sf}\nFor evasion attacks, $\\mathcal{A}$ aims to maximize the following with respect to a given malware example-label pair $(z,+)$: $$\\max_{z'} L_\\mathcal{A}\\left(\\hat{F}_{\\hat\\theta}(\\hat{\\phi}(\\hat{S},z')),+\\right)~~\\text{s.t.}~~(z'\\leftarrow \\mathcal{A}(z,\\delta_z)) \\land (\\delta_z\\subseteq \\mathcal{M}) \\land (|\\delta_z|\\leq m),$$ \nwhere $m$ is the maximum degree of manipulations.\nFor poisoning attacks, $\\mathcal{A}$ aims to maximize the following with respect to the given $D_{poison}$ and $D_{target}$, \n\\begin{align}\n \\max_{D'_{poison}}\\mathbb{E}_{(\\dot{z},\\dot{y})\\in D_{target}}L_\\mathcal{A}(\\hat{F}_{\\hat{\\theta}'}(\\hat{\\phi}_c(\\hat{S},\\dot{z}),\\dot{y})),\n\\end{align}\nwhere $\\hat{\\theta}'$ is learned from $\\hat{D}_{train}\\cup D'_{poison}$ such that $\\forall z'\\in D'_{poison}$ is obtained via the perturbation $\\delta_z$ with obeying $(z'\\leftarrow \\mathcal{A}(z,\\delta_z)) \\land (\\delta_z\\subseteq \\mathcal{M}) \\land (z\\in D_{poison}) \\land (|\\delta_z|\\leq m)$.\n\\end{definition} \nThe attack technique specified by Definition \\ref{definition:mi} is that $\\mathcal{A}$ perturbs malware example $z$ by mimicking a benign example, while noting that this attack technique can be algorithm-agnostic.\n\\begin{definition}[MImicry or {\\sf MI}, adapted from ] \\label{definition:mi}\nGiven a set of benign examples $D_{ben}$ and a malware example $z$, $\\mathcal{A}$ aims to achieve the following minimization:\n\t\\begin{eqnarray}\n\t\t\\min_{\\delta_z\\in\\mathcal{M}} \\Gamma(\\mathcal{A}(z,\\delta_z),z_{ben})~~\\text{s.t.}~~ (\\exists~z_{ben}\\in D_{ben}). \\label{eq:mimicry}\n\t\\end{eqnarray}\n\\end{definition}\nThe attack technique specified by Definition \\ref{definition:mi} can be extended to accommodate the similarity between representations in the feature space .\nThe attack technique specified by Definition \\ref{definition:tr} is that $\\mathcal{A}$ generates adversarial examples against a surrogate model $\\hat{f}$.\n\\begin{definition}[TRansferability or {\\sf TR}, adapted from ] \\label{definition:tr}\n$\\mathcal{A}$ learns a surrogate model $\\hat{f}$ of $f$ from $\\hat{D}_{train}$, $\\hat{S}$, $\\hat{\\phi_c}$ and $\\hat{F}$. For evasion attacks,\n$\\mathcal{A}$ achieves $-\\leftarrow\\hat{f}(z')$ by perturbing malware example $z$ to $z'$ and then attacks $f$ with $z'$. For poisoning attacks, $\\mathcal{A}$ contaminates $\\hat{f}$ to $\\hat{f}'$ such that $-\\leftarrow\\hat{f}'(\\dot{z})~\\text{for}~\\forall(\\dot{z},\\dot{y}) \\in D_{target}$, by poisoning the training set $\\hat{D}_{train}$ with $D'_{poison}$ and the attacks $f$ with $D'_{poison}$.\n\\end{definition}\nThe attack technique specified by\nDefinition \\ref{definition:hs} is that $\\mathcal{A}$ searches perturbations in $\\mathcal{M}$ via some heuristics, while leveraging oracle $\\mathcal{O}$'s responses to $\\mathcal{A}$'s queries and $f$'s responses to $\\mathcal{A}$'s queries. Since $\\mathcal{M}$ is defined with respect to the problem space, this attack technique does not need the invertibility Assumption \\ref{assumption:inverse}.\n\\begin{definition}[Heuristic Search or {\\sf HS}] \\label{definition:hs}\nLet $h$ be a function taking $\\mathcal{O}$'s response and $f$'s response as input. Given a malware example $z$, $\\mathcal{A}$ looks for an $m$-length manipulation path \n\\begin{align}\n\\langle z_{(0)}, z_{(1)},\\ldots,z_{(m)}\\rangle \n~~\\text{s.t.}~~z_{(i+1)}=\\mathcal{A}(z_{(i)},\\delta_{z,(i)}) \\land (\\delta_{z,(i)}& \\in \\mathcal{M}) \\land (h(\\mathcal{O}, f,z,z_{(i)})\\leq h(\\mathcal{O}, f,z,z_{(i+1)})) \\nonumber\n\\end{align}\nwhere $z_{(0)}=z$.\n\\end{definition}\nThe attack technique specified by Definition \\ref{def:generative-model} is that $\\mathcal{A}$ uses a generative model $G$ with parameters $\\theta_g$ to perturb malware representation vectors and then leverages the invertibility Assumption \\ref{assumption:inverse} to turn the perturbed vector into an adversarial malware example.\n\\begin{definition}[Generative Model or {\\sf GM}]\\label{def:generative-model}\nGiven a malware representation vector $\\mathbf{x}=\\hat{\\phi}_c(\\hat{S},z)$, $\\mathcal{A}$ achieves\n$$\\max_{\\theta_g} L_\\mathcal{A}\\left(\\hat{F}_{\\hat{\\theta}}(G_{\\theta_g}(\\mathbf{x})),+\\right)~~\\text{s.t.}~~ G_{\\theta_g}(\\mathbf{x})\\in [\\underline{\\mathbf{u}}-\\mathbf{x}, \\overline{\\mathbf{u}}-\\mathbf{x}]$$\nand leverages the invertibility Assumption \\ref{assumption:inverse} to obtain an adversarial example $z'=\\hat\\phi_c^{-1}(G_{\\theta_g}(\\mathbf{x}))$.\n\\end{definition}\nThe attack technique specified by\nDefinition \\ref{definition:ms} is that $\\mathcal{A}$ combines multiple perturbation methods to perturb an example. \n\\begin{definition}[Mixture Strategy or {\\sf MS} ] \\label{definition:ms}\nLet $\\mathcal{H}_A$ denote the space of generative methods and $\\mathcal{W}_a=\\{\\mathbf{w}_a:\\mathbf{w}_a=(w_{a,1},\\ldots,w_{a,K}),w_{a,i}\\geq 0\\}$ with $i=1,\\ldots,K$ denote the weights space. Given a malware example $z$, $\\mathcal{A}$ aims to achieve $$\\max_{\\mathbf{w}_a} L_\\mathcal{A}(\\hat{F}_{\\hat{\\theta}}(\\hat\\phi(z')),+) ~~\\text{s.t.}~~(z'=\\sum_{i=1}^K{w_{a,i}}g_i(z))\\land (\\mathcal{O}(z,z')={\\sf true})) \\land (g_i\\in \\mathcal{H}_A) \\land (\\mathbf{w}_a\\in \\mathcal{W}_a).$$\n\\end{definition}\n({\\bf ix}) $A_9$: it corresponds to $\\mathcal{A}$'s adversarial files. Given file manipulation set $\\mathcal{M}$, the corresponding set of adversarial files is defined as $\\mathcal{Z}_{\\mathcal M}=\\{\\mathcal{A}(z, \\delta_z):(z \\in \\mathcal{Z}) \\land (\\delta_z\\subseteq\\mathcal{M})\\}$. Given feature manipulation set $\\mathbf{M}$, the set of adversarial feature vectors is: $\\mathcal{X}_{\\mathbf{M}}=\\{\\mathbf{x}':(\\mathbf{x}'=\\mathbf{x}+\\delta_{\\mathbf{x}}) \\land (\\delta_\\mathbf{x}\\in\\mathbf{M})\\}.$\n\\begin{figure}[htbp!]\n\t\\centering\n\t\\begin{tikzpicture}[thick,scale=0.95, every node/.style={transform shape}]\n\t\\node(top) [label=right:{White-box} attack, label=left:{}] at (0,6) {$(1,1,1,1,1)$};\n\t\\node(l5n0)[label=below:{},align=center] at (0,5) {$(a_1,a_2,1,1,a_5)$};\n\t\\node(l4n0)[label=below:{},align=center] at (2.2,4) {$(a_1,a_2,a_3,1,a_5)$};\n\t\\node(l4n1)[label=below:{},align=center] at (0,4) {$(0,0,1,1,0)$};\n\t\\node(l4n2)[label=below:{},align=center] at (-2.2,4) {$(a_1,a_2,1,a_4,a_5)$};\n\t\\node(l3n0)[label=below:{},align=center] at (4.4,3) {$(a_1,a_2,0,1,a_5)$};\n\t\\node(l3n1)[label=below:{},align=center] at (2.2,3) {$(0,0,a_3,1,0)$};\n\t\\node(l3n2)[label=below:{},align=center] at (0,3) {$(a_1,a_2,a_3,a_4,a_5)$};\n\t\\node(l3n3)[label=below:{},align=center] at (-2.2,3) {$(0,0,1,a_4,0)$};\n\t\\node(l3n4)[label=below:{},align=center] at (-4.4,3) {$(a_1,a_2,1,0,a_5)$};\n\t\\node(l2n0)[label=below:{},align=center] at (4.4,2) {$(0,0,0,1,0 )$};\n\t\\node(l2n1)[label=below:{},align=center] at (2.2,2) {$(a_1,a_2,0,a_4,a_5)$};\n\t\\node(l2n2)[label=below:{},align=center] at (0,2) {$(0,0,a_3,a_4,0)$};\n\t\\node(l2n3)[label=below:{},align=center] at (-2.2,2) {$(a_1,a_2,a_3,0,a_5)$};\n\t\\node(l2n4)[label=below:{},align=center] at (-4.4,2) {$(0,0,1,0,0)$};\n\t\\node(l1n0)[label=below:{},align=center] at (2.2,1) {$(0,0,0,a_4,0 )$};\n\t\\node(l1n1)[label=below:{},align=center] at (0,1) {$(a_1,a_2,0,0,a_5)$};\n\t\\node(l1n2)[label=below:{},align=center] at (-2.2,1) {$(0,0,a_3,0,0 )$};\n\t\\node(bot)[label=right:{Black-box attack}, label=left:{}] at (0,0) {$ (0,0,0,0,0)$};\n\t\\draw(l5n0) -- (top);\n\t\\draw(l4n0) -- (l5n0);\n\t\\draw(l4n1) -- (l5n0);\n\t\\draw(l4n2) -- (l5n0);\n\t\\draw(l3n0) -- (l4n0);\n\t\\draw(l3n1) -- (l4n0);\n\t\\draw(l3n1) -- (l4n1);\n\t\\draw(l3n2) -- (l4n0);\n\t\\draw(l3n2) -- (l4n2);\n\t\\draw(l3n3) -- (l4n1);\n\t\\draw(l3n3) -- (l4n2);\n\t\\draw(l3n4) -- (l4n2);\n\t\\draw(l2n0) -- (l3n0);\n\t\\draw(l2n0) -- (l3n1);\n\t\\draw(l2n1) -- (l3n0);\n\t\\draw(l2n1) -- (l3n2);\n\t\\draw(l2n2) -- (l3n1);\n\t\\draw(l2n2) -- (l3n2);\n\t\\draw(l2n2) -- (l3n3);\n\t\\draw(l2n3) -- (l3n2);\n\t\\draw(l2n3) -- (l3n4);\n\t\\draw(l2n4) -- (l3n3);\n\t\\draw(l2n4) -- (l3n4);\n\t\\draw(l1n0) -- (l2n0);\n\t\\draw(l1n0) -- (l2n1);\n\t\\draw(l1n0) -- (l2n2);\n\t\\draw(l1n1) -- (l2n1);\n\t\\draw(l1n1) -- (l2n3);\n\t\\draw(l1n2) -- (l2n2);\n\t\\draw(l1n2) -- (l2n3);\n\t\\draw(l1n2) -- (l2n4);\n\t\\draw(bot) -- (l1n2);\n\t\\draw(bot) -- (l1n1);\n\t\\draw(bot) -- (l1n0);\n\t\\end{tikzpicture}\n\t\\caption{A portion of the partial order defined over $(a_1,\\ldots,a_5)$.\n\t} \n\t\\label{fig:attack_input}\n\\end{figure}\n\\smallskip\n\\noindent{\\bf On the Usefulness of the Preceding Specification}. The preceding specification can be applied to formulate a partial order in the attribute space, which allows to compare attacks unambiguously. Figure \\ref{fig:attack_input} depicts how vector $(a_1, \\cdots, a_5)$ formulates a partial order between the widely-used informal notions of {\\em black-box} attack, namely $(a_1,a_2,a_3,a_4,a_5)=(0,0,0,0,0)$, \nand {\\em white-box} attack, namely $(a_1,a_2,a_3,a_4,a_5)=(1,1,1,1,1)$;\nthere are many kinds of grey-box attacks in between.","id":"5f85312b-af7a-44ed-87ac-7e94a51dd8cc","level":"subsubsection","origin_cites_number":45,"parent_id":"8e32af6c-587d-43e1-aabc-dc5d8345628d","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Survey and Systematization Methodology"],["subsection","Framework"],["subsubsection","Systematizing Attacks"]],"subsections":[],"title":"Systematizing Attacks"},{"cite_extract_rate":0,"cites":[],"content":"Similarly, we systematize defenses from two perspectives: {\\em defender's objective} (i.e., what the defender aims to achieve) and {\\em defender's input} (i.e., what leverages the defender can use). \nWe also discuss how to apply the specification to formulate a partial-order structure for comparing defenses.\n\\smallskip\n\\noindent{\\bf Defender's Objectives}.\n$\\mathcal{I}$ aims to detect ideally all of the malicious files, adversarial and non-adversarial alike, while suffering from small side-effects (e.g., increasing false-positives).\n\\smallskip\n\\noindent{\\bf Defender's Input}.\nAs highlighted in Table \\ref{tab:attr}, $\\mathcal{I}$'s input includes attributes $A_1,\\ldots,A_5$, which are under $\\mathcal{I}$'s control, and the extent $a_6,\\ldots,a_9$ at which $\\mathcal{I}$ respectively knows about attributes $A_6,\\ldots,A_9$, which are under $\\mathcal{A}$'s control. Note that $A_1,\\ldots,A_9$ have been defined above.\n\\begin{itemize}\n\\item We define $a_6 \\in [0,1]$ to represent the extent at which $\\mathcal{I}$ knows $\\mathcal{A}$'s manipulation set $\\mathcal{M}$. Let $\\hat{\\mathcal{M}}\\subseteq \\mathcal{M}$ denote the subset of $\\mathcal{A}$'s manipulation set known to $\\mathcal{I}$. \nThen, we set $a_6=|\\hat{\\mathcal{M}}|/|\\mathcal{M}|$.\n\\item We define $a_7\\in\\{0,1\\}$ such that $a_7=0$ means $\\mathcal{I}$ does not know $\\mathcal{A}$'s attack tactic $A_7\\in \\{{\\sf BE},{\\sf OE1},{\\sf OE2},{\\sf BP},$ ${\\sf OP}\\}$ and $a_7=1$ means $\\mathcal{I}$ knows $\\mathcal{A}$'s tactic. \n\\item We define $a_8\\in\\{0,1\\}$ such that $a_8=1~(a_8=0)$ means the defender does (not) know $\\mathcal{I}$'s attack technique $A_8\\in\\{{\\sf GO}, {\\sf SF}, {\\sf TR}, {\\sf MI}, {\\sf HS}, {\\sf GM}, {\\sf MS}\\}$.\n\\item We use $a_9=|\\hat{\\mathcal{Z}}_{\\mathcal{M}}| / |\\mathcal{Z}_{\\mathcal{M}}|$ to represent the extent at which $\\mathcal{I}$ knows about $\\mathcal{A}$'s adversarial files, where $a_9\\in [0,1]$ and $\\mathcal{Z}_{\\mathcal{M}}$ is $\\mathcal{A}$'s adversarial files and $\\hat{\\mathcal{Z}}_{\\mathcal{M}}\\subseteq\\mathcal{Z}_{\\mathcal{M}}$ is known to $\\mathcal{I}$.\n\\end{itemize}\n\\smallskip\n\\noindent{\\bf On the Usefulness of the Preceding Specification}. \nSimilarly, the defense specification can be used to formulate a partial order in the attribute space, paving the way for comparing defenses unambiguously. Figure \\ref{fig:defense_input} depicts how vector $(a_6,\\ldots,a_9)$ formulates a partial order between the widely-used informal notions of black-box defense $(a_6,a_7,a_8,a_9)=(0,0,0,0)$ and white-box defense $(a_6,a_7,a_8,a_9)=(1,1,1,1,1)$; there are many kinds of grey-box defenses in between.\n\\begin{figure}[!htbp]\n\t\\centering\n\t\\begin{tikzpicture}[thick,scale=0.95, every node/.style={transform shape}]\n\t\t\\node(top)[label=right:{White-box defense}, label=left:{}] at (0,6) {$( 1,1,1,1)$};\n\t\t\\node(l3n0)[label=below:{},align=center] at (1.8,5) {$( a_6,a_7,a_8,1)$};\n\t\t\\node(l3n1)[label=below:{},align=center] at (0,5) {$( 1,0,0,1)$};\n\t\t\\node(l3n2)[label=below:{},align=center] at (-1.8,5) {$( 1,a_7,a_8,a_9)$};\n\t\t\\node(l2n0)[label=below:{},align=center] at (3.6,4) {$( 0,a_7,a_8, 1)$};\n\t\t\\node(l2n1)[label=below:{},align=center] at (1.85,4) {$( a_6,0,0,1)$};\n\t\t\\node(l2n2)[label=below:{},align=center] at (0,4) {$( a_6,a_7,a_8,a_9)$};\n\t\t\\node(l2n3)[label=below:{},align=center] at (-1.85,4) {$( 1,0,0,a_8)$};\n\t\t\\node(l2n4)[label=below:{},align=center] at (-3.6,4) {$( 1,a_7,a_8,0)$};\n\t\t\\node(l1n0)[label=below:{},align=center] at (3.6,3) {$( 0,0,0,1)$};\n\t\t\\node(l1n1)[label=below:{},align=center] at (1.85,3) {$( 0,a_7,a_8,a_9)$};\n\t\t\\node(l1n2)[label=below:{},align=center] at (0,3) {$( a_6,0,0,a_8)$};\n\t\t\\node(l1n3)[label=below:{},align=center] at (-1.85,3) {$( a_6,a_7,a_8,0)$};\n\t\t\\node(l1n4)[label=below:{},align=center] at (-3.6,3) {$( 1,0,0,0)$};\n\t\t\\node(l0n0)[label=below:{},align=center] at (1.85,2) {$( 0,0,0,a_9)$};\n\t\t\\node(l0n1)[label=below:{},align=center] at (0,2) {$( 0,a_7,a_8,0)$};\n\t\t\\node(l0n2)[label=below:{},align=center] at (-1.85,2) {$( a_6,0,0,0)$};\n\t\t\\node(bot)[label=right:{Black-box defense}, label=left:{}] at (0,1) {$( 0,0,0,0)$};\n\t\t\\draw(bot) -- (l0n0);\n\t\t\\draw(bot) -- (l0n1);\n\t\t\\draw(bot) -- (l0n2);\n\t\t\\draw(l0n0) -- (l1n0);\n\t\t\\draw(l0n0) -- (l1n1);\n\t\t\\draw(l0n0) -- (l1n2);\n\t\t\\draw(l0n1) -- (l1n1);\n\t\t\\draw(l0n1) -- (l1n3);\n\t\t\\draw(l0n2) -- (l1n2);\n\t\t\\draw(l0n2) -- (l1n3);\n\t\t\\draw(l0n2) -- (l1n4);\n\t\t\\draw(l1n0) -- (l2n0);\n\t\t\\draw(l1n0) -- (l2n1);\n\t\t\\draw(l1n1) -- (l2n0);\n\t\t\\draw(l1n1) -- (l2n2);\n\t\t\\draw(l1n2) -- (l2n1);\n\t\t\\draw(l1n2) -- (l2n2);\n\t\t\\draw(l1n2) -- (l2n3);\n\t\t\\draw(l1n3) -- (l2n2);\n\t\t\\draw(l1n3) -- (l2n4);\n\t\t\\draw(l1n4) -- (l2n3);\n\t\t\\draw(l1n4) -- (l2n4);\n\t\t\\draw(l2n0) -- (l3n0);\n\t\t\\draw(l2n1) -- (l3n0);\n\t\t\\draw(l2n1) -- (l3n1);\n\t\t\\draw(l2n2) -- (l3n0);\n\t\t\\draw(l2n2) -- (l3n2);\n\t\t\\draw(l2n3) -- (l3n1);\n\t\t\\draw(l2n3) -- (l3n2);\n\t\t\\draw(l2n4) -- (l3n2);\n\t\t\\draw(l3n0) -- (top);\n\t\t\\draw(l3n1) -- (top);\n\t\t\\draw(l3n2) -- (top);\n\t\\end{tikzpicture}\n\t\\caption{A portion of the partial order defined over $(a_6,\\ldots,a_9)$.\n\t}\n\t\\label{fig:defense_input}\n\\end{figure}","id":"1e0de86d-a9b2-453b-b977-3d5abab5bd99","level":"subsubsection","origin_cites_number":0,"parent_id":"8e32af6c-587d-43e1-aabc-dc5d8345628d","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Survey and Systematization Methodology"],["subsection","Framework"],["subsubsection","Systematizing Defenses"]],"subsections":[],"title":"Systematizing Defenses"},{"cite_extract_rate":0.8,"cites":[917,3866,3870,3854],"content":"\\label{sec:ml_model_property}\nSince $f=\\varphi(\\phi(\\cdot))$, we decompose $f$'s security properties into $\\varphi$'s and $\\phi$'s. We consider: Representation Robustness ({\\sf RR}), meaning that two similar files have similar feature representations; Classification Robustness ({\\sf CR}), meaning that two similar feature representations lead to the same label; Detection Robustness ({\\sf DR}), meaning that feature extraction function $\\phi$ returns similar representations for two files with the same functionality; Training Robustness ({\\sf TR}), meaning that small changes to the training set does not cause any significant change to the learned classifier. \nWith respect to small perturbations, Definitions \\ref{definition:representation-robustness} and\n\\ref{definition:classification-robustness} below collectively say that when two files $z$ and $z'$ are similar, they would be classified as the same label with a high probability. Since the classification function $\\varphi$ is linear, we can obtain a $\\epsilon$-robust $\\varphi$ analytically, where $\\epsilon$ is a small scalar that bounds the perturbations applied to feature vectors . This means that the main challenge is to achieve robust feature extraction.\n\\begin{definition}[{\\sf RR} or $(\\epsilon,\\eta)$-robust feature extraction; adapted from ] \n\\label{definition:representation-robustness}\nGiven constants $\\epsilon,\\eta\\in [0,1]$, and files $z,z' \\in \\mathcal{Z}$ such that $(\\Gamma(z,z')\\approx 0) \\land ({\\tt true}\\leftarrow \\mathcal{O}(z,z'))$,\nwe say feature extraction function $\\phi$ is $(\\epsilon,\\eta)$-robust if\n$$\\mathbb{P}(C(\\mathbf{x},\\mathbf{x}')\\leq \\epsilon)=\\mathbb{P}(C(\\phi(z),\\phi(z'))\\leq \\epsilon)>1 - \\eta.$$\n\\end{definition}\n\\begin{definition}[{\\sf CR} or {$\\epsilon$-robust classification} ]\n\\label{definition:classification-robustness} \nGiven constant $\\epsilon\\in [0,1]$ as in Definition \n\\ref{definition:representation-robustness} and\nany feature vectors $\\mathbf{x},{\\bf x}'\\in \\mathcal{X}$, we say classification function $\\varphi$ is $\\epsilon$-robust if $$(C(\\mathbf{x},\\mathbf{x}') \\leq \\epsilon) \\rightarrow ((\\varphi({\\bf x})>\\tau)\\land (\\varphi({\\bf x}')>\\tau)).$$\n\\end{definition}\nDefinition \\ref{definition:representation-robustness2} \nspecifies detection robustness, which says that feature extraction function $\\phi$ returns similar representations for two different files as long as they have the same functionality.\nNote that Definitions \\ref{definition:classification-robustness} and\n\\ref{definition:representation-robustness2} collectively produce a malware detector with detection robustness.\n\\begin{definition}[{\\sf DR} or $(\\mathcal{O},\\eta)$-robust feature extraction; adapted from ] \n\\label{definition:representation-robustness2}\nGiven constant $\\eta\\in [0,1]$ and two files $z,z' \\in \\mathcal{Z}$ such that $(\\Gamma(z,z')>>0)\\land({\\tt true}\\leftarrow \\mathcal{O}(z,z'))$, we say feature extraction $\\phi$ is $(\\mathcal{O}, \\eta)$-robust if $\\mathbb{P}(C(\\phi(z),\\phi(z'))\\leq \\epsilon)>1 - \\eta.$\n\\end{definition}\nSuppose we impose a restriction on the adversarial files set $D'_{poison}$ such that $|D'_{poison}|\\leq \\gamma|D_{train}|$ for some constant $\\gamma\\in[0,1]$. Let classifier $f'$ be learned from $D_{train} \\cup D'_{poison}$. \nDefinition \\ref{def:robust-training} says that a classifier $f'$ learned from poisoned training set can predict as accurately as $f$ learned from $D_{train}$ with a high probability. \n\\begin{definition}[{\\sf TR} or $(\\gamma,\\zeta)$-robust training; adapted from ]\\label{def:robust-training}\nGiven classifiers $f$ learned from $D_{train}$ and $f'$ learned from $D_{train} \\cup D'_{poison}$ where\n$|D_{poison}'| \\leq \\gamma |D_{train}|$, and small constants $\\zeta\\in [0,1]$, we say\n$f'$ is $(\\gamma,\\zeta)$-robust if $\\forall z\\in\\mathcal{Z}$\n$$\\left((f(z)>\\tau) \\land (|D_{poison}'| \\leq \\gamma |D_{train}|)\\right) \\rightarrow \\left(\\mathbb{P}(f'(z) > \\tau)>1-\\zeta\\right).$$\n\\end{definition}","id":"b05fe1bc-da2e-4a87-8f17-d43f1acc81fc","level":"subsection","origin_cites_number":5,"parent_id":"a1283bee-37c1-4371-bed3-ceeb8af64d75","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Survey and Systematization Methodology"],["subsection","Systematizing Security Properties"]],"subsections":[],"title":"Systematizing Security Properties"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:systematizstion-AMD-studies}\nWe systematize attacks according to $\\mathcal{A}$'s objective, input, assumptions, the security properties that are broken, and the types of victim malware detectors (e.g., Windows vs. Android). Similarly, we systematize defenses according to $\\mathcal{I}$'s objective, input, assumptions, the security properties that are achieved, and the types of enhanced malware detectors (e.g., Windows vs. Android).\nWe group attacks (defenses) according to the attacker's (defender's) techniques and then summarize them in a table according to the publication date in chronological order. \nFor convenience, we will use wildcard $*$ to indicate any value in a domain (e.g., $[0,1]$); we will use $\\lor$ to describe $\\mathcal{A}$'s and $\\mathcal{I}$'s ``broader'' input (if applicable). For example, $(0,1,0,1,0|A_6,\\ldots,A_9)\\lor (1,0,1,1,1|A_6,\\ldots,A_9)$ means that $\\mathcal{A}$ has either $(a_1,a_2,a_3,a_4,a_5)=(0,1,0,1,0)$ or\n$(a_1,a_2,a_3,a_4,a_5)=(1,0,1,1,1)$. Finally, we will present the attack-defense escalation.","id":"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87","level":"section","origin_cites_number":0,"parent_id":"4fe25043-06f8-4577-bd62-3fe42c7ab899","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"]],"subsections":["108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","ecf055b9-207b-4f7f-9c11-8bb57d05480e","76e4c68c-dec2-4af2-b3fa-ce426944cf52"],"title":"Systematizing AMD Arms Race"},{"cite_extract_rate":0.423076923076923,"cites":[3867,3857,6172,7756,3864,3862,937,3871,9099,3854,7754],"content":"\\label{sec:attack}\n\\begin{table}[htbp]\n\\caption{Summary of AMD attacks (\\checkmark means applicable, \\protect\\fullcirc means 0, \\protect\\emptycirc means 1, \\protect\\dotcirc means a value in $[0,1]$).}\n\t\\centering\n\t\\resizebox{0.9\\columnwidth}{!}{\n\t\\setlength{\\tabcolsep}{0.4em}\n\t\\begin{tabular}{l|ccc|ccccc|cccc|ccccc|cccc|ccc}\n\t\\hline\n\t\\multicolumn{1}{c|}{\\specialcell{Attack \\\\ (in chronological order)}}& \\multicolumn{3}{c|}{\\specialcell{Attack \\\\ Objective}} & \\multicolumn{9}{c|}{\\specialcell{Attack Input}} & \n\t\\multicolumn{5}{c|}{Assumptions} &\n\t\\multicolumn{4}{c|}{\\specialcell{Broken \\\\ Properties}} & \\multicolumn{3}{c}{\\specialcell{Malware \\\\ detector}} \\\\\n\t& \\vthead{\\sf Indiscriminate} \n\t& \\vthead{\\sf Targeted}\n\t& \\vthead{\\sf Availability}\n\t& \\vthead{$A_1$: Training set $D_{train}$}\n\t& \\vthead{$A_2$: Defense technique}\n\t& \\vthead{$A_3$: Feature set}\n\t& \\vthead{$A_4$: Learning algorithm}\n\t& \\vthead{$A_5$: Response}\n\t& \\vthead{$A_6$: Manipulation set}\n\t& \\vthead{$A_7$: Attack tactic}\n\t& \\vthead{$A_8$: Attack technique}\n\t& \\vthead{$A_9$: Adversarial example set}\n\t& \\vthead{{\\sf IID} assumption}\n\t& \\vthead{{\\sf Oracle} assumption}\n\t& \\vthead{{\\sf Measurability} assumption}\n\t& \\vthead{{\\sf Smoothness} assumption}\n\t& \\vthead{{\\sf Invertibility} assumption}\n\t& \\vthead{{\\sf RR}: Representation Robustness}\n\t& \\vthead{{\\sf CR}: Classification Robustness}\n\t& \\vthead{{\\sf DR}: Detection Robustness}\n\t& \\vthead{{\\sf TR}: Training Robustness}\n\t& \\vthead{Windows Program}\n\t& \\vthead{Android Package}\n\t& \\vthead{PDF}\n\t\\\\\\hline\\hline\n \\rowcolor{lightgray!30}\n Smutz and Stavrou & \n \\checkmark &&&\n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} &\n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} &\n $\\mathbf{M}$ &\n {\\sf OE2} &\n {\\sf MI} &\n $\\mathcal{X}_\\mathbf{M}$ &\n & & \\checkmark & & &\n & \\checkmark & & &\n & & \\checkmark\n \\\\\n Biggio et al. &\n \\checkmark &&& \n \\specialcell{\\pie{0} \\\\ \\pie{4}} & \n \\specialcell{\\pie{0} \\\\ \\pie{4}} & \n \\specialcell{\\pie{0} \\\\ \\pie{0}} & \n \\specialcell{\\pie{0} \\\\ \\pieg} & \n \\specialcell{\\pie{0} \\\\ \\pie{4}} &\n $\\mathbf{M}$ &\n ${\\sf OE2}$ &\n ${\\sf GO}$ &\n $\\mathcal{X}_{\\mathbf{M}}$ &\n & \\checkmark & \\checkmark && \\checkmark &\n & \\checkmark & &&\n & & \\checkmark \n \\\\\n \\rowcolor{lightgray!30}\n Maiorca et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} &\n \\specialcell{\\pie{4}} &\n \t$\\mathcal{M}$ &\n \t${\\sf BE}$ &\n \t{\\sf MI} &\n \t$\\mathcal{Z}_{\\mathcal{M}}$ &\n \t& \\checkmark & & & &\n \t& & \\checkmark & &\n \t& & \\checkmark \n \t\\\\\n {\\v{S}}rndi\\'{c} and Laskov &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{0} \\\\ \\pie{0}} & \n \\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{4} \\\\ \\pie{4}} & \n \\specialcell{\\pieg \\\\ \\pieg \\\\ \\pieg \\\\ \\pieg} & \n \\specialcell{\\pie{4} \\\\ \\pieg \\\\ \\pie{4} \\\\ \\pieg} &\n \\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{4} \\\\ \\pie{4}} &\n \t\\specialcell{$\\mathbf{M}$ \\\\ $\\mathcal{M}$ \\\\ $\\mathbf{M}$ \\\\ $\\mathcal{M}$} &\n \t\\specialcell{${\\sf BE}$ \\\\ ${\\sf BE}$ \\\\ ${\\sf OE2}$ \\\\ ${\\sf OE2}$ } &\n \t\\specialcell{${\\sf TR}$ \\\\ ${\\sf TR}$ \\\\ ${\\sf TR}$ \\\\ ${\\sf TR}$} &\n \t\\specialcell{$\\mathcal{X}_{\\mathbf{M}}$ \\\\ $\\mathcal{Z}_{\\mathcal{M}}$ \\\\ $\\mathcal{X}_{\\mathbf{M}}$ \\\\ $\\mathcal{Z}_{\\mathcal{M}}$} &\n \t& \\checkmark &&&\\checkmark& \n \t&&\\checkmark&&\n \t& & \\checkmark \n \t\\\\\n \\rowcolor{lightgray!30}\n Xu et al.~ &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} &\n \\specialcell{\\pie{0}} &\n \t$\\mathcal{M}$ &\n \t${\\sf BE}$ &\n \t{\\sf HS} &\n \t$\\mathcal{Z}_{\\mathcal{M}}$ &\n \t& & & & &\n \t& & \\checkmark &&\n \t& & \\checkmark \n \t\\\\\n \tCarmony et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \t\\specialcell{\\pie{4}} & \n \t\\specialcell{\\pie{4}} & \n \t\\specialcell{\\pie{4}} &\n \t\\specialcell{\\pie{4}} &\n \t$\\mathcal{M}$ &\n \t${\\sf BE}$ &\n \t{\\sf MI} &\n \t$\\mathcal{Z}_{\\mathcal{M}}$ &\n \t& \\checkmark & & & &\n \t& & \\checkmark & &\n \t& & \\checkmark \n \t\\\\\n \\rowcolor{lightgray!30}\n Hu and Tan & \n \\checkmark &&& \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{4}} &\n \\specialcell{\\pie{0}} &\n $\\mathbf{M}$ &\n ${\\sf BE}$ &\n {\\sf GM} &\n $\\mathcal{X}_{\\mathbf{M}}$ &\n \\checkmark& \\checkmark & & & \\checkmark& \n & & \\checkmark &&\n \\checkmark & & \n \\\\\n Hu and Tan & \n \\checkmark &&& \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{4}} &\n \\specialcell{\\pie{0}} &\n $\\mathbf{M}$ &\n ${\\sf BE}$ &\n {\\sf GM} &\n $\\mathcal{X}_{\\mathbf{M}}$ &\n \\checkmark& \\checkmark & & & \\checkmark &\n & & \\checkmark &&\n \\checkmark & & \n \\\\\n \\rowcolor{lightgray!30}\n Demontis et al. &\n \\checkmark &&& \n \\specialcell{\\pie{0} \\\\ \\pie{4}} & \n \\specialcell{\\pie{0} \\\\ \\pie{4}} &\n \\specialcell{\\pie{0} \\\\ \\pie{0}} & \n \\specialcell{\\pie{0} \\\\ \\pieg} &\n \\specialcell{\\pie{0} \\\\ \\pie{4}} &\n $\\mathbf{M}$ &\n ${\\sf OE2}$ &\n {\\sf SF} &\n $\\mathcal{X}_{\\mathbf{M}}$ &\n & \\checkmark & \\checkmark & & \\checkmark &\n & \\checkmark & &&\n & \\checkmark & \n \\\\\n Grosse et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} &\n \t$\\mathbf{M}$ &\n \t${\\sf OE2}$ &\n \t{\\sf SF} &\n \t$\\mathcal{X}_{\\mathbf{M}}$ &\n \t& \\checkmark &\\checkmark & & \\checkmark &\n \t\\checkmark & \\checkmark & &&\n \t& \\checkmark & \n \t\\\\\n \\rowcolor{lightgray!30}\n Chen et al. &\n \\checkmark &&& \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} &\n $\\mathbf{M}$ &\n ${\\sf OE2}$ &\n {\\sf SF} &\n $\\mathcal{X}_{\\mathbf{M}}$ &\n & \\checkmark & \\checkmark & & \\checkmark & \n & \\checkmark&&&\n \\checkmark & & \n \\\\\n Khasawneh et al. &\n \\checkmark &&& \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pieg} &\n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} &\n \\specialcell{ $\\mathbf{M}$ \\\\ $\\mathcal{M}$} &\n \\specialcell{${\\sf BE}$ \\\\ ${\\sf BE}$} &\n \\specialcell{{\\sf TR} \\\\ {\\sf TR}} &\n \\specialcell{$\\mathcal{X}_{\\mathbf{M}}$ \\\\ $\\mathcal{Z}_{\\mathcal{M}}$} &\n & \\checkmark & & & \\checkmark & \n & \\checkmark & &&\n \\checkmark & & \n \\\\\n \\rowcolor{lightgray!30}\n Dang et al. &\n \\checkmark &&& \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} &\n \\specialcell{\\pie{4}} &\n $\\mathcal{M}$ &\n ${\\sf BE}$ &\n {\\sf HS} &\n $\\mathcal{Z}_{\\mathcal{M}}$ &\n & & & & &\n & & \\checkmark &&\n & & \\checkmark \n \\\\\n Mu{\\~n}oz-Gonz{\\'a}lez et al. & \n &&\\checkmark& \n \\specialcell{\\pie{0} \\\\ \\pie{0}} & \n \\specialcell{\\pie{0} \\\\ \\pie{4}} & \n \\specialcell{\\pie{0} \\\\ \\pie{0}} & \n \\specialcell{\\pie{0} \\\\ \\pie{4}} & \n \\specialcell{\\pie{0} \\\\ \\pie{4}} &\n $\\mathbf{M}$ &\n ${\\sf OP}$ &\n {\\sf GO} &\n $\\mathcal{X}_{\\mathbf{M}}$ &\n & \\checkmark & & &\\checkmark &\n & & & \\checkmark &\n \\checkmark & & \n \\\\\n \\rowcolor{lightgray!30}\n Yang et al. &\n \\checkmark &&& \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} &\n \\specialcell{\\pie{0}} &\n $\\mathcal{M}$ &\n ${\\sf BE}$ &\n ${\\sf HS}$ &\n $\\mathcal{Z}_{\\mathcal{M}}$ &\n & \\checkmark & \\checkmark & & &\n & & \\checkmark &&\n & \\checkmark & \n \\\\\n Rosenberg et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} &\n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} &\n \t\\specialcell{$\\mathbf{M}$ \\\\ $\\mathcal{M}$} &\n \t\\specialcell{${\\sf BE}$ \\\\ ${\\sf BE}$} &\n \t\\specialcell{{\\sf TR} \\\\ {\\sf TR}} &\n \t\\specialcell{$\\mathcal{X}_{\\mathbf{M}}$ \\\\ $\\mathcal{Z}_{\\mathcal{M}}$} &\n \t& \\checkmark & & & \\checkmark &\n \t\\checkmark & \\checkmark & &&\n \t\\checkmark & & \n \t\\\\\n \\rowcolor{lightgray!30}\n Anderson et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} & \n \\specialcell{\\pie{4}} &\n \\specialcell{\\pie{0}} &\n \t$\\mathcal{M}$ &\n \t${\\sf OE2}$ &\n \t{\\sf GM} &\n \t$\\mathcal{Z}_{\\mathcal{M}}$ \n \t&\\checkmark& \\checkmark& \\checkmark & & \n \t& \\checkmark & \\checkmark &&\n \t& \\checkmark & & \n \t\\\\\n \tKreuk et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \t\\specialcell{\\pie{0}} & \n \t\\specialcell{\\pie{0}} & \n \t\\specialcell{\\pie{0}} & \n \t\\specialcell{\\pie{0}} &\n \t$\\mathbf{M}$ &\n \t${\\sf OE2}$ &\n \t{\\sf GO} &\n \t$\\mathcal{X}_{\\mathbf{M}}$ \n \t& &\\checkmark& & & \\checkmark\n \t& \\checkmark & \\checkmark &&\n \t& \\checkmark & & \n \t\\\\\n \t\\rowcolor{lightgray!30}\n \tChen et al. & \n \t\\checkmark &&& \n \t\\specialcell{\\pie{0}\\\\\\pie{4}\\\\ \\pie{0}} & \n \t\\specialcell{\\pie{0} \\\\ \\pie{4} \\\\ \\pie{4}} & \n \t\\specialcell{\\pie{0} \\\\ \\pie{4} \\\\ \\pieg} & \n \t\\specialcell{\\pie{0} \\\\ \\pie{0} \\\\ \\pie{0}} & \n \t\\specialcell{\\pie{0}\\\\ \\pie{0} \\\\ \\pie{0}} &\n \t$\\mathbf{M}$ &\n \t{\\sf BP} &\n \t{\\sf SF} &\n \t$\\mathcal{X}_\\mathbf{M}$ &\n \t& \\checkmark & \\checkmark & & \\checkmark & \n \t& & & \\checkmark &\n \t& \\checkmark & \n \t\\\\\n Al-Dujaili et al. & \n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} &\n \t$\\mathbf{M}$ &\n \t${\\sf OE2}$ &\n \t{\\sf GO} &\n \t$\\mathcal{X}_{\\mathbf{M}}$ &\n \t& \\checkmark & & & \\checkmark &\n \t&& \\checkmark &&\n \t\\checkmark & & \n \t\\\\\n \t\\rowcolor{lightgray!30}\n \tSuciu et al. & \n \t&\\checkmark&& \n \t\\specialcell{\\pieg \\\\ \\pie{0} \\\\ \\pie{0} \\\\ \\pie{0}} & \n \t\\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{4} \\\\ \\pie{0}} & \n \t\\specialcell{\\pie{0} \\\\ \\pieg \\\\ \\pie{0} \\\\ \\pie{0}} & \n \t\\specialcell{\\pieg \\\\ \\pieg \\\\ \\pie{4} \\\\ \\pie{0}} & \n \t\\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{4} \\\\ \\pie{0}} &\n \t\\specialcell{$\\mathbf{M}$ \\\\ $\\mathcal{M}$} &\n \t\\specialcell{${\\sf BP}$ \\\\ ${\\sf BP}$} &\n \t\\specialcell{{\\sf SF} \\\\ {\\sf SF}} &\n \t\\specialcell{$\\mathcal{X}_{\\mathbf{M}}$ \\\\ $\\mathcal{Z}_{\\mathcal{M}}$} &\n \t& \\checkmark & \\checkmark & & \\checkmark & \n \t& & &\\checkmark&\n \t& \\checkmark & \n \t\\\\\n Kolosnjaji et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} &\n \t$\\mathbf{M}$ &\n \t${\\sf OE2}$ &\n \t{\\sf GO} &\n \t$\\mathcal{X}_{\\mathbf{M}}$ \n \t&& \\checkmark&& & \\checkmark\n \t& \\checkmark & \\checkmark &&\n \t& \\checkmark & & \n \t\\\\\n \\rowcolor{lightgray!30}\n Suciu et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} & \n \\specialcell{\\pie{0}} &\n \t$\\mathbf{M}$ &\n \t${\\sf OE2}$ &\n \t{\\sf GO} &\n \t$\\mathcal{X}_{\\mathbf{M}}$ &\n \t& \\checkmark & & & \\checkmark &\n \t\\checkmark & \\checkmark &&&\n \t\\checkmark & & \n \t\\\\\n Chen et al. &\n \t\\checkmark &&& \n \t\\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{0} \\\\ \\pie{0}} &\n \\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{4} \\\\ \\pie{4}} &\n \\specialcell{\\pie{0} \\\\ \\pie{0} \\\\ \\pie{0} \\\\ \\pie{0}} &\n \\specialcell{\\pie{4} \\\\ \\pie{4} \\\\ \\pie{4} \\\\ \\pie{4}} &\n \\specialcell{\\pie{4} \\\\ \\pie{0} \\\\ \\pie{4} \\\\ \\pie{0}} &\n \t\\specialcell{$\\mathbf{M}$ \\\\ $\\mathbf{M}$} &\n \t\\specialcell{${\\sf OE1}$ \\\\ ${\\sf OE2}$} &\n \t\\specialcell{{\\sf GO} \\\\ {\\sf SF}} &\n \t\\specialcell{$\\mathcal{X}_{\\mathbf{M}}$\\\\ $\\mathcal{X}_{\\mathbf{M}}$} &\n \t& \\specialcell{\\checkmark } & \\specialcell{\\checkmark } & & \\checkmark &\n \t& \\checkmark & & &\n \t& \\checkmark & \n \t\\\\\n\t\\rowcolor{lightgray!30}\n\tPierazzi et al. &\n\t\\checkmark &&& \n\t\\specialcell{\\pie{0}} & \n\t\\specialcell{\\pie{0}} &\n\t\\specialcell{\\pie{0}} & \n\t\\specialcell{\\pie{0}} &\n\t\\specialcell{\\pie{0}} &\n\t$\\mathcal{M}$ &\n\t${\\sf OE2}$ &\n\t{\\sf MI} &\n\t$\\mathcal{Z}_{\\mathbf{M}}$ &\n\t& \\checkmark & & & &\n\t& & \\checkmark & & \n\t& \\checkmark & \n\t\\\\ \t \n\tLi and Li & \n\t\\checkmark &&& \n\t\\specialcell{\\pie{4}} & \n\t\\specialcell{\\pie{0}} & \n\t\\specialcell{\\pie{0}} & \n\t\\specialcell{\\pie{0}} & \n\t\\specialcell{\\pie{0}} &\n\t\\specialcell{$\\mathbf{M}$ \\\\ $\\mathcal{M}$} &\n\t\\specialcell{${\\sf OE2}$ \\\\ ${\\sf OE2}$} &\n\t\\specialcell{{\\sf MS} \\\\ {\\sf MS}} &\n\t\\specialcell{$\\mathcal{X}_{\\mathbf{M}}$ \\\\ $\\mathcal{Z}_{\\mathcal{M}}$} &\n\t& \\checkmark & & & \\checkmark &\n\t&& \\checkmark &&\n\t& \\checkmark & \n\t\\\\\\hline\n\t\\end{tabular}\n}\n\t\\label{tbl:attacks}\n\\end{table}","id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","level":"subsection","origin_cites_number":26,"parent_id":"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"]],"subsections":["f3db0c13-9b45-4545-9019-43245444f75d","6dae1197-fcde-489c-835b-7117fc4737e8","6bd1592d-a0b4-4061-b526-553c6ebc409a","4e7e5878-d20a-4c28-acf4-a41004766a2c","b8ab1fcb-f749-4c70-9893-28b772af455c","33314056-9caf-48b3-90ff-0ca708a5a80c","c51df6e6-79f5-400d-8643-cef009ff014f","d4bb8301-3585-4f94-b9fa-12ca7a7b0164"],"title":"Systematizing Attack Literature"},{"cite_extract_rate":0.8235294117647051,"cites":[892,1193,923,3867,3873,3862,3872,937,3871,3863,9099,3855,3854,3861],"content":")}\nBiggio et al.~ propose solving the problem of optimal evasion attacks by leveraging gradient-based optimization techniques. They focus on high-confidence evasion attacks with small perturbations (cf. Definition \\ref{def:opt_evasion2}). Given a malware representation-label pair $(\\mathbf{x},y=+)$, the optimization problem specified in Eq.\\eqref{eq:a8:evs2} with respect to {\\sf GO} is instantiated as:\n\\begin{align*}\n\t\\max_{\\delta_{\\mathbf{x}}} L_\\mathcal{A}(\\hat{F}_{\\hat\\theta}(\\mathbf{x}+\\delta_{\\mathbf{x}}),y=+) & = \\min_{\\delta_{\\mathbf{x}}} \\left(L(F_\\theta(\\mathbf{x}+\\delta_{\\mathbf{x}}, y=-)) - \\beta_a\\mathcal{K}(\\mathbf{x} + \\delta_{\\mathbf{x}})\\right) \\\\ ~~\\text{s.t.}~~(\\delta_{\\mathbf{x}}\\in[\\mathbf{0},\\overline{\\mathbf{u}}]) &\\land (C(\\mathbf{x},\\mathbf{x}+\\delta_{\\mathbf{x}})\\leq m),\n\\end{align*}\nwhere $\\beta_a\\geq 0$ is a balance factor and $\\mathcal{K}$ is a density estimation function for lifting $\\mathbf{x}+\\delta_{\\mathbf{x}}$ to the populated region of benign examples. Since $\\delta_{\\mathbf{x}}\\geq \\mathbf{0}$, the manipulation only permits object injections to meet the requirement of preserving malicious functionalities. The attack is validated by using the PDF malware detector and the feature representation is the number of appearances of hand-selected keywords (e.g., JavaScript). Because the perturbation is continuous, the authors suggest searching a discrete point close to the continuous one and aligning the point with $\\nabla L_\\mathcal{A}(\\mathbf{x}+\\delta_{\\mathbf{x}},y=+)$. \nThis attack makes the invertibility Assumption \\ref{assumption:inverse} because it operates in the feature space. Experimental results show that when $\\mathcal{I}$ employs no countermeasures, knowing $\\mathcal{I}$'s feature set $S$ and learning algorithm $F$ are sufficient for $\\mathcal{A}$ to evade $\\mathcal{I}$'s detector. This attack and its variants have been shown to evade PDF malware detectors , PE malware detectors , Android malware detectors , and Flash malware detectors . The kernel density estimation item makes the perturbed representation $\\mathbf{x}+\\delta_\\mathbf{x}$ similar to the representations of benign examples, explaining the successful evasion. In summary, the attack works under the ${\\sf Oracle}$, ${\\sf Measurability}$, and {\\sf Invertibility} assumptions. $\\mathcal{A}$'s input is, or $\\mathcal{A}$ can be characterized as, $(a_1,\\ldots,a_5|A_6,\\cdots,A_9) = (1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_{\\mathbf{M}}) \\lor (0,0,1,*,0|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_{\\mathbf{M}})$ and $\\mathcal{A}$ breaks the {\\sf CR} property.\nAl-Dujaili et al.~ propose evasion attacks against DNN-based malware detectors in the feature space. In this attack,\n$\\mathcal{A}$ generates adversarial examples with possibly large perturbations in the feature space. More precisely, given a representation-label pair $(\\mathbf{x},y=+)$, the optimization problem of Eq.\\eqref{eq:a8:evs2} with respect to {\\sf GO} is instantiated as:\n\t$\\max_{\\delta_{\\mathbf{x}}} L(F_{\\theta}(\\mathbf{x}+\\delta_{\\mathbf{x}}),y=+) ~~\\text{s.t.}~~ \\delta_{\\mathbf{x}} \\in [\\mathbf{0}, \\mathbf{1}-\\mathbf{x}].$\nThe attack has four variants, with each perturbing the representation in a different direction (e.g., normalized gradient of the loss function using the $\\ell_\\infty$ norm). A ``random'' rounding operation is used to map continuous perturbations into a discrete domain. When compared with the basic rounding (which returns 0 if the input is smaller than 0.5, and returns 1 otherwise), the ``random'' rounding means that the threshold of rounding is sampled from the interval $[0,1]$ uniformly. For binary feature representation, the manipulation set $\\mathbf{M}_\\mathbf{x}=[\\mathbf{0}, \\mathbf{1}-\\mathbf{x}]$ assures the flipping of 0 to 1. The effectiveness of the attack is validated using Windows malware detector in the feature space. In summary, the attack works under the {\\sf Oracle} and {\\sf Invertibility} assumptions with $\\mathcal{A}$ input $(a_1,\\ldots,a_5|A_6,\\cdots,A_9) =(0,1,1,1,1|\\mathbf{M}, {\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$ and breaks the {\\sf DR} property.\nKreuk et al. propose an evasion attack in the feature space against MalConv, which is an end-to-end Windows PE malware detector (as reviewed in Section \\ref{sec:ml}) . Given a malware embedding code $\\mathbf{x}$, the optimization problem of Eq.\\eqref{eq:a8:evs2} with respect to {\\sf GO} is instantiated as:\n\t$\\min_{\\delta_{\\mathbf{x}}} L(F_{\\theta}([\\mathbf{x}|\\delta_{\\mathbf{x}}]),y=-) ~~\\text{s.t.}~~ \\Vert\\delta_{\\mathbf{x}}\\Vert_p \\leq \\epsilon,$\nwhere | means concatenation and $\\Vert\\cdot\\Vert_p$ is the $p$ norm where $p\\geq 1$. Because MalConv is learned from sequential data, perturbation means appending some content to the end of a PE file. Perturbations are generated in a single step by following the direction of the $\\ell_\\infty$ or $\\ell_2$ normalized gradients of the loss function . For instance, the attack based on the $\\ell_{\\infty}$ norm is\n$\\tilde{\\mathbf{x}}'=\\mathbf{x} - \\epsilon\\cdot\\sign(\\nabla_{\\bf x}(L(F_\\theta(\\mathbf{x}), -))$, where $\\sign(x)=+1~(-1)$ if $x\\geq0~(x<0)$.\nSince the embedding operation uses a look-up table to map discrete values (0, 1, $\\ldots$, 255) to the learned real-value vectors, the attack uses a nearest neighbor search to look for the learned embedding code close to $\\tilde{\\mathbf{x}}'$. \nIn summary, the attack works under the {\\sf Oracle} and {\\sf Invertibility} assumptions with input $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$ and breaks the {\\sf RR} and {\\sf CR} properties.\nKolosnjaji et al. and Suciu et al. independently propose gradient-based attacks in the feature space to evade MalConv . Both studies also use the loss function exploited by Kreuk et al. . Kolosnjaji et al. use the manipulation set $\\mathcal{M}$ corresponding to appending instructions at the end of a file. This attack proceeds iteratively and starts with randomly initialized perturbations. In each iteration, continuous perturbations are updated in the direction of the $\\ell_2$ normalized gradient of the loss function with respect to the input, and then a nearest neighbor search is applied to obtain discrete perturbations. Suciu et al. perturb embedding codes in the direction of the $\\ell_\\infty$ normalized gradient of the loss function, while adding instructions in the mid of a PE file (e.g., between PE {\\em sections}) while noting that appended content could be truncated by MalConv. Both attacks work under the {\\sf Oracle} and {\\sf Invertibility} assumptions with input $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$ and break the {\\sf RR} and {\\sf CR} properties.\nMu{\\~n}oz-Gonz{\\'a}lez et al. propose the optimal poisoning {\\sf OP} attack in the feature space (Definition \\ref{def:optimal poisoning attack}), which is NP-hard. In this case, the optimization problem of Eq.\\eqref{eq:a8:poi} is relaxed by supposing that the classifier is linear to render the optimization problem tractable . The attack is waged against Windows PE malware detectors. Feature set includes API calls, actions and modifications in the file system; each file is represented by a binary vector. The attack has two variants: one uses white-box input, where $\\mathcal{A}$ derives $D_{poison}'$ from $\\mathcal{I}$'s detector $f$; the other uses grey-box input, where $\\mathcal{A}$ knows $\\mathcal{I}$'s training set as well as feature set and trains a surrogate detector. The attack works under the {\\sf Oracle} and {\\sf Invertibility} assumptions with input $(a_1,\\ldots,a_5|A_6,\\cdots,A_9) =(1,1,1,1,1|\\mathbf{M},{\\sf OP}, {\\sf GO}, \\mathcal{X}_\\mathbf{M})\\lor(1,0,1,0,0|\\mathbf{M},{\\sf OP}, {\\sf GO},\\mathcal{X}_\\mathbf{M})$ and breaks {\\sf TR}.","id":"f3db0c13-9b45-4545-9019-43245444f75d","level":"subsubsection","origin_cites_number":17,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Attacks using Gradient-based Optimization ({\\sf GO"]],"subsections":[],"title":"Attacks using Gradient-based Optimization ({\\sf GO"},{"cite_extract_rate":0.375,"cites":[894,890,3864,7754,3863,7154],"content":")}\nDemontis et al. propose the optimal evasion {\\sf OE2} in the feature space to perturb important features in terms of their weights in the linear function $\\varphi(\\mathbf{x})={\\mathbf{w}}^\\top\\mathbf{x} + b$, where $\\mathbf{w}=(w_1,w_2,\\cdots,w_d)$ is a weight vector, $b$ is the bias, and $d$ is the dimension of feature space. The attack is waged against Drebin malware detector (which is reviewed in Section \\ref{sec:ml}). $\\mathcal{A}$ manipulates the $x_i$'s with largest $|w_i|$'s as follows:\nflip $x_i=1$ to $x_i=0$ if $w_i > 0$, flip $x_i=0$ to $x_i=1$ if $w_i < 0$, and do nothing otherwise, while obeying the manipulation set $\\mathbf{M}$ corresponding to the injection or removal of features. The attack works under the {\\sf Oracle}, {\\sf Measurability}, and {\\sf Invertiblity} assumptions with\ninput $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M}) \\lor (0,0,1,*,0|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$ and breaks the {\\sf CR} property.\nGrosse et al. propose a variant of the Jacobian-based Saliency Map Attack (JSMA) in the feature space against the Drebin malware detector (which is reviewed in Section \\ref{sec:ml}). Instead of using SVM, Deep Neural Network (DNN) is used to build a detector. Important features are identified by leveraging the gradients of the softmax output of a malware example with respect to the input. A large gradient value indicates a high important feature. $\\mathcal{A}$ only injects manifest features to manipulate Android Packages and generates adversarial files from $\\mathcal{I}$'s detector $f$. The attack works under the {\\sf Oracle}, {\\sf Measurability}, and {\\sf Invertiblity} assumptions\nwith input $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$ and breaks the {\\sf RR} and {\\sf CR} properties.\nChen et al. propose an evasion attack in the feature space by perturbing the important features derived from a wrapper-based feature selection algorithm . The attacker's loss function $L_\\mathcal{A}$ has two parts: (i) the classification error in the mean squared loss and (ii) the manipulation cost $C(\\mathbf{x}, \\mathbf{x}')=\\sum_{i=1}^{d}c_i|x_i - x_i'|$, where $\\mathbf{x}=(x_1,\\ldots,x_d)$, $\\mathbf{x}'=(x_1', \\ldots,x_d')$, and $c_i$ is the hardness of perturbing the $i$th feature while preserving malware's functionality. The attack is waged against Windows PE malware detector that uses hand-crafted Windows API calls as features and the binary feature representation. However, there are no details about the composition of manipulation set. This attack works under the {\\sf Oracle}, {\\sf Measurability}, and {\\sf Invertibility} assumptions with input $(a_1,\\ldots,a_5|A_7,\\cdots,A_9)=(1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf SF}, \\mathcal{X}_\\mathbf{M}))$ and breaks {\\sf CR}.\nChen et al. propose evasion attacks in the feature space against two Android malware detectors, MaMaDroid and Drebin . The manipulation set $\\mathbf{M}$ corresponds to the injection of manifest features (e.g., {\\em activities}) and \nAPI calls. $\\mathcal{A}$ evades MaMaDroid by using the optimal evasion ${\\sf OE1}$ (Definition \\ref{def:opt_evasion1}) and ${\\sf OE2}$ (Definition \\ref{def:opt_evasion2}), and evades Drebin by using ${\\sf OE2}$. The optimization problem of ${\\sf OE1}$ Eq.\\eqref{eq:a8:evs1} is solved using an advanced gradient-based method known as C\\&W . ${\\sf OE2}$ is solved using JSMA . Because JSMA perturbs sensitive features, we categorize this attack into the {\\sf SF} group. The ${\\sf OE2}$ attack works under the {\\sf Oracle}, {\\sf Measurability} and {\\sf Invertibility} assumptions,\nwith four kinds of input \n$(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,0,1,0,0|\\mathbf{M},{\\sf OE2}, {\\sf SF},\\mathcal{X}_\\mathbf{M}) \\lor (0,0,1,0,1|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M}) \\lor (1,0,1,0,0|\\mathbf{M},{\\sf OE2},{\\sf SF}, \\mathcal{X}_\\mathbf{M}) \\lor (1,0,1,0,1|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$,\nand breaks the {\\sf CR} property. The ${\\sf OE1}$ attack works under the same assumptions with the same input except using attack technique {\\sf GO}, and breaks the {\\sf CR} property.\nChen et al. propose a basic poisoning {\\sf BP} attack in the feature space against Android malware detectors. The feature set contains syntax features (e.g., permission, hardware, {API}) and semantic features (e.g., sequence of pre-determined program behaviors such as {\\tt getDevicedID}$\\rightarrow$ URL$\\rightarrow${\\tt openConnection}). The ML algorithm used is SVM, random forest, or $K$-Nearest Neighbor (KNN) . The malware representations are perturbed using a JSMA variant against the SVM-based classifier (while noting JSMA is applicable neither to random forests nor to KNN because they are gradient-free). Feature manipulation set ${\\bf M}$ corresponds to the injection of syntax features. $\\mathcal{A}$ poisons $\\mathcal{I}$'s training set by injecting perturbed perturbations with label $-$. The attack works under the {\\sf Oracle}, {\\sf Measurability}, and {\\sf Invertibility} assumptions with input\n$(a_1,\\ldots,a_5|A_6,\\cdots,A_9)=(1,1,1,1,1|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M}) \\lor (0,0,0,1,1|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M}) \\lor (1,0,*,1,1|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M})$ and breaks {\\sf TR}.\nSuciu et al. propose a basic poisoning attack in both feature and problem spaces. The authors obtain $D'_{poison}$ by applying small manipulations to non-adversarial benign files and then obtain their labels as given by VirusTotal service . \n$\\mathcal{A}$'s objective is to make $\\mathcal{I}$'s classifier $f$ mis-classify a targeted malware file $z_{mal}$ as benign. $\\mathcal{A}$ proceeds as follow: (i) obtain an initial benign file $z_{ben}$, where $z_{ben} \\approx z_{mal}$ in the feature space with respect to the $\\ell_1$ norm; (ii) use the JSMA method to manipulate $z_{ben}$ to $z'_{ben}$ by making a small perturbation so that they have similar feature representations; (iii) add $z'_{ben}$ and its label obtained from VirusTotal to $D'_{poison}$ and use $D_{train} \\cup D'_{poison}$ to train classifier $f'$ (Definition \\ref{definition:sf}); (iv) undo the addition if $z'_{ben}$ lowers the classification accuracy significantly, and accept it otherwise. The attack is waged against the Drebin malware detector and the manipulation set corresponds to the feature injection of permission, API, and strings. This attack works under the {\\sf Oracle}, {\\sf Measurability}, and {\\sf Inversiability} assumptions with input \n$(a_1,\\ldots,a_5|A_6,\\ldots,A_9) =(*,0,1,*,0|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M})\\lor(1,1,1,1,1|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M})\\lor(1,0,*,*,0|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M}) \\lor (1,0,1,0,0|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M})$\nand breaks the {\\sf TR} property. The study generates adversarial malware examples, but does not test their malicious functionalities.","id":"6dae1197-fcde-489c-835b-7117fc4737e8","level":"subsubsection","origin_cites_number":16,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Attacks using Sensitive Features ({\\sf SF"]],"subsections":[],"title":"Attacks using Sensitive Features ({\\sf SF"},{"cite_extract_rate":0.16666666666666602,"cites":[7754],"content":")} Smutz and Stavrou propose a mimicry attack in the feature space to modify features of a malicious file to mimic benign ones, where $\\mathcal{A}$ knows $\\mathcal{I}$'s classifier $f$. Discriminative features are identified by observing their impact on classification accuracy. The attack perturbs features of malware examples by replacing their value with the mean of the benign examples. The attack is leveraged to estimate the robustness of PDF malware detectors without considering the preservation of malware functionality. The attack works under the {\\sf Measurability} assumption with input $(a_1,\\ldots,a_5|A_6,\\cdots,A_9) = (1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf MI},\\mathcal{X})$ and breaks the {\\sf CR} property.\nMaiorca et al. propose a reverse mimicry attack against PDF malware detectors in the problem space. Instead of modifying malicious files to mimic benign ones, $\\mathcal{A}$ embeds malicious payload (e.g., JavaScript code) into a benign file. The attack can be enhanced by using parser confusion strategies, which make the injected objects being neglected by feature extractors when rendered by PDF readers . The attack works under the {\\sf Oracle} assumption with input $(a_1,\\ldots,a_5|A_6,\\cdots,A_9) = (0,0,0,0,0|$ $\\mathcal{M},{\\sf BE},{\\sf MI},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf DR} property. \nPierazzi et al. propose a white-box evasion attack against the Drebin malware detector and then an enhanced version of the detector in the problem space . They intend to bridge the gap between the attacks in the problem space and the attacks in the feature space.\nIn addition, four realistic constraints are imposed on the manipulation set $\\mathcal{M}$, including available transformation, preserved semantics, robustness to preprocessing, and plausibility. In order to cope with the side-effect features when incorporating gradient information of $\\mathcal{I}$'s classifier, the attacker first harvests a set of manipulations from benign files; \nManipulations in the problem space are used to query $\\mathcal{I}$'s feature extraction for obtaining perturbations in the feature space;\nan adversarial malware example is obtained by using the manipulations corresponding to the perturbations that have a high impact on the classification accuracy. This attack works under the {\\sf Oracle} assumption with input $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(1,1,1,1,1|\\mathcal{M},{\\sf OE2},{\\sf MI},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf DR} property.","id":"6bd1592d-a0b4-4061-b526-553c6ebc409a","level":"subsubsection","origin_cites_number":6,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Attacks using MImicry ({\\sf MI"]],"subsections":[],"title":"Attacks using MImicry ({\\sf MI"},{"cite_extract_rate":0,"cites":[],"content":")} \n{\\v{S}}rndi\\'{c} and Laskov investigate the mimicry attack and the aforementioned gradient descent and kernel density estimation attack against the PDFrate service, where $\\mathcal{A}$ knows some features used by $\\mathcal{I}$. $\\mathcal{A}$ makes the representation of an adversarial malware example similar to a benign one. Manipulation set $\\mathcal{M}$ corresponds to adding objects into PDF files. Both attacks perturb feature vectors against a surrogate model, and then map the perturbed feature representations to the problem space by injecting manipulations between the body and the trailer of a PDF file. For the mimicry attack, $\\mathcal{A}$ uses $N_{ben}>0$ benign examples to guide manipulations, resulting in $N_{ben}$ perturbed examples. The example incurring the highest classification error is used as an adversarial example. The attack works under the {\\sf Oracle} and {\\sf Invertibility} assumptions with input \n$( a_1,\\ldots,a_5|A_6,\\ldots,A_9) =(0,0,*,0,0|\\mathbf{M},{\\sf BE},{\\sf TR},\\mathcal{X}_\\mathbf{M})\\lor(0,0,*,*,0|\\mathbf{M},{\\sf BE},{\\sf TR},\\mathcal{X}_\\mathbf{M})\\lor(1,0,*,0,0|\\mathbf{M},{\\sf BE},{\\sf TR},\\mathcal{X}_\\mathbf{M})\\lor(1,0,*,*,0|\\mathbf{M},{\\sf BE},{\\sf TR},\\mathcal{X}_\\mathbf{M})\\lor(0,0,*,0,0|\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathcal{M})\\lor(0,0,*,*,0|\\mathcal{M},$ ${\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathcal{M})\\lor(1,0,*,0,0|\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathcal{M})\\lor(1,0,*,*,0|\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathcal{M})$\nand breaks {\\sf DR}. The gradient-based attack neglects the constraint of small perturbations and works under the same assumptions with the same input except using the attack technique {\\sf OE2}.\nKhasawneh et al. propose an evasion attack in both the feature and problem spaces against malware detectors learned from dynamic hardware features (e.g., instruction frequency), where $\\mathcal{A}$ knows some features used by $\\mathcal{I}$. The attack proceeds as follows. $\\mathcal{A}$ first queries $\\mathcal{I}$'s classifier to obtain a surrogate model and then generates adversarial files against the surrogate model. Manipulation set $\\mathcal{M}$ corresponds to the injection of some features because the others (e.g., memory access) are uncontrollable. Perturbations are conducted to the important features that are identified by large weights in the model. The attack works under the {\\sf Oracle} and {\\sf Invertibility} assumptions with input $( a_1,\\ldots,a_5|A_6,\\cdots,A_9) =(0,0,*,0,1|\\mathbf{M},{\\sf BE},{\\sf TR},\\mathcal{X}_\\mathbf{M})\\lor(0,0,*,0,1|\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf CR} property.\nRosenberg et al. propose an evasion attack in both the feature and problem spaces against a Recurrent Neural Network (RNN) based surrogate model, which is learned from API call sequences. In this attack, $\\mathcal{A}$'s training data is different from $\\mathcal{I}$'s, but the labels are obtained by querying $\\mathcal{I}$'s detector. In order to reduce the number of queries to $\\mathcal{I}$'s detector, $\\mathcal{A}$ augments its training data using the Jacobian-based augmentation technique and modifies the API sequence of an example in the direction of the $\\ell_\\infty$ normalized gradient of the loss function. Manipulation set ${\\bf M}$ corresponds to inserting no-op API calls. Experimental results show that adversarial examples generated from a surrogate RNN model can evade SVM, DNN, and RNN detectors. The attack works under the {\\sf Oracle} and {\\sf Invertibility} assumptions with input $( a_1,\\ldots,a_5|A_6,\\cdots,A_9) =(0,0,1,0,1|\\mathbf{M},{\\sf BE},{\\sf TR},\\mathcal{X}_\\mathbf{M}) \\lor (0,0,1,0,1|\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf RR} and {\\sf CR} properties.","id":"4e7e5878-d20a-4c28-acf4-a41004766a2c","level":"subsubsection","origin_cites_number":4,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Attacks using TRansferability ({\\sf TR"]],"subsections":[],"title":"Attacks using TRansferability ({\\sf TR"},{"cite_extract_rate":0.2,"cites":[3857],"content":")}\nXu et al.~ propose black-box evasion attacks in the problem space against two PDF malware detectors known as PDFrate and Hidost , respectively. Given a malicious file $z$, $\\mathcal{A}$ uses a genetic algorithm to iteratively generate $z'$ from $z$ as follows: (i) $\\mathcal{A}$ manipulates a set of candidates (or $z$ in the initial iteration) via object deletion, insertion, or replacement. (ii) $\\mathcal{A}$ queries these variants to $\\mathcal{O}$ and $f$. (iii) $\\mathcal{A}$ succeeds when obtaining a successful adversarial example $z'$, namely $({\\tt True} \\leftarrow \\mathcal{O}(z,z')) \\land (-\\leftarrow f(z'))$; otherwise, $\\mathcal{A}$ uses a score function to select candidates for the next iteration or aborts after reaching a threshold number of iterations. The score function $h$ varies with classifiers; for PDFrate, $h(\\mathcal{O},f,z,z')=0.5-f(z')$ if $\\mathcal{O}(z,z')={\\sf true}$, and returns -0.5 if $\\mathcal{O}(z,z')={\\sf false}$. This attack models an {\\sf Oracle} and works with the input $(a_1,\\cdots, a_5|A_6,\\cdots,A_9) = (0,0,0,0,1|\\mathcal{M},{\\sf BE},{\\sf HS},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf DR} property.\nYang et al. propose evasion attacks against Android malware detectors in the problem space. In this attack, $\\mathcal{A}$ also uses a genetic algorithm to perturb a malware example $z$ iteratively. In each iteration, $\\mathcal{A}$ extracts some features and calculates similarity scores between the malicious APKs in the feature space; the features that have high impact on the similarity scores are selected; the manipulations are to perturb the selected features.\nThe attack works under the {\\sf Oracle} and {\\sf Measurability} assumptions with input $(a_1,\\cdots, a_5|A_6,\\cdots,A_9) = (0,0,0,0,1|\\mathcal{M},\\mathcal{Z},{\\sf BE}, {\\sf HS},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf DR} property.\nDang et al. propose a black-box evasion attack against malware detectors (e.g., PDFrate) in the problem space. Given a malicious file $z$, $\\mathcal{A}$ uses the hill-climbing algorithm to iteratively generate adversarial file $z'$ from $z$. In each iteration, $\\mathcal{A}$ generates a path of variants sequentially, each of which is perturbed from its predecessor using manipulations corresponding to object deletion, insertion, or replacement. A score function $h$ is leveraged to select candidates, such as $h(\\mathcal{O},f,z,z')={\\sf mal}_{z'}-{\\sf clf}_{z'}$ or $h(\\mathcal{O},f,z,z')={\\sf mal}_{z'}/{\\sf clf}_{z'}$, where ${\\sf mal}_{z'}$ denotes the length of the first example turned from malicious to benign (obtaining by using an oracle) on the manipulation path (cf. Definition \\ref{definition:hs}) and ${\\sf clf}_{z'}$ denotes the length of the first malware example that has successfully misled the classifier $f$. Both examples of interest are obtained by a binary search, effectively reducing the number of queries to oracle $\\mathcal{O}$ and $f$. The attack models an {\\sf Oracle} and works with input $(a_1,\\cdots, a_5|A_6,\\cdots,A_9) = (0,0,0,0,0|\\mathcal{M},{\\sf BE},{\\sf HS},\\mathcal{Z}_{\\mathcal{M}})$ and breaks the {\\sf DR} property.","id":"b8ab1fcb-f749-4c70-9893-28b772af455c","level":"subsubsection","origin_cites_number":5,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Attacks using Heuristic Search ({\\sf HS"]],"subsections":[],"title":"Attacks using Heuristic Search ({\\sf HS"},{"cite_extract_rate":0.5,"cites":[243,6172],"content":")}\nHu and Tan \\cite {Hu2017} propose an evasion attack against Windows malware detectors in the feature space, by using Generative Adversarial Networks (GAN) . In this attack, $\\mathcal{A}$ modifies the binary representation of Windows API calls made by malicious files, namely flipping some feature values from 0 to 1. $\\mathcal{A}$ learns a generator $G_{\\theta_{g}}$ and a discriminator from $\\mathcal{A}$'s training dataset. The discriminator is a surrogate detector learned from feature vectors corresponding to $\\mathcal{A}$'s benign files and those produced by $G_{\\theta_{g}}$, along with labels obtained by querying $\\mathcal{I}$'s detector $f$. An adversarial example feature vector is generated by using\n$\\mathbf{x}' = \\max(\\mathbf{x}, \\round(G_{\\theta_{g}}(\\mathbf{x}, \\mathbf{a})))$,\nwhere $\\mathbf{a}$ is a vector of noises, $\\round$ is the round function, and $\\max$ means element-wise maximum.\nHu and Tan~ also propose another evasion attack using the Seq2Seq model . Both attacks work under the {\\sf IID}, {\\sf Oracle} and {\\sf Invertibility} assumptions with input $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,0,1,0,1|\\mathbf{M},{\\sf BE},{\\sf GM},\\mathcal{X}_\\mathbf{M})$ and break the {\\sf DR} property.\nAnderson et al. propose a Reinforcement Learning (RL)-based evasion attack against Windows PE malware detectors in the problem space. Manipulation set $\\mathcal{M}$ is the RL action space, which includes some bytecode injections (e.g., API insertion) and some bytecode deletion. Attacker $\\mathcal{A}$ learns an RL agent on $\\mathcal{A}$'s data, with labels obtained by querying defender $\\mathcal{I}$'s detector $f$. The learned agent predicts manipulations sequentially for a given malware example. Moreover, $\\mathcal{A}$ is restricted by only applying a small number of manipulations to a malicious PE file. Experimental results show that the attack is not as effective as others (e.g., gradient-based methods). The attack works under the {\\sf Oracle} and {\\sf Measurability} assumptions with input $(a_1,\\cdots, a_5|A_6,\\cdots,A_9) = (0,0,0,0,1|\\mathcal{M},{\\sf OE2},{\\sf GM},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf RR} and {\\sf CR} properties.","id":"33314056-9caf-48b3-90ff-0ca708a5a80c","level":"subsubsection","origin_cites_number":4,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Attacks using Generative Model ({\\sf GM"]],"subsections":[],"title":"Attacks using Generative Model ({\\sf GM"},{"cite_extract_rate":1,"cites":[9099],"content":")} \nLi and Li propose evasion attacks against DNN-based Android malware detectors in both feature and problem spaces. Given four gradient-based attack methods, the attack looks for the best one to perturb malware representations. $\\mathcal{A}$ can iteratively perform this strategy to modify the example obtained in the previous iteration. Experimental results show that the mixture of attacks can evade malware detectors effectively. The attack works under the {\\sf IID}, {\\sf Oracle} and {\\sf Invertibility} assumptions with input $(a_1,\\cdots, a_5|A_6,\\cdots,A_9) = (1,1,1,1,1|\\mathbf{M},{\\sf BE},{\\sf MS},\\mathcal{X}_\\mathbf{M})\\lor(1,1,1,1,1|\\mathcal{M},{\\sf BE},{\\sf MS},\\mathcal{Z}_\\mathcal{M})$ and breaks the {\\sf DR} property.","id":"c51df6e6-79f5-400d-8643-cef009ff014f","level":"subsubsection","origin_cites_number":1,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Attacks using Mixture Strategy ({\\sf MS"]],"subsections":[],"title":"Attacks using Mixture Strategy ({\\sf MS"},{"cite_extract_rate":0,"cites":[],"content":"We summarize the preceding reviews with the following observations.\n(i) Indiscriminate attacks have been much more extensively investigated than targeted attacks and availability attacks. (ii) Evasion attacks have been much more extensively studied than poisoning attacks. (iii) The {\\sf Oracle} assumption has been widely made. \nIn addition, we draw the following insights.\n\\begin{insight} \n(i) Knowing the defender's feature set is critical to the success of transfer attacks, highlighting the importance of keeping the defender's feature set secret (e.g., randomizing the defender's feature set).\n(ii) The effectiveness of evasion attacks largely depends on the attacker's degree of freedom in conducting manipulations in the problem space (i.e., a smaller degree of freedom means it is harder for the attack to succeed).\n\\end{insight}\n\\ignore{\nIn order to draw further insights, we propose applying ML to the systematized structured data in Table \\ref{tbl:attacks}. We propose formulating {\\em insights learning} as a multiclass classification problem. We use attributes of attack objective, attack input and assumption as features and treat the broken properties as the labels to learn. We preprocess data as follows:\nfor features corresponding to attack objective and assumption, we use 1 to represent ``applicable'' and\n0 otherwise; for features corresponding to $A_1,\\ldots,A_5$, we set $a_i =0.5$ for $1\\leq i \\leq 5$; for features corresponding to categorical $A_6,\\dots,A_9$, we use integers to represent the elements (e.g., 0,1 respectively representing $\\mathcal{M},\\mathbf{M}$ for $A_6$). This leads to a dataset of 63 attacks, each of which has 17 dimensions. We use the dataset to train a {\\em random forest} model with default hyper-parameters in scikit-learn . {\\color{blue} The random forest is ML algorithm, capable of learning from un-normalized features and producing the important features in terms of classification error .}\nWe focus on learning important attributes.\n\\begin{figure}[!htbp]\n\t\\centering\n\t\\scalebox{0.25}{\n\t\\includegraphics{figures/attack-important-feat.eps}\n\t}\n\\caption{Insights learning from systematized AMD attacks via attribute importance.}\n\t\\label{fig:attack-attribute}\n\\end{figure}\nFigure \\ref{fig:attack-attribute} highlights the most important features, from which we make the following observations. (i) Attack tactic is the most important feature in determining the security property that is broken. This can be explained attacks using sensitive features ({\\sf SF}) usually breaks {\\sf CR} or {\\sf TR}, and attacks using heuristic search ({\\sf HS}) usually breaks the {\\sf DR} property. (ii) Attack scenario is the second important feature. This is because ${\\sf BE}$ breaks ${\\sf DR}$, ${\\sf OE2}$ usually breaks {\\sf CR} and {\\sf RR}, and poisoning attacks ({\\sf BP} and {\\sf OP}) break {\\sf TR}. (iii) Feature set is the third important feature, resonating the manually-drawn insight. (iv) The {\\sf Measurability} assumption is the four important feature. This can be explained because {\\sf RR} and {\\sf CR} are defined with respect to small-manipulation attacks, which need this assumption.\n\\begin{insight}[Insights learned via ML]\nTechnique-wise, attack tactic and attack scenario (core part of the threat model) largely determines what security property is broken.\nMethodology-wise, ML is an effective method for insights learning in survey and systematization studies.\n\\end{insight}\n}","id":"d4bb8301-3585-4f94-b9fa-12ca7a7b0164","level":"subsubsection","origin_cites_number":2,"parent_id":"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Attack Literature"],["subsubsection","Drawing Observations and Insights"]],"subsections":[],"title":"Drawing Observations and Insights"},{"cite_extract_rate":0.38095238095238004,"cites":[3857,3853,9099,3854,9100,7754,3874,3863],"content":"\\label{sec:defense}\n\\begin{table}[htbp]\n\t\\caption{Summary of AMD defenses (\\checkmark means applicable, \\protect\\fullcirc means 0, \\protect\\emptycirc means 1, \\protect\\dotcirc means a value in $[0,1]$)}\n\t\\resizebox{1.\\columnwidth}{!}{\n\t\\begin{threeparttable}\n\t\\centering\n\t\\setlength{\\tabcolsep}{0.4em}\n\t\\begin{tabular}{l|c|ccccc|cccc|ccccc|cccc|ccc}\n\t\\hline\n\t\\multicolumn{1}{c|}{\\specialcell{Defense \\\\ (in chronological order)}}& \n\t\\multicolumn{1}{c|}{\\specialcell{Defense \\\\ Objective}} &\n\t\\multicolumn{9}{c|}{\\specialcell{Defense Input}} & \n\t\\multicolumn{5}{c|}{\\specialcell{Assumptions}} & \n\t\\multicolumn{4}{c|}{\\specialcell{Achieved \\\\ Properties}} & \n\t\\multicolumn{3}{c}{\\specialcell{Malware\\\\Detector}}\n\t\\\\\n\t& \\vthead{malware detection} \n\t& \\vthead{$A_1$: Training set $D_{train}$}\n\t& \\vthead{$A_2$: Defense technique}\n\t& \\vthead{$A_3$: Feature set}\n\t& \\vthead{$A_4$: Learning algorithm}\n\t& \\vthead{$A_5$: Response}\n\t& \\vthead{$A_6$: Manipulation set}\n\t& \\vthead{$A_7$: Attack tactic}\n\t& \\vthead{$A_8$: Attack technique}\n\t& \\vthead{$A_9$: Adversarial example set}\n\t& \\vthead{${\\sf IID}$ assumption}\n\t& \\vthead{${\\sf Oracle}$ assumption}\n\t& \\vthead{${\\sf Measurability}$ assumption}\n\t& \\vthead{${\\sf Smoothness}$ assumption}\n\t& \\vthead{${\\sf Invertibility}$ assumption}\n\t& \\vthead{{\\sf RR}: Representation Robustness}\n\t& \\vthead{{\\sf CR}: Classification Robustness}\n\t& \\vthead{{\\sf DR}: Detection Robustness}\n\t& \\vthead{{\\sf TR}: Training Robustness}\n\t& \\vthead{Windows Program}\n\t& \\vthead{Android Package}\n\t& \\vthead{PDF}\n\t\\\\\n\t\\hline\\hline\n \\rowcolor{lightgray!30}\n Biggio et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf EL}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark& & & & &\n \t&\\checkmark&&&\n \t& &\\checkmark\n \t\\\\\n Smutz and Stavrou &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf SE}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t&&\\checkmark&&\n \t& & \\checkmark \n \t\\\\\n \\rowcolor{lightgray!30}\n Zhang et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf RF}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&\\checkmark&\\checkmark&&\n \t\\checkmark&\\checkmark&&&\n \t& & \\checkmark\n \t\\\\\n Demontis et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf WR}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark& &\\checkmark& & &\n \t&\\checkmark&&&\n \t& \\checkmark &\n \t\\\\\n \t\\rowcolor{lightgray!30}\n \tWang et al. &\n \t\\checkmark&\n \t$D_{train}$ &\n \t${\\sf IT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t&\\checkmark&&&\n \t\\checkmark & &\n \t\\\\\n Grosse et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf WR}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t\\checkmark&\\checkmark&&&\n \t& \\checkmark &\n \t\\\\\n \\rowcolor{lightgray!30}\n Grosse et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf AT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{0} &\n \t\\pie{4} &\n \t\\checkmark& &\\checkmark&&&\n \t\\checkmark&\\checkmark&&&\n \t& \\checkmark &\n \t\\\\\n Chen et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf AT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{0} &\n \t\\pie{4} &\n \t\\checkmark& &\\checkmark&&&\n \t&\\checkmark&&&\n \t\\checkmark & &\n \t\\\\\n \\rowcolor{lightgray!30}\n Khasawneh et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf CD}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{0} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t&\\checkmark&&&\n \t\\checkmark & &\n \t\\\\ \n \tDang et al. &\n \t\\checkmark&\n \t$D_{train}$ &\n \t${\\sf SE}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf LQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t&&\\checkmark&&\n \t& & \\checkmark \n \t\\\\ \n \\rowcolor{lightgray!30}\n Yang et al. &\n \\checkmark&\n \t\\specialcell{$D^\\ast_{train}$} &\n \t${\\sf AT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pieg &\n \t\\checkmark&&&&&\n \t&&\\checkmark&&\n \t& \\checkmark &\n \t\\\\\n Yang et al. &\n \\checkmark&\n \t\\specialcell{$D_{train}$} &\n \t${\\sf SE}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{0} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t&&\\checkmark&&\n \t& \\checkmark &\n \t\\\\\n \\rowcolor{lightgray!30}\n Chen et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf RF}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark& &\\checkmark&\\checkmark& &\n \t\\checkmark&\\checkmark&&&\n \t& \\checkmark & \n \t\\\\\n Incer et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf VL}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark& & & & &\n \t\\checkmark&\\checkmark&\\checkmark& & \n \t\\checkmark& & \n \t\\\\ \n \t\\rowcolor{lightgray!30}\n \tChen et al. ~ &\n \t\\checkmark&\n \t$D_{train}$ &\n \t${\\sf SE}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t&&\\checkmark&&&\n \t&&&\\checkmark&\n \t&\\checkmark &\n \t\\\\\n Al-Dujaili et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf AT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&& &\n \t&&\\checkmark&&\n \t\\checkmark & &\n \t\\\\\n \\rowcolor{lightgray!30}\n Chen et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf IT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t&\\checkmark&&&\n \t& \\checkmark &\n \t\\\\\n Jordan et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf RF}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t&&&&&\n \t\\checkmark&\\checkmark&\\checkmark&&\n \t& & \\checkmark\n \t\\\\ \n \\rowcolor{lightgray!30}\n Li et al. &\n \\checkmark&\n \t$D_{train}$ &\n \t${\\sf AT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{4} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&\\checkmark&\\checkmark& &\n \t\\checkmark&\\checkmark&&&\n \t\\checkmark & &\n \t\\\\\n \tTong et al. &\n \t\\checkmark&\n \t$D_{train}$ &\n \t${\\sf RF}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&&&\n \t\\checkmark&\\checkmark&\\checkmark&&\n \t& & \\checkmark\n \t\\\\ \n \t\\rowcolor{lightgray!30}\n \tLi and Li &\n \t\\checkmark&\n \t$D_{train}$ &\n \t${\\sf AT}$ &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&&& &\n \t&&\\checkmark&&\n \t& \\checkmark &\n \t\\\\\n Chen et al. &\n \\checkmark&\n $D_{train}$ &\n ${\\sf VL}$ &\n $S$ &\n $F_\\theta$ &\n {\\sf FQ} &\n \\pie{0} & \n \\pie{0} & \n \\pie{4} &\n \\pie{4} &\n \\checkmark& &\\checkmark & \\checkmark& &\n \\checkmark&\\checkmark& & & \n & & \\checkmark\n \\\\\n \\rowcolor{lightgray!30}\n Li et al. & \n \t\\checkmark &\n \t$D_{train}$ &\n \t\\small{{\\sf EL}+{\\sf AT}+{\\sf RF}} &\n \t$S$ &\n \t$F_\\theta$ &\n \t{\\sf FQ} &\n \t\\pie{0} & \n \t\\pie{0} & \n \t\\pie{4} &\n \t\\pie{4} &\n \t\\checkmark&&\\checkmark&& &\n \t&&\\checkmark&&\n \t& \\checkmark &\n \t\\\\\n \\hline\n\t\\end{tabular}\n\t\\begin{tablenotes}\n\t\t\\small\n\t\t\\item $D^\\ast_{train}$ contains $D_{train}$ and a portion of $\\mathcal{A}$'s adversarial examples.\n\t\\end{tablenotes}\n \\end{threeparttable}\n\t}\n\t\\label{tbl:defenses}\n\\end{table}","id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","level":"subsection","origin_cites_number":21,"parent_id":"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"]],"subsections":["6f9fa41a-b356-42e4-b332-196f10febba0","1e7bc569-0fe6-4f78-b486-391ed3ea051e","272e13c4-c17a-4f6c-ab41-e2e83ef3c588","89604438-0b71-470a-93c0-468dd4ea65af","48b02090-502b-4798-92fb-c646a315077a","a5b08aa1-d4f9-4077-b1e1-6b8c15efc187","45225fbe-29bd-4000-8b60-553581208d2f","362ab0d6-2e4b-4607-b141-4a89fdd71ebf","7446c022-f8d7-4a15-911d-d2f735cce556"],"title":"Systematizing Defense Literature"},{"cite_extract_rate":0.5,"cites":[937,7754],"content":")} \\label{sec:defense-el}\nBiggio et al. propose a one-and-a-half-class SVM classifier against evasion attacks, by leveraging an interesting observation (i.e., decision boundaries of one-class SVM classifiers are tighter than that of two-class SVM classifiers) to facilitate outlier detection. Specifically, the authors propose an ensemble of a two-class classifier and two one-class classifiers, and then combine them using another one-class classifier. \nThe defense can enhance PDF malware detectors against gradient-based attacks , which can be characterized as $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(1,1,1,1,1|\\mathbf{M},{\\sf OP2}, {\\sf GO},\\mathcal{X}_\\mathbf{M})$. However, the defense cannot thwart attacks incurring large perturbations. Independent of this study, other researchers propose using the random subspace and bagging techniques to enhance SVM-based malware detectors, dubbed Multiple Classifier System SVM (MCS-SVM), which leads to evenly distributed weights .\nThese defenses work under the {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf EL},S,F_\\theta,{\\sf FQ}|0,1,0,0)$ and achieves the {\\sf CR} property.","id":"6f9fa41a-b356-42e4-b332-196f10febba0","level":"subsubsection","origin_cites_number":4,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Ensemble Learning ({\\sf EL"]],"subsections":[],"title":"Defenses using Ensemble Learning ({\\sf EL"},{"cite_extract_rate":0.33333333333333304,"cites":[314,7754],"content":")} \\label{sec:attack-wr}\nDemontis et al. propose enhancing the Drebin malware detector $\\varphi(\\mathbf{x})=\\mathbf{w}^\\top\\mathbf{x} + b$ by using box-constraint weights. The inspiration is that the classifier's sensitivity to perturbations based on the $\\ell_1$ norm is bounded by the $\\ell_\\infty$ norm of the weights. This defense\nhardens the Drebin detector against a mimicry attack with input $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,0,1,0,0|\\mathbf{M},{\\sf BE},{\\sf MI},\\mathcal{X}_\\mathbf{M})$, obfuscation attack with input $(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,0,0,0,$ $0|\\mathcal{M},{\\sf BE},-, \\mathcal{Z}_\\mathcal{M})$, and the attack that modifies important features with input $(a_1,\\cdots,a_5|A_6,\\cdots,$ $A_9)=(1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$, here `$-$' means inapplicable. Experimental results show this defense outperforms MSC-SVM . The defense works under the {\\sf IID}, {\\sf Oracle} and {\\sf Measurability} assumptions with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf WR},S,F_\\theta,{\\sf FQ}|1,1,0,0)$ and achieves {\\sf CR}.\nGrosse et al. investigate how to apply two defense techniques known as distillation and retraining to enhance the DNN-based Drebin malware detector. The distillation technique can decrease a model's generalization error by leveraging a teacher to relabel the training data represented by real-value vectors (rather than one-hot encoding). It uses retraining to tune a learned model with respect to an augmented training set with adversarial examples. Both defenses are estimated against a variant of JSMA and can be characterized by their input as $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(0,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf SF}, \\mathcal{X}_\\mathbf{M})$. Experimental results show the two defenses achieve limited success. The defense based on the distillation technique works under the {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf WR},S,F_\\theta,{\\sf FQ}|0,1,0,0)$ and achieves the {\\sf RR} and {\\sf CR} properties. The defense based on the retraining technique works under the {\\sf IID} and {\\sf Measurability} assumptions with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ}|1,$ $1,1,0)$ and achieves the {\\sf RR} and {\\sf CR} properties.","id":"1e7bc569-0fe6-4f78-b486-391ed3ea051e","level":"subsubsection","origin_cites_number":6,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Weight Regularization ({\\sf WR"]],"subsections":[],"title":"Defenses using Weight Regularization ({\\sf WR"},{"cite_extract_rate":0.5,"cites":[3875,9100,7754,9099,3854],"content":")}\nChen et al. adapt a generic retraining framework proposed in the AML context to enhance linear malware detectors. The defense uses a label smoothness regularization technique to mitigate the side-effect of adversarial training . The defense is evaluated using Windows malware detectors against ``feature selection''-based evasion attacks, which can be characterized as $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$. The defense works under the {\\sf IID} and {\\sf Measurability} assumptions and can be characterized as $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ}|1,1,1,0)$, while assuring the {\\sf CR} property.\nYang et al. propose a defense against genetic algorithm-based evasion attacks that can be characterized as $(A_1,\\cdots,A_5|a_6,$ $\\cdots,a_9)=(0,0,0,0,1|\\mathcal{M},{\\sf BE},{\\sf HS},\\mathcal{Z}_\\mathcal{M})$. The defense uses three techniques: adversarial training, sanitizing examples, and weight regularization . The adversarial training uses one half of $\\mathcal{A}$'s adversarial examples. The defense works under the {\\sf IID} assumption and can be characterized as\n$(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D'_{train},{\\sf AT},S,F_\\theta,{\\sf FQ}|0,1,0,*)$, where $D'_{train}$ is the union of $D_{train}$ and a portion (e.g., one half) of $\\mathcal{A}$'s adversarial examples. The defense of sanitizing examples is learned from manipulations used by the attacker and works under the IID assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf SE},S,F_\\theta,{\\sf FQ}|1,1,1,0)$. Both defenses achieve the {\\sf DR} property. The defense of wight regularization is reviewed in Section \\ref{sec:attack-wr}.\nAl-Dujaili et al.~ adapt the idea of minmax adversarial training (proposed in the AML context) to enhance DNN-based malware detectors. In this defense, the inner-layer optimization generates adversarial files by maximizing the classifier's loss function; the outer-layer optimization searches for the parameters $\\theta$ (of DNN $F_\\theta$) that minimize the classifier's loss with respect to the adversarial files.\nThe defense enhances Windows malware detectors against attacks with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(0,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$. \nExperimental results show that malware detectors that are hardened to resist one attack may not be able to defend against other attacks. By observing this phenomenon, researchers propose using a mixture of attacks to harden DNN-based malware detectors .\nThe defense works under the {\\sf IID} assumption and can be characterized as $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ}|1,1,0,0)$. The defense assures the {\\sf DR} property. \nLi et al. propose a DNN-based attack-agnostic framework to enhance adversarial malware detectors. The key idea is dubbed adversarial regularization, which enhances malware detectors via the (approximately) optimally small perturbation. The framework wins the AICS'2019 adversarial malware classification challenge organized by MIT Lincoln Lab researcher , without knowing anything about the attack. \nThe defense works under the {\\sf IID}, {\\sf Measurability}, and {\\sf Smoothness} assumptions with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ}|0,1,0,0)$ and assures the {\\sf RR} and {\\sf CR} properties. In the extended study , the authors further enhance the framework with 6 defense principles, including ensemble learning, adversarial training, and robust representation learning. The enhanced defense is validated with 20 attacks (including 11 grey-box attacks and 9 white-box attacks) against Android malware detectors. The enhanced defense works under the {\\sf IID} and {\\sf Measurability} assumptions with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf AT}+{\\sf EL}+{\\sf RF},S,F_\\theta,{\\sf FQ}|1,1,0,0)$ and assures the {\\sf DR} property.","id":"272e13c4-c17a-4f6c-ab41-e2e83ef3c588","level":"subsubsection","origin_cites_number":10,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Adversarial Training ({\\sf AT"]],"subsections":[],"title":"Defenses using Adversarial Training ({\\sf AT"},{"cite_extract_rate":0.5,"cites":[3877,3876],"content":")}\nIncer et al. propose using monotonic malware classifiers to defend against evasion attacks, where {\\em monotonic} means\n$\\varphi(\\mathbf{x}) \\leq \\varphi(\\mathbf{x}')$ when ${\\mathbf{x}} \\leq {\\mathbf{x}}'$ . Technically, this can be achieved by using (i) robust features that can only be removed or added but not both and (ii) monotonic classification function (e.g., linear models with non-negative weights). The resulting classifier can thwart any attack that perturbs feature values monotonically. The defense works under the {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf VL},S,F_\\theta,{\\sf FQ}|1,1,0,0)$ and assures the {\\sf RR}, {\\sf CR} and {\\sf DR} properties.\nChen et al. propose a defense to enhance PDF malware detectors against evasion attacks, by leveraging the observation that manipulations on PDF files are subtree additions and/or removals. They also propose new metrics for quantifying such structural perturbations. This allows to adapt the \n{\\em symbolic interval analysis} technique proposed in the AML context to enhance the PDF malware detectors. The defense can cope with attacks leveraging small perturbations in the training phase. This defense works under the {\\sf IID}, {\\sf Measurability}, and {\\sf Smoothness} assumptions with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf VL},S,F_\\theta,{\\sf FQ}|1,1,0,0)$ and achieves the {\\sf RR} and {\\sf CR} properties.","id":"89604438-0b71-470a-93c0-468dd4ea65af","level":"subsubsection","origin_cites_number":4,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Verifiable Learning ({\\sf VL"]],"subsections":[],"title":"Defenses using Verifiable Learning ({\\sf VL"},{"cite_extract_rate":0.4,"cites":[3874,3863],"content":")}\nZhang et al. propose leveraging optimal adversarial attacks for feature selection. The defense enhances PDF malware detectors against gradient-based attacks, which can be characterized as $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_{\\mathbf{M}})$. The defense works under the {\\sf IID}, {\\sf Measurability}, and {\\sf Smoothness} assumptions and can be characterized as $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ}|1,1,0,0)$. The defense assures the {\\sf RR} and {\\sf CR} properties.\nTong et al. propose refining features into invariant ones to defend against genetic algorithm-based evasion attacks with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(0,0,0,0,1|\\mathcal{M},{\\sf BE},{\\sf HS},\\mathcal{Z}_\\mathcal{M})$. Experimental results show that adversarial training can be further leveraged to enhance the robustness of the defense. The defense works under {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)$ $=(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ}|1,1,0,0)$, and achieves {\\sf RR}, {\\sf CR} and {\\sf DR}.\nChen et al. propose mitigating evasive attacks by filtering features according to their importance $|w_i|/c_i$ with respect to the linear function $\\varphi(\\mathbf{x})=\\mathbf{w}^\\top\\mathbf{x} + b$, where $x_i$, $w_i$ and $c_i$ denote respectively the $i$th component of $\\mathbf{x}$, $\\mathbf{w}$ and the constraint on manipulation cost $\\mathbf{c}$. The defense enhances Android malware detectors against three attacks: a random attack with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(0,0,1,0,0|\\mathbf{M},{\\sf BE},-,\\mathcal{X}_\\mathbf{M})$, a variant of the mimicry attack with input $(0,0,1,0,0|\\mathbf{M},{\\sf BE},{\\sf MI},$ $\\mathcal{X}_\\mathbf{M})$, and the attack that modifies important features with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(1,1,1,1,1|\\mathbf{M},{\\sf BE},{\\sf SF},$ $\\mathcal{X}_\\mathbf{M})$, where `$-$' means inapplicable. The defense works under the {\\sf IID}, {\\sf Measurability}, and {\\sf Smoothness} assumptions with input $(A_1,\\cdots,A_5|a_6,$ $\\cdots,a_9)=(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ}|1,1,0,0)$ and achieves {\\sf RR} and {\\sf CR}.\nJordan et al. propose a robust PDF malware detector against evasion attacks by interpreting JavaScript behaviors using static analysis. A PDF file is classified as malicious when it calls a vulnerable API method or when it exhibits potentially malicious or unknown behaviors. The defense is validated against the {\\em reverse mimicry} attack with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(0,0,0,0,0|\\mathcal{M},{\\sf BE},{\\sf MI},\\mathcal{Z}_\\mathcal{M})$. The defense has input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ}|1,$ $1,0,0)$ and achieves {\\sf RR}, {\\sf CR} and {\\sf DR}.","id":"48b02090-502b-4798-92fb-c646a315077a","level":"subsubsection","origin_cites_number":5,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Robust Features ({\\sf RF"]],"subsections":[],"title":"Defenses using Robust Features ({\\sf RF"},{"cite_extract_rate":0.5,"cites":[892,3853],"content":")}\nWang et al. propose the {\\em random feature nullification} to enhance DNN-based malware detectors against the attack of Fast Gradient Sign Method (FGSM) by nullifying (or dropping) features randomly in both training and testing phases. This offers a probabilistic assurance in preventing a white-box attacker from deriving adversarial files by using gradients of the loss function with respect to the input. \nThe defense enhances Windows malware detectors against the FGSM attack with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(0,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$. The defense works under {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,$ $\\cdots,a_9)=(D_{train},{\\sf IT},S,F_\\theta,{\\sf FQ}|0,1,0,0)$ and achieves {\\sf CR}.\nDroidEye defends Android malware detectors against evasion attacks by quantizing binary representations, namely transforming binary representations into real values and then using compression to reduce the effect of adversarial manipulations. The defense enhances linear malware detectors against a ``feature selection''-based attack with input\n$(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(1,1,1,1,1|\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$ and the FGSM attack with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(0,1,1,1,$ $1|\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$ . The defense works under {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)$ $=(D_{train},{\\sf IT},S,F_\\theta,{\\sf FQ}|0,1,0,0)$ and achieves {\\sf CR}.","id":"a5b08aa1-d4f9-4077-b1e1-6b8c15efc187","level":"subsubsection","origin_cites_number":4,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Input Transformation ({\\sf IT"]],"subsections":[],"title":"Defenses using Input Transformation ({\\sf IT"},{"cite_extract_rate":0,"cites":[],"content":")}\nKhasawneh et al. propose randomizing classifiers (i.e., using one randomly chosen from a pool of classifiers that use heterogeneous features) to defend against transfer attacks. The defense is validated against an attack which perturbs important features with input\n$(a_1,\\dots,a_5|a_6,\\cdots,a_9)=(0,0,*,0,1|\\mathcal{M},{\\sf BE},{\\sf SF},\\mathcal{Z}_\\mathcal{M})$.\nThe defense works under the {\\sf IID} assumption with input $(a_1,\\dots,a_5|a_6,\\cdots,a_9)=(D_{train},{\\sf CD},S,F_\\theta,$ ${\\sf FQ}|0,1,1,0)$ and achieves the {\\sf CR} property.","id":"45225fbe-29bd-4000-8b60-553581208d2f","level":"subsubsection","origin_cites_number":1,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Classifier Randomization ({\\sf CD"]],"subsections":[],"title":"Defenses using Classifier Randomization ({\\sf CD"},{"cite_extract_rate":0.2,"cites":[3857],"content":")}\nSmutz and Stavrou propose an ensemble classifier to defend against grey-box evasion attacks by returning classification results as benign, uncertain and malicious according to the voting result (e.g., $[0\\%,25\\%]$ classifiers saying malicious can be treated as benign, $[25\\%,75\\%]$ saying malicious can be treated as uncertain, and $[75\\%,100\\%]$ saying malicious can be treated as malicious). The defense enhances a PDF malware detector against three types of evasion attacks: gradient-based attack with input $(a_1,\\ldots,a_5|A_6,\\cdots,A_9) =(1,0,*,*,0|\\mathcal{M},{\\sf OE2},{\\sf TR},\\mathcal{Z}_\\mathbf{M})$, mimicry attack with input $(1,0,*,*,0|\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathbf{M})$, and reverse mimicry attack with input $(0,0,0,0,0|\\mathcal{M},{\\sf BE},{\\sf MI},\\mathcal{Z}_\\mathbf{M})$ .\nThe defense works under the {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)$ $=(D_{train},{\\sf SE},S,F_\\theta,{\\sf FQ}|0,1,0,0)$ and achieves {\\sf DR}.\nDang et al. propose enhancing PDF malware detectors by lowering the classification threshold $\\tau$ and restricting the maximum query times, rendering genetic algorithm-based evasion attacks harder to succeed. This defense works under the {\\sf IID} assumption with input $(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf SE},S,F_\\theta,{\\sf LQ}|0,1,0,0)$ and achieves {\\sf DR}.\nChen et al. investigate defending Android malware detectors against poisoning attacks with input $(a_1,\\ldots,a_5|A_6,\\cdots,A_9) =(1,1,1,1,1|\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M})$\nThe idea is to filter adversarial files that are distant from non-adversarial ones, where distance is measured by the Jaccard index, Jaccard-weight similarity and cosine similarity.\nThe defense works under the {\\sf Measurability} assumption with input \n$(A_1,\\cdots,A_5|a_6,\\cdots,a_9)=(D_{train},{\\sf SE},S,F_\\theta,$ ${\\sf FQ}|0,1,0,0)$\nand achieves {\\sf TR}.","id":"362ab0d6-2e4b-4607-b141-4a89fdd71ebf","level":"subsubsection","origin_cites_number":5,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Defenses using Sanitizing Examples ({\\sf SE"]],"subsections":[],"title":"Defenses using Sanitizing Examples ({\\sf SE"},{"cite_extract_rate":0.5,"cites":[966],"content":"Summarizing the preceding discussions, we draw the following observations.\n(i) Most studies focus on black-box defenses (i.e., the defender knows little about the attacker), which is against the principle of ``knowing yourself and knowing your enemy\". (ii) Most studies focus on defenses against evasion attacks rather than poisoning attacks. (iii) There is no silver bullet defense against evasion attacks or poisoning attacks, at least for now. (iv) Sanitizing adversarial files as outliers is effective against black-box and grey-box attacks, but not white-box attacks. (v) The security properties achieved by defenses have been evaluated empirically rather than rigorously proven (despite that provable security is emerging on the small degree of perturbations; see for example ). (vi) There is no theoretical evidence to support that the effectiveness of defense tactics on the training set (e.g., adversarial training and verifiable learning) can generalize to other adversarial examples.\nIn addition, we draw the following insights:\n\\begin{insight}\n(i) Effective defenses often require the defender to know the attacker's manipulation set. In the real world, it is hard to achieve this, explaining from one perspective why it is hard to design effective defenses.\n(ii) The effectiveness of adversarial training depends on the defender's capability in identifying the most powerful attack.\n\\end{insight}\n\\ignore{\nIn order to draw further insights, we train another {\\em random forest} model from the systematized structure data presented in Table \\ref{tbl:defenses}. We use attributes of defense objective, defense input, and assumptions as features and treat the achieved properties as the labels to learn. The data preprocessing is similar to insight learning from AMD attacks. This leads to a dataset of 22 defenses, each of which has 15 dimensions. \n\\begin{figure}[!htbp]\n\t\\centering\n\t\\scalebox{0.25}{\n\t\\includegraphics{figures/defense-important-feat.eps}\n\t}\n\\caption{Insights learning from systematized AMD defenses via attribute importance.\n}\n\t\\label{fig:defense-attribute}\n\\end{figure}\nFigure \\ref{fig:defense-attribute} shows the most important features on AMD defenses, from which we make the following observations. (i) Defense technique is the most important feature in determining the empirically achieved security property. This can be explained because ${\\sf EL},{\\sf IT},{\\sf CD}, {\\sf VL}, {\\sf WR}$ enhance the classifier against small manipulations to achieve the {\\sf CR} property; {\\sf SE} usually achieves the {\\sf DR} property against large manipulations, which could be detected as outliers; ${\\sf AT},{\\sf RF}$ enhance classifiers against certain attacks specified by the defender. (ii) The {\\sf Measurability} assumption is the second important feature.\n(iii) Manipulation set is the third important feature because knowing this attribute enables the defender to specify {\\sf RF} and {\\sf AT} techniques, which can enhance classifiers to achieve empirical {\\sf RR}, {\\sf CR} and {\\sf DR} properties.\n\\begin{insight}[Insights automatically learned]\nTechnique-wise, defense technique largely determines what security properties can be (empirically) achieved and knowing attacker's manipulation set is critical to defender's success.\n\\end{insight}\n}","id":"7446c022-f8d7-4a15-911d-d2f735cce556","level":"subsubsection","origin_cites_number":2,"parent_id":"ecf055b9-207b-4f7f-9c11-8bb57d05480e","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing Defense Literature"],["subsubsection","Drawing Observations and Insights"]],"subsections":[],"title":"Drawing Observations and Insights"},{"cite_extract_rate":0.4,"cites":[3867,3857,6172,3864,7756,3862,937,3871,3853,9099,3854,7754,3874,3863,3861,7154],"content":"\\afterpage{\n\\begin{rotatepage}\n\\begin{figure*}[ht!]\n\\centering\n\\rotatebox{90}{\n\\scalebox{0.8}{\n\\begin{tikzpicture}\n\\node(rbot1)[label=below:{}, label=left:{},align=center] at (20,3.6) {$ (*,0,1,*,0)$\\\\$ (1,0,1,0,0)$\\\\$ (1,0,*,*,0)$};\n\\node(rbot1_attr)[label=below:{},align=center] at (20,2.6) {$(\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M})$ \\\\ $(\\mathcal{M},{\\sf BP},{\\sf SF},\\mathcal{Z}_\\mathcal{M})$};\n\\node(rbot1_tag)[label=below:{},align=center] at (20,2.1) {};\n\\node(rbot)[label=below:{}, label=left:{}] at (20,1) {$ (1,1,1,1,1)$};\n\\node(rbot_attr)[label=below:{},align=center] at (20,0.6) {$(\\mathbf{M},{\\sf OE2},{\\sf MI},\\mathcal{X})$};\n\\node(rbot_tag)[label=below:{},align=center] at (20,0.2) {};\n\\node(rbot0)[label=below:{}, label=left:{}] at (20,-2) {$ (1,1,1,1,1)$};\n\\node(rbot0_attr)[label=below:{},align=center] at (20,-2.4) {$(\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$};\n\\node(rbot0_refer)[label=below:{},align=center] at (20,-2.8) {};\n\\node(lbot3)[label=below:{}, label=left:{}] at (0,3) {$ (0,0,0,0,1)$};\n\\node(lbot3_attr)[label=below:{},align=center] at (0,2.6) {$(\\mathcal{M},{\\sf BE},{\\sf HS},\\mathcal{Z}_\\mathcal{M})$};\n\\node(lbot3_refer)[label=below:{},align=center] at (0,2.2) {};\n\\node(lbot2)[label=below:{}, label=left:{}] at (0,1) {$ (0,0,0,0,0)$};\n\\node(lbot2_attr)[label=below:{},align=center] at (0,0.6) {$(\\mathcal{M},{\\sf BE},{\\sf HS},\\mathcal{Z}_\\mathcal{M})$};\n\\node(lbot2_refer)[label=below:{},align=center] at (0,0.2) {};\n\\node(lbot1)[label=below:{}, label=left:{}] at (0,-1) {$ (0,0,0,0,0)$};\n\\node(lbot1_attr)[label=below:{},align=center] at (0,-1.4) {$(\\mathcal{M},{\\sf BE},{\\sf MI},\\mathcal{Z}_\\mathcal{M})$};\n\\node(lbot1_refer)[label=below:{},align=center] at (0,-1.8) {};\n\\node(lbot0)[label=below:{}, label=left:{}] at (0,-3) {$ (0,0,0,0,0)$};\n\\node(lbot0_attr)[label=below:{},align=center] at (0,-3.4) {$(\\mathcal{M},{\\sf BE},{\\sf MI},\\mathcal{Z}_\\mathcal{M})$};\n\\node(lbot0_refer)[label=below:{},align=center] at (0,-3.8) {};\n\\node(above7_attr)[label=below:{},align=center] at (20,6) {$(\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$};\n\\node(above7)[label=below:{}, label=left:{}] at (20,6.4) {$ (0,1,1,1,1)$};\n\\node(above7_refer)[label=below:{},align=center] at (20,6.8) {};\n\\node(above6_attr)[label=below:{},align=center] at (16.8,6) {$(\\mathbf{M},{\\sf OE2},{\\sf MS},\\mathcal{X}_\\mathbf{M})$};\n\\node(above6)[label=below:{}, label=left:{}] at (16.8,6.4) {$ (0,1,1,1,1)$};\n\\node(above6_refer)[label=below:{},align=center] at (16.8,6.8) {};\n\\node(above5_attr)[label=below:{},align=center] at (14,6) {$(\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$};\n\\node(above5)[label=below:{}, label=left:{}] at (14,6.4) {$ (0,1,1,1,1)$};\n\\node(above5_refer)[label=below:{},align=center] at (14,6.8) {};\n\\node(above4_attr)[label=below:{},align=center] at (11.2,6) {$(\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$}; \n\\node(above4)[label=below:{}, label=left:{}] at (11.2,6.4) {$ (0,1,1,1,1)$};\n\\node(above4_refer)[label=below:{},align=center] at (11.2,6.8) {};\n\\node(above3_attr)[label=below:{},align=center] at (8.4,6) {$(\\mathcal{M},{\\sf OE2},{\\sf MI},\\mathcal{Z}_\\mathcal{M})$};\n\\node(above3)[label=below:{}, label=left:{}] at (8.4,6.4) {$ (1,1,1,1,1)$};\n\\node(above3_refer)[label=below:{},align=center] at (8.4,6.8) {};\n\\node(above2_attr)[label=below:{},align=center] at (5.6,6) {$(\\mathcal{M},{\\sf BE},{\\sf HS},\\mathcal{Z}_\\mathcal{M})$};\n\\node(above2)[label=below:{}, label=left:{}] at (5.6,6.4) {$ (0,0,0,0,1)$};\n\\node(above2_refer)[label=below:{},align=center] at (5.6,6.8) {};\n\\node(above1_attr)[label=below:{},align=center] at (2.8,6) {$(\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$};\n\\node(above1)[label=below:{}, label=left:{}] at (2.8,6.4) {$ (0,0,1,*,0)$};\n\\node(above1_refer)[label=below:{},align=center] at (2.8,6.8) {};\n\\node(above0_attr)[label=below:{},align=center] at (0,5.8) {$(\\mathbf{M},{\\sf OE1},{\\sf GO},\\mathcal{X}_\\mathbf{M})$ \\\\ $(\\mathbf{M},{\\sf OE2},{\\sf SF},\\mathcal{X}_\\mathbf{M})$ };\n\\node(above0)[label=below:{}, label=left:{}] at (0,6.4) {$ (0,0,1,0,0)$};\n\\node(above0_refer)[label=below:{},align=center] at (0,6.8) {};\n\\node(below7)[label=below:{}, label=left:{}] at (20,-6.4) {$ (1,0,1,0,0)$};\n\\node(below7_attr)[label=below:{},align=center] at (20,-6) {$(\\mathbf{M},{\\sf OP},{\\sf GO},\\mathcal{X}_\\mathbf{M})$};\n\\node(below7_refer)[label=below:{},align=center] at (20,-6.8) {};\n\\node(below6)[label=below:{}, label=left:{}] at (16.8,-6) {$ (0,0,0,1,1)$};\n\\node(below6_attr)[label=below:{},align=center] at (16.8,-6.4) {$(\\mathbf{M},{\\sf BP},{\\sf SF},\\mathcal{X}_\\mathbf{M})$};\n\\node(below6_tag)[label=below:{},align=center] at (16.8,-6.8) {};\n\\node(below5)[label=below:{}, label=left:{}] at (14,-6) {$ (0,0,1,0,1)$};\n\\node(below5_attr)[label=below:{},align=center] at (14,-6.4) {$(\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathcal{M})$};\n\\node(below5_refer)[label=below:{},align=center] at (14,-6.8) {};\n\\node(below4)[label=below:{}, label=left:{}] at (11.2,-6) {$ (0,0,1,0,1)$};\n\\node(below4_attr)[label=below:{},align=center] at (11.2,-6.4) {$(\\mathbf{M},{\\sf BE},{\\sf GM},\\mathcal{X}_\\mathbf{M})$};\n\\node(below4_refer)[label=below:{},align=center] at (11.2,-6.8) {};\n\\node(below3)[label=below:{}, label=left:{}] at (8.4,-6) {$ (0,0,1,*,0)$};\n\\node(below3_attr)[label=below:{},align=center] at (8.4,-6.4) {$(\\mathbf{M},{\\sf OE2},{\\sf GO},\\mathcal{X}_\\mathbf{M})$};\n\\node(below3_refer)[label=below:{},align=center] at (8.4,-6.8) {};\n\\node(below2)[label=below:{}, label=left:{}] at (5.6,-6) {$ (0,0,*,0,1)$};\n\\node(below2_attr)[label=below:{},align=center] at (5.6,-6.4) {$(\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathbf{M})$};\n\\node(below2_refer)[label=below:{},align=center] at (5.6,-6.8) {};\n\\node(below1)[label=below:{}, label=left:{}] at (2.8,-6) {$ (0,0,0,0,1)$};\n\\node(below1_attr)[label=below:{},align=center] at (2.8,-6.4) {$(\\mathcal{M},{\\sf OE2},{\\sf GO},\\mathcal{Z}_\\mathcal{M})$};\n\\node(below1_refer)[label=below:{},align=center] at (2.8,-6.8) {};\n\\node(below0)[label=below:{}, label=left:{}] at (0,-6) {$ (0,0,*,0,0)$};\n\\node(below0_attr)[label=below:{},align=center] at (0,-6.4) {$(\\mathcal{M},{\\sf BE},{\\sf TR},\\mathcal{Z}_\\mathbf{M})$};\n\\node(below0_refer)[label=below:{},align=center] at (0,-6.8) {};\n\\fill [lightgray!30] (1.5,5) rectangle (18.6,-5);\n\\node(mid4)[label=below:{},align=left] at (15.5,0) {:$(D_{train},{\\sf SE},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ})$\\\\\n:$(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ})$\n};\n\\node(mid4-attr)[align=center] at (15.5,-0.85) {$(1,1,1,0)$};\n\\node(mid3)[label=below:{$(1,1,0,0)$},align=left] at (10,1.88) {\n:$(D_{train},{\\sf WR},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf RF},S,F_\\theta,{\\sf FQ})$\n}; \n\\node(mid2)[label=below:{$(0,1,1,0)$},align=left] at (10,-1.5) {:$(D_{train},{\\sf CD},S,F_\\theta,{\\sf FQ})$};\n\\node(mid1)[label=below:{$(0,1,0,*)$},align=center] at (10,-3) {:$(D^\\ast_{train},{\\sf AT},S,F_\\theta,{\\sf FQ})$};\n\\node(mid0)[label=below:{$(0,1,0,0)$},align=left] at (4.5,0) {\n:$(D_{train},{\\sf SE},S,F_\\theta,{\\sf LQ})$\\\\\n:$(D_{train},{\\sf AT},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf IT},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf WR},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf SE},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf SE},S,F_\\theta,{\\sf FQ})$ \\\\\n:$(D_{train},{\\sf IT},S,F_\\theta,{\\sf FQ})$ \\\\ \n:$(D_{train},{\\sf EL},S,F_\\theta,{\\sf FQ})$\n};\n\\node(det3)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (18.2,3) {};\n\\node(det2)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (18.1,4.65) {};\n\\node(dnn1)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (15.2, 3.7) {DNN-based malware \\\\ detector };\n\\node(dnn1_label)[align=center,color=blue!50] at (15.2,2.8) {$(D_{train},\\emptyset,S,F_\\theta,{\\sf FQ})$\\\\$(0,0,0,0)$};\n\\node(apk1)[label=right:{},align=center, fill=blue!20,fill opacity=0.8] at (6, 3.9) {Drebin };\n\\node(apk1_label) [align=center,color=blue!50] at (6.2,3.2) {$(D_{train},\\emptyset,S,F_\\theta,{\\sf FQ})$\\\\$(0,0,0,0)$};\n\\node(det1)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (3.05,3.9) {};\n\\node(pdf2)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (3.6, 3.2) {PDFrate };\n\\node(pdf2_label)[align=center,color=blue!50] at (3.6,2.4) {$(D_{train},\\emptyset,S,F_\\theta,{\\sf FQ})$\\\\$(0,0,0,0)$};\n\\node(pdf1)[label=below:{},label=below:{},align=left, fill=blue!20,fill opacity=0.8] at (3.6, -3.8) {PDFrate };\n\\node(pdf1_label)[align=center,color=blue!50] at (3.6,-4.6) {$(D_{train},\\emptyset,S,F_\\theta,{\\sf FQ})$\\\\$(0,0,0,0)$};\n{\n\\draw [-latex,line width=1] (3.3,-5.7) -- (3.3, -5) -- (2.3, -5) -- (2.3, -5.7);\n\\draw [-latex,line width=1] (11.7,-5.7) -- (11.7, -5) -- (10.7, -5) -- (10.7, -5.7);\n\\draw [-latex,line width=1] (14.5,-5.7) -- (14.5, -5) -- (13.5, -5) -- (13.5, -5.7);\n\\draw [-latex,line width=1] (20.5,-5.7) -- (20.5,-5) -- (19.5,-5) -- (19.5,-5.7);\n\\draw [-latex,line width=1] (19.5,1.2) -- (19.5,1.4) -- (18.6,1.4) -- (18.6,0) -- (19.5,0) -- (19.5,0.4); \n\\draw [-latex,line width=1] (above0_attr) -- (0, 3.9) -- (det1);\n\\draw [-latex,line width=1] (20,5.7) -- (20,4.65) -- (det2);\n\\draw [-latex,line width=1] (19,3.) -- (det3);\n\\draw [-latex,line width=1] (6.1,-5.7) -- (6.1, -5) -- (5.1, -5) -- (5.1, -5.7);\n\\draw [-latex,line width=1] (17.3,-5.7) -- (17.3, -5) -- (16.3, -5) -- (16.3, -5.7);\n\\ignore{\n\\draw [-latex,line width=1] (6.4,-0.2)node[red,right]{}--(5.8,-0.2);\n\\draw [-latex,line width=1] (6.4,0.2)node[red,right]{}--(5.9,0.2);\n\\draw [-latex,line width=1] (9.2,0.7)node[red,above]{}--(9.2,0.2);\n\\draw [-latex,line width=1] (6.4,-3.2)node[red,left]{}--(6.9,-3.2);\n\\draw [-latex,line width=1] (13.6,-3.2)node[red,left]{}--(14.1,-3.2);\n}\n\\draw [-latex,line width=1] (below3) -- (8.4,-5.0) -- (7.25,-5.0) -- (7.25,-1.5) -- (6.25,-1.5);\n\\draw [-latex,line width=1] (7.25,-1.5) -- (7.25,0.2) -- (8.1,0.2);\n\\draw [line width=1] (18.7,-2.4) -- (18.6,-2.4) -- (18.6,-1.05) -- (7.35,-1.05); \n\\draw [-latex,line width=1] (7.15,-1.05) --(6.25,-1.05);\n\\draw [-latex,line width=1] (12.55, -1.05) -- (12.55, 0.65) -- (11.75, 0.65);\n\\tkzDefPoint(7.25,-1.05){c1}\n\\tkzDefPoint(7.15,-1.05){c2}\n\\tkzDrawArc[rotate,color=black,line width=1](c1,c2)(180);\n\\draw [-latex,line width=1] (11.1,5.7) -- (11.1,4.3) -- (13.15,4.3) -- (13.15,3.7) -- (13.6, 3.7); \n\\draw [blue,densely dashdotted,-latex,line width=1] (16.8,3.7) -- (17.5,3.7) -- (17.5,0.1) -- (17.2,0.1); \n\\draw [red,dashed,-latex,line width=1] (11.3,5.7) -- (11.3,5.0) -- (13.9,5.0) -- (13.9,5.7);\n\\draw [-latex,line width=1] (14.1,5.7) -- (14.1,4.3) -- (17.7,4.3) -- (17.7,-0.1) -- (17.2,-0.1); \n\\draw [blue,densely dashdotted,-latex,line width=1] (13.55,0) -- (12.75,0) -- (12.75,1.8) -- (11.6,1.8);\n\\draw [red,dashed,-latex,line width=1] (14.3,5.7) -- (14.3,5.0) -- (16.7,5.0) -- (16.7,5.7);\n\\draw [line width=1] (17.3,5.7) -- (17.3,4.4);\n\\draw [-latex,line width=1] (17.3,4.2) -- (17.3,2) -- (11.6,2);\n\\draw [blue,densely dashdotted,-latex,line width=1] (8.05, 1.9) -- (7.25,1.9) -- (7.25,1.45) -- (8.1,1.45) ;\n\\tkzDefPoint(17.3,4.3){c3}\n\\tkzDefPoint(17.3,4.2){c4}\n\\tkzDrawArc[rotate,color=black,line width=1](c3,c4)(180);\n\\draw [-latex,line width=1] (2.7,5.7) -- (2.7,4.3) -- (4.7,4.3) -- (4.7,3.9) -- (5.2,3.9); \n\\draw [blue,densely dashdotted,-latex,line width=1] (6.9,3.9) -- (7.05,3.9) -- (7.05,3.58) -- (8.1,3.58); \n\\draw [red,dashed,-latex,line width=1] (2.9,5.7) -- (2.9,5.0) -- (8.5,5.0) --(8.5,5.7); \n\\draw [red,dashed,-latex,line width=1] (5.7,5.0) -- (5.7,5.7);\n\\draw [-latex,line width=1] (5.5,5.7) -- (5.5,4.3) -- (8.3,4.3) -- (8.3,3.7); \n\\draw [line width=1] (8.3,5.7) -- (8.3,4.3); \n\\draw [blue,densely dashdotted,-latex,line width=1] (11.9,3.7) -- (12.95,3.7) -- (12.95,0.43) -- (13.55,0.43);\n\\draw [blue,densely dashdotted,-latex,line width=1] (11.9,3.5) -- (12.75,3.5) -- (12.75,2.7) -- (11.8,2.7);\n\\draw [-latex,line width=1] (0.1, 3.1) -- (0.1,3.3) -- (2.5,3.3); \n\\draw [blue,densely dashdotted,-latex,line width=1] (2.4,3.1) -- (1.7,3.1) -- (1.7, 1.55) -- (2.5,1.55); \n\\draw [red,dashed,-latex,line width=1] (-0.1,2.1) -- (-0.1,1.1);\n\\draw [-latex,line width=1] (0.1,1.1) -- (0.1, 1.35) -- (2.5,1.35); \n\\draw [blue,densely dashdotted,-latex,line width=1] (4.7,3.2) -- (8.1,3.2);\n\\draw [-latex,line width=1] (0.1,-5.7) -- (0.1,-5) -- (1.5,-5) -- (1.5,-5) -- (1.5,-3.9) -- (2.5,-3.9); \n\\draw [blue,densely dashdotted,-latex,line width=1] (2.4,-3.7) -- (1.7,-3.7) -- (1.7, -0.75) -- (2.5, -0.75); \n\\draw [red,dashed,-latex,line width=1] (-0.1,-5.7) -- (-0.1,-3.9); \n\\draw [-latex,line width=1](0.1,-2.9) -- (0.1,-2) -- (1.5,-2) -- (1.5,-0.55) -- (2.5,-0.55); \n\\draw [blue,densely dashdotted,-latex,line width=1] (6.4,-0.6) -- (7.05,-0.6) -- (7.05,2.3) -- (8.1,2.3); \n\\draw [red,dashed,-latex,line width=1] (-0.1,-2.9) -- (-0.1,-2); \n\\draw [-latex,line width=1](0.1,-0.9) -- (0.1,-0.55) -- (2.5,-0.55);\n}\n\\draw [-latex,line width=1] (12.3,7.45) -- (12.9,7.45) node[right,align=center]{Waging \\\\ Attack};\n\\draw [blue,densely dashdotted,-latex,line width=1] (14.5,7.45) -- (15.1,7.45);\n\\node [right,align=center] at (15.1,7.45) {Defense \\\\ Escalation};\n\\draw [red,dashed,-latex,line width=1] (17,7.45) -- (17.6,7.45);\n\\node [right,align=center] at (17.6, 7.45) {Attack \\\\ Escalation};\n\\draw (11,7) -- (20,7) -- (20,7.9) -- (11,7.9) -- node [right]{Legend:} (11,7);\n\\end{tikzpicture}\n}\n}\n\\caption{Arms race in AMD attack and defense escalations.}\n\\label{fig:attacks-defenses}\n\\end{figure*}\n\\end{rotatepage}\n}\nFigure \\ref{fig:attacks-defenses} displays AMD attack-defense arms race surrounding three malware detectors: PDFrate, Drebin, and DNN-based detector. For a better visual effect, we group papers that proposed defense methods in terms of a common input $(a_6,\\ldots,a_9)$. For example, we group\n, and together because the defenders in both papers have input $(a_6,\\ldots,a_9)=(1,1,1,0)$, while noting that their input on $(A_1,\\ldots,A_5)$ may or may not be different.\nWe also simplify attack and defense inputs while preserving the {\\em critical information} when an attack (defense) works for multiple inputs. \nFor example, $(a_1,\\ldots,a_5|A_6,\\cdots,A_9)=(0,0,1,0,0|\\mathbf{M},{\\sf OE2}, {\\sf SF}, \\mathcal{X}_\\mathbf{M})$ is the {\\em critical information} for attack input \n$(a_1,\\cdots,a_5|A_6,\\cdots,A_9)=(0,0,1,0,0|\\mathbf{M},{\\sf OE2}, {\\sf SF}, \\mathcal{X}_\\mathbf{M}) \\lor (0,0,1,0,1|\\mathbf{M},{\\sf OE2}, {\\sf SF}, \\mathcal{X}_\\mathbf{M}) \\lor (1,0,1,0,0|\\mathbf{M},{\\sf OE2}, {\\sf SF}, \\mathcal{X}_\\mathbf{M}) \\lor (1,0,1,0,1|\\mathbf{M},{\\sf OE2}, {\\sf SF},\\mathcal{X}_\\mathbf{M})$\nbecause it is the weakest attack input in the partial order formulated by these $(a_1,\\ldots,a_5)$'s. This suggests us to focus on attack input $(0,0,1,0,0|\\mathbf{M},{\\sf OE2}, {\\sf SF},\\mathcal{X}_\\mathbf{M})$ because it is already able to break some defense and automatically implies that a stronger input can achieve the same (while noting some special cases, see discussion in Section \\ref{sec:defense-el}). Multiple defense inputs are simplified in the same manner.\n{\\em Arms race in PDF malware detection}: We summarize two sequences of escalations caused by PDFrate . In one sequence, PDFrate is defeated by transfer attacks, which are realized by gradient-based and mimicry methods against surrogate models . These attacks trigger the defense escalation to an ensemble detector built on top of some diversified classifiers . This defense triggers attack escalation to {\\em reverse mimicry attacks} , \nwhich trigger the defense escalation of using robust hand-crafted features . This defense represents the state-of-the-art PDF malware detector, but still incurs a high false-positive rate. In the other sequence of arms race, PDFrate is defeated by genetic algorithm-based attacks . These attacks trigger the defense escalation to and . The former defense restricts the responses to attacker queries, but can be defeated by the escalated attack that leverages the hill-climbing algorithm (also shown in ). The latter defense uses invariant features to thwart the attacks and represents another state-of-the-art PDF malware detectors.\n{\\em Arms race in android malware detection}: Drebin is defeated by the attack that modifies a limited number of important features , which also proposes the new defense to defeat the escalated attack. This defense triggers the attack escalation to, and is defeated by, the genetic algorithm-based attack and the mimicry-alike attack . The former attack triggers the escalated defense (also presented in ) that leverages attack mutations to detect adversarial examples . The latter attack injects objects in APKs and (in principle) can be defeated by the monotonic classifier . These escalated defenses represent the state-of-the-art Android malware detectors, but still incur a high false-positive rate.\n{\\em Arms race in DNN-based malware detection}: The DNN-based detector triggers four gradient-based evasion attacks presented in , which also hardens the DNN malware detector by using an {\\em minmax adversarial training} instantiation to incorporates the $\\ell_\\infty$ normalized gradient-based attack. This escalated defense triggers the mixture of attacks presented in . The defense of {\\em minmax adversarial training} incorporating a mixture of attacks can defeat a broad range of attacks, but still suffers from the mimicry attack and other mixtures of attacks . \nAs a consequence, there are no effective defenses can thwart all kinds of attacks.\n{\\em Independent arms race}:\nThere are studies that have yet to trigger cascading arms races, including: (i) Studies propose independent attacks and then show how to defeat these attacks.\n(ii) Studies propose attacks to defeat naive malware detectors. (iii) Studies propose defenses to counter some attacks .","id":"76e4c68c-dec2-4af2-b3fa-ce426944cf52","level":"subsection","origin_cites_number":40,"parent_id":"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Systematizing AMD Arms Race"],["subsection","Systematizing AMD Arms Race"]],"subsections":[],"title":"Systematizing AMD Arms Race"},{"cite_extract_rate":0.818181818181818,"cites":[8399,8696,314,892,3878,966,937,7754,7312],"content":"\\label{sec:cd}\n\\noindent{\\bf FRD 1}: {\\em Pinning down the root cause(s) of adversarial malware examples}.\nSpeculations on root cause(s) include: (i) invalidity of the {\\sf IID} assumption because of distribution drifting, namely that testing files and training files are drawn from different distributions );\n(ii) incompetent feature extraction ;\n(iii) high dimensionality of malware representations ;\n(iv) insufficient scale of training data ;\n(v) low-probability ``pockets'' in data manifolds ;\n(vi) linearity of DNNs ; and\n(vii) large curvature of decision boundaries .Although these speculations may be true, more studies are needed in order to (in)validate them. \n\\smallskip\n\\noindent{\\bf FRD 2}: {\\em Characterizing the relationship between transferability and vulnerability}.\nIn the AMD context, an attacker may use a surrogate model to generate adversarial examples and a defender may use a surrogate model for adversarial training. Transferability is related to the extent at which knowledge gained by a surrogate model may be the same as, or similar to, what is accommodated by a target model. The wide use of surrogate models in the AMD context suggests that there may be a fundamental connection between knowledge transferability and model vulnerability.\n\\smallskip\n\\noindent{\\bf FRD 3}: {\\em Investigating adversarial malware examples in the wild}. In the AMD context, it is challenging to generate practical adversarial malware examples to correspond to perturbations conducted in the feature space, owing to realistic constraints. On the other hand, an attacker can directly search for manipulations in the problem space. This may cause large perturbations, putting the value of studies on small perturbations in question. This represents a fundamental open problem that distinguishes the field of AMD from its counterparts in other application settings. \nThis issue is largely unaddressed by assuming that there is an oracle for telling whether manipulated or perturbed features indeed correspond to a malware sample or not.\n\\smallskip\n\\noindent{\\bf FRD 4}: {\\em Quantifying the robustness and resilience of malware detectors}. Robustness and resilience of malware detectors against adversarial examples need to be quantified, ideally with a provable guarantee. For this purpose, one may adapt the reduction-based paradigm underlying the provable security of cryptographic primitives and protocols.\n\\smallskip\n\\noindent{\\bf FRD 5}: {\\em Designing malware detectors with provable robustness and resilience guarantees}.\nHaving understood the root cause(s) of adversarial examples, characterized the effect of transferability, investigated the effectiveness of practical attacks, and designed metrics for quantifying the robustness and resilience of malware detectors, it is imperative to investigate robust malware detectors with provable robustness, ideally as rigorous as what has been achieved in the field of cryptography. In this regard, robust feature extraction, adversarial learning, and verifiable learning are promising candidates for making breakthroughs.\n\\smallskip\n\\noindent{\\bf FRD 6}: {\\em Forecasting the arms race in malware detection}. Arms race is a fundamental phenomenon inherent to the cybersecurity domain. In order to effectively defend against adversarial malware, one approach is to deploy proactive defense, which requires the capability to forecast the arms race between malware writers and defenders. For instance, it is important to predict how attacks will evolve and what kinds of information would be necessary in order to defeat such attacks.","id":"965da080-653e-47f5-949b-de42288cf6a5","level":"section","origin_cites_number":11,"parent_id":"4fe25043-06f8-4577-bd62-3fe42c7ab899","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Future Research Directions (FDRs)"]],"subsections":[],"title":"Future Research Directions (FDRs)"},{"cite_extract_rate":1,"cites":[7155],"content":"\\label{sec:conc}\nWe have presented a framework for systematizing the field of AMD through the lens of assumptions, attacks, defenses and security properties. This paves the way for precisely relating attacks and defenses. We have also shown how to apply the framework to systematize the AMD literature, including the arms race between AMD attacks and defenses. We have reported a number of insights.\nThe study leads to a set of future research directions. In addition to the ones described in Section \\ref{sec:cd}, we mention the following two, which are discussed here because there are rarely studies on these aspects. (i) To what extent explainability (or interpretability) of ML models can be leveraged to cope with adversarial malware examples? It is intuitive that explainability could be leveraged to recognize adversarial examples because they may not be explainable . \n(ii) To what extent uncertainty quantification can be leveraged to cope with adversarial malware examples? If the uncertainty associated with detectors' predictions on adversarial malware examples are inherently and substantially higher than the uncertainty associated with non-adversarial malware examples, this fact can be leveraged to recognize adversarial malware examples.\nFinally, we reiterate that the research community should seek to establish a solid foundation for AMD. Although this foundation can leverage ideas and techniques from AML, the unique characteristics of AMD warrant the need of a unique foundation.\n\\bibliographystyle{ACM-Reference-Format}\n\\bibliography{mal_attacks,mal_defenses,aml,cyber_security,data_mining,mal_detection}\n\\end{document}","id":"82494750-c0aa-44ab-add1-1d1d59481f4a","level":"section","origin_cites_number":1,"parent_id":"4fe25043-06f8-4577-bd62-3fe42c7ab899","prefix_titles":[["title","Arms Race in Adversarial Malware Detection: A Survey"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"4fe25043-06f8-4577-bd62-3fe42c7ab899\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ]\n ],\n \"subsections\": [\n \"74b358c3-1042-4f31-a16f-380254516292\",\n \"a1283bee-37c1-4371-bed3-ceeb8af64d75\",\n \"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87\",\n \"965da080-653e-47f5-949b-de42288cf6a5\",\n \"82494750-c0aa-44ab-add1-1d1d59481f4a\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.32692307692307604,\n \"cites\": [\n 314,\n 7755,\n 7337,\n 893,\n 3856,\n 3857,\n 937,\n 3853,\n 3854,\n 7754,\n 7153,\n 3855,\n 3861,\n 7154,\n 3860,\n 3859,\n 3858\n ],\n \"content\": \"\\\\label{sec:intro}\\nMalware (malicious software) is a big cyber threat and has received a due amount of attention. For instance, Kaspersky reports that 21,643,946 unique malicious files were detected in the year 2018, 24,610,126 in 2019, and 33,412,568 in 2020 .\\nA popular defense against malware is to use {\\\\em signature-based} detectors , where a signature is often extracted by malware analysts from known malware examples. This approach has two drawbacks: signatures are tedious to extract and can be evaded by a range of techniques (e.g., encryption, repacking, polymorphism ). This incompetence has motivated the use of Machine Learning (ML) based malware detectors, which can be automated to some degree and can possibly detect \\\\emph{new} malware examples (via model generalization or knowledge adaptation ). More recently, Deep Learning (DL) has been used for malware detection (see, e.g., ). \\nWhile promising, ML-based malware detectors are vulnerable to attacks known as {\\\\em adversarial examples} . \\nThere are two kinds of attacks. One is {\\\\em evasion attack}, where the attacker perturbs test examples to {\\\\em adversarial examples} to evade malware detectors . The other is {\\\\em poisoning attack}, where the attacker manipulates the training dataset for learning malware detectors . These attacks usher in the new field of Adversarial Malware Detection (AMD) . \\nThe state-of-the-art in AMD is that there are some specific results scattered in the literature but there is no systematic understanding.\\nThis is true despite that there have been attempts at systematizing the related field of Adversarial Machine Learning (AML) , which however cannot be automatically translated to AMD. This is so because malware detection \\nhas three {\\\\em unique} characteristics which are not exhibited by the other application domains (e.g., image or audio processing). (i) There are no common, standard feature definitions because both attackers and defenders can define their own features to represent computer files. As a consequence, attackers can leverage this ``freedom'' in feature definition to craft adversarial examples. (ii) Malware features are often discrete rather than continuous and program files are often highly structured with multiple modalities. This means that arbitrarily perturbing malware files or their feature representations might make the perturbed files no more executable. \\nThis also means that the discrete domain makes perturbation a {\\\\em non-differentiable} and {\\\\em non-convex} task. \\n(iii) Any meaningful perturbation to a malware example or its feature representation must preserve its malicious functionality. For example, the Android Package Kit (APK) requires that the used permissions are publicized in the AndroidManifest.xml, meaning that removing permissions in this manifest file would incur a runtime error. The preceding (ii) and (iii) make both the attacker's and defender's tasks more challenging than their \\ncounterparts where small perturbations are not noticeable (e.g., images).\\n\\\\noindent{\\\\bf Our Contributions}. We propose a conceptual framework for systematizing the AMD field through the lens of {\\\\em assumptions}, {\\\\em attacks}, {\\\\em defenses}, and {\\\\em security properties}. In specifying these, we seek rigorous definitions whenever possible, while noting that these definitions have been scattered in the literature. Rigorous definitions are important because they can serve as a common reference for future studies. The framework allows us to map the known attacks and defenses into some partial order structures and systematize the AMD attack-defense arms race. \\nWe make a number of observations, including: (i) the indiscriminate attack that treats malicious examples as equally important has been extensively investigated, but targeted and availability attacks are much less investigated; (ii) the evasion attack is much more extensively studied than the poisoning attack; (iii) there is no silver-bullet defense against evasion and poisoning attacks; (iv) sanitizing examples is effective against black-box and grey-box attacks, but not white-box attacks; (v) AMD security properties have been evaluated empirically rather than rigorously; (vi) there is no theoretical evidence to support that the effectiveness of defense techniques on the training set can generalize to other adversarial examples.\\nWe draw a number of insights, including: (i) knowing defender's feature set is critical to the success of transfer attacks, highlighting the importance of keeping defender's feature set secret (e.g., by randomizing defender's feature set);\\n(ii) the effectiveness of practical evasion attacks largely depends on the attacker's degree of freedom in conducting manipulations in the problem space (i.e., a small degree of freedom means harder to succeed); (iii) effective defenses often require the defender to know the attacker's manipulation set, explaining from one perspective why it is hard to design effective defenses;\\n(iv) effectiveness of adversarial training depends on the defender's capability in identifying the most powerful attack.\\nFinally, we discuss a number of future research directions, which hopefully will inspire and encourage many researchers to explore them.\\n\\\\smallskip\\n\\\\noindent{\\\\bf Related Work}.\\nThe closely related prior work is Maiorca et al. , which surveys previous studies in adversarial malicious PDF document detection. In contrast, we consider the broader context of AMD and propose novel partial orders to accommodate AMD assumptions, attacks, defenses, and properties.\\nThere are loosely-related prior studies, which survey prior AML studies (but not focusing on AMD), including .\\nFor example, Yuan et al. survey attack methods for generating adversarial examples, while briefly discussing evasion attacks in the AMD context; \\nBarreno et al. propose a taxonomy of AML attacks\\n(causative vs. exploratory attacks, integrity vs. availability attacks, and targeted vs. indiscriminate attacks); Biggio et al. propose a defense framework for protecting \\nSupport Vector Machines (SVMs) from evasion attacks, poisoning attacks and privacy violations; Papernot et al. systematize AML security and privacy with emphasis on demonstrating the trade-off between detection accuracy and robustness.\\n\\\\smallskip\\n\\\\noindent{\\\\bf Paper Outline}.\\nSection \\\\ref{sec:fw} describes our survey and systematization methodology and framework.\\nSection \\\\ref{sec:systematizstion-AMD-studies} applies our framework to systematize the literature AMD studies.\\nSection \\\\ref{sec:cd} discusses future research directions. \\nSection \\\\ref{sec:conc} concludes the paper.\",\n \"id\": \"74b358c3-1042-4f31-a16f-380254516292\",\n \"level\": \"section\",\n \"origin_cites_number\": 52,\n \"parent_id\": \"4fe25043-06f8-4577-bd62-3fe42c7ab899\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:fw}\\n\\\\begin{table}[htbp!]\\n\\t\\\\centering\\\\caption{Main notations used in the paper \\\\label{table:notations}}\\n\\t\\\\begin{adjustbox}{width=0.99\\\\textwidth}\\n\\t\\\\begin{tabular}{l|p{.8\\\\textwidth}}\\n\\t\\t\\\\hline\\n\\t\\tNotation & Meaning\\\\\\\\\\\\hline\\n\\t\\t$\\\\mathbb R$ ($\\\\mathbb R_{+}$) & the set of (positive) real numbers \\\\\\\\\\n\\t\\t$\\\\mathcal{A},\\\\mathcal{I}$ & attacker and defender (treated as algorithms)\\\\\\\\\\n\\t\\t$\\\\mathbb{P}$ & the probability function \\\\\\\\\\n\\t\\t$z, z'\\\\in\\\\mathcal{Z}$ & $\\\\mathcal{Z}$ is example space; $z'$ is obtained by perturbing $z$\\\\\\\\\\n\\t\\t$S$ & defender $\\\\mathcal{I}$'s feature set for representing files \\\\\\\\\\n\\t\\t${\\\\mathbf x}, {\\\\mathbf x}'\\\\in \\\\mathcal X$ & $\\\\mathcal X=\\\\mathbb{R}^d$ is $d$-dimensional feature space; ${\\\\mathbf x},{\\\\mathbf x}' \\\\in \\\\mathcal{X}$ are respectively feature representations of $z,z'\\\\in \\\\mathcal{Z}$ \\\\\\\\\\n\\t\\t$\\\\mathcal{Y}$, $y$ & $\\\\mathcal{Y}$ is the label space of binary classification, $\\\\mathcal{Y}=\\\\{+/1,-/0\\\\}$; $y\\\\in \\\\mathcal{Y}$ \\\\\\\\\\n\\t\\t$\\\\mathcal{D}=\\\\mathcal Z \\\\times \\\\mathcal Y$ & the file-label (i.e., example-label) space \\\\\\\\\\n\\t\\t$D_{train}\\\\subset \\\\mathcal D,n$ & the training set in file-label space; $n=|D_{train}|$ \\\\\\\\\\n\\t\\t$D_{test}$ & the test set in file-label space \\\\\\\\\\n\\t\\t$D_{poison},D'_{poison}$& $D'_{poison}$ is set of adversarial file-label pairs obtained by perturbing non-adversarial files in $D_{poison}\\\\subset\\\\mathcal{D}$ \\\\\\\\\\n\\t\\t$\\\\mathcal{O}(z,z')$ & $\\\\mathcal{O}(z,z'):\\\\mathcal{Z}\\\\times\\\\mathcal{Z}\\\\to\\\\{{\\\\tt true}, {\\\\tt false}\\\\}$ is an oracle telling if two files have the same functionality or not \\\\\\\\\\n\\t\\t$\\\\delta$ & a manipulation for perturbing files {\\\\em with} preserving their functionalities \\\\\\\\\\n\\t\\t$\\\\mathcal{M}$, $\\\\mathcal{Z}_\\\\mathcal{M}\\\\subseteq \\\\mathcal{Z}$ & $\\\\mathcal{M}$ is manipulation set in the problem space;\\n\\t\\t$\\\\mathcal{Z}_\\\\mathcal{M}$ is set of adversarial files generated using $\\\\mathcal{M}$\\\\\\\\\\n\\t\\t$\\\\mathbf{M}$, $\\\\mathcal{X}_\\\\mathbf{M}\\\\subseteq \\\\mathcal{X}$ & $\\\\mathbf{M}$ is feature manipulation set;\\n\\t\\t$\\\\mathcal{X}_\\\\mathbf{M} $ is set of adversarial feature vectors generated using $\\\\mathbf{M}$ \\\\\\\\\\n\\t\\t$\\\\Gamma(z,z')$ & $\\\\Gamma(z,z'):\\\\mathcal Z \\\\times \\\\mathcal{Z} \\\\to \\\\mathbb R_+$ measures the degree of manipulation for perturbing $z \\\\in \\\\mathcal{Z}$ into $z'\\\\in \\\\mathcal{Z}$ \\\\\\\\\\n\\t\\t$C(\\\\mathbf{x},\\\\mathbf{x}')$& $C(\\\\mathbf{x},\\\\mathbf{x}'):\\\\mathcal X \\\\times \\\\mathcal X \\\\to \\\\mathbb R_+$ is the function measuring the cost incurred by changing feature vector $\\\\mathbf{x}$ to $\\\\mathbf{x}'$ \\\\\\\\\\n\\t\\t$\\\\delta_z \\\\in \\\\mathcal{M}$ & $\\\\delta_z$ is a set of manipulations of $z$ w.r.t. $z'$ \\\\\\\\\\n\\t\\t$\\\\delta_{\\\\mathbf{x}} \\\\in \\\\mathbf{M}$ & $\\\\delta_{\\\\mathbf{x}}={\\\\mathbf x}' - {\\\\mathbf x}$ is a perturbation vector of $\\\\mathbf{x}$ w.r.t. $\\\\mathbf{x}'$\\\\\\\\\\n\\t\\t$\\\\phi:\\\\mathcal Z \\\\to \\\\mathcal{X} $ & feature extraction function; $\\\\mathbf{x}\\\\leftarrow\\\\phi(z)$, $\\\\mathbf{x}'\\\\leftarrow\\\\phi(z')$ \\\\\\\\\\n\\t\\t$\\\\varphi,f$ & $\\\\varphi:\\\\mathcal{X} \\\\to \\\\mathbb{R}$ is classification function; $f:\\\\mathcal Z \\\\to \\\\mathbb{R}$ is classifier $f=\\\\varphi(\\\\phi(\\\\cdot))$; by abusing notation a little bit, we also use ``$+\\\\leftarrow f(z)$'' to mean that $f$ predicts $z$ as malicious when\\n\\t\\t$f(z)\\\\geq \\\\tau$ for a threshold $\\\\tau$\\\\\\\\\\n\\t\\t$F_{\\\\theta}:\\\\mathcal X \\\\to \\\\mathbb R$& machine learning algorithm \\n\\t\\twith parameters $\\\\theta$ \\\\\\\\\\n\\t\\t$L:\\\\mathbb{R}\\n\\t\\t\\\\times\\\\mathcal{Y}\\\\to\\\\mathbb{R}$ & loss function measuring prediction error of $F_\\\\theta$ \\\\\\\\\\n\\t\\t${\\\\sf EL},{\\\\sf WR},{\\\\sf AT}$ & defense techniques: Ensemble Learning, Weight Regularization, Adversarial Training \\\\\\\\\\n\\t\\t${\\\\sf VL},{\\\\sf RF},{\\\\sf IT},{\\\\sf CD},{\\\\sf SE}$ & defense techniques: Verifiable Learning, Robust Feature, Input Transformation, Classifier ranDomization, Sanitizing Examples \\\\\\\\\\n\\t\\t${\\\\sf BE},{\\\\sf OE},{\\\\sf BP},{\\\\sf OP}$ & attack tactics: basic and optimal evasion; basic and optimal poisoning\\\\\\\\\\n\\t\\t{\\\\sf GO},{\\\\sf SF},{\\\\sf MI} & attack techniques: Gradient-based Optimization, Sensitive Features, MImicry\\\\\\\\\\n\\t\\t{\\\\sf TR},{\\\\sf HS},{\\\\sf GM},{\\\\sf MS} & attack techniques: TRansferability, Heuristic Search, Generative Model, Mixture Strategy \\\\\\\\\\n\\t\\t$A_1,\\\\ldots,A_5$ & the 5 attributes under $\\\\mathcal{I}$'s control; they are known to $\\\\mathcal{A}$ at respective degrees\\n\\t\\t$a_1,\\\\ldots,a_5$ \\\\\\\\\\n\\t\\t$A_6,\\\\ldots,A_9$ & the 4 attributes under $\\\\mathcal{A}$'s control; they are known to $\\\\mathcal{I}$ at respective degree $a_6,\\\\ldots,a_9$ \\\\\\\\\\n\\t\\t${\\\\sf RR},{\\\\sf CR},{\\\\sf DR},{\\\\sf TR}$ & security properties: Representation Robustness, Classification Robustness, Detection Robustness, Training Robustness\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\\\end{tabular}\\n\\t\\\\end{adjustbox}\\n\\\\end{table}\\n\\\\noindent{\\\\bf Terminology, Scope and Notations}. \\nIn the AMD context, a defender $\\\\mathcal{I}$ aims to use ML to detect or classify computer files as benign or malicious; i.e., we focus on {\\\\em binary classification}. An attacker $\\\\mathcal{A}$ attempts to make malicious files evade $\\\\mathcal{I}$'s detection by leveraging {\\\\em adversarial files} (interchangeably, {\\\\em adversarial examples}). Adversarial malware examples are often generated by perturbing or manipulating malware examples, explaining why we will use the two terms, perturbation and manipulation, interchangeably. Adversarial attacks can be waged in the training phase of a ML model (a.k.a., poisoning attack) or in the test phase (a.k.a., evasion attack). It is worth mentioning that the {\\\\em privacy violation} attack is waged in addition to the preceding two attacks because $\\\\mathcal{A}$ can always probe $\\\\mathcal{I}$'s detectors.\\nA file, benign and malicious alike, is adversarial if it is intentionally crafted to (help malicious files) evade $\\\\mathcal{I}$'s detection, and non-adversarial otherwise.\\nWe focus on $\\\\mathcal{I}$ using supervised learning to detect malicious files, which may be adversarial or non-adversarial because they co-exist in the real world with no self-identification. This means that we do not consider the large body of malware detection literature that does not cope with AMD, which has been addressed elsewhere (e.g., ). \\nTable \\\\ref{table:notations} summarizes the main notations used in the paper.\",\n \"id\": \"a1283bee-37c1-4371-bed3-ceeb8af64d75\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"4fe25043-06f8-4577-bd62-3fe42c7ab899\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Survey and Systematization Methodology\"\n ]\n ],\n \"subsections\": [\n \"b7dbff6a-c1f8-49f8-8c03-efdb7ba23041\",\n \"8e32af6c-587d-43e1-aabc-dc5d8345628d\",\n \"b05fe1bc-da2e-4a87-8f17-d43f1acc81fc\"\n ],\n \"title\": \"Survey and Systematization Methodology\"\n },\n {\n \"cite_extract_rate\": 0.277777777777777,\n \"cites\": [\n 1159,\n 8565,\n 318,\n 166,\n 3861\n ],\n \"content\": \"\\\\label{sec:ml}\\nLet $\\\\mathcal Z$ be the {\\\\em example space} of benign/malicious adversarial/non-adversarial files. Let $\\\\mathcal{Y}=\\\\{+,-\\\\}$ or $\\\\mathcal{Y}=\\\\{1,0\\\\}$ be the {\\\\em label space} of binary classification, where $+$/$1$ ($-$/$0$) means a file is malicious (benign). Let $\\\\mathcal D =\\\\mathcal Z \\\\times \\\\mathcal Y$ be the file-label (example-label) space. For training and evaluating a classifier in the absence of adversarial files, $\\\\mathcal I$ is given a set $D\\\\subset \\\\mathcal{D}$ of non-adversarial benign/malicious files as well as their {\\\\em ground-truth} labels. $\\\\mathcal I$ splits $D$ into three disjoint sets: a training set $D_{train}=\\\\{(z_i,y_i)\\\\}_{i=1}^n$, a validation set for model selection, and a test set for evaluation. \\nEach file $z_i\\\\in \\\\mathcal{Z}$ is characterized by a set $S$ of features and represented by a numerical vector $\\\\mathbf{x}_i=(x_{i,1},\\\\ldots,x_{i,d})$ in the $d$-dimensional {\\\\em feature space} $\\\\mathcal X=\\\\mathbb R^d$, which accommodates both continuous and discrete feature representations .\\nThe process for obtaining feature representation $\\\\mathbf{x}_i$ of $z_i\\\\in \\\\mathcal{Z}$ is called {\\\\it feature extraction}, denoted by a function $\\\\phi:\\\\mathcal{Z}\\\\to \\\\mathcal{X}$ with ${\\\\bf x}_i\\\\leftarrow \\\\phi(S,z_i)$. Because $\\\\phi$ can be hand-crafted (denoted by $\\\\phi_c$), automatically learned (denoted by $\\\\phi_a$), or a hybrid of both , we unify them into $\\\\phi$ such that $\\\\phi(S,z)=\\\\phi_a(\\\\phi_c(S,z))$; when only manual (automatic) feature extraction is involved, we can set $\\\\phi_a$ ($\\\\phi_c$) as the identity map.\\nThere are two kinds of features: {\\\\em static features} are extracted via static analysis (e.g., strings, API calls\\n); {\\\\em dynamic features} are extracted via dynamic analysis (e.g., instructions, registry activities\\n).\\n\\\\begin{figure}[!htbp]\\n\\t\\\\centering\\n\\t\\\\scalebox{0.32}{\\n\\t\\\\includegraphics{figures/ml.pdf}\\n\\t}\\n\\\\caption{Illustration of ML-based malware detector.\\n}\\n\\t\\\\label{fig:ml}\\n\\\\end{figure}\\nAs highlighted in Figure \\\\ref{fig:ml},\\n$\\\\mathcal{I}$ uses $\\\\{(z_i,y_i)\\\\}_{i=1}^n$ to learn a malware detector or classifier $f:\\\\mathcal{Z}\\\\to [0,1]$, where $f(z)=\\\\varphi(\\\\phi(S,z))$ is composed of feature extraction function $\\\\phi: \\\\mathcal{Z}\\\\to \\\\mathcal{X}$ and classification function $\\\\varphi:\\\\mathcal{X} \\\\to [0,1]$.\\nNote that $f(z)\\\\in [0,1]$, namely $\\\\varphi({\\\\bf x})\\\\in [0,1]$ with ${\\\\bf x}\\\\leftarrow \\\\phi(S,z)$, can be interpreted as the probability that $z$ is malicious (while noting that calibration may be needed ). For a given threshold $\\\\tau\\\\in[0,1]$, we further say (by slightly abusing notations) $z$ is labeled by $f$ as $+$, or $+\\\\leftarrow f(z)$, if $f(z)\\\\geq \\\\tau$, and labeled as $-$ or $-\\\\leftarrow f(z)$ otherwise. In practice, $f$ is often specified by a learning algorithm $F$ with learnable parameter $\\\\theta$ (e.g., weights) and a hand-crafted feature extraction $\\\\phi_c$; then, $\\\\theta$ is tuned to minimize the empirical risk associated with a loss function $L:[0,1] \\\\times \\\\mathcal Y \\\\to \\\\mathbb{R}$ measuring the prediction error of $F_\\\\theta$ (e.g., cross-entropy ), namely\\n\\\\begin{equation}\\n\\\\min \\\\limits_{\\\\theta}~\\\\mathcal{L}(\\\\theta, D_{train}) = \\\\min \\\\limits_{\\\\theta}~\\\\frac{1}{n}\\\\sum_{(z_i, y_i)\\\\in D_{train}}\\\\left(L(F_\\\\theta(\\\\phi_c(S,z_i)), y_i)\\\\right). \\\\label{eq:rsk_emp}\\n\\\\end{equation}\\n\\\\noindent{\\\\bf Example 1: The Drebin malware detector}. Drebin is an Android malware detector trained from static features . Table \\\\ref{table:drebin} summarizes Drebin's feature set, which includes 4 subsets \\n\\\\begin{table}[htbp!]\\n\\t\\\\centering\\\\caption{Drebin features \\\\label{table:drebin}}\\n\\t\\\\begin{tabular}{c|p{.32\\\\textwidth}}\\n\\t\\t\\\\hline\\n\\t\\t\\\\multicolumn{2}{c}{Feature set} \\\\\\\\\\\\hline\\n\\t\\t\\\\multirow{4}{*}{Manifest} \\n\\t\\t& $S_1$~~Hardware components \\\\\\\\\\n\\t\\t& $S_2$~~Requested permissions\\\\\\\\\\n\\t\\t& $S_3$~~App components \\\\\\\\\\n\\t\\t& $S_4$~~Filtered intents \\\\\\\\\\\\hline\\n\\t\\t\\\\multirow{4}{*}{Dexcode}\\n\\t\\t& $S_5$~~Restricted API calls\\\\\\\\\\n\\t\\t& $S_6$~~Used permissions\\\\\\\\\\n\\t\\t& $S_7$~~Suspicious API calls\\\\\\\\\\n\\t\\t& $S_8$~~Network addresses\\\\\\\\\\n\\t\\t\\\\hline\\n\\t\\\\end{tabular}\\n\\\\end{table}\\nof features $S_1, S_2, S_3, S_4$ extracted from the AndroidManifest.xml and another 4 subsets of features $S_5, S_6, S_7, S_8$ extracted from the disassembled DEX code files (recalling that DEX code is compiled from a program written in some language and can be understood by the Android Runtime). Specifically, $S_1$ contains features related to the access of an Android package (APK) to smartphone hardware (e.g., camera, touchscreen, or GPS module); $S_2$ contains features related to APK's requested permissions listed in the manifest prior to installation; $S_3$ contains features related to application components (e.g., {\\\\em activities}, {\\\\em service}, {\\\\em receivers}, and {\\\\em providers}.); $S_4$ contains features related to APK's communications with the other APKs; $S_5$ contains features related to critical system API calls, which cannot run without appropriate permissions or the {\\\\em root} privilege; $S_6$ contains features corresponding to the used permissions; $S_7$ contains features related to API calls that can access sensitive data or resources in a smartphone; and $S_8$ contains features related to IP addresses, hostnames and URLs found in the disassembled codes.\\nThe feature representation is binary, meaning $\\\\phi=\\\\phi_c:\\\\mathcal{Z} \\\\mapsto \\\\{0,1\\\\}^d$ with $|S|=d$ and ${\\\\bf x}=(x_1,\\\\ldots,x_d)$, where $x_i=1$ if the corresponding feature is present in the APK $z$ and $x_i=0$ otherwise. A file $z$ in the feature space looks like the following: \\n\\\\[\\n\\\\mathbf x = \\\\phi(z) \\\\to\\n\\\\begin{pmatrix}\\n\\\\cdots \\\\\\\\ 0\\\\\\\\ 1\\\\\\\\\\n\\\\cdots \\\\\\\\ 1\\\\\\\\ 0\\\\\\\\\\n\\\\cdots\\\\\\\\\\n\\\\end{pmatrix}\\n\\\\begin{array}{ll}\\n\\\\cdots & \\\\multirow{4}{*}{\\\\hspace{-1mm}\\\\bigg \\\\} $S_2$ }\\\\\\\\\\n\\\\texttt{\\\\small permission::SEND\\\\_SMS} \\\\\\\\\\n\\\\texttt{\\\\small permission::READ\\\\_CONTACTS}\\\\\\\\\\n\\\\cdots & \\\\multirow{4}{*}{\\\\hspace{-1mm}\\\\bigg \\\\} $S_5$ }\\\\\\\\\\n\\\\texttt{\\\\small api\\\\_call::getDeviceID}\\\\\\\\\\n\\\\texttt{\\\\small api\\\\_call::setWifiEnabled}\\\\\\\\\\n\\\\cdots & \\\\\\\\\\n\\\\end{array}\\n\\\\]\\nDrebin uses a linear Support Vector Machine (SVM) to learn classifiers.\\n\\\\smallskip\\n\\\\noindent{\\\\bf Example 2: The MalConv malware detector}.\\nMalConv is Convolutional Neural Network (CNN)-based Windows Portable Executable (PE) malware detector learned from raw binary programs (i.e., end-to-end detection)\\n. \\nFigure \\\\ref{fig:malconv} depicts its architecture.\\nThe sequence of binary code is transformed into byte values (between 0 to 255) with the maximum length bounded by $N_{max}$ (e.g., $N_{max}=2^{21}$ bytes or 2MB). Each byte is further mapped into a real-valued vector using the embedding . The CNN layer and pooling layer learn abstract representations. The embedding, CNN and pooling layers belong to feature extraction $\\\\phi_a$, and the fully-connected and softmax layers belong to the classification operation $\\\\varphi$.\\n\\\\begin{figure}[!htbp]\\n\\t\\\\centering\\n\\t\\\\includegraphics[width=.7\\\\textwidth]{figures/malconv.pdf}\\n\\t\\\\caption{MalConv architecture .}\\n\\t\\\\label{fig:malconv}\\n\\\\end{figure}\",\n \"id\": \"b7dbff6a-c1f8-49f8-8c03-efdb7ba23041\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"a1283bee-37c1-4371-bed3-ceeb8af64d75\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Survey and Systematization Methodology\"\n ],\n [\n \"subsection\",\n \"Brief Review on ML-based Malware Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Brief Review on ML-based Malware Detection\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"We systematize AMD studies through the lens of four aspects: (i) the assumptions that are made; (ii) the attack or threat model in terms of attacker ${\\\\mathcal A}$'s {\\\\em objective} and ${\\\\mathcal A}$'s {\\\\em input}, with the latter including ${\\\\mathcal A}$'s information about the defender $\\\\mathcal{I}$ and ${\\\\mathcal A}$'s own; (iii) the defense in terms of $\\\\mathcal{I}$'s {\\\\em objective} and $\\\\mathcal{I}$'s {\\\\em input}, with the latter including $\\\\mathcal{I}$'s information about $\\\\mathcal{A}$ and $\\\\mathcal{I}$'s own; (iv) the security properties that are at stake. These four aspects are respectively elaborated below.\",\n \"id\": \"8e32af6c-587d-43e1-aabc-dc5d8345628d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"a1283bee-37c1-4371-bed3-ceeb8af64d75\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Survey and Systematization Methodology\"\n ],\n [\n \"subsection\",\n \"Framework\"\n ]\n ],\n \"subsections\": [\n \"18cf934c-a897-4b88-8a38-4cd32c20a2a0\",\n \"5f85312b-af7a-44ed-87ac-7e94a51dd8cc\",\n \"1e0de86d-a9b2-453b-b977-3d5abab5bd99\"\n ],\n \"title\": \"Framework\"\n },\n {\n \"cite_extract_rate\": 0.4375,\n \"cites\": [\n 318,\n 3857,\n 3862,\n 937,\n 7312,\n 9099,\n 3854\n ],\n \"content\": \"\\\\label{sec:assumption}\\nFive assumptions have been made in the AMD literature. Assumption \\\\ref{assumption:iid} below says that the data samples in $D$ are Independent and Identically Distributed (IID), which is a strong assumption and researchers have started to weaken it . \\n\\\\begin{assumption}[{\\\\sf IID} assumption; see, e.g., ]\\n\\\\label{assumption:iid}\\nComputer files in training data and testing data are independently drawn from the same distribution.\\n\\\\end{assumption}\\nAssumption \\\\ref{assumption:oracle} below is adapted from AML context, where humans can serve as an oracle $\\\\mathcal{O}$ for determining whether two images are the same . In the AMD context, $\\\\mathcal{O}$ can be instantiated as (or approximated by) malware analysts or automated tools (e.g., Sandbox ), with the latter often using heuristic rules produced by malware analysts (e.g., YARA ).\\n\\\\begin{assumption}[{\\\\sf Oracle} assumption; adapted from ]\\n\\\\label{assumption:oracle}\\nThere is an oracle $\\\\mathcal{O}:\\\\mathcal{Z}\\\\times\\\\mathcal{Z}\\\\to\\\\{{\\\\tt true},{\\\\tt false}\\\\}$ that tells if two files $z,z'\\\\in\\\\mathcal{Z}$ have the same functionality or not; ${\\\\tt true}\\\\leftarrow {\\\\mathcal{O}}(z,z')$ if and only if $z$ and $z'$ have the same functionality. \\n\\\\end{assumption}\\nAssumption \\\\ref{assumption:measurability} below says that there is a way to measure the degree of manipulations by which one file \\nis transformed to another.\\n\\\\begin{assumption}[{\\\\sf Measurability} assumption ]\\n\\\\label{assumption:measurability}\\nThere is a function $\\\\Gamma(z,z'):\\\\mathcal Z \\\\times \\\\mathcal Z \\\\to \\\\mathbb R_+$ that measures the degree of manipulations according to which a file $z'\\\\in \\\\mathcal{Z}$ can be derived from the file $z \\\\in \\\\mathcal{Z}$.\\n\\\\end{assumption}\\nSince Assumption \\\\ref{assumption:measurability} is often difficult to validate, $\\\\Gamma(z,z')$ may be replaced by a function that quantifies the degree of manipulation that can turn feature representation $\\\\mathbf{x}$ into $\\\\mathbf{x}'$, where\\n$\\\\mathbf{x}=\\\\phi(S,z)$ and $\\\\mathbf{x}'=\\\\phi(S,z')$. This leads to:\\n\\\\begin{assumption}[{\\\\sf Smoothness} assumption ]\\n\\\\label{assumption:smoothness}\\n\\\\ignore{\\nThis assumption says that for any $z,z'\\\\in \\\\mathcal{Z}$ and $\\\\delta \\\\in \\\\Delta$, it holds that $C(\\\\phi(z),\\\\phi(z'))\\\\approx 0$ when $(z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta)) \\\\wedge (\\\\Gamma(z,z')\\\\approx 0) \\\\wedge ({\\\\tt true}\\\\leftarrow {\\\\mathcal{O}(z,z')})$.\\n}\\nThere is a function $C(\\\\mathbf{x},\\\\mathbf{x}'):\\\\mathcal X \\\\times \\\\mathcal X \\\\to \\\\mathbb R_+$ such that $C(\\\\phi(S,z),\\\\phi(S,z'))\\\\approx 0$ when $ (\\\\Gamma(z,z')\\\\approx 0) \\\\wedge ({\\\\tt true}\\\\leftarrow {\\\\mathcal{O}(z,z')})$.\\n\\\\end{assumption}\\nAssumption \\\\ref{assumption:inverse} below says that the inverse of feature extraction, $\\\\phi^{-1}$, is solvable so that a perturbed representation $\\\\mathbf{x}'$ can be mapped back to a legitimate file. \\n\\\\begin{assumption}[{\\\\sf Invertibility} assumption ] \\\\label{assumption:inverse}\\nFeature extraction $\\\\phi$ is invertible, meaning that given $\\\\mathbf{x}'$, the function $\\\\phi^{-1}:\\\\mathcal{X}\\\\to\\\\mathcal{Z}$ produces $z'=\\\\phi^{-1}(\\\\mathbf{x}')$. \\n\\\\end{assumption}\\nRecall that the feature extraction function $\\\\phi$ may be composed of a hand-crafted $\\\\phi_c$ and an automated $\\\\phi_a$, where $\\\\phi_c$ may be neither differentiable nor invertible . This means $\\\\mathbf{x}'$ may not be mapped to a legitimate file. Researchers tend to relax the assumption by overlooking the interdependent features , while suffering from the side-effect $\\\\mathbf{x}'\\\\neq \\\\phi(\\\\phi^{-1}(\\\\mathbf{x}'))$ .\",\n \"id\": \"18cf934c-a897-4b88-8a38-4cd32c20a2a0\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"8e32af6c-587d-43e1-aabc-dc5d8345628d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Survey and Systematization Methodology\"\n ],\n [\n \"subsection\",\n \"Framework\"\n ],\n [\n \"subsubsection\",\n \"Systematizing Assumptions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Systematizing Assumptions\"\n },\n {\n \"cite_extract_rate\": 0.5111111111111111,\n \"cites\": [\n 8695,\n 890,\n 3866,\n 7156,\n 888,\n 3868,\n 1193,\n 3867,\n 3865,\n 166,\n 6172,\n 3864,\n 7155,\n 3862,\n 937,\n 3853,\n 9099,\n 9100,\n 7754,\n 3863,\n 7153,\n 3869\n ],\n \"content\": \"\\\\label{sec:sys-attack}\\nWe systematize attacks from two perspectives: {\\\\em attacker's objective} (i.e., what the attacker attempts to accomplish) and {\\\\em attacker's input} (i.e., what leverages the attacker can use). Whenever possible, we seek rigorous definitions to specify the attacker's input, while noting that these definitions have been scattered in the literature. We believe this specification is important because it can serve as a common reference for future studies. To demonstrate this, we discuss how to apply it to formulate a partial-order structure for comparing attacks.\\n\\\\smallskip\\n\\\\noindent{\\\\bf Attacker's Objective}.\\nThere are three kinds of objectives: (i)\\n{\\\\sf Indiscriminate}, meaning $\\\\mathcal{A}$ attempts to cause as many false-negatives as possible ; \\n(ii) {\\\\sf Targeted}, meaning $\\\\mathcal{A}$ attempts to cause specific false-negatives (i.e., making certain malicious files evade the detection );\\n(iii) {\\\\sf Availability}, meaning $\\\\mathcal{A}$ attempts to frustrate defender $\\\\mathcal{I}$ by rendering $\\\\mathcal{I}$'s classifier $f$ unusable (e.g., causing substantially high false-positives ).\\n\\\\begin{table}[!htbp]\\n\\\\caption{Attributes for specifying $\\\\mathcal{A}$'s and $\\\\mathcal{I}$'s input.}\\n\\t\\\\begin{minipage}{\\\\columnwidth}\\n\\t\\t\\\\centering\\n\\t\\t\\\\begin{tabular}{l|c|c}\\n\\t\\t\\\\hline\\nAttributes & Attacker $\\\\mathcal{A}$'s input & Defender $\\\\mathcal{I}$'s input \\\\\\\\\\\\hline\\\\hline\\n\\\\multicolumn{3}{l}{Attributes under $\\\\mathcal{I}$'s control but may be known to $\\\\mathcal{A}$ to some extent} \\\\\\\\\\\\hline\\n$A_1$: Training set $D_{train}$ & $a_1\\\\in [0,1]$ & 1 \\\\\\\\\\\\hline\\n$A_2$: Defense technique & $a_2\\\\in \\\\{0,1\\\\}$ & 1 \\\\\\\\\\\\hline\\n$A_3$: Feature set $S$ & $a_3\\\\in [0,1]$ & 1\\\\\\\\\\\\hline\\n$A_4$: Learning algorithm $F_\\\\theta$ & $a_4\\\\in [0,1]$ & 1 \\\\\\\\\\\\hline\\n$A_5$: Response & $a_5\\\\in \\\\{0,1\\\\}$ & 1 \\\\\\\\\\\\hline\\\\hline\\n\\\\multicolumn{3}{l}{Attributes under $\\\\mathcal{A}$'s control but may be known to $\\\\mathcal{I}$ to some extent} \\\\\\\\\\\\hline\\n$A_6$: Manipulation set $\\\\mathcal{M}$ & 1 & $a_6\\\\in [0,1]$ \\\\\\\\\\\\hline\\n$A_7$: Attack tactic & 1 & $a_7\\\\in \\\\{0,1\\\\}$ \\\\\\\\\\\\hline\\n$A_8$: Attack technique & 1& $a_8\\\\in \\\\{0,1\\\\}$ \\\\\\\\\\\\hline\\n$A_9$: Adversarial examples & 1 & $a_9\\\\in [0,1]$ \\\\\\\\\\\\hline\\n\\t\\t\\\\end{tabular}\\n\\t\\\\end{minipage}\\n\\t\\\\label{tab:attr}\\n\\\\end{table}\\n\\\\smallskip\\n\\\\noindent{\\\\bf Attacker's Input}.\\nTable \\\\ref{tab:attr} highlights the attributes we define to describe $\\\\mathcal{A}$'s input, including: five attributes $A_1,\\\\ldots,A_5$ that are under $\\\\mathcal{I}$'s control (indicated by 1) but may be known to $\\\\mathcal{A}$ at some extent $a_1,\\\\ldots,a_5$, respectively; and four attributes $A_6,\\\\ldots,A_9$ that are under $\\\\mathcal{A}$'s control (indicated by 1). These attributes are elaborated below.\\n({\\\\bf i}) $A_1$: it describes $\\\\mathcal{I}$'s training set $D_{train}$ for learning classifier $f$. \\nWe use $a_1\\\\in [0,1]$ to represent the extent at which $D_{train}$ is known to $\\\\mathcal{A}$.\\nLet $\\\\hat{D}_{train}$ be the training files that are known to $\\\\mathcal{A}$. Then, $a_1=|\\\\hat{D}_{train}\\\\cap{D}_{train}|/|{D}_{train}|$.\\n({\\\\bf ii}) $A_2$: it describes $\\\\mathcal{I}$'s techniques, which can be Ensemble Learning (${\\\\sf EL}$), Weight Regularization (${\\\\sf WR}$), Adversarial Training (${\\\\sf AT}$), Verifiable Learning (${\\\\sf VL}$), Robust Feature ({\\\\sf RF}), Input Transformation (${\\\\sf IT}$), Classifier ranDomization (${\\\\sf CD}$), Sanitizing Examples ({\\\\sf SE}). Let $A_2\\\\in\\\\{${\\\\sf EL},{\\\\sf WR},{\\\\sf AT},{\\\\sf VL},{\\\\sf RF},{\\\\sf IT},{\\\\sf CD},{\\\\sf SE}$\\\\}$ and $a_2\\\\in \\\\{0,1\\\\}$ such that $a_2=0$ means $\\\\mathcal{A}$ does not know $\\\\mathcal{I}$'s techniques and $a_2=1$ means $\\\\mathcal{A}$ knows $\\\\mathcal{I}$'s technique. The techniques are defined as follows.\\nDefinition \\\\ref{definition:el} says that $\\\\mathcal{I}$ constructs multiple classifiers and uses them collectively in malware detection.\\n\\\\begin{definition} [ensemble learning or {\\\\sf EL} ] \\\\label{definition:el}\\nLet $\\\\mathcal{H}$ be $\\\\mathcal{I}$'s classifier space.\\nGiven $K$ classifiers $\\\\{f_i\\\\}_{i=1}^{K}$ where $f_i\\\\in\\\\mathcal{H}$ and $f_i:\\\\mathcal{Z}\\\\to[0,1]$, let $f_i$ be assigned with weight $\\\\omega_i$ with $\\\\sum_{i=1}^K{\\\\omega_i}=1$ and $\\\\omega_i\\\\geq 0$. Then, $f=\\\\sum_{i=1}^{K}\\\\omega_i f_i$.\\n\\\\end{definition}\\nDefinition \\\\ref{definition:wr} says that $\\\\mathcal{I}$ uses regularization (e.g., $\\\\ell_2$ regularization or {\\\\em dropout} ) to decrease model's sensitivity to adversarial examples.\\n\\\\begin{definition} [weight regularization or {\\\\sf WR} ] \\\\label{definition:wr}\\nGiven a regularization item $\\\\Omega$ (e.g., constraints imposed on the learnable parameters), the empirical risk is\\n$\\\\min \\\\limits_{\\\\theta}~\\\\left[\\\\mathcal{L}(\\\\theta, D_{train})+\\\\Omega(\\\\theta)\\\\right]$, where $\\\\mathcal{L}$ is defined in Eq. \\\\eqref{eq:rsk_emp}.\\n\\\\end{definition}\\nDefinition \\\\ref{definition:at} says that $\\\\mathcal{I}$ proactively makes its classifier $f$ perceive some information about adversarial files. That is, $\\\\mathcal{I}$ augments the training set by incorporating adversarial examples that may be produced by $\\\\mathcal{I}$, $\\\\mathcal{A}$, or both.\\n\\\\begin{definition} [adversarial training or {\\\\sf AT} ] \\\\label{definition:at}\\nLet $D'$ denote a set of adversarial file-label pairs.\\nThen, $\\\\mathcal{I}$ tunes model parameters by minimizing the empirical risk: $\\\\min \\\\limits_{\\\\theta}~\\\\left[\\\\mathcal{L}(\\\\theta, D_{train}) + \\\\beta \\\\mathcal{L}(\\\\theta, D')\\\\right]$, where $\\\\beta\\\\geq 0$ denotes a balance factor.\\n\\\\end{definition}\\nDefinition \\\\ref{definiton:vl} says that $\\\\mathcal{I}$ intentionally over-estimates the error incurred by $\\\\mathcal{A}$'s manipulations and then minimizes it.\\n\\\\begin{definition}[verifiable learning or {\\\\sf VL} ] \\\\label{definiton:vl}\\nGiven $(z,y)\\\\in D_{train}$ and a manipulation set $\\\\hat{\\\\mathcal{M}}$ known by $\\\\mathcal{I}$, let $z(\\\\hat{\\\\mathcal{M}})$ denote the upper and lower boundaries on $\\\\hat{\\\\mathcal{M}}$. Then, this defense technique minimizes the following loss function derived from Eq.\\\\eqref{eq:rsk_emp}:\\n$L(F_\\\\theta(\\\\phi_c(S,z)), y)+\\\\beta L(F_\\\\theta(\\\\phi_c(S,z(\\\\hat{\\\\mathcal{M}}))), y).$\\n\\\\end{definition}\\nDefinition \\\\ref{eq:robust-feature-defense-technique} says that $\\\\mathcal{I}$ uses a set of features $S^*\\\\subseteq S$ that can lead to higher detection capability against adversarial example attacks.\\n\\\\begin{definition} [robust feature or {\\\\sf RF}; adapted from ]\\n\\\\label{eq:robust-feature-defense-technique}\\nGiven a training set $D_{train} \\\\cup D'$ that contains (adversarial) file-label pairs, the set of robust feature set $S^*$ is $$S^*=\\\\argmin_{\\\\tilde{S}\\\\subset S}\\\\sum_{(z,y)\\\\in {D_{train}\\\\cup D'}}L(\\\\widetilde{F}_\\\\theta(\\\\phi_c(\\\\tilde{S},z)),y),$$ \\nwhere $\\\\widetilde{F}_\\\\theta$ is $F_\\\\theta$ or a simplified learning algorithm that is computationally faster than $F_\\\\theta$ .\\n\\\\end{definition}\\nDefinition \\\\ref{def:it} says that $\\\\mathcal{I}$ aims to use non-learning methods (e.g., de-obfuscation as shown in Proguard ) to offset $\\\\mathcal{A}$'s manipulations.\\n\\\\begin{definition} [input transformation or {\\\\sf IT}, adapted from ] \\\\label{def:it}\\nLet ${\\\\tt IT}:\\\\mathcal{Z}\\\\to\\\\mathcal{Z}$ denote an input transformation in the file space. Given file $z$ and transformation ${\\\\tt IT}$, the classifier is $f=\\\\varphi(\\\\phi({\\\\tt IT}(z)))$.\\n\\\\end{definition}\\nDefinition \\\\ref{definition:cd} says that $\\\\mathcal{I}$ randomly chooses $m$ classifiers and uses their results for prediction. That is, $\\\\mathcal{I}$ aims to randomize the feature representation used by $f$, the learning algorithm, and/or response to $\\\\mathcal{A}$'s queries (to prevent $\\\\mathcal{A}$ from inferring information about $f$). \\n\\\\begin{definition} [classifier randomization or {\\\\sf CD}; adapted from ] \\\\label{definition:cd}\\nGiven $\\\\mathcal{I}$'s classifier space $\\\\mathcal{H}$ and an input file $z$, $\\\\mathcal{I}$ randomly selects $m$ classifiers from $\\\\mathcal{H}$ with replacement, say $\\\\{f_i\\\\}_{i=1}^m$. Then,\\n\\t$f=\\\\frac{1}{m}\\\\sum_{i=1}^m~f_i(z)$. \\n\\\\end{definition}\\nInstead of enhancing malware detectors, Definition \\\\ref{definition:se} provides an alternative that detects the adversarial examples for further analysis.\\n\\\\begin{definition}[sanitizing examples or ${\\\\sf SE}$; adapted from ] \\\\label{definition:se}\\n$\\\\mathcal{I}$ aims to detect adversarial files by using function \\n${\\\\sf flag}:\\\\mathcal{Z}\\\\to \\\\{{\\\\tt yes},{\\\\tt no}\\\\}$ to flag a file as adversarial ({\\\\tt yes}) or not ({\\\\tt no}).\\n\\\\end{definition}\\n({\\\\bf iii}) $A_3$: it describes $\\\\mathcal{I}$'s feature set $S$. We use $a_3\\\\in [0,1]$ to represent the extent at which $\\\\mathcal{A}$ knows about $S$. Let $\\\\hat{S}$\\ndenote the features that are known to $\\\\mathcal{A}$. Then,\\n$a_3=|\\\\hat{S}\\\\cap S|/|S|$.\\n({\\\\bf iv}) $A_4$: it describes $\\\\mathcal{I}$'s learning algorithm $F_\\\\theta$, the set of trainable parameters $\\\\theta$, and hyperparameters (which are set manually, e.g., $\\\\beta$ in Definition \\\\ref{definition:at})\\n. We use $a_4\\\\in [0,1]$ to represent that $\\\\mathcal{A}$ knows an $a_4$ degree about $A_4$, where $a_4=0$ means $\\\\mathcal{A}$ knows nothing and $a_4=1$ means $\\\\mathcal{A}$ knows everything.\\n({\\\\bf v}) $A_5$: it describes $\\\\mathcal{I}$'s response to $\\\\mathcal{A}$'s query to $f$ (if applicable), which is relevant because $\\\\mathcal{A}$ can learn useful information about $f$ by observing $f$'s responses .\\nWe define $a_5\\\\in \\\\{0,1\\\\}$ such that $a_5=0$ means there is a limit on the response that can be made by $\\\\mathcal{A}$ to $f$\\n(referred as ${\\\\sf LQ}$) and $a_5=1$ means there is no limit (referred as ${\\\\sf FQ}$).\\n({\\\\bf vi}) $A_6$: it describes $\\\\mathcal{A}$'s manipulation set in the problem space, which describes perturbations for generating adversarial files (adapted from {\\\\em perturbation set} in the AML literature ):\\n\\\\begin{eqnarray*}\\n\\\\mathcal{M}=\\\\{\\\\delta: (z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta)) \\\\wedge ({\\\\tt true}\\\\leftarrow \\\\mathcal{O}(z,z')\\n) \\\\wedge (z \\\\in \\\\mathcal{Z}) \\\\wedge (z'\\\\neq z)\\\\}.\\n\\\\end{eqnarray*}\\n$\\\\mathcal{M}$ is application-specific. For instance, an Android Package Kit (APK) permits adding codes or renaming class names , a Windows Portable Executable (PE) permits adding codes or changing PE {\\\\em section} names , and a Portable Document Format (PDF) file permits appending dead-code at its end or add new instructions . This means that a perturbation $\\\\delta\\\\in\\\\mathcal{M}$ can be a tuple specifying an operator (e.g., addition or removal), an object (e.g., a feature used by $\\\\mathcal{I}$), and other kinds of information (e.g., perturbation location in a file). \\nSince it is often infeasible to enumerate the entire manipulation set, $\\\\mathcal{A}$ may leverage an empirical one $\\\\widetilde{\\\\mathcal{M}}$ , which can be defined in the problem or feature space. Manipulations in the problem space must not violate the relevant constraints (e.g., adding APIs in an APK should not cause the request of unauthorized permissions). Manipulations in the feature space facilitate efficient computing via gradient-based methods as long as the inverse feature mapping $\\\\phi^{-1}$ is available.\\nFurthermore, we can use the manipulation set $\\\\mathcal{M}$ to define a {\\\\em feature manipulation set} $\\\\mathbf{M}$:\\n\\\\begin{eqnarray}\\n\\\\mathbf{M}=\\\\{\\\\delta_{\\\\mathbf{x}}=\\\\mathbf{x}'-\\\\mathbf{x}: (\\\\mathbf{x}=\\\\phi(z)) \\\\land (\\\\mathbf{x}'=\\\\phi(z')) \\\\land (z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta)) \\\\land (\\\\delta \\\\in \\\\mathcal{M}) \\\\land (z \\\\in \\\\mathcal{Z})\\n\\\\}. \\\\label{eq:manipulations}\\n\\\\end{eqnarray}\\nIn order to compute $\\\\mathbf{M}$ efficiently, one strategy is to estimate a feature-space analog of $\\\\widetilde{\\\\mathcal{M}}$, denoted by $\\\\widetilde{\\\\mathbf{M}}$ . This however demands resolving the invertibility Assumption \\\\ref{assumption:inverse}.\\n({\\\\bf vii}) $A_7$: it describes $\\\\mathcal{A}$'s attack tactics. We consider two tactics: classifier {\\\\em evasion} and classifier {\\\\em poisoning}. For the evasion attack, we consider three variants: basic evasion (${\\\\sf BE}$), optimal evasion 1 (${\\\\sf OE1}$) and optimal evasion 2 (${\\\\sf OE2}$).\\nFor the poisoning attack, we consider two variants: basic poisoning (${\\\\sf BP}$) and optimal poisoning (${\\\\sf OP}$). Correspondingly, we have $A_7\\\\in\\\\{{\\\\sf BE},{\\\\sf OE1},{\\\\sf OE 2},{\\\\sf BP},{\\\\sf OP}\\\\}$.\\nThese tactics are elaborated below, while noting that they do not explicitly call oracle $\\\\mathcal{O}$ because definitions of manipulation sets $\\\\mathcal{M}$ already assure that manipulations preserve functionalities of non-adversarial files.\\nAs shown in Definition \\\\ref{defintion:basic_evs}, the basic evasion attack is that $\\\\mathcal{A}$ uses a set of perturbations $\\\\delta_z\\\\subseteq\\\\mathcal{M}$ to manipulate a malicious file $z$, which is classified by $\\\\mathcal{I}$'s classifier $f$ as $+\\\\leftarrow f(z)$, to an adversarial file $z'$ such that $-\\\\leftarrow f(z')$.\\n\\\\begin{definition}[basic evasion or ${\\\\sf BE}$ ] \\\\label{defintion:basic_evs}\\n$\\\\mathcal{A}$ looks for $\\\\delta_z\\\\subseteq\\\\mathcal{M}$ to achieve the following for $z\\\\in \\\\mathcal{Z}$ with $+\\\\leftarrow f(z)$:\\n\\\\begin{eqnarray*}\\n-\\\\leftarrow f(z') ~\\\\text{where}~\\n(z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta_z)) \\\\wedge\\n(\\\\delta_z\\\\subseteq\\\\mathcal{M}).\\n\\\\end{eqnarray*}\\n\\\\end{definition}\\nAs shown in Definition \\\\ref{def:opt_evasion1}, the attacker attempts to minimize the degree of perturbations. In other words, this attack tactic is the same as {\\\\sf BE}, except that $\\\\mathcal{A}$ attempts to minimize the manipulation when perturbing a non-adversarial file $z\\\\in \\\\mathcal{Z}$ into an adversarial file $z'\\\\in \\\\mathcal{Z}$.\\n\\\\begin{definition}[optimal evasion 1 or ${\\\\sf OE1}$; adapted from ] \\\\label{def:opt_evasion1}\\n$\\\\mathcal{A}$ attempts to achieve the following for $z\\\\in \\\\mathcal{Z}$ with $+\\\\leftarrow f(z)$:\\n\\\\begin{eqnarray*}\\n\\\\min \\\\limits_{z'}\\n\\\\Gamma(z', z) \\n~\\\\text{s.t.}~(z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta_z))\\\\land (\\\\delta_z\\\\subseteq\\\\mathcal{M}) \\\\land (-\\\\leftarrow f(z')). \\n\\\\end{eqnarray*}\\n\\\\end{definition}\\nAs shown in Definition \\\\ref{def:opt_evasion2}, the attacker attempts to maximize $\\\\mathcal{I}$'s loss for waging high-confidence evasion attacks, while noting the small perturbations may be incorporated.\\n\\\\begin{definition}[optimal evasion 2 or ${\\\\sf OE2}$; adapted from ] \\\\label{def:opt_evasion2}\\n$\\\\mathcal{A}$ attempts to achieve the following for $z\\\\in \\\\mathcal{Z}$ with $+\\\\leftarrow f(z)$:\\n\\\\begin{eqnarray*}\\n\\\\max\\\\limits_{z'}\\nL(F_\\\\theta(\\\\phi_c(S,z')),+)\\n~\\\\text{s.t.}~(z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta_z))\\\\land (\\\\delta_z\\\\subseteq\\\\mathcal{M}) \\\\land (-\\\\leftarrow f(z')).\\n\\\\end{eqnarray*}\\n\\\\end{definition}\\nLet $D'_{poison}\\\\subset \\\\mathcal{D}$ be a set of adversarial file-label pairs obtained by manipulating non-adversarial files in $D_{poison}$. Let $D'_{train}\\\\leftarrow D_{train} \\\\cup D'_{poison}$ be the contaminated training data for learning a classifier $f'$ with parameters $\\\\theta'$. As shown in Definition \\\\ref{def:basic poisoning attack}, the basic poisoning attack is that the attacker aims to make $f'$ mis-classify the files in a dataset $D_{target}$, while accommodating the attacks that $\\\\mathcal{A}$ manipulates labels of the files in $D_{poison}$ .\\n\\\\begin{definition}[basic poisoning or ${\\\\sf BP}$ ]\\n\\\\label{def:basic poisoning attack}\\nGiven a set $D_{target}$ of files where $+\\\\leftarrow f(\\\\dot{z})$ for $\\\\dot{z}\\\\in D_{target}$ and a set $D_{poison}$ of non-adversarial files, $\\\\mathcal{A}$ attempts to perturb files in $D_{poison}$ to adversarial ones\\n$D'_{poison}=\\n\\\\{(\\\\mathcal{A}(z,\\\\delta_z), \\\\mathcal{A}(y)): ((z,y)\\\\in D_{poison})\\\\land (\\\\delta_z\\\\subseteq \\\\mathcal{M})\\\\land $ $ (\\\\mathcal{A}(y)\\\\in\\\\{+,-\\\\}) \\\\}$ such that classifier\\n$f'$ learned from $D'_{train}\\\\leftarrow D_{train} \\\\cup D'_{poison}$\\nmis-classifies the files in $D_{target}$.\\nFormally, the attacker intents to achieve the following for $\\\\forall~\\\\dot{z} \\\\in D_{target}$: \\n$-\\\\leftarrow f'(\\\\dot{z})$ where $f'$ is learned from $D'_{train}\\\\leftarrow D_{train}\\\\cup D'_{poison}$.\\n\\\\end{definition}\\nAs shown in Definition \\\\ref{def:optimal poisoning attack}, the optimal poisoning attack is the same as {\\\\sf BF}, except that $\\\\mathcal{A}$ attempts to\\nmaximize the loss when using classifier $f'$ with parameter $\\\\theta'$ to classify\\nfiles in $D_{target}$. \\nDefinition \\\\ref{def:optimal poisoning attack} can have multiple variants by considering bounds on $|D'_{poison}|$ or bounds on the degree of perturbations $|\\\\delta_z|$ . \\n\\\\begin{definition}[optimal poisoning or ${\\\\sf OP}$ ]\\n\\\\label{def:optimal poisoning attack}\\nGiven $D_{poison}$, $\\\\mathcal{A}$ perturbs $D_{poison}$ into $D'_{poison}$ for achieving:\\n\\\\begin{align*}\\n\\\\max \\\\limits_{D'_{poison}}~\\\\mathcal{L}(\\\\theta', D_{target})\\n~\\\\text{where}~{\\\\theta'} \\\\leftarrow \\\\operatorname*{\\\\arg\\\\min}_{\\\\theta} \\\\mathcal{L}(\\\\theta, D_{train} \\\\cup D'_{poison}).\\n\\\\end{align*}\\n\\\\end{definition}\\n{({\\\\bf viii})} $A_8$: it describes $\\\\mathcal{A}$'s attack techniques, such as Gradient-based Optimization ({\\\\sf GO}), Sensitive Features ({\\\\sf SF}), MImicry ({\\\\sf MI}), TRansferability ({\\\\sf TR}), Heuristic Search ({\\\\sf HS}), Generative Model ({\\\\sf GM}), and Mixture Strategy ({\\\\sf MS}). We denote this by $A_8\\\\in\\\\{${\\\\sf GO}, {\\\\sf SF}, {\\\\sf MI}, {\\\\sf TR}, {\\\\sf HS}, {\\\\sf GM}, {\\\\sf MS}$\\\\}$. Let $\\\\mathcal{A}$ have a classifier $\\\\hat{f}$, which consists of a hand-crafted feature extraction $\\\\hat{\\\\phi}_c$ and a parameterized model $\\\\hat{F}_{\\\\hat\\\\theta}$. Let $\\\\mathcal{A}$ also have an objective function $L_\\\\mathcal{A}:[0,1]\\\\times\\\\mathcal{Y}\\\\to\\\\mathcal{R}$, which measures $\\\\hat{f}$'s error or $\\\\mathcal{A}$' failure in evasion . Note that $\\\\hat{f}$ and $L_\\\\mathcal{A}$ can be the same as, or can mimic (by leveraging $\\\\mathcal{A}$'s knowledge about $\\\\mathcal{I}$'s attributes $A_1,\\\\ldots,A_5$), $\\\\mathcal{I}$'s classifier $f$ and loss function $L$, respectively. \\nThe attack technique specified by Definition \\\\ref{definition:go} is that $\\\\mathcal{A}$ solves the feature-space optimization problems described in Definitions \\\\ref{def:opt_evasion1}, \\\\ref{def:opt_evasion2} and \\\\ref{def:optimal poisoning attack} by using some gradient-based optimization method and then leverages the invertibility Assumption \\\\ref{assumption:inverse} to generate adversarial malware examples.\\n\\\\begin{definition}[Gradient-based Optimization or {\\\\sf GO}, adapted from ] \\\\label{definition:go}\\nLet $\\\\mathbf{x}\\\\leftarrow\\\\hat{\\\\phi}_c(\\\\hat{S},z)$ and $\\\\mathbf{x}'\\\\leftarrow\\\\mathbf{x}+\\\\delta_\\\\mathbf{x}$. The feature-space optimization problem in Definition \\\\ref{def:opt_evasion1} can be written as\\n\\t\\\\begin{equation}\\n\\t\\t\\\\min_{\\\\delta_\\\\mathbf{x}} C(\\\\mathbf{x}, \\\\mathbf{x}+\\\\delta_\\\\mathbf{x})~~~\\\\text{s.t.}~~~ (\\\\delta_\\\\mathbf{x}\\\\in[\\\\underline{\\\\mathbf{u}},\\\\overline{\\\\mathbf{u}}]) \\\\land (\\\\hat{F}_{\\\\hat{\\\\theta}}(\\\\mathbf{x}')<\\\\tau), \\\\label{eq:a8:evs1}\\n\\t\\\\end{equation}\\nwhere $\\\\underline{\\\\mathbf{u}}$ and $\\\\overline{\\\\mathbf{u}}$ are respectively the lower and upper bounds on $\\\\mathbf{M}$ (e.g., $\\\\delta_\\\\mathbf{x}\\\\in[-\\\\mathbf{x}, 1-\\\\mathbf{x}]$ for binary representation $\\\\mathbf{x}$). \\nThe feature-space optimization problem in Definition \\\\ref{def:opt_evasion2} can be written as \\n \\\\begin{equation}\\n \\t\\\\max_{\\\\delta_\\\\mathbf{x}} L_\\\\mathcal{A}\\\\left(\\\\hat{F}_{\\\\hat{\\\\theta}}(\\\\mathbf{x}+\\\\delta_\\\\mathbf{x}), +\\\\right)~~\\\\text{s.t.}~~(\\\\delta_\\\\mathbf{x}\\\\in[\\\\underline{\\\\mathbf{u}},\\\\overline{\\\\mathbf{u}}]). \\\\label{eq:a8:evs2}\\n \\\\end{equation}\\nThe feature-space optimization problem specified in Definition \\\\ref{def:optimal poisoning attack} can be written as \\n \\\\begin{align}\\n \\t&\\\\max_{\\\\delta_\\\\mathbf{x}\\\\in[\\\\underline{\\\\mathbf{u}},\\\\overline{\\\\mathbf{u}}]}\\\\mathbb{E}_{(\\\\dot{z},\\\\dot{y})\\\\in D_{target}}L_\\\\mathcal{A}(\\\\hat{F}_{\\\\hat{\\\\theta}'}(\\\\hat{\\\\phi}_c(\\\\hat{S},\\\\dot{z}),\\\\dot{y})),~~ \\\\forall (z,y)\\\\in D_{poison} \\\\label{eq:a8:poi}\\\\\\\\\\n \\t~~\\\\text{where}~~&{\\\\hat\\\\theta}'\\\\leftarrow\\\\argmin_{\\\\hat\\\\theta}\\\\mathbb{E}_{(z_t,y_t)\\\\in{\\\\hat{D}_{train}\\\\cup\\\\{(\\\\hat{\\\\phi}_c^{-1}(\\\\hat{\\\\phi}_c(z)+\\\\delta_\\\\mathbf{x}),y')\\\\}}}L_\\\\mathcal{A}\\\\left(\\\\hat{F}_{\\\\hat\\\\theta}(\\\\hat{\\\\phi}_c(\\\\hat{S},z_t),y_t)\\\\right). \\\\nonumber\\n \\\\end{align}\\n\\\\end{definition}\\nIn order to calculate the gradients of loss function $L_\\\\mathcal{A}$ with respect to $\\\\delta_\\\\mathbf{x}$ in Eqs.\\\\eqref{eq:a8:evs1} and \\\\eqref{eq:a8:evs2}, inequality constraints can be handled by appending penalty items to the loss function in question and box-constraints can be coped with by using gradient projection . Since $\\\\mathbf{x}+\\\\delta_\\\\mathbf{x}$ is continuous, the {\\\\sf GO} attack technique needs to map $\\\\delta_\\\\mathbf{x}$ to a discrete perturbation vector in $\\\\mathbf{M}$, for instance by using the nearest neighbor search .\\nThe gradients of loss function $L_\\\\mathcal{A}$ with respect to $\\\\delta_\\\\mathbf{x}$ in Eq. \\\\eqref{eq:a8:poi} are delicate to deal with. One issue is the indirect relation between $L_\\\\mathcal{A}$ and $\\\\delta_\\\\mathbf{x}$, which can be handled by the chain rule . \\nAnother issue is the difficulty that is encountered when computing the partial derivatives $\\\\partial{\\\\hat\\\\theta'}/\\\\partial\\\\delta_{\\\\mathbf{x}}$ . \\nFor dealing with this, researchers often relax the underlying constraints (e.g., by supposing that $\\\\hat{F}_{\\\\hat\\\\theta}$ is a linear model). \\nThe attack technique specified by Definition \\\\ref{definition:sf} is that $\\\\mathcal{A}$ perturbs malware examples by injecting or removing a small number of features to decrease the classification error measured by the loss function ${L}_\\\\mathcal{A}$ as much as possible.\\n\\\\begin{definition}[Sensitive Features or {\\\\sf SF}, adapted from ] \\\\label{definition:sf}\\nFor evasion attacks, $\\\\mathcal{A}$ aims to maximize the following with respect to a given malware example-label pair $(z,+)$: $$\\\\max_{z'} L_\\\\mathcal{A}\\\\left(\\\\hat{F}_{\\\\hat\\\\theta}(\\\\hat{\\\\phi}(\\\\hat{S},z')),+\\\\right)~~\\\\text{s.t.}~~(z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta_z)) \\\\land (\\\\delta_z\\\\subseteq \\\\mathcal{M}) \\\\land (|\\\\delta_z|\\\\leq m),$$ \\nwhere $m$ is the maximum degree of manipulations.\\nFor poisoning attacks, $\\\\mathcal{A}$ aims to maximize the following with respect to the given $D_{poison}$ and $D_{target}$, \\n\\\\begin{align}\\n \\\\max_{D'_{poison}}\\\\mathbb{E}_{(\\\\dot{z},\\\\dot{y})\\\\in D_{target}}L_\\\\mathcal{A}(\\\\hat{F}_{\\\\hat{\\\\theta}'}(\\\\hat{\\\\phi}_c(\\\\hat{S},\\\\dot{z}),\\\\dot{y})),\\n\\\\end{align}\\nwhere $\\\\hat{\\\\theta}'$ is learned from $\\\\hat{D}_{train}\\\\cup D'_{poison}$ such that $\\\\forall z'\\\\in D'_{poison}$ is obtained via the perturbation $\\\\delta_z$ with obeying $(z'\\\\leftarrow \\\\mathcal{A}(z,\\\\delta_z)) \\\\land (\\\\delta_z\\\\subseteq \\\\mathcal{M}) \\\\land (z\\\\in D_{poison}) \\\\land (|\\\\delta_z|\\\\leq m)$.\\n\\\\end{definition} \\nThe attack technique specified by Definition \\\\ref{definition:mi} is that $\\\\mathcal{A}$ perturbs malware example $z$ by mimicking a benign example, while noting that this attack technique can be algorithm-agnostic.\\n\\\\begin{definition}[MImicry or {\\\\sf MI}, adapted from ] \\\\label{definition:mi}\\nGiven a set of benign examples $D_{ben}$ and a malware example $z$, $\\\\mathcal{A}$ aims to achieve the following minimization:\\n\\t\\\\begin{eqnarray}\\n\\t\\t\\\\min_{\\\\delta_z\\\\in\\\\mathcal{M}} \\\\Gamma(\\\\mathcal{A}(z,\\\\delta_z),z_{ben})~~\\\\text{s.t.}~~ (\\\\exists~z_{ben}\\\\in D_{ben}). \\\\label{eq:mimicry}\\n\\t\\\\end{eqnarray}\\n\\\\end{definition}\\nThe attack technique specified by Definition \\\\ref{definition:mi} can be extended to accommodate the similarity between representations in the feature space .\\nThe attack technique specified by Definition \\\\ref{definition:tr} is that $\\\\mathcal{A}$ generates adversarial examples against a surrogate model $\\\\hat{f}$.\\n\\\\begin{definition}[TRansferability or {\\\\sf TR}, adapted from ] \\\\label{definition:tr}\\n$\\\\mathcal{A}$ learns a surrogate model $\\\\hat{f}$ of $f$ from $\\\\hat{D}_{train}$, $\\\\hat{S}$, $\\\\hat{\\\\phi_c}$ and $\\\\hat{F}$. For evasion attacks,\\n$\\\\mathcal{A}$ achieves $-\\\\leftarrow\\\\hat{f}(z')$ by perturbing malware example $z$ to $z'$ and then attacks $f$ with $z'$. For poisoning attacks, $\\\\mathcal{A}$ contaminates $\\\\hat{f}$ to $\\\\hat{f}'$ such that $-\\\\leftarrow\\\\hat{f}'(\\\\dot{z})~\\\\text{for}~\\\\forall(\\\\dot{z},\\\\dot{y}) \\\\in D_{target}$, by poisoning the training set $\\\\hat{D}_{train}$ with $D'_{poison}$ and the attacks $f$ with $D'_{poison}$.\\n\\\\end{definition}\\nThe attack technique specified by\\nDefinition \\\\ref{definition:hs} is that $\\\\mathcal{A}$ searches perturbations in $\\\\mathcal{M}$ via some heuristics, while leveraging oracle $\\\\mathcal{O}$'s responses to $\\\\mathcal{A}$'s queries and $f$'s responses to $\\\\mathcal{A}$'s queries. Since $\\\\mathcal{M}$ is defined with respect to the problem space, this attack technique does not need the invertibility Assumption \\\\ref{assumption:inverse}.\\n\\\\begin{definition}[Heuristic Search or {\\\\sf HS}] \\\\label{definition:hs}\\nLet $h$ be a function taking $\\\\mathcal{O}$'s response and $f$'s response as input. Given a malware example $z$, $\\\\mathcal{A}$ looks for an $m$-length manipulation path \\n\\\\begin{align}\\n\\\\langle z_{(0)}, z_{(1)},\\\\ldots,z_{(m)}\\\\rangle \\n~~\\\\text{s.t.}~~z_{(i+1)}=\\\\mathcal{A}(z_{(i)},\\\\delta_{z,(i)}) \\\\land (\\\\delta_{z,(i)}& \\\\in \\\\mathcal{M}) \\\\land (h(\\\\mathcal{O}, f,z,z_{(i)})\\\\leq h(\\\\mathcal{O}, f,z,z_{(i+1)})) \\\\nonumber\\n\\\\end{align}\\nwhere $z_{(0)}=z$.\\n\\\\end{definition}\\nThe attack technique specified by Definition \\\\ref{def:generative-model} is that $\\\\mathcal{A}$ uses a generative model $G$ with parameters $\\\\theta_g$ to perturb malware representation vectors and then leverages the invertibility Assumption \\\\ref{assumption:inverse} to turn the perturbed vector into an adversarial malware example.\\n\\\\begin{definition}[Generative Model or {\\\\sf GM}]\\\\label{def:generative-model}\\nGiven a malware representation vector $\\\\mathbf{x}=\\\\hat{\\\\phi}_c(\\\\hat{S},z)$, $\\\\mathcal{A}$ achieves\\n$$\\\\max_{\\\\theta_g} L_\\\\mathcal{A}\\\\left(\\\\hat{F}_{\\\\hat{\\\\theta}}(G_{\\\\theta_g}(\\\\mathbf{x})),+\\\\right)~~\\\\text{s.t.}~~ G_{\\\\theta_g}(\\\\mathbf{x})\\\\in [\\\\underline{\\\\mathbf{u}}-\\\\mathbf{x}, \\\\overline{\\\\mathbf{u}}-\\\\mathbf{x}]$$\\nand leverages the invertibility Assumption \\\\ref{assumption:inverse} to obtain an adversarial example $z'=\\\\hat\\\\phi_c^{-1}(G_{\\\\theta_g}(\\\\mathbf{x}))$.\\n\\\\end{definition}\\nThe attack technique specified by\\nDefinition \\\\ref{definition:ms} is that $\\\\mathcal{A}$ combines multiple perturbation methods to perturb an example. \\n\\\\begin{definition}[Mixture Strategy or {\\\\sf MS} ] \\\\label{definition:ms}\\nLet $\\\\mathcal{H}_A$ denote the space of generative methods and $\\\\mathcal{W}_a=\\\\{\\\\mathbf{w}_a:\\\\mathbf{w}_a=(w_{a,1},\\\\ldots,w_{a,K}),w_{a,i}\\\\geq 0\\\\}$ with $i=1,\\\\ldots,K$ denote the weights space. Given a malware example $z$, $\\\\mathcal{A}$ aims to achieve $$\\\\max_{\\\\mathbf{w}_a} L_\\\\mathcal{A}(\\\\hat{F}_{\\\\hat{\\\\theta}}(\\\\hat\\\\phi(z')),+) ~~\\\\text{s.t.}~~(z'=\\\\sum_{i=1}^K{w_{a,i}}g_i(z))\\\\land (\\\\mathcal{O}(z,z')={\\\\sf true})) \\\\land (g_i\\\\in \\\\mathcal{H}_A) \\\\land (\\\\mathbf{w}_a\\\\in \\\\mathcal{W}_a).$$\\n\\\\end{definition}\\n({\\\\bf ix}) $A_9$: it corresponds to $\\\\mathcal{A}$'s adversarial files. Given file manipulation set $\\\\mathcal{M}$, the corresponding set of adversarial files is defined as $\\\\mathcal{Z}_{\\\\mathcal M}=\\\\{\\\\mathcal{A}(z, \\\\delta_z):(z \\\\in \\\\mathcal{Z}) \\\\land (\\\\delta_z\\\\subseteq\\\\mathcal{M})\\\\}$. Given feature manipulation set $\\\\mathbf{M}$, the set of adversarial feature vectors is: $\\\\mathcal{X}_{\\\\mathbf{M}}=\\\\{\\\\mathbf{x}':(\\\\mathbf{x}'=\\\\mathbf{x}+\\\\delta_{\\\\mathbf{x}}) \\\\land (\\\\delta_\\\\mathbf{x}\\\\in\\\\mathbf{M})\\\\}.$\\n\\\\begin{figure}[htbp!]\\n\\t\\\\centering\\n\\t\\\\begin{tikzpicture}[thick,scale=0.95, every node/.style={transform shape}]\\n\\t\\\\node(top) [label=right:{White-box} attack, label=left:{}] at (0,6) {$(1,1,1,1,1)$};\\n\\t\\\\node(l5n0)[label=below:{},align=center] at (0,5) {$(a_1,a_2,1,1,a_5)$};\\n\\t\\\\node(l4n0)[label=below:{},align=center] at (2.2,4) {$(a_1,a_2,a_3,1,a_5)$};\\n\\t\\\\node(l4n1)[label=below:{},align=center] at (0,4) {$(0,0,1,1,0)$};\\n\\t\\\\node(l4n2)[label=below:{},align=center] at (-2.2,4) {$(a_1,a_2,1,a_4,a_5)$};\\n\\t\\\\node(l3n0)[label=below:{},align=center] at (4.4,3) {$(a_1,a_2,0,1,a_5)$};\\n\\t\\\\node(l3n1)[label=below:{},align=center] at (2.2,3) {$(0,0,a_3,1,0)$};\\n\\t\\\\node(l3n2)[label=below:{},align=center] at (0,3) {$(a_1,a_2,a_3,a_4,a_5)$};\\n\\t\\\\node(l3n3)[label=below:{},align=center] at (-2.2,3) {$(0,0,1,a_4,0)$};\\n\\t\\\\node(l3n4)[label=below:{},align=center] at (-4.4,3) {$(a_1,a_2,1,0,a_5)$};\\n\\t\\\\node(l2n0)[label=below:{},align=center] at (4.4,2) {$(0,0,0,1,0 )$};\\n\\t\\\\node(l2n1)[label=below:{},align=center] at (2.2,2) {$(a_1,a_2,0,a_4,a_5)$};\\n\\t\\\\node(l2n2)[label=below:{},align=center] at (0,2) {$(0,0,a_3,a_4,0)$};\\n\\t\\\\node(l2n3)[label=below:{},align=center] at (-2.2,2) {$(a_1,a_2,a_3,0,a_5)$};\\n\\t\\\\node(l2n4)[label=below:{},align=center] at (-4.4,2) {$(0,0,1,0,0)$};\\n\\t\\\\node(l1n0)[label=below:{},align=center] at (2.2,1) {$(0,0,0,a_4,0 )$};\\n\\t\\\\node(l1n1)[label=below:{},align=center] at (0,1) {$(a_1,a_2,0,0,a_5)$};\\n\\t\\\\node(l1n2)[label=below:{},align=center] at (-2.2,1) {$(0,0,a_3,0,0 )$};\\n\\t\\\\node(bot)[label=right:{Black-box attack}, label=left:{}] at (0,0) {$ (0,0,0,0,0)$};\\n\\t\\\\draw(l5n0) -- (top);\\n\\t\\\\draw(l4n0) -- (l5n0);\\n\\t\\\\draw(l4n1) -- (l5n0);\\n\\t\\\\draw(l4n2) -- (l5n0);\\n\\t\\\\draw(l3n0) -- (l4n0);\\n\\t\\\\draw(l3n1) -- (l4n0);\\n\\t\\\\draw(l3n1) -- (l4n1);\\n\\t\\\\draw(l3n2) -- (l4n0);\\n\\t\\\\draw(l3n2) -- (l4n2);\\n\\t\\\\draw(l3n3) -- (l4n1);\\n\\t\\\\draw(l3n3) -- (l4n2);\\n\\t\\\\draw(l3n4) -- (l4n2);\\n\\t\\\\draw(l2n0) -- (l3n0);\\n\\t\\\\draw(l2n0) -- (l3n1);\\n\\t\\\\draw(l2n1) -- (l3n0);\\n\\t\\\\draw(l2n1) -- (l3n2);\\n\\t\\\\draw(l2n2) -- (l3n1);\\n\\t\\\\draw(l2n2) -- (l3n2);\\n\\t\\\\draw(l2n2) -- (l3n3);\\n\\t\\\\draw(l2n3) -- (l3n2);\\n\\t\\\\draw(l2n3) -- (l3n4);\\n\\t\\\\draw(l2n4) -- (l3n3);\\n\\t\\\\draw(l2n4) -- (l3n4);\\n\\t\\\\draw(l1n0) -- (l2n0);\\n\\t\\\\draw(l1n0) -- (l2n1);\\n\\t\\\\draw(l1n0) -- (l2n2);\\n\\t\\\\draw(l1n1) -- (l2n1);\\n\\t\\\\draw(l1n1) -- (l2n3);\\n\\t\\\\draw(l1n2) -- (l2n2);\\n\\t\\\\draw(l1n2) -- (l2n3);\\n\\t\\\\draw(l1n2) -- (l2n4);\\n\\t\\\\draw(bot) -- (l1n2);\\n\\t\\\\draw(bot) -- (l1n1);\\n\\t\\\\draw(bot) -- (l1n0);\\n\\t\\\\end{tikzpicture}\\n\\t\\\\caption{A portion of the partial order defined over $(a_1,\\\\ldots,a_5)$.\\n\\t} \\n\\t\\\\label{fig:attack_input}\\n\\\\end{figure}\\n\\\\smallskip\\n\\\\noindent{\\\\bf On the Usefulness of the Preceding Specification}. The preceding specification can be applied to formulate a partial order in the attribute space, which allows to compare attacks unambiguously. Figure \\\\ref{fig:attack_input} depicts how vector $(a_1, \\\\cdots, a_5)$ formulates a partial order between the widely-used informal notions of {\\\\em black-box} attack, namely $(a_1,a_2,a_3,a_4,a_5)=(0,0,0,0,0)$, \\nand {\\\\em white-box} attack, namely $(a_1,a_2,a_3,a_4,a_5)=(1,1,1,1,1)$;\\nthere are many kinds of grey-box attacks in between.\",\n \"id\": \"5f85312b-af7a-44ed-87ac-7e94a51dd8cc\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 45,\n \"parent_id\": \"8e32af6c-587d-43e1-aabc-dc5d8345628d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Survey and Systematization Methodology\"\n ],\n [\n \"subsection\",\n \"Framework\"\n ],\n [\n \"subsubsection\",\n \"Systematizing Attacks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Systematizing Attacks\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Similarly, we systematize defenses from two perspectives: {\\\\em defender's objective} (i.e., what the defender aims to achieve) and {\\\\em defender's input} (i.e., what leverages the defender can use). \\nWe also discuss how to apply the specification to formulate a partial-order structure for comparing defenses.\\n\\\\smallskip\\n\\\\noindent{\\\\bf Defender's Objectives}.\\n$\\\\mathcal{I}$ aims to detect ideally all of the malicious files, adversarial and non-adversarial alike, while suffering from small side-effects (e.g., increasing false-positives).\\n\\\\smallskip\\n\\\\noindent{\\\\bf Defender's Input}.\\nAs highlighted in Table \\\\ref{tab:attr}, $\\\\mathcal{I}$'s input includes attributes $A_1,\\\\ldots,A_5$, which are under $\\\\mathcal{I}$'s control, and the extent $a_6,\\\\ldots,a_9$ at which $\\\\mathcal{I}$ respectively knows about attributes $A_6,\\\\ldots,A_9$, which are under $\\\\mathcal{A}$'s control. Note that $A_1,\\\\ldots,A_9$ have been defined above.\\n\\\\begin{itemize}\\n\\\\item We define $a_6 \\\\in [0,1]$ to represent the extent at which $\\\\mathcal{I}$ knows $\\\\mathcal{A}$'s manipulation set $\\\\mathcal{M}$. Let $\\\\hat{\\\\mathcal{M}}\\\\subseteq \\\\mathcal{M}$ denote the subset of $\\\\mathcal{A}$'s manipulation set known to $\\\\mathcal{I}$. \\nThen, we set $a_6=|\\\\hat{\\\\mathcal{M}}|/|\\\\mathcal{M}|$.\\n\\\\item We define $a_7\\\\in\\\\{0,1\\\\}$ such that $a_7=0$ means $\\\\mathcal{I}$ does not know $\\\\mathcal{A}$'s attack tactic $A_7\\\\in \\\\{{\\\\sf BE},{\\\\sf OE1},{\\\\sf OE2},{\\\\sf BP},$ ${\\\\sf OP}\\\\}$ and $a_7=1$ means $\\\\mathcal{I}$ knows $\\\\mathcal{A}$'s tactic. \\n\\\\item We define $a_8\\\\in\\\\{0,1\\\\}$ such that $a_8=1~(a_8=0)$ means the defender does (not) know $\\\\mathcal{I}$'s attack technique $A_8\\\\in\\\\{{\\\\sf GO}, {\\\\sf SF}, {\\\\sf TR}, {\\\\sf MI}, {\\\\sf HS}, {\\\\sf GM}, {\\\\sf MS}\\\\}$.\\n\\\\item We use $a_9=|\\\\hat{\\\\mathcal{Z}}_{\\\\mathcal{M}}| / |\\\\mathcal{Z}_{\\\\mathcal{M}}|$ to represent the extent at which $\\\\mathcal{I}$ knows about $\\\\mathcal{A}$'s adversarial files, where $a_9\\\\in [0,1]$ and $\\\\mathcal{Z}_{\\\\mathcal{M}}$ is $\\\\mathcal{A}$'s adversarial files and $\\\\hat{\\\\mathcal{Z}}_{\\\\mathcal{M}}\\\\subseteq\\\\mathcal{Z}_{\\\\mathcal{M}}$ is known to $\\\\mathcal{I}$.\\n\\\\end{itemize}\\n\\\\smallskip\\n\\\\noindent{\\\\bf On the Usefulness of the Preceding Specification}. \\nSimilarly, the defense specification can be used to formulate a partial order in the attribute space, paving the way for comparing defenses unambiguously. Figure \\\\ref{fig:defense_input} depicts how vector $(a_6,\\\\ldots,a_9)$ formulates a partial order between the widely-used informal notions of black-box defense $(a_6,a_7,a_8,a_9)=(0,0,0,0)$ and white-box defense $(a_6,a_7,a_8,a_9)=(1,1,1,1,1)$; there are many kinds of grey-box defenses in between.\\n\\\\begin{figure}[!htbp]\\n\\t\\\\centering\\n\\t\\\\begin{tikzpicture}[thick,scale=0.95, every node/.style={transform shape}]\\n\\t\\t\\\\node(top)[label=right:{White-box defense}, label=left:{}] at (0,6) {$( 1,1,1,1)$};\\n\\t\\t\\\\node(l3n0)[label=below:{},align=center] at (1.8,5) {$( a_6,a_7,a_8,1)$};\\n\\t\\t\\\\node(l3n1)[label=below:{},align=center] at (0,5) {$( 1,0,0,1)$};\\n\\t\\t\\\\node(l3n2)[label=below:{},align=center] at (-1.8,5) {$( 1,a_7,a_8,a_9)$};\\n\\t\\t\\\\node(l2n0)[label=below:{},align=center] at (3.6,4) {$( 0,a_7,a_8, 1)$};\\n\\t\\t\\\\node(l2n1)[label=below:{},align=center] at (1.85,4) {$( a_6,0,0,1)$};\\n\\t\\t\\\\node(l2n2)[label=below:{},align=center] at (0,4) {$( a_6,a_7,a_8,a_9)$};\\n\\t\\t\\\\node(l2n3)[label=below:{},align=center] at (-1.85,4) {$( 1,0,0,a_8)$};\\n\\t\\t\\\\node(l2n4)[label=below:{},align=center] at (-3.6,4) {$( 1,a_7,a_8,0)$};\\n\\t\\t\\\\node(l1n0)[label=below:{},align=center] at (3.6,3) {$( 0,0,0,1)$};\\n\\t\\t\\\\node(l1n1)[label=below:{},align=center] at (1.85,3) {$( 0,a_7,a_8,a_9)$};\\n\\t\\t\\\\node(l1n2)[label=below:{},align=center] at (0,3) {$( a_6,0,0,a_8)$};\\n\\t\\t\\\\node(l1n3)[label=below:{},align=center] at (-1.85,3) {$( a_6,a_7,a_8,0)$};\\n\\t\\t\\\\node(l1n4)[label=below:{},align=center] at (-3.6,3) {$( 1,0,0,0)$};\\n\\t\\t\\\\node(l0n0)[label=below:{},align=center] at (1.85,2) {$( 0,0,0,a_9)$};\\n\\t\\t\\\\node(l0n1)[label=below:{},align=center] at (0,2) {$( 0,a_7,a_8,0)$};\\n\\t\\t\\\\node(l0n2)[label=below:{},align=center] at (-1.85,2) {$( a_6,0,0,0)$};\\n\\t\\t\\\\node(bot)[label=right:{Black-box defense}, label=left:{}] at (0,1) {$( 0,0,0,0)$};\\n\\t\\t\\\\draw(bot) -- (l0n0);\\n\\t\\t\\\\draw(bot) -- (l0n1);\\n\\t\\t\\\\draw(bot) -- (l0n2);\\n\\t\\t\\\\draw(l0n0) -- (l1n0);\\n\\t\\t\\\\draw(l0n0) -- (l1n1);\\n\\t\\t\\\\draw(l0n0) -- (l1n2);\\n\\t\\t\\\\draw(l0n1) -- (l1n1);\\n\\t\\t\\\\draw(l0n1) -- (l1n3);\\n\\t\\t\\\\draw(l0n2) -- (l1n2);\\n\\t\\t\\\\draw(l0n2) -- (l1n3);\\n\\t\\t\\\\draw(l0n2) -- (l1n4);\\n\\t\\t\\\\draw(l1n0) -- (l2n0);\\n\\t\\t\\\\draw(l1n0) -- (l2n1);\\n\\t\\t\\\\draw(l1n1) -- (l2n0);\\n\\t\\t\\\\draw(l1n1) -- (l2n2);\\n\\t\\t\\\\draw(l1n2) -- (l2n1);\\n\\t\\t\\\\draw(l1n2) -- (l2n2);\\n\\t\\t\\\\draw(l1n2) -- (l2n3);\\n\\t\\t\\\\draw(l1n3) -- (l2n2);\\n\\t\\t\\\\draw(l1n3) -- (l2n4);\\n\\t\\t\\\\draw(l1n4) -- (l2n3);\\n\\t\\t\\\\draw(l1n4) -- (l2n4);\\n\\t\\t\\\\draw(l2n0) -- (l3n0);\\n\\t\\t\\\\draw(l2n1) -- (l3n0);\\n\\t\\t\\\\draw(l2n1) -- (l3n1);\\n\\t\\t\\\\draw(l2n2) -- (l3n0);\\n\\t\\t\\\\draw(l2n2) -- (l3n2);\\n\\t\\t\\\\draw(l2n3) -- (l3n1);\\n\\t\\t\\\\draw(l2n3) -- (l3n2);\\n\\t\\t\\\\draw(l2n4) -- (l3n2);\\n\\t\\t\\\\draw(l3n0) -- (top);\\n\\t\\t\\\\draw(l3n1) -- (top);\\n\\t\\t\\\\draw(l3n2) -- (top);\\n\\t\\\\end{tikzpicture}\\n\\t\\\\caption{A portion of the partial order defined over $(a_6,\\\\ldots,a_9)$.\\n\\t}\\n\\t\\\\label{fig:defense_input}\\n\\\\end{figure}\",\n \"id\": \"1e0de86d-a9b2-453b-b977-3d5abab5bd99\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"8e32af6c-587d-43e1-aabc-dc5d8345628d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Survey and Systematization Methodology\"\n ],\n [\n \"subsection\",\n \"Framework\"\n ],\n [\n \"subsubsection\",\n \"Systematizing Defenses\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Systematizing Defenses\"\n },\n {\n \"cite_extract_rate\": 0.8,\n \"cites\": [\n 917,\n 3866,\n 3870,\n 3854\n ],\n \"content\": \"\\\\label{sec:ml_model_property}\\nSince $f=\\\\varphi(\\\\phi(\\\\cdot))$, we decompose $f$'s security properties into $\\\\varphi$'s and $\\\\phi$'s. We consider: Representation Robustness ({\\\\sf RR}), meaning that two similar files have similar feature representations; Classification Robustness ({\\\\sf CR}), meaning that two similar feature representations lead to the same label; Detection Robustness ({\\\\sf DR}), meaning that feature extraction function $\\\\phi$ returns similar representations for two files with the same functionality; Training Robustness ({\\\\sf TR}), meaning that small changes to the training set does not cause any significant change to the learned classifier. \\nWith respect to small perturbations, Definitions \\\\ref{definition:representation-robustness} and\\n\\\\ref{definition:classification-robustness} below collectively say that when two files $z$ and $z'$ are similar, they would be classified as the same label with a high probability. Since the classification function $\\\\varphi$ is linear, we can obtain a $\\\\epsilon$-robust $\\\\varphi$ analytically, where $\\\\epsilon$ is a small scalar that bounds the perturbations applied to feature vectors . This means that the main challenge is to achieve robust feature extraction.\\n\\\\begin{definition}[{\\\\sf RR} or $(\\\\epsilon,\\\\eta)$-robust feature extraction; adapted from ] \\n\\\\label{definition:representation-robustness}\\nGiven constants $\\\\epsilon,\\\\eta\\\\in [0,1]$, and files $z,z' \\\\in \\\\mathcal{Z}$ such that $(\\\\Gamma(z,z')\\\\approx 0) \\\\land ({\\\\tt true}\\\\leftarrow \\\\mathcal{O}(z,z'))$,\\nwe say feature extraction function $\\\\phi$ is $(\\\\epsilon,\\\\eta)$-robust if\\n$$\\\\mathbb{P}(C(\\\\mathbf{x},\\\\mathbf{x}')\\\\leq \\\\epsilon)=\\\\mathbb{P}(C(\\\\phi(z),\\\\phi(z'))\\\\leq \\\\epsilon)>1 - \\\\eta.$$\\n\\\\end{definition}\\n\\\\begin{definition}[{\\\\sf CR} or {$\\\\epsilon$-robust classification} ]\\n\\\\label{definition:classification-robustness} \\nGiven constant $\\\\epsilon\\\\in [0,1]$ as in Definition \\n\\\\ref{definition:representation-robustness} and\\nany feature vectors $\\\\mathbf{x},{\\\\bf x}'\\\\in \\\\mathcal{X}$, we say classification function $\\\\varphi$ is $\\\\epsilon$-robust if $$(C(\\\\mathbf{x},\\\\mathbf{x}') \\\\leq \\\\epsilon) \\\\rightarrow ((\\\\varphi({\\\\bf x})>\\\\tau)\\\\land (\\\\varphi({\\\\bf x}')>\\\\tau)).$$\\n\\\\end{definition}\\nDefinition \\\\ref{definition:representation-robustness2} \\nspecifies detection robustness, which says that feature extraction function $\\\\phi$ returns similar representations for two different files as long as they have the same functionality.\\nNote that Definitions \\\\ref{definition:classification-robustness} and\\n\\\\ref{definition:representation-robustness2} collectively produce a malware detector with detection robustness.\\n\\\\begin{definition}[{\\\\sf DR} or $(\\\\mathcal{O},\\\\eta)$-robust feature extraction; adapted from ] \\n\\\\label{definition:representation-robustness2}\\nGiven constant $\\\\eta\\\\in [0,1]$ and two files $z,z' \\\\in \\\\mathcal{Z}$ such that $(\\\\Gamma(z,z')>>0)\\\\land({\\\\tt true}\\\\leftarrow \\\\mathcal{O}(z,z'))$, we say feature extraction $\\\\phi$ is $(\\\\mathcal{O}, \\\\eta)$-robust if $\\\\mathbb{P}(C(\\\\phi(z),\\\\phi(z'))\\\\leq \\\\epsilon)>1 - \\\\eta.$\\n\\\\end{definition}\\nSuppose we impose a restriction on the adversarial files set $D'_{poison}$ such that $|D'_{poison}|\\\\leq \\\\gamma|D_{train}|$ for some constant $\\\\gamma\\\\in[0,1]$. Let classifier $f'$ be learned from $D_{train} \\\\cup D'_{poison}$. \\nDefinition \\\\ref{def:robust-training} says that a classifier $f'$ learned from poisoned training set can predict as accurately as $f$ learned from $D_{train}$ with a high probability. \\n\\\\begin{definition}[{\\\\sf TR} or $(\\\\gamma,\\\\zeta)$-robust training; adapted from ]\\\\label{def:robust-training}\\nGiven classifiers $f$ learned from $D_{train}$ and $f'$ learned from $D_{train} \\\\cup D'_{poison}$ where\\n$|D_{poison}'| \\\\leq \\\\gamma |D_{train}|$, and small constants $\\\\zeta\\\\in [0,1]$, we say\\n$f'$ is $(\\\\gamma,\\\\zeta)$-robust if $\\\\forall z\\\\in\\\\mathcal{Z}$\\n$$\\\\left((f(z)>\\\\tau) \\\\land (|D_{poison}'| \\\\leq \\\\gamma |D_{train}|)\\\\right) \\\\rightarrow \\\\left(\\\\mathbb{P}(f'(z) > \\\\tau)>1-\\\\zeta\\\\right).$$\\n\\\\end{definition}\",\n \"id\": \"b05fe1bc-da2e-4a87-8f17-d43f1acc81fc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"a1283bee-37c1-4371-bed3-ceeb8af64d75\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Survey and Systematization Methodology\"\n ],\n [\n \"subsection\",\n \"Systematizing Security Properties\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Systematizing Security Properties\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:systematizstion-AMD-studies}\\nWe systematize attacks according to $\\\\mathcal{A}$'s objective, input, assumptions, the security properties that are broken, and the types of victim malware detectors (e.g., Windows vs. Android). Similarly, we systematize defenses according to $\\\\mathcal{I}$'s objective, input, assumptions, the security properties that are achieved, and the types of enhanced malware detectors (e.g., Windows vs. Android).\\nWe group attacks (defenses) according to the attacker's (defender's) techniques and then summarize them in a table according to the publication date in chronological order. \\nFor convenience, we will use wildcard $*$ to indicate any value in a domain (e.g., $[0,1]$); we will use $\\\\lor$ to describe $\\\\mathcal{A}$'s and $\\\\mathcal{I}$'s ``broader'' input (if applicable). For example, $(0,1,0,1,0|A_6,\\\\ldots,A_9)\\\\lor (1,0,1,1,1|A_6,\\\\ldots,A_9)$ means that $\\\\mathcal{A}$ has either $(a_1,a_2,a_3,a_4,a_5)=(0,1,0,1,0)$ or\\n$(a_1,a_2,a_3,a_4,a_5)=(1,0,1,1,1)$. Finally, we will present the attack-defense escalation.\",\n \"id\": \"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"4fe25043-06f8-4577-bd62-3fe42c7ab899\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ]\n ],\n \"subsections\": [\n \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"76e4c68c-dec2-4af2-b3fa-ce426944cf52\"\n ],\n \"title\": \"Systematizing AMD Arms Race\"\n },\n {\n \"cite_extract_rate\": 0.423076923076923,\n \"cites\": [\n 3867,\n 3857,\n 6172,\n 7756,\n 3864,\n 3862,\n 937,\n 3871,\n 9099,\n 3854,\n 7754\n ],\n \"content\": \"\\\\label{sec:attack}\\n\\\\begin{table}[htbp]\\n\\\\caption{Summary of AMD attacks (\\\\checkmark means applicable, \\\\protect\\\\fullcirc means 0, \\\\protect\\\\emptycirc means 1, \\\\protect\\\\dotcirc means a value in $[0,1]$).}\\n\\t\\\\centering\\n\\t\\\\resizebox{0.9\\\\columnwidth}{!}{\\n\\t\\\\setlength{\\\\tabcolsep}{0.4em}\\n\\t\\\\begin{tabular}{l|ccc|ccccc|cccc|ccccc|cccc|ccc}\\n\\t\\\\hline\\n\\t\\\\multicolumn{1}{c|}{\\\\specialcell{Attack \\\\\\\\ (in chronological order)}}& \\\\multicolumn{3}{c|}{\\\\specialcell{Attack \\\\\\\\ Objective}} & \\\\multicolumn{9}{c|}{\\\\specialcell{Attack Input}} & \\n\\t\\\\multicolumn{5}{c|}{Assumptions} &\\n\\t\\\\multicolumn{4}{c|}{\\\\specialcell{Broken \\\\\\\\ Properties}} & \\\\multicolumn{3}{c}{\\\\specialcell{Malware \\\\\\\\ detector}} \\\\\\\\\\n\\t& \\\\vthead{\\\\sf Indiscriminate} \\n\\t& \\\\vthead{\\\\sf Targeted}\\n\\t& \\\\vthead{\\\\sf Availability}\\n\\t& \\\\vthead{$A_1$: Training set $D_{train}$}\\n\\t& \\\\vthead{$A_2$: Defense technique}\\n\\t& \\\\vthead{$A_3$: Feature set}\\n\\t& \\\\vthead{$A_4$: Learning algorithm}\\n\\t& \\\\vthead{$A_5$: Response}\\n\\t& \\\\vthead{$A_6$: Manipulation set}\\n\\t& \\\\vthead{$A_7$: Attack tactic}\\n\\t& \\\\vthead{$A_8$: Attack technique}\\n\\t& \\\\vthead{$A_9$: Adversarial example set}\\n\\t& \\\\vthead{{\\\\sf IID} assumption}\\n\\t& \\\\vthead{{\\\\sf Oracle} assumption}\\n\\t& \\\\vthead{{\\\\sf Measurability} assumption}\\n\\t& \\\\vthead{{\\\\sf Smoothness} assumption}\\n\\t& \\\\vthead{{\\\\sf Invertibility} assumption}\\n\\t& \\\\vthead{{\\\\sf RR}: Representation Robustness}\\n\\t& \\\\vthead{{\\\\sf CR}: Classification Robustness}\\n\\t& \\\\vthead{{\\\\sf DR}: Detection Robustness}\\n\\t& \\\\vthead{{\\\\sf TR}: Training Robustness}\\n\\t& \\\\vthead{Windows Program}\\n\\t& \\\\vthead{Android Package}\\n\\t& \\\\vthead{PDF}\\n\\t\\\\\\\\\\\\hline\\\\hline\\n \\\\rowcolor{lightgray!30}\\n Smutz and Stavrou & \\n \\\\checkmark &&&\\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} &\\n $\\\\mathbf{M}$ &\\n {\\\\sf OE2} &\\n {\\\\sf MI} &\\n $\\\\mathcal{X}_\\\\mathbf{M}$ &\\n & & \\\\checkmark & & &\\n & \\\\checkmark & & &\\n & & \\\\checkmark\\n \\\\\\\\\\n Biggio et al. &\\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pieg} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} &\\n $\\\\mathbf{M}$ &\\n ${\\\\sf OE2}$ &\\n ${\\\\sf GO}$ &\\n $\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n & \\\\checkmark & \\\\checkmark && \\\\checkmark &\\n & \\\\checkmark & &&\\n & & \\\\checkmark \\n \\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Maiorca et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} &\\n \\\\specialcell{\\\\pie{4}} &\\n \\t$\\\\mathcal{M}$ &\\n \\t${\\\\sf BE}$ &\\n \\t{\\\\sf MI} &\\n \\t$\\\\mathcal{Z}_{\\\\mathcal{M}}$ &\\n \\t& \\\\checkmark & & & &\\n \\t& & \\\\checkmark & &\\n \\t& & \\\\checkmark \\n \\t\\\\\\\\\\n {\\\\v{S}}rndi\\\\'{c} and Laskov &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4}} & \\n \\\\specialcell{\\\\pieg \\\\\\\\ \\\\pieg \\\\\\\\ \\\\pieg \\\\\\\\ \\\\pieg} & \\n \\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pieg \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pieg} &\\n \\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4}} &\\n \\t\\\\specialcell{$\\\\mathbf{M}$ \\\\\\\\ $\\\\mathcal{M}$ \\\\\\\\ $\\\\mathbf{M}$ \\\\\\\\ $\\\\mathcal{M}$} &\\n \\t\\\\specialcell{${\\\\sf BE}$ \\\\\\\\ ${\\\\sf BE}$ \\\\\\\\ ${\\\\sf OE2}$ \\\\\\\\ ${\\\\sf OE2}$ } &\\n \\t\\\\specialcell{${\\\\sf TR}$ \\\\\\\\ ${\\\\sf TR}$ \\\\\\\\ ${\\\\sf TR}$ \\\\\\\\ ${\\\\sf TR}$} &\\n \\t\\\\specialcell{$\\\\mathcal{X}_{\\\\mathbf{M}}$ \\\\\\\\ $\\\\mathcal{Z}_{\\\\mathcal{M}}$ \\\\\\\\ $\\\\mathcal{X}_{\\\\mathbf{M}}$ \\\\\\\\ $\\\\mathcal{Z}_{\\\\mathcal{M}}$} &\\n \\t& \\\\checkmark &&&\\\\checkmark& \\n \\t&&\\\\checkmark&&\\n \\t& & \\\\checkmark \\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Xu et al.~ &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} &\\n \\\\specialcell{\\\\pie{0}} &\\n \\t$\\\\mathcal{M}$ &\\n \\t${\\\\sf BE}$ &\\n \\t{\\\\sf HS} &\\n \\t$\\\\mathcal{Z}_{\\\\mathcal{M}}$ &\\n \\t& & & & &\\n \\t& & \\\\checkmark &&\\n \\t& & \\\\checkmark \\n \\t\\\\\\\\\\n \\tCarmony et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\t\\\\specialcell{\\\\pie{4}} &\\n \\t\\\\specialcell{\\\\pie{4}} &\\n \\t$\\\\mathcal{M}$ &\\n \\t${\\\\sf BE}$ &\\n \\t{\\\\sf MI} &\\n \\t$\\\\mathcal{Z}_{\\\\mathcal{M}}$ &\\n \\t& \\\\checkmark & & & &\\n \\t& & \\\\checkmark & &\\n \\t& & \\\\checkmark \\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Hu and Tan & \\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{4}} &\\n \\\\specialcell{\\\\pie{0}} &\\n $\\\\mathbf{M}$ &\\n ${\\\\sf BE}$ &\\n {\\\\sf GM} &\\n $\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n \\\\checkmark& \\\\checkmark & & & \\\\checkmark& \\n & & \\\\checkmark &&\\n \\\\checkmark & & \\n \\\\\\\\\\n Hu and Tan & \\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{4}} &\\n \\\\specialcell{\\\\pie{0}} &\\n $\\\\mathbf{M}$ &\\n ${\\\\sf BE}$ &\\n {\\\\sf GM} &\\n $\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n \\\\checkmark& \\\\checkmark & & & \\\\checkmark &\\n & & \\\\checkmark &&\\n \\\\checkmark & & \\n \\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Demontis et al. &\\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} &\\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pieg} &\\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} &\\n $\\\\mathbf{M}$ &\\n ${\\\\sf OE2}$ &\\n {\\\\sf SF} &\\n $\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n & \\\\checkmark & \\\\checkmark & & \\\\checkmark &\\n & \\\\checkmark & &&\\n & \\\\checkmark & \\n \\\\\\\\\\n Grosse et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\t$\\\\mathbf{M}$ &\\n \\t${\\\\sf OE2}$ &\\n \\t{\\\\sf SF} &\\n \\t$\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n \\t& \\\\checkmark &\\\\checkmark & & \\\\checkmark &\\n \\t\\\\checkmark & \\\\checkmark & &&\\n \\t& \\\\checkmark & \\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Chen et al. &\\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} &\\n $\\\\mathbf{M}$ &\\n ${\\\\sf OE2}$ &\\n {\\\\sf SF} &\\n $\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n & \\\\checkmark & \\\\checkmark & & \\\\checkmark & \\n & \\\\checkmark&&&\\n \\\\checkmark & & \\n \\\\\\\\\\n Khasawneh et al. &\\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pieg} &\\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\\\specialcell{ $\\\\mathbf{M}$ \\\\\\\\ $\\\\mathcal{M}$} &\\n \\\\specialcell{${\\\\sf BE}$ \\\\\\\\ ${\\\\sf BE}$} &\\n \\\\specialcell{{\\\\sf TR} \\\\\\\\ {\\\\sf TR}} &\\n \\\\specialcell{$\\\\mathcal{X}_{\\\\mathbf{M}}$ \\\\\\\\ $\\\\mathcal{Z}_{\\\\mathcal{M}}$} &\\n & \\\\checkmark & & & \\\\checkmark & \\n & \\\\checkmark & &&\\n \\\\checkmark & & \\n \\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Dang et al. &\\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} &\\n \\\\specialcell{\\\\pie{4}} &\\n $\\\\mathcal{M}$ &\\n ${\\\\sf BE}$ &\\n {\\\\sf HS} &\\n $\\\\mathcal{Z}_{\\\\mathcal{M}}$ &\\n & & & & &\\n & & \\\\checkmark &&\\n & & \\\\checkmark \\n \\\\\\\\\\n Mu{\\\\~n}oz-Gonz{\\\\'a}lez et al. & \\n &&\\\\checkmark& \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4}} &\\n $\\\\mathbf{M}$ &\\n ${\\\\sf OP}$ &\\n {\\\\sf GO} &\\n $\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n & \\\\checkmark & & &\\\\checkmark &\\n & & & \\\\checkmark &\\n \\\\checkmark & & \\n \\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Yang et al. &\\n \\\\checkmark &&& \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} &\\n \\\\specialcell{\\\\pie{0}} &\\n $\\\\mathcal{M}$ &\\n ${\\\\sf BE}$ &\\n ${\\\\sf HS}$ &\\n $\\\\mathcal{Z}_{\\\\mathcal{M}}$ &\\n & \\\\checkmark & \\\\checkmark & & &\\n & & \\\\checkmark &&\\n & \\\\checkmark & \\n \\\\\\\\\\n Rosenberg et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\t\\\\specialcell{$\\\\mathbf{M}$ \\\\\\\\ $\\\\mathcal{M}$} &\\n \\t\\\\specialcell{${\\\\sf BE}$ \\\\\\\\ ${\\\\sf BE}$} &\\n \\t\\\\specialcell{{\\\\sf TR} \\\\\\\\ {\\\\sf TR}} &\\n \\t\\\\specialcell{$\\\\mathcal{X}_{\\\\mathbf{M}}$ \\\\\\\\ $\\\\mathcal{Z}_{\\\\mathcal{M}}$} &\\n \\t& \\\\checkmark & & & \\\\checkmark &\\n \\t\\\\checkmark & \\\\checkmark & &&\\n \\t\\\\checkmark & & \\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Anderson et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{4}} &\\n \\\\specialcell{\\\\pie{0}} &\\n \\t$\\\\mathcal{M}$ &\\n \\t${\\\\sf OE2}$ &\\n \\t{\\\\sf GM} &\\n \\t$\\\\mathcal{Z}_{\\\\mathcal{M}}$ \\n \\t&\\\\checkmark& \\\\checkmark& \\\\checkmark & & \\n \\t& \\\\checkmark & \\\\checkmark &&\\n \\t& \\\\checkmark & & \\n \\t\\\\\\\\\\n \\tKreuk et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\t\\\\specialcell{\\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{0}} &\\n \\t$\\\\mathbf{M}$ &\\n \\t${\\\\sf OE2}$ &\\n \\t{\\\\sf GO} &\\n \\t$\\\\mathcal{X}_{\\\\mathbf{M}}$ \\n \\t& &\\\\checkmark& & & \\\\checkmark\\n \\t& \\\\checkmark & \\\\checkmark &&\\n \\t& \\\\checkmark & & \\n \\t\\\\\\\\\\n \\t\\\\rowcolor{lightgray!30}\\n \\tChen et al. & \\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{0}\\\\\\\\\\\\pie{4}\\\\\\\\ \\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4}} & \\n \\t\\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pieg} & \\n \\t\\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{0}\\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0}} &\\n \\t$\\\\mathbf{M}$ &\\n \\t{\\\\sf BP} &\\n \\t{\\\\sf SF} &\\n \\t$\\\\mathcal{X}_\\\\mathbf{M}$ &\\n \\t& \\\\checkmark & \\\\checkmark & & \\\\checkmark & \\n \\t& & & \\\\checkmark &\\n \\t& \\\\checkmark & \\n \\t\\\\\\\\\\n Al-Dujaili et al. & \\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\t$\\\\mathbf{M}$ &\\n \\t${\\\\sf OE2}$ &\\n \\t{\\\\sf GO} &\\n \\t$\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n \\t& \\\\checkmark & & & \\\\checkmark &\\n \\t&& \\\\checkmark &&\\n \\t\\\\checkmark & & \\n \\t\\\\\\\\\\n \\t\\\\rowcolor{lightgray!30}\\n \\tSuciu et al. & \\n \\t&\\\\checkmark&& \\n \\t\\\\specialcell{\\\\pieg \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pieg \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pieg \\\\\\\\ \\\\pieg \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{0}} & \\n \\t\\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{0}} &\\n \\t\\\\specialcell{$\\\\mathbf{M}$ \\\\\\\\ $\\\\mathcal{M}$} &\\n \\t\\\\specialcell{${\\\\sf BP}$ \\\\\\\\ ${\\\\sf BP}$} &\\n \\t\\\\specialcell{{\\\\sf SF} \\\\\\\\ {\\\\sf SF}} &\\n \\t\\\\specialcell{$\\\\mathcal{X}_{\\\\mathbf{M}}$ \\\\\\\\ $\\\\mathcal{Z}_{\\\\mathcal{M}}$} &\\n \\t& \\\\checkmark & \\\\checkmark & & \\\\checkmark & \\n \\t& & &\\\\checkmark&\\n \\t& \\\\checkmark & \\n \\t\\\\\\\\\\n Kolosnjaji et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\t$\\\\mathbf{M}$ &\\n \\t${\\\\sf OE2}$ &\\n \\t{\\\\sf GO} &\\n \\t$\\\\mathcal{X}_{\\\\mathbf{M}}$ \\n \\t&& \\\\checkmark&& & \\\\checkmark\\n \\t& \\\\checkmark & \\\\checkmark &&\\n \\t& \\\\checkmark & & \\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Suciu et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} & \\n \\\\specialcell{\\\\pie{0}} &\\n \\t$\\\\mathbf{M}$ &\\n \\t${\\\\sf OE2}$ &\\n \\t{\\\\sf GO} &\\n \\t$\\\\mathcal{X}_{\\\\mathbf{M}}$ &\\n \\t& \\\\checkmark & & & \\\\checkmark &\\n \\t\\\\checkmark & \\\\checkmark &&&\\n \\t\\\\checkmark & & \\n \\t\\\\\\\\\\n Chen et al. &\\n \\t\\\\checkmark &&& \\n \\t\\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0}} &\\n \\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4}} &\\n \\\\specialcell{\\\\pie{0} \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{0}} &\\n \\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{4}} &\\n \\\\specialcell{\\\\pie{4} \\\\\\\\ \\\\pie{0} \\\\\\\\ \\\\pie{4} \\\\\\\\ \\\\pie{0}} &\\n \\t\\\\specialcell{$\\\\mathbf{M}$ \\\\\\\\ $\\\\mathbf{M}$} &\\n \\t\\\\specialcell{${\\\\sf OE1}$ \\\\\\\\ ${\\\\sf OE2}$} &\\n \\t\\\\specialcell{{\\\\sf GO} \\\\\\\\ {\\\\sf SF}} &\\n \\t\\\\specialcell{$\\\\mathcal{X}_{\\\\mathbf{M}}$\\\\\\\\ $\\\\mathcal{X}_{\\\\mathbf{M}}$} &\\n \\t& \\\\specialcell{\\\\checkmark } & \\\\specialcell{\\\\checkmark } & & \\\\checkmark &\\n \\t& \\\\checkmark & & &\\n \\t& \\\\checkmark & \\n \\t\\\\\\\\\\n\\t\\\\rowcolor{lightgray!30}\\n\\tPierazzi et al. &\\n\\t\\\\checkmark &&& \\n\\t\\\\specialcell{\\\\pie{0}} & \\n\\t\\\\specialcell{\\\\pie{0}} &\\n\\t\\\\specialcell{\\\\pie{0}} & \\n\\t\\\\specialcell{\\\\pie{0}} &\\n\\t\\\\specialcell{\\\\pie{0}} &\\n\\t$\\\\mathcal{M}$ &\\n\\t${\\\\sf OE2}$ &\\n\\t{\\\\sf MI} &\\n\\t$\\\\mathcal{Z}_{\\\\mathbf{M}}$ &\\n\\t& \\\\checkmark & & & &\\n\\t& & \\\\checkmark & & \\n\\t& \\\\checkmark & \\n\\t\\\\\\\\ \\t \\n\\tLi and Li & \\n\\t\\\\checkmark &&& \\n\\t\\\\specialcell{\\\\pie{4}} & \\n\\t\\\\specialcell{\\\\pie{0}} & \\n\\t\\\\specialcell{\\\\pie{0}} & \\n\\t\\\\specialcell{\\\\pie{0}} & \\n\\t\\\\specialcell{\\\\pie{0}} &\\n\\t\\\\specialcell{$\\\\mathbf{M}$ \\\\\\\\ $\\\\mathcal{M}$} &\\n\\t\\\\specialcell{${\\\\sf OE2}$ \\\\\\\\ ${\\\\sf OE2}$} &\\n\\t\\\\specialcell{{\\\\sf MS} \\\\\\\\ {\\\\sf MS}} &\\n\\t\\\\specialcell{$\\\\mathcal{X}_{\\\\mathbf{M}}$ \\\\\\\\ $\\\\mathcal{Z}_{\\\\mathcal{M}}$} &\\n\\t& \\\\checkmark & & & \\\\checkmark &\\n\\t&& \\\\checkmark &&\\n\\t& \\\\checkmark & \\n\\t\\\\\\\\\\\\hline\\n\\t\\\\end{tabular}\\n}\\n\\t\\\\label{tbl:attacks}\\n\\\\end{table}\",\n \"id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 26,\n \"parent_id\": \"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ]\n ],\n \"subsections\": [\n \"f3db0c13-9b45-4545-9019-43245444f75d\",\n \"6dae1197-fcde-489c-835b-7117fc4737e8\",\n \"6bd1592d-a0b4-4061-b526-553c6ebc409a\",\n \"4e7e5878-d20a-4c28-acf4-a41004766a2c\",\n \"b8ab1fcb-f749-4c70-9893-28b772af455c\",\n \"33314056-9caf-48b3-90ff-0ca708a5a80c\",\n \"c51df6e6-79f5-400d-8643-cef009ff014f\",\n \"d4bb8301-3585-4f94-b9fa-12ca7a7b0164\"\n ],\n \"title\": \"Systematizing Attack Literature\"\n },\n {\n \"cite_extract_rate\": 0.8235294117647051,\n \"cites\": [\n 892,\n 1193,\n 923,\n 3867,\n 3873,\n 3862,\n 3872,\n 937,\n 3871,\n 3863,\n 9099,\n 3855,\n 3854,\n 3861\n ],\n \"content\": \")}\\nBiggio et al.~ propose solving the problem of optimal evasion attacks by leveraging gradient-based optimization techniques. They focus on high-confidence evasion attacks with small perturbations (cf. Definition \\\\ref{def:opt_evasion2}). Given a malware representation-label pair $(\\\\mathbf{x},y=+)$, the optimization problem specified in Eq.\\\\eqref{eq:a8:evs2} with respect to {\\\\sf GO} is instantiated as:\\n\\\\begin{align*}\\n\\t\\\\max_{\\\\delta_{\\\\mathbf{x}}} L_\\\\mathcal{A}(\\\\hat{F}_{\\\\hat\\\\theta}(\\\\mathbf{x}+\\\\delta_{\\\\mathbf{x}}),y=+) & = \\\\min_{\\\\delta_{\\\\mathbf{x}}} \\\\left(L(F_\\\\theta(\\\\mathbf{x}+\\\\delta_{\\\\mathbf{x}}, y=-)) - \\\\beta_a\\\\mathcal{K}(\\\\mathbf{x} + \\\\delta_{\\\\mathbf{x}})\\\\right) \\\\\\\\ ~~\\\\text{s.t.}~~(\\\\delta_{\\\\mathbf{x}}\\\\in[\\\\mathbf{0},\\\\overline{\\\\mathbf{u}}]) &\\\\land (C(\\\\mathbf{x},\\\\mathbf{x}+\\\\delta_{\\\\mathbf{x}})\\\\leq m),\\n\\\\end{align*}\\nwhere $\\\\beta_a\\\\geq 0$ is a balance factor and $\\\\mathcal{K}$ is a density estimation function for lifting $\\\\mathbf{x}+\\\\delta_{\\\\mathbf{x}}$ to the populated region of benign examples. Since $\\\\delta_{\\\\mathbf{x}}\\\\geq \\\\mathbf{0}$, the manipulation only permits object injections to meet the requirement of preserving malicious functionalities. The attack is validated by using the PDF malware detector and the feature representation is the number of appearances of hand-selected keywords (e.g., JavaScript). Because the perturbation is continuous, the authors suggest searching a discrete point close to the continuous one and aligning the point with $\\\\nabla L_\\\\mathcal{A}(\\\\mathbf{x}+\\\\delta_{\\\\mathbf{x}},y=+)$. \\nThis attack makes the invertibility Assumption \\\\ref{assumption:inverse} because it operates in the feature space. Experimental results show that when $\\\\mathcal{I}$ employs no countermeasures, knowing $\\\\mathcal{I}$'s feature set $S$ and learning algorithm $F$ are sufficient for $\\\\mathcal{A}$ to evade $\\\\mathcal{I}$'s detector. This attack and its variants have been shown to evade PDF malware detectors , PE malware detectors , Android malware detectors , and Flash malware detectors . The kernel density estimation item makes the perturbed representation $\\\\mathbf{x}+\\\\delta_\\\\mathbf{x}$ similar to the representations of benign examples, explaining the successful evasion. In summary, the attack works under the ${\\\\sf Oracle}$, ${\\\\sf Measurability}$, and {\\\\sf Invertibility} assumptions. $\\\\mathcal{A}$'s input is, or $\\\\mathcal{A}$ can be characterized as, $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) = (1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_{\\\\mathbf{M}}) \\\\lor (0,0,1,*,0|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_{\\\\mathbf{M}})$ and $\\\\mathcal{A}$ breaks the {\\\\sf CR} property.\\nAl-Dujaili et al.~ propose evasion attacks against DNN-based malware detectors in the feature space. In this attack,\\n$\\\\mathcal{A}$ generates adversarial examples with possibly large perturbations in the feature space. More precisely, given a representation-label pair $(\\\\mathbf{x},y=+)$, the optimization problem of Eq.\\\\eqref{eq:a8:evs2} with respect to {\\\\sf GO} is instantiated as:\\n\\t$\\\\max_{\\\\delta_{\\\\mathbf{x}}} L(F_{\\\\theta}(\\\\mathbf{x}+\\\\delta_{\\\\mathbf{x}}),y=+) ~~\\\\text{s.t.}~~ \\\\delta_{\\\\mathbf{x}} \\\\in [\\\\mathbf{0}, \\\\mathbf{1}-\\\\mathbf{x}].$\\nThe attack has four variants, with each perturbing the representation in a different direction (e.g., normalized gradient of the loss function using the $\\\\ell_\\\\infty$ norm). A ``random'' rounding operation is used to map continuous perturbations into a discrete domain. When compared with the basic rounding (which returns 0 if the input is smaller than 0.5, and returns 1 otherwise), the ``random'' rounding means that the threshold of rounding is sampled from the interval $[0,1]$ uniformly. For binary feature representation, the manipulation set $\\\\mathbf{M}_\\\\mathbf{x}=[\\\\mathbf{0}, \\\\mathbf{1}-\\\\mathbf{x}]$ assures the flipping of 0 to 1. The effectiveness of the attack is validated using Windows malware detector in the feature space. In summary, the attack works under the {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with $\\\\mathcal{A}$ input $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) =(0,1,1,1,1|\\\\mathbf{M}, {\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$ and breaks the {\\\\sf DR} property.\\nKreuk et al. propose an evasion attack in the feature space against MalConv, which is an end-to-end Windows PE malware detector (as reviewed in Section \\\\ref{sec:ml}) . Given a malware embedding code $\\\\mathbf{x}$, the optimization problem of Eq.\\\\eqref{eq:a8:evs2} with respect to {\\\\sf GO} is instantiated as:\\n\\t$\\\\min_{\\\\delta_{\\\\mathbf{x}}} L(F_{\\\\theta}([\\\\mathbf{x}|\\\\delta_{\\\\mathbf{x}}]),y=-) ~~\\\\text{s.t.}~~ \\\\Vert\\\\delta_{\\\\mathbf{x}}\\\\Vert_p \\\\leq \\\\epsilon,$\\nwhere | means concatenation and $\\\\Vert\\\\cdot\\\\Vert_p$ is the $p$ norm where $p\\\\geq 1$. Because MalConv is learned from sequential data, perturbation means appending some content to the end of a PE file. Perturbations are generated in a single step by following the direction of the $\\\\ell_\\\\infty$ or $\\\\ell_2$ normalized gradients of the loss function . For instance, the attack based on the $\\\\ell_{\\\\infty}$ norm is\\n$\\\\tilde{\\\\mathbf{x}}'=\\\\mathbf{x} - \\\\epsilon\\\\cdot\\\\sign(\\\\nabla_{\\\\bf x}(L(F_\\\\theta(\\\\mathbf{x}), -))$, where $\\\\sign(x)=+1~(-1)$ if $x\\\\geq0~(x<0)$.\\nSince the embedding operation uses a look-up table to map discrete values (0, 1, $\\\\ldots$, 255) to the learned real-value vectors, the attack uses a nearest neighbor search to look for the learned embedding code close to $\\\\tilde{\\\\mathbf{x}}'$. \\nIn summary, the attack works under the {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$ and breaks the {\\\\sf RR} and {\\\\sf CR} properties.\\nKolosnjaji et al. and Suciu et al. independently propose gradient-based attacks in the feature space to evade MalConv . Both studies also use the loss function exploited by Kreuk et al. . Kolosnjaji et al. use the manipulation set $\\\\mathcal{M}$ corresponding to appending instructions at the end of a file. This attack proceeds iteratively and starts with randomly initialized perturbations. In each iteration, continuous perturbations are updated in the direction of the $\\\\ell_2$ normalized gradient of the loss function with respect to the input, and then a nearest neighbor search is applied to obtain discrete perturbations. Suciu et al. perturb embedding codes in the direction of the $\\\\ell_\\\\infty$ normalized gradient of the loss function, while adding instructions in the mid of a PE file (e.g., between PE {\\\\em sections}) while noting that appended content could be truncated by MalConv. Both attacks work under the {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$ and break the {\\\\sf RR} and {\\\\sf CR} properties.\\nMu{\\\\~n}oz-Gonz{\\\\'a}lez et al. propose the optimal poisoning {\\\\sf OP} attack in the feature space (Definition \\\\ref{def:optimal poisoning attack}), which is NP-hard. In this case, the optimization problem of Eq.\\\\eqref{eq:a8:poi} is relaxed by supposing that the classifier is linear to render the optimization problem tractable . The attack is waged against Windows PE malware detectors. Feature set includes API calls, actions and modifications in the file system; each file is represented by a binary vector. The attack has two variants: one uses white-box input, where $\\\\mathcal{A}$ derives $D_{poison}'$ from $\\\\mathcal{I}$'s detector $f$; the other uses grey-box input, where $\\\\mathcal{A}$ knows $\\\\mathcal{I}$'s training set as well as feature set and trains a surrogate detector. The attack works under the {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) =(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OP}, {\\\\sf GO}, \\\\mathcal{X}_\\\\mathbf{M})\\\\lor(1,0,1,0,0|\\\\mathbf{M},{\\\\sf OP}, {\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$ and breaks {\\\\sf TR}.\",\n \"id\": \"f3db0c13-9b45-4545-9019-43245444f75d\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Attacks using Gradient-based Optimization ({\\\\sf GO\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks using Gradient-based Optimization ({\\\\sf GO\"\n },\n {\n \"cite_extract_rate\": 0.375,\n \"cites\": [\n 894,\n 890,\n 3864,\n 7754,\n 3863,\n 7154\n ],\n \"content\": \")}\\nDemontis et al. propose the optimal evasion {\\\\sf OE2} in the feature space to perturb important features in terms of their weights in the linear function $\\\\varphi(\\\\mathbf{x})={\\\\mathbf{w}}^\\\\top\\\\mathbf{x} + b$, where $\\\\mathbf{w}=(w_1,w_2,\\\\cdots,w_d)$ is a weight vector, $b$ is the bias, and $d$ is the dimension of feature space. The attack is waged against Drebin malware detector (which is reviewed in Section \\\\ref{sec:ml}). $\\\\mathcal{A}$ manipulates the $x_i$'s with largest $|w_i|$'s as follows:\\nflip $x_i=1$ to $x_i=0$ if $w_i > 0$, flip $x_i=0$ to $x_i=1$ if $w_i < 0$, and do nothing otherwise, while obeying the manipulation set $\\\\mathbf{M}$ corresponding to the injection or removal of features. The attack works under the {\\\\sf Oracle}, {\\\\sf Measurability}, and {\\\\sf Invertiblity} assumptions with\\ninput $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (0,0,1,*,0|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$ and breaks the {\\\\sf CR} property.\\nGrosse et al. propose a variant of the Jacobian-based Saliency Map Attack (JSMA) in the feature space against the Drebin malware detector (which is reviewed in Section \\\\ref{sec:ml}). Instead of using SVM, Deep Neural Network (DNN) is used to build a detector. Important features are identified by leveraging the gradients of the softmax output of a malware example with respect to the input. A large gradient value indicates a high important feature. $\\\\mathcal{A}$ only injects manifest features to manipulate Android Packages and generates adversarial files from $\\\\mathcal{I}$'s detector $f$. The attack works under the {\\\\sf Oracle}, {\\\\sf Measurability}, and {\\\\sf Invertiblity} assumptions\\nwith input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$ and breaks the {\\\\sf RR} and {\\\\sf CR} properties.\\nChen et al. propose an evasion attack in the feature space by perturbing the important features derived from a wrapper-based feature selection algorithm . The attacker's loss function $L_\\\\mathcal{A}$ has two parts: (i) the classification error in the mean squared loss and (ii) the manipulation cost $C(\\\\mathbf{x}, \\\\mathbf{x}')=\\\\sum_{i=1}^{d}c_i|x_i - x_i'|$, where $\\\\mathbf{x}=(x_1,\\\\ldots,x_d)$, $\\\\mathbf{x}'=(x_1', \\\\ldots,x_d')$, and $c_i$ is the hardness of perturbing the $i$th feature while preserving malware's functionality. The attack is waged against Windows PE malware detector that uses hand-crafted Windows API calls as features and the binary feature representation. However, there are no details about the composition of manipulation set. This attack works under the {\\\\sf Oracle}, {\\\\sf Measurability}, and {\\\\sf Invertibility} assumptions with input $(a_1,\\\\ldots,a_5|A_7,\\\\cdots,A_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF}, \\\\mathcal{X}_\\\\mathbf{M}))$ and breaks {\\\\sf CR}.\\nChen et al. propose evasion attacks in the feature space against two Android malware detectors, MaMaDroid and Drebin . The manipulation set $\\\\mathbf{M}$ corresponds to the injection of manifest features (e.g., {\\\\em activities}) and \\nAPI calls. $\\\\mathcal{A}$ evades MaMaDroid by using the optimal evasion ${\\\\sf OE1}$ (Definition \\\\ref{def:opt_evasion1}) and ${\\\\sf OE2}$ (Definition \\\\ref{def:opt_evasion2}), and evades Drebin by using ${\\\\sf OE2}$. The optimization problem of ${\\\\sf OE1}$ Eq.\\\\eqref{eq:a8:evs1} is solved using an advanced gradient-based method known as C\\\\&W . ${\\\\sf OE2}$ is solved using JSMA . Because JSMA perturbs sensitive features, we categorize this attack into the {\\\\sf SF} group. The ${\\\\sf OE2}$ attack works under the {\\\\sf Oracle}, {\\\\sf Measurability} and {\\\\sf Invertibility} assumptions,\\nwith four kinds of input \\n$(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,0,1,0,0|\\\\mathbf{M},{\\\\sf OE2}, {\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (0,0,1,0,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (1,0,1,0,0|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF}, \\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (1,0,1,0,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$,\\nand breaks the {\\\\sf CR} property. The ${\\\\sf OE1}$ attack works under the same assumptions with the same input except using attack technique {\\\\sf GO}, and breaks the {\\\\sf CR} property.\\nChen et al. propose a basic poisoning {\\\\sf BP} attack in the feature space against Android malware detectors. The feature set contains syntax features (e.g., permission, hardware, {API}) and semantic features (e.g., sequence of pre-determined program behaviors such as {\\\\tt getDevicedID}$\\\\rightarrow$ URL$\\\\rightarrow${\\\\tt openConnection}). The ML algorithm used is SVM, random forest, or $K$-Nearest Neighbor (KNN) . The malware representations are perturbed using a JSMA variant against the SVM-based classifier (while noting JSMA is applicable neither to random forests nor to KNN because they are gradient-free). Feature manipulation set ${\\\\bf M}$ corresponds to the injection of syntax features. $\\\\mathcal{A}$ poisons $\\\\mathcal{I}$'s training set by injecting perturbed perturbations with label $-$. The attack works under the {\\\\sf Oracle}, {\\\\sf Measurability}, and {\\\\sf Invertibility} assumptions with input\\n$(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (0,0,0,1,1|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (1,0,*,1,1|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$ and breaks {\\\\sf TR}.\\nSuciu et al. propose a basic poisoning attack in both feature and problem spaces. The authors obtain $D'_{poison}$ by applying small manipulations to non-adversarial benign files and then obtain their labels as given by VirusTotal service . \\n$\\\\mathcal{A}$'s objective is to make $\\\\mathcal{I}$'s classifier $f$ mis-classify a targeted malware file $z_{mal}$ as benign. $\\\\mathcal{A}$ proceeds as follow: (i) obtain an initial benign file $z_{ben}$, where $z_{ben} \\\\approx z_{mal}$ in the feature space with respect to the $\\\\ell_1$ norm; (ii) use the JSMA method to manipulate $z_{ben}$ to $z'_{ben}$ by making a small perturbation so that they have similar feature representations; (iii) add $z'_{ben}$ and its label obtained from VirusTotal to $D'_{poison}$ and use $D_{train} \\\\cup D'_{poison}$ to train classifier $f'$ (Definition \\\\ref{definition:sf}); (iv) undo the addition if $z'_{ben}$ lowers the classification accuracy significantly, and accept it otherwise. The attack is waged against the Drebin malware detector and the manipulation set corresponds to the feature injection of permission, API, and strings. This attack works under the {\\\\sf Oracle}, {\\\\sf Measurability}, and {\\\\sf Inversiability} assumptions with input \\n$(a_1,\\\\ldots,a_5|A_6,\\\\ldots,A_9) =(*,0,1,*,0|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(1,1,1,1,1|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(1,0,*,*,0|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (1,0,1,0,0|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$\\nand breaks the {\\\\sf TR} property. The study generates adversarial malware examples, but does not test their malicious functionalities.\",\n \"id\": \"6dae1197-fcde-489c-835b-7117fc4737e8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Attacks using Sensitive Features ({\\\\sf SF\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks using Sensitive Features ({\\\\sf SF\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 7754\n ],\n \"content\": \")} Smutz and Stavrou propose a mimicry attack in the feature space to modify features of a malicious file to mimic benign ones, where $\\\\mathcal{A}$ knows $\\\\mathcal{I}$'s classifier $f$. Discriminative features are identified by observing their impact on classification accuracy. The attack perturbs features of malware examples by replacing their value with the mean of the benign examples. The attack is leveraged to estimate the robustness of PDF malware detectors without considering the preservation of malware functionality. The attack works under the {\\\\sf Measurability} assumption with input $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) = (1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf MI},\\\\mathcal{X})$ and breaks the {\\\\sf CR} property.\\nMaiorca et al. propose a reverse mimicry attack against PDF malware detectors in the problem space. Instead of modifying malicious files to mimic benign ones, $\\\\mathcal{A}$ embeds malicious payload (e.g., JavaScript code) into a benign file. The attack can be enhanced by using parser confusion strategies, which make the injected objects being neglected by feature extractors when rendered by PDF readers . The attack works under the {\\\\sf Oracle} assumption with input $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) = (0,0,0,0,0|$ $\\\\mathcal{M},{\\\\sf BE},{\\\\sf MI},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf DR} property. \\nPierazzi et al. propose a white-box evasion attack against the Drebin malware detector and then an enhanced version of the detector in the problem space . They intend to bridge the gap between the attacks in the problem space and the attacks in the feature space.\\nIn addition, four realistic constraints are imposed on the manipulation set $\\\\mathcal{M}$, including available transformation, preserved semantics, robustness to preprocessing, and plausibility. In order to cope with the side-effect features when incorporating gradient information of $\\\\mathcal{I}$'s classifier, the attacker first harvests a set of manipulations from benign files; \\nManipulations in the problem space are used to query $\\\\mathcal{I}$'s feature extraction for obtaining perturbations in the feature space;\\nan adversarial malware example is obtained by using the manipulations corresponding to the perturbations that have a high impact on the classification accuracy. This attack works under the {\\\\sf Oracle} assumption with input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(1,1,1,1,1|\\\\mathcal{M},{\\\\sf OE2},{\\\\sf MI},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf DR} property.\",\n \"id\": \"6bd1592d-a0b4-4061-b526-553c6ebc409a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Attacks using MImicry ({\\\\sf MI\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks using MImicry ({\\\\sf MI\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \")} \\n{\\\\v{S}}rndi\\\\'{c} and Laskov investigate the mimicry attack and the aforementioned gradient descent and kernel density estimation attack against the PDFrate service, where $\\\\mathcal{A}$ knows some features used by $\\\\mathcal{I}$. $\\\\mathcal{A}$ makes the representation of an adversarial malware example similar to a benign one. Manipulation set $\\\\mathcal{M}$ corresponds to adding objects into PDF files. Both attacks perturb feature vectors against a surrogate model, and then map the perturbed feature representations to the problem space by injecting manipulations between the body and the trailer of a PDF file. For the mimicry attack, $\\\\mathcal{A}$ uses $N_{ben}>0$ benign examples to guide manipulations, resulting in $N_{ben}$ perturbed examples. The example incurring the highest classification error is used as an adversarial example. The attack works under the {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input \\n$( a_1,\\\\ldots,a_5|A_6,\\\\ldots,A_9) =(0,0,*,0,0|\\\\mathbf{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(0,0,*,*,0|\\\\mathbf{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(1,0,*,0,0|\\\\mathbf{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(1,0,*,*,0|\\\\mathbf{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(0,0,*,0,0|\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathcal{M})\\\\lor(0,0,*,*,0|\\\\mathcal{M},$ ${\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathcal{M})\\\\lor(1,0,*,0,0|\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathcal{M})\\\\lor(1,0,*,*,0|\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathcal{M})$\\nand breaks {\\\\sf DR}. The gradient-based attack neglects the constraint of small perturbations and works under the same assumptions with the same input except using the attack technique {\\\\sf OE2}.\\nKhasawneh et al. propose an evasion attack in both the feature and problem spaces against malware detectors learned from dynamic hardware features (e.g., instruction frequency), where $\\\\mathcal{A}$ knows some features used by $\\\\mathcal{I}$. The attack proceeds as follows. $\\\\mathcal{A}$ first queries $\\\\mathcal{I}$'s classifier to obtain a surrogate model and then generates adversarial files against the surrogate model. Manipulation set $\\\\mathcal{M}$ corresponds to the injection of some features because the others (e.g., memory access) are uncontrollable. Perturbations are conducted to the important features that are identified by large weights in the model. The attack works under the {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input $( a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) =(0,0,*,0,1|\\\\mathbf{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(0,0,*,0,1|\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf CR} property.\\nRosenberg et al. propose an evasion attack in both the feature and problem spaces against a Recurrent Neural Network (RNN) based surrogate model, which is learned from API call sequences. In this attack, $\\\\mathcal{A}$'s training data is different from $\\\\mathcal{I}$'s, but the labels are obtained by querying $\\\\mathcal{I}$'s detector. In order to reduce the number of queries to $\\\\mathcal{I}$'s detector, $\\\\mathcal{A}$ augments its training data using the Jacobian-based augmentation technique and modifies the API sequence of an example in the direction of the $\\\\ell_\\\\infty$ normalized gradient of the loss function. Manipulation set ${\\\\bf M}$ corresponds to inserting no-op API calls. Experimental results show that adversarial examples generated from a surrogate RNN model can evade SVM, DNN, and RNN detectors. The attack works under the {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input $( a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) =(0,0,1,0,1|\\\\mathbf{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (0,0,1,0,1|\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf RR} and {\\\\sf CR} properties.\",\n \"id\": \"4e7e5878-d20a-4c28-acf4-a41004766a2c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Attacks using TRansferability ({\\\\sf TR\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks using TRansferability ({\\\\sf TR\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 3857\n ],\n \"content\": \")}\\nXu et al.~ propose black-box evasion attacks in the problem space against two PDF malware detectors known as PDFrate and Hidost , respectively. Given a malicious file $z$, $\\\\mathcal{A}$ uses a genetic algorithm to iteratively generate $z'$ from $z$ as follows: (i) $\\\\mathcal{A}$ manipulates a set of candidates (or $z$ in the initial iteration) via object deletion, insertion, or replacement. (ii) $\\\\mathcal{A}$ queries these variants to $\\\\mathcal{O}$ and $f$. (iii) $\\\\mathcal{A}$ succeeds when obtaining a successful adversarial example $z'$, namely $({\\\\tt True} \\\\leftarrow \\\\mathcal{O}(z,z')) \\\\land (-\\\\leftarrow f(z'))$; otherwise, $\\\\mathcal{A}$ uses a score function to select candidates for the next iteration or aborts after reaching a threshold number of iterations. The score function $h$ varies with classifiers; for PDFrate, $h(\\\\mathcal{O},f,z,z')=0.5-f(z')$ if $\\\\mathcal{O}(z,z')={\\\\sf true}$, and returns -0.5 if $\\\\mathcal{O}(z,z')={\\\\sf false}$. This attack models an {\\\\sf Oracle} and works with the input $(a_1,\\\\cdots, a_5|A_6,\\\\cdots,A_9) = (0,0,0,0,1|\\\\mathcal{M},{\\\\sf BE},{\\\\sf HS},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf DR} property.\\nYang et al. propose evasion attacks against Android malware detectors in the problem space. In this attack, $\\\\mathcal{A}$ also uses a genetic algorithm to perturb a malware example $z$ iteratively. In each iteration, $\\\\mathcal{A}$ extracts some features and calculates similarity scores between the malicious APKs in the feature space; the features that have high impact on the similarity scores are selected; the manipulations are to perturb the selected features.\\nThe attack works under the {\\\\sf Oracle} and {\\\\sf Measurability} assumptions with input $(a_1,\\\\cdots, a_5|A_6,\\\\cdots,A_9) = (0,0,0,0,1|\\\\mathcal{M},\\\\mathcal{Z},{\\\\sf BE}, {\\\\sf HS},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf DR} property.\\nDang et al. propose a black-box evasion attack against malware detectors (e.g., PDFrate) in the problem space. Given a malicious file $z$, $\\\\mathcal{A}$ uses the hill-climbing algorithm to iteratively generate adversarial file $z'$ from $z$. In each iteration, $\\\\mathcal{A}$ generates a path of variants sequentially, each of which is perturbed from its predecessor using manipulations corresponding to object deletion, insertion, or replacement. A score function $h$ is leveraged to select candidates, such as $h(\\\\mathcal{O},f,z,z')={\\\\sf mal}_{z'}-{\\\\sf clf}_{z'}$ or $h(\\\\mathcal{O},f,z,z')={\\\\sf mal}_{z'}/{\\\\sf clf}_{z'}$, where ${\\\\sf mal}_{z'}$ denotes the length of the first example turned from malicious to benign (obtaining by using an oracle) on the manipulation path (cf. Definition \\\\ref{definition:hs}) and ${\\\\sf clf}_{z'}$ denotes the length of the first malware example that has successfully misled the classifier $f$. Both examples of interest are obtained by a binary search, effectively reducing the number of queries to oracle $\\\\mathcal{O}$ and $f$. The attack models an {\\\\sf Oracle} and works with input $(a_1,\\\\cdots, a_5|A_6,\\\\cdots,A_9) = (0,0,0,0,0|\\\\mathcal{M},{\\\\sf BE},{\\\\sf HS},\\\\mathcal{Z}_{\\\\mathcal{M}})$ and breaks the {\\\\sf DR} property.\",\n \"id\": \"b8ab1fcb-f749-4c70-9893-28b772af455c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Attacks using Heuristic Search ({\\\\sf HS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks using Heuristic Search ({\\\\sf HS\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 243,\n 6172\n ],\n \"content\": \")}\\nHu and Tan \\\\cite {Hu2017} propose an evasion attack against Windows malware detectors in the feature space, by using Generative Adversarial Networks (GAN) . In this attack, $\\\\mathcal{A}$ modifies the binary representation of Windows API calls made by malicious files, namely flipping some feature values from 0 to 1. $\\\\mathcal{A}$ learns a generator $G_{\\\\theta_{g}}$ and a discriminator from $\\\\mathcal{A}$'s training dataset. The discriminator is a surrogate detector learned from feature vectors corresponding to $\\\\mathcal{A}$'s benign files and those produced by $G_{\\\\theta_{g}}$, along with labels obtained by querying $\\\\mathcal{I}$'s detector $f$. An adversarial example feature vector is generated by using\\n$\\\\mathbf{x}' = \\\\max(\\\\mathbf{x}, \\\\round(G_{\\\\theta_{g}}(\\\\mathbf{x}, \\\\mathbf{a})))$,\\nwhere $\\\\mathbf{a}$ is a vector of noises, $\\\\round$ is the round function, and $\\\\max$ means element-wise maximum.\\nHu and Tan~ also propose another evasion attack using the Seq2Seq model . Both attacks work under the {\\\\sf IID}, {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,0,1,0,1|\\\\mathbf{M},{\\\\sf BE},{\\\\sf GM},\\\\mathcal{X}_\\\\mathbf{M})$ and break the {\\\\sf DR} property.\\nAnderson et al. propose a Reinforcement Learning (RL)-based evasion attack against Windows PE malware detectors in the problem space. Manipulation set $\\\\mathcal{M}$ is the RL action space, which includes some bytecode injections (e.g., API insertion) and some bytecode deletion. Attacker $\\\\mathcal{A}$ learns an RL agent on $\\\\mathcal{A}$'s data, with labels obtained by querying defender $\\\\mathcal{I}$'s detector $f$. The learned agent predicts manipulations sequentially for a given malware example. Moreover, $\\\\mathcal{A}$ is restricted by only applying a small number of manipulations to a malicious PE file. Experimental results show that the attack is not as effective as others (e.g., gradient-based methods). The attack works under the {\\\\sf Oracle} and {\\\\sf Measurability} assumptions with input $(a_1,\\\\cdots, a_5|A_6,\\\\cdots,A_9) = (0,0,0,0,1|\\\\mathcal{M},{\\\\sf OE2},{\\\\sf GM},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf RR} and {\\\\sf CR} properties.\",\n \"id\": \"33314056-9caf-48b3-90ff-0ca708a5a80c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Attacks using Generative Model ({\\\\sf GM\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks using Generative Model ({\\\\sf GM\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 9099\n ],\n \"content\": \")} \\nLi and Li propose evasion attacks against DNN-based Android malware detectors in both feature and problem spaces. Given four gradient-based attack methods, the attack looks for the best one to perturb malware representations. $\\\\mathcal{A}$ can iteratively perform this strategy to modify the example obtained in the previous iteration. Experimental results show that the mixture of attacks can evade malware detectors effectively. The attack works under the {\\\\sf IID}, {\\\\sf Oracle} and {\\\\sf Invertibility} assumptions with input $(a_1,\\\\cdots, a_5|A_6,\\\\cdots,A_9) = (1,1,1,1,1|\\\\mathbf{M},{\\\\sf BE},{\\\\sf MS},\\\\mathcal{X}_\\\\mathbf{M})\\\\lor(1,1,1,1,1|\\\\mathcal{M},{\\\\sf BE},{\\\\sf MS},\\\\mathcal{Z}_\\\\mathcal{M})$ and breaks the {\\\\sf DR} property.\",\n \"id\": \"c51df6e6-79f5-400d-8643-cef009ff014f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Attacks using Mixture Strategy ({\\\\sf MS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attacks using Mixture Strategy ({\\\\sf MS\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"We summarize the preceding reviews with the following observations.\\n(i) Indiscriminate attacks have been much more extensively investigated than targeted attacks and availability attacks. (ii) Evasion attacks have been much more extensively studied than poisoning attacks. (iii) The {\\\\sf Oracle} assumption has been widely made. \\nIn addition, we draw the following insights.\\n\\\\begin{insight} \\n(i) Knowing the defender's feature set is critical to the success of transfer attacks, highlighting the importance of keeping the defender's feature set secret (e.g., randomizing the defender's feature set).\\n(ii) The effectiveness of evasion attacks largely depends on the attacker's degree of freedom in conducting manipulations in the problem space (i.e., a smaller degree of freedom means it is harder for the attack to succeed).\\n\\\\end{insight}\\n\\\\ignore{\\nIn order to draw further insights, we propose applying ML to the systematized structured data in Table \\\\ref{tbl:attacks}. We propose formulating {\\\\em insights learning} as a multiclass classification problem. We use attributes of attack objective, attack input and assumption as features and treat the broken properties as the labels to learn. We preprocess data as follows:\\nfor features corresponding to attack objective and assumption, we use 1 to represent ``applicable'' and\\n0 otherwise; for features corresponding to $A_1,\\\\ldots,A_5$, we set $a_i =0.5$ for $1\\\\leq i \\\\leq 5$; for features corresponding to categorical $A_6,\\\\dots,A_9$, we use integers to represent the elements (e.g., 0,1 respectively representing $\\\\mathcal{M},\\\\mathbf{M}$ for $A_6$). This leads to a dataset of 63 attacks, each of which has 17 dimensions. We use the dataset to train a {\\\\em random forest} model with default hyper-parameters in scikit-learn . {\\\\color{blue} The random forest is ML algorithm, capable of learning from un-normalized features and producing the important features in terms of classification error .}\\nWe focus on learning important attributes.\\n\\\\begin{figure}[!htbp]\\n\\t\\\\centering\\n\\t\\\\scalebox{0.25}{\\n\\t\\\\includegraphics{figures/attack-important-feat.eps}\\n\\t}\\n\\\\caption{Insights learning from systematized AMD attacks via attribute importance.}\\n\\t\\\\label{fig:attack-attribute}\\n\\\\end{figure}\\nFigure \\\\ref{fig:attack-attribute} highlights the most important features, from which we make the following observations. (i) Attack tactic is the most important feature in determining the security property that is broken. This can be explained attacks using sensitive features ({\\\\sf SF}) usually breaks {\\\\sf CR} or {\\\\sf TR}, and attacks using heuristic search ({\\\\sf HS}) usually breaks the {\\\\sf DR} property. (ii) Attack scenario is the second important feature. This is because ${\\\\sf BE}$ breaks ${\\\\sf DR}$, ${\\\\sf OE2}$ usually breaks {\\\\sf CR} and {\\\\sf RR}, and poisoning attacks ({\\\\sf BP} and {\\\\sf OP}) break {\\\\sf TR}. (iii) Feature set is the third important feature, resonating the manually-drawn insight. (iv) The {\\\\sf Measurability} assumption is the four important feature. This can be explained because {\\\\sf RR} and {\\\\sf CR} are defined with respect to small-manipulation attacks, which need this assumption.\\n\\\\begin{insight}[Insights learned via ML]\\nTechnique-wise, attack tactic and attack scenario (core part of the threat model) largely determines what security property is broken.\\nMethodology-wise, ML is an effective method for insights learning in survey and systematization studies.\\n\\\\end{insight}\\n}\",\n \"id\": \"d4bb8301-3585-4f94-b9fa-12ca7a7b0164\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"108d281b-7e1a-46ce-b750-a8b5a5ef8ec0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Attack Literature\"\n ],\n [\n \"subsubsection\",\n \"Drawing Observations and Insights\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Drawing Observations and Insights\"\n },\n {\n \"cite_extract_rate\": 0.38095238095238004,\n \"cites\": [\n 3857,\n 3853,\n 9099,\n 3854,\n 9100,\n 7754,\n 3874,\n 3863\n ],\n \"content\": \"\\\\label{sec:defense}\\n\\\\begin{table}[htbp]\\n\\t\\\\caption{Summary of AMD defenses (\\\\checkmark means applicable, \\\\protect\\\\fullcirc means 0, \\\\protect\\\\emptycirc means 1, \\\\protect\\\\dotcirc means a value in $[0,1]$)}\\n\\t\\\\resizebox{1.\\\\columnwidth}{!}{\\n\\t\\\\begin{threeparttable}\\n\\t\\\\centering\\n\\t\\\\setlength{\\\\tabcolsep}{0.4em}\\n\\t\\\\begin{tabular}{l|c|ccccc|cccc|ccccc|cccc|ccc}\\n\\t\\\\hline\\n\\t\\\\multicolumn{1}{c|}{\\\\specialcell{Defense \\\\\\\\ (in chronological order)}}& \\n\\t\\\\multicolumn{1}{c|}{\\\\specialcell{Defense \\\\\\\\ Objective}} &\\n\\t\\\\multicolumn{9}{c|}{\\\\specialcell{Defense Input}} & \\n\\t\\\\multicolumn{5}{c|}{\\\\specialcell{Assumptions}} & \\n\\t\\\\multicolumn{4}{c|}{\\\\specialcell{Achieved \\\\\\\\ Properties}} & \\n\\t\\\\multicolumn{3}{c}{\\\\specialcell{Malware\\\\\\\\Detector}}\\n\\t\\\\\\\\\\n\\t& \\\\vthead{malware detection} \\n\\t& \\\\vthead{$A_1$: Training set $D_{train}$}\\n\\t& \\\\vthead{$A_2$: Defense technique}\\n\\t& \\\\vthead{$A_3$: Feature set}\\n\\t& \\\\vthead{$A_4$: Learning algorithm}\\n\\t& \\\\vthead{$A_5$: Response}\\n\\t& \\\\vthead{$A_6$: Manipulation set}\\n\\t& \\\\vthead{$A_7$: Attack tactic}\\n\\t& \\\\vthead{$A_8$: Attack technique}\\n\\t& \\\\vthead{$A_9$: Adversarial example set}\\n\\t& \\\\vthead{${\\\\sf IID}$ assumption}\\n\\t& \\\\vthead{${\\\\sf Oracle}$ assumption}\\n\\t& \\\\vthead{${\\\\sf Measurability}$ assumption}\\n\\t& \\\\vthead{${\\\\sf Smoothness}$ assumption}\\n\\t& \\\\vthead{${\\\\sf Invertibility}$ assumption}\\n\\t& \\\\vthead{{\\\\sf RR}: Representation Robustness}\\n\\t& \\\\vthead{{\\\\sf CR}: Classification Robustness}\\n\\t& \\\\vthead{{\\\\sf DR}: Detection Robustness}\\n\\t& \\\\vthead{{\\\\sf TR}: Training Robustness}\\n\\t& \\\\vthead{Windows Program}\\n\\t& \\\\vthead{Android Package}\\n\\t& \\\\vthead{PDF}\\n\\t\\\\\\\\\\n\\t\\\\hline\\\\hline\\n \\\\rowcolor{lightgray!30}\\n Biggio et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf EL}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark& & & & &\\n \\t&\\\\checkmark&&&\\n \\t& &\\\\checkmark\\n \\t\\\\\\\\\\n Smutz and Stavrou &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf SE}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t&&\\\\checkmark&&\\n \\t& & \\\\checkmark \\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Zhang et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf RF}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&\\\\checkmark&\\\\checkmark&&\\n \\t\\\\checkmark&\\\\checkmark&&&\\n \\t& & \\\\checkmark\\n \\t\\\\\\\\\\n Demontis et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf WR}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark& &\\\\checkmark& & &\\n \\t&\\\\checkmark&&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n \\t\\\\rowcolor{lightgray!30}\\n \\tWang et al. &\\n \\t\\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf IT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t&\\\\checkmark&&&\\n \\t\\\\checkmark & &\\n \\t\\\\\\\\\\n Grosse et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf WR}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t\\\\checkmark&\\\\checkmark&&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Grosse et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf AT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark& &\\\\checkmark&&&\\n \\t\\\\checkmark&\\\\checkmark&&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n Chen et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf AT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark& &\\\\checkmark&&&\\n \\t&\\\\checkmark&&&\\n \\t\\\\checkmark & &\\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Khasawneh et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf CD}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t&\\\\checkmark&&&\\n \\t\\\\checkmark & &\\n \\t\\\\\\\\ \\n \\tDang et al. &\\n \\t\\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf SE}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf LQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t&&\\\\checkmark&&\\n \\t& & \\\\checkmark \\n \\t\\\\\\\\ \\n \\\\rowcolor{lightgray!30}\\n Yang et al. &\\n \\\\checkmark&\\n \\t\\\\specialcell{$D^\\\\ast_{train}$} &\\n \\t${\\\\sf AT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pieg &\\n \\t\\\\checkmark&&&&&\\n \\t&&\\\\checkmark&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n Yang et al. &\\n \\\\checkmark&\\n \\t\\\\specialcell{$D_{train}$} &\\n \\t${\\\\sf SE}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t&&\\\\checkmark&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Chen et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf RF}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark& &\\\\checkmark&\\\\checkmark& &\\n \\t\\\\checkmark&\\\\checkmark&&&\\n \\t& \\\\checkmark & \\n \\t\\\\\\\\\\n Incer et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf VL}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark& & & & &\\n \\t\\\\checkmark&\\\\checkmark&\\\\checkmark& & \\n \\t\\\\checkmark& & \\n \\t\\\\\\\\ \\n \\t\\\\rowcolor{lightgray!30}\\n \\tChen et al. ~ &\\n \\t\\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf SE}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t&&\\\\checkmark&&&\\n \\t&&&\\\\checkmark&\\n \\t&\\\\checkmark &\\n \\t\\\\\\\\\\n Al-Dujaili et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf AT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&& &\\n \\t&&\\\\checkmark&&\\n \\t\\\\checkmark & &\\n \\t\\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Chen et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf IT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t&\\\\checkmark&&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n Jordan et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf RF}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t&&&&&\\n \\t\\\\checkmark&\\\\checkmark&\\\\checkmark&&\\n \\t& & \\\\checkmark\\n \\t\\\\\\\\ \\n \\\\rowcolor{lightgray!30}\\n Li et al. &\\n \\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf AT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{4} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&\\\\checkmark&\\\\checkmark& &\\n \\t\\\\checkmark&\\\\checkmark&&&\\n \\t\\\\checkmark & &\\n \\t\\\\\\\\\\n \\tTong et al. &\\n \\t\\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf RF}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&&&\\n \\t\\\\checkmark&\\\\checkmark&\\\\checkmark&&\\n \\t& & \\\\checkmark\\n \\t\\\\\\\\ \\n \\t\\\\rowcolor{lightgray!30}\\n \\tLi and Li &\\n \\t\\\\checkmark&\\n \\t$D_{train}$ &\\n \\t${\\\\sf AT}$ &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&&& &\\n \\t&&\\\\checkmark&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n Chen et al. &\\n \\\\checkmark&\\n $D_{train}$ &\\n ${\\\\sf VL}$ &\\n $S$ &\\n $F_\\\\theta$ &\\n {\\\\sf FQ} &\\n \\\\pie{0} & \\n \\\\pie{0} & \\n \\\\pie{4} &\\n \\\\pie{4} &\\n \\\\checkmark& &\\\\checkmark & \\\\checkmark& &\\n \\\\checkmark&\\\\checkmark& & & \\n & & \\\\checkmark\\n \\\\\\\\\\n \\\\rowcolor{lightgray!30}\\n Li et al. & \\n \\t\\\\checkmark &\\n \\t$D_{train}$ &\\n \\t\\\\small{{\\\\sf EL}+{\\\\sf AT}+{\\\\sf RF}} &\\n \\t$S$ &\\n \\t$F_\\\\theta$ &\\n \\t{\\\\sf FQ} &\\n \\t\\\\pie{0} & \\n \\t\\\\pie{0} & \\n \\t\\\\pie{4} &\\n \\t\\\\pie{4} &\\n \\t\\\\checkmark&&\\\\checkmark&& &\\n \\t&&\\\\checkmark&&\\n \\t& \\\\checkmark &\\n \\t\\\\\\\\\\n \\\\hline\\n\\t\\\\end{tabular}\\n\\t\\\\begin{tablenotes}\\n\\t\\t\\\\small\\n\\t\\t\\\\item $D^\\\\ast_{train}$ contains $D_{train}$ and a portion of $\\\\mathcal{A}$'s adversarial examples.\\n\\t\\\\end{tablenotes}\\n \\\\end{threeparttable}\\n\\t}\\n\\t\\\\label{tbl:defenses}\\n\\\\end{table}\",\n \"id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ]\n ],\n \"subsections\": [\n \"6f9fa41a-b356-42e4-b332-196f10febba0\",\n \"1e7bc569-0fe6-4f78-b486-391ed3ea051e\",\n \"272e13c4-c17a-4f6c-ab41-e2e83ef3c588\",\n \"89604438-0b71-470a-93c0-468dd4ea65af\",\n \"48b02090-502b-4798-92fb-c646a315077a\",\n \"a5b08aa1-d4f9-4077-b1e1-6b8c15efc187\",\n \"45225fbe-29bd-4000-8b60-553581208d2f\",\n \"362ab0d6-2e4b-4607-b141-4a89fdd71ebf\",\n \"7446c022-f8d7-4a15-911d-d2f735cce556\"\n ],\n \"title\": \"Systematizing Defense Literature\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 937,\n 7754\n ],\n \"content\": \")} \\\\label{sec:defense-el}\\nBiggio et al. propose a one-and-a-half-class SVM classifier against evasion attacks, by leveraging an interesting observation (i.e., decision boundaries of one-class SVM classifiers are tighter than that of two-class SVM classifiers) to facilitate outlier detection. Specifically, the authors propose an ensemble of a two-class classifier and two one-class classifiers, and then combine them using another one-class classifier. \\nThe defense can enhance PDF malware detectors against gradient-based attacks , which can be characterized as $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OP2}, {\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$. However, the defense cannot thwart attacks incurring large perturbations. Independent of this study, other researchers propose using the random subspace and bagging techniques to enhance SVM-based malware detectors, dubbed Multiple Classifier System SVM (MCS-SVM), which leads to evenly distributed weights .\\nThese defenses work under the {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf EL},S,F_\\\\theta,{\\\\sf FQ}|0,1,0,0)$ and achieves the {\\\\sf CR} property.\",\n \"id\": \"6f9fa41a-b356-42e4-b332-196f10febba0\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Ensemble Learning ({\\\\sf EL\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Ensemble Learning ({\\\\sf EL\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 314,\n 7754\n ],\n \"content\": \")} \\\\label{sec:attack-wr}\\nDemontis et al. propose enhancing the Drebin malware detector $\\\\varphi(\\\\mathbf{x})=\\\\mathbf{w}^\\\\top\\\\mathbf{x} + b$ by using box-constraint weights. The inspiration is that the classifier's sensitivity to perturbations based on the $\\\\ell_1$ norm is bounded by the $\\\\ell_\\\\infty$ norm of the weights. This defense\\nhardens the Drebin detector against a mimicry attack with input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,0,1,0,0|\\\\mathbf{M},{\\\\sf BE},{\\\\sf MI},\\\\mathcal{X}_\\\\mathbf{M})$, obfuscation attack with input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,0,0,0,$ $0|\\\\mathcal{M},{\\\\sf BE},-, \\\\mathcal{Z}_\\\\mathcal{M})$, and the attack that modifies important features with input $(a_1,\\\\cdots,a_5|A_6,\\\\cdots,$ $A_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$, here `$-$' means inapplicable. Experimental results show this defense outperforms MSC-SVM . The defense works under the {\\\\sf IID}, {\\\\sf Oracle} and {\\\\sf Measurability} assumptions with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf WR},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$ and achieves {\\\\sf CR}.\\nGrosse et al. investigate how to apply two defense techniques known as distillation and retraining to enhance the DNN-based Drebin malware detector. The distillation technique can decrease a model's generalization error by leveraging a teacher to relabel the training data represented by real-value vectors (rather than one-hot encoding). It uses retraining to tune a learned model with respect to an augmented training set with adversarial examples. Both defenses are estimated against a variant of JSMA and can be characterized by their input as $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(0,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF}, \\\\mathcal{X}_\\\\mathbf{M})$. Experimental results show the two defenses achieve limited success. The defense based on the distillation technique works under the {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf WR},S,F_\\\\theta,{\\\\sf FQ}|0,1,0,0)$ and achieves the {\\\\sf RR} and {\\\\sf CR} properties. The defense based on the retraining technique works under the {\\\\sf IID} and {\\\\sf Measurability} assumptions with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ}|1,$ $1,1,0)$ and achieves the {\\\\sf RR} and {\\\\sf CR} properties.\",\n \"id\": \"1e7bc569-0fe6-4f78-b486-391ed3ea051e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Weight Regularization ({\\\\sf WR\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Weight Regularization ({\\\\sf WR\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 3875,\n 9100,\n 7754,\n 9099,\n 3854\n ],\n \"content\": \")}\\nChen et al. adapt a generic retraining framework proposed in the AML context to enhance linear malware detectors. The defense uses a label smoothness regularization technique to mitigate the side-effect of adversarial training . The defense is evaluated using Windows malware detectors against ``feature selection''-based evasion attacks, which can be characterized as $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$. The defense works under the {\\\\sf IID} and {\\\\sf Measurability} assumptions and can be characterized as $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ}|1,1,1,0)$, while assuring the {\\\\sf CR} property.\\nYang et al. propose a defense against genetic algorithm-based evasion attacks that can be characterized as $(A_1,\\\\cdots,A_5|a_6,$ $\\\\cdots,a_9)=(0,0,0,0,1|\\\\mathcal{M},{\\\\sf BE},{\\\\sf HS},\\\\mathcal{Z}_\\\\mathcal{M})$. The defense uses three techniques: adversarial training, sanitizing examples, and weight regularization . The adversarial training uses one half of $\\\\mathcal{A}$'s adversarial examples. The defense works under the {\\\\sf IID} assumption and can be characterized as\\n$(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D'_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ}|0,1,0,*)$, where $D'_{train}$ is the union of $D_{train}$ and a portion (e.g., one half) of $\\\\mathcal{A}$'s adversarial examples. The defense of sanitizing examples is learned from manipulations used by the attacker and works under the IID assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf SE},S,F_\\\\theta,{\\\\sf FQ}|1,1,1,0)$. Both defenses achieve the {\\\\sf DR} property. The defense of wight regularization is reviewed in Section \\\\ref{sec:attack-wr}.\\nAl-Dujaili et al.~ adapt the idea of minmax adversarial training (proposed in the AML context) to enhance DNN-based malware detectors. In this defense, the inner-layer optimization generates adversarial files by maximizing the classifier's loss function; the outer-layer optimization searches for the parameters $\\\\theta$ (of DNN $F_\\\\theta$) that minimize the classifier's loss with respect to the adversarial files.\\nThe defense enhances Windows malware detectors against attacks with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(0,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$. \\nExperimental results show that malware detectors that are hardened to resist one attack may not be able to defend against other attacks. By observing this phenomenon, researchers propose using a mixture of attacks to harden DNN-based malware detectors .\\nThe defense works under the {\\\\sf IID} assumption and can be characterized as $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$. The defense assures the {\\\\sf DR} property. \\nLi et al. propose a DNN-based attack-agnostic framework to enhance adversarial malware detectors. The key idea is dubbed adversarial regularization, which enhances malware detectors via the (approximately) optimally small perturbation. The framework wins the AICS'2019 adversarial malware classification challenge organized by MIT Lincoln Lab researcher , without knowing anything about the attack. \\nThe defense works under the {\\\\sf IID}, {\\\\sf Measurability}, and {\\\\sf Smoothness} assumptions with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ}|0,1,0,0)$ and assures the {\\\\sf RR} and {\\\\sf CR} properties. In the extended study , the authors further enhance the framework with 6 defense principles, including ensemble learning, adversarial training, and robust representation learning. The enhanced defense is validated with 20 attacks (including 11 grey-box attacks and 9 white-box attacks) against Android malware detectors. The enhanced defense works under the {\\\\sf IID} and {\\\\sf Measurability} assumptions with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf AT}+{\\\\sf EL}+{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$ and assures the {\\\\sf DR} property.\",\n \"id\": \"272e13c4-c17a-4f6c-ab41-e2e83ef3c588\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Adversarial Training ({\\\\sf AT\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Adversarial Training ({\\\\sf AT\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 3877,\n 3876\n ],\n \"content\": \")}\\nIncer et al. propose using monotonic malware classifiers to defend against evasion attacks, where {\\\\em monotonic} means\\n$\\\\varphi(\\\\mathbf{x}) \\\\leq \\\\varphi(\\\\mathbf{x}')$ when ${\\\\mathbf{x}} \\\\leq {\\\\mathbf{x}}'$ . Technically, this can be achieved by using (i) robust features that can only be removed or added but not both and (ii) monotonic classification function (e.g., linear models with non-negative weights). The resulting classifier can thwart any attack that perturbs feature values monotonically. The defense works under the {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf VL},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$ and assures the {\\\\sf RR}, {\\\\sf CR} and {\\\\sf DR} properties.\\nChen et al. propose a defense to enhance PDF malware detectors against evasion attacks, by leveraging the observation that manipulations on PDF files are subtree additions and/or removals. They also propose new metrics for quantifying such structural perturbations. This allows to adapt the \\n{\\\\em symbolic interval analysis} technique proposed in the AML context to enhance the PDF malware detectors. The defense can cope with attacks leveraging small perturbations in the training phase. This defense works under the {\\\\sf IID}, {\\\\sf Measurability}, and {\\\\sf Smoothness} assumptions with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf VL},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$ and achieves the {\\\\sf RR} and {\\\\sf CR} properties.\",\n \"id\": \"89604438-0b71-470a-93c0-468dd4ea65af\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Verifiable Learning ({\\\\sf VL\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Verifiable Learning ({\\\\sf VL\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 3874,\n 3863\n ],\n \"content\": \")}\\nZhang et al. propose leveraging optimal adversarial attacks for feature selection. The defense enhances PDF malware detectors against gradient-based attacks, which can be characterized as $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_{\\\\mathbf{M}})$. The defense works under the {\\\\sf IID}, {\\\\sf Measurability}, and {\\\\sf Smoothness} assumptions and can be characterized as $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$. The defense assures the {\\\\sf RR} and {\\\\sf CR} properties.\\nTong et al. propose refining features into invariant ones to defend against genetic algorithm-based evasion attacks with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(0,0,0,0,1|\\\\mathcal{M},{\\\\sf BE},{\\\\sf HS},\\\\mathcal{Z}_\\\\mathcal{M})$. Experimental results show that adversarial training can be further leveraged to enhance the robustness of the defense. The defense works under {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)$ $=(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$, and achieves {\\\\sf RR}, {\\\\sf CR} and {\\\\sf DR}.\\nChen et al. propose mitigating evasive attacks by filtering features according to their importance $|w_i|/c_i$ with respect to the linear function $\\\\varphi(\\\\mathbf{x})=\\\\mathbf{w}^\\\\top\\\\mathbf{x} + b$, where $x_i$, $w_i$ and $c_i$ denote respectively the $i$th component of $\\\\mathbf{x}$, $\\\\mathbf{w}$ and the constraint on manipulation cost $\\\\mathbf{c}$. The defense enhances Android malware detectors against three attacks: a random attack with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(0,0,1,0,0|\\\\mathbf{M},{\\\\sf BE},-,\\\\mathcal{X}_\\\\mathbf{M})$, a variant of the mimicry attack with input $(0,0,1,0,0|\\\\mathbf{M},{\\\\sf BE},{\\\\sf MI},$ $\\\\mathcal{X}_\\\\mathbf{M})$, and the attack that modifies important features with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf BE},{\\\\sf SF},$ $\\\\mathcal{X}_\\\\mathbf{M})$, where `$-$' means inapplicable. The defense works under the {\\\\sf IID}, {\\\\sf Measurability}, and {\\\\sf Smoothness} assumptions with input $(A_1,\\\\cdots,A_5|a_6,$ $\\\\cdots,a_9)=(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ}|1,1,0,0)$ and achieves {\\\\sf RR} and {\\\\sf CR}.\\nJordan et al. propose a robust PDF malware detector against evasion attacks by interpreting JavaScript behaviors using static analysis. A PDF file is classified as malicious when it calls a vulnerable API method or when it exhibits potentially malicious or unknown behaviors. The defense is validated against the {\\\\em reverse mimicry} attack with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(0,0,0,0,0|\\\\mathcal{M},{\\\\sf BE},{\\\\sf MI},\\\\mathcal{Z}_\\\\mathcal{M})$. The defense has input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ}|1,$ $1,0,0)$ and achieves {\\\\sf RR}, {\\\\sf CR} and {\\\\sf DR}.\",\n \"id\": \"48b02090-502b-4798-92fb-c646a315077a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Robust Features ({\\\\sf RF\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Robust Features ({\\\\sf RF\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 892,\n 3853\n ],\n \"content\": \")}\\nWang et al. propose the {\\\\em random feature nullification} to enhance DNN-based malware detectors against the attack of Fast Gradient Sign Method (FGSM) by nullifying (or dropping) features randomly in both training and testing phases. This offers a probabilistic assurance in preventing a white-box attacker from deriving adversarial files by using gradients of the loss function with respect to the input. \\nThe defense enhances Windows malware detectors against the FGSM attack with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(0,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$. The defense works under {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,$ $\\\\cdots,a_9)=(D_{train},{\\\\sf IT},S,F_\\\\theta,{\\\\sf FQ}|0,1,0,0)$ and achieves {\\\\sf CR}.\\nDroidEye defends Android malware detectors against evasion attacks by quantizing binary representations, namely transforming binary representations into real values and then using compression to reduce the effect of adversarial manipulations. The defense enhances linear malware detectors against a ``feature selection''-based attack with input\\n$(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(1,1,1,1,1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$ and the FGSM attack with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(0,1,1,1,$ $1|\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$ . The defense works under {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)$ $=(D_{train},{\\\\sf IT},S,F_\\\\theta,{\\\\sf FQ}|0,1,0,0)$ and achieves {\\\\sf CR}.\",\n \"id\": \"a5b08aa1-d4f9-4077-b1e1-6b8c15efc187\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Input Transformation ({\\\\sf IT\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Input Transformation ({\\\\sf IT\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \")}\\nKhasawneh et al. propose randomizing classifiers (i.e., using one randomly chosen from a pool of classifiers that use heterogeneous features) to defend against transfer attacks. The defense is validated against an attack which perturbs important features with input\\n$(a_1,\\\\dots,a_5|a_6,\\\\cdots,a_9)=(0,0,*,0,1|\\\\mathcal{M},{\\\\sf BE},{\\\\sf SF},\\\\mathcal{Z}_\\\\mathcal{M})$.\\nThe defense works under the {\\\\sf IID} assumption with input $(a_1,\\\\dots,a_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf CD},S,F_\\\\theta,$ ${\\\\sf FQ}|0,1,1,0)$ and achieves the {\\\\sf CR} property.\",\n \"id\": \"45225fbe-29bd-4000-8b60-553581208d2f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Classifier Randomization ({\\\\sf CD\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Classifier Randomization ({\\\\sf CD\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 3857\n ],\n \"content\": \")}\\nSmutz and Stavrou propose an ensemble classifier to defend against grey-box evasion attacks by returning classification results as benign, uncertain and malicious according to the voting result (e.g., $[0\\\\%,25\\\\%]$ classifiers saying malicious can be treated as benign, $[25\\\\%,75\\\\%]$ saying malicious can be treated as uncertain, and $[75\\\\%,100\\\\%]$ saying malicious can be treated as malicious). The defense enhances a PDF malware detector against three types of evasion attacks: gradient-based attack with input $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) =(1,0,*,*,0|\\\\mathcal{M},{\\\\sf OE2},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathbf{M})$, mimicry attack with input $(1,0,*,*,0|\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathbf{M})$, and reverse mimicry attack with input $(0,0,0,0,0|\\\\mathcal{M},{\\\\sf BE},{\\\\sf MI},\\\\mathcal{Z}_\\\\mathbf{M})$ .\\nThe defense works under the {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)$ $=(D_{train},{\\\\sf SE},S,F_\\\\theta,{\\\\sf FQ}|0,1,0,0)$ and achieves {\\\\sf DR}.\\nDang et al. propose enhancing PDF malware detectors by lowering the classification threshold $\\\\tau$ and restricting the maximum query times, rendering genetic algorithm-based evasion attacks harder to succeed. This defense works under the {\\\\sf IID} assumption with input $(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf SE},S,F_\\\\theta,{\\\\sf LQ}|0,1,0,0)$ and achieves {\\\\sf DR}.\\nChen et al. investigate defending Android malware detectors against poisoning attacks with input $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9) =(1,1,1,1,1|\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$\\nThe idea is to filter adversarial files that are distant from non-adversarial ones, where distance is measured by the Jaccard index, Jaccard-weight similarity and cosine similarity.\\nThe defense works under the {\\\\sf Measurability} assumption with input \\n$(A_1,\\\\cdots,A_5|a_6,\\\\cdots,a_9)=(D_{train},{\\\\sf SE},S,F_\\\\theta,$ ${\\\\sf FQ}|0,1,0,0)$\\nand achieves {\\\\sf TR}.\",\n \"id\": \"362ab0d6-2e4b-4607-b141-4a89fdd71ebf\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Defenses using Sanitizing Examples ({\\\\sf SE\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Defenses using Sanitizing Examples ({\\\\sf SE\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 966\n ],\n \"content\": \"Summarizing the preceding discussions, we draw the following observations.\\n(i) Most studies focus on black-box defenses (i.e., the defender knows little about the attacker), which is against the principle of ``knowing yourself and knowing your enemy\\\". (ii) Most studies focus on defenses against evasion attacks rather than poisoning attacks. (iii) There is no silver bullet defense against evasion attacks or poisoning attacks, at least for now. (iv) Sanitizing adversarial files as outliers is effective against black-box and grey-box attacks, but not white-box attacks. (v) The security properties achieved by defenses have been evaluated empirically rather than rigorously proven (despite that provable security is emerging on the small degree of perturbations; see for example ). (vi) There is no theoretical evidence to support that the effectiveness of defense tactics on the training set (e.g., adversarial training and verifiable learning) can generalize to other adversarial examples.\\nIn addition, we draw the following insights:\\n\\\\begin{insight}\\n(i) Effective defenses often require the defender to know the attacker's manipulation set. In the real world, it is hard to achieve this, explaining from one perspective why it is hard to design effective defenses.\\n(ii) The effectiveness of adversarial training depends on the defender's capability in identifying the most powerful attack.\\n\\\\end{insight}\\n\\\\ignore{\\nIn order to draw further insights, we train another {\\\\em random forest} model from the systematized structure data presented in Table \\\\ref{tbl:defenses}. We use attributes of defense objective, defense input, and assumptions as features and treat the achieved properties as the labels to learn. The data preprocessing is similar to insight learning from AMD attacks. This leads to a dataset of 22 defenses, each of which has 15 dimensions. \\n\\\\begin{figure}[!htbp]\\n\\t\\\\centering\\n\\t\\\\scalebox{0.25}{\\n\\t\\\\includegraphics{figures/defense-important-feat.eps}\\n\\t}\\n\\\\caption{Insights learning from systematized AMD defenses via attribute importance.\\n}\\n\\t\\\\label{fig:defense-attribute}\\n\\\\end{figure}\\nFigure \\\\ref{fig:defense-attribute} shows the most important features on AMD defenses, from which we make the following observations. (i) Defense technique is the most important feature in determining the empirically achieved security property. This can be explained because ${\\\\sf EL},{\\\\sf IT},{\\\\sf CD}, {\\\\sf VL}, {\\\\sf WR}$ enhance the classifier against small manipulations to achieve the {\\\\sf CR} property; {\\\\sf SE} usually achieves the {\\\\sf DR} property against large manipulations, which could be detected as outliers; ${\\\\sf AT},{\\\\sf RF}$ enhance classifiers against certain attacks specified by the defender. (ii) The {\\\\sf Measurability} assumption is the second important feature.\\n(iii) Manipulation set is the third important feature because knowing this attribute enables the defender to specify {\\\\sf RF} and {\\\\sf AT} techniques, which can enhance classifiers to achieve empirical {\\\\sf RR}, {\\\\sf CR} and {\\\\sf DR} properties.\\n\\\\begin{insight}[Insights automatically learned]\\nTechnique-wise, defense technique largely determines what security properties can be (empirically) achieved and knowing attacker's manipulation set is critical to defender's success.\\n\\\\end{insight}\\n}\",\n \"id\": \"7446c022-f8d7-4a15-911d-d2f735cce556\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"ecf055b9-207b-4f7f-9c11-8bb57d05480e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing Defense Literature\"\n ],\n [\n \"subsubsection\",\n \"Drawing Observations and Insights\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Drawing Observations and Insights\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 3867,\n 3857,\n 6172,\n 3864,\n 7756,\n 3862,\n 937,\n 3871,\n 3853,\n 9099,\n 3854,\n 7754,\n 3874,\n 3863,\n 3861,\n 7154\n ],\n \"content\": \"\\\\afterpage{\\n\\\\begin{rotatepage}\\n\\\\begin{figure*}[ht!]\\n\\\\centering\\n\\\\rotatebox{90}{\\n\\\\scalebox{0.8}{\\n\\\\begin{tikzpicture}\\n\\\\node(rbot1)[label=below:{}, label=left:{},align=center] at (20,3.6) {$ (*,0,1,*,0)$\\\\\\\\$ (1,0,1,0,0)$\\\\\\\\$ (1,0,*,*,0)$};\\n\\\\node(rbot1_attr)[label=below:{},align=center] at (20,2.6) {$(\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$ \\\\\\\\ $(\\\\mathcal{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(rbot1_tag)[label=below:{},align=center] at (20,2.1) {};\\n\\\\node(rbot)[label=below:{}, label=left:{}] at (20,1) {$ (1,1,1,1,1)$};\\n\\\\node(rbot_attr)[label=below:{},align=center] at (20,0.6) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf MI},\\\\mathcal{X})$};\\n\\\\node(rbot_tag)[label=below:{},align=center] at (20,0.2) {};\\n\\\\node(rbot0)[label=below:{}, label=left:{}] at (20,-2) {$ (1,1,1,1,1)$};\\n\\\\node(rbot0_attr)[label=below:{},align=center] at (20,-2.4) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(rbot0_refer)[label=below:{},align=center] at (20,-2.8) {};\\n\\\\node(lbot3)[label=below:{}, label=left:{}] at (0,3) {$ (0,0,0,0,1)$};\\n\\\\node(lbot3_attr)[label=below:{},align=center] at (0,2.6) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf HS},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(lbot3_refer)[label=below:{},align=center] at (0,2.2) {};\\n\\\\node(lbot2)[label=below:{}, label=left:{}] at (0,1) {$ (0,0,0,0,0)$};\\n\\\\node(lbot2_attr)[label=below:{},align=center] at (0,0.6) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf HS},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(lbot2_refer)[label=below:{},align=center] at (0,0.2) {};\\n\\\\node(lbot1)[label=below:{}, label=left:{}] at (0,-1) {$ (0,0,0,0,0)$};\\n\\\\node(lbot1_attr)[label=below:{},align=center] at (0,-1.4) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf MI},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(lbot1_refer)[label=below:{},align=center] at (0,-1.8) {};\\n\\\\node(lbot0)[label=below:{}, label=left:{}] at (0,-3) {$ (0,0,0,0,0)$};\\n\\\\node(lbot0_attr)[label=below:{},align=center] at (0,-3.4) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf MI},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(lbot0_refer)[label=below:{},align=center] at (0,-3.8) {};\\n\\\\node(above7_attr)[label=below:{},align=center] at (20,6) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(above7)[label=below:{}, label=left:{}] at (20,6.4) {$ (0,1,1,1,1)$};\\n\\\\node(above7_refer)[label=below:{},align=center] at (20,6.8) {};\\n\\\\node(above6_attr)[label=below:{},align=center] at (16.8,6) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf MS},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(above6)[label=below:{}, label=left:{}] at (16.8,6.4) {$ (0,1,1,1,1)$};\\n\\\\node(above6_refer)[label=below:{},align=center] at (16.8,6.8) {};\\n\\\\node(above5_attr)[label=below:{},align=center] at (14,6) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(above5)[label=below:{}, label=left:{}] at (14,6.4) {$ (0,1,1,1,1)$};\\n\\\\node(above5_refer)[label=below:{},align=center] at (14,6.8) {};\\n\\\\node(above4_attr)[label=below:{},align=center] at (11.2,6) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$}; \\n\\\\node(above4)[label=below:{}, label=left:{}] at (11.2,6.4) {$ (0,1,1,1,1)$};\\n\\\\node(above4_refer)[label=below:{},align=center] at (11.2,6.8) {};\\n\\\\node(above3_attr)[label=below:{},align=center] at (8.4,6) {$(\\\\mathcal{M},{\\\\sf OE2},{\\\\sf MI},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(above3)[label=below:{}, label=left:{}] at (8.4,6.4) {$ (1,1,1,1,1)$};\\n\\\\node(above3_refer)[label=below:{},align=center] at (8.4,6.8) {};\\n\\\\node(above2_attr)[label=below:{},align=center] at (5.6,6) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf HS},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(above2)[label=below:{}, label=left:{}] at (5.6,6.4) {$ (0,0,0,0,1)$};\\n\\\\node(above2_refer)[label=below:{},align=center] at (5.6,6.8) {};\\n\\\\node(above1_attr)[label=below:{},align=center] at (2.8,6) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(above1)[label=below:{}, label=left:{}] at (2.8,6.4) {$ (0,0,1,*,0)$};\\n\\\\node(above1_refer)[label=below:{},align=center] at (2.8,6.8) {};\\n\\\\node(above0_attr)[label=below:{},align=center] at (0,5.8) {$(\\\\mathbf{M},{\\\\sf OE1},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$ \\\\\\\\ $(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$ };\\n\\\\node(above0)[label=below:{}, label=left:{}] at (0,6.4) {$ (0,0,1,0,0)$};\\n\\\\node(above0_refer)[label=below:{},align=center] at (0,6.8) {};\\n\\\\node(below7)[label=below:{}, label=left:{}] at (20,-6.4) {$ (1,0,1,0,0)$};\\n\\\\node(below7_attr)[label=below:{},align=center] at (20,-6) {$(\\\\mathbf{M},{\\\\sf OP},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(below7_refer)[label=below:{},align=center] at (20,-6.8) {};\\n\\\\node(below6)[label=below:{}, label=left:{}] at (16.8,-6) {$ (0,0,0,1,1)$};\\n\\\\node(below6_attr)[label=below:{},align=center] at (16.8,-6.4) {$(\\\\mathbf{M},{\\\\sf BP},{\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(below6_tag)[label=below:{},align=center] at (16.8,-6.8) {};\\n\\\\node(below5)[label=below:{}, label=left:{}] at (14,-6) {$ (0,0,1,0,1)$};\\n\\\\node(below5_attr)[label=below:{},align=center] at (14,-6.4) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(below5_refer)[label=below:{},align=center] at (14,-6.8) {};\\n\\\\node(below4)[label=below:{}, label=left:{}] at (11.2,-6) {$ (0,0,1,0,1)$};\\n\\\\node(below4_attr)[label=below:{},align=center] at (11.2,-6.4) {$(\\\\mathbf{M},{\\\\sf BE},{\\\\sf GM},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(below4_refer)[label=below:{},align=center] at (11.2,-6.8) {};\\n\\\\node(below3)[label=below:{}, label=left:{}] at (8.4,-6) {$ (0,0,1,*,0)$};\\n\\\\node(below3_attr)[label=below:{},align=center] at (8.4,-6.4) {$(\\\\mathbf{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{X}_\\\\mathbf{M})$};\\n\\\\node(below3_refer)[label=below:{},align=center] at (8.4,-6.8) {};\\n\\\\node(below2)[label=below:{}, label=left:{}] at (5.6,-6) {$ (0,0,*,0,1)$};\\n\\\\node(below2_attr)[label=below:{},align=center] at (5.6,-6.4) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathbf{M})$};\\n\\\\node(below2_refer)[label=below:{},align=center] at (5.6,-6.8) {};\\n\\\\node(below1)[label=below:{}, label=left:{}] at (2.8,-6) {$ (0,0,0,0,1)$};\\n\\\\node(below1_attr)[label=below:{},align=center] at (2.8,-6.4) {$(\\\\mathcal{M},{\\\\sf OE2},{\\\\sf GO},\\\\mathcal{Z}_\\\\mathcal{M})$};\\n\\\\node(below1_refer)[label=below:{},align=center] at (2.8,-6.8) {};\\n\\\\node(below0)[label=below:{}, label=left:{}] at (0,-6) {$ (0,0,*,0,0)$};\\n\\\\node(below0_attr)[label=below:{},align=center] at (0,-6.4) {$(\\\\mathcal{M},{\\\\sf BE},{\\\\sf TR},\\\\mathcal{Z}_\\\\mathbf{M})$};\\n\\\\node(below0_refer)[label=below:{},align=center] at (0,-6.8) {};\\n\\\\fill [lightgray!30] (1.5,5) rectangle (18.6,-5);\\n\\\\node(mid4)[label=below:{},align=left] at (15.5,0) {:$(D_{train},{\\\\sf SE},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ})$\\\\\\\\\\n:$(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ})$\\n};\\n\\\\node(mid4-attr)[align=center] at (15.5,-0.85) {$(1,1,1,0)$};\\n\\\\node(mid3)[label=below:{$(1,1,0,0)$},align=left] at (10,1.88) {\\n:$(D_{train},{\\\\sf WR},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf RF},S,F_\\\\theta,{\\\\sf FQ})$\\n}; \\n\\\\node(mid2)[label=below:{$(0,1,1,0)$},align=left] at (10,-1.5) {:$(D_{train},{\\\\sf CD},S,F_\\\\theta,{\\\\sf FQ})$};\\n\\\\node(mid1)[label=below:{$(0,1,0,*)$},align=center] at (10,-3) {:$(D^\\\\ast_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ})$};\\n\\\\node(mid0)[label=below:{$(0,1,0,0)$},align=left] at (4.5,0) {\\n:$(D_{train},{\\\\sf SE},S,F_\\\\theta,{\\\\sf LQ})$\\\\\\\\\\n:$(D_{train},{\\\\sf AT},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf IT},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf WR},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf SE},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf SE},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\\\n:$(D_{train},{\\\\sf IT},S,F_\\\\theta,{\\\\sf FQ})$ \\\\\\\\ \\n:$(D_{train},{\\\\sf EL},S,F_\\\\theta,{\\\\sf FQ})$\\n};\\n\\\\node(det3)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (18.2,3) {};\\n\\\\node(det2)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (18.1,4.65) {};\\n\\\\node(dnn1)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (15.2, 3.7) {DNN-based malware \\\\\\\\ detector };\\n\\\\node(dnn1_label)[align=center,color=blue!50] at (15.2,2.8) {$(D_{train},\\\\emptyset,S,F_\\\\theta,{\\\\sf FQ})$\\\\\\\\$(0,0,0,0)$};\\n\\\\node(apk1)[label=right:{},align=center, fill=blue!20,fill opacity=0.8] at (6, 3.9) {Drebin };\\n\\\\node(apk1_label) [align=center,color=blue!50] at (6.2,3.2) {$(D_{train},\\\\emptyset,S,F_\\\\theta,{\\\\sf FQ})$\\\\\\\\$(0,0,0,0)$};\\n\\\\node(det1)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (3.05,3.9) {};\\n\\\\node(pdf2)[label=below:{},align=center, fill=blue!20,fill opacity=0.8] at (3.6, 3.2) {PDFrate };\\n\\\\node(pdf2_label)[align=center,color=blue!50] at (3.6,2.4) {$(D_{train},\\\\emptyset,S,F_\\\\theta,{\\\\sf FQ})$\\\\\\\\$(0,0,0,0)$};\\n\\\\node(pdf1)[label=below:{},label=below:{},align=left, fill=blue!20,fill opacity=0.8] at (3.6, -3.8) {PDFrate };\\n\\\\node(pdf1_label)[align=center,color=blue!50] at (3.6,-4.6) {$(D_{train},\\\\emptyset,S,F_\\\\theta,{\\\\sf FQ})$\\\\\\\\$(0,0,0,0)$};\\n{\\n\\\\draw [-latex,line width=1] (3.3,-5.7) -- (3.3, -5) -- (2.3, -5) -- (2.3, -5.7);\\n\\\\draw [-latex,line width=1] (11.7,-5.7) -- (11.7, -5) -- (10.7, -5) -- (10.7, -5.7);\\n\\\\draw [-latex,line width=1] (14.5,-5.7) -- (14.5, -5) -- (13.5, -5) -- (13.5, -5.7);\\n\\\\draw [-latex,line width=1] (20.5,-5.7) -- (20.5,-5) -- (19.5,-5) -- (19.5,-5.7);\\n\\\\draw [-latex,line width=1] (19.5,1.2) -- (19.5,1.4) -- (18.6,1.4) -- (18.6,0) -- (19.5,0) -- (19.5,0.4); \\n\\\\draw [-latex,line width=1] (above0_attr) -- (0, 3.9) -- (det1);\\n\\\\draw [-latex,line width=1] (20,5.7) -- (20,4.65) -- (det2);\\n\\\\draw [-latex,line width=1] (19,3.) -- (det3);\\n\\\\draw [-latex,line width=1] (6.1,-5.7) -- (6.1, -5) -- (5.1, -5) -- (5.1, -5.7);\\n\\\\draw [-latex,line width=1] (17.3,-5.7) -- (17.3, -5) -- (16.3, -5) -- (16.3, -5.7);\\n\\\\ignore{\\n\\\\draw [-latex,line width=1] (6.4,-0.2)node[red,right]{}--(5.8,-0.2);\\n\\\\draw [-latex,line width=1] (6.4,0.2)node[red,right]{}--(5.9,0.2);\\n\\\\draw [-latex,line width=1] (9.2,0.7)node[red,above]{}--(9.2,0.2);\\n\\\\draw [-latex,line width=1] (6.4,-3.2)node[red,left]{}--(6.9,-3.2);\\n\\\\draw [-latex,line width=1] (13.6,-3.2)node[red,left]{}--(14.1,-3.2);\\n}\\n\\\\draw [-latex,line width=1] (below3) -- (8.4,-5.0) -- (7.25,-5.0) -- (7.25,-1.5) -- (6.25,-1.5);\\n\\\\draw [-latex,line width=1] (7.25,-1.5) -- (7.25,0.2) -- (8.1,0.2);\\n\\\\draw [line width=1] (18.7,-2.4) -- (18.6,-2.4) -- (18.6,-1.05) -- (7.35,-1.05); \\n\\\\draw [-latex,line width=1] (7.15,-1.05) --(6.25,-1.05);\\n\\\\draw [-latex,line width=1] (12.55, -1.05) -- (12.55, 0.65) -- (11.75, 0.65);\\n\\\\tkzDefPoint(7.25,-1.05){c1}\\n\\\\tkzDefPoint(7.15,-1.05){c2}\\n\\\\tkzDrawArc[rotate,color=black,line width=1](c1,c2)(180);\\n\\\\draw [-latex,line width=1] (11.1,5.7) -- (11.1,4.3) -- (13.15,4.3) -- (13.15,3.7) -- (13.6, 3.7); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (16.8,3.7) -- (17.5,3.7) -- (17.5,0.1) -- (17.2,0.1); \\n\\\\draw [red,dashed,-latex,line width=1] (11.3,5.7) -- (11.3,5.0) -- (13.9,5.0) -- (13.9,5.7);\\n\\\\draw [-latex,line width=1] (14.1,5.7) -- (14.1,4.3) -- (17.7,4.3) -- (17.7,-0.1) -- (17.2,-0.1); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (13.55,0) -- (12.75,0) -- (12.75,1.8) -- (11.6,1.8);\\n\\\\draw [red,dashed,-latex,line width=1] (14.3,5.7) -- (14.3,5.0) -- (16.7,5.0) -- (16.7,5.7);\\n\\\\draw [line width=1] (17.3,5.7) -- (17.3,4.4);\\n\\\\draw [-latex,line width=1] (17.3,4.2) -- (17.3,2) -- (11.6,2);\\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (8.05, 1.9) -- (7.25,1.9) -- (7.25,1.45) -- (8.1,1.45) ;\\n\\\\tkzDefPoint(17.3,4.3){c3}\\n\\\\tkzDefPoint(17.3,4.2){c4}\\n\\\\tkzDrawArc[rotate,color=black,line width=1](c3,c4)(180);\\n\\\\draw [-latex,line width=1] (2.7,5.7) -- (2.7,4.3) -- (4.7,4.3) -- (4.7,3.9) -- (5.2,3.9); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (6.9,3.9) -- (7.05,3.9) -- (7.05,3.58) -- (8.1,3.58); \\n\\\\draw [red,dashed,-latex,line width=1] (2.9,5.7) -- (2.9,5.0) -- (8.5,5.0) --(8.5,5.7); \\n\\\\draw [red,dashed,-latex,line width=1] (5.7,5.0) -- (5.7,5.7);\\n\\\\draw [-latex,line width=1] (5.5,5.7) -- (5.5,4.3) -- (8.3,4.3) -- (8.3,3.7); \\n\\\\draw [line width=1] (8.3,5.7) -- (8.3,4.3); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (11.9,3.7) -- (12.95,3.7) -- (12.95,0.43) -- (13.55,0.43);\\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (11.9,3.5) -- (12.75,3.5) -- (12.75,2.7) -- (11.8,2.7);\\n\\\\draw [-latex,line width=1] (0.1, 3.1) -- (0.1,3.3) -- (2.5,3.3); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (2.4,3.1) -- (1.7,3.1) -- (1.7, 1.55) -- (2.5,1.55); \\n\\\\draw [red,dashed,-latex,line width=1] (-0.1,2.1) -- (-0.1,1.1);\\n\\\\draw [-latex,line width=1] (0.1,1.1) -- (0.1, 1.35) -- (2.5,1.35); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (4.7,3.2) -- (8.1,3.2);\\n\\\\draw [-latex,line width=1] (0.1,-5.7) -- (0.1,-5) -- (1.5,-5) -- (1.5,-5) -- (1.5,-3.9) -- (2.5,-3.9); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (2.4,-3.7) -- (1.7,-3.7) -- (1.7, -0.75) -- (2.5, -0.75); \\n\\\\draw [red,dashed,-latex,line width=1] (-0.1,-5.7) -- (-0.1,-3.9); \\n\\\\draw [-latex,line width=1](0.1,-2.9) -- (0.1,-2) -- (1.5,-2) -- (1.5,-0.55) -- (2.5,-0.55); \\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (6.4,-0.6) -- (7.05,-0.6) -- (7.05,2.3) -- (8.1,2.3); \\n\\\\draw [red,dashed,-latex,line width=1] (-0.1,-2.9) -- (-0.1,-2); \\n\\\\draw [-latex,line width=1](0.1,-0.9) -- (0.1,-0.55) -- (2.5,-0.55);\\n}\\n\\\\draw [-latex,line width=1] (12.3,7.45) -- (12.9,7.45) node[right,align=center]{Waging \\\\\\\\ Attack};\\n\\\\draw [blue,densely dashdotted,-latex,line width=1] (14.5,7.45) -- (15.1,7.45);\\n\\\\node [right,align=center] at (15.1,7.45) {Defense \\\\\\\\ Escalation};\\n\\\\draw [red,dashed,-latex,line width=1] (17,7.45) -- (17.6,7.45);\\n\\\\node [right,align=center] at (17.6, 7.45) {Attack \\\\\\\\ Escalation};\\n\\\\draw (11,7) -- (20,7) -- (20,7.9) -- (11,7.9) -- node [right]{Legend:} (11,7);\\n\\\\end{tikzpicture}\\n}\\n}\\n\\\\caption{Arms race in AMD attack and defense escalations.}\\n\\\\label{fig:attacks-defenses}\\n\\\\end{figure*}\\n\\\\end{rotatepage}\\n}\\nFigure \\\\ref{fig:attacks-defenses} displays AMD attack-defense arms race surrounding three malware detectors: PDFrate, Drebin, and DNN-based detector. For a better visual effect, we group papers that proposed defense methods in terms of a common input $(a_6,\\\\ldots,a_9)$. For example, we group\\n, and together because the defenders in both papers have input $(a_6,\\\\ldots,a_9)=(1,1,1,0)$, while noting that their input on $(A_1,\\\\ldots,A_5)$ may or may not be different.\\nWe also simplify attack and defense inputs while preserving the {\\\\em critical information} when an attack (defense) works for multiple inputs. \\nFor example, $(a_1,\\\\ldots,a_5|A_6,\\\\cdots,A_9)=(0,0,1,0,0|\\\\mathbf{M},{\\\\sf OE2}, {\\\\sf SF}, \\\\mathcal{X}_\\\\mathbf{M})$ is the {\\\\em critical information} for attack input \\n$(a_1,\\\\cdots,a_5|A_6,\\\\cdots,A_9)=(0,0,1,0,0|\\\\mathbf{M},{\\\\sf OE2}, {\\\\sf SF}, \\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (0,0,1,0,1|\\\\mathbf{M},{\\\\sf OE2}, {\\\\sf SF}, \\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (1,0,1,0,0|\\\\mathbf{M},{\\\\sf OE2}, {\\\\sf SF}, \\\\mathcal{X}_\\\\mathbf{M}) \\\\lor (1,0,1,0,1|\\\\mathbf{M},{\\\\sf OE2}, {\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$\\nbecause it is the weakest attack input in the partial order formulated by these $(a_1,\\\\ldots,a_5)$'s. This suggests us to focus on attack input $(0,0,1,0,0|\\\\mathbf{M},{\\\\sf OE2}, {\\\\sf SF},\\\\mathcal{X}_\\\\mathbf{M})$ because it is already able to break some defense and automatically implies that a stronger input can achieve the same (while noting some special cases, see discussion in Section \\\\ref{sec:defense-el}). Multiple defense inputs are simplified in the same manner.\\n{\\\\em Arms race in PDF malware detection}: We summarize two sequences of escalations caused by PDFrate . In one sequence, PDFrate is defeated by transfer attacks, which are realized by gradient-based and mimicry methods against surrogate models . These attacks trigger the defense escalation to an ensemble detector built on top of some diversified classifiers . This defense triggers attack escalation to {\\\\em reverse mimicry attacks} , \\nwhich trigger the defense escalation of using robust hand-crafted features . This defense represents the state-of-the-art PDF malware detector, but still incurs a high false-positive rate. In the other sequence of arms race, PDFrate is defeated by genetic algorithm-based attacks . These attacks trigger the defense escalation to and . The former defense restricts the responses to attacker queries, but can be defeated by the escalated attack that leverages the hill-climbing algorithm (also shown in ). The latter defense uses invariant features to thwart the attacks and represents another state-of-the-art PDF malware detectors.\\n{\\\\em Arms race in android malware detection}: Drebin is defeated by the attack that modifies a limited number of important features , which also proposes the new defense to defeat the escalated attack. This defense triggers the attack escalation to, and is defeated by, the genetic algorithm-based attack and the mimicry-alike attack . The former attack triggers the escalated defense (also presented in ) that leverages attack mutations to detect adversarial examples . The latter attack injects objects in APKs and (in principle) can be defeated by the monotonic classifier . These escalated defenses represent the state-of-the-art Android malware detectors, but still incur a high false-positive rate.\\n{\\\\em Arms race in DNN-based malware detection}: The DNN-based detector triggers four gradient-based evasion attacks presented in , which also hardens the DNN malware detector by using an {\\\\em minmax adversarial training} instantiation to incorporates the $\\\\ell_\\\\infty$ normalized gradient-based attack. This escalated defense triggers the mixture of attacks presented in . The defense of {\\\\em minmax adversarial training} incorporating a mixture of attacks can defeat a broad range of attacks, but still suffers from the mimicry attack and other mixtures of attacks . \\nAs a consequence, there are no effective defenses can thwart all kinds of attacks.\\n{\\\\em Independent arms race}:\\nThere are studies that have yet to trigger cascading arms races, including: (i) Studies propose independent attacks and then show how to defeat these attacks.\\n(ii) Studies propose attacks to defeat naive malware detectors. (iii) Studies propose defenses to counter some attacks .\",\n \"id\": \"76e4c68c-dec2-4af2-b3fa-ce426944cf52\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 40,\n \"parent_id\": \"e26ef155-9b9c-4a92-9ffa-e334fbe6ea87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Systematizing AMD Arms Race\"\n ],\n [\n \"subsection\",\n \"Systematizing AMD Arms Race\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Systematizing AMD Arms Race\"\n },\n {\n \"cite_extract_rate\": 0.818181818181818,\n \"cites\": [\n 8399,\n 8696,\n 314,\n 892,\n 3878,\n 966,\n 937,\n 7754,\n 7312\n ],\n \"content\": \"\\\\label{sec:cd}\\n\\\\noindent{\\\\bf FRD 1}: {\\\\em Pinning down the root cause(s) of adversarial malware examples}.\\nSpeculations on root cause(s) include: (i) invalidity of the {\\\\sf IID} assumption because of distribution drifting, namely that testing files and training files are drawn from different distributions );\\n(ii) incompetent feature extraction ;\\n(iii) high dimensionality of malware representations ;\\n(iv) insufficient scale of training data ;\\n(v) low-probability ``pockets'' in data manifolds ;\\n(vi) linearity of DNNs ; and\\n(vii) large curvature of decision boundaries .Although these speculations may be true, more studies are needed in order to (in)validate them. \\n\\\\smallskip\\n\\\\noindent{\\\\bf FRD 2}: {\\\\em Characterizing the relationship between transferability and vulnerability}.\\nIn the AMD context, an attacker may use a surrogate model to generate adversarial examples and a defender may use a surrogate model for adversarial training. Transferability is related to the extent at which knowledge gained by a surrogate model may be the same as, or similar to, what is accommodated by a target model. The wide use of surrogate models in the AMD context suggests that there may be a fundamental connection between knowledge transferability and model vulnerability.\\n\\\\smallskip\\n\\\\noindent{\\\\bf FRD 3}: {\\\\em Investigating adversarial malware examples in the wild}. In the AMD context, it is challenging to generate practical adversarial malware examples to correspond to perturbations conducted in the feature space, owing to realistic constraints. On the other hand, an attacker can directly search for manipulations in the problem space. This may cause large perturbations, putting the value of studies on small perturbations in question. This represents a fundamental open problem that distinguishes the field of AMD from its counterparts in other application settings. \\nThis issue is largely unaddressed by assuming that there is an oracle for telling whether manipulated or perturbed features indeed correspond to a malware sample or not.\\n\\\\smallskip\\n\\\\noindent{\\\\bf FRD 4}: {\\\\em Quantifying the robustness and resilience of malware detectors}. Robustness and resilience of malware detectors against adversarial examples need to be quantified, ideally with a provable guarantee. For this purpose, one may adapt the reduction-based paradigm underlying the provable security of cryptographic primitives and protocols.\\n\\\\smallskip\\n\\\\noindent{\\\\bf FRD 5}: {\\\\em Designing malware detectors with provable robustness and resilience guarantees}.\\nHaving understood the root cause(s) of adversarial examples, characterized the effect of transferability, investigated the effectiveness of practical attacks, and designed metrics for quantifying the robustness and resilience of malware detectors, it is imperative to investigate robust malware detectors with provable robustness, ideally as rigorous as what has been achieved in the field of cryptography. In this regard, robust feature extraction, adversarial learning, and verifiable learning are promising candidates for making breakthroughs.\\n\\\\smallskip\\n\\\\noindent{\\\\bf FRD 6}: {\\\\em Forecasting the arms race in malware detection}. Arms race is a fundamental phenomenon inherent to the cybersecurity domain. In order to effectively defend against adversarial malware, one approach is to deploy proactive defense, which requires the capability to forecast the arms race between malware writers and defenders. For instance, it is important to predict how attacks will evolve and what kinds of information would be necessary in order to defeat such attacks.\",\n \"id\": \"965da080-653e-47f5-949b-de42288cf6a5\",\n \"level\": \"section\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"4fe25043-06f8-4577-bd62-3fe42c7ab899\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Future Research Directions (FDRs)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Future Research Directions (FDRs)\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7155\n ],\n \"content\": \"\\\\label{sec:conc}\\nWe have presented a framework for systematizing the field of AMD through the lens of assumptions, attacks, defenses and security properties. This paves the way for precisely relating attacks and defenses. We have also shown how to apply the framework to systematize the AMD literature, including the arms race between AMD attacks and defenses. We have reported a number of insights.\\nThe study leads to a set of future research directions. In addition to the ones described in Section \\\\ref{sec:cd}, we mention the following two, which are discussed here because there are rarely studies on these aspects. (i) To what extent explainability (or interpretability) of ML models can be leveraged to cope with adversarial malware examples? It is intuitive that explainability could be leveraged to recognize adversarial examples because they may not be explainable . \\n(ii) To what extent uncertainty quantification can be leveraged to cope with adversarial malware examples? If the uncertainty associated with detectors' predictions on adversarial malware examples are inherently and substantially higher than the uncertainty associated with non-adversarial malware examples, this fact can be leveraged to recognize adversarial malware examples.\\nFinally, we reiterate that the research community should seek to establish a solid foundation for AMD. Although this foundation can leverage ideas and techniques from AML, the unique characteristics of AMD warrant the need of a unique foundation.\\n\\\\bibliographystyle{ACM-Reference-Format}\\n\\\\bibliography{mal_attacks,mal_defenses,aml,cyber_security,data_mining,mal_detection}\\n\\\\end{document}\",\n \"id\": \"82494750-c0aa-44ab-add1-1d1d59481f4a\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"4fe25043-06f8-4577-bd62-3fe42c7ab899\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Arms Race in Adversarial Malware Detection: A Survey\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":94,"string":"94"},"all_cites":{"kind":"list like","value":[314,7755,7337,893,3856,3857,937,3853,3854,7754,7153,3855,3861,7154,3860,3859,3858,1159,8565,318,166,3862,7312,9099,8695,890,3866,7156,888,3868,1193,3867,3865,6172,3864,7155,9100,3863,3869,917,3870,7756,3871,892,923,3873,3872,894,243,3874,3875,3877,3876,966,8399,8696,3878],"string":"[\n 314,\n 7755,\n 7337,\n 893,\n 3856,\n 3857,\n 937,\n 3853,\n 3854,\n 7754,\n 7153,\n 3855,\n 3861,\n 7154,\n 3860,\n 3859,\n 3858,\n 1159,\n 8565,\n 318,\n 166,\n 3862,\n 7312,\n 9099,\n 8695,\n 890,\n 3866,\n 7156,\n 888,\n 3868,\n 1193,\n 3867,\n 3865,\n 6172,\n 3864,\n 7155,\n 9100,\n 3863,\n 3869,\n 917,\n 3870,\n 7756,\n 3871,\n 892,\n 923,\n 3873,\n 3872,\n 894,\n 243,\n 3874,\n 3875,\n 3877,\n 3876,\n 966,\n 8399,\n 8696,\n 3878\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.9592913179769931,"string":"0.959291"}}},{"rowIdx":62,"cells":{"authors":{"kind":"list like","value":["Xiaoxiao Ma","Jia Wu","Shan Xue","Jian Yang","Chuan Zhou","Quan Z. Sheng","Hui Xiong","Leman Akoglu"],"string":"[\n \"Xiaoxiao Ma\",\n \"Jia Wu\",\n \"Shan Xue\",\n \"Jian Yang\",\n \"Chuan Zhou\",\n \"Quan Z. Sheng\",\n \"Hui Xiong\",\n \"Leman Akoglu\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-06-14T06:04:57Z","string":"2021-06-14T06:04:57Z"},"category":{"kind":"string","value":"cs.LG"},"abstract":{"kind":"string","value":" Anomalies are rare observations (\\eg data records or events) that deviate significantly from the others in the sample. Over the past few decades, research on anomaly mining has received increasing interests due to the implications of these occurrences in a wide range of disciplines - for instance, security, finance, and medicine. For this reason, anomaly detection, which aims to identify these rare observations, has become one of the most vital tasks in the world and has shown its power in preventing detrimental events, such as financial fraud, network intrusions, and social spam. The detection task is typically solved by identifying outlying data points in the feature space, which, inherently, overlooks the relational information in real-world data. At the same time, graphs have been prevalently used to represent the structural/relational information, which raises the \\textit{graph anomaly detection problem} - identifying anomalous graph objects (\\ie nodes, edges and sub-graphs) in a single graph, or anomalous graphs in a set/database of graphs. Conventional anomaly detection techniques cannot tackle this problem well because of the complexity of graph data (\\eg irregular structures, relational dependencies, node/edge types/attributes/directions/multiplicities/weights, large scale, etc.). However, thanks to the advent of deep learning in breaking these limitations, graph anomaly detection with deep learning has received a growing attention recently. In this survey, we aim to provide a systematic and comprehensive review of the contemporary deep learning techniques for graph anomaly detection. Specifically, we provide a taxonomy that follows a task-driven strategy and categorizes existing work according to the anomalous graph objects that they can detect. We especially focus on the challenges in this research area and discuss the key intuitions, technical details as well as relative strengths and weaknesses of various techniques in each category. From the survey results, we highlight 12 future research directions spanning unsolved and emerging problems introduced by graph data, anomaly detection, deep learning and real-world applications. Additionally, to provide a wealth of useful resources for future studies, we have compiled a set of open-source implementations, public datasets, and commonly-used evaluation metrics. With this survey, our goal is to create a ``one-stop-shop'' that provides a unified understanding of the problem categories and existing approaches, publicly available hands-on resources, and high-impact open challenges for graph anomaly detection using deep learning. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"5020eb09-4f12-4d1a-8dae-076499c8173a","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"]],"subsections":["ecbafaf3-1565-4746-bf9c-a9182b235789","8ea06e54-3d89-4768-967e-6b68c2d22f53","59eed45d-0f6f-4983-a1d0-48c2fb48a606","12f2fc41-29d6-4189-b62f-ff980f20a825","678bd7af-5bf2-44f6-ae32-df413829eb80","78c71045-f884-4e7f-b2a8-d7769a5bc16b","5da2d684-654a-4048-95ec-2ae406b319b2","9ba15010-a3db-4588-8e7f-56a7def289f5","ae3bbf88-f1cb-408b-8ded-ef8445178dfa","8147e459-f864-454f-86f6-52d0b0cd6b52","e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","32ddd0ab-f741-4a6c-a1ff-380ee935d4a2","9a4d2cc0-0208-4633-a227-f43cf17e07e4","12e1194f-a7ed-4e46-889d-a06a013eaeaf","dfc7ea98-0750-495a-b3c3-8183220b6042","a37030ed-a13e-4c9b-bdf9-32f760d9d73a","ef880176-a742-4e90-bb20-77c1c0e4b593","6dcc483a-2e1d-4f81-abe8-97ae152ee771"],"title":"root"},{"cite_extract_rate":0.24193548387096703,"cites":[6272,3207,1671,553,3344,6271,6277,6275,217,1662,6274,6276,6270,6273,6278],"content":"\\label{sec:introduction}}\n\\IEEEPARstart{A}{nomalies} were first defined by Grubbs in 1969~ as \\textit{``one that appears to deviate markedly from other members of the sample in which it occurs''} and the studies on anomaly detection were initiated by the statistics community in the 19th century.\nTo us, anomalies might appear as social spammers or misinformation in social media; fraudsters, bot users or sexual predators in social networks; network intruders or malware in computer networks and broken devices or malfunctioning blocks in industry systems, and they often introduce huge damage to the real-world systems they appear in.\nAccording to FBI's 2014 Internet Crime Report\\footnote{https://www.fbi.gov/file-repository/2014\\_ic3report.pdf/view}, the financial loss due to crime on social media reached more than \\$60 million in the second half of the year alone and a more up-to-date report\\footnote{https://www.zdnet.com/article/online-fake-news-costing-us-78-billion-globally-each-year/} indicates that the global economic cost of online fake news reached around \\$78 billion a year in 2020.\nIn computer science, the research on anomaly detection dates back to the 1980s, and detecting anomalies on graph data has been an important data mining paradigm since the beginning.\nHowever, the extensive presence of connections between real-world objects and advances in graph data mining in the last decade have revolutionized our understanding of the graph anomaly detection problems such that this research field has received a dramatic increase in interest over the past five years.\nOne of the most significant changes is that graph anomaly detection has evolved from relying heavily on human experts' domain knowledge into machine learning techniques that eliminate human intervention, and more recently, to various deep learning technologies.\nThese deep learning techniques are not only capable of identifying potential anomalies in graphs far more accurately than ever before, but they can also do so in real-time.\n\\begin{figure}[!t]\n\\setlength{\\belowcaptionskip}{-0.5cm}\n\\centering\n\\subfigure[Conventional Anomaly Detection]{\n\\label{Fig1a}\n\\includegraphics[scale=0.45]{pics/Introduction/Conventional_AD.pdf}}\n\\subfigure[Graph Anomaly Detection]{\n\\label{Fig1b}\n\\includegraphics[scale=0.45]{pics/Introduction/Toy_Example.pdf}} \n\\caption{Toy Examples of Conventional Anomaly Detection and Graph Anomaly Detection. Apart from anomalies shown in (b), graph anomaly detection also identifies graph-level anomalies, detailed in Sections~\\ref{sec:anosgd:db} and~\\ref{sec:anosgd:dynamic}.} \n\\label{Toy}\n\\end{figure}\nFor our purposes today, anomalies, which are also known as outliers, exceptions, peculiarities, rarities, novelties, etc., in different application fields, refer to abnormal objects that are significantly different from the standard, normal, or expected. \nAlthough these objects rarely occur in real-world, they contain critical information to support downstream applications. For example, the behaviors of fraudsters provide evidences for anti-fraud detection and abnormal network traffics reveal signals for network intrusion protection. Anomalies, in many cases, may also have real and adverse impacts, for instance, fake news in social media can create panic and chaos with misleading beliefs~, untrustworthy reviews in online review systems can affect customers' shopping choices~, network intrusions might leak private personal information to hackers~, and financial frauds can cause huge damage to economic systems~.\nAnomaly detection is the data mining process that aims to identify the unusual patterns that deviate from the majorities in a dataset~. In order to detect anomalies, conventional techniques typically represent real-world objects as feature vectors (\\eg news in social media are represented as bag-of-words~, and images in web pages are represented as color histograms~), and then detect outlying data points in the vector space~, as shown in Fig.~\\ref{Fig1a}. \nAlthough these techniques have shown power in locating deviating data points under tabulated data format, they inherently discard the complex relationships between objects~.\nYet, in reality, many objects have rich relationships with each other, which can provide valuable complementary information for anomaly detection.\nTake online social networks as an example, fake users can be created using valid information from normal users or they can camouflage themselves by mimicking benign users' attributes~. \nIn such situations, fake users and benign users would have near-identical features, and conventional anomaly detection techniques might not be able to identify them using feature information only. \nMeanwhile, fake users always build relationships with a large number of benign users to increase their reputation and influence so they can get unexpected benefits, whereas benign users rarely exhibit such activities~.\nHence, these dense and unexpected connections formed by fake users denote their deviations to the benigns and more comprehensive detection techniques should take these structural information into account to pinpoint the deviating patterns of anomalies.\nTo represent the structural information, \\textit{Graphs}, in which nodes/vertices denote real objects, and the edges denote their relationships, have been prevalently used in a range of application fields~, including social activities, e-commerce, biology, academia and communication. With the structural information contained in graphs, detecting anomalies in graphs raises a more complex anomaly detection problem in non-Euclidean space - graph anomaly detection (GAD) that aims to identify anomalous graph objects (\\ie nodes, edges or sub-graphs) in a single graph as well as anomalous graphs among a set/database of graphs~.\nAs a toy example shown in Fig.~\\ref{Fig1b}, given an online social network, graph anomaly detection aims to identify anomalous nodes (\\ie malicious users), anomalous edges (\\ie abnormal relations) and anomalous sub-graphs (\\ie malicious user groups).\nBut, because the copious types of graph anomalies cannot be directly represented in Euclidean feature space, it is not feasible to directly apply traditional anomaly detection techniques to graph anomaly detection, and researchers have intensified their efforts to GAD recently.\nAmongst earlier works in this area, the detection methods relied heavily on handcrafted feature engineering or statistical models built by domain experts~.\nThis inherently limits these techniques' capability to detect unknown anomalies, and the exercise tended to be very labor-intensive. \nMany machine learning techniques, such as matrix factorization~ and SVM~, have also been applied to detect graph anomalies.\nHowever, real-world networks often contain millions of nodes and edges that result in extremely high dimensional and large-scale data, and these techniques do not easily scale up to such data efficiently.\nPractically, they exhibit high computational overhead in both the storage and execution time~.\nThese general challenges associated with graph data are significant for the detection techniques, and we categorize them as data-specific challenges (Data-CHs) in this survey.\nA summary of them is provided in Appendix~\\ref{appendix:data-challenges}.\n\\begin{table*}[!h]\n\\centering \n\\footnotesize\n\\renewcommand\\arraystretch{1}\n\\setlength{\\tabcolsep}{2.8mm}{\n\\caption{A Comparison Between Existing Surveys on Anomaly Detection. We mark edge, sub-graph and graph detection as \\includegraphics[scale=0.2]{pics/Introduction/4.pdf} in our survey because we review more deep learning based works than any previous surveys.} \n\\resizebox{0.96\\textwidth}{!}{\n\\begin{tabular}{l|c|c|c|c|c|c|c|c|c|c}\n\\toprule[1 pt]\n\\multirow{2}{*}{\\textbf{Surveys}} & \\multirow{2}{*}{\\textbf{AD}} & \\multirow{2}{*}{\\textbf{DAD}} & \\multirow{2}{*}{\\textbf{GAD}} \n& \\multicolumn{4}{c|}{\\textbf{GADL}} & \\multirow{2}{*}{\\textbf{Source Code}} & \\multicolumn{2}{c}{\\textbf{Dataset}} \\\\ \\cline{5-8} \\cline{10-11}\n& & & &\\textbf{Node} &\\textbf{Edge} & \\textbf{Sub-graph} & \\textbf{Graph} & &\\textbf{Real-world} &\\textbf{Synthetic} \\\\ \n\\midrule[1 pt]\nOur Survey & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} \\\\\n\\midrule[1 pt]\nChandola \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - & - & - \\\\\nBoukerche \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & - & - & - & - & - & - & - & - \\\\\nBulusu \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - & - \\\\\nThudumu \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - \\\\\n\\midrule[1 pt]\nPang \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - \\\\\nChalapathy and Chawla~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - \\\\\n\\midrule[1 pt]\nAkoglu \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\nRanshous \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & -& - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - \\\\\nJennifer and Kumar~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\n\\midrule[1 pt]\nEltanbouly \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - \\\\\nFernandes \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\nKwon \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & - & - & - & - & - & - & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - \\\\\nGogoi \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - & - \\\\\n\\midrule[1 pt]\nSavage \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\nYu \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\nHunkelmann \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - \\\\\nPourhabibi \\etal~ & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & - & - & - & - & - & - \\\\\n\\bottomrule[1 pt]\n\\multicolumn{11}{l}{* AD: Anomaly Detection, DAD: Anomaly Detection with Deep Learning, GAD: Graph Anomaly Detection.} \\\\\n\\multicolumn{11}{l}{* GADL: Graph Anomaly Detection with Deep Learning.}\\\\\n\\multicolumn{11}{l}{* -: not included, \\includegraphics[scale=0.2]{pics/Introduction/1.pdf} (1-2 references included), \\includegraphics[scale=0.2]{pics/Introduction/2.pdf} (3-10 references included), \\includegraphics[scale=0.2]{pics/Introduction/4.pdf} (10+ references included).}\n\\end{tabular}\n\\label{table:comparison}}}\n\\end{table*}\nNon-deep learning based techniques also lack the capability to capture the non-linear properties of real objects~.\nHence, the representations of objects learned by them are not expressive enough to fully support graph anomaly detection.\nTo tackle these problems, more recent studies seek the potential of adopting deep learning techniques to identify anomalous graph objects. \nAs a powerful tool for data mining, deep learning has achieved great success in data representation and pattern recognition~.\nIts deep architecture with layers of parameters and transformations appear to suit the aforementioned problems well. The more recent studies, such as deep graph representation learning and graph neural networks (GNNs), further enrich the capability of deep learning for graph data mining~. \nBy extracting expressive representations such that graph anomalies and normal objects can be easily separated, or the deviating patterns of anomalies can be learned directly through deep learning techniques, graph anomaly detection with deep learning (GADL) is starting to take the lead in the forefront of anomaly detection.\nAs a frontier technology, graph anomaly detection with deep learning, hence, is expected to generate more fruitful results on detecting anomalies and secure a more convenient life for the society.","id":"ecbafaf3-1565-4746-bf9c-a9182b235789","level":"section","origin_cites_number":62,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Introduction"]],"subsections":["2ea37eb2-3734-4b08-ad7a-6c1226af2e9b","b87f5810-e203-4606-8bb5-00623ec5335b","7291e51c-caa8-4bd7-9c08-2d393eaf83dc"],"title":"Introduction"},{"cite_extract_rate":0.33333333333333304,"cites":[6279,4539,6280],"content":"\\label{sec:introdution:challenges}\nDue to the complexity of anomaly detection and graph data mining~, in addition to the prior mentioned data-specific challenges, adopting deep learning techniques for graph anomaly detection also faces a number of challenges from the technical side. \nThese challenges associated with deep learning are categorized as technique-specific challenges (Tech-CHs), and they are summarized as follows.\n\\textbf{Tech-CH1. Anomaly-aware training objectives.} Deep learning models rely heavily on the training objectives to fine-tune all the trainable parameters. \nFor graph anomaly detection, this necessitates appropriate training objectives or loss functions such that the GADL models can effectively capture the differences between benign and anomalous objects.\nDesigning anomaly-aware objectives is very challenging because there is no prior knowledge about the ground-truth anomalies as well as their deviating patterns versus the majority.\nHow to effectively separate anomalies from normal objects through training remains critical for deep learning-based models.\n\\textbf{Tech-CH2. Anomaly interpretability.} In real-world scenarios, the interpretability of detected anomalies is also vital because we need to provide convincing evidence to support the subsequent anomaly handling process. For example, the risk management department of a financial organization must provide lawful evidence before blocking the accounts of identified anomalous users. \nAs deep learning has been limited for its interpretability~, how to justify the detected graph anomalies remains a big challenge for deep learning techniques.\n\\textbf{Tech-CH3. High training cost.} Although D(G)NNs are capable of digesting rich information (\\eg structural information and attributes) in graph data for anomaly detection, these GADL models are more complex than conventional deep neural networks or machine learning methods due to the anomaly-aware training objectives. Such complexity inherently leads to high training costs in both time and computing resources.\n\\textbf{Tech-CH4. Hyperparameter tuning.} D(G)NNs naturally exhibit a large set of hyperparameters, such as the number of neurons in each neural network layer, the learning rate, the weight decay and the number of training epochs. Their learning performance is significantly affected by the values of these hyperparameters. However, it remains a serious challenge to effectively select the optimal/sub-optimal settings for the detection models due to the lack of labeled data in real scenarios.\nBecause deep learning models are sensitive to their associated hyperparameters, setting well-performing values for the hyperparameters is vital to the success of a task.\nTuning hyperparameter is relatively trivial in supervised learning when labeled data are available. For instance, users can find an optimal/sub-optimal set of hyperparameters (\\eg through random search, grid search) by comparing the model's outputs with the ground-truth. However, unsupervised anomaly detection has no accessible labeled data to judge the model's performance under different hyperparameter settings~. Selecting the ideal hyperparameter values for unsupervised detection models persists as a critical obstacle to applying them in a wide range of real scenarios.","id":"2ea37eb2-3734-4b08-ad7a-6c1226af2e9b","level":"subsection","origin_cites_number":9,"parent_id":"ecbafaf3-1565-4746-bf9c-a9182b235789","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Introduction"],["subsection","Challenges in GAD with Deep Learning"]],"subsections":[],"title":"Challenges in GAD with Deep Learning"},{"cite_extract_rate":0.29411764705882304,"cites":[6275,6274,3207,3206,3344],"content":"\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n \\centerline{\\includegraphics[width=0.98\\textwidth]{pics/Introduction/Timeline.pdf}}\n \\caption{A Timeline of Graph Anomaly Detection and Reviewed Techniques.}\n \\label{pic:timeline}\n\\end{figure*} \nRecognizing the significance of anomaly detection, many review works have been conducted in the last ten years covering a range of anomaly detection topics: anomaly detection with deep learning, graph anomaly detection, graph anomaly detection with deep learning, and particular applications of graph anomaly detection such as social media, social networks, fraud detection and network security, etc.\nThere are some representative surveys on generalized anomaly detection techniques -~,~ and~.\nBut only the most up-to-date work in Thudumu \\etal~ covers the topic of graph anomaly detection.\nRecognizing the power of deep learning, the three contemporary surveys, Ruff \\etal~, Pang \\etal~ and Chalapathy and Chawla~ specifically review deep learning based anomaly detection techniques specifically.\nAs for graph anomaly detection, Akoglu \\etal~, Ranshous \\etal~, and Jennifer and Kumar~ put their concentration on graph anomaly detection, reviewing many conventional approaches in this area, including statistical models and machine learning techniques.\nOther surveys are dedicated to particular applications of graph anomaly detection, such as computer network intrusion detection and anomaly detection in online social networks, \\eg~, and~.\nThese works provided solid reviews of the application of anomaly detection/graph anomaly detection techniques in these high demand and vital domains. \nHowever, none of the mentioned surveys are dedicated to techniques on graph anomaly detection with deep learning, as shown in Table~\\ref{table:comparison}, and hence do not provide a systematic and comprehensive review of these techniques.","id":"b87f5810-e203-4606-8bb5-00623ec5335b","level":"subsection","origin_cites_number":17,"parent_id":"ecbafaf3-1565-4746-bf9c-a9182b235789","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Introduction"],["subsection","Existing Anomaly Detection Surveys"]],"subsections":[],"title":"Existing Anomaly Detection Surveys"},{"cite_extract_rate":0,"cites":[],"content":"Our contributions are summarized as follows:\n\\begin{itemize}\n \\item \\textbf{The first survey in graph anomaly detection with deep learning.} To the best of our knowledge, our survey is the first to review the state-of-the-art deep learning techniques for graph anomaly detection. Most of the relevant surveys focus either on conventional graph anomaly detection methods using non-deep learning techniques or on generalized anomaly detection techniques (for tabular/point data, time series, etc.). Until now, there has been no dedicated and comprehensive survey on graph anomaly detection with deep learning. Our work bridges this gap, and we expect that an organized and systematic survey will help push forward research in this area. \n \\item \\textbf{A systematic and comprehensive review.} In this survey, we review the most up-to-date deep learning techniques for graph anomaly detection published in influential international conferences and journals in the area of deep learning, data mining, web services, and artificial intelligence, including: TKDE, TKDD, TPAMI, NeurIPS, SIGKDD, ICDM, WSDM, SDM, SIGMOD, IJCAI, AAAI, ICDE, CIKM, ICML, WWW, CVPR, and others. We first summarize seven data-specific and four technique-specific challenges in graph anomaly detection with deep learning. We then comprehensively review existing works from the perspectives of: 1) the motivations behind the deep methods; 2) the main ideas for identifying graph anomalies; 3) a brief introduction to conventional non-deep learning techniques; and 4) the technical details of deep learning algorithms. A brief timeline of graph anomaly detection and reviewed works is given in Fig.~\\ref{pic:timeline}.\n \\item \\textbf{Future directions.} From the survey results, we highlight 12 future research directions covering emerging problems introduced by graph data, anomaly detection, deep learning models, and real-world applications. These future opportunities indicate challenges that have not been adequately tackled, and so more effort is needed in the future.\n \\item \\textbf{Affluent resources.} Our survey also provides an extensive collection of open-sourced anomaly detection algorithms, public datasets, synthetic dataset generating techniques, as well as commonly used evaluation metrics to push forward the state-of-the-art in graph anomaly detection. These published resources offer benchmark datasets and baselines for future research. \n \\item \\textbf{A new taxonomy.} We have organized this survey with regard to different types of anomalies (\\ie nodes, edges, sub-graphs, and graphs) existing in graphs or graph databases. We also pinpoint the differences and similarities between different types of graph anomalies.\n\\end{itemize}\nThe rest of this survey is organized as follows. In Section~\\ref{sec:preliminaries}, we provide preliminaries about the different types of settings. \nFrom Section~\\ref{sec:AND} to Section~\\ref{sec:anosgd:dynamic}, we review existing techniques for detecting anomalous nodes, edges, sub-graphs and graphs, respectively.\nIn Section~\\ref{sec:resources}, we first provide a collection of published graph anomaly detection algorithms and datasets and then summarize commonly used evaluation metrics and synthetic data generation strategies.\nWe highlight 12 future directions concerning deep learning in graph anomaly detection in Section~\\ref{sec:futures} and summarize our survey in Section~\\ref{sec:conclusion}.\nA concrete taxonomy of our survey is given in Appendix~\\ref{appendix:taxonomy}.\n\\vspace{-0.1cm}","id":"7291e51c-caa8-4bd7-9c08-2d393eaf83dc","level":"subsection","origin_cites_number":0,"parent_id":"ecbafaf3-1565-4746-bf9c-a9182b235789","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Introduction"],["subsection","Contributions"]],"subsections":[],"title":"Contributions"},{"cite_extract_rate":0.33333333333333304,"cites":[3344],"content":"\\label{sec:preliminaries}\nIn this section, we provide definitions of different types of graphs mostly used in node/edge/sub-graph-level anomaly detection (Section~\\ref{sec:AND} to Section~\\ref{sec:subgraph}).\nFor consistency, we have followed the conventional categorization of graphs as in existing works~ and categorize them as static graphs, dynamic graphs, and graph databases. Unless otherwise specified, all graphs mentioned in the following sections are static. Meanwhile, as graph-level anomaly detection is discussed far away on page 13, to enhance readability, the definition for the graph database is given closer to the material in Section~\\ref{sec:anosgd:db}.\n\\textbf{\\textit{Definition 1 (Plain Graph)}}. A static plain graph $G = \\{V, E\\}$ comprises a node set $V = \\{ v_{i} \\}_{1}^{n}$ and an edge set $E = \\{e_{i,j}\\}$ where $n$ is the number of nodes and $e_{i,j} = (v_i,v_j)$ denotes an edge between nodes $v_{i}$ and $v_{j}$. The adjacency matrix $A = [a_{i,j}]_{n\\times n}$ restores the graph structure, where $a_{i,j} = 1$ if node $v_{i}$ and $v_{j}$ is connected, otherwise $a_{i,j} = 0$.\n\\textbf{\\textit{Definition 2 (Attributed Graph)}}. A static attributed graph $G = \\{V, E, X\\}$ comprises a node set $V$, an edge set $E$ and an attribute set $X$. In an attributed graph, the graph structure follows the definition in Definition 1. The attribute matrix $X = [\\mathbf{x}_{i}]_{n\\times k}$ consists of nodes' attribute vectors, where $\\mathbf{x}_{i}$ is the attribute vector associated with node $v_{i}$ and $k$ is the vector's dimension. Hereafter, the terms attribute and feature are used interchangeably.\n\\textbf{\\textit{Definition 3 (Dynamic Graph)}}. A dynamic graph $G(t) = \\{V(t), E(t), X_{v}(t), X_{e}(t) \\}$ comprises nodes and edges changing overtime. $V(t)$ is the nodes set in the graph at a specific time step $t$, $E(t)$ is the corresponding edge set, $X_{v}(t)$ and $X_{e}(t)$ are the node attribute matrix and edge attribute matrix at time step $t$ in the graph if existed.\nIn reality, the nodes or edges might also be associated with numerical or categorical labels to indicate their classes (\\eg normal or abnormal). When label information is available/partially-available, supervised/semi-supervised detection models could be effectively trained.","id":"8ea06e54-3d89-4768-967e-6b68c2d22f53","level":"section","origin_cites_number":3,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Preliminaries"]],"subsections":[],"title":"Preliminaries"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:AND}\nAnomalous nodes are commonly recognized as individual nodes that are significantly different from others.\nIn real-world applications, these nodes often represent abnormal objects that appear individually, such as a single network intruder in computer networks, an independent fraudulent user in online social networks or a specific fake news on social media.\nIn this section, we specifically focus on anomalous node detection in static graphs.\nThe reviews on dynamic graphs can be found in Section~\\ref{sec:node:dg}. Table~\\ref{table:ANOSND} at the end of Section~\\ref{sec:node:dg} provides a summary of techniques reviewed for ANOS ND.\nWhen detecting anomalous nodes in static graphs, the differences between anomalies and regular nodes are mainly drawn from the graph structural information and nodes/edges' attributes~.\nGiven prior knowledge (\\ie community structure, attributes) about a static graph, anomalous nodes can be further categorized into the following three types:\n\\begin{itemize}\n \\item \\textbf{Global anomalies} only consider the node attributes. They are nodes that have attributes significantly different from all other nodes in the graph. \n \\item \\textbf{Structural anomalies} only consider the graph structural information. They are abnormal nodes that have different connection patterns (\\eg connecting different communities, forming dense links with others).\n \\item \\textbf{Community anomalies} consider both node attributes and graph structural information. They are defined as nodes that have different attribute values compared to other nodes in the same community.\n\\end{itemize}\nIn Fig.~\\ref{pic:node_classes}, node 14 is a global anomaly because its 4th feature value is 1 while all other nodes in the graph have the value of 0 for the corresponding feature. Nodes 5, 6, and 11 are identified as structural anomalies because they have links with other communities while other nodes in their community do not form cross-community links. Nodes 2 and 7 are community anomalies because their feature values are different from others in the communities they belong to. \n\\begin{figure}[!t]\n \\setlength{\\belowcaptionskip}{-0.5cm} \n \\centerline{\\includegraphics[scale=0.47]{pics/Node/node_classes_new.pdf}}\n \\caption{Three Types of Anomalous Nodes: Structural Anomalies, Community Anomalies and Global Anomalies.}\n \\label{pic:node_classes}\n\\end{figure}","id":"59eed45d-0f6f-4983-a1d0-48c2fb48a606","level":"section","origin_cites_number":4,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"]],"subsections":["5e100893-099a-4201-b53e-ba07115d4307","9b59f8e2-fd33-4bf3-bf31-9d82848223fa"],"title":"Anomalous node detection (ANOS ND)"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:node:spg}\nPlain graphs are dedicated to representing the structural information in real-world networks.\nTo detect anomalous nodes in plain graphs, the graph structure has been extensively exploited from various angles. \nHere, we first summarize the representative traditional non-deep learning approaches, followed by a more recent, advanced detection technique based on representation learning.\n\\begin{figure*}[!t]\n \\setlength{\\belowcaptionskip}{-0.25cm} \n \\centerline{\\includegraphics[width=0.98\\textwidth]{pics/Node/Traditional.pdf}}\n \\caption{ANOS ND on attributed graphs -- Deep NN based approaches. As an example, the autoencoder is used to capture the graph structure and node attributes. With a specially-designed anomaly aware loss function, anomaly scores will be assigned to every node, and the top-k nodes are anomalies (\\eg nodes 9, 1, and 2 at top-3).}\n \\label{pic:conventional_autoencoder_framework}\n\\end{figure*}","id":"5e100893-099a-4201-b53e-ba07115d4307","level":"subsection","origin_cites_number":0,"parent_id":"59eed45d-0f6f-4983-a1d0-48c2fb48a606","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Plain Graphs"]],"subsections":["658ca9f1-8bc4-490f-ac7d-c460db855b3b","699af85c-49af-4ca0-a394-878a65f78e4e","bc2ebce7-d1c1-4f0b-881b-6a0ace25800e"],"title":"ANOS ND on Plain Graphs"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:anosnd:tndp}\nPrior to the recent advances in deep learning and other state-of-the-art data mining technologies, traditional non-deep learning techniques have been widely used in many real-world networks to identify anomalous entities.\nA key idea behind these techniques was to transform the graph anomaly detection into a traditional anomaly detection problem, because the graph data with rich structure information can not be handled by the traditional detection techniques (for tabular data only) directly. \nTo bridge the gap, many approaches~ used the statistical features associated with each node, such as in/out degree, to detect anomalous nodes. \nFor instance, OddBall~ employs the statistical features (\\eg the number of 1-hop neighbors and edges, the total weight of edges) extracted from each node and its 1-hop neighbors to detect particular structural anomalies that: 1) form local structures in shape of near-cliques or stars; 2) have heavy links with neighbors such that the total weight is extremely large; or 3) have a single dominant heavy link with one of the neighbors. \nWith properly selected statistical features, anomalous nodes can be identified with respect to their deviating feature patterns.\nBut, in real scenarios, it is very hard to choose the most suitable features from a large number of candidates, and domain experts can always design new statistics, \\eg the maximum/minimum weight of edges.\nAs a result, these techniques often carry prohibitive cost for assessing the most significant features and do not effectively capture the structural information.","id":"658ca9f1-8bc4-490f-ac7d-c460db855b3b","level":"subsubsection","origin_cites_number":3,"parent_id":"5e100893-099a-4201-b53e-ba07115d4307","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Plain Graphs"],["subsubsection","Traditional Non-Deep Learning Techniques"]],"subsections":[],"title":"Traditional Non-Deep Learning Techniques"},{"cite_extract_rate":0.27272727272727204,"cites":[282,6281,218],"content":"\\label{sec:spg:nr}\nTo capture more valuable information from the graph structure for anomaly detection, network representation techniques have been widely exploited. \nTypically, these techniques encode the graph structure into an embedded vector space and identify anomalous nodes through further analysis. Hu \\etal~, for example, proposed an effective embedding method to detect structural anomalies that are connecting with many communities. \nIt first adopts a graph partitioning algorithm (\\eg METIS~) to group nodes into $d$ communities ($d$ is a user-specified number). \nThen, the method employs a specially designed embedding procedure to learn node embeddings that could capture the link information between each node and $d$ communities.\nDenoting the embedding for node $i$ as $Z_i = \\{z_{i}^{1},\\cdots,z_{i}^{d}\\}$, the procedure initializes each $z_{i}^{c} \\in Z_i$ with regard to the membership of node $i$ to community $c$ (if node $i$ belongs to the community, then $z_{i}^{c} = \\frac{1}{\\sqrt{2}}$; otherwise, 0.) and optimizes node embeddings such that directly linked nodes have similar embeddings and unconnected nodes are dissimilar.\nAfter generating the node embeddings, the link information between node $i$ and $d$ communities is quantified for further anomaly detection analysis.\nFor a given node $i$, such information is represented as:\n\\begin{equation} \\label{eq:hu:nbi}\n \\overline{NB(i)} = (y_{i}^{1},...,y_{i}^{d}) = \\mathop{\\sum}\\limits_{j\\in NB(i)}(1- \\|Z_{i} - Z_{j}\\|)\\cdot Z_{j},\n\\end{equation}\nwhere $NB(i)$ comprises node $i$'s neighbors.\nIf $i$ has many links with community $c$, then the value in the corresponding dimension $y_{i}^{c}$ will be large.\nIn the last step, Hu \\etal~ formulate a scoring function to assign anomalousness scores, calculated as:\n\\begin{equation} \\label{eq:hu:as}\n AScore(i) = \\mathop{\\sum}\\limits_{k=1}^{d}\\frac{y_{i}^{k}}{y_{i}^{*}}, y_{i}^{*}=\\max\\{y_{i}^{1},...,y_{i}^{d}\\}.\n\\end{equation}\nAs expected, structural anomalies receive higher scores as they connect to different communities.\nIndeed, given a predefined threshold, nodes with above-threshold scores are identified as anomalies.\nTo date, many plain network representation methods such as Deepwalk~, Node2Vec~ and LINE~ have shown their effectiveness in generating node representations and been used for anomaly detection performance validation~.\nBy pairing the conventional anomaly detection techniques such as density-based techniques~ and distance-based techniques~ with node embedding techniques, anomalous nodes can be identified with regard to their distinguishable locations (\\ie low-density areas or far away from the majorities) in the embedding space.","id":"699af85c-49af-4ca0-a394-878a65f78e4e","level":"subsubsection","origin_cites_number":11,"parent_id":"5e100893-099a-4201-b53e-ba07115d4307","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Plain Graphs"],["subsubsection","Network Representation Based Techniques"]],"subsections":[],"title":"Network Representation Based Techniques"},{"cite_extract_rate":1,"cites":[6282],"content":"The success of reinforcement learning (RL) in tackling real-world decision making problems has attracted substantial interests from the anomaly detection community. Detecting anomalous nodes can be naturally regarded as a problem of deciding which class a node belongs to - anomalous or benign. As a special scenario of the general selective harvesting task, the anomalous node detection problem can be approached by a recent work in~ that intuitively combines reinforcement learning and network embedding techniques for selective harvesting. The proposed model, NAC, is trained with labeled data without any human intervention. Specifically, it first selects a seed network consisting of partially observed nodes and edges. Then, starting from the seed network, NAC adopts reinforcement learning to learn a node selection plan such that anomalous nodes in the undiscovered area can be identified. This is achieved by rewarding selection plans that can choose labeled anomalies with higher gains. Through offline training, NAC will learn an optimal/suboptimal anomalous node selection strategy and discover potential anomalies in the undiscovered graph step by step.","id":"bc2ebce7-d1c1-4f0b-881b-6a0ace25800e","level":"subsubsection","origin_cites_number":1,"parent_id":"5e100893-099a-4201-b53e-ba07115d4307","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Plain Graphs"],["subsubsection","Reinforcement Learning Based Techniques"]],"subsections":[],"title":"Reinforcement Learning Based Techniques"},{"cite_extract_rate":1,"cites":[7007,180,242],"content":"\\label{lb:node:sag}\nIn addition to the structural information, real-world networks also contain rich attribute information affiliated with nodes~. These attributes provide complementary information about real objects and together with graph structure, more hidden anomalies that are non-trivial can now be detected. \n\\begin{figure*}[!t]\n \\setlength{\\belowcaptionskip}{-0.25cm} \n \\centerline{\\includegraphics[width=0.98\\textwidth]{pics/Node/GCN.pdf}}\n \\caption{ANOS ND on attributed graphs -- GCN based approaches. Node representations are generated through GCN layers. Anomalies are then detected according to their reconstruction loss (\\ding{172}) or embedding distribution in the embedding space (\\ding{173}).}\n \\label{pic:GCNbased_framework}\n\\end{figure*}\nFor clarity, we distinguish between deep neural networks and graph neural networks in this survey. \nWe review deep neural network (Deep NN) based techniques, GCN based techniques, and reinforcement learning based techniques for ANOS ND as follows. \nDue to page limitations, other existing works including traditional non-deep learning techniques, GAT~ based techniques, GAN based techniques, and network representation based techniques are surveyed in Appendix~\\ref{appendix:node:static}.","id":"9b59f8e2-fd33-4bf3-bf31-9d82848223fa","level":"subsection","origin_cites_number":3,"parent_id":"59eed45d-0f6f-4983-a1d0-48c2fb48a606","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Attributed Graphs"]],"subsections":["9201dd97-4956-4c75-840c-212d3ec5e5ae","73bf642c-6661-4550-b563-f83087e0c906","be74c2e6-73e6-42d6-9154-063b0fce612a"],"title":"ANOS ND on Attributed Graphs"},{"cite_extract_rate":0,"cites":[],"content":"The deep learning models such as autoencoder and deep neural networks provide solid basis for learning data representations. \nAdopting these models for more effective anomalous node detection have drawn substantial interest recently.\nFor example, Bandyopadhyay \\etal~ developed an unsupervised deep model, DONE, to detect global anomalies, structural anomalies and community anomalies in attributed graphs.\nSpecifically, this work measures three anomaly scores for each node that indicate the likelihood of the situations where 1) it has similar attributes with nodes in different communities ($o_i^a$); or 2) it connects with other communities ($o_i^s$); or 3) it belongs to one community structurally but the attributes follow the pattern of another community ($o_i^{com}$).\nIf a particular node exhibits any of these characteristics, then it is assigned a higher score and is anomalous.\nTo acquire these scores, DONE adopts two separate autoencoders (AE), \\ie a structure AE and an attribute AE, as shown in Fig.~\\ref{pic:conventional_autoencoder_framework}. \nBoth are trained by minimizing the reconstruction errors and preserving the homophily that assumes connected nodes have similar representations in the graph. \nWhen training the AEs, nodes exhibiting the predefined characteristics are hard to reconstruct and therefore introduce more reconstruction errors because their structure or attribute patterns do not conform to the standard behavior.\nHence, the adverse impact of anomalies should be alleviated to achieve the minimized error.\nAccordingly, DONE specially designs an anomaly-aware loss function with five terms: $\\mathcal{L}_{str}^{Recs}$, $\\mathcal{L}_{attr}^{Recs}$, $\\mathcal{L}_{str}^{Hom}$, $\\mathcal{L}_{attr}^{Hom}$, and $\\mathcal{L}^{Com}$.\n$\\mathcal{L}_{str}^{Resc}$ and $\\mathcal{L}_{attr}^{Resc}$ are the structure reconstruction error and attribute reconstruction error that can be written as: \n\\begin{equation} \\label{DONE:ls1}\n \\begin{split}\n \\mathcal{L}_{str}^{Recs} = \\frac{1}{N}\\sum_{i=1}^{N}\\log(\\frac{1}{o_{i}^{s}})\\|\\mathbf{t}_{i} - \\mathbf{\\hat{t}}_{i}\\|_{2}^{2},\n \\end{split}\n\\end{equation}\nand \n\\begin{equation} \\label{DONE:ls2}\n \\begin{split}\n \\mathcal{L}_{attr}^{Recs} \\frac{1}{N}\\sum_{i=1}^{N}\\log(\\frac{1}{o_{i}^{a}})\\|\\mathbf{x}_{i} - \\mathbf{\\hat{x}}_{i}\\|_{2}^{2},\n \\end{split}\n\\end{equation}\nwhere $N$ is the number of nodes, $\\mathbf{t}_{i}$ and $\\mathbf{x}_{i}$ store the structure information and attributes of node $i$, $\\mathbf{\\hat{t}}_{i}$ and $\\mathbf{\\hat{x}}_{i}$ are the reconstructed vectors. \n$\\mathcal{L}_{str}^{Hom}$ and $\\mathcal{L}_{attr}^{Hom}$ are proposed to maintain the homophily and they are formulated as:\n\\begin{equation} \\label{DONE:ls3}\n \\begin{split}\n \\mathcal{L}_{str}^{Hom} = \\frac{1}{N}\\sum_{i=1}^{N}\\log(\\frac{1}{o_{i}^{s}})\\frac{1}{|N(i)|}\\sum_{j\\in N(i)}||\\mathbf{h}_{i}^{s} - \\mathbf{h}_{j}^{s}||_{2}^{2},\n \\end{split}\n\\end{equation}\nand \n\\begin{equation} \\label{DONE:ls4}\n \\begin{split}\n \\mathcal{L}_{attr}^{Hom} = \\frac{1}{N}\\sum_{i=1}^{N}\\log(\\frac{1}{o_{i}^{a}})\\frac{1}{|N(i)|}\\sum_{j\\in N(i)}||\\mathbf{h}_{i}^{a} - \\mathbf{h}_{j}^{a}||_{2}^{2},\n \\end{split}\n\\end{equation}\nwhere $\\mathbf{h}_{i}^{s}$ and $\\mathbf{h}_{i}^{a}$ are the learned latent representations from the structure AE and attribute AE, respectively.\n$\\mathcal{L}^{Com}$ poses further restrictions on the generated representations for each node by the two AEs such that the graph structure and node attributes complement each other. It is formulated as:\n\\begin{equation} \\label{DONE:ls5}\n \\begin{split}\n \\mathcal{L}^{Com} = \\frac{1}{N}\\sum_{i=1}^{N}\\log(\\frac{1}{o_{i}^{com}})||\\mathbf{h}_{i}^{s} - \\mathbf{h}_{i}^{a}||_{2}^{2},\n \\end{split}\n\\end{equation}\nBy minimizing the sum of these loss functions, the anomaly scores of each node are quantified, and the top-k nodes with higher scores are identified as anomalies.","id":"9201dd97-4956-4c75-840c-212d3ec5e5ae","level":"subsubsection","origin_cites_number":1,"parent_id":"9b59f8e2-fd33-4bf3-bf31-9d82848223fa","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Attributed Graphs"],["subsubsection","Deep NN Based Techniques"]],"subsections":[],"title":"Deep NN Based Techniques"},{"cite_extract_rate":0.22222222222222202,"cites":[6283,6284],"content":"\\label{sec:AND:GCN}\nGraph convolutional neural networks (GCNs)~ have accomplished decent success in many graph data mining tasks (\\eg link prediction, node classification, and recommendation) owing to its capability of capturing comprehensive information in the graph structure and node attributes.\nTherefore, many anomalous node detection techniques start to investigate GCNs. \nFig.~\\ref{pic:GCNbased_framework} illustrates a general framework of existing works in this line.\nIn~, Ding \\etal measured an anomaly score for each node using the network reconstruction errors of both the structure and attribute.\nThe proposed method, DOMINANT, comprises three parts, namely, the graph convolutional encoder, the structure reconstruction decoder, and the attribute reconstruction decoder. The graph convolutional encoder generates node embeddings through multiple graph convolutional layers. The structure reconstruction decoder tends to reconstruct the network structure from the learned node embeddings, while the attribute reconstruction decoder reconstructs the node attribute matrix.\nThe whole neural network is trained to minimize the following loss function:\n\\begin{equation} \\label{eq:ad}\n \\begin{split}\n \\mathcal{L}_{DOMINANT} &= (1-\\alpha)\\mathcal{R}_{S} + \\alpha \\mathcal{R}_{A}\\\\\n &= (1-\\alpha)||A - \\hat{A}||_{F}^{2} + \\alpha||X - \\hat{X}||_{F}^{2},\n \\end{split}\n\\end{equation}\nwhere $\\alpha$ is the coefficient, $A$ depicts the adjacency matrix of the graph, $\\mathcal{R}_{S}$ and $\\mathcal{R}_{A}$ quantify reconstruction errors with regard to the graph structure and node attributes, respectively.\nWhen the training is finished, an anomaly score is then assigned to each node according to its contribution to the total reconstruction error, which is calculated by:\n\\begin{equation} \\label{eq:nodescore}\n \\textit{score}(i) \n = (1-\\alpha)||\\mathbf{a}_{i} - \\mathbf{\\hat{a}}_{i}||_{2} + \\alpha||\\mathbf{x}_{i} - \\mathbf{\\hat{x}}_{i}||_{2},\n\\end{equation}\nwhere $\\mathbf{a}_{i}$ and $\\mathbf{x}_{i}$ are the structure vector and attribute vector of node $i$, $\\mathbf{\\hat{a}}_{i}$ and $\\mathbf{\\hat{x}}_{i}$ are their corresponding reconstructed vectors. \nThe nodes are then ranked according to their anomaly scores in descending order, and the top-k nodes are recognized as anomalies.\nTo enhance the performance of anomalous node detection, later work by Peng \\etal~ further explores node attributes from multiple attributed views to detect anomalies.\nThe multiple attributed views are employed to describe different perspectives of the objects~.\nFor example, in online social networks, user's demographic information and posted contents are two different attributed views, and they characterize the personal information and social activities, respectively.\nThe underlying intuition of investigating different views is that anomalies might appear to be normal in one view but abnormal in another view.\nFor the purpose of capturing these signals, the proposed method, ALARM, applies multiple GCNs to encode information in different views and adopts a weighted aggregation of them to generate node representations.\nThis model's training strategy is similar to DOMINANT~ in that it aims to minimize the network reconstruction loss and attribute reconstruction loss and can be formulated as:\n\\begin{equation} \\label{eq:alarmloss}\n\\begin{split}\n\\mathcal{L}_{ALARM} = & \\sum_{i=1}^n \\sum_{j=1}^n - [\\gamma A_{ij}\\log\\hat{A}_{ij} + (1-A_{ij})\\log(1-\\hat{A}_{ij})] \\\\\n & + ||X - \\tilde{X}||^{F}_{2},\n\\end{split}\n\\end{equation}\nwhere $\\gamma$ is coefficient to balance the errors, $A_{ij}$ is the element at coordinate $(i,j)$ in the adjacency matrix $A$, $\\hat{A}_{ij}$ is the corresponding element in the reconstructed adjacency matrix $\\hat{A}$, $X$ is the original node feature matrix and $\\tilde{X}$ is the reconstructed node feature matrix.\nLastly, ALARM adopts the same scoring function as~, and nodes with top-k highest scores are anomalous.\nInstead of spotting unexpected nodes using their reconstruction errors, Li \\etal~ proposed SpecAE to detect global anomalies and community anomalies via a density estimation approach, Gaussian Mixture Model (GMM).\nGlobal anomalies can be identified by only considering the node attributes. \nFor community anomalies, the structure and attributes need to be jointly considered because of their distinctive attributes to the neighbors.\nAccordingly, SpecAE investigates a graph convolutional encoder to learn node representations and reconstruct the nodal attributes through a deconvolution decoder.\nThe parameters in the GMM are then estimated using the node representations. \nDue to the deviating attribute patterns of global and community anomalies, normal nodes are expected to exhibit greater energies in GMM, and the k nodes with the lowest probabilities are deemed to be anomalies.\nIn~, Wang \\etal developed a novel detection model that identify fraudsters using their relations and features. Their proposed method, Fdgars, first models online users' reviews and visited items as their features, and then identifies a small portion of significant fraudsters based on these features. In the last step, a GCN is trained in a semi-supervised manner by using the user-user network, user features, and labeled users. After training, the model can directly label unseen users.\nA more recent work, GraphRfi~, also explores the potential of combining anomaly detection with other downstream graph analysis tasks.\nIt targets on leveraging anomaly detection to identify malicious users and provide more accurate recommendations to service benign users by alleviating the impact of these untrustworthy users. \nSpecifically, a GCN framework is deployed to encode users and items into a shared embedding space for recommendation and users are classified as fraudsters or normal users through an additional neural random forest using their embeddings. \nFor rating prediction between users and items, the framework reduces the corresponding impact of suspicious users by assigning less weights to their training loss. \nAt the same time, the rating behavior of users also provides auxiliary information for fraudster detection. \nThe mutually beneficial relationship between these two applications (anomaly detection and recommendation) indicates the potential of information sharing among multiple graph learning tasks.","id":"73bf642c-6661-4550-b563-f83087e0c906","level":"subsubsection","origin_cites_number":9,"parent_id":"9b59f8e2-fd33-4bf3-bf31-9d82848223fa","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Attributed Graphs"],["subsubsection","GCN Based Techniques"]],"subsections":[],"title":"GCN Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"In contrast to NAC, Ding \\etal~ investigated to the use of reinforcement learning for anomalous node detection in attributed graphs.\nTheir proposed algorithm, GraphUCB, models both attribute information and structural information, and inherits the merits of the contextual multi-armed bandit technology~ to output potential anomalies.\nBy grouping nodes into $k$ clusters based on their features, GraphUCB forms a $k$-armed bandit model and measures the payoff of selecting a specific node as a potential anomaly for expert evaluation.\nWith experts' feedback on the predicted anomalies, the decision-making strategy is continuously optimized.\nEventually, the most potential anomalies can be selected. \n\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n\\centerline{\\includegraphics[width=0.98\\textwidth]{pics/Edge/UGED.pdf}}\n\\caption{ANOS ED on static graphs -- Deep NN based approaches. For example, the detection technique employs the autoencoder and fully connected network to learn the presence probability of each edge. Anomaly scores are assigned with regard to the probabilities, and the top-k edges are anomalous (\\eg $E_{8,9}$, $E_{1,3}$, and $E_{2,4}$ at top-3).}\n\\label{pic:UGED}\n\\end{figure*}","id":"be74c2e6-73e6-42d6-9154-063b0fce612a","level":"subsubsection","origin_cites_number":2,"parent_id":"9b59f8e2-fd33-4bf3-bf31-9d82848223fa","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous node detection (ANOS ND)"],["subsection","ANOS ND on Attributed Graphs"],["subsubsection","Reinforcement Learning Based Techniques"]],"subsections":[],"title":"Reinforcement Learning Based Techniques"},{"cite_extract_rate":0.38235294117647,"cites":[6285,6290,6288,6287,6283,6289,6282,3344,6271,6284,6291,8991,6286],"content":"\\label{sec:node:dg}\nReal-world networks can be modeled as dynamic graphs to represent evolving objects and the relationships among them.\nIn addition to structural information and node attributes, dynamic graphs also contain rich temporal signals~, \\eg the evolving patterns of the graph structure and node attributes. \nOn the one hand, these information inherently makes anomalous node detection on dynamic graphs more challenging. \nThis is because dynamic graphs usually introduce large volume of data and temporal signals should also be captured for anomaly detection.\nBut, on the other hand, they could provide more details about anomalies~. \nIn fact, some anomalies might appear to be normal in the graph snapshot at each time stamp, and, only when the changes in a graph's structure are considered, do they become noticeable.\nIn this section, we review the network representation based techniques and GAN based techniques as follows. Relevant techniques from traditional non-deep learning approaches are reviewed in Appendix~\\ref{appendix:node:dynamic}.\n\\begin{table*}[!t]\n\\centering\n\\footnotesize\n\\renewcommand\\arraystretch{1}\n\\setlength{\\tabcolsep}{2.8mm}\n\\caption{Summary of Anomalous Node Detection Techniques.}\n\\resizebox{0.96\\textwidth}{!}{\n\\begin{tabular}{m{2.1cm}<{\\centering}|m{2.3cm}<{\\centering}|m{1.1cm}<{\\centering}|m{4cm}<{\\centering}|m{1.8cm}<{\\centering}|m{2.9cm}<{\\centering}}\n\\toprule[1 pt]\n\\textbf{Graph Type} & \\textbf{Approach} & \\textbf{Category} & \\textbf{Objective Function} & \\textbf{Measurement} & \\textbf{Outputs} \\\\ \n\\midrule[1 pt]\n\\multirow{3}{*}{\\shortstack{Static Graph - \\\\ Plain}} & ~ & NR & $\\mathop{\\sum}\\limits_{(i,j)\\in E} \\|\\mathbf{Z_{i}} - \\mathbf{Z_{j}}\\|^{2} + \\alpha \\mathop{\\sum}\\limits_{(i,j)\\notin E}(\\|\\mathbf{Z_{i}} - \\mathbf{Z_{j}}\\| -1)^{2}$ & Anomaly Score & $ \\mathop{\\sum}\\limits_{k=1}^{d}\\frac{y_{i}^{k}}{y_{i}^{*}}$ \\\\ \\cline{2-6}\n& DCI~ & NR &$\\frac{1}{K}\\mathop{\\sum}\\limits_{k=1}^K\\mathcal{L}_{DCI}^{k}$ & Anomaly Prediction & Predicted Label \\\\\\cline{2-6}\n&NAC~ & RL & Cumulative reward & - & Anomalies \\\\ \n\\midrule[1 pt]\n\\multirow{24}{*}{\\shortstack{Static Graph - \\\\ Attributed}} &ALAD~ & Non-DP & $\\min\\limits_{W,H} \\|A-WW^{T}\\|_{F}^{2} + \\alpha \\|X-WH\\|_{F}^{2} + \\gamma (\\|W\\|_{F}^{2} + \\|H\\|_{F}^{2})$ & Anomaly Score & $\\frac{\\mathbf{W}_{n,c}}{\\sum_{c}\\mathbf{W}_{n,c}} cos(\\mathbf{A}_{n*}, \\mathbf{H}_{c*})$ \\\\ \\cline{2-6}\n&Radar~ & Non-DP & $\\min\\limits_{W,R}\\|X - W^{T} X - R\\|_{F}^{2}+ \\alpha \\|W\\|_{2,1} +\\beta \\|R\\|_{2,1} +\\gamma tr(R^{T}LR)$ & Residual Analysis & Residual Value \\\\ \\cline{2-6}\n&ANOMALOUS~ & Non-DP & $\\min\\limits_{W,\\tilde{R}}||X - XWX - \\tilde{R}||_{F}^{2} + \\alpha ||W||_{2,1} + \\beta ||W^{T}||_{2,1} + \\gamma ||\\tilde{R}^{T}||_{2,1} + \\varphi tr(\\tilde{R}L\\tilde{R}^{T})$ & Residual Analysis & Residual Value \\\\ \\cline{2-6}\n&SGASD~ & Non-DP & $\\min\\limits_{\\mathbf{w,c}}\\frac{1}{2}\\sum_{i=1}^{m}\\mathbf{c_{i}}(V_{i,*}\\mathbf{w} - y_i)^2 + \\frac{\\lambda_{1}}{2}||\\mathbf{w}||_{2}^{2} + \\lambda_{2} \\sum_{i=0}^{d}\\sum_{j=1}^{n_{i}}||\\mathbf{c}_{G_{j}^{i}}||_{2}$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n&DONE~ & DNN & $\\alpha_{1}\\mathcal{L}_{str}^{Recs} + \\alpha_{2}\\mathcal{L}_{attr}^{Recs} + \\alpha_{3}\\mathcal{L}_{str}^{Hom} + \\alpha_{4}\\mathcal{L}_{attr}^{Hom} + \\alpha_{5}\\mathcal{L}^{Com}$ & Anomaly Scores & $o_{i}^{s},o_{i}^{a},o_{i}^{com}$ \\\\ \\cline{2-6}\n&DOMINANT~ & GCN & $(1-\\alpha)\\mathcal{R}_{S} + \\alpha \\mathcal{R}_{A}$ & Anomaly Score & $(1-\\alpha)||\\mathbf{a}_{i} - \\mathbf{\\hat{a}}_{i}||_{2} + \\alpha||\\mathbf{x}_{i} - \\mathbf{\\hat{x}}_{i}||_{2}$ \\\\ \\cline{2-6}\n&ALARM~ & GCN & $\\sum_{i=1}^n \\sum_{j=1}^n - [\\gamma A_{ij}\\log\\hat{A}_{ij} + (1-A_{ij})\\log(1-\\hat{A}_{ij})] + \\mathcal{L}_a$ & Anomaly Score & $(1-\\alpha)||\\mathbf{a}_{i} - \\mathbf{\\hat{a}}_{i}||^2_{2} + \\alpha||\\mathbf{x}_{i} - \\mathbf{\\hat{x}}_{i}||^2_{2}$ \\\\ \\cline{2-6}\n&SpecAE~ & GCN & $\\mathbb{E}[dis(X,\\hat{X})] + \\mathbb{E}[dis(X,\\tilde{X})] + \\lambda_{1}\\mathbb{E}(E(Z)) + \\lambda_{2}KL$ & Density Estimation & Anomalousness Rank \\\\ \\cline{2-6}\n&Fdgars~ & GCN & $\\mathcal{L}_{GCN}$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n&GraphRfi~ & GCN & $\\mathcal{L}_{rating} + \\lambda \\mathcal{L}_{fraudster}$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n&ResGCN~ & GCN & $(1-\\alpha)||A - \\hat{A}||^2_{F} + \\alpha||X - \\hat{X} - \\lambda R||^2_{F}$ & Anomaly Score & $||R_{i,:}||_{2}$ \\\\ \\cline{2-6}\n&GraphUCB~ & RL & Expert Judgment & - & Anomalies \\\\ \\cline{2-6}\n&AnomalyDAE~ & GAT & $\\alpha ||(A - \\hat{A})\\odot \\bm{\\theta}||_{F}^{2} + (1-\\alpha)||(X - \\hat{X})\\odot \\bm{\\eta}||_{F}^{2}$ & Reconstruction Loss & Anomalousness Rank \\\\ \\cline{2-6}\n&SemiGNN~ & GAT & $\\alpha \\mathcal{L}_{sup} + (1- \\alpha) \\mathcal{L}_{unsup} + \\lambda \\mathcal{L}_{reg}$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n&AEGIS~ & GAN & $\\mathcal{L}_{AE} + \\mathcal{L}_{GAN}$ & Anomaly Score & $1 - D(\\mathbf{z}_i)$ \\\\ \\cline{2-6}\n&REMAD~ & NR & $\\mathcal{L}_{res} + \\beta\\| R^{T}\\|_{2,1}$ & Residual Analysis & Residual Value \\\\ \\cline{2-6}\n&CARE-GNN~ & NR & $\\mathcal{L}_{GNN} + \\lambda_{1}\\mathcal{L}_{Simi}^{(1)} + \\lambda_{2}\\mathcal{L}_{reg}$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n&SEANO~ & NR & $-\\sum_{i \\in V_{L}} \\log p(y_{i}|\\mathbf{x}_{i},\\Bar{\\mathbf{x}}_{N_{i}}) - \\sum_{i \\in V}\\sum_{v\\prime \\in C_i} \\log p(v\\prime|\\mathbf{x}_{i},\\Bar{\\mathbf{x}}_{N_{i}})$ & Anomaly Score & Discriminator's Output \\\\ \\cline{2-6}\n&OCGNN~ & NR & $\\frac{1}{\\beta K}\\sum\\limits_{v_{i}\\in\\mathbf{V}_{tr}}[||g(X,A;\\mathcal{W})_{v_i}-c||^2 -r^2]^{+}+ r^2 +\\frac{\\lambda}{2}\\sum\\limits_{l=1}^{L}||W^{(l)}||^2$ & Location in Embedding Space & Distance to Hypersphere Center \\\\ \\cline{2-6}\n&GAL~ & NR & $\\max\\{0,\\max\\limits_{y_{v^{\\prime}} \\neq y_{u}}g(u,v^\\prime) - \\min\\limits_{y_{v} = y_{u}}g(u,v) + \\Delta_{y_u}\\}$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n&CoLA~& NR & $-\\sum\\limits_{i=1}^{N}y_{i}\\log(CLM(v_i,\\mathcal{G}_{i})) + (1-y_i)\\log(1-CLM(v_i,\\mathcal{G}_{i}))$ & Anomaly Score & $\\frac{\\sum\\limits_{r=1}^{R}(s_{i,r}^{(-)}-s_{i,r}^{(+)})}{R}$ \\\\ \\cline{2-6}\n&COMMANDER~ & NR & $-\\mathcal{L}_{D} + \\mathcal{L}_{C} + \\mathcal{L}_{R}$ & Anomaly Score & $\\bar{y_i}||\\mathbf{\\tilde{x}_i} - \\mathbf{x}_i||_2^2$ \\\\ \\cline{2-6}\n&FRAUDRE~& NR & $\\sum\\limits_{i=1}^{n}f^{*}(y_i,\\mathbf{h}_{i}^{(final)}\\mathbf{W}_2)$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n&Meta-GDN~ & NR & $(1-y_i)\\cdot|dev(v_i)| + y_i\\cdot\\max(0,dev(v_i))$ & Anomaly Score & $\\mathbf{u}_{s}^{T}\\mathbf{o}_{i} + b_s$ \\\\ \n\\midrule[1 pt]\nDynamic Graph - Plain & NetWalk~ & DNN & $\\gamma\\mathcal{L}_{AE} + \\mathcal{L}_{Clique} + \\lambda\\|W\\|_{F}^{2} + \\beta KL$ & Anomaly Score & Nearest Distance to Cluster Centers \\\\ [2ex]\n\\midrule[1 pt]\n\\multirow{2}{*}{\\shortstack{Dynamic Graph - \\\\ Attributed}} &MTHL~ & Non-DP & $\\mathop{min}_{\\mathcal{P}}f(\\mathcal{P})$ & Anomaly Score & Distance to Hypersphere Centroid \\\\ \\cline{2-6}\n&OCAN~ & GAN & $\\mathcal{L}_{LSTM-AE} + \\mathcal{L}_{GAN}$ & Anomaly Score & Discriminator's Output \\\\\n\\bottomrule[1 pt]\n\\multicolumn{6}{l}{* Non-DP: Non-Deep Learning Techniques, DNN: Deep NN Based Techniques, GCN: GCN Based Techniques, RL: Reinforcement Learning Based Techniques.} \\\\\n\\multicolumn{6}{l}{* GAT: GAT Based Techniques, NR: Network Representation Based Techniques, GAN: Generative Adversarial Network Based Techniques.}\n\\end{tabular}\n}\n\\label{table:ANOSND}\n\\end{table*}\n\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm} \n\\centerline{\\includegraphics[width=0.98\\textwidth]{pics/Edge/AddGraph.pdf}}\n\\caption{ANOS ED on dynamic graphs -- GCN based approaches. GCN is employed to learn node embeddings from the temporal graph at each timestamp. The attention-based GRU generates the current hidden state using the node embeddings and previous hidden states. The edge scoring function, such as a FCN, is learned to assign anomaly scores, and the top-k edges are depicted as anomalies.}\n\\label{pic:Dynamic_Edge_Comparison}\n\\end{figure*}","id":"12f2fc41-29d6-4189-b62f-ff980f20a825","level":"section","origin_cites_number":34,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Dynamic Graphs"]],"subsections":["0af4432d-d88f-455d-b4ca-b342865e37f7","75faec9f-6952-423b-8944-196564d6aaa7"],"title":"ANOS ND on Dynamic Graphs"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:ND:NR}\nFollowing the research line of encoding a graph into an embedding space, after which anomaly detection is performed, dynamic network representation techniques have been investigated in the more recent works.\nSpecifically, in~, Yu \\etal presented a flexible deep representation technique, called NetWalk, for detecting anomalous nodes in dynamic (plain) graphs using only the structure information.\nIt adopts an autoencoder to learn node representations on the initial graph and incrementally updates them when new edges are added or existing edges are deleted.\nTo detect anomalies, NetWalk first executes the streaming $k$-means clustering algorithm~ to group existing nodes in the current time stamp into different clusters. \nThen, each node's anomaly score is measured with regard to its closest distance to the $k$ clusters. \nWhen the node representations are updated, the cluster centers and anomaly scores are recalculated accordingly.","id":"0af4432d-d88f-455d-b4ca-b342865e37f7","level":"subsection","origin_cites_number":2,"parent_id":"12f2fc41-29d6-4189-b62f-ff980f20a825","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Dynamic Graphs"],["subsection","Network Representation Based Techniques"]],"subsections":[],"title":"Network Representation Based Techniques"},{"cite_extract_rate":0.6000000000000001,"cites":[2626,6285,6292],"content":"\\label{sec:ND:GAN}\nIn practice, anomaly detection is facing great challenges from the shortage of ground-truth anomalies.\nConsequently, many research efforts have been invested in modeling the features of anomalies or regular objects such that anomalies can be identified effectively.\nAmong these techniques, generative adversarial networks (GAN)~ have received extensive attention because of its impressive performance in capturing real data distribution and generating simulated data.\nMotivated by the recent advances in ``bad\" GAN~, Zheng \\etal~ circumvented the fraudster detection problem using only the observed benign users' attributes.\nThe basic idea is to seize the normal activity patterns and detect anomalies that behave significantly differently.\nThe proposed method, OCAN, starts by extracting the benign users' content features using their historical social behaviors (\\eg historical posts, posts' URL), for which this method is classified into the dynamic category.\nA long short-term memory (LSTM) based autoencoder~ is employed to achieve this and as assumed, benign users and malicious users are in separate regions in the feature space.\nNext, a novel one-class adversarial net comprising a generator and a discriminator is trained.\nSpecifically, the generator produces complementary data points that locate in the relatively low density areas of benign users.\nThe discriminator, accordingly, aims to distinguish the generated samples from the benign users. \nAfter training, benign users' regions are learned by the discriminator and anomalies can hence be identified with regard to their locations.\nBoth NetWalk~ and OCAN~ approach the anomalous node detection problem promisingly, however, they respectively only consider the structure or attributes.\nBy the success of static graph anomaly detection techniques that analyze both aspects, when the structure and attribute information in dynamic graphs are jointly considered, an enhanced detection performance can be foreseen.\nWe therefore highlight this unexplored area for future works in Section~\\ref{sec:futures}.","id":"75faec9f-6952-423b-8944-196564d6aaa7","level":"subsection","origin_cites_number":5,"parent_id":"12f2fc41-29d6-4189-b62f-ff980f20a825","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Dynamic Graphs"],["subsection","GAN Based Techniques"]],"subsections":[],"title":"GAN Based Techniques"},{"cite_extract_rate":1,"cites":[6293],"content":"\\label{sec:edge:static}\nIn contrast to anomalous node detection, which targets individual nodes, ANOS ED aims to identify abnormal links.\nThese links often inform the unexpected or unusual relationships between real objects~, such as the abnormal interactions between fraudsters and benign users shown in Fig.~\\ref{Toy}, or suspicious interactions between attacker nodes and benign user machines in computer networks. Following the previous taxonomy, in this section, we review the state-of-the-art ANOS ED methods for static graphs, and Section~\\ref{sec:edge:dynamic} summarizes the techniques for dynamic graphs. A summary is provided in Table~\\ref{tb:edgeandsubgraph}. This section includes methods based on deep NNs, GCNs and network representations. The non-deep learning techniques are reviewed in Appendix~\\ref{appendix:edge}.","id":"678bd7af-5bf2-44f6-ae32-df413829eb80","level":"section","origin_cites_number":1,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous edge detection (ANOS ED) "]],"subsections":["509789f2-b5ef-4288-92b5-0d64443ed4af","e519d916-fe62-4029-8d13-0a7fc3954141","231d1246-99ce-4de5-9270-25f30b2ceb5d"],"title":"Anomalous edge detection (ANOS ED) "},{"cite_extract_rate":0,"cites":[],"content":"Similar to deep NN based ANOS ND techniques, autoencoder and fully connected network (FCN) have also been used for anomalous edge detection.\nAs an example, Ouyang \\etal~ approached the problem by modeling the distribution of edges through deep models to identify the existing edges that are least likely to appear as anomalies (as shown in Fig.~\\ref{pic:UGED}). The probability of each edge $e_{u,v}$ is decided by $P(v|u,N(u))$ and $P(u|v,N(v))$ which measure the edge probability using node $u$ with its neighbors $N(u)$ and node $v$ with its neighbors $N(v)$, respectively.\nTo calculate $P(v|u,N(u))$, the proposed method, UGED, first encodes each node into a lower-dimensional vector through a FCN layer and generates node $u$'s representation by a mean aggregation of itself and its neighbors' vectors.\nNext, the node representations are fed into another FCN to estimate $P(v|u,N(u))$.\nThe prediction is expressed as $\\hat{P}(v|u,N(u)) = \\text{Softmax}(W\\cdot H(u))|_v$, where $W$ represents the trainable parameters, and $H(u)$ is $u$'s representation.\nUGED's training scheme aims to maximize the prediction of existing edges via a cross-entropy-based loss function, $\\text{CE}(\\hat{P}(v|u,N(u)), v)$.\nAfter training, an anomaly score is assigned to each edge using the average of $1-P(v|u,N(u))$ and $1-P(u|v,N(v))$. \nAs such, existing edges that have a lower probability will get higher scores and the top-k edges are reported as anomalous.","id":"509789f2-b5ef-4288-92b5-0d64443ed4af","level":"subsection","origin_cites_number":1,"parent_id":"678bd7af-5bf2-44f6-ae32-df413829eb80","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous edge detection (ANOS ED) "],["subsection","Deep NN Based Techniques"]],"subsections":[],"title":"Deep NN Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"Following the line of modeling edge distributions, some studies leverage GCNs to better capture the graph structure information.\nDuan \\etal~ demonstrated that the existence of anomalous edges in the training data prevents traditional GCN based models from capturing real edge distributions, which leads to sub-optimal detection performance.\nThis inherently raises a problem: to achieve better detection performance, the node embedding process should alleviate the negative impact of anomalous edges, but these edges are detected using the learned embeddings.\nTo tackle this, the proposed method, AANE, jointly considers these two issues by iteratively updating the embeddings and detection results during training.\nIn each training iteration, AANE generates node embeddings $Z$ through GCN layers and learns an indicator matrix $I$ to spot potential anomalous edges.\nGiven an input graph $G$ with adjacency matrix $A$, each term $I_{uv}$ in $I$ is 1 if $\\hat{A}_{uv} < \\text{mean}_{v^{\\prime}\\in N_{u}} \\hat{A}_{uv^{\\prime}} - \\mu\\cdot\\text{std}_{v^{\\prime}\\in N_{u}} \\hat{A}_{uv^{\\prime}}$, and 0 otherwise.\nHere, $\\hat{A}_{uv}$ is the predicted link probability between nodes $u$ and $v$, which is calculated as the hyperbolic tangent of $u$ and $v$'s embeddings, and $\\mu$ is a predefined threshold.\nBy this, an edge $uv$ is identified as anomalous when its predicted probability is less than the average of all links associated with the node $u$ by a predefined threshold.\nThe total loss function of AANE contains two parts: an anomaly-aware loss ($\\mathcal{L}_{aal}$) and an adjusted fitting loss ($\\mathcal{L}_{afl}$). \n$\\mathcal{L}_{aal}$ is proposed to penalize the link prediction results and the indicator matrix $I$ such that anomalous edges will have lower prediction probabilities when they are marked as 1 in $I$. This is formulated as:\n\\begin{equation}\n \\mathcal{L}_{aal} = \\sqrt{\\sum_{u\\in V} \\sum_{v\\in N(u)}((1-\\hat{A}^2_{uv})(1-I_{uv}) + \\hat{A}^2_{uv}I_{uv})},\n\\end{equation}\nwhere $V$ is the node set, $N(u)$ is the set of $u$'s neighbors.\n$\\mathcal{L}_{afl}$ quantifies the reconstruction loss with regard to the removal of potential anomalous edges, denoted as:\n\\begin{equation}\n \\mathcal{L}_{afl} = \\| B - \\hat{A}\\|^{2}_{2},\n\\end{equation}\nwhere $B$ is an adjusted adjacency matrix that removes all predicted anomalies from the input adjacency matrix $A$.\nBy minimizing these two losses, AANE identifies the top-k edges with lowest probabilities as anomalies.","id":"e519d916-fe62-4029-8d13-0a7fc3954141","level":"subsection","origin_cites_number":1,"parent_id":"678bd7af-5bf2-44f6-ae32-df413829eb80","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous edge detection (ANOS ED) "],["subsection","GCN Based Techniques"]],"subsections":[],"title":"GCN Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"Instead of using node embeddings for ANOS ED, edge representations learned directly from the graph are also feasible for distinguishing anomalies.\nIf the edge representations well-preserve the graph structure and interaction content (\\eg messages in online social networks, co-authored papers in citation networks) between pairs of nodes, an enhanced detection performance can then be expected.\nTo date, several studies, such as Xu \\etal~, have shown promising results in generating edge representations.\nAlthough they are not specifically designed for graph anomaly detection, they pinpoint a potential approach to ANOS ED. This is highlighted as a potential future direction in Section~\\ref{sec:future:ED}.","id":"231d1246-99ce-4de5-9270-25f30b2ceb5d","level":"subsection","origin_cites_number":1,"parent_id":"678bd7af-5bf2-44f6-ae32-df413829eb80","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous edge detection (ANOS ED) "],["subsection","Network Representation Based Techniques"]],"subsections":[],"title":"Network Representation Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:edge:dynamic}\n\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n\\centerline{\\includegraphics[width=0.98\\textwidth]{pics/subgraph/SGD_Static.pdf}}\n\\caption{ANOS SGD. Real-world networks are usually represented as bipartite graphs for reflecting the interactions between two different types of objects. To detect ANOS SGD, source nodes and sink nodes are embedded using two autoencoders (linked by a shared loss function), respectively. Anomalous sub-graphs are identified by applying dense region detection algorithms in the embedding space.}\n\\label{pic:Staic_SG}\n\\end{figure*}\nDynamic graphs are powerful in reflecting the appearance/disappearance of edges over time~.\nAnomalous edges can be distinguished by modeling the changes in graph structure and capturing the edge distributions at each time step.\nRecent approaches to ANOS ED on dynamic graphs are reviewed in this section.","id":"78c71045-f884-4e7f-b2a8-d7769a5bc16b","level":"section","origin_cites_number":1,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ED on Dynamic Graphs"]],"subsections":["af8f0ff9-09a3-46b3-96d8-edda541a9029","946969f4-52c1-445f-99a0-7e162c350a60"],"title":"ANOS ED on Dynamic Graphs"},{"cite_extract_rate":0,"cites":[],"content":"The intuition of network representation based techniques is to encode the dynamic graph structure information into edge representations and apply the aforementioned traditional anomaly detection techniques to spot irregular edges.\nThis is quite straightforward, but there remain vital challenges in generating/updating informative edge representations when the graph structure evolves.\nTo mitigate this challenge, the ANOS ND model NetWalk~ is also capable of detecting anomalous edges in dynamic graphs. \nFollowing the line of distance-based anomaly detection, NetWalk encodes edges into a shared latent space using node embeddings, and anomalies are identified based on their distances to the nearest edge-cluster centers in the latent space.\nPractically, Netwalk generates edge representations as the Hadamard product of the source and destination nodes' representations, denoted as: $\\mathbf{z}_{u,v} = \\mathbf{z}_u \\odot \\mathbf{z}_v$.\nWhen new edges arrive or existing edges disappear, the node and edge representations are updated from random walks in the temporary graphs at each time stamp, after which the edge-cluster centers and edge anomaly scores are recalculated. \nFinally, the top-k farthest edges to the edge-clusters are reported as anomalies.","id":"af8f0ff9-09a3-46b3-96d8-edda541a9029","level":"subsection","origin_cites_number":1,"parent_id":"78c71045-f884-4e7f-b2a8-d7769a5bc16b","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ED on Dynamic Graphs"],["subsection","Network Representation Based Techniques"]],"subsections":[],"title":"Network Representation Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"Although NetWalk is capable of detecting anomalies in dynamic graphs, it simply updates edge representations without considering the evolving patterns of long/short-term nodes and the graph's structure.\nFor more effective ANOS ED, Zheng \\etal~ intuitively combined temporal, structural and attribute information to measure the anomalousness of edges in dynamic graphs. \nThey propose a semi-supervised model, AddGraph, which comprises a GCN and Gated Recurrent Units (GRU) with attention~ to capture more representative structural information from the temporal graph in each time stamp and dependencies between them, respectively.\nAt each time stamp $t$, GCN takes the output hidden state ($H^{t-1}$) at time $t-1$ to generate node embeddings, after which the GRU learns the current hidden state $H^{t}$ from the node embeddings and attentions on previous hidden states (as shown in Fig.~\\ref{pic:Dynamic_Edge_Comparison}).\nAfter getting the hidden state $H^{t}$ of all nodes, AddGraph assigns an anomaly score to each edge in the temporal graph based on the nodes associated with it. The proposed anomaly scoring function is formulated as:\n\\begin{equation}\\label{eq:AddGraph:scoring}\n f(u,v,w) = w\\cdot\\sigma(\\beta\\cdot (||\\mathbf{a}\\odot \\mathbf{h}_{u} + \\mathbf{b}\\odot \\mathbf{h}_{v}||^{2}_{2} - \\mu)),\n\\end{equation}\nwhere $u$ and $v$ are the corresponding nodes, $w$ is the weight of the edge, $\\mathbf{a}$ and $\\mathbf{b}$ are trainable parameters, $\\beta$ and $\\mu$ are hyper-parameters, and $\\sigma(\\cdot)$ is the non-linear activation function.\nTo learn $\\mathbf{a}$ and $\\mathbf{b}$, Zheng \\etal further assumed that all existing edges in the dynamic graph are normal in the training stage, and sampled non-existing edges as anomalies. Specifically, they form the loss function as:\n\\begin{equation}\\label{eq:AddGraph:loss}\n \\begin{split}\n \\mathcal{L}_{AddGraph} = &\\min\\sum_{{(u,v,w)\\in \\varepsilon^{t}}}\\sum_{{(u^{\\prime},v^{\\prime},w)\\notin \\varepsilon^{t}}} \\\\\n & \\max\\{0, \\gamma + f(u,v,w) - f(u^{\\prime},v^{\\prime},w)\\} + \\lambda\\mathcal{L}_{reg},\n \\end{split}\n\\end{equation}\nwhere $\\varepsilon^{t}$ is the edge set, $(u^{\\prime},v^{\\prime})$ are sampled non-existing edges at time stamp $t$, $\\lambda$ is a hyper-parameter, and $\\mathcal{L}_{reg}$ regularizes all trainable parameters in the model. \nAfter training, the scoring function identifies anomalous edges in the test data by assigning higher anomaly scores to them based on Eq.~\\ref{eq:AddGraph:scoring}.","id":"946969f4-52c1-445f-99a0-7e162c350a60","level":"subsection","origin_cites_number":2,"parent_id":"78c71045-f884-4e7f-b2a8-d7769a5bc16b","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ED on Dynamic Graphs"],["subsection","GCN Based Techniques"]],"subsections":[],"title":"GCN Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:subgraph}\nIn real life, anomalies might also collude and behave collectively with others to garner benefits.\nFor instance, fraudulent user groups in an online review network, as shown in Fig.~\\ref{Toy}, may post misleading reviews to promote or besmirch certain merchandise. When these data are represented as graphs, anomalies and their interactions usually form suspicious sub-graphs, and ANOS SGD is proposed to distinguish them from the benign.\n\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n\\centerline{\\includegraphics[width=0.98\\textwidth]{pics/graph/graph.pdf}}\n\\caption{ANOS GD on Graph Database. Generally, the graph-level anomaly detection techniques take a graph database as input. By generating embeddings/hidden state for each single graph in the database through D(G)NNs, anomalous graphs can be depicted by One-class classifiers or MLP.}\n\\label{pic:graph_level}\n\\end{figure*}\nUnlike individual and independent graph anomalies, \\ie single nodes or edges, each node and edge in a suspicious sub-graph might be normal.\nHowever, when considered as a collection, they turn out to be anomalous.\nMoreover, these sub-graphs also vary in size and inner structure, making anomalous sub-graph detection more challenging than ANOS ND/ED~.\nAlthough extensive effort has been placed on circumventing this problem, deep-learning techniques have only begun to address this problem in the last five years.\nFor reference, traditional non-deep learning based techniques are briefly introduced in Appendix~\\ref{appendix:subgraph}, and a summary of techniques reviewed for ANOS SGD is provided in Table~\\ref{tb:edgeandsubgraph} at the end of Section~\\ref{sec:anosgd:dynamic}.\nDue to the flexibility of heterogeneous graphs in representing the complex relationships between different kinds of real objects, several recent works have taken advantage of deep network representation techniques to detect real-world anomalies through ANOS SGD.\nFor instance, Wang \\etal~ represented online shopping networks as bipartite graphs (a specific type of heterogeneous graph that has two types of nodes and one type of edge), in which users are source nodes and items are sink nodes.\nFraudulent groups are then detected based on suspicious dense blocks that form in these graphs.\nWang \\etal~ aimed to learn anomaly-aware representations of users such that suspicious users in the same group will be located closely in the vector space, while benign users will be far away (as shown in the embedding space in Fig.~\\ref{pic:Staic_SG}). \nAccording to the observation that user nodes belonging to one fraudulent group are more likely to connect with the same item nodes, the developed model, DeepFD, measures similarities in the behavior of two users, $sim_{ij}$, as the percentage of items shared among all the items they have reviewed.\nUser representations are then generated through a traditional autoencoder, which is trained using three losses and follows the encoding-decoding process.\nThe first loss is the reconstruction loss $\\mathcal{L}_{res}$ that ensures the bipartite graph structure can be reconstructed properly using the learned user representations and item representations.\nThe second term $\\mathcal{L}_{sim}$ preserves the user similarity information in the learned user representations. That is, if two users have similar behaviors, their representations should also be similar. This loss is formulated as:\n\\begin{equation}\\label{eq:DeepFD:simloss}\n \\mathcal{L}_{sim} = \\sum_{i,j=1}^{m} sim_{ij} \\cdot \\| \\widehat{sim}_{ij} - sim_{ij} \\|_{2}^{2},\n\\end{equation}\nwhere $m$ is the number of user nodes, $\\widehat{sim}_{ij}$ measures the similarity of user $i$ and $j$'s representations using an RBF kernel or other alternative.\nThe third loss $\\mathcal{L}_{reg}$ regularizes all trainable parameters.\nFinally, the suspicious dense blocks, which are expected to form dense regions in the vector space, are detected using DBSCAN~.\nAnother work, FraudNE~, also models online review networks as bipartite graphs and further detects both malicious users and associated manipulated items following the dense block detection principle.\nUnlike DeepFD, FraudNE aspires to encode both types of nodes into a shared latent space where suspicious users and items belonging to the same dense block are very close to each other while others distribute uniformly (as shown in Fig.~\\ref{pic:Staic_SG}).\nFraudNE adopts two traditional autoencoders, namely, a source node autoencoder and a sink node autoencoder, to learn user representations and item representations, respectively.\nBoth autoencoders are trained to jointly minimize their corresponding reconstruction losses and a shared loss function, and the total loss can be formulated as:\n\\begin{equation}\n \\mathcal{L}_{FraudNE} = \\mathcal{L}_{res}^{source} + \\mathcal{L}_{res}^{sink} + \\alpha \\mathcal{L}_{share} + \\eta \\mathcal{L}_{reg},\n\\end{equation}\nwhere $\\alpha$ and $\\eta$ are hyperparameters, and $\\mathcal{L}_{reg}$ regularizes all trainable parameters.\nSpecifically, the reconstruction losses (\\ie $\\mathcal{L}_{res}^{source}$ and $\\mathcal{L}_{res}^{sink}$) measure the gap between the input user/item features (extracted from the graph structure) and their decoded features. \nThe shared loss function is proposed to restrict the representation learning process such that each linked pair of users and items get similar representations.\nAs the DBSCAN~ algorithm is convenient to apply for dense region detection, FraudNE also uses it to distinguish the dense sub-graphs formed by suspicious users and items.\nTo date, only a few works have put their efforts into using deep learning techniques for ANOS SGD. However, with intensifying research interest in sub-graph representation learning, we encourage more studies on ANOS SGD and highlight this as a potential future in Section~\\ref{sec:future:ED}.\n\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n \\centerline{\\includegraphics[width=0.98\\textwidth]{pics/graph/dynamic.pdf}}\n \\caption{ANOS GD on dynamic graphs. For each graph snapshot in the dynamic graph, the LSTM autoencoder generates its hidden state using its adjacency matrix and previous hidden states. Through hypersphere learning, a hypersphere with centroid $a$ and radius $r$ is learned such that anomalous snapshots lay outside.}\n \\label{pic:GDdynamic}\n\\end{figure*}","id":"5da2d684-654a-4048-95ec-2ae406b319b2","level":"section","origin_cites_number":4,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous sub-graph detection (ANOS SGD)"]],"subsections":[],"title":"Anomalous sub-graph detection (ANOS SGD)"},{"cite_extract_rate":0.2,"cites":[8992],"content":"\\label{sec:anosgd:db}\nBeyond anomalous node, edge, and sub-graph, graph anomalies might also appear as abnormal graphs in a set/database of graphs.\nTypically, a graph database is defined as:\n\\textbf{\\textit{Definition 4 (Graph Database)}}. A graph database $\\mathcal{G} = \\{ G_i = (V_i,E_i,X_v(i), X_e(i)) \\}^N_{i=1}$ contains $N$ individual graphs. Here, each graph $G_i$ is comprised of a node set $V_i$ and an edge set $E_i$. $X_v(i)$ and $X_e(i)$ are the node attribute matrix and edge attribute matrix of $G_i$ if it is an attributed graph.\nThis graph-level ANOS GD aims to detect individual graphs that deviate significantly from the others.\nA concrete example of ANOS GD is unusual molecule detection.\nWhen chemical compounds are represented as molecular/chemical graphs where the atoms and bonds are represented as nodes and edges~, unusual molecules can be identified because their corresponding graphs have structures and/or features that deviate from the others.\nBrain disorders detection is another example.\nA brain disorder can be diagnosed by analyzing the dynamics of brain graphs at different stages of aging in sequence and finding an inconsistent snapshot at a specific time stamp.\nThe prior reviewed techniques (\\ie ANOS ND/ED/SGD) are not compatible with ANOS GD because they are dedicated to detecting anomalies in a single graph, whereas ANOS GD is directed at detecting graph-level anomalies.\nThis problem is commonly approached by: 1) measuring the pairwise proximities of graphs using graph kernels~; 2) detecting the appearance of anomalous graph signals created by abnormal groups of nodes~; or 3) encoding graphs using frequent motifs~.\nHowever, none of these methods are deep learning-based.\nAs the time of writing, very few studies in ANOS GD with deep learning have been undertaken. As such, this is highlighted as a potential future direction in Section~\\ref{sec:future:ED}.","id":"9ba15010-a3db-4588-8e7f-56a7def289f5","level":"section","origin_cites_number":5,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous graph detection (ANOS GD)"]],"subsections":["49385046-24ac-4a52-8807-655e25c7942e","c5bcea84-0572-429e-a5e7-37efbed81759"],"title":"Anomalous graph detection (ANOS GD)"},{"cite_extract_rate":0.33333333333333304,"cites":[8149],"content":"Motivated by the success of GNNs in various graph classification tasks, the most recent works in ANOS GD employ GNNs to classify single graphs as normal/abnormal in the given graph database. Specifically, Dou \\etal~ transformed fake news detection into an ANOS GD problem by modeling news as tree-structured propagation graphs where the root nodes denote pieces of news, and child nodes denote users who interact with the root news. Their end-to-end framework, UPFD, extracts two embeddings for the news piece and users, respectively, via a text embedding model (e.g. word2vec, BERT) and a user engagement embedding process. For each news graph, its latent representation is a flattened concatenation of these two embeddings, which is input to train a neural classifier with the label of the news. \nCorresponding propagation graphs that are labeled as fake by the trained model are regarded as anomalous.\nAnother representative work by Zhao and Akoglu~ employed a GIN model and one-class classification (\\ie DeepSVDD~) loss to train a graph-level anomaly detection framework in an end-to-end manner.\nFor each individual graph in the graph database, its graph-level embedding is generated by applying mean-pooling over its nodes' node-level embeddings.\nA graph is eventually depicted as anomalous if it lies outside the learned hypersphere, as shown in Fig.~\\ref{pic:graph_level}.","id":"49385046-24ac-4a52-8807-655e25c7942e","level":"subsection","origin_cites_number":3,"parent_id":"9ba15010-a3db-4588-8e7f-56a7def289f5","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous graph detection (ANOS GD)"],["subsection","GNN Based Techniques"]],"subsections":[],"title":"GNN Based Techniques"},{"cite_extract_rate":0.5,"cites":[3953],"content":"It is also possible to apply general graph-level network representation techniques to ANOS GD.\nWith these methods, the detection problem is transformed into a conventional outlier detection problem in the embedding space.\nIn contrast to D(G)NN based techniques that can detect graph anomalies in an end-to-end manner, adopting these representation techniques for anomaly detection is two-staged.\nFirst, graphs in the database are encoded into a shared latent space using graph-level representation techniques, such as Graph2Vec~, FGSD~.\nThen, the anomalousness of each single graph is measured by an off-the-shelf outlier detector.\nEssentially, this kind of approach involves pairing existing methods in both stages, yet, the stages are disconnected from each other and, hence, the detection performance can be subpar since the embedding similarities are not necessarily designed for the sake of anomaly detection.","id":"c5bcea84-0572-429e-a5e7-37efbed81759","level":"subsection","origin_cites_number":2,"parent_id":"9ba15010-a3db-4588-8e7f-56a7def289f5","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Anomalous graph detection (ANOS GD)"],["subsection","Network Representation Based Techniques"]],"subsections":[],"title":"Network Representation Based Techniques"},{"cite_extract_rate":0.412698412698412,"cites":[8993,6285,6281,6302,6297,3333,6283,6303,6296,6289,3700,6294,6304,6299,6271,6301,1420,6277,6295,6300,6293,5645,8149,6291,6286,6298],"content":"\\label{sec:anosgd:dynamic}\nFor dynamic graph environments, graph-level anomaly detection endeavors to identify abnormal graph snapshots/temporal graphs. Similar to ANOS ND and ED on dynamic graphs, given a sequence of graphs, anomalous graphs can be distinguished regarding their unusual evolving patterns, abnormal graph-level features, or other characteristics.\n\\begin{table*}[!h]\n\\centering \n\\footnotesize\n\\renewcommand\\arraystretch{1}\n\\setlength{\\tabcolsep}{2.8mm}\n\\caption{Summary of Anomalous Edge, Sub-graph and Graph Detection Techniques.} \n\\resizebox{0.96\\textwidth}{!}{\n\\begin{tabular}{m{2.1cm}<{\\centering}|m{2.3cm}<{\\centering}|m{1.1cm}<{\\centering}|m{4cm}<{\\centering}|m{1.8cm}<{\\centering}|m{2.9cm}<{\\centering}}\n\\toprule[1 pt]\n\\textbf{Graph Type} & \\textbf{Approach} & \\textbf{Category} & \\textbf{Objective Function} & \\textbf{Measurement} & \\textbf{Outputs} \\\\ \n\\midrule[1 pt]\n\\multicolumn{6}{c}{Anomalous Edge Detection Techniques} \\\\\n\\midrule[1 pt]\n\\multirow{2}{*}{\\shortstack{Static Graph - \\\\ Plain}} &UGED~ & DNN & $\\text{cross-entropy}(f(u,N(u)), v)$ & Anomaly Score & $\\text{mean}(1-P(v|u,N(u)), 1-P(u|v,N(v)))$ \\\\ \\cline{2-6}\n&AANE~ & GCN & $\\mathcal{L} = \\mathcal{L}_{afl} + \\gamma\\mathcal{L}_{aal}$ & Anomaly Ranking & Edge Existing Probability \\\\ \n\\midrule[1 pt]\nStatic Graph - Attributed &eFraudCom~ & NR & $\\mathcal{L}_{MAIN} - \\lambda\\mathcal{L}_{MI}$ & Anomaly Prediction & Predicted Label \\\\ \n\\midrule[1 pt]\nDynamic Graph - Plain & NetWalk~ & NR & $\\gamma\\mathcal{L}_{AE} + \\mathcal{L}_{Clique} + \\lambda\\|W\\|_{F}^{2} + \\beta KL$ & Anomaly Score & Nearest Distance to Cluster Centers \\\\\n\\midrule[1 pt]\nDynamic Graph - Attributed & AddGraph~ & GCN & $\\text{min} \\sum_{{e\\in \\varepsilon^{t}}}\\sum_{{e^{\\prime}\\notin \\varepsilon^{t}}} \\text{max}\\{0, \\gamma + f(i,j,w) - f(i^{\\prime},j^{\\prime},w)\\} + \\lambda\\mathcal{L}_{reg}$ & Anomaly Score & $f(i,j,w) = w\\cdot\\sigma(\\beta\\cdot (||\\mathbf{a} \\odot \\mathbf{h}_{i} + \\mathbf{b} \\odot \\mathbf{h}_{j}||^{2}_{2} - \\mu))$ \\\\ \n\\midrule[1 pt]\n\\midrule[1 pt]\n\\multicolumn{6}{c}{Anomalous Sub-graph Detection Techniques} \\\\\n\\midrule[1 pt]\n\\multirow{2}{*}{\\shortstack{Static Graph - \\\\ Plain}} &DeepFD~ & NR & $\\mathcal{L}_{recon} +\\alpha \\mathcal{L}_{sim} + \\gamma \\mathcal{L}_{reg}$& Density-based Method (DBSCAN) & Dense sub-graphs \\\\ \\cline{2-6}\n&FraudNE~ & NR & $\\mathcal{L}_{res}^{source} + \\mathcal{L}_{res}^{sink} + \\alpha \\mathcal{L}_{share} + \\eta \\mathcal{L}_{reg}$ & Density-based Method (DBSCAN) & Dense sub-graphs \\\\\n\\midrule[1 pt]\n\\midrule[1 pt]\n\\multicolumn{6}{c}{Anomalous Graph Detection Techniques} \\\\\n\\midrule[1 pt]\n\\multirow{2}{*}{\\shortstack{Graph Database - \\\\ Attributed}}&UPFD~ & NR & $ -(y\\log(p) + (1-y)\\log(1-p))$ & Anomaly Prediction & Predicted Label \\\\ \\cline{2-6}\n& OCGIN~ & GNN & $ \\min\\limits_{W}\\frac{1}{N}\\sum\\limits_{i=1}^{N}||GIN(G_i,W)-c||^2 + \\frac{\\lambda}{2}\\sum\\limits_{l=1}^{L}||W^l||^2_{F}$ & Location in Embedding Space & Distance to Hypersphere Center \\\\\n\\midrule[1 pt]\n\\multirow{2}{*}{\\shortstack{Dynamic Graph - \\\\ Plain}} & DeepSphere~ & DNN & $\\mathcal{L} = \\mathcal{L}_{h} + \\lambda\\mathcal{L}_{res}$ & Location in Embedding Space & Anomalous Label \\\\ \\cline{2-6}\n& GLAD-PAW~ & GNN & $\\text{cross-entropy}(\\mathbf{y},\\mathbf{\\hat{y}})$ & Anomaly Prediction & Predicted Label \\\\ \n\\bottomrule[1 pt]\n\\multicolumn{6}{l}{* DNN: Deep NN Based techniques, GCN: GCN Based Techniques, NR: Network Representation Based Techniques.}\\\\\n\\multicolumn{6}{l}{* GNN: Graph Neural Network Based Techniques.}\n\\end{tabular}\n\\label{tb:edgeandsubgraph}\n}\n\\end{table*}\nIn order to derive each graph snapshot/temporal graph’s characteristics, the commonly used GNN, LSTM and autoencoder are feasible to apply. For instance, Teng \\etal~ applied a LSTM-autoencoder to detect abnormal graph snapshots, as shown in Fig.~\\ref{pic:GDdynamic}.\nIn their proposed model, DeepSphere, a dynamic graph is described as a collection of three-order tensors, $\\{\\mathcal{X}_k, k=1,2...\\}$ where each $\\mathcal{X} \\in \\mathcal{R}^{N \\times N \\times T}$, and the slices along the time dimension are the adjacency matrices of graph snapshots.\nTo identify abnormal tensors, DeepSphere first embeds each graph snapshot into a latent space using an LSTM autoencoder, and then leverages a one-class classification objective~ that learns a hypersphere such that normal snapshots are covered, and anomalous snapshots lay outside.\nThe LSTM autoencoder takes the adjacency matrices as input sequentially and attempts to reconstruct these input matrices through training.\nThe hypersphere is learned through a single neural network layer and its objective function is formulated as:\n\\begin{equation} \\label{eq:ds:hsp}\n \\mathcal{L}_{h} = r^{2} + \\gamma\\sum_{k=1}^{m} \\epsilon_{k} + \\frac{1}{m}\\sum_{k=1}^{m} \\|\\textbf{z}_{k} - \\textbf{a}\\|^{2},\n\\end{equation}\nwhere $\\textbf{z}_{k}$ is the latent representation generated by the LSTM autoencoder, $\\textbf{a}$ is the centroid of the hypersphere, $r$ is the radius, $\\epsilon_{k}$ is the outlier penalty ($\\epsilon_{k} = \\|\\textbf{z}_{k} - \\textbf{a}\\|^{2} - r^2$), $m$ is the number of training graph snapshots, and $\\gamma$ is a hyperparameter. \nThe overall objective function of DeepSphere is represented as:\n\\begin{equation} \\label{eq:ds:obj}\n \\mathcal{L} = \\mathcal{L}_{h} + \\lambda\\mathcal{L}_{res},\n\\end{equation}\nwhere $\\mathcal{L}_{res}$ is the reconstruction loss of the LSTM autoencoder.\nWhen the training is finished, DeepSphere spots a given unseen data $\\mathcal{X}$ as anomalous if its embedding lies outside the learned hypersphere with a radius of $r$.\nIn addition to all ANOS ND, ED, SGD, and GD techniques reviewed above, it is worth mentioning that perturbed graphs, which adversarial models generate to attack graph classification algorithms or GNNs~, can also be regarded as (intensional) anomalies. In a perturbed graph, the nodes and edges are modified deliberately to deviate from the others. We have not reviewed these in this survey because their main purpose is to attack a GNN model. The key idea behind these methods is the attacking/perturbation strategy, and studies in this sphere seldom focus on a detection or reasoning module to identify the perturbed graph or its sub-structures, \\ie anomalous nodes, edges, sub-graphs, or graphs.\n\\begin{table*}[!t]\n\\centering \n\\footnotesize\n\\renewcommand\\arraystretch{1}\n\\setlength{\\tabcolsep}{2.8mm}\n\\caption{Published Algorithms and Models} \n\\resizebox{0.96\\textwidth}{!}{\n\\begin{tabular}{c|c|c|c|l}\n\\toprule[1 pt]\n\\textbf{Model} & \\textbf{Language} & \\textbf{Platform} & \\textbf{Graph} & \\textbf{Code Repository} \\\\ \n\\midrule[1 pt]\nAnomalyDAE~ & Python & Tensorflow & Static Attributed Graph & https://github.com/haoyfan/AnomalyDAE \\\\ \\hline\nMADAN~ & Python & - & Static Attributed Graph & https://github.com/leoguti85/MADAN \\\\ \\hline\nPAICAN~ & Python & Tensorflow & Static Attributed Graph & http://www.kdd.in.tum.de/PAICAN/ \\\\ \\hline\nONE~ & Python & - & Static Attributed Graph & https://github.com/sambaranban/ONE \\\\ \\hline\nDONE\\&AdONE~ & Python & Tensorflow & Static Attributed Graph & https://bit.ly/35A2xHs \\\\ \\hline\nSLICENDICE~ & Python & - & Static Attributed Graph & http://github.com/hamedn/SliceNDice/ \\\\ \\hline\nFRAUDRE~ & Python & Pytorch & Static Attributed Graph & https://github.com/FraudDetection/FRAUDRE \\\\ \\hline\nSemiGNN~ & Python & Tensorflow & Static Attributed Graph & https://github.com/safe-graph/DGFraud \\\\ \\hline\nCARE-GNN~ & Python & Pytorch & Static Attributed Graph & https://github.com/YingtongDou/CARE-GNN \\\\ \\hline\nGraphConsis~ & Python & Tensorflow & Static Attributed Graph & https://github.com/safe-graph/DGFraud \\\\ \\hline\nGLOD~ & Python & Pytorch & Static Attributed Graph & https://github.com/LingxiaoShawn/GLOD-Issues \\\\ \\hline\nOCAN~ & Python & Tensorflow & Static Graph &https://github.com/PanpanZheng/OCAN \\\\ \\hline\nDeFrauder~ & Python & - & Static Graph & https://github.com/LCS2-IIITD/DeFrauder \\\\ \\hline\nGCAN~ & Python & Keras & Heterogeneous Graph & https://github.com/l852888/GCAN \\\\ \\hline\nHGATRD~ & Python & Pytorch & Heterogeneous Graph & https://github.com/201518018629031/HGATRD \\\\\\hline\nGLAN~ & Python & Pytorch & Heterogeneous Graph & https://github.com/chunyuanY/RumorDetection \\\\ \\hline\nGEM~ & Python & - & Heterogeneous Graph & https://github.com/safe-graph/DGFraud/tree/master/algorithms/GEM \\\\ \\hline\neFraudCom~ & Python & Pytorch & Heterogeneous Graph & https://github.com/GeZhangMQ/eFraudCom \\\\ \\hline\nDeepFD~ & Python & Pytorch & Bipartite Graph & https://github.com/JiaWu-Repository/DeepFD-pyTorch \\\\ \\hline\nANOMRANK~ & C++ & - & Dynamic Graph & https://github.com/minjiyoon/anomrank \\\\ \\hline\nMIDAS~ & C++ & - & Dynamic Graph & https://github.com/Stream-AD/MIDAS \\\\ \\hline\nSedanspot~ & C++ & - & Dynamic Graph & https://www.github.com/dhivyaeswaran/sedanspot \\\\ \\hline\nF-FADE~ & Python & Pytorch & Dynamic Graph & http://snap.stanford.edu/f-fade/ \\\\ \\hline\nDeepSphere~ & Python & Tensorflow & Dynamic Graph & https://github.com/picsolab/DeepSphere \\\\ \\hline\nChangedar~ & Matlab & - & Dynamic Graph & https://bhooi.github.io/changedar/ \\\\\\hline\nUPFD~ & Python & Pytorch & Graph Database & https://github.com/safe-graph/GNN-FakeNews \\\\ \\hline\nOCGIN~ & Python & Pytorch & Graph Database & https://github.com/LingxiaoShawn/GLOD-Issues \\\\ \\hline\nDAGMM~ & Python & Pytorch & Non Graph & https://github.com/danieltan07/dagmm \\\\ \\hline\nDevNet~ & Python & Tensorflow & Non Graph &https://github.com/GuansongPang/deviation-network \\\\ \\hline\nRDA~ & Python & Tensorflow & Non Graph &https://github.com/zc8340311/RobustAutoencoder \\\\ \\hline\nGAD~ & Python & Tensorflow & Non Graph &https://github.com/raghavchalapathy/gad \\\\ \\hline\nDeep SAD~ &Python & Pytorch & Non Graph & https://github.com/lukasruff/Deep-SAD-PyTorch \\\\ \\hline\nDATE~ &Python & Pytorch &Non Graph & https://github.com/Roytsai27/Dual-Attentive-Tree-aware-Embedding \\\\ \\hline\nSTS-NN~ & Python & Pytorch & Non Graph & https://github.com/JiaWu-Repository/STS-NN \\\\\n\\bottomrule[1 pt]\n\\multicolumn{5}{l}{* -: No Dedicated Platforms.}\n\\end{tabular}\n\\label{tb:publishedalgs}\n}\n\\end{table*}\n\\begin{table*}[!t]\n\\centering \n\\footnotesize\n\\renewcommand\\arraystretch{1}\n\\setlength{\\tabcolsep}{2.8mm}\n\\caption{Published Datasets.}\n\\resizebox{0.96\\textwidth}{!}{\n\\begin{tabular}{m{1.6cm}<{\\centering}|m{1.4cm}<{\\centering}|m{0.5cm}<{\\centering}|m{0.5cm}<{\\centering}|m{0.5cm}<{\\centering}|m{0.5cm}<{\\centering}|m{0.5cm}<{\\centering}|m{2cm}<{\\centering}|l}\n\\toprule[1 pt]\n\\textbf{Category} & \\textbf{Dataset} & \\textbf{\\#G} & \\textbf{\\#N} & \\textbf{\\#E} & \\textbf{\\#FT} & \\textbf{\\#AN} & \\textbf{REF} & \\textbf{URL} \\\\ \n\\midrule[1 pt]\n\\multirow{5}{*}{\\shortstack{Citation \\\\ Networks}} \n& ACM & 1 & 16K & 71K & 8.3K & - & &\\text{http://www.arnetminer.org/open-academic-graph} \\\\ \\cline{2-9}\n& Cora & 1 & 2.7K & 5.2K & 1.4K & - & &http://linqs.cs.umd.edu/projects/projects/lbc \\\\ \\cline{2-9}\n& Citeseer & 1& 3.3K & 4.7K & 3.7K & - & & http://linqs.cs.umd.edu/projects/projects/lbc \\\\ \\cline{2-9}\n& Pubmed & 1& 19K & 44K & 500 & - & &http://linqs.cs.umd.edu/projects/projects/lbc \\\\ \\cline{2-9}\n& DBLP & 1& - & - & - & - & &http://www.informatik.uni-trier.de/˜ley/db/ \\\\\n\\midrule[1 pt]\n\\multirow{10}{*}{\\shortstack{Social \\\\ Networks}}\n&Enron & - & 80K & - & - & - & & http://odds.cs.stonybrook.edu/\\#table2 \\\\ \\cline{2-9}\n&UCI Message & 1& 5K & - & - & - & &http://archive.ics.uci.edu/ml \\\\ \\cline{2-9}\n&Google+ & 4 & 75M & 11G & - & - & - & https://wangbinghui.net/dataset.html \\\\ \\cline{2-9}\n&Twitter Sybil & 3 & 41M & - & - & 100K & - & https://wangbinghui.net/dataset.html \\\\ \\cline{2-9}\n&Twitter WorldCup2014 & - & 54K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\ \\cline{2-9}\n&Twitter Security2014 & - & 130K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\ \\cline{2-9}\n&Reality Mining & - & 9.1K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\ \\cline{2-9}\n&NYTNews & - & 320K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\ \\cline{2-9}\n&Politifact & 314 & 41K & 40K & - & 157 & &https://github.com/safe-graph/GNN-FakeNews \\\\ \\cline{2-9}\n&Gossipcop & 5.4K & 314K & 308K & - & 2.7K & &https://github.com/safe-graph/GNN-FakeNews \\\\ \n\\midrule[1 pt]\n\\multirow{5}{*}{\\shortstack{Co-purchasing \\\\ Networks}}\n& Disney & 1 & 124 & 334 & 30 & 6 & &https://www.ipd.kit.edu/mitarbeiter/muellere/consub/ \\\\ \\cline{2-9}\n&Amazon-v1 & 1 & 314K & 882K & 28 & 6.2K & &https://www.ipd.kit.edu/mitarbeiter/muellere/consub/ \\\\ \\cline{2-9}\n&Amazon-v2 & 1 & 11K & - & 25 & 821 & -&https://github.com/dmlc/dgl/blob/master/python/dgl/data/fraud.py \\\\ \\cline{2-9}\n&Elliptic & 1 & 203K & 234K & 166 & 4.5K & - &https://www.kaggle.com/ellipticco/elliptic-data-set \\\\\\cline{2-9}\n&Yelp & 1 & 45K & - & 32 & 6.6K & -&https://github.com/dmlc/dgl/blob/master/python/dgl/data/fraud.py \\\\ \n\\midrule[1 pt]\nTransportation Networks & New York City Taxi& - & - & - & - & - & &http://www.nyc.gov/html/tlc/html/about/triprecorddata.shtml \\\\\n\\bottomrule[1 pt]\n\\multicolumn{9}{l}{* -: Not Given, \\#G: Number of Graphs, \\#N: Number of Nodes, \\#E: Number of Edges, \\#FT: Number of Features, \\#AN: Number of Anomalies, REF: References.}\n\\end{tabular}\n\\label{tb:publisheddataset}\n}\n\\end{table*}","id":"ae3bbf88-f1cb-408b-8ded-ef8445178dfa","level":"section","origin_cites_number":63,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS GD on Dynamic Graphs"]],"subsections":[],"title":"ANOS GD on Dynamic Graphs"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:resources}\nAcquiring open-sourced implementations and real-world datasets with real-world anomalies are far from trivial in academic research on graph anomaly detection.\nHere, we first list the published algorithms with publicly available implementations, then we provide a collection of public benchmark datasets and summarize the commonly used evaluation metrics. \nLastly, due to the shortage of labeled anomalies in real-world datasets, we review three synthetic dataset generation strategies used in the existing works.\nAll the resources are available at: https://github.com/XiaoxiaoMa-MQ/Awesome-Deep-Graph-Anomaly-Detection/.","id":"8147e459-f864-454f-86f6-52d0b0cd6b52","level":"section","origin_cites_number":0,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Published Algorithms and Datasets"]],"subsections":["9907d812-2cb7-4bfc-ba97-f1d8f46efce2","152cd0db-3934-4cfa-8fbb-64d458d75863","4da63f48-4501-4170-981d-8fe6cdbe1f52","6b67eb52-a545-45eb-aa8e-2135a2a408e8"],"title":"Published Algorithms and Datasets"},{"cite_extract_rate":0,"cites":[],"content":"The published implementations of algorithms and models contribute to baseline experiments. Table~\\ref{tb:publishedalgs} provides a summary of published implementations outlining the language and platforms, graphs that they can admit, and URLs to code repositories.","id":"9907d812-2cb7-4bfc-ba97-f1d8f46efce2","level":"subsection","origin_cites_number":0,"parent_id":"8147e459-f864-454f-86f6-52d0b0cd6b52","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Published Algorithms and Datasets"],["subsection","Published Algorithms"]],"subsections":[],"title":"Published Algorithms"},{"cite_extract_rate":0,"cites":[],"content":"Table~\\ref{tb:publisheddataset} summarizes the most commonly used datasets, categorizing them into different groups with regard to their application fields. Notably, there is a lack of anomaly ground truth with labeled anomalies only provided in the Enron, Twitter Sybil, Disney, Amazon, Elliptic and Yelp datasets. Details of the DBLP, UCI message, Digg, Wikipedia, and New York city taxi datasets are not given because these public datasets only contain the raw data, and in most existing works, they are further processed to build different graphs (\\eg homogeneous graphs, bipartite graphs).\nThe well-known citation networks are often used to generate synthetic datasets by injecting anomalies into them - the number of anomalies varies from study to study.\nIn addition to these anomaly detection datasets, Table~\\ref{tb:publisheddataset} also lists eight graph classification datasets.\nAs mentioned in~\\ref{sec:synthetic}, through downsampling, these datasets can be used as benchmarks for evaluating the anomaly detection performance.","id":"152cd0db-3934-4cfa-8fbb-64d458d75863","level":"subsection","origin_cites_number":0,"parent_id":"8147e459-f864-454f-86f6-52d0b0cd6b52","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Published Algorithms and Datasets"],["subsection","Published Datasets"]],"subsections":[],"title":"Published Datasets"},{"cite_extract_rate":0.14285714285714202,"cites":[6281],"content":"\\label{sec:synthetic}\nGiven the rarity of ground-truth anomalies, many researchers have employed synthetic datasets to investigate the effectiveness of their proposed methods~. \nTypically, these datasets can be categorized as follows: \n\\begin{itemize}\n \\item \\textit{Synthetic graphs with injected anomalies.} Pursuing this strategy, graphs are created to simulate real-world networks. All the nodes and edges are manually added with well-known benchmarks (\\eg Lanchinetti-Fornunato-Radicchi (LFR)~, small-world~, scale-free graphs~). Once built, ground-truth anomalies are planted into the network. For the feasibility of generating expected scale of networks, this strategy is mostly used by previous works to validate their underlying intuitions in anomaly detection.\n \\item \\textit{Real-world datasets with injected anomalies.} These datasets are built based on the real-world networks. In particular, anomalies are created either by modifying the topological structure or the attributes of existing nodes/edges/sub-graphs, or by inserting non-existent graph objects. \n \\item \\textit{Downsampled graph classification datasets.} The widely-used graph classification datasets (\\eg NCI1, IMDB, ENZYMES in~) can be easily converted into sets suitable for anomaly detection through two steps. Firstly, one particular class and its data records are chosen to represent normal objects. Then, the other data records are downsampled as anomalies at a specified downsampling rate. By this, the generated graph anomaly detection dataset is, in fact, a subset of the original dataset. The most significant strength of this strategy is that no single data record has been modified.\n\\end{itemize}","id":"4da63f48-4501-4170-981d-8fe6cdbe1f52","level":"subsection","origin_cites_number":7,"parent_id":"8147e459-f864-454f-86f6-52d0b0cd6b52","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Published Algorithms and Datasets"],["subsection","Synthetic Dataset Generation"]],"subsections":[],"title":"Synthetic Dataset Generation"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:em}\nTo date, the most widely used metrics for evaluating anomaly detection performance evaluation include accuracy, precision, recall rate, F1-score and AUC-AP (Average Precision). Their formulas/descriptions are given in Table~\\ref{tb:evluationmetrics}. However, a more dedicated analysis with some new evaluation metrics is needed for further performance examination \nbecause anomaly detection has different requirements for different applications~, \\eg false negative and false positive.\nFor instance, network intrusion prevention systems are more sensitive to false negative errors, while false positive errors are considered relatively harmless. This is because any risky connections should be shut down to prevent information leaks.\nBy contrast, other applications concentrate more on the false positives, \\eg in auditing domain, companies often set a budget for an auditor to look at flagged anomalies, and they want high precision/small false positive rate such that the auditor's time can be best used.\nHence, when evaluating detection performance, we suggest reviewing the specific requirements of the application domain for fair and suitable comparisons.\n\\begin{table}\n\\centering \n\\footnotesize\n\\renewcommand\\arraystretch{1}\n\\setlength{\\tabcolsep}{2.8mm}\n\\caption{Evaluation Metrics. $tp$: true positives; $tn$: true negatives; $fp$: false positives; $fn$: false negatives.}\n\\resizebox{0.46\\textwidth}{!}{\n\\begin{tabular}{l|c}\n\\toprule[1 pt]\n\\textbf{Evaluation Metric} & \\textbf{Formula/Description} \\\\ \n\\midrule[1 pt]\nAccuracy & $\\frac{tp + tn}{tp + tn + fp + fn}$ \\\\\\hline\nPrecision & $\\frac{tp}{tp + fp}$ \\\\\\hline\nRecall & $\\frac{tp}{tp + fn}$ \\\\\\hline\nF1 Score & $2 * \\frac{Recall * Precision}{Recall + Precision}$ \\\\\\hline\nAUC-ROC & The Area Under the ROC Curve\\\\ \\hline\nAUC-AP & The Area Under Precision-Recall Curve \\\\\n\\bottomrule[1 pt]\n\\end{tabular}\n\\label{tb:evluationmetrics}\n}\n\\end{table}","id":"6b67eb52-a545-45eb-aa8e-2135a2a408e8","level":"subsection","origin_cites_number":3,"parent_id":"8147e459-f864-454f-86f6-52d0b0cd6b52","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Published Algorithms and Datasets"],["subsection","Evaluation Metrics"]],"subsections":[],"title":"Evaluation Metrics"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:futures}\nSo far, we have reviewed the contemporary deep learning techniques that are devoted to graph anomaly detection.\nAn apparent observation from our survey is there remain many compounded challenges imposed by the complexity of anomaly detection, graph data, and the immaturity of deep learning techniques for graph data mining. Another observation is that deep learning techniques in graph anomaly detection are still confined to a relatively small number of studies, and most of these focus on anomalous node detection. For a gauge, simply compare the length of Tables~\\ref{table:ANOSND} and~\\ref{tb:edgeandsubgraph}. Edge, sub-graph, and graph-level anomaly detection have clearly received much less attention.\nTo bridge the gaps and push forward future work, we have identified 12 directions of future research for graph anomaly detection with deep learning.","id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","level":"section","origin_cites_number":0,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"]],"subsections":["32fa0363-7074-47b3-9e03-5211f2f264be","239988e3-c594-47d4-adb9-69b272543369","aff7afb2-0d0e-4d85-baf6-f135a679f1c2","094e633e-5d3b-45d7-bca3-171ad2d8492b","e773dc9f-5c7d-4700-bcf9-1e7e0772172c","cc214041-b949-4f3c-95ae-5deb83ec1573","85e8f422-2b40-4284-aca9-eaeef926d8d1","1224de5d-ffc7-41d5-bee2-9722227a3ca0","d42a5aab-3ff4-4f52-8576-f381cddddc12","4eb29545-f452-443d-b7b6-8accd75e515c","9d5eeb0b-b689-4f33-bf46-0aa58f25c4d7","b1a4c699-5c1b-4fe6-a282-ae8f4a92410d"],"title":"Future Directions"},{"cite_extract_rate":0.5,"cites":[6305],"content":"\\label{sec:future:ED}\nIn real-world graphs, anomalies also appear as unusual relationships between objects, sub-structures formed by abnormal groups, or abnormal graphs, which are known as anomalous edges, sub-graphs, and graphs respectively.\nAs indicated in our review, there is a huge gap between the existing anomalous edge/sub-graph/graph detection techniques and the emerging demands for more advanced solutions in various application domains (\\eg social networks, computer networks, financial networks).\nWhen detecting anomalous edges/sub-graphs/graphs, the proposed methods should be capable of leveraging the rich information contained in graphs to find clues and characteristics that can distinguish normal objects and anomalies in specific applications.\nTypically, this involves extracting edge/sub-graph/graph-level features, modeling the patterns of these features, and measuring the abnormalities accordingly.\nHowever, current deep learning based graph anomaly detection techniques put forward very little effort in this regard.\n\\textit{Opportunities}: We believe more research efforts can be done on anomalous edge, sub-graph, and graph detection with regard to their significance in real-world applications. Possible solutions to this gap might to be first consider the application domain and explore domain knowledge to find complementary clues as a basis for these problems. Then, motivated by recent advances in deep learning for edge, sub-graph, and graph-level representation learning~, extensive work can be done to learn an anomaly-aware embedding space such that it is feasible to extract abnormal patterns of anomalies. Although this direction seems quite straightforward, the true challenge lies in the specific application domains. Hence, domain knowledge, anomalous pattern recognition and anomaly-aware deep learning techniques should be enforced simultaneously.","id":"32fa0363-7074-47b3-9e03-5211f2f264be","level":"subsection","origin_cites_number":2,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Anomalous Edge, Sub-graph, and Graph Detection"]],"subsections":[],"title":"Anomalous Edge, Sub-graph, and Graph Detection"},{"cite_extract_rate":0.28571428571428503,"cites":[217,218],"content":"Dynamic graphs provide powerful machinery with which to capture the evolving relationships between real objects and their attributes.\nTheir ever-changing structure and attribute information inherently make anomaly detection very challenging in these scenarios, leading to two primary concerns for the task.\nThe first is to consider the spatial and temporal information contained in each graph snapshot at different time stamps, and the second is to explore the evolving patterns of nodes, edges, sub-graphs and graphs, as well as their interaction with the node/edge attributes over time.\nWhen these challenges have been tackled with mature solutions, detection techniques will achieve better results.\n\\textit{Opportunities}: From our observations, most of deep learning based dynamic graph anomaly detection techniques are built on DeepWalk~, GCN~ or other deep models that are intuitively designed for static graphs. This means other information, like evolving patterns in attributes~), are not adequately used in the detection task.\nWe can therefore identify the following directions for future studies to target.\n\\begin{itemize}\n \\item \\textit{Using dynamic graph mining tools.} As a popular research topic, deep learning for dynamic graph data mining~ has shown its effectiveness in supporting dynamic graph analysis, such as node clustering and graph classification~. More future works can be foreseen that adopt these techniques for anomaly detection.\n \\item \\textit{Deriving solid evidence for anomaly detection.} The rich structural, attribute and temporal information in dynamic graphs are valuable resources for identifying anomalies. Apart from the indicators widely used in current works, such as burst of connections between node pairs or suddenly vanishing connections, we suggest exploring structural and attribute changes in depth. From such studies, we may derive additional information to enhance the detection performance, such as the appearance of abnormal attributes.\n \\item \\textit{Handling complex dynamics.} Real-world networks always exhibit changes in both the network structure and node attributes, but only very few studies address this circumstance. Most of the state-of-arts only consider changes in one of these aspects. Although this `double' scenario is extremely complex and detecting anomalies in this kind of dynamic graph is very challenging, it is worth studying because these graphs are highly reflective of real network data.\n\\end{itemize}","id":"239988e3-c594-47d4-adb9-69b272543369","level":"subsection","origin_cites_number":7,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Anomaly Detection in Dynamic Graphs"]],"subsections":[],"title":"Anomaly Detection in Dynamic Graphs"},{"cite_extract_rate":1,"cites":[6296,6299,6300],"content":"Heterogeneous graphs are a specific type of graph that contain diverse types of nodes and edges.\nFor instance, Twitter can be intuitively modeled as a heterogeneous graph comprised of tweets, users, words, etc.\n\\textit{Opportunities}: To use the complex relationships between different types of nodes in heterogeneous graphs for anomaly detection, representative works, such as HGATRD~, GCAN~ and GLAN~, typically decompose a heterogeneous graph into individual graphs according to meta-paths, \\eg one with tweets and users, and another with tweets and words. They then use D(G)NNs to learn the embeddings for graph anomaly detection. Such a decomposition inherently overlooks the direct inter-relations among diverse types of nodes/edges and downgrades the effectiveness of the embeddings. A possible solution is to reveal the complex relations between different types of nodes and edges, and encode them into a unique representation for boosted detection performance.","id":"aff7afb2-0d0e-4d85-baf6-f135a679f1c2","level":"subsection","origin_cites_number":3,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Anomaly Detection in Heterogeneous Graphs"]],"subsections":[],"title":"Anomaly Detection in Heterogeneous Graphs"},{"cite_extract_rate":1,"cites":[242],"content":"The scalability of methods to high-dimensional and large-scale data is an ongoing and significant challenge to anomaly detection techniques.\nIn face of large-scale networks, such as Facebook and Twitter that contain billions of users and friendship links, the size of data in terms of both graph size and number of node attributes is extremely high.\nHowever, most of the existing works lack the ability to detect anomalies in such large-scale data because they are transductive models and need to take the whole graph as input for further analysis.\nComputation time and memory cost increase dramatically as the network scales up, and this stops existing techniques from being used on large-scale networks.\n\\textit{Opportunities}: Accordingly, there is a need for scalable graph anomaly detection techniques.\nOne possible approach would be an inductive learning scheme that first trains a detection model on part of the whole graph and then applies the model to detect anomalies in the unseen data.\nAs some inductive learning models, such as GraphSAGE~, have shown their effectiveness on link prediction and node classification in large-scale graphs, this approach is expected to provide a basis for graph anomaly detection in large-scale graphs and similar techniques can be investigated in the future.","id":"094e633e-5d3b-45d7-bca3-171ad2d8492b","level":"subsection","origin_cites_number":1,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Anomaly Detection in Large-scale Graphs"]],"subsections":[],"title":"Anomaly Detection in Large-scale Graphs"},{"cite_extract_rate":0.14285714285714202,"cites":[8994],"content":"In real-world networks, objects might form different kinds of relationships with others (\\eg user's followership and friendship on Twiter). And their attribute information might be collected from different resources, such as user's profile, historical posts.\nThis results in two types of multi-view graphs: 1) multi-graph that contains more than one type of edges between two nodes~; and 2) multi-attributed-view graph that stores node attributes in different attributed views~.\n\\textit{Opportunities}: These multi-views basically allow us to analyze real objects' characteristics from different perspectives.\nEach view also provides complementary information to other views, and they might have different significance on anomaly detection. \nFor instance, anomalies might be indistinguishable in one view but are obviously divergent from the majority in another view.\nThere are a variety of work in data mining on multi-view learning~. However, work that can accommodate multi-view graphs along with multi-view attributes on nodes for anomaly detection purposes is nascent.\nMoreover, the rich information contained in multiple views and the inconsistency among them are overlooked in these works. \nTo this end, we believe more research effort in this direction is needed. Digesting the relationships between views will be vital to their success, as two views might provide contrary/supplementary information for anomaly detection.","id":"e773dc9f-5c7d-4700-bcf9-1e7e0772172c","level":"subsection","origin_cites_number":7,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Multi-view Graph Anomaly Detection"]],"subsections":[],"title":"Multi-view Graph Anomaly Detection"},{"cite_extract_rate":0.5,"cites":[6307,6306,6271,8995],"content":"The easy accessibility of online platforms has made them convenient targets for fraudsters, attackers and other malevolent agents to carry out malicious activities.\nAlthough various anomaly detection systems have been deployed to protect benign objects, anomalies can still conceal themselves to evade detection~.\nKnown as camouflaged anomalies, these entities typically disguise themselves as regular objects.\nIf the detection techniques are not robust against such cases, \\ie if they cannot quickly and effectively adapt to the evolving behavior of evasion-seeking attackers, the anomalies are simply left to cause their damage.\n\\textit{Opportunities}: In the face of camouflage, the boundary between anomalies and regular objects is blurred, making anomalies much harder to be identified.\nWe believe extensive effort should be placed on detecting these anomalies because, as yet, very few studies have looked at handling camouflaged anomalies in graphs~.\nTo fulfill this gap, one major direction might be to jointly analyze the attributes, co-relations, such as the triadic, tetradic, or high-order relationships between objects in hypergraphs~, and other information comprised in graphs.\nBy this, anomalies that only camouflage their local structures or attributes can be identified effectively.\nEnhancing existing techniques might be another direction. \nThis involves incorporating additional detection mechanisms or function blocks particularly designed for distinguishing camouflaged anomalies with existing detection techniques.\nConsequently, these techniques will bridge most existing works and camouflaged anomaly detection.","id":"cc214041-b949-4f3c-95ae-5deb83ec1573","level":"subsection","origin_cites_number":8,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Camouflaged/Adversarial Anomaly Detection"]],"subsections":[],"title":"Camouflaged/Adversarial Anomaly Detection"},{"cite_extract_rate":0,"cites":[],"content":"Anomalies are rare, which means anomaly detection always occurs coexists with class imbalance in the training data. As deep learning models rely heavily on the training data, such imbalance pose great challenges to graph anomaly detection, and this remains a significant obstacle to deep learning techniques. \nTypically, the imbalanced class distributions will downgrade the detection techniques' ability to capture the differences between anomalies and non-anomalies. It might even cause over-fitting on the anomalous class because there are too few anomalies in the data.\nIf the detection model overlooks this critical fact and is trained improperly, detection performance will be sub-optimal.\n\\textit{Opportunities}: In fact, class imbalance has been widely explored in various research areas~.\nAdvances such as under-sampling the majority class or modifying the algorithms shed important light on solving training problems with imbalances.\nYet, contemporary graph anomaly detection methods rarely incorporate these techniques. For more effective detection techniques, biased models that pay more attention to anomalies, such as penalizing additional training losses on misclassified anomalies, would be a possible direction to circumvent the problem.\nMoreover, when adopting graph neural networks that aggregate neighboring information to the target node, like GCN or GraphSAGE, the over-smoothing between the features of connected nodes should be prevented such that the distinguishable features of the anomalies can be preserved to support anomaly detection.","id":"85e8f422-2b40-4284-aca9-eaeef926d8d1","level":"subsection","origin_cites_number":2,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Imbalanced Graph Anomaly Detection"]],"subsections":[],"title":"Imbalanced Graph Anomaly Detection"},{"cite_extract_rate":0.6666666666666661,"cites":[6308,1671,6309,6284],"content":"Graph anomaly detection has close relations with other graph mining tasks including community detection~ and node classification~, and link prediction~. \nFor a concrete example, when detecting community anomalies, community detection techniques are usually used to extract the community structures prior to anomaly detection. Meanwhile, the anomaly detection results can be used to optimize the community structure.\nSuch mutually beneficial collaborations between anomaly detection and other tasks inherently suggest an opportunity for multi-task learning that can handle diverse tasks simultaneously and share information among tasks.\n\\textit{Opportunities}: Multi-task learning provides effective machinery with which incorporate associated tasks~. Its utmost advantage is that the training signal from another task could yield complementary information to distinguish anomalies from non-anomalies. The result would be enhanced detection performance. However, very few attempts focus on this at present.\nBeyond current works, such as~ that jointly perform anomalous node detection and personalized recommendation, explorations into combining other learning tasks with graph anomaly detection are likely to emerge as a fruitful future direction.","id":"1224de5d-ffc7-41d5-bee2-9722227a3ca0","level":"subsection","origin_cites_number":6,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Multi-task Anomaly Detection"]],"subsections":[],"title":"Multi-task Anomaly Detection"},{"cite_extract_rate":0.5,"cites":[8996,6310],"content":"The interpretability of anomaly detection techniques is vital to the subsequent anomaly handling process.\nWhen applying these techniques to real applications, such as financial and insurance systems, it is essential to provide explainable and lawful evidence to support the detection results.\nHowever, most of the existing works lack the ability to provide such evidence.\nTo identify the anomalies, the most commonly used metrics are top-k rankings and simple anomaly scoring functions.\nThese metrics are flexible enough to label objects as being either an anomaly or not an anomaly, but they cannot derive solid explanations.\nMoreover, as deep learning techniques have also been criticized for their low interpretability, future works on graph anomaly detection with deep learning should pay much more attention to this~.\n\\textit{Opportunities}: To bridge this gap, integrating specially designed interpretation algorithms or mechanisms~ into the detection framework would be a possible solution, noting that this would inherently induce a higher computational cost.\nFuture works should therefore balance the cost of anomaly detection performance and interpretability.\nVisualization-based approaches, \\eg dashboards, charts, might also be feasible for showing the distinction between anomalies and non-anomalies in a human-friendly manner.\nFurther research in this direction will be successful if interpretable visualization results can be given~.","id":"d42a5aab-3ff4-4f52-8576-f381cddddc12","level":"subsection","origin_cites_number":4,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Graph Anomaly Interpretability"]],"subsections":[],"title":"Graph Anomaly Interpretability"},{"cite_extract_rate":0.36363636363636304,"cites":[6289,6287,6283,8997],"content":"Amongst existing unsupervised graph anomaly detection techniques, anomalies are mainly identified based on residual analysis~, reconstruction loss~, distance-based statistics~, density-based statistics~, graph scan statistics~, and one-class classification~.\nThe underlying intuition of these identification strategies is that anomalies have inconsistent data patterns with regular objects, and they will, therefore: 1) introduce more residual errors or be harder to reconstruct; or 2) be located in low-density areas or far away from the majority class in an anomaly-aware feature space.\nEffort toward designing novel loss functions for GNNs for anomaly detection is currently quite limited~\n\\textit{Opportunities}: Although these strategies could capture the deviating data patterns of anomalies, they also have different limitations. Specifically, the residual analysis, one-class classification and reconstruction loss strategies are sensitive to noisy training data. Noisy nodes, edges or sub-graphs also exhibit large residuals, a large distance to the origin/hypersphere center and high reconstruction losses. Meanwhile, the distance-based and density-based strategies can only be applied when anomalies and non-anomalies are well separated in lower-dimensional space. Detection performance also downgrades dramatically if the gap between anomalies and non-anomalies is not that evident.\nIt calls for extensive future efforts to break these limitations and explore new anomaly identification strategies.","id":"4eb29545-f452-443d-b7b6-8accd75e515c","level":"subsection","origin_cites_number":11,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Graph Anomaly Identification Strategies"]],"subsections":[],"title":"Graph Anomaly Identification Strategies"},{"cite_extract_rate":0,"cites":[],"content":"Systematic benchmarking is key to evaluating the performance of graph anomaly detection techniques.\nAs indicated in our analysis in Section~\\ref{sec:em}, recent studies are constantly raising attention to more comprehensive and effective benchmarking~. \nTypically, the benchmarking framework consists of benchmark datasets, baseline methods, evaluation metrics, and further analysis tools.\nWhen evaluating a techniques' performance with other baselines, the evaluation dataset and metrics become very important because the performance of each model may vary depending on the setting.\nThe shortage of public datasets and (public available) baseline methods also imposes great challenges for effective evaluation.\nAlthough one of the aims of our survey is to provide extensive materials for this purpose, like open-sourced implementations, datasets, and evaluation metrics, this work can only serve as a basis for future investigations into systematic benchmarking. We invite more efforts from the anomaly detection community toward this important case.\nCertainly, rigorous attention to designing better frameworks for bechmarking would help to disclose the advances and shortcomings of various detection techniques and essentially track a unprejudiced and accurate progress record in this field.","id":"9d5eeb0b-b689-4f33-bf46-0aa58f25c4d7","level":"subsection","origin_cites_number":4,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Systematic Benchmarking"]],"subsections":[],"title":"Systematic Benchmarking"},{"cite_extract_rate":0,"cites":[],"content":"A graph anomaly can be categorized as an anomalous node, edge, and sub-graph in a single graph or an anomalous graph in a graph database.\nThese anomalies usually coexist in real-world datasets.\nFor instance, individual fraudsters, abnormal relationships, and fraud groups exist concurrently in online social network, as shown in Fig.~\\ref{Toy}.\nMoreover, there may be different ways to define anomalies of a certain type, such as community outliers versus anomalous communities or attribute-based versus structural anomalies.\nWhen deploying detection techniques in real applications, it is expected that all types of anomalies can be identified while consuming the least resources and time.\nA straightforward approach would be to integrate independent anomalous node, edge, and sub-graph detection techniques.\nAlthough this is convenient to apply to relatively small networks, its high computational cost will surely prevent the approach from scaling to large networks, such as Facebook and Twitter, because the same graph data has to be loaded and processed more than once by different techniques.\n\\textit{Opportunities}: Unified frameworks that can detect diverse types of anomalies together~ may provide feasible solutions to bridge the gap.\nTo build such frameworks, one possible direction is to capture all the information needed by different detection techniques simultaneously so that these techniques can be applied.\nThe idea seems to be non-challenging, but in deep learning, methods of designing neural network layers and learning strategies that can fulfill this need will require extensive effort.","id":"b1a4c699-5c1b-4fe6-a282-ae8f4a92410d","level":"subsection","origin_cites_number":2,"parent_id":"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Future Directions"],["subsection","Unified Anomaly Detection Framework"]],"subsections":[],"title":"Unified Anomaly Detection Framework"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:conclusion}\nDue to the complex relationships between real-world objects and recent advances in deep learning, especially graph neural networks, graph anomaly detection with deep learning is currently at the forefront of anomaly detection.\nTo the best of our knowledge, this is the first survey to present a comprehensive review dedicated to graph anomaly detection with modern deep learning techniques.\nSpecifically, we have reviewed and categorized contemporary deep learning techniques according to the types of graph anomalies they can detect, including: (1) anomalous node detection; (2) anomalous edge detection; (3) anomalous sub-graph detection; and, finally, (4) anomalous graph detection.\nClear summarizations and comparisons between different works are given to offer a complete and thorough picture of the current work and progress of graph anomaly detection as a field.\nMoreover, to push forward future research in this area, we have provided a basis for systematic benchmarking by compiling a wide range of commonly used datasets, open-sourced implementations and synthetic dataset generation techniques.\nWe further highlight 12 potential directions for future work based on the survey results.\nIt is our firm belief that graph anomaly detection with deep learning is indeed more than a burst of temporary interest, and numerous applications from diverse domains are to surely benefit from it for the years to come.\n\\bibliographystyle{IEEEtran}\n\\bibliography{main.bib}\n\\clearpage\n\\setcounter{page}{1}\n\\appendices","id":"32ddd0ab-f741-4a6c-a1ff-380ee935d4a2","level":"section","origin_cites_number":0,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"},{"cite_extract_rate":0.31707317073170704,"cites":[6285,6281,6283,6289,6307,6279,6271,6306,6293,5645,8149,6291,6286],"content":"\\label{appendix:data-challenges}\nDue to the complexity of anomaly detection and graph data mining, adopting deep learning technologies for graph anomaly detection faces a number of challenges:\n\\textbf{Data-CH1. Ground-truth is scarce.} In most cases, there is little or no prior knowledge about the features or patterns of anomalies in real applications. Ground-truth anomalies are often identified by domain experts, and this generally cost-prohibitive. As a result, labeled ground-truth anomalies are often unavailable for analysis in a wide range of disciplines.\n\\textbf{Data-CH2. Various types of graphs.} Different graphs model different real-world data. For instance, plain graphs contain only structural information, attributed graphs contain both structural and attribute information, and heterogeneous graphs represent the complex relations between different types of objects.\nThese graphs reflect the real-world data in different forms, and graph anomalies will show different deviating patterns in different types of graphs.\n\\textbf{Data-CH3. Various types of graph anomalies.} Given a specific type of graph, graph anomalies could appear as a specific node, edge, sub-graph, or an entire graph, and each type of these anomalies is significantly different from others. This means detection methods must involve concise definitions of anomalies and be able to identify concrete clues about the deviating patterns of anomalies.\n\\textbf{Data-CH4. High dimensionality and large scale.} Representing the structure information of real-world networks usually results in high dimensional and large-scale data~ because real-world network often contain millions or billions of nodes. Graph anomaly detection techniques, hence, should be capable of handling such high dimensional and large scale data; this includes the ability to extract anomalous patterns under the constraints of execution time and feasible computing resources. \n\\textbf{Data-CH5. Interdependencies and dynamics.} The relationships between real objects reveal their interdependencies and they can no longer be treated individually for anomaly detection. That is to say, the detection techniques need to consider the deviating patterns of anomalies by assessing the pairwise, triadic, and higher relationships among objects restored in conventional graphs or hypergraphs~.\nIn addition, the dynamic nature of real-world networks makes detection problems much more challenging.\n\\textbf{Data-CH6. Class imbalance.} As anomalies are rare occurrences, only a very small proportion of the real-world data might be anomalous. This naturally introduces a critical class imbalance problem to anomaly detection because the number of normal objects is far greater than anomalies in the training data. If no further actions are taken to tackle this challenge, learning-based anomaly detection techniques might overlook the patterns of anomalies, leading to sub-optimal results.\n\\textbf{Data-CH7. Unknown and camouflage of anomalies.} In reality, knowledge about anomalies mainly stems from human expertise. There are still many unknown anomalies across different application domains, and new types of anomalies might appear in the future. Nevertheless, real-world anomalies can hide or be camouflaged as benign objects to bypass existing detection systems. In graphs, anomalies might hide themselves by connecting with many normal nodes or by mimicking their attributes. Detection methods, therefore, need to be adaptive to unknown and novel anomalies and robust to camouflaged anomalies.\nThese data-specific challenges and technical-specific challenges (discussed in Section~\\ref{sec:introdution:challenges}) are summarized in Table.~\\ref{tb:challenges}, along with the corresponding articles that aim to address them.\n\\begin{table}\n\\centering \n\\footnotesize\n\\renewcommand\\arraystretch{1}\n\\setlength{\\tabcolsep}{2.8mm}\n\\caption{Challenges and Methods.}\n\\resizebox{0.48\\textwidth}{!}{\n\\begin{tabular}{m{1.1cm}<{\\centering}|m{2cm}<{\\centering}|m{2.9cm}<{\\centering}}\n\\toprule[1 pt]\n\\textbf{Challenges} & \\textbf{Details} & \\textbf{Methods} \\\\ \n\\midrule[1 pt]\n\\multicolumn{3}{c}{\\textbf{Data-specific Challenges}} \\\\\n\\midrule[1 pt]\nData-CH1 & Ground-truth is scarce & \\\\ \n\\cline{1-3}\nData-CH2 & Various types of graphs & \\\\ \n\\cline{1-3}\nData-CH3 & Various types of graph anomalies & \\\\ \n\\cline{1-3}\nData-CH4 & High dimensionality and large scale & \\\\ \n\\cline{1-3}\nData-CH5 & Interdependencies and dynamics & \\\\ \n\\cline{1-3}\nData-CH6 & Class imbalance & \\\\ \n\\cline{1-3}\nData-CH7 & Unknown and camouflage of anomalies & \\\\ \n\\midrule[1 pt]\n\\midrule[1 pt]\n\\multicolumn{3}{c}{\\textbf{Techniques-specific Challenges}} \\\\\n\\midrule[1 pt]\nTech-CH1 & Anomaly-aware training objectives & \\\\\\cline{1-3}\nTech-CH2 & Anomaly interpretability & \\\\\\cline{1-3}\nTech-CH3 & High training cost & \\\\\\cline{1-3}\nTech-CH4 & Hyperparameter tuning & \\\\\n\\bottomrule[1 pt]\n\\end{tabular}\n\\label{tb:challenges}\n}\n\\end{table}","id":"9a4d2cc0-0208-4633-a227-f43cf17e07e4","level":"section","origin_cites_number":41,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Challenges in Graph Anomaly Detection"]],"subsections":[],"title":"Challenges in Graph Anomaly Detection"},{"cite_extract_rate":0,"cites":[],"content":"\\label{appendix:taxonomy}\nThe taxonomy of this survey is shown in Fig.~\\ref{pic:taxonomy}.","id":"12e1194f-a7ed-4e46-889d-a06a013eaeaf","level":"section","origin_cites_number":0,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","Taxonomy"]],"subsections":[],"title":"Taxonomy"},{"cite_extract_rate":0,"cites":[],"content":"\\label{appendix:node:static}\nFollowing the taxonomy of Section~\\ref{lb:node:sag}, this section reviews traditional non-deep learning techniques designed for ANOS ND on static attributed graphs, followed by techniques based on GATs, GANs, and network representation.\n\\onecolumn\n\\begin{sidewaysfigure*}[h]\n \\centering\n \\includegraphics[width=\\textwidth,keepaspectratio]{pics/taxonomy.pdf}\n \\captionof{figure}{The Taxonomy of Our Survey.}\n \\label{pic:taxonomy}\n\\end{sidewaysfigure*}\n\\twocolumn","id":"dfc7ea98-0750-495a-b3c3-8183220b6042","level":"section","origin_cites_number":0,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Static Graphs"]],"subsections":["92e0465c-2bc8-4d8d-9625-6c4f4b200390","ac5b6d2a-da5e-45c0-bd9e-1a63dbed603a","ec58be64-9072-46de-98e9-29a0e0542d63","99113f23-1403-4b3d-bae2-8db2bf051152"],"title":"ANOS ND on Static Graphs"},{"cite_extract_rate":0,"cites":[],"content":"Traditional techniques, such as statistical models, matrix factorization, and KNN, have been widely applied to extract the structural/attribute patterns of anomalous nodes for a subsequent detection process.\nAmong these, matrix factorization (MF) based techniques have shown power at capturing both topological structures and node attributes, achieving promising detection performance.\nAn early attempt at this type of method was Liu \\etal~.\nThey aimed to detect community anomalies (defined in Section~\\ref{sec:AND}) through the developed model, ALAD.\nAs shown in Fig.~\\ref{pic:ALAD}, through non-negative matrix factorization, ALAD incorporates both the graph structure $A$ and node attributes $X$ to derive community structures $W$ and their attribute distribution vectors $H$.\nWhen the matrices are decomposed, ALAD measures the normality of each node according to the attribute similarity between it and the community it belongs to.\nBy ranking the nodes' normality scores in ascending order, the top-k nodes are identified as community anomalies. \nLi \\etal~ approached ANOS ND from a different perspective using residual analysis.\nAs assumed, anomalies will lead to larger attribute reconstruction residual errors because they do not conform to the attribute patterns of the majorities.\nAccordingly, the proposed model, Radar, learns the residual errors $R$, as shown in Fig.~\\ref{pic:Radar}, by factorizing the node attributes $X$. To incorporate the structural information for obtaining these errors, Radar puts explicit restrictions on the learned residuals such that directly linked nodes will have similar residual patterns in $R$ (known as the homophily effect).\nFinally, the top-k nodes with larger norms in $R$ are identified as anomalies.\nAlthough the homophily hypothesis provides strong support for exploiting structural information, it might not always be hold.\nIn fact, real objects may have distinct attributes from their connected neighbors and it is non-trivial to regulate all connected objects share similar values in each dimension in the feature space. By this, Peng \\etal~ indicated that there are structurally irrelevant node attributes that do not satisfy the homophily hypothesis.\nIndeed, these structurally irrelevant node attributes would have adverse effects on anomaly detection techniques that are developed based on this hypothesis. \nTo tackle this problem, their developed model, ANOMALOUS, uses CUR~ decomposition to select attributes that are closely related to the network structure and then spot anomalies through residual analysis following Radar (as shown in Fig.~\\ref{pic:ANOMALOUS}). \n\\begin{figure}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n \\centering\n \\subfigure[The Framework of ALAD~]{\n \\label{pic:ALAD} \n \\includegraphics[width=0.48\\textwidth]{pics/Node/ALAD.pdf}}\n \\subfigure[The Framework of Radar~]{\n \\label{pic:Radar} \n \\includegraphics[width=0.48\\textwidth]{pics/Node/Radar.pdf}}\n \\subfigure[The Framework of ANOMALOUS~]{\n \\label{pic:ANOMALOUS}\n \\includegraphics[width=0.48\\textwidth]{pics/Node/Anomalous.pdf}}\n \\caption{ANOS ND on attributed graphs -- Matrix Factorization based approaches. The three representative MF techniques adopt different decomposition strategies to fuse the graph structural information and node attributes. Anomalous nodes are then spotted through scoring functions or residual analysis.}\n \\label{pic:MFtechniques}\n\\end{figure}\nBeyond matrix factorization, linear regression models have also been designed to train anomaly classifiers given labeled training data.\nA representative work is Wu \\etal~.\nTheir supervised model, SGASD, has yielded encouraging results with identifying social spammers using the social network structure, the content information in social media and user labels.\nThese non-deep learning techniques are able to capture valuable information from graph topology and node attributes, but their application and generalizability to real-world networks (which are usually in large-scale) is strictly limited due to the high computational cost of matrix decomposition operations and regression models.","id":"92e0465c-2bc8-4d8d-9625-6c4f4b200390","level":"subsection","origin_cites_number":5,"parent_id":"dfc7ea98-0750-495a-b3c3-8183220b6042","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Static Graphs"],["subsection","Traditional Non-Deep Learning Techniques"]],"subsections":[],"title":"Traditional Non-Deep Learning Techniques"},{"cite_extract_rate":1,"cites":[6291,180,6289],"content":"\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n \\centerline{\\includegraphics[width=0.98\\textwidth]{pics/Node/GAT.pdf}}\n \\caption{ANOS ND on attributed graphs -- GAT based approaches. Given the input graph, these techniques employ graph attention neural network to learn node embeddings. The unsupervised technique, AnomalyDAE~, scores each node based on the reconstruction loss and mark the top-k nodes as anomalies, while the semi-supervised technique, SemiGNN~, trains a classifier to predict node labels.}\n \\label{pic:GAT}\n\\end{figure*}\nAlthough GCN provides an effective solution to incorporating graph structure with node attributes for ANOS ND (reviewed in Section~\\ref{sec:AND:GCN}), its ability to capture the most relevant information from neighboring is subpar.\nThis is due to the simple convolution operation that aggregates neighbor information equally to the target node.\nRecently, graph attention mechanism (GAT)~ is employed to replace the traditional graph convolution.\nFor instance, Fan \\etal~ applied graph attention neural network to encode the network structure information (structure encoding).\nThe method, AnomalyDAE, also adopts a separate attribute autoencoder to embed the node attributes (attribute encoding).\nThrough an unsupervised encoding-decoding process, each node is ranked according to its corresponding reconstruction loss, and the top-k nodes introducing the greatest losses are identified as anomalies.\nSpecifically, the attribute decoding process takes both node embeddings learned through the structure and attribute encoding processes to reconstruct node attributes, as shown in Fig.~\\ref{pic:GAT}, while the graph topology is reconstructed only using the embeddings output by the GAT.\nTo acquire better reconstruction results, AnomalyDAE is trained to minimize the overall loss function, denoted as:\n\\begin{equation} \\label{eq:anomalydae}\n\\begin{split}\n \\mathcal{L}_{AnomalyDAE} = & \\alpha ||(A - \\hat{A})\\odot \\bm{\\theta}||_{2}^{2} + \\\\\n &(1-\\alpha)||(X - \\hat{X})\\odot \\bm{\\eta}||_{2}^{2},\n\\end{split}\n\\end{equation}\nwhere $\\alpha$ is the coefficient, $A$ and $X$ is the input adjacency matrix and attribute matrix, $\\hat{A}$ and $\\hat{X}$ are the reconstructed matrices. Each $\\theta_{i,j} \\in \\bm{\\theta}$ and $\\eta_{i,j} \\in \\bm{\\eta}$ is 1, if the corresponding element $A_{ij}$ and $X_{ij}$ equals 0, otherwise, their values are defined by hyperparameters greater than 1.\nAnother decent work is SemiGNN~, in which Wang \\etal proposed a semi-supervised attention-based graph neural network for detecting fraudulent users in online payment platforms.\nThis work further explores user information collected from various sources (\\eg transaction information and user profiles), and represents real-networks as multi-view graphs.\nEach view in the graph is modeled to reflect the relationship between users or the correlation between user attributes.\nFor anomaly detection, SemiGNN first generates node embedding $h_u^v$ from each view $v$ by aggregating neighbor information through a node-level attention mechanism.\nIt then employs view-level attention to aggregate node embeddings from each view and generates a unified representation $a_u$ for each node.\nLastly, the class of each node is predicted through a softmax classifier.\nIndeed, Wang \\etal designed a supervised classification loss and an unsupervised graph reconstruction loss to jointly optimize the model by fully utilizing labeled and unlabeled data.\nThe classification loss can be denoted as:\n\\begin{equation} \n \\mathcal{L}_{sup} = - \\frac{1}{|U_L|}\\sum_{u \\in U_L} \\sum_{i=1}^{k}I(y_u=i)\\log\\frac{\\exp(a_u\\cdot\\theta_i)}{\\sum_{j=1}^{k}\\exp(a_u\\cdot\\theta_j)},\n\\end{equation}\nwhere $U_L$ is the labeled user set and its size is $|U_L|$, $I(\\cdot)$ is an indicator function, $k$ is the number of labels to be predicted (in most cases, the label is either anomalies or non-anomalies, and $k=2$), and $\\theta$ represents the trainable variables.\nMeanwhile, the unsupervised loss encourages unlabeled nodes (users) that can be reached by labeled nodes through random walks to obtain similar representations and vice versa.\nThis is achieved by negative sampling (unlabeled nodes that cannot be reached by random walks are negative samples) and the loss can be formulated as:\n\\begin{equation}\n \\begin{split}\n \\mathcal{L}_{unsup} & = \\sum_{u\\in U}\\sum_{v\\in N_u \\cup Neg_{u}} -\\log(\\sigma(a_{u}^{T}a_{v})) \\\\\n &- 3\\cdot E_{q\\sim P_{neg}(u)}\\log(\\sigma(a_{u}^{T}a_{q})),\n \\end{split}\n\\end{equation}\nwhere $U$ denotes the user set, $N_u$ denotes the neighbor set of $u$, $Neg_{u}$ represents negative samples, $P$ is the sampling distribution, and $\\sigma(\\cdot)$ is the sigmoid function. The total loss takes the sum of them and is formulated as:\n\\begin{equation} \n \\mathcal{L}_{SemiGNN} = \\alpha \\mathcal{L}_{sup} + (1- \\alpha) \\mathcal{L}_{unsup} + \\lambda \\mathcal{L}_{reg},\n\\end{equation}\nwhere $\\alpha$ is a balancing parameter and $\\mathcal{L}_{reg}$ regularizes all trainable variables.","id":"ac5b6d2a-da5e-45c0-bd9e-1a63dbed603a","level":"subsection","origin_cites_number":3,"parent_id":"dfc7ea98-0750-495a-b3c3-8183220b6042","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Static Graphs"],["subsection","GAT Based Techniques"]],"subsections":[],"title":"GAT Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"\\begin{figure*}[!t]\n\\setlength{\\belowcaptionskip}{-0.25cm}\n \\centerline{\\includegraphics[width=0.98\\textwidth]{pics/Node/GAN.pdf}}\n \\caption{ANOS ND on static graphs -- GAN based approaches. The generator, G, generates potential anomalies by sampling noisy data from a prior distribution to fool the discriminator. The discriminator, D, distinguishes the generated anomalies and normal nodes. The scoring function then quantifies the anomaly score of each node according to D's output.}\n \\label{pic:GAN}\n\\end{figure*}\nBecause GAN is effective at capturing anomalous/regular data distributions (as reviewed in Section~\\ref{sec:ND:GAN}), Ding \\etal~ used GAN in their developed model, AEGIS, for improved anomaly discriminability on unseen data.\nAs shown in Fig.~\\ref{pic:GAN}, this model first generates node embeddings through a GNN from the input attributed graph, and then a generator and a discriminator are trained to identify anomalies. In the first phase, anomalous nodes and regular nodes are mapped to distinctive areas in the embedding space such that the GAN is able to learn a boundary between them. Accordingly, Ding \\etal built an autoencoder network with graph differentiative layers to capture the attribute difference between each node and its $k$-th neighbors. Such difference information enables anomalies to be distinguished easily. The embeddings are encoded as follows:\n\\begin{equation} \n h_{i}^{l} = \\sum_{k=1}^{K} \\beta_{i}^{k} \\sigma \\left( W_1h_{i}^{l-1} + \\sum_{j\\in N_{k}(i)}\\alpha_{ij}W_{2}\\Delta_{ij}^{l-1}\\right),\n\\end{equation}\nwhere $h_{i}^{l}$ is embedded features of node $i$ through the $l$-th layer, $\\beta_{i}^{k}$ is the attention for each hop, $N_{k}(i)$ is the set of $k$-th order neighbors, $\\alpha_{ij}$ is the attention for each neighbor, and $\\Delta_{ij}^{l-1}$ is the difference between $i$ and $j$, $W_1$ and $W_2$ are the trainable variables.\nThe autoencoder is fine-tuned until the node attributes can be best reconstructed using the learned embeddings, after which the GAN is trained.\nIn the second phase, the generator follows a prior distribution $p(\\tilde{z})$ to generate anomalies by sampling noisy data, while the discriminator struggles to distinguish between the embeddings of the normal nodes and those of the generated anomalies. The training process is formulated as a mini-max game between the generator and discriminator as follows:\n\\begin{equation}\n \\min_{G}\\max_{D}\\mathbb{E}_{z\\sim Z}[\\log(D(z))] + \\mathbb{E}_{\\tilde{z}\\sim \\tilde{Z}}[\\log(1-D(G(\\tilde{z})))],\n\\end{equation}\nwhere $z$ is the node embeddings, and $\\tilde{z}$ are the generated anomalies.\nAfter training, AEGIS directly learns embedding $z_u$ for a test node $u$, and quantifies its anomaly score with regard to the discriminator's outputs, \\ie the possibility that node is normal. The scoring function is formulated as:\n\\begin{equation} \n Ascore(u) = 1 - D(z_u),\n\\end{equation}\nand the top-k nodes are deemed to be anomalous.","id":"ec58be64-9072-46de-98e9-29a0e0542d63","level":"subsection","origin_cites_number":1,"parent_id":"dfc7ea98-0750-495a-b3c3-8183220b6042","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Static Graphs"],["subsection","GAN Based Techniques"]],"subsections":[],"title":"GAN Based Techniques"},{"cite_extract_rate":0.6000000000000001,"cites":[6286,6298,6271],"content":"With network representation, graphs are first encoded into a vector space before the anomalies detection procedure takes place.\nAs outlined in Section~\\ref{sec:spg:nr}, numerous studies on ANOS ND in attributed graphs have exploited deep network representation techniques.\nFor instance, Zhang \\etal~ detected abnormal nodes that have attributes significantly deviating from their neighbors through a 3-layer neural network, REMAD, and residual analysis.\nThey explicitly divide the original node attribute matrix into a residual attribute matrix $R$ that captures the abnormal characters of anomalies and a structurally relevant attribute matrix $\\hat{X}$ for network representation learning. Both matrices are jointly updated throughout the representation learning process so nearby nodes are encouraged to have similar representations.\nSpecifically, these node embeddings are generated by aggregating neighbor information with each node's own attributes, formulated as:\n\\begin{equation} \n h_{i}^{l} = \\sigma \\left( W^l \\cdot \\text{CONCAT}\\{h_{i}^{l-1},h_{N_i}^{l}\\} + b^2\\right),\n\\end{equation}\nwhere $h_{i}^{l}$ is node $i$'s representation generated by the $l$-th layer ($h_{i}^{0} = \\hat{X}$), $N_i$ contains $i$'s neighbors, $\\sigma()$ is the activation function, $W^{l}$ and $b$ are the trainable variables. Finally, the residual matrix $R$ will contain the abnormal information of each node and the top-k nodes with the largest norms are considered anomalies.\nGiven partial node labels, Liang \\etal~ developed a semi-supervised representation model, SEANO, that incorporates graph structure, node attributes and label information.\nSimilar to REMAD, SEANO also aggregates neighbor information to center nodes, and the node representations are obtained through an embedding layer, formulated as:\n\\begin{equation} \\label{SEANO:embedding}\n z_{i} = \\lambda_{i}h^{l_{1}}(x_{i}) + (1-\\lambda_{i})h^{l_{1}}(\\Bar{x}_{N_{i}}),\n\\end{equation} \nwhere $z_{i}$ is $i$'s representation, $\\lambda_{i}$ is a trainable variable that identifies the weight of $i$'s own attributes ($x_{i}$), $\\Bar{x}_{N_{i}}$ is the average of node $i$'s neighbors' representations, and the function $h^{k}(x_{i}) = \\phi(W^kh^{k-1}(x_i)+b^k)$ maps original node attributes into lower dimensional vectors.\nThen, a supervised component, which takes the representations as input, predicts node labels through a softmax classifier, and an unsupervised component is trained to reconstruct node contexts (node sequences).\nThe context of each node is not only generated through random walks on the graph but also from the labeled nodes that belong to the same class. After training, SEANO interprets $\\lambda_{i}$ as the normality score of node $i$ and the top-k nodes with the highest scores are classed as anomalies.\nLearning node representations via aggregating neighbor information has proven effective for capturing comprehensive information from graph structure and node attributes. \nBut, Liu \\etal~ demonstrated such an approach can help anomalies aggregate features from regular nodes, making them look normal and leading to sub-optimal detection performance. \nThey identified three concrete issues that should be considered when applying aggregation operations for anomaly detection: \n1) Anomalies are rare objects in a network. Hence, directly aggregating neighborhood information will smooth the difference between the anomalies and normal instances, blurring boundaries between them. \n2) Directly connected nodes have distinctive features, and the assumption that connected nodes share similar features, which serves as the basis for feature aggregation, no longer holds in this scenario.\n3) Real objects also form multiple types of relations with others, which means aggregation results for different types of relations will be distinctive.\nWith regard to these concerns, their proposed method, GraphConsis, follows a sampling strategy to avoid potential anomalous neighbors when aggregating node features.\nThis method also adopts an attention mechanism to aggregate neighbor information following different links.\nThe learned node representations, therefore, are more robust to anomalies. As such, GraphConsis takes them as input to train a classifier for predicting labels.\nDou \\etal~ further considered camouflage behaviors of fraudsters in their proposed model CARE-GNN to enhance detection performance.\nAs specified, the camouflages can be categorized as either feature camouflage or relation camouflage.\nRespectively, anomalies either adjust their feature information or form connections with many benign objects to gloss over suspicious information.\nHence, directly employing aggregation will overlook the camouflages and smooth the abnormal patterns of anomalies, eliminating the distinctions between anomalies and normal objects. \nTo alleviate over-smoothness, CARE-GNN also adopts a neighbor sampling strategy, as is the case with GraphConsis, to filter camouflaged anomalies and explores different types of relations formed between users.\nSpecifically, under each relation, Dou \\etal employed a MLP to predict node labels using their features and measure the similarity ($l1$ distance) between each node and its neighbors according to the MLP's output.\nThen, the top-k most similar neighbors are selected for feature aggregation, and CARE-GNN generates each node's representation through a combination of latent representations that are learned under different relations.\nA classifier is eventually trained using the representations to predict the node labels.\nAs can be seen, the performance of these network representation based techniques is decided by their training objectives/loss functions.\nEnhanced detection performance is probable if the loss function is able to separate normal nodes from abnormal nodes reasonably well. \nMotivated by this, a more recent work in~ emphasizes the importance of anomaly-aware loss functions.\nIn order to adjust margins for the anomalies, the authors proposed a novel loss function to guide the representation learning process.\nSpecifically, this loss function is designed to find the relative scales between the margins of outlier nodes and normal nodes.\nAn MLP-based classifier is finally trained using the node representations generated by the anomaly-aware loss-guided GNNs and node labels. \nFor unseen nodes, the classifier will label them upon their representations.","id":"99113f23-1403-4b3d-bae2-8db2bf051152","level":"subsection","origin_cites_number":5,"parent_id":"dfc7ea98-0750-495a-b3c3-8183220b6042","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Static Graphs"],["subsection","Network Representation Based Techniques"]],"subsections":[],"title":"Network Representation Based Techniques"},{"cite_extract_rate":0,"cites":[],"content":"\\label{appendix:node:dynamic}\nTo detect anomalous nodes in dynamic plain graphs, traditional non-deep learning techniques rely heavily on modeling the structural evolving patterns of nodes.\nRepresentative works like~ and~ assume that the evolutionary behaviors of regular nodes' (\\ie generate or remove connections with others) usually follow stable patterns, and their changes introduce less impact on the graph structure compared to anomalies.\nSpecifically, in~, Wang \\etal proposed a novel link prediction method to fit the evolutionary patterns of most nodes such that anomalies can be identified because their observations significantly conflict with the prediction results.\nThey further quantified the impact of anomalous behaviors by assessing the perturbation imposed on the graph adjacency matrix.\nOther traditional works also exploit node/edge attributes and their changes. For examples, Teng \\etal~ took node and edge attributes as two different views to describe each node. By encoding both types of information into a shared latent space, their proposed model learns a hypersphere from historical records. When new observations of existing nodes are given, the model distinguishes between the benign and the anomalies according to their distances to the hypersphere centroid. Points lying outside the hypersphere are spotted as anomalies.\nUnlike embedding techniques, Nguyen \\etal~ proposed a non-parametric method to detect anomalous users, tweets, hashtags, and links on social platforms. Specifically, they modeled social platforms as heterogeneous social graphs such that the affluent relationships between users, tweets, hashtags and links were effectively captured. Through extensive analysis of the features, such as user registration information, keywords in tweets, the linguistic style of links and the popularity scores of hashtags, anomalous objects are spotted based on their deviating features. This work also uses relationships between individual objects as well as the detected anomalies, and detects groups of abnormal objects.\n\\begin{figure}[!t]\n \\centerline{\\includegraphics[width=0.48\\textwidth]{pics/Node/NetWalk.pdf}}\n \\caption{The Framework of NetWalk~. The input dynamic graph is modeled as an initial graph with an incoming edge stream. Given the initial graph, a trunk of node sequences is generated using random walks on the graph and the deep autoencoder encodes each node into an embedding space following process (1). The node sequences and node embeddings are updated based on the incoming edge stream following process (2). Finally, NetWalk assigns an anomaly score to each node according to its distance to the cluster centroid in the embedding space.}\n \\label{pic:netwalk}\n\\end{figure}","id":"a37030ed-a13e-4c9b-bdf9-32f760d9d73a","level":"section","origin_cites_number":5,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ND on Dynamic Graphs With Traditional Non-Deep Learning Techniques"]],"subsections":[],"title":"ANOS ND on Dynamic Graphs With Traditional Non-Deep Learning Techniques"},{"cite_extract_rate":0.25,"cites":[6293],"content":"\\label{appendix:edge}\nTraditional non-deep learning based approaches mainly focus on using temporal signals (\\eg changes in graph structure), and applying specially designed statistical metrics to detect anomalous edges on dynamic graphs~. \nAs a concrete example, Eswaran and Faloutsos~ modeled a dynamic graph as a stream of edges and exploited the graph structure as well as the structure evolving patterns. They identified two signs of anomalous edges: 1) connecting regions of the graph that were disconnected; and 2) connections that appear in bursts. For incoming edges, their model assigns anomaly scores to each edge, and the top-k edges with highest scores are anomalies. Another most recent work by Chang \\etal~ proposed a novel frequency factorization algorithm, aiming to spot anomalous incoming edges based on their likelihood of observed frequency. This method merges the advantages of both probabilistic models and matrix factorization for capturing both temporal and structural changes of nodes, and as reported, it only requires constant memory to handle edge streams.","id":"ef880176-a742-4e90-bb20-77c1c0e4b593","level":"section","origin_cites_number":4,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS ED With Traditional Non-Deep Learning Techniques"]],"subsections":[],"title":"ANOS ED With Traditional Non-Deep Learning Techniques"},{"cite_extract_rate":0,"cites":[],"content":"\\label{appendix:subgraph}","id":"6dcc483a-2e1d-4f81-abe8-97ae152ee771","level":"section","origin_cites_number":0,"parent_id":"5020eb09-4f12-4d1a-8dae-076499c8173a","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS SGD With Traditional Non-Deep Learning Techniques"]],"subsections":["a68d74d8-bc4c-4876-9c25-1421d3078bb4","ab11e3f8-fc96-4814-ac49-21cd01dfc491"],"title":"ANOS SGD With Traditional Non-Deep Learning Techniques"},{"cite_extract_rate":0.5,"cites":[6277,6297],"content":"\\label{appendix:subgraph:static}\nOne motivation of ANOS SGD in static graphs is that anomalous sub-graphs often exhibit significantly different attribute distributions.\nTherefore, traditional non-deep learning techniques, such as gAnomaly~, AMEN~, and SLICENDICE~, focus on modeling the attribute distributions and measuring the normality of sub-graphs.\nAnother line of investigation is graph residual analysis. The rich attribute information contained in real-world networks provides insight into the relationships formed between objects. Thus, the motivation behind several studies to spot anomalous sub-graphs has been to measure the residual between the expected structures and observed structures~.","id":"a68d74d8-bc4c-4876-9c25-1421d3078bb4","level":"subsection","origin_cites_number":4,"parent_id":"6dcc483a-2e1d-4f81-abe8-97ae152ee771","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS SGD With Traditional Non-Deep Learning Techniques"],["subsection","ANOS SGD on Static Graphs"]],"subsections":[],"title":"ANOS SGD on Static Graphs"},{"cite_extract_rate":0.18181818181818102,"cites":[6272,8997],"content":"\\label{appendix:subgraph:dynamic}\nDevising metrics for ANOS SGD has been the subject of many traditional works. \nFor instance, Chen \\etal~ introduced six metrics to identify community-based anomalies, namely: grown community, shrunken community, merged community, split community, born community and vanished community.\nAlthough these hand-crafted features or statistical patterns well fit some particular types of existing anomalies, their abilities to detect unseen and camouflage anomalies are limited and applying them directly might introduce high false negative rate, which is not optimal for applications like financial security. \nOther works, such as SPOTLIGHT by Eswaran \\etal~ and another by Liu \\etal~, explore sudden changes in dynamic graphs and identify anomalous sub-graphs that are related to such changes.\nMotivated by the phenomena that social spam and fraud groups often form dense temporal sub-graphs in online social networks, plenty of works, including,~, use manually-extracted features and spot anomalous dense sub-graphs that have evolved significantly different from the reset of the graph. \nIn addition to these studies, a large number of works discuss uses of various graph scan statistics for anomalous sub-graph detection, such as the Kulldorff statistic~, Poisson statistic~, elevated mean scan statistic~ and Berk-Jones statistic~. Specifically, Shao \\etal~ proposed a non-parametric method to detect anomalous sub-graphs in dynamic graphs where the network structure is constant, but the node attributes change overtime. This approach measures the anomalous score of each sub-graph with regard to the p-values of nodes it comprises. Sub-graphs with higher scores are more anomalous. Another work, GBGP~, instead, adopts the elevated mean scan statistic to identify nodes that might form anomalous sub-graphs and detects anomalous groups that follow predefined irregular structures.\n\\end{document}","id":"ab11e3f8-fc96-4814-ac49-21cd01dfc491","level":"subsection","origin_cites_number":11,"parent_id":"6dcc483a-2e1d-4f81-abe8-97ae152ee771","prefix_titles":[["title","A Comprehensive Survey on\\\\ Graph Anomaly Detection with Deep Learning"],["section","ANOS SGD With Traditional Non-Deep Learning Techniques"],["subsection","ANOS SGD on Dynamic Graphs"]],"subsections":[],"title":"ANOS SGD on Dynamic Graphs"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ]\n ],\n \"subsections\": [\n \"ecbafaf3-1565-4746-bf9c-a9182b235789\",\n \"8ea06e54-3d89-4768-967e-6b68c2d22f53\",\n \"59eed45d-0f6f-4983-a1d0-48c2fb48a606\",\n \"12f2fc41-29d6-4189-b62f-ff980f20a825\",\n \"678bd7af-5bf2-44f6-ae32-df413829eb80\",\n \"78c71045-f884-4e7f-b2a8-d7769a5bc16b\",\n \"5da2d684-654a-4048-95ec-2ae406b319b2\",\n \"9ba15010-a3db-4588-8e7f-56a7def289f5\",\n \"ae3bbf88-f1cb-408b-8ded-ef8445178dfa\",\n \"8147e459-f864-454f-86f6-52d0b0cd6b52\",\n \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"32ddd0ab-f741-4a6c-a1ff-380ee935d4a2\",\n \"9a4d2cc0-0208-4633-a227-f43cf17e07e4\",\n \"12e1194f-a7ed-4e46-889d-a06a013eaeaf\",\n \"dfc7ea98-0750-495a-b3c3-8183220b6042\",\n \"a37030ed-a13e-4c9b-bdf9-32f760d9d73a\",\n \"ef880176-a742-4e90-bb20-77c1c0e4b593\",\n \"6dcc483a-2e1d-4f81-abe8-97ae152ee771\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.24193548387096703,\n \"cites\": [\n 6272,\n 3207,\n 1671,\n 553,\n 3344,\n 6271,\n 6277,\n 6275,\n 217,\n 1662,\n 6274,\n 6276,\n 6270,\n 6273,\n 6278\n ],\n \"content\": \"\\\\label{sec:introduction}}\\n\\\\IEEEPARstart{A}{nomalies} were first defined by Grubbs in 1969~ as \\\\textit{``one that appears to deviate markedly from other members of the sample in which it occurs''} and the studies on anomaly detection were initiated by the statistics community in the 19th century.\\nTo us, anomalies might appear as social spammers or misinformation in social media; fraudsters, bot users or sexual predators in social networks; network intruders or malware in computer networks and broken devices or malfunctioning blocks in industry systems, and they often introduce huge damage to the real-world systems they appear in.\\nAccording to FBI's 2014 Internet Crime Report\\\\footnote{https://www.fbi.gov/file-repository/2014\\\\_ic3report.pdf/view}, the financial loss due to crime on social media reached more than \\\\$60 million in the second half of the year alone and a more up-to-date report\\\\footnote{https://www.zdnet.com/article/online-fake-news-costing-us-78-billion-globally-each-year/} indicates that the global economic cost of online fake news reached around \\\\$78 billion a year in 2020.\\nIn computer science, the research on anomaly detection dates back to the 1980s, and detecting anomalies on graph data has been an important data mining paradigm since the beginning.\\nHowever, the extensive presence of connections between real-world objects and advances in graph data mining in the last decade have revolutionized our understanding of the graph anomaly detection problems such that this research field has received a dramatic increase in interest over the past five years.\\nOne of the most significant changes is that graph anomaly detection has evolved from relying heavily on human experts' domain knowledge into machine learning techniques that eliminate human intervention, and more recently, to various deep learning technologies.\\nThese deep learning techniques are not only capable of identifying potential anomalies in graphs far more accurately than ever before, but they can also do so in real-time.\\n\\\\begin{figure}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.5cm}\\n\\\\centering\\n\\\\subfigure[Conventional Anomaly Detection]{\\n\\\\label{Fig1a}\\n\\\\includegraphics[scale=0.45]{pics/Introduction/Conventional_AD.pdf}}\\n\\\\subfigure[Graph Anomaly Detection]{\\n\\\\label{Fig1b}\\n\\\\includegraphics[scale=0.45]{pics/Introduction/Toy_Example.pdf}} \\n\\\\caption{Toy Examples of Conventional Anomaly Detection and Graph Anomaly Detection. Apart from anomalies shown in (b), graph anomaly detection also identifies graph-level anomalies, detailed in Sections~\\\\ref{sec:anosgd:db} and~\\\\ref{sec:anosgd:dynamic}.} \\n\\\\label{Toy}\\n\\\\end{figure}\\nFor our purposes today, anomalies, which are also known as outliers, exceptions, peculiarities, rarities, novelties, etc., in different application fields, refer to abnormal objects that are significantly different from the standard, normal, or expected. \\nAlthough these objects rarely occur in real-world, they contain critical information to support downstream applications. For example, the behaviors of fraudsters provide evidences for anti-fraud detection and abnormal network traffics reveal signals for network intrusion protection. Anomalies, in many cases, may also have real and adverse impacts, for instance, fake news in social media can create panic and chaos with misleading beliefs~, untrustworthy reviews in online review systems can affect customers' shopping choices~, network intrusions might leak private personal information to hackers~, and financial frauds can cause huge damage to economic systems~.\\nAnomaly detection is the data mining process that aims to identify the unusual patterns that deviate from the majorities in a dataset~. In order to detect anomalies, conventional techniques typically represent real-world objects as feature vectors (\\\\eg news in social media are represented as bag-of-words~, and images in web pages are represented as color histograms~), and then detect outlying data points in the vector space~, as shown in Fig.~\\\\ref{Fig1a}. \\nAlthough these techniques have shown power in locating deviating data points under tabulated data format, they inherently discard the complex relationships between objects~.\\nYet, in reality, many objects have rich relationships with each other, which can provide valuable complementary information for anomaly detection.\\nTake online social networks as an example, fake users can be created using valid information from normal users or they can camouflage themselves by mimicking benign users' attributes~. \\nIn such situations, fake users and benign users would have near-identical features, and conventional anomaly detection techniques might not be able to identify them using feature information only. \\nMeanwhile, fake users always build relationships with a large number of benign users to increase their reputation and influence so they can get unexpected benefits, whereas benign users rarely exhibit such activities~.\\nHence, these dense and unexpected connections formed by fake users denote their deviations to the benigns and more comprehensive detection techniques should take these structural information into account to pinpoint the deviating patterns of anomalies.\\nTo represent the structural information, \\\\textit{Graphs}, in which nodes/vertices denote real objects, and the edges denote their relationships, have been prevalently used in a range of application fields~, including social activities, e-commerce, biology, academia and communication. With the structural information contained in graphs, detecting anomalies in graphs raises a more complex anomaly detection problem in non-Euclidean space - graph anomaly detection (GAD) that aims to identify anomalous graph objects (\\\\ie nodes, edges or sub-graphs) in a single graph as well as anomalous graphs among a set/database of graphs~.\\nAs a toy example shown in Fig.~\\\\ref{Fig1b}, given an online social network, graph anomaly detection aims to identify anomalous nodes (\\\\ie malicious users), anomalous edges (\\\\ie abnormal relations) and anomalous sub-graphs (\\\\ie malicious user groups).\\nBut, because the copious types of graph anomalies cannot be directly represented in Euclidean feature space, it is not feasible to directly apply traditional anomaly detection techniques to graph anomaly detection, and researchers have intensified their efforts to GAD recently.\\nAmongst earlier works in this area, the detection methods relied heavily on handcrafted feature engineering or statistical models built by domain experts~.\\nThis inherently limits these techniques' capability to detect unknown anomalies, and the exercise tended to be very labor-intensive. \\nMany machine learning techniques, such as matrix factorization~ and SVM~, have also been applied to detect graph anomalies.\\nHowever, real-world networks often contain millions of nodes and edges that result in extremely high dimensional and large-scale data, and these techniques do not easily scale up to such data efficiently.\\nPractically, they exhibit high computational overhead in both the storage and execution time~.\\nThese general challenges associated with graph data are significant for the detection techniques, and we categorize them as data-specific challenges (Data-CHs) in this survey.\\nA summary of them is provided in Appendix~\\\\ref{appendix:data-challenges}.\\n\\\\begin{table*}[!h]\\n\\\\centering \\n\\\\footnotesize\\n\\\\renewcommand\\\\arraystretch{1}\\n\\\\setlength{\\\\tabcolsep}{2.8mm}{\\n\\\\caption{A Comparison Between Existing Surveys on Anomaly Detection. We mark edge, sub-graph and graph detection as \\\\includegraphics[scale=0.2]{pics/Introduction/4.pdf} in our survey because we review more deep learning based works than any previous surveys.} \\n\\\\resizebox{0.96\\\\textwidth}{!}{\\n\\\\begin{tabular}{l|c|c|c|c|c|c|c|c|c|c}\\n\\\\toprule[1 pt]\\n\\\\multirow{2}{*}{\\\\textbf{Surveys}} & \\\\multirow{2}{*}{\\\\textbf{AD}} & \\\\multirow{2}{*}{\\\\textbf{DAD}} & \\\\multirow{2}{*}{\\\\textbf{GAD}} \\n& \\\\multicolumn{4}{c|}{\\\\textbf{GADL}} & \\\\multirow{2}{*}{\\\\textbf{Source Code}} & \\\\multicolumn{2}{c}{\\\\textbf{Dataset}} \\\\\\\\ \\\\cline{5-8} \\\\cline{10-11}\\n& & & &\\\\textbf{Node} &\\\\textbf{Edge} & \\\\textbf{Sub-graph} & \\\\textbf{Graph} & &\\\\textbf{Real-world} &\\\\textbf{Synthetic} \\\\\\\\ \\n\\\\midrule[1 pt]\\nOur Survey & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf}& \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} \\\\\\\\\\n\\\\midrule[1 pt]\\nChandola \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - & - & - \\\\\\\\\\nBoukerche \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & - & - & - & - & - & - & - & - \\\\\\\\\\nBulusu \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - & - \\\\\\\\\\nThudumu \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - \\\\\\\\\\n\\\\midrule[1 pt]\\nPang \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - \\\\\\\\\\nChalapathy and Chawla~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - \\\\\\\\\\n\\\\midrule[1 pt]\\nAkoglu \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\\\\\\nRanshous \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & -& - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - \\\\\\\\\\nJennifer and Kumar~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\\\\\\n\\\\midrule[1 pt]\\nEltanbouly \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - \\\\\\\\\\nFernandes \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\\\\\\nKwon \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & - & - & - & - & - & - & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - \\\\\\\\\\nGogoi \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - & - \\\\\\\\\\n\\\\midrule[1 pt]\\nSavage \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\\\\\\nYu \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & - & - & - & - & - & - \\\\\\\\\\nHunkelmann \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & - & \\\\includegraphics[scale=0.15]{pics/Introduction/1.pdf} & - & - & - & - & - & - & - \\\\\\\\\\nPourhabibi \\\\etal~ & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/4.pdf} & \\\\includegraphics[scale=0.15]{pics/Introduction/2.pdf} & - & - & - & - & - & - \\\\\\\\\\n\\\\bottomrule[1 pt]\\n\\\\multicolumn{11}{l}{* AD: Anomaly Detection, DAD: Anomaly Detection with Deep Learning, GAD: Graph Anomaly Detection.} \\\\\\\\\\n\\\\multicolumn{11}{l}{* GADL: Graph Anomaly Detection with Deep Learning.}\\\\\\\\\\n\\\\multicolumn{11}{l}{* -: not included, \\\\includegraphics[scale=0.2]{pics/Introduction/1.pdf} (1-2 references included), \\\\includegraphics[scale=0.2]{pics/Introduction/2.pdf} (3-10 references included), \\\\includegraphics[scale=0.2]{pics/Introduction/4.pdf} (10+ references included).}\\n\\\\end{tabular}\\n\\\\label{table:comparison}}}\\n\\\\end{table*}\\nNon-deep learning based techniques also lack the capability to capture the non-linear properties of real objects~.\\nHence, the representations of objects learned by them are not expressive enough to fully support graph anomaly detection.\\nTo tackle these problems, more recent studies seek the potential of adopting deep learning techniques to identify anomalous graph objects. \\nAs a powerful tool for data mining, deep learning has achieved great success in data representation and pattern recognition~.\\nIts deep architecture with layers of parameters and transformations appear to suit the aforementioned problems well. The more recent studies, such as deep graph representation learning and graph neural networks (GNNs), further enrich the capability of deep learning for graph data mining~. \\nBy extracting expressive representations such that graph anomalies and normal objects can be easily separated, or the deviating patterns of anomalies can be learned directly through deep learning techniques, graph anomaly detection with deep learning (GADL) is starting to take the lead in the forefront of anomaly detection.\\nAs a frontier technology, graph anomaly detection with deep learning, hence, is expected to generate more fruitful results on detecting anomalies and secure a more convenient life for the society.\",\n \"id\": \"ecbafaf3-1565-4746-bf9c-a9182b235789\",\n \"level\": \"section\",\n \"origin_cites_number\": 62,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [\n \"2ea37eb2-3734-4b08-ad7a-6c1226af2e9b\",\n \"b87f5810-e203-4606-8bb5-00623ec5335b\",\n \"7291e51c-caa8-4bd7-9c08-2d393eaf83dc\"\n ],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 6279,\n 4539,\n 6280\n ],\n \"content\": \"\\\\label{sec:introdution:challenges}\\nDue to the complexity of anomaly detection and graph data mining~, in addition to the prior mentioned data-specific challenges, adopting deep learning techniques for graph anomaly detection also faces a number of challenges from the technical side. \\nThese challenges associated with deep learning are categorized as technique-specific challenges (Tech-CHs), and they are summarized as follows.\\n\\\\textbf{Tech-CH1. Anomaly-aware training objectives.} Deep learning models rely heavily on the training objectives to fine-tune all the trainable parameters. \\nFor graph anomaly detection, this necessitates appropriate training objectives or loss functions such that the GADL models can effectively capture the differences between benign and anomalous objects.\\nDesigning anomaly-aware objectives is very challenging because there is no prior knowledge about the ground-truth anomalies as well as their deviating patterns versus the majority.\\nHow to effectively separate anomalies from normal objects through training remains critical for deep learning-based models.\\n\\\\textbf{Tech-CH2. Anomaly interpretability.} In real-world scenarios, the interpretability of detected anomalies is also vital because we need to provide convincing evidence to support the subsequent anomaly handling process. For example, the risk management department of a financial organization must provide lawful evidence before blocking the accounts of identified anomalous users. \\nAs deep learning has been limited for its interpretability~, how to justify the detected graph anomalies remains a big challenge for deep learning techniques.\\n\\\\textbf{Tech-CH3. High training cost.} Although D(G)NNs are capable of digesting rich information (\\\\eg structural information and attributes) in graph data for anomaly detection, these GADL models are more complex than conventional deep neural networks or machine learning methods due to the anomaly-aware training objectives. Such complexity inherently leads to high training costs in both time and computing resources.\\n\\\\textbf{Tech-CH4. Hyperparameter tuning.} D(G)NNs naturally exhibit a large set of hyperparameters, such as the number of neurons in each neural network layer, the learning rate, the weight decay and the number of training epochs. Their learning performance is significantly affected by the values of these hyperparameters. However, it remains a serious challenge to effectively select the optimal/sub-optimal settings for the detection models due to the lack of labeled data in real scenarios.\\nBecause deep learning models are sensitive to their associated hyperparameters, setting well-performing values for the hyperparameters is vital to the success of a task.\\nTuning hyperparameter is relatively trivial in supervised learning when labeled data are available. For instance, users can find an optimal/sub-optimal set of hyperparameters (\\\\eg through random search, grid search) by comparing the model's outputs with the ground-truth. However, unsupervised anomaly detection has no accessible labeled data to judge the model's performance under different hyperparameter settings~. Selecting the ideal hyperparameter values for unsupervised detection models persists as a critical obstacle to applying them in a wide range of real scenarios.\",\n \"id\": \"2ea37eb2-3734-4b08-ad7a-6c1226af2e9b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"ecbafaf3-1565-4746-bf9c-a9182b235789\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Challenges in GAD with Deep Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Challenges in GAD with Deep Learning\"\n },\n {\n \"cite_extract_rate\": 0.29411764705882304,\n \"cites\": [\n 6275,\n 6274,\n 3207,\n 3206,\n 3344\n ],\n \"content\": \"\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n \\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/Introduction/Timeline.pdf}}\\n \\\\caption{A Timeline of Graph Anomaly Detection and Reviewed Techniques.}\\n \\\\label{pic:timeline}\\n\\\\end{figure*} \\nRecognizing the significance of anomaly detection, many review works have been conducted in the last ten years covering a range of anomaly detection topics: anomaly detection with deep learning, graph anomaly detection, graph anomaly detection with deep learning, and particular applications of graph anomaly detection such as social media, social networks, fraud detection and network security, etc.\\nThere are some representative surveys on generalized anomaly detection techniques -~,~ and~.\\nBut only the most up-to-date work in Thudumu \\\\etal~ covers the topic of graph anomaly detection.\\nRecognizing the power of deep learning, the three contemporary surveys, Ruff \\\\etal~, Pang \\\\etal~ and Chalapathy and Chawla~ specifically review deep learning based anomaly detection techniques specifically.\\nAs for graph anomaly detection, Akoglu \\\\etal~, Ranshous \\\\etal~, and Jennifer and Kumar~ put their concentration on graph anomaly detection, reviewing many conventional approaches in this area, including statistical models and machine learning techniques.\\nOther surveys are dedicated to particular applications of graph anomaly detection, such as computer network intrusion detection and anomaly detection in online social networks, \\\\eg~, and~.\\nThese works provided solid reviews of the application of anomaly detection/graph anomaly detection techniques in these high demand and vital domains. \\nHowever, none of the mentioned surveys are dedicated to techniques on graph anomaly detection with deep learning, as shown in Table~\\\\ref{table:comparison}, and hence do not provide a systematic and comprehensive review of these techniques.\",\n \"id\": \"b87f5810-e203-4606-8bb5-00623ec5335b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"ecbafaf3-1565-4746-bf9c-a9182b235789\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Existing Anomaly Detection Surveys\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Existing Anomaly Detection Surveys\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Our contributions are summarized as follows:\\n\\\\begin{itemize}\\n \\\\item \\\\textbf{The first survey in graph anomaly detection with deep learning.} To the best of our knowledge, our survey is the first to review the state-of-the-art deep learning techniques for graph anomaly detection. Most of the relevant surveys focus either on conventional graph anomaly detection methods using non-deep learning techniques or on generalized anomaly detection techniques (for tabular/point data, time series, etc.). Until now, there has been no dedicated and comprehensive survey on graph anomaly detection with deep learning. Our work bridges this gap, and we expect that an organized and systematic survey will help push forward research in this area. \\n \\\\item \\\\textbf{A systematic and comprehensive review.} In this survey, we review the most up-to-date deep learning techniques for graph anomaly detection published in influential international conferences and journals in the area of deep learning, data mining, web services, and artificial intelligence, including: TKDE, TKDD, TPAMI, NeurIPS, SIGKDD, ICDM, WSDM, SDM, SIGMOD, IJCAI, AAAI, ICDE, CIKM, ICML, WWW, CVPR, and others. We first summarize seven data-specific and four technique-specific challenges in graph anomaly detection with deep learning. We then comprehensively review existing works from the perspectives of: 1) the motivations behind the deep methods; 2) the main ideas for identifying graph anomalies; 3) a brief introduction to conventional non-deep learning techniques; and 4) the technical details of deep learning algorithms. A brief timeline of graph anomaly detection and reviewed works is given in Fig.~\\\\ref{pic:timeline}.\\n \\\\item \\\\textbf{Future directions.} From the survey results, we highlight 12 future research directions covering emerging problems introduced by graph data, anomaly detection, deep learning models, and real-world applications. These future opportunities indicate challenges that have not been adequately tackled, and so more effort is needed in the future.\\n \\\\item \\\\textbf{Affluent resources.} Our survey also provides an extensive collection of open-sourced anomaly detection algorithms, public datasets, synthetic dataset generating techniques, as well as commonly used evaluation metrics to push forward the state-of-the-art in graph anomaly detection. These published resources offer benchmark datasets and baselines for future research. \\n \\\\item \\\\textbf{A new taxonomy.} We have organized this survey with regard to different types of anomalies (\\\\ie nodes, edges, sub-graphs, and graphs) existing in graphs or graph databases. We also pinpoint the differences and similarities between different types of graph anomalies.\\n\\\\end{itemize}\\nThe rest of this survey is organized as follows. In Section~\\\\ref{sec:preliminaries}, we provide preliminaries about the different types of settings. \\nFrom Section~\\\\ref{sec:AND} to Section~\\\\ref{sec:anosgd:dynamic}, we review existing techniques for detecting anomalous nodes, edges, sub-graphs and graphs, respectively.\\nIn Section~\\\\ref{sec:resources}, we first provide a collection of published graph anomaly detection algorithms and datasets and then summarize commonly used evaluation metrics and synthetic data generation strategies.\\nWe highlight 12 future directions concerning deep learning in graph anomaly detection in Section~\\\\ref{sec:futures} and summarize our survey in Section~\\\\ref{sec:conclusion}.\\nA concrete taxonomy of our survey is given in Appendix~\\\\ref{appendix:taxonomy}.\\n\\\\vspace{-0.1cm}\",\n \"id\": \"7291e51c-caa8-4bd7-9c08-2d393eaf83dc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"ecbafaf3-1565-4746-bf9c-a9182b235789\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Contributions\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Contributions\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 3344\n ],\n \"content\": \"\\\\label{sec:preliminaries}\\nIn this section, we provide definitions of different types of graphs mostly used in node/edge/sub-graph-level anomaly detection (Section~\\\\ref{sec:AND} to Section~\\\\ref{sec:subgraph}).\\nFor consistency, we have followed the conventional categorization of graphs as in existing works~ and categorize them as static graphs, dynamic graphs, and graph databases. Unless otherwise specified, all graphs mentioned in the following sections are static. Meanwhile, as graph-level anomaly detection is discussed far away on page 13, to enhance readability, the definition for the graph database is given closer to the material in Section~\\\\ref{sec:anosgd:db}.\\n\\\\textbf{\\\\textit{Definition 1 (Plain Graph)}}. A static plain graph $G = \\\\{V, E\\\\}$ comprises a node set $V = \\\\{ v_{i} \\\\}_{1}^{n}$ and an edge set $E = \\\\{e_{i,j}\\\\}$ where $n$ is the number of nodes and $e_{i,j} = (v_i,v_j)$ denotes an edge between nodes $v_{i}$ and $v_{j}$. The adjacency matrix $A = [a_{i,j}]_{n\\\\times n}$ restores the graph structure, where $a_{i,j} = 1$ if node $v_{i}$ and $v_{j}$ is connected, otherwise $a_{i,j} = 0$.\\n\\\\textbf{\\\\textit{Definition 2 (Attributed Graph)}}. A static attributed graph $G = \\\\{V, E, X\\\\}$ comprises a node set $V$, an edge set $E$ and an attribute set $X$. In an attributed graph, the graph structure follows the definition in Definition 1. The attribute matrix $X = [\\\\mathbf{x}_{i}]_{n\\\\times k}$ consists of nodes' attribute vectors, where $\\\\mathbf{x}_{i}$ is the attribute vector associated with node $v_{i}$ and $k$ is the vector's dimension. Hereafter, the terms attribute and feature are used interchangeably.\\n\\\\textbf{\\\\textit{Definition 3 (Dynamic Graph)}}. A dynamic graph $G(t) = \\\\{V(t), E(t), X_{v}(t), X_{e}(t) \\\\}$ comprises nodes and edges changing overtime. $V(t)$ is the nodes set in the graph at a specific time step $t$, $E(t)$ is the corresponding edge set, $X_{v}(t)$ and $X_{e}(t)$ are the node attribute matrix and edge attribute matrix at time step $t$ in the graph if existed.\\nIn reality, the nodes or edges might also be associated with numerical or categorical labels to indicate their classes (\\\\eg normal or abnormal). When label information is available/partially-available, supervised/semi-supervised detection models could be effectively trained.\",\n \"id\": \"8ea06e54-3d89-4768-967e-6b68c2d22f53\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Preliminaries\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Preliminaries\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:AND}\\nAnomalous nodes are commonly recognized as individual nodes that are significantly different from others.\\nIn real-world applications, these nodes often represent abnormal objects that appear individually, such as a single network intruder in computer networks, an independent fraudulent user in online social networks or a specific fake news on social media.\\nIn this section, we specifically focus on anomalous node detection in static graphs.\\nThe reviews on dynamic graphs can be found in Section~\\\\ref{sec:node:dg}. Table~\\\\ref{table:ANOSND} at the end of Section~\\\\ref{sec:node:dg} provides a summary of techniques reviewed for ANOS ND.\\nWhen detecting anomalous nodes in static graphs, the differences between anomalies and regular nodes are mainly drawn from the graph structural information and nodes/edges' attributes~.\\nGiven prior knowledge (\\\\ie community structure, attributes) about a static graph, anomalous nodes can be further categorized into the following three types:\\n\\\\begin{itemize}\\n \\\\item \\\\textbf{Global anomalies} only consider the node attributes. They are nodes that have attributes significantly different from all other nodes in the graph. \\n \\\\item \\\\textbf{Structural anomalies} only consider the graph structural information. They are abnormal nodes that have different connection patterns (\\\\eg connecting different communities, forming dense links with others).\\n \\\\item \\\\textbf{Community anomalies} consider both node attributes and graph structural information. They are defined as nodes that have different attribute values compared to other nodes in the same community.\\n\\\\end{itemize}\\nIn Fig.~\\\\ref{pic:node_classes}, node 14 is a global anomaly because its 4th feature value is 1 while all other nodes in the graph have the value of 0 for the corresponding feature. Nodes 5, 6, and 11 are identified as structural anomalies because they have links with other communities while other nodes in their community do not form cross-community links. Nodes 2 and 7 are community anomalies because their feature values are different from others in the communities they belong to. \\n\\\\begin{figure}[!t]\\n \\\\setlength{\\\\belowcaptionskip}{-0.5cm} \\n \\\\centerline{\\\\includegraphics[scale=0.47]{pics/Node/node_classes_new.pdf}}\\n \\\\caption{Three Types of Anomalous Nodes: Structural Anomalies, Community Anomalies and Global Anomalies.}\\n \\\\label{pic:node_classes}\\n\\\\end{figure}\",\n \"id\": \"59eed45d-0f6f-4983-a1d0-48c2fb48a606\",\n \"level\": \"section\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ]\n ],\n \"subsections\": [\n \"5e100893-099a-4201-b53e-ba07115d4307\",\n \"9b59f8e2-fd33-4bf3-bf31-9d82848223fa\"\n ],\n \"title\": \"Anomalous node detection (ANOS ND)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:node:spg}\\nPlain graphs are dedicated to representing the structural information in real-world networks.\\nTo detect anomalous nodes in plain graphs, the graph structure has been extensively exploited from various angles. \\nHere, we first summarize the representative traditional non-deep learning approaches, followed by a more recent, advanced detection technique based on representation learning.\\n\\\\begin{figure*}[!t]\\n \\\\setlength{\\\\belowcaptionskip}{-0.25cm} \\n \\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/Node/Traditional.pdf}}\\n \\\\caption{ANOS ND on attributed graphs -- Deep NN based approaches. As an example, the autoencoder is used to capture the graph structure and node attributes. With a specially-designed anomaly aware loss function, anomaly scores will be assigned to every node, and the top-k nodes are anomalies (\\\\eg nodes 9, 1, and 2 at top-3).}\\n \\\\label{pic:conventional_autoencoder_framework}\\n\\\\end{figure*}\",\n \"id\": \"5e100893-099a-4201-b53e-ba07115d4307\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"59eed45d-0f6f-4983-a1d0-48c2fb48a606\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Plain Graphs\"\n ]\n ],\n \"subsections\": [\n \"658ca9f1-8bc4-490f-ac7d-c460db855b3b\",\n \"699af85c-49af-4ca0-a394-878a65f78e4e\",\n \"bc2ebce7-d1c1-4f0b-881b-6a0ace25800e\"\n ],\n \"title\": \"ANOS ND on Plain Graphs\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:anosnd:tndp}\\nPrior to the recent advances in deep learning and other state-of-the-art data mining technologies, traditional non-deep learning techniques have been widely used in many real-world networks to identify anomalous entities.\\nA key idea behind these techniques was to transform the graph anomaly detection into a traditional anomaly detection problem, because the graph data with rich structure information can not be handled by the traditional detection techniques (for tabular data only) directly. \\nTo bridge the gap, many approaches~ used the statistical features associated with each node, such as in/out degree, to detect anomalous nodes. \\nFor instance, OddBall~ employs the statistical features (\\\\eg the number of 1-hop neighbors and edges, the total weight of edges) extracted from each node and its 1-hop neighbors to detect particular structural anomalies that: 1) form local structures in shape of near-cliques or stars; 2) have heavy links with neighbors such that the total weight is extremely large; or 3) have a single dominant heavy link with one of the neighbors. \\nWith properly selected statistical features, anomalous nodes can be identified with respect to their deviating feature patterns.\\nBut, in real scenarios, it is very hard to choose the most suitable features from a large number of candidates, and domain experts can always design new statistics, \\\\eg the maximum/minimum weight of edges.\\nAs a result, these techniques often carry prohibitive cost for assessing the most significant features and do not effectively capture the structural information.\",\n \"id\": \"658ca9f1-8bc4-490f-ac7d-c460db855b3b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"5e100893-099a-4201-b53e-ba07115d4307\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Plain Graphs\"\n ],\n [\n \"subsubsection\",\n \"Traditional Non-Deep Learning Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Traditional Non-Deep Learning Techniques\"\n },\n {\n \"cite_extract_rate\": 0.27272727272727204,\n \"cites\": [\n 282,\n 6281,\n 218\n ],\n \"content\": \"\\\\label{sec:spg:nr}\\nTo capture more valuable information from the graph structure for anomaly detection, network representation techniques have been widely exploited. \\nTypically, these techniques encode the graph structure into an embedded vector space and identify anomalous nodes through further analysis. Hu \\\\etal~, for example, proposed an effective embedding method to detect structural anomalies that are connecting with many communities. \\nIt first adopts a graph partitioning algorithm (\\\\eg METIS~) to group nodes into $d$ communities ($d$ is a user-specified number). \\nThen, the method employs a specially designed embedding procedure to learn node embeddings that could capture the link information between each node and $d$ communities.\\nDenoting the embedding for node $i$ as $Z_i = \\\\{z_{i}^{1},\\\\cdots,z_{i}^{d}\\\\}$, the procedure initializes each $z_{i}^{c} \\\\in Z_i$ with regard to the membership of node $i$ to community $c$ (if node $i$ belongs to the community, then $z_{i}^{c} = \\\\frac{1}{\\\\sqrt{2}}$; otherwise, 0.) and optimizes node embeddings such that directly linked nodes have similar embeddings and unconnected nodes are dissimilar.\\nAfter generating the node embeddings, the link information between node $i$ and $d$ communities is quantified for further anomaly detection analysis.\\nFor a given node $i$, such information is represented as:\\n\\\\begin{equation} \\\\label{eq:hu:nbi}\\n \\\\overline{NB(i)} = (y_{i}^{1},...,y_{i}^{d}) = \\\\mathop{\\\\sum}\\\\limits_{j\\\\in NB(i)}(1- \\\\|Z_{i} - Z_{j}\\\\|)\\\\cdot Z_{j},\\n\\\\end{equation}\\nwhere $NB(i)$ comprises node $i$'s neighbors.\\nIf $i$ has many links with community $c$, then the value in the corresponding dimension $y_{i}^{c}$ will be large.\\nIn the last step, Hu \\\\etal~ formulate a scoring function to assign anomalousness scores, calculated as:\\n\\\\begin{equation} \\\\label{eq:hu:as}\\n AScore(i) = \\\\mathop{\\\\sum}\\\\limits_{k=1}^{d}\\\\frac{y_{i}^{k}}{y_{i}^{*}}, y_{i}^{*}=\\\\max\\\\{y_{i}^{1},...,y_{i}^{d}\\\\}.\\n\\\\end{equation}\\nAs expected, structural anomalies receive higher scores as they connect to different communities.\\nIndeed, given a predefined threshold, nodes with above-threshold scores are identified as anomalies.\\nTo date, many plain network representation methods such as Deepwalk~, Node2Vec~ and LINE~ have shown their effectiveness in generating node representations and been used for anomaly detection performance validation~.\\nBy pairing the conventional anomaly detection techniques such as density-based techniques~ and distance-based techniques~ with node embedding techniques, anomalous nodes can be identified with regard to their distinguishable locations (\\\\ie low-density areas or far away from the majorities) in the embedding space.\",\n \"id\": \"699af85c-49af-4ca0-a394-878a65f78e4e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"5e100893-099a-4201-b53e-ba07115d4307\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Plain Graphs\"\n ],\n [\n \"subsubsection\",\n \"Network Representation Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Representation Based Techniques\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6282\n ],\n \"content\": \"The success of reinforcement learning (RL) in tackling real-world decision making problems has attracted substantial interests from the anomaly detection community. Detecting anomalous nodes can be naturally regarded as a problem of deciding which class a node belongs to - anomalous or benign. As a special scenario of the general selective harvesting task, the anomalous node detection problem can be approached by a recent work in~ that intuitively combines reinforcement learning and network embedding techniques for selective harvesting. The proposed model, NAC, is trained with labeled data without any human intervention. Specifically, it first selects a seed network consisting of partially observed nodes and edges. Then, starting from the seed network, NAC adopts reinforcement learning to learn a node selection plan such that anomalous nodes in the undiscovered area can be identified. This is achieved by rewarding selection plans that can choose labeled anomalies with higher gains. Through offline training, NAC will learn an optimal/suboptimal anomalous node selection strategy and discover potential anomalies in the undiscovered graph step by step.\",\n \"id\": \"bc2ebce7-d1c1-4f0b-881b-6a0ace25800e\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"5e100893-099a-4201-b53e-ba07115d4307\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Plain Graphs\"\n ],\n [\n \"subsubsection\",\n \"Reinforcement Learning Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Reinforcement Learning Based Techniques\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 7007,\n 180,\n 242\n ],\n \"content\": \"\\\\label{lb:node:sag}\\nIn addition to the structural information, real-world networks also contain rich attribute information affiliated with nodes~. These attributes provide complementary information about real objects and together with graph structure, more hidden anomalies that are non-trivial can now be detected. \\n\\\\begin{figure*}[!t]\\n \\\\setlength{\\\\belowcaptionskip}{-0.25cm} \\n \\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/Node/GCN.pdf}}\\n \\\\caption{ANOS ND on attributed graphs -- GCN based approaches. Node representations are generated through GCN layers. Anomalies are then detected according to their reconstruction loss (\\\\ding{172}) or embedding distribution in the embedding space (\\\\ding{173}).}\\n \\\\label{pic:GCNbased_framework}\\n\\\\end{figure*}\\nFor clarity, we distinguish between deep neural networks and graph neural networks in this survey. \\nWe review deep neural network (Deep NN) based techniques, GCN based techniques, and reinforcement learning based techniques for ANOS ND as follows. \\nDue to page limitations, other existing works including traditional non-deep learning techniques, GAT~ based techniques, GAN based techniques, and network representation based techniques are surveyed in Appendix~\\\\ref{appendix:node:static}.\",\n \"id\": \"9b59f8e2-fd33-4bf3-bf31-9d82848223fa\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"59eed45d-0f6f-4983-a1d0-48c2fb48a606\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Attributed Graphs\"\n ]\n ],\n \"subsections\": [\n \"9201dd97-4956-4c75-840c-212d3ec5e5ae\",\n \"73bf642c-6661-4550-b563-f83087e0c906\",\n \"be74c2e6-73e6-42d6-9154-063b0fce612a\"\n ],\n \"title\": \"ANOS ND on Attributed Graphs\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The deep learning models such as autoencoder and deep neural networks provide solid basis for learning data representations. \\nAdopting these models for more effective anomalous node detection have drawn substantial interest recently.\\nFor example, Bandyopadhyay \\\\etal~ developed an unsupervised deep model, DONE, to detect global anomalies, structural anomalies and community anomalies in attributed graphs.\\nSpecifically, this work measures three anomaly scores for each node that indicate the likelihood of the situations where 1) it has similar attributes with nodes in different communities ($o_i^a$); or 2) it connects with other communities ($o_i^s$); or 3) it belongs to one community structurally but the attributes follow the pattern of another community ($o_i^{com}$).\\nIf a particular node exhibits any of these characteristics, then it is assigned a higher score and is anomalous.\\nTo acquire these scores, DONE adopts two separate autoencoders (AE), \\\\ie a structure AE and an attribute AE, as shown in Fig.~\\\\ref{pic:conventional_autoencoder_framework}. \\nBoth are trained by minimizing the reconstruction errors and preserving the homophily that assumes connected nodes have similar representations in the graph. \\nWhen training the AEs, nodes exhibiting the predefined characteristics are hard to reconstruct and therefore introduce more reconstruction errors because their structure or attribute patterns do not conform to the standard behavior.\\nHence, the adverse impact of anomalies should be alleviated to achieve the minimized error.\\nAccordingly, DONE specially designs an anomaly-aware loss function with five terms: $\\\\mathcal{L}_{str}^{Recs}$, $\\\\mathcal{L}_{attr}^{Recs}$, $\\\\mathcal{L}_{str}^{Hom}$, $\\\\mathcal{L}_{attr}^{Hom}$, and $\\\\mathcal{L}^{Com}$.\\n$\\\\mathcal{L}_{str}^{Resc}$ and $\\\\mathcal{L}_{attr}^{Resc}$ are the structure reconstruction error and attribute reconstruction error that can be written as: \\n\\\\begin{equation} \\\\label{DONE:ls1}\\n \\\\begin{split}\\n \\\\mathcal{L}_{str}^{Recs} = \\\\frac{1}{N}\\\\sum_{i=1}^{N}\\\\log(\\\\frac{1}{o_{i}^{s}})\\\\|\\\\mathbf{t}_{i} - \\\\mathbf{\\\\hat{t}}_{i}\\\\|_{2}^{2},\\n \\\\end{split}\\n\\\\end{equation}\\nand \\n\\\\begin{equation} \\\\label{DONE:ls2}\\n \\\\begin{split}\\n \\\\mathcal{L}_{attr}^{Recs} \\\\frac{1}{N}\\\\sum_{i=1}^{N}\\\\log(\\\\frac{1}{o_{i}^{a}})\\\\|\\\\mathbf{x}_{i} - \\\\mathbf{\\\\hat{x}}_{i}\\\\|_{2}^{2},\\n \\\\end{split}\\n\\\\end{equation}\\nwhere $N$ is the number of nodes, $\\\\mathbf{t}_{i}$ and $\\\\mathbf{x}_{i}$ store the structure information and attributes of node $i$, $\\\\mathbf{\\\\hat{t}}_{i}$ and $\\\\mathbf{\\\\hat{x}}_{i}$ are the reconstructed vectors. \\n$\\\\mathcal{L}_{str}^{Hom}$ and $\\\\mathcal{L}_{attr}^{Hom}$ are proposed to maintain the homophily and they are formulated as:\\n\\\\begin{equation} \\\\label{DONE:ls3}\\n \\\\begin{split}\\n \\\\mathcal{L}_{str}^{Hom} = \\\\frac{1}{N}\\\\sum_{i=1}^{N}\\\\log(\\\\frac{1}{o_{i}^{s}})\\\\frac{1}{|N(i)|}\\\\sum_{j\\\\in N(i)}||\\\\mathbf{h}_{i}^{s} - \\\\mathbf{h}_{j}^{s}||_{2}^{2},\\n \\\\end{split}\\n\\\\end{equation}\\nand \\n\\\\begin{equation} \\\\label{DONE:ls4}\\n \\\\begin{split}\\n \\\\mathcal{L}_{attr}^{Hom} = \\\\frac{1}{N}\\\\sum_{i=1}^{N}\\\\log(\\\\frac{1}{o_{i}^{a}})\\\\frac{1}{|N(i)|}\\\\sum_{j\\\\in N(i)}||\\\\mathbf{h}_{i}^{a} - \\\\mathbf{h}_{j}^{a}||_{2}^{2},\\n \\\\end{split}\\n\\\\end{equation}\\nwhere $\\\\mathbf{h}_{i}^{s}$ and $\\\\mathbf{h}_{i}^{a}$ are the learned latent representations from the structure AE and attribute AE, respectively.\\n$\\\\mathcal{L}^{Com}$ poses further restrictions on the generated representations for each node by the two AEs such that the graph structure and node attributes complement each other. It is formulated as:\\n\\\\begin{equation} \\\\label{DONE:ls5}\\n \\\\begin{split}\\n \\\\mathcal{L}^{Com} = \\\\frac{1}{N}\\\\sum_{i=1}^{N}\\\\log(\\\\frac{1}{o_{i}^{com}})||\\\\mathbf{h}_{i}^{s} - \\\\mathbf{h}_{i}^{a}||_{2}^{2},\\n \\\\end{split}\\n\\\\end{equation}\\nBy minimizing the sum of these loss functions, the anomaly scores of each node are quantified, and the top-k nodes with higher scores are identified as anomalies.\",\n \"id\": \"9201dd97-4956-4c75-840c-212d3ec5e5ae\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"9b59f8e2-fd33-4bf3-bf31-9d82848223fa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Attributed Graphs\"\n ],\n [\n \"subsubsection\",\n \"Deep NN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep NN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0.22222222222222202,\n \"cites\": [\n 6283,\n 6284\n ],\n \"content\": \"\\\\label{sec:AND:GCN}\\nGraph convolutional neural networks (GCNs)~ have accomplished decent success in many graph data mining tasks (\\\\eg link prediction, node classification, and recommendation) owing to its capability of capturing comprehensive information in the graph structure and node attributes.\\nTherefore, many anomalous node detection techniques start to investigate GCNs. \\nFig.~\\\\ref{pic:GCNbased_framework} illustrates a general framework of existing works in this line.\\nIn~, Ding \\\\etal measured an anomaly score for each node using the network reconstruction errors of both the structure and attribute.\\nThe proposed method, DOMINANT, comprises three parts, namely, the graph convolutional encoder, the structure reconstruction decoder, and the attribute reconstruction decoder. The graph convolutional encoder generates node embeddings through multiple graph convolutional layers. The structure reconstruction decoder tends to reconstruct the network structure from the learned node embeddings, while the attribute reconstruction decoder reconstructs the node attribute matrix.\\nThe whole neural network is trained to minimize the following loss function:\\n\\\\begin{equation} \\\\label{eq:ad}\\n \\\\begin{split}\\n \\\\mathcal{L}_{DOMINANT} &= (1-\\\\alpha)\\\\mathcal{R}_{S} + \\\\alpha \\\\mathcal{R}_{A}\\\\\\\\\\n &= (1-\\\\alpha)||A - \\\\hat{A}||_{F}^{2} + \\\\alpha||X - \\\\hat{X}||_{F}^{2},\\n \\\\end{split}\\n\\\\end{equation}\\nwhere $\\\\alpha$ is the coefficient, $A$ depicts the adjacency matrix of the graph, $\\\\mathcal{R}_{S}$ and $\\\\mathcal{R}_{A}$ quantify reconstruction errors with regard to the graph structure and node attributes, respectively.\\nWhen the training is finished, an anomaly score is then assigned to each node according to its contribution to the total reconstruction error, which is calculated by:\\n\\\\begin{equation} \\\\label{eq:nodescore}\\n \\\\textit{score}(i) \\n = (1-\\\\alpha)||\\\\mathbf{a}_{i} - \\\\mathbf{\\\\hat{a}}_{i}||_{2} + \\\\alpha||\\\\mathbf{x}_{i} - \\\\mathbf{\\\\hat{x}}_{i}||_{2},\\n\\\\end{equation}\\nwhere $\\\\mathbf{a}_{i}$ and $\\\\mathbf{x}_{i}$ are the structure vector and attribute vector of node $i$, $\\\\mathbf{\\\\hat{a}}_{i}$ and $\\\\mathbf{\\\\hat{x}}_{i}$ are their corresponding reconstructed vectors. \\nThe nodes are then ranked according to their anomaly scores in descending order, and the top-k nodes are recognized as anomalies.\\nTo enhance the performance of anomalous node detection, later work by Peng \\\\etal~ further explores node attributes from multiple attributed views to detect anomalies.\\nThe multiple attributed views are employed to describe different perspectives of the objects~.\\nFor example, in online social networks, user's demographic information and posted contents are two different attributed views, and they characterize the personal information and social activities, respectively.\\nThe underlying intuition of investigating different views is that anomalies might appear to be normal in one view but abnormal in another view.\\nFor the purpose of capturing these signals, the proposed method, ALARM, applies multiple GCNs to encode information in different views and adopts a weighted aggregation of them to generate node representations.\\nThis model's training strategy is similar to DOMINANT~ in that it aims to minimize the network reconstruction loss and attribute reconstruction loss and can be formulated as:\\n\\\\begin{equation} \\\\label{eq:alarmloss}\\n\\\\begin{split}\\n\\\\mathcal{L}_{ALARM} = & \\\\sum_{i=1}^n \\\\sum_{j=1}^n - [\\\\gamma A_{ij}\\\\log\\\\hat{A}_{ij} + (1-A_{ij})\\\\log(1-\\\\hat{A}_{ij})] \\\\\\\\\\n & + ||X - \\\\tilde{X}||^{F}_{2},\\n\\\\end{split}\\n\\\\end{equation}\\nwhere $\\\\gamma$ is coefficient to balance the errors, $A_{ij}$ is the element at coordinate $(i,j)$ in the adjacency matrix $A$, $\\\\hat{A}_{ij}$ is the corresponding element in the reconstructed adjacency matrix $\\\\hat{A}$, $X$ is the original node feature matrix and $\\\\tilde{X}$ is the reconstructed node feature matrix.\\nLastly, ALARM adopts the same scoring function as~, and nodes with top-k highest scores are anomalous.\\nInstead of spotting unexpected nodes using their reconstruction errors, Li \\\\etal~ proposed SpecAE to detect global anomalies and community anomalies via a density estimation approach, Gaussian Mixture Model (GMM).\\nGlobal anomalies can be identified by only considering the node attributes. \\nFor community anomalies, the structure and attributes need to be jointly considered because of their distinctive attributes to the neighbors.\\nAccordingly, SpecAE investigates a graph convolutional encoder to learn node representations and reconstruct the nodal attributes through a deconvolution decoder.\\nThe parameters in the GMM are then estimated using the node representations. \\nDue to the deviating attribute patterns of global and community anomalies, normal nodes are expected to exhibit greater energies in GMM, and the k nodes with the lowest probabilities are deemed to be anomalies.\\nIn~, Wang \\\\etal developed a novel detection model that identify fraudsters using their relations and features. Their proposed method, Fdgars, first models online users' reviews and visited items as their features, and then identifies a small portion of significant fraudsters based on these features. In the last step, a GCN is trained in a semi-supervised manner by using the user-user network, user features, and labeled users. After training, the model can directly label unseen users.\\nA more recent work, GraphRfi~, also explores the potential of combining anomaly detection with other downstream graph analysis tasks.\\nIt targets on leveraging anomaly detection to identify malicious users and provide more accurate recommendations to service benign users by alleviating the impact of these untrustworthy users. \\nSpecifically, a GCN framework is deployed to encode users and items into a shared embedding space for recommendation and users are classified as fraudsters or normal users through an additional neural random forest using their embeddings. \\nFor rating prediction between users and items, the framework reduces the corresponding impact of suspicious users by assigning less weights to their training loss. \\nAt the same time, the rating behavior of users also provides auxiliary information for fraudster detection. \\nThe mutually beneficial relationship between these two applications (anomaly detection and recommendation) indicates the potential of information sharing among multiple graph learning tasks.\",\n \"id\": \"73bf642c-6661-4550-b563-f83087e0c906\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"9b59f8e2-fd33-4bf3-bf31-9d82848223fa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Attributed Graphs\"\n ],\n [\n \"subsubsection\",\n \"GCN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GCN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In contrast to NAC, Ding \\\\etal~ investigated to the use of reinforcement learning for anomalous node detection in attributed graphs.\\nTheir proposed algorithm, GraphUCB, models both attribute information and structural information, and inherits the merits of the contextual multi-armed bandit technology~ to output potential anomalies.\\nBy grouping nodes into $k$ clusters based on their features, GraphUCB forms a $k$-armed bandit model and measures the payoff of selecting a specific node as a potential anomaly for expert evaluation.\\nWith experts' feedback on the predicted anomalies, the decision-making strategy is continuously optimized.\\nEventually, the most potential anomalies can be selected. \\n\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n\\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/Edge/UGED.pdf}}\\n\\\\caption{ANOS ED on static graphs -- Deep NN based approaches. For example, the detection technique employs the autoencoder and fully connected network to learn the presence probability of each edge. Anomaly scores are assigned with regard to the probabilities, and the top-k edges are anomalous (\\\\eg $E_{8,9}$, $E_{1,3}$, and $E_{2,4}$ at top-3).}\\n\\\\label{pic:UGED}\\n\\\\end{figure*}\",\n \"id\": \"be74c2e6-73e6-42d6-9154-063b0fce612a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"9b59f8e2-fd33-4bf3-bf31-9d82848223fa\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous node detection (ANOS ND)\"\n ],\n [\n \"subsection\",\n \"ANOS ND on Attributed Graphs\"\n ],\n [\n \"subsubsection\",\n \"Reinforcement Learning Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Reinforcement Learning Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0.38235294117647,\n \"cites\": [\n 6285,\n 6290,\n 6288,\n 6287,\n 6283,\n 6289,\n 6282,\n 3344,\n 6271,\n 6284,\n 6291,\n 8991,\n 6286\n ],\n \"content\": \"\\\\label{sec:node:dg}\\nReal-world networks can be modeled as dynamic graphs to represent evolving objects and the relationships among them.\\nIn addition to structural information and node attributes, dynamic graphs also contain rich temporal signals~, \\\\eg the evolving patterns of the graph structure and node attributes. \\nOn the one hand, these information inherently makes anomalous node detection on dynamic graphs more challenging. \\nThis is because dynamic graphs usually introduce large volume of data and temporal signals should also be captured for anomaly detection.\\nBut, on the other hand, they could provide more details about anomalies~. \\nIn fact, some anomalies might appear to be normal in the graph snapshot at each time stamp, and, only when the changes in a graph's structure are considered, do they become noticeable.\\nIn this section, we review the network representation based techniques and GAN based techniques as follows. Relevant techniques from traditional non-deep learning approaches are reviewed in Appendix~\\\\ref{appendix:node:dynamic}.\\n\\\\begin{table*}[!t]\\n\\\\centering\\n\\\\footnotesize\\n\\\\renewcommand\\\\arraystretch{1}\\n\\\\setlength{\\\\tabcolsep}{2.8mm}\\n\\\\caption{Summary of Anomalous Node Detection Techniques.}\\n\\\\resizebox{0.96\\\\textwidth}{!}{\\n\\\\begin{tabular}{m{2.1cm}<{\\\\centering}|m{2.3cm}<{\\\\centering}|m{1.1cm}<{\\\\centering}|m{4cm}<{\\\\centering}|m{1.8cm}<{\\\\centering}|m{2.9cm}<{\\\\centering}}\\n\\\\toprule[1 pt]\\n\\\\textbf{Graph Type} & \\\\textbf{Approach} & \\\\textbf{Category} & \\\\textbf{Objective Function} & \\\\textbf{Measurement} & \\\\textbf{Outputs} \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\multirow{3}{*}{\\\\shortstack{Static Graph - \\\\\\\\ Plain}} & ~ & NR & $\\\\mathop{\\\\sum}\\\\limits_{(i,j)\\\\in E} \\\\|\\\\mathbf{Z_{i}} - \\\\mathbf{Z_{j}}\\\\|^{2} + \\\\alpha \\\\mathop{\\\\sum}\\\\limits_{(i,j)\\\\notin E}(\\\\|\\\\mathbf{Z_{i}} - \\\\mathbf{Z_{j}}\\\\| -1)^{2}$ & Anomaly Score & $ \\\\mathop{\\\\sum}\\\\limits_{k=1}^{d}\\\\frac{y_{i}^{k}}{y_{i}^{*}}$ \\\\\\\\ \\\\cline{2-6}\\n& DCI~ & NR &$\\\\frac{1}{K}\\\\mathop{\\\\sum}\\\\limits_{k=1}^K\\\\mathcal{L}_{DCI}^{k}$ & Anomaly Prediction & Predicted Label \\\\\\\\\\\\cline{2-6}\\n&NAC~ & RL & Cumulative reward & - & Anomalies \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\multirow{24}{*}{\\\\shortstack{Static Graph - \\\\\\\\ Attributed}} &ALAD~ & Non-DP & $\\\\min\\\\limits_{W,H} \\\\|A-WW^{T}\\\\|_{F}^{2} + \\\\alpha \\\\|X-WH\\\\|_{F}^{2} + \\\\gamma (\\\\|W\\\\|_{F}^{2} + \\\\|H\\\\|_{F}^{2})$ & Anomaly Score & $\\\\frac{\\\\mathbf{W}_{n,c}}{\\\\sum_{c}\\\\mathbf{W}_{n,c}} cos(\\\\mathbf{A}_{n*}, \\\\mathbf{H}_{c*})$ \\\\\\\\ \\\\cline{2-6}\\n&Radar~ & Non-DP & $\\\\min\\\\limits_{W,R}\\\\|X - W^{T} X - R\\\\|_{F}^{2}+ \\\\alpha \\\\|W\\\\|_{2,1} +\\\\beta \\\\|R\\\\|_{2,1} +\\\\gamma tr(R^{T}LR)$ & Residual Analysis & Residual Value \\\\\\\\ \\\\cline{2-6}\\n&ANOMALOUS~ & Non-DP & $\\\\min\\\\limits_{W,\\\\tilde{R}}||X - XWX - \\\\tilde{R}||_{F}^{2} + \\\\alpha ||W||_{2,1} + \\\\beta ||W^{T}||_{2,1} + \\\\gamma ||\\\\tilde{R}^{T}||_{2,1} + \\\\varphi tr(\\\\tilde{R}L\\\\tilde{R}^{T})$ & Residual Analysis & Residual Value \\\\\\\\ \\\\cline{2-6}\\n&SGASD~ & Non-DP & $\\\\min\\\\limits_{\\\\mathbf{w,c}}\\\\frac{1}{2}\\\\sum_{i=1}^{m}\\\\mathbf{c_{i}}(V_{i,*}\\\\mathbf{w} - y_i)^2 + \\\\frac{\\\\lambda_{1}}{2}||\\\\mathbf{w}||_{2}^{2} + \\\\lambda_{2} \\\\sum_{i=0}^{d}\\\\sum_{j=1}^{n_{i}}||\\\\mathbf{c}_{G_{j}^{i}}||_{2}$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n&DONE~ & DNN & $\\\\alpha_{1}\\\\mathcal{L}_{str}^{Recs} + \\\\alpha_{2}\\\\mathcal{L}_{attr}^{Recs} + \\\\alpha_{3}\\\\mathcal{L}_{str}^{Hom} + \\\\alpha_{4}\\\\mathcal{L}_{attr}^{Hom} + \\\\alpha_{5}\\\\mathcal{L}^{Com}$ & Anomaly Scores & $o_{i}^{s},o_{i}^{a},o_{i}^{com}$ \\\\\\\\ \\\\cline{2-6}\\n&DOMINANT~ & GCN & $(1-\\\\alpha)\\\\mathcal{R}_{S} + \\\\alpha \\\\mathcal{R}_{A}$ & Anomaly Score & $(1-\\\\alpha)||\\\\mathbf{a}_{i} - \\\\mathbf{\\\\hat{a}}_{i}||_{2} + \\\\alpha||\\\\mathbf{x}_{i} - \\\\mathbf{\\\\hat{x}}_{i}||_{2}$ \\\\\\\\ \\\\cline{2-6}\\n&ALARM~ & GCN & $\\\\sum_{i=1}^n \\\\sum_{j=1}^n - [\\\\gamma A_{ij}\\\\log\\\\hat{A}_{ij} + (1-A_{ij})\\\\log(1-\\\\hat{A}_{ij})] + \\\\mathcal{L}_a$ & Anomaly Score & $(1-\\\\alpha)||\\\\mathbf{a}_{i} - \\\\mathbf{\\\\hat{a}}_{i}||^2_{2} + \\\\alpha||\\\\mathbf{x}_{i} - \\\\mathbf{\\\\hat{x}}_{i}||^2_{2}$ \\\\\\\\ \\\\cline{2-6}\\n&SpecAE~ & GCN & $\\\\mathbb{E}[dis(X,\\\\hat{X})] + \\\\mathbb{E}[dis(X,\\\\tilde{X})] + \\\\lambda_{1}\\\\mathbb{E}(E(Z)) + \\\\lambda_{2}KL$ & Density Estimation & Anomalousness Rank \\\\\\\\ \\\\cline{2-6}\\n&Fdgars~ & GCN & $\\\\mathcal{L}_{GCN}$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n&GraphRfi~ & GCN & $\\\\mathcal{L}_{rating} + \\\\lambda \\\\mathcal{L}_{fraudster}$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n&ResGCN~ & GCN & $(1-\\\\alpha)||A - \\\\hat{A}||^2_{F} + \\\\alpha||X - \\\\hat{X} - \\\\lambda R||^2_{F}$ & Anomaly Score & $||R_{i,:}||_{2}$ \\\\\\\\ \\\\cline{2-6}\\n&GraphUCB~ & RL & Expert Judgment & - & Anomalies \\\\\\\\ \\\\cline{2-6}\\n&AnomalyDAE~ & GAT & $\\\\alpha ||(A - \\\\hat{A})\\\\odot \\\\bm{\\\\theta}||_{F}^{2} + (1-\\\\alpha)||(X - \\\\hat{X})\\\\odot \\\\bm{\\\\eta}||_{F}^{2}$ & Reconstruction Loss & Anomalousness Rank \\\\\\\\ \\\\cline{2-6}\\n&SemiGNN~ & GAT & $\\\\alpha \\\\mathcal{L}_{sup} + (1- \\\\alpha) \\\\mathcal{L}_{unsup} + \\\\lambda \\\\mathcal{L}_{reg}$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n&AEGIS~ & GAN & $\\\\mathcal{L}_{AE} + \\\\mathcal{L}_{GAN}$ & Anomaly Score & $1 - D(\\\\mathbf{z}_i)$ \\\\\\\\ \\\\cline{2-6}\\n&REMAD~ & NR & $\\\\mathcal{L}_{res} + \\\\beta\\\\| R^{T}\\\\|_{2,1}$ & Residual Analysis & Residual Value \\\\\\\\ \\\\cline{2-6}\\n&CARE-GNN~ & NR & $\\\\mathcal{L}_{GNN} + \\\\lambda_{1}\\\\mathcal{L}_{Simi}^{(1)} + \\\\lambda_{2}\\\\mathcal{L}_{reg}$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n&SEANO~ & NR & $-\\\\sum_{i \\\\in V_{L}} \\\\log p(y_{i}|\\\\mathbf{x}_{i},\\\\Bar{\\\\mathbf{x}}_{N_{i}}) - \\\\sum_{i \\\\in V}\\\\sum_{v\\\\prime \\\\in C_i} \\\\log p(v\\\\prime|\\\\mathbf{x}_{i},\\\\Bar{\\\\mathbf{x}}_{N_{i}})$ & Anomaly Score & Discriminator's Output \\\\\\\\ \\\\cline{2-6}\\n&OCGNN~ & NR & $\\\\frac{1}{\\\\beta K}\\\\sum\\\\limits_{v_{i}\\\\in\\\\mathbf{V}_{tr}}[||g(X,A;\\\\mathcal{W})_{v_i}-c||^2 -r^2]^{+}+ r^2 +\\\\frac{\\\\lambda}{2}\\\\sum\\\\limits_{l=1}^{L}||W^{(l)}||^2$ & Location in Embedding Space & Distance to Hypersphere Center \\\\\\\\ \\\\cline{2-6}\\n&GAL~ & NR & $\\\\max\\\\{0,\\\\max\\\\limits_{y_{v^{\\\\prime}} \\\\neq y_{u}}g(u,v^\\\\prime) - \\\\min\\\\limits_{y_{v} = y_{u}}g(u,v) + \\\\Delta_{y_u}\\\\}$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n&CoLA~& NR & $-\\\\sum\\\\limits_{i=1}^{N}y_{i}\\\\log(CLM(v_i,\\\\mathcal{G}_{i})) + (1-y_i)\\\\log(1-CLM(v_i,\\\\mathcal{G}_{i}))$ & Anomaly Score & $\\\\frac{\\\\sum\\\\limits_{r=1}^{R}(s_{i,r}^{(-)}-s_{i,r}^{(+)})}{R}$ \\\\\\\\ \\\\cline{2-6}\\n&COMMANDER~ & NR & $-\\\\mathcal{L}_{D} + \\\\mathcal{L}_{C} + \\\\mathcal{L}_{R}$ & Anomaly Score & $\\\\bar{y_i}||\\\\mathbf{\\\\tilde{x}_i} - \\\\mathbf{x}_i||_2^2$ \\\\\\\\ \\\\cline{2-6}\\n&FRAUDRE~& NR & $\\\\sum\\\\limits_{i=1}^{n}f^{*}(y_i,\\\\mathbf{h}_{i}^{(final)}\\\\mathbf{W}_2)$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n&Meta-GDN~ & NR & $(1-y_i)\\\\cdot|dev(v_i)| + y_i\\\\cdot\\\\max(0,dev(v_i))$ & Anomaly Score & $\\\\mathbf{u}_{s}^{T}\\\\mathbf{o}_{i} + b_s$ \\\\\\\\ \\n\\\\midrule[1 pt]\\nDynamic Graph - Plain & NetWalk~ & DNN & $\\\\gamma\\\\mathcal{L}_{AE} + \\\\mathcal{L}_{Clique} + \\\\lambda\\\\|W\\\\|_{F}^{2} + \\\\beta KL$ & Anomaly Score & Nearest Distance to Cluster Centers \\\\\\\\ [2ex]\\n\\\\midrule[1 pt]\\n\\\\multirow{2}{*}{\\\\shortstack{Dynamic Graph - \\\\\\\\ Attributed}} &MTHL~ & Non-DP & $\\\\mathop{min}_{\\\\mathcal{P}}f(\\\\mathcal{P})$ & Anomaly Score & Distance to Hypersphere Centroid \\\\\\\\ \\\\cline{2-6}\\n&OCAN~ & GAN & $\\\\mathcal{L}_{LSTM-AE} + \\\\mathcal{L}_{GAN}$ & Anomaly Score & Discriminator's Output \\\\\\\\\\n\\\\bottomrule[1 pt]\\n\\\\multicolumn{6}{l}{* Non-DP: Non-Deep Learning Techniques, DNN: Deep NN Based Techniques, GCN: GCN Based Techniques, RL: Reinforcement Learning Based Techniques.} \\\\\\\\\\n\\\\multicolumn{6}{l}{* GAT: GAT Based Techniques, NR: Network Representation Based Techniques, GAN: Generative Adversarial Network Based Techniques.}\\n\\\\end{tabular}\\n}\\n\\\\label{table:ANOSND}\\n\\\\end{table*}\\n\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm} \\n\\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/Edge/AddGraph.pdf}}\\n\\\\caption{ANOS ED on dynamic graphs -- GCN based approaches. GCN is employed to learn node embeddings from the temporal graph at each timestamp. The attention-based GRU generates the current hidden state using the node embeddings and previous hidden states. The edge scoring function, such as a FCN, is learned to assign anomaly scores, and the top-k edges are depicted as anomalies.}\\n\\\\label{pic:Dynamic_Edge_Comparison}\\n\\\\end{figure*}\",\n \"id\": \"12f2fc41-29d6-4189-b62f-ff980f20a825\",\n \"level\": \"section\",\n \"origin_cites_number\": 34,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Dynamic Graphs\"\n ]\n ],\n \"subsections\": [\n \"0af4432d-d88f-455d-b4ca-b342865e37f7\",\n \"75faec9f-6952-423b-8944-196564d6aaa7\"\n ],\n \"title\": \"ANOS ND on Dynamic Graphs\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:ND:NR}\\nFollowing the research line of encoding a graph into an embedding space, after which anomaly detection is performed, dynamic network representation techniques have been investigated in the more recent works.\\nSpecifically, in~, Yu \\\\etal presented a flexible deep representation technique, called NetWalk, for detecting anomalous nodes in dynamic (plain) graphs using only the structure information.\\nIt adopts an autoencoder to learn node representations on the initial graph and incrementally updates them when new edges are added or existing edges are deleted.\\nTo detect anomalies, NetWalk first executes the streaming $k$-means clustering algorithm~ to group existing nodes in the current time stamp into different clusters. \\nThen, each node's anomaly score is measured with regard to its closest distance to the $k$ clusters. \\nWhen the node representations are updated, the cluster centers and anomaly scores are recalculated accordingly.\",\n \"id\": \"0af4432d-d88f-455d-b4ca-b342865e37f7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"12f2fc41-29d6-4189-b62f-ff980f20a825\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Dynamic Graphs\"\n ],\n [\n \"subsection\",\n \"Network Representation Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Representation Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 2626,\n 6285,\n 6292\n ],\n \"content\": \"\\\\label{sec:ND:GAN}\\nIn practice, anomaly detection is facing great challenges from the shortage of ground-truth anomalies.\\nConsequently, many research efforts have been invested in modeling the features of anomalies or regular objects such that anomalies can be identified effectively.\\nAmong these techniques, generative adversarial networks (GAN)~ have received extensive attention because of its impressive performance in capturing real data distribution and generating simulated data.\\nMotivated by the recent advances in ``bad\\\" GAN~, Zheng \\\\etal~ circumvented the fraudster detection problem using only the observed benign users' attributes.\\nThe basic idea is to seize the normal activity patterns and detect anomalies that behave significantly differently.\\nThe proposed method, OCAN, starts by extracting the benign users' content features using their historical social behaviors (\\\\eg historical posts, posts' URL), for which this method is classified into the dynamic category.\\nA long short-term memory (LSTM) based autoencoder~ is employed to achieve this and as assumed, benign users and malicious users are in separate regions in the feature space.\\nNext, a novel one-class adversarial net comprising a generator and a discriminator is trained.\\nSpecifically, the generator produces complementary data points that locate in the relatively low density areas of benign users.\\nThe discriminator, accordingly, aims to distinguish the generated samples from the benign users. \\nAfter training, benign users' regions are learned by the discriminator and anomalies can hence be identified with regard to their locations.\\nBoth NetWalk~ and OCAN~ approach the anomalous node detection problem promisingly, however, they respectively only consider the structure or attributes.\\nBy the success of static graph anomaly detection techniques that analyze both aspects, when the structure and attribute information in dynamic graphs are jointly considered, an enhanced detection performance can be foreseen.\\nWe therefore highlight this unexplored area for future works in Section~\\\\ref{sec:futures}.\",\n \"id\": \"75faec9f-6952-423b-8944-196564d6aaa7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"12f2fc41-29d6-4189-b62f-ff980f20a825\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Dynamic Graphs\"\n ],\n [\n \"subsection\",\n \"GAN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GAN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6293\n ],\n \"content\": \"\\\\label{sec:edge:static}\\nIn contrast to anomalous node detection, which targets individual nodes, ANOS ED aims to identify abnormal links.\\nThese links often inform the unexpected or unusual relationships between real objects~, such as the abnormal interactions between fraudsters and benign users shown in Fig.~\\\\ref{Toy}, or suspicious interactions between attacker nodes and benign user machines in computer networks. Following the previous taxonomy, in this section, we review the state-of-the-art ANOS ED methods for static graphs, and Section~\\\\ref{sec:edge:dynamic} summarizes the techniques for dynamic graphs. A summary is provided in Table~\\\\ref{tb:edgeandsubgraph}. This section includes methods based on deep NNs, GCNs and network representations. The non-deep learning techniques are reviewed in Appendix~\\\\ref{appendix:edge}.\",\n \"id\": \"678bd7af-5bf2-44f6-ae32-df413829eb80\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous edge detection (ANOS ED) \"\n ]\n ],\n \"subsections\": [\n \"509789f2-b5ef-4288-92b5-0d64443ed4af\",\n \"e519d916-fe62-4029-8d13-0a7fc3954141\",\n \"231d1246-99ce-4de5-9270-25f30b2ceb5d\"\n ],\n \"title\": \"Anomalous edge detection (ANOS ED) \"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Similar to deep NN based ANOS ND techniques, autoencoder and fully connected network (FCN) have also been used for anomalous edge detection.\\nAs an example, Ouyang \\\\etal~ approached the problem by modeling the distribution of edges through deep models to identify the existing edges that are least likely to appear as anomalies (as shown in Fig.~\\\\ref{pic:UGED}). The probability of each edge $e_{u,v}$ is decided by $P(v|u,N(u))$ and $P(u|v,N(v))$ which measure the edge probability using node $u$ with its neighbors $N(u)$ and node $v$ with its neighbors $N(v)$, respectively.\\nTo calculate $P(v|u,N(u))$, the proposed method, UGED, first encodes each node into a lower-dimensional vector through a FCN layer and generates node $u$'s representation by a mean aggregation of itself and its neighbors' vectors.\\nNext, the node representations are fed into another FCN to estimate $P(v|u,N(u))$.\\nThe prediction is expressed as $\\\\hat{P}(v|u,N(u)) = \\\\text{Softmax}(W\\\\cdot H(u))|_v$, where $W$ represents the trainable parameters, and $H(u)$ is $u$'s representation.\\nUGED's training scheme aims to maximize the prediction of existing edges via a cross-entropy-based loss function, $\\\\text{CE}(\\\\hat{P}(v|u,N(u)), v)$.\\nAfter training, an anomaly score is assigned to each edge using the average of $1-P(v|u,N(u))$ and $1-P(u|v,N(v))$. \\nAs such, existing edges that have a lower probability will get higher scores and the top-k edges are reported as anomalous.\",\n \"id\": \"509789f2-b5ef-4288-92b5-0d64443ed4af\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"678bd7af-5bf2-44f6-ae32-df413829eb80\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous edge detection (ANOS ED) \"\n ],\n [\n \"subsection\",\n \"Deep NN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep NN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Following the line of modeling edge distributions, some studies leverage GCNs to better capture the graph structure information.\\nDuan \\\\etal~ demonstrated that the existence of anomalous edges in the training data prevents traditional GCN based models from capturing real edge distributions, which leads to sub-optimal detection performance.\\nThis inherently raises a problem: to achieve better detection performance, the node embedding process should alleviate the negative impact of anomalous edges, but these edges are detected using the learned embeddings.\\nTo tackle this, the proposed method, AANE, jointly considers these two issues by iteratively updating the embeddings and detection results during training.\\nIn each training iteration, AANE generates node embeddings $Z$ through GCN layers and learns an indicator matrix $I$ to spot potential anomalous edges.\\nGiven an input graph $G$ with adjacency matrix $A$, each term $I_{uv}$ in $I$ is 1 if $\\\\hat{A}_{uv} < \\\\text{mean}_{v^{\\\\prime}\\\\in N_{u}} \\\\hat{A}_{uv^{\\\\prime}} - \\\\mu\\\\cdot\\\\text{std}_{v^{\\\\prime}\\\\in N_{u}} \\\\hat{A}_{uv^{\\\\prime}}$, and 0 otherwise.\\nHere, $\\\\hat{A}_{uv}$ is the predicted link probability between nodes $u$ and $v$, which is calculated as the hyperbolic tangent of $u$ and $v$'s embeddings, and $\\\\mu$ is a predefined threshold.\\nBy this, an edge $uv$ is identified as anomalous when its predicted probability is less than the average of all links associated with the node $u$ by a predefined threshold.\\nThe total loss function of AANE contains two parts: an anomaly-aware loss ($\\\\mathcal{L}_{aal}$) and an adjusted fitting loss ($\\\\mathcal{L}_{afl}$). \\n$\\\\mathcal{L}_{aal}$ is proposed to penalize the link prediction results and the indicator matrix $I$ such that anomalous edges will have lower prediction probabilities when they are marked as 1 in $I$. This is formulated as:\\n\\\\begin{equation}\\n \\\\mathcal{L}_{aal} = \\\\sqrt{\\\\sum_{u\\\\in V} \\\\sum_{v\\\\in N(u)}((1-\\\\hat{A}^2_{uv})(1-I_{uv}) + \\\\hat{A}^2_{uv}I_{uv})},\\n\\\\end{equation}\\nwhere $V$ is the node set, $N(u)$ is the set of $u$'s neighbors.\\n$\\\\mathcal{L}_{afl}$ quantifies the reconstruction loss with regard to the removal of potential anomalous edges, denoted as:\\n\\\\begin{equation}\\n \\\\mathcal{L}_{afl} = \\\\| B - \\\\hat{A}\\\\|^{2}_{2},\\n\\\\end{equation}\\nwhere $B$ is an adjusted adjacency matrix that removes all predicted anomalies from the input adjacency matrix $A$.\\nBy minimizing these two losses, AANE identifies the top-k edges with lowest probabilities as anomalies.\",\n \"id\": \"e519d916-fe62-4029-8d13-0a7fc3954141\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"678bd7af-5bf2-44f6-ae32-df413829eb80\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous edge detection (ANOS ED) \"\n ],\n [\n \"subsection\",\n \"GCN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GCN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Instead of using node embeddings for ANOS ED, edge representations learned directly from the graph are also feasible for distinguishing anomalies.\\nIf the edge representations well-preserve the graph structure and interaction content (\\\\eg messages in online social networks, co-authored papers in citation networks) between pairs of nodes, an enhanced detection performance can then be expected.\\nTo date, several studies, such as Xu \\\\etal~, have shown promising results in generating edge representations.\\nAlthough they are not specifically designed for graph anomaly detection, they pinpoint a potential approach to ANOS ED. This is highlighted as a potential future direction in Section~\\\\ref{sec:future:ED}.\",\n \"id\": \"231d1246-99ce-4de5-9270-25f30b2ceb5d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"678bd7af-5bf2-44f6-ae32-df413829eb80\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous edge detection (ANOS ED) \"\n ],\n [\n \"subsection\",\n \"Network Representation Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Representation Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:edge:dynamic}\\n\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n\\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/subgraph/SGD_Static.pdf}}\\n\\\\caption{ANOS SGD. Real-world networks are usually represented as bipartite graphs for reflecting the interactions between two different types of objects. To detect ANOS SGD, source nodes and sink nodes are embedded using two autoencoders (linked by a shared loss function), respectively. Anomalous sub-graphs are identified by applying dense region detection algorithms in the embedding space.}\\n\\\\label{pic:Staic_SG}\\n\\\\end{figure*}\\nDynamic graphs are powerful in reflecting the appearance/disappearance of edges over time~.\\nAnomalous edges can be distinguished by modeling the changes in graph structure and capturing the edge distributions at each time step.\\nRecent approaches to ANOS ED on dynamic graphs are reviewed in this section.\",\n \"id\": \"78c71045-f884-4e7f-b2a8-d7769a5bc16b\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ED on Dynamic Graphs\"\n ]\n ],\n \"subsections\": [\n \"af8f0ff9-09a3-46b3-96d8-edda541a9029\",\n \"946969f4-52c1-445f-99a0-7e162c350a60\"\n ],\n \"title\": \"ANOS ED on Dynamic Graphs\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The intuition of network representation based techniques is to encode the dynamic graph structure information into edge representations and apply the aforementioned traditional anomaly detection techniques to spot irregular edges.\\nThis is quite straightforward, but there remain vital challenges in generating/updating informative edge representations when the graph structure evolves.\\nTo mitigate this challenge, the ANOS ND model NetWalk~ is also capable of detecting anomalous edges in dynamic graphs. \\nFollowing the line of distance-based anomaly detection, NetWalk encodes edges into a shared latent space using node embeddings, and anomalies are identified based on their distances to the nearest edge-cluster centers in the latent space.\\nPractically, Netwalk generates edge representations as the Hadamard product of the source and destination nodes' representations, denoted as: $\\\\mathbf{z}_{u,v} = \\\\mathbf{z}_u \\\\odot \\\\mathbf{z}_v$.\\nWhen new edges arrive or existing edges disappear, the node and edge representations are updated from random walks in the temporary graphs at each time stamp, after which the edge-cluster centers and edge anomaly scores are recalculated. \\nFinally, the top-k farthest edges to the edge-clusters are reported as anomalies.\",\n \"id\": \"af8f0ff9-09a3-46b3-96d8-edda541a9029\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"78c71045-f884-4e7f-b2a8-d7769a5bc16b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ED on Dynamic Graphs\"\n ],\n [\n \"subsection\",\n \"Network Representation Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Representation Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Although NetWalk is capable of detecting anomalies in dynamic graphs, it simply updates edge representations without considering the evolving patterns of long/short-term nodes and the graph's structure.\\nFor more effective ANOS ED, Zheng \\\\etal~ intuitively combined temporal, structural and attribute information to measure the anomalousness of edges in dynamic graphs. \\nThey propose a semi-supervised model, AddGraph, which comprises a GCN and Gated Recurrent Units (GRU) with attention~ to capture more representative structural information from the temporal graph in each time stamp and dependencies between them, respectively.\\nAt each time stamp $t$, GCN takes the output hidden state ($H^{t-1}$) at time $t-1$ to generate node embeddings, after which the GRU learns the current hidden state $H^{t}$ from the node embeddings and attentions on previous hidden states (as shown in Fig.~\\\\ref{pic:Dynamic_Edge_Comparison}).\\nAfter getting the hidden state $H^{t}$ of all nodes, AddGraph assigns an anomaly score to each edge in the temporal graph based on the nodes associated with it. The proposed anomaly scoring function is formulated as:\\n\\\\begin{equation}\\\\label{eq:AddGraph:scoring}\\n f(u,v,w) = w\\\\cdot\\\\sigma(\\\\beta\\\\cdot (||\\\\mathbf{a}\\\\odot \\\\mathbf{h}_{u} + \\\\mathbf{b}\\\\odot \\\\mathbf{h}_{v}||^{2}_{2} - \\\\mu)),\\n\\\\end{equation}\\nwhere $u$ and $v$ are the corresponding nodes, $w$ is the weight of the edge, $\\\\mathbf{a}$ and $\\\\mathbf{b}$ are trainable parameters, $\\\\beta$ and $\\\\mu$ are hyper-parameters, and $\\\\sigma(\\\\cdot)$ is the non-linear activation function.\\nTo learn $\\\\mathbf{a}$ and $\\\\mathbf{b}$, Zheng \\\\etal further assumed that all existing edges in the dynamic graph are normal in the training stage, and sampled non-existing edges as anomalies. Specifically, they form the loss function as:\\n\\\\begin{equation}\\\\label{eq:AddGraph:loss}\\n \\\\begin{split}\\n \\\\mathcal{L}_{AddGraph} = &\\\\min\\\\sum_{{(u,v,w)\\\\in \\\\varepsilon^{t}}}\\\\sum_{{(u^{\\\\prime},v^{\\\\prime},w)\\\\notin \\\\varepsilon^{t}}} \\\\\\\\\\n & \\\\max\\\\{0, \\\\gamma + f(u,v,w) - f(u^{\\\\prime},v^{\\\\prime},w)\\\\} + \\\\lambda\\\\mathcal{L}_{reg},\\n \\\\end{split}\\n\\\\end{equation}\\nwhere $\\\\varepsilon^{t}$ is the edge set, $(u^{\\\\prime},v^{\\\\prime})$ are sampled non-existing edges at time stamp $t$, $\\\\lambda$ is a hyper-parameter, and $\\\\mathcal{L}_{reg}$ regularizes all trainable parameters in the model. \\nAfter training, the scoring function identifies anomalous edges in the test data by assigning higher anomaly scores to them based on Eq.~\\\\ref{eq:AddGraph:scoring}.\",\n \"id\": \"946969f4-52c1-445f-99a0-7e162c350a60\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"78c71045-f884-4e7f-b2a8-d7769a5bc16b\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ED on Dynamic Graphs\"\n ],\n [\n \"subsection\",\n \"GCN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GCN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:subgraph}\\nIn real life, anomalies might also collude and behave collectively with others to garner benefits.\\nFor instance, fraudulent user groups in an online review network, as shown in Fig.~\\\\ref{Toy}, may post misleading reviews to promote or besmirch certain merchandise. When these data are represented as graphs, anomalies and their interactions usually form suspicious sub-graphs, and ANOS SGD is proposed to distinguish them from the benign.\\n\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n\\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/graph/graph.pdf}}\\n\\\\caption{ANOS GD on Graph Database. Generally, the graph-level anomaly detection techniques take a graph database as input. By generating embeddings/hidden state for each single graph in the database through D(G)NNs, anomalous graphs can be depicted by One-class classifiers or MLP.}\\n\\\\label{pic:graph_level}\\n\\\\end{figure*}\\nUnlike individual and independent graph anomalies, \\\\ie single nodes or edges, each node and edge in a suspicious sub-graph might be normal.\\nHowever, when considered as a collection, they turn out to be anomalous.\\nMoreover, these sub-graphs also vary in size and inner structure, making anomalous sub-graph detection more challenging than ANOS ND/ED~.\\nAlthough extensive effort has been placed on circumventing this problem, deep-learning techniques have only begun to address this problem in the last five years.\\nFor reference, traditional non-deep learning based techniques are briefly introduced in Appendix~\\\\ref{appendix:subgraph}, and a summary of techniques reviewed for ANOS SGD is provided in Table~\\\\ref{tb:edgeandsubgraph} at the end of Section~\\\\ref{sec:anosgd:dynamic}.\\nDue to the flexibility of heterogeneous graphs in representing the complex relationships between different kinds of real objects, several recent works have taken advantage of deep network representation techniques to detect real-world anomalies through ANOS SGD.\\nFor instance, Wang \\\\etal~ represented online shopping networks as bipartite graphs (a specific type of heterogeneous graph that has two types of nodes and one type of edge), in which users are source nodes and items are sink nodes.\\nFraudulent groups are then detected based on suspicious dense blocks that form in these graphs.\\nWang \\\\etal~ aimed to learn anomaly-aware representations of users such that suspicious users in the same group will be located closely in the vector space, while benign users will be far away (as shown in the embedding space in Fig.~\\\\ref{pic:Staic_SG}). \\nAccording to the observation that user nodes belonging to one fraudulent group are more likely to connect with the same item nodes, the developed model, DeepFD, measures similarities in the behavior of two users, $sim_{ij}$, as the percentage of items shared among all the items they have reviewed.\\nUser representations are then generated through a traditional autoencoder, which is trained using three losses and follows the encoding-decoding process.\\nThe first loss is the reconstruction loss $\\\\mathcal{L}_{res}$ that ensures the bipartite graph structure can be reconstructed properly using the learned user representations and item representations.\\nThe second term $\\\\mathcal{L}_{sim}$ preserves the user similarity information in the learned user representations. That is, if two users have similar behaviors, their representations should also be similar. This loss is formulated as:\\n\\\\begin{equation}\\\\label{eq:DeepFD:simloss}\\n \\\\mathcal{L}_{sim} = \\\\sum_{i,j=1}^{m} sim_{ij} \\\\cdot \\\\| \\\\widehat{sim}_{ij} - sim_{ij} \\\\|_{2}^{2},\\n\\\\end{equation}\\nwhere $m$ is the number of user nodes, $\\\\widehat{sim}_{ij}$ measures the similarity of user $i$ and $j$'s representations using an RBF kernel or other alternative.\\nThe third loss $\\\\mathcal{L}_{reg}$ regularizes all trainable parameters.\\nFinally, the suspicious dense blocks, which are expected to form dense regions in the vector space, are detected using DBSCAN~.\\nAnother work, FraudNE~, also models online review networks as bipartite graphs and further detects both malicious users and associated manipulated items following the dense block detection principle.\\nUnlike DeepFD, FraudNE aspires to encode both types of nodes into a shared latent space where suspicious users and items belonging to the same dense block are very close to each other while others distribute uniformly (as shown in Fig.~\\\\ref{pic:Staic_SG}).\\nFraudNE adopts two traditional autoencoders, namely, a source node autoencoder and a sink node autoencoder, to learn user representations and item representations, respectively.\\nBoth autoencoders are trained to jointly minimize their corresponding reconstruction losses and a shared loss function, and the total loss can be formulated as:\\n\\\\begin{equation}\\n \\\\mathcal{L}_{FraudNE} = \\\\mathcal{L}_{res}^{source} + \\\\mathcal{L}_{res}^{sink} + \\\\alpha \\\\mathcal{L}_{share} + \\\\eta \\\\mathcal{L}_{reg},\\n\\\\end{equation}\\nwhere $\\\\alpha$ and $\\\\eta$ are hyperparameters, and $\\\\mathcal{L}_{reg}$ regularizes all trainable parameters.\\nSpecifically, the reconstruction losses (\\\\ie $\\\\mathcal{L}_{res}^{source}$ and $\\\\mathcal{L}_{res}^{sink}$) measure the gap between the input user/item features (extracted from the graph structure) and their decoded features. \\nThe shared loss function is proposed to restrict the representation learning process such that each linked pair of users and items get similar representations.\\nAs the DBSCAN~ algorithm is convenient to apply for dense region detection, FraudNE also uses it to distinguish the dense sub-graphs formed by suspicious users and items.\\nTo date, only a few works have put their efforts into using deep learning techniques for ANOS SGD. However, with intensifying research interest in sub-graph representation learning, we encourage more studies on ANOS SGD and highlight this as a potential future in Section~\\\\ref{sec:future:ED}.\\n\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n \\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/graph/dynamic.pdf}}\\n \\\\caption{ANOS GD on dynamic graphs. For each graph snapshot in the dynamic graph, the LSTM autoencoder generates its hidden state using its adjacency matrix and previous hidden states. Through hypersphere learning, a hypersphere with centroid $a$ and radius $r$ is learned such that anomalous snapshots lay outside.}\\n \\\\label{pic:GDdynamic}\\n\\\\end{figure*}\",\n \"id\": \"5da2d684-654a-4048-95ec-2ae406b319b2\",\n \"level\": \"section\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous sub-graph detection (ANOS SGD)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Anomalous sub-graph detection (ANOS SGD)\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 8992\n ],\n \"content\": \"\\\\label{sec:anosgd:db}\\nBeyond anomalous node, edge, and sub-graph, graph anomalies might also appear as abnormal graphs in a set/database of graphs.\\nTypically, a graph database is defined as:\\n\\\\textbf{\\\\textit{Definition 4 (Graph Database)}}. A graph database $\\\\mathcal{G} = \\\\{ G_i = (V_i,E_i,X_v(i), X_e(i)) \\\\}^N_{i=1}$ contains $N$ individual graphs. Here, each graph $G_i$ is comprised of a node set $V_i$ and an edge set $E_i$. $X_v(i)$ and $X_e(i)$ are the node attribute matrix and edge attribute matrix of $G_i$ if it is an attributed graph.\\nThis graph-level ANOS GD aims to detect individual graphs that deviate significantly from the others.\\nA concrete example of ANOS GD is unusual molecule detection.\\nWhen chemical compounds are represented as molecular/chemical graphs where the atoms and bonds are represented as nodes and edges~, unusual molecules can be identified because their corresponding graphs have structures and/or features that deviate from the others.\\nBrain disorders detection is another example.\\nA brain disorder can be diagnosed by analyzing the dynamics of brain graphs at different stages of aging in sequence and finding an inconsistent snapshot at a specific time stamp.\\nThe prior reviewed techniques (\\\\ie ANOS ND/ED/SGD) are not compatible with ANOS GD because they are dedicated to detecting anomalies in a single graph, whereas ANOS GD is directed at detecting graph-level anomalies.\\nThis problem is commonly approached by: 1) measuring the pairwise proximities of graphs using graph kernels~; 2) detecting the appearance of anomalous graph signals created by abnormal groups of nodes~; or 3) encoding graphs using frequent motifs~.\\nHowever, none of these methods are deep learning-based.\\nAs the time of writing, very few studies in ANOS GD with deep learning have been undertaken. As such, this is highlighted as a potential future direction in Section~\\\\ref{sec:future:ED}.\",\n \"id\": \"9ba15010-a3db-4588-8e7f-56a7def289f5\",\n \"level\": \"section\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous graph detection (ANOS GD)\"\n ]\n ],\n \"subsections\": [\n \"49385046-24ac-4a52-8807-655e25c7942e\",\n \"c5bcea84-0572-429e-a5e7-37efbed81759\"\n ],\n \"title\": \"Anomalous graph detection (ANOS GD)\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 8149\n ],\n \"content\": \"Motivated by the success of GNNs in various graph classification tasks, the most recent works in ANOS GD employ GNNs to classify single graphs as normal/abnormal in the given graph database. Specifically, Dou \\\\etal~ transformed fake news detection into an ANOS GD problem by modeling news as tree-structured propagation graphs where the root nodes denote pieces of news, and child nodes denote users who interact with the root news. Their end-to-end framework, UPFD, extracts two embeddings for the news piece and users, respectively, via a text embedding model (e.g. word2vec, BERT) and a user engagement embedding process. For each news graph, its latent representation is a flattened concatenation of these two embeddings, which is input to train a neural classifier with the label of the news. \\nCorresponding propagation graphs that are labeled as fake by the trained model are regarded as anomalous.\\nAnother representative work by Zhao and Akoglu~ employed a GIN model and one-class classification (\\\\ie DeepSVDD~) loss to train a graph-level anomaly detection framework in an end-to-end manner.\\nFor each individual graph in the graph database, its graph-level embedding is generated by applying mean-pooling over its nodes' node-level embeddings.\\nA graph is eventually depicted as anomalous if it lies outside the learned hypersphere, as shown in Fig.~\\\\ref{pic:graph_level}.\",\n \"id\": \"49385046-24ac-4a52-8807-655e25c7942e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"9ba15010-a3db-4588-8e7f-56a7def289f5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous graph detection (ANOS GD)\"\n ],\n [\n \"subsection\",\n \"GNN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GNN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 3953\n ],\n \"content\": \"It is also possible to apply general graph-level network representation techniques to ANOS GD.\\nWith these methods, the detection problem is transformed into a conventional outlier detection problem in the embedding space.\\nIn contrast to D(G)NN based techniques that can detect graph anomalies in an end-to-end manner, adopting these representation techniques for anomaly detection is two-staged.\\nFirst, graphs in the database are encoded into a shared latent space using graph-level representation techniques, such as Graph2Vec~, FGSD~.\\nThen, the anomalousness of each single graph is measured by an off-the-shelf outlier detector.\\nEssentially, this kind of approach involves pairing existing methods in both stages, yet, the stages are disconnected from each other and, hence, the detection performance can be subpar since the embedding similarities are not necessarily designed for the sake of anomaly detection.\",\n \"id\": \"c5bcea84-0572-429e-a5e7-37efbed81759\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"9ba15010-a3db-4588-8e7f-56a7def289f5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Anomalous graph detection (ANOS GD)\"\n ],\n [\n \"subsection\",\n \"Network Representation Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Representation Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0.412698412698412,\n \"cites\": [\n 8993,\n 6285,\n 6281,\n 6302,\n 6297,\n 3333,\n 6283,\n 6303,\n 6296,\n 6289,\n 3700,\n 6294,\n 6304,\n 6299,\n 6271,\n 6301,\n 1420,\n 6277,\n 6295,\n 6300,\n 6293,\n 5645,\n 8149,\n 6291,\n 6286,\n 6298\n ],\n \"content\": \"\\\\label{sec:anosgd:dynamic}\\nFor dynamic graph environments, graph-level anomaly detection endeavors to identify abnormal graph snapshots/temporal graphs. Similar to ANOS ND and ED on dynamic graphs, given a sequence of graphs, anomalous graphs can be distinguished regarding their unusual evolving patterns, abnormal graph-level features, or other characteristics.\\n\\\\begin{table*}[!h]\\n\\\\centering \\n\\\\footnotesize\\n\\\\renewcommand\\\\arraystretch{1}\\n\\\\setlength{\\\\tabcolsep}{2.8mm}\\n\\\\caption{Summary of Anomalous Edge, Sub-graph and Graph Detection Techniques.} \\n\\\\resizebox{0.96\\\\textwidth}{!}{\\n\\\\begin{tabular}{m{2.1cm}<{\\\\centering}|m{2.3cm}<{\\\\centering}|m{1.1cm}<{\\\\centering}|m{4cm}<{\\\\centering}|m{1.8cm}<{\\\\centering}|m{2.9cm}<{\\\\centering}}\\n\\\\toprule[1 pt]\\n\\\\textbf{Graph Type} & \\\\textbf{Approach} & \\\\textbf{Category} & \\\\textbf{Objective Function} & \\\\textbf{Measurement} & \\\\textbf{Outputs} \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\multicolumn{6}{c}{Anomalous Edge Detection Techniques} \\\\\\\\\\n\\\\midrule[1 pt]\\n\\\\multirow{2}{*}{\\\\shortstack{Static Graph - \\\\\\\\ Plain}} &UGED~ & DNN & $\\\\text{cross-entropy}(f(u,N(u)), v)$ & Anomaly Score & $\\\\text{mean}(1-P(v|u,N(u)), 1-P(u|v,N(v)))$ \\\\\\\\ \\\\cline{2-6}\\n&AANE~ & GCN & $\\\\mathcal{L} = \\\\mathcal{L}_{afl} + \\\\gamma\\\\mathcal{L}_{aal}$ & Anomaly Ranking & Edge Existing Probability \\\\\\\\ \\n\\\\midrule[1 pt]\\nStatic Graph - Attributed &eFraudCom~ & NR & $\\\\mathcal{L}_{MAIN} - \\\\lambda\\\\mathcal{L}_{MI}$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\n\\\\midrule[1 pt]\\nDynamic Graph - Plain & NetWalk~ & NR & $\\\\gamma\\\\mathcal{L}_{AE} + \\\\mathcal{L}_{Clique} + \\\\lambda\\\\|W\\\\|_{F}^{2} + \\\\beta KL$ & Anomaly Score & Nearest Distance to Cluster Centers \\\\\\\\\\n\\\\midrule[1 pt]\\nDynamic Graph - Attributed & AddGraph~ & GCN & $\\\\text{min} \\\\sum_{{e\\\\in \\\\varepsilon^{t}}}\\\\sum_{{e^{\\\\prime}\\\\notin \\\\varepsilon^{t}}} \\\\text{max}\\\\{0, \\\\gamma + f(i,j,w) - f(i^{\\\\prime},j^{\\\\prime},w)\\\\} + \\\\lambda\\\\mathcal{L}_{reg}$ & Anomaly Score & $f(i,j,w) = w\\\\cdot\\\\sigma(\\\\beta\\\\cdot (||\\\\mathbf{a} \\\\odot \\\\mathbf{h}_{i} + \\\\mathbf{b} \\\\odot \\\\mathbf{h}_{j}||^{2}_{2} - \\\\mu))$ \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\midrule[1 pt]\\n\\\\multicolumn{6}{c}{Anomalous Sub-graph Detection Techniques} \\\\\\\\\\n\\\\midrule[1 pt]\\n\\\\multirow{2}{*}{\\\\shortstack{Static Graph - \\\\\\\\ Plain}} &DeepFD~ & NR & $\\\\mathcal{L}_{recon} +\\\\alpha \\\\mathcal{L}_{sim} + \\\\gamma \\\\mathcal{L}_{reg}$& Density-based Method (DBSCAN) & Dense sub-graphs \\\\\\\\ \\\\cline{2-6}\\n&FraudNE~ & NR & $\\\\mathcal{L}_{res}^{source} + \\\\mathcal{L}_{res}^{sink} + \\\\alpha \\\\mathcal{L}_{share} + \\\\eta \\\\mathcal{L}_{reg}$ & Density-based Method (DBSCAN) & Dense sub-graphs \\\\\\\\\\n\\\\midrule[1 pt]\\n\\\\midrule[1 pt]\\n\\\\multicolumn{6}{c}{Anomalous Graph Detection Techniques} \\\\\\\\\\n\\\\midrule[1 pt]\\n\\\\multirow{2}{*}{\\\\shortstack{Graph Database - \\\\\\\\ Attributed}}&UPFD~ & NR & $ -(y\\\\log(p) + (1-y)\\\\log(1-p))$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\\\cline{2-6}\\n& OCGIN~ & GNN & $ \\\\min\\\\limits_{W}\\\\frac{1}{N}\\\\sum\\\\limits_{i=1}^{N}||GIN(G_i,W)-c||^2 + \\\\frac{\\\\lambda}{2}\\\\sum\\\\limits_{l=1}^{L}||W^l||^2_{F}$ & Location in Embedding Space & Distance to Hypersphere Center \\\\\\\\\\n\\\\midrule[1 pt]\\n\\\\multirow{2}{*}{\\\\shortstack{Dynamic Graph - \\\\\\\\ Plain}} & DeepSphere~ & DNN & $\\\\mathcal{L} = \\\\mathcal{L}_{h} + \\\\lambda\\\\mathcal{L}_{res}$ & Location in Embedding Space & Anomalous Label \\\\\\\\ \\\\cline{2-6}\\n& GLAD-PAW~ & GNN & $\\\\text{cross-entropy}(\\\\mathbf{y},\\\\mathbf{\\\\hat{y}})$ & Anomaly Prediction & Predicted Label \\\\\\\\ \\n\\\\bottomrule[1 pt]\\n\\\\multicolumn{6}{l}{* DNN: Deep NN Based techniques, GCN: GCN Based Techniques, NR: Network Representation Based Techniques.}\\\\\\\\\\n\\\\multicolumn{6}{l}{* GNN: Graph Neural Network Based Techniques.}\\n\\\\end{tabular}\\n\\\\label{tb:edgeandsubgraph}\\n}\\n\\\\end{table*}\\nIn order to derive each graph snapshot/temporal graph’s characteristics, the commonly used GNN, LSTM and autoencoder are feasible to apply. For instance, Teng \\\\etal~ applied a LSTM-autoencoder to detect abnormal graph snapshots, as shown in Fig.~\\\\ref{pic:GDdynamic}.\\nIn their proposed model, DeepSphere, a dynamic graph is described as a collection of three-order tensors, $\\\\{\\\\mathcal{X}_k, k=1,2...\\\\}$ where each $\\\\mathcal{X} \\\\in \\\\mathcal{R}^{N \\\\times N \\\\times T}$, and the slices along the time dimension are the adjacency matrices of graph snapshots.\\nTo identify abnormal tensors, DeepSphere first embeds each graph snapshot into a latent space using an LSTM autoencoder, and then leverages a one-class classification objective~ that learns a hypersphere such that normal snapshots are covered, and anomalous snapshots lay outside.\\nThe LSTM autoencoder takes the adjacency matrices as input sequentially and attempts to reconstruct these input matrices through training.\\nThe hypersphere is learned through a single neural network layer and its objective function is formulated as:\\n\\\\begin{equation} \\\\label{eq:ds:hsp}\\n \\\\mathcal{L}_{h} = r^{2} + \\\\gamma\\\\sum_{k=1}^{m} \\\\epsilon_{k} + \\\\frac{1}{m}\\\\sum_{k=1}^{m} \\\\|\\\\textbf{z}_{k} - \\\\textbf{a}\\\\|^{2},\\n\\\\end{equation}\\nwhere $\\\\textbf{z}_{k}$ is the latent representation generated by the LSTM autoencoder, $\\\\textbf{a}$ is the centroid of the hypersphere, $r$ is the radius, $\\\\epsilon_{k}$ is the outlier penalty ($\\\\epsilon_{k} = \\\\|\\\\textbf{z}_{k} - \\\\textbf{a}\\\\|^{2} - r^2$), $m$ is the number of training graph snapshots, and $\\\\gamma$ is a hyperparameter. \\nThe overall objective function of DeepSphere is represented as:\\n\\\\begin{equation} \\\\label{eq:ds:obj}\\n \\\\mathcal{L} = \\\\mathcal{L}_{h} + \\\\lambda\\\\mathcal{L}_{res},\\n\\\\end{equation}\\nwhere $\\\\mathcal{L}_{res}$ is the reconstruction loss of the LSTM autoencoder.\\nWhen the training is finished, DeepSphere spots a given unseen data $\\\\mathcal{X}$ as anomalous if its embedding lies outside the learned hypersphere with a radius of $r$.\\nIn addition to all ANOS ND, ED, SGD, and GD techniques reviewed above, it is worth mentioning that perturbed graphs, which adversarial models generate to attack graph classification algorithms or GNNs~, can also be regarded as (intensional) anomalies. In a perturbed graph, the nodes and edges are modified deliberately to deviate from the others. We have not reviewed these in this survey because their main purpose is to attack a GNN model. The key idea behind these methods is the attacking/perturbation strategy, and studies in this sphere seldom focus on a detection or reasoning module to identify the perturbed graph or its sub-structures, \\\\ie anomalous nodes, edges, sub-graphs, or graphs.\\n\\\\begin{table*}[!t]\\n\\\\centering \\n\\\\footnotesize\\n\\\\renewcommand\\\\arraystretch{1}\\n\\\\setlength{\\\\tabcolsep}{2.8mm}\\n\\\\caption{Published Algorithms and Models} \\n\\\\resizebox{0.96\\\\textwidth}{!}{\\n\\\\begin{tabular}{c|c|c|c|l}\\n\\\\toprule[1 pt]\\n\\\\textbf{Model} & \\\\textbf{Language} & \\\\textbf{Platform} & \\\\textbf{Graph} & \\\\textbf{Code Repository} \\\\\\\\ \\n\\\\midrule[1 pt]\\nAnomalyDAE~ & Python & Tensorflow & Static Attributed Graph & https://github.com/haoyfan/AnomalyDAE \\\\\\\\ \\\\hline\\nMADAN~ & Python & - & Static Attributed Graph & https://github.com/leoguti85/MADAN \\\\\\\\ \\\\hline\\nPAICAN~ & Python & Tensorflow & Static Attributed Graph & http://www.kdd.in.tum.de/PAICAN/ \\\\\\\\ \\\\hline\\nONE~ & Python & - & Static Attributed Graph & https://github.com/sambaranban/ONE \\\\\\\\ \\\\hline\\nDONE\\\\&AdONE~ & Python & Tensorflow & Static Attributed Graph & https://bit.ly/35A2xHs \\\\\\\\ \\\\hline\\nSLICENDICE~ & Python & - & Static Attributed Graph & http://github.com/hamedn/SliceNDice/ \\\\\\\\ \\\\hline\\nFRAUDRE~ & Python & Pytorch & Static Attributed Graph & https://github.com/FraudDetection/FRAUDRE \\\\\\\\ \\\\hline\\nSemiGNN~ & Python & Tensorflow & Static Attributed Graph & https://github.com/safe-graph/DGFraud \\\\\\\\ \\\\hline\\nCARE-GNN~ & Python & Pytorch & Static Attributed Graph & https://github.com/YingtongDou/CARE-GNN \\\\\\\\ \\\\hline\\nGraphConsis~ & Python & Tensorflow & Static Attributed Graph & https://github.com/safe-graph/DGFraud \\\\\\\\ \\\\hline\\nGLOD~ & Python & Pytorch & Static Attributed Graph & https://github.com/LingxiaoShawn/GLOD-Issues \\\\\\\\ \\\\hline\\nOCAN~ & Python & Tensorflow & Static Graph &https://github.com/PanpanZheng/OCAN \\\\\\\\ \\\\hline\\nDeFrauder~ & Python & - & Static Graph & https://github.com/LCS2-IIITD/DeFrauder \\\\\\\\ \\\\hline\\nGCAN~ & Python & Keras & Heterogeneous Graph & https://github.com/l852888/GCAN \\\\\\\\ \\\\hline\\nHGATRD~ & Python & Pytorch & Heterogeneous Graph & https://github.com/201518018629031/HGATRD \\\\\\\\\\\\hline\\nGLAN~ & Python & Pytorch & Heterogeneous Graph & https://github.com/chunyuanY/RumorDetection \\\\\\\\ \\\\hline\\nGEM~ & Python & - & Heterogeneous Graph & https://github.com/safe-graph/DGFraud/tree/master/algorithms/GEM \\\\\\\\ \\\\hline\\neFraudCom~ & Python & Pytorch & Heterogeneous Graph & https://github.com/GeZhangMQ/eFraudCom \\\\\\\\ \\\\hline\\nDeepFD~ & Python & Pytorch & Bipartite Graph & https://github.com/JiaWu-Repository/DeepFD-pyTorch \\\\\\\\ \\\\hline\\nANOMRANK~ & C++ & - & Dynamic Graph & https://github.com/minjiyoon/anomrank \\\\\\\\ \\\\hline\\nMIDAS~ & C++ & - & Dynamic Graph & https://github.com/Stream-AD/MIDAS \\\\\\\\ \\\\hline\\nSedanspot~ & C++ & - & Dynamic Graph & https://www.github.com/dhivyaeswaran/sedanspot \\\\\\\\ \\\\hline\\nF-FADE~ & Python & Pytorch & Dynamic Graph & http://snap.stanford.edu/f-fade/ \\\\\\\\ \\\\hline\\nDeepSphere~ & Python & Tensorflow & Dynamic Graph & https://github.com/picsolab/DeepSphere \\\\\\\\ \\\\hline\\nChangedar~ & Matlab & - & Dynamic Graph & https://bhooi.github.io/changedar/ \\\\\\\\\\\\hline\\nUPFD~ & Python & Pytorch & Graph Database & https://github.com/safe-graph/GNN-FakeNews \\\\\\\\ \\\\hline\\nOCGIN~ & Python & Pytorch & Graph Database & https://github.com/LingxiaoShawn/GLOD-Issues \\\\\\\\ \\\\hline\\nDAGMM~ & Python & Pytorch & Non Graph & https://github.com/danieltan07/dagmm \\\\\\\\ \\\\hline\\nDevNet~ & Python & Tensorflow & Non Graph &https://github.com/GuansongPang/deviation-network \\\\\\\\ \\\\hline\\nRDA~ & Python & Tensorflow & Non Graph &https://github.com/zc8340311/RobustAutoencoder \\\\\\\\ \\\\hline\\nGAD~ & Python & Tensorflow & Non Graph &https://github.com/raghavchalapathy/gad \\\\\\\\ \\\\hline\\nDeep SAD~ &Python & Pytorch & Non Graph & https://github.com/lukasruff/Deep-SAD-PyTorch \\\\\\\\ \\\\hline\\nDATE~ &Python & Pytorch &Non Graph & https://github.com/Roytsai27/Dual-Attentive-Tree-aware-Embedding \\\\\\\\ \\\\hline\\nSTS-NN~ & Python & Pytorch & Non Graph & https://github.com/JiaWu-Repository/STS-NN \\\\\\\\\\n\\\\bottomrule[1 pt]\\n\\\\multicolumn{5}{l}{* -: No Dedicated Platforms.}\\n\\\\end{tabular}\\n\\\\label{tb:publishedalgs}\\n}\\n\\\\end{table*}\\n\\\\begin{table*}[!t]\\n\\\\centering \\n\\\\footnotesize\\n\\\\renewcommand\\\\arraystretch{1}\\n\\\\setlength{\\\\tabcolsep}{2.8mm}\\n\\\\caption{Published Datasets.}\\n\\\\resizebox{0.96\\\\textwidth}{!}{\\n\\\\begin{tabular}{m{1.6cm}<{\\\\centering}|m{1.4cm}<{\\\\centering}|m{0.5cm}<{\\\\centering}|m{0.5cm}<{\\\\centering}|m{0.5cm}<{\\\\centering}|m{0.5cm}<{\\\\centering}|m{0.5cm}<{\\\\centering}|m{2cm}<{\\\\centering}|l}\\n\\\\toprule[1 pt]\\n\\\\textbf{Category} & \\\\textbf{Dataset} & \\\\textbf{\\\\#G} & \\\\textbf{\\\\#N} & \\\\textbf{\\\\#E} & \\\\textbf{\\\\#FT} & \\\\textbf{\\\\#AN} & \\\\textbf{REF} & \\\\textbf{URL} \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\multirow{5}{*}{\\\\shortstack{Citation \\\\\\\\ Networks}} \\n& ACM & 1 & 16K & 71K & 8.3K & - & &\\\\text{http://www.arnetminer.org/open-academic-graph} \\\\\\\\ \\\\cline{2-9}\\n& Cora & 1 & 2.7K & 5.2K & 1.4K & - & &http://linqs.cs.umd.edu/projects/projects/lbc \\\\\\\\ \\\\cline{2-9}\\n& Citeseer & 1& 3.3K & 4.7K & 3.7K & - & & http://linqs.cs.umd.edu/projects/projects/lbc \\\\\\\\ \\\\cline{2-9}\\n& Pubmed & 1& 19K & 44K & 500 & - & &http://linqs.cs.umd.edu/projects/projects/lbc \\\\\\\\ \\\\cline{2-9}\\n& DBLP & 1& - & - & - & - & &http://www.informatik.uni-trier.de/˜ley/db/ \\\\\\\\\\n\\\\midrule[1 pt]\\n\\\\multirow{10}{*}{\\\\shortstack{Social \\\\\\\\ Networks}}\\n&Enron & - & 80K & - & - & - & & http://odds.cs.stonybrook.edu/\\\\#table2 \\\\\\\\ \\\\cline{2-9}\\n&UCI Message & 1& 5K & - & - & - & &http://archive.ics.uci.edu/ml \\\\\\\\ \\\\cline{2-9}\\n&Google+ & 4 & 75M & 11G & - & - & - & https://wangbinghui.net/dataset.html \\\\\\\\ \\\\cline{2-9}\\n&Twitter Sybil & 3 & 41M & - & - & 100K & - & https://wangbinghui.net/dataset.html \\\\\\\\ \\\\cline{2-9}\\n&Twitter WorldCup2014 & - & 54K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\\\\\ \\\\cline{2-9}\\n&Twitter Security2014 & - & 130K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\\\\\ \\\\cline{2-9}\\n&Reality Mining & - & 9.1K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\\\\\ \\\\cline{2-9}\\n&NYTNews & - & 320K & - & - & - & & http://shebuti.com/SelectiveAnomalyEnsemble/ \\\\\\\\ \\\\cline{2-9}\\n&Politifact & 314 & 41K & 40K & - & 157 & &https://github.com/safe-graph/GNN-FakeNews \\\\\\\\ \\\\cline{2-9}\\n&Gossipcop & 5.4K & 314K & 308K & - & 2.7K & &https://github.com/safe-graph/GNN-FakeNews \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\multirow{5}{*}{\\\\shortstack{Co-purchasing \\\\\\\\ Networks}}\\n& Disney & 1 & 124 & 334 & 30 & 6 & &https://www.ipd.kit.edu/mitarbeiter/muellere/consub/ \\\\\\\\ \\\\cline{2-9}\\n&Amazon-v1 & 1 & 314K & 882K & 28 & 6.2K & &https://www.ipd.kit.edu/mitarbeiter/muellere/consub/ \\\\\\\\ \\\\cline{2-9}\\n&Amazon-v2 & 1 & 11K & - & 25 & 821 & -&https://github.com/dmlc/dgl/blob/master/python/dgl/data/fraud.py \\\\\\\\ \\\\cline{2-9}\\n&Elliptic & 1 & 203K & 234K & 166 & 4.5K & - &https://www.kaggle.com/ellipticco/elliptic-data-set \\\\\\\\\\\\cline{2-9}\\n&Yelp & 1 & 45K & - & 32 & 6.6K & -&https://github.com/dmlc/dgl/blob/master/python/dgl/data/fraud.py \\\\\\\\ \\n\\\\midrule[1 pt]\\nTransportation Networks & New York City Taxi& - & - & - & - & - & &http://www.nyc.gov/html/tlc/html/about/triprecorddata.shtml \\\\\\\\\\n\\\\bottomrule[1 pt]\\n\\\\multicolumn{9}{l}{* -: Not Given, \\\\#G: Number of Graphs, \\\\#N: Number of Nodes, \\\\#E: Number of Edges, \\\\#FT: Number of Features, \\\\#AN: Number of Anomalies, REF: References.}\\n\\\\end{tabular}\\n\\\\label{tb:publisheddataset}\\n}\\n\\\\end{table*}\",\n \"id\": \"ae3bbf88-f1cb-408b-8ded-ef8445178dfa\",\n \"level\": \"section\",\n \"origin_cites_number\": 63,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS GD on Dynamic Graphs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ANOS GD on Dynamic Graphs\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:resources}\\nAcquiring open-sourced implementations and real-world datasets with real-world anomalies are far from trivial in academic research on graph anomaly detection.\\nHere, we first list the published algorithms with publicly available implementations, then we provide a collection of public benchmark datasets and summarize the commonly used evaluation metrics. \\nLastly, due to the shortage of labeled anomalies in real-world datasets, we review three synthetic dataset generation strategies used in the existing works.\\nAll the resources are available at: https://github.com/XiaoxiaoMa-MQ/Awesome-Deep-Graph-Anomaly-Detection/.\",\n \"id\": \"8147e459-f864-454f-86f6-52d0b0cd6b52\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Published Algorithms and Datasets\"\n ]\n ],\n \"subsections\": [\n \"9907d812-2cb7-4bfc-ba97-f1d8f46efce2\",\n \"152cd0db-3934-4cfa-8fbb-64d458d75863\",\n \"4da63f48-4501-4170-981d-8fe6cdbe1f52\",\n \"6b67eb52-a545-45eb-aa8e-2135a2a408e8\"\n ],\n \"title\": \"Published Algorithms and Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The published implementations of algorithms and models contribute to baseline experiments. Table~\\\\ref{tb:publishedalgs} provides a summary of published implementations outlining the language and platforms, graphs that they can admit, and URLs to code repositories.\",\n \"id\": \"9907d812-2cb7-4bfc-ba97-f1d8f46efce2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"8147e459-f864-454f-86f6-52d0b0cd6b52\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Published Algorithms and Datasets\"\n ],\n [\n \"subsection\",\n \"Published Algorithms\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Published Algorithms\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Table~\\\\ref{tb:publisheddataset} summarizes the most commonly used datasets, categorizing them into different groups with regard to their application fields. Notably, there is a lack of anomaly ground truth with labeled anomalies only provided in the Enron, Twitter Sybil, Disney, Amazon, Elliptic and Yelp datasets. Details of the DBLP, UCI message, Digg, Wikipedia, and New York city taxi datasets are not given because these public datasets only contain the raw data, and in most existing works, they are further processed to build different graphs (\\\\eg homogeneous graphs, bipartite graphs).\\nThe well-known citation networks are often used to generate synthetic datasets by injecting anomalies into them - the number of anomalies varies from study to study.\\nIn addition to these anomaly detection datasets, Table~\\\\ref{tb:publisheddataset} also lists eight graph classification datasets.\\nAs mentioned in~\\\\ref{sec:synthetic}, through downsampling, these datasets can be used as benchmarks for evaluating the anomaly detection performance.\",\n \"id\": \"152cd0db-3934-4cfa-8fbb-64d458d75863\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"8147e459-f864-454f-86f6-52d0b0cd6b52\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Published Algorithms and Datasets\"\n ],\n [\n \"subsection\",\n \"Published Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Published Datasets\"\n },\n {\n \"cite_extract_rate\": 0.14285714285714202,\n \"cites\": [\n 6281\n ],\n \"content\": \"\\\\label{sec:synthetic}\\nGiven the rarity of ground-truth anomalies, many researchers have employed synthetic datasets to investigate the effectiveness of their proposed methods~. \\nTypically, these datasets can be categorized as follows: \\n\\\\begin{itemize}\\n \\\\item \\\\textit{Synthetic graphs with injected anomalies.} Pursuing this strategy, graphs are created to simulate real-world networks. All the nodes and edges are manually added with well-known benchmarks (\\\\eg Lanchinetti-Fornunato-Radicchi (LFR)~, small-world~, scale-free graphs~). Once built, ground-truth anomalies are planted into the network. For the feasibility of generating expected scale of networks, this strategy is mostly used by previous works to validate their underlying intuitions in anomaly detection.\\n \\\\item \\\\textit{Real-world datasets with injected anomalies.} These datasets are built based on the real-world networks. In particular, anomalies are created either by modifying the topological structure or the attributes of existing nodes/edges/sub-graphs, or by inserting non-existent graph objects. \\n \\\\item \\\\textit{Downsampled graph classification datasets.} The widely-used graph classification datasets (\\\\eg NCI1, IMDB, ENZYMES in~) can be easily converted into sets suitable for anomaly detection through two steps. Firstly, one particular class and its data records are chosen to represent normal objects. Then, the other data records are downsampled as anomalies at a specified downsampling rate. By this, the generated graph anomaly detection dataset is, in fact, a subset of the original dataset. The most significant strength of this strategy is that no single data record has been modified.\\n\\\\end{itemize}\",\n \"id\": \"4da63f48-4501-4170-981d-8fe6cdbe1f52\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"8147e459-f864-454f-86f6-52d0b0cd6b52\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Published Algorithms and Datasets\"\n ],\n [\n \"subsection\",\n \"Synthetic Dataset Generation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Synthetic Dataset Generation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:em}\\nTo date, the most widely used metrics for evaluating anomaly detection performance evaluation include accuracy, precision, recall rate, F1-score and AUC-AP (Average Precision). Their formulas/descriptions are given in Table~\\\\ref{tb:evluationmetrics}. However, a more dedicated analysis with some new evaluation metrics is needed for further performance examination \\nbecause anomaly detection has different requirements for different applications~, \\\\eg false negative and false positive.\\nFor instance, network intrusion prevention systems are more sensitive to false negative errors, while false positive errors are considered relatively harmless. This is because any risky connections should be shut down to prevent information leaks.\\nBy contrast, other applications concentrate more on the false positives, \\\\eg in auditing domain, companies often set a budget for an auditor to look at flagged anomalies, and they want high precision/small false positive rate such that the auditor's time can be best used.\\nHence, when evaluating detection performance, we suggest reviewing the specific requirements of the application domain for fair and suitable comparisons.\\n\\\\begin{table}\\n\\\\centering \\n\\\\footnotesize\\n\\\\renewcommand\\\\arraystretch{1}\\n\\\\setlength{\\\\tabcolsep}{2.8mm}\\n\\\\caption{Evaluation Metrics. $tp$: true positives; $tn$: true negatives; $fp$: false positives; $fn$: false negatives.}\\n\\\\resizebox{0.46\\\\textwidth}{!}{\\n\\\\begin{tabular}{l|c}\\n\\\\toprule[1 pt]\\n\\\\textbf{Evaluation Metric} & \\\\textbf{Formula/Description} \\\\\\\\ \\n\\\\midrule[1 pt]\\nAccuracy & $\\\\frac{tp + tn}{tp + tn + fp + fn}$ \\\\\\\\\\\\hline\\nPrecision & $\\\\frac{tp}{tp + fp}$ \\\\\\\\\\\\hline\\nRecall & $\\\\frac{tp}{tp + fn}$ \\\\\\\\\\\\hline\\nF1 Score & $2 * \\\\frac{Recall * Precision}{Recall + Precision}$ \\\\\\\\\\\\hline\\nAUC-ROC & The Area Under the ROC Curve\\\\\\\\ \\\\hline\\nAUC-AP & The Area Under Precision-Recall Curve \\\\\\\\\\n\\\\bottomrule[1 pt]\\n\\\\end{tabular}\\n\\\\label{tb:evluationmetrics}\\n}\\n\\\\end{table}\",\n \"id\": \"6b67eb52-a545-45eb-aa8e-2135a2a408e8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"8147e459-f864-454f-86f6-52d0b0cd6b52\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Published Algorithms and Datasets\"\n ],\n [\n \"subsection\",\n \"Evaluation Metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Evaluation Metrics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:futures}\\nSo far, we have reviewed the contemporary deep learning techniques that are devoted to graph anomaly detection.\\nAn apparent observation from our survey is there remain many compounded challenges imposed by the complexity of anomaly detection, graph data, and the immaturity of deep learning techniques for graph data mining. Another observation is that deep learning techniques in graph anomaly detection are still confined to a relatively small number of studies, and most of these focus on anomalous node detection. For a gauge, simply compare the length of Tables~\\\\ref{table:ANOSND} and~\\\\ref{tb:edgeandsubgraph}. Edge, sub-graph, and graph-level anomaly detection have clearly received much less attention.\\nTo bridge the gaps and push forward future work, we have identified 12 directions of future research for graph anomaly detection with deep learning.\",\n \"id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ]\n ],\n \"subsections\": [\n \"32fa0363-7074-47b3-9e03-5211f2f264be\",\n \"239988e3-c594-47d4-adb9-69b272543369\",\n \"aff7afb2-0d0e-4d85-baf6-f135a679f1c2\",\n \"094e633e-5d3b-45d7-bca3-171ad2d8492b\",\n \"e773dc9f-5c7d-4700-bcf9-1e7e0772172c\",\n \"cc214041-b949-4f3c-95ae-5deb83ec1573\",\n \"85e8f422-2b40-4284-aca9-eaeef926d8d1\",\n \"1224de5d-ffc7-41d5-bee2-9722227a3ca0\",\n \"d42a5aab-3ff4-4f52-8576-f381cddddc12\",\n \"4eb29545-f452-443d-b7b6-8accd75e515c\",\n \"9d5eeb0b-b689-4f33-bf46-0aa58f25c4d7\",\n \"b1a4c699-5c1b-4fe6-a282-ae8f4a92410d\"\n ],\n \"title\": \"Future Directions\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6305\n ],\n \"content\": \"\\\\label{sec:future:ED}\\nIn real-world graphs, anomalies also appear as unusual relationships between objects, sub-structures formed by abnormal groups, or abnormal graphs, which are known as anomalous edges, sub-graphs, and graphs respectively.\\nAs indicated in our review, there is a huge gap between the existing anomalous edge/sub-graph/graph detection techniques and the emerging demands for more advanced solutions in various application domains (\\\\eg social networks, computer networks, financial networks).\\nWhen detecting anomalous edges/sub-graphs/graphs, the proposed methods should be capable of leveraging the rich information contained in graphs to find clues and characteristics that can distinguish normal objects and anomalies in specific applications.\\nTypically, this involves extracting edge/sub-graph/graph-level features, modeling the patterns of these features, and measuring the abnormalities accordingly.\\nHowever, current deep learning based graph anomaly detection techniques put forward very little effort in this regard.\\n\\\\textit{Opportunities}: We believe more research efforts can be done on anomalous edge, sub-graph, and graph detection with regard to their significance in real-world applications. Possible solutions to this gap might to be first consider the application domain and explore domain knowledge to find complementary clues as a basis for these problems. Then, motivated by recent advances in deep learning for edge, sub-graph, and graph-level representation learning~, extensive work can be done to learn an anomaly-aware embedding space such that it is feasible to extract abnormal patterns of anomalies. Although this direction seems quite straightforward, the true challenge lies in the specific application domains. Hence, domain knowledge, anomalous pattern recognition and anomaly-aware deep learning techniques should be enforced simultaneously.\",\n \"id\": \"32fa0363-7074-47b3-9e03-5211f2f264be\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Anomalous Edge, Sub-graph, and Graph Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Anomalous Edge, Sub-graph, and Graph Detection\"\n },\n {\n \"cite_extract_rate\": 0.28571428571428503,\n \"cites\": [\n 217,\n 218\n ],\n \"content\": \"Dynamic graphs provide powerful machinery with which to capture the evolving relationships between real objects and their attributes.\\nTheir ever-changing structure and attribute information inherently make anomaly detection very challenging in these scenarios, leading to two primary concerns for the task.\\nThe first is to consider the spatial and temporal information contained in each graph snapshot at different time stamps, and the second is to explore the evolving patterns of nodes, edges, sub-graphs and graphs, as well as their interaction with the node/edge attributes over time.\\nWhen these challenges have been tackled with mature solutions, detection techniques will achieve better results.\\n\\\\textit{Opportunities}: From our observations, most of deep learning based dynamic graph anomaly detection techniques are built on DeepWalk~, GCN~ or other deep models that are intuitively designed for static graphs. This means other information, like evolving patterns in attributes~), are not adequately used in the detection task.\\nWe can therefore identify the following directions for future studies to target.\\n\\\\begin{itemize}\\n \\\\item \\\\textit{Using dynamic graph mining tools.} As a popular research topic, deep learning for dynamic graph data mining~ has shown its effectiveness in supporting dynamic graph analysis, such as node clustering and graph classification~. More future works can be foreseen that adopt these techniques for anomaly detection.\\n \\\\item \\\\textit{Deriving solid evidence for anomaly detection.} The rich structural, attribute and temporal information in dynamic graphs are valuable resources for identifying anomalies. Apart from the indicators widely used in current works, such as burst of connections between node pairs or suddenly vanishing connections, we suggest exploring structural and attribute changes in depth. From such studies, we may derive additional information to enhance the detection performance, such as the appearance of abnormal attributes.\\n \\\\item \\\\textit{Handling complex dynamics.} Real-world networks always exhibit changes in both the network structure and node attributes, but only very few studies address this circumstance. Most of the state-of-arts only consider changes in one of these aspects. Although this `double' scenario is extremely complex and detecting anomalies in this kind of dynamic graph is very challenging, it is worth studying because these graphs are highly reflective of real network data.\\n\\\\end{itemize}\",\n \"id\": \"239988e3-c594-47d4-adb9-69b272543369\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Anomaly Detection in Dynamic Graphs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Anomaly Detection in Dynamic Graphs\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6296,\n 6299,\n 6300\n ],\n \"content\": \"Heterogeneous graphs are a specific type of graph that contain diverse types of nodes and edges.\\nFor instance, Twitter can be intuitively modeled as a heterogeneous graph comprised of tweets, users, words, etc.\\n\\\\textit{Opportunities}: To use the complex relationships between different types of nodes in heterogeneous graphs for anomaly detection, representative works, such as HGATRD~, GCAN~ and GLAN~, typically decompose a heterogeneous graph into individual graphs according to meta-paths, \\\\eg one with tweets and users, and another with tweets and words. They then use D(G)NNs to learn the embeddings for graph anomaly detection. Such a decomposition inherently overlooks the direct inter-relations among diverse types of nodes/edges and downgrades the effectiveness of the embeddings. A possible solution is to reveal the complex relations between different types of nodes and edges, and encode them into a unique representation for boosted detection performance.\",\n \"id\": \"aff7afb2-0d0e-4d85-baf6-f135a679f1c2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Anomaly Detection in Heterogeneous Graphs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Anomaly Detection in Heterogeneous Graphs\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 242\n ],\n \"content\": \"The scalability of methods to high-dimensional and large-scale data is an ongoing and significant challenge to anomaly detection techniques.\\nIn face of large-scale networks, such as Facebook and Twitter that contain billions of users and friendship links, the size of data in terms of both graph size and number of node attributes is extremely high.\\nHowever, most of the existing works lack the ability to detect anomalies in such large-scale data because they are transductive models and need to take the whole graph as input for further analysis.\\nComputation time and memory cost increase dramatically as the network scales up, and this stops existing techniques from being used on large-scale networks.\\n\\\\textit{Opportunities}: Accordingly, there is a need for scalable graph anomaly detection techniques.\\nOne possible approach would be an inductive learning scheme that first trains a detection model on part of the whole graph and then applies the model to detect anomalies in the unseen data.\\nAs some inductive learning models, such as GraphSAGE~, have shown their effectiveness on link prediction and node classification in large-scale graphs, this approach is expected to provide a basis for graph anomaly detection in large-scale graphs and similar techniques can be investigated in the future.\",\n \"id\": \"094e633e-5d3b-45d7-bca3-171ad2d8492b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Anomaly Detection in Large-scale Graphs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Anomaly Detection in Large-scale Graphs\"\n },\n {\n \"cite_extract_rate\": 0.14285714285714202,\n \"cites\": [\n 8994\n ],\n \"content\": \"In real-world networks, objects might form different kinds of relationships with others (\\\\eg user's followership and friendship on Twiter). And their attribute information might be collected from different resources, such as user's profile, historical posts.\\nThis results in two types of multi-view graphs: 1) multi-graph that contains more than one type of edges between two nodes~; and 2) multi-attributed-view graph that stores node attributes in different attributed views~.\\n\\\\textit{Opportunities}: These multi-views basically allow us to analyze real objects' characteristics from different perspectives.\\nEach view also provides complementary information to other views, and they might have different significance on anomaly detection. \\nFor instance, anomalies might be indistinguishable in one view but are obviously divergent from the majority in another view.\\nThere are a variety of work in data mining on multi-view learning~. However, work that can accommodate multi-view graphs along with multi-view attributes on nodes for anomaly detection purposes is nascent.\\nMoreover, the rich information contained in multiple views and the inconsistency among them are overlooked in these works. \\nTo this end, we believe more research effort in this direction is needed. Digesting the relationships between views will be vital to their success, as two views might provide contrary/supplementary information for anomaly detection.\",\n \"id\": \"e773dc9f-5c7d-4700-bcf9-1e7e0772172c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Multi-view Graph Anomaly Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-view Graph Anomaly Detection\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6307,\n 6306,\n 6271,\n 8995\n ],\n \"content\": \"The easy accessibility of online platforms has made them convenient targets for fraudsters, attackers and other malevolent agents to carry out malicious activities.\\nAlthough various anomaly detection systems have been deployed to protect benign objects, anomalies can still conceal themselves to evade detection~.\\nKnown as camouflaged anomalies, these entities typically disguise themselves as regular objects.\\nIf the detection techniques are not robust against such cases, \\\\ie if they cannot quickly and effectively adapt to the evolving behavior of evasion-seeking attackers, the anomalies are simply left to cause their damage.\\n\\\\textit{Opportunities}: In the face of camouflage, the boundary between anomalies and regular objects is blurred, making anomalies much harder to be identified.\\nWe believe extensive effort should be placed on detecting these anomalies because, as yet, very few studies have looked at handling camouflaged anomalies in graphs~.\\nTo fulfill this gap, one major direction might be to jointly analyze the attributes, co-relations, such as the triadic, tetradic, or high-order relationships between objects in hypergraphs~, and other information comprised in graphs.\\nBy this, anomalies that only camouflage their local structures or attributes can be identified effectively.\\nEnhancing existing techniques might be another direction. \\nThis involves incorporating additional detection mechanisms or function blocks particularly designed for distinguishing camouflaged anomalies with existing detection techniques.\\nConsequently, these techniques will bridge most existing works and camouflaged anomaly detection.\",\n \"id\": \"cc214041-b949-4f3c-95ae-5deb83ec1573\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Camouflaged/Adversarial Anomaly Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Camouflaged/Adversarial Anomaly Detection\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Anomalies are rare, which means anomaly detection always occurs coexists with class imbalance in the training data. As deep learning models rely heavily on the training data, such imbalance pose great challenges to graph anomaly detection, and this remains a significant obstacle to deep learning techniques. \\nTypically, the imbalanced class distributions will downgrade the detection techniques' ability to capture the differences between anomalies and non-anomalies. It might even cause over-fitting on the anomalous class because there are too few anomalies in the data.\\nIf the detection model overlooks this critical fact and is trained improperly, detection performance will be sub-optimal.\\n\\\\textit{Opportunities}: In fact, class imbalance has been widely explored in various research areas~.\\nAdvances such as under-sampling the majority class or modifying the algorithms shed important light on solving training problems with imbalances.\\nYet, contemporary graph anomaly detection methods rarely incorporate these techniques. For more effective detection techniques, biased models that pay more attention to anomalies, such as penalizing additional training losses on misclassified anomalies, would be a possible direction to circumvent the problem.\\nMoreover, when adopting graph neural networks that aggregate neighboring information to the target node, like GCN or GraphSAGE, the over-smoothing between the features of connected nodes should be prevented such that the distinguishable features of the anomalies can be preserved to support anomaly detection.\",\n \"id\": \"85e8f422-2b40-4284-aca9-eaeef926d8d1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Imbalanced Graph Anomaly Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Imbalanced Graph Anomaly Detection\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 6308,\n 1671,\n 6309,\n 6284\n ],\n \"content\": \"Graph anomaly detection has close relations with other graph mining tasks including community detection~ and node classification~, and link prediction~. \\nFor a concrete example, when detecting community anomalies, community detection techniques are usually used to extract the community structures prior to anomaly detection. Meanwhile, the anomaly detection results can be used to optimize the community structure.\\nSuch mutually beneficial collaborations between anomaly detection and other tasks inherently suggest an opportunity for multi-task learning that can handle diverse tasks simultaneously and share information among tasks.\\n\\\\textit{Opportunities}: Multi-task learning provides effective machinery with which incorporate associated tasks~. Its utmost advantage is that the training signal from another task could yield complementary information to distinguish anomalies from non-anomalies. The result would be enhanced detection performance. However, very few attempts focus on this at present.\\nBeyond current works, such as~ that jointly perform anomalous node detection and personalized recommendation, explorations into combining other learning tasks with graph anomaly detection are likely to emerge as a fruitful future direction.\",\n \"id\": \"1224de5d-ffc7-41d5-bee2-9722227a3ca0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Multi-task Anomaly Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Multi-task Anomaly Detection\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8996,\n 6310\n ],\n \"content\": \"The interpretability of anomaly detection techniques is vital to the subsequent anomaly handling process.\\nWhen applying these techniques to real applications, such as financial and insurance systems, it is essential to provide explainable and lawful evidence to support the detection results.\\nHowever, most of the existing works lack the ability to provide such evidence.\\nTo identify the anomalies, the most commonly used metrics are top-k rankings and simple anomaly scoring functions.\\nThese metrics are flexible enough to label objects as being either an anomaly or not an anomaly, but they cannot derive solid explanations.\\nMoreover, as deep learning techniques have also been criticized for their low interpretability, future works on graph anomaly detection with deep learning should pay much more attention to this~.\\n\\\\textit{Opportunities}: To bridge this gap, integrating specially designed interpretation algorithms or mechanisms~ into the detection framework would be a possible solution, noting that this would inherently induce a higher computational cost.\\nFuture works should therefore balance the cost of anomaly detection performance and interpretability.\\nVisualization-based approaches, \\\\eg dashboards, charts, might also be feasible for showing the distinction between anomalies and non-anomalies in a human-friendly manner.\\nFurther research in this direction will be successful if interpretable visualization results can be given~.\",\n \"id\": \"d42a5aab-3ff4-4f52-8576-f381cddddc12\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Graph Anomaly Interpretability\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Graph Anomaly Interpretability\"\n },\n {\n \"cite_extract_rate\": 0.36363636363636304,\n \"cites\": [\n 6289,\n 6287,\n 6283,\n 8997\n ],\n \"content\": \"Amongst existing unsupervised graph anomaly detection techniques, anomalies are mainly identified based on residual analysis~, reconstruction loss~, distance-based statistics~, density-based statistics~, graph scan statistics~, and one-class classification~.\\nThe underlying intuition of these identification strategies is that anomalies have inconsistent data patterns with regular objects, and they will, therefore: 1) introduce more residual errors or be harder to reconstruct; or 2) be located in low-density areas or far away from the majority class in an anomaly-aware feature space.\\nEffort toward designing novel loss functions for GNNs for anomaly detection is currently quite limited~\\n\\\\textit{Opportunities}: Although these strategies could capture the deviating data patterns of anomalies, they also have different limitations. Specifically, the residual analysis, one-class classification and reconstruction loss strategies are sensitive to noisy training data. Noisy nodes, edges or sub-graphs also exhibit large residuals, a large distance to the origin/hypersphere center and high reconstruction losses. Meanwhile, the distance-based and density-based strategies can only be applied when anomalies and non-anomalies are well separated in lower-dimensional space. Detection performance also downgrades dramatically if the gap between anomalies and non-anomalies is not that evident.\\nIt calls for extensive future efforts to break these limitations and explore new anomaly identification strategies.\",\n \"id\": \"4eb29545-f452-443d-b7b6-8accd75e515c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Graph Anomaly Identification Strategies\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Graph Anomaly Identification Strategies\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Systematic benchmarking is key to evaluating the performance of graph anomaly detection techniques.\\nAs indicated in our analysis in Section~\\\\ref{sec:em}, recent studies are constantly raising attention to more comprehensive and effective benchmarking~. \\nTypically, the benchmarking framework consists of benchmark datasets, baseline methods, evaluation metrics, and further analysis tools.\\nWhen evaluating a techniques' performance with other baselines, the evaluation dataset and metrics become very important because the performance of each model may vary depending on the setting.\\nThe shortage of public datasets and (public available) baseline methods also imposes great challenges for effective evaluation.\\nAlthough one of the aims of our survey is to provide extensive materials for this purpose, like open-sourced implementations, datasets, and evaluation metrics, this work can only serve as a basis for future investigations into systematic benchmarking. We invite more efforts from the anomaly detection community toward this important case.\\nCertainly, rigorous attention to designing better frameworks for bechmarking would help to disclose the advances and shortcomings of various detection techniques and essentially track a unprejudiced and accurate progress record in this field.\",\n \"id\": \"9d5eeb0b-b689-4f33-bf46-0aa58f25c4d7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Systematic Benchmarking\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Systematic Benchmarking\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"A graph anomaly can be categorized as an anomalous node, edge, and sub-graph in a single graph or an anomalous graph in a graph database.\\nThese anomalies usually coexist in real-world datasets.\\nFor instance, individual fraudsters, abnormal relationships, and fraud groups exist concurrently in online social network, as shown in Fig.~\\\\ref{Toy}.\\nMoreover, there may be different ways to define anomalies of a certain type, such as community outliers versus anomalous communities or attribute-based versus structural anomalies.\\nWhen deploying detection techniques in real applications, it is expected that all types of anomalies can be identified while consuming the least resources and time.\\nA straightforward approach would be to integrate independent anomalous node, edge, and sub-graph detection techniques.\\nAlthough this is convenient to apply to relatively small networks, its high computational cost will surely prevent the approach from scaling to large networks, such as Facebook and Twitter, because the same graph data has to be loaded and processed more than once by different techniques.\\n\\\\textit{Opportunities}: Unified frameworks that can detect diverse types of anomalies together~ may provide feasible solutions to bridge the gap.\\nTo build such frameworks, one possible direction is to capture all the information needed by different detection techniques simultaneously so that these techniques can be applied.\\nThe idea seems to be non-challenging, but in deep learning, methods of designing neural network layers and learning strategies that can fulfill this need will require extensive effort.\",\n \"id\": \"b1a4c699-5c1b-4fe6-a282-ae8f4a92410d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"e97bdf9c-475c-4ebc-8114-7af7a8b0aa87\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Future Directions\"\n ],\n [\n \"subsection\",\n \"Unified Anomaly Detection Framework\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Unified Anomaly Detection Framework\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:conclusion}\\nDue to the complex relationships between real-world objects and recent advances in deep learning, especially graph neural networks, graph anomaly detection with deep learning is currently at the forefront of anomaly detection.\\nTo the best of our knowledge, this is the first survey to present a comprehensive review dedicated to graph anomaly detection with modern deep learning techniques.\\nSpecifically, we have reviewed and categorized contemporary deep learning techniques according to the types of graph anomalies they can detect, including: (1) anomalous node detection; (2) anomalous edge detection; (3) anomalous sub-graph detection; and, finally, (4) anomalous graph detection.\\nClear summarizations and comparisons between different works are given to offer a complete and thorough picture of the current work and progress of graph anomaly detection as a field.\\nMoreover, to push forward future research in this area, we have provided a basis for systematic benchmarking by compiling a wide range of commonly used datasets, open-sourced implementations and synthetic dataset generation techniques.\\nWe further highlight 12 potential directions for future work based on the survey results.\\nIt is our firm belief that graph anomaly detection with deep learning is indeed more than a burst of temporary interest, and numerous applications from diverse domains are to surely benefit from it for the years to come.\\n\\\\bibliographystyle{IEEEtran}\\n\\\\bibliography{main.bib}\\n\\\\clearpage\\n\\\\setcounter{page}{1}\\n\\\\appendices\",\n \"id\": \"32ddd0ab-f741-4a6c-a1ff-380ee935d4a2\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n },\n {\n \"cite_extract_rate\": 0.31707317073170704,\n \"cites\": [\n 6285,\n 6281,\n 6283,\n 6289,\n 6307,\n 6279,\n 6271,\n 6306,\n 6293,\n 5645,\n 8149,\n 6291,\n 6286\n ],\n \"content\": \"\\\\label{appendix:data-challenges}\\nDue to the complexity of anomaly detection and graph data mining, adopting deep learning technologies for graph anomaly detection faces a number of challenges:\\n\\\\textbf{Data-CH1. Ground-truth is scarce.} In most cases, there is little or no prior knowledge about the features or patterns of anomalies in real applications. Ground-truth anomalies are often identified by domain experts, and this generally cost-prohibitive. As a result, labeled ground-truth anomalies are often unavailable for analysis in a wide range of disciplines.\\n\\\\textbf{Data-CH2. Various types of graphs.} Different graphs model different real-world data. For instance, plain graphs contain only structural information, attributed graphs contain both structural and attribute information, and heterogeneous graphs represent the complex relations between different types of objects.\\nThese graphs reflect the real-world data in different forms, and graph anomalies will show different deviating patterns in different types of graphs.\\n\\\\textbf{Data-CH3. Various types of graph anomalies.} Given a specific type of graph, graph anomalies could appear as a specific node, edge, sub-graph, or an entire graph, and each type of these anomalies is significantly different from others. This means detection methods must involve concise definitions of anomalies and be able to identify concrete clues about the deviating patterns of anomalies.\\n\\\\textbf{Data-CH4. High dimensionality and large scale.} Representing the structure information of real-world networks usually results in high dimensional and large-scale data~ because real-world network often contain millions or billions of nodes. Graph anomaly detection techniques, hence, should be capable of handling such high dimensional and large scale data; this includes the ability to extract anomalous patterns under the constraints of execution time and feasible computing resources. \\n\\\\textbf{Data-CH5. Interdependencies and dynamics.} The relationships between real objects reveal their interdependencies and they can no longer be treated individually for anomaly detection. That is to say, the detection techniques need to consider the deviating patterns of anomalies by assessing the pairwise, triadic, and higher relationships among objects restored in conventional graphs or hypergraphs~.\\nIn addition, the dynamic nature of real-world networks makes detection problems much more challenging.\\n\\\\textbf{Data-CH6. Class imbalance.} As anomalies are rare occurrences, only a very small proportion of the real-world data might be anomalous. This naturally introduces a critical class imbalance problem to anomaly detection because the number of normal objects is far greater than anomalies in the training data. If no further actions are taken to tackle this challenge, learning-based anomaly detection techniques might overlook the patterns of anomalies, leading to sub-optimal results.\\n\\\\textbf{Data-CH7. Unknown and camouflage of anomalies.} In reality, knowledge about anomalies mainly stems from human expertise. There are still many unknown anomalies across different application domains, and new types of anomalies might appear in the future. Nevertheless, real-world anomalies can hide or be camouflaged as benign objects to bypass existing detection systems. In graphs, anomalies might hide themselves by connecting with many normal nodes or by mimicking their attributes. Detection methods, therefore, need to be adaptive to unknown and novel anomalies and robust to camouflaged anomalies.\\nThese data-specific challenges and technical-specific challenges (discussed in Section~\\\\ref{sec:introdution:challenges}) are summarized in Table.~\\\\ref{tb:challenges}, along with the corresponding articles that aim to address them.\\n\\\\begin{table}\\n\\\\centering \\n\\\\footnotesize\\n\\\\renewcommand\\\\arraystretch{1}\\n\\\\setlength{\\\\tabcolsep}{2.8mm}\\n\\\\caption{Challenges and Methods.}\\n\\\\resizebox{0.48\\\\textwidth}{!}{\\n\\\\begin{tabular}{m{1.1cm}<{\\\\centering}|m{2cm}<{\\\\centering}|m{2.9cm}<{\\\\centering}}\\n\\\\toprule[1 pt]\\n\\\\textbf{Challenges} & \\\\textbf{Details} & \\\\textbf{Methods} \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\multicolumn{3}{c}{\\\\textbf{Data-specific Challenges}} \\\\\\\\\\n\\\\midrule[1 pt]\\nData-CH1 & Ground-truth is scarce & \\\\\\\\ \\n\\\\cline{1-3}\\nData-CH2 & Various types of graphs & \\\\\\\\ \\n\\\\cline{1-3}\\nData-CH3 & Various types of graph anomalies & \\\\\\\\ \\n\\\\cline{1-3}\\nData-CH4 & High dimensionality and large scale & \\\\\\\\ \\n\\\\cline{1-3}\\nData-CH5 & Interdependencies and dynamics & \\\\\\\\ \\n\\\\cline{1-3}\\nData-CH6 & Class imbalance & \\\\\\\\ \\n\\\\cline{1-3}\\nData-CH7 & Unknown and camouflage of anomalies & \\\\\\\\ \\n\\\\midrule[1 pt]\\n\\\\midrule[1 pt]\\n\\\\multicolumn{3}{c}{\\\\textbf{Techniques-specific Challenges}} \\\\\\\\\\n\\\\midrule[1 pt]\\nTech-CH1 & Anomaly-aware training objectives & \\\\\\\\\\\\cline{1-3}\\nTech-CH2 & Anomaly interpretability & \\\\\\\\\\\\cline{1-3}\\nTech-CH3 & High training cost & \\\\\\\\\\\\cline{1-3}\\nTech-CH4 & Hyperparameter tuning & \\\\\\\\\\n\\\\bottomrule[1 pt]\\n\\\\end{tabular}\\n\\\\label{tb:challenges}\\n}\\n\\\\end{table}\",\n \"id\": \"9a4d2cc0-0208-4633-a227-f43cf17e07e4\",\n \"level\": \"section\",\n \"origin_cites_number\": 41,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Challenges in Graph Anomaly Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Challenges in Graph Anomaly Detection\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{appendix:taxonomy}\\nThe taxonomy of this survey is shown in Fig.~\\\\ref{pic:taxonomy}.\",\n \"id\": \"12e1194f-a7ed-4e46-889d-a06a013eaeaf\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"Taxonomy\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Taxonomy\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{appendix:node:static}\\nFollowing the taxonomy of Section~\\\\ref{lb:node:sag}, this section reviews traditional non-deep learning techniques designed for ANOS ND on static attributed graphs, followed by techniques based on GATs, GANs, and network representation.\\n\\\\onecolumn\\n\\\\begin{sidewaysfigure*}[h]\\n \\\\centering\\n \\\\includegraphics[width=\\\\textwidth,keepaspectratio]{pics/taxonomy.pdf}\\n \\\\captionof{figure}{The Taxonomy of Our Survey.}\\n \\\\label{pic:taxonomy}\\n\\\\end{sidewaysfigure*}\\n\\\\twocolumn\",\n \"id\": \"dfc7ea98-0750-495a-b3c3-8183220b6042\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Static Graphs\"\n ]\n ],\n \"subsections\": [\n \"92e0465c-2bc8-4d8d-9625-6c4f4b200390\",\n \"ac5b6d2a-da5e-45c0-bd9e-1a63dbed603a\",\n \"ec58be64-9072-46de-98e9-29a0e0542d63\",\n \"99113f23-1403-4b3d-bae2-8db2bf051152\"\n ],\n \"title\": \"ANOS ND on Static Graphs\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Traditional techniques, such as statistical models, matrix factorization, and KNN, have been widely applied to extract the structural/attribute patterns of anomalous nodes for a subsequent detection process.\\nAmong these, matrix factorization (MF) based techniques have shown power at capturing both topological structures and node attributes, achieving promising detection performance.\\nAn early attempt at this type of method was Liu \\\\etal~.\\nThey aimed to detect community anomalies (defined in Section~\\\\ref{sec:AND}) through the developed model, ALAD.\\nAs shown in Fig.~\\\\ref{pic:ALAD}, through non-negative matrix factorization, ALAD incorporates both the graph structure $A$ and node attributes $X$ to derive community structures $W$ and their attribute distribution vectors $H$.\\nWhen the matrices are decomposed, ALAD measures the normality of each node according to the attribute similarity between it and the community it belongs to.\\nBy ranking the nodes' normality scores in ascending order, the top-k nodes are identified as community anomalies. \\nLi \\\\etal~ approached ANOS ND from a different perspective using residual analysis.\\nAs assumed, anomalies will lead to larger attribute reconstruction residual errors because they do not conform to the attribute patterns of the majorities.\\nAccordingly, the proposed model, Radar, learns the residual errors $R$, as shown in Fig.~\\\\ref{pic:Radar}, by factorizing the node attributes $X$. To incorporate the structural information for obtaining these errors, Radar puts explicit restrictions on the learned residuals such that directly linked nodes will have similar residual patterns in $R$ (known as the homophily effect).\\nFinally, the top-k nodes with larger norms in $R$ are identified as anomalies.\\nAlthough the homophily hypothesis provides strong support for exploiting structural information, it might not always be hold.\\nIn fact, real objects may have distinct attributes from their connected neighbors and it is non-trivial to regulate all connected objects share similar values in each dimension in the feature space. By this, Peng \\\\etal~ indicated that there are structurally irrelevant node attributes that do not satisfy the homophily hypothesis.\\nIndeed, these structurally irrelevant node attributes would have adverse effects on anomaly detection techniques that are developed based on this hypothesis. \\nTo tackle this problem, their developed model, ANOMALOUS, uses CUR~ decomposition to select attributes that are closely related to the network structure and then spot anomalies through residual analysis following Radar (as shown in Fig.~\\\\ref{pic:ANOMALOUS}). \\n\\\\begin{figure}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n \\\\centering\\n \\\\subfigure[The Framework of ALAD~]{\\n \\\\label{pic:ALAD} \\n \\\\includegraphics[width=0.48\\\\textwidth]{pics/Node/ALAD.pdf}}\\n \\\\subfigure[The Framework of Radar~]{\\n \\\\label{pic:Radar} \\n \\\\includegraphics[width=0.48\\\\textwidth]{pics/Node/Radar.pdf}}\\n \\\\subfigure[The Framework of ANOMALOUS~]{\\n \\\\label{pic:ANOMALOUS}\\n \\\\includegraphics[width=0.48\\\\textwidth]{pics/Node/Anomalous.pdf}}\\n \\\\caption{ANOS ND on attributed graphs -- Matrix Factorization based approaches. The three representative MF techniques adopt different decomposition strategies to fuse the graph structural information and node attributes. Anomalous nodes are then spotted through scoring functions or residual analysis.}\\n \\\\label{pic:MFtechniques}\\n\\\\end{figure}\\nBeyond matrix factorization, linear regression models have also been designed to train anomaly classifiers given labeled training data.\\nA representative work is Wu \\\\etal~.\\nTheir supervised model, SGASD, has yielded encouraging results with identifying social spammers using the social network structure, the content information in social media and user labels.\\nThese non-deep learning techniques are able to capture valuable information from graph topology and node attributes, but their application and generalizability to real-world networks (which are usually in large-scale) is strictly limited due to the high computational cost of matrix decomposition operations and regression models.\",\n \"id\": \"92e0465c-2bc8-4d8d-9625-6c4f4b200390\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"dfc7ea98-0750-495a-b3c3-8183220b6042\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Static Graphs\"\n ],\n [\n \"subsection\",\n \"Traditional Non-Deep Learning Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Traditional Non-Deep Learning Techniques\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 6291,\n 180,\n 6289\n ],\n \"content\": \"\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n \\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/Node/GAT.pdf}}\\n \\\\caption{ANOS ND on attributed graphs -- GAT based approaches. Given the input graph, these techniques employ graph attention neural network to learn node embeddings. The unsupervised technique, AnomalyDAE~, scores each node based on the reconstruction loss and mark the top-k nodes as anomalies, while the semi-supervised technique, SemiGNN~, trains a classifier to predict node labels.}\\n \\\\label{pic:GAT}\\n\\\\end{figure*}\\nAlthough GCN provides an effective solution to incorporating graph structure with node attributes for ANOS ND (reviewed in Section~\\\\ref{sec:AND:GCN}), its ability to capture the most relevant information from neighboring is subpar.\\nThis is due to the simple convolution operation that aggregates neighbor information equally to the target node.\\nRecently, graph attention mechanism (GAT)~ is employed to replace the traditional graph convolution.\\nFor instance, Fan \\\\etal~ applied graph attention neural network to encode the network structure information (structure encoding).\\nThe method, AnomalyDAE, also adopts a separate attribute autoencoder to embed the node attributes (attribute encoding).\\nThrough an unsupervised encoding-decoding process, each node is ranked according to its corresponding reconstruction loss, and the top-k nodes introducing the greatest losses are identified as anomalies.\\nSpecifically, the attribute decoding process takes both node embeddings learned through the structure and attribute encoding processes to reconstruct node attributes, as shown in Fig.~\\\\ref{pic:GAT}, while the graph topology is reconstructed only using the embeddings output by the GAT.\\nTo acquire better reconstruction results, AnomalyDAE is trained to minimize the overall loss function, denoted as:\\n\\\\begin{equation} \\\\label{eq:anomalydae}\\n\\\\begin{split}\\n \\\\mathcal{L}_{AnomalyDAE} = & \\\\alpha ||(A - \\\\hat{A})\\\\odot \\\\bm{\\\\theta}||_{2}^{2} + \\\\\\\\\\n &(1-\\\\alpha)||(X - \\\\hat{X})\\\\odot \\\\bm{\\\\eta}||_{2}^{2},\\n\\\\end{split}\\n\\\\end{equation}\\nwhere $\\\\alpha$ is the coefficient, $A$ and $X$ is the input adjacency matrix and attribute matrix, $\\\\hat{A}$ and $\\\\hat{X}$ are the reconstructed matrices. Each $\\\\theta_{i,j} \\\\in \\\\bm{\\\\theta}$ and $\\\\eta_{i,j} \\\\in \\\\bm{\\\\eta}$ is 1, if the corresponding element $A_{ij}$ and $X_{ij}$ equals 0, otherwise, their values are defined by hyperparameters greater than 1.\\nAnother decent work is SemiGNN~, in which Wang \\\\etal proposed a semi-supervised attention-based graph neural network for detecting fraudulent users in online payment platforms.\\nThis work further explores user information collected from various sources (\\\\eg transaction information and user profiles), and represents real-networks as multi-view graphs.\\nEach view in the graph is modeled to reflect the relationship between users or the correlation between user attributes.\\nFor anomaly detection, SemiGNN first generates node embedding $h_u^v$ from each view $v$ by aggregating neighbor information through a node-level attention mechanism.\\nIt then employs view-level attention to aggregate node embeddings from each view and generates a unified representation $a_u$ for each node.\\nLastly, the class of each node is predicted through a softmax classifier.\\nIndeed, Wang \\\\etal designed a supervised classification loss and an unsupervised graph reconstruction loss to jointly optimize the model by fully utilizing labeled and unlabeled data.\\nThe classification loss can be denoted as:\\n\\\\begin{equation} \\n \\\\mathcal{L}_{sup} = - \\\\frac{1}{|U_L|}\\\\sum_{u \\\\in U_L} \\\\sum_{i=1}^{k}I(y_u=i)\\\\log\\\\frac{\\\\exp(a_u\\\\cdot\\\\theta_i)}{\\\\sum_{j=1}^{k}\\\\exp(a_u\\\\cdot\\\\theta_j)},\\n\\\\end{equation}\\nwhere $U_L$ is the labeled user set and its size is $|U_L|$, $I(\\\\cdot)$ is an indicator function, $k$ is the number of labels to be predicted (in most cases, the label is either anomalies or non-anomalies, and $k=2$), and $\\\\theta$ represents the trainable variables.\\nMeanwhile, the unsupervised loss encourages unlabeled nodes (users) that can be reached by labeled nodes through random walks to obtain similar representations and vice versa.\\nThis is achieved by negative sampling (unlabeled nodes that cannot be reached by random walks are negative samples) and the loss can be formulated as:\\n\\\\begin{equation}\\n \\\\begin{split}\\n \\\\mathcal{L}_{unsup} & = \\\\sum_{u\\\\in U}\\\\sum_{v\\\\in N_u \\\\cup Neg_{u}} -\\\\log(\\\\sigma(a_{u}^{T}a_{v})) \\\\\\\\\\n &- 3\\\\cdot E_{q\\\\sim P_{neg}(u)}\\\\log(\\\\sigma(a_{u}^{T}a_{q})),\\n \\\\end{split}\\n\\\\end{equation}\\nwhere $U$ denotes the user set, $N_u$ denotes the neighbor set of $u$, $Neg_{u}$ represents negative samples, $P$ is the sampling distribution, and $\\\\sigma(\\\\cdot)$ is the sigmoid function. The total loss takes the sum of them and is formulated as:\\n\\\\begin{equation} \\n \\\\mathcal{L}_{SemiGNN} = \\\\alpha \\\\mathcal{L}_{sup} + (1- \\\\alpha) \\\\mathcal{L}_{unsup} + \\\\lambda \\\\mathcal{L}_{reg},\\n\\\\end{equation}\\nwhere $\\\\alpha$ is a balancing parameter and $\\\\mathcal{L}_{reg}$ regularizes all trainable variables.\",\n \"id\": \"ac5b6d2a-da5e-45c0-bd9e-1a63dbed603a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"dfc7ea98-0750-495a-b3c3-8183220b6042\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Static Graphs\"\n ],\n [\n \"subsection\",\n \"GAT Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GAT Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\begin{figure*}[!t]\\n\\\\setlength{\\\\belowcaptionskip}{-0.25cm}\\n \\\\centerline{\\\\includegraphics[width=0.98\\\\textwidth]{pics/Node/GAN.pdf}}\\n \\\\caption{ANOS ND on static graphs -- GAN based approaches. The generator, G, generates potential anomalies by sampling noisy data from a prior distribution to fool the discriminator. The discriminator, D, distinguishes the generated anomalies and normal nodes. The scoring function then quantifies the anomaly score of each node according to D's output.}\\n \\\\label{pic:GAN}\\n\\\\end{figure*}\\nBecause GAN is effective at capturing anomalous/regular data distributions (as reviewed in Section~\\\\ref{sec:ND:GAN}), Ding \\\\etal~ used GAN in their developed model, AEGIS, for improved anomaly discriminability on unseen data.\\nAs shown in Fig.~\\\\ref{pic:GAN}, this model first generates node embeddings through a GNN from the input attributed graph, and then a generator and a discriminator are trained to identify anomalies. In the first phase, anomalous nodes and regular nodes are mapped to distinctive areas in the embedding space such that the GAN is able to learn a boundary between them. Accordingly, Ding \\\\etal built an autoencoder network with graph differentiative layers to capture the attribute difference between each node and its $k$-th neighbors. Such difference information enables anomalies to be distinguished easily. The embeddings are encoded as follows:\\n\\\\begin{equation} \\n h_{i}^{l} = \\\\sum_{k=1}^{K} \\\\beta_{i}^{k} \\\\sigma \\\\left( W_1h_{i}^{l-1} + \\\\sum_{j\\\\in N_{k}(i)}\\\\alpha_{ij}W_{2}\\\\Delta_{ij}^{l-1}\\\\right),\\n\\\\end{equation}\\nwhere $h_{i}^{l}$ is embedded features of node $i$ through the $l$-th layer, $\\\\beta_{i}^{k}$ is the attention for each hop, $N_{k}(i)$ is the set of $k$-th order neighbors, $\\\\alpha_{ij}$ is the attention for each neighbor, and $\\\\Delta_{ij}^{l-1}$ is the difference between $i$ and $j$, $W_1$ and $W_2$ are the trainable variables.\\nThe autoencoder is fine-tuned until the node attributes can be best reconstructed using the learned embeddings, after which the GAN is trained.\\nIn the second phase, the generator follows a prior distribution $p(\\\\tilde{z})$ to generate anomalies by sampling noisy data, while the discriminator struggles to distinguish between the embeddings of the normal nodes and those of the generated anomalies. The training process is formulated as a mini-max game between the generator and discriminator as follows:\\n\\\\begin{equation}\\n \\\\min_{G}\\\\max_{D}\\\\mathbb{E}_{z\\\\sim Z}[\\\\log(D(z))] + \\\\mathbb{E}_{\\\\tilde{z}\\\\sim \\\\tilde{Z}}[\\\\log(1-D(G(\\\\tilde{z})))],\\n\\\\end{equation}\\nwhere $z$ is the node embeddings, and $\\\\tilde{z}$ are the generated anomalies.\\nAfter training, AEGIS directly learns embedding $z_u$ for a test node $u$, and quantifies its anomaly score with regard to the discriminator's outputs, \\\\ie the possibility that node is normal. The scoring function is formulated as:\\n\\\\begin{equation} \\n Ascore(u) = 1 - D(z_u),\\n\\\\end{equation}\\nand the top-k nodes are deemed to be anomalous.\",\n \"id\": \"ec58be64-9072-46de-98e9-29a0e0542d63\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"dfc7ea98-0750-495a-b3c3-8183220b6042\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Static Graphs\"\n ],\n [\n \"subsection\",\n \"GAN Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"GAN Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 6286,\n 6298,\n 6271\n ],\n \"content\": \"With network representation, graphs are first encoded into a vector space before the anomalies detection procedure takes place.\\nAs outlined in Section~\\\\ref{sec:spg:nr}, numerous studies on ANOS ND in attributed graphs have exploited deep network representation techniques.\\nFor instance, Zhang \\\\etal~ detected abnormal nodes that have attributes significantly deviating from their neighbors through a 3-layer neural network, REMAD, and residual analysis.\\nThey explicitly divide the original node attribute matrix into a residual attribute matrix $R$ that captures the abnormal characters of anomalies and a structurally relevant attribute matrix $\\\\hat{X}$ for network representation learning. Both matrices are jointly updated throughout the representation learning process so nearby nodes are encouraged to have similar representations.\\nSpecifically, these node embeddings are generated by aggregating neighbor information with each node's own attributes, formulated as:\\n\\\\begin{equation} \\n h_{i}^{l} = \\\\sigma \\\\left( W^l \\\\cdot \\\\text{CONCAT}\\\\{h_{i}^{l-1},h_{N_i}^{l}\\\\} + b^2\\\\right),\\n\\\\end{equation}\\nwhere $h_{i}^{l}$ is node $i$'s representation generated by the $l$-th layer ($h_{i}^{0} = \\\\hat{X}$), $N_i$ contains $i$'s neighbors, $\\\\sigma()$ is the activation function, $W^{l}$ and $b$ are the trainable variables. Finally, the residual matrix $R$ will contain the abnormal information of each node and the top-k nodes with the largest norms are considered anomalies.\\nGiven partial node labels, Liang \\\\etal~ developed a semi-supervised representation model, SEANO, that incorporates graph structure, node attributes and label information.\\nSimilar to REMAD, SEANO also aggregates neighbor information to center nodes, and the node representations are obtained through an embedding layer, formulated as:\\n\\\\begin{equation} \\\\label{SEANO:embedding}\\n z_{i} = \\\\lambda_{i}h^{l_{1}}(x_{i}) + (1-\\\\lambda_{i})h^{l_{1}}(\\\\Bar{x}_{N_{i}}),\\n\\\\end{equation} \\nwhere $z_{i}$ is $i$'s representation, $\\\\lambda_{i}$ is a trainable variable that identifies the weight of $i$'s own attributes ($x_{i}$), $\\\\Bar{x}_{N_{i}}$ is the average of node $i$'s neighbors' representations, and the function $h^{k}(x_{i}) = \\\\phi(W^kh^{k-1}(x_i)+b^k)$ maps original node attributes into lower dimensional vectors.\\nThen, a supervised component, which takes the representations as input, predicts node labels through a softmax classifier, and an unsupervised component is trained to reconstruct node contexts (node sequences).\\nThe context of each node is not only generated through random walks on the graph but also from the labeled nodes that belong to the same class. After training, SEANO interprets $\\\\lambda_{i}$ as the normality score of node $i$ and the top-k nodes with the highest scores are classed as anomalies.\\nLearning node representations via aggregating neighbor information has proven effective for capturing comprehensive information from graph structure and node attributes. \\nBut, Liu \\\\etal~ demonstrated such an approach can help anomalies aggregate features from regular nodes, making them look normal and leading to sub-optimal detection performance. \\nThey identified three concrete issues that should be considered when applying aggregation operations for anomaly detection: \\n1) Anomalies are rare objects in a network. Hence, directly aggregating neighborhood information will smooth the difference between the anomalies and normal instances, blurring boundaries between them. \\n2) Directly connected nodes have distinctive features, and the assumption that connected nodes share similar features, which serves as the basis for feature aggregation, no longer holds in this scenario.\\n3) Real objects also form multiple types of relations with others, which means aggregation results for different types of relations will be distinctive.\\nWith regard to these concerns, their proposed method, GraphConsis, follows a sampling strategy to avoid potential anomalous neighbors when aggregating node features.\\nThis method also adopts an attention mechanism to aggregate neighbor information following different links.\\nThe learned node representations, therefore, are more robust to anomalies. As such, GraphConsis takes them as input to train a classifier for predicting labels.\\nDou \\\\etal~ further considered camouflage behaviors of fraudsters in their proposed model CARE-GNN to enhance detection performance.\\nAs specified, the camouflages can be categorized as either feature camouflage or relation camouflage.\\nRespectively, anomalies either adjust their feature information or form connections with many benign objects to gloss over suspicious information.\\nHence, directly employing aggregation will overlook the camouflages and smooth the abnormal patterns of anomalies, eliminating the distinctions between anomalies and normal objects. \\nTo alleviate over-smoothness, CARE-GNN also adopts a neighbor sampling strategy, as is the case with GraphConsis, to filter camouflaged anomalies and explores different types of relations formed between users.\\nSpecifically, under each relation, Dou \\\\etal employed a MLP to predict node labels using their features and measure the similarity ($l1$ distance) between each node and its neighbors according to the MLP's output.\\nThen, the top-k most similar neighbors are selected for feature aggregation, and CARE-GNN generates each node's representation through a combination of latent representations that are learned under different relations.\\nA classifier is eventually trained using the representations to predict the node labels.\\nAs can be seen, the performance of these network representation based techniques is decided by their training objectives/loss functions.\\nEnhanced detection performance is probable if the loss function is able to separate normal nodes from abnormal nodes reasonably well. \\nMotivated by this, a more recent work in~ emphasizes the importance of anomaly-aware loss functions.\\nIn order to adjust margins for the anomalies, the authors proposed a novel loss function to guide the representation learning process.\\nSpecifically, this loss function is designed to find the relative scales between the margins of outlier nodes and normal nodes.\\nAn MLP-based classifier is finally trained using the node representations generated by the anomaly-aware loss-guided GNNs and node labels. \\nFor unseen nodes, the classifier will label them upon their representations.\",\n \"id\": \"99113f23-1403-4b3d-bae2-8db2bf051152\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"dfc7ea98-0750-495a-b3c3-8183220b6042\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Static Graphs\"\n ],\n [\n \"subsection\",\n \"Network Representation Based Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network Representation Based Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{appendix:node:dynamic}\\nTo detect anomalous nodes in dynamic plain graphs, traditional non-deep learning techniques rely heavily on modeling the structural evolving patterns of nodes.\\nRepresentative works like~ and~ assume that the evolutionary behaviors of regular nodes' (\\\\ie generate or remove connections with others) usually follow stable patterns, and their changes introduce less impact on the graph structure compared to anomalies.\\nSpecifically, in~, Wang \\\\etal proposed a novel link prediction method to fit the evolutionary patterns of most nodes such that anomalies can be identified because their observations significantly conflict with the prediction results.\\nThey further quantified the impact of anomalous behaviors by assessing the perturbation imposed on the graph adjacency matrix.\\nOther traditional works also exploit node/edge attributes and their changes. For examples, Teng \\\\etal~ took node and edge attributes as two different views to describe each node. By encoding both types of information into a shared latent space, their proposed model learns a hypersphere from historical records. When new observations of existing nodes are given, the model distinguishes between the benign and the anomalies according to their distances to the hypersphere centroid. Points lying outside the hypersphere are spotted as anomalies.\\nUnlike embedding techniques, Nguyen \\\\etal~ proposed a non-parametric method to detect anomalous users, tweets, hashtags, and links on social platforms. Specifically, they modeled social platforms as heterogeneous social graphs such that the affluent relationships between users, tweets, hashtags and links were effectively captured. Through extensive analysis of the features, such as user registration information, keywords in tweets, the linguistic style of links and the popularity scores of hashtags, anomalous objects are spotted based on their deviating features. This work also uses relationships between individual objects as well as the detected anomalies, and detects groups of abnormal objects.\\n\\\\begin{figure}[!t]\\n \\\\centerline{\\\\includegraphics[width=0.48\\\\textwidth]{pics/Node/NetWalk.pdf}}\\n \\\\caption{The Framework of NetWalk~. The input dynamic graph is modeled as an initial graph with an incoming edge stream. Given the initial graph, a trunk of node sequences is generated using random walks on the graph and the deep autoencoder encodes each node into an embedding space following process (1). The node sequences and node embeddings are updated based on the incoming edge stream following process (2). Finally, NetWalk assigns an anomaly score to each node according to its distance to the cluster centroid in the embedding space.}\\n \\\\label{pic:netwalk}\\n\\\\end{figure}\",\n \"id\": \"a37030ed-a13e-4c9b-bdf9-32f760d9d73a\",\n \"level\": \"section\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ND on Dynamic Graphs With Traditional Non-Deep Learning Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ANOS ND on Dynamic Graphs With Traditional Non-Deep Learning Techniques\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 6293\n ],\n \"content\": \"\\\\label{appendix:edge}\\nTraditional non-deep learning based approaches mainly focus on using temporal signals (\\\\eg changes in graph structure), and applying specially designed statistical metrics to detect anomalous edges on dynamic graphs~. \\nAs a concrete example, Eswaran and Faloutsos~ modeled a dynamic graph as a stream of edges and exploited the graph structure as well as the structure evolving patterns. They identified two signs of anomalous edges: 1) connecting regions of the graph that were disconnected; and 2) connections that appear in bursts. For incoming edges, their model assigns anomaly scores to each edge, and the top-k edges with highest scores are anomalies. Another most recent work by Chang \\\\etal~ proposed a novel frequency factorization algorithm, aiming to spot anomalous incoming edges based on their likelihood of observed frequency. This method merges the advantages of both probabilistic models and matrix factorization for capturing both temporal and structural changes of nodes, and as reported, it only requires constant memory to handle edge streams.\",\n \"id\": \"ef880176-a742-4e90-bb20-77c1c0e4b593\",\n \"level\": \"section\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS ED With Traditional Non-Deep Learning Techniques\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ANOS ED With Traditional Non-Deep Learning Techniques\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{appendix:subgraph}\",\n \"id\": \"6dcc483a-2e1d-4f81-abe8-97ae152ee771\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5020eb09-4f12-4d1a-8dae-076499c8173a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS SGD With Traditional Non-Deep Learning Techniques\"\n ]\n ],\n \"subsections\": [\n \"a68d74d8-bc4c-4876-9c25-1421d3078bb4\",\n \"ab11e3f8-fc96-4814-ac49-21cd01dfc491\"\n ],\n \"title\": \"ANOS SGD With Traditional Non-Deep Learning Techniques\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6277,\n 6297\n ],\n \"content\": \"\\\\label{appendix:subgraph:static}\\nOne motivation of ANOS SGD in static graphs is that anomalous sub-graphs often exhibit significantly different attribute distributions.\\nTherefore, traditional non-deep learning techniques, such as gAnomaly~, AMEN~, and SLICENDICE~, focus on modeling the attribute distributions and measuring the normality of sub-graphs.\\nAnother line of investigation is graph residual analysis. The rich attribute information contained in real-world networks provides insight into the relationships formed between objects. Thus, the motivation behind several studies to spot anomalous sub-graphs has been to measure the residual between the expected structures and observed structures~.\",\n \"id\": \"a68d74d8-bc4c-4876-9c25-1421d3078bb4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"6dcc483a-2e1d-4f81-abe8-97ae152ee771\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS SGD With Traditional Non-Deep Learning Techniques\"\n ],\n [\n \"subsection\",\n \"ANOS SGD on Static Graphs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ANOS SGD on Static Graphs\"\n },\n {\n \"cite_extract_rate\": 0.18181818181818102,\n \"cites\": [\n 6272,\n 8997\n ],\n \"content\": \"\\\\label{appendix:subgraph:dynamic}\\nDevising metrics for ANOS SGD has been the subject of many traditional works. \\nFor instance, Chen \\\\etal~ introduced six metrics to identify community-based anomalies, namely: grown community, shrunken community, merged community, split community, born community and vanished community.\\nAlthough these hand-crafted features or statistical patterns well fit some particular types of existing anomalies, their abilities to detect unseen and camouflage anomalies are limited and applying them directly might introduce high false negative rate, which is not optimal for applications like financial security. \\nOther works, such as SPOTLIGHT by Eswaran \\\\etal~ and another by Liu \\\\etal~, explore sudden changes in dynamic graphs and identify anomalous sub-graphs that are related to such changes.\\nMotivated by the phenomena that social spam and fraud groups often form dense temporal sub-graphs in online social networks, plenty of works, including,~, use manually-extracted features and spot anomalous dense sub-graphs that have evolved significantly different from the reset of the graph. \\nIn addition to these studies, a large number of works discuss uses of various graph scan statistics for anomalous sub-graph detection, such as the Kulldorff statistic~, Poisson statistic~, elevated mean scan statistic~ and Berk-Jones statistic~. Specifically, Shao \\\\etal~ proposed a non-parametric method to detect anomalous sub-graphs in dynamic graphs where the network structure is constant, but the node attributes change overtime. This approach measures the anomalous score of each sub-graph with regard to the p-values of nodes it comprises. Sub-graphs with higher scores are more anomalous. Another work, GBGP~, instead, adopts the elevated mean scan statistic to identify nodes that might form anomalous sub-graphs and detects anomalous groups that follow predefined irregular structures.\\n\\\\end{document}\",\n \"id\": \"ab11e3f8-fc96-4814-ac49-21cd01dfc491\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"6dcc483a-2e1d-4f81-abe8-97ae152ee771\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Comprehensive Survey on\\\\\\\\ Graph Anomaly Detection with Deep Learning\"\n ],\n [\n \"section\",\n \"ANOS SGD With Traditional Non-Deep Learning Techniques\"\n ],\n [\n \"subsection\",\n \"ANOS SGD on Dynamic Graphs\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"ANOS SGD on Dynamic Graphs\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":95,"string":"95"},"all_cites":{"kind":"list like","value":[6272,3207,1671,553,3344,6271,6277,6275,217,1662,6274,6276,6270,6273,6278,6279,4539,6280,3206,282,6281,218,6282,7007,180,242,6283,6284,6285,6290,6288,6287,6289,6291,8991,6286,2626,6292,6293,8992,8149,3953,8993,6302,6297,3333,6303,6296,3700,6294,6304,6299,6301,1420,6295,6300,5645,6298,6305,8994,6307,6306,8995,6308,6309,8996,6310,8997],"string":"[\n 6272,\n 3207,\n 1671,\n 553,\n 3344,\n 6271,\n 6277,\n 6275,\n 217,\n 1662,\n 6274,\n 6276,\n 6270,\n 6273,\n 6278,\n 6279,\n 4539,\n 6280,\n 3206,\n 282,\n 6281,\n 218,\n 6282,\n 7007,\n 180,\n 242,\n 6283,\n 6284,\n 6285,\n 6290,\n 6288,\n 6287,\n 6289,\n 6291,\n 8991,\n 6286,\n 2626,\n 6292,\n 6293,\n 8992,\n 8149,\n 3953,\n 8993,\n 6302,\n 6297,\n 3333,\n 6303,\n 6296,\n 3700,\n 6294,\n 6304,\n 6299,\n 6301,\n 1420,\n 6295,\n 6300,\n 5645,\n 6298,\n 6305,\n 8994,\n 6307,\n 6306,\n 8995,\n 6308,\n 6309,\n 8996,\n 6310,\n 8997\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.136778198259202,"string":"1.136778"}}},{"rowIdx":63,"cells":{"authors":{"kind":"list like","value":["Wei Chen","Yu Liu","Weiping Wang","Erwin Bakker","Theodoros Georgiou","Paul Fieguth","Li Liu","Michael S. Lew"],"string":"[\n \"Wei Chen\",\n \"Yu Liu\",\n \"Weiping Wang\",\n \"Erwin Bakker\",\n \"Theodoros Georgiou\",\n \"Paul Fieguth\",\n \"Li Liu\",\n \"Michael S. Lew\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Deep Learning for Instance Retrieval: A Survey"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-01-27T09:32:58Z","string":"2021-01-27T09:32:58Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" In recent years a vast amount of visual content has been generated and shared from many fields, such as social media platforms, medical imaging, and robotics. This abundance of content creation and sharing has introduced new challenges, particularly that of searching databases for similar content --- Content Based Image Retrieval (CBIR) --- a long-established research area in which improved efficiency and accuracy are needed for real-time retrieval. Artificial intelligence has made progress in CBIR and has significantly facilitated the process of instance search. In this survey we review recent instance retrieval works that are developed based on deep learning algorithms and techniques, with the survey organized by deep feature extraction, feature embedding and aggregation methods, and network fine-tuning strategies. Our survey considers a wide variety of recent methods, whereby we identify milestone work, reveal connections among various methods and present the commonly used benchmarks, evaluation results, common challenges, and propose promising future directions. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"515509a6-319a-4385-b1b5-4418a83d2bc2","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"]],"subsections":["87ccd777-d058-4a76-9575-0c136cb10697","8a512096-a741-4540-8715-b12b364b3d9e","1b3ca748-df26-4b7f-85f3-3eb438a055f5","acffc1fe-bf03-450b-857b-db143905236e","a6c69e00-d8a1-492b-8514-fab570f109d2","6d16c13c-e552-486e-8c1d-7e8590b69688"],"title":"root"},{"cite_extract_rate":0.35714285714285704,"cites":[2093,2092,2091,2094,2090,97,8490,2089,2095,209],"content":"\\label{sec:intro}}\n\\IEEEPARstart{C}{ontent} Based Image Retrieval (CBIR) is the problem of searching for relevant images in an image gallery by analyzing the visual content (colors, textures, shapes, objects \\emph{etc.}), given a query image ,. CBIR has been a longstanding research topic in the fields of computer vision and multimedia ,. With the exponential growth of image data, the development of appropriate information systems that efficiently manage such large image collections is of utmost importance, with image searching being one of the most indispensable techniques. Thus there is a nearly endless potential for applications of CBIR, such as person/vehicle reidentification ,, landmark retrieval , \nremote sensing , online product searching .\nGenerally, CBIR methods can be grouped into two different tasks ,: Category level Image Retrieval (CIR) and Instance level Image Retrieval (IIR). The goal of CIR is to find an arbitrary image representative of the same category as the query (\\eg dogs, cars) ,. By contrast, in the IIR task, a query image of a particular instance (\\eg the Eiffel Tower, my neighbor's dog) is given and the goal is to find images containing the same instance that may be captured under different conditions like different imaging distances, viewing angles, backgrounds, illuminations, and weather conditions (reidentifying exemplars of the same instance) ,. The focus of this survey is the IIR task\\footnote{If not further specified, ``image retrieval'', ''IIR'', and ``instance retrieval'' are considered equivalent and will be used interchangeably.}.\nIn many real world applications, IIR is usually to find a desired image requiring a search among thousands, millions, or even billions of images. Hence, searching efficiently is as critical as searching accurately, to which continued efforts have been devoted ,,. To enable accurate and efficient retrieval over a large-scale image collection, \\emph{developing compact yet discriminative feature representations} is at the core of IIR.\n\\iffalse\n\\begin {figure}[!t]\n\\centering\n\\begin{tabular}{c}\n\\includegraphics[width=0.9\\columnwidth]{./Figures/TheProblem.pdf} \\\\\n(\\textit{a}) The CBIR problem \\\\\n\\includegraphics[width=0.9\\columnwidth]{./Figures/InstancevsCategory.pdf} \\\\\n(\\textit{b}) instance-level retrieval versus category-level retrieval\n \\end{tabular}\n\\caption{ Illustration of (a) the CBIR problem, and (b)\nCBIR categorization. The images in green frame are retrieved correctly, while the ones in red frame are matched incorrectly.\n}\n\\label{category_vs_instance}\n\\vspace{-1em}\n\\end {figure}\n\\fi\n\\begin{table*}[!ht]\n\\caption{A summary and comparison of the primary surveys in the field of image retrieval.}\\label{SurveyCompare}\n\\vspace{-1em}\n\\centering\n\\setlength{\\tabcolsep}{3.0mm}\\footnotesize\n\\begin{tabular}{!{\\vrule width0.3bp}p{5cm}|p{0.5cm}|p{1.6cm}|p{9.2cm}!{\\vrule width0.3bp}}\n\\Xhline{1.0pt}\n\\multicolumn{1}{|c|}{ \\multirow{2}*{ Title} } & \\multirow{2}*{ $\\!\\!\\!\\!$ Year} & \\multirow{2}*{ Published in} & \\multicolumn{1}{c|}{ \\multirow{2}*{ Main Content } } \\\\ \n\\hline\n\\multirowcell{2}{Content-Based Image Retrieval at the \\\\ End of the Early Years } & \\multirow{2}*{2000} & \\multirowcell{2}{ TPAMI } & \\multirowcell{2}{This paper discusses the steps for image retrieval systems, including image \\\\ processing, feature extraction, user interaction, and similarity evaluation. }\\\\\n\\hline\n\\multirowcell{2}{Image Search from Thousands to Billions \\\\in 20 Years } & \\multirow{2}*{2013} & \\multirowcell{2}{ TOMM} & \\multirowcell{2}{This paper gives a good presentation of image search achievements from \\\\ 1970 to 2013, but the methods are not deep learning-based.} \\\\\n\\hline\n\\multirowcell{2}{Deep Learning for Content-Based Image \\\\ Retrieval: A Comprehensive Study } & \\multirow{2}*{2014} & \\multirowcell{2}{ ACM MM} & \\multirowcell{2}{This paper introduces supervised metric learning methods for fine-tuning \\\\ AlexNet. Details of instance-based image retrieval are limited.}\\\\\n\\hline\n\\multirowcell{2}{Semantic Content-based Image Retrieval:\\\\ A Comprehensive Study } & \\multirow{2}*{2015} & \\multirowcell{2}{ JVCI } & \\multirowcell{2}{This paper presents a comprehensive study about CBIR using traditional \\\\ methods; deep learning is introduced as a section with limited details.}\\\\\n\\hline\n\\multirowcell{3}{Socializing the Semantic Gap: A Compa-\\\\rative Survey on Image Tag Assignment, \\\\Refinement, and Retrieval } & \\multirowcell{1}{2016} & \\multirowcell{3}{CSUR} & \\multirowcell{3}{A taxonomy is introduced to structure the growing literature of image \\\\ retrieval. Deep learning methods for feature learning is introduced as \\\\future work.} \\\\\n\\hline\n\\multirowcell{2}{Recent Advance in Content based Image \\\\ Retrieval: A Literature Survey } & \\multirow{2}*{2017} & \\multirowcell{2}{ arXiv } & \\multirowcell{2}{This survey presents image retrieval from 2003 to 2016. Neural networks \\\\ are introduced in a section and mainly discussed as a future direction.}\\\\\n\\hline\n\\multirowcell{3}{Information Fusion in Content-based \\\\Image Retrieval: A Comprehensive \\\\ Overview } & \\multirowcell{1}{2017} & \\multirowcell{3}{Information\\\\Fusion} & \\multirowcell{3}{This paper presents information fusion strategies in CBIR. \\\\ Deep convolutional networks for feature learning are introduced \\\\ briefly but not covered thoroughly.} \\\\\n\\hline\n\\multirowcell{2} { $\\!\\!\\!$ A Survey on Learning to Hash } & \\multirow{2}*{2018} & \\multirowcell{2}{ TPAMI} & \\multirowcell{2}{This paper focuses on hash learning algorithms and introduces the \\\\ similarity-preserving methods and discusses their relationships. } \\\\\n\\hline\n\\multirowcell{2}{$\\!\\!\\!$ SIFT Meets CNN: A Decade Survey of \\\\ Instance Retrieval } & \\multirow{2}*{2018} & \\multirowcell{2}{ TPAMI} & \\multirowcell{2}{This paper presents a comprehensive review of instance retrieval based on \\\\SIFT and CNN methods.}\\\\\n\\hline\n\\multirowcell{3}{ \\bf Deep Learning for Instance \\\\ \\bf Retrieval: A Survey } & \\multirowcell{1}{\\bf 2021} & \\multirowcell{3}{ \\bf Ours} & \\multirowcell{3}{\\bf Our survey focuses on deep learning methods. We expand the review \\\\ \\bf with indepth details on IIR, including methods of feature extraction, \\\\ \\bf feature embedding and aggregation, and network fine-tuning. }\n \\\\\n\\Xhline{1.0pt}\n\\end{tabular}\n\\end{table*}\nDuring the past two decades, startling progress has been witnessed in image representation which mainly consists of two important periods, \\ie feature engineering and deep learning. In the feature engineering era, the field was dominated by various milestone handcrafted image representations like SIFT and Bag of visual Words (BoW) . The deep learning era was reignited in 2012 when a Deep Convolutional Neural Network (DCNN) referred as ``AlexNet'' won the first place in the ImageNet classification contest with a breakthrough reduction in classification error rate. Since then, the dominating role of SIFT like local descriptors has been replaced by data driven Deep Neural Networks (DNNs) which can learn powerful feature representations with multiple levels of abstraction directly from data. During the past decade, DNNs have set the state of the art in various classical computer vision tasks, including image classification~,, object detection~, semantic segmentation , \nand image retrieval~.\nGiven this period of rapid evolution, the goal of this paper is to provide a comprehensive survey of the recent achievements in IIR. In comparison with existing excellent surveys on traditional image retrieval ,,,, as contrasted in Table~\\ref{SurveyCompare}, our focus in this paper is reviewing deep learning based methods for IIR, particularly on issues of retrieval accuracy and efficiency.","id":"87ccd777-d058-4a76-9575-0c136cb10697","level":"section","origin_cites_number":28,"parent_id":"515509a6-319a-4385-b1b5-4418a83d2bc2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Introduction"]],"subsections":["d77f2bd6-4077-4eb5-ba4e-7c6288149714","846c213c-7851-4f7e-b5f1-17d26ed0cdb1"],"title":"Introduction"},{"cite_extract_rate":0.5789473684210521,"cites":[7543,2101,514,2096,2098,2102,2105,2108,2094,2104,2103,8524,7544,2109,2099,2097,2106,97,2107,629,2100,7545],"content":"\\label{Summary_of_Progress}\nAfter the highly successful image classification implementation of AlexNet~, significant exploration of DCNNs for instance retrieval tasks was undertaken and representative methods are shown in Figure \\ref{fig1}. Building on these methods, more recent progress for IIR can be achieved from the perspectives of off-the-shelf models and fine-tuned models, which form the basis for this survey.\nOff-the-shelf models, based on DCNNs with fixed parameters ,,, extract features at image scales or patch scales, which correspond to single-pass and multiple-pass schemes, respectively. These methods focus on effective feature use, for which researchers have proposed embedding and aggregation methods, such as R-MAC , CroW , and SPoC to promote the discriminativity of the extracted features. Fine-tuned models, based on DCNNs in which parameters are updated , are more popular since deep networks themselves have been investigated extensively. To learn better retrieval features, researchers have proposed to improve the network architectures and/or update the pre-stored parameters . \nThis survey will explore recent progress in detail in the context of the three following themes:\n\\begin{figure*}[!t]\n\\centering\n\\includegraphics[width=\\textwidth]{./Figures/RepresentativeMethods.pdf} \n\\vspace{-1.5em}\n\\caption{Representative methods in IIR. Off-the-shelf models have model parameters which are not further updated or tuned when extracting retrieval features. For single-pass schemes, the key step is the feature embedding and aggregation to promote the discriminativity of the extracted image-level activations ,,,, whereas for multiple-pass schemes the goal is to extract instance features at region scales and eliminate image clutter as much as possible ,,,. In contrast, for fine-tuned models, the model parameters are tuned towards the retrieval task and address the issue of domain shifts. For supervised fine-tuning, the key step lies in the design of objective functions and sample sampling strategies ,,,,, while the success of unsupervised fine-tuning is to mine the relevance among training samples ,,,,. See Sections \\ref{Retrieval_with_Off_the_Shelf_DCNN_Models} and \\ref{Retrieval_via_Learning_DCNN_Representations} for details.}\n\\label{fig1}\n\\end{figure*}\n\\medskip\n\\noindent (1) \\emph{Deep Feature Extraction} \\hfill (Section \\ref{Deep_Feature_Extraction})\n\\smallskip\n\\noindent \nOne key step in IIR is to make the descriptors as semantic-aware , as possible. For this, some recent work focus on the input data of DCNNs, thereby instance features can be extracted from the whole image, \\eg CroW , VLAD-CNN or from image patches, \\eg MOP-CNN , FAemb . For instance, evaluated on the Holidays dataset , patch-level input scheme can improve mAP by 8.29\\% compared to the results (70.53\\%) achieved using image-level input . Others focus on exploring different feature extractors, \\eg one layer of a given DCNN, to get the output activations. Initially, fully-connected layers are usually chosen to extract features ,, and then convolutional layers are popularly used ,. Afterwards, some work leverage the complementarity of different extractors to explore layer-level fusion, such as MoF , and model-level fusion, such as DeepIndex to promote the retrieval performance.\n\\medskip\n\\noindent \n(2) \\emph{Feature Embedding and Aggregation} \\hfill (Section \\ref{Deep_Feature_Enhancement})\n\\smallskip\n\\noindent \nRecent work revisit the classical embedding and aggregation methods and apply on deep features. Most work have a preference towards mapping individual vector from convolutional layer ,, and then aggregating into a global feature. The mapping process can be realized by using a pre-trained codebook (\\ie built separately), such as VLAD-CNN , DeepIndex or learned as parameters during training (built simultaneously), such as NetVLAD , GeM-DSM . Some work aggregate local features into a global one by direct pooling or sophisticated pooling-based methods without the aggregation operation, such as R-MAC .\n\\medskip\n\\noindent \n(3) \\emph{Network Fine-tuning for Learning Representations} \\hfill (Section \\ref{Retrieval_via_Learning_DCNN_Representations})\n\\smallskip\n\\noindent \nDCNNs pretrained on source datasets for image classification are influenced by domain shifts when performing retrieval tasks on new datasets. Thus, it is necessary to fine-tune deep networks to the specific domain , by using supervised or unsupervised fine-tuning methods. As depicted in Figure \\ref{fig1}, recent supervised fine-tuning methods focus on designing objective functions (\\eg Listwise loss ) and sample sampling strategies, such as NetVLAD , Triplet Network . Unsupervised methods focus on mining the relevance among training samples by using clustering, such as SfM-GeM or manifold learning, such as AILIR . Recently, convolution-free models that only rely on transformer layers have shown competitive performance and been used as a powerful alternative to DCNNs, such as IRT , reranking Transformers .","id":"d77f2bd6-4077-4eb5-ba4e-7c6288149714","level":"subsection","origin_cites_number":38,"parent_id":"87ccd777-d058-4a76-9575-0c136cb10697","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Introduction"],["subsection","Summary of Progress since 2012"]],"subsections":[],"title":"Summary of Progress since 2012"},{"cite_extract_rate":0.75,"cites":[2095,2110,2112,2111,2109,2105],"content":"\\label{Keychallenges}\nThe goal of each of the preceding three themes is to address the competing objectives of \\emph{accuracy} and \\emph{efficiency}, with both objectives continuing to present challenges:\n\\textbf{A) Accuracy related challenges} depend on the input data, the feature extractor, and the way in which the extracted features are processed:\n\\begin{itemize}\n\\item Invariance challenge: The instance in an image may be rotated, translated, or scaled differently, so the final features are affected by these transformations and retrieval accuracy may be degraded . It is necessary to incorporate invariance into the IIR pipeline ,.\n\\item Distraction challenge: IIR systems may need to focus on only a certain object, or even only a small portion. DCNNs may be affected by image clutter or background, so multiple-pass schemes have been examined where region proposals are studied before feature extraction.\n\\item Discriminativity challenge: Deep features for IIR are needed to be as discriminative as possible to distinguish instances with small differences, leading to many explorations in feature processing. These include feature embedding and aggregation methods, to promote feature discriminativity; and attention mechanisms, to highlight the most relevant regions within the extracted features or to enable deep networks to focus on regions of interest. \n\\item Fine-tune challenge: DCNNs can be fine-tuned as powerful extractors to capture fine semantic distinctions among instances. These explorations have offered improved accuracy, however the area remains a major challenge. \n\\end{itemize}\n\\begin{figure*}[!t]\n\\centering\n\\includegraphics[width=0.95\\textwidth]{./Figures/PipelineofImageRetrieval.pdf}\n \\vspace{-0.5em}\n\\caption{General framework of IIR, which includes feature extraction on image or image patches, followed by feature embedding and aggregation methods to improve feature discriminativity. Feature matching can be performed by using global features (initial filtering) or use local features to rerank the top-ranked images matched by global features. } \n\\label{PipelineofImageRetrieval}\n\\end{figure*}\n\\textbf{B) Efficiency related challenges} are important, especially for large-scale datasets . Retrieval systems should respond quickly when given a query image. Deep features are high-dimensional and contain semantic-aware information to support higher accuracy, yet is often at the expense of efficiency. \nOn the one hand, the efficiency is related to the \nformat of features, \\ie real valued or binary. Hash codes have advantages in storage and searching ,, however for hashing methods one needs to carefully consider the loss function design ,, to obtain optimal codes for high retrieval accuracy. \nOn the other hand, efficiency is also related to the mechanism of feature matching. For example, instead of time-consuming cross-matching between local features, one can choose to use global features to perform an initial ranking and then a post-step re-ranking via the features of top-ranked images. \n\\vspace{-1em}","id":"846c213c-7851-4f7e-b5f1-17d26ed0cdb1","level":"subsection","origin_cites_number":8,"parent_id":"87ccd777-d058-4a76-9575-0c136cb10697","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Introduction"],["subsection","Key Challenges"]],"subsections":[],"title":"Key Challenges"},{"cite_extract_rate":0.36842105263157804,"cites":[2093,2113,2094,2115,2114,2110,7545],"content":"\\label{General_Framework_of_IIR}\nFigure \\ref{PipelineofImageRetrieval} offers an overview of the general framework for deep-learning-based IIR, involving three main stages. \n\\medskip\n\\noindent \n\\textbf{1) Deep feature extraction}: \\hfill (Section \\ref{Deep_Feature_Extraction})\n\\smallskip\n\\noindent\nFeature extraction is the first step of IIR and can be realized in a single-pass or multiple-pass way. Single-pass methods take as input the whole image, whereas multiple-pass methods depend on region extraction, as depicted in Figure \\ref{MultiplePass}.\nThe activations from fully-connected layers of a given DCNN can be used as retrieval features whether based on a whole image or on patches. The tensors from convolutional layers can be used when further processed by sophisticated pooling, as shown in Figure \\ref{PipelineofImageRetrieval}. Different layers of the same deep network can be combined as a more powerful extractor \n,. Furthermore, it is possible to fuse the activations from layers of different models \n,. Feature extraction is the step to produce vanilla network activations (\\ie 3D tensors or a single vector), these activations, in most cases, are needed to be further processed. \n\\medskip\n\\noindent \n\\textbf{2) Embedding and aggregation}: \\hfill (Section \\ref{Deep_Feature_Enhancement})\n\\smallskip\n\\noindent\nFeature embedding and aggregation are two essential steps to produce global or local features. Feature embedding maps individual local features into higher-dimensional space whereas feature aggregation summarizes the multiple mapped vectors or all individual features into a global vector. Global features may come from pooling convolutional feature maps directly , or using some sophisticated weighting methods , (\\ie both without feature embedding). Feature embedding method using a pre-generated codebook can be performed to encode individual convolutional vectors and then aggregated ,,. For local features, the well-embedded representations for all regions of interest are stored individually and used for cross-matching in the reranking stage without aggregation.\n\\medskip\n\\noindent \n\\textbf{3) Feature matching}:\n\\smallskip\n\\noindent\nFeature matching is a process to measure the feature similarity between images and then return a ranked list. Global matches can be computed efficiently via such as Euclidean distance. For local features ,, the image similarity is usually evaluated by summarizing the similarities across local features, using classical RANSAC or more recent variations ,. Storing local features separately and then estimating their similarity individually lead to additional memory and search costs ,, therefore in most cases local features are used to re-rank the initial ranking image matched by global features ,,,.\nThe three preceding stages for IIR rely on DCNNs as backbone architectures. In almost all cases, pre-stored parameters in these backbones can be fine-tuned (Section \\ref{Retrieval_via_Learning_DCNN_Representations}) to be better suited for instance retrieval and to contribute to better performance. \nThe detailed categorization of the material of the following sections is shown in Figure~\\ref{FourAspectsofprogress}.","id":"8a512096-a741-4540-8715-b12b364b3d9e","level":"section","origin_cites_number":19,"parent_id":"515509a6-319a-4385-b1b5-4418a83d2bc2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","General Framework of IIR"]],"subsections":[],"title":"General Framework of IIR"},{"cite_extract_rate":0.5769230769230761,"cites":[2118,514,2096,2113,2105,2094,2103,7544,2119,2097,2109,2116,629,7545,2117],"content":"\\label{Retrieval_with_Off_the_Shelf_DCNN_Models}\n\\textcolor{black}{Because of their size, DCNNs need to be trained, initially for classification tasks, on exceptionally large-scale datasets and be able to recognize images from different classes.} One possible scheme, then, is that DCNNs effectively trained for classification directly serve as off-the-shelf feature detectors for the image retrieval task, the topic of this survey. That is, one can propose to undertake image retrieval on the basis of DCNNs, trained for classification and with their pre-trained parameters frozen. \n\\begin{figure}[!htbp]\n\\centering\n\\footnotesize\n\\begin{tikzpicture}[xscale=0.8, yscale=0.4]\n\\draw [thick, -] (0, 16.5) -- (0, -4); \\node [right] at (-0.5, 17) {\\bf \\em Deep Learning for Instance-level Image Retrieval (overall survey)};\n\\draw [thick, -] (0, 16) -- (0.5, 16);\\node [right] at (0.5, 16) {Retrieval with Off-the-Shelf DCNN Models (Section \\ref{Retrieval_with_Off_the_Shelf_DCNN_Models})};\n\\draw [thick, -] (1, 15.5) -- (1, 5);\n\\draw [thick, -] (1, 15) -- (1.5, 15);\\node [right] at (1.5, 15) {\\bf Deep Feature Extraction (Section \\ref{Deep_Feature_Extraction})};\n\\draw [thick, -] (2, 14.5) -- (2, 8); \n\\draw [thick, -] (2, 14) -- (2.5, 14);\\node [right] at (2.5, 14) {Network Feedforward Scheme (Section \\ref{Network_Feedforward_Scheme})};\n\\draw [thick, -] (3, 13.5) -- (3, 12); \n\\draw [thick, -] (3, 13) -- (3.5, 13);\\node [right] at (3.5, 13) {Single Feedforward Pass: MAC , R-MAC };\n\\draw [thick, -] (3, 12) -- (3.5, 12);\\node [right] at (3.5, 12) {Multiple Feedforward Pass: SPM , RPNs };\n\\draw [thick, -] (2, 11) -- (2.5, 11);\\node [right] at (2.5, 11) {Deep Feature Selection (Section \\ref{Deep_Feature_Selection}) };\n\\draw [thick, -] (3, 10.5) -- (3, 9); \n\\draw [thick, -] (3.0, 10) -- (3.5, 10);\\node [right] at (3.5, 10) { Fully-connected Layer: Neural codes };\n\\draw [thick, -] (3.0, 9) -- (3.5, 9);\\node [right] at (3.5, 9) { Convolutional Layer: SPoC , CroW };\n\\draw [thick, -] (2.0, 8) -- (2.5, 8);\\node [right] at (2.5, 8) {Feature Fusion Strategy (Section \\ref{Feature_Fusion_Strategy}) };\n\\draw [thick, -] (3.0, 7.5) -- (3.0, 6);\n\\draw [thick, -] (3.0, 7) -- (3.5, 7);\\node [right] at (3.5, 7) { Layer-level Fusion: MoF , MOP };\n\\draw [thick, -] (3.0, 6) -- (3.5, 6);\\node [right] at (3.5, 6) { Model-level Fusion: ConvNet fusion };\n\\draw [thick, -] (1, 5) -- (1.5, 5);\\node [right] at (1.5, 5) {\\bf Feature Embedding and Aggregation (Section \\ref{Deep_Feature_Enhancement}) };\n\\draw [thick, -] (2, 4.5) -- (2, -1); \n\\draw [thick, -] (2.0, 4) -- (2.5, 4);\\node [right] at (2.5, 4) { Matching with Global Features (Section \\ref{Matching_with_Global_Feature})};\n\\draw [thick, -] (2.0, 3) -- (2.5, 3);\\node [right] at (2.5, 3) {Matching with Local Features (Section \\ref{Matching_with_Local_Feature})};\n\\draw [thick, -] (2.0, 2) -- (2.5, 2);\\node [right] at (2.5, 2) { Attention Mechanism (Section \\ref{Attention_Mechanism}) };\n\\draw [thick, -] (3.0, 1.5) -- (3.0, 0);\n\\draw [thick, -] (3.0, 1) -- (3.5, 1);\\node [right] at (3.5, 1) { Non-parameteric: CroW , SPoC , SWVF };\n\\draw [thick, -] (3.0, 0) -- (3.5, 0);\\node [right] at (3.5, 0) { Parameteric: CRN , DeepFixNet+SAM , };\n\\draw [thick, -] (2.0, -1) -- (2.5, -1);\\node [right] at (2.5, -1) {Deep Hash Embedding (Section \\ref{Deep_Hash_Embedding}) };\n\\draw [thick, -] (3.0, -1.5) -- (3.0, -3);\n\\draw [thick, -] (3.0, -2) -- (3.5, -2);\\node [right] at (3.5, -2) { Supervised Hashing: SSDH };\n\\draw [thick, -] (3.0, -3) -- (3.5, -3);\\node [right] at (3.5, -3) { Unsupervised Hashing: DeepBit , DSTH };\n\\draw [thick, -] (0, -4) -- (0.5, -4);\\node [right] at (0.5, -4) {\\bf Fine-Tuning for Learning DCNN Representations (Section \\ref{Retrieval_via_Learning_DCNN_Representations})};\n\\draw [thick, -] (1, -4.5) -- (1, -11);\n\\draw [thick, -] (1, -5) -- (1.5, -5);\\node [right] at (1.5, -5) {Supervised Fine-tuning (Section \\ref{Supervised_Fine-tuning}) };\n\\draw [thick, -] (2, -5.5) -- (2, -7); \n\\draw [thick, -] (2.0, -6) -- (2.5, -6);\\node [right] at (2.5, -6) {Fine-tuning via classification loss (Section \\ref{Classification-based_Fine-tuning}) }; \n\\draw [thick, -] (2.0, -7) -- (2.5, -7);\\node [right] at (2.5, -7) {Fine-tuning via pairwise ranking loss (Section \\ref{Verification_based_Learning}) };\n\\draw [thick, -] (3.0, -7.5) -- (3.0, -10);\n\\draw [thick, -] (3.0, -8) -- (3.5, -8);\\node [right] at (3.5, -8) { Transformation Matrix: Non-metric };\n\\draw [thick, -] (3.0, -9) -- (3.5, -9);\\node [right] at (3.5, -9) { Siamese Networks };\n\\draw [thick, -] (3.0, -10) -- (3.5, -10);\\node [right] at (3.5, -10) { Triplet Networks };\n\\draw [thick, -] (1.0, -11) -- (1.5, -11);\\node [right] at (1.5, -11) { Unsupervised Fine-tuning (Section \\ref{Unsupervised_Fine-tuning})};\n\\draw [thick, -] (2, -11.5) -- (2, -13); \n\\draw [thick, -] (2.0, -12) -- (2.5, -12);\\node [right] at (2.5, -12) { Manifold Learning Samples Mining: Diffusion Net };\n\\draw [thick, -] (2.0, -13) -- (2.5, -13);\\node [right] at (2.5, -13) {Mining Samples by Clustering: SfM-GeM , };\n\\end{tikzpicture}\n\\vspace{-2em}\n\\caption{This survey is organized around three key themes in instance-level image retrieval, shown in bold.}\n\\label{FourAspectsofprogress}\n\\end{figure}\nThere are limitations to this approach, most fundamentally that there is a model-transfer or domain-shift challenge between tasks~,,, meaning that models trained for classification do not necessarily extract features well suited to image retrieval. In particular, a classification decision can be successful as long as features remain within classification boundaries, however features from such models may show insufficient capacity for retrieval where feature matching itself is more important than classification. This section will survey the strategies which have been developed to improve the quality of feature representations, particularly based on feature extraction / fusion (Section \\ref{Deep_Feature_Extraction}) and feature embedding / aggregation (Section \\ref{Deep_Feature_Enhancement}).","id":"1b3ca748-df26-4b7f-85f3-3eb438a055f5","level":"section","origin_cites_number":26,"parent_id":"515509a6-319a-4385-b1b5-4418a83d2bc2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"]],"subsections":["f1ce1975-10d7-4d60-8e5d-320750dbf247","f6eab6de-c628-4a89-bfb0-43b4f87b1151"],"title":"Retrieval with Off-the-Shelf DCNN Models"},{"cite_extract_rate":0,"cites":[],"content":"\\label{Deep_Feature_Extraction}\nFeature extraction is about the mechanism by which retrieval features can be extracted from off-the-shelf DCNNs. \nFor an input image $x$ and a network $f(\\cdot; \\bm{\\theta})$, we denote its features from a convolutional layer as $ \\boldsymbol{A} := f_{conv}(x) \\in \\mathbb{R}^{H \\times W \\times C}$ with height $H$, width $W$, and channels $C$ while that from a fully-connected layer as $ \\boldsymbol{B} := f_{fc}(x) \\in \\mathbb{R}^{D \\times 1}$ with the dimensional $D$.","id":"f1ce1975-10d7-4d60-8e5d-320750dbf247","level":"subsection","origin_cites_number":0,"parent_id":"1b3ca748-df26-4b7f-85f3-3eb438a055f5","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection"," Deep Feature Extraction"]],"subsections":["d9000dca-9ea2-4e50-b3cd-ac34b0c46c16","4be11049-ece2-473a-babd-9ac8ce82a05f","f5333f99-42b6-4668-a652-9777bc115d47"],"title":" Deep Feature Extraction"},{"cite_extract_rate":0.518518518518518,"cites":[7543,2105,8525,7546,2094,2103,2109,2120,2121,2114,2107,2112,209,7545],"content":"\\label{Network_Feedforward_Scheme}\nNetwork feedforward schemes focus on how images are fed into a DCNN, which includes single-pass and multiple-pass.\n\\begin{flushleft}\n\\emph{a. Single Feedforward Pass Methods}. \n\\end{flushleft}\nSingle feedforward pass methods take the whole image and feed it into an off-the-shelf model to extract features. The approach is relatively efficient since the input image is fed only once. For these methods, both the fully-connected layer and last convolutional layer can be used as feature extractors .\nEarly network-based IIR work focused on leveraging DCNNs as a fixed extractor to obtain global features, especially based on the fully-connected layers ,, requiring close to zero engineering effort. However, extracting features in this way affects retrieval accuracy since the extracted features may include background information or activations for irrelevant objects.\nThe key to single-pass schemes is to embed and aggregate features to improve their discriminativity, such that features of two related images (\\ie including the same object) are more similar than these of two unrelated images . For this purpose, it is possible to first map the features $ \\boldsymbol{B} $ into a high-dimensional space and then to aggregate them into a final global feature . Another direction is to treat regions in convolutional features $ \\boldsymbol{A} $ as different sub-vectors, such that a combination of sub-vectors of all feature maps are used to represent the input image ,.\n\\begin{flushleft}\n\\emph{b. Multiple Feedforward Pass Methods}. \n\\end{flushleft}\nCompared to single-pass schemes, multiple-pass methods are more time-consuming because several patches are generated and then fed into the network, \n\\textcolor{black}{\nmultiple-pass schemes are more helpful for addressing the ``\\textit{invariance challenges} and ``\\textit{distraction challenges}'' in Section \\ref{Keychallenges}. Local patches at multiple scales become more robust for image translation, scaling and rotation ,. Also, these patches are helpful to filter several irrelevant background information. \n}\nThe representations are usually produced from two stages: patch detection and patch description. Multi-scale image patches are obtained using sliding windows , or spatial pyramid model (SPM) ,,, as shown in Figure \\ref{MultiplePass}. For example, Zheng \\etal partition an image by using SPM and extract features at increasing scales, thus enabling the integration of global, regional, local contextual information.\nPatch detection methods lack retrieval efficiency since irrelevant patches are also detected . For example, Cao \\etal propose to merge image patches into larger regions with different hyper-parameters, where the hyper-parameter selection is viewed as an optimization problem to maximize the similarity between query and candidate features.\n\\begin {figure}[!t]\n\\centering\n { \n \\includegraphics[width=\\columnwidth]{./Figures/MultiplePass.pdf} \n }\n \\vspace{-2em}\n\\caption{Image patch generation schemes: (a) Sliding windows ,; (b) Spatial pyramid modeling ; (c) Dense sampling ,; (d) Region proposals from region proposal networks ,.}\n\\vspace{-1em}\n\\label{MultiplePass}\n\\end {figure}\nInstead of generating multi-scale image patches randomly or densely, region proposal methods introduce a degree of purpose. Region proposals can be generated using object detectors, such as selective search , edge boxes ,, and BING . For example, Yu \\etal propose fuzzy object matching (FOM) for instance search in which the fuzzy objects are generated from 300 object proposals and then clustered to filter out overlapping proposals. Region proposals can also be learned using such as region proposal networks (RPNs) , and convolutional kernel networks (CKNs) , and then to apply these networks into end-to-end fine-tuning for learning similarity . \\textcolor{black}{This usually requires the datasets provide well-localized bounding boxes as supervision, \\eg the datasets INSTRE , Oxford-5k , Paris-6k , GLD-v2 variant . Also, in the off-the-shelf scenarios, the way that using the bounding boxes to crop the query images and use as input the DCNNs has been shown to provide better retrieval performance since only the information relevant to the instance is extracted ,.\n}","id":"d9000dca-9ea2-4e50-b3cd-ac34b0c46c16","level":"subsubsection","origin_cites_number":27,"parent_id":"f1ce1975-10d7-4d60-8e5d-320750dbf247","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection"," Deep Feature Extraction"],["subsubsection"," Network Feedforward Scheme"]],"subsections":[],"title":" Network Feedforward Scheme"},{"cite_extract_rate":0.6666666666666661,"cites":[7547,7543,2122,2102,2113,2105,2108,2094,2104,2109,2110,2112,7548,7545],"content":"\\label{Deep_Feature_Selection}\nFeature selection decides the receptive field of the extracted features, \\ie global-level from fully-connected layers and regional-level from convolutional layers.\n\\begin{flushleft}\n\\emph{a. Extracted from Fully-connected Layers}\n\\end{flushleft}\nIt is straightforward to select a fully-connected layer as a global feature extractor ,,. With PCA dimensionality reduction and normalization~ image similarity can be measured. Extracting features $ \\boldsymbol{B} $ from fully-connected layer leads to two obvious limitations for IIR: including irrelevant information, and a lack of local geometric invariance .\nWith regards to the first limitation, image-level global descriptors may include irrelevant patterns or background clutter, especially when a target instance is only a small portion of an image. It may then be more reasonable to extract region-level features at finer scales, \\ie using multiple passes~,,.\nFor the second limitation, \nan alternative is to extract multi-scale features on a convolutional layer ,. Further, it makes the global features incompatible with techniques such as spatial verification and re-ranking. Several methods then choose to leverage intermediate convolutional layers ,,,.\n\\begin{flushleft}\n\\emph{b. Extracted from Convolutional Layers}\n\\end{flushleft}\nThe neurons in a convolutional layer are connected only to a local region of the input image, and this smaller receptive field ensures that the produced features $ \\boldsymbol{A} $, usually from the last layer, preserve more local structural information , and are more robust to image transformations \\textcolor{black}{ thereby address the ``\\textit{invariance challenge}''. For instance, Razavian \\etal extract multi-scale features on the last convolutional layer and Mor{\\`e}re \\etal incorporate a series of nested pooling layers into CNN. Both of them provide higher feature invariance. Thus, many image retrieval methods use convolutional layers as feature extractors~ ,,,.}\nSum/average and max pooling are two simple aggregation methods to produce global features . For a pooled layer, the last convolutional layer usually yields superior accuracy over other shallower or later fully-connected layers . There is no other operation on the feature maps before pooling, so we illustrate these methods as ``direct pooling'' in Figure~\\ref{PipelineofImageRetrieval}.\n\\begin {figure}[!t]\n\\centering\n { \n \\includegraphics[width=\\columnwidth]{./Figures/SinglePass.pdf} \n }\n \\vspace{-1.5em}\n\\caption{Representative methods in single-pass methods, focusing on convolutional feature tensor $ \\boldsymbol{A} $. We denote the entry in $ \\boldsymbol{A} $ corresponding to channel $ c $, at spatial location ($i, j$) as $ A_{ijc} $:\nMAC , R-MAC , SPoC with the per-channel Gaussian weighting $\\alpha^{\\prime}_{ij} A_{ij}$ where $\\alpha^{\\prime}_{ij} = \\exp\\left\\{-{\\textstyle\\frac{\\left(i-\\frac H2\\right)^2+\\left(j-\\frac W2\\right)^2}{2\\sigma^2}}\\right\\} $ , CroW with $ \\alpha^{\\prime\\prime} $ computed by summing all $C$ feature maps at location ($i, j$) and $ \\beta $ computed by summing the $H\\times W$-array at each feature map $c$ , GeM with channel-wise powers operation , and CAM+CroW by performing $ M_{ij}^{(l)}=\\sum_{c=1}^C\\omega_{lc}A_{ijc} $ where $\\omega_{lc}$ are weights activated by $l$-th class . \n}\n\\vspace{-1em}\n\\label{SinglePass}\n\\end {figure}\nInstead of direct pooling, many sophisticated aggregation methods have been explored, such as channel-wise or spatial-wise feature weighting on the convolutional feature maps ,,. These aggregation methods aim to highlight feature importance or reduce the undesirable influence of bursty descriptors of some regions ,. For clarity, we illustrate the representative strategies in Figure \\ref{SinglePass}. \nNote that these feature aggregation methods are usually performed before channel-wise sum/max pooling and does not embed features into a higher dimensional space.\nOne rationale behind using convolutional features is that each such vector can act as a ``dense SIFT'' feature since each vector corresponds to a region in the input image. Inspired by this perception, many works leverage embedding methods (\\eg BoW) used for SIFT features on the regional feature vectors and then aggregate them (\\eg by sum pooling) into a global descriptor. \n\\textcolor{black}{Feature embedding methods address the discriminativity challenge via mapping individual features into a high-dimensional space and make them distinguishable .} Feature embedding is followed by PCA to reduce feature dimensionality and whitening to down-weight co-occurrence between features.","id":"4be11049-ece2-473a-babd-9ac8ce82a05f","level":"subsubsection","origin_cites_number":21,"parent_id":"f1ce1975-10d7-4d60-8e5d-320750dbf247","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection"," Deep Feature Extraction"],["subsubsection","Deep Feature Selection"]],"subsections":[],"title":"Deep Feature Selection"},{"cite_extract_rate":0.388888888888888,"cites":[2124,2115,2123,7548,2125,514,2105],"content":"\\label{Feature_Fusion_Strategy}\nFusion studies the complementarity of different features which includes layer-level and model-level fusion explorations.\n\\begin{flushleft}\n\\emph{a. Layer-level Fusion}\n\\end{flushleft}\nWith layer-level fusion it is possible to fuse multiple fully-connected layers in a deep network ,. For instance, Liu \\etal introduce DeepIndex to incorporate multiple global features from different fully connected layers. The activation from the first fully-connected layer is taken as column indexing, and that from the second layer serves as row indexing. Similarly, it is also possible to fuse the activations from multiple convolutional layers. For instance, Li \\etal apply the R-MAC encoding scheme on five convolutional layers of VGG-16 and then concatenate them into a multi-scale feature vector.\nFeatures from fully-connected layers retain global high-level semantics, whereas features from convolutional layers can present local low- and intermediate-level cues. Global and local features therefore complement each other when measuring semantic similarity and can, to some extent, guarantee retrieval performance ,. Such features can be concatenated directly ,, with convolutional features normally filtered by sliding windows or region proposal nets. Direct concatenation can also be replaced by other advanced methods, such as orthogonal operations or pooling-based methods, such as Multi-layer Orderless Fusion (MOF) of Li \\etal , which is inspired by Multi-layer Orderless Pooling (MOP) . However local features cannot play a decisive role in distinguishing subtle feature differences if global and local features are treated identically. Yu \\etal use a mapping function to assert local features in refining the return ranking lists, via an exponential mapping function for tapping the complementary strengths of convolutional and fully-connected layers. Similarly, \nLiu \\etal design two sub-networks on top of convolutional layers to obtain global and local features and then learn to fuse these features, thereby adaptively adjusting the fusion weights. Instead of directly fusing the layer activations, Zhang \\etal fuse the index matrices which are generated based on the two feature types extracted from the same CNN, a feature fusion which has low computational complexity.\nIt is worth considering which layer combinations are better for fusion given their differences and complementarity. Yu \\etal compare the performance of different combinations between fully-connected and convolutional layers on the Oxford 5k, Holiday, and UKBench datasets. The results show that the combinations including the first fully-connected layer always perform better. Li \\etal demonstrate that fusing convolutional and fully-connected layers outperforms the fusion of only convolutional layers. Fusing two convolutional layers with one fully-connected layer achieves the best performance on the Holiday and UKBench datasets.\n\\begin{flushleft}\n\\emph{b. Model-level Fusion}\n\\end{flushleft}\nIt is possible to combine features from different models; such a fusion focuses more on model complementarity, with methods categorized into \\emph{intra-model} and \\emph{inter-model}.\nIntra-model fusion suggests multiple deep models having similar or highly compatible structures, while inter-model fusion involves models with differing structures. For example, \nSimonyan \\etal introduce a ConvNet intra-model fusion strategy to improve the feature learning capacity of VGG where VGG-16 and VGG-19 are fused. \nTo attend to different parts of an image object, Wang \\etal realize the multi-feature fusion by selecting all convolutional layers of VGG-16 to extract image representations, which is demonstrated to be more robust than using only single-layer features.\nInter-model fusion is a way to bridge different features given the fact that different deep networks have different receptive fields ,,,. For instance, a two-stream attention network is introduced to implement image retrieval where the main network for semantic prediction is VGG-16 while an auxiliary network is used for predicting attention maps. Similarly, considering the importance and necessity of inter-model fusion to bridge the gap between mid-level and high-level features, Liu \\etal and Zheng \\etal combine VGG-19 and AlexNet to learn combined features, while Ozaki \\etal concatenate descriptors from six different models. \nInter-model and intra-model fusion are relevant to model selection. There are some strategies to determine how to combine the features from two models. It is straightforward to fuse all features from the candidate models and then learn a metric based on the concatenated features ,, which is a kind of ``\\emph{early fusion}'' strategy. Alternatively, it is also possible to learn optimal metrics separately for the features from each model, and then to combine these metrics for final retrieval ranking ,, which is a kind of ``\\emph{late fusion}'' strategy.\n\\textbf{Discussion.} Layer-level fusion and model-level fusion are conditioned on the fact that the associated layers or networks have different feature description capacities. For these fusion strategies, the key question is \\textit{what features are the best to be combined?} Some explorations have been made on the basis of off-the-shelf models, such as Xuan \\etal , who illustrates the effect of combining different numbers of features and different sizes within the ensemble. Chen \\etal analyze the performance of embedded features from off-the-shelf image classification and object detection models with respect to image retrieval.","id":"f5333f99-42b6-4668-a652-9777bc115d47","level":"subsubsection","origin_cites_number":18,"parent_id":"f1ce1975-10d7-4d60-8e5d-320750dbf247","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection"," Deep Feature Extraction"],["subsubsection","Feature Fusion Strategies"]],"subsections":[],"title":"Feature Fusion Strategies"},{"cite_extract_rate":0,"cites":[],"content":"\\label{Deep_Feature_Enhancement}\n\\textcolor{black}{The primary aim of feature embedding and aggregation is to further promote feature discriminativity, targeting for the ``\\textit{discriminativity challenge}'', and obtain final global and/or local features for retrieving specific instances.}","id":"f6eab6de-c628-4a89-bfb0-43b4f87b1151","level":"subsection","origin_cites_number":0,"parent_id":"1b3ca748-df26-4b7f-85f3-3eb438a055f5","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection","Feature Embedding and Aggregation"]],"subsections":["96cef0e9-d2bb-419e-8aa3-3757a002290b","22557b43-bfa7-48d0-9dad-fad3f1f25551","3362dbb1-5a27-4b56-85a4-413bc5c384f6","b5eed66c-a402-4570-be6b-8b9987293fc6"],"title":"Feature Embedding and Aggregation"},{"cite_extract_rate":0.448275862068965,"cites":[7543,2122,2102,2113,2105,8525,2108,2109,2097,2123,2126,7548,7545],"content":"\\label{Matching_with_Global_Feature}\nGlobal features can be extracted from fully-connected layers, followed by dimensionality reduction and normalization ,. They are easy to implement and there is no further aggregation process. Gong \\etal extract fully-connected activations for local image patches at three scale levels and embed patch-level activations individually using VLAD. Thus, the final concatenated features significantly tackle the invariance challenge caused by image rotations.\nConvolutional features can also be aggregated into compact a global feature. Simple aggregation methods are sum/average or max pooling ,. \nSum/average pooling is less discriminative, because it takes into account all activated convolutional outputs, thereby weakening the effect of highly activated features . As a result, max pooling is particularly well suited for sparse features having a low probability of being active, however max pooling may be inferior to sum/average pooling when image features are whitened .\nFigure \\ref{SinglePass} illustrates sophisticated feature aggregation methods using channel-wise or spatial-wise weighting ,. For example, Babenko \\etal propose sum-pooling convolutional features (SPoC) to obtain compact descriptors weighted by $ \\alpha^{\\prime} $ with a Gaussian center prior. Similarly, \nit is possible to treat regions in feature maps as different sub-vectors ,,, thus combinations of $R$ sub-vectors are used to represent the input image, such as R-MAC . Since convolutional features may include repetitive patterns and each vector may correspond to identical regions, the resulting descriptors may be bursty, which makes the final aggregated global feature less distinguishable. As a solution, Pang \\etal leverage heat diffusion to weigh convolutional features at the aggregation stage, and reduce the undesirable influence of burstiness. \nConvolutional features have an interpretation as descriptors of local regions, thus many works leverage embedding methods, including BoW, VLAD, and FV, to encode regional feature vectors and then aggregate them into a global descriptor. Note that BoW and VLAD can be extended by using other metrics, such as a Hamming distance . Here we briefly describe the principle of Euclidean embeddings.\nBoW is a widely used feature embedding which leads to a sparse vector of occurrence. Let $ \\boldsymbol{a} = \\left\\{a_{1}, a_{2},...,a_{R} \\right\\} $ be a set of $R$ local features, each of dimensionality $d$. BoW requires a pre-defined codebook $ \\boldsymbol{c} = \\left\\{c_{1}, c_{2},...,c_{K} \\right\\} $ with $ K $ centroids, usually learned offline, to cluster these local descriptors, and maps each descriptor $a_{t}$ to the nearest centroid $c_{k}$. For each centroid, one can count and normalize the number of occurrences as\n\\begin{equation}\ng(c_{k}) = \\frac{1}{R} \\sum_{r=1}^{R}\\phi(a_{r}, c_{k})\n\\end{equation}\n\\begin{equation}\n\\phi(a_{r}, c_{k}) = \\left\\{ \\begin{array}{ll}\n1 & \\textrm{if $ c_{k} $ is the closest codeword for $a_{r}$ }\\\\\n0 & \\textrm{otherwise}\\\\\n\\end{array} \\right.\n\\label{eqn-phi}\n\\end{equation}\nThus BoW considers the number of descriptors belonging to each $c_{k}$ (\\emph{i.e.} 0-order feature statistics), so the BoW representation is the concatenation of all mapped vectors:\n\\begin{equation}\n\\label{BoW}\n G_{_{BoW}}(\\boldsymbol{a}) =\n \\left[\n \\begin{array}{ccc}\n g(c_{1}), \\cdots ,g(c_{k}) ,\\cdots, g(c_{K})\n \\end{array}\n \\right] \\rm ^{\\top}\n\\end{equation}\nBoW is simple to implement the encoding of local descriptors, such as convolutional feature maps , or fully-connected activations ,, or to encode regional descriptors ,. Mukherjee \\etal extract image patches based on information entropy and feed into a pre-trained VGG-16, then use BoW to embed and aggregate the patch-level descriptors from a fully-connected layer. Embedded BoW vectors are typically high-dimensional and sparse, so not well suited to large-scale datasets in terms of the mentioned efficiency challenge.\nVLAD stores the sum of residuals for each visual word. Similar to BoW, it generates $K$ visual word centroids, then each feature $a_{r}$ is assigned to its nearest visual centroid $ c_{k} $:\n\\begin{equation}\ng(c_{k}) = \\frac{1}{R} \\sum_{r=1}^{R}\\phi(a_{r} , c_{k})(a_{r} - c_{k})\n\\end{equation}\nThe VLAD representation is stacked by the residuals for all centroids, with dimension ($ d \\times K $), \\ie \n\\begin{equation}\nG_{_{VLAD}}(\\boldsymbol{a}) \\! = \\! \\left[\n \\begin{array}{ccc}\n \\cdots, g(c_{k}) \\rm ^{\\top},\\cdots \n \\end{array}\n \\right] \\rm ^{\\top} .\n\\end{equation}\nVLAD captures first-order feature statistics, \\ie ($a_{r} - c_{k}$). Similar to BoW, the performance of VLAD is affected by the number of clusters: more centroids produce larger vectors that are harder to index. For instance-level image retrieval, \nGong \\etal concatenate the activations of a fully-connected layer with VLAD applied to image-level and patch-level inputs . Ng \\etal replace BoW~ with VLAD~, and are the first to encode local features into VLAD representations. This idea inspired another milestone work where, for the first time, VLAD is plugged into the last convolutional layer, which allows end-to-end training via back-propagation.\nFV extends BoW by encoding the first and second order statistics. FV clusters the set of local descriptors by a Gaussian Mixture Model (GMM) with $K$ components to generate a dictionary $ \\boldsymbol{c} = \\left\\{ \\mu_{k}; \\Sigma_{k}; w_{k}\\right\\}_{k=1}^{K} $ made up of mean / covariance / weight triples , where the covariance may be simplified by keeping only its diagonal elements. For each local feature $ a_{r} $, a GMM is given by \n\\begin{equation}\n\\begin{aligned}\n\\gamma_{k}(a_{r}) = w_{k}\\times p_{k}(a_{r})/\\Big(\\sum_{k=1}^{K}w_{k}p_{k}(a_{r})\\Big) \\quad s.t. \\sum_{k=1}^K w_{k} = 1\n\\end{aligned}\n\\end{equation}\nwhere $ p_{k}(a_{r})=\\mathcal{N}(a_{r}, \\mu_{k}, \\sigma_{k}^2 )$. All local features are assigned into each component $k$ in the dictionary, which is computed as\n\\begin{equation}\n\\begin{aligned}\n &g_{w_{k}} = \\frac{1}{R \\sqrt{w_{k}}} \\sum_{r=1}^{R} \\Big(\\gamma_{k}(a_{r}) - w_{k}\\Big) \\\\\n& g_{u_{k}} = \\frac{\\gamma_{k}(a_{r})}{R \\sqrt{w_{k}}} \\sum_{r=1}^{R} \\left( \\frac{a_{r} - \\mu_{k}}{\\sigma_{k}} \\right),\\\\ \n& g_{\\sigma_{k}^2} = \\frac{\\gamma_{k}(a_{r})}{R \\sqrt{2w_{k}}} \\sum_{r=1}^{R}\\left[ {\\left( \\frac{a_{r} - \\mu_{k}}{\\sigma_{k}} \\right)}^{2} - 1 \\right]\n\\end{aligned}\n\\end{equation}\nThe FV representation is produced by concatenating vectors from the $K$ components: \n\\begin{equation}\n\\begin{aligned}\n\\!\\!\\! G_{_{FV}}(\\boldsymbol{a}) \\! = \\! \\left[\n \\begin{array}{ccc}\n g_{w_{1}}, \\cdots , g_{w_{K}}, g_{u_{1}}, \\cdots, g_{u_{K}}, g_{\\sigma_{1}^2}, \\cdots, g_{\\sigma_{K}^2} \n \\end{array}\n \\right] \\rm ^{\\top}\n\\end{aligned}\n\\end{equation}\nThe FV representation defines a kernel from a generative process and captures more statistics than BoW and VLAD. FV vectors do not increase the computational cost significantly but require more memory. Applying FV without memory controls may lead to suboptimal performance . \n\\textbf{Discussion.} Traditionally, pooling-based aggregation methods (\\eg in Figure \\ref{SinglePass}) are directly plugged into deep networks and then the whole model is used end-to-end. The three embedding methods (BoW, VLAD, FV) are initially trained with large pre-defined vocabularies ,. One needs to pay attention on their properties before choosing an embedding: BoW and VLAD are computed in the rigid Euclidean space where performance is closely related to the number of centroids, whereas FV can capture higher-order statistics and improves the effectiveness of feature embedding at the expense of a higher memory cost. Further, although vocabularies are usually built separately and pre-trained before encoding deep features, it is necessary to integrate the training of networks and the learning of vocabulary parameters into a unified framework so as to guarantee training and testing efficiency. For example, VLAD is integrated into deep networks where each spatial column feature is used to construct clusters via k-means . This idea led to NetVLAD , where deep networks are fine-tuned with the VLAD vectors. The FV method is also combined with deep networks for retrieval tasks ,.","id":"96cef0e9-d2bb-419e-8aa3-3757a002290b","level":"subsubsection","origin_cites_number":29,"parent_id":"f6eab6de-c628-4a89-bfb0-43b4f87b1151","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection","Feature Embedding and Aggregation"],["subsubsection","Matching with Global Features"]],"subsections":[],"title":"Matching with Global Features"},{"cite_extract_rate":0.6000000000000001,"cites":[7543,2093,2101,7546,8525,2127,2103,8524,2121,2106,2114,2126,2110,209,7545],"content":"\\label{Matching_with_Local_Feature}\nAlthough matching with global features has high efficiency for both feature extraction and similarity computation, global features are not compatible with spatial verification and correspondence estimation, which are important procedures for instance-level retrieval tasks, motivating work on matching with local features. In terms of the matching process, global features are matched only once while local feature matching is evaluated by summarizing the similarity across all individual local features (\\ie many-to-many matching).\nOne important aspect of local features is to detect the keypoints for an instance within an image, and then to describe the detected keypoints as a set of local descriptors. Inspired by , the common strategies of this whole procedure for IIR can be categorized as \\textit{detect-then-describe} and \\textit{describe-then-detect}.\nIn terms of \\textit{detect-then-describe}, we regard the descriptors around keypoints as local features, similar to ,. Coarse regions can be detected, for example, by using the methods depicted in Figure~\\ref{MultiplePass}, and regions of interest in an image can be detected by using region proposal networks (RPNs) ,. The extracted coarse regions around the keypoints are fed into a DCNN, followed by feature description. Traditional detectors can also be used to detect fine regions around a keypoint. For instance, Zheng \\etal employ the popular Hessian-Affine detector to get an affine-invariant local region. Paulin \\etal and Mishchuk \\etal detect regions using the Hessian-Affine detector and feed into patch-convolutional kernel networks (Patch-CKNs) . \\textcolor{black}{Note that it becomes more convenient for the case where bounding boxes annotations have been provided by datasets (see Section \\ref{Datasets_and_Evaluation_Criteria}), and then the image regions can be cropped directly for further reranking . }\nRather than performing keypoint detection early on, it is possible to postpone the detection stage on the convolutional feature maps, \\ie \\textit{describe-then-detect}. One can select regions on the convolutional feature maps to obtain a set of local features ,,; the local maxima of the feature maps are then detected as keypoints . A similar strategy is also used in network fine-tuning ,,,,, where the keypoints on the convolutional feature maps can be selected based on attention scores predicted by an attention network ,, or based on single-head and multi-head attention modules in transformers ,. This approach to keypoint selection is better for achieving computational efficiency. \nAfter keypoint detection and description, a large number of local features are used in the matching stage to perform instance-level retrieval, and the image similarity is evaluated by matching across all local features. Local matching techniques include spatial verification and selective match kernels (SMK) . Spatial verification assumes object instances are rigid so that local matches between images can be estimated as an affine transformation using RANdom SAmple Consensus (RANSAC) . One limitation of RANSAC is its high computational complexity of estimating the transformation model when all local descriptors are considered; instead, it is possible to apply RANSAC to a small number of top-ranked local descriptors, such as those selected by approximate nearest neighbor . \nSMK weighs the contributions of individual matches with a non-linear selective function, but is still memory intensive. Its extension, the Aggregated Selective Match Kernel (ASMK), focuses more on aggregating similarities between local features without explicitly modeling the geometric alignment, which can produce a more compact representation ,. Recently, Teichmann \\etal introduced Regional Aggregated Selective Match Kernel (R-ASMK) to combine information from detected regions, boosting image retrieval accuracy compared to the ASMK.\n\\textbf{Discussion.} Using local descriptors to perform instance retrieval tasks has two limitations. First, the local descriptors for an image are stored individually and independently, which is memory-intensive, and not well-suited for large-scale scenarios.\nSecond, estimating the similarity between the query and database images depends on cross-matching all local descriptor pairs, which incurs additional searching cost and then a low retrieval efficiency. Therefore, most instance retrieval systems using local features follow a two-stage paradigm: initial filtering and re-ranking ,,,,, as in Figure \\ref{PipelineofImageRetrieval}. The initial filtering stage is to employ a global descriptor to select a set of candidate matching images, thereby reducing the solution space; the re-ranking stage is to use local descriptors to re-rank the top-ranked images from the global descriptor.","id":"22557b43-bfa7-48d0-9dad-fad3f1f25551","level":"subsubsection","origin_cites_number":25,"parent_id":"f6eab6de-c628-4a89-bfb0-43b4f87b1151","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection","Feature Embedding and Aggregation"],["subsubsection","Matching with Local Features"]],"subsections":[],"title":"Matching with Local Features"},{"cite_extract_rate":0.5,"cites":[2124,7547,2094,2104,2093,2118,2099],"content":"\\label{Attention_Mechanism}\nAttention mechanism can be regarded as a kind of feature aggregation, whose aim is to \\textcolor{black}{highlight the most relevant feature parts. It can effectively address the ``\\textit{distraction challenge}'' and also promote feature discriminativity , realized by computing an attention map.} Approaches to obtaining attention maps can be categorized into non-parametric and parametric groups, as shown in Figure~\\ref{fig3}, where the main difference is whether the importance weights in the attention map are learnable.\nNon-parametric weighting is a straightforward method to highlight feature importance, and the corresponding attention maps can be obtained by channel-wise or spatial-wise pooling, as in Figure \\ref{fig3} (a,b). For spatial-wise pooling, \nKalantidis \\etal propose an effective CroW method to weight and pool feature maps, which concentrate on weighting activations at different spatial locations, without considering the relations between these activations. In contrast, Ng \\etal explore the correlations among activations at different spatial locations on the convolutional feature maps. \nChannel-wise weighting methods are also popular non-parametric attention mechanisms ,. Xu \\etal rank the weighted feature maps to build ``probabilistic proposals'' to select regional features. Jimenez \\etal combine CroW and R-MAC to propose Classes Activation Maps (CAM) to weigh the feature map per class. Xiang \\etal employ a Gram matrix to analyze the correlations between different channels and then obtain channel sensitivity information to tune the importance of each feature map. Channel-wise and spatial-wise weighting methods are usually integrated into a deep model to highlight feature importance ,.\nParametric attention maps, shown in Figure \\ref{fig3} (c,d), can be learned via deep networks, where the input can be either image patches or feature maps ,,, approaches which are commonly used in supervised metric learning . Kim \\etal make the first attempt to propose a shallow network (CRN) to take as input the feature maps of convolutional layers and outputs a weighted mask indicating the importance of spatial regions in the feature maps. The resulting mask modulates feature aggregation to create a global representation of the input image. \nNoh \\etal design a 2-layer CNN with a softplus output layer to compute scores which indicate the importance of different image regions. Inspired by R-MAC, Kim \\etal employ a pre-trained ResNet101 to train a context-aware attention network using multi-scale feature maps.\nApart from using feature maps as inputs, a whole image can be used to learn feature importance, for which specific networks are needed ,,. Mohedano explores different saliency models, including DeepFixNet and Saliency Attentive Model. Yang \\etal and Wei \\etal introduce a two-stream network for image retrieval in which the auxiliary stream, DeepFixNet, is used specifically for predicting attention maps, which are then fused with the feature maps produced by the main network.\nFor image retrieval, attention mechanisms can be combined with supervised metric learning .","id":"3362dbb1-5a27-4b56-85a4-413bc5c384f6","level":"subsubsection","origin_cites_number":14,"parent_id":"f6eab6de-c628-4a89-bfb0-43b4f87b1151","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection","Feature Embedding and Aggregation"],["subsubsection","Attention Mechanism"]],"subsections":[],"title":"Attention Mechanism"},{"cite_extract_rate":0.45,"cites":[7547,2118,2094,2104,2099,2116,2128,2095,2110],"content":"\\label{Deep_Hash_Embedding}\nReal-valued features extracted by deep networks are typically high-dimensional, and therefore are not well-suited to retrieval efficiency. As a result, there is significant motivation to transform deep features into more compact codes. \\textcolor{black}{Since their computational and storage efficiency are beneficial for the ``\\textit{efficiency challenge}'', hashing algorithms have been widely used for global , and local descriptors ,,.}\n\\begin{figure}[!t]\n\\centering\n\\includegraphics[width=0.5\\textwidth]{./Figures/AttentionMachanism.pdf}\n\\vspace{-1.5em}\n\\caption{Illustration of attention mechanisms. (a)-(b) Non-parametric schemes: The attention is based on convolutional feature $ \\boldsymbol{A} $. Channel-wise attention in (a) produces a $C$-dimensional importance vector \\textbf{$\\beta_{1}$} ,; Spatial-wise attention in (b) computes a 2-dimensional attention map \\textbf{$\\alpha$} ,,. (c)-(d) Parametric schemes: The attention weights are learned by a trainable network. In (c), \\textbf{$\\beta_2$} are provided by a sub-network with parameters $\\theta_{\\gamma}$ ,,,. In (d), the attention maps, as a tensor, are predicted by some auxiliary saliency extraction models from the input image directly ,,.} \\label{fig3}\n\\end{figure}\nHash functions can be plugged as a layer into deep networks, so that hash codes and deep networks can be simultaneously trained and optimized, either supervised or unsupervised . During hash function training, the hash codes of originally similar images are embedded as closely as possible, and the hash codes of dissimilar images are as separated as possible. $d$-dim hash codes from a hash function $h(\\cdot)$ for an image $x$ can be formulated as $ b_{x} = h(x) = h\\big(f(x;\\bm{\\theta})\\big) \\in \\{+1, -1\\}^d $. Because hash codes are non-differentiable their optimization is difficult, so $h(\\cdot)$ can be relaxed to be differentiable by using \\textit{tanh} or \\textit{sigmoid} functions .\nWhen binarizing real-valued features, it is crucial to preserve image similarity and to improve hash code quality . These two aspects are at the heart of hashing algorithms to maximize retrieval accuracy. \n\\begin{flushleft}\n\\emph{a. Hash Functions to Preserve Image Similarity}\n\\end{flushleft}\nPreserving similarity seeks to minimize the inconsistencies between real-valued features and corresponding hash codes, for which a variety of strategies have been adopted. \nLoss functions can significantly influence similarity preservation, which includes both supervised and unsupervised methods. With class labels available, many loss functions are designed to learn hash codes in a Hamming space. As a straightforward method, one can optimize either the difference between matrices computed from the binary codes and their supervision labels , or the difference between the hash codes and real-valued deep features ,. Song \\etal propose to learn hash codes for regional features in which each local feature is converted to a set of binary codes by multiplying a hash function and the raw RoI features, then the differences between RoI features and hash codes are characterized by an L$_2$ loss. Do \\etal regularize hash codes with a reconstruction loss, which ensure that codes can be reconstructed to their inputs so that similar/dissimilar inputs are mapped to similar/dissimilar hash codes. Lin \\etal learn hash codes and address the ``\\textit{invariance challenge}'' by introducing an objective function which characterize the difference between the binary codes which are computed from the original image and the geometric transformed one.\n\\begin{flushleft}\n\\emph{b. Improving Hash Function Quality}\n\\end{flushleft}\nA good hash function seeks to have binary codes uniformly distributed; that is, maximally filling and using the hash code space, normally on the basis of bit uncorrelation and bit balance ,. Bit uncorrelation implies that different bits are as independent as possible, so that a given set of bits can aggregate more information within a given code length . Bit balance means that each bit should have a 50\\% chance of being +1 or -1, thereby maximizing code variance and information . Mor{\\`e}re \\etal use the uniform distribution $U$(0,1) to build a regularization term to make hash codes distribute evenly where the codes are learned by a Restricted Boltzmann Machine layer. Likewise, Lin \\etal optimize the mean of learned hash codes to be close to 0.5 to prevent any bit bias towards zero or one.","id":"b5eed66c-a402-4570-be6b-8b9987293fc6","level":"subsubsection","origin_cites_number":20,"parent_id":"f6eab6de-c628-4a89-bfb0-43b4f87b1151","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Retrieval with Off-the-Shelf DCNN Models"],["subsection","Feature Embedding and Aggregation"],["subsubsection"," Hashing Embedding "]],"subsections":[],"title":" Hashing Embedding "},{"cite_extract_rate":0.6923076923076921,"cites":[2117,7546,2110,2103,2129,2096,629,2109,2097],"content":"\\label{Retrieval_via_Learning_DCNN_Representations}\nThe off-the-shelf DCNNs pre-trained on source datasets for classification are quite robust to inter-class variability. However, in most cases, deep features extracted based on off-the-shelf models may not be sufficient for accurate retrieval, even with the strategies discussed in Section \\ref{Retrieval_with_Off_the_Shelf_DCNN_Models}. In order for models to be more effective for retrieval, a common practice is network fine-tuning, \\ie updating the pre-stored parameters . Fine-tuning methods have been studied extensively to learn better features, \\textcolor{black}{whose primary aim is to address the ``\\textit{fine-tune challenge}''.} A standard dataset with clear and well-defined ground-truth labels is indispensable for the supervised fine-tuning and subsequently pair-wise supervisory information is incorporated into ranking loss to update networks by regularizing on retrieval representations, otherwise it is necessary to develop unsupervised fine-tuned methods. After network fine-tuning, features can be organized as global or local to perform retrieval. \n\\textcolor{black}{For the most feature strategies we presented in Section \\ref{Retrieval_with_Off_the_Shelf_DCNN_Models}, including feature extraction, feature embedding and feature aggregation. Note that fine-tuning does not contradict or render irrelevant these feature processing methods; indeed, these strategies are complementary and can be equivalently incorporated as part of network fine-tuning. To this end, this section will survey the strategies which have been developed, based on the patch-level, image-level, or class-level supervision, to fine-tune deep networks for better instance retrieval.}\n\\begin{figure*}[!t]\n\\centering\n\\includegraphics[width=0.9\\textwidth]{./Figures/All_Metric_Learning.pdf}\n\\vspace{-0.5em}\n\\caption{Schemes of supervised fine-tuning. Anchor, positive, and negative images are indicated by $x_{a}$, $x_{p}$, $x_{n}$, respectively. (a) classification loss ; (b) similarity learning by using a transformation matrix ; (c) Siamese loss ,,,; (d) triplet loss ; (e) an attention block into DCNNs to highlight regions ; (f) combining classification loss and pairwise ranking loss ,; (g) region proposal networks (RPNs) to locate the RoI and highlight specific regions or instances ; (h) inserting the RPNs of (g) into DCNNs, such that the RPNs extract regions or instances at the convolutional layer ,. }\n\\vspace{-1em}\n\\label{All_Metric_Learning}\n\\end{figure*}","id":"acffc1fe-bf03-450b-857b-db143905236e","level":"section","origin_cites_number":13,"parent_id":"515509a6-319a-4385-b1b5-4418a83d2bc2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "]],"subsections":["45dba903-566e-4f9f-a48c-8f8adf0593ca","871a2087-f4a4-4968-91d1-0752906a7612"],"title":"$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "},{"cite_extract_rate":0,"cites":[],"content":"\\label{Supervised_Fine-tuning}\nThe way to realize supervised fine-tuning can be determined by the given class labels or pairwise supervisory information.","id":"45dba903-566e-4f9f-a48c-8f8adf0593ca","level":"subsection","origin_cites_number":0,"parent_id":"acffc1fe-bf03-450b-857b-db143905236e","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "],["subsection","Supervised Fine-tuning "]],"subsections":["c05b41ca-005d-493d-96fa-98315460aeed","db1c70e4-39b7-4baf-8fa7-4ccb702563d0","14afec8b-a454-4ba0-a24a-2c2e37482b34"],"title":"Supervised Fine-tuning "},{"cite_extract_rate":0.75,"cites":[2114,309,2130,2093,2111,2109],"content":"\\label{Classification-based_Fine-tuning}\nWhen class labels of a new dataset are available (\\eg INSTRE , GLDv2 ,), it is preferable to begin with a previously-trained DCNN, trained on a separate dataset, with the backbone DCNN typically chosen from one of AlexNet, VGG, GoogLeNet, or ResNet.\nThe DCNN can then be fine-tuned, as shown in Figure \\ref{All_Metric_Learning} (a), by optimizing its parameters on the basis of a cross entropy loss \n\\begin{equation}\n\\mathcal{L}_{CE}(\\hat{p_{i}},y_{i})= - \\! \\sum^{c}_{i} \\! \\big(y_{i} \\! \\times \\!log(\\hat{p}_{i}) \\big)\n\\label{cross_entropy_loss}\n\\end{equation}\nHere $y_{i}$ and $\\hat{p}_{i}$ are the ground-truth labels and the predicted logits, respectively, and $c$ is the total number of categories. The milestone work in such fine-tuning is , in which AlexNet is re-trained on the Landmarks dataset. \\textcolor{black}{\nAccording to the class labels, the image-level features are required to compute the logits. Thus, the descriptors extracted from local regions on convolutional feature maps , or image patch inputs are further needed to be aggregated. }\n\\textcolor{black}{A classification-based fine-tuning method enables to enforce higher similarity for intra-class samples and diversity for inter-class samples. Cao \\etal employ the ArcFace loss , which uses the margin-adjusted cosine similarity in the form of softmax loss, to induce smaller intra-class variance and show excellent results for instance retrieval.} Recently, Boudiaf \\etal claim that cross entropy loss can minimize intra-class distances while maximizing inter-class distances. Cross entropy loss is, in essence, maximizing a common mutual information between the retrieval features and the ground-truth labels. Therefore, it can be regarded as an upper bound on a new pairwise loss, which has a structure similar to various pairwise ranking losses, of which representatives are introduced below.","id":"c05b41ca-005d-493d-96fa-98315460aeed","level":"subsubsection","origin_cites_number":8,"parent_id":"45dba903-566e-4f9f-a48c-8f8adf0593ca","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "],["subsection","Supervised Fine-tuning "],["subsubsection","Fine-tuning via Classification Loss"]],"subsections":[],"title":"Fine-tuning via Classification Loss"},{"cite_extract_rate":0.631578947368421,"cites":[2101,2096,2098,7546,7549,2103,2097,2106,2129,2110,209,2117],"content":"\\label{Verification_based_Learning}\nWith affinity information (\\eg samples from the same group) indicating similar and dissimilar pairs, fine-tuning methods based on pairwise ranking loss learn an optimal metric which minimizes or maximizes the distance of pairs to maintain their similarity. Network fine-tuning via ranking loss involves two types of information :\n\\begin{enumerate}\n\\item A pair-wise constraint, corresponding to a Siamese network as in Figure \\ref{All_Metric_Learning} (c), in which input images are paired with either a positive or negative sample;\n\\item A triplet constraint, associated with triplet networks as in Figure \\ref{All_Metric_Learning} (e), in which anchor images are paired with both similar and dissimilar samples .\n\\end{enumerate}\nThese pairwise ranking loss based methods are categorized into globally supervised approaches (Figure \\ref{All_Metric_Learning} (c,d)) and locally supervised approaches (Figure \\ref{All_Metric_Learning} (g,h)), where the former ones learn a metric on global features by satisfying all constraints, whereas the latter ones focus on local areas by only satisfying the given local constraints (\\emph{e.g.} region proposals).\nTo be specific, consider a triplet set $ X \\!\\! = \\!\\! \\lbrace(x_{a}, x_{p}, x_{n})\\} $ in a mini-batch, where $ (x_{a}, x_{p} )$ indicates a similar pair and $ (x_{a}, x_{n} )$ a dissimilar pair. Features $ f(x;\\bm{\\theta}) $ of one image are extracted by a network $f(\\cdot)$ with parameters $\\bm{\\theta}$, for which we can represent the affinity information for each similar or dissimilar pair as\n\\begin{equation}\n\\begin{aligned}\n\\mathcal{D}_{ij} = \\mathcal{D}(x_{i}, x_{j}) = || f(x_{i}; \\bm{\\theta}) - f(x_{j}; \\bm{\\theta})||_{2}^{2}\n\\end{aligned}\\label{Distance_pre_defined}\n\\end{equation}\n\\begin{flushleft}\n\\emph{a. Refining with Transformation Matrix}. \n\\end{flushleft}\nLearning the similarity among input samples can be implemented by optimizing the weights of a linear transformation matrix . It transforms the concatenated feature pairs into a common latent space using a transformation matrix $ \\bm{W} \\!\\! \\in \\!\\! \\mathbb{R}^{2d \\times 1} $, where $d$ is the final feature dimension. The similarity score of these pairs are predicted via a sub-network $ \\mathcal{S}_{W}(x_{i}, x_{j}) = f_{W}(f(x_{i}; \\bm{\\theta})\\cup f(x_{j};\\bm{\\theta}); \\bm{W}) $ . In other words, the sub-network $f_{W}$ predicts how similar the feature pairs are. Given the affinity information of feature pairs $ \\mathcal{S}_{ij}=\\mathcal{S}(x_{i}, x_{j}) \\! \\in \\! \\{0,1\\} $, the binary labels 0 and 1 indicate the similar (positive) or dissimilar (negative) pairs, respectively. The training of function $f_{W}$ can be achieved by using a regression loss: \n\\begin{equation}\n\\begin{aligned}\n\\!\\!\\!\\mathcal{L}_{W}(x_{i}, x_{j}) = & | \\mathcal{S}_{W}(x_{i}, x_{j}) - \\mathcal{S}_{ij}\\big(\\text{sim}(x_{i}, x_{j}) + m\\big) - \\\\\n&(1-\\mathcal{S}_{ij})\\big(\\text{sim}(x_{i}, x_{j})-m\\big) | \\label{regressionLoss} \n\\end{aligned}\n\\end{equation}\nwhere $\\text{sim}(x_{i}, x_{j}) $ can be the cosine function for guiding the training of $\\bm{W}$ and $m$ is a margin. By optimizing the regression loss and updating $ \\bm{W} $, deep networks maximize the similarity of similar pairs and minimize that of dissimilar pairs. It is worth noting that the pre-stored parameters in the deep models are frozen when optimizing $ \\bm{W} $. The pipeline of this approach is depicted in Figure \\ref{All_Metric_Learning} (b). \n\\begin{flushleft}\n\\emph{b. Fine-tuning with Siamese Networks}. \n\\end{flushleft}\nSiamese networks represent important options in implementing metric learning for fine-tuning, as in Figure \\ref{All_Metric_Learning} (c) and Figure \\ref{mAP_loss_illustration} (a). It is a structure composed of two branches that share the same weights across layers. Siamese networks are trained on paired data, consisting of an image pair $ (x_{i}, x_{j}) $ such that $ \\mathcal{S}(x_{i}, x_{j}) \\! \\in \\! \\{0,1\\} $. A Siamese loss \nis formulated as\n\\begin{equation}\n\\begin{aligned}\n\\mathcal{L}_{Siam}(x_{i}, x_{j})& = \\frac{1}{2}\\mathcal{S}(x_{i}, x_{j})\\mathcal{D}(x_{i}, x_{j}) \\; + \\\\\n&\\frac{1}{2}\\big(1 - \\mathcal{S}(x_{i}, x_{j})\\big)\\max\\big(0,\\; m - \\mathcal{D}(x_{i}, x_{j})\\big)\n\\end{aligned}\\label{contrastive}\n\\end{equation}\nSiamese loss has recently been reaffirmed as a very effective metric in category-level image retrieval, outperforming many more sophisticated losses if implemented carefully . Enabled by the standard Siamese network, this objective function is used to learn the similarity between semantically relevant samples under different scenarios ,. For example, Radenovi{\\'c} \\etal employ a Siamese network on matching and non-matching global feature pairs which are aggregated by GeM-based pooling. The deep network fine-tuned by the Siamese loss generalizes better and converges at higher retrieval performance. Ong \\etal leverage the Siamese network to learn image features which are then fed into the Fisher Vector model for further encoding. Siamese networks can also be applied to hashing learning in which the Euclidean distance $\\mathcal{D}(\\cdot)$ in Eq. \\ref{contrastive} is computed for binary codes .\nAn implicit drawback of the Siamese loss is that it may penalize similar image pairs even if the margin between these pairs is small or zero , if the constraint is too strong and unbalanced. At the same time, it is hard to map the features of similar pairs to the same point when images contain complex contents or scenes. To tackle this limitation, Cao \\etal adopt a double-margin Siamese loss to relax the penalty for similar pairs by setting a margin $ m_1 $ instead of zero, in which case the original single-margin Siamese loss is re-formulated as \n\\begin{equation}\n\\begin{aligned}\n\\!\\!\\!\\! \\mathcal{L}_{\\mathcal{D}\\_Siam}(x_{i}, x_{j}) & = \\frac{1}{2}\\mathcal{S}(x_{i}, x_{j})\\max \\big(0, \\mathcal{D}(x_{i}, x_{j}) - m_{1} \\big) + \\\\\n& \\!\\! \\frac{1}{2}\\big(1 - \\mathcal{S}(x_{i}, x_{j})\\big)\\max\\big(0, m_{2} - \\mathcal{D}(x_{i}, x_{j})\\big)\n\\end{aligned}\\label{doublemargincontrastive}\n\\end{equation}\nwhere $ m_{1} \\!\\! > \\!\\! 0 $ and $ m_{2} \\!\\! > \\!\\! 0 $ are the margins affecting the similar and dissimilar pairs, respectively, as in Figure \\ref{mAP_loss_illustration} (b), meaning that the double margin Siamese loss only applies a contrastive force when the distance of a similar pair is larger than $ m_{1} $. The mAP metric of retrieval is improved when using the double margin Siamese loss . \nMore recently, transformers have been trained under the regularization of cross entropy and Siamese loss for instance-level retrieval and achieved competitive performance, positioning it as an alternative to convolutional architectures. As observed by , the transformer-based architecture is less impacted than convolutional networks by feature collapse since each input feature is projected to different sub-spaces before the multi-headed attention. Moreover, the transformer backbone operates as a learned aggregation operator, thereby avoiding the design of sophisticated feature aggregation methods.\n\\begin{figure*}[!t]\n\\centering\n\\includegraphics[width=\\linewidth]{./Figures/mAP_loss_illustration.pdf}\n\\vspace{-2em}\n\\caption{Illustrations of different losses for network fine-tuning. The same shape with different colors denotes images that include the same instance. (a)-(c) have been introduced in the text ,,. (d) Listwise AP loss considers a mini-batch of $N$ features simultaneously and directly optimizes the Average-Precision computed from these features ,.\n} \\label{mAP_loss_illustration}\n\\end{figure*}\n\\begin{flushleft}\n\\emph{c. Fine-tuning with Triplet Networks}. \n\\end{flushleft}\nTriplet networks optimize similar and dissimilar pairs simultaneously. As shown in Figure \\ref{All_Metric_Learning} (d) and Figure \\ref{mAP_loss_illustration} (c), the plain triplet networks adopt a ranking loss for training:\n\\begin{equation}\n\\begin{aligned}\n\\mathcal{L}_{Triplet}(x_{a}, x_{p}, x_{n}) = \\max \\big( 0, m + \\mathcal{D}(x_{a}, x_{p}) - \n \\mathcal{D}(x_{a}, x_{n})\\big)\\label{triplet}\n\\end{aligned}\n\\end{equation}\nwhich indicates that the distance of an anchor-negative pair $ \\mathcal{D}(x_{a}, x_{n}) $ should be larger than that of an anchor-positive pair $ \\mathcal{D}(x_{a}, x_{p}) $ by a certain margin $ m $. \n\\textcolor{black}{Given the datasets that provide bounding box annotations, such as INSTRE, Oxford-5k, Paris-6k, and their variants, the bounding box annotations are used as patch-level supervision to train a region detector which enables the final DCNNs to locate specific regions or objects. As an example, region proposal networks (RPNs) is fine-tuned and subsequently plugged into DCNNs and trained end-to-end , as shown in Figure \\ref{All_Metric_Learning} (g). RPNs yield the regressed bounding box coordinates of objects and are trained by the multi-class classification loss. Once fine-tuned, RPNs can produce regional features for each detected region by RoI pooling and perform better instance search. }\n\\textcolor{black}{Further, local supervised metric learning has been explored based on the fact that RPNs enable deep models to learn regional features for particular instance objects ,,,. RPNs used in the triplet formulation are shown in Figure \\ref{All_Metric_Learning} (h). Firstly, regression loss (RPNs loss) is used to minimize the regressed bounding box relative to ground-truth. Then, the regional features for all detected RoIs are aggregated into a global one and L2-normalized for the triplet loss.} Note that, in some cases, jointly training an RPN loss and triplet loss leads to unstable results, a problem addressed in by first training a CNN to produce R-MAC using a rigid grid, after which the parameters in convolutional layers are fixed and RPNs are trained to replace the rigid grid.\nAttention mechanisms can also be combined with metric learning for fine-tuning , as in Figure \\ref{All_Metric_Learning} (e), where the attention module is typically end-to-end trainable and takes as input the convolutional feature maps. Song \\etal introduce a convolutional attention layer to explore spatial-semantic information, highlighting regions in images to significantly improve the discrimination for image retrieval.\nRecent studies , have jointly optimized the triplet loss and classification loss to further improve network capacity, as shown in Figure \\ref{All_Metric_Learning} (f). The overall joint function is \n\\begin{equation}\n\\begin{aligned}\n\\!\\!\\!\\mathcal{L}_{Joint} = &\\lambda_{1} \\! \\cdot \\! \\mathcal{L}_{Triplet}(x_{i,a}, x_{i,p}, x_{i,n}) \\! + \\! \\lambda_{2} \\!\\cdot \\! \\mathcal{L}_{CE}(\\hat{p_{i}},y_{i})\n\\end{aligned}\n\\end{equation}\nwhere the cross entropy loss (CE loss) $ \\mathcal{L}_{CE} $ is defined in Eq. (\\ref{cross_entropy_loss}) and the triplet loss $ \\mathcal{L}_{Triplet} $ in Eq. (\\ref{triplet}). $ \\lambda_{1} $ and $ \\lambda_{2} $ are hyper-parameters tuning the tradeoff between the two loss functions.","id":"db1c70e4-39b7-4baf-8fa7-4ccb702563d0","level":"subsubsection","origin_cites_number":19,"parent_id":"45dba903-566e-4f9f-a48c-8f8adf0593ca","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "],["subsection","Supervised Fine-tuning "],["subsubsection","Fine-tuning via Pairwise Ranking Loss"]],"subsections":[],"title":"Fine-tuning via Pairwise Ranking Loss"},{"cite_extract_rate":0.6666666666666661,"cites":[2126,2127,2130,2129,7549,2098],"content":"\\label{Discussion_ranking_loss}\nIn some cases, pairwise ranking loss cannot effectively learn the variations between samples and still suffers from a weaker generalization capability if the training set is not ordered correctly. Therefore, pairwise ranking loss requires careful sample mining and weighting strategies to obtain the most informative pairs, especially when considering mini-batches. The hard negative mining strategy is commonly used ,,, however further sophisticated mining strategies have recently been developed. Mishchuk \\etal calculate a pair-wise distance matrix on all mini-batch samples to select two closest negative and one anchor-positive pair to form a triplet. \nInstead of traversing all possible two-tuple or three-tuple combinations, it is possible to consider all positive samples in one cluster and negative samples together. Liu \\etal introduce a group-group loss to decrease the intra-group distance and increase the inter-group distance. Considering all samples may be beneficial for stabilizing optimization and promoting generalization due to a larger data diversity, however the extra computational cost remains an issue to be addressed.\nSubstantial research has been devoted to pair-wise ranking loss, while cross entropy loss, mainly used for classification, has been largely overlooked. Recently, Boudiaf \\etal claim that cross entropy loss can match and even surpass the pair-wise ranking loss when carefully tuned on fine-grained category-level retrieval tasks. In fact, the greatest improvements have come from enhanced training schemes (\\eg data augmentation, learning rate polices, batch normalization freeze) rather than intrinsic properties of pairwise ranking loss. Further, although several sophisticated ranking losses have been explored and validated for category-level retrieval, Musgrave \\etal revisited these losses and found that most of them perform on par to vanilla Siamese loss and triplet loss, so there is merit to consider these losses also for instance-level image retrieval tasks.\nBoth cross entropy loss and pair-wise ranking loss regularize on the embedded features and the corresponding labels so as to maximize their mutual information . Their effectiveness is not guaranteed to give retrieval results that also optimize mAP . To tackle this limitation one can directly optimize the average precision (AP) metric using the listwise AP loss,\n\\begin{equation}\n\\begin{aligned}\n\\!\\!\\!\\mathcal{L}_{mAP} = 1 - \\frac{1}{N} \\sum^{N}_{i=1} {\\rm AP}(x_{i}^{\\top}X_{N}, Y_{i})\n\\label{AP_loss}\n\\end{aligned}\n\\end{equation}\nwhich optimizes the global ranking of thousands of images simultaneously, instead of only a few images at a time. Here $Y_{i}$ is the binary label to evaluate the relevance between batch images. $X_{N} = \\{x_{1}, x_{2},...x_{j},...,x_{N}\\}$ denotes the features of all images, where each $x_{i}$ is used as a potential query to rank the remaining batch images. Each similarity score $x_{i}^{\\top}x_{j}$ can be measured by a cosine function. \nIt is demonstrated that training with AP-based loss improves retrieval performance ,. However average precision, as a metric, is normally non-differentiable. To directly optimize the AP loss during back-propagation, the key is that the indicator function for AP computing needs to be relaxed using methods such as triangular kernel-based soft assignment or sigmoid function , as shown in Figure \\ref{mAP_loss_illustration} (d).","id":"14afec8b-a454-4ba0-a24a-2c2e37482b34","level":"subsubsection","origin_cites_number":9,"parent_id":"45dba903-566e-4f9f-a48c-8f8adf0593ca","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "],["subsection","Supervised Fine-tuning "],["subsubsection","Discussion"]],"subsections":[],"title":"Discussion"},{"cite_extract_rate":1,"cites":[2131],"content":"\\label{Unsupervised_Fine-tuning}\nSupervised network fine-tuning becomes infeasible when there is insufficient supervisory information, normally because of cost or unavailability. Therefore unsupervised fine-tuning methods for image retrieval are quite necessary, but less studied . \nFor unsupervised fine-tuning, two directions are to mine relevance among features via manifold learning, and via clustering techniques, each discussed below.","id":"871a2087-f4a4-4968-91d1-0752906a7612","level":"subsection","origin_cites_number":1,"parent_id":"acffc1fe-bf03-450b-857b-db143905236e","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "],["subsection","Unsupervised Fine-tuning "]],"subsections":["f6c0b11c-f5b0-47c6-9e48-358712ee7aee","7ab93d94-79dd-47b1-b0d5-6e73b13d8f40"],"title":"Unsupervised Fine-tuning "},{"cite_extract_rate":0.33333333333333304,"cites":[2131,2103,2132,2100],"content":"\\label{Mining_Samples_with_Manifold_Learning}\nManifold learning focuses on capturing intrinsic correlations on a manifold structure to mine or deduce relevance, as illustrated in Figure \\ref{manifold_learning_pipeline}. Initial similarities between the extracted global features or local features , are used to construct an affinity matrix, which is then re-evaluated and updated using manifold learning . According to the manifold similarity in the updated affinity matrix, positive and hard negative samples are selected for metric learning using pairwise ranking loss based functions such as pair loss , or triplet loss ,. Note that this is different from the aforementioned methods for pairwise ranking loss based fine-tuning methods, where the hard positive and negative samples are explicitly selected from an ordered dataset according to the given affinity information.\nIt is important to capture the geometry of the manifold of deep features, generally involving two steps , known as diffusion. First, the affinity matrix (Figure \\ref{manifold_learning_pipeline}) is interpreted as a weighted kNN graph, where each vector is represented by a node, and edges are defined by the pairwise affinities of two connected nodes. Then, the pairwise affinities are re-evaluated in the context of all other elements by diffusing the similarity values through the graph ,,,, with recent strategies proposed such as regularized diffusion (RDP) and regional diffusion . For more details on diffusion methods refer to survey .\nMost algorithms follow the two steps of ; the differences among methods lie primarily in three aspects: \n\\begin{enumerate}\n\\item {\\bf Similarity initialization},\nwhich affects the subsequent kNN graph construction in an affinity matrix. Usually, an inner product or Euclidean distance is directly computed for the affinities. A Gaussian kernel function can be used ,, or consider regional similarity from image patches .\n\\item {\\bf Transition matrix definition}, a row-stochastic matrix , determines the probabilities of transiting from one node to another in the graph. These probabilities are proportional to the affinities between nodes, which can be measured by Geodesic distance (\\emph{e.g.} the summation of weights of relevant edges).\n\\item {\\bf Iteration scheme},\nto re-valuate and update the values in the affinity matrix by the manifold similarity until some convergence is achieved. Most algorithms are iteration-based ,, as illustrated in Figure \\ref{manifold_learning_pipeline}. \n\\end{enumerate}\nDiffusion process algorithms are indispensable for unsupervised fine-tuning. Better image similarity is guaranteed when it is improved based on initialization (\\emph{e.g.} regional similarity or higher order information ). Diffusion is normally iterative and is computationally demanding , a limitation which cannot meet the efficiency requirements of image retrieval. To reduce the computational complexity, Bai \\etal propose a regularized diffusion process, facilitated by an efficient iteration-based solver. Zhao \\etal regard the diffusion process as a non-linear kernel mapping function, which is then modelled by a deep neural network. Other studies replace the diffusion process on a kNN graph with a diffusion network , which is derived from graph convolution networks , an end-to-end trainable framework which allows efficient computation during training and testing.\n\\begin{figure}[!t]\n\\centering\n\\includegraphics[width=0.5\\textwidth]{./Figures/manifoldlearningpipeline.pdf}\n\\caption{Paradigm of manifold learning for unsupervised metric learning, based on triplet loss ,.} \n\\label{manifold_learning_pipeline}\n\\end{figure}\nOnce the manifold space is learned, samples are mined by computing geodesic distances based on the Floyd-Warshall algorithm or by comparing the set difference . The selected samples are fed into deep networks to perform fine-tuning.","id":"f6c0b11c-f5b0-47c6-9e48-358712ee7aee","level":"subsubsection","origin_cites_number":12,"parent_id":"871a2087-f4a4-4968-91d1-0752906a7612","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "],["subsection","Unsupervised Fine-tuning "],["subsubsection","Mining Samples with Manifold Learning"]],"subsections":[],"title":"Mining Samples with Manifold Learning"},{"cite_extract_rate":0.4,"cites":[9094,2135,2133,2128,7544,2134],"content":"\\label{Mining_Samples_by_Clustering}\nClustering is used to explore proximity information that has been studied in instance-level retrieval ,,,,. The rationale behind these methods is that samples in a cluster are likely to satisfy a degree of similarity. \nOne class of methods for clustering deep features is via k-means. Given $k$ cluster centroids, during each training epoch a deep network alternates between two steps: first, a soft assignment between the feature representations and the cluster centroids; second, the cluster centroids are refined and, at the same time, the deep network is updated by learning from current high confidence assignments using a certain regularization. These two steps are repeated until a convergence criterion is met, at which point the cluster assignments are used as pseudo-labels ,. Alternatively, the pseudo-labels can be calculated from the samples in a cluster, \\eg the mean values. For example, Tzelepi \\etal compute $ k $ nearest feature representations with respect to a query feature and then compute their mean vectors, which is used as a target for the query feature. In this case, fine-tuning is performed by minimizing the squared distance between each query feature and the mean of its $ k $ nearest features. Liu \\etal propose a self-taught hashing algorithm using a kNN graph construction to generate pseudo labels that are used to analyze and guide network training. Shen \\etal and Radenovi{\\'c} \\etal , use Structure-from-Motion (SfM) for each image cluster to explore sample reconstructions to select images for triplet loss. Clustering methods depend on the Euclidean distance, making it difficult to reveal the intrinsic relationship between objects.\nThere are further techniques for instance retrieval, such as by using AutoEncoder ,, generative adversarial networks (GANs) , convolutional kernel networks ,, and graph convolutional networks . For these methods, they focus on devising novel unsupervised frameworks to realize unsupervised learning, instead of iterative similarity diffusion or cluster refinement on feature space. For example, instead of performing iterative traversal on a set of nearest neighbors defined by kNN graph, Liu \\etal employ graph convolutional networks to directly encode the neighbor information into image descriptors and then train the deep models to learn a new feature space. This method is demonstrated to significantly improve retrieval accuracy while maintaining efficiency. GANs are also explored, for the first time, for instance-level retrieval in an unsupervised fashion . The generator retrieves images that contain similar instances as a given image, while the discriminator judges whether the retrieved images have the specified instance which appeared in the query image. During training, the discriminator and the generator play a min-max game via an adversarial reward which is computed based on the cosine distance between the query image and the images retrieved by the generator.","id":"7ab93d94-79dd-47b1-b0d5-6e73b13d8f40","level":"subsubsection","origin_cites_number":15,"parent_id":"871a2087-f4a4-4968-91d1-0752906a7612","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","$\\!\\!\\!\\!\\!$Retrieval via Learning DCNN Representations "],["subsection","Unsupervised Fine-tuning "],["subsubsection","Mining Samples by Clustering"]],"subsections":[],"title":"Mining Samples by Clustering"},{"cite_extract_rate":0,"cites":[],"content":"\\label{State_of_the_Art_Performance}","id":"a6c69e00-d8a1-492b-8514-fab570f109d2","level":"section","origin_cites_number":0,"parent_id":"515509a6-319a-4385-b1b5-4418a83d2bc2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","State of the Art Performance"]],"subsections":["347ae713-adb4-418f-9e62-4191f141d96c","78dcf4cf-a971-4517-838a-4552eb9d65a5","4289168f-a812-4899-b8a0-52669cd0ba36"],"title":"State of the Art Performance"},{"cite_extract_rate":0.384615384615384,"cites":[2114,2136,2137,2093,2111],"content":"\\label{Datasets_and_Evaluation_Criteria}\nTo demonstrate the effectiveness of methods, we choose the following commonly-used datasets for performance comparison:\n\\textbf{UKBench (UKB)} consists of 10,200 images of objects. This dataset has 2,550 groups of images, each group having four images of the same object from different viewpoints or illumination conditions, which can be regarded as a kind of class-level supervision information. All images can be used as a query.\n\\textbf{Holidays} consists of 1,491 images collected from personal holiday albums. Most images are scene-related. The dataset comprises 500 groups of similar images with a single query image for each group. The dataset also provides position information of the interest regions for each image.\n\\textbf{Oxford-5k} consists of 5,062 images for 11 Oxford buildings. Each building is associated with five hand-drawn bounding box queries. \\textcolor{black}{According to the relevance level, each image of the same building is assigned a label \\textit{Good} (\\ie \\textit{positive}), \\textit{OK} (\\ie \\textit{positive}), \\textit{Junk}, or \\textit{Bad} (\\ie \\textit{negative}). \\textit{Junk} images can be \ndiscarded or regarded as \\textit{negative} examples ,.\nTo build a tuple for each given query, one can select a positive example whose label corresponds to \\textit{Good} or \\textit{OK} in the same category, and select one negative example from each of the remaining building categories. Furthermore, an additional disjoint set of 100,000 distractor images is added to obtain Oxford-105k.}\n\\textbf{Paris-6k} includes 6,412 images and is categorized into 12 groups by architecture. \nThe supervision information can be used like that of Oxford-5k. Likewise, an additional disjoint set of 100,000 distractor images is added to obtain Paris-106k.\n\\textbf{INSTRE} consists of 28,543 images from 250 different object classes, including three disjoint subsets\\footnote{https://github.com/imatge-upc/salbow}: INSTRE-S1, INSTRE-S2, INSTRE-M. \n\\textcolor{black}{INSTRE dataset has bounding box annotations, providing single-labelled and double-labelled class information for single- and multiple-object retrieval, respectively. One can use the class information to build a tuple, with two positive examples from the same class and one negative from one of the remaining classes. The performance evaluation on INSTRE in our experiments follows the protocol in . }\n\\textbf{Google Landmarks Dataset (GLD)} , consists of GLD-v1 and GLD-v2. \\textcolor{black}{GLD-v2 is mainly recommended to use and it has the advantage of stability where all images have permissive licenses . GLD-v2 is divided into three subsets: (i) 118k query images with ground-truth annotations, (ii) 4.1M training images of 203k landmarks with labels, and (iii) 762k index images of 101k landmarks. Due to its large scale, GLD-v2 provides class-level ground-truth which can be used to build training tuples. Due to its image diversity, it may produce clutter images for each landmark so it is necessary to introduce pre-processing methods to select the more relevant images . Finally, the training set is cleaned by removing these clutters, consisting of a subset ``GLD-v2-clean'' containing 1.6M images of 81k landmarks. Since Google landmarks dataset stills lack bounding box for objects of interest, Teichmann \\etal provide a new dataset of landmark bounding boxes, based on GLD. This patch-level supervision information can help locate the most relevant regions. }\n\\textcolor{black}{Note that, additional queries and distractor images have been added into Oxford-5k and Paris-6k, producing the Revisited Oxford ($\\mathcal{R}$Oxford) and Revisited Paris ($\\mathcal{R}$Paris) datasets where each image of the same building is assigned a label \\textit{Easy}, \\textit{Hard}, \\textit{Unclear}, or \\textit{Negative} . Different label combinations are used as \\textit{positive} according to the difficulty level of different setups. During testing, if there are no positive images for a query, then that query is excluded from the evaluation. For details, we refer the reader to . We undertake partial comparisons under the \\textit{hard} evaluation protocol on these revisited datasets. }","id":"347ae713-adb4-418f-9e62-4191f141d96c","level":"subsection","origin_cites_number":13,"parent_id":"a6c69e00-d8a1-492b-8514-fab570f109d2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","State of the Art Performance"],["subsection","Datasets"]],"subsections":[],"title":"Datasets"},{"cite_extract_rate":0,"cites":[],"content":"\\textbf{Average precision} (AP) refers to the coverage area under the precision-recall (PR) curve. A larger AP implies a higher PR curve and better retrieval accuracy. AP can be calculated as $ AP=\\frac{1}{R} \\sum_{k=1}^N P(k)\\cdot rel(k) $,\nwhere $ R $ denotes the number of relevant results for the query image from the total number $ N $ of images. $ P(k) $ is the precision of the top $ k $ retrieved images, and $ rel(k) $ is an indicator function equal to 1 if the item within rank $ k $ is a relevant image and 0 otherwise. Mean average precision (mAP) is adopted for the evaluation over all query images, \\ie $mAP=\\frac1Q{\\sum_{q=1}^Q}AP(q) $,\nwhere $ Q $ is the number of query images.\nThe \\textbf{N-S score} is a metric used for UKBench ; the N-S score is the average for the top-4 precision over the dataset.","id":"78dcf4cf-a971-4517-838a-4552eb9d65a5","level":"subsection","origin_cites_number":1,"parent_id":"a6c69e00-d8a1-492b-8514-fab570f109d2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","State of the Art Performance"],["subsection","Evaluation Metrics"]],"subsections":[],"title":"Evaluation Metrics"},{"cite_extract_rate":0.50943396226415,"cites":[2136,2140,7543,2093,514,2096,2105,2113,8525,8526,7546,2094,2137,2131,2103,2139,2138,7544,2119,2109,2097,2132,2100,2117,2112,7545,2111],"content":"\\begin{figure}[t]\n\\centering\n\\begin{subfigure}{0.248\\textwidth}\n\\resizebox{\\columnwidth}{!}{\n\\begin{tikzpicture} \n\\scriptsize\n\\begin{axis}[legend style={at={(0.95,0.3)}, scale=0.5, anchor=north east, font= \\scriptsize}, symbolic x coords={2014, 2015, 2016, 2017, 2018, 2019, 2020 }, grid=both, ylabel={mAP(\\%)}, ymin=77.5, ymax=99, xmin=2014, xmax=2020, xtick=data, x tick label style={rotate=45,anchor=east} ];\n\\addlegendentry{Holidays}\n\\addplot[mark=square*,thick, line width=0.45mm, color={rgb, 255:red,152;green,179;blue,134}, on layer=main] coordinates { (2014, 80.18) (2015,89.7) (2016,94.2) (2017,95.13) (2018,95.67) (2019, 95.5) (2020, 94.0) };\n\\node [above right=0cm, font=\\tiny] at (axis cs:2014, 80.5) {80.18 }; \\node [ font=\\tiny] at (axis cs:2015, 90.4) {89.7 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2016, 93.4) {94.2 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2017, 94.8) {95.13 }; \\node [above right=-0.4cm, font=\\tiny] at (axis cs:2018, 95.6) {95.7 }; \\node [above right=-0.5cm, font=\\tiny] at (axis cs:2019, 95.9) {95.5 }; \\node [above right=-0.8cm, font=\\tiny] at (axis cs:2020, 96) {94.0 };\n\\addlegendentry{UKBench}\n\\addplot[mark=diamond*,thick,line width=0.45mm, color={rgb,255:red,141;green,170;blue,196}, on layer=main,] coordinates {(2014,91.1) (2015,91.3) (2016,96.3) (2018,98.08) (2020,98.84) }; \\node [above right=0cm, font=\\tiny] at (axis cs:2014, 91) {91.1 }; \\node [above right=-0.4cm, font=\\tiny] at (axis cs:2015, 92.8) {91.3 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2016, 97.4) {96.3 }; \\node [above right=-0.4cm, font=\\tiny] at (axis cs:2018, 99.5) {98.1 }; \\node [above right=-0.8cm, font=\\tiny] at (axis cs:2020, 100.8) {98.8 };\n\\addlegendentry{Oxford-5k}\n\\addplot[mark=*,thick, on layer=background, line width=0.45mm, color={rgb, 255:red,237;green,160;blue,155}] coordinates { (2014,78.34) (2015,84.4) (2016,88.95) \n(2018,95.8) (2019,96.2) (2020,96.2) }; \\node [above right=0cm, font=\\tiny] at (axis cs:2014, 77.5) {78.34 }; \\node [ font=\\tiny] at (axis cs:2015, 85.15) {84.4 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2016, 88.6) {88.95 }; \n\\node [above right=-0.4cm, font=\\tiny] at (axis cs:2018, 97.2) {95.8 }; \\node [above right=-0.5cm, font=\\tiny] at (axis cs:2019, 98) {96.2 }; \\node [above right=-0.8cm, font=\\tiny] at (axis cs:2020, 99.1) {96.2 };\n\\addlegendentry{Paris-6k}\n\\addplot[mark=triangle*,thick,line width=0.45mm, color={rgb, 255:red,217;green,188;blue,102}, on layer=main,] coordinates { (2014,86.83) (2015,86.5) (2016,95.88) (2017,96.0) (2018, 97.0) (2019, 97.8) (2020, 97.4) };\n\\node [above right=0cm, font=\\tiny] at (axis cs:2014, 86.8) {86.83 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2015, 87.8) {86.5 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2016, 95.6) {95.8 };\\node [above right=-0.3cm, font=\\tiny] at (axis cs:2017, 97.0) {96.0 }; \\node [above right=-0.4cm, font=\\tiny] at (axis cs:2018, 98.4) {97.0 }; \\node [above right=-0.5cm, font=\\tiny] at (axis cs:2019, 99.3) {97.8 }; \\node [above right=-0.8cm, font=\\tiny] at (axis cs:2020, 100) {97.4 };\n\\end{axis}\n\\end{tikzpicture}\n} \n\\vspace{-2.1em}\n\\end{subfigure}\n\\begin{subfigure}{0.235\\textwidth}\n\\resizebox{\\columnwidth}{!}{\n\\begin{tikzpicture} \n\\scriptsize\n\\begin{axis}[legend style={at={(0.425,0.3)}, scale=0.5, anchor=north east, font=\\scriptsize}, symbolic x coords={2017, 2018, 2019, 2020}, grid=both, ylabel={}, ymin=15, ymax=95, xmin=2017, xmax=2020, xtick=data, x tick label style={rotate=45, anchor=east} ];\n\\addlegendentry{$\\mathcal{R}$Oxford-5k}\n\\addplot[mark=*,thick,line width=0.45mm, color={rgb, 255:red,181;green,144;blue,202}, on layer=main,] coordinates { (2018, 54.8) (2019, 65.8) (2020, 64.0) };\n\\node [above right=-0.3cm, font=\\tiny] at (axis cs:2018, 53.8) {54.8 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2019, 72.5) {65.8 }; \\node [above right=-0.8cm, font=\\tiny] at (axis cs:2020, 74.0) {64.0 };\n\\addlegendentry{$\\mathcal{R}$Paris-6k}\n\\addplot[mark=triangle*,thick,line width=0.45mm, color={rgb, 255:red,117;green,207;blue,184}, on layer=main,] coordinates { (2018, 84.0) (2019, 85.3) (2020, 80.4) };\n\\node [above right=-0.3cm, font=\\tiny] at (axis cs:2018, 92.5) {84.0 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2019, 84.0) {85.3 }; \\node [above right=-0.9cm, font=\\tiny] at (axis cs:2020, 93.5) {80.4 };\n\\addlegendentry{INSTRE}\n\\addplot[mark=diamond*,thick,line width=0.45mm, color={rgb, 255:red,83;green,82;blue,4}, on layer=main,] coordinates { (2017, 80) (2018, 80.5) (2019, 92.4)};\n\\node [above right=-0.05cm, font=\\tiny] at (axis cs:2017, 71.5) {80.0 }; \\node [above right=-0.3cm, font=\\tiny] at (axis cs:2018, 79.5) {80.5 }; \\node [above right=0cm, font=\\tiny] at (axis cs:2019, 86) {92.4 } ;\n\\addlegendentry{GLD-v2}\n\\addplot[mark=square*,thick,line width=0.45mm, color={rgb, 255:red,130;green,54;blue,203}, on layer=main,] coordinates {(2020, 26.8) };\n\\node [above right=-0.8cm, font=\\tiny] at (axis cs:2020, 38) {26.8 };\n\\end{axis};\n\\end{tikzpicture}\n}\n\\vspace{-2.1em}\n\\end{subfigure}\n\\caption{Performance improved from 2014 to 2020.} \n\\label{datasets_progress_2014_2_2020}\n\\end{figure}\n\\textbf{Overview.} Figure \\ref{datasets_progress_2014_2_2020} summarizes the performance over 6 datasets from 2014 to 2020. Early on, the powerful feature extraction of DCNNs led to rapid improvements. Subsequent key ideas have been to extract instance features at the region level to reduce image clutter , and to improve feature discriminativity by using methods including feature fusion ,,, feature aggregation ,, and feature embedding . Fine-tuning is an important strategy to improve performance by tuning deep networks specific for learning instance features ,. For instance, the accuracy increases steadily from 78.34\\% to 96.2\\% on the Oxford-5k dataset when manifold learning is used to fine-tune deep networks. The mAP on $\\mathcal{R}$Paris-6k and $\\mathcal{R}$Oxford-5k is smaller than Paris-6k and Oxford-5k, leaving room for improvement.\nWe report results using off-the-shelf models (Table \\ref{Table_retrieval_off_the_shelf}) and fine-tuning networks (Table \\ref{Metric_learning_table_B}). In Table \\ref{Table_retrieval_off_the_shelf}, single-pass and multiple-pass are analyzed, while supervised and unsupervised fine-tuning are compared in Table \\ref{Metric_learning_table_B}. Since there are many aspects that vary across the different methods, making them not directly comparable, we mainly draw some general claims or trends based on the collected results.\n\\begin{table*}[t]\n\\centering \n\\caption{\\footnotesize Performance evaluation of off-the-shelf DCNN models. ``$ \\bullet $'' indicates that the models or layers are combined to learn features; ``PCA$_{w}$ indicates PCA with whitening on the extracted features to improve robustness; ``MP'' means Max Pooling; ``SP'' means Sum Pooling. The CNN-M network with ``$\\ast$'' has an architecture similar to that of AlexNet. ``-'' means that the results were not reported. }\n\\vspace{-1em}\n\\label{Table_retrieval_off_the_shelf}\n\\renewcommand{\\arraystretch}{1.0}\n\\resizebox{19.5cm}{!}{\n\\begin{tabular}{!{\\vrule width1.2bp}c|c|c|c|c|c|c|c|c|c|p{9cm}!{\\vrule width1.2bp}}\n\\Xhline{1pt}\n\\footnotesize Type & \\footnotesize \\shortstack [c] { Method } & \\footnotesize \\shortstack [c] {Backbone \\\\ DCNN}\t& \\footnotesize \\shortstack [c] {Output \\\\ Layer} & \\footnotesize \\shortstack [c] {Embed. \\\\ Aggre.}\t& \\footnotesize \\shortstack [c] { Feat. \\\\ Dim } & \\footnotesize \\shortstack [c] {Holidays} & \\footnotesize \\shortstack [c] {UKB} & \\footnotesize \\shortstack [c] {Oxford5k \\\\ (+100k)} & \\footnotesize \\shortstack [c] {Paris6k \\\\ (+100k)} & \\footnotesize Brief Conclusions and Highlights \\\\\n\\Xhline{1.0pt}\n\\multirow{5}{*}{ \\raisebox{-15ex}[0pt]{\\begin{tabular}[c]{@{}c@{}} \\rotatebox{90}{Single-pass} \\end{tabular}}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { Neural \\\\ codes } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ AlexNet} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { FC6}}\t& \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { PCA}} & \\footnotesize \\raisebox{-2ex}[0pt]{ $128$} & \\footnotesize \\raisebox{-2ex}[0pt]{ $74.7$} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { $ 3.42 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}}& \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $43.3$\\\\ (38.6)}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize Compressed neural codes of different layers are explored. AlexNet is also fine-tuned for retrieval.\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { SPoC } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { Conv5}}\t& \\footnotesize \\raisebox{-2.5ex}[0pt]{\\shortstack [c] { SPoC + \\\\ PCA$_{w} $}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 256} & \\footnotesize \\raisebox{-2ex}[0pt]{ 80.2} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { $ 3.65 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $58.9$\\\\ (57.8)}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize Exploring Gassian weighting scheme \\ie the centering prior, to improve the discrimination of\nfeatures. \\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { CroW } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{VGG16} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { Conv5}}\t& \\footnotesize \\raisebox{-2.5ex}[0pt]{\\shortstack [c] { CroW + \\\\ PCA$_{w} $}} & \\footnotesize \\raisebox{-2ex}[0pt]{256} & \\footnotesize \\raisebox{-2ex}[0pt]{ $85.1$} & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { $-$}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] {$68.4$\\\\ (63.7)}} & \\footnotesize\t\\raisebox{-3.0ex}[0pt]{\\shortstack [c] { $ 76.5 $\\\\ (69.1) }} & \\footnotesize The spatialwise and channelwise weighting mechanisms are utilized to highlight crucial convolutional features. \\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { R-MAC } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { Conv5}}\t& \\footnotesize \\raisebox{-2.5ex}[0pt]{\\shortstack [c] { R-MAC + \\\\ PCA$_{w} $}} & \\footnotesize \\raisebox{-2ex}[0pt]{512} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$ }} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $66.9$\\\\ (61.6)}} & \\footnotesize\t\\raisebox{-3.0ex}[0pt]{\\shortstack [c] { $-$ \\\\ (75.7) }} & \\footnotesize Sliding windows with different scales on convolutional feature maps to encode multiple image regions.\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { Multi-layer \\\\ CNN } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{VGG16} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { FC6 $\\bullet$ \\\\ Conv4$\\sim$5 }}\t& \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { SP}} & \\footnotesize \\raisebox{-2ex}[0pt]{4096} & \\footnotesize \\raisebox{-2ex}[0pt]{ 91.4} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { $ 3.68 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $61.5$\\\\($-$)}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize Layer-level feature fusion and the complementary properties of different layers are explored.\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { BLCF } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { Conv5}}\t& \\footnotesize \\raisebox{-2.5ex}[0pt]{\\shortstack [c] { BoW + \\\\ PCA$_{w} $}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 25k} & \\footnotesize \\raisebox{-2ex}[0pt]{$-$} & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { $-$}}& \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $73.9$\\\\ (59.3)}} & \\footnotesize\t\\raisebox{-3.0ex}[0pt]{\\shortstack [c] { $ 82.0 $\\\\ (64.8) }} & \\footnotesize Both global features and local features are explored, demonstrating that local features have higher accuracy. \\\\ \\cline{2-11} \n\\Xhline{1.0pt}\n\\multirow{7}{*}{ \\raisebox{-20ex}[0pt]{\\begin{tabular}[c]{@{}c@{}} \\rotatebox{90}{Multiple-pass} \\end{tabular}}}\n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { OLDFP } } & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] {AlexNet }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { FC6}}\t& \\footnotesize \\raisebox{-2.6ex}[0pt]{\\shortstack [c] { MP \\\\ + PCA$_{w} $ }} & \\footnotesize \\raisebox{-2ex}[0pt]{ 512} & \\footnotesize \\raisebox{-2ex}[0pt]{ 88.5} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] {$ 3.81 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $60.7$\\\\ ($-$)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $66.2$\\\\ ($-$)}} & \\footnotesize Exploring the impact of proposal number. Patches are extracted by RPNs (see Figure \\ref{MultiplePass} (d)) and the features are encoded in an orderless way.\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-1.6ex}[0pt]{ \\shortstack [c] { MOP-CNN } } & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { AlexNet }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { FC7}} & \\footnotesize \\raisebox{-2.5ex}[0pt]{\\shortstack [c] { VLAD \\\\ + PCA$_{w} $ }} & \\footnotesize \\raisebox{-2ex}[0pt]{ 2048} & \\footnotesize \\raisebox{-2ex}[0pt]{ 80.2} & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { $-$}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize Image patches are extracted densely, as shown in Figure \\ref{MultiplePass} (c). Multi-scale patch features are further embedded into VLAD descriptors. \\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { CNNaug-ss } } & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { Overfeat\\\\ }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { FC}}\t\t& \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { PCA$_{w} $ }} & \\footnotesize \\raisebox{-2ex}[0pt]{ 15k} & \\footnotesize \\raisebox{-2ex}[0pt]{ 84.3} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { $ 91.1 $ \\\\ \\textcolor[rgb]{0,0,0}{(mAP)}}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $68.0$\\\\($-$)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $79.5$\\\\ ($-$)}} & \\footnotesize Image patches are extracted densely, as shown in Figure \\ref{MultiplePass} (c). Image regions at different locations with different sizes are included.\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { MOF } } & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { CNN-M$^{\\ast}$ \\\\ }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { FC7 $\\bullet$ \\\\ Conv }}\t& \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { SP or MP\\\\ + BoW }} & \\footnotesize \\raisebox{-2ex}[0pt]{ 20k} & \\footnotesize \\raisebox{-2ex}[0pt]{ 76.8} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { $ 3.00 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize Exploring layer-level fusion scheme. Image patches are extracted using spatial pyramid modeling, as shown in Figure \\ref{MultiplePass} (b).\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2.7ex}[0pt]{ \\shortstack [c] { Multi-scale \\\\ CNN } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { Conv5}} & \\footnotesize \\raisebox{-2.7ex}[0pt]{\\shortstack [c] { SP or MP\\\\ + PCA$_{w} $ }} & \\footnotesize \\raisebox{-2ex}[0pt]{ 32k} & \\footnotesize \\raisebox{-2ex}[0pt]{ 89.6} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { $ 95.1 $ \\\\ \\textcolor[rgb]{0,0,0}{(mAP)}}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $84.3$\\\\ ($-$)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $87.9$\\\\ ($-$)}} & \\footnotesize Image patches are extracted in a dense manner, as shown in Figure \\ref{MultiplePass} (c). Geometric invariance is considered when aggregating patch features.\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { LDD } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG19} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { Conv5}}\t& \\footnotesize \\raisebox{-2.6ex}[0pt]{\\shortstack [c] {BoW \\\\ + PCA$_{w} $ }} & \\footnotesize \\raisebox{-2ex}[0pt]{500k} & \\footnotesize \\raisebox{-2ex}[0pt]{ 84.6 } & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$ }} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $83.3$ \\\\ ($-$)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $87.2$ \\\\ ($-$)}} & \\footnotesize Image patches are obtained using a uniform square mesh, as shown in Figure \\ref{MultiplePass} (a). Patch features are encoded into BoW descriptors.\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-1.7ex}[0pt]{ \\shortstack [c] { DeepIndex } } & \\footnotesize\t\\raisebox{-2.7ex}[0pt]{ \\shortstack [c] { AlexNet \\\\ $\\bullet$ VGG19 }} & \\footnotesize\t\\raisebox{-2.7ex}[0pt]{ \\shortstack [c] { FC6-7 $\\bullet$\\\\ FC17-18 }}\t& \\footnotesize \\raisebox{-2.7ex}[0pt]{\\shortstack [c] {BoW + \\\\ PCA }} & \\footnotesize \\raisebox{-2ex}[0pt]{ 512} & \\footnotesize \\raisebox{-2ex}[0pt]{ 81.7} & \\footnotesize \\raisebox{-2.7ex}[0pt]{\\shortstack [c] { $ 3.32 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize\t\\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-2.7ex}[0pt]{ \\shortstack [c] { $75.4$\\\\ ($-$)}} & \\footnotesize Exploring layer-level and model-level fusion methods. Image patches are extracted using spatial pyramid modeling, as shown in Figure \\ref{MultiplePass} (b).\\\\ \\cline{2-11} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { CCS } } & \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { GoogLeNet }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { Conv}}\t& \\footnotesize \\raisebox{-2.7ex}[0pt]{\\shortstack [c] { VLAD \\\\ + PCA$_{w} $ }} & \\footnotesize \\raisebox{-2ex}[0pt]{ 128} & \\footnotesize \\raisebox{-2ex}[0pt]{ 84.1} & \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { $ 3.81 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $64.8$\\\\ ($-$)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $76.8$\\\\ ($-$)}} & \\footnotesize Object proposals are extracted by RPNs, as shown in Figure \\ref{MultiplePass} (d). Object level and point level feature concatenation schemes are explored.\\\\ \\cline{2-11} \n\\Xhline{1.0pt}\n\\end{tabular}\n}\n\\vspace{-0.5em}\n\\end{table*}\n\\begin{table*}[!t]\n\\centering \n \\caption{ \\footnotesize Performance evaluation of methods in which DCNN models are fine-tuned, in a supervised or an unsupervised manner. ``CE Loss'' means the models are fine-tuned using the classification-based loss function in the form of Eq. \\ref{cross_entropy_loss}. ``Siamese Loss'' is in the form of Eq. \\ref{contrastive}. ``Regression Loss'' is in the form of Eq. \\ref{regressionLoss}. ``Triplet Loss'' is in the form of Eq. \\ref{triplet}. }\n \\vspace{-1em}\n \\label{Metric_learning_table_B} \n \\renewcommand{\\arraystretch}{1.0}\n\\resizebox{19.5cm}{!}{\n\\begin{tabular}{!{\\vrule width1.2bp}c|p{2.0cm}|c|p{1.35cm}|c|p{1.1cm}|p{1.25cm}|c|c|p{0.7cm}|c|c|p{7.8cm}!{\\vrule width1.2bp}}\n\\Xhline{1pt}\n\\footnotesize Type & \\footnotesize \\shortstack [c] { Method } & \\footnotesize \\shortstack [c] {Backbone \\\\ DCNN} & \\footnotesize \\shortstack [c] {Training \\\\ set}\t& \\footnotesize \\shortstack [c] {Output \\\\ Layer} & \\footnotesize \\shortstack [c] {Embed. \\\\ Aggre.}\t& \\footnotesize \\shortstack [c] { Loss \\\\ Function } & \\footnotesize \\shortstack [c] { Feat. \\\\ Dim } & \\footnotesize \\shortstack [c] {Holidays} & \\footnotesize \\shortstack [c] {UKB} & \\footnotesize \\shortstack [c] {Oxford5k \\\\ (+100k)} & \\footnotesize \\shortstack [c] {Paris6k \\\\ (+100k)} & \\footnotesize Brief Conclusions and Highlights \\\\ \\Xhline{1.0pt}\n\\multirow{5}{*}{ \\raisebox{-24ex}[0pt]{\\begin{tabular}[c]{@{}c@{}} \\rotatebox{90}{Supervised Fine-tuning}\\end{tabular}}} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\!\\!\\! $ Neural codes } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ AlexNet} & \\footnotesize \\raisebox{-3.5ex}[0pt]{\\shortstack [c] { Landmarks \\\\ }} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { FC6}}\t& \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { PCA }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { CE Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 128} & \\footnotesize \\raisebox{-2ex}[0pt]{ 78.9 } & \\footnotesize \\raisebox{-3.5ex}[0pt]{\\shortstack [c] { $ 3.29 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $55.7$\\\\ (52.3)}} & \\footnotesize \\raisebox{-2.0ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize The first work which fine-tunes deep networks for IIR. Compressed neural codes and different layers are explored. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { Nonmetric } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{VGG16} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ $\\!\\!\\! $ Landmarks } & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { Conv5}}\t& \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { PCA$_w$ }} & \\footnotesize\t\\raisebox{-3.2ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\! $ Regression \\\\ $\\!\\!\\!\\! $ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{512} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] {$-$} } & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $88.2$\\\\ (82.1)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $88.2$\\\\ (82.9)}} & \\footnotesize Similarity learning of similar and dissimilar pairs is performed by a neural network, optimized using regression loss. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { SIAM-FV } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ $\\!\\!\\!\\! $ Landmarks } & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { Conv5}}\t& \\footnotesize\t\\raisebox{-2.7ex}[0pt]{ \\shortstack [c] { FV + \\\\ PCA$_w$ }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Siamese \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 512} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] {$-$} } & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $81.5$\\\\ (76.6)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $82.4$\\\\ ($-$)}} & \\footnotesize Fisher Vector is integrated on top of VGG and is trained with VGG simultaneously. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { Faster \\\\ R-CNN } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { Oxford5k \\\\ Paris6k } } & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { Conv5}}\t& \\footnotesize \\raisebox{-2ex}[0pt]{\\shortstack [c] { MP / SP }} & \\footnotesize\t\\raisebox{-3.2ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\! $ Regression \\\\ $\\!\\!\\!\\! $ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 512} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] {$-$} } & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $75.1$\\\\ $(-)$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $80.7$\\\\ ($-$)}} & \\footnotesize RPN is fine-tuned, based on bounding box coordinates and class scores for specific region query which is region-targeted. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { SIFT-CNN } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{VGG16} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Holidays \\\\ UKB }} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { Conv5}}\t& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { SP }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Siamese \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 512} & \\footnotesize \\raisebox{-2ex}[0pt]{ 88.4 } & \\footnotesize \\raisebox{-3.5ex}[0pt]{\\shortstack [c] { $ 3.91 $ \\\\ \\textcolor[rgb]{0,0,0}{(N-S)}}} & \\footnotesize\t \\raisebox{-2ex}[0pt]{ \\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] {$-$}} & \\footnotesize SIFT features are used as supervisory information for mining positive and negative samples. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\!\\! $ Quartet-Net } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize\t\\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { GeoPair \\\\ }} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { FC6}}\t& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { PCA }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Siamese \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 128} & \\footnotesize \\raisebox{-2ex}[0pt]{71.2 } & \\footnotesize \\raisebox{-3.5ex}[0pt]{\\shortstack [c] { $ 87.5 $ \\\\ \\textcolor[rgb]{0,0,0}{(mAP)}}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $48.5$\\\\ ($-$)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] {$48.8$\\\\($-$)}} & \\footnotesize Quartet-net learning is explored to improve feature discrimination where double-margin contrastive loss is used. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { NetVLAD } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ VGG16} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\!\\! $ Tokyo Time \\\\ Machine }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { VLAD \\\\ Layer }}\t& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { PCA$_{w}$ }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Triplet \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 256} & \\footnotesize \\raisebox{-2ex}[0pt]{ 79.9 } & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $62.5$\\\\ ($-$)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $72.0$\\\\ ($-$)}} & \\footnotesize VLAD is integrated at the last convolutional layer of VGG16 network as a plugged layer. \\\\ \\cline{2-13} \\textbf{}\n& \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { Deep \\\\ Retrieval } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ ResNet-101} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ $\\!\\!\\!\\!\\! $ Landmarks } & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Conv5 \\\\ Block }}\t& \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { MP + \\\\ PCA$_w$ }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Triplet \\\\Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 2048} & \\footnotesize \\raisebox{-2ex}[0pt]{ 90.3 } & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $86.1$\\\\ (82.8)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $94.5$\\\\ (90.6)}} & \\footnotesize Dataset is cleaned automatically. Features are encoded by R-MAC. RPN is used to extract the most relevant regions. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { DELF } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ ResNet-101} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ GLDv1 } & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Conv4 \\\\ Block }}\t& \\footnotesize\t\\raisebox{-2.6ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\! $ Attention \\\\ + PCA$_w$ }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { CE Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 2048} & \\footnotesize \\raisebox{-2ex}[0pt]{$-$} & \\footnotesize \\raisebox{-2ex}[0pt]{$-$} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $83.8$\\\\ (82.6)}} & \\footnotesize \\raisebox{-3ex}[0pt]{ \\shortstack [c] { $85.0$\\\\ (81.7)}} & \\footnotesize Exploring the FCN to extract region-level features and construct feature pyramids of different sizes. \\\\ \\cline{2-13} \n \\Xhline{1.0pt}\n\\multirow{7}{*}{ \\raisebox{-13ex}[0pt]{\\begin{tabular}[c]{@{}c@{}} \\rotatebox{90}{Unsupervised Fine-tuning} \\end{tabular}}}\n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { MoM } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { VGG16 }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ Flickr 7M } & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { Conv5}}\t& \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { MP + \\\\ PCA$_{w} $ }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Siamese \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 64} & \\footnotesize \\raisebox{-2ex}[0pt]{ 87.5} & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $78.2$\\\\ (72.6)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $85.1$\\\\ (78.0)}} & \\footnotesize Exploring manifold learning for mining dis/similar samples. Features are tested globally and regionally.\\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { GeM } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { VGG16 }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ Flickr 7M } & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] { Conv5}}\t& \\footnotesize \\raisebox{-3ex}[0pt]{\\shortstack [c] { GeM \\\\ Pooling }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Siamese \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 512} & \\footnotesize \\raisebox{-2ex}[0pt]{ 83.1} & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $82.0$\\\\ (76.9)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $79.7$\\\\ (72.6)}} & \\footnotesize Fine-tuning CNNs on an unordered dataset. Samples are selected from an automated 3D reconstruction system. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { SfM-CNN } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ \\shortstack [c] { VGG16 }} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ Flickr 7M } & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [l] {Conv5}}\t& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { PCA$_{w}$ }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Siamese \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{512} & \\footnotesize \\raisebox{-2ex}[0pt]{ 82.5} & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $77.0$\\\\ (69.2)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $83.8$\\\\ (76.4)}} & \\footnotesize Employing Structure-from-Motion to select positive and negative samples from unordered images. \\\\ \\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { MDP-CNN } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ ResNet-101} & \\footnotesize\t\\raisebox{-2ex}[0pt]{ $\\!\\!\\!\\!\\! $ Landmarks } & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Conv5 \\\\ Block }}\t& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { SP }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { Triplet \\\\ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 2048} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] {$-$} } & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $85.4$\\\\ (85.1)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $96.3$\\\\ (94.7)}} & \\footnotesize Exploring global feature structure by modeling the manifold learning to select positive and negative pairs. \\\\ \n\\cline{2-13} \n& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { IME-CNN } } & \\footnotesize\t\\raisebox{-2ex}[0pt]{ ResNet-101} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\! $ Oxford105k \\\\ Paris106k }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { IME \\\\ Layer }}\t& \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] { MP }} & \\footnotesize\t\\raisebox{-3ex}[0pt]{ \\shortstack [c] { $\\!\\!\\!\\! $ Regression \\\\ $\\!\\!\\!\\!\\! $ Loss}} & \\footnotesize \\raisebox{-2ex}[0pt]{ 2048} & \\footnotesize \\raisebox{-2ex}[0pt]{ \\shortstack [c] {$-$} } & \\footnotesize\t\\raisebox{-2ex}[0pt]{\\shortstack [c] {$-$}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $92.0$\\\\ (87.2)}} & \\footnotesize \\raisebox{-3.5ex}[0pt]{ \\shortstack [c] { $96.6$\\\\ (93.3)}} & \\footnotesize Graph-based manifold learning is explored to mine the matching and non-matching pairs in unordered datasets. \\\\ \\cline{2-13} \n\\Xhline{1.0pt}\n\\end{tabular}\n}\n\\end{table*}\n\\textbf{Evaluation for single feedforward pass.} \\textcolor{black}{In general, we observe that fully-connected layers used as feature extractors may give a lower accuracy (\\eg 74.7\\% on Holidays in ), compared to using convolutional layers in Table \\ref{Table_retrieval_off_the_shelf}. For the case where the same VGG net is used, the way to embed or aggregate features is critical. The methods shown in Figure \\ref{SinglePass} improve the discrimination of convolutional feature maps and perform differently in Table \\ref{Table_retrieval_off_the_shelf}, 66.9\\% of R-MAC and 58.9\\% of SPoC on Oxford-5k, differences which we see as critical factors for further analysis. If embedded by a BoW model, the results are competitive on Oxford-5k and Paris-6k (73.9\\% and 82.0\\%, respectively), while its codebook size is 25k, which may affect retrieval efficiency. Moreover, layer-level feature fusion improves retrieval accuracy. Yu \\etal combine three layers (mAP of 91.4\\% on Holidays), outperforming the performance of non-fusion method (mAP of 80.2\\%).}\n\\textbf{Evaluation for multiple feedforward pass.}\nResults for the methods of Figure \\ref{MultiplePass} are reported in Table \\ref{Table_retrieval_off_the_shelf}. Among them, extracting image patches densely using VGG has the highest performance on the 4 datasets , and rigid grid with BoW encoding is competitive (mAP of 87.2\\% on Paris-6k). These two methods consider more patches, even background information, when used for feature extraction. Instead of generating patches densely, region proposals and spatial pyramid modeling introduce a degree of purpose and efficiency in processing image objects. Spatial information is better maintained using multiple-pass schemes than with single-pass. For example, a shallower network (AlexNet) and region proposal networks in have a UKBench N-Score of 3.81, higher than using deeper networks ,,. Besides feeding image patches into the same network, model-level fusion also exploits complementary spatial information to improve accuracy. For instance, as reported in , which combines AlexNet and VGG, the results on Holidays (81.7\\% of mAP) and UKBench (3.32 of N-Score) are better than these in (76.75\\% and 3.00, respectively). \n\\textbf{Evaluation for supervised fine-tuning.} Compared to off-the-shelf models, fine-tuning deep networks usually improves accuracy, see Table \\ref{Metric_learning_table_B}. For instance, the result on Oxford-5k by using a pre-trained VGG is improved from 66.9\\% to 81.5\\% in when a single-margin Siamese loss is used. Similar trends can be also observed on the Paris-6k dataset. For classification-based fine-tuning, its performance may be improved by using powerful DCNNs and feature enhancement methods such as the attention mechanism in , with an mAP increased from 55.7\\% in to 83.8\\% in on Oxford-5k. As for pairwise ranking loss fine-tuning, in some cases the loss used for fine-tuning is essential for performance improvement. For example, RPN is re-trained using regression loss on Oxford-5k and Paris-6k (75.1\\% and 80.7\\%, respectively) . Its results are lower than the results from (88.2\\% and 88.2\\%, respectively) where a transformation matrix is used to learn visual similarity. However, when RPN is trained by using triplet loss such as , the effectiveness of retrieval is improved significantly where the results are 86.1\\% (on Oxford-5k) and 94.5\\% (on Paris-6k). Feature embedding methods are important for retrieval accuracy; Ong \\etal embedded \\textit{Conv5} feature maps by Fisher Vector and achieved an mAP of 81.5\\% on Oxford-5k, while embedding feature maps by using VLAD achieves an mAP of 62.5\\% on this dataset ,.\n\\textbf{Evaluation for unsupervised fine-tuning.} \nCompared to supervised fine-tuning, unsupervised fine-tuning methods are relatively less explored. The difficulty for unsupervised fine-tuning is to mine sample relevance without ground-truth labels. In general, unsupervised fine-tuning methods should be expected to have lower performance than supervised. For instance, supervised fine-tuning using Siamese loss achieves an mAP 88.4\\% on Holidays, while unsupervised fine-tuning using the same loss function in ,, achieves 82.5\\%, 83.1\\%, and 87.5\\%, respectively. However, unsupervised fine-tuning methods can achieve a similar accuracy, even outperform the supervised fine-tuning, if a suitable feature embedding method is used. For instance, Zhao \\etal explore global feature structure modeling the manifold learning, producing an mAP of 85.4\\% (on Oxford-5k) and 96.3\\% (on Paris-6k), which is similar to supervised results of 86.1\\% (on Oxford-5k) and 94.5\\% (on Paris-6k). As another example, the precision of ResNet-101 fine-tuned by cross entropy loss achieves 83.8\\% on Oxford-5k , while the precision is further improved to 92.0\\% when an IME layer is used to embed features and fine-tuned in an unsupervised way . Note that fine-tuning strategies are related to the type of the target retrieval datasets. As demonstrated in Table \\ref{GLD_evaluation_on_training_dataset} and , fine-tuning on different datasets may produce a different final retrieval performance. \n\\textbf{Network depth.} We compare the efficacy of DCNNs by depth, following the fine-tuning protocols\\footnote{https://github.com/filipradenovic/cnnimageretrieval-pytorch} in . For fair comparisons, all convolutional features from these backbone DCNNs are aggregated by MAC , and fine-tuned by using the same loss function with the same learning rate, thus the adopted methods are the same except for the DCNN depth. We use the default feature dimension (\\emph{i.e.} AlexNet (256), VGG (512), GoogLeNet (1024), ResNet-50/101 (2048)). The results are reported in Figure \\ref{CNNandaggregation_depth_dim_dataset} (a). We observe that the deeper networks consistently lead to better accuracy due to extracting more discriminative features.\n\\textbf{Feature aggregation methods.} The methods of embedding convolutional feature maps were illustrated in Figure \\ref{SinglePass}. We use the off-the-shelf VGG (without updating parameters) on the Oxford and Paris datasets. The results are reported in Figure \\ref{CNNandaggregation_depth_dim_dataset} (b). We observe that the different ways to aggregate the same off-the-shelf DCNN leads to differences in retrieval performance. These reported results provide a reference for feature aggregation when one uses convolutional layers for performing retrieval tasks.\n\\textbf{Global feature dimension.} We add fully-connected layers on the top of pooled convolutional features of ResNet-50 to obtain global descriptors with their dimensions varying from 32 to 8192. The results of 5 datasets are shown in Figure \\ref{CNNandaggregation_depth_dim_dataset} (c). It is expected that higher-dimension features usually capture more semantics and are helpful for retrieval. The performance\ntends to be stable when the dimension is very large.\n\\textcolor{black}{\\textbf{Number of image regions.} We compare the retrieval performance when different number of regions are fed and other components are kept the same, as depicted in Figure \\ref{CNNandaggregation_depth_dim_dataset} (d). Convolutional features of each region are pooled as 2048-dim regional features by MAC and then aggregated into a global one. Note that the final memory requirement is identical for the case that a holistic image is used as input (\\ie regarded as the case where only one region is used). Regional inputs on an image are extracted with a 40\\% overlap of neighboring regions and the number varying from 1 to 41. For Oxford-5k, the best result is given by the case where 9 image regions are used. For the rest datasets, 3 image regions give the best results. Finally, more regions extracted from one image decline the retrieval mAP. A reason is that features of background or irrelevant regions have also been aggregated, and negatively affect the performance.}\n\\begin{table}[t]\n\\caption{Evaluations of training sets and retrieval reranking. Numerical results are cited from .\n}\n\\vspace{-1em}\n\\label{GLD_evaluation_on_training_dataset}\n\\footnotesize\n\\setlength{\\tabcolsep}{0.6mm}\n\\begin{tabular}{|cc|c|c|cc|c|c|c|}\n\\hline\n\\multicolumn{2}{|c|}{\\multirow{3}{*}{Conditions}} \n & \\multirow{3}{*}{Global} & \\multirow{3}{*}{\\begin{tabular}[c]{@{}c@{}}Local\\\\ reranking\\end{tabular}} & \\multicolumn{2}{c|}{Training set} & \\multirow{3}{*}{$\\mathcal{R}$Oxf} & \\multirow{3}{*}{$\\mathcal{R}$Par} & \\multirow{3}{*}{\\begin{tabular}[c]{@{}c@{}}GLD-v2\\\\ testing\\end{tabular}} \\\\ \\cline{5-6}\n & & & & \\multicolumn{1}{c|}{GLD-v1} & \\begin{tabular}[c]{@{}c@{}}GLD-v2\\\\ -clean\\end{tabular} & & & \\\\ \\hline\n\\multicolumn{1}{|c|}{\\multirow{4}{*}{\\begin{tabular}[c]{@{}c@{}}ResNet\\\\ -50\\end{tabular}}} & Case 1 & \\tickYes & \\tickNo & \\multicolumn{1}{c|}{\\tickYes} & \\tickNo & 45.1 & 63.4 & 20.4 \\\\ \\cline{2-9} \n\\multicolumn{1}{|c|}{} & Case 2 & \\tickYes & \\tickNo & \\multicolumn{1}{c|}{\\tickNo} & \\tickYes & 51.0 & 71.5 & 24.1 \\\\ \\cline{2-9} \n\\multicolumn{1}{|c|}{} & Case 3 & \\tickYes & \\tickYes & \\multicolumn{1}{c|}{\\tickYes} & \\tickNo & 54.2 & 64.9 & 22.3 \\\\ \\cline{2-9} \n\\multicolumn{1}{|c|}{}& Case 4 & \\tickYes & \\tickYes & \\multicolumn{1}{c|}{\\tickNo} & \\tickYes & 57.9 & 71.0 & 24.3 \\\\ \\hline\n\\multicolumn{1}{|c|}{\\multirow{4}{*}{\\begin{tabular}[c]{@{}c@{}}ResNet\\\\ -101 \\end{tabular}}} & Case 5 & \\tickYes & \\tickNo & \\multicolumn{1}{c|}{\\tickYes} & \\tickNo & 51.2 & 64.7 & 21.7 \\\\ \\cline{2-9} \n\\multicolumn{1}{|c|}{} & Case 6 & \\tickYes & \\tickNo & \\multicolumn{1}{c|}{\\tickNo} & \\tickYes & 55.6 & 72.4 & 26.0 \\\\ \\cline{2-9} \n\\multicolumn{1}{|c|}{} & Case 7 & \\tickYes & \\tickYes & \\multicolumn{1}{c|}{\\tickYes} & \\tickNo & 59.3 & 65.5 & 24.3 \\\\ \\cline{2-9} \n\\multicolumn{1}{|c|}{} & Case 8 & \\tickYes & \\tickYes & \\multicolumn{1}{c|}{\\tickNo} & \\tickYes & 64.0 & 72.8 & 26.8 \\\\ \\hline\n\\end{tabular}\n\\end{table}\n\\textbf{Fine-tuning datasets and retrieval reranking.} We compare performance on $\\mathcal{R}$Oxford-5k, $\\mathcal{R}$Paris-6k, and GLD-v2, aiming at comparing the role of different fine-tuning training sets and the effectiveness of retrieval reranking. Table \\ref{GLD_evaluation_on_training_dataset} lists 8 experimental scenarios using two network backbones, as in . \n\\textcolor{black}{Since GLD-v2 provides class-level ground-truth, its including images show large context diversity and may pose challenges to the network fine-tuning. Thus, the pre-processing steps, as proposed in ,, are necessary to select the more coherent images, referring to the GLD-v2-clean subset. As a result, when using the global features only, this cleaned version of the training set improves the performance, as observed in Cases 1/5 and Cases 2/6 for ResNet-50/ResNet-101, respectively.} As an important postprocessing strategy, reranking further boosts the accuracy after the initial filtering step by using global features.\n \\begin{figure*}[!t]\n\\centering \n {\n \\includegraphics[width=\\textwidth]{./Figures/CNNandaggregation_depth_dim_dataset.pdf} \n } \n \\vspace{-2em}\n \\caption{(a) The effectiveness of different DCNNs; (b) Comparison of the feature aggregation methods in Figure \\ref{SinglePass}; (c) The impact of global feature dimension by using ResNet-50; (d) Performance comparison when aggregating different numbers of image regions.} \n \\label{CNNandaggregation_depth_dim_dataset} \n \\end{figure*}","id":"4289168f-a812-4899-b8a0-52669cd0ba36","level":"subsection","origin_cites_number":53,"parent_id":"a6c69e00-d8a1-492b-8514-fab570f109d2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","State of the Art Performance"],["subsection","Performance Comparison and Analysis"]],"subsections":[],"title":"Performance Comparison and Analysis"},{"cite_extract_rate":0.7272727272727271,"cites":[2141,2136,2144,2142,2110,2093,2111,2143],"content":"\\label{Conclusion_and_Future_Directions}\n\\textcolor{black}{As a comprehensive yet timely survey on instance retrieval using deep learning, this paper has discussed the main challenges, presented a taxonomy of recent developments according to their roles in implementing instance retrieval, highlighted the recent representative methods and analyzed their merits and demerits, discussed the datasets, evaluation protocols, and SOTA performance. Nowadays the exponentially increasing amount of image and video data due to surveillance, e-commerce, medical images, handheld devices, robotics, \\etc, offers an endless potential for applications of instance retrieval. \nAlthough significant progress has been made, as discussed in Section \\ref{Keychallenges}, the main challenges in instance retrieval have not been fully addressed. Below we identify a number of promising directions for\nfuture research. }\n\\textcolor{black}{\\textbf{(1) Accurate and robust feature representations.} \nOne of the main challenges in instance retrieval is the large intra-class variations due to changes in viewpoint, scale, illumination, weather condition and background clutter \\emph{etc.}, as we discussed in Section~\\ref{Keychallenges}. However, DCNN representations have very little invariance, even though trained with lots of augmented data . Fortunately, before deep learning, with instance retrieval there are lots of important ideas in handling such intra-class variations like local interest point detectors and local invariant descriptors. Therefore, it is worth enabling DCNN to learn more accurate and robust representations via leveraging such traditional wisdom to design better DCNNs. In addition, unlike most objects in existing benchmarks which are rigid, planar and textured, textureless objects, 3D objects, transparent objects, reflective surfaces, \\etc are still very hard to find.}\n\\textcolor{black}{In addition, pursuing accuracy alone is not sufficient, as instance retrieval systems should be able to resist potential adversarial attacks. \nRecently, deep networks have been proven to be fooled rather easily by adversarial examples , \n\\ie images added with intentionally designed yet nearly imperceptible perturbations, which raises serious safety and robustness concerns. However, adversarial robustness in instance retrieval , has received very little attention, and should merit further effort.}\n\\textcolor{black}{\\textbf{(2) Compact and efficient deep representations.} In instance retrieval, \nsearching efficiently is as critical as searching accurately, especially for the pervasive \nmobile or wearable devices with very limited computing resources. However, existing methods adopt large scale, energy hungry DCNNs that are very difficult to be deployed in mobile devices. Hence, there has been pressing needs to develop compact, efficient, yet reusable deep representations tailored to the resource limited devices, like using binary neural networks ,,.}\n\\textcolor{black}{\\textbf{(3) Learning with fewer labels.} Deep learning require a large amount of high-quality labeled dataset to achieve high accuracy. The presence of labels errors or the limited amount of labeled data can greatly degrade DCNN's accuracy. However, collecting massive amounts of accurately labeled data is costly. In practical scenarios, datasets like GLDv2 , have long-tailed distributions, and noisy labels. Thus, to address such limitations, few shot learning , self-supervised learning , \nimbalanced learning , \n noisy label aware learning \\etc should be paid more attention in instance retrieval in the future. }\n\\textcolor{black}{\\textbf{(4) Continual learning for instance retrieval.}\nIn specific, the current IIR methods make restrictive assumptions, such as the training data being enough and stationary, retraining from scratch being possible when new data becomes available, which is problematic in realistic conditions. Our living world is continuously varying, and in\ngeneral data distributions are often non-stationary, new data may be added, and\npreviously unseen classes may be encountered. Thus, continual learning plays a vital role in continuously updating the IIR systems. The key issues are how to retain and utilize the previously learned knowledge, how to update the retrieval system as new images becomes available, and how to learn and improve over time.\n}\n\\textbf{(5) Privacy-aware instance retrieval.} Most IIR systems concentrate on improving the accuracy or efficiency performance, and the higher performance might come at the cost of users' privacy. Therefore, in some cases, such as personalized search systems, the privacy protection problem is also an important issue to be considered. Deep models should be privacy-aware and protect users' personalized searching experience to avoid their worries about using such IIR systems.\n\\textbf{(6) Video instance retrieval.} Searching a specific instance in an image cannot always meet the requirements in some scenarios such as the video surveillance system in the field of searching criminals. Currently, with the rapid growth of video data, retrieving a certain object, place, or action in videos has become more and more important and highly necessary in the future. For video instance retrieval, 3D-CNNs models need to be built to learn video’s spatio-temporal representations to compute the semantic similarity of instances.\n\\footnotesize\n\\bibliographystyle{IEEEtran}\n\\bibliography{IEEEabrv,myreference}\n\\vspace{-2.0 cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/WeiChen.pdf}}]{Wei Chen} received the Ph.D. degree at Leiden University, in 2021. Before starting with PhD study in Leiden University, he received his Master degree from the National University of Defense Technology, China, in 2016. His research interest focuses on cross-modal retrieval with deep learning methods, and also in the context of incremental learning. He has published papers in international conferences and journal including CVPR, ACM MM, PR, Neurocomputing, and IEEE TMM \\etc\n\\end{IEEEbiography}\n\\vspace{-1.7 cm} \n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Yu_Liu.pdf}}]{Yu Liu} is currently an Associate Professor in International School\nof Information Science and Engineering at Dalian University of Technology, China.\nHe was a post-doctoral researcher in the PSI group of KU Leuven, Belgium. \nIn 2018, he received the PhD degree from Leiden University, Netherlands.\nHis research interests include object recognition and retrieval in the context of continual learning and zero-shot learning. \nHe has co-organized several workshops at ICCV, CVPR and ECCV, respectively. \nHe has published papers in CVPR, ICCV, ECCV, ACM MM, IEEE TIP, etc, and \nreceived the best paper award at MMM2017.\n\\end{IEEEbiography}\n\\vspace{-1.7 cm} \n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Weiping_Wang.pdf}}]{Weiping Wang} received the Ph.D. degree in systems engineering from the National University of Defense Technology, Changsha, China. He is a Professor at Academy of Advanced Technology Research of Hunan. He has more than 200 papers published on journals and conferences including IEEE Transactions multimedia, Transactions on Vehicular Technology, \\etc His current research interests focus on intelligent decision experimentation, multi-modal knowledge extraction and knowledge integrated computation.\n\\end{IEEEbiography}\n\\vspace{-2.0 cm} \n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Erwin.pdf}}]{Erwin M. Bakker} is co-director of the LIACS Media Lab at Leiden University. He has published widely in the fields of image retrieval, audio analysis and retrieval and bioinformatics. He was closely involved with the start of the International Conference on Image and Video Retrieval (CIVR). Moreover, he regularly serves as a program committee member or organizing committee member for scientific multimedia and human-computer interaction conferences and workshops.\n\\end{IEEEbiography}\n\\vspace{-2.0 cm} \n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Theodoros_Georgiou.pdf}}]{Theodoros Georgiou} received the Ph.D. degree at Leiden University, in 2021. His research interest focuses on deep learning methods applied on higher than two dimensional data. Before starting with PhD study in Leiden University, he received his Master degree from the Leiden University in 2016. He has published papers in international conferences and journal including ICPR, CBMI, WCCI and IJMIR.\n\\end{IEEEbiography}\n\\vspace{-2.0 cm} \n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Paul_Fieguth.pdf}}]{Paul Fieguth} is co-director of the Vision \\& Image Processing Group in Systems Design Engineering at the University of Waterloo, where he is Professor and Associate Dean. He received the Ph.D.\\ degree from the Massachusetts Institute of Technology, Cambridge, in 1995, and has held visiting appointments at the University of Heidelberg in Germany, at INRIA/Sophia in France, at the Cambridge Research Laboratory in Boston, at Oxford University and the Rutherford Appleton Laboratory in England. His research interests include statistical signal and image processing, hierarchical algorithms, data fusion, machine learning, and the interdisciplinary applications of such methods. He has published textbooks on Statistical Image Processing and on Complex Systems theory.\n\\end{IEEEbiography}\n\\vspace{-1.7 cm} \n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/LiSmileGoodFace.pdf}}]{Li Liu} received the Ph.D. degree in information and communication engineering from the National University of Defense Technology, China, in 2012. She is currently a Professor with the College of System Engineering. She has held visiting appointments at the University of Waterloo, Canada, at the Chinese University of Hong Kong, and at the University of Oulu, Finland. Her current research interests include computer vision, pattern recognition and machine learning. Her papers have currently over 7800 citations in Google Scholar.\n\\end{IEEEbiography}\n\\vspace{-1.7 cm}\n\\begin{IEEEbiography}[{\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Michael_Lew.pdf}}]{Michael S. Lew} is the head of the Deep Learning and Computer Vision Research Group and a full Professor at LIACS, Leiden University. He has published over a dozen books and 190 peer-reviewed scientific articles in the areas of image retrieval, computer vision, and deep learning. Notably, he had the most cited paper in the ACM Transactions on Multimedia and one of the top 10 most cited articles in the history (out of more than 16,000 articles) of the ACM SIGMM. He was also a founding member of the advisory committee for the TRECVID video retrieval evaluation project, chair of the steering committee for the ACM International Conference on Multimedia Retrieval and a member of the ACM SIGMM Executive Committee. \n\\end{IEEEbiography}\n\\vspace{-1.7 cm} \n \\newpage\n\\end{document}","id":"6d16c13c-e552-486e-8c1d-7e8590b69688","level":"section","origin_cites_number":11,"parent_id":"515509a6-319a-4385-b1b5-4418a83d2bc2","prefix_titles":[["title","Deep Learning for Instance Retrieval: A Survey"],["section","Conclusions and Outlooks"]],"subsections":[],"title":"Conclusions and Outlooks"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"515509a6-319a-4385-b1b5-4418a83d2bc2\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ]\n ],\n \"subsections\": [\n \"87ccd777-d058-4a76-9575-0c136cb10697\",\n \"8a512096-a741-4540-8715-b12b364b3d9e\",\n \"1b3ca748-df26-4b7f-85f3-3eb438a055f5\",\n \"acffc1fe-bf03-450b-857b-db143905236e\",\n \"a6c69e00-d8a1-492b-8514-fab570f109d2\",\n \"6d16c13c-e552-486e-8c1d-7e8590b69688\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.35714285714285704,\n \"cites\": [\n 2093,\n 2092,\n 2091,\n 2094,\n 2090,\n 97,\n 8490,\n 2089,\n 2095,\n 209\n ],\n \"content\": \"\\\\label{sec:intro}}\\n\\\\IEEEPARstart{C}{ontent} Based Image Retrieval (CBIR) is the problem of searching for relevant images in an image gallery by analyzing the visual content (colors, textures, shapes, objects \\\\emph{etc.}), given a query image ,. CBIR has been a longstanding research topic in the fields of computer vision and multimedia ,. With the exponential growth of image data, the development of appropriate information systems that efficiently manage such large image collections is of utmost importance, with image searching being one of the most indispensable techniques. Thus there is a nearly endless potential for applications of CBIR, such as person/vehicle reidentification ,, landmark retrieval , \\nremote sensing , online product searching .\\nGenerally, CBIR methods can be grouped into two different tasks ,: Category level Image Retrieval (CIR) and Instance level Image Retrieval (IIR). The goal of CIR is to find an arbitrary image representative of the same category as the query (\\\\eg dogs, cars) ,. By contrast, in the IIR task, a query image of a particular instance (\\\\eg the Eiffel Tower, my neighbor's dog) is given and the goal is to find images containing the same instance that may be captured under different conditions like different imaging distances, viewing angles, backgrounds, illuminations, and weather conditions (reidentifying exemplars of the same instance) ,. The focus of this survey is the IIR task\\\\footnote{If not further specified, ``image retrieval'', ''IIR'', and ``instance retrieval'' are considered equivalent and will be used interchangeably.}.\\nIn many real world applications, IIR is usually to find a desired image requiring a search among thousands, millions, or even billions of images. Hence, searching efficiently is as critical as searching accurately, to which continued efforts have been devoted ,,. To enable accurate and efficient retrieval over a large-scale image collection, \\\\emph{developing compact yet discriminative feature representations} is at the core of IIR.\\n\\\\iffalse\\n\\\\begin {figure}[!t]\\n\\\\centering\\n\\\\begin{tabular}{c}\\n\\\\includegraphics[width=0.9\\\\columnwidth]{./Figures/TheProblem.pdf} \\\\\\\\\\n(\\\\textit{a}) The CBIR problem \\\\\\\\\\n\\\\includegraphics[width=0.9\\\\columnwidth]{./Figures/InstancevsCategory.pdf} \\\\\\\\\\n(\\\\textit{b}) instance-level retrieval versus category-level retrieval\\n \\\\end{tabular}\\n\\\\caption{ Illustration of (a) the CBIR problem, and (b)\\nCBIR categorization. The images in green frame are retrieved correctly, while the ones in red frame are matched incorrectly.\\n}\\n\\\\label{category_vs_instance}\\n\\\\vspace{-1em}\\n\\\\end {figure}\\n\\\\fi\\n\\\\begin{table*}[!ht]\\n\\\\caption{A summary and comparison of the primary surveys in the field of image retrieval.}\\\\label{SurveyCompare}\\n\\\\vspace{-1em}\\n\\\\centering\\n\\\\setlength{\\\\tabcolsep}{3.0mm}\\\\footnotesize\\n\\\\begin{tabular}{!{\\\\vrule width0.3bp}p{5cm}|p{0.5cm}|p{1.6cm}|p{9.2cm}!{\\\\vrule width0.3bp}}\\n\\\\Xhline{1.0pt}\\n\\\\multicolumn{1}{|c|}{ \\\\multirow{2}*{ Title} } & \\\\multirow{2}*{ $\\\\!\\\\!\\\\!\\\\!$ Year} & \\\\multirow{2}*{ Published in} & \\\\multicolumn{1}{c|}{ \\\\multirow{2}*{ Main Content } } \\\\\\\\ \\n\\\\hline\\n\\\\multirowcell{2}{Content-Based Image Retrieval at the \\\\\\\\ End of the Early Years } & \\\\multirow{2}*{2000} & \\\\multirowcell{2}{ TPAMI } & \\\\multirowcell{2}{This paper discusses the steps for image retrieval systems, including image \\\\\\\\ processing, feature extraction, user interaction, and similarity evaluation. }\\\\\\\\\\n\\\\hline\\n\\\\multirowcell{2}{Image Search from Thousands to Billions \\\\\\\\in 20 Years } & \\\\multirow{2}*{2013} & \\\\multirowcell{2}{ TOMM} & \\\\multirowcell{2}{This paper gives a good presentation of image search achievements from \\\\\\\\ 1970 to 2013, but the methods are not deep learning-based.} \\\\\\\\\\n\\\\hline\\n\\\\multirowcell{2}{Deep Learning for Content-Based Image \\\\\\\\ Retrieval: A Comprehensive Study } & \\\\multirow{2}*{2014} & \\\\multirowcell{2}{ ACM MM} & \\\\multirowcell{2}{This paper introduces supervised metric learning methods for fine-tuning \\\\\\\\ AlexNet. Details of instance-based image retrieval are limited.}\\\\\\\\\\n\\\\hline\\n\\\\multirowcell{2}{Semantic Content-based Image Retrieval:\\\\\\\\ A Comprehensive Study } & \\\\multirow{2}*{2015} & \\\\multirowcell{2}{ JVCI } & \\\\multirowcell{2}{This paper presents a comprehensive study about CBIR using traditional \\\\\\\\ methods; deep learning is introduced as a section with limited details.}\\\\\\\\\\n\\\\hline\\n\\\\multirowcell{3}{Socializing the Semantic Gap: A Compa-\\\\\\\\rative Survey on Image Tag Assignment, \\\\\\\\Refinement, and Retrieval } & \\\\multirowcell{1}{2016} & \\\\multirowcell{3}{CSUR} & \\\\multirowcell{3}{A taxonomy is introduced to structure the growing literature of image \\\\\\\\ retrieval. Deep learning methods for feature learning is introduced as \\\\\\\\future work.} \\\\\\\\\\n\\\\hline\\n\\\\multirowcell{2}{Recent Advance in Content based Image \\\\\\\\ Retrieval: A Literature Survey } & \\\\multirow{2}*{2017} & \\\\multirowcell{2}{ arXiv } & \\\\multirowcell{2}{This survey presents image retrieval from 2003 to 2016. Neural networks \\\\\\\\ are introduced in a section and mainly discussed as a future direction.}\\\\\\\\\\n\\\\hline\\n\\\\multirowcell{3}{Information Fusion in Content-based \\\\\\\\Image Retrieval: A Comprehensive \\\\\\\\ Overview } & \\\\multirowcell{1}{2017} & \\\\multirowcell{3}{Information\\\\\\\\Fusion} & \\\\multirowcell{3}{This paper presents information fusion strategies in CBIR. \\\\\\\\ Deep convolutional networks for feature learning are introduced \\\\\\\\ briefly but not covered thoroughly.} \\\\\\\\\\n\\\\hline\\n\\\\multirowcell{2} { $\\\\!\\\\!\\\\!$ A Survey on Learning to Hash } & \\\\multirow{2}*{2018} & \\\\multirowcell{2}{ TPAMI} & \\\\multirowcell{2}{This paper focuses on hash learning algorithms and introduces the \\\\\\\\ similarity-preserving methods and discusses their relationships. } \\\\\\\\\\n\\\\hline\\n\\\\multirowcell{2}{$\\\\!\\\\!\\\\!$ SIFT Meets CNN: A Decade Survey of \\\\\\\\ Instance Retrieval } & \\\\multirow{2}*{2018} & \\\\multirowcell{2}{ TPAMI} & \\\\multirowcell{2}{This paper presents a comprehensive review of instance retrieval based on \\\\\\\\SIFT and CNN methods.}\\\\\\\\\\n\\\\hline\\n\\\\multirowcell{3}{ \\\\bf Deep Learning for Instance \\\\\\\\ \\\\bf Retrieval: A Survey } & \\\\multirowcell{1}{\\\\bf 2021} & \\\\multirowcell{3}{ \\\\bf Ours} & \\\\multirowcell{3}{\\\\bf Our survey focuses on deep learning methods. We expand the review \\\\\\\\ \\\\bf with indepth details on IIR, including methods of feature extraction, \\\\\\\\ \\\\bf feature embedding and aggregation, and network fine-tuning. }\\n \\\\\\\\\\n\\\\Xhline{1.0pt}\\n\\\\end{tabular}\\n\\\\end{table*}\\nDuring the past two decades, startling progress has been witnessed in image representation which mainly consists of two important periods, \\\\ie feature engineering and deep learning. In the feature engineering era, the field was dominated by various milestone handcrafted image representations like SIFT and Bag of visual Words (BoW) . The deep learning era was reignited in 2012 when a Deep Convolutional Neural Network (DCNN) referred as ``AlexNet'' won the first place in the ImageNet classification contest with a breakthrough reduction in classification error rate. Since then, the dominating role of SIFT like local descriptors has been replaced by data driven Deep Neural Networks (DNNs) which can learn powerful feature representations with multiple levels of abstraction directly from data. During the past decade, DNNs have set the state of the art in various classical computer vision tasks, including image classification~,, object detection~, semantic segmentation , \\nand image retrieval~.\\nGiven this period of rapid evolution, the goal of this paper is to provide a comprehensive survey of the recent achievements in IIR. In comparison with existing excellent surveys on traditional image retrieval ,,,, as contrasted in Table~\\\\ref{SurveyCompare}, our focus in this paper is reviewing deep learning based methods for IIR, particularly on issues of retrieval accuracy and efficiency.\",\n \"id\": \"87ccd777-d058-4a76-9575-0c136cb10697\",\n \"level\": \"section\",\n \"origin_cites_number\": 28,\n \"parent_id\": \"515509a6-319a-4385-b1b5-4418a83d2bc2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [\n \"d77f2bd6-4077-4eb5-ba4e-7c6288149714\",\n \"846c213c-7851-4f7e-b5f1-17d26ed0cdb1\"\n ],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.5789473684210521,\n \"cites\": [\n 7543,\n 2101,\n 514,\n 2096,\n 2098,\n 2102,\n 2105,\n 2108,\n 2094,\n 2104,\n 2103,\n 8524,\n 7544,\n 2109,\n 2099,\n 2097,\n 2106,\n 97,\n 2107,\n 629,\n 2100,\n 7545\n ],\n \"content\": \"\\\\label{Summary_of_Progress}\\nAfter the highly successful image classification implementation of AlexNet~, significant exploration of DCNNs for instance retrieval tasks was undertaken and representative methods are shown in Figure \\\\ref{fig1}. Building on these methods, more recent progress for IIR can be achieved from the perspectives of off-the-shelf models and fine-tuned models, which form the basis for this survey.\\nOff-the-shelf models, based on DCNNs with fixed parameters ,,, extract features at image scales or patch scales, which correspond to single-pass and multiple-pass schemes, respectively. These methods focus on effective feature use, for which researchers have proposed embedding and aggregation methods, such as R-MAC , CroW , and SPoC to promote the discriminativity of the extracted features. Fine-tuned models, based on DCNNs in which parameters are updated , are more popular since deep networks themselves have been investigated extensively. To learn better retrieval features, researchers have proposed to improve the network architectures and/or update the pre-stored parameters . \\nThis survey will explore recent progress in detail in the context of the three following themes:\\n\\\\begin{figure*}[!t]\\n\\\\centering\\n\\\\includegraphics[width=\\\\textwidth]{./Figures/RepresentativeMethods.pdf} \\n\\\\vspace{-1.5em}\\n\\\\caption{Representative methods in IIR. Off-the-shelf models have model parameters which are not further updated or tuned when extracting retrieval features. For single-pass schemes, the key step is the feature embedding and aggregation to promote the discriminativity of the extracted image-level activations ,,,, whereas for multiple-pass schemes the goal is to extract instance features at region scales and eliminate image clutter as much as possible ,,,. In contrast, for fine-tuned models, the model parameters are tuned towards the retrieval task and address the issue of domain shifts. For supervised fine-tuning, the key step lies in the design of objective functions and sample sampling strategies ,,,,, while the success of unsupervised fine-tuning is to mine the relevance among training samples ,,,,. See Sections \\\\ref{Retrieval_with_Off_the_Shelf_DCNN_Models} and \\\\ref{Retrieval_via_Learning_DCNN_Representations} for details.}\\n\\\\label{fig1}\\n\\\\end{figure*}\\n\\\\medskip\\n\\\\noindent (1) \\\\emph{Deep Feature Extraction} \\\\hfill (Section \\\\ref{Deep_Feature_Extraction})\\n\\\\smallskip\\n\\\\noindent \\nOne key step in IIR is to make the descriptors as semantic-aware , as possible. For this, some recent work focus on the input data of DCNNs, thereby instance features can be extracted from the whole image, \\\\eg CroW , VLAD-CNN or from image patches, \\\\eg MOP-CNN , FAemb . For instance, evaluated on the Holidays dataset , patch-level input scheme can improve mAP by 8.29\\\\% compared to the results (70.53\\\\%) achieved using image-level input . Others focus on exploring different feature extractors, \\\\eg one layer of a given DCNN, to get the output activations. Initially, fully-connected layers are usually chosen to extract features ,, and then convolutional layers are popularly used ,. Afterwards, some work leverage the complementarity of different extractors to explore layer-level fusion, such as MoF , and model-level fusion, such as DeepIndex to promote the retrieval performance.\\n\\\\medskip\\n\\\\noindent \\n(2) \\\\emph{Feature Embedding and Aggregation} \\\\hfill (Section \\\\ref{Deep_Feature_Enhancement})\\n\\\\smallskip\\n\\\\noindent \\nRecent work revisit the classical embedding and aggregation methods and apply on deep features. Most work have a preference towards mapping individual vector from convolutional layer ,, and then aggregating into a global feature. The mapping process can be realized by using a pre-trained codebook (\\\\ie built separately), such as VLAD-CNN , DeepIndex or learned as parameters during training (built simultaneously), such as NetVLAD , GeM-DSM . Some work aggregate local features into a global one by direct pooling or sophisticated pooling-based methods without the aggregation operation, such as R-MAC .\\n\\\\medskip\\n\\\\noindent \\n(3) \\\\emph{Network Fine-tuning for Learning Representations} \\\\hfill (Section \\\\ref{Retrieval_via_Learning_DCNN_Representations})\\n\\\\smallskip\\n\\\\noindent \\nDCNNs pretrained on source datasets for image classification are influenced by domain shifts when performing retrieval tasks on new datasets. Thus, it is necessary to fine-tune deep networks to the specific domain , by using supervised or unsupervised fine-tuning methods. As depicted in Figure \\\\ref{fig1}, recent supervised fine-tuning methods focus on designing objective functions (\\\\eg Listwise loss ) and sample sampling strategies, such as NetVLAD , Triplet Network . Unsupervised methods focus on mining the relevance among training samples by using clustering, such as SfM-GeM or manifold learning, such as AILIR . Recently, convolution-free models that only rely on transformer layers have shown competitive performance and been used as a powerful alternative to DCNNs, such as IRT , reranking Transformers .\",\n \"id\": \"d77f2bd6-4077-4eb5-ba4e-7c6288149714\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 38,\n \"parent_id\": \"87ccd777-d058-4a76-9575-0c136cb10697\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Summary of Progress since 2012\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary of Progress since 2012\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 2095,\n 2110,\n 2112,\n 2111,\n 2109,\n 2105\n ],\n \"content\": \"\\\\label{Keychallenges}\\nThe goal of each of the preceding three themes is to address the competing objectives of \\\\emph{accuracy} and \\\\emph{efficiency}, with both objectives continuing to present challenges:\\n\\\\textbf{A) Accuracy related challenges} depend on the input data, the feature extractor, and the way in which the extracted features are processed:\\n\\\\begin{itemize}\\n\\\\item Invariance challenge: The instance in an image may be rotated, translated, or scaled differently, so the final features are affected by these transformations and retrieval accuracy may be degraded . It is necessary to incorporate invariance into the IIR pipeline ,.\\n\\\\item Distraction challenge: IIR systems may need to focus on only a certain object, or even only a small portion. DCNNs may be affected by image clutter or background, so multiple-pass schemes have been examined where region proposals are studied before feature extraction.\\n\\\\item Discriminativity challenge: Deep features for IIR are needed to be as discriminative as possible to distinguish instances with small differences, leading to many explorations in feature processing. These include feature embedding and aggregation methods, to promote feature discriminativity; and attention mechanisms, to highlight the most relevant regions within the extracted features or to enable deep networks to focus on regions of interest. \\n\\\\item Fine-tune challenge: DCNNs can be fine-tuned as powerful extractors to capture fine semantic distinctions among instances. These explorations have offered improved accuracy, however the area remains a major challenge. \\n\\\\end{itemize}\\n\\\\begin{figure*}[!t]\\n\\\\centering\\n\\\\includegraphics[width=0.95\\\\textwidth]{./Figures/PipelineofImageRetrieval.pdf}\\n \\\\vspace{-0.5em}\\n\\\\caption{General framework of IIR, which includes feature extraction on image or image patches, followed by feature embedding and aggregation methods to improve feature discriminativity. Feature matching can be performed by using global features (initial filtering) or use local features to rerank the top-ranked images matched by global features. } \\n\\\\label{PipelineofImageRetrieval}\\n\\\\end{figure*}\\n\\\\textbf{B) Efficiency related challenges} are important, especially for large-scale datasets . Retrieval systems should respond quickly when given a query image. Deep features are high-dimensional and contain semantic-aware information to support higher accuracy, yet is often at the expense of efficiency. \\nOn the one hand, the efficiency is related to the \\nformat of features, \\\\ie real valued or binary. Hash codes have advantages in storage and searching ,, however for hashing methods one needs to carefully consider the loss function design ,, to obtain optimal codes for high retrieval accuracy. \\nOn the other hand, efficiency is also related to the mechanism of feature matching. For example, instead of time-consuming cross-matching between local features, one can choose to use global features to perform an initial ranking and then a post-step re-ranking via the features of top-ranked images. \\n\\\\vspace{-1em}\",\n \"id\": \"846c213c-7851-4f7e-b5f1-17d26ed0cdb1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"87ccd777-d058-4a76-9575-0c136cb10697\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ],\n [\n \"subsection\",\n \"Key Challenges\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Key Challenges\"\n },\n {\n \"cite_extract_rate\": 0.36842105263157804,\n \"cites\": [\n 2093,\n 2113,\n 2094,\n 2115,\n 2114,\n 2110,\n 7545\n ],\n \"content\": \"\\\\label{General_Framework_of_IIR}\\nFigure \\\\ref{PipelineofImageRetrieval} offers an overview of the general framework for deep-learning-based IIR, involving three main stages. \\n\\\\medskip\\n\\\\noindent \\n\\\\textbf{1) Deep feature extraction}: \\\\hfill (Section \\\\ref{Deep_Feature_Extraction})\\n\\\\smallskip\\n\\\\noindent\\nFeature extraction is the first step of IIR and can be realized in a single-pass or multiple-pass way. Single-pass methods take as input the whole image, whereas multiple-pass methods depend on region extraction, as depicted in Figure \\\\ref{MultiplePass}.\\nThe activations from fully-connected layers of a given DCNN can be used as retrieval features whether based on a whole image or on patches. The tensors from convolutional layers can be used when further processed by sophisticated pooling, as shown in Figure \\\\ref{PipelineofImageRetrieval}. Different layers of the same deep network can be combined as a more powerful extractor \\n,. Furthermore, it is possible to fuse the activations from layers of different models \\n,. Feature extraction is the step to produce vanilla network activations (\\\\ie 3D tensors or a single vector), these activations, in most cases, are needed to be further processed. \\n\\\\medskip\\n\\\\noindent \\n\\\\textbf{2) Embedding and aggregation}: \\\\hfill (Section \\\\ref{Deep_Feature_Enhancement})\\n\\\\smallskip\\n\\\\noindent\\nFeature embedding and aggregation are two essential steps to produce global or local features. Feature embedding maps individual local features into higher-dimensional space whereas feature aggregation summarizes the multiple mapped vectors or all individual features into a global vector. Global features may come from pooling convolutional feature maps directly , or using some sophisticated weighting methods , (\\\\ie both without feature embedding). Feature embedding method using a pre-generated codebook can be performed to encode individual convolutional vectors and then aggregated ,,. For local features, the well-embedded representations for all regions of interest are stored individually and used for cross-matching in the reranking stage without aggregation.\\n\\\\medskip\\n\\\\noindent \\n\\\\textbf{3) Feature matching}:\\n\\\\smallskip\\n\\\\noindent\\nFeature matching is a process to measure the feature similarity between images and then return a ranked list. Global matches can be computed efficiently via such as Euclidean distance. For local features ,, the image similarity is usually evaluated by summarizing the similarities across local features, using classical RANSAC or more recent variations ,. Storing local features separately and then estimating their similarity individually lead to additional memory and search costs ,, therefore in most cases local features are used to re-rank the initial ranking image matched by global features ,,,.\\nThe three preceding stages for IIR rely on DCNNs as backbone architectures. In almost all cases, pre-stored parameters in these backbones can be fine-tuned (Section \\\\ref{Retrieval_via_Learning_DCNN_Representations}) to be better suited for instance retrieval and to contribute to better performance. \\nThe detailed categorization of the material of the following sections is shown in Figure~\\\\ref{FourAspectsofprogress}.\",\n \"id\": \"8a512096-a741-4540-8715-b12b364b3d9e\",\n \"level\": \"section\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"515509a6-319a-4385-b1b5-4418a83d2bc2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"General Framework of IIR\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"General Framework of IIR\"\n },\n {\n \"cite_extract_rate\": 0.5769230769230761,\n \"cites\": [\n 2118,\n 514,\n 2096,\n 2113,\n 2105,\n 2094,\n 2103,\n 7544,\n 2119,\n 2097,\n 2109,\n 2116,\n 629,\n 7545,\n 2117\n ],\n \"content\": \"\\\\label{Retrieval_with_Off_the_Shelf_DCNN_Models}\\n\\\\textcolor{black}{Because of their size, DCNNs need to be trained, initially for classification tasks, on exceptionally large-scale datasets and be able to recognize images from different classes.} One possible scheme, then, is that DCNNs effectively trained for classification directly serve as off-the-shelf feature detectors for the image retrieval task, the topic of this survey. That is, one can propose to undertake image retrieval on the basis of DCNNs, trained for classification and with their pre-trained parameters frozen. \\n\\\\begin{figure}[!htbp]\\n\\\\centering\\n\\\\footnotesize\\n\\\\begin{tikzpicture}[xscale=0.8, yscale=0.4]\\n\\\\draw [thick, -] (0, 16.5) -- (0, -4); \\\\node [right] at (-0.5, 17) {\\\\bf \\\\em Deep Learning for Instance-level Image Retrieval (overall survey)};\\n\\\\draw [thick, -] (0, 16) -- (0.5, 16);\\\\node [right] at (0.5, 16) {Retrieval with Off-the-Shelf DCNN Models (Section \\\\ref{Retrieval_with_Off_the_Shelf_DCNN_Models})};\\n\\\\draw [thick, -] (1, 15.5) -- (1, 5);\\n\\\\draw [thick, -] (1, 15) -- (1.5, 15);\\\\node [right] at (1.5, 15) {\\\\bf Deep Feature Extraction (Section \\\\ref{Deep_Feature_Extraction})};\\n\\\\draw [thick, -] (2, 14.5) -- (2, 8); \\n\\\\draw [thick, -] (2, 14) -- (2.5, 14);\\\\node [right] at (2.5, 14) {Network Feedforward Scheme (Section \\\\ref{Network_Feedforward_Scheme})};\\n\\\\draw [thick, -] (3, 13.5) -- (3, 12); \\n\\\\draw [thick, -] (3, 13) -- (3.5, 13);\\\\node [right] at (3.5, 13) {Single Feedforward Pass: MAC , R-MAC };\\n\\\\draw [thick, -] (3, 12) -- (3.5, 12);\\\\node [right] at (3.5, 12) {Multiple Feedforward Pass: SPM , RPNs };\\n\\\\draw [thick, -] (2, 11) -- (2.5, 11);\\\\node [right] at (2.5, 11) {Deep Feature Selection (Section \\\\ref{Deep_Feature_Selection}) };\\n\\\\draw [thick, -] (3, 10.5) -- (3, 9); \\n\\\\draw [thick, -] (3.0, 10) -- (3.5, 10);\\\\node [right] at (3.5, 10) { Fully-connected Layer: Neural codes };\\n\\\\draw [thick, -] (3.0, 9) -- (3.5, 9);\\\\node [right] at (3.5, 9) { Convolutional Layer: SPoC , CroW };\\n\\\\draw [thick, -] (2.0, 8) -- (2.5, 8);\\\\node [right] at (2.5, 8) {Feature Fusion Strategy (Section \\\\ref{Feature_Fusion_Strategy}) };\\n\\\\draw [thick, -] (3.0, 7.5) -- (3.0, 6);\\n\\\\draw [thick, -] (3.0, 7) -- (3.5, 7);\\\\node [right] at (3.5, 7) { Layer-level Fusion: MoF , MOP };\\n\\\\draw [thick, -] (3.0, 6) -- (3.5, 6);\\\\node [right] at (3.5, 6) { Model-level Fusion: ConvNet fusion };\\n\\\\draw [thick, -] (1, 5) -- (1.5, 5);\\\\node [right] at (1.5, 5) {\\\\bf Feature Embedding and Aggregation (Section \\\\ref{Deep_Feature_Enhancement}) };\\n\\\\draw [thick, -] (2, 4.5) -- (2, -1); \\n\\\\draw [thick, -] (2.0, 4) -- (2.5, 4);\\\\node [right] at (2.5, 4) { Matching with Global Features (Section \\\\ref{Matching_with_Global_Feature})};\\n\\\\draw [thick, -] (2.0, 3) -- (2.5, 3);\\\\node [right] at (2.5, 3) {Matching with Local Features (Section \\\\ref{Matching_with_Local_Feature})};\\n\\\\draw [thick, -] (2.0, 2) -- (2.5, 2);\\\\node [right] at (2.5, 2) { Attention Mechanism (Section \\\\ref{Attention_Mechanism}) };\\n\\\\draw [thick, -] (3.0, 1.5) -- (3.0, 0);\\n\\\\draw [thick, -] (3.0, 1) -- (3.5, 1);\\\\node [right] at (3.5, 1) { Non-parameteric: CroW , SPoC , SWVF };\\n\\\\draw [thick, -] (3.0, 0) -- (3.5, 0);\\\\node [right] at (3.5, 0) { Parameteric: CRN , DeepFixNet+SAM , };\\n\\\\draw [thick, -] (2.0, -1) -- (2.5, -1);\\\\node [right] at (2.5, -1) {Deep Hash Embedding (Section \\\\ref{Deep_Hash_Embedding}) };\\n\\\\draw [thick, -] (3.0, -1.5) -- (3.0, -3);\\n\\\\draw [thick, -] (3.0, -2) -- (3.5, -2);\\\\node [right] at (3.5, -2) { Supervised Hashing: SSDH };\\n\\\\draw [thick, -] (3.0, -3) -- (3.5, -3);\\\\node [right] at (3.5, -3) { Unsupervised Hashing: DeepBit , DSTH };\\n\\\\draw [thick, -] (0, -4) -- (0.5, -4);\\\\node [right] at (0.5, -4) {\\\\bf Fine-Tuning for Learning DCNN Representations (Section \\\\ref{Retrieval_via_Learning_DCNN_Representations})};\\n\\\\draw [thick, -] (1, -4.5) -- (1, -11);\\n\\\\draw [thick, -] (1, -5) -- (1.5, -5);\\\\node [right] at (1.5, -5) {Supervised Fine-tuning (Section \\\\ref{Supervised_Fine-tuning}) };\\n\\\\draw [thick, -] (2, -5.5) -- (2, -7); \\n\\\\draw [thick, -] (2.0, -6) -- (2.5, -6);\\\\node [right] at (2.5, -6) {Fine-tuning via classification loss (Section \\\\ref{Classification-based_Fine-tuning}) }; \\n\\\\draw [thick, -] (2.0, -7) -- (2.5, -7);\\\\node [right] at (2.5, -7) {Fine-tuning via pairwise ranking loss (Section \\\\ref{Verification_based_Learning}) };\\n\\\\draw [thick, -] (3.0, -7.5) -- (3.0, -10);\\n\\\\draw [thick, -] (3.0, -8) -- (3.5, -8);\\\\node [right] at (3.5, -8) { Transformation Matrix: Non-metric };\\n\\\\draw [thick, -] (3.0, -9) -- (3.5, -9);\\\\node [right] at (3.5, -9) { Siamese Networks };\\n\\\\draw [thick, -] (3.0, -10) -- (3.5, -10);\\\\node [right] at (3.5, -10) { Triplet Networks };\\n\\\\draw [thick, -] (1.0, -11) -- (1.5, -11);\\\\node [right] at (1.5, -11) { Unsupervised Fine-tuning (Section \\\\ref{Unsupervised_Fine-tuning})};\\n\\\\draw [thick, -] (2, -11.5) -- (2, -13); \\n\\\\draw [thick, -] (2.0, -12) -- (2.5, -12);\\\\node [right] at (2.5, -12) { Manifold Learning Samples Mining: Diffusion Net };\\n\\\\draw [thick, -] (2.0, -13) -- (2.5, -13);\\\\node [right] at (2.5, -13) {Mining Samples by Clustering: SfM-GeM , };\\n\\\\end{tikzpicture}\\n\\\\vspace{-2em}\\n\\\\caption{This survey is organized around three key themes in instance-level image retrieval, shown in bold.}\\n\\\\label{FourAspectsofprogress}\\n\\\\end{figure}\\nThere are limitations to this approach, most fundamentally that there is a model-transfer or domain-shift challenge between tasks~,,, meaning that models trained for classification do not necessarily extract features well suited to image retrieval. In particular, a classification decision can be successful as long as features remain within classification boundaries, however features from such models may show insufficient capacity for retrieval where feature matching itself is more important than classification. This section will survey the strategies which have been developed to improve the quality of feature representations, particularly based on feature extraction / fusion (Section \\\\ref{Deep_Feature_Extraction}) and feature embedding / aggregation (Section \\\\ref{Deep_Feature_Enhancement}).\",\n \"id\": \"1b3ca748-df26-4b7f-85f3-3eb438a055f5\",\n \"level\": \"section\",\n \"origin_cites_number\": 26,\n \"parent_id\": \"515509a6-319a-4385-b1b5-4418a83d2bc2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ]\n ],\n \"subsections\": [\n \"f1ce1975-10d7-4d60-8e5d-320750dbf247\",\n \"f6eab6de-c628-4a89-bfb0-43b4f87b1151\"\n ],\n \"title\": \"Retrieval with Off-the-Shelf DCNN Models\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{Deep_Feature_Extraction}\\nFeature extraction is about the mechanism by which retrieval features can be extracted from off-the-shelf DCNNs. \\nFor an input image $x$ and a network $f(\\\\cdot; \\\\bm{\\\\theta})$, we denote its features from a convolutional layer as $ \\\\boldsymbol{A} := f_{conv}(x) \\\\in \\\\mathbb{R}^{H \\\\times W \\\\times C}$ with height $H$, width $W$, and channels $C$ while that from a fully-connected layer as $ \\\\boldsymbol{B} := f_{fc}(x) \\\\in \\\\mathbb{R}^{D \\\\times 1}$ with the dimensional $D$.\",\n \"id\": \"f1ce1975-10d7-4d60-8e5d-320750dbf247\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1b3ca748-df26-4b7f-85f3-3eb438a055f5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \" Deep Feature Extraction\"\n ]\n ],\n \"subsections\": [\n \"d9000dca-9ea2-4e50-b3cd-ac34b0c46c16\",\n \"4be11049-ece2-473a-babd-9ac8ce82a05f\",\n \"f5333f99-42b6-4668-a652-9777bc115d47\"\n ],\n \"title\": \" Deep Feature Extraction\"\n },\n {\n \"cite_extract_rate\": 0.518518518518518,\n \"cites\": [\n 7543,\n 2105,\n 8525,\n 7546,\n 2094,\n 2103,\n 2109,\n 2120,\n 2121,\n 2114,\n 2107,\n 2112,\n 209,\n 7545\n ],\n \"content\": \"\\\\label{Network_Feedforward_Scheme}\\nNetwork feedforward schemes focus on how images are fed into a DCNN, which includes single-pass and multiple-pass.\\n\\\\begin{flushleft}\\n\\\\emph{a. Single Feedforward Pass Methods}. \\n\\\\end{flushleft}\\nSingle feedforward pass methods take the whole image and feed it into an off-the-shelf model to extract features. The approach is relatively efficient since the input image is fed only once. For these methods, both the fully-connected layer and last convolutional layer can be used as feature extractors .\\nEarly network-based IIR work focused on leveraging DCNNs as a fixed extractor to obtain global features, especially based on the fully-connected layers ,, requiring close to zero engineering effort. However, extracting features in this way affects retrieval accuracy since the extracted features may include background information or activations for irrelevant objects.\\nThe key to single-pass schemes is to embed and aggregate features to improve their discriminativity, such that features of two related images (\\\\ie including the same object) are more similar than these of two unrelated images . For this purpose, it is possible to first map the features $ \\\\boldsymbol{B} $ into a high-dimensional space and then to aggregate them into a final global feature . Another direction is to treat regions in convolutional features $ \\\\boldsymbol{A} $ as different sub-vectors, such that a combination of sub-vectors of all feature maps are used to represent the input image ,.\\n\\\\begin{flushleft}\\n\\\\emph{b. Multiple Feedforward Pass Methods}. \\n\\\\end{flushleft}\\nCompared to single-pass schemes, multiple-pass methods are more time-consuming because several patches are generated and then fed into the network, \\n\\\\textcolor{black}{\\nmultiple-pass schemes are more helpful for addressing the ``\\\\textit{invariance challenges} and ``\\\\textit{distraction challenges}'' in Section \\\\ref{Keychallenges}. Local patches at multiple scales become more robust for image translation, scaling and rotation ,. Also, these patches are helpful to filter several irrelevant background information. \\n}\\nThe representations are usually produced from two stages: patch detection and patch description. Multi-scale image patches are obtained using sliding windows , or spatial pyramid model (SPM) ,,, as shown in Figure \\\\ref{MultiplePass}. For example, Zheng \\\\etal partition an image by using SPM and extract features at increasing scales, thus enabling the integration of global, regional, local contextual information.\\nPatch detection methods lack retrieval efficiency since irrelevant patches are also detected . For example, Cao \\\\etal propose to merge image patches into larger regions with different hyper-parameters, where the hyper-parameter selection is viewed as an optimization problem to maximize the similarity between query and candidate features.\\n\\\\begin {figure}[!t]\\n\\\\centering\\n { \\n \\\\includegraphics[width=\\\\columnwidth]{./Figures/MultiplePass.pdf} \\n }\\n \\\\vspace{-2em}\\n\\\\caption{Image patch generation schemes: (a) Sliding windows ,; (b) Spatial pyramid modeling ; (c) Dense sampling ,; (d) Region proposals from region proposal networks ,.}\\n\\\\vspace{-1em}\\n\\\\label{MultiplePass}\\n\\\\end {figure}\\nInstead of generating multi-scale image patches randomly or densely, region proposal methods introduce a degree of purpose. Region proposals can be generated using object detectors, such as selective search , edge boxes ,, and BING . For example, Yu \\\\etal propose fuzzy object matching (FOM) for instance search in which the fuzzy objects are generated from 300 object proposals and then clustered to filter out overlapping proposals. Region proposals can also be learned using such as region proposal networks (RPNs) , and convolutional kernel networks (CKNs) , and then to apply these networks into end-to-end fine-tuning for learning similarity . \\\\textcolor{black}{This usually requires the datasets provide well-localized bounding boxes as supervision, \\\\eg the datasets INSTRE , Oxford-5k , Paris-6k , GLD-v2 variant . Also, in the off-the-shelf scenarios, the way that using the bounding boxes to crop the query images and use as input the DCNNs has been shown to provide better retrieval performance since only the information relevant to the instance is extracted ,.\\n}\",\n \"id\": \"d9000dca-9ea2-4e50-b3cd-ac34b0c46c16\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 27,\n \"parent_id\": \"f1ce1975-10d7-4d60-8e5d-320750dbf247\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \" Deep Feature Extraction\"\n ],\n [\n \"subsubsection\",\n \" Network Feedforward Scheme\"\n ]\n ],\n \"subsections\": [],\n \"title\": \" Network Feedforward Scheme\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 7547,\n 7543,\n 2122,\n 2102,\n 2113,\n 2105,\n 2108,\n 2094,\n 2104,\n 2109,\n 2110,\n 2112,\n 7548,\n 7545\n ],\n \"content\": \"\\\\label{Deep_Feature_Selection}\\nFeature selection decides the receptive field of the extracted features, \\\\ie global-level from fully-connected layers and regional-level from convolutional layers.\\n\\\\begin{flushleft}\\n\\\\emph{a. Extracted from Fully-connected Layers}\\n\\\\end{flushleft}\\nIt is straightforward to select a fully-connected layer as a global feature extractor ,,. With PCA dimensionality reduction and normalization~ image similarity can be measured. Extracting features $ \\\\boldsymbol{B} $ from fully-connected layer leads to two obvious limitations for IIR: including irrelevant information, and a lack of local geometric invariance .\\nWith regards to the first limitation, image-level global descriptors may include irrelevant patterns or background clutter, especially when a target instance is only a small portion of an image. It may then be more reasonable to extract region-level features at finer scales, \\\\ie using multiple passes~,,.\\nFor the second limitation, \\nan alternative is to extract multi-scale features on a convolutional layer ,. Further, it makes the global features incompatible with techniques such as spatial verification and re-ranking. Several methods then choose to leverage intermediate convolutional layers ,,,.\\n\\\\begin{flushleft}\\n\\\\emph{b. Extracted from Convolutional Layers}\\n\\\\end{flushleft}\\nThe neurons in a convolutional layer are connected only to a local region of the input image, and this smaller receptive field ensures that the produced features $ \\\\boldsymbol{A} $, usually from the last layer, preserve more local structural information , and are more robust to image transformations \\\\textcolor{black}{ thereby address the ``\\\\textit{invariance challenge}''. For instance, Razavian \\\\etal extract multi-scale features on the last convolutional layer and Mor{\\\\`e}re \\\\etal incorporate a series of nested pooling layers into CNN. Both of them provide higher feature invariance. Thus, many image retrieval methods use convolutional layers as feature extractors~ ,,,.}\\nSum/average and max pooling are two simple aggregation methods to produce global features . For a pooled layer, the last convolutional layer usually yields superior accuracy over other shallower or later fully-connected layers . There is no other operation on the feature maps before pooling, so we illustrate these methods as ``direct pooling'' in Figure~\\\\ref{PipelineofImageRetrieval}.\\n\\\\begin {figure}[!t]\\n\\\\centering\\n { \\n \\\\includegraphics[width=\\\\columnwidth]{./Figures/SinglePass.pdf} \\n }\\n \\\\vspace{-1.5em}\\n\\\\caption{Representative methods in single-pass methods, focusing on convolutional feature tensor $ \\\\boldsymbol{A} $. We denote the entry in $ \\\\boldsymbol{A} $ corresponding to channel $ c $, at spatial location ($i, j$) as $ A_{ijc} $:\\nMAC , R-MAC , SPoC with the per-channel Gaussian weighting $\\\\alpha^{\\\\prime}_{ij} A_{ij}$ where $\\\\alpha^{\\\\prime}_{ij} = \\\\exp\\\\left\\\\{-{\\\\textstyle\\\\frac{\\\\left(i-\\\\frac H2\\\\right)^2+\\\\left(j-\\\\frac W2\\\\right)^2}{2\\\\sigma^2}}\\\\right\\\\} $ , CroW with $ \\\\alpha^{\\\\prime\\\\prime} $ computed by summing all $C$ feature maps at location ($i, j$) and $ \\\\beta $ computed by summing the $H\\\\times W$-array at each feature map $c$ , GeM with channel-wise powers operation , and CAM+CroW by performing $ M_{ij}^{(l)}=\\\\sum_{c=1}^C\\\\omega_{lc}A_{ijc} $ where $\\\\omega_{lc}$ are weights activated by $l$-th class . \\n}\\n\\\\vspace{-1em}\\n\\\\label{SinglePass}\\n\\\\end {figure}\\nInstead of direct pooling, many sophisticated aggregation methods have been explored, such as channel-wise or spatial-wise feature weighting on the convolutional feature maps ,,. These aggregation methods aim to highlight feature importance or reduce the undesirable influence of bursty descriptors of some regions ,. For clarity, we illustrate the representative strategies in Figure \\\\ref{SinglePass}. \\nNote that these feature aggregation methods are usually performed before channel-wise sum/max pooling and does not embed features into a higher dimensional space.\\nOne rationale behind using convolutional features is that each such vector can act as a ``dense SIFT'' feature since each vector corresponds to a region in the input image. Inspired by this perception, many works leverage embedding methods (\\\\eg BoW) used for SIFT features on the regional feature vectors and then aggregate them (\\\\eg by sum pooling) into a global descriptor. \\n\\\\textcolor{black}{Feature embedding methods address the discriminativity challenge via mapping individual features into a high-dimensional space and make them distinguishable .} Feature embedding is followed by PCA to reduce feature dimensionality and whitening to down-weight co-occurrence between features.\",\n \"id\": \"4be11049-ece2-473a-babd-9ac8ce82a05f\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"f1ce1975-10d7-4d60-8e5d-320750dbf247\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \" Deep Feature Extraction\"\n ],\n [\n \"subsubsection\",\n \"Deep Feature Selection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Deep Feature Selection\"\n },\n {\n \"cite_extract_rate\": 0.388888888888888,\n \"cites\": [\n 2124,\n 2115,\n 2123,\n 7548,\n 2125,\n 514,\n 2105\n ],\n \"content\": \"\\\\label{Feature_Fusion_Strategy}\\nFusion studies the complementarity of different features which includes layer-level and model-level fusion explorations.\\n\\\\begin{flushleft}\\n\\\\emph{a. Layer-level Fusion}\\n\\\\end{flushleft}\\nWith layer-level fusion it is possible to fuse multiple fully-connected layers in a deep network ,. For instance, Liu \\\\etal introduce DeepIndex to incorporate multiple global features from different fully connected layers. The activation from the first fully-connected layer is taken as column indexing, and that from the second layer serves as row indexing. Similarly, it is also possible to fuse the activations from multiple convolutional layers. For instance, Li \\\\etal apply the R-MAC encoding scheme on five convolutional layers of VGG-16 and then concatenate them into a multi-scale feature vector.\\nFeatures from fully-connected layers retain global high-level semantics, whereas features from convolutional layers can present local low- and intermediate-level cues. Global and local features therefore complement each other when measuring semantic similarity and can, to some extent, guarantee retrieval performance ,. Such features can be concatenated directly ,, with convolutional features normally filtered by sliding windows or region proposal nets. Direct concatenation can also be replaced by other advanced methods, such as orthogonal operations or pooling-based methods, such as Multi-layer Orderless Fusion (MOF) of Li \\\\etal , which is inspired by Multi-layer Orderless Pooling (MOP) . However local features cannot play a decisive role in distinguishing subtle feature differences if global and local features are treated identically. Yu \\\\etal use a mapping function to assert local features in refining the return ranking lists, via an exponential mapping function for tapping the complementary strengths of convolutional and fully-connected layers. Similarly, \\nLiu \\\\etal design two sub-networks on top of convolutional layers to obtain global and local features and then learn to fuse these features, thereby adaptively adjusting the fusion weights. Instead of directly fusing the layer activations, Zhang \\\\etal fuse the index matrices which are generated based on the two feature types extracted from the same CNN, a feature fusion which has low computational complexity.\\nIt is worth considering which layer combinations are better for fusion given their differences and complementarity. Yu \\\\etal compare the performance of different combinations between fully-connected and convolutional layers on the Oxford 5k, Holiday, and UKBench datasets. The results show that the combinations including the first fully-connected layer always perform better. Li \\\\etal demonstrate that fusing convolutional and fully-connected layers outperforms the fusion of only convolutional layers. Fusing two convolutional layers with one fully-connected layer achieves the best performance on the Holiday and UKBench datasets.\\n\\\\begin{flushleft}\\n\\\\emph{b. Model-level Fusion}\\n\\\\end{flushleft}\\nIt is possible to combine features from different models; such a fusion focuses more on model complementarity, with methods categorized into \\\\emph{intra-model} and \\\\emph{inter-model}.\\nIntra-model fusion suggests multiple deep models having similar or highly compatible structures, while inter-model fusion involves models with differing structures. For example, \\nSimonyan \\\\etal introduce a ConvNet intra-model fusion strategy to improve the feature learning capacity of VGG where VGG-16 and VGG-19 are fused. \\nTo attend to different parts of an image object, Wang \\\\etal realize the multi-feature fusion by selecting all convolutional layers of VGG-16 to extract image representations, which is demonstrated to be more robust than using only single-layer features.\\nInter-model fusion is a way to bridge different features given the fact that different deep networks have different receptive fields ,,,. For instance, a two-stream attention network is introduced to implement image retrieval where the main network for semantic prediction is VGG-16 while an auxiliary network is used for predicting attention maps. Similarly, considering the importance and necessity of inter-model fusion to bridge the gap between mid-level and high-level features, Liu \\\\etal and Zheng \\\\etal combine VGG-19 and AlexNet to learn combined features, while Ozaki \\\\etal concatenate descriptors from six different models. \\nInter-model and intra-model fusion are relevant to model selection. There are some strategies to determine how to combine the features from two models. It is straightforward to fuse all features from the candidate models and then learn a metric based on the concatenated features ,, which is a kind of ``\\\\emph{early fusion}'' strategy. Alternatively, it is also possible to learn optimal metrics separately for the features from each model, and then to combine these metrics for final retrieval ranking ,, which is a kind of ``\\\\emph{late fusion}'' strategy.\\n\\\\textbf{Discussion.} Layer-level fusion and model-level fusion are conditioned on the fact that the associated layers or networks have different feature description capacities. For these fusion strategies, the key question is \\\\textit{what features are the best to be combined?} Some explorations have been made on the basis of off-the-shelf models, such as Xuan \\\\etal , who illustrates the effect of combining different numbers of features and different sizes within the ensemble. Chen \\\\etal analyze the performance of embedded features from off-the-shelf image classification and object detection models with respect to image retrieval.\",\n \"id\": \"f5333f99-42b6-4668-a652-9777bc115d47\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"f1ce1975-10d7-4d60-8e5d-320750dbf247\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \" Deep Feature Extraction\"\n ],\n [\n \"subsubsection\",\n \"Feature Fusion Strategies\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Feature Fusion Strategies\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{Deep_Feature_Enhancement}\\n\\\\textcolor{black}{The primary aim of feature embedding and aggregation is to further promote feature discriminativity, targeting for the ``\\\\textit{discriminativity challenge}'', and obtain final global and/or local features for retrieving specific instances.}\",\n \"id\": \"f6eab6de-c628-4a89-bfb0-43b4f87b1151\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"1b3ca748-df26-4b7f-85f3-3eb438a055f5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \"Feature Embedding and Aggregation\"\n ]\n ],\n \"subsections\": [\n \"96cef0e9-d2bb-419e-8aa3-3757a002290b\",\n \"22557b43-bfa7-48d0-9dad-fad3f1f25551\",\n \"3362dbb1-5a27-4b56-85a4-413bc5c384f6\",\n \"b5eed66c-a402-4570-be6b-8b9987293fc6\"\n ],\n \"title\": \"Feature Embedding and Aggregation\"\n },\n {\n \"cite_extract_rate\": 0.448275862068965,\n \"cites\": [\n 7543,\n 2122,\n 2102,\n 2113,\n 2105,\n 8525,\n 2108,\n 2109,\n 2097,\n 2123,\n 2126,\n 7548,\n 7545\n ],\n \"content\": \"\\\\label{Matching_with_Global_Feature}\\nGlobal features can be extracted from fully-connected layers, followed by dimensionality reduction and normalization ,. They are easy to implement and there is no further aggregation process. Gong \\\\etal extract fully-connected activations for local image patches at three scale levels and embed patch-level activations individually using VLAD. Thus, the final concatenated features significantly tackle the invariance challenge caused by image rotations.\\nConvolutional features can also be aggregated into compact a global feature. Simple aggregation methods are sum/average or max pooling ,. \\nSum/average pooling is less discriminative, because it takes into account all activated convolutional outputs, thereby weakening the effect of highly activated features . As a result, max pooling is particularly well suited for sparse features having a low probability of being active, however max pooling may be inferior to sum/average pooling when image features are whitened .\\nFigure \\\\ref{SinglePass} illustrates sophisticated feature aggregation methods using channel-wise or spatial-wise weighting ,. For example, Babenko \\\\etal propose sum-pooling convolutional features (SPoC) to obtain compact descriptors weighted by $ \\\\alpha^{\\\\prime} $ with a Gaussian center prior. Similarly, \\nit is possible to treat regions in feature maps as different sub-vectors ,,, thus combinations of $R$ sub-vectors are used to represent the input image, such as R-MAC . Since convolutional features may include repetitive patterns and each vector may correspond to identical regions, the resulting descriptors may be bursty, which makes the final aggregated global feature less distinguishable. As a solution, Pang \\\\etal leverage heat diffusion to weigh convolutional features at the aggregation stage, and reduce the undesirable influence of burstiness. \\nConvolutional features have an interpretation as descriptors of local regions, thus many works leverage embedding methods, including BoW, VLAD, and FV, to encode regional feature vectors and then aggregate them into a global descriptor. Note that BoW and VLAD can be extended by using other metrics, such as a Hamming distance . Here we briefly describe the principle of Euclidean embeddings.\\nBoW is a widely used feature embedding which leads to a sparse vector of occurrence. Let $ \\\\boldsymbol{a} = \\\\left\\\\{a_{1}, a_{2},...,a_{R} \\\\right\\\\} $ be a set of $R$ local features, each of dimensionality $d$. BoW requires a pre-defined codebook $ \\\\boldsymbol{c} = \\\\left\\\\{c_{1}, c_{2},...,c_{K} \\\\right\\\\} $ with $ K $ centroids, usually learned offline, to cluster these local descriptors, and maps each descriptor $a_{t}$ to the nearest centroid $c_{k}$. For each centroid, one can count and normalize the number of occurrences as\\n\\\\begin{equation}\\ng(c_{k}) = \\\\frac{1}{R} \\\\sum_{r=1}^{R}\\\\phi(a_{r}, c_{k})\\n\\\\end{equation}\\n\\\\begin{equation}\\n\\\\phi(a_{r}, c_{k}) = \\\\left\\\\{ \\\\begin{array}{ll}\\n1 & \\\\textrm{if $ c_{k} $ is the closest codeword for $a_{r}$ }\\\\\\\\\\n0 & \\\\textrm{otherwise}\\\\\\\\\\n\\\\end{array} \\\\right.\\n\\\\label{eqn-phi}\\n\\\\end{equation}\\nThus BoW considers the number of descriptors belonging to each $c_{k}$ (\\\\emph{i.e.} 0-order feature statistics), so the BoW representation is the concatenation of all mapped vectors:\\n\\\\begin{equation}\\n\\\\label{BoW}\\n G_{_{BoW}}(\\\\boldsymbol{a}) =\\n \\\\left[\\n \\\\begin{array}{ccc}\\n g(c_{1}), \\\\cdots ,g(c_{k}) ,\\\\cdots, g(c_{K})\\n \\\\end{array}\\n \\\\right] \\\\rm ^{\\\\top}\\n\\\\end{equation}\\nBoW is simple to implement the encoding of local descriptors, such as convolutional feature maps , or fully-connected activations ,, or to encode regional descriptors ,. Mukherjee \\\\etal extract image patches based on information entropy and feed into a pre-trained VGG-16, then use BoW to embed and aggregate the patch-level descriptors from a fully-connected layer. Embedded BoW vectors are typically high-dimensional and sparse, so not well suited to large-scale datasets in terms of the mentioned efficiency challenge.\\nVLAD stores the sum of residuals for each visual word. Similar to BoW, it generates $K$ visual word centroids, then each feature $a_{r}$ is assigned to its nearest visual centroid $ c_{k} $:\\n\\\\begin{equation}\\ng(c_{k}) = \\\\frac{1}{R} \\\\sum_{r=1}^{R}\\\\phi(a_{r} , c_{k})(a_{r} - c_{k})\\n\\\\end{equation}\\nThe VLAD representation is stacked by the residuals for all centroids, with dimension ($ d \\\\times K $), \\\\ie \\n\\\\begin{equation}\\nG_{_{VLAD}}(\\\\boldsymbol{a}) \\\\! = \\\\! \\\\left[\\n \\\\begin{array}{ccc}\\n \\\\cdots, g(c_{k}) \\\\rm ^{\\\\top},\\\\cdots \\n \\\\end{array}\\n \\\\right] \\\\rm ^{\\\\top} .\\n\\\\end{equation}\\nVLAD captures first-order feature statistics, \\\\ie ($a_{r} - c_{k}$). Similar to BoW, the performance of VLAD is affected by the number of clusters: more centroids produce larger vectors that are harder to index. For instance-level image retrieval, \\nGong \\\\etal concatenate the activations of a fully-connected layer with VLAD applied to image-level and patch-level inputs . Ng \\\\etal replace BoW~ with VLAD~, and are the first to encode local features into VLAD representations. This idea inspired another milestone work where, for the first time, VLAD is plugged into the last convolutional layer, which allows end-to-end training via back-propagation.\\nFV extends BoW by encoding the first and second order statistics. FV clusters the set of local descriptors by a Gaussian Mixture Model (GMM) with $K$ components to generate a dictionary $ \\\\boldsymbol{c} = \\\\left\\\\{ \\\\mu_{k}; \\\\Sigma_{k}; w_{k}\\\\right\\\\}_{k=1}^{K} $ made up of mean / covariance / weight triples , where the covariance may be simplified by keeping only its diagonal elements. For each local feature $ a_{r} $, a GMM is given by \\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\gamma_{k}(a_{r}) = w_{k}\\\\times p_{k}(a_{r})/\\\\Big(\\\\sum_{k=1}^{K}w_{k}p_{k}(a_{r})\\\\Big) \\\\quad s.t. \\\\sum_{k=1}^K w_{k} = 1\\n\\\\end{aligned}\\n\\\\end{equation}\\nwhere $ p_{k}(a_{r})=\\\\mathcal{N}(a_{r}, \\\\mu_{k}, \\\\sigma_{k}^2 )$. All local features are assigned into each component $k$ in the dictionary, which is computed as\\n\\\\begin{equation}\\n\\\\begin{aligned}\\n &g_{w_{k}} = \\\\frac{1}{R \\\\sqrt{w_{k}}} \\\\sum_{r=1}^{R} \\\\Big(\\\\gamma_{k}(a_{r}) - w_{k}\\\\Big) \\\\\\\\\\n& g_{u_{k}} = \\\\frac{\\\\gamma_{k}(a_{r})}{R \\\\sqrt{w_{k}}} \\\\sum_{r=1}^{R} \\\\left( \\\\frac{a_{r} - \\\\mu_{k}}{\\\\sigma_{k}} \\\\right),\\\\\\\\ \\n& g_{\\\\sigma_{k}^2} = \\\\frac{\\\\gamma_{k}(a_{r})}{R \\\\sqrt{2w_{k}}} \\\\sum_{r=1}^{R}\\\\left[ {\\\\left( \\\\frac{a_{r} - \\\\mu_{k}}{\\\\sigma_{k}} \\\\right)}^{2} - 1 \\\\right]\\n\\\\end{aligned}\\n\\\\end{equation}\\nThe FV representation is produced by concatenating vectors from the $K$ components: \\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\!\\\\!\\\\! G_{_{FV}}(\\\\boldsymbol{a}) \\\\! = \\\\! \\\\left[\\n \\\\begin{array}{ccc}\\n g_{w_{1}}, \\\\cdots , g_{w_{K}}, g_{u_{1}}, \\\\cdots, g_{u_{K}}, g_{\\\\sigma_{1}^2}, \\\\cdots, g_{\\\\sigma_{K}^2} \\n \\\\end{array}\\n \\\\right] \\\\rm ^{\\\\top}\\n\\\\end{aligned}\\n\\\\end{equation}\\nThe FV representation defines a kernel from a generative process and captures more statistics than BoW and VLAD. FV vectors do not increase the computational cost significantly but require more memory. Applying FV without memory controls may lead to suboptimal performance . \\n\\\\textbf{Discussion.} Traditionally, pooling-based aggregation methods (\\\\eg in Figure \\\\ref{SinglePass}) are directly plugged into deep networks and then the whole model is used end-to-end. The three embedding methods (BoW, VLAD, FV) are initially trained with large pre-defined vocabularies ,. One needs to pay attention on their properties before choosing an embedding: BoW and VLAD are computed in the rigid Euclidean space where performance is closely related to the number of centroids, whereas FV can capture higher-order statistics and improves the effectiveness of feature embedding at the expense of a higher memory cost. Further, although vocabularies are usually built separately and pre-trained before encoding deep features, it is necessary to integrate the training of networks and the learning of vocabulary parameters into a unified framework so as to guarantee training and testing efficiency. For example, VLAD is integrated into deep networks where each spatial column feature is used to construct clusters via k-means . This idea led to NetVLAD , where deep networks are fine-tuned with the VLAD vectors. The FV method is also combined with deep networks for retrieval tasks ,.\",\n \"id\": \"96cef0e9-d2bb-419e-8aa3-3757a002290b\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 29,\n \"parent_id\": \"f6eab6de-c628-4a89-bfb0-43b4f87b1151\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \"Feature Embedding and Aggregation\"\n ],\n [\n \"subsubsection\",\n \"Matching with Global Features\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Matching with Global Features\"\n },\n {\n \"cite_extract_rate\": 0.6000000000000001,\n \"cites\": [\n 7543,\n 2093,\n 2101,\n 7546,\n 8525,\n 2127,\n 2103,\n 8524,\n 2121,\n 2106,\n 2114,\n 2126,\n 2110,\n 209,\n 7545\n ],\n \"content\": \"\\\\label{Matching_with_Local_Feature}\\nAlthough matching with global features has high efficiency for both feature extraction and similarity computation, global features are not compatible with spatial verification and correspondence estimation, which are important procedures for instance-level retrieval tasks, motivating work on matching with local features. In terms of the matching process, global features are matched only once while local feature matching is evaluated by summarizing the similarity across all individual local features (\\\\ie many-to-many matching).\\nOne important aspect of local features is to detect the keypoints for an instance within an image, and then to describe the detected keypoints as a set of local descriptors. Inspired by , the common strategies of this whole procedure for IIR can be categorized as \\\\textit{detect-then-describe} and \\\\textit{describe-then-detect}.\\nIn terms of \\\\textit{detect-then-describe}, we regard the descriptors around keypoints as local features, similar to ,. Coarse regions can be detected, for example, by using the methods depicted in Figure~\\\\ref{MultiplePass}, and regions of interest in an image can be detected by using region proposal networks (RPNs) ,. The extracted coarse regions around the keypoints are fed into a DCNN, followed by feature description. Traditional detectors can also be used to detect fine regions around a keypoint. For instance, Zheng \\\\etal employ the popular Hessian-Affine detector to get an affine-invariant local region. Paulin \\\\etal and Mishchuk \\\\etal detect regions using the Hessian-Affine detector and feed into patch-convolutional kernel networks (Patch-CKNs) . \\\\textcolor{black}{Note that it becomes more convenient for the case where bounding boxes annotations have been provided by datasets (see Section \\\\ref{Datasets_and_Evaluation_Criteria}), and then the image regions can be cropped directly for further reranking . }\\nRather than performing keypoint detection early on, it is possible to postpone the detection stage on the convolutional feature maps, \\\\ie \\\\textit{describe-then-detect}. One can select regions on the convolutional feature maps to obtain a set of local features ,,; the local maxima of the feature maps are then detected as keypoints . A similar strategy is also used in network fine-tuning ,,,,, where the keypoints on the convolutional feature maps can be selected based on attention scores predicted by an attention network ,, or based on single-head and multi-head attention modules in transformers ,. This approach to keypoint selection is better for achieving computational efficiency. \\nAfter keypoint detection and description, a large number of local features are used in the matching stage to perform instance-level retrieval, and the image similarity is evaluated by matching across all local features. Local matching techniques include spatial verification and selective match kernels (SMK) . Spatial verification assumes object instances are rigid so that local matches between images can be estimated as an affine transformation using RANdom SAmple Consensus (RANSAC) . One limitation of RANSAC is its high computational complexity of estimating the transformation model when all local descriptors are considered; instead, it is possible to apply RANSAC to a small number of top-ranked local descriptors, such as those selected by approximate nearest neighbor . \\nSMK weighs the contributions of individual matches with a non-linear selective function, but is still memory intensive. Its extension, the Aggregated Selective Match Kernel (ASMK), focuses more on aggregating similarities between local features without explicitly modeling the geometric alignment, which can produce a more compact representation ,. Recently, Teichmann \\\\etal introduced Regional Aggregated Selective Match Kernel (R-ASMK) to combine information from detected regions, boosting image retrieval accuracy compared to the ASMK.\\n\\\\textbf{Discussion.} Using local descriptors to perform instance retrieval tasks has two limitations. First, the local descriptors for an image are stored individually and independently, which is memory-intensive, and not well-suited for large-scale scenarios.\\nSecond, estimating the similarity between the query and database images depends on cross-matching all local descriptor pairs, which incurs additional searching cost and then a low retrieval efficiency. Therefore, most instance retrieval systems using local features follow a two-stage paradigm: initial filtering and re-ranking ,,,,, as in Figure \\\\ref{PipelineofImageRetrieval}. The initial filtering stage is to employ a global descriptor to select a set of candidate matching images, thereby reducing the solution space; the re-ranking stage is to use local descriptors to re-rank the top-ranked images from the global descriptor.\",\n \"id\": \"22557b43-bfa7-48d0-9dad-fad3f1f25551\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 25,\n \"parent_id\": \"f6eab6de-c628-4a89-bfb0-43b4f87b1151\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \"Feature Embedding and Aggregation\"\n ],\n [\n \"subsubsection\",\n \"Matching with Local Features\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Matching with Local Features\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 2124,\n 7547,\n 2094,\n 2104,\n 2093,\n 2118,\n 2099\n ],\n \"content\": \"\\\\label{Attention_Mechanism}\\nAttention mechanism can be regarded as a kind of feature aggregation, whose aim is to \\\\textcolor{black}{highlight the most relevant feature parts. It can effectively address the ``\\\\textit{distraction challenge}'' and also promote feature discriminativity , realized by computing an attention map.} Approaches to obtaining attention maps can be categorized into non-parametric and parametric groups, as shown in Figure~\\\\ref{fig3}, where the main difference is whether the importance weights in the attention map are learnable.\\nNon-parametric weighting is a straightforward method to highlight feature importance, and the corresponding attention maps can be obtained by channel-wise or spatial-wise pooling, as in Figure \\\\ref{fig3} (a,b). For spatial-wise pooling, \\nKalantidis \\\\etal propose an effective CroW method to weight and pool feature maps, which concentrate on weighting activations at different spatial locations, without considering the relations between these activations. In contrast, Ng \\\\etal explore the correlations among activations at different spatial locations on the convolutional feature maps. \\nChannel-wise weighting methods are also popular non-parametric attention mechanisms ,. Xu \\\\etal rank the weighted feature maps to build ``probabilistic proposals'' to select regional features. Jimenez \\\\etal combine CroW and R-MAC to propose Classes Activation Maps (CAM) to weigh the feature map per class. Xiang \\\\etal employ a Gram matrix to analyze the correlations between different channels and then obtain channel sensitivity information to tune the importance of each feature map. Channel-wise and spatial-wise weighting methods are usually integrated into a deep model to highlight feature importance ,.\\nParametric attention maps, shown in Figure \\\\ref{fig3} (c,d), can be learned via deep networks, where the input can be either image patches or feature maps ,,, approaches which are commonly used in supervised metric learning . Kim \\\\etal make the first attempt to propose a shallow network (CRN) to take as input the feature maps of convolutional layers and outputs a weighted mask indicating the importance of spatial regions in the feature maps. The resulting mask modulates feature aggregation to create a global representation of the input image. \\nNoh \\\\etal design a 2-layer CNN with a softplus output layer to compute scores which indicate the importance of different image regions. Inspired by R-MAC, Kim \\\\etal employ a pre-trained ResNet101 to train a context-aware attention network using multi-scale feature maps.\\nApart from using feature maps as inputs, a whole image can be used to learn feature importance, for which specific networks are needed ,,. Mohedano explores different saliency models, including DeepFixNet and Saliency Attentive Model. Yang \\\\etal and Wei \\\\etal introduce a two-stream network for image retrieval in which the auxiliary stream, DeepFixNet, is used specifically for predicting attention maps, which are then fused with the feature maps produced by the main network.\\nFor image retrieval, attention mechanisms can be combined with supervised metric learning .\",\n \"id\": \"3362dbb1-5a27-4b56-85a4-413bc5c384f6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"f6eab6de-c628-4a89-bfb0-43b4f87b1151\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \"Feature Embedding and Aggregation\"\n ],\n [\n \"subsubsection\",\n \"Attention Mechanism\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Attention Mechanism\"\n },\n {\n \"cite_extract_rate\": 0.45,\n \"cites\": [\n 7547,\n 2118,\n 2094,\n 2104,\n 2099,\n 2116,\n 2128,\n 2095,\n 2110\n ],\n \"content\": \"\\\\label{Deep_Hash_Embedding}\\nReal-valued features extracted by deep networks are typically high-dimensional, and therefore are not well-suited to retrieval efficiency. As a result, there is significant motivation to transform deep features into more compact codes. \\\\textcolor{black}{Since their computational and storage efficiency are beneficial for the ``\\\\textit{efficiency challenge}'', hashing algorithms have been widely used for global , and local descriptors ,,.}\\n\\\\begin{figure}[!t]\\n\\\\centering\\n\\\\includegraphics[width=0.5\\\\textwidth]{./Figures/AttentionMachanism.pdf}\\n\\\\vspace{-1.5em}\\n\\\\caption{Illustration of attention mechanisms. (a)-(b) Non-parametric schemes: The attention is based on convolutional feature $ \\\\boldsymbol{A} $. Channel-wise attention in (a) produces a $C$-dimensional importance vector \\\\textbf{$\\\\beta_{1}$} ,; Spatial-wise attention in (b) computes a 2-dimensional attention map \\\\textbf{$\\\\alpha$} ,,. (c)-(d) Parametric schemes: The attention weights are learned by a trainable network. In (c), \\\\textbf{$\\\\beta_2$} are provided by a sub-network with parameters $\\\\theta_{\\\\gamma}$ ,,,. In (d), the attention maps, as a tensor, are predicted by some auxiliary saliency extraction models from the input image directly ,,.} \\\\label{fig3}\\n\\\\end{figure}\\nHash functions can be plugged as a layer into deep networks, so that hash codes and deep networks can be simultaneously trained and optimized, either supervised or unsupervised . During hash function training, the hash codes of originally similar images are embedded as closely as possible, and the hash codes of dissimilar images are as separated as possible. $d$-dim hash codes from a hash function $h(\\\\cdot)$ for an image $x$ can be formulated as $ b_{x} = h(x) = h\\\\big(f(x;\\\\bm{\\\\theta})\\\\big) \\\\in \\\\{+1, -1\\\\}^d $. Because hash codes are non-differentiable their optimization is difficult, so $h(\\\\cdot)$ can be relaxed to be differentiable by using \\\\textit{tanh} or \\\\textit{sigmoid} functions .\\nWhen binarizing real-valued features, it is crucial to preserve image similarity and to improve hash code quality . These two aspects are at the heart of hashing algorithms to maximize retrieval accuracy. \\n\\\\begin{flushleft}\\n\\\\emph{a. Hash Functions to Preserve Image Similarity}\\n\\\\end{flushleft}\\nPreserving similarity seeks to minimize the inconsistencies between real-valued features and corresponding hash codes, for which a variety of strategies have been adopted. \\nLoss functions can significantly influence similarity preservation, which includes both supervised and unsupervised methods. With class labels available, many loss functions are designed to learn hash codes in a Hamming space. As a straightforward method, one can optimize either the difference between matrices computed from the binary codes and their supervision labels , or the difference between the hash codes and real-valued deep features ,. Song \\\\etal propose to learn hash codes for regional features in which each local feature is converted to a set of binary codes by multiplying a hash function and the raw RoI features, then the differences between RoI features and hash codes are characterized by an L$_2$ loss. Do \\\\etal regularize hash codes with a reconstruction loss, which ensure that codes can be reconstructed to their inputs so that similar/dissimilar inputs are mapped to similar/dissimilar hash codes. Lin \\\\etal learn hash codes and address the ``\\\\textit{invariance challenge}'' by introducing an objective function which characterize the difference between the binary codes which are computed from the original image and the geometric transformed one.\\n\\\\begin{flushleft}\\n\\\\emph{b. Improving Hash Function Quality}\\n\\\\end{flushleft}\\nA good hash function seeks to have binary codes uniformly distributed; that is, maximally filling and using the hash code space, normally on the basis of bit uncorrelation and bit balance ,. Bit uncorrelation implies that different bits are as independent as possible, so that a given set of bits can aggregate more information within a given code length . Bit balance means that each bit should have a 50\\\\% chance of being +1 or -1, thereby maximizing code variance and information . Mor{\\\\`e}re \\\\etal use the uniform distribution $U$(0,1) to build a regularization term to make hash codes distribute evenly where the codes are learned by a Restricted Boltzmann Machine layer. Likewise, Lin \\\\etal optimize the mean of learned hash codes to be close to 0.5 to prevent any bit bias towards zero or one.\",\n \"id\": \"b5eed66c-a402-4570-be6b-8b9987293fc6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"f6eab6de-c628-4a89-bfb0-43b4f87b1151\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Retrieval with Off-the-Shelf DCNN Models\"\n ],\n [\n \"subsection\",\n \"Feature Embedding and Aggregation\"\n ],\n [\n \"subsubsection\",\n \" Hashing Embedding \"\n ]\n ],\n \"subsections\": [],\n \"title\": \" Hashing Embedding \"\n },\n {\n \"cite_extract_rate\": 0.6923076923076921,\n \"cites\": [\n 2117,\n 7546,\n 2110,\n 2103,\n 2129,\n 2096,\n 629,\n 2109,\n 2097\n ],\n \"content\": \"\\\\label{Retrieval_via_Learning_DCNN_Representations}\\nThe off-the-shelf DCNNs pre-trained on source datasets for classification are quite robust to inter-class variability. However, in most cases, deep features extracted based on off-the-shelf models may not be sufficient for accurate retrieval, even with the strategies discussed in Section \\\\ref{Retrieval_with_Off_the_Shelf_DCNN_Models}. In order for models to be more effective for retrieval, a common practice is network fine-tuning, \\\\ie updating the pre-stored parameters . Fine-tuning methods have been studied extensively to learn better features, \\\\textcolor{black}{whose primary aim is to address the ``\\\\textit{fine-tune challenge}''.} A standard dataset with clear and well-defined ground-truth labels is indispensable for the supervised fine-tuning and subsequently pair-wise supervisory information is incorporated into ranking loss to update networks by regularizing on retrieval representations, otherwise it is necessary to develop unsupervised fine-tuned methods. After network fine-tuning, features can be organized as global or local to perform retrieval. \\n\\\\textcolor{black}{For the most feature strategies we presented in Section \\\\ref{Retrieval_with_Off_the_Shelf_DCNN_Models}, including feature extraction, feature embedding and feature aggregation. Note that fine-tuning does not contradict or render irrelevant these feature processing methods; indeed, these strategies are complementary and can be equivalently incorporated as part of network fine-tuning. To this end, this section will survey the strategies which have been developed, based on the patch-level, image-level, or class-level supervision, to fine-tune deep networks for better instance retrieval.}\\n\\\\begin{figure*}[!t]\\n\\\\centering\\n\\\\includegraphics[width=0.9\\\\textwidth]{./Figures/All_Metric_Learning.pdf}\\n\\\\vspace{-0.5em}\\n\\\\caption{Schemes of supervised fine-tuning. Anchor, positive, and negative images are indicated by $x_{a}$, $x_{p}$, $x_{n}$, respectively. (a) classification loss ; (b) similarity learning by using a transformation matrix ; (c) Siamese loss ,,,; (d) triplet loss ; (e) an attention block into DCNNs to highlight regions ; (f) combining classification loss and pairwise ranking loss ,; (g) region proposal networks (RPNs) to locate the RoI and highlight specific regions or instances ; (h) inserting the RPNs of (g) into DCNNs, such that the RPNs extract regions or instances at the convolutional layer ,. }\\n\\\\vspace{-1em}\\n\\\\label{All_Metric_Learning}\\n\\\\end{figure*}\",\n \"id\": \"acffc1fe-bf03-450b-857b-db143905236e\",\n \"level\": \"section\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"515509a6-319a-4385-b1b5-4418a83d2bc2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ]\n ],\n \"subsections\": [\n \"45dba903-566e-4f9f-a48c-8f8adf0593ca\",\n \"871a2087-f4a4-4968-91d1-0752906a7612\"\n ],\n \"title\": \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{Supervised_Fine-tuning}\\nThe way to realize supervised fine-tuning can be determined by the given class labels or pairwise supervisory information.\",\n \"id\": \"45dba903-566e-4f9f-a48c-8f8adf0593ca\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"acffc1fe-bf03-450b-857b-db143905236e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ],\n [\n \"subsection\",\n \"Supervised Fine-tuning \"\n ]\n ],\n \"subsections\": [\n \"c05b41ca-005d-493d-96fa-98315460aeed\",\n \"db1c70e4-39b7-4baf-8fa7-4ccb702563d0\",\n \"14afec8b-a454-4ba0-a24a-2c2e37482b34\"\n ],\n \"title\": \"Supervised Fine-tuning \"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 2114,\n 309,\n 2130,\n 2093,\n 2111,\n 2109\n ],\n \"content\": \"\\\\label{Classification-based_Fine-tuning}\\nWhen class labels of a new dataset are available (\\\\eg INSTRE , GLDv2 ,), it is preferable to begin with a previously-trained DCNN, trained on a separate dataset, with the backbone DCNN typically chosen from one of AlexNet, VGG, GoogLeNet, or ResNet.\\nThe DCNN can then be fine-tuned, as shown in Figure \\\\ref{All_Metric_Learning} (a), by optimizing its parameters on the basis of a cross entropy loss \\n\\\\begin{equation}\\n\\\\mathcal{L}_{CE}(\\\\hat{p_{i}},y_{i})= - \\\\! \\\\sum^{c}_{i} \\\\! \\\\big(y_{i} \\\\! \\\\times \\\\!log(\\\\hat{p}_{i}) \\\\big)\\n\\\\label{cross_entropy_loss}\\n\\\\end{equation}\\nHere $y_{i}$ and $\\\\hat{p}_{i}$ are the ground-truth labels and the predicted logits, respectively, and $c$ is the total number of categories. The milestone work in such fine-tuning is , in which AlexNet is re-trained on the Landmarks dataset. \\\\textcolor{black}{\\nAccording to the class labels, the image-level features are required to compute the logits. Thus, the descriptors extracted from local regions on convolutional feature maps , or image patch inputs are further needed to be aggregated. }\\n\\\\textcolor{black}{A classification-based fine-tuning method enables to enforce higher similarity for intra-class samples and diversity for inter-class samples. Cao \\\\etal employ the ArcFace loss , which uses the margin-adjusted cosine similarity in the form of softmax loss, to induce smaller intra-class variance and show excellent results for instance retrieval.} Recently, Boudiaf \\\\etal claim that cross entropy loss can minimize intra-class distances while maximizing inter-class distances. Cross entropy loss is, in essence, maximizing a common mutual information between the retrieval features and the ground-truth labels. Therefore, it can be regarded as an upper bound on a new pairwise loss, which has a structure similar to various pairwise ranking losses, of which representatives are introduced below.\",\n \"id\": \"c05b41ca-005d-493d-96fa-98315460aeed\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"45dba903-566e-4f9f-a48c-8f8adf0593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ],\n [\n \"subsection\",\n \"Supervised Fine-tuning \"\n ],\n [\n \"subsubsection\",\n \"Fine-tuning via Classification Loss\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Fine-tuning via Classification Loss\"\n },\n {\n \"cite_extract_rate\": 0.631578947368421,\n \"cites\": [\n 2101,\n 2096,\n 2098,\n 7546,\n 7549,\n 2103,\n 2097,\n 2106,\n 2129,\n 2110,\n 209,\n 2117\n ],\n \"content\": \"\\\\label{Verification_based_Learning}\\nWith affinity information (\\\\eg samples from the same group) indicating similar and dissimilar pairs, fine-tuning methods based on pairwise ranking loss learn an optimal metric which minimizes or maximizes the distance of pairs to maintain their similarity. Network fine-tuning via ranking loss involves two types of information :\\n\\\\begin{enumerate}\\n\\\\item A pair-wise constraint, corresponding to a Siamese network as in Figure \\\\ref{All_Metric_Learning} (c), in which input images are paired with either a positive or negative sample;\\n\\\\item A triplet constraint, associated with triplet networks as in Figure \\\\ref{All_Metric_Learning} (e), in which anchor images are paired with both similar and dissimilar samples .\\n\\\\end{enumerate}\\nThese pairwise ranking loss based methods are categorized into globally supervised approaches (Figure \\\\ref{All_Metric_Learning} (c,d)) and locally supervised approaches (Figure \\\\ref{All_Metric_Learning} (g,h)), where the former ones learn a metric on global features by satisfying all constraints, whereas the latter ones focus on local areas by only satisfying the given local constraints (\\\\emph{e.g.} region proposals).\\nTo be specific, consider a triplet set $ X \\\\!\\\\! = \\\\!\\\\! \\\\lbrace(x_{a}, x_{p}, x_{n})\\\\} $ in a mini-batch, where $ (x_{a}, x_{p} )$ indicates a similar pair and $ (x_{a}, x_{n} )$ a dissimilar pair. Features $ f(x;\\\\bm{\\\\theta}) $ of one image are extracted by a network $f(\\\\cdot)$ with parameters $\\\\bm{\\\\theta}$, for which we can represent the affinity information for each similar or dissimilar pair as\\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\mathcal{D}_{ij} = \\\\mathcal{D}(x_{i}, x_{j}) = || f(x_{i}; \\\\bm{\\\\theta}) - f(x_{j}; \\\\bm{\\\\theta})||_{2}^{2}\\n\\\\end{aligned}\\\\label{Distance_pre_defined}\\n\\\\end{equation}\\n\\\\begin{flushleft}\\n\\\\emph{a. Refining with Transformation Matrix}. \\n\\\\end{flushleft}\\nLearning the similarity among input samples can be implemented by optimizing the weights of a linear transformation matrix . It transforms the concatenated feature pairs into a common latent space using a transformation matrix $ \\\\bm{W} \\\\!\\\\! \\\\in \\\\!\\\\! \\\\mathbb{R}^{2d \\\\times 1} $, where $d$ is the final feature dimension. The similarity score of these pairs are predicted via a sub-network $ \\\\mathcal{S}_{W}(x_{i}, x_{j}) = f_{W}(f(x_{i}; \\\\bm{\\\\theta})\\\\cup f(x_{j};\\\\bm{\\\\theta}); \\\\bm{W}) $ . In other words, the sub-network $f_{W}$ predicts how similar the feature pairs are. Given the affinity information of feature pairs $ \\\\mathcal{S}_{ij}=\\\\mathcal{S}(x_{i}, x_{j}) \\\\! \\\\in \\\\! \\\\{0,1\\\\} $, the binary labels 0 and 1 indicate the similar (positive) or dissimilar (negative) pairs, respectively. The training of function $f_{W}$ can be achieved by using a regression loss: \\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\!\\\\!\\\\!\\\\mathcal{L}_{W}(x_{i}, x_{j}) = & | \\\\mathcal{S}_{W}(x_{i}, x_{j}) - \\\\mathcal{S}_{ij}\\\\big(\\\\text{sim}(x_{i}, x_{j}) + m\\\\big) - \\\\\\\\\\n&(1-\\\\mathcal{S}_{ij})\\\\big(\\\\text{sim}(x_{i}, x_{j})-m\\\\big) | \\\\label{regressionLoss} \\n\\\\end{aligned}\\n\\\\end{equation}\\nwhere $\\\\text{sim}(x_{i}, x_{j}) $ can be the cosine function for guiding the training of $\\\\bm{W}$ and $m$ is a margin. By optimizing the regression loss and updating $ \\\\bm{W} $, deep networks maximize the similarity of similar pairs and minimize that of dissimilar pairs. It is worth noting that the pre-stored parameters in the deep models are frozen when optimizing $ \\\\bm{W} $. The pipeline of this approach is depicted in Figure \\\\ref{All_Metric_Learning} (b). \\n\\\\begin{flushleft}\\n\\\\emph{b. Fine-tuning with Siamese Networks}. \\n\\\\end{flushleft}\\nSiamese networks represent important options in implementing metric learning for fine-tuning, as in Figure \\\\ref{All_Metric_Learning} (c) and Figure \\\\ref{mAP_loss_illustration} (a). It is a structure composed of two branches that share the same weights across layers. Siamese networks are trained on paired data, consisting of an image pair $ (x_{i}, x_{j}) $ such that $ \\\\mathcal{S}(x_{i}, x_{j}) \\\\! \\\\in \\\\! \\\\{0,1\\\\} $. A Siamese loss \\nis formulated as\\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\mathcal{L}_{Siam}(x_{i}, x_{j})& = \\\\frac{1}{2}\\\\mathcal{S}(x_{i}, x_{j})\\\\mathcal{D}(x_{i}, x_{j}) \\\\; + \\\\\\\\\\n&\\\\frac{1}{2}\\\\big(1 - \\\\mathcal{S}(x_{i}, x_{j})\\\\big)\\\\max\\\\big(0,\\\\; m - \\\\mathcal{D}(x_{i}, x_{j})\\\\big)\\n\\\\end{aligned}\\\\label{contrastive}\\n\\\\end{equation}\\nSiamese loss has recently been reaffirmed as a very effective metric in category-level image retrieval, outperforming many more sophisticated losses if implemented carefully . Enabled by the standard Siamese network, this objective function is used to learn the similarity between semantically relevant samples under different scenarios ,. For example, Radenovi{\\\\'c} \\\\etal employ a Siamese network on matching and non-matching global feature pairs which are aggregated by GeM-based pooling. The deep network fine-tuned by the Siamese loss generalizes better and converges at higher retrieval performance. Ong \\\\etal leverage the Siamese network to learn image features which are then fed into the Fisher Vector model for further encoding. Siamese networks can also be applied to hashing learning in which the Euclidean distance $\\\\mathcal{D}(\\\\cdot)$ in Eq. \\\\ref{contrastive} is computed for binary codes .\\nAn implicit drawback of the Siamese loss is that it may penalize similar image pairs even if the margin between these pairs is small or zero , if the constraint is too strong and unbalanced. At the same time, it is hard to map the features of similar pairs to the same point when images contain complex contents or scenes. To tackle this limitation, Cao \\\\etal adopt a double-margin Siamese loss to relax the penalty for similar pairs by setting a margin $ m_1 $ instead of zero, in which case the original single-margin Siamese loss is re-formulated as \\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\!\\\\!\\\\!\\\\! \\\\mathcal{L}_{\\\\mathcal{D}\\\\_Siam}(x_{i}, x_{j}) & = \\\\frac{1}{2}\\\\mathcal{S}(x_{i}, x_{j})\\\\max \\\\big(0, \\\\mathcal{D}(x_{i}, x_{j}) - m_{1} \\\\big) + \\\\\\\\\\n& \\\\!\\\\! \\\\frac{1}{2}\\\\big(1 - \\\\mathcal{S}(x_{i}, x_{j})\\\\big)\\\\max\\\\big(0, m_{2} - \\\\mathcal{D}(x_{i}, x_{j})\\\\big)\\n\\\\end{aligned}\\\\label{doublemargincontrastive}\\n\\\\end{equation}\\nwhere $ m_{1} \\\\!\\\\! > \\\\!\\\\! 0 $ and $ m_{2} \\\\!\\\\! > \\\\!\\\\! 0 $ are the margins affecting the similar and dissimilar pairs, respectively, as in Figure \\\\ref{mAP_loss_illustration} (b), meaning that the double margin Siamese loss only applies a contrastive force when the distance of a similar pair is larger than $ m_{1} $. The mAP metric of retrieval is improved when using the double margin Siamese loss . \\nMore recently, transformers have been trained under the regularization of cross entropy and Siamese loss for instance-level retrieval and achieved competitive performance, positioning it as an alternative to convolutional architectures. As observed by , the transformer-based architecture is less impacted than convolutional networks by feature collapse since each input feature is projected to different sub-spaces before the multi-headed attention. Moreover, the transformer backbone operates as a learned aggregation operator, thereby avoiding the design of sophisticated feature aggregation methods.\\n\\\\begin{figure*}[!t]\\n\\\\centering\\n\\\\includegraphics[width=\\\\linewidth]{./Figures/mAP_loss_illustration.pdf}\\n\\\\vspace{-2em}\\n\\\\caption{Illustrations of different losses for network fine-tuning. The same shape with different colors denotes images that include the same instance. (a)-(c) have been introduced in the text ,,. (d) Listwise AP loss considers a mini-batch of $N$ features simultaneously and directly optimizes the Average-Precision computed from these features ,.\\n} \\\\label{mAP_loss_illustration}\\n\\\\end{figure*}\\n\\\\begin{flushleft}\\n\\\\emph{c. Fine-tuning with Triplet Networks}. \\n\\\\end{flushleft}\\nTriplet networks optimize similar and dissimilar pairs simultaneously. As shown in Figure \\\\ref{All_Metric_Learning} (d) and Figure \\\\ref{mAP_loss_illustration} (c), the plain triplet networks adopt a ranking loss for training:\\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\mathcal{L}_{Triplet}(x_{a}, x_{p}, x_{n}) = \\\\max \\\\big( 0, m + \\\\mathcal{D}(x_{a}, x_{p}) - \\n \\\\mathcal{D}(x_{a}, x_{n})\\\\big)\\\\label{triplet}\\n\\\\end{aligned}\\n\\\\end{equation}\\nwhich indicates that the distance of an anchor-negative pair $ \\\\mathcal{D}(x_{a}, x_{n}) $ should be larger than that of an anchor-positive pair $ \\\\mathcal{D}(x_{a}, x_{p}) $ by a certain margin $ m $. \\n\\\\textcolor{black}{Given the datasets that provide bounding box annotations, such as INSTRE, Oxford-5k, Paris-6k, and their variants, the bounding box annotations are used as patch-level supervision to train a region detector which enables the final DCNNs to locate specific regions or objects. As an example, region proposal networks (RPNs) is fine-tuned and subsequently plugged into DCNNs and trained end-to-end , as shown in Figure \\\\ref{All_Metric_Learning} (g). RPNs yield the regressed bounding box coordinates of objects and are trained by the multi-class classification loss. Once fine-tuned, RPNs can produce regional features for each detected region by RoI pooling and perform better instance search. }\\n\\\\textcolor{black}{Further, local supervised metric learning has been explored based on the fact that RPNs enable deep models to learn regional features for particular instance objects ,,,. RPNs used in the triplet formulation are shown in Figure \\\\ref{All_Metric_Learning} (h). Firstly, regression loss (RPNs loss) is used to minimize the regressed bounding box relative to ground-truth. Then, the regional features for all detected RoIs are aggregated into a global one and L2-normalized for the triplet loss.} Note that, in some cases, jointly training an RPN loss and triplet loss leads to unstable results, a problem addressed in by first training a CNN to produce R-MAC using a rigid grid, after which the parameters in convolutional layers are fixed and RPNs are trained to replace the rigid grid.\\nAttention mechanisms can also be combined with metric learning for fine-tuning , as in Figure \\\\ref{All_Metric_Learning} (e), where the attention module is typically end-to-end trainable and takes as input the convolutional feature maps. Song \\\\etal introduce a convolutional attention layer to explore spatial-semantic information, highlighting regions in images to significantly improve the discrimination for image retrieval.\\nRecent studies , have jointly optimized the triplet loss and classification loss to further improve network capacity, as shown in Figure \\\\ref{All_Metric_Learning} (f). The overall joint function is \\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\!\\\\!\\\\!\\\\mathcal{L}_{Joint} = &\\\\lambda_{1} \\\\! \\\\cdot \\\\! \\\\mathcal{L}_{Triplet}(x_{i,a}, x_{i,p}, x_{i,n}) \\\\! + \\\\! \\\\lambda_{2} \\\\!\\\\cdot \\\\! \\\\mathcal{L}_{CE}(\\\\hat{p_{i}},y_{i})\\n\\\\end{aligned}\\n\\\\end{equation}\\nwhere the cross entropy loss (CE loss) $ \\\\mathcal{L}_{CE} $ is defined in Eq. (\\\\ref{cross_entropy_loss}) and the triplet loss $ \\\\mathcal{L}_{Triplet} $ in Eq. (\\\\ref{triplet}). $ \\\\lambda_{1} $ and $ \\\\lambda_{2} $ are hyper-parameters tuning the tradeoff between the two loss functions.\",\n \"id\": \"db1c70e4-39b7-4baf-8fa7-4ccb702563d0\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"45dba903-566e-4f9f-a48c-8f8adf0593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ],\n [\n \"subsection\",\n \"Supervised Fine-tuning \"\n ],\n [\n \"subsubsection\",\n \"Fine-tuning via Pairwise Ranking Loss\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Fine-tuning via Pairwise Ranking Loss\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 2126,\n 2127,\n 2130,\n 2129,\n 7549,\n 2098\n ],\n \"content\": \"\\\\label{Discussion_ranking_loss}\\nIn some cases, pairwise ranking loss cannot effectively learn the variations between samples and still suffers from a weaker generalization capability if the training set is not ordered correctly. Therefore, pairwise ranking loss requires careful sample mining and weighting strategies to obtain the most informative pairs, especially when considering mini-batches. The hard negative mining strategy is commonly used ,,, however further sophisticated mining strategies have recently been developed. Mishchuk \\\\etal calculate a pair-wise distance matrix on all mini-batch samples to select two closest negative and one anchor-positive pair to form a triplet. \\nInstead of traversing all possible two-tuple or three-tuple combinations, it is possible to consider all positive samples in one cluster and negative samples together. Liu \\\\etal introduce a group-group loss to decrease the intra-group distance and increase the inter-group distance. Considering all samples may be beneficial for stabilizing optimization and promoting generalization due to a larger data diversity, however the extra computational cost remains an issue to be addressed.\\nSubstantial research has been devoted to pair-wise ranking loss, while cross entropy loss, mainly used for classification, has been largely overlooked. Recently, Boudiaf \\\\etal claim that cross entropy loss can match and even surpass the pair-wise ranking loss when carefully tuned on fine-grained category-level retrieval tasks. In fact, the greatest improvements have come from enhanced training schemes (\\\\eg data augmentation, learning rate polices, batch normalization freeze) rather than intrinsic properties of pairwise ranking loss. Further, although several sophisticated ranking losses have been explored and validated for category-level retrieval, Musgrave \\\\etal revisited these losses and found that most of them perform on par to vanilla Siamese loss and triplet loss, so there is merit to consider these losses also for instance-level image retrieval tasks.\\nBoth cross entropy loss and pair-wise ranking loss regularize on the embedded features and the corresponding labels so as to maximize their mutual information . Their effectiveness is not guaranteed to give retrieval results that also optimize mAP . To tackle this limitation one can directly optimize the average precision (AP) metric using the listwise AP loss,\\n\\\\begin{equation}\\n\\\\begin{aligned}\\n\\\\!\\\\!\\\\!\\\\mathcal{L}_{mAP} = 1 - \\\\frac{1}{N} \\\\sum^{N}_{i=1} {\\\\rm AP}(x_{i}^{\\\\top}X_{N}, Y_{i})\\n\\\\label{AP_loss}\\n\\\\end{aligned}\\n\\\\end{equation}\\nwhich optimizes the global ranking of thousands of images simultaneously, instead of only a few images at a time. Here $Y_{i}$ is the binary label to evaluate the relevance between batch images. $X_{N} = \\\\{x_{1}, x_{2},...x_{j},...,x_{N}\\\\}$ denotes the features of all images, where each $x_{i}$ is used as a potential query to rank the remaining batch images. Each similarity score $x_{i}^{\\\\top}x_{j}$ can be measured by a cosine function. \\nIt is demonstrated that training with AP-based loss improves retrieval performance ,. However average precision, as a metric, is normally non-differentiable. To directly optimize the AP loss during back-propagation, the key is that the indicator function for AP computing needs to be relaxed using methods such as triangular kernel-based soft assignment or sigmoid function , as shown in Figure \\\\ref{mAP_loss_illustration} (d).\",\n \"id\": \"14afec8b-a454-4ba0-a24a-2c2e37482b34\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"45dba903-566e-4f9f-a48c-8f8adf0593ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ],\n [\n \"subsection\",\n \"Supervised Fine-tuning \"\n ],\n [\n \"subsubsection\",\n \"Discussion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Discussion\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 2131\n ],\n \"content\": \"\\\\label{Unsupervised_Fine-tuning}\\nSupervised network fine-tuning becomes infeasible when there is insufficient supervisory information, normally because of cost or unavailability. Therefore unsupervised fine-tuning methods for image retrieval are quite necessary, but less studied . \\nFor unsupervised fine-tuning, two directions are to mine relevance among features via manifold learning, and via clustering techniques, each discussed below.\",\n \"id\": \"871a2087-f4a4-4968-91d1-0752906a7612\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"acffc1fe-bf03-450b-857b-db143905236e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ],\n [\n \"subsection\",\n \"Unsupervised Fine-tuning \"\n ]\n ],\n \"subsections\": [\n \"f6c0b11c-f5b0-47c6-9e48-358712ee7aee\",\n \"7ab93d94-79dd-47b1-b0d5-6e73b13d8f40\"\n ],\n \"title\": \"Unsupervised Fine-tuning \"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 2131,\n 2103,\n 2132,\n 2100\n ],\n \"content\": \"\\\\label{Mining_Samples_with_Manifold_Learning}\\nManifold learning focuses on capturing intrinsic correlations on a manifold structure to mine or deduce relevance, as illustrated in Figure \\\\ref{manifold_learning_pipeline}. Initial similarities between the extracted global features or local features , are used to construct an affinity matrix, which is then re-evaluated and updated using manifold learning . According to the manifold similarity in the updated affinity matrix, positive and hard negative samples are selected for metric learning using pairwise ranking loss based functions such as pair loss , or triplet loss ,. Note that this is different from the aforementioned methods for pairwise ranking loss based fine-tuning methods, where the hard positive and negative samples are explicitly selected from an ordered dataset according to the given affinity information.\\nIt is important to capture the geometry of the manifold of deep features, generally involving two steps , known as diffusion. First, the affinity matrix (Figure \\\\ref{manifold_learning_pipeline}) is interpreted as a weighted kNN graph, where each vector is represented by a node, and edges are defined by the pairwise affinities of two connected nodes. Then, the pairwise affinities are re-evaluated in the context of all other elements by diffusing the similarity values through the graph ,,,, with recent strategies proposed such as regularized diffusion (RDP) and regional diffusion . For more details on diffusion methods refer to survey .\\nMost algorithms follow the two steps of ; the differences among methods lie primarily in three aspects: \\n\\\\begin{enumerate}\\n\\\\item {\\\\bf Similarity initialization},\\nwhich affects the subsequent kNN graph construction in an affinity matrix. Usually, an inner product or Euclidean distance is directly computed for the affinities. A Gaussian kernel function can be used ,, or consider regional similarity from image patches .\\n\\\\item {\\\\bf Transition matrix definition}, a row-stochastic matrix , determines the probabilities of transiting from one node to another in the graph. These probabilities are proportional to the affinities between nodes, which can be measured by Geodesic distance (\\\\emph{e.g.} the summation of weights of relevant edges).\\n\\\\item {\\\\bf Iteration scheme},\\nto re-valuate and update the values in the affinity matrix by the manifold similarity until some convergence is achieved. Most algorithms are iteration-based ,, as illustrated in Figure \\\\ref{manifold_learning_pipeline}. \\n\\\\end{enumerate}\\nDiffusion process algorithms are indispensable for unsupervised fine-tuning. Better image similarity is guaranteed when it is improved based on initialization (\\\\emph{e.g.} regional similarity or higher order information ). Diffusion is normally iterative and is computationally demanding , a limitation which cannot meet the efficiency requirements of image retrieval. To reduce the computational complexity, Bai \\\\etal propose a regularized diffusion process, facilitated by an efficient iteration-based solver. Zhao \\\\etal regard the diffusion process as a non-linear kernel mapping function, which is then modelled by a deep neural network. Other studies replace the diffusion process on a kNN graph with a diffusion network , which is derived from graph convolution networks , an end-to-end trainable framework which allows efficient computation during training and testing.\\n\\\\begin{figure}[!t]\\n\\\\centering\\n\\\\includegraphics[width=0.5\\\\textwidth]{./Figures/manifoldlearningpipeline.pdf}\\n\\\\caption{Paradigm of manifold learning for unsupervised metric learning, based on triplet loss ,.} \\n\\\\label{manifold_learning_pipeline}\\n\\\\end{figure}\\nOnce the manifold space is learned, samples are mined by computing geodesic distances based on the Floyd-Warshall algorithm or by comparing the set difference . The selected samples are fed into deep networks to perform fine-tuning.\",\n \"id\": \"f6c0b11c-f5b0-47c6-9e48-358712ee7aee\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"871a2087-f4a4-4968-91d1-0752906a7612\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ],\n [\n \"subsection\",\n \"Unsupervised Fine-tuning \"\n ],\n [\n \"subsubsection\",\n \"Mining Samples with Manifold Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Mining Samples with Manifold Learning\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 9094,\n 2135,\n 2133,\n 2128,\n 7544,\n 2134\n ],\n \"content\": \"\\\\label{Mining_Samples_by_Clustering}\\nClustering is used to explore proximity information that has been studied in instance-level retrieval ,,,,. The rationale behind these methods is that samples in a cluster are likely to satisfy a degree of similarity. \\nOne class of methods for clustering deep features is via k-means. Given $k$ cluster centroids, during each training epoch a deep network alternates between two steps: first, a soft assignment between the feature representations and the cluster centroids; second, the cluster centroids are refined and, at the same time, the deep network is updated by learning from current high confidence assignments using a certain regularization. These two steps are repeated until a convergence criterion is met, at which point the cluster assignments are used as pseudo-labels ,. Alternatively, the pseudo-labels can be calculated from the samples in a cluster, \\\\eg the mean values. For example, Tzelepi \\\\etal compute $ k $ nearest feature representations with respect to a query feature and then compute their mean vectors, which is used as a target for the query feature. In this case, fine-tuning is performed by minimizing the squared distance between each query feature and the mean of its $ k $ nearest features. Liu \\\\etal propose a self-taught hashing algorithm using a kNN graph construction to generate pseudo labels that are used to analyze and guide network training. Shen \\\\etal and Radenovi{\\\\'c} \\\\etal , use Structure-from-Motion (SfM) for each image cluster to explore sample reconstructions to select images for triplet loss. Clustering methods depend on the Euclidean distance, making it difficult to reveal the intrinsic relationship between objects.\\nThere are further techniques for instance retrieval, such as by using AutoEncoder ,, generative adversarial networks (GANs) , convolutional kernel networks ,, and graph convolutional networks . For these methods, they focus on devising novel unsupervised frameworks to realize unsupervised learning, instead of iterative similarity diffusion or cluster refinement on feature space. For example, instead of performing iterative traversal on a set of nearest neighbors defined by kNN graph, Liu \\\\etal employ graph convolutional networks to directly encode the neighbor information into image descriptors and then train the deep models to learn a new feature space. This method is demonstrated to significantly improve retrieval accuracy while maintaining efficiency. GANs are also explored, for the first time, for instance-level retrieval in an unsupervised fashion . The generator retrieves images that contain similar instances as a given image, while the discriminator judges whether the retrieved images have the specified instance which appeared in the query image. During training, the discriminator and the generator play a min-max game via an adversarial reward which is computed based on the cosine distance between the query image and the images retrieved by the generator.\",\n \"id\": \"7ab93d94-79dd-47b1-b0d5-6e73b13d8f40\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"871a2087-f4a4-4968-91d1-0752906a7612\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"$\\\\!\\\\!\\\\!\\\\!\\\\!$Retrieval via Learning DCNN Representations \"\n ],\n [\n \"subsection\",\n \"Unsupervised Fine-tuning \"\n ],\n [\n \"subsubsection\",\n \"Mining Samples by Clustering\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Mining Samples by Clustering\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{State_of_the_Art_Performance}\",\n \"id\": \"a6c69e00-d8a1-492b-8514-fab570f109d2\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"515509a6-319a-4385-b1b5-4418a83d2bc2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"State of the Art Performance\"\n ]\n ],\n \"subsections\": [\n \"347ae713-adb4-418f-9e62-4191f141d96c\",\n \"78dcf4cf-a971-4517-838a-4552eb9d65a5\",\n \"4289168f-a812-4899-b8a0-52669cd0ba36\"\n ],\n \"title\": \"State of the Art Performance\"\n },\n {\n \"cite_extract_rate\": 0.384615384615384,\n \"cites\": [\n 2114,\n 2136,\n 2137,\n 2093,\n 2111\n ],\n \"content\": \"\\\\label{Datasets_and_Evaluation_Criteria}\\nTo demonstrate the effectiveness of methods, we choose the following commonly-used datasets for performance comparison:\\n\\\\textbf{UKBench (UKB)} consists of 10,200 images of objects. This dataset has 2,550 groups of images, each group having four images of the same object from different viewpoints or illumination conditions, which can be regarded as a kind of class-level supervision information. All images can be used as a query.\\n\\\\textbf{Holidays} consists of 1,491 images collected from personal holiday albums. Most images are scene-related. The dataset comprises 500 groups of similar images with a single query image for each group. The dataset also provides position information of the interest regions for each image.\\n\\\\textbf{Oxford-5k} consists of 5,062 images for 11 Oxford buildings. Each building is associated with five hand-drawn bounding box queries. \\\\textcolor{black}{According to the relevance level, each image of the same building is assigned a label \\\\textit{Good} (\\\\ie \\\\textit{positive}), \\\\textit{OK} (\\\\ie \\\\textit{positive}), \\\\textit{Junk}, or \\\\textit{Bad} (\\\\ie \\\\textit{negative}). \\\\textit{Junk} images can be \\ndiscarded or regarded as \\\\textit{negative} examples ,.\\nTo build a tuple for each given query, one can select a positive example whose label corresponds to \\\\textit{Good} or \\\\textit{OK} in the same category, and select one negative example from each of the remaining building categories. Furthermore, an additional disjoint set of 100,000 distractor images is added to obtain Oxford-105k.}\\n\\\\textbf{Paris-6k} includes 6,412 images and is categorized into 12 groups by architecture. \\nThe supervision information can be used like that of Oxford-5k. Likewise, an additional disjoint set of 100,000 distractor images is added to obtain Paris-106k.\\n\\\\textbf{INSTRE} consists of 28,543 images from 250 different object classes, including three disjoint subsets\\\\footnote{https://github.com/imatge-upc/salbow}: INSTRE-S1, INSTRE-S2, INSTRE-M. \\n\\\\textcolor{black}{INSTRE dataset has bounding box annotations, providing single-labelled and double-labelled class information for single- and multiple-object retrieval, respectively. One can use the class information to build a tuple, with two positive examples from the same class and one negative from one of the remaining classes. The performance evaluation on INSTRE in our experiments follows the protocol in . }\\n\\\\textbf{Google Landmarks Dataset (GLD)} , consists of GLD-v1 and GLD-v2. \\\\textcolor{black}{GLD-v2 is mainly recommended to use and it has the advantage of stability where all images have permissive licenses . GLD-v2 is divided into three subsets: (i) 118k query images with ground-truth annotations, (ii) 4.1M training images of 203k landmarks with labels, and (iii) 762k index images of 101k landmarks. Due to its large scale, GLD-v2 provides class-level ground-truth which can be used to build training tuples. Due to its image diversity, it may produce clutter images for each landmark so it is necessary to introduce pre-processing methods to select the more relevant images . Finally, the training set is cleaned by removing these clutters, consisting of a subset ``GLD-v2-clean'' containing 1.6M images of 81k landmarks. Since Google landmarks dataset stills lack bounding box for objects of interest, Teichmann \\\\etal provide a new dataset of landmark bounding boxes, based on GLD. This patch-level supervision information can help locate the most relevant regions. }\\n\\\\textcolor{black}{Note that, additional queries and distractor images have been added into Oxford-5k and Paris-6k, producing the Revisited Oxford ($\\\\mathcal{R}$Oxford) and Revisited Paris ($\\\\mathcal{R}$Paris) datasets where each image of the same building is assigned a label \\\\textit{Easy}, \\\\textit{Hard}, \\\\textit{Unclear}, or \\\\textit{Negative} . Different label combinations are used as \\\\textit{positive} according to the difficulty level of different setups. During testing, if there are no positive images for a query, then that query is excluded from the evaluation. For details, we refer the reader to . We undertake partial comparisons under the \\\\textit{hard} evaluation protocol on these revisited datasets. }\",\n \"id\": \"347ae713-adb4-418f-9e62-4191f141d96c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"a6c69e00-d8a1-492b-8514-fab570f109d2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"State of the Art Performance\"\n ],\n [\n \"subsection\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\textbf{Average precision} (AP) refers to the coverage area under the precision-recall (PR) curve. A larger AP implies a higher PR curve and better retrieval accuracy. AP can be calculated as $ AP=\\\\frac{1}{R} \\\\sum_{k=1}^N P(k)\\\\cdot rel(k) $,\\nwhere $ R $ denotes the number of relevant results for the query image from the total number $ N $ of images. $ P(k) $ is the precision of the top $ k $ retrieved images, and $ rel(k) $ is an indicator function equal to 1 if the item within rank $ k $ is a relevant image and 0 otherwise. Mean average precision (mAP) is adopted for the evaluation over all query images, \\\\ie $mAP=\\\\frac1Q{\\\\sum_{q=1}^Q}AP(q) $,\\nwhere $ Q $ is the number of query images.\\nThe \\\\textbf{N-S score} is a metric used for UKBench ; the N-S score is the average for the top-4 precision over the dataset.\",\n \"id\": \"78dcf4cf-a971-4517-838a-4552eb9d65a5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"a6c69e00-d8a1-492b-8514-fab570f109d2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"State of the Art Performance\"\n ],\n [\n \"subsection\",\n \"Evaluation Metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Evaluation Metrics\"\n },\n {\n \"cite_extract_rate\": 0.50943396226415,\n \"cites\": [\n 2136,\n 2140,\n 7543,\n 2093,\n 514,\n 2096,\n 2105,\n 2113,\n 8525,\n 8526,\n 7546,\n 2094,\n 2137,\n 2131,\n 2103,\n 2139,\n 2138,\n 7544,\n 2119,\n 2109,\n 2097,\n 2132,\n 2100,\n 2117,\n 2112,\n 7545,\n 2111\n ],\n \"content\": \"\\\\begin{figure}[t]\\n\\\\centering\\n\\\\begin{subfigure}{0.248\\\\textwidth}\\n\\\\resizebox{\\\\columnwidth}{!}{\\n\\\\begin{tikzpicture} \\n\\\\scriptsize\\n\\\\begin{axis}[legend style={at={(0.95,0.3)}, scale=0.5, anchor=north east, font= \\\\scriptsize}, symbolic x coords={2014, 2015, 2016, 2017, 2018, 2019, 2020 }, grid=both, ylabel={mAP(\\\\%)}, ymin=77.5, ymax=99, xmin=2014, xmax=2020, xtick=data, x tick label style={rotate=45,anchor=east} ];\\n\\\\addlegendentry{Holidays}\\n\\\\addplot[mark=square*,thick, line width=0.45mm, color={rgb, 255:red,152;green,179;blue,134}, on layer=main] coordinates { (2014, 80.18) (2015,89.7) (2016,94.2) (2017,95.13) (2018,95.67) (2019, 95.5) (2020, 94.0) };\\n\\\\node [above right=0cm, font=\\\\tiny] at (axis cs:2014, 80.5) {80.18 }; \\\\node [ font=\\\\tiny] at (axis cs:2015, 90.4) {89.7 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2016, 93.4) {94.2 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2017, 94.8) {95.13 }; \\\\node [above right=-0.4cm, font=\\\\tiny] at (axis cs:2018, 95.6) {95.7 }; \\\\node [above right=-0.5cm, font=\\\\tiny] at (axis cs:2019, 95.9) {95.5 }; \\\\node [above right=-0.8cm, font=\\\\tiny] at (axis cs:2020, 96) {94.0 };\\n\\\\addlegendentry{UKBench}\\n\\\\addplot[mark=diamond*,thick,line width=0.45mm, color={rgb,255:red,141;green,170;blue,196}, on layer=main,] coordinates {(2014,91.1) (2015,91.3) (2016,96.3) (2018,98.08) (2020,98.84) }; \\\\node [above right=0cm, font=\\\\tiny] at (axis cs:2014, 91) {91.1 }; \\\\node [above right=-0.4cm, font=\\\\tiny] at (axis cs:2015, 92.8) {91.3 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2016, 97.4) {96.3 }; \\\\node [above right=-0.4cm, font=\\\\tiny] at (axis cs:2018, 99.5) {98.1 }; \\\\node [above right=-0.8cm, font=\\\\tiny] at (axis cs:2020, 100.8) {98.8 };\\n\\\\addlegendentry{Oxford-5k}\\n\\\\addplot[mark=*,thick, on layer=background, line width=0.45mm, color={rgb, 255:red,237;green,160;blue,155}] coordinates { (2014,78.34) (2015,84.4) (2016,88.95) \\n(2018,95.8) (2019,96.2) (2020,96.2) }; \\\\node [above right=0cm, font=\\\\tiny] at (axis cs:2014, 77.5) {78.34 }; \\\\node [ font=\\\\tiny] at (axis cs:2015, 85.15) {84.4 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2016, 88.6) {88.95 }; \\n\\\\node [above right=-0.4cm, font=\\\\tiny] at (axis cs:2018, 97.2) {95.8 }; \\\\node [above right=-0.5cm, font=\\\\tiny] at (axis cs:2019, 98) {96.2 }; \\\\node [above right=-0.8cm, font=\\\\tiny] at (axis cs:2020, 99.1) {96.2 };\\n\\\\addlegendentry{Paris-6k}\\n\\\\addplot[mark=triangle*,thick,line width=0.45mm, color={rgb, 255:red,217;green,188;blue,102}, on layer=main,] coordinates { (2014,86.83) (2015,86.5) (2016,95.88) (2017,96.0) (2018, 97.0) (2019, 97.8) (2020, 97.4) };\\n\\\\node [above right=0cm, font=\\\\tiny] at (axis cs:2014, 86.8) {86.83 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2015, 87.8) {86.5 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2016, 95.6) {95.8 };\\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2017, 97.0) {96.0 }; \\\\node [above right=-0.4cm, font=\\\\tiny] at (axis cs:2018, 98.4) {97.0 }; \\\\node [above right=-0.5cm, font=\\\\tiny] at (axis cs:2019, 99.3) {97.8 }; \\\\node [above right=-0.8cm, font=\\\\tiny] at (axis cs:2020, 100) {97.4 };\\n\\\\end{axis}\\n\\\\end{tikzpicture}\\n} \\n\\\\vspace{-2.1em}\\n\\\\end{subfigure}\\n\\\\begin{subfigure}{0.235\\\\textwidth}\\n\\\\resizebox{\\\\columnwidth}{!}{\\n\\\\begin{tikzpicture} \\n\\\\scriptsize\\n\\\\begin{axis}[legend style={at={(0.425,0.3)}, scale=0.5, anchor=north east, font=\\\\scriptsize}, symbolic x coords={2017, 2018, 2019, 2020}, grid=both, ylabel={}, ymin=15, ymax=95, xmin=2017, xmax=2020, xtick=data, x tick label style={rotate=45, anchor=east} ];\\n\\\\addlegendentry{$\\\\mathcal{R}$Oxford-5k}\\n\\\\addplot[mark=*,thick,line width=0.45mm, color={rgb, 255:red,181;green,144;blue,202}, on layer=main,] coordinates { (2018, 54.8) (2019, 65.8) (2020, 64.0) };\\n\\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2018, 53.8) {54.8 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2019, 72.5) {65.8 }; \\\\node [above right=-0.8cm, font=\\\\tiny] at (axis cs:2020, 74.0) {64.0 };\\n\\\\addlegendentry{$\\\\mathcal{R}$Paris-6k}\\n\\\\addplot[mark=triangle*,thick,line width=0.45mm, color={rgb, 255:red,117;green,207;blue,184}, on layer=main,] coordinates { (2018, 84.0) (2019, 85.3) (2020, 80.4) };\\n\\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2018, 92.5) {84.0 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2019, 84.0) {85.3 }; \\\\node [above right=-0.9cm, font=\\\\tiny] at (axis cs:2020, 93.5) {80.4 };\\n\\\\addlegendentry{INSTRE}\\n\\\\addplot[mark=diamond*,thick,line width=0.45mm, color={rgb, 255:red,83;green,82;blue,4}, on layer=main,] coordinates { (2017, 80) (2018, 80.5) (2019, 92.4)};\\n\\\\node [above right=-0.05cm, font=\\\\tiny] at (axis cs:2017, 71.5) {80.0 }; \\\\node [above right=-0.3cm, font=\\\\tiny] at (axis cs:2018, 79.5) {80.5 }; \\\\node [above right=0cm, font=\\\\tiny] at (axis cs:2019, 86) {92.4 } ;\\n\\\\addlegendentry{GLD-v2}\\n\\\\addplot[mark=square*,thick,line width=0.45mm, color={rgb, 255:red,130;green,54;blue,203}, on layer=main,] coordinates {(2020, 26.8) };\\n\\\\node [above right=-0.8cm, font=\\\\tiny] at (axis cs:2020, 38) {26.8 };\\n\\\\end{axis};\\n\\\\end{tikzpicture}\\n}\\n\\\\vspace{-2.1em}\\n\\\\end{subfigure}\\n\\\\caption{Performance improved from 2014 to 2020.} \\n\\\\label{datasets_progress_2014_2_2020}\\n\\\\end{figure}\\n\\\\textbf{Overview.} Figure \\\\ref{datasets_progress_2014_2_2020} summarizes the performance over 6 datasets from 2014 to 2020. Early on, the powerful feature extraction of DCNNs led to rapid improvements. Subsequent key ideas have been to extract instance features at the region level to reduce image clutter , and to improve feature discriminativity by using methods including feature fusion ,,, feature aggregation ,, and feature embedding . Fine-tuning is an important strategy to improve performance by tuning deep networks specific for learning instance features ,. For instance, the accuracy increases steadily from 78.34\\\\% to 96.2\\\\% on the Oxford-5k dataset when manifold learning is used to fine-tune deep networks. The mAP on $\\\\mathcal{R}$Paris-6k and $\\\\mathcal{R}$Oxford-5k is smaller than Paris-6k and Oxford-5k, leaving room for improvement.\\nWe report results using off-the-shelf models (Table \\\\ref{Table_retrieval_off_the_shelf}) and fine-tuning networks (Table \\\\ref{Metric_learning_table_B}). In Table \\\\ref{Table_retrieval_off_the_shelf}, single-pass and multiple-pass are analyzed, while supervised and unsupervised fine-tuning are compared in Table \\\\ref{Metric_learning_table_B}. Since there are many aspects that vary across the different methods, making them not directly comparable, we mainly draw some general claims or trends based on the collected results.\\n\\\\begin{table*}[t]\\n\\\\centering \\n\\\\caption{\\\\footnotesize Performance evaluation of off-the-shelf DCNN models. ``$ \\\\bullet $'' indicates that the models or layers are combined to learn features; ``PCA$_{w}$ indicates PCA with whitening on the extracted features to improve robustness; ``MP'' means Max Pooling; ``SP'' means Sum Pooling. The CNN-M network with ``$\\\\ast$'' has an architecture similar to that of AlexNet. ``-'' means that the results were not reported. }\\n\\\\vspace{-1em}\\n\\\\label{Table_retrieval_off_the_shelf}\\n\\\\renewcommand{\\\\arraystretch}{1.0}\\n\\\\resizebox{19.5cm}{!}{\\n\\\\begin{tabular}{!{\\\\vrule width1.2bp}c|c|c|c|c|c|c|c|c|c|p{9cm}!{\\\\vrule width1.2bp}}\\n\\\\Xhline{1pt}\\n\\\\footnotesize Type & \\\\footnotesize \\\\shortstack [c] { Method } & \\\\footnotesize \\\\shortstack [c] {Backbone \\\\\\\\ DCNN}\\t& \\\\footnotesize \\\\shortstack [c] {Output \\\\\\\\ Layer} & \\\\footnotesize \\\\shortstack [c] {Embed. \\\\\\\\ Aggre.}\\t& \\\\footnotesize \\\\shortstack [c] { Feat. \\\\\\\\ Dim } & \\\\footnotesize \\\\shortstack [c] {Holidays} & \\\\footnotesize \\\\shortstack [c] {UKB} & \\\\footnotesize \\\\shortstack [c] {Oxford5k \\\\\\\\ (+100k)} & \\\\footnotesize \\\\shortstack [c] {Paris6k \\\\\\\\ (+100k)} & \\\\footnotesize Brief Conclusions and Highlights \\\\\\\\\\n\\\\Xhline{1.0pt}\\n\\\\multirow{5}{*}{ \\\\raisebox{-15ex}[0pt]{\\\\begin{tabular}[c]{@{}c@{}} \\\\rotatebox{90}{Single-pass} \\\\end{tabular}}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Neural \\\\\\\\ codes } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ AlexNet} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { FC6}}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { PCA}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ $128$} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ $74.7$} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { $ 3.42 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}}& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $43.3$\\\\\\\\ (38.6)}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize Compressed neural codes of different layers are explored. AlexNet is also fine-tuned for retrieval.\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { SPoC } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2.5ex}[0pt]{\\\\shortstack [c] { SPoC + \\\\\\\\ PCA$_{w} $}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 256} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 80.2} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { $ 3.65 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $58.9$\\\\\\\\ (57.8)}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize Exploring Gassian weighting scheme \\\\ie the centering prior, to improve the discrimination of\\nfeatures. \\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { CroW } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{VGG16} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2.5ex}[0pt]{\\\\shortstack [c] { CroW + \\\\\\\\ PCA$_{w} $}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{256} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ $85.1$} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { $-$}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] {$68.4$\\\\\\\\ (63.7)}} & \\\\footnotesize\\t\\\\raisebox{-3.0ex}[0pt]{\\\\shortstack [c] { $ 76.5 $\\\\\\\\ (69.1) }} & \\\\footnotesize The spatialwise and channelwise weighting mechanisms are utilized to highlight crucial convolutional features. \\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { R-MAC } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2.5ex}[0pt]{\\\\shortstack [c] { R-MAC + \\\\\\\\ PCA$_{w} $}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{512} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$ }} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $66.9$\\\\\\\\ (61.6)}} & \\\\footnotesize\\t\\\\raisebox{-3.0ex}[0pt]{\\\\shortstack [c] { $-$ \\\\\\\\ (75.7) }} & \\\\footnotesize Sliding windows with different scales on convolutional feature maps to encode multiple image regions.\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Multi-layer \\\\\\\\ CNN } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{VGG16} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { FC6 $\\\\bullet$ \\\\\\\\ Conv4$\\\\sim$5 }}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { SP}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{4096} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 91.4} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { $ 3.68 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $61.5$\\\\\\\\($-$)}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize Layer-level feature fusion and the complementary properties of different layers are explored.\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { BLCF } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2.5ex}[0pt]{\\\\shortstack [c] { BoW + \\\\\\\\ PCA$_{w} $}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 25k} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{$-$} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { $-$}}& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $73.9$\\\\\\\\ (59.3)}} & \\\\footnotesize\\t\\\\raisebox{-3.0ex}[0pt]{\\\\shortstack [c] { $ 82.0 $\\\\\\\\ (64.8) }} & \\\\footnotesize Both global features and local features are explored, demonstrating that local features have higher accuracy. \\\\\\\\ \\\\cline{2-11} \\n\\\\Xhline{1.0pt}\\n\\\\multirow{7}{*}{ \\\\raisebox{-20ex}[0pt]{\\\\begin{tabular}[c]{@{}c@{}} \\\\rotatebox{90}{Multiple-pass} \\\\end{tabular}}}\\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { OLDFP } } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {AlexNet }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { FC6}}\\t& \\\\footnotesize \\\\raisebox{-2.6ex}[0pt]{\\\\shortstack [c] { MP \\\\\\\\ + PCA$_{w} $ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 88.5} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] {$ 3.81 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $60.7$\\\\\\\\ ($-$)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $66.2$\\\\\\\\ ($-$)}} & \\\\footnotesize Exploring the impact of proposal number. Patches are extracted by RPNs (see Figure \\\\ref{MultiplePass} (d)) and the features are encoded in an orderless way.\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-1.6ex}[0pt]{ \\\\shortstack [c] { MOP-CNN } } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { AlexNet }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { FC7}} & \\\\footnotesize \\\\raisebox{-2.5ex}[0pt]{\\\\shortstack [c] { VLAD \\\\\\\\ + PCA$_{w} $ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 2048} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 80.2} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { $-$}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize Image patches are extracted densely, as shown in Figure \\\\ref{MultiplePass} (c). Multi-scale patch features are further embedded into VLAD descriptors. \\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { CNNaug-ss } } & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { Overfeat\\\\\\\\ }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { FC}}\\t\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { PCA$_{w} $ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 15k} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 84.3} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { $ 91.1 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(mAP)}}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $68.0$\\\\\\\\($-$)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $79.5$\\\\\\\\ ($-$)}} & \\\\footnotesize Image patches are extracted densely, as shown in Figure \\\\ref{MultiplePass} (c). Image regions at different locations with different sizes are included.\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { MOF } } & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { CNN-M$^{\\\\ast}$ \\\\\\\\ }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { FC7 $\\\\bullet$ \\\\\\\\ Conv }}\\t& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { SP or MP\\\\\\\\ + BoW }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 20k} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 76.8} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { $ 3.00 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize Exploring layer-level fusion scheme. Image patches are extracted using spatial pyramid modeling, as shown in Figure \\\\ref{MultiplePass} (b).\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2.7ex}[0pt]{ \\\\shortstack [c] { Multi-scale \\\\\\\\ CNN } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Conv5}} & \\\\footnotesize \\\\raisebox{-2.7ex}[0pt]{\\\\shortstack [c] { SP or MP\\\\\\\\ + PCA$_{w} $ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 32k} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 89.6} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { $ 95.1 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(mAP)}}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $84.3$\\\\\\\\ ($-$)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $87.9$\\\\\\\\ ($-$)}} & \\\\footnotesize Image patches are extracted in a dense manner, as shown in Figure \\\\ref{MultiplePass} (c). Geometric invariance is considered when aggregating patch features.\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { LDD } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG19} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2.6ex}[0pt]{\\\\shortstack [c] {BoW \\\\\\\\ + PCA$_{w} $ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{500k} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 84.6 } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$ }} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $83.3$ \\\\\\\\ ($-$)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $87.2$ \\\\\\\\ ($-$)}} & \\\\footnotesize Image patches are obtained using a uniform square mesh, as shown in Figure \\\\ref{MultiplePass} (a). Patch features are encoded into BoW descriptors.\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-1.7ex}[0pt]{ \\\\shortstack [c] { DeepIndex } } & \\\\footnotesize\\t\\\\raisebox{-2.7ex}[0pt]{ \\\\shortstack [c] { AlexNet \\\\\\\\ $\\\\bullet$ VGG19 }} & \\\\footnotesize\\t\\\\raisebox{-2.7ex}[0pt]{ \\\\shortstack [c] { FC6-7 $\\\\bullet$\\\\\\\\ FC17-18 }}\\t& \\\\footnotesize \\\\raisebox{-2.7ex}[0pt]{\\\\shortstack [c] {BoW + \\\\\\\\ PCA }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 81.7} & \\\\footnotesize \\\\raisebox{-2.7ex}[0pt]{\\\\shortstack [c] { $ 3.32 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize\\t\\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-2.7ex}[0pt]{ \\\\shortstack [c] { $75.4$\\\\\\\\ ($-$)}} & \\\\footnotesize Exploring layer-level and model-level fusion methods. Image patches are extracted using spatial pyramid modeling, as shown in Figure \\\\ref{MultiplePass} (b).\\\\\\\\ \\\\cline{2-11} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { CCS } } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { GoogLeNet }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Conv}}\\t& \\\\footnotesize \\\\raisebox{-2.7ex}[0pt]{\\\\shortstack [c] { VLAD \\\\\\\\ + PCA$_{w} $ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 128} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 84.1} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { $ 3.81 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $64.8$\\\\\\\\ ($-$)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $76.8$\\\\\\\\ ($-$)}} & \\\\footnotesize Object proposals are extracted by RPNs, as shown in Figure \\\\ref{MultiplePass} (d). Object level and point level feature concatenation schemes are explored.\\\\\\\\ \\\\cline{2-11} \\n\\\\Xhline{1.0pt}\\n\\\\end{tabular}\\n}\\n\\\\vspace{-0.5em}\\n\\\\end{table*}\\n\\\\begin{table*}[!t]\\n\\\\centering \\n \\\\caption{ \\\\footnotesize Performance evaluation of methods in which DCNN models are fine-tuned, in a supervised or an unsupervised manner. ``CE Loss'' means the models are fine-tuned using the classification-based loss function in the form of Eq. \\\\ref{cross_entropy_loss}. ``Siamese Loss'' is in the form of Eq. \\\\ref{contrastive}. ``Regression Loss'' is in the form of Eq. \\\\ref{regressionLoss}. ``Triplet Loss'' is in the form of Eq. \\\\ref{triplet}. }\\n \\\\vspace{-1em}\\n \\\\label{Metric_learning_table_B} \\n \\\\renewcommand{\\\\arraystretch}{1.0}\\n\\\\resizebox{19.5cm}{!}{\\n\\\\begin{tabular}{!{\\\\vrule width1.2bp}c|p{2.0cm}|c|p{1.35cm}|c|p{1.1cm}|p{1.25cm}|c|c|p{0.7cm}|c|c|p{7.8cm}!{\\\\vrule width1.2bp}}\\n\\\\Xhline{1pt}\\n\\\\footnotesize Type & \\\\footnotesize \\\\shortstack [c] { Method } & \\\\footnotesize \\\\shortstack [c] {Backbone \\\\\\\\ DCNN} & \\\\footnotesize \\\\shortstack [c] {Training \\\\\\\\ set}\\t& \\\\footnotesize \\\\shortstack [c] {Output \\\\\\\\ Layer} & \\\\footnotesize \\\\shortstack [c] {Embed. \\\\\\\\ Aggre.}\\t& \\\\footnotesize \\\\shortstack [c] { Loss \\\\\\\\ Function } & \\\\footnotesize \\\\shortstack [c] { Feat. \\\\\\\\ Dim } & \\\\footnotesize \\\\shortstack [c] {Holidays} & \\\\footnotesize \\\\shortstack [c] {UKB} & \\\\footnotesize \\\\shortstack [c] {Oxford5k \\\\\\\\ (+100k)} & \\\\footnotesize \\\\shortstack [c] {Paris6k \\\\\\\\ (+100k)} & \\\\footnotesize Brief Conclusions and Highlights \\\\\\\\ \\\\Xhline{1.0pt}\\n\\\\multirow{5}{*}{ \\\\raisebox{-24ex}[0pt]{\\\\begin{tabular}[c]{@{}c@{}} \\\\rotatebox{90}{Supervised Fine-tuning}\\\\end{tabular}}} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\!\\\\!\\\\! $ Neural codes } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ AlexNet} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{\\\\shortstack [c] { Landmarks \\\\\\\\ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { FC6}}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { PCA }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { CE Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 128} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 78.9 } & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{\\\\shortstack [c] { $ 3.29 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $55.7$\\\\\\\\ (52.3)}} & \\\\footnotesize \\\\raisebox{-2.0ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize The first work which fine-tunes deep networks for IIR. Compressed neural codes and different layers are explored. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { Nonmetric } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{VGG16} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ $\\\\!\\\\!\\\\! $ Landmarks } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { PCA$_w$ }} & \\\\footnotesize\\t\\\\raisebox{-3.2ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\! $ Regression \\\\\\\\ $\\\\!\\\\!\\\\!\\\\! $ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] {$-$} } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $88.2$\\\\\\\\ (82.1)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $88.2$\\\\\\\\ (82.9)}} & \\\\footnotesize Similarity learning of similar and dissimilar pairs is performed by a neural network, optimized using regression loss. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { SIAM-FV } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ $\\\\!\\\\!\\\\!\\\\! $ Landmarks } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { Conv5}}\\t& \\\\footnotesize\\t\\\\raisebox{-2.7ex}[0pt]{ \\\\shortstack [c] { FV + \\\\\\\\ PCA$_w$ }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Siamese \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] {$-$} } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $81.5$\\\\\\\\ (76.6)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $82.4$\\\\\\\\ ($-$)}} & \\\\footnotesize Fisher Vector is integrated on top of VGG and is trained with VGG simultaneously. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Faster \\\\\\\\ R-CNN } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Oxford5k \\\\\\\\ Paris6k } } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] { MP / SP }} & \\\\footnotesize\\t\\\\raisebox{-3.2ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\! $ Regression \\\\\\\\ $\\\\!\\\\!\\\\!\\\\! $ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] {$-$} } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $75.1$\\\\\\\\ $(-)$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $80.7$\\\\\\\\ ($-$)}} & \\\\footnotesize RPN is fine-tuned, based on bounding box coordinates and class scores for specific region query which is region-targeted. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { SIFT-CNN } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{VGG16} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Holidays \\\\\\\\ UKB }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { SP }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Siamese \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 88.4 } & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{\\\\shortstack [c] { $ 3.91 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(N-S)}}} & \\\\footnotesize\\t \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] {$-$}} & \\\\footnotesize SIFT features are used as supervisory information for mining positive and negative samples. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\!\\\\! $ Quartet-Net } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize\\t\\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { GeoPair \\\\\\\\ }} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { FC6}}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { PCA }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Siamese \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 128} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{71.2 } & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{\\\\shortstack [c] { $ 87.5 $ \\\\\\\\ \\\\textcolor[rgb]{0,0,0}{(mAP)}}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $48.5$\\\\\\\\ ($-$)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] {$48.8$\\\\\\\\($-$)}} & \\\\footnotesize Quartet-net learning is explored to improve feature discrimination where double-margin contrastive loss is used. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { NetVLAD } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ VGG16} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\!\\\\! $ Tokyo Time \\\\\\\\ Machine }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { VLAD \\\\\\\\ Layer }}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { PCA$_{w}$ }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Triplet \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 256} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 79.9 } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $62.5$\\\\\\\\ ($-$)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $72.0$\\\\\\\\ ($-$)}} & \\\\footnotesize VLAD is integrated at the last convolutional layer of VGG16 network as a plugged layer. \\\\\\\\ \\\\cline{2-13} \\\\textbf{}\\n& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Deep \\\\\\\\ Retrieval } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ ResNet-101} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ $\\\\!\\\\!\\\\!\\\\!\\\\! $ Landmarks } & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Conv5 \\\\\\\\ Block }}\\t& \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { MP + \\\\\\\\ PCA$_w$ }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Triplet \\\\\\\\Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 2048} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 90.3 } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $86.1$\\\\\\\\ (82.8)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $94.5$\\\\\\\\ (90.6)}} & \\\\footnotesize Dataset is cleaned automatically. Features are encoded by R-MAC. RPN is used to extract the most relevant regions. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { DELF } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ ResNet-101} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ GLDv1 } & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Conv4 \\\\\\\\ Block }}\\t& \\\\footnotesize\\t\\\\raisebox{-2.6ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\! $ Attention \\\\\\\\ + PCA$_w$ }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { CE Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 2048} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{$-$} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{$-$} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $83.8$\\\\\\\\ (82.6)}} & \\\\footnotesize \\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $85.0$\\\\\\\\ (81.7)}} & \\\\footnotesize Exploring the FCN to extract region-level features and construct feature pyramids of different sizes. \\\\\\\\ \\\\cline{2-13} \\n \\\\Xhline{1.0pt}\\n\\\\multirow{7}{*}{ \\\\raisebox{-13ex}[0pt]{\\\\begin{tabular}[c]{@{}c@{}} \\\\rotatebox{90}{Unsupervised Fine-tuning} \\\\end{tabular}}}\\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { MoM } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { VGG16 }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ Flickr 7M } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { MP + \\\\\\\\ PCA$_{w} $ }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Siamese \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 64} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 87.5} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $78.2$\\\\\\\\ (72.6)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $85.1$\\\\\\\\ (78.0)}} & \\\\footnotesize Exploring manifold learning for mining dis/similar samples. Features are tested globally and regionally.\\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { GeM } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { VGG16 }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ Flickr 7M } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] { Conv5}}\\t& \\\\footnotesize \\\\raisebox{-3ex}[0pt]{\\\\shortstack [c] { GeM \\\\\\\\ Pooling }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Siamese \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 83.1} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $82.0$\\\\\\\\ (76.9)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $79.7$\\\\\\\\ (72.6)}} & \\\\footnotesize Fine-tuning CNNs on an unordered dataset. Samples are selected from an automated 3D reconstruction system. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { SfM-CNN } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { VGG16 }} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ Flickr 7M } & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [l] {Conv5}}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { PCA$_{w}$ }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Siamese \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{512} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 82.5} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $77.0$\\\\\\\\ (69.2)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $83.8$\\\\\\\\ (76.4)}} & \\\\footnotesize Employing Structure-from-Motion to select positive and negative samples from unordered images. \\\\\\\\ \\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { MDP-CNN } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ ResNet-101} & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ $\\\\!\\\\!\\\\!\\\\!\\\\! $ Landmarks } & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Conv5 \\\\\\\\ Block }}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { SP }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { Triplet \\\\\\\\ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 2048} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] {$-$} } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $85.4$\\\\\\\\ (85.1)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $96.3$\\\\\\\\ (94.7)}} & \\\\footnotesize Exploring global feature structure by modeling the manifold learning to select positive and negative pairs. \\\\\\\\ \\n\\\\cline{2-13} \\n& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { IME-CNN } } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{ ResNet-101} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\! $ Oxford105k \\\\\\\\ Paris106k }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { IME \\\\\\\\ Layer }}\\t& \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] { MP }} & \\\\footnotesize\\t\\\\raisebox{-3ex}[0pt]{ \\\\shortstack [c] { $\\\\!\\\\!\\\\!\\\\! $ Regression \\\\\\\\ $\\\\!\\\\!\\\\!\\\\!\\\\! $ Loss}} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ 2048} & \\\\footnotesize \\\\raisebox{-2ex}[0pt]{ \\\\shortstack [c] {$-$} } & \\\\footnotesize\\t\\\\raisebox{-2ex}[0pt]{\\\\shortstack [c] {$-$}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $92.0$\\\\\\\\ (87.2)}} & \\\\footnotesize \\\\raisebox{-3.5ex}[0pt]{ \\\\shortstack [c] { $96.6$\\\\\\\\ (93.3)}} & \\\\footnotesize Graph-based manifold learning is explored to mine the matching and non-matching pairs in unordered datasets. \\\\\\\\ \\\\cline{2-13} \\n\\\\Xhline{1.0pt}\\n\\\\end{tabular}\\n}\\n\\\\end{table*}\\n\\\\textbf{Evaluation for single feedforward pass.} \\\\textcolor{black}{In general, we observe that fully-connected layers used as feature extractors may give a lower accuracy (\\\\eg 74.7\\\\% on Holidays in ), compared to using convolutional layers in Table \\\\ref{Table_retrieval_off_the_shelf}. For the case where the same VGG net is used, the way to embed or aggregate features is critical. The methods shown in Figure \\\\ref{SinglePass} improve the discrimination of convolutional feature maps and perform differently in Table \\\\ref{Table_retrieval_off_the_shelf}, 66.9\\\\% of R-MAC and 58.9\\\\% of SPoC on Oxford-5k, differences which we see as critical factors for further analysis. If embedded by a BoW model, the results are competitive on Oxford-5k and Paris-6k (73.9\\\\% and 82.0\\\\%, respectively), while its codebook size is 25k, which may affect retrieval efficiency. Moreover, layer-level feature fusion improves retrieval accuracy. Yu \\\\etal combine three layers (mAP of 91.4\\\\% on Holidays), outperforming the performance of non-fusion method (mAP of 80.2\\\\%).}\\n\\\\textbf{Evaluation for multiple feedforward pass.}\\nResults for the methods of Figure \\\\ref{MultiplePass} are reported in Table \\\\ref{Table_retrieval_off_the_shelf}. Among them, extracting image patches densely using VGG has the highest performance on the 4 datasets , and rigid grid with BoW encoding is competitive (mAP of 87.2\\\\% on Paris-6k). These two methods consider more patches, even background information, when used for feature extraction. Instead of generating patches densely, region proposals and spatial pyramid modeling introduce a degree of purpose and efficiency in processing image objects. Spatial information is better maintained using multiple-pass schemes than with single-pass. For example, a shallower network (AlexNet) and region proposal networks in have a UKBench N-Score of 3.81, higher than using deeper networks ,,. Besides feeding image patches into the same network, model-level fusion also exploits complementary spatial information to improve accuracy. For instance, as reported in , which combines AlexNet and VGG, the results on Holidays (81.7\\\\% of mAP) and UKBench (3.32 of N-Score) are better than these in (76.75\\\\% and 3.00, respectively). \\n\\\\textbf{Evaluation for supervised fine-tuning.} Compared to off-the-shelf models, fine-tuning deep networks usually improves accuracy, see Table \\\\ref{Metric_learning_table_B}. For instance, the result on Oxford-5k by using a pre-trained VGG is improved from 66.9\\\\% to 81.5\\\\% in when a single-margin Siamese loss is used. Similar trends can be also observed on the Paris-6k dataset. For classification-based fine-tuning, its performance may be improved by using powerful DCNNs and feature enhancement methods such as the attention mechanism in , with an mAP increased from 55.7\\\\% in to 83.8\\\\% in on Oxford-5k. As for pairwise ranking loss fine-tuning, in some cases the loss used for fine-tuning is essential for performance improvement. For example, RPN is re-trained using regression loss on Oxford-5k and Paris-6k (75.1\\\\% and 80.7\\\\%, respectively) . Its results are lower than the results from (88.2\\\\% and 88.2\\\\%, respectively) where a transformation matrix is used to learn visual similarity. However, when RPN is trained by using triplet loss such as , the effectiveness of retrieval is improved significantly where the results are 86.1\\\\% (on Oxford-5k) and 94.5\\\\% (on Paris-6k). Feature embedding methods are important for retrieval accuracy; Ong \\\\etal embedded \\\\textit{Conv5} feature maps by Fisher Vector and achieved an mAP of 81.5\\\\% on Oxford-5k, while embedding feature maps by using VLAD achieves an mAP of 62.5\\\\% on this dataset ,.\\n\\\\textbf{Evaluation for unsupervised fine-tuning.} \\nCompared to supervised fine-tuning, unsupervised fine-tuning methods are relatively less explored. The difficulty for unsupervised fine-tuning is to mine sample relevance without ground-truth labels. In general, unsupervised fine-tuning methods should be expected to have lower performance than supervised. For instance, supervised fine-tuning using Siamese loss achieves an mAP 88.4\\\\% on Holidays, while unsupervised fine-tuning using the same loss function in ,, achieves 82.5\\\\%, 83.1\\\\%, and 87.5\\\\%, respectively. However, unsupervised fine-tuning methods can achieve a similar accuracy, even outperform the supervised fine-tuning, if a suitable feature embedding method is used. For instance, Zhao \\\\etal explore global feature structure modeling the manifold learning, producing an mAP of 85.4\\\\% (on Oxford-5k) and 96.3\\\\% (on Paris-6k), which is similar to supervised results of 86.1\\\\% (on Oxford-5k) and 94.5\\\\% (on Paris-6k). As another example, the precision of ResNet-101 fine-tuned by cross entropy loss achieves 83.8\\\\% on Oxford-5k , while the precision is further improved to 92.0\\\\% when an IME layer is used to embed features and fine-tuned in an unsupervised way . Note that fine-tuning strategies are related to the type of the target retrieval datasets. As demonstrated in Table \\\\ref{GLD_evaluation_on_training_dataset} and , fine-tuning on different datasets may produce a different final retrieval performance. \\n\\\\textbf{Network depth.} We compare the efficacy of DCNNs by depth, following the fine-tuning protocols\\\\footnote{https://github.com/filipradenovic/cnnimageretrieval-pytorch} in . For fair comparisons, all convolutional features from these backbone DCNNs are aggregated by MAC , and fine-tuned by using the same loss function with the same learning rate, thus the adopted methods are the same except for the DCNN depth. We use the default feature dimension (\\\\emph{i.e.} AlexNet (256), VGG (512), GoogLeNet (1024), ResNet-50/101 (2048)). The results are reported in Figure \\\\ref{CNNandaggregation_depth_dim_dataset} (a). We observe that the deeper networks consistently lead to better accuracy due to extracting more discriminative features.\\n\\\\textbf{Feature aggregation methods.} The methods of embedding convolutional feature maps were illustrated in Figure \\\\ref{SinglePass}. We use the off-the-shelf VGG (without updating parameters) on the Oxford and Paris datasets. The results are reported in Figure \\\\ref{CNNandaggregation_depth_dim_dataset} (b). We observe that the different ways to aggregate the same off-the-shelf DCNN leads to differences in retrieval performance. These reported results provide a reference for feature aggregation when one uses convolutional layers for performing retrieval tasks.\\n\\\\textbf{Global feature dimension.} We add fully-connected layers on the top of pooled convolutional features of ResNet-50 to obtain global descriptors with their dimensions varying from 32 to 8192. The results of 5 datasets are shown in Figure \\\\ref{CNNandaggregation_depth_dim_dataset} (c). It is expected that higher-dimension features usually capture more semantics and are helpful for retrieval. The performance\\ntends to be stable when the dimension is very large.\\n\\\\textcolor{black}{\\\\textbf{Number of image regions.} We compare the retrieval performance when different number of regions are fed and other components are kept the same, as depicted in Figure \\\\ref{CNNandaggregation_depth_dim_dataset} (d). Convolutional features of each region are pooled as 2048-dim regional features by MAC and then aggregated into a global one. Note that the final memory requirement is identical for the case that a holistic image is used as input (\\\\ie regarded as the case where only one region is used). Regional inputs on an image are extracted with a 40\\\\% overlap of neighboring regions and the number varying from 1 to 41. For Oxford-5k, the best result is given by the case where 9 image regions are used. For the rest datasets, 3 image regions give the best results. Finally, more regions extracted from one image decline the retrieval mAP. A reason is that features of background or irrelevant regions have also been aggregated, and negatively affect the performance.}\\n\\\\begin{table}[t]\\n\\\\caption{Evaluations of training sets and retrieval reranking. Numerical results are cited from .\\n}\\n\\\\vspace{-1em}\\n\\\\label{GLD_evaluation_on_training_dataset}\\n\\\\footnotesize\\n\\\\setlength{\\\\tabcolsep}{0.6mm}\\n\\\\begin{tabular}{|cc|c|c|cc|c|c|c|}\\n\\\\hline\\n\\\\multicolumn{2}{|c|}{\\\\multirow{3}{*}{Conditions}} \\n & \\\\multirow{3}{*}{Global} & \\\\multirow{3}{*}{\\\\begin{tabular}[c]{@{}c@{}}Local\\\\\\\\ reranking\\\\end{tabular}} & \\\\multicolumn{2}{c|}{Training set} & \\\\multirow{3}{*}{$\\\\mathcal{R}$Oxf} & \\\\multirow{3}{*}{$\\\\mathcal{R}$Par} & \\\\multirow{3}{*}{\\\\begin{tabular}[c]{@{}c@{}}GLD-v2\\\\\\\\ testing\\\\end{tabular}} \\\\\\\\ \\\\cline{5-6}\\n & & & & \\\\multicolumn{1}{c|}{GLD-v1} & \\\\begin{tabular}[c]{@{}c@{}}GLD-v2\\\\\\\\ -clean\\\\end{tabular} & & & \\\\\\\\ \\\\hline\\n\\\\multicolumn{1}{|c|}{\\\\multirow{4}{*}{\\\\begin{tabular}[c]{@{}c@{}}ResNet\\\\\\\\ -50\\\\end{tabular}}} & Case 1 & \\\\tickYes & \\\\tickNo & \\\\multicolumn{1}{c|}{\\\\tickYes} & \\\\tickNo & 45.1 & 63.4 & 20.4 \\\\\\\\ \\\\cline{2-9} \\n\\\\multicolumn{1}{|c|}{} & Case 2 & \\\\tickYes & \\\\tickNo & \\\\multicolumn{1}{c|}{\\\\tickNo} & \\\\tickYes & 51.0 & 71.5 & 24.1 \\\\\\\\ \\\\cline{2-9} \\n\\\\multicolumn{1}{|c|}{} & Case 3 & \\\\tickYes & \\\\tickYes & \\\\multicolumn{1}{c|}{\\\\tickYes} & \\\\tickNo & 54.2 & 64.9 & 22.3 \\\\\\\\ \\\\cline{2-9} \\n\\\\multicolumn{1}{|c|}{}& Case 4 & \\\\tickYes & \\\\tickYes & \\\\multicolumn{1}{c|}{\\\\tickNo} & \\\\tickYes & 57.9 & 71.0 & 24.3 \\\\\\\\ \\\\hline\\n\\\\multicolumn{1}{|c|}{\\\\multirow{4}{*}{\\\\begin{tabular}[c]{@{}c@{}}ResNet\\\\\\\\ -101 \\\\end{tabular}}} & Case 5 & \\\\tickYes & \\\\tickNo & \\\\multicolumn{1}{c|}{\\\\tickYes} & \\\\tickNo & 51.2 & 64.7 & 21.7 \\\\\\\\ \\\\cline{2-9} \\n\\\\multicolumn{1}{|c|}{} & Case 6 & \\\\tickYes & \\\\tickNo & \\\\multicolumn{1}{c|}{\\\\tickNo} & \\\\tickYes & 55.6 & 72.4 & 26.0 \\\\\\\\ \\\\cline{2-9} \\n\\\\multicolumn{1}{|c|}{} & Case 7 & \\\\tickYes & \\\\tickYes & \\\\multicolumn{1}{c|}{\\\\tickYes} & \\\\tickNo & 59.3 & 65.5 & 24.3 \\\\\\\\ \\\\cline{2-9} \\n\\\\multicolumn{1}{|c|}{} & Case 8 & \\\\tickYes & \\\\tickYes & \\\\multicolumn{1}{c|}{\\\\tickNo} & \\\\tickYes & 64.0 & 72.8 & 26.8 \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\end{table}\\n\\\\textbf{Fine-tuning datasets and retrieval reranking.} We compare performance on $\\\\mathcal{R}$Oxford-5k, $\\\\mathcal{R}$Paris-6k, and GLD-v2, aiming at comparing the role of different fine-tuning training sets and the effectiveness of retrieval reranking. Table \\\\ref{GLD_evaluation_on_training_dataset} lists 8 experimental scenarios using two network backbones, as in . \\n\\\\textcolor{black}{Since GLD-v2 provides class-level ground-truth, its including images show large context diversity and may pose challenges to the network fine-tuning. Thus, the pre-processing steps, as proposed in ,, are necessary to select the more coherent images, referring to the GLD-v2-clean subset. As a result, when using the global features only, this cleaned version of the training set improves the performance, as observed in Cases 1/5 and Cases 2/6 for ResNet-50/ResNet-101, respectively.} As an important postprocessing strategy, reranking further boosts the accuracy after the initial filtering step by using global features.\\n \\\\begin{figure*}[!t]\\n\\\\centering \\n {\\n \\\\includegraphics[width=\\\\textwidth]{./Figures/CNNandaggregation_depth_dim_dataset.pdf} \\n } \\n \\\\vspace{-2em}\\n \\\\caption{(a) The effectiveness of different DCNNs; (b) Comparison of the feature aggregation methods in Figure \\\\ref{SinglePass}; (c) The impact of global feature dimension by using ResNet-50; (d) Performance comparison when aggregating different numbers of image regions.} \\n \\\\label{CNNandaggregation_depth_dim_dataset} \\n \\\\end{figure*}\",\n \"id\": \"4289168f-a812-4899-b8a0-52669cd0ba36\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 53,\n \"parent_id\": \"a6c69e00-d8a1-492b-8514-fab570f109d2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"State of the Art Performance\"\n ],\n [\n \"subsection\",\n \"Performance Comparison and Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Performance Comparison and Analysis\"\n },\n {\n \"cite_extract_rate\": 0.7272727272727271,\n \"cites\": [\n 2141,\n 2136,\n 2144,\n 2142,\n 2110,\n 2093,\n 2111,\n 2143\n ],\n \"content\": \"\\\\label{Conclusion_and_Future_Directions}\\n\\\\textcolor{black}{As a comprehensive yet timely survey on instance retrieval using deep learning, this paper has discussed the main challenges, presented a taxonomy of recent developments according to their roles in implementing instance retrieval, highlighted the recent representative methods and analyzed their merits and demerits, discussed the datasets, evaluation protocols, and SOTA performance. Nowadays the exponentially increasing amount of image and video data due to surveillance, e-commerce, medical images, handheld devices, robotics, \\\\etc, offers an endless potential for applications of instance retrieval. \\nAlthough significant progress has been made, as discussed in Section \\\\ref{Keychallenges}, the main challenges in instance retrieval have not been fully addressed. Below we identify a number of promising directions for\\nfuture research. }\\n\\\\textcolor{black}{\\\\textbf{(1) Accurate and robust feature representations.} \\nOne of the main challenges in instance retrieval is the large intra-class variations due to changes in viewpoint, scale, illumination, weather condition and background clutter \\\\emph{etc.}, as we discussed in Section~\\\\ref{Keychallenges}. However, DCNN representations have very little invariance, even though trained with lots of augmented data . Fortunately, before deep learning, with instance retrieval there are lots of important ideas in handling such intra-class variations like local interest point detectors and local invariant descriptors. Therefore, it is worth enabling DCNN to learn more accurate and robust representations via leveraging such traditional wisdom to design better DCNNs. In addition, unlike most objects in existing benchmarks which are rigid, planar and textured, textureless objects, 3D objects, transparent objects, reflective surfaces, \\\\etc are still very hard to find.}\\n\\\\textcolor{black}{In addition, pursuing accuracy alone is not sufficient, as instance retrieval systems should be able to resist potential adversarial attacks. \\nRecently, deep networks have been proven to be fooled rather easily by adversarial examples , \\n\\\\ie images added with intentionally designed yet nearly imperceptible perturbations, which raises serious safety and robustness concerns. However, adversarial robustness in instance retrieval , has received very little attention, and should merit further effort.}\\n\\\\textcolor{black}{\\\\textbf{(2) Compact and efficient deep representations.} In instance retrieval, \\nsearching efficiently is as critical as searching accurately, especially for the pervasive \\nmobile or wearable devices with very limited computing resources. However, existing methods adopt large scale, energy hungry DCNNs that are very difficult to be deployed in mobile devices. Hence, there has been pressing needs to develop compact, efficient, yet reusable deep representations tailored to the resource limited devices, like using binary neural networks ,,.}\\n\\\\textcolor{black}{\\\\textbf{(3) Learning with fewer labels.} Deep learning require a large amount of high-quality labeled dataset to achieve high accuracy. The presence of labels errors or the limited amount of labeled data can greatly degrade DCNN's accuracy. However, collecting massive amounts of accurately labeled data is costly. In practical scenarios, datasets like GLDv2 , have long-tailed distributions, and noisy labels. Thus, to address such limitations, few shot learning , self-supervised learning , \\nimbalanced learning , \\n noisy label aware learning \\\\etc should be paid more attention in instance retrieval in the future. }\\n\\\\textcolor{black}{\\\\textbf{(4) Continual learning for instance retrieval.}\\nIn specific, the current IIR methods make restrictive assumptions, such as the training data being enough and stationary, retraining from scratch being possible when new data becomes available, which is problematic in realistic conditions. Our living world is continuously varying, and in\\ngeneral data distributions are often non-stationary, new data may be added, and\\npreviously unseen classes may be encountered. Thus, continual learning plays a vital role in continuously updating the IIR systems. The key issues are how to retain and utilize the previously learned knowledge, how to update the retrieval system as new images becomes available, and how to learn and improve over time.\\n}\\n\\\\textbf{(5) Privacy-aware instance retrieval.} Most IIR systems concentrate on improving the accuracy or efficiency performance, and the higher performance might come at the cost of users' privacy. Therefore, in some cases, such as personalized search systems, the privacy protection problem is also an important issue to be considered. Deep models should be privacy-aware and protect users' personalized searching experience to avoid their worries about using such IIR systems.\\n\\\\textbf{(6) Video instance retrieval.} Searching a specific instance in an image cannot always meet the requirements in some scenarios such as the video surveillance system in the field of searching criminals. Currently, with the rapid growth of video data, retrieving a certain object, place, or action in videos has become more and more important and highly necessary in the future. For video instance retrieval, 3D-CNNs models need to be built to learn video’s spatio-temporal representations to compute the semantic similarity of instances.\\n\\\\footnotesize\\n\\\\bibliographystyle{IEEEtran}\\n\\\\bibliography{IEEEabrv,myreference}\\n\\\\vspace{-2.0 cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/WeiChen.pdf}}]{Wei Chen} received the Ph.D. degree at Leiden University, in 2021. Before starting with PhD study in Leiden University, he received his Master degree from the National University of Defense Technology, China, in 2016. His research interest focuses on cross-modal retrieval with deep learning methods, and also in the context of incremental learning. He has published papers in international conferences and journal including CVPR, ACM MM, PR, Neurocomputing, and IEEE TMM \\\\etc\\n\\\\end{IEEEbiography}\\n\\\\vspace{-1.7 cm} \\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Yu_Liu.pdf}}]{Yu Liu} is currently an Associate Professor in International School\\nof Information Science and Engineering at Dalian University of Technology, China.\\nHe was a post-doctoral researcher in the PSI group of KU Leuven, Belgium. \\nIn 2018, he received the PhD degree from Leiden University, Netherlands.\\nHis research interests include object recognition and retrieval in the context of continual learning and zero-shot learning. \\nHe has co-organized several workshops at ICCV, CVPR and ECCV, respectively. \\nHe has published papers in CVPR, ICCV, ECCV, ACM MM, IEEE TIP, etc, and \\nreceived the best paper award at MMM2017.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-1.7 cm} \\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Weiping_Wang.pdf}}]{Weiping Wang} received the Ph.D. degree in systems engineering from the National University of Defense Technology, Changsha, China. He is a Professor at Academy of Advanced Technology Research of Hunan. He has more than 200 papers published on journals and conferences including IEEE Transactions multimedia, Transactions on Vehicular Technology, \\\\etc His current research interests focus on intelligent decision experimentation, multi-modal knowledge extraction and knowledge integrated computation.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-2.0 cm} \\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Erwin.pdf}}]{Erwin M. Bakker} is co-director of the LIACS Media Lab at Leiden University. He has published widely in the fields of image retrieval, audio analysis and retrieval and bioinformatics. He was closely involved with the start of the International Conference on Image and Video Retrieval (CIVR). Moreover, he regularly serves as a program committee member or organizing committee member for scientific multimedia and human-computer interaction conferences and workshops.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-2.0 cm} \\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Theodoros_Georgiou.pdf}}]{Theodoros Georgiou} received the Ph.D. degree at Leiden University, in 2021. His research interest focuses on deep learning methods applied on higher than two dimensional data. Before starting with PhD study in Leiden University, he received his Master degree from the Leiden University in 2016. He has published papers in international conferences and journal including ICPR, CBMI, WCCI and IJMIR.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-2.0 cm} \\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Paul_Fieguth.pdf}}]{Paul Fieguth} is co-director of the Vision \\\\& Image Processing Group in Systems Design Engineering at the University of Waterloo, where he is Professor and Associate Dean. He received the Ph.D.\\\\ degree from the Massachusetts Institute of Technology, Cambridge, in 1995, and has held visiting appointments at the University of Heidelberg in Germany, at INRIA/Sophia in France, at the Cambridge Research Laboratory in Boston, at Oxford University and the Rutherford Appleton Laboratory in England. His research interests include statistical signal and image processing, hierarchical algorithms, data fusion, machine learning, and the interdisciplinary applications of such methods. He has published textbooks on Statistical Image Processing and on Complex Systems theory.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-1.7 cm} \\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/LiSmileGoodFace.pdf}}]{Li Liu} received the Ph.D. degree in information and communication engineering from the National University of Defense Technology, China, in 2012. She is currently a Professor with the College of System Engineering. She has held visiting appointments at the University of Waterloo, Canada, at the Chinese University of Hong Kong, and at the University of Oulu, Finland. Her current research interests include computer vision, pattern recognition and machine learning. Her papers have currently over 7800 citations in Google Scholar.\\n\\\\end{IEEEbiography}\\n\\\\vspace{-1.7 cm}\\n\\\\begin{IEEEbiography}[{\\\\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{./Figures/Michael_Lew.pdf}}]{Michael S. Lew} is the head of the Deep Learning and Computer Vision Research Group and a full Professor at LIACS, Leiden University. He has published over a dozen books and 190 peer-reviewed scientific articles in the areas of image retrieval, computer vision, and deep learning. Notably, he had the most cited paper in the ACM Transactions on Multimedia and one of the top 10 most cited articles in the history (out of more than 16,000 articles) of the ACM SIGMM. He was also a founding member of the advisory committee for the TRECVID video retrieval evaluation project, chair of the steering committee for the ACM International Conference on Multimedia Retrieval and a member of the ACM SIGMM Executive Committee. \\n\\\\end{IEEEbiography}\\n\\\\vspace{-1.7 cm} \\n \\\\newpage\\n\\\\end{document}\",\n \"id\": \"6d16c13c-e552-486e-8c1d-7e8590b69688\",\n \"level\": \"section\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"515509a6-319a-4385-b1b5-4418a83d2bc2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Instance Retrieval: A Survey\"\n ],\n [\n \"section\",\n \"Conclusions and Outlooks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusions and Outlooks\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":96,"string":"96"},"all_cites":{"kind":"list like","value":[2093,2092,2091,2094,2090,97,8490,2089,2095,209,7543,2101,514,2096,2098,2102,2105,2108,2104,2103,8524,7544,2109,2099,2097,2106,2107,629,2100,7545,2110,2112,2111,2113,2115,2114,2118,2119,2116,2117,8525,7546,2120,2121,7547,2122,7548,2124,2123,2125,2126,2127,2128,2129,309,2130,7549,2131,2132,9094,2135,2133,2134,2136,2137,2140,8526,2139,2138,2141,2144,2142,2143],"string":"[\n 2093,\n 2092,\n 2091,\n 2094,\n 2090,\n 97,\n 8490,\n 2089,\n 2095,\n 209,\n 7543,\n 2101,\n 514,\n 2096,\n 2098,\n 2102,\n 2105,\n 2108,\n 2104,\n 2103,\n 8524,\n 7544,\n 2109,\n 2099,\n 2097,\n 2106,\n 2107,\n 629,\n 2100,\n 7545,\n 2110,\n 2112,\n 2111,\n 2113,\n 2115,\n 2114,\n 2118,\n 2119,\n 2116,\n 2117,\n 8525,\n 7546,\n 2120,\n 2121,\n 7547,\n 2122,\n 7548,\n 2124,\n 2123,\n 2125,\n 2126,\n 2127,\n 2128,\n 2129,\n 309,\n 2130,\n 7549,\n 2131,\n 2132,\n 9094,\n 2135,\n 2133,\n 2134,\n 2136,\n 2137,\n 2140,\n 8526,\n 2139,\n 2138,\n 2141,\n 2144,\n 2142,\n 2143\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.761755724759363,"string":"0.761756"}}},{"rowIdx":64,"cells":{"authors":{"kind":"list like","value":["Amir Rasouli"],"string":"[\n \"Amir Rasouli\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Deep Learning for Vision-based Prediction: A Survey"},"year":{"kind":"string","value":"2020"},"date":{"kind":"timestamp","value":"2020-06-30T20:26:46Z","string":"2020-06-30T20:26:46Z"},"category":{"kind":"string","value":"cs.CV"},"abstract":{"kind":"string","value":" Vision-based prediction algorithms have a wide range of applications including autonomous driving, surveillance, human-robot interaction, weather prediction. The objective of this paper is to provide an overview of the field in the past five years with a particular focus on deep learning approaches. For this purpose, we categorize these algorithms into video prediction, action prediction, trajectory prediction, body motion prediction, and other prediction applications. For each category, we highlight the common architectures, training methods and types of data used. In addition, we discuss the common evaluation metrics and datasets used for vision-based prediction tasks. A database of all the information presented in this survey, cross-referenced according to papers, datasets and metrics, can be found online at \\url{https://github.com/aras62/vision-based-prediction}. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"]],"subsections":["2fec0d8e-0911-43a3-a218-2d210969c7e7","cbf20d18-4b97-4aff-b65c-6034395a9fa5","80143ee1-3a53-4ec6-96bf-ec81cc495f85","040ae750-d27c-4e06-99d2-48f8312f00ca","991d18f2-ffd8-45b8-b2e8-b4c21fb20720","ea670b52-fc46-4dcb-ac08-efd4b23b03c3","77155bc6-d46c-48ec-a4ba-c2d738e13396","374ffbdc-d4ff-461d-a063-67f03516e35e","5bb03330-9009-4e51-abdd-12e3b8d3505d","4c69ee31-a62c-468f-a731-a0d63c6e9fe2","6ab7b3e0-1f9f-4a59-bcf5-66c554040f7e","2dbff660-6f4e-44d8-90a2-fa6081f28574","23d750d7-3a7a-460d-8e6b-4f54e379529f","63a0cc61-98a6-4be7-b982-ebf42e368ec1","14c8fcc2-4517-4a60-9512-dfff6198b563","e82c2ffc-f5dc-4d01-9184-399825652bbe","a33fb5e9-81a5-4a37-8f37-6dd6031d77ae"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:introduction}}\n\\IEEEPARstart{T}{he} ability to predict the changes in the environment and the behavior of objects is fundamental in many applications such as surveillance, autonomous driving, scene understanding, etc. Prediction is a widely studied field in various artificial intelligence communities. A subset of these algorithms relies primarily on visual appearances of the objects and the scene to reason about the future. Other approaches use different forms of sensors such as wearable or environmental sensors to learn about the past states of the environment or objects. \nThe focus of this report is on vision-based prediction algorithms, which primarily use visual information to observe the changes in the environment and predict the future. In this context, prediction can be in the form of generating future scenes or reasoning about specific aspects of the objects, e.g. their trajectories, poses, etc.\nFor this review, we divide the prediction algorithms into five groups, namely video prediction, action prediction, trajectory prediction, motion (pose) prediction, and others which involve various applications of prediction such as trend prediction, visual weather prediction, map prediction, semantic prediction, etc. In addition, we briefly discuss algorithms that use a form of prediction as an intermediate step to perform tasks such as object detection, action detection, and recognition, etc. Moreover, for each group of prediction algorithms, we will talk about the common datasets and metrics and discuss of their characteristics. It should be noted that due to the broad scope of this review and the large body of work on the vision-based prediction, this review will only focus on works that had been published since five years ago in major computer vision, robotics and machine learning venues. In addition, as the title of the paper suggests, the main focus of the discussion will be on deep learning methods given their popularity in recent years.","id":"2fec0d8e-0911-43a3-a218-2d210969c7e7","level":"section","origin_cites_number":0,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"Before reviewing the works on vision-based prediction algorithms, there are a number of points that should be considered.","id":"cbf20d18-4b97-4aff-b65c-6034395a9fa5","level":"section","origin_cites_number":0,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Vision-based Prediction"]],"subsections":["ffddf8b2-9c16-4742-b890-9fae3c51a950","d51e120c-5adf-414f-9f19-52ce0c4182c7"],"title":"Vision-based Prediction"},{"cite_extract_rate":0,"cites":[],"content":"Based on our review, we have identified four major vision-based applications namely, video prediction, action prediction, trajectory prediction, and motion prediction. We discuss each of the studies in each category in a dedicated section. Some of the prediction works, such as visual weather prediction, semantic prediction, contests outcome prediction, that do not fit to any of the four major categories are presented in other application section. \nSome works address multiple prediction tasks, e.g. predicting trajectories and actions simultaneously, and therefore might fall in more than one category. It should be noted that we only include an algorithm in each category if the corresponding task is directly evaluated. For instance, if an algorithm performs video prediction for future action classification, and only evaluates the accuracy of predicted actions, it will only appear in the action prediction category. Furthermore, some works that are reviewed in this paper propose multiple architectures, e.g. recurrent and feedforward, for solving the same problem. In architecture-based categorizations, these algorithms may appear more than once.","id":"ffddf8b2-9c16-4742-b890-9fae3c51a950","level":"subsection","origin_cites_number":0,"parent_id":"cbf20d18-4b97-4aff-b65c-6034395a9fa5","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Vision-based Prediction"],["subsection","Applications"]],"subsections":[],"title":"Applications"},{"cite_extract_rate":0,"cites":[],"content":"","id":"d51e120c-5adf-414f-9f19-52ce0c4182c7","level":"subsection","origin_cites_number":0,"parent_id":"cbf20d18-4b97-4aff-b65c-6034395a9fa5","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Vision-based Prediction"],["subsection","Methods"]],"subsections":["f46c6d90-8842-4e11-a89d-e5ce565f4be9","56ce5010-908f-43e2-ac25-ca9c33b72d32","fa1f4460-f37c-4c55-8b41-c1ea861edc31"],"title":"Methods"},{"cite_extract_rate":0,"cites":[],"content":"This work focuses on vision-based algorithms, which use some form of visual input such as RGB camera images or active sensors such as LIDAR. It should be noted that many algorithms, especially trajectory prediction ones, only use ground truth data such as object trajectories without actual visual processing, e.g. for detection of objects. However, as long as these algorithms are evaluated on vision datasets, they are included in this paper. Note that a completer list of papers with published code can be found in Appendix \\ref{papers_with_code}.","id":"f46c6d90-8842-4e11-a89d-e5ce565f4be9","level":"subsubsection","origin_cites_number":0,"parent_id":"d51e120c-5adf-414f-9f19-52ce0c4182c7","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Vision-based Prediction"],["subsection","Methods"],["subsubsection","Algorithms"]],"subsections":[],"title":"Algorithms"},{"cite_extract_rate":0,"cites":[],"content":"As mentioned earlier, we focus on algorithms that have a deep learning component, either in the stage of visual representation generation (e.g. using convolutional features) or reasoning (e.g. using an MultiLayer Preceptron (MLP) for classification). We will, however, acknowledge the classical methods by mentioning some of the main techniques and including them in the datasets and metrics sections of this paper.\nWe classify the algorithms in terms of training techniques and architectures. In practice, this is very challenging as the majority of algorithms use a combination of different approaches. For example, recurrent networks often rely on a form of Convolutional Neural Networks (CNNs) to generate feature representations for scenes, poses of agents, etc. To better distinguish between different classes of algorithms, we only focus on the core component of each algorithm, i.e. the parts that are used for reasoning about the future. Hence, for example, if an algorithm uses a CNN model for pre-processing input data and a recurrent network for temporal reasoning, we consider this algorithm as recurrent. On the other hand, if the features are used with a fully connected network, we categorize this algorithm as feedforward or one-shot method. A few algorithms propose the use of both architectures for reasoning. We address those methods as hybrid. In addition, it should be noted that many works propose alternative approaches using each architecture. Therefore, we categorize them in more than one group.","id":"56ce5010-908f-43e2-ac25-ca9c33b72d32","level":"subsubsection","origin_cites_number":0,"parent_id":"d51e120c-5adf-414f-9f19-52ce0c4182c7","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Vision-based Prediction"],["subsection","Methods"],["subsubsection","Architectures"]],"subsections":[],"title":"Architectures"},{"cite_extract_rate":0,"cites":[],"content":"As one would expect, vision-based algorithms primarily rely on visual information. However, many algorithms use pre-trained off-the-shelf algorithms to transform the input to some explicit feature spaces, e.g. poses, trajectories, action labels and perform reasoning in those feature spaces. If pre-processing is not part of the main algorithm, we consider those secondary features as different types of data inputs to the algorithms. If some basic processing, e.g. generating convolutional features for a scene is used, we consider the data type of the original input, e.g. RGB images.","id":"fa1f4460-f37c-4c55-8b41-c1ea861edc31","level":"subsubsection","origin_cites_number":0,"parent_id":"d51e120c-5adf-414f-9f19-52ce0c4182c7","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Vision-based Prediction"],["subsection","Methods"],["subsubsection","Data type"]],"subsections":[],"title":"Data type"},{"cite_extract_rate":0.5897435897435891,"cites":[2804,2814,2817,2807,2808,2816,5680,2819,2815,2806,2809,2818,2811,2802,2820,2810,2805,2812,8590,7617,2803,2813,2073],"content":"Video or future scene prediction can be regarded as the most generic form of prediction. The objective of video prediction algorithms is to generate future scenes, often in the form of RGB images and/or optical flow maps . The generated images in turn can be used for various tasks such as action prediction , event prediction , flow estimation , semantic segmentation , etc.\nVideo prediction applications rely on generative models whose task is to predict future scene(s) based on a short observation of input sequences (or in some cases only a single image ). Although many approaches use feedforard architectures , the majority of algorithms take advantage of Recurrent Neural Networks (RNNs) such as Gated Recurrent Units (GRUs) , Long-Short Term Memory (LSTM) networks , its variation Convolutional LSTMs (ConvLSTMs) or a combination of these .\nGenerative Adversarial Networks (GANs) are particularly popular in the video prediction community. . In these adversarial training frameworks, there are two compoents: A generative network that produces future representations and a discriminator whose objective is to distinguish between the predicted representations (e.g. optical flow , frames , motion ) or their temporal consistency and the actual ground truth data by producing a binary classification score that indicates whether the prediction is real or fake. While many algorithms use discriminators to judge how realistic the final generated images are or intermediate features (e.g. poses ), others use multiple discriminators at different stages of processing. For example, the authors of use two discriminators, one is responsible for judging the temporal consistency of the generated frames (i.e. whether the order of generated frames is real) and the other assesses whether the generated frames are real or not. Lee et al. use three discriminators to assess the quality of generated frames and the intermediate motion and content features. Using a two-stream approach, the method in produces both the next frame and optical flow and each stream is trained with a separate discriminator. The prediction network of uses a discriminator for intermediate features generated from input scenes and another discriminator for the final results.\nVariational Autoencoders (VAEs) or Conditional VAEs (CVAEs) are also used in some approaches . VAEs model uncertainty in generated future frames by defining a posterior distribution over some latent variable space . In CVAEs, the posterior is conditioned on an additional parameter such as the observed action in the scenes or initial observation . Using VAEs, at inference time, a random sample is drawn from the posterior to generate the future frame.\nMany video prediction algorithms operate solely on input images and propose various architectural innovations for encoding the content and generating future images . For example, the method in performs a two-way prediction, forward and backward. Each prediction relies on two discriminators for assessing the quality of the generated images and temporal consistency. The model presented in trains a context network by inputting an image sequence into a ConvLSTM whose output is used to initialize convolutional networks responsible for generating the next frames. Xu et al. , in addition to raw pixel values, encode the output of a high pass filter applied to the image as a means of maintaining the structural integrity of the objects in the scene. In , the authors use a two-step approach in which they first perform a coarse frame prediction followed by a fine frame prediction. In , the algorithm learns in two stages. A discriminator is applied after features are generated from the scenes and another one after the final generated frames. \nOptical flow prediction has been widely used as an intermediate step in video prediction algorithms . For example, to deal with occlusion in dynamic scenes, Gao et al. disentangle flow and pixel-level predictions into two steps: the algorithm first predicts the flow of the scene, and then uses it, in conjunction with the input frames, to predict the future. Similar multi-step approaches have also been used in . In , the authors use two separate branches: one branch receives two consecutive frames $(t, t+1)$ and produces context information. The second branch produces motion information by receiving two frames that are $k$ steps apart (i.e. $t+1$, $t+k$). The outputs of these two branches are fused and fed into the final scene generator. The method in simultaneously produces the next future frame and the corresponding optical flow map. In this architecture, two additional networks are used: A flow estimator which uses the output of the frame generator and the last observation to estimate a flow map and a warping layer which performs differential 2D spatial transformation to warp the last observed image into the future predicted frame according to the predicted flow map. \nSome algorithms rely on various intermediate steps for video prediction. For instance, the method in , reasons about the locations and features of individual entities (e.g. cubes) for final scene predictions. Kim et al. first identify keypoints, which may correspond to important structures such as joints, and then predict their motion. For videos involving humans, in the authors identify and reason about the changes in poses, and use this information to generate future frames. In , in addition to raw input images, the differences between consecutive frames used in the learning process.\nPrediction networks can also be provided with additional information to guide future frame generation. In an optical flow network and in a pose estimation network are used in addition to RGB images. Using a CVAE architecture, in the authors use the action lables as conditional input for frame generation. In the context of active tasks, e.g. object manipulation with robotic arms, in which the consequences of actions influence the future scene configuration, it is common to condition the future scene generation on the current or future intended actions .","id":"80143ee1-3a53-4ec6-96bf-ec81cc495f85","level":"section","origin_cites_number":39,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Video prediction"]],"subsections":["f163c039-4fc8-43c5-8ff4-fc6e32ae5edf"],"title":"Video prediction"},{"cite_extract_rate":0,"cites":[],"content":"Video prediction algorithms are based on generative models that produce future images given a short observation, or in extreme cases a single view of the scene. Both recurrent and feedforward models are widely used in the field, with recurrent ones being slightly more favorable. The architectural designs and training strategies such as the VAEs or GANs are very common. However, it is hard to establish which one of these approaches is superior given that the majority of the video prediction algorithms are application-agnostic, meaning that they are evaluated on a wide range of video datasets with very different characteristics such as traffic scenes, activities, games, object manipulations, etc. (more on this in Section \\ref{datasets}).\nDespite the great progress in the field, video prediction algorithms are still facing some major challenges. One of them is the ability to hallucinate, which is to generate visual representations for parts of the scenes that were not visible during observation phase, e.g. due to occlusions. This is particularly an issue for more complex images such as traffic scenes, movies, etc. The complexity of the scenes also determines how fast the generated images would degrade. Although these algorithms show promising results in simple synthetic videos or action sequences, they still struggle in real practical applications. In addition, many of these algorithms cannot reason about the expected presence or absence of objects in the future. For example, if a moving object is present in the observations and is about to exit the field of view in near future, the algorithms account for it in the future scenes as long as parts of it are visible in the observation stage. This can be an issue for safety-critical applications such as autonomous driving in which the presence or absence of traffic elements and the interactions between them are essential for action planning.","id":"f163c039-4fc8-43c5-8ff4-fc6e32ae5edf","level":"subsection","origin_cites_number":0,"parent_id":"80143ee1-3a53-4ec6-96bf-ec81cc495f85","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Video prediction"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0.17391304347826,"cites":[2821,2823,2822,2824],"content":"Action prediction algorithms can be categorized into two groups: Next action or event prediction (or action anticipation) and early action prediction. In the former category, the algorithms use the observation of current activities or scene configurations and predict what will happen next. Early action prediction algorithms, on the other hand, observe parts of the current action in progress and predict what this action is. The classical learning approaches such as Conditional Random Fields (CRFs) , Support Vector Machines (SVMs) with hand-crafted features , Markov models , Bayesian networks and other statistical methods have been widely used in recent years. However, as mentioned earlier, we will only focus on deep learning approaches.","id":"040ae750-d27c-4e06-99d2-48f8312f00ca","level":"section","origin_cites_number":23,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Action prediction"]],"subsections":["b0a73071-8988-459c-960a-fbc201a48a45","2ebd83ed-2292-4ab8-9334-3163c17703ba"],"title":"Action prediction"},{"cite_extract_rate":0.486486486486486,"cites":[2829,2837,7117,2838,2832,2828,2825,2806,2834,2835,2827,2826,2831,2836,2833,7618,2839,2830],"content":"Action prediction algorithms are used in a wide range of applications including cooking activities , traffic understanding , accident prediction , sports and other forms of activities . Although the majority of these algorithms use sequences in which the objects and agents are fully observable, a number of methods rely on egocentric scenes which are recorded from the point of view of the acting agents and only parts of their bodies (e.g. hands) are observable.\nAction prediction methods predominantly use a variation of RNN-based architectures including LSTMs , GRUs , ConvLSTMs , and Quasi-RNNs (QRNNs) . For instance, in the authors use a graph-based RNN architecture in which the nodes represent actions and the edges of the graph represent the transitions between the actions. The method in employs a two-step approach: using a recognition algorithm, the observed actions and their durations are recognized. These form a one-hot encoding vector which is fed into GRUs for the prediction of the future activities, their corresponding start time and length. In the context of vehicle behavior prediction, Ding et al. uses a two-stream GRU-based architecture to encode the trajectory of two vehicles and a shared activation unit to encode the vehicles mutual interactions. Scheel et al. encode the relationship between the ego-vehicle and surrounding vehicles in terms of their mutual distances. The vectorized encoding is then fed into a bi-directional LSTM. At each time step, the output of the LSTM is classified, using a softmax activation, into a binary value indicating whether it is safe for the ego-vehicle to change lane. In the authors use a QRNN network to capture the relationships between road users in order to predict the likelihood of a traffic accident. To train the model, the authors propose an adaptive loss function that assigns penalty weights depending on how early the model can predict accidents. \nAs an alternative to recurrent architectures, some algorithms use feedforward architectures using both 3D and 2D convolutional networks. For example, in the context of pedestrian crossing prediction, in the authors use a generative 3D CNN model that produces future scenes and is followed by a classifier. The method of detects and tracks pedestrians in the scenes, and then feeds the visual representations of the tracks, in the form of an image sequence, into a 3D CNN architecture, which directly classifies how likely the pedestrian will cross the road. To predict the time of traffic accidents, the method in processes each input image using a 2D CNN model and then combines the representations followed by a fully-conntected (\\textit{fc}) layer for prediction. Farha et al. create a 2D matrix by stacking one-hot encodings of actions for each segment of observation and use a 2D convolutional net to generate future actions encodings. Casas et al. use a two-stream 2D CNN, each processing the stacked voxelized LIDAR scans and the scene map. The feature maps obtained from each stream are fused and fed into a backbone network followed by three headers responsible for the detection of the vehicles and predicting their intentions and trajectories. For sports forecasting, Felsen et al. concatenate 5 image observations channel-wise and feed the resulting output into a 2D CNN network comprised of 4 convolutional layers and an \\textit{fc} layer.\nAlthough some algorithms rely on a single source of information, e.g. a set of pre-processed features from RGB images or trajectories , many algorithms use a multimodal approach by using various sources of information such as optical flow maps , poses , scene attributes (e.g. road structure, semantics) , text , action labels , length of actions , speed (e.g. ego-vehicle or surrounding agents) , gaze , current activities and the time of the actions. For example, the method in uses a multi-stream LSTM in which two LSTMs encode visual features and cooking recipes and an LSTM decodes them for final predictions. To capture the relationships within and between sequences, Gammulle et al. propose a two-stream LSTM network with external neural memory units. Each stream is responsible for encoding visual features and action labels. In , a multi-layer GRU structure is used in which features with different modalities enter the network at different levels and are fused with the previous level encodings. The fusion process is taking place according to the complexity of the data modality, e.g. more complex features such as encodings of pedestrian appearances enter the network at the bottom layer, whereas location and speed features enter at the second-last and last layers respectively. Farha et al. use a two-layer stacked GRU architecture which receives as input a feature tuple of the length of the activity and its corresponding one-hot vector encoding. In , the method uses a two-stage architecture: First information regarding the appearance of the scene, optical flow (pre-processed using a CNN) and vehicle dynamics are fed into individual LSTM units. Then, the output of these units is combined and passed through an \\textit{fc} layer to create a representation of the context. This representation is used by another LSTM network to predict future traffic actions. In the context of human-robot interaction, the authors of combine the information regarding the gaze and pose of the humans using an encoder-decoder LSTM architecture to predict their next actions. Jain et al. use a fusion network to combine head pose information of the driver, outside scene features, GPS information, and vehicle dynamics to predict the driver's next action.\nBefore concluding this section, it is important to discuss the use of attention modules which have gained popularity in recent years . As the name implies, the objective of attention modules is to determine what has to be given more importance at a given time. These modules can come in different froms and can be applied to different dimensions of data and at various processing. Some of these modules are temporal attention for identifying keyframes, modality attention to prioritize between different modalities of data input, spatial attention for highlighting the important parts of the scenes, and graph attention for weighting nodes of the graph. In some cases, a combination of different attention mechanisms is used .","id":"b0a73071-8988-459c-960a-fbc201a48a45","level":"subsection","origin_cites_number":37,"parent_id":"040ae750-d27c-4e06-99d2-48f8312f00ca","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Action prediction"],["subsection","Action anticipation"]],"subsections":["41132876-21fb-45f4-86a1-c905e9d97c64"],"title":"Action anticipation"},{"cite_extract_rate":0,"cites":[],"content":"In the field of action anticipation, RNN architectures are strongly preferred. Compared to feedforward algorithms, recurrent methods have the flexibility of dealing with variable observation lengths and multi-modal data, in particular, when they are significantly different, e.g. trajectories and RGB images. However, basic recurrent architectures such as LSTMs and GRUs rely on some forms of pre-processing, especially when dealing with high dimensional data such as RGB images, which requires the use of various convolutional networks, a process that can be computationally costly. Feedforward models, on the other hand, can perform prediction in one shot, meaning that they can simultaneously perform temporal reasoning and spatial feature generation in a single framework, and as a result, potentially have a shorter processing time. \nMany of the approaches mentioned earlier are generative in nature. They generate representations in some feature space and then using these representations predict what will happen next. Some algorithms go one step further and generate the actual future images and use them for prediction. Although such an approach seems effective for single actor events, e.g. cooking scenes, human-robot interaction, it is not a feasible approach for multi-agent predictions such as reasoning about behaviors of pedestrians or cars in traffic scenes.\nThe majority of the methods reviewed in this section use multi-modal data input. This seems to be a very effective approach, especially in high dimensional problems such as predicting road user behavior where the state (e.g. location and speed) of the road user, observer, and other agents, as well as scene structure, lighting conditions, and many other factors, play a role in predicting the future behavior.\nMulti-tasking, e.g. predicting actions and trajectories, are an effective way to predict future actions. For instance, trajectories can imply the possibility of certain actions, e.g. moving towards the road implies the possibility that the person might cross the street. As a result, the simultaneous learning of different tasks can be beneficial. \nLast but not least is the use of attention modules. These modules are deemed to be very effective , in particular for tasks with high complexity in terms of the modality of input data, the scene structure and temporal relations.","id":"41132876-21fb-45f4-86a1-c905e9d97c64","level":"subsubsection","origin_cites_number":0,"parent_id":"b0a73071-8988-459c-960a-fbc201a48a45","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Action prediction"],["subsection","Action anticipation"],["subsubsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0.41176470588235203,"cites":[2829,2841,2845,2844,2840,2843,2842],"content":"Similar to action anticipation methods, early action prediction algorithms widely use recurrent network architectures . Although many of these algorithms use similar aproaches to action anticipation algorithms, some propose new approaches. For example, in the authors use a teacher-student learning scheme where the teacher learns to recognize full sequences using a bi-directional LSTM and the student relies on partial videos using an LSTM network. They perform knowledge distillation by linking feature representations of both networks. Using GAN frameworks, the methods in predict future feature representations of videos in order to predict actions. Zhao et al. implement a Kalman filter using an LSTM architecture. In this method, actions are predcted after observing each frame and corrections are made if the next observation provide additional information. The method of uses a two-step LSTM architecture which first generates an encoding of the context using context-aware convolutional features and then combines these encodings with action-aware convolutional features to predict the action. The authors of this method propose a new loss function that penalizes false negatives with the same strength at any point in time and false positives with a strength that increases linearly over time, to reach the same weight as that on false negatives. In the authors perform coarse-to-fine-grained predictions depending on the amount of evidence available for the type of action. \nMany early action prediction methods adopt feedforward architectures . The authors of predict actions from a single image by transforming it into a new representation called Ranked Saliency Map and Predicted Optical Flow. This representation is passed through a 2D convent and \\textit{fc} layers for final prediction. In , the authors use temporal CNN (TCNNs) architectures, which are a series of 1D dilated convolutional layers designed to capture the temporal dependencies of feature representations, e.g. in the form of a vector representation of poses or word-embeddings describing the video frames. Chen et al. use features of body parts generated by a CNN model and train an attention module whose objective is to activate only features that contribute to the prediction of the action. In , the authors use an action detection framework, which incrementally predicts the locations of the actors and the action classes based on the current detections.\nThe majority of the early action prediction algorithms pre-process the entire observed scenes using, e.g. different forms of convolutional neural networks or other forms of features . Some algorithms complement these features using optical flow maps . Another group of action prediction methods focuses on specific parts of the observations. For example, in the authors use the changes in the poses of actors to predict their actions. The method in uses body parts extracted by cropping a local patch around identified joints of the actors. The authors of only use the visual appearances of actors extracted from detected bounding boxes, and the relationship between them.","id":"2ebd83ed-2292-4ab8-9334-3163c17703ba","level":"subsection","origin_cites_number":17,"parent_id":"040ae750-d27c-4e06-99d2-48f8312f00ca","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Action prediction"],["subsection","Early action prediction"]],"subsections":["2bcad629-77b1-4f2d-ae76-43a654b49c71"],"title":"Early action prediction"},{"cite_extract_rate":0,"cites":[],"content":"Early action detection methods have many commonalities with action anticipation algorithms in terms of architectural design. However, there are some exceptions that are more applicable in this domain. These exceptions are teacher-student training schemes for knowledge distillation, identifying discriminative features, and recursive prediction/correction mechanisms. In addition, early action prediction algorithm, with a few exceptions, rely on single modal data for prediction and rarely use refinement frameworks such as attention modules. Adopting these techniques and operating on multi-modal feature spaces can further improve the performance of early action prediction algorithms. Unlike the action anticipation methods, there is no strong preference for recurrent or feedforward approaches. Some approaches take advantage of architectures such as temporal CNNs which are popular in the language processing domain and show their effectiveness for early action prediction tasks.","id":"2bcad629-77b1-4f2d-ae76-43a654b49c71","level":"subsubsection","origin_cites_number":0,"parent_id":"2ebd83ed-2292-4ab8-9334-3163c17703ba","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Action prediction"],["subsection","Early action prediction"],["subsubsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0.42465753424657504,"cites":[2867,2855,2847,2860,2854,2853,2868,2857,2869,2856,2846,2862,2852,2866,979,988,2858,2859,2848,2850,2865,2831,2851,8591,1261,2861,2863,2849,7379,2830,2864],"content":"As the name implies, trajectory prediction algorithms predict the future trajectories of objects, i.e. the future positions of the objects over time. These approaches are particularly popular for applications such as intelligent driving and surveillance. Predicted trajectories can be used directly, e.g. in route planning for autonomous vehicles, or used for predicting future events, anomalies, or actions. \nIn this section, we follow the same routine as before and focus on algorithms that have a deep learning component while acknowledging many recent works that have used classical approaches including Gaussian mixture models and processes , Markov decision processes (MDPs) , Markov chains and other techniques .\nTrajectory prediction applications like many other sequence prediction tasks heavily rely on recurrent architectures such as LSTMs , and GRUs . These methods often use an encoder-decoder architecture in which a network, e.g. an LSTM encodes single- or multi-modal observations of the scenes for some time, and another network generates future trajectories given the encoding of the observations. Depending on the complexity of input data, these algorithms may rely on some form of pre-processing for generating features or embedding mechanisms to minimize the dimensionality of the data.\nThe feedforward algorithms often use whole views of the scenes (i.e. the environment and moving objects) and encode them using convolutional layers followed by a regression layer to predict trajectories. A few algorithms use hybrid approaches in which both convolutional and recurrent reasoning are also used . \nDepending on the prediction task, algorithms may rely on single- or multi-modal observations. For example, in the context of visual surveillance where a fixed camera provide a top-down or bird-eye viewing angle, many algorithms only use past trajectories of the agents in either actual 2D frame coordinates or their velocities calculated by the changes from each time step to another . In addition to observations of individual trajectories of agents, these algorithms focus on modeling the interaction between the agents and how they impact each other. For example, Zhang et al. use a state refinement module that aligns all pedestrians in the scene with a message passing mechanism that receives as input the current locations of the subjects and their encodings from an LSTM unit. In a graph-based approach is used where pedestrians are considered as nodes and the interactions between them as edges of the graph. By aggregating information from neighboring nodes, the network learns to assign a different level of importance to each node for a given subject. The authors of perform a pooling operation on the generated representations by sharing the state of individual LSTMs that have spatial proximity. \nAs shown in some works, other sources of information are used in surveilling objects . For example, in addition to encoding the interactions with the environment, Liang et al. use the semantic information of the scene as well as changes in the poses of the pedestrians. In the visual representations of the layout of the environment and the appearances of the subjects are included. The authors of use an occupancy map which highlights the potential traversable locations for the subjects. The method in takes into account pedestrians' head orientations to estimate their fields of view in order to predict which subjects would potentially interact with one another. To predict interactions between humans, in the authors use both poses and trajectories of the agents. Ma et al. go one step further and take into account the pedestrians' characteristics (e.g. age, gender) within a game-theoretic perspective to determine how the trajectory of one pedestrians impact each other.\nIn the context of traffic understanding, predicting trajectories can be more challenging due to the fact that there is camera ego-motion involved (e.g. the prediction is from the perspective of a moving vehicle), there are interactions between different types of objects (e.g. vehicles and pedestrians), and there are certain constraints involved such as traffic rules, signals, etc. To achieve robustness, many methods in this domain take advantage of multi-modal data for trajectory prediction . In addition to using past trajectories, all these algorithms account for the road structure (whether it is from the perspective of the ego-vehicle or a top-down view) often in the form of raw visual inputs or, in some cases, as an occupancy map . The scene layout can implicitly capture the structure of the road, the appearances of the objects (e.g. shape) and the dynamics (e.g. velocity or locations of subjects). Such implicit information can be further augmented by explicit data such as the shapes of the objects (in the case of vehicles) , the speed and steering angle of the ego-vehicle, the distance between the objects , traffic rules and signals , and kinematic constraints . For example, the method in uses a two-stream LSTM encoder-decoder scheme: first stream encodes the current ego-vehicle's odometry (steering angle and speed) and the last observation of the scene and predicts future odometry of the vehicle. The second stream is a trajectory stream that jointly encodes location information of pedestrians and the ego-vehicle's odometry and then combines the encoding with the prediction of the odometry stream to predict the future trajectories of the pedestrians. The method in further extends this approach and adds an intention prediction stream which outputs how likely the observed pedestrian intends to cross the street. The intention likelihood is produced using an LSTM network that encodes the dynamics of the pedestrian, their appearances and their surroundings. Chandra et al. create embeddings of contextual information by taking into account the shape and velocity of the road users and their spatial coordinates within a neighboring region. These embeddings are then fed into some LSTM networks followed by a number of convolutional layers to capture the dynamics of the scenes. In the authors use separate LSTMs for encoding the trajectories of pedestrians and vehicles (as orientated bounding boxes) and then combine them into a unified framework by generating an occupancy map of the scene centered at each agent, followed by a pooling operation to capture the interactions between different subjects. Lee et al. predict the future trajectories of vehicles in two steps: First, an encoder-decoder GRU architecture predicts future trajectories by observing the past ones. Then a refinement network adjusts the predicted trajectories by taking into account the contextual information in the forms of social interactions, dynamics of the agents involved, and the road structure.\nSimilar to action prediction algorithms, attention modules have been widely used in trajectory prediction methods . For example, in , the attention module jointly measures spatial and temporal interactions. The authors of propose the use of social attention modules which estimate the relative importance of interactions between the subject of interest and its neighboring subjects. The method in uses two attention mechanisms, a temporal attention module that measures the importance of each timestep of the observed trajectories and a series of self-attention modules which are applied to encodings of observations prior to predictions. Xue et al. propose an attention mechanism to measure the relative importance between different data modalities, namely the locations and velocities of subjects.\nOne of the major challenges in trajectory prediction is to model uncertainty of predictions, especially in scenarios where many possibilities exist (e.g. pedestrians at intersections). Some algorithms such as train the models by directly measuring the error between the ground truth trajectories and predicted ones, e.g. by using an \\textit{L2} objective function. At the inference time, these algorithms generate deterministic predictions. To measure the uncertainty of predictions, these models are trained multiple times with randomly initialized parameters. Alternatively, uncertainty can be estimated via probabilistic objective functions, e.g. Gaussian log-likelihood, as in Instead of a single point in space, these algorithms predict a distribution that captures the uncertainty of the predictions. VAEs are another technique that can be used to estimate uncertainty . Using these methods, at training time a posterior distribution over some latent space $z$ is learned by conditioning, for example, over future trajectories. At the inference time, a random sample is drawn from the latent space for the predictions.\nSince trajectory prediction algorithms are generative by nature, many approaches rely on adversarial training techniques in which at training time a discriminator is used to predict whether the generated trajectories are real or fake. Kosaraju et al. extend this approach by using two discriminators: A local discriminator which predicts the results on the output of the prediction using only past trajectories, and one global discriminator which operates on the output of the entire network, i.e. the prediction results based on trajectories and scene information.","id":"991d18f2-ffd8-45b8-b2e8-b4c21fb20720","level":"section","origin_cites_number":73,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Trajectory prediction"]],"subsections":["66a51a9f-6cef-4c6e-8e6b-9a97e4e8a052"],"title":"Trajectory prediction"},{"cite_extract_rate":0,"cites":[],"content":"Trajectory prediction is a widely studied field in the computer vision community. Although these works dominantly use recurrent network architectures, many approaches, such as those used in the field of traffic scene understanding, use feedforward networks. Trajectory prediction algorithms rely on one or more sources of information such as the past trajectories of subjects, surrounding visual context, object attributes, vehicle sensor readings, etc. One factor that is common in many of trajectory prediction algorithms is modeling the interactions between dynamic or dynamic and static objects. Relationships are captured explicitly or implicitly via encoding the scenes as a whole. Like many other prediction approaches, trajectory prediction algorithms benefit from various forms of attention mechanisms to learn the importance of spatial, temporal or social interactions between objects. To model uncertainty, techniques such as probabilistic objectives and variational encodings are used. \nTrajectory prediction algorithms are predominantly rely on past trajectory information to predict the future. Although past motion observations are very informative, in some context, e.g. traffic scenes, they are simply not enough. There is a need for a more explicit encoding of contextual information such as road conditions, the subject's attributes, rules and constraints, scene structure, etc. A number of approaches successfully have included a subset of these factors, but a more comprehensive approach should be considered.","id":"66a51a9f-6cef-4c6e-8e6b-9a97e4e8a052","level":"subsection","origin_cites_number":0,"parent_id":"991d18f2-ffd8-45b8-b2e8-b4c21fb20720","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Trajectory prediction"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0.565217391304347,"cites":[2814,2821,2871,8592,2874,2841,2875,2873,2870,2878,2876,2877,2872],"content":"Although the term \"motion prediction\" in many cases is used to refer to future trajectory prediction, here we only consider the algorithms that are designed to predict changes in human pose. Motion prediction play a fundamental role in all prediction approaches as an intermediate step, e.g. to reflect how the future visual representations would look like or the types of actions to anticipate. Like many other prediction applications, this field is dominated by deep learning models, even though some methods still rely on classical techniques .\nSimilar to other prediction algorithms, motion prediction methods widely use recurrent architectures such as LSTMs and GRUs or a combination of both . For example, in the authors use a two-layer GRU model in which the top layer operates backward to learn noise processes and the bottom level is used to predict the poses given the past pose observations and the output of the top layer. Chiu et al. propose a hierarchical LSTM architecture in which each layer of the network encodes the observed poses at different time-scales. In the context of 3D pose prediction, some algorithms rely on a two-stage process where the visual inputs, either as a single image or a sequence of images , are fed into a recurrent network to predict 2D poses of the agent. This is followed by a refinement procedure that transforms the 2D poses into 3D. \nSome approaches adopt feedforward architectures . For example, the method in uses two feedforward networks in a two-stage process. First, the input poses are fed into an autoencoder which is comprised of fully connected layerss (implemented by 1D convolutions with a kernel size of 1) and self-attention blocks. The encodings are then used by multi-level 2D convolutional blocks for final predictions. Zhang et al. predict 3D poses from RGB videos. In their method, the images are converted to a feature space using a convolutional network, and then the features are used to learn a latent 3D representation of 3D human dynamics. The representation is used by a network to predict future 3D poses. To capture movement patterns, a method proposed in converts poses into a trajectory space using discrete cosine transformation. The newly formed representations are then used in a Graph-CNN framework to learn the dependencies between different joint trajectories. \nTo train motion prediction models, some authors use adversarial training methods in which a discriminator is used to classify whether the predicted poses are real or fake . The discrimination procedure can also be applied to evaluating the continuity, i.e. correct order of, predictions as demonstrated in . In the discrimination score is used to generate a policy for future action predictions in the context of imitation learning.","id":"ea670b52-fc46-4dcb-ac08-efd4b23b03c3","level":"section","origin_cites_number":23,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Motion prediction"]],"subsections":["304422d0-f0b0-4723-862a-ed7a4565bca2"],"title":"Motion prediction"},{"cite_extract_rate":0,"cites":[],"content":"Motion prediction algorithms primarily focus on the prediction of changes in the dynamics (i.e. poses) of observed agents. Such predictions can be fundamental to many other applications such as video or trajectory prediction tasks some of which were discussed previously. \nIn recent works, recurrent network architectures are strongly preferred. The architecture of the choice often depends on the representation of the input data, e.g. whether joint coordinates are directly used or are encoded into a high-dimensional representation. \nDespite the development of many successful motion prediction algorithms, the majority of these methods rely on a single source of information, for example, poses or scenes. Encoding higher-level contextual information, such as scene semantics, interactions, etc. can potentially result in more robust predictions, as shown in other prediction applications. Attention modules also, except for one instance, haven't been adopted within motion prediction algorithms. Given the success of using attention in other prediction applications, motion prediction algorithms may benefit from the use of attention mechanisms.","id":"304422d0-f0b0-4723-862a-ed7a4565bca2","level":"subsection","origin_cites_number":0,"parent_id":"ea670b52-fc46-4dcb-ac08-efd4b23b03c3","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Motion prediction"],["subsection","Summary"]],"subsections":[],"title":"Summary"},{"cite_extract_rate":0.565217391304347,"cites":[2882,2879,2880,2885,2888,2806,2883,2881,2887,2886,9096,2884,2803],"content":"In the context of autonomous robotics, some algorithms are designed to predict occupancy grid maps (OGMs) that are grayscale representations of the robot's surroundings showing which parts of the environment are traversable. These approaches are often object-agnostic and are concerned with generating future OGMs which are used by an autonomous agent to perform path planning. In recent years both classical and deep learning methods are used. The deep learning approaches, in essence, are similar to video prediction methods in which the model receives as input a sequence of OGMs and predicts the future ones over some period of time. In this context both recurrent and feedforward methods were common. Another group of generative approaches is semantic map prediction algorithms . These algorithms receive as inputs RGB images of the scenes and predict future segmentation maps.\nSome of the other vision-based prediction applications include weather and Solar irradiance forecasting , steering angle prediction , predicting the popularities of tweets based on tweeted images used and the users' histories , forecasting fashion trends , storyline prediction , pain anticipation , predicting the effect of force after manipulating objects , forecasting the winner of Miss Universe given the appearances of contestants' gowns , and predicting election results given the facial attributes of candidates . These algorithms rely on a combination of techniques discussed earlier in this paper.","id":"77155bc6-d46c-48ec-a4ba-c2d738e13396","level":"section","origin_cites_number":23,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Other applications"]],"subsections":[],"title":"Other applications"},{"cite_extract_rate":0.5714285714285711,"cites":[8593,2889,7619,6980],"content":"Before concluding our discussion on vision-prediction methods, it is worth mentioning that prediction techniques are also widely used in other visual processing tasks such as video summarization , anomaly detection , tracking , active object recognition , action detection and recognition . For example, tracking algorithms are very closely related to trajectory prediction ones and often rely on short term predictions to deal with gaps, e.g. due to occlusions, in tracking. For example, in the method uses a recurrent framework to generate future frames in order to localize pedestrians in next frames. In the context of action detection, some methods rely on a future frame or trajectory prediction of objects to detect actions . In , a method is used for detecting an object in 3D by relying on predicting next best viewing angle of the object. Liu et al. uses a video prediction framework to predict future motion flow maps and images. The future predictions that do not conform with expectations will be identified as abnormal.","id":"374ffbdc-d4ff-461d-a063-67f03516e35e","level":"section","origin_cites_number":7,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Prediction in other vision applications"]],"subsections":[],"title":"Prediction in other vision applications"},{"cite_extract_rate":0,"cites":[],"content":"When it comes to the evaluation of algorithms, there are two important factors: metrics and datasets. They highlight the strengths and weaknesses of the algorithms and provide a means to compare the relative performances of the methods. Given the importance of these two factors in the design of algorithms, we dedicate the following sections to discussing the common metrics and datasets used for vision-based prediction tasks. Since the datasets and metrics used in these applications are highly diverse, we will focus our discussion on some of the main ones for each prediction category while providing visual aids to summarize what the past works used for evaluation purposes.","id":"5bb03330-9009-4e51-abdd-12e3b8d3505d","level":"section","origin_cites_number":0,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","The evaluation of state-of-the-art"]],"subsections":[],"title":"The evaluation of state-of-the-art"},{"cite_extract_rate":0,"cites":[],"content":"\\begin{figure*}\n\\centering\n \\includegraphics[width=0.21\\textwidth]{video_metrics.png}\n \\hspace{0.1cm}\n \\includegraphics[width=0.21\\textwidth]{action_metrics.png}\n \\hspace{0.1cm}\n \\includegraphics[width=0.22\\textwidth]{trajectory_metrics.png}\n \\hspace{0.1cm}\n \\includegraphics[width=0.18\\textwidth]{motion_metrics.png}\n \\caption{Metrics used in vision-based prediction applications. From left to right: Video, action, trajectory and motion prediction. }\n\\label{metrics_fig}\n\\end{figure*}\nIn this section, we follow the same routine as the discussion on the past works and divide the metrics into different categories. A summary of the metrics can be found in Figure \\ref{metrics_fig}. The interested readers can also refer to Appendix \\ref{metrics_papers} for further information regarding the metrics and the the papers that used them. Note while we discuss the metrics in each category and we only provide mathematical expressions of the most popular metrics in Appendix \\ref{appendix_metrics}.","id":"4c69ee31-a62c-468f-a731-a0d63c6e9fe2","level":"section","origin_cites_number":0,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"]],"subsections":["a3e52edb-4aa4-4bb7-abe0-4be3c8454b3e","f0c26ecc-6bb8-4c9f-9512-d54a1c7c4a8c","cbe38bc2-9ec9-4841-bbf1-11b25e444a6d","d0bc398b-84fd-40e6-a4bf-c6c2267c1d9c","6b621d84-5e91-4190-9b89-35e7d1a31664"],"title":"Metrics"},{"cite_extract_rate":0.585365853658536,"cites":[2804,2814,2817,2807,2808,2816,2819,2815,2806,117,2809,2818,7620,2811,2802,2820,2810,491,2805,2812,8590,7617,2813,2890],"content":"Video prediction is about generating realistic images, hence the best performance is achieved when the disparities between the generated images and groundtruth images are minimal. The most straightforward way of computing the disparity is to measure pixel-wise error using a Mean Square Error (MSE) , which computes average squared intensity differences between pixels. Another more popular metric related to MSE is Peak Signal-to-Noise Ratio (PSNR) . PSNR is the ratio of the maximum pixel value (i.e. possible signal power), e.g. 255 in 8-bit images, divided by the MSE (or power of distorting noise) measure of two images. The lower the error between two images, the higher the value of PSNR, and consequently, the higher the quality of the generated images. Because of the wide dynamic range of signals, PSNR value is expressed in the logarithmic decibel scale. \nAlthough MSE, and PSNR metrics are easy to calculate, they cannot measure the perceived visual quality of a generated image. An alternative metric to address this issue is Structural SIMilarity (SSIM) index () which is designed to model image distortion. To capture the structural differences between the two images, SSIM separates illumination information as it is independent of objects' structures. As a result, the similarity is measured by a combination of three comparisons, namely luminance, contrast, and structure. \nHigher-level contextual similarities may not be captured by distance measures on pixel values. More recently proposed metric, Learned Perceptual Image Patch Similarity (LPIPS), () measures the similarity between two images by comparing internal activations of convolutional networks trained for high-level classification tasks. The value is calculated by an average L2 distance over normalized deep features.\nSome other metrics that have been used in the literature are qualitative human judgment , Frechet Video Distance (FVD) () , Maximum Mean Discrepancy (MMD) , Inception Scores (IS) () , Binary Cross Entropy (BCE) , L1 , and Root MSE (RMSE) .","id":"a3e52edb-4aa4-4bb7-abe0-4be3c8454b3e","level":"subsection","origin_cites_number":41,"parent_id":"4c69ee31-a62c-468f-a731-a0d63c6e9fe2","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"],["subsection","Video prediction"]],"subsections":[],"title":"Video prediction"},{"cite_extract_rate":0.378378378378378,"cites":[2821,2837,2829,7117,2844,2838,2832,2841,2823,2822,2828,2825,2806,2834,2840,2824,2845,2835,2827,2826,2831,2842,2833,2836,7618,2839,2830,2843],"content":"Similar to classification tasks, many action prediction algorithms use accuracy measure to report on the performance that is the ratio of the correct predictions with respect to the total number of predictions . Despite being used widely, accuracy on its own is not a strong indicator of performance, especially when we are dealing with class-imbalance data. This is because, for example, the model can simply favor the more represented class and predict every input as that class. This then would result in a high accuracy measure because the metric only considers the ratio of correct predictions. To address these shortcomings, some works use complimentary metrics that, in addition to correct predictions, account for different types of false predictions. These metrics are precision , recall , and Area Under the Curve (AUC) of precision-recall graph. Precision and recall also form the basis for the calculation of some higher level metrics such as F1-score , Average Precision (AP) and its variations mean AP (mAP) and calibrated AP (cAP) . Some of the less common performance metrics are Matthews Correlation Coefficient (MCC) , False positive (FP), True Positive Rate (TPR) and False Positive Rate (FPR) \n, Prediction Power (PP) , and Mean Reciprocal Rank (MRR) .\nDepending on the application, some algorithms evaluate the timing factor in terms of the Run Time (RT) of the model or time of the event, e.g. beginning of the next activity , Time To Accident (or collision) (TTA) , and, in the context of driving, Time To Maneuver (TTM) .","id":"f0c26ecc-6bb8-4c9f-9512-d54a1c7c4a8c","level":"subsection","origin_cites_number":74,"parent_id":"4c69ee31-a62c-468f-a731-a0d63c6e9fe2","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"],["subsection","Action prediction"]],"subsections":[],"title":"Action prediction"},{"cite_extract_rate":0.42465753424657504,"cites":[2860,2855,2867,2847,2854,2853,2868,2857,2869,2862,2856,2846,2852,2866,979,988,2858,2859,2848,2850,2865,2831,2851,8591,1261,2849,2863,2861,7379,2830,8594],"content":"Perhaps the most popular performance measure for trajectory prediction is Average Displacement Error (ADE) calculated as the average error between the prediction location and the ground truth over all time steps. Some methods complement ADE measure with its extension Final Displacement Error (FDE) . As the name suggests, FDE only measures the error between the ground truth and the generated trajectory for the final time step.\nMany other works use the same metric as ADE or ADE/FDE without using the same terminology. It is also a common practice that instead of using average or final time step measures, to calculate the error at different time steps over a period of time .\nTo measure displacement error in probablistic trajectory prediction algorithms, some works generate a set number of samples and report the best measure (i.e. minimum error) or average over all samples . Depending on the error metric used, some refer to these measures as Minimum ADE/FDE (MinADE/FDE) (using Euclidean distance) or Mean/Minimum Mean Square Displacement (Mean/MinMSD) (using MSE). Some of the other probabilistic measures are Log-Likelihood (LL) , Negative Log-Likelihood (NLL) , Kullback–Leibler Divergence (KLD) , Negative Log-Probability (NLP), Cross Entropy (CE) , Average Prediction Probability (APP) .\nPerformance can also be evaluated using common classification metrics. For example, in Hit Rate (HR) and in Miss Rate (MR) metrics are used. In these cases, if the predicted trajectory is below (or above) a certain distance threshold from the groundtruth, it is considered as a hit or miss. Following a similar approach, some authors calculate accuracy or precision of predictions.\nSome of the other metrics used in the literature are Run Time (RT) , Average Non-linear Displacement Error (ANDE)\n, Maximum Distance (MaxD) , State collision rate (SCR) , Percentage Deviated (PD) , Distance to Goal (DtG) , Fraction of Near Misses (FNM) , Expected Calibration Error (ECE) , and qualitative (Q) . A few works predict the orientations of pedestrians or vehicles , therefore also report performance using Mean angular error (MAnE).","id":"cbe38bc2-9ec9-4841-bbf1-11b25e444a6d","level":"subsection","origin_cites_number":73,"parent_id":"4c69ee31-a62c-468f-a731-a0d63c6e9fe2","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"],["subsection","Trajectory prediction"]],"subsections":["6226a0b6-b6ff-448f-a412-1d2941716ad2"],"title":"Trajectory prediction"},{"cite_extract_rate":0.43076923076923,"cites":[2860,2855,2867,2847,2853,2857,2869,2862,2856,2846,2866,2852,979,988,2858,2859,2850,2865,2831,2851,1261,8591,2861,2863,2849,2830,2864,8594],"content":"Unlike video and action prediction fields, performance measures for trajectory prediction algorithms are not standardized in terms error metrics used and units of measure. For example, for measuring displacement error, although many algorithms use Euclidean Distance (ED) (aka L2-distance, L2-norm, Euclidean norm) , many others rely on different error metrics including MSE , RMSE , Weighted RMSE , Mean Absolute Error (MAE), Hausdorff Distance (HD), Modified HD (MHD) , and discrete Fr\\'echet distance (DFD) . Moreover, trajectory prediction algorithms use different units for measuring displacement error. These are meter , pixel , normalized pixel and feet .\nAlthough such discrepancy between error metrics and units is expected across different applications, the problem arises when the proposed works do not specify the error metric , the unit of measure or both . Despite the fact that the reported results might imply the choice of the metrics and units, the lack of specification can cause erroneous comparisons, specially because many authors use the results of previous works directly as reported in the papers.\nUnfortunately, metric and unit discrepancy exists within the same applications and the same error measuring techniques. For instance, in the case of ADE measure, this metric is originally proposed in in terms of ED, and was referred to as ADE by the authors of despite the fact that they used MSE instead. This is also apparent in many subsequent works that employed ADE measure. For example, the majority of methods use the original metric and report the results in terms of ED whereas some works use MSE and RMSE or do not specify the metric . Although the formulation of these metrics look similar, they produce different results. ADE using ED, for example, is square-root of squared differences averaged over all samples and time steps. Unlike ED, in RMSE, the averaging takes place inside square-root operation. MSE, on the other hand, is very different from the other two metrics, and does not calculate the root of the error. As we can also see in some of the past works, this discrepancy may cause confusion about the intended and actual metric that is used. For example, in the authors propose to use MAE metric while presenting mathematical formulation of Euclidean distance. The authors of make a similar mistake and define ED formulation but refer to it as MSE.\nIn addition, some algorithms within the same applications and using the same datasets use different measuring unit. For instance, in the context of surveillance, ETH is one of the most commonly used datasets. Many works use this dataset for benchmarking the performance of their proposed algorithms, however, they either use different units, e.g. meter , pixel , normalized pixel , or do not specify the unit used .\nLast but not least, another potential source of error in performance evaluation is in the design of the experiments. Taking surveillance applications as an example, it is a common practice to evaluate algorithms with 8 frames observations of the past and prediction 12 steps in the future . However, in some cases the performance of state-of-the-art is reported under the standard 8/12 condition, but the proposed algorithms are tested under different conditions. For instance, in the authors incorrectly compared the performance of their proposed algorithm using 5 observations and 5 predictions with the results of the previous works evaluated under the standard 8/12 condition.","id":"6226a0b6-b6ff-448f-a412-1d2941716ad2","level":"subsubsection","origin_cites_number":65,"parent_id":"cbe38bc2-9ec9-4841-bbf1-11b25e444a6d","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"],["subsection","Trajectory prediction"],["subsubsection","Pitfalls of trajectory prediction metrics"]],"subsections":[],"title":"Pitfalls of trajectory prediction metrics"},{"cite_extract_rate":0.565217391304347,"cites":[2814,2871,2821,8592,2874,2841,2875,2873,2870,2878,2876,2877,2872],"content":"Due to the inherent stochasticity of human body movement, motion prediction algorithms often limit their prediction horizon to approximately $500ms$. To measure the error between corresponding ground truth and predicted poses, these algorithms use mean average error, either in angle space (MAnE) or joint space (MJE) in terms of joint coordinates . In the 3D motion prediction domain, a metric known as Mean Per Joint Prediction Error (MPJPE) is used which is the error over joints normalized with respect to the root joint. \nAs an alternative to distance error metrics, Percentage of Correct Keypoints (PCK) measures how many of the keypoints (e.g. joints) are predicted correctly. The correct predictions are those that are below a certain error threshold (e.g. 0.05). Some works also use the accuracy metric to report on how well the algorithm can localize the position of a particular joint within an error tolerance region .\nOther metrics used in the literature include Normalized Power Spectrum Similarity (NPSS) , Reconstruction Error (RE) , Limb Orientation (LO) , PoSe Entropy (PSEnt), PoSe KL (PSKL) , qualitative human judgment and method Run Time (RT) .","id":"d0bc398b-84fd-40e6-a4bf-c6c2267c1d9c","level":"subsection","origin_cites_number":23,"parent_id":"4c69ee31-a62c-468f-a731-a0d63c6e9fe2","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"],["subsection","Motion prediction"]],"subsections":["aa256766-5674-49c7-8798-111d626f386c"],"title":"Motion prediction"},{"cite_extract_rate":0.5454545454545451,"cites":[2814,2871,2821,8592,2874,2841,2875,2873,2878,2876,2877,2872],"content":"Similar to trajectory methods, motion prediction algorithms are evaluated using distance-based methods that calculate the error between pose vectors. In the case of MAnE measure, some methods use ED metric while others use MSE . Sometimes no metric is specified . The same holds for MJE measure where metrics used include MSE , RMSE , ED , MAE , or no metric is specified .\nThe added challenge in coordinate-based error measures, e.g. MJE, MPJPE, is the error unit. While many approaches do not specify the unit explicitly , others clearly state whether the unit is in pixel , centimeter , meter or millimeter . As was the case before, here many algorithms that benchmark on the same datasets, may use different performance metrics, e.g. using popular human datasset Human 3.6M , MAnE (ED), MAnE (MSE) and MJE (ED) are used.","id":"aa256766-5674-49c7-8798-111d626f386c","level":"subsubsection","origin_cites_number":22,"parent_id":"d0bc398b-84fd-40e6-a4bf-c6c2267c1d9c","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"],["subsection","Motion prediction"],["subsubsection","Pitfalls of motion prediction metrics"]],"subsections":[],"title":"Pitfalls of motion prediction metrics"},{"cite_extract_rate":0.56,"cites":[2882,2879,2880,2885,8595,2806,2888,2883,2881,2887,2886,9096,2884,2803],"content":"Depending on the task objectives, the metrics used in other prediction applications are similar to the ones discussed thus far. For instance, the applications that classify future events or outcomes, e.g. contest or an election winner, next image index for storytelling, severe weather, and pain, use common metrics such as accuracy , precision, recall , percentage of correct predictions (PCP) , and Matthews correlation coefficient (MCC) which predicts the quality of binary classification by taking into account both false and true predictions.\nRegression based methods, such as temperature, trends, or steering prediction, use distance metrics including Euclidean Distance (ED) , RMSE , MSE , MAE , Mean Absolute Percentage Error (MAPE) , normalized MAPE (nMAPE) , and the Spearman's ranking Correlation (SRC) which measures the strength and direction of relationship between two variables.\nOf particular interest are metrics used for evaluating generative models that predict Occupancy Grid Maps (OGMs) and segmentation maps. OGMs are grayscale images that highlight the likelihood of a certain region (represented as a cell in the grid) that is occupied. The generated map can be compared to ground truth by using image similarity metrics such as SSIM , PSNR or psi ($\\psi$) . Alternatively, OGM can be evaluated using a binary classification metric. Here, the grid cells are classified as occupied or free by applying a threshold and then can be evaluated as a whole by using metrics such as True Positive (TP), True Negative (TN) , Receiver Operator Characteristic (ROC) curve over TP and TN , F1-score , precision, recall, and their corresponding AUC . Given that OGM prediction algorithms are mainly used in safety-critical applications such as autonomous driving, some algorithms are also evaluated in terms of their Run Time (RT) .\nImage similarity metrics such as PSNR and SSIM can also be used in the segmentation prediction domain . The most common metric, however, is Intersection over Union (IoU) which measures the average overlap of segmented instances with the ground truth segments. In addition, by applying a threshold to IoU scores, the true matches can be identified and used to calculate the Average Precision (AP) scores as in . Other metrics used for segmentation prediction tasks include EndPoint error (EPE) , Probabilistic Rand Index (RI), Global Consistency Error (GCE), and Variation of Information (VoI) .","id":"6b621d84-5e91-4190-9b89-35e7d1a31664","level":"subsection","origin_cites_number":25,"parent_id":"4c69ee31-a62c-468f-a731-a0d63c6e9fe2","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics"],["subsection","Other prediction applications"]],"subsections":[],"title":"Other prediction applications"},{"cite_extract_rate":0.33628318584070804,"cites":[2899,2838,2808,2869,2824,2892,8596,2835,2897,1517,2903,8590,2894,6980,2891,2900,2626,2908,2906,2896,2905,2898,2902,8597,1733,2904,2907,2893,2878,2851,7379,2901,2880,2895],"content":"\\label{datasets}\n\\begin{table*}[]\n\\centering\n\\resizebox{1\\textwidth}{!}{\n\\begin{tabular}{|l|l|l|l|c|c|c|c|c|}\n\\hline\n\\multirow{2}{*}{\\textbf{Year}} & \\multicolumn{1}{c|}{\\multirow{2}{*}{\\textbf{Dataset}}} & \\multicolumn{1}{c|}{\\multirow{2}{*}{\\textbf{Type}}} & \\multicolumn{1}{c|}{\\multirow{2}{*}{\\textbf{Annotations}}} & \\multicolumn{5}{c|}{\\textbf{Application}} \\\\ \\cline{5-9} \n & \\multicolumn{1}{c|}{} & \\multicolumn{1}{c|}{} & \\multicolumn{1}{c|}{} & \\textit{\\textbf{V}} & \\textit{\\textbf{A}} & \\textit{\\textbf{T}} & \\textit{\\textbf{M}} & \\textit{\\textbf{O}} \\\\ \\hline\n\\multirow{12}{*}{\\textit{\\textbf{2019}}} & ARGOVerse & Traffic & RGB, LIDAR, 3D BB & & & x & & \\\\ \\cline{2-9} \n & CARLA & Traffic (sim) & RGB & & x & & & \\\\ \\cline{2-9} \n & EgoPose & Pose (ego) & RGB, 3D Pose & & & & x & \\\\ \\cline{2-9} \n & Future Motion (FM) & Mix & RGB, BB, Attrib. & & & x & & \\\\ \\cline{2-9} \n & InstaVariety & Activities & RGB, BB, Pose & & & & x & \\\\ \\cline{2-9} \n & INTEARCTION & Traffic & Map, Traj. & & & x & & \\\\ \\cline{2-9} \n & Luggage & Robot & Stereo RGB, BB & & x & & & \\\\ \\cline{2-9} \n & MGIF & Activities & RGB & x & & & & \\\\ \\cline{2-9} \n & Pedestrian Intention Estimation (PIE) & Traffic & RGB, BB, Class, Attrib., Temporal seg., Vehicle sensors & & x & x & & \\\\ \\cline{2-9} \n & nuScenes & Traffic & RGB, LIDAR, 3D BB, Vehicle sensors & & & x & & \\\\ \\cline{2-9} \n & Vehicle-Pedestrian-Mixed (VPM) & Traffic & RGB, BB & & & x & & \\\\ \\cline{2-9} \n & TRAF & Traffic & RGB, BN, Class, Time-of-day & & & x & & \\\\ \\hline\n\\multirow{10}{*}{\\textit{\\textbf{2018}}} & 3D POSES IN THE WILD (3DPW) & Outdoor & RGB, 2D/3D Pose, Models & & & & x & \\\\ \\cline{2-9} \n & ActEV/VIRAT & Surveillance & RGB, BB, Activity, Temporal seg. & & x & x & & \\\\ \\cline{2-9} \n & ACTICIPATE & Interaction & RGB, Gaze, Pose & & x & & & \\\\ \\cline{2-9} \n & Atomic Visual Actions (AVA) & Activities & RGB, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & Epic-Kitchen & Cooking (ego) & RGB, Audio, BB, Class, Text, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & EGTEA Gaze+ & Cooking (ego) & RGB, Gaze, Mask, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & ShanghaiTech Campus (STC) & Surveillance & RGB, Anomaly & x & & & & \\\\ \\cline{2-9} \n & ShapeStack & Objects (sim) & RGBD, Mask, Stability & x & & & & \\\\ \\cline{2-9} \n & VIENA & Traffic (sim) & RGB, Activity, Vehicle sensors & & x & & & \\\\ \\cline{2-9} \n & YouCook2 & Cooking & RGB, Audio, Text, Activity, Temporal seg. & & x & & & \\\\ \\hline\n\\multirow{10}{*}{\\textit{\\textbf{2017}}} & BU Action (BUA) & Activities & RGB (image), Activity & & x & & & \\\\ \\cline{2-9} \n & CityPerson & Traffic & Stereo RGB, BB, Semantic seg. & & & x & & \\\\ \\cline{2-9} \n & Epic-Fail & Risk assessment & RGB, BB, Traj., Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & Joint Attention in Autonomous Driving (JAAD) & Traffic & RGB, BB, Attrib., Temporal seg. & x & x & x & & x \\\\ \\cline{2-9} \n & L-CAS & Traffic & LIDAR, 3D BB, Attrib. & & & x & & \\\\ \\cline{2-9} \n & Mouse Fish & Animals & Depth, 3D Pose & & & & x & \\\\ \\cline{2-9} \n & Oxford Robot Car (ORC) & Traffic & Stereo RGB, LIDAR, Vehicle sensors & & & x & & \\\\ \\cline{2-9} \n & PKU-MMD & Activities, interactions & RGBD, IR, 3D Pose, Multiview, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & Recipe1M & Cooking & RGB(image), Text & & x & & & \\\\ \\cline{2-9} \n & STRANDS & Traffic & RGBD, 3DBB & & & x & & \\\\ \\hline\n\\multirow{13}{*}{\\textit{\\textbf{2016}}} & BAIR Push & Object manipulation & RGB & x & & & & \\\\ \\cline{2-9} \n & Bouncing Ball (BB) & Simulation & RGB & x & & & & \\\\ \\cline{2-9} \n & Miss Universe (MU) & Miss universe & RGB, BB, Scores & & & & & x \\\\ \\cline{2-9} \n & Cityscapes & Traffic & Stereo RGB, BB, Semantic seg., Vehicle Sensors & x & & x & & x \\\\ \\cline{2-9} \n & CMU Mocap & Activities & 3D Pose, Activity & & x & & x & \\\\ \\cline{2-9} \n & Dashcam Accident Dataset (DAD) & Traffic, accidents & RGB, BB, Class, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & NTU RGB-D & Activities & RGBD, IR, 3D Pose, Activity & & x & & & \\\\ \\cline{2-9} \n & Ongoing Activity (OA) & Actvities & RGB, Activity & & x & & & \\\\ \\cline{2-9} \n & OAD & Activities & RGBD, 3D Pose, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & Stanford Drone (SD) & Surveillance & RGB, BB, Class & & & x & & \\\\ \\cline{2-9} \n & TV Series & Activities & RGB, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & Visual StoryTelling (VIST) & Visual story & RGB, Text & & & & & x \\\\ \\cline{2-9} \n & Youtube-8M & Activities & RGB, Activity, Temporal seg. & x & & & & \\\\ \\hline\n\\multirow{14}{*}{\\textit{\\textbf{2015}}} & Amazon & Fashion & Features, Attrib., Text & & & & & x \\\\ \\cline{2-9} \n & Atari & Games & RGB & x & & & & \\\\ \\cline{2-9} \n & Brain4Cars & Traffic, Driver & RGB, BB, Attrib., Temporal seg., Vehicle sensors & & x & & & \\\\ \\cline{2-9} \n & CMU Panoptic & Interaction & RGBD, Multiview, 3D Pose, 3D facial landmark & x & x & & x & \\\\ \\cline{2-9} \n & First Person Personalized Activities (FPPA) & Activities (ego) & RGB, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & GTEA Gaze + & Cooking (ego) & RGB, Gaze, Mask, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & MicroBlog-Images (MBI-1M) & Tweets & RGB (image), Attrib., Text & & & & & x \\\\ \\cline{2-9} \n & MOT & Surveillance & RGB, BB & & & x & & \\\\ \\cline{2-9} \n & Moving MNIST (MMNIST) & Digits & Grayscale & x & & & & \\\\ \\cline{2-9} \n & SUN RGB-D & Places & RGBD, 3D BB , Class & & & & & x \\\\ \\cline{2-9} \n & SYSU 3DHOI & Object interaction & RGBD, 3D Pose, Activity & & x & & & \\\\ \\cline{2-9} \n & THUMOS & Activities & RGB, Activity, Temporal seg. & x & x & & & \\\\ \\cline{2-9} \n & Watch-n-Push (WnP) & Activities & RGBD, 3D Pose, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & Wider & Activities & RGB (image), Activity & & x & & & \\\\ \\hline\n\\multirow{5}{*}{\\textit{\\textbf{2014}}} & Breakfast & Cooking & RGB, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & Human3.6M & Activities & RGB, 3D Pose, Activity & x & x & & x & \\\\ \\cline{2-9} \n & MPII Human Pose & Activities & RGB, Pose, Activity & & & & x & \\\\ \\cline{2-9} \n & Online RGBD Action Dataset (ORGBD) & Activities & RGBD, BB, 3D Pose, Activity & & x & & & \\\\ \\cline{2-9} \n & Sports-1M & Sports & RGB, Activity & x & x & & & \\\\ \\hline\n\\end{tabular}\n}\n\\caption{A summary of common datasets from years 2014-2019 used in vision-based prediction applications, namely video (V), action (A), trajectory (T), motion (M) and others (O). The annotation column specifies the type of data (e.g. RGB, Infrared(IR)) and annotation types. All datasets contain image sequences unless specified by \"image\". As for annotations, BB stands for bounding box. Attributes include any object characteristics (e.g. for pedestrians demographics, behavior). Vehicle sensors may include speed, steering angle, GPS, etc. Temporal seg. identifies the datasets that specify the start and end of the events.}\n\\label{dataset_table1}\n\\end{table*}\n\\begin{table*}[]\n\\centering\n\\resizebox{1\\textwidth}{!}{\n\\begin{tabular}{|l|l|l|l|c|c|c|c|c|}\n\\hline\n\\multirow{2}{*}{\\textbf{Year}} & \\multicolumn{1}{c|}{\\multirow{2}{*}{\\textbf{Dataset}}} & \\multicolumn{1}{c|}{\\multirow{2}{*}{\\textbf{Type}}} & \\multicolumn{1}{c|}{\\multirow{2}{*}{\\textbf{Annotations}}} & \\multicolumn{5}{c|}{\\textbf{Application}} \\\\ \\cline{5-9} \n & \\multicolumn{1}{c|}{} & \\multicolumn{1}{c|}{} & \\multicolumn{1}{c|}{} & \\textit{\\textbf{V}} & \\textit{\\textbf{A}} & \\textit{\\textbf{T}} & \\textit{\\textbf{M}} & \\textit{\\textbf{O}} \\\\ \\hline\n \\multirow{7}{*}{\\textit{\\textbf{2013}}} & 50Salads & Cooking (ego) & RGBD, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & ATC & Surveillance & RGB, Traj., Attrib. & & & x & & \\\\ \\cline{2-9} \n & CAD-120 & Activities & RGBD, 3D Pose, Activity & & x & & & \\\\ \\cline{2-9} \n & CHUK Avenue & Surveillance & RGB, BB, Anomaly, Temporal seg. & x & & x & & \\\\ \\cline{2-9} \n & Daimler Path & Traffic & Stereo Grayscale, BB, Temporal seg. , Vehicle sensors & & x & & & \\\\ \\cline{2-9} \n & Joint-annotated HMDB (JHMDB) & Activities & RGB, Mask, Activity, Pose, Optical flow & x & x & & & \\\\ \\cline{2-9} \n & Penn Action & Activities & RGB, BB, Pose, Activity & x & & & x & \\\\ \\hline\n\\multirow{11}{*}{\\textit{\\textbf{2012}}} & BIT & Interaction & RGB, Activity & & x & & & \\\\ \\cline{2-9} \n & GTEA Gaze & Cooking (ego) & RGB, Gaze, Mask, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & KITTI & Traffic & Stereo RGB, LIDAR, BB, Optical flow, Vehicle sensors & x & & x & & x \\\\ \\cline{2-9} \n & MANIAC & Object manipulation & RGBD, Semantic seg., Activity & & x & & & \\\\ \\cline{2-9} \n & MPII-Cooking & Cooking & RGB, 3D Pose, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & MSR Daily Activity (MSRDA) & Activities & Depth, Activity & & x & & & \\\\ \\cline{2-9} \n & New York Grand Central (GC) & Surveillance & RGB, Traj. & & & x & & \\\\ \\cline{2-9} \n & SBU Kinetic Interction (SBUKI) & Interaction & RGBD, 3D Pose, Activity & & x & x & x & \\\\ \\cline{2-9} \n & UCF-101 & Activities & RGB, Activity & x & x & & x & \\\\ \\cline{2-9} \n & UTKinect-Action (UTKA) & Activities & RGBD, 3D Pose, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & UvA-NEMO & Smiles & RGB & x & & & & \\\\ \\hline\n\\multirow{5}{*}{\\textit{\\textbf{2011}}} & Ford campus vision LiDAR (FCVL) & Traffic & RGB, LIDAR, Vehicle sensors & & & & & x \\\\ \\cline{2-9} \n & Human Motion Database (HMDB) & Activities & RGB, BB, Mask, Activity & & x & & & \\\\ \\cline{2-9} \n & Stanford 40 & Activities & RGB (image), BB, Activity & & x & & & \\\\ \\cline{2-9} \n & Town Center & Surveillance & RGB, BB & & & x & & \\\\ \\cline{2-9} \n & VIRAT & Surveillance, Activities & RGB, BB, Activity, Temporal seg. & & x & x & & \\\\ \\hline\n\\multirow{8}{*}{\\textit{\\textbf{2010}}} & DISPLECS & Traffic & RGB, Vehicle sensors & & & & & x \\\\ \\cline{2-9} \n & MSR & Activities & Depth, Activity & x & & & & \\\\ \\cline{2-9} \n & MUG & Facial expressions & RGB, Keypoints, Label & \\multicolumn{1}{l|}{x} & \\multicolumn{1}{l|}{} & \\multicolumn{1}{l|}{} & \\multicolumn{1}{l|}{} & \\multicolumn{1}{l|}{} \\\\ \\cline{2-9} \n & PROST & Objects & RGB, BB & x & & & & \\\\ \\cline{2-9} \n & TV Human Interaction (THI) & Interaction & RGB, BB, Head pose, Activity & & x & & & \\\\ \\cline{2-9} \n & UT Interaction (UTI) & Interaction & RGB, BB, Activity, Temporal seg. & & x & & & \\\\ \\cline{2-9} \n & VISOR & Surveillance & RGB, BB, Pose, Attrib. & x & & & & \\\\ \\cline{2-9} \n & Willow Action & Activiites & RGB (image), Activity & & x & & & \\\\ \\hline\n\\multirow{9}{*}{\\textit{\\textbf{2009}}} & Caltech Pedestrian & Traffic & RGB, BB & x & x & & & \\\\ \\cline{2-9} \n & Collective Activity (CA) & Interaction & RGB, BB, Attrib., Activity, Temporal seg. & & x & x & x & \\\\ \\cline{2-9} \n & Edinburgh IFP & Surveillance & RGB, BB & & & x & & \\\\ \\cline{2-9} \n & ETH & Surveillance & RGB, Traj. & & & x & & \\\\ \\cline{2-9} \n & OSU & Sports & RGB, BB, Attrib. & & & x & & \\\\ \\cline{2-9} \n & PETS2009 & Surveillance & RGB, BB & & & x & & \\\\ \\cline{2-9} \n & QMUL & Traffic, anomaly & RGB, Traj. & & & x & & \\\\ \\cline{2-9} \n & TUM Kitchen & Activities & RGB, RFID, 3D Pose, Activity, Temporal seg. & & & x & & \\\\ \\cline{2-9} \n & YUV Videos & Mix videos & RGB & x & & & & \\\\ \\hline\n\\multirow{2}{*}{\\textit{\\textbf{2008}}} & Daimler & Traffic & Grayscale, BB & & x & & & \\\\ \\cline{2-9} \n & MIT Trajectory (MITT) & Surveillance & RGB, Traj. & & & x & & \\\\ \\hline\n\\multirow{5}{*}{\\textit{\\textbf{2007}}} & AMOS & Weather & RGB, Temperature, Time & & & & & x \\\\ \\cline{2-9} \n & ETH Pedestrian & Traffic & RGB, BB & & x & & & \\\\ \\cline{2-9} \n & Lankershim Boulevard & Traffic & RGB, Traj. & & & x & & \\\\ \\cline{2-9} \n & Next Generation Simulation (NGSIM) & Traffic & Map, Traj. & & x & x & & \\\\ \\cline{2-9} \n & UCY & Surveillance & RGB, Traj., Gaze & & & x & & \\\\ \\hline\n\\textit{\\textbf{2006}} & Tuscan Arizona & Weather & RGB & & & & & x \\\\ \\hline\n\\textit{\\textbf{2004}} & KTH & Activities & Grayscale, Activity & x & & & & \\\\ \\hline\n\\textit{\\textbf{1981}} & Golden Colorado & Weather & RGB & & & & & x \\\\ \\hline\n\\end{tabular}\n}\n\\caption{A summary of common datasets from years 2013 and earlier used in vision-based prediction applications, namely video (V), action (A), trajectory (T), motion (M) and others (O). The annotation column specifies the type of data (e.g. RGB, Infrared(IR)) and annotation types. All datasets contain sequences unless specified by \"image\". As for annotations, BB stands for bounding box. Attributes include any object characteristics (e.g. for pedestrians demographics, behavior). Vehicle sensors may include speed, steering angle, GPS, etc. Temporal seg. identifies the datasets that specify the start and end of the events.}\n\\label{dataset_table2}\n\\end{table*}\n\\begin{figure*}\n\\centering\n\\includegraphics[width=1\\textwidth]{dataset_paper.png}\n\\caption{An illustration of datasets and papers that use them.}\n\\label{datasets_fig}\n\\end{figure*}\nWe have identified more than 100 datasets that are used in the vision-based prediction literature. Discussing all datasets in detail is beyond the scope of this paper. We provide a summary of the datasets and their characteristics in Tables \\ref{dataset_table1} and \\ref{dataset_table2} and briefly discuss more popular datasets in each field. Figure \\ref{datasets_fig} illustrates the list of papers and corresponding datasets used for evaluation. Note that the papers that do not use publicly available datasets are not listed in this figure. For further information, the readers can also refer to Appendices \\ref{links_datasets} and \\ref{papers_datasets}. \n\\textbf{Video prediction.} Almost any forms of sequential RGB images can be used for evaluation of video prediction algorithms. Among the most common datasets are traffic datasets such a KITTI , and Caltech Pedestrians . KITTI is a dataset recorded from inside of a vehicle and contains images of urban roads annotated with bounding box information. It also contains depth maps, LIDAR point clouds and semantic segmentation maps. Caltech Pedestrian is a similar dataset with the difference of only containing RGB images and bounding boxes for pedestrians. It also contains occlusion bounding boxes highlighting the visible portions of the pedestrians. Activity datasets such as UCF-101 and Human3.6M are also widely used. UCF-101 contains videos of various types of activities such as sports, applying makeup, playing music instruments annotated with activity labels per video. Human3.6M consists of 3.6 million 3D human poses and corresponding images recorded from 11 professional actors. This dataset contains 17 generic scenarios such as discussion, smoking, and taking photos.\n\\textbf{Action prediction. } The algorithms in this domain are evaluated on a wide range of datasets. For anticipation tasks, traffic datasets such as Next Generation Simulation (NGSIM) and Joint Attention in Autonomous Driving (JAAD) are used. NGSIM contains trajectories of vehicles driving on highways in the United States. The trajectories are accompanied by the top-down views of the corresponding road structures. The JAAD dataset contains videos of pedestrians crossing the road recorded using an on-board camera. This dataset contains the frame-wise pedestrian bounding boxes, and action labels as well as pedestrians' and roads' attributes. A similar dataset to JAAD is Pedestrian Intention Estimation (PIE) which, in addition, provides the ego-vehicle sensor data and spatial annotations for traffic elements.\nAnother popular category of datasets in this domain is those containing videos of cooking activities. These datasets are Epic-Kitchen , 50salads , Breakfast and MPII-Cooking . These datasets contain videos showing sequences of different cooking actions of preparing meals. All videos in the datasets have temporal segments with corresponding activity labels. Some datasets also provide additional annotations such as object bounding boxes, voice and text in Epic-Kitchen, and the poses of the actors in MPII-Cooking. \nEarly action prediction works widely use the popular UCF-101 dataset and interaction datasets such as UT Interaction (UTI) and BIT . UTI and BIT contain videos of people engaged in interaction with the corresponding label for the types of interactions. In addition, UTI has the added temporal segment annotations detailing different stages of interactions.\n\\textbf{Trajectory prediction. } The most common datasets in this domain are ETH and UCY which contain surveillance videos of pedestrians walking on sidewalks annotated with their position coordinates. UCY also provides the gaze directions to capture the viewing angle of pedestrians. Another popular dataset is Stanford Aerial Pedestrian (SAP), also known as Stanford Drone (SD) . This dataset has the footage of road users from a top-down view recorded by a drone. The annotations include bounding boxes and object class labels. \n\\textbf{Motion prediction. } The algorithms in this domain are mainly evaluated on the widely popular dataset Human 3.6M described earlier. This dataset is particularly suitable for these applications because it contains accurate 3D poses of the actors recorded by a high-speed motion capture system. Using this dataset, the background can be accurately removed allowing the algorithms to focus purely on changes in the poses. Another popular dataset in this field is Penn Action which contains RGB videos of various activities with corresponding activity labels and poses of the actors involved. \n\\textbf{Other applications. } The most notable datasets are KITTI which is used by the OGM prediction algorithms and CityScapes that is used by the segmentation prediction algorithms. CityScapes contains video footage of urban environments recorded by an on-board camera. The data is annotated with semantic masks of the traffic objects and corresponding bounding boxes.","id":"6ab7b3e0-1f9f-4a59-bcf5-66c554040f7e","level":"section","origin_cites_number":113,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Datasets"]],"subsections":[],"title":"Datasets"},{"cite_extract_rate":0,"cites":[],"content":"","id":"2dbff660-6f4e-44d8-90a2-fa6081f28574","level":"section","origin_cites_number":0,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Summary and Discussion"]],"subsections":["52732619-477d-4fcc-bf24-a81fa432764f","24dca31c-1523-4bb9-9bbe-85a5ca0d065e","dd7e38b9-46fb-48ee-b838-9bbba4067a29","6521f90c-6d80-4566-ab87-2b5811291280"],"title":"Summary and Discussion"},{"cite_extract_rate":0,"cites":[],"content":"There are many factors that define the choice of architecture for vision-based prediction tasks. These factors include the types of input data and expected output, computational efficiency, application-specific constrains, etc. For instance, in terms of network choice, whether it is feedforward and recurrent, no preference is observed in video applications. However, in the case of action, trajectory and motion predictions, recurrent architectures are strongly preferred. This can be due to the fact that these applications often rely on multi-modal data which can be combined easier in a recurrent framework. In the case of trajectory prediction, recurrent architectures give the flexibility of varying observation and prediction lengths without the need for architectural modifications. \nGenerative Adversarial Networks (GANs) are widely used in video prediction applications and to some extent in trajectory prediction methods. Some of the major challenges using generative models is to deal with inherent uncertainty of future representations, in particular, this an issue in the context of trajectory prediction due to high unpredictability of human movement. To remedy this issue and to capture uncertainty of movement, techniques such as variational auto encoders, in which uncertainty is modeled as a latent distribution, and the use of probabilistic objective functions are common. \nA more recent trend in the field of vision-based prediction (and perhaps in other computer vision applications) is the use of attention modules. These modules can be applied at spatial or temporal level or even to adjust the impact of different modalities of data.","id":"52732619-477d-4fcc-bf24-a81fa432764f","level":"subsection","origin_cites_number":0,"parent_id":"2dbff660-6f4e-44d8-90a2-fa6081f28574","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Summary and Discussion"],["subsection","Architecture"]],"subsections":[],"title":"Architecture"},{"cite_extract_rate":0,"cites":[],"content":"The type of data and methods of processing vary across different applications. For instance, video applications mainly rely on images but also take advantage of alternative representations, such as optical flow, poses, object-based keypoints, and report improved results. Similarly, many action prediction algorithms use different sources of information such as optical flow, poses, scene attributes, text, complementary sensor readings (e.g. speed in vehicles), gaze and time of the actions.\nTrajectory prediction algorithms, predominantly rely on trajectory information, with some exceptions that use scene layouts, complimentary sensors' readings or other constrains. One of the main applications in this domain, in particular surveillance, is modeling the social interaction between the dynamic agents. Unlike other vision-based applications, motion prediction algorithms are mainly single-modal and use only poses and the images of agents as inputs.","id":"24dca31c-1523-4bb9-9bbe-85a5ca0d065e","level":"subsection","origin_cites_number":0,"parent_id":"2dbff660-6f4e-44d8-90a2-fa6081f28574","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Summary and Discussion"],["subsection","Data representation and processing"]],"subsections":[],"title":"Data representation and processing"},{"cite_extract_rate":0,"cites":[],"content":"","id":"dd7e38b9-46fb-48ee-b838-9bbba4067a29","level":"subsection","origin_cites_number":0,"parent_id":"2dbff660-6f4e-44d8-90a2-fa6081f28574","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Summary and Discussion"],["subsection","Evaluation"]],"subsections":["eb39ad05-ca7d-4eac-ac9e-5926623f5423","f4e6a673-f09d-443b-b219-ea24268d39ec"],"title":"Evaluation"},{"cite_extract_rate":0,"cites":[],"content":"Metrics may vary across different applications of vision-based prediction. Video prediction algorithms, for instance, are mainly evaluated using MSE, SSIM, and PSNR, whereas in the case of action prediction algorithms the main metrics are accuracy, precision, and recall. Trajectory prediction works often measure the average distance (ADE) or final distance (FDE) between the actual and predicted locations of the agents. The models with probabilistic outputs are also evaluated using NLL and KLD metrics. Distance-based metrics are used in motion prediction methods where the error in joint prediction is either calculated on average (MJE) or per joint (MPJPE). In addition, joint accuracy can be reported in terms of the percentage of correct prediction using PCK metric. In this case, a tolerance threshold is defined to determine whether a predicted joint is correct.\nWhile calculating performance error for video and action prediction algorithms are fairly standardized, there are major discrepancies across different works in the way error is computed for trajectory and motion prediction algorithms. For example, in trajectory prediction, distance error is calculated by using metrics such as MSE, RMSE, ED, etc. and units such as pixels and meters. Such discrepancy, and the fact that many works omit mentioning the choice of error metrics and units, increases the chance of incorrect comparisons between models.","id":"eb39ad05-ca7d-4eac-ac9e-5926623f5423","level":"subsubsection","origin_cites_number":0,"parent_id":"dd7e38b9-46fb-48ee-b838-9bbba4067a29","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Summary and Discussion"],["subsection","Evaluation"],["subsubsection","Metrics"]],"subsections":[],"title":"Metrics"},{"cite_extract_rate":0,"cites":[],"content":"The choice of datasets depends on the objective of the applications. For example, action prediction algorithms for cooking activities are evaluated on datasets such as Epic-Kitchen, 50Salads, Breakfast, and MPII-Cooking and the ones for traffic events evaluated on JAAD, NSGIM, and PIE. Similarly, trajectory prediction works for surveillance widely use UCY, ETH, and SD and for traffic NGSIM. Motion prediction algorithms are more focusing on individual movements in diverse context, therefore predominantly use Human3.6M and Penn Action datasets. \nCompared to other applications, video prediction is an exception. The algorithms in this group are evaluated on almost any datasets with video content. The algorithms in this domain are often task agnostic meaning that the same approaches are evaluated on datasets with traffic scenes (e.g. KITTI, Caltech Pedestrian), general activities (e.g. UCF-101, Penn Action), basic actions (e.g. Human3.6M, KTH) and synthetic data (e.g. MMNIST, Atari games). Although such generalizability across different domains is a desirable feature in video prediction algorithms, it is often the case that the reason behind the choice of datasets is not discussed raising the question of whether the decision for selecting particular datasets is motivated by the limitations of the algorithms.","id":"f4e6a673-f09d-443b-b219-ea24268d39ec","level":"subsubsection","origin_cites_number":0,"parent_id":"dd7e38b9-46fb-48ee-b838-9bbba4067a29","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Summary and Discussion"],["subsection","Evaluation"],["subsubsection","Datasets"]],"subsections":[],"title":"Datasets"},{"cite_extract_rate":0,"cites":[],"content":"In recent years we have witnessed an exponential growth in the number of works published in the field of vision-based prediction. There are still, however, many open research problems in the field that need to be addressed. \nThe ability to hallucinate or generate parts of the image that were not previously observed is still a major challenge in video prediction applications. In addition, the algorithms in this domain cannot deal with cases where some objects go out of view in future time frames. The performances of action prediction algorithms are still sub-optimal, especially in safety critical and complex tasks such as event prediction in traffic scenes. To make predictions in such cases, many modalities of data and the relationships between them should be considered which is often not the case in the proposed approaches.\nTrajectory prediction algorithms mainly rely on changes in the location of the agents to predict their future states. Although, this might be an effective approach for tasks such as surveillance, in many other cases it might not be sufficient. For example, in order to predict trajectories of pedestrians in traffic scenes, many other sources of information, such as their poses and orientation, road structure, interactions, road conditions, traffic flow, etc., are potentially relevant. Such contextual analysis can also be beneficial for motion prediction algorithms which manly rely on the changes in poses to predict the future.\nIn terms of the choice of learning architectures and training schemes, a systematic comparison of different approaches, e.g. using feedforward vs recurrent networks, the benefits of using adversarial training schemes, various uncertainty modeling approaches, etc. is missing. Such information can be partially extracted from the existing literature, however, in many cases it is not possible due to the lack of standard evaluation procedures and metrics, unavailability of corresponding implementation code and the datasets used for comparison. \n\\ifCLASSOPTIONcaptionsoff\n \\newpage\n\\fi\n\\bibliographystyle{IEEEtran}\n\\bibliography{vision_based_prediction}\n\\appendices","id":"6521f90c-6d80-4566-ab87-2b5811291280","level":"subsection","origin_cites_number":0,"parent_id":"2dbff660-6f4e-44d8-90a2-fa6081f28574","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Summary and Discussion"],["subsection","What's next"]],"subsections":[],"title":"What's next"},{"cite_extract_rate":0.7272727272727271,"cites":[2871,7117,2817,2838,8592,2808,2874,2869,2819,2856,2866,2809,979,2834,2802,2845,2881,2873,2870,2805,2831,2878,2851,2876,8590,7618,2863,2849,2861,2872,2813,2890],"content":"\\label{papers_with_code}\n\\begin{table*}[]\n\\centering\n\\resizebox{1\\textwidth}{!}{\n\\begin{tabular}{|c|l|l|}\n\\hline\n\\textbf{App} & \\multicolumn{1}{c|}{\\textbf{Paper}} & \\multicolumn{1}{c|}{\\textbf{Link}} \\\\ \\hline\n\\multirow{14}{*}{\\textbf{Video}}\n & & \\url{https://github.com/andrewjywang/SEENet} \\\\ \\cline{2-3} \n & & \\url{https://github.com/Yijunmaverick/FlowGrounded-VideoPrediction} \\\\ \\cline{2-3} \n & & \\url{https://github.com/liuem607/DYAN} \\\\ \\cline{2-3} \n & & \\url{https://github.com/garyzhao/FRGAN} \\\\ \\cline{2-3} \n & & \\url{https://github.com/jthsieh/DDPAE-video-prediction} \\\\ \\cline{2-3} \n & & \\url{https://github.com/xjwxjw/VPSS} \\\\ \\cline{2-3} \n & & \\url{https://github.com/jinbeibei/VarNet} \\\\ \\cline{2-3} \n & & \\url{https://bit.ly/2HqiHqx} \\\\ \\cline{2-3} \n & & \\url{https://github.com/ujjax/pred-rnn} \\\\ \\cline{2-3} \n & & \\url{https://github.com/rubenvillegas/icml2017hierchvid} \\\\ \\cline{2-3} \n & & \\url{https://github.com/tensorflow/models/tree/master/research/video_prediction} \\\\ \\cline{2-3} \n & & \\url{https://github.com/junhyukoh/nips2015-action-conditional-video-prediction} \\\\ \\hline\n\\multirow{8}{*}{\\textbf{Action}} & & \\url{https://github.com/google/next-prediction} \\\\ \\cline{2-3} \n & & \\url{https://github.com/fpv-iplab/rulstm} \\\\ \\cline{2-3} \n & & \\url{https://github.com/aras62/SF-GRU} \\\\ \\cline{2-3} \n & & \\url{https://github.com/aashi7/NearCollision} \\\\ \\cline{2-3} \n & & \\url{https://github.com/yabufarha/anticipating-activities} \\\\ \\cline{2-3} \n & & \\url{https://github.com/gurkirt/realtime-action-detection} \\\\ \\cline{2-3} \n & & \\url{https://github.com/asheshjain399/RNNexp} \\\\ \\cline{2-3} \n & & \\url{https://github.com/aditya7874/Activity-Prediction-in-EgoCentric-Videos} \\\\ \\cline{2-3} \n & & \\url{https://github.com/JoeHEZHAO/Spatiotemporal-Residual-Propagation} \\\\ \\hline\n\\multirow{15}{*}{\\textbf{Trajectory}} & & \\url{https://go.umd.edu/TraPHic} \\\\ \\cline{2-3} \n & & \\url{https://github.com/google/next-prediction} \\\\ \\cline{2-3} \n & & \\url{https://github.com/zhangpur/SR-LSTM} \\\\ \\cline{2-3} \n & & \\url{https://github.com/aras62/PIEPredict} \\\\ \\cline{2-3} \n & & \\url{https://sites.google.com/view/precog} \\\\ \\cline{2-3} \n & & \\url{https://rebrand.ly/INFER-results} \\\\ \\cline{2-3} \n & & \\url{https://github.com/wzhi/KernelTrajectoryMaps} \\\\ \\cline{2-3} \n & & \\url{https://github.com/apratimbhattacharyya18/onboard_long_term_prediction} \\\\ \\cline{2-3} \n & & \\url{https://github.com/agrimgupta92/sgan} \\\\ \\cline{2-3} \n & & \\url{https://github.com/svip-lab/CIDNN} \\\\ \\cline{2-3} \n & & \\url{https://github.com/yfzhang/vehicle-motion-forecasting} \\\\ \\cline{2-3} \n & & \\url{https://github.com/yadrimz/DESIRE} \\\\ \\cline{2-3} \n & & \\url{https://github.com/quancore/social-lstm} \\\\ \\hline\n\\multirow{11}{*}{\\textbf{Motion}} & & \\url{https://github.com/cr7anand/neural_temporal_models} \\\\ \\cline{2-3} \n & & \\url{https://github.com/BII-wushuang/Lie-Group-Motion-Prediction} \\\\ \\cline{2-3} \n & & \\url{https://github.com/wei-mao-2019/LearnTrajDep} \\\\ \\cline{2-3} \n & & \\url{https://github.com/magnux/MotionGAN} \\\\ \\cline{2-3} \n & & \\url{https://github.com/Khrylx/EgoPose} \\\\ \\cline{2-3} \n & & \\url{https://jasonyzhang.com/phd/} \\\\ \\cline{2-3} \n & & \\url{https://github.com/eddyhkchiu/pose_forecast_wacv/.} \\\\ \\cline{2-3} \n & & \\url{https://github.com/ywchao/image-play} \\\\ \\cline{2-3} \n & & \\url{https://github.com/una-dinosauria/human-motion-prediction} \\\\ \\cline{2-3} \\hline\n\\multicolumn{1}{|l|}{\\multirow{2}{*}{\\textbf{Others}}} & & \\url{https://bitbucket.org/vguizilini/cvpp/src} \\\\ \\cline{2-3} \n\\multicolumn{1}{|l|}{} & & \\url{https://github.com/facebookresearch/SegmPred} \\\\ \\hline\n\\end{tabular}\n}\n\\caption{A summary of vision-based prediction papers with published code.}\n\\label{table_published_code}\n\\end{table*}\nA list of papers with official published code can be found in Table \\ref{table_published_code}.","id":"23d750d7-3a7a-460d-8e6b-4f54e379529f","level":"section","origin_cites_number":44,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Papers with code"]],"subsections":[],"title":"Papers with code"},{"cite_extract_rate":0.461187214611872,"cites":[2829,2871,2844,2807,2808,2838,2869,2841,2862,2822,979,2840,2824,2811,2848,2845,2850,2835,2873,2870,2827,2842,2836,2887,9096,1261,8590,2849,2877,2813,2814,2855,2860,7117,2853,2879,2816,2823,2856,2866,2852,2888,2809,2858,2802,2865,2805,2826,2833,2876,7618,2839,2830,2843,2847,2854,8592,2882,2868,2832,2874,2857,2828,2806,8595,2834,2859,2875,2881,2812,2886,2878,2851,2863,7617,7379,2803,2890,8594,2804,2821,2837,2867,2817,2880,2885,2819,2846,2815,2825,2883,988,2818,2820,2810,2831,2884,8591,2861,2872,2864],"content":"\\label{metrics_papers}\n\\begin{table*}[ht]\n\\centering\n\\begin{tabular}{|l|p{0.8\\textwidth}|}\n\\hline\n\\textbf{Metric} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{BCE}} & \\\\ \\hline\n\\textit{\\textbf{FVD}} & , \\\\ \\hline\n\\textit{\\textbf{Human}} & ,,, \\\\ \\hline\n\\textit{\\textbf{IS}} & \\\\ \\hline\n\\textit{\\textbf{L1}} & , \\\\ \\hline\n\\textit{\\textbf{LPIPS}} & ,,, \\\\ \\hline\n\\textit{\\textbf{MMD}} & \\\\ \\hline\n\\textit{\\textbf{MSE}} & ,,,,, ,,,,, ,, \\\\ \\hline\n\\textit{\\textbf{PSNR}} & ,,,,, ,,,,, ,,,,, ,,,,, ,,,, \\\\ \\hline\n\\textit{\\textbf{RMSE}} & \\\\ \\hline\n\\textit{\\textbf{SSIM}} & ,,,,, ,,,,, ,,,,, ,,,,, ,, \\\\ \\hline\n\\end{tabular}\n\\caption{Metrics used in \\textbf{video prediction} applications.}\n\\label{video_metrics_table}\n\\end{table*}\n\\begin{table*}[ht]\n\\centering\n\\begin{tabular}{|l|p{0.8\\textwidth}|}\n\\hline\n\\textbf{Metrics} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{AP}} & ,,,,, \\\\ \\hline\n\\textit{\\textbf{ATTA}} & \\\\ \\hline\n\\textit{\\textbf{ATTC}} & \\\\ \\hline\n\\textit{\\textbf{AUC}} & ,, \\\\ \\hline\n\\textit{\\textbf{Accuracy}} & ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,,,,,,, , \\\\ \\hline\n\\textit{\\textbf{F1}} & ,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{FP}} & \\\\ \\hline\n\\textit{\\textbf{MAE}} & \\\\ \\hline\n\\textit{\\textbf{MCC}} & \\\\ \\hline\n\\textit{\\textbf{MRR}} & \\\\ \\hline\n\\textit{\\textbf{PP}} & \\\\ \\hline\n\\textit{\\textbf{Precision}} & ,,,,, ,,,,, ,,,,, \\\\ \\hline\n\\textit{\\textbf{RMSE}} & ,, \\\\ \\hline\n\\textit{\\textbf{Recall}} & ,,,,, ,,,,, ,,,,, ,, \\\\ \\hline\n\\textit{\\textbf{Run time}} & , \\\\ \\hline\n\\textit{\\textbf{TNR}} & \\\\ \\hline\n\\textit{\\textbf{TPR}} & \\\\ \\hline\n\\textit{\\textbf{TTA}} & \\\\ \\hline\n\\textit{\\textbf{TTM}} & ,,,, \\\\ \\hline\n\\textit{\\textbf{cAP}} & \\\\ \\hline\n\\textit{\\textbf{mAP}} & ,,,,,,, \\\\ \\hline\n\\textit{\\textbf{recall}} & , \\\\ \\hline\n\\end{tabular}\n\\caption{Metrics used in \\textbf{action prediction} applications.}\n\\label{action_metrics_table}\n\\end{table*}\n\\begin{table*}[ht]\n\\centering\n\\begin{tabular}{|l|p{0.8\\textwidth}|}\n\\hline\n\\textbf{Metrics} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{ADE}} & ,,,,, ,,,,,,,,,, ,,,,, ,,,,,,,,,, ,,,,, ,,,,,,,,,, ,\n\\\\ \\hline\n\\textit{\\textbf{AEDE}} & \\\\ \\hline\n\\textit{\\textbf{ANDE}} & , \\\\ \\hline\n\\textit{\\textbf{APP}} & \\\\ \\hline\n\\textit{\\textbf{Accuracy}} & , \\\\ \\hline\n\\textit{\\textbf{CE}} & \\\\ \\hline\n\\textit{\\textbf{DtG}} & \\\\ \\hline\n\\textit{\\textbf{ECE}} & \\\\ \\hline\n\\textit{\\textbf{ED}} & ,,,,,,,,,\n\\\\ \\hline\n\\textit{\\textbf{FDE}} & ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,\n\\\\ \\hline\n\\textit{\\textbf{FNM}} & \\\\ \\hline\n\\textit{\\textbf{Hit rate}} & , \\\\ \\hline\n\\textit{\\textbf{KLD}} & ,, \\\\ \\hline\n\\textit{\\textbf{L1}} & \\\\ \\hline\n\\textit{\\textbf{LL}} & ,,, \\\\ \\hline\n\\textit{\\textbf{MAnE}} & \\\\ \\hline\n\\textit{\\textbf{MHD}} & ,, \\\\ \\hline\n\\textit{\\textbf{MSE}} & , \\\\ \\hline\n\\textit{\\textbf{Miss rate}} & , \\\\ \\hline\n\\textit{\\textbf{NLL}} & ,,,,, \\\\ \\hline\n\\textit{\\textbf{NLP}} & , \\\\ \\hline\n\\textit{\\textbf{None}} & \\\\ \\hline\n\\textit{\\textbf{PD}} & \\\\ \\hline\n\\textit{\\textbf{Precision}} & \\\\ \\hline\n\\textit{\\textbf{RMSE}} & , \\\\ \\hline\n\\textit{\\textbf{Run time}} & ,,,, \\\\ \\hline\n\\textit{\\textbf{SCR}} & \\\\ \\hline\n\\textit{\\textbf{WRMSE}} & \\\\ \\hline\n\\textit{\\textbf{maxD}} & , \\\\ \\hline\n\\textit{\\textbf{meanMSD}} & , \\\\ \\hline\n\\textit{\\textbf{minADE}} & ,,, \\\\ \\hline\n\\textit{\\textbf{minED}} & , \\\\ \\hline\n\\textit{\\textbf{minFDE}} & , \\\\ \\hline\n\\textit{\\textbf{minMSD}} & , \\\\ \\hline\n\\end{tabular}\n\\caption{Metrics used in \\textbf{trajectory prediction} applications.}\n\\label{trajectory_metrics_table}\n\\end{table*}\n\\begin{table*}[h]\n\\centering\n\\begin{tabular}{|l|l|}\n\\hline\n\\textbf{Metrics} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{Accuracy}} & , \\\\ \\hline\n\\textit{\\textbf{Human}} & , \\\\ \\hline\n\\textit{\\textbf{LO}} & \\\\ \\hline\n\\textit{\\textbf{MAnE}} & ,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{MJE}} & ,,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{MPJPE}} & , \\\\ \\hline\n\\textit{\\textbf{NPSS}} & \\\\ \\hline\n\\textit{\\textbf{PCK}} & ,,, \\\\ \\hline\n\\textit{\\textbf{PSEnt}} & \\\\ \\hline\n\\textit{\\textbf{PSKL}} & \\\\ \\hline\n\\textit{\\textbf{RE}} & \\\\ \\hline\n\\textit{\\textbf{Run time}} & , \\\\ \\hline\n\\end{tabular}\n\\caption{Metrics used in motion prediction applications.}\n\\label{motion_metrics_table}\n\\end{table*}\nLists of metrics and corresponding papers can be found in Tables \\ref{video_metrics_table}, \\ref{action_metrics_table}, \\ref{trajectory_metrics_table}, \\ref{motion_metrics_table}, and \\ref{other_metrics_table}. \n\\begin{table*}[]\n\\centering\n\\begin{tabular}{|l|l|}\n\\hline\n\\textbf{Metrics} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{AUC}} & \\\\ \\hline\n\\textit{\\textbf{Accuracy}} & ,, \\\\ \\hline\n\\textit{\\textbf{ED}} & , \\\\ \\hline\n\\textit{\\textbf{EPE}} & \\\\ \\hline\n\\textit{\\textbf{F1}} & , \\\\ \\hline\n\\textit{\\textbf{GCE}} & \\\\ \\hline\n\\textit{\\textbf{ISM}} & \\\\ \\hline\n\\textit{\\textbf{IoU}} & ,, \\\\ \\hline\n\\textit{\\textbf{MAE}} & ,, \\\\ \\hline\n\\textit{\\textbf{MAPE}} & , \\\\ \\hline\n\\textit{\\textbf{MCC}} & \\\\ \\hline\n\\textit{\\textbf{MIoU}} & \\\\ \\hline\n\\textit{\\textbf{MSE}} & \\\\ \\hline\n\\textit{\\textbf{PCP}} & \\\\ \\hline\n\\textit{\\textbf{PSNR}} & , \\\\ \\hline\n\\textit{\\textbf{Precision}} & ,, \\\\ \\hline\n\\textit{\\textbf{Psi}} & \\\\ \\hline\n\\textit{\\textbf{RI}} & \\\\ \\hline\n\\textit{\\textbf{RMSE}} & \\\\ \\hline\n\\textit{\\textbf{ROC}} & , \\\\ \\hline\n\\textit{\\textbf{Recall}} & ,, \\\\ \\hline\n\\textit{\\textbf{Run time}} & ,, \\\\ \\hline\n\\textit{\\textbf{SRC}} & \\\\ \\hline\n\\textit{\\textbf{SSIM}} & ,, \\\\ \\hline\n\\textit{\\textbf{TN}} & \\\\ \\hline\n\\textit{\\textbf{TP}} & \\\\ \\hline\n\\textit{\\textbf{VoI}} & \\\\ \\hline\n\\textit{\\textbf{nMAPE}} & \\\\ \\hline\n\\end{tabular}\n\\caption{Metrics used in \\textbf{other prediction} applications.}\n\\label{other_metrics_table}\n\\end{table*}\n\\clearpage\n\\clearpage","id":"63a0cc61-98a6-4be7-b982-ebf42e368ec1","level":"section","origin_cites_number":219,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metrics and corresponding papers"]],"subsections":[],"title":"Metrics and corresponding papers"},{"cite_extract_rate":0,"cites":[],"content":"\\label{appendix_metrics}","id":"14c8fcc2-4517-4a60-9512-dfff6198b563","level":"section","origin_cites_number":0,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metric formulas"]],"subsections":["e9d8a36b-2d5e-4533-a6ab-b63d369319dc","ddfb59d7-61bc-4e3f-8381-c962428d999f","ed1fa5f0-0752-4fc1-89e1-31159f822bee","327259be-0b0d-405f-9880-1b39cd8f50fa"],"title":"Metric formulas"},{"cite_extract_rate":0,"cites":[],"content":"$$ MSE= \\frac{1}{MN} \\sum_{i=1}^M\\sum_{j=1}^N(I(i,j) -\\tilde{I}(i,j))^2$$\n $$PSNR = 20\\log\\left(\\frac{MAX_I}{\\sqrt{MSE}}\\right)$$\n\\textbf{Structural Similarity (SSIM)} \n$$luminance (l)(x,y) = \\frac{2\\mu_x\\mu_y + C_1}{\\mu_x^2 + \\mu_y^2 + C_1}$$\n$$\\mu_x=\\frac{1}{N}\\sum_{i=1}^Nx_i$$\n$$C_1=(K_1L)^2$$\n\\noindent where $L$ is dynamic range of pixel values (e.g. 255) and $K_1 \\ll 1$ is a small constant.\n$$contrast (c)(x,y) = \\frac{2\\sigma_x\\sigma_y+C_2}{\\sigma_x^2+\\sigma_y^2+C_2}$$\n$$\\sigma_x=\\left(\\frac{1}{N-1}\\sum_{i=1}^N(x_i-\\mu_x)^2\\right)^{1/2}$$\n$$C_2=(K_2L)^2$$ $$K_2 \\ll 1$$\n$$structure(s)(x,y) = \\frac{\\sigma_{xy}+C_3}{\\sigma_x\\sigma_y+C_3}$$\n$$C_3=(K_3L)^2$$ $$K_3 \\ll 1$$\n$$SSIM(x,y)=[l(x,y)]^\\alpha.[c(x,y)]^\\beta.[s(x,y)]^\\gamma$$\n\\noindent where $\\alpha ,\\beta, \\gamma >0$ are parameters to choose in order to adjust the importance.\n\\textbf{Learned Perceptual Image Patch Similarity (LPIPS)}\nAssume features are extracted from $L$ layers and unit-normalized in channel dimension, for layer l\n$$\\hat{y}^l,\\hat{y}^l \\in R^{H_l\\times W_l \\times C_l}$$.\nThe distance between reference $x$ and distorted patches $x_0$ is given by,\n$$d(x,x_0)=\\sum_l \\frac{1}{H_lW_l}\\sum_{w,l}\\parallel w_l \\odot (\\hat{y}^l_{hw},\\hat{y}^l_{0hw})\\parallel ^2_2$$","id":"e9d8a36b-2d5e-4533-a6ab-b63d369319dc","level":"subsection","origin_cites_number":0,"parent_id":"14c8fcc2-4517-4a60-9512-dfff6198b563","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metric formulas"],["subsection","Video prediction"]],"subsections":[],"title":"Video prediction"},{"cite_extract_rate":0,"cites":[],"content":"There are 4 possibilities for classification: True positive (TP) and True Negative (TN) when the algorithm correctly classifies positive and negative samples, and False Positive (FP) and False Negative (FN) when the algorithm incorrectly classifies negative samples as positive and vice versa.\n$$Accuracy = \\frac{TN+TP}{TP+TN+FP+FN}$$\n$$Precision = \\frac{TP}{TP+FP}$$\n$$Recall = \\frac{TP}{TP+FN}$$\n$$F1-score = 2 \\times \\frac{Precision \\times Recall}{Precision + Recall}$$\n$$RMSE = \\sqrt{\\frac{1}{n} \\sum_{i=1}^n(y_i -\\tilde{y_i})}$$\nLet $p(r)$ be the precision-recall curve. Then,\n$$AP = \\int_0^1 p(r)dr$$","id":"ddfb59d7-61bc-4e3f-8381-c962428d999f","level":"subsection","origin_cites_number":0,"parent_id":"14c8fcc2-4517-4a60-9512-dfff6198b563","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metric formulas"],["subsection","Action prediction"]],"subsections":[],"title":"Action prediction"},{"cite_extract_rate":0,"cites":[],"content":"","id":"ed1fa5f0-0752-4fc1-89e1-31159f822bee","level":"subsection","origin_cites_number":0,"parent_id":"14c8fcc2-4517-4a60-9512-dfff6198b563","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metric formulas"],["subsection","Trajectory prediction"]],"subsections":["4a5b290c-66e4-4bd3-b5e3-68e6a0a15129"],"title":"Trajectory prediction"},{"cite_extract_rate":0,"cites":[],"content":"$$\\text{Euclidean Distance}(ED) = \\parallel y-\\tilde{y}\\parallel = \\parallel y-\\tilde{y}\\parallel_2$$\n$$ = \\sqrt{\\sum_{i=1}^n(y_i -\\tilde{y_i})^2} $$\n$$ \\text{Mean Absolute Error}(MAE) = \\frac{1}{n} \\sum_{i=1}^n|y_i -\\tilde{y_i}| $$\n$$\\text{Mean Square Error} (MSE) = \\parallel y-\\tilde{y}\\parallel^2$$\n$$ =\\frac{1}{n} \\sum_{i=1}^n(y_i -\\tilde{y_i})^2$$\n$$\\text{Root MSE} (RMSE) = \\sqrt{\\frac{1}{n} \\sum_{i=1}^n(y_i -\\tilde{y_i})^2}$$\n$$\\text{Hausdorff Distance}(HD) = max_{y\\in Y}min_{\\tilde{y}\\in\\tilde{Y}} \\parallel y-\\tilde{y}\\parallel$$\n$$\\text{Modified HD}(MHD) = max(d(Y,\\tilde{Y}),d(\\tilde{Y}, Y))$$\n$$ d(Y,\\tilde{Y}) = \\frac{1}{N_y}\\sum_{y\\in Y} min_{\\tilde{y}\\in \\tilde{Y}} \\parallel y - \\tilde{y}\\parallel$$\n$$ ADE = \\frac{\\sum^N_{i=1}\\sum^{T_{pred}}_{t=1}\\parallel \\tilde{y}^i_t - y^i_t \\parallel}{N \\times T_{pred}}$$\nwhere $N$ is the number of samples and $T_{pred}$ is the prediction steps.\n$$FDE = \\frac{ \\sum^N_{i=1} \\parallel \\tilde{y}^i_{T_{pred}} - y^i_{T_{pred}} \\parallel}{N}$$\n$$ minMSD = \\mathbb{E}_{\\tilde{Y}_k\\sim q_\\theta}min_{\\tilde{y}\\in\\tilde{Y}_k}\\parallel y-\\tilde{y}\\parallel^2$$\nwhere $q_\\theta$ is the sampling space and $K$ number of samples. \n$$ meanMSD = \\frac{1}{K}\\sum_{k=1}^K\\parallel y-\\tilde{y}\\parallel^2$$\n$$NLL = \\mathbb{E}_{p(Y|X)}\\left[-\\log \\prod_{t=1}^{T_{pred}} p(y_t|X)\\right]$$","id":"4a5b290c-66e4-4bd3-b5e3-68e6a0a15129","level":"subsubsection","origin_cites_number":0,"parent_id":"ed1fa5f0-0752-4fc1-89e1-31159f822bee","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metric formulas"],["subsection","Trajectory prediction"],["subsubsection","Distance metrics"]],"subsections":[],"title":"Distance metrics"},{"cite_extract_rate":0,"cites":[],"content":"$$MPJPE = \\frac{1}{N \\times T_{pred}}\\sum_{t=1}^{T_{pred}}\\sum_{i=1}^N \\parallel (J_i^t-J_{root}^t)-(\\tilde{J}_i^t-\\tilde{J}_{root}^t)\\parallel$$\n\\clearpage","id":"327259be-0b0d-405f-9880-1b39cd8f50fa","level":"subsection","origin_cites_number":0,"parent_id":"14c8fcc2-4517-4a60-9512-dfff6198b563","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Metric formulas"],["subsection","Motion prediction"]],"subsections":[],"title":"Motion prediction"},{"cite_extract_rate":0.33628318584070804,"cites":[2899,2838,2808,2869,2824,2892,8596,2835,2897,1517,2903,8590,2894,6980,2891,2900,2626,2908,2906,2896,2905,2898,2902,8597,1733,2904,2907,2893,2878,2851,7379,2901,2880,2895],"content":"\\label{links_datasets}\n\\begin{table*}[]\n\\centering\n\\resizebox{1\\textwidth}{!}{\n\\begin{tabular}{|l|l|l|}\n\\hline\nYear & Dataset & Links \\\\ \\hline\n\\multirow{12}{*}{2019} & ARGOVerse & \\url{https://www.argoverse.org/data.html} \\\\ \\cline{2-3} \n & CARLA & \\url{https://sites.google.com/view/precog} \\\\ \\cline{2-3} \n & EgoPose & \\url{https://github.com/Khrylx/EgoPose} \\\\ \\cline{2-3} \n & FM & \\url{https://mcl.korea.ac.kr/$\\sim$krkim/iccv2019/index.html} \\\\ \\cline{2-3} \n & InstaVariety & \\url{https://github.com/akanazawa/human_dynamics} \\\\ \\cline{2-3} \n & INTEARCTION & \\url{https://interaction-dataset.com} \\\\ \\cline{2-3} \n & Luggage & \\url{https://aashi7.github.io/NearCollision.html} \\\\ \\cline{2-3} \n & MGIF & \\url{https://github.com/AliaksandrSiarohin/monkey-net} \\\\ \\cline{2-3} \n & PIE & \\url{http://data.nvision2.eecs.yorku.ca/PIE_dataset/} \\\\ \\cline{2-3} \n & nuScenes & \\url{https://www.nuscenes.org/} \\\\ \\cline{2-3} \n & Vehicle-Pedestrian-Mixed (VPM) & \\url{http://vr.ict.ac.cn/vp-lstm.} \\\\ \\cline{2-3} \n & TRAF & \\url{https://drive.google.com/drive/folders/1LqzJuRkx5yhOcjWFORO5WZ97v6jg8RHN} \\\\ \\hline\n\\multirow{10}{*}{2018} & 3DPW & \\url{https://virtualhumans.mpi-inf.mpg.de/3DPW/} \\\\ \\cline{2-3} \n & ActEV/VIRAT & \\url{https://actev.nist.gov/trecvid19} \\\\ \\cline{2-3} \n & ACTICIPATE & \\url{http://vislab.isr.tecnico.ulisboa.pt/datasets/} \\\\ \\cline{2-3} \n & AVA & \\url{https://research.google.com/ava/} \\\\ \\cline{2-3} \n & Epic-Kitchen & \\url{https://epic-kitchens.github.io/2019} \\\\ \\cline{2-3} \n & EGTEA Gaze+ & \\url{http://www.cbi.gatech.edu/fpv/} \\\\ \\cline{2-3} \n & STC & \\url{https://svip-lab.github.io/dataset/campus_dataset.html} \\\\ \\cline{2-3} \n & ShapeStack & \\url{https://shapestacks.robots.ox.ac.uk/} \\\\ \\cline{2-3} \n & VIENA & \\url{https://sites.google.com/view/viena2-project/home} \\\\ \\cline{2-3} \n & YouCook2 & \\url{http://youcook2.eecs.umich.edu/} \\\\ \\hline\n\\multirow{10}{*}{2017} & BUA & \\url{http://cs-people.bu.edu/sbargal/BU-action/} \\\\ \\cline{2-3} \n & CityPerson & \\url{https://bitbucket.org/shanshanzhang/citypersons/src/default/} \\\\ \\cline{2-3} \n & Epic-fail & \\url{http://aliensunmin.github.io/project/video-Forecasting/} \\\\ \\cline{2-3} \n & JAAD & \\url{http://data.nvision2.eecs.yorku.ca/JAAD_dataset/} \\\\ \\cline{2-3} \n & L-CAS & \\url{https://lcas.lincoln.ac.uk/wp/research/data-sets-software/l-cas-3d-point-cloud-people-dataset/} \\\\ \\cline{2-3} \n & Mouse Fish & \\url{https://web.bii.a-star.edu.sg/archive/machine_learning/Projects/behaviorAnalysis/Lie-X/Lie-X.html} \\\\ \\cline{2-3} \n & ORC & \\url{https://robotcar-dataset.robots.ox.ac.uk/} \\\\ \\cline{2-3} \n & PKU-MMD & \\url{http://www.icst.pku.edu.cn/struct/Projects/PKUMMD.html} \\\\ \\cline{2-3} \n & Recipe1M & \\url{http://pic2recipe.csail.mit.edu/} \\\\ \\cline{2-3} \n & STRANDS & \\url{https://strands.readthedocs.io/en/latest/datasets/} \\\\ \\hline\n\\multirow{13}{*}{2016} & BAIR Push & \\url{https://sites.google.com/site/brainrobotdata/home/push-dataset} \\\\ \\cline{2-3} \n & BB & \\url{https://github.com/mbchang/dynamics} \\\\ \\cline{2-3} \n & MU & \\url{http://staff.itee.uq.edu.au/lovell/MissUniverse/} \\\\ \\cline{2-3} \n & Cityscapes & \\url{https://www.cityscapes-dataset.com/} \\\\ \\cline{2-3} \n & CMU mocap & \\url{http://mocap.cs.cmu.edu/} \\\\ \\cline{2-3} \n & DAD & \\url{https://aliensunmin.github.io/project/dashcam/} \\\\ \\cline{2-3} \n & NTU RGB-D & \\url{http://rose1.ntu.edu.sg/Datasets/actionRecognition.asp} \\\\ \\cline{2-3} \n & OA & \\url{http://www.mpii.de/ongoing-activity} \\\\ \\cline{2-3} \n & OAD & \\url{http://www.icst.pku.edu.cn/struct/Projects/OAD.html} \\\\ \\cline{2-3} \n & SD & \\url{http://cvgl.stanford.edu/projects/uav_data/} \\\\ \\cline{2-3} \n & TV Series & \\url{https://github.com/zhenyangli/online_action} \\\\ \\cline{2-3} \n & VIST & \\url{http://visionandlanguage.net/VIST/} \\\\ \\cline{2-3} \n & Youtube-8M & \\url{https://research.google.com/youtube8m/} \\\\ \\hline\n\\multirow{14}{*}{2015} & Amazon & \\url{http://jmcauley.ucsd.edu/data/amazon/index_2014.html} \\\\ \\cline{2-3} \n & Atari & \\url{https://github.com/junhyukoh/nips2015-action-conditional-video-prediction} \\\\ \\cline{2-3} \n & Brain4Cars & \\url{https://github.com/asheshjain399/ICCV2015_Brain4Cars} \\\\ \\cline{2-3} \n & CMU Panoptic & \\url{http://domedb.perception.cs.cmu.edu/dataset.html} \\\\ \\cline{2-3} \n & FPPA & \\url{http://bvision11.cs.unc.edu/bigpen/yipin/ICCV2015/prediction_webpage/Prediction.html} \\\\ \\cline{2-3} \n & GTEA Gaze+ & \\url{http://www.cbi.gatech.edu/fpv/} \\\\ \\cline{2-3} \n & MBI-1M & \\url{http://academic.mywebsiteontheinternet.com/data/} \\\\ \\cline{2-3} \n & MOT & \\url{https://motchallenge.net/} \\\\ \\cline{2-3} \n & MMNIST & \\url{http://www.cs.toronto.edu/$\\sim$nitish/unsupervised_video/} \\\\ \\cline{2-3} \n & SUN RGB-D & \\url{http://rgbd.cs.princeton.edu/} \\\\ \\cline{2-3} \n & SYSU 3DHOI & \\url{http://www.isee-ai.cn/$\\sim$hujianfang/ProjectJOULE.html} \\\\ \\cline{2-3} \n & THUMOS & \\url{http://www.thumos.info/home.html} \\\\ \\cline{2-3} \n & WnP & \\url{http://watchnpatch.cs.cornell.edu/} \\\\ \\cline{2-3} \n & Wider & \\url{http://yjxiong.me/event_recog/WIDER/} \\\\ \\hline\n\\multicolumn{1}{|c|}{\\multirow{5}{*}{2014}} & Breakfast & \\url{http://serre-lab.clps.brown.edu/resource/breakfast-actions-dataset/} \\\\ \\cline{2-3} \n\\multicolumn{1}{|c|}{} & Human3.6M & \\url{http://vision.imar.ro/human3.6m/description.php} \\\\ \\cline{2-3} \n\\multicolumn{1}{|c|}{} & MPII Human Pose & \\url{http://human-pose.mpi-inf.mpg.de/} \\\\ \\cline{2-3} \n\\multicolumn{1}{|c|}{} & ORGBD & \\url{https://sites.google.com/site/skicyyu/orgbd} \\\\ \\cline{2-3} \n\\multicolumn{1}{|c|}{} & Sports-1M & \\url{https://cs.stanford.edu/people/karpathy/deepvideo/} \\\\ \\hline\n\\end{tabular}\n}\n\\caption{A summary of datasets (from year 2014-2019) used in vision-based prediction papers and corresponding links.}\n\\label{data_set_links1}\n\\end{table*}\n\\begin{table*}[]\n\\centering\n\\resizebox{1\\textwidth}{!}{\n\\begin{tabular}{|l|l|l|}\n\\hline\nYear & Dataset & Links \\\\ \\hline\n\\multirow{7}{*}{2013} & 50 salads & \\url{https://cvip.computing.dundee.ac.uk/datasets/foodpreparation/50salads/} \\\\ \\cline{2-3} \n & ATC & \\url{https://irc.atr.jp/crest2010_HRI/ATC_dataset/} \\\\ \\cline{2-3} \n & CAD-120 & \\url{http://pr.cs.cornell.edu/humanactivities/data.php} \\\\ \\cline{2-3} \n & CHUK Avenue & \\url{http://www.cse.cuhk.edu.hk/leojia/projects/detectabnormal/dataset.html} \\\\ \\cline{2-3} \n & Daimler path & \\url{http://www.gavrila.net/Datasets/Daimler_Pedestrian_Benchmark_D/Pedestrian_Path_Predict_GCPR_1/pedestrian_path_predict_gcpr_1.html} \\\\ \\cline{2-3} \n & JHMDB & \\url{http://jhmdb.is.tue.mpg.de/} \\\\ \\cline{2-3} \n & Penn Action & \\url{http://dreamdragon.github.io/PennAction/} \\\\ \\hline\n\\multirow{11}{*}{2012} & BIT & \\url{https://sites.google.com/site/alexkongy/software} \\\\ \\cline{2-3} \n & GTEA Gaze & \\url{http://www.cbi.gatech.edu/fpv/} \\\\ \\cline{2-3} \n & KITTI & \\url{http://www.cvlibs.net/datasets/kitti/} \\\\ \\cline{2-3} \n & MANIAC & \\url{https://alexandria.physik3.uni-goettingen.de/cns-group/datasets/maniac/} \\\\ \\cline{2-3} \n & MPII-cooking & \\url{https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/research/human-activity-recognition/mpii-cooking-activities-dataset/} \\\\ \\cline{2-3} \n & MSRDA & \\url{https://documents.uow.edu.au/$\\sim$wanqing/\\#MSRAction3DDatasets} \\\\ \\cline{2-3} \n & GC & \\url{http://www.ee.cuhk.edu.hk/$\\sim$xgwang/grandcentral.html} \\\\ \\cline{2-3} \n & SBUKI & \\url{https://www3.cs.stonybrook.edu/$\\sim$kyun/research/kinect_interaction/index.html} \\\\ \\cline{2-3} \n & UCF-101 & \\url{https://www.crcv.ucf.edu/data/UCF101.php} \\\\ \\cline{2-3} \n & UTKA & \\url{http://cvrc.ece.utexas.edu/KinectDatasets/HOJ3D.html} \\\\ \\cline{2-3} \n & UvA-NEMO & \\url{https://www.uva-nemo.org/} \\\\ \\hline\n\\multirow{5}{*}{2011} & FCVL & \\url{http://robots.engin.umich.edu/SoftwareData/Ford} \\\\ \\cline{2-3} \n & HMDB & \\url{http://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/} \\\\ \\cline{2-3} \n & Stanford 40 & \\url{http://vision.stanford.edu/Datasets/40actions.html} \\\\ \\cline{2-3} \n & Town Center & \\url{http://www.robots.ox.ac.uk/ActiveVision/Research/Projects/2009bbenfold_headpose/project.html\\#datasets} \\\\ \\cline{2-3} \n & VIRAT & \\url{http://viratdata.org/} \\\\ \\hline\n\\multirow{8}{*}{2010} & DISPLECS & \\url{https://cvssp.org/data/diplecs/} \\\\ \\cline{2-3} \n & MSR & \\url{https://www.microsoft.com/en-us/download/details.aspx?id=52315} \\\\ \\cline{2-3} \n & MUG Facial Expression & \\url{https://mug.ee.auth.gr/fed/} \\\\ \\cline{2-3} \n & PROST & \\url{www.gpu4vision.com } \\\\ \\cline{2-3} \n & THI & \\url{http://www.robots.ox.ac.uk/$\\sim$alonso/tv_human_interactions.html} \\\\ \\cline{2-3} \n & UTI & \\url{http://cvrc.ece.utexas.edu/SDHA2010/Human_Interaction.html} \\\\ \\cline{2-3} \n & VISOR & \\url{imagelab.ing.unimore.it/visor} \\\\ \\cline{2-3} \n & Willow Action & \\url{https://www.di.ens.fr/willow/research/stillactions/} \\\\ \\hline\n\\multirow{9}{*}{2009} & Caltech Pedestrian & \\url{http://www.vision.caltech.edu/Image_Datasets/CaltechPedestrians/} \\\\ \\cline{2-3} \n & Collective Activity (CA) & \\url{http://www-personal.umich.edu/$\\sim$wgchoi/eccv12/wongun_eccv12.html} \\\\ \\cline{2-3} \n & EIFP & \\url{http://homepages.inf.ed.ac.uk/rbf/FORUMTRACKING/} \\\\ \\cline{2-3} \n & ETH & \\url{http://www.vision.ee.ethz.ch/en/datasets/} \\\\ \\cline{2-3} \n & OSU & \\url{http://eecs.oregonstate.edu/football/tracking/dataset} \\\\ \\cline{2-3} \n & PETS2009 & \\url{http://www.cvg.reading.ac.uk/PETS2009/a.html} \\\\ \\cline{2-3} \n & QMUL & \\url{http://personal.ie.cuhk.edu.hk/$\\sim$ccloy/downloads_qmul_junction.html} \\\\ \\cline{2-3} \n & TUM Kitchen & \\url{https://ias.in.tum.de/dokuwiki/software/kitchen-activity-data} \\\\ \\cline{2-3} \n & YUV Videos & \\url{http://trace.kom.aau.dk/yuv/index.html} \\\\ \\hline\n\\multirow{2}{*}{2008} & Daimler & \\url{http://www.gavrila.net/Datasets/Daimler_Pedestrian_Benchmark_D/Daimler_Mono_Ped__Detection_Be/daimler_mono_ped__detection_be.html} \\\\ \\cline{2-3} \n & MITT & \\url{http://www.ee.cuhk.edu.hk/$\\sim$xgwang/MITtrajsingle.html} \\\\ \\hline\n\\multirow{5}{*}{2007} & AMOS & \\url{http://amos.cse.wustl.edu/} \\\\ \\cline{2-3} \n & ETH pedestrian & \\url{https://data.vision.ee.ethz.ch/cvl/aess/} \\\\ \\cline{2-3} \n & Lankershim Boulevard & \\url{https://www.fhwa.dot.gov/publications/research/operations/07029/index.cfm} \\\\ \\cline{2-3} \n & NGSIM & \\url{https://ops.fhwa.dot.gov/trafficanalysistools/ngsim.htm} \\\\ \\cline{2-3} \n & UCY & \\url{https://graphics.cs.ucy.ac.cy/research/downloads/crowd-data} \\\\ \\hline\n2006 & Tuscan Arizona & \\url{http://www.mmto.org/} \\\\ \\hline\n2004 & KTH & \\url{http://www.nada.kth.se/cvap/actions/} \\\\ \\hline\n1981 & Golden Colorado & \\url{https://www.osti.gov/dataexplorer/biblio/dataset/1052221} \\\\ \\hline\n\\end{tabular}\n}\n\\caption{A summary of datasets (from year 2013 and earlier) used in vision-based prediction papers and corresponding links.}\n\\label{data_set_links2}\n\\end{table*}\nLists of datasets with associated repository links can be found in Tables \\ref{data_set_links1} and \\ref{data_set_links2}.\n\\clearpage","id":"e82c2ffc-f5dc-4d01-9184-399825652bbe","level":"section","origin_cites_number":113,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Links to the datasets"]],"subsections":[],"title":"Links to the datasets"},{"cite_extract_rate":0.505813953488372,"cites":[2829,2871,2844,2838,2808,2807,2869,2841,2822,979,2840,2824,2811,2845,2850,2835,2873,2870,2827,2842,2836,2887,9096,8590,2849,2877,2813,2814,7117,2853,2879,2816,2856,2866,2888,2809,2858,2802,2805,2826,2833,2876,7618,2839,2843,2847,2854,8592,2868,2832,2874,2828,2806,8595,2834,2859,2875,2881,2812,2886,2878,2851,2863,7617,7379,2803,2890,8594,2804,2821,2837,2867,2817,2880,2885,2819,2846,2815,988,2818,2820,2810,2831,2884,2861,2872,2864],"content":"\\label{papers_datasets}\nLists of datasets and corresponding papers can be found in Tables \\ref{video_datasets_table}, \\ref{action_datasets_table}, \\ref{trajectory_datasets_table}, \\ref{motion_datasets_table}, and \\ref{other_datasets_table}. \n\\begin{table*}[]\n\\centering\n\\begin{tabular}{|l|l|}\n\\hline\n\\textbf{Datasets} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{Atari}} & \\\\ \\hline\n\\textit{\\textbf{BAIR Push}} & ,, \n \\\\ \\hline\n\\textit{\\textbf{Bouncing Ball}} & \\\\ \\hline\n\\textit{\\textbf{CHUK Avenue}} & \\\\ \\hline\n\\textit{\\textbf{Caltech Pedestrian}} & , , ,,,,, \\\\ \\hline\n\\textit{\\textbf{Cityscapes}} & , \\\\ \\hline\n\\textit{\\textbf{Human 3.6M}} & ,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{JAAD}} & \\\\ \\hline\n\\textit{\\textbf{JHMDB}} & \\\\ \\hline\n\\textit{\\textbf{KITTI}} & ,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{KTH}} & ,,,,,, \\\\ \\hline\n\\textit{\\textbf{MGIF}} & \\\\ \\hline\n\\textit{\\textbf{MMNIST}} & ,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{MSR}} & \\\\ \\hline\n\\textit{\\textbf{MUG}} & \\\\ \\hline\n\\textit{\\textbf{Own}} & , \\\\ \\hline\n\\textit{\\textbf{PROST}} & \\\\ \\hline\n\\textit{\\textbf{Penn Action}} & ,, \\\\ \\hline\n\\textit{\\textbf{Penn action}} & ,, \\\\ \\hline\n\\textit{\\textbf{ShanghaiTech Campus}} & \\\\ \\hline\n\\textit{\\textbf{ShapeStack}} & \\\\ \\hline\n\\textit{\\textbf{Sports-1M}} & , \\\\ \\hline\n\\textit{\\textbf{THUMOS}} & \\\\ \\hline\n\\textit{\\textbf{UCF-101}} & ,,,,,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{UvA-NEMO}} & \\\\ \\hline\n\\textit{\\textbf{ViSOR}} & \\\\ \\hline\n\\textit{\\textbf{YUV}} & , \\\\ \\hline\n\\textit{\\textbf{Youtube-8M}} & \\\\ \\hline\n\\textit{\\textbf{pedestrian}} & \\\\ \\hline\n\\end{tabular}\n\\caption{Datasets used in \\textbf{video prediction} applications.}\n\\label{video_datasets_table}\n\\end{table*}\n\\begin{table*}[]\n\\centering\n\\begin{tabular}{|l|l|}\n\\hline\n\\textbf{Datasets} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{50Salad}} & ,, \\\\ \\hline\n\\textit{\\textbf{ANTICIPATE}} & \\\\ \\hline\n\\textit{\\textbf{AVA}} & , \\\\ \\hline\n\\textit{\\textbf{ActEV/VIRAT}} & \\\\ \\hline\n\\textit{\\textbf{BIT}} & ,,, \\\\ \\hline\n\\textit{\\textbf{BU Action}} & \\\\ \\hline\n\\textit{\\textbf{Brain4Cars}} & ,, \\\\ \\hline\n\\textit{\\textbf{Breakfast}} & ,, \\\\ \\hline\n\\textit{\\textbf{CA}} & \\\\ \\hline\n\\textit{\\textbf{CAD-120}} & ,,,, \\\\ \\hline\n\\textit{\\textbf{CMU Panoptic}} & \\\\ \\hline\n\\textit{\\textbf{CMU Mocap}} & \\\\ \\hline\n\\textit{\\textbf{Caltech Pedestrian}} & \\\\ \\hline\n\\textit{\\textbf{DAD}} & , \\\\ \\hline\n\\textit{\\textbf{Daimler}} & \\\\ \\hline\n\\textit{\\textbf{Daimler Path}} & \\\\ \\hline\n\\textit{\\textbf{EGTEA Gaze+}} & \\\\ \\hline\n\\textit{\\textbf{ETH Pedestrian}} & \\\\ \\hline\n\\textit{\\textbf{Epic-fail}} & \\\\ \\hline\n\\textit{\\textbf{Epic-Kitchen}} & ,, \\\\ \\hline\n\\textit{\\textbf{FPPA}} & \\\\ \\hline\n\\textit{\\textbf{GTEA Gaze}} & \\\\ \\hline\n\\textit{\\textbf{GTEA Gaze+}} & \\\\ \\hline\n\\textit{\\textbf{HMDB}} & \\\\ \\hline\n\\textit{\\textbf{Human 3.6M}} & \\\\ \\hline\n\\textit{\\textbf{JAAD}} & ,,, \\\\ \\hline\n\\textit{\\textbf{JHMDB}} & ,,,,, \\\\ \\hline\n\\textit{\\textbf{Luggage}} & \\\\ \\hline\n\\textit{\\textbf{MANIAC}} & \\\\ \\hline\n\\textit{\\textbf{MPII Cooking}} & ,, \\\\ \\hline\n\\textit{\\textbf{MSRDA}} & \\\\ \\hline\n\\textit{\\textbf{NGSIM}} & ,, \\\\ \\hline\n\\textit{\\textbf{NTU RGB-D}} & \\\\ \\hline\n\\textit{\\textbf{OA}} & \\\\ \\hline\n\\textit{\\textbf{OAD}} & \\\\ \\hline\n\\textit{\\textbf{ORGBD}} & \\\\ \\hline\n\\textit{\\textbf{PIE}} & \\\\ \\hline\n\\textit{\\textbf{PKU-MMD}} & \\\\ \\hline\n\\textit{\\textbf{Recipe1M}} & \\\\ \\hline\n\\textit{\\textbf{SBUIK}} & \\\\ \\hline\n\\textit{\\textbf{SYSU 3DHOI}} & , \\\\ \\hline\n\\textit{\\textbf{Stanford-40}} & \\\\ \\hline\n\\textit{\\textbf{THUMOS}} & ,, \\\\ \\hline\n\\textit{\\textbf{TV Human Interaction}} & , ,,, \\\\ \\hline\n\\textit{\\textbf{TV Series}} & \\\\ \\hline\n\\textit{\\textbf{UCF-101}} & ,,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{UTI}} & ,,,,,, \\\\ \\hline\n\\textit{\\textbf{UTKA}} & \\\\ \\hline\n\\textit{\\textbf{VIENA}} & \\\\ \\hline\n\\textit{\\textbf{VIRAT}} & \\\\ \\hline\n\\textit{\\textbf{WIDER}} & \\\\ \\hline\n\\textit{\\textbf{Willow Action}} & \\\\ \\hline\n\\textit{\\textbf{WnP}} & \\\\ \\hline\n\\textit{\\textbf{YouCook2}} & \\\\ \\hline\n\\textit{\\textbf{Sports-1M}} & \\\\ \\hline\n\\end{tabular}\n\\caption{Datasets used in \\textbf{action prediction} applications.}\n\\label{action_datasets_table}\n\\end{table*}\n\\begin{table*}[]\n\\centering\n\\begin{tabular}{|l|p{0.8\\textwidth}|}\n\\hline\n\\textbf{Datasets} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{ARGOVerse}} & \\\\ \\hline\n\\textit{\\textbf{ATC}} & \\\\ \\hline\n\\textit{\\textbf{ActEV/VIRAT}} & \\\\ \\hline\n\\textit{\\textbf{CA}} & \\\\ \\hline\n\\textit{\\textbf{CARLA}} & , \\\\ \\hline\n\\textit{\\textbf{CHUK}} & \\\\ \\hline\n\\textit{\\textbf{CityPerson}} & \\\\ \\hline\n\\textit{\\textbf{Cityscapes}} & \\\\ \\hline\n\\textit{\\textbf{Daimler Path}} & \\\\ \\hline\n\\textit{\\textbf{ETH}} & ,,,,,,,,,,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{Edinburgh (IFP)}} & , \\\\ \\hline\n\\textit{\\textbf{FM}} & \\\\ \\hline\n\\textit{\\textbf{GC}} & ,,, \\\\ \\hline\n\\textit{\\textbf{INTEARCTION}} & \\\\ \\hline\n\\textit{\\textbf{JAAD}} & \\\\ \\hline\n\\textit{\\textbf{KITTI}} & ,, \\\\ \\hline\n\\textit{\\textbf{L-CAS}} & \\\\ \\hline\n\\textit{\\textbf{Lankershim Boulevard}} & \\\\ \\hline\n\\textit{\\textbf{MITT}} & \\\\ \\hline\n\\textit{\\textbf{MOT}} & \\\\ \\hline\n\\textit{\\textbf{NGSIM}} & ,,,,,, \\\\ \\hline\n\\textit{\\textbf{OSU}} & \\\\ \\hline\n\\textit{\\textbf{Oxford}} & \\\\ \\hline\n\\textit{\\textbf{PETS2009}} & \\\\ \\hline\n\\textit{\\textbf{PIE}} & \\\\ \\hline\n\\textit{\\textbf{QMUL}} & \\\\ \\hline\n\\textit{\\textbf{SBUIK}} & \\\\ \\hline\n\\textit{\\textbf{SD}} & ,,,,,, \\\\ \\hline\n\\textit{\\textbf{STRANDS}} & \\\\ \\hline\n\\textit{\\textbf{TRAF}} & \\\\ \\hline\n\\textit{\\textbf{TUM Kitchen}} & \\\\ \\hline\n\\textit{\\textbf{Town Center}} & ,, \\\\ \\hline\n\\textit{\\textbf{UCY}} & ,,,,,,,,,,,,,,,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{VIRAT}} & \\\\ \\hline\n\\textit{\\textbf{VPM}} & \\\\ \\hline\n\\textit{\\textbf{nuScenes}} & \\\\ \\hline\n\\end{tabular}\n\\caption{Datasets used in \\textbf{trajectory prediction} applications.}\n\\label{trajectory_datasets_table}\n\\end{table*}\n\\begin{table*}[]\n\\centering\n\\begin{tabular}{|l|l|}\n\\hline\n\\textbf{Datasets} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{3DPW}} & \\\\ \\hline\n\\textit{\\textbf{CA}} & \\\\ \\hline\n\\textit{\\textbf{CMU Mocap}} & \\\\ \\hline\n\\textit{\\textbf{CMU Panoptic}} & \\\\ \\hline\n\\textit{\\textbf{Egopose}} & \\\\ \\hline\n\\textit{\\textbf{Human 3.6M}} & ,,,,,,,,,,,,,, \\\\ \\hline\n\\textit{\\textbf{InstaVariety}} & \\\\ \\hline\n\\textit{\\textbf{MPII Human Pose}} & \\\\ \\hline\n\\textit{\\textbf{Mouse Fish}} & \\\\ \\hline\n\\textit{\\textbf{Own}} & ,,,, \\\\ \\hline\n\\textit{\\textbf{Penn Action}} & ,,, \\\\ \\hline\n\\textit{\\textbf{SBUIK}} & \\\\ \\hline\n\\textit{\\textbf{UCF-101}} & \\\\ \\hline\n\\end{tabular}\n\\caption{Datasets used in \\textbf{motion prediction} applications.}\n\\label{motion_datasets_table}\n\\end{table*}\n\\begin{table*}[]\n\\centering\n\\begin{tabular}{|l|l|}\n\\hline\n\\textbf{Datasets} & \\textbf{Papers} \\\\ \\hline\n\\textit{\\textbf{AMOS}} & \\\\ \\hline\n\\textit{\\textbf{Amazon}} & \\\\ \\hline\n\\textit{\\textbf{Cityscapes}} & ,,, \\\\ \\hline\n\\textit{\\textbf{DIPLECS}} & \\\\ \\hline\n\\textit{\\textbf{FCVL}} & \\\\ \\hline\n\\textit{\\textbf{Golden Colorado}} & \\\\ \\hline\n\\textit{\\textbf{JAAD}} & \\\\ \\hline\n\\textit{\\textbf{KITTI}} & , \\\\ \\hline\n\\textit{\\textbf{MBI-1M}} & \\\\ \\hline\n\\textit{\\textbf{MU}} & \\\\ \\hline\n\\textit{\\textbf{SUN RGB-D}} & \\\\ \\hline\n\\textit{\\textbf{Tuscan Arizona}} & \\\\ \\hline\n\\textit{\\textbf{VIST}} & \\\\ \\hline\n\\end{tabular}\n\\caption{Datasets used in \\textbf{other prediction} applications.}\n\\label{other_datasets_table}\n\\end{table*}\n\\end{document}","id":"a33fb5e9-81a5-4a37-8f37-6dd6031d77ae","level":"section","origin_cites_number":172,"parent_id":"54da09bf-16a7-43ca-8a25-4aafacb8f53e","prefix_titles":[["title","Deep Learning for Vision-based Prediction: A Survey"],["section","Datasets and corresponding papers"]],"subsections":[],"title":"Datasets and corresponding papers"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ]\n ],\n \"subsections\": [\n \"2fec0d8e-0911-43a3-a218-2d210969c7e7\",\n \"cbf20d18-4b97-4aff-b65c-6034395a9fa5\",\n \"80143ee1-3a53-4ec6-96bf-ec81cc495f85\",\n \"040ae750-d27c-4e06-99d2-48f8312f00ca\",\n \"991d18f2-ffd8-45b8-b2e8-b4c21fb20720\",\n \"ea670b52-fc46-4dcb-ac08-efd4b23b03c3\",\n \"77155bc6-d46c-48ec-a4ba-c2d738e13396\",\n \"374ffbdc-d4ff-461d-a063-67f03516e35e\",\n \"5bb03330-9009-4e51-abdd-12e3b8d3505d\",\n \"4c69ee31-a62c-468f-a731-a0d63c6e9fe2\",\n \"6ab7b3e0-1f9f-4a59-bcf5-66c554040f7e\",\n \"2dbff660-6f4e-44d8-90a2-fa6081f28574\",\n \"23d750d7-3a7a-460d-8e6b-4f54e379529f\",\n \"63a0cc61-98a6-4be7-b982-ebf42e368ec1\",\n \"14c8fcc2-4517-4a60-9512-dfff6198b563\",\n \"e82c2ffc-f5dc-4d01-9184-399825652bbe\",\n \"a33fb5e9-81a5-4a37-8f37-6dd6031d77ae\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:introduction}}\\n\\\\IEEEPARstart{T}{he} ability to predict the changes in the environment and the behavior of objects is fundamental in many applications such as surveillance, autonomous driving, scene understanding, etc. Prediction is a widely studied field in various artificial intelligence communities. A subset of these algorithms relies primarily on visual appearances of the objects and the scene to reason about the future. Other approaches use different forms of sensors such as wearable or environmental sensors to learn about the past states of the environment or objects. \\nThe focus of this report is on vision-based prediction algorithms, which primarily use visual information to observe the changes in the environment and predict the future. In this context, prediction can be in the form of generating future scenes or reasoning about specific aspects of the objects, e.g. their trajectories, poses, etc.\\nFor this review, we divide the prediction algorithms into five groups, namely video prediction, action prediction, trajectory prediction, motion (pose) prediction, and others which involve various applications of prediction such as trend prediction, visual weather prediction, map prediction, semantic prediction, etc. In addition, we briefly discuss algorithms that use a form of prediction as an intermediate step to perform tasks such as object detection, action detection, and recognition, etc. Moreover, for each group of prediction algorithms, we will talk about the common datasets and metrics and discuss of their characteristics. It should be noted that due to the broad scope of this review and the large body of work on the vision-based prediction, this review will only focus on works that had been published since five years ago in major computer vision, robotics and machine learning venues. In addition, as the title of the paper suggests, the main focus of the discussion will be on deep learning methods given their popularity in recent years.\",\n \"id\": \"2fec0d8e-0911-43a3-a218-2d210969c7e7\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Before reviewing the works on vision-based prediction algorithms, there are a number of points that should be considered.\",\n \"id\": \"cbf20d18-4b97-4aff-b65c-6034395a9fa5\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Vision-based Prediction\"\n ]\n ],\n \"subsections\": [\n \"ffddf8b2-9c16-4742-b890-9fae3c51a950\",\n \"d51e120c-5adf-414f-9f19-52ce0c4182c7\"\n ],\n \"title\": \"Vision-based Prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Based on our review, we have identified four major vision-based applications namely, video prediction, action prediction, trajectory prediction, and motion prediction. We discuss each of the studies in each category in a dedicated section. Some of the prediction works, such as visual weather prediction, semantic prediction, contests outcome prediction, that do not fit to any of the four major categories are presented in other application section. \\nSome works address multiple prediction tasks, e.g. predicting trajectories and actions simultaneously, and therefore might fall in more than one category. It should be noted that we only include an algorithm in each category if the corresponding task is directly evaluated. For instance, if an algorithm performs video prediction for future action classification, and only evaluates the accuracy of predicted actions, it will only appear in the action prediction category. Furthermore, some works that are reviewed in this paper propose multiple architectures, e.g. recurrent and feedforward, for solving the same problem. In architecture-based categorizations, these algorithms may appear more than once.\",\n \"id\": \"ffddf8b2-9c16-4742-b890-9fae3c51a950\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"cbf20d18-4b97-4aff-b65c-6034395a9fa5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Vision-based Prediction\"\n ],\n [\n \"subsection\",\n \"Applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Applications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"d51e120c-5adf-414f-9f19-52ce0c4182c7\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"cbf20d18-4b97-4aff-b65c-6034395a9fa5\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Vision-based Prediction\"\n ],\n [\n \"subsection\",\n \"Methods\"\n ]\n ],\n \"subsections\": [\n \"f46c6d90-8842-4e11-a89d-e5ce565f4be9\",\n \"56ce5010-908f-43e2-ac25-ca9c33b72d32\",\n \"fa1f4460-f37c-4c55-8b41-c1ea861edc31\"\n ],\n \"title\": \"Methods\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"This work focuses on vision-based algorithms, which use some form of visual input such as RGB camera images or active sensors such as LIDAR. It should be noted that many algorithms, especially trajectory prediction ones, only use ground truth data such as object trajectories without actual visual processing, e.g. for detection of objects. However, as long as these algorithms are evaluated on vision datasets, they are included in this paper. Note that a completer list of papers with published code can be found in Appendix \\\\ref{papers_with_code}.\",\n \"id\": \"f46c6d90-8842-4e11-a89d-e5ce565f4be9\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"d51e120c-5adf-414f-9f19-52ce0c4182c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Vision-based Prediction\"\n ],\n [\n \"subsection\",\n \"Methods\"\n ],\n [\n \"subsubsection\",\n \"Algorithms\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Algorithms\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As mentioned earlier, we focus on algorithms that have a deep learning component, either in the stage of visual representation generation (e.g. using convolutional features) or reasoning (e.g. using an MultiLayer Preceptron (MLP) for classification). We will, however, acknowledge the classical methods by mentioning some of the main techniques and including them in the datasets and metrics sections of this paper.\\nWe classify the algorithms in terms of training techniques and architectures. In practice, this is very challenging as the majority of algorithms use a combination of different approaches. For example, recurrent networks often rely on a form of Convolutional Neural Networks (CNNs) to generate feature representations for scenes, poses of agents, etc. To better distinguish between different classes of algorithms, we only focus on the core component of each algorithm, i.e. the parts that are used for reasoning about the future. Hence, for example, if an algorithm uses a CNN model for pre-processing input data and a recurrent network for temporal reasoning, we consider this algorithm as recurrent. On the other hand, if the features are used with a fully connected network, we categorize this algorithm as feedforward or one-shot method. A few algorithms propose the use of both architectures for reasoning. We address those methods as hybrid. In addition, it should be noted that many works propose alternative approaches using each architecture. Therefore, we categorize them in more than one group.\",\n \"id\": \"56ce5010-908f-43e2-ac25-ca9c33b72d32\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"d51e120c-5adf-414f-9f19-52ce0c4182c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Vision-based Prediction\"\n ],\n [\n \"subsection\",\n \"Methods\"\n ],\n [\n \"subsubsection\",\n \"Architectures\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Architectures\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"As one would expect, vision-based algorithms primarily rely on visual information. However, many algorithms use pre-trained off-the-shelf algorithms to transform the input to some explicit feature spaces, e.g. poses, trajectories, action labels and perform reasoning in those feature spaces. If pre-processing is not part of the main algorithm, we consider those secondary features as different types of data inputs to the algorithms. If some basic processing, e.g. generating convolutional features for a scene is used, we consider the data type of the original input, e.g. RGB images.\",\n \"id\": \"fa1f4460-f37c-4c55-8b41-c1ea861edc31\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"d51e120c-5adf-414f-9f19-52ce0c4182c7\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Vision-based Prediction\"\n ],\n [\n \"subsection\",\n \"Methods\"\n ],\n [\n \"subsubsection\",\n \"Data type\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data type\"\n },\n {\n \"cite_extract_rate\": 0.5897435897435891,\n \"cites\": [\n 2804,\n 2814,\n 2817,\n 2807,\n 2808,\n 2816,\n 5680,\n 2819,\n 2815,\n 2806,\n 2809,\n 2818,\n 2811,\n 2802,\n 2820,\n 2810,\n 2805,\n 2812,\n 8590,\n 7617,\n 2803,\n 2813,\n 2073\n ],\n \"content\": \"Video or future scene prediction can be regarded as the most generic form of prediction. The objective of video prediction algorithms is to generate future scenes, often in the form of RGB images and/or optical flow maps . The generated images in turn can be used for various tasks such as action prediction , event prediction , flow estimation , semantic segmentation , etc.\\nVideo prediction applications rely on generative models whose task is to predict future scene(s) based on a short observation of input sequences (or in some cases only a single image ). Although many approaches use feedforard architectures , the majority of algorithms take advantage of Recurrent Neural Networks (RNNs) such as Gated Recurrent Units (GRUs) , Long-Short Term Memory (LSTM) networks , its variation Convolutional LSTMs (ConvLSTMs) or a combination of these .\\nGenerative Adversarial Networks (GANs) are particularly popular in the video prediction community. . In these adversarial training frameworks, there are two compoents: A generative network that produces future representations and a discriminator whose objective is to distinguish between the predicted representations (e.g. optical flow , frames , motion ) or their temporal consistency and the actual ground truth data by producing a binary classification score that indicates whether the prediction is real or fake. While many algorithms use discriminators to judge how realistic the final generated images are or intermediate features (e.g. poses ), others use multiple discriminators at different stages of processing. For example, the authors of use two discriminators, one is responsible for judging the temporal consistency of the generated frames (i.e. whether the order of generated frames is real) and the other assesses whether the generated frames are real or not. Lee et al. use three discriminators to assess the quality of generated frames and the intermediate motion and content features. Using a two-stream approach, the method in produces both the next frame and optical flow and each stream is trained with a separate discriminator. The prediction network of uses a discriminator for intermediate features generated from input scenes and another discriminator for the final results.\\nVariational Autoencoders (VAEs) or Conditional VAEs (CVAEs) are also used in some approaches . VAEs model uncertainty in generated future frames by defining a posterior distribution over some latent variable space . In CVAEs, the posterior is conditioned on an additional parameter such as the observed action in the scenes or initial observation . Using VAEs, at inference time, a random sample is drawn from the posterior to generate the future frame.\\nMany video prediction algorithms operate solely on input images and propose various architectural innovations for encoding the content and generating future images . For example, the method in performs a two-way prediction, forward and backward. Each prediction relies on two discriminators for assessing the quality of the generated images and temporal consistency. The model presented in trains a context network by inputting an image sequence into a ConvLSTM whose output is used to initialize convolutional networks responsible for generating the next frames. Xu et al. , in addition to raw pixel values, encode the output of a high pass filter applied to the image as a means of maintaining the structural integrity of the objects in the scene. In , the authors use a two-step approach in which they first perform a coarse frame prediction followed by a fine frame prediction. In , the algorithm learns in two stages. A discriminator is applied after features are generated from the scenes and another one after the final generated frames. \\nOptical flow prediction has been widely used as an intermediate step in video prediction algorithms . For example, to deal with occlusion in dynamic scenes, Gao et al. disentangle flow and pixel-level predictions into two steps: the algorithm first predicts the flow of the scene, and then uses it, in conjunction with the input frames, to predict the future. Similar multi-step approaches have also been used in . In , the authors use two separate branches: one branch receives two consecutive frames $(t, t+1)$ and produces context information. The second branch produces motion information by receiving two frames that are $k$ steps apart (i.e. $t+1$, $t+k$). The outputs of these two branches are fused and fed into the final scene generator. The method in simultaneously produces the next future frame and the corresponding optical flow map. In this architecture, two additional networks are used: A flow estimator which uses the output of the frame generator and the last observation to estimate a flow map and a warping layer which performs differential 2D spatial transformation to warp the last observed image into the future predicted frame according to the predicted flow map. \\nSome algorithms rely on various intermediate steps for video prediction. For instance, the method in , reasons about the locations and features of individual entities (e.g. cubes) for final scene predictions. Kim et al. first identify keypoints, which may correspond to important structures such as joints, and then predict their motion. For videos involving humans, in the authors identify and reason about the changes in poses, and use this information to generate future frames. In , in addition to raw input images, the differences between consecutive frames used in the learning process.\\nPrediction networks can also be provided with additional information to guide future frame generation. In an optical flow network and in a pose estimation network are used in addition to RGB images. Using a CVAE architecture, in the authors use the action lables as conditional input for frame generation. In the context of active tasks, e.g. object manipulation with robotic arms, in which the consequences of actions influence the future scene configuration, it is common to condition the future scene generation on the current or future intended actions .\",\n \"id\": \"80143ee1-3a53-4ec6-96bf-ec81cc495f85\",\n \"level\": \"section\",\n \"origin_cites_number\": 39,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Video prediction\"\n ]\n ],\n \"subsections\": [\n \"f163c039-4fc8-43c5-8ff4-fc6e32ae5edf\"\n ],\n \"title\": \"Video prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Video prediction algorithms are based on generative models that produce future images given a short observation, or in extreme cases a single view of the scene. Both recurrent and feedforward models are widely used in the field, with recurrent ones being slightly more favorable. The architectural designs and training strategies such as the VAEs or GANs are very common. However, it is hard to establish which one of these approaches is superior given that the majority of the video prediction algorithms are application-agnostic, meaning that they are evaluated on a wide range of video datasets with very different characteristics such as traffic scenes, activities, games, object manipulations, etc. (more on this in Section \\\\ref{datasets}).\\nDespite the great progress in the field, video prediction algorithms are still facing some major challenges. One of them is the ability to hallucinate, which is to generate visual representations for parts of the scenes that were not visible during observation phase, e.g. due to occlusions. This is particularly an issue for more complex images such as traffic scenes, movies, etc. The complexity of the scenes also determines how fast the generated images would degrade. Although these algorithms show promising results in simple synthetic videos or action sequences, they still struggle in real practical applications. In addition, many of these algorithms cannot reason about the expected presence or absence of objects in the future. For example, if a moving object is present in the observations and is about to exit the field of view in near future, the algorithms account for it in the future scenes as long as parts of it are visible in the observation stage. This can be an issue for safety-critical applications such as autonomous driving in which the presence or absence of traffic elements and the interactions between them are essential for action planning.\",\n \"id\": \"f163c039-4fc8-43c5-8ff4-fc6e32ae5edf\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"80143ee1-3a53-4ec6-96bf-ec81cc495f85\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Video prediction\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0.17391304347826,\n \"cites\": [\n 2821,\n 2823,\n 2822,\n 2824\n ],\n \"content\": \"Action prediction algorithms can be categorized into two groups: Next action or event prediction (or action anticipation) and early action prediction. In the former category, the algorithms use the observation of current activities or scene configurations and predict what will happen next. Early action prediction algorithms, on the other hand, observe parts of the current action in progress and predict what this action is. The classical learning approaches such as Conditional Random Fields (CRFs) , Support Vector Machines (SVMs) with hand-crafted features , Markov models , Bayesian networks and other statistical methods have been widely used in recent years. However, as mentioned earlier, we will only focus on deep learning approaches.\",\n \"id\": \"040ae750-d27c-4e06-99d2-48f8312f00ca\",\n \"level\": \"section\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Action prediction\"\n ]\n ],\n \"subsections\": [\n \"b0a73071-8988-459c-960a-fbc201a48a45\",\n \"2ebd83ed-2292-4ab8-9334-3163c17703ba\"\n ],\n \"title\": \"Action prediction\"\n },\n {\n \"cite_extract_rate\": 0.486486486486486,\n \"cites\": [\n 2829,\n 2837,\n 7117,\n 2838,\n 2832,\n 2828,\n 2825,\n 2806,\n 2834,\n 2835,\n 2827,\n 2826,\n 2831,\n 2836,\n 2833,\n 7618,\n 2839,\n 2830\n ],\n \"content\": \"Action prediction algorithms are used in a wide range of applications including cooking activities , traffic understanding , accident prediction , sports and other forms of activities . Although the majority of these algorithms use sequences in which the objects and agents are fully observable, a number of methods rely on egocentric scenes which are recorded from the point of view of the acting agents and only parts of their bodies (e.g. hands) are observable.\\nAction prediction methods predominantly use a variation of RNN-based architectures including LSTMs , GRUs , ConvLSTMs , and Quasi-RNNs (QRNNs) . For instance, in the authors use a graph-based RNN architecture in which the nodes represent actions and the edges of the graph represent the transitions between the actions. The method in employs a two-step approach: using a recognition algorithm, the observed actions and their durations are recognized. These form a one-hot encoding vector which is fed into GRUs for the prediction of the future activities, their corresponding start time and length. In the context of vehicle behavior prediction, Ding et al. uses a two-stream GRU-based architecture to encode the trajectory of two vehicles and a shared activation unit to encode the vehicles mutual interactions. Scheel et al. encode the relationship between the ego-vehicle and surrounding vehicles in terms of their mutual distances. The vectorized encoding is then fed into a bi-directional LSTM. At each time step, the output of the LSTM is classified, using a softmax activation, into a binary value indicating whether it is safe for the ego-vehicle to change lane. In the authors use a QRNN network to capture the relationships between road users in order to predict the likelihood of a traffic accident. To train the model, the authors propose an adaptive loss function that assigns penalty weights depending on how early the model can predict accidents. \\nAs an alternative to recurrent architectures, some algorithms use feedforward architectures using both 3D and 2D convolutional networks. For example, in the context of pedestrian crossing prediction, in the authors use a generative 3D CNN model that produces future scenes and is followed by a classifier. The method of detects and tracks pedestrians in the scenes, and then feeds the visual representations of the tracks, in the form of an image sequence, into a 3D CNN architecture, which directly classifies how likely the pedestrian will cross the road. To predict the time of traffic accidents, the method in processes each input image using a 2D CNN model and then combines the representations followed by a fully-conntected (\\\\textit{fc}) layer for prediction. Farha et al. create a 2D matrix by stacking one-hot encodings of actions for each segment of observation and use a 2D convolutional net to generate future actions encodings. Casas et al. use a two-stream 2D CNN, each processing the stacked voxelized LIDAR scans and the scene map. The feature maps obtained from each stream are fused and fed into a backbone network followed by three headers responsible for the detection of the vehicles and predicting their intentions and trajectories. For sports forecasting, Felsen et al. concatenate 5 image observations channel-wise and feed the resulting output into a 2D CNN network comprised of 4 convolutional layers and an \\\\textit{fc} layer.\\nAlthough some algorithms rely on a single source of information, e.g. a set of pre-processed features from RGB images or trajectories , many algorithms use a multimodal approach by using various sources of information such as optical flow maps , poses , scene attributes (e.g. road structure, semantics) , text , action labels , length of actions , speed (e.g. ego-vehicle or surrounding agents) , gaze , current activities and the time of the actions. For example, the method in uses a multi-stream LSTM in which two LSTMs encode visual features and cooking recipes and an LSTM decodes them for final predictions. To capture the relationships within and between sequences, Gammulle et al. propose a two-stream LSTM network with external neural memory units. Each stream is responsible for encoding visual features and action labels. In , a multi-layer GRU structure is used in which features with different modalities enter the network at different levels and are fused with the previous level encodings. The fusion process is taking place according to the complexity of the data modality, e.g. more complex features such as encodings of pedestrian appearances enter the network at the bottom layer, whereas location and speed features enter at the second-last and last layers respectively. Farha et al. use a two-layer stacked GRU architecture which receives as input a feature tuple of the length of the activity and its corresponding one-hot vector encoding. In , the method uses a two-stage architecture: First information regarding the appearance of the scene, optical flow (pre-processed using a CNN) and vehicle dynamics are fed into individual LSTM units. Then, the output of these units is combined and passed through an \\\\textit{fc} layer to create a representation of the context. This representation is used by another LSTM network to predict future traffic actions. In the context of human-robot interaction, the authors of combine the information regarding the gaze and pose of the humans using an encoder-decoder LSTM architecture to predict their next actions. Jain et al. use a fusion network to combine head pose information of the driver, outside scene features, GPS information, and vehicle dynamics to predict the driver's next action.\\nBefore concluding this section, it is important to discuss the use of attention modules which have gained popularity in recent years . As the name implies, the objective of attention modules is to determine what has to be given more importance at a given time. These modules can come in different froms and can be applied to different dimensions of data and at various processing. Some of these modules are temporal attention for identifying keyframes, modality attention to prioritize between different modalities of data input, spatial attention for highlighting the important parts of the scenes, and graph attention for weighting nodes of the graph. In some cases, a combination of different attention mechanisms is used .\",\n \"id\": \"b0a73071-8988-459c-960a-fbc201a48a45\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 37,\n \"parent_id\": \"040ae750-d27c-4e06-99d2-48f8312f00ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Action prediction\"\n ],\n [\n \"subsection\",\n \"Action anticipation\"\n ]\n ],\n \"subsections\": [\n \"41132876-21fb-45f4-86a1-c905e9d97c64\"\n ],\n \"title\": \"Action anticipation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In the field of action anticipation, RNN architectures are strongly preferred. Compared to feedforward algorithms, recurrent methods have the flexibility of dealing with variable observation lengths and multi-modal data, in particular, when they are significantly different, e.g. trajectories and RGB images. However, basic recurrent architectures such as LSTMs and GRUs rely on some forms of pre-processing, especially when dealing with high dimensional data such as RGB images, which requires the use of various convolutional networks, a process that can be computationally costly. Feedforward models, on the other hand, can perform prediction in one shot, meaning that they can simultaneously perform temporal reasoning and spatial feature generation in a single framework, and as a result, potentially have a shorter processing time. \\nMany of the approaches mentioned earlier are generative in nature. They generate representations in some feature space and then using these representations predict what will happen next. Some algorithms go one step further and generate the actual future images and use them for prediction. Although such an approach seems effective for single actor events, e.g. cooking scenes, human-robot interaction, it is not a feasible approach for multi-agent predictions such as reasoning about behaviors of pedestrians or cars in traffic scenes.\\nThe majority of the methods reviewed in this section use multi-modal data input. This seems to be a very effective approach, especially in high dimensional problems such as predicting road user behavior where the state (e.g. location and speed) of the road user, observer, and other agents, as well as scene structure, lighting conditions, and many other factors, play a role in predicting the future behavior.\\nMulti-tasking, e.g. predicting actions and trajectories, are an effective way to predict future actions. For instance, trajectories can imply the possibility of certain actions, e.g. moving towards the road implies the possibility that the person might cross the street. As a result, the simultaneous learning of different tasks can be beneficial. \\nLast but not least is the use of attention modules. These modules are deemed to be very effective , in particular for tasks with high complexity in terms of the modality of input data, the scene structure and temporal relations.\",\n \"id\": \"41132876-21fb-45f4-86a1-c905e9d97c64\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"b0a73071-8988-459c-960a-fbc201a48a45\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Action prediction\"\n ],\n [\n \"subsection\",\n \"Action anticipation\"\n ],\n [\n \"subsubsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0.41176470588235203,\n \"cites\": [\n 2829,\n 2841,\n 2845,\n 2844,\n 2840,\n 2843,\n 2842\n ],\n \"content\": \"Similar to action anticipation methods, early action prediction algorithms widely use recurrent network architectures . Although many of these algorithms use similar aproaches to action anticipation algorithms, some propose new approaches. For example, in the authors use a teacher-student learning scheme where the teacher learns to recognize full sequences using a bi-directional LSTM and the student relies on partial videos using an LSTM network. They perform knowledge distillation by linking feature representations of both networks. Using GAN frameworks, the methods in predict future feature representations of videos in order to predict actions. Zhao et al. implement a Kalman filter using an LSTM architecture. In this method, actions are predcted after observing each frame and corrections are made if the next observation provide additional information. The method of uses a two-step LSTM architecture which first generates an encoding of the context using context-aware convolutional features and then combines these encodings with action-aware convolutional features to predict the action. The authors of this method propose a new loss function that penalizes false negatives with the same strength at any point in time and false positives with a strength that increases linearly over time, to reach the same weight as that on false negatives. In the authors perform coarse-to-fine-grained predictions depending on the amount of evidence available for the type of action. \\nMany early action prediction methods adopt feedforward architectures . The authors of predict actions from a single image by transforming it into a new representation called Ranked Saliency Map and Predicted Optical Flow. This representation is passed through a 2D convent and \\\\textit{fc} layers for final prediction. In , the authors use temporal CNN (TCNNs) architectures, which are a series of 1D dilated convolutional layers designed to capture the temporal dependencies of feature representations, e.g. in the form of a vector representation of poses or word-embeddings describing the video frames. Chen et al. use features of body parts generated by a CNN model and train an attention module whose objective is to activate only features that contribute to the prediction of the action. In , the authors use an action detection framework, which incrementally predicts the locations of the actors and the action classes based on the current detections.\\nThe majority of the early action prediction algorithms pre-process the entire observed scenes using, e.g. different forms of convolutional neural networks or other forms of features . Some algorithms complement these features using optical flow maps . Another group of action prediction methods focuses on specific parts of the observations. For example, in the authors use the changes in the poses of actors to predict their actions. The method in uses body parts extracted by cropping a local patch around identified joints of the actors. The authors of only use the visual appearances of actors extracted from detected bounding boxes, and the relationship between them.\",\n \"id\": \"2ebd83ed-2292-4ab8-9334-3163c17703ba\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"040ae750-d27c-4e06-99d2-48f8312f00ca\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Action prediction\"\n ],\n [\n \"subsection\",\n \"Early action prediction\"\n ]\n ],\n \"subsections\": [\n \"2bcad629-77b1-4f2d-ae76-43a654b49c71\"\n ],\n \"title\": \"Early action prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Early action detection methods have many commonalities with action anticipation algorithms in terms of architectural design. However, there are some exceptions that are more applicable in this domain. These exceptions are teacher-student training schemes for knowledge distillation, identifying discriminative features, and recursive prediction/correction mechanisms. In addition, early action prediction algorithm, with a few exceptions, rely on single modal data for prediction and rarely use refinement frameworks such as attention modules. Adopting these techniques and operating on multi-modal feature spaces can further improve the performance of early action prediction algorithms. Unlike the action anticipation methods, there is no strong preference for recurrent or feedforward approaches. Some approaches take advantage of architectures such as temporal CNNs which are popular in the language processing domain and show their effectiveness for early action prediction tasks.\",\n \"id\": \"2bcad629-77b1-4f2d-ae76-43a654b49c71\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"2ebd83ed-2292-4ab8-9334-3163c17703ba\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Action prediction\"\n ],\n [\n \"subsection\",\n \"Early action prediction\"\n ],\n [\n \"subsubsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0.42465753424657504,\n \"cites\": [\n 2867,\n 2855,\n 2847,\n 2860,\n 2854,\n 2853,\n 2868,\n 2857,\n 2869,\n 2856,\n 2846,\n 2862,\n 2852,\n 2866,\n 979,\n 988,\n 2858,\n 2859,\n 2848,\n 2850,\n 2865,\n 2831,\n 2851,\n 8591,\n 1261,\n 2861,\n 2863,\n 2849,\n 7379,\n 2830,\n 2864\n ],\n \"content\": \"As the name implies, trajectory prediction algorithms predict the future trajectories of objects, i.e. the future positions of the objects over time. These approaches are particularly popular for applications such as intelligent driving and surveillance. Predicted trajectories can be used directly, e.g. in route planning for autonomous vehicles, or used for predicting future events, anomalies, or actions. \\nIn this section, we follow the same routine as before and focus on algorithms that have a deep learning component while acknowledging many recent works that have used classical approaches including Gaussian mixture models and processes , Markov decision processes (MDPs) , Markov chains and other techniques .\\nTrajectory prediction applications like many other sequence prediction tasks heavily rely on recurrent architectures such as LSTMs , and GRUs . These methods often use an encoder-decoder architecture in which a network, e.g. an LSTM encodes single- or multi-modal observations of the scenes for some time, and another network generates future trajectories given the encoding of the observations. Depending on the complexity of input data, these algorithms may rely on some form of pre-processing for generating features or embedding mechanisms to minimize the dimensionality of the data.\\nThe feedforward algorithms often use whole views of the scenes (i.e. the environment and moving objects) and encode them using convolutional layers followed by a regression layer to predict trajectories. A few algorithms use hybrid approaches in which both convolutional and recurrent reasoning are also used . \\nDepending on the prediction task, algorithms may rely on single- or multi-modal observations. For example, in the context of visual surveillance where a fixed camera provide a top-down or bird-eye viewing angle, many algorithms only use past trajectories of the agents in either actual 2D frame coordinates or their velocities calculated by the changes from each time step to another . In addition to observations of individual trajectories of agents, these algorithms focus on modeling the interaction between the agents and how they impact each other. For example, Zhang et al. use a state refinement module that aligns all pedestrians in the scene with a message passing mechanism that receives as input the current locations of the subjects and their encodings from an LSTM unit. In a graph-based approach is used where pedestrians are considered as nodes and the interactions between them as edges of the graph. By aggregating information from neighboring nodes, the network learns to assign a different level of importance to each node for a given subject. The authors of perform a pooling operation on the generated representations by sharing the state of individual LSTMs that have spatial proximity. \\nAs shown in some works, other sources of information are used in surveilling objects . For example, in addition to encoding the interactions with the environment, Liang et al. use the semantic information of the scene as well as changes in the poses of the pedestrians. In the visual representations of the layout of the environment and the appearances of the subjects are included. The authors of use an occupancy map which highlights the potential traversable locations for the subjects. The method in takes into account pedestrians' head orientations to estimate their fields of view in order to predict which subjects would potentially interact with one another. To predict interactions between humans, in the authors use both poses and trajectories of the agents. Ma et al. go one step further and take into account the pedestrians' characteristics (e.g. age, gender) within a game-theoretic perspective to determine how the trajectory of one pedestrians impact each other.\\nIn the context of traffic understanding, predicting trajectories can be more challenging due to the fact that there is camera ego-motion involved (e.g. the prediction is from the perspective of a moving vehicle), there are interactions between different types of objects (e.g. vehicles and pedestrians), and there are certain constraints involved such as traffic rules, signals, etc. To achieve robustness, many methods in this domain take advantage of multi-modal data for trajectory prediction . In addition to using past trajectories, all these algorithms account for the road structure (whether it is from the perspective of the ego-vehicle or a top-down view) often in the form of raw visual inputs or, in some cases, as an occupancy map . The scene layout can implicitly capture the structure of the road, the appearances of the objects (e.g. shape) and the dynamics (e.g. velocity or locations of subjects). Such implicit information can be further augmented by explicit data such as the shapes of the objects (in the case of vehicles) , the speed and steering angle of the ego-vehicle, the distance between the objects , traffic rules and signals , and kinematic constraints . For example, the method in uses a two-stream LSTM encoder-decoder scheme: first stream encodes the current ego-vehicle's odometry (steering angle and speed) and the last observation of the scene and predicts future odometry of the vehicle. The second stream is a trajectory stream that jointly encodes location information of pedestrians and the ego-vehicle's odometry and then combines the encoding with the prediction of the odometry stream to predict the future trajectories of the pedestrians. The method in further extends this approach and adds an intention prediction stream which outputs how likely the observed pedestrian intends to cross the street. The intention likelihood is produced using an LSTM network that encodes the dynamics of the pedestrian, their appearances and their surroundings. Chandra et al. create embeddings of contextual information by taking into account the shape and velocity of the road users and their spatial coordinates within a neighboring region. These embeddings are then fed into some LSTM networks followed by a number of convolutional layers to capture the dynamics of the scenes. In the authors use separate LSTMs for encoding the trajectories of pedestrians and vehicles (as orientated bounding boxes) and then combine them into a unified framework by generating an occupancy map of the scene centered at each agent, followed by a pooling operation to capture the interactions between different subjects. Lee et al. predict the future trajectories of vehicles in two steps: First, an encoder-decoder GRU architecture predicts future trajectories by observing the past ones. Then a refinement network adjusts the predicted trajectories by taking into account the contextual information in the forms of social interactions, dynamics of the agents involved, and the road structure.\\nSimilar to action prediction algorithms, attention modules have been widely used in trajectory prediction methods . For example, in , the attention module jointly measures spatial and temporal interactions. The authors of propose the use of social attention modules which estimate the relative importance of interactions between the subject of interest and its neighboring subjects. The method in uses two attention mechanisms, a temporal attention module that measures the importance of each timestep of the observed trajectories and a series of self-attention modules which are applied to encodings of observations prior to predictions. Xue et al. propose an attention mechanism to measure the relative importance between different data modalities, namely the locations and velocities of subjects.\\nOne of the major challenges in trajectory prediction is to model uncertainty of predictions, especially in scenarios where many possibilities exist (e.g. pedestrians at intersections). Some algorithms such as train the models by directly measuring the error between the ground truth trajectories and predicted ones, e.g. by using an \\\\textit{L2} objective function. At the inference time, these algorithms generate deterministic predictions. To measure the uncertainty of predictions, these models are trained multiple times with randomly initialized parameters. Alternatively, uncertainty can be estimated via probabilistic objective functions, e.g. Gaussian log-likelihood, as in Instead of a single point in space, these algorithms predict a distribution that captures the uncertainty of the predictions. VAEs are another technique that can be used to estimate uncertainty . Using these methods, at training time a posterior distribution over some latent space $z$ is learned by conditioning, for example, over future trajectories. At the inference time, a random sample is drawn from the latent space for the predictions.\\nSince trajectory prediction algorithms are generative by nature, many approaches rely on adversarial training techniques in which at training time a discriminator is used to predict whether the generated trajectories are real or fake. Kosaraju et al. extend this approach by using two discriminators: A local discriminator which predicts the results on the output of the prediction using only past trajectories, and one global discriminator which operates on the output of the entire network, i.e. the prediction results based on trajectories and scene information.\",\n \"id\": \"991d18f2-ffd8-45b8-b2e8-b4c21fb20720\",\n \"level\": \"section\",\n \"origin_cites_number\": 73,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Trajectory prediction\"\n ]\n ],\n \"subsections\": [\n \"66a51a9f-6cef-4c6e-8e6b-9a97e4e8a052\"\n ],\n \"title\": \"Trajectory prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Trajectory prediction is a widely studied field in the computer vision community. Although these works dominantly use recurrent network architectures, many approaches, such as those used in the field of traffic scene understanding, use feedforward networks. Trajectory prediction algorithms rely on one or more sources of information such as the past trajectories of subjects, surrounding visual context, object attributes, vehicle sensor readings, etc. One factor that is common in many of trajectory prediction algorithms is modeling the interactions between dynamic or dynamic and static objects. Relationships are captured explicitly or implicitly via encoding the scenes as a whole. Like many other prediction approaches, trajectory prediction algorithms benefit from various forms of attention mechanisms to learn the importance of spatial, temporal or social interactions between objects. To model uncertainty, techniques such as probabilistic objectives and variational encodings are used. \\nTrajectory prediction algorithms are predominantly rely on past trajectory information to predict the future. Although past motion observations are very informative, in some context, e.g. traffic scenes, they are simply not enough. There is a need for a more explicit encoding of contextual information such as road conditions, the subject's attributes, rules and constraints, scene structure, etc. A number of approaches successfully have included a subset of these factors, but a more comprehensive approach should be considered.\",\n \"id\": \"66a51a9f-6cef-4c6e-8e6b-9a97e4e8a052\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"991d18f2-ffd8-45b8-b2e8-b4c21fb20720\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Trajectory prediction\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0.565217391304347,\n \"cites\": [\n 2814,\n 2821,\n 2871,\n 8592,\n 2874,\n 2841,\n 2875,\n 2873,\n 2870,\n 2878,\n 2876,\n 2877,\n 2872\n ],\n \"content\": \"Although the term \\\"motion prediction\\\" in many cases is used to refer to future trajectory prediction, here we only consider the algorithms that are designed to predict changes in human pose. Motion prediction play a fundamental role in all prediction approaches as an intermediate step, e.g. to reflect how the future visual representations would look like or the types of actions to anticipate. Like many other prediction applications, this field is dominated by deep learning models, even though some methods still rely on classical techniques .\\nSimilar to other prediction algorithms, motion prediction methods widely use recurrent architectures such as LSTMs and GRUs or a combination of both . For example, in the authors use a two-layer GRU model in which the top layer operates backward to learn noise processes and the bottom level is used to predict the poses given the past pose observations and the output of the top layer. Chiu et al. propose a hierarchical LSTM architecture in which each layer of the network encodes the observed poses at different time-scales. In the context of 3D pose prediction, some algorithms rely on a two-stage process where the visual inputs, either as a single image or a sequence of images , are fed into a recurrent network to predict 2D poses of the agent. This is followed by a refinement procedure that transforms the 2D poses into 3D. \\nSome approaches adopt feedforward architectures . For example, the method in uses two feedforward networks in a two-stage process. First, the input poses are fed into an autoencoder which is comprised of fully connected layerss (implemented by 1D convolutions with a kernel size of 1) and self-attention blocks. The encodings are then used by multi-level 2D convolutional blocks for final predictions. Zhang et al. predict 3D poses from RGB videos. In their method, the images are converted to a feature space using a convolutional network, and then the features are used to learn a latent 3D representation of 3D human dynamics. The representation is used by a network to predict future 3D poses. To capture movement patterns, a method proposed in converts poses into a trajectory space using discrete cosine transformation. The newly formed representations are then used in a Graph-CNN framework to learn the dependencies between different joint trajectories. \\nTo train motion prediction models, some authors use adversarial training methods in which a discriminator is used to classify whether the predicted poses are real or fake . The discrimination procedure can also be applied to evaluating the continuity, i.e. correct order of, predictions as demonstrated in . In the discrimination score is used to generate a policy for future action predictions in the context of imitation learning.\",\n \"id\": \"ea670b52-fc46-4dcb-ac08-efd4b23b03c3\",\n \"level\": \"section\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Motion prediction\"\n ]\n ],\n \"subsections\": [\n \"304422d0-f0b0-4723-862a-ed7a4565bca2\"\n ],\n \"title\": \"Motion prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Motion prediction algorithms primarily focus on the prediction of changes in the dynamics (i.e. poses) of observed agents. Such predictions can be fundamental to many other applications such as video or trajectory prediction tasks some of which were discussed previously. \\nIn recent works, recurrent network architectures are strongly preferred. The architecture of the choice often depends on the representation of the input data, e.g. whether joint coordinates are directly used or are encoded into a high-dimensional representation. \\nDespite the development of many successful motion prediction algorithms, the majority of these methods rely on a single source of information, for example, poses or scenes. Encoding higher-level contextual information, such as scene semantics, interactions, etc. can potentially result in more robust predictions, as shown in other prediction applications. Attention modules also, except for one instance, haven't been adopted within motion prediction algorithms. Given the success of using attention in other prediction applications, motion prediction algorithms may benefit from the use of attention mechanisms.\",\n \"id\": \"304422d0-f0b0-4723-862a-ed7a4565bca2\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"ea670b52-fc46-4dcb-ac08-efd4b23b03c3\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Motion prediction\"\n ],\n [\n \"subsection\",\n \"Summary\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Summary\"\n },\n {\n \"cite_extract_rate\": 0.565217391304347,\n \"cites\": [\n 2882,\n 2879,\n 2880,\n 2885,\n 2888,\n 2806,\n 2883,\n 2881,\n 2887,\n 2886,\n 9096,\n 2884,\n 2803\n ],\n \"content\": \"In the context of autonomous robotics, some algorithms are designed to predict occupancy grid maps (OGMs) that are grayscale representations of the robot's surroundings showing which parts of the environment are traversable. These approaches are often object-agnostic and are concerned with generating future OGMs which are used by an autonomous agent to perform path planning. In recent years both classical and deep learning methods are used. The deep learning approaches, in essence, are similar to video prediction methods in which the model receives as input a sequence of OGMs and predicts the future ones over some period of time. In this context both recurrent and feedforward methods were common. Another group of generative approaches is semantic map prediction algorithms . These algorithms receive as inputs RGB images of the scenes and predict future segmentation maps.\\nSome of the other vision-based prediction applications include weather and Solar irradiance forecasting , steering angle prediction , predicting the popularities of tweets based on tweeted images used and the users' histories , forecasting fashion trends , storyline prediction , pain anticipation , predicting the effect of force after manipulating objects , forecasting the winner of Miss Universe given the appearances of contestants' gowns , and predicting election results given the facial attributes of candidates . These algorithms rely on a combination of techniques discussed earlier in this paper.\",\n \"id\": \"77155bc6-d46c-48ec-a4ba-c2d738e13396\",\n \"level\": \"section\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Other applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Other applications\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 8593,\n 2889,\n 7619,\n 6980\n ],\n \"content\": \"Before concluding our discussion on vision-prediction methods, it is worth mentioning that prediction techniques are also widely used in other visual processing tasks such as video summarization , anomaly detection , tracking , active object recognition , action detection and recognition . For example, tracking algorithms are very closely related to trajectory prediction ones and often rely on short term predictions to deal with gaps, e.g. due to occlusions, in tracking. For example, in the method uses a recurrent framework to generate future frames in order to localize pedestrians in next frames. In the context of action detection, some methods rely on a future frame or trajectory prediction of objects to detect actions . In , a method is used for detecting an object in 3D by relying on predicting next best viewing angle of the object. Liu et al. uses a video prediction framework to predict future motion flow maps and images. The future predictions that do not conform with expectations will be identified as abnormal.\",\n \"id\": \"374ffbdc-d4ff-461d-a063-67f03516e35e\",\n \"level\": \"section\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Prediction in other vision applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Prediction in other vision applications\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"When it comes to the evaluation of algorithms, there are two important factors: metrics and datasets. They highlight the strengths and weaknesses of the algorithms and provide a means to compare the relative performances of the methods. Given the importance of these two factors in the design of algorithms, we dedicate the following sections to discussing the common metrics and datasets used for vision-based prediction tasks. Since the datasets and metrics used in these applications are highly diverse, we will focus our discussion on some of the main ones for each prediction category while providing visual aids to summarize what the past works used for evaluation purposes.\",\n \"id\": \"5bb03330-9009-4e51-abdd-12e3b8d3505d\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"The evaluation of state-of-the-art\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"The evaluation of state-of-the-art\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\begin{figure*}\\n\\\\centering\\n \\\\includegraphics[width=0.21\\\\textwidth]{video_metrics.png}\\n \\\\hspace{0.1cm}\\n \\\\includegraphics[width=0.21\\\\textwidth]{action_metrics.png}\\n \\\\hspace{0.1cm}\\n \\\\includegraphics[width=0.22\\\\textwidth]{trajectory_metrics.png}\\n \\\\hspace{0.1cm}\\n \\\\includegraphics[width=0.18\\\\textwidth]{motion_metrics.png}\\n \\\\caption{Metrics used in vision-based prediction applications. From left to right: Video, action, trajectory and motion prediction. }\\n\\\\label{metrics_fig}\\n\\\\end{figure*}\\nIn this section, we follow the same routine as the discussion on the past works and divide the metrics into different categories. A summary of the metrics can be found in Figure \\\\ref{metrics_fig}. The interested readers can also refer to Appendix \\\\ref{metrics_papers} for further information regarding the metrics and the the papers that used them. Note while we discuss the metrics in each category and we only provide mathematical expressions of the most popular metrics in Appendix \\\\ref{appendix_metrics}.\",\n \"id\": \"4c69ee31-a62c-468f-a731-a0d63c6e9fe2\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ]\n ],\n \"subsections\": [\n \"a3e52edb-4aa4-4bb7-abe0-4be3c8454b3e\",\n \"f0c26ecc-6bb8-4c9f-9512-d54a1c7c4a8c\",\n \"cbe38bc2-9ec9-4841-bbf1-11b25e444a6d\",\n \"d0bc398b-84fd-40e6-a4bf-c6c2267c1d9c\",\n \"6b621d84-5e91-4190-9b89-35e7d1a31664\"\n ],\n \"title\": \"Metrics\"\n },\n {\n \"cite_extract_rate\": 0.585365853658536,\n \"cites\": [\n 2804,\n 2814,\n 2817,\n 2807,\n 2808,\n 2816,\n 2819,\n 2815,\n 2806,\n 117,\n 2809,\n 2818,\n 7620,\n 2811,\n 2802,\n 2820,\n 2810,\n 491,\n 2805,\n 2812,\n 8590,\n 7617,\n 2813,\n 2890\n ],\n \"content\": \"Video prediction is about generating realistic images, hence the best performance is achieved when the disparities between the generated images and groundtruth images are minimal. The most straightforward way of computing the disparity is to measure pixel-wise error using a Mean Square Error (MSE) , which computes average squared intensity differences between pixels. Another more popular metric related to MSE is Peak Signal-to-Noise Ratio (PSNR) . PSNR is the ratio of the maximum pixel value (i.e. possible signal power), e.g. 255 in 8-bit images, divided by the MSE (or power of distorting noise) measure of two images. The lower the error between two images, the higher the value of PSNR, and consequently, the higher the quality of the generated images. Because of the wide dynamic range of signals, PSNR value is expressed in the logarithmic decibel scale. \\nAlthough MSE, and PSNR metrics are easy to calculate, they cannot measure the perceived visual quality of a generated image. An alternative metric to address this issue is Structural SIMilarity (SSIM) index () which is designed to model image distortion. To capture the structural differences between the two images, SSIM separates illumination information as it is independent of objects' structures. As a result, the similarity is measured by a combination of three comparisons, namely luminance, contrast, and structure. \\nHigher-level contextual similarities may not be captured by distance measures on pixel values. More recently proposed metric, Learned Perceptual Image Patch Similarity (LPIPS), () measures the similarity between two images by comparing internal activations of convolutional networks trained for high-level classification tasks. The value is calculated by an average L2 distance over normalized deep features.\\nSome other metrics that have been used in the literature are qualitative human judgment , Frechet Video Distance (FVD) () , Maximum Mean Discrepancy (MMD) , Inception Scores (IS) () , Binary Cross Entropy (BCE) , L1 , and Root MSE (RMSE) .\",\n \"id\": \"a3e52edb-4aa4-4bb7-abe0-4be3c8454b3e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 41,\n \"parent_id\": \"4c69ee31-a62c-468f-a731-a0d63c6e9fe2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ],\n [\n \"subsection\",\n \"Video prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Video prediction\"\n },\n {\n \"cite_extract_rate\": 0.378378378378378,\n \"cites\": [\n 2821,\n 2837,\n 2829,\n 7117,\n 2844,\n 2838,\n 2832,\n 2841,\n 2823,\n 2822,\n 2828,\n 2825,\n 2806,\n 2834,\n 2840,\n 2824,\n 2845,\n 2835,\n 2827,\n 2826,\n 2831,\n 2842,\n 2833,\n 2836,\n 7618,\n 2839,\n 2830,\n 2843\n ],\n \"content\": \"Similar to classification tasks, many action prediction algorithms use accuracy measure to report on the performance that is the ratio of the correct predictions with respect to the total number of predictions . Despite being used widely, accuracy on its own is not a strong indicator of performance, especially when we are dealing with class-imbalance data. This is because, for example, the model can simply favor the more represented class and predict every input as that class. This then would result in a high accuracy measure because the metric only considers the ratio of correct predictions. To address these shortcomings, some works use complimentary metrics that, in addition to correct predictions, account for different types of false predictions. These metrics are precision , recall , and Area Under the Curve (AUC) of precision-recall graph. Precision and recall also form the basis for the calculation of some higher level metrics such as F1-score , Average Precision (AP) and its variations mean AP (mAP) and calibrated AP (cAP) . Some of the less common performance metrics are Matthews Correlation Coefficient (MCC) , False positive (FP), True Positive Rate (TPR) and False Positive Rate (FPR) \\n, Prediction Power (PP) , and Mean Reciprocal Rank (MRR) .\\nDepending on the application, some algorithms evaluate the timing factor in terms of the Run Time (RT) of the model or time of the event, e.g. beginning of the next activity , Time To Accident (or collision) (TTA) , and, in the context of driving, Time To Maneuver (TTM) .\",\n \"id\": \"f0c26ecc-6bb8-4c9f-9512-d54a1c7c4a8c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 74,\n \"parent_id\": \"4c69ee31-a62c-468f-a731-a0d63c6e9fe2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ],\n [\n \"subsection\",\n \"Action prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Action prediction\"\n },\n {\n \"cite_extract_rate\": 0.42465753424657504,\n \"cites\": [\n 2860,\n 2855,\n 2867,\n 2847,\n 2854,\n 2853,\n 2868,\n 2857,\n 2869,\n 2862,\n 2856,\n 2846,\n 2852,\n 2866,\n 979,\n 988,\n 2858,\n 2859,\n 2848,\n 2850,\n 2865,\n 2831,\n 2851,\n 8591,\n 1261,\n 2849,\n 2863,\n 2861,\n 7379,\n 2830,\n 8594\n ],\n \"content\": \"Perhaps the most popular performance measure for trajectory prediction is Average Displacement Error (ADE) calculated as the average error between the prediction location and the ground truth over all time steps. Some methods complement ADE measure with its extension Final Displacement Error (FDE) . As the name suggests, FDE only measures the error between the ground truth and the generated trajectory for the final time step.\\nMany other works use the same metric as ADE or ADE/FDE without using the same terminology. It is also a common practice that instead of using average or final time step measures, to calculate the error at different time steps over a period of time .\\nTo measure displacement error in probablistic trajectory prediction algorithms, some works generate a set number of samples and report the best measure (i.e. minimum error) or average over all samples . Depending on the error metric used, some refer to these measures as Minimum ADE/FDE (MinADE/FDE) (using Euclidean distance) or Mean/Minimum Mean Square Displacement (Mean/MinMSD) (using MSE). Some of the other probabilistic measures are Log-Likelihood (LL) , Negative Log-Likelihood (NLL) , Kullback–Leibler Divergence (KLD) , Negative Log-Probability (NLP), Cross Entropy (CE) , Average Prediction Probability (APP) .\\nPerformance can also be evaluated using common classification metrics. For example, in Hit Rate (HR) and in Miss Rate (MR) metrics are used. In these cases, if the predicted trajectory is below (or above) a certain distance threshold from the groundtruth, it is considered as a hit or miss. Following a similar approach, some authors calculate accuracy or precision of predictions.\\nSome of the other metrics used in the literature are Run Time (RT) , Average Non-linear Displacement Error (ANDE)\\n, Maximum Distance (MaxD) , State collision rate (SCR) , Percentage Deviated (PD) , Distance to Goal (DtG) , Fraction of Near Misses (FNM) , Expected Calibration Error (ECE) , and qualitative (Q) . A few works predict the orientations of pedestrians or vehicles , therefore also report performance using Mean angular error (MAnE).\",\n \"id\": \"cbe38bc2-9ec9-4841-bbf1-11b25e444a6d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 73,\n \"parent_id\": \"4c69ee31-a62c-468f-a731-a0d63c6e9fe2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ],\n [\n \"subsection\",\n \"Trajectory prediction\"\n ]\n ],\n \"subsections\": [\n \"6226a0b6-b6ff-448f-a412-1d2941716ad2\"\n ],\n \"title\": \"Trajectory prediction\"\n },\n {\n \"cite_extract_rate\": 0.43076923076923,\n \"cites\": [\n 2860,\n 2855,\n 2867,\n 2847,\n 2853,\n 2857,\n 2869,\n 2862,\n 2856,\n 2846,\n 2866,\n 2852,\n 979,\n 988,\n 2858,\n 2859,\n 2850,\n 2865,\n 2831,\n 2851,\n 1261,\n 8591,\n 2861,\n 2863,\n 2849,\n 2830,\n 2864,\n 8594\n ],\n \"content\": \"Unlike video and action prediction fields, performance measures for trajectory prediction algorithms are not standardized in terms error metrics used and units of measure. For example, for measuring displacement error, although many algorithms use Euclidean Distance (ED) (aka L2-distance, L2-norm, Euclidean norm) , many others rely on different error metrics including MSE , RMSE , Weighted RMSE , Mean Absolute Error (MAE), Hausdorff Distance (HD), Modified HD (MHD) , and discrete Fr\\\\'echet distance (DFD) . Moreover, trajectory prediction algorithms use different units for measuring displacement error. These are meter , pixel , normalized pixel and feet .\\nAlthough such discrepancy between error metrics and units is expected across different applications, the problem arises when the proposed works do not specify the error metric , the unit of measure or both . Despite the fact that the reported results might imply the choice of the metrics and units, the lack of specification can cause erroneous comparisons, specially because many authors use the results of previous works directly as reported in the papers.\\nUnfortunately, metric and unit discrepancy exists within the same applications and the same error measuring techniques. For instance, in the case of ADE measure, this metric is originally proposed in in terms of ED, and was referred to as ADE by the authors of despite the fact that they used MSE instead. This is also apparent in many subsequent works that employed ADE measure. For example, the majority of methods use the original metric and report the results in terms of ED whereas some works use MSE and RMSE or do not specify the metric . Although the formulation of these metrics look similar, they produce different results. ADE using ED, for example, is square-root of squared differences averaged over all samples and time steps. Unlike ED, in RMSE, the averaging takes place inside square-root operation. MSE, on the other hand, is very different from the other two metrics, and does not calculate the root of the error. As we can also see in some of the past works, this discrepancy may cause confusion about the intended and actual metric that is used. For example, in the authors propose to use MAE metric while presenting mathematical formulation of Euclidean distance. The authors of make a similar mistake and define ED formulation but refer to it as MSE.\\nIn addition, some algorithms within the same applications and using the same datasets use different measuring unit. For instance, in the context of surveillance, ETH is one of the most commonly used datasets. Many works use this dataset for benchmarking the performance of their proposed algorithms, however, they either use different units, e.g. meter , pixel , normalized pixel , or do not specify the unit used .\\nLast but not least, another potential source of error in performance evaluation is in the design of the experiments. Taking surveillance applications as an example, it is a common practice to evaluate algorithms with 8 frames observations of the past and prediction 12 steps in the future . However, in some cases the performance of state-of-the-art is reported under the standard 8/12 condition, but the proposed algorithms are tested under different conditions. For instance, in the authors incorrectly compared the performance of their proposed algorithm using 5 observations and 5 predictions with the results of the previous works evaluated under the standard 8/12 condition.\",\n \"id\": \"6226a0b6-b6ff-448f-a412-1d2941716ad2\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 65,\n \"parent_id\": \"cbe38bc2-9ec9-4841-bbf1-11b25e444a6d\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ],\n [\n \"subsection\",\n \"Trajectory prediction\"\n ],\n [\n \"subsubsection\",\n \"Pitfalls of trajectory prediction metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pitfalls of trajectory prediction metrics\"\n },\n {\n \"cite_extract_rate\": 0.565217391304347,\n \"cites\": [\n 2814,\n 2871,\n 2821,\n 8592,\n 2874,\n 2841,\n 2875,\n 2873,\n 2870,\n 2878,\n 2876,\n 2877,\n 2872\n ],\n \"content\": \"Due to the inherent stochasticity of human body movement, motion prediction algorithms often limit their prediction horizon to approximately $500ms$. To measure the error between corresponding ground truth and predicted poses, these algorithms use mean average error, either in angle space (MAnE) or joint space (MJE) in terms of joint coordinates . In the 3D motion prediction domain, a metric known as Mean Per Joint Prediction Error (MPJPE) is used which is the error over joints normalized with respect to the root joint. \\nAs an alternative to distance error metrics, Percentage of Correct Keypoints (PCK) measures how many of the keypoints (e.g. joints) are predicted correctly. The correct predictions are those that are below a certain error threshold (e.g. 0.05). Some works also use the accuracy metric to report on how well the algorithm can localize the position of a particular joint within an error tolerance region .\\nOther metrics used in the literature include Normalized Power Spectrum Similarity (NPSS) , Reconstruction Error (RE) , Limb Orientation (LO) , PoSe Entropy (PSEnt), PoSe KL (PSKL) , qualitative human judgment and method Run Time (RT) .\",\n \"id\": \"d0bc398b-84fd-40e6-a4bf-c6c2267c1d9c\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 23,\n \"parent_id\": \"4c69ee31-a62c-468f-a731-a0d63c6e9fe2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ],\n [\n \"subsection\",\n \"Motion prediction\"\n ]\n ],\n \"subsections\": [\n \"aa256766-5674-49c7-8798-111d626f386c\"\n ],\n \"title\": \"Motion prediction\"\n },\n {\n \"cite_extract_rate\": 0.5454545454545451,\n \"cites\": [\n 2814,\n 2871,\n 2821,\n 8592,\n 2874,\n 2841,\n 2875,\n 2873,\n 2878,\n 2876,\n 2877,\n 2872\n ],\n \"content\": \"Similar to trajectory methods, motion prediction algorithms are evaluated using distance-based methods that calculate the error between pose vectors. In the case of MAnE measure, some methods use ED metric while others use MSE . Sometimes no metric is specified . The same holds for MJE measure where metrics used include MSE , RMSE , ED , MAE , or no metric is specified .\\nThe added challenge in coordinate-based error measures, e.g. MJE, MPJPE, is the error unit. While many approaches do not specify the unit explicitly , others clearly state whether the unit is in pixel , centimeter , meter or millimeter . As was the case before, here many algorithms that benchmark on the same datasets, may use different performance metrics, e.g. using popular human datasset Human 3.6M , MAnE (ED), MAnE (MSE) and MJE (ED) are used.\",\n \"id\": \"aa256766-5674-49c7-8798-111d626f386c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"d0bc398b-84fd-40e6-a4bf-c6c2267c1d9c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ],\n [\n \"subsection\",\n \"Motion prediction\"\n ],\n [\n \"subsubsection\",\n \"Pitfalls of motion prediction metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pitfalls of motion prediction metrics\"\n },\n {\n \"cite_extract_rate\": 0.56,\n \"cites\": [\n 2882,\n 2879,\n 2880,\n 2885,\n 8595,\n 2806,\n 2888,\n 2883,\n 2881,\n 2887,\n 2886,\n 9096,\n 2884,\n 2803\n ],\n \"content\": \"Depending on the task objectives, the metrics used in other prediction applications are similar to the ones discussed thus far. For instance, the applications that classify future events or outcomes, e.g. contest or an election winner, next image index for storytelling, severe weather, and pain, use common metrics such as accuracy , precision, recall , percentage of correct predictions (PCP) , and Matthews correlation coefficient (MCC) which predicts the quality of binary classification by taking into account both false and true predictions.\\nRegression based methods, such as temperature, trends, or steering prediction, use distance metrics including Euclidean Distance (ED) , RMSE , MSE , MAE , Mean Absolute Percentage Error (MAPE) , normalized MAPE (nMAPE) , and the Spearman's ranking Correlation (SRC) which measures the strength and direction of relationship between two variables.\\nOf particular interest are metrics used for evaluating generative models that predict Occupancy Grid Maps (OGMs) and segmentation maps. OGMs are grayscale images that highlight the likelihood of a certain region (represented as a cell in the grid) that is occupied. The generated map can be compared to ground truth by using image similarity metrics such as SSIM , PSNR or psi ($\\\\psi$) . Alternatively, OGM can be evaluated using a binary classification metric. Here, the grid cells are classified as occupied or free by applying a threshold and then can be evaluated as a whole by using metrics such as True Positive (TP), True Negative (TN) , Receiver Operator Characteristic (ROC) curve over TP and TN , F1-score , precision, recall, and their corresponding AUC . Given that OGM prediction algorithms are mainly used in safety-critical applications such as autonomous driving, some algorithms are also evaluated in terms of their Run Time (RT) .\\nImage similarity metrics such as PSNR and SSIM can also be used in the segmentation prediction domain . The most common metric, however, is Intersection over Union (IoU) which measures the average overlap of segmented instances with the ground truth segments. In addition, by applying a threshold to IoU scores, the true matches can be identified and used to calculate the Average Precision (AP) scores as in . Other metrics used for segmentation prediction tasks include EndPoint error (EPE) , Probabilistic Rand Index (RI), Global Consistency Error (GCE), and Variation of Information (VoI) .\",\n \"id\": \"6b621d84-5e91-4190-9b89-35e7d1a31664\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 25,\n \"parent_id\": \"4c69ee31-a62c-468f-a731-a0d63c6e9fe2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics\"\n ],\n [\n \"subsection\",\n \"Other prediction applications\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Other prediction applications\"\n },\n {\n \"cite_extract_rate\": 0.33628318584070804,\n \"cites\": [\n 2899,\n 2838,\n 2808,\n 2869,\n 2824,\n 2892,\n 8596,\n 2835,\n 2897,\n 1517,\n 2903,\n 8590,\n 2894,\n 6980,\n 2891,\n 2900,\n 2626,\n 2908,\n 2906,\n 2896,\n 2905,\n 2898,\n 2902,\n 8597,\n 1733,\n 2904,\n 2907,\n 2893,\n 2878,\n 2851,\n 7379,\n 2901,\n 2880,\n 2895\n ],\n \"content\": \"\\\\label{datasets}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\resizebox{1\\\\textwidth}{!}{\\n\\\\begin{tabular}{|l|l|l|l|c|c|c|c|c|}\\n\\\\hline\\n\\\\multirow{2}{*}{\\\\textbf{Year}} & \\\\multicolumn{1}{c|}{\\\\multirow{2}{*}{\\\\textbf{Dataset}}} & \\\\multicolumn{1}{c|}{\\\\multirow{2}{*}{\\\\textbf{Type}}} & \\\\multicolumn{1}{c|}{\\\\multirow{2}{*}{\\\\textbf{Annotations}}} & \\\\multicolumn{5}{c|}{\\\\textbf{Application}} \\\\\\\\ \\\\cline{5-9} \\n & \\\\multicolumn{1}{c|}{} & \\\\multicolumn{1}{c|}{} & \\\\multicolumn{1}{c|}{} & \\\\textit{\\\\textbf{V}} & \\\\textit{\\\\textbf{A}} & \\\\textit{\\\\textbf{T}} & \\\\textit{\\\\textbf{M}} & \\\\textit{\\\\textbf{O}} \\\\\\\\ \\\\hline\\n\\\\multirow{12}{*}{\\\\textit{\\\\textbf{2019}}} & ARGOVerse & Traffic & RGB, LIDAR, 3D BB & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & CARLA & Traffic (sim) & RGB & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & EgoPose & Pose (ego) & RGB, 3D Pose & & & & x & \\\\\\\\ \\\\cline{2-9} \\n & Future Motion (FM) & Mix & RGB, BB, Attrib. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & InstaVariety & Activities & RGB, BB, Pose & & & & x & \\\\\\\\ \\\\cline{2-9} \\n & INTEARCTION & Traffic & Map, Traj. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & Luggage & Robot & Stereo RGB, BB & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & MGIF & Activities & RGB & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & Pedestrian Intention Estimation (PIE) & Traffic & RGB, BB, Class, Attrib., Temporal seg., Vehicle sensors & & x & x & & \\\\\\\\ \\\\cline{2-9} \\n & nuScenes & Traffic & RGB, LIDAR, 3D BB, Vehicle sensors & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & Vehicle-Pedestrian-Mixed (VPM) & Traffic & RGB, BB & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & TRAF & Traffic & RGB, BN, Class, Time-of-day & & & x & & \\\\\\\\ \\\\hline\\n\\\\multirow{10}{*}{\\\\textit{\\\\textbf{2018}}} & 3D POSES IN THE WILD (3DPW) & Outdoor & RGB, 2D/3D Pose, Models & & & & x & \\\\\\\\ \\\\cline{2-9} \\n & ActEV/VIRAT & Surveillance & RGB, BB, Activity, Temporal seg. & & x & x & & \\\\\\\\ \\\\cline{2-9} \\n & ACTICIPATE & Interaction & RGB, Gaze, Pose & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Atomic Visual Actions (AVA) & Activities & RGB, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Epic-Kitchen & Cooking (ego) & RGB, Audio, BB, Class, Text, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & EGTEA Gaze+ & Cooking (ego) & RGB, Gaze, Mask, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & ShanghaiTech Campus (STC) & Surveillance & RGB, Anomaly & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & ShapeStack & Objects (sim) & RGBD, Mask, Stability & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & VIENA & Traffic (sim) & RGB, Activity, Vehicle sensors & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & YouCook2 & Cooking & RGB, Audio, Text, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\hline\\n\\\\multirow{10}{*}{\\\\textit{\\\\textbf{2017}}} & BU Action (BUA) & Activities & RGB (image), Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & CityPerson & Traffic & Stereo RGB, BB, Semantic seg. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & Epic-Fail & Risk assessment & RGB, BB, Traj., Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Joint Attention in Autonomous Driving (JAAD) & Traffic & RGB, BB, Attrib., Temporal seg. & x & x & x & & x \\\\\\\\ \\\\cline{2-9} \\n & L-CAS & Traffic & LIDAR, 3D BB, Attrib. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & Mouse Fish & Animals & Depth, 3D Pose & & & & x & \\\\\\\\ \\\\cline{2-9} \\n & Oxford Robot Car (ORC) & Traffic & Stereo RGB, LIDAR, Vehicle sensors & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & PKU-MMD & Activities, interactions & RGBD, IR, 3D Pose, Multiview, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Recipe1M & Cooking & RGB(image), Text & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & STRANDS & Traffic & RGBD, 3DBB & & & x & & \\\\\\\\ \\\\hline\\n\\\\multirow{13}{*}{\\\\textit{\\\\textbf{2016}}} & BAIR Push & Object manipulation & RGB & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & Bouncing Ball (BB) & Simulation & RGB & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & Miss Universe (MU) & Miss universe & RGB, BB, Scores & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & Cityscapes & Traffic & Stereo RGB, BB, Semantic seg., Vehicle Sensors & x & & x & & x \\\\\\\\ \\\\cline{2-9} \\n & CMU Mocap & Activities & 3D Pose, Activity & & x & & x & \\\\\\\\ \\\\cline{2-9} \\n & Dashcam Accident Dataset (DAD) & Traffic, accidents & RGB, BB, Class, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & NTU RGB-D & Activities & RGBD, IR, 3D Pose, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Ongoing Activity (OA) & Actvities & RGB, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & OAD & Activities & RGBD, 3D Pose, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Stanford Drone (SD) & Surveillance & RGB, BB, Class & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & TV Series & Activities & RGB, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Visual StoryTelling (VIST) & Visual story & RGB, Text & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & Youtube-8M & Activities & RGB, Activity, Temporal seg. & x & & & & \\\\\\\\ \\\\hline\\n\\\\multirow{14}{*}{\\\\textit{\\\\textbf{2015}}} & Amazon & Fashion & Features, Attrib., Text & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & Atari & Games & RGB & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & Brain4Cars & Traffic, Driver & RGB, BB, Attrib., Temporal seg., Vehicle sensors & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & CMU Panoptic & Interaction & RGBD, Multiview, 3D Pose, 3D facial landmark & x & x & & x & \\\\\\\\ \\\\cline{2-9} \\n & First Person Personalized Activities (FPPA) & Activities (ego) & RGB, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & GTEA Gaze + & Cooking (ego) & RGB, Gaze, Mask, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & MicroBlog-Images (MBI-1M) & Tweets & RGB (image), Attrib., Text & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & MOT & Surveillance & RGB, BB & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & Moving MNIST (MMNIST) & Digits & Grayscale & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & SUN RGB-D & Places & RGBD, 3D BB , Class & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & SYSU 3DHOI & Object interaction & RGBD, 3D Pose, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & THUMOS & Activities & RGB, Activity, Temporal seg. & x & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Watch-n-Push (WnP) & Activities & RGBD, 3D Pose, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Wider & Activities & RGB (image), Activity & & x & & & \\\\\\\\ \\\\hline\\n\\\\multirow{5}{*}{\\\\textit{\\\\textbf{2014}}} & Breakfast & Cooking & RGB, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Human3.6M & Activities & RGB, 3D Pose, Activity & x & x & & x & \\\\\\\\ \\\\cline{2-9} \\n & MPII Human Pose & Activities & RGB, Pose, Activity & & & & x & \\\\\\\\ \\\\cline{2-9} \\n & Online RGBD Action Dataset (ORGBD) & Activities & RGBD, BB, 3D Pose, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Sports-1M & Sports & RGB, Activity & x & x & & & \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n}\\n\\\\caption{A summary of common datasets from years 2014-2019 used in vision-based prediction applications, namely video (V), action (A), trajectory (T), motion (M) and others (O). The annotation column specifies the type of data (e.g. RGB, Infrared(IR)) and annotation types. All datasets contain image sequences unless specified by \\\"image\\\". As for annotations, BB stands for bounding box. Attributes include any object characteristics (e.g. for pedestrians demographics, behavior). Vehicle sensors may include speed, steering angle, GPS, etc. Temporal seg. identifies the datasets that specify the start and end of the events.}\\n\\\\label{dataset_table1}\\n\\\\end{table*}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\resizebox{1\\\\textwidth}{!}{\\n\\\\begin{tabular}{|l|l|l|l|c|c|c|c|c|}\\n\\\\hline\\n\\\\multirow{2}{*}{\\\\textbf{Year}} & \\\\multicolumn{1}{c|}{\\\\multirow{2}{*}{\\\\textbf{Dataset}}} & \\\\multicolumn{1}{c|}{\\\\multirow{2}{*}{\\\\textbf{Type}}} & \\\\multicolumn{1}{c|}{\\\\multirow{2}{*}{\\\\textbf{Annotations}}} & \\\\multicolumn{5}{c|}{\\\\textbf{Application}} \\\\\\\\ \\\\cline{5-9} \\n & \\\\multicolumn{1}{c|}{} & \\\\multicolumn{1}{c|}{} & \\\\multicolumn{1}{c|}{} & \\\\textit{\\\\textbf{V}} & \\\\textit{\\\\textbf{A}} & \\\\textit{\\\\textbf{T}} & \\\\textit{\\\\textbf{M}} & \\\\textit{\\\\textbf{O}} \\\\\\\\ \\\\hline\\n \\\\multirow{7}{*}{\\\\textit{\\\\textbf{2013}}} & 50Salads & Cooking (ego) & RGBD, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & ATC & Surveillance & RGB, Traj., Attrib. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & CAD-120 & Activities & RGBD, 3D Pose, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & CHUK Avenue & Surveillance & RGB, BB, Anomaly, Temporal seg. & x & & x & & \\\\\\\\ \\\\cline{2-9} \\n & Daimler Path & Traffic & Stereo Grayscale, BB, Temporal seg. , Vehicle sensors & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Joint-annotated HMDB (JHMDB) & Activities & RGB, Mask, Activity, Pose, Optical flow & x & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Penn Action & Activities & RGB, BB, Pose, Activity & x & & & x & \\\\\\\\ \\\\hline\\n\\\\multirow{11}{*}{\\\\textit{\\\\textbf{2012}}} & BIT & Interaction & RGB, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & GTEA Gaze & Cooking (ego) & RGB, Gaze, Mask, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & KITTI & Traffic & Stereo RGB, LIDAR, BB, Optical flow, Vehicle sensors & x & & x & & x \\\\\\\\ \\\\cline{2-9} \\n & MANIAC & Object manipulation & RGBD, Semantic seg., Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & MPII-Cooking & Cooking & RGB, 3D Pose, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & MSR Daily Activity (MSRDA) & Activities & Depth, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & New York Grand Central (GC) & Surveillance & RGB, Traj. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & SBU Kinetic Interction (SBUKI) & Interaction & RGBD, 3D Pose, Activity & & x & x & x & \\\\\\\\ \\\\cline{2-9} \\n & UCF-101 & Activities & RGB, Activity & x & x & & x & \\\\\\\\ \\\\cline{2-9} \\n & UTKinect-Action (UTKA) & Activities & RGBD, 3D Pose, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & UvA-NEMO & Smiles & RGB & x & & & & \\\\\\\\ \\\\hline\\n\\\\multirow{5}{*}{\\\\textit{\\\\textbf{2011}}} & Ford campus vision LiDAR (FCVL) & Traffic & RGB, LIDAR, Vehicle sensors & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & Human Motion Database (HMDB) & Activities & RGB, BB, Mask, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Stanford 40 & Activities & RGB (image), BB, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Town Center & Surveillance & RGB, BB & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & VIRAT & Surveillance, Activities & RGB, BB, Activity, Temporal seg. & & x & x & & \\\\\\\\ \\\\hline\\n\\\\multirow{8}{*}{\\\\textit{\\\\textbf{2010}}} & DISPLECS & Traffic & RGB, Vehicle sensors & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & MSR & Activities & Depth, Activity & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & MUG & Facial expressions & RGB, Keypoints, Label & \\\\multicolumn{1}{l|}{x} & \\\\multicolumn{1}{l|}{} & \\\\multicolumn{1}{l|}{} & \\\\multicolumn{1}{l|}{} & \\\\multicolumn{1}{l|}{} \\\\\\\\ \\\\cline{2-9} \\n & PROST & Objects & RGB, BB & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & TV Human Interaction (THI) & Interaction & RGB, BB, Head pose, Activity & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & UT Interaction (UTI) & Interaction & RGB, BB, Activity, Temporal seg. & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & VISOR & Surveillance & RGB, BB, Pose, Attrib. & x & & & & \\\\\\\\ \\\\cline{2-9} \\n & Willow Action & Activiites & RGB (image), Activity & & x & & & \\\\\\\\ \\\\hline\\n\\\\multirow{9}{*}{\\\\textit{\\\\textbf{2009}}} & Caltech Pedestrian & Traffic & RGB, BB & x & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Collective Activity (CA) & Interaction & RGB, BB, Attrib., Activity, Temporal seg. & & x & x & x & \\\\\\\\ \\\\cline{2-9} \\n & Edinburgh IFP & Surveillance & RGB, BB & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & ETH & Surveillance & RGB, Traj. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & OSU & Sports & RGB, BB, Attrib. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & PETS2009 & Surveillance & RGB, BB & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & QMUL & Traffic, anomaly & RGB, Traj. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & TUM Kitchen & Activities & RGB, RFID, 3D Pose, Activity, Temporal seg. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & YUV Videos & Mix videos & RGB & x & & & & \\\\\\\\ \\\\hline\\n\\\\multirow{2}{*}{\\\\textit{\\\\textbf{2008}}} & Daimler & Traffic & Grayscale, BB & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & MIT Trajectory (MITT) & Surveillance & RGB, Traj. & & & x & & \\\\\\\\ \\\\hline\\n\\\\multirow{5}{*}{\\\\textit{\\\\textbf{2007}}} & AMOS & Weather & RGB, Temperature, Time & & & & & x \\\\\\\\ \\\\cline{2-9} \\n & ETH Pedestrian & Traffic & RGB, BB & & x & & & \\\\\\\\ \\\\cline{2-9} \\n & Lankershim Boulevard & Traffic & RGB, Traj. & & & x & & \\\\\\\\ \\\\cline{2-9} \\n & Next Generation Simulation (NGSIM) & Traffic & Map, Traj. & & x & x & & \\\\\\\\ \\\\cline{2-9} \\n & UCY & Surveillance & RGB, Traj., Gaze & & & x & & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{2006}} & Tuscan Arizona & Weather & RGB & & & & & x \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{2004}} & KTH & Activities & Grayscale, Activity & x & & & & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{1981}} & Golden Colorado & Weather & RGB & & & & & x \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n}\\n\\\\caption{A summary of common datasets from years 2013 and earlier used in vision-based prediction applications, namely video (V), action (A), trajectory (T), motion (M) and others (O). The annotation column specifies the type of data (e.g. RGB, Infrared(IR)) and annotation types. All datasets contain sequences unless specified by \\\"image\\\". As for annotations, BB stands for bounding box. Attributes include any object characteristics (e.g. for pedestrians demographics, behavior). Vehicle sensors may include speed, steering angle, GPS, etc. Temporal seg. identifies the datasets that specify the start and end of the events.}\\n\\\\label{dataset_table2}\\n\\\\end{table*}\\n\\\\begin{figure*}\\n\\\\centering\\n\\\\includegraphics[width=1\\\\textwidth]{dataset_paper.png}\\n\\\\caption{An illustration of datasets and papers that use them.}\\n\\\\label{datasets_fig}\\n\\\\end{figure*}\\nWe have identified more than 100 datasets that are used in the vision-based prediction literature. Discussing all datasets in detail is beyond the scope of this paper. We provide a summary of the datasets and their characteristics in Tables \\\\ref{dataset_table1} and \\\\ref{dataset_table2} and briefly discuss more popular datasets in each field. Figure \\\\ref{datasets_fig} illustrates the list of papers and corresponding datasets used for evaluation. Note that the papers that do not use publicly available datasets are not listed in this figure. For further information, the readers can also refer to Appendices \\\\ref{links_datasets} and \\\\ref{papers_datasets}. \\n\\\\textbf{Video prediction.} Almost any forms of sequential RGB images can be used for evaluation of video prediction algorithms. Among the most common datasets are traffic datasets such a KITTI , and Caltech Pedestrians . KITTI is a dataset recorded from inside of a vehicle and contains images of urban roads annotated with bounding box information. It also contains depth maps, LIDAR point clouds and semantic segmentation maps. Caltech Pedestrian is a similar dataset with the difference of only containing RGB images and bounding boxes for pedestrians. It also contains occlusion bounding boxes highlighting the visible portions of the pedestrians. Activity datasets such as UCF-101 and Human3.6M are also widely used. UCF-101 contains videos of various types of activities such as sports, applying makeup, playing music instruments annotated with activity labels per video. Human3.6M consists of 3.6 million 3D human poses and corresponding images recorded from 11 professional actors. This dataset contains 17 generic scenarios such as discussion, smoking, and taking photos.\\n\\\\textbf{Action prediction. } The algorithms in this domain are evaluated on a wide range of datasets. For anticipation tasks, traffic datasets such as Next Generation Simulation (NGSIM) and Joint Attention in Autonomous Driving (JAAD) are used. NGSIM contains trajectories of vehicles driving on highways in the United States. The trajectories are accompanied by the top-down views of the corresponding road structures. The JAAD dataset contains videos of pedestrians crossing the road recorded using an on-board camera. This dataset contains the frame-wise pedestrian bounding boxes, and action labels as well as pedestrians' and roads' attributes. A similar dataset to JAAD is Pedestrian Intention Estimation (PIE) which, in addition, provides the ego-vehicle sensor data and spatial annotations for traffic elements.\\nAnother popular category of datasets in this domain is those containing videos of cooking activities. These datasets are Epic-Kitchen , 50salads , Breakfast and MPII-Cooking . These datasets contain videos showing sequences of different cooking actions of preparing meals. All videos in the datasets have temporal segments with corresponding activity labels. Some datasets also provide additional annotations such as object bounding boxes, voice and text in Epic-Kitchen, and the poses of the actors in MPII-Cooking. \\nEarly action prediction works widely use the popular UCF-101 dataset and interaction datasets such as UT Interaction (UTI) and BIT . UTI and BIT contain videos of people engaged in interaction with the corresponding label for the types of interactions. In addition, UTI has the added temporal segment annotations detailing different stages of interactions.\\n\\\\textbf{Trajectory prediction. } The most common datasets in this domain are ETH and UCY which contain surveillance videos of pedestrians walking on sidewalks annotated with their position coordinates. UCY also provides the gaze directions to capture the viewing angle of pedestrians. Another popular dataset is Stanford Aerial Pedestrian (SAP), also known as Stanford Drone (SD) . This dataset has the footage of road users from a top-down view recorded by a drone. The annotations include bounding boxes and object class labels. \\n\\\\textbf{Motion prediction. } The algorithms in this domain are mainly evaluated on the widely popular dataset Human 3.6M described earlier. This dataset is particularly suitable for these applications because it contains accurate 3D poses of the actors recorded by a high-speed motion capture system. Using this dataset, the background can be accurately removed allowing the algorithms to focus purely on changes in the poses. Another popular dataset in this field is Penn Action which contains RGB videos of various activities with corresponding activity labels and poses of the actors involved. \\n\\\\textbf{Other applications. } The most notable datasets are KITTI which is used by the OGM prediction algorithms and CityScapes that is used by the segmentation prediction algorithms. CityScapes contains video footage of urban environments recorded by an on-board camera. The data is annotated with semantic masks of the traffic objects and corresponding bounding boxes.\",\n \"id\": \"6ab7b3e0-1f9f-4a59-bcf5-66c554040f7e\",\n \"level\": \"section\",\n \"origin_cites_number\": 113,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"2dbff660-6f4e-44d8-90a2-fa6081f28574\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Discussion\"\n ]\n ],\n \"subsections\": [\n \"52732619-477d-4fcc-bf24-a81fa432764f\",\n \"24dca31c-1523-4bb9-9bbe-85a5ca0d065e\",\n \"dd7e38b9-46fb-48ee-b838-9bbba4067a29\",\n \"6521f90c-6d80-4566-ab87-2b5811291280\"\n ],\n \"title\": \"Summary and Discussion\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"There are many factors that define the choice of architecture for vision-based prediction tasks. These factors include the types of input data and expected output, computational efficiency, application-specific constrains, etc. For instance, in terms of network choice, whether it is feedforward and recurrent, no preference is observed in video applications. However, in the case of action, trajectory and motion predictions, recurrent architectures are strongly preferred. This can be due to the fact that these applications often rely on multi-modal data which can be combined easier in a recurrent framework. In the case of trajectory prediction, recurrent architectures give the flexibility of varying observation and prediction lengths without the need for architectural modifications. \\nGenerative Adversarial Networks (GANs) are widely used in video prediction applications and to some extent in trajectory prediction methods. Some of the major challenges using generative models is to deal with inherent uncertainty of future representations, in particular, this an issue in the context of trajectory prediction due to high unpredictability of human movement. To remedy this issue and to capture uncertainty of movement, techniques such as variational auto encoders, in which uncertainty is modeled as a latent distribution, and the use of probabilistic objective functions are common. \\nA more recent trend in the field of vision-based prediction (and perhaps in other computer vision applications) is the use of attention modules. These modules can be applied at spatial or temporal level or even to adjust the impact of different modalities of data.\",\n \"id\": \"52732619-477d-4fcc-bf24-a81fa432764f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"2dbff660-6f4e-44d8-90a2-fa6081f28574\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Discussion\"\n ],\n [\n \"subsection\",\n \"Architecture\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Architecture\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The type of data and methods of processing vary across different applications. For instance, video applications mainly rely on images but also take advantage of alternative representations, such as optical flow, poses, object-based keypoints, and report improved results. Similarly, many action prediction algorithms use different sources of information such as optical flow, poses, scene attributes, text, complementary sensor readings (e.g. speed in vehicles), gaze and time of the actions.\\nTrajectory prediction algorithms, predominantly rely on trajectory information, with some exceptions that use scene layouts, complimentary sensors' readings or other constrains. One of the main applications in this domain, in particular surveillance, is modeling the social interaction between the dynamic agents. Unlike other vision-based applications, motion prediction algorithms are mainly single-modal and use only poses and the images of agents as inputs.\",\n \"id\": \"24dca31c-1523-4bb9-9bbe-85a5ca0d065e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"2dbff660-6f4e-44d8-90a2-fa6081f28574\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Discussion\"\n ],\n [\n \"subsection\",\n \"Data representation and processing\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data representation and processing\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"dd7e38b9-46fb-48ee-b838-9bbba4067a29\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"2dbff660-6f4e-44d8-90a2-fa6081f28574\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Discussion\"\n ],\n [\n \"subsection\",\n \"Evaluation\"\n ]\n ],\n \"subsections\": [\n \"eb39ad05-ca7d-4eac-ac9e-5926623f5423\",\n \"f4e6a673-f09d-443b-b219-ea24268d39ec\"\n ],\n \"title\": \"Evaluation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Metrics may vary across different applications of vision-based prediction. Video prediction algorithms, for instance, are mainly evaluated using MSE, SSIM, and PSNR, whereas in the case of action prediction algorithms the main metrics are accuracy, precision, and recall. Trajectory prediction works often measure the average distance (ADE) or final distance (FDE) between the actual and predicted locations of the agents. The models with probabilistic outputs are also evaluated using NLL and KLD metrics. Distance-based metrics are used in motion prediction methods where the error in joint prediction is either calculated on average (MJE) or per joint (MPJPE). In addition, joint accuracy can be reported in terms of the percentage of correct prediction using PCK metric. In this case, a tolerance threshold is defined to determine whether a predicted joint is correct.\\nWhile calculating performance error for video and action prediction algorithms are fairly standardized, there are major discrepancies across different works in the way error is computed for trajectory and motion prediction algorithms. For example, in trajectory prediction, distance error is calculated by using metrics such as MSE, RMSE, ED, etc. and units such as pixels and meters. Such discrepancy, and the fact that many works omit mentioning the choice of error metrics and units, increases the chance of incorrect comparisons between models.\",\n \"id\": \"eb39ad05-ca7d-4eac-ac9e-5926623f5423\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"dd7e38b9-46fb-48ee-b838-9bbba4067a29\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Discussion\"\n ],\n [\n \"subsection\",\n \"Evaluation\"\n ],\n [\n \"subsubsection\",\n \"Metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Metrics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"The choice of datasets depends on the objective of the applications. For example, action prediction algorithms for cooking activities are evaluated on datasets such as Epic-Kitchen, 50Salads, Breakfast, and MPII-Cooking and the ones for traffic events evaluated on JAAD, NSGIM, and PIE. Similarly, trajectory prediction works for surveillance widely use UCY, ETH, and SD and for traffic NGSIM. Motion prediction algorithms are more focusing on individual movements in diverse context, therefore predominantly use Human3.6M and Penn Action datasets. \\nCompared to other applications, video prediction is an exception. The algorithms in this group are evaluated on almost any datasets with video content. The algorithms in this domain are often task agnostic meaning that the same approaches are evaluated on datasets with traffic scenes (e.g. KITTI, Caltech Pedestrian), general activities (e.g. UCF-101, Penn Action), basic actions (e.g. Human3.6M, KTH) and synthetic data (e.g. MMNIST, Atari games). Although such generalizability across different domains is a desirable feature in video prediction algorithms, it is often the case that the reason behind the choice of datasets is not discussed raising the question of whether the decision for selecting particular datasets is motivated by the limitations of the algorithms.\",\n \"id\": \"f4e6a673-f09d-443b-b219-ea24268d39ec\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"dd7e38b9-46fb-48ee-b838-9bbba4067a29\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Discussion\"\n ],\n [\n \"subsection\",\n \"Evaluation\"\n ],\n [\n \"subsubsection\",\n \"Datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In recent years we have witnessed an exponential growth in the number of works published in the field of vision-based prediction. There are still, however, many open research problems in the field that need to be addressed. \\nThe ability to hallucinate or generate parts of the image that were not previously observed is still a major challenge in video prediction applications. In addition, the algorithms in this domain cannot deal with cases where some objects go out of view in future time frames. The performances of action prediction algorithms are still sub-optimal, especially in safety critical and complex tasks such as event prediction in traffic scenes. To make predictions in such cases, many modalities of data and the relationships between them should be considered which is often not the case in the proposed approaches.\\nTrajectory prediction algorithms mainly rely on changes in the location of the agents to predict their future states. Although, this might be an effective approach for tasks such as surveillance, in many other cases it might not be sufficient. For example, in order to predict trajectories of pedestrians in traffic scenes, many other sources of information, such as their poses and orientation, road structure, interactions, road conditions, traffic flow, etc., are potentially relevant. Such contextual analysis can also be beneficial for motion prediction algorithms which manly rely on the changes in poses to predict the future.\\nIn terms of the choice of learning architectures and training schemes, a systematic comparison of different approaches, e.g. using feedforward vs recurrent networks, the benefits of using adversarial training schemes, various uncertainty modeling approaches, etc. is missing. Such information can be partially extracted from the existing literature, however, in many cases it is not possible due to the lack of standard evaluation procedures and metrics, unavailability of corresponding implementation code and the datasets used for comparison. \\n\\\\ifCLASSOPTIONcaptionsoff\\n \\\\newpage\\n\\\\fi\\n\\\\bibliographystyle{IEEEtran}\\n\\\\bibliography{vision_based_prediction}\\n\\\\appendices\",\n \"id\": \"6521f90c-6d80-4566-ab87-2b5811291280\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"2dbff660-6f4e-44d8-90a2-fa6081f28574\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Summary and Discussion\"\n ],\n [\n \"subsection\",\n \"What's next\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"What's next\"\n },\n {\n \"cite_extract_rate\": 0.7272727272727271,\n \"cites\": [\n 2871,\n 7117,\n 2817,\n 2838,\n 8592,\n 2808,\n 2874,\n 2869,\n 2819,\n 2856,\n 2866,\n 2809,\n 979,\n 2834,\n 2802,\n 2845,\n 2881,\n 2873,\n 2870,\n 2805,\n 2831,\n 2878,\n 2851,\n 2876,\n 8590,\n 7618,\n 2863,\n 2849,\n 2861,\n 2872,\n 2813,\n 2890\n ],\n \"content\": \"\\\\label{papers_with_code}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\resizebox{1\\\\textwidth}{!}{\\n\\\\begin{tabular}{|c|l|l|}\\n\\\\hline\\n\\\\textbf{App} & \\\\multicolumn{1}{c|}{\\\\textbf{Paper}} & \\\\multicolumn{1}{c|}{\\\\textbf{Link}} \\\\\\\\ \\\\hline\\n\\\\multirow{14}{*}{\\\\textbf{Video}}\\n & & \\\\url{https://github.com/andrewjywang/SEENet} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/Yijunmaverick/FlowGrounded-VideoPrediction} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/liuem607/DYAN} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/garyzhao/FRGAN} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/jthsieh/DDPAE-video-prediction} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/xjwxjw/VPSS} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/jinbeibei/VarNet} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://bit.ly/2HqiHqx} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/ujjax/pred-rnn} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/rubenvillegas/icml2017hierchvid} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/tensorflow/models/tree/master/research/video_prediction} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/junhyukoh/nips2015-action-conditional-video-prediction} \\\\\\\\ \\\\hline\\n\\\\multirow{8}{*}{\\\\textbf{Action}} & & \\\\url{https://github.com/google/next-prediction} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/fpv-iplab/rulstm} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/aras62/SF-GRU} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/aashi7/NearCollision} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/yabufarha/anticipating-activities} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/gurkirt/realtime-action-detection} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/asheshjain399/RNNexp} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/aditya7874/Activity-Prediction-in-EgoCentric-Videos} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/JoeHEZHAO/Spatiotemporal-Residual-Propagation} \\\\\\\\ \\\\hline\\n\\\\multirow{15}{*}{\\\\textbf{Trajectory}} & & \\\\url{https://go.umd.edu/TraPHic} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/google/next-prediction} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/zhangpur/SR-LSTM} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/aras62/PIEPredict} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://sites.google.com/view/precog} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://rebrand.ly/INFER-results} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/wzhi/KernelTrajectoryMaps} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/apratimbhattacharyya18/onboard_long_term_prediction} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/agrimgupta92/sgan} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/svip-lab/CIDNN} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/yfzhang/vehicle-motion-forecasting} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/yadrimz/DESIRE} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/quancore/social-lstm} \\\\\\\\ \\\\hline\\n\\\\multirow{11}{*}{\\\\textbf{Motion}} & & \\\\url{https://github.com/cr7anand/neural_temporal_models} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/BII-wushuang/Lie-Group-Motion-Prediction} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/wei-mao-2019/LearnTrajDep} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/magnux/MotionGAN} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/Khrylx/EgoPose} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://jasonyzhang.com/phd/} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/eddyhkchiu/pose_forecast_wacv/.} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/ywchao/image-play} \\\\\\\\ \\\\cline{2-3} \\n & & \\\\url{https://github.com/una-dinosauria/human-motion-prediction} \\\\\\\\ \\\\cline{2-3} \\\\hline\\n\\\\multicolumn{1}{|l|}{\\\\multirow{2}{*}{\\\\textbf{Others}}} & & \\\\url{https://bitbucket.org/vguizilini/cvpp/src} \\\\\\\\ \\\\cline{2-3} \\n\\\\multicolumn{1}{|l|}{} & & \\\\url{https://github.com/facebookresearch/SegmPred} \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n}\\n\\\\caption{A summary of vision-based prediction papers with published code.}\\n\\\\label{table_published_code}\\n\\\\end{table*}\\nA list of papers with official published code can be found in Table \\\\ref{table_published_code}.\",\n \"id\": \"23d750d7-3a7a-460d-8e6b-4f54e379529f\",\n \"level\": \"section\",\n \"origin_cites_number\": 44,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Papers with code\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Papers with code\"\n },\n {\n \"cite_extract_rate\": 0.461187214611872,\n \"cites\": [\n 2829,\n 2871,\n 2844,\n 2807,\n 2808,\n 2838,\n 2869,\n 2841,\n 2862,\n 2822,\n 979,\n 2840,\n 2824,\n 2811,\n 2848,\n 2845,\n 2850,\n 2835,\n 2873,\n 2870,\n 2827,\n 2842,\n 2836,\n 2887,\n 9096,\n 1261,\n 8590,\n 2849,\n 2877,\n 2813,\n 2814,\n 2855,\n 2860,\n 7117,\n 2853,\n 2879,\n 2816,\n 2823,\n 2856,\n 2866,\n 2852,\n 2888,\n 2809,\n 2858,\n 2802,\n 2865,\n 2805,\n 2826,\n 2833,\n 2876,\n 7618,\n 2839,\n 2830,\n 2843,\n 2847,\n 2854,\n 8592,\n 2882,\n 2868,\n 2832,\n 2874,\n 2857,\n 2828,\n 2806,\n 8595,\n 2834,\n 2859,\n 2875,\n 2881,\n 2812,\n 2886,\n 2878,\n 2851,\n 2863,\n 7617,\n 7379,\n 2803,\n 2890,\n 8594,\n 2804,\n 2821,\n 2837,\n 2867,\n 2817,\n 2880,\n 2885,\n 2819,\n 2846,\n 2815,\n 2825,\n 2883,\n 988,\n 2818,\n 2820,\n 2810,\n 2831,\n 2884,\n 8591,\n 2861,\n 2872,\n 2864\n ],\n \"content\": \"\\\\label{metrics_papers}\\n\\\\begin{table*}[ht]\\n\\\\centering\\n\\\\begin{tabular}{|l|p{0.8\\\\textwidth}|}\\n\\\\hline\\n\\\\textbf{Metric} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{BCE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{FVD}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Human}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{IS}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{L1}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{LPIPS}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MMD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MSE}} & ,,,,, ,,,,, ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PSNR}} & ,,,,, ,,,,, ,,,,, ,,,,, ,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{RMSE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SSIM}} & ,,,,, ,,,,, ,,,,, ,,,,, ,, \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Metrics used in \\\\textbf{video prediction} applications.}\\n\\\\label{video_metrics_table}\\n\\\\end{table*}\\n\\\\begin{table*}[ht]\\n\\\\centering\\n\\\\begin{tabular}{|l|p{0.8\\\\textwidth}|}\\n\\\\hline\\n\\\\textbf{Metrics} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{AP}} & ,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ATTA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ATTC}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{AUC}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Accuracy}} & ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,,,,,,,,,, , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{F1}} & ,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{FP}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MAE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MCC}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MRR}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PP}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Precision}} & ,,,,, ,,,,, ,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{RMSE}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Recall}} & ,,,,, ,,,,, ,,,,, ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Run time}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TNR}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TPR}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TTA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TTM}} & ,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{cAP}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{mAP}} & ,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{recall}} & , \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Metrics used in \\\\textbf{action prediction} applications.}\\n\\\\label{action_metrics_table}\\n\\\\end{table*}\\n\\\\begin{table*}[ht]\\n\\\\centering\\n\\\\begin{tabular}{|l|p{0.8\\\\textwidth}|}\\n\\\\hline\\n\\\\textbf{Metrics} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ADE}} & ,,,,, ,,,,,,,,,, ,,,,, ,,,,,,,,,, ,,,,, ,,,,,,,,,, ,\\n\\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{AEDE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ANDE}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{APP}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Accuracy}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{DtG}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ECE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ED}} & ,,,,,,,,,\\n\\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{FDE}} & ,,,,, ,,,,, ,,,,, ,,,,, ,,,,, ,\\n\\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{FNM}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Hit rate}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{KLD}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{L1}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{LL}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MAnE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MHD}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MSE}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Miss rate}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{NLL}} & ,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{NLP}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{None}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Precision}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{RMSE}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Run time}} & ,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SCR}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{WRMSE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{maxD}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{meanMSD}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{minADE}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{minED}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{minFDE}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{minMSD}} & , \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Metrics used in \\\\textbf{trajectory prediction} applications.}\\n\\\\label{trajectory_metrics_table}\\n\\\\end{table*}\\n\\\\begin{table*}[h]\\n\\\\centering\\n\\\\begin{tabular}{|l|l|}\\n\\\\hline\\n\\\\textbf{Metrics} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Accuracy}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Human}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{LO}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MAnE}} & ,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MJE}} & ,,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MPJPE}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{NPSS}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PCK}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PSEnt}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PSKL}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{RE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Run time}} & , \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Metrics used in motion prediction applications.}\\n\\\\label{motion_metrics_table}\\n\\\\end{table*}\\nLists of metrics and corresponding papers can be found in Tables \\\\ref{video_metrics_table}, \\\\ref{action_metrics_table}, \\\\ref{trajectory_metrics_table}, \\\\ref{motion_metrics_table}, and \\\\ref{other_metrics_table}. \\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\begin{tabular}{|l|l|}\\n\\\\hline\\n\\\\textbf{Metrics} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{AUC}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Accuracy}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ED}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{EPE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{F1}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{GCE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ISM}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{IoU}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MAE}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MAPE}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MCC}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MIoU}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MSE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PCP}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PSNR}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Precision}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Psi}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{RI}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{RMSE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ROC}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Recall}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Run time}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SRC}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SSIM}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TN}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TP}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{VoI}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{nMAPE}} & \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Metrics used in \\\\textbf{other prediction} applications.}\\n\\\\label{other_metrics_table}\\n\\\\end{table*}\\n\\\\clearpage\\n\\\\clearpage\",\n \"id\": \"63a0cc61-98a6-4be7-b982-ebf42e368ec1\",\n \"level\": \"section\",\n \"origin_cites_number\": 219,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metrics and corresponding papers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Metrics and corresponding papers\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{appendix_metrics}\",\n \"id\": \"14c8fcc2-4517-4a60-9512-dfff6198b563\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metric formulas\"\n ]\n ],\n \"subsections\": [\n \"e9d8a36b-2d5e-4533-a6ab-b63d369319dc\",\n \"ddfb59d7-61bc-4e3f-8381-c962428d999f\",\n \"ed1fa5f0-0752-4fc1-89e1-31159f822bee\",\n \"327259be-0b0d-405f-9880-1b39cd8f50fa\"\n ],\n \"title\": \"Metric formulas\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"$$ MSE= \\\\frac{1}{MN} \\\\sum_{i=1}^M\\\\sum_{j=1}^N(I(i,j) -\\\\tilde{I}(i,j))^2$$\\n $$PSNR = 20\\\\log\\\\left(\\\\frac{MAX_I}{\\\\sqrt{MSE}}\\\\right)$$\\n\\\\textbf{Structural Similarity (SSIM)} \\n$$luminance (l)(x,y) = \\\\frac{2\\\\mu_x\\\\mu_y + C_1}{\\\\mu_x^2 + \\\\mu_y^2 + C_1}$$\\n$$\\\\mu_x=\\\\frac{1}{N}\\\\sum_{i=1}^Nx_i$$\\n$$C_1=(K_1L)^2$$\\n\\\\noindent where $L$ is dynamic range of pixel values (e.g. 255) and $K_1 \\\\ll 1$ is a small constant.\\n$$contrast (c)(x,y) = \\\\frac{2\\\\sigma_x\\\\sigma_y+C_2}{\\\\sigma_x^2+\\\\sigma_y^2+C_2}$$\\n$$\\\\sigma_x=\\\\left(\\\\frac{1}{N-1}\\\\sum_{i=1}^N(x_i-\\\\mu_x)^2\\\\right)^{1/2}$$\\n$$C_2=(K_2L)^2$$ $$K_2 \\\\ll 1$$\\n$$structure(s)(x,y) = \\\\frac{\\\\sigma_{xy}+C_3}{\\\\sigma_x\\\\sigma_y+C_3}$$\\n$$C_3=(K_3L)^2$$ $$K_3 \\\\ll 1$$\\n$$SSIM(x,y)=[l(x,y)]^\\\\alpha.[c(x,y)]^\\\\beta.[s(x,y)]^\\\\gamma$$\\n\\\\noindent where $\\\\alpha ,\\\\beta, \\\\gamma >0$ are parameters to choose in order to adjust the importance.\\n\\\\textbf{Learned Perceptual Image Patch Similarity (LPIPS)}\\nAssume features are extracted from $L$ layers and unit-normalized in channel dimension, for layer l\\n$$\\\\hat{y}^l,\\\\hat{y}^l \\\\in R^{H_l\\\\times W_l \\\\times C_l}$$.\\nThe distance between reference $x$ and distorted patches $x_0$ is given by,\\n$$d(x,x_0)=\\\\sum_l \\\\frac{1}{H_lW_l}\\\\sum_{w,l}\\\\parallel w_l \\\\odot (\\\\hat{y}^l_{hw},\\\\hat{y}^l_{0hw})\\\\parallel ^2_2$$\",\n \"id\": \"e9d8a36b-2d5e-4533-a6ab-b63d369319dc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"14c8fcc2-4517-4a60-9512-dfff6198b563\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metric formulas\"\n ],\n [\n \"subsection\",\n \"Video prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Video prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"There are 4 possibilities for classification: True positive (TP) and True Negative (TN) when the algorithm correctly classifies positive and negative samples, and False Positive (FP) and False Negative (FN) when the algorithm incorrectly classifies negative samples as positive and vice versa.\\n$$Accuracy = \\\\frac{TN+TP}{TP+TN+FP+FN}$$\\n$$Precision = \\\\frac{TP}{TP+FP}$$\\n$$Recall = \\\\frac{TP}{TP+FN}$$\\n$$F1-score = 2 \\\\times \\\\frac{Precision \\\\times Recall}{Precision + Recall}$$\\n$$RMSE = \\\\sqrt{\\\\frac{1}{n} \\\\sum_{i=1}^n(y_i -\\\\tilde{y_i})}$$\\nLet $p(r)$ be the precision-recall curve. Then,\\n$$AP = \\\\int_0^1 p(r)dr$$\",\n \"id\": \"ddfb59d7-61bc-4e3f-8381-c962428d999f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"14c8fcc2-4517-4a60-9512-dfff6198b563\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metric formulas\"\n ],\n [\n \"subsection\",\n \"Action prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Action prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"ed1fa5f0-0752-4fc1-89e1-31159f822bee\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"14c8fcc2-4517-4a60-9512-dfff6198b563\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metric formulas\"\n ],\n [\n \"subsection\",\n \"Trajectory prediction\"\n ]\n ],\n \"subsections\": [\n \"4a5b290c-66e4-4bd3-b5e3-68e6a0a15129\"\n ],\n \"title\": \"Trajectory prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"$$\\\\text{Euclidean Distance}(ED) = \\\\parallel y-\\\\tilde{y}\\\\parallel = \\\\parallel y-\\\\tilde{y}\\\\parallel_2$$\\n$$ = \\\\sqrt{\\\\sum_{i=1}^n(y_i -\\\\tilde{y_i})^2} $$\\n$$ \\\\text{Mean Absolute Error}(MAE) = \\\\frac{1}{n} \\\\sum_{i=1}^n|y_i -\\\\tilde{y_i}| $$\\n$$\\\\text{Mean Square Error} (MSE) = \\\\parallel y-\\\\tilde{y}\\\\parallel^2$$\\n$$ =\\\\frac{1}{n} \\\\sum_{i=1}^n(y_i -\\\\tilde{y_i})^2$$\\n$$\\\\text{Root MSE} (RMSE) = \\\\sqrt{\\\\frac{1}{n} \\\\sum_{i=1}^n(y_i -\\\\tilde{y_i})^2}$$\\n$$\\\\text{Hausdorff Distance}(HD) = max_{y\\\\in Y}min_{\\\\tilde{y}\\\\in\\\\tilde{Y}} \\\\parallel y-\\\\tilde{y}\\\\parallel$$\\n$$\\\\text{Modified HD}(MHD) = max(d(Y,\\\\tilde{Y}),d(\\\\tilde{Y}, Y))$$\\n$$ d(Y,\\\\tilde{Y}) = \\\\frac{1}{N_y}\\\\sum_{y\\\\in Y} min_{\\\\tilde{y}\\\\in \\\\tilde{Y}} \\\\parallel y - \\\\tilde{y}\\\\parallel$$\\n$$ ADE = \\\\frac{\\\\sum^N_{i=1}\\\\sum^{T_{pred}}_{t=1}\\\\parallel \\\\tilde{y}^i_t - y^i_t \\\\parallel}{N \\\\times T_{pred}}$$\\nwhere $N$ is the number of samples and $T_{pred}$ is the prediction steps.\\n$$FDE = \\\\frac{ \\\\sum^N_{i=1} \\\\parallel \\\\tilde{y}^i_{T_{pred}} - y^i_{T_{pred}} \\\\parallel}{N}$$\\n$$ minMSD = \\\\mathbb{E}_{\\\\tilde{Y}_k\\\\sim q_\\\\theta}min_{\\\\tilde{y}\\\\in\\\\tilde{Y}_k}\\\\parallel y-\\\\tilde{y}\\\\parallel^2$$\\nwhere $q_\\\\theta$ is the sampling space and $K$ number of samples. \\n$$ meanMSD = \\\\frac{1}{K}\\\\sum_{k=1}^K\\\\parallel y-\\\\tilde{y}\\\\parallel^2$$\\n$$NLL = \\\\mathbb{E}_{p(Y|X)}\\\\left[-\\\\log \\\\prod_{t=1}^{T_{pred}} p(y_t|X)\\\\right]$$\",\n \"id\": \"4a5b290c-66e4-4bd3-b5e3-68e6a0a15129\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"ed1fa5f0-0752-4fc1-89e1-31159f822bee\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metric formulas\"\n ],\n [\n \"subsection\",\n \"Trajectory prediction\"\n ],\n [\n \"subsubsection\",\n \"Distance metrics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Distance metrics\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"$$MPJPE = \\\\frac{1}{N \\\\times T_{pred}}\\\\sum_{t=1}^{T_{pred}}\\\\sum_{i=1}^N \\\\parallel (J_i^t-J_{root}^t)-(\\\\tilde{J}_i^t-\\\\tilde{J}_{root}^t)\\\\parallel$$\\n\\\\clearpage\",\n \"id\": \"327259be-0b0d-405f-9880-1b39cd8f50fa\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"14c8fcc2-4517-4a60-9512-dfff6198b563\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Metric formulas\"\n ],\n [\n \"subsection\",\n \"Motion prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Motion prediction\"\n },\n {\n \"cite_extract_rate\": 0.33628318584070804,\n \"cites\": [\n 2899,\n 2838,\n 2808,\n 2869,\n 2824,\n 2892,\n 8596,\n 2835,\n 2897,\n 1517,\n 2903,\n 8590,\n 2894,\n 6980,\n 2891,\n 2900,\n 2626,\n 2908,\n 2906,\n 2896,\n 2905,\n 2898,\n 2902,\n 8597,\n 1733,\n 2904,\n 2907,\n 2893,\n 2878,\n 2851,\n 7379,\n 2901,\n 2880,\n 2895\n ],\n \"content\": \"\\\\label{links_datasets}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\resizebox{1\\\\textwidth}{!}{\\n\\\\begin{tabular}{|l|l|l|}\\n\\\\hline\\nYear & Dataset & Links \\\\\\\\ \\\\hline\\n\\\\multirow{12}{*}{2019} & ARGOVerse & \\\\url{https://www.argoverse.org/data.html} \\\\\\\\ \\\\cline{2-3} \\n & CARLA & \\\\url{https://sites.google.com/view/precog} \\\\\\\\ \\\\cline{2-3} \\n & EgoPose & \\\\url{https://github.com/Khrylx/EgoPose} \\\\\\\\ \\\\cline{2-3} \\n & FM & \\\\url{https://mcl.korea.ac.kr/$\\\\sim$krkim/iccv2019/index.html} \\\\\\\\ \\\\cline{2-3} \\n & InstaVariety & \\\\url{https://github.com/akanazawa/human_dynamics} \\\\\\\\ \\\\cline{2-3} \\n & INTEARCTION & \\\\url{https://interaction-dataset.com} \\\\\\\\ \\\\cline{2-3} \\n & Luggage & \\\\url{https://aashi7.github.io/NearCollision.html} \\\\\\\\ \\\\cline{2-3} \\n & MGIF & \\\\url{https://github.com/AliaksandrSiarohin/monkey-net} \\\\\\\\ \\\\cline{2-3} \\n & PIE & \\\\url{http://data.nvision2.eecs.yorku.ca/PIE_dataset/} \\\\\\\\ \\\\cline{2-3} \\n & nuScenes & \\\\url{https://www.nuscenes.org/} \\\\\\\\ \\\\cline{2-3} \\n & Vehicle-Pedestrian-Mixed (VPM) & \\\\url{http://vr.ict.ac.cn/vp-lstm.} \\\\\\\\ \\\\cline{2-3} \\n & TRAF & \\\\url{https://drive.google.com/drive/folders/1LqzJuRkx5yhOcjWFORO5WZ97v6jg8RHN} \\\\\\\\ \\\\hline\\n\\\\multirow{10}{*}{2018} & 3DPW & \\\\url{https://virtualhumans.mpi-inf.mpg.de/3DPW/} \\\\\\\\ \\\\cline{2-3} \\n & ActEV/VIRAT & \\\\url{https://actev.nist.gov/trecvid19} \\\\\\\\ \\\\cline{2-3} \\n & ACTICIPATE & \\\\url{http://vislab.isr.tecnico.ulisboa.pt/datasets/} \\\\\\\\ \\\\cline{2-3} \\n & AVA & \\\\url{https://research.google.com/ava/} \\\\\\\\ \\\\cline{2-3} \\n & Epic-Kitchen & \\\\url{https://epic-kitchens.github.io/2019} \\\\\\\\ \\\\cline{2-3} \\n & EGTEA Gaze+ & \\\\url{http://www.cbi.gatech.edu/fpv/} \\\\\\\\ \\\\cline{2-3} \\n & STC & \\\\url{https://svip-lab.github.io/dataset/campus_dataset.html} \\\\\\\\ \\\\cline{2-3} \\n & ShapeStack & \\\\url{https://shapestacks.robots.ox.ac.uk/} \\\\\\\\ \\\\cline{2-3} \\n & VIENA & \\\\url{https://sites.google.com/view/viena2-project/home} \\\\\\\\ \\\\cline{2-3} \\n & YouCook2 & \\\\url{http://youcook2.eecs.umich.edu/} \\\\\\\\ \\\\hline\\n\\\\multirow{10}{*}{2017} & BUA & \\\\url{http://cs-people.bu.edu/sbargal/BU-action/} \\\\\\\\ \\\\cline{2-3} \\n & CityPerson & \\\\url{https://bitbucket.org/shanshanzhang/citypersons/src/default/} \\\\\\\\ \\\\cline{2-3} \\n & Epic-fail & \\\\url{http://aliensunmin.github.io/project/video-Forecasting/} \\\\\\\\ \\\\cline{2-3} \\n & JAAD & \\\\url{http://data.nvision2.eecs.yorku.ca/JAAD_dataset/} \\\\\\\\ \\\\cline{2-3} \\n & L-CAS & \\\\url{https://lcas.lincoln.ac.uk/wp/research/data-sets-software/l-cas-3d-point-cloud-people-dataset/} \\\\\\\\ \\\\cline{2-3} \\n & Mouse Fish & \\\\url{https://web.bii.a-star.edu.sg/archive/machine_learning/Projects/behaviorAnalysis/Lie-X/Lie-X.html} \\\\\\\\ \\\\cline{2-3} \\n & ORC & \\\\url{https://robotcar-dataset.robots.ox.ac.uk/} \\\\\\\\ \\\\cline{2-3} \\n & PKU-MMD & \\\\url{http://www.icst.pku.edu.cn/struct/Projects/PKUMMD.html} \\\\\\\\ \\\\cline{2-3} \\n & Recipe1M & \\\\url{http://pic2recipe.csail.mit.edu/} \\\\\\\\ \\\\cline{2-3} \\n & STRANDS & \\\\url{https://strands.readthedocs.io/en/latest/datasets/} \\\\\\\\ \\\\hline\\n\\\\multirow{13}{*}{2016} & BAIR Push & \\\\url{https://sites.google.com/site/brainrobotdata/home/push-dataset} \\\\\\\\ \\\\cline{2-3} \\n & BB & \\\\url{https://github.com/mbchang/dynamics} \\\\\\\\ \\\\cline{2-3} \\n & MU & \\\\url{http://staff.itee.uq.edu.au/lovell/MissUniverse/} \\\\\\\\ \\\\cline{2-3} \\n & Cityscapes & \\\\url{https://www.cityscapes-dataset.com/} \\\\\\\\ \\\\cline{2-3} \\n & CMU mocap & \\\\url{http://mocap.cs.cmu.edu/} \\\\\\\\ \\\\cline{2-3} \\n & DAD & \\\\url{https://aliensunmin.github.io/project/dashcam/} \\\\\\\\ \\\\cline{2-3} \\n & NTU RGB-D & \\\\url{http://rose1.ntu.edu.sg/Datasets/actionRecognition.asp} \\\\\\\\ \\\\cline{2-3} \\n & OA & \\\\url{http://www.mpii.de/ongoing-activity} \\\\\\\\ \\\\cline{2-3} \\n & OAD & \\\\url{http://www.icst.pku.edu.cn/struct/Projects/OAD.html} \\\\\\\\ \\\\cline{2-3} \\n & SD & \\\\url{http://cvgl.stanford.edu/projects/uav_data/} \\\\\\\\ \\\\cline{2-3} \\n & TV Series & \\\\url{https://github.com/zhenyangli/online_action} \\\\\\\\ \\\\cline{2-3} \\n & VIST & \\\\url{http://visionandlanguage.net/VIST/} \\\\\\\\ \\\\cline{2-3} \\n & Youtube-8M & \\\\url{https://research.google.com/youtube8m/} \\\\\\\\ \\\\hline\\n\\\\multirow{14}{*}{2015} & Amazon & \\\\url{http://jmcauley.ucsd.edu/data/amazon/index_2014.html} \\\\\\\\ \\\\cline{2-3} \\n & Atari & \\\\url{https://github.com/junhyukoh/nips2015-action-conditional-video-prediction} \\\\\\\\ \\\\cline{2-3} \\n & Brain4Cars & \\\\url{https://github.com/asheshjain399/ICCV2015_Brain4Cars} \\\\\\\\ \\\\cline{2-3} \\n & CMU Panoptic & \\\\url{http://domedb.perception.cs.cmu.edu/dataset.html} \\\\\\\\ \\\\cline{2-3} \\n & FPPA & \\\\url{http://bvision11.cs.unc.edu/bigpen/yipin/ICCV2015/prediction_webpage/Prediction.html} \\\\\\\\ \\\\cline{2-3} \\n & GTEA Gaze+ & \\\\url{http://www.cbi.gatech.edu/fpv/} \\\\\\\\ \\\\cline{2-3} \\n & MBI-1M & \\\\url{http://academic.mywebsiteontheinternet.com/data/} \\\\\\\\ \\\\cline{2-3} \\n & MOT & \\\\url{https://motchallenge.net/} \\\\\\\\ \\\\cline{2-3} \\n & MMNIST & \\\\url{http://www.cs.toronto.edu/$\\\\sim$nitish/unsupervised_video/} \\\\\\\\ \\\\cline{2-3} \\n & SUN RGB-D & \\\\url{http://rgbd.cs.princeton.edu/} \\\\\\\\ \\\\cline{2-3} \\n & SYSU 3DHOI & \\\\url{http://www.isee-ai.cn/$\\\\sim$hujianfang/ProjectJOULE.html} \\\\\\\\ \\\\cline{2-3} \\n & THUMOS & \\\\url{http://www.thumos.info/home.html} \\\\\\\\ \\\\cline{2-3} \\n & WnP & \\\\url{http://watchnpatch.cs.cornell.edu/} \\\\\\\\ \\\\cline{2-3} \\n & Wider & \\\\url{http://yjxiong.me/event_recog/WIDER/} \\\\\\\\ \\\\hline\\n\\\\multicolumn{1}{|c|}{\\\\multirow{5}{*}{2014}} & Breakfast & \\\\url{http://serre-lab.clps.brown.edu/resource/breakfast-actions-dataset/} \\\\\\\\ \\\\cline{2-3} \\n\\\\multicolumn{1}{|c|}{} & Human3.6M & \\\\url{http://vision.imar.ro/human3.6m/description.php} \\\\\\\\ \\\\cline{2-3} \\n\\\\multicolumn{1}{|c|}{} & MPII Human Pose & \\\\url{http://human-pose.mpi-inf.mpg.de/} \\\\\\\\ \\\\cline{2-3} \\n\\\\multicolumn{1}{|c|}{} & ORGBD & \\\\url{https://sites.google.com/site/skicyyu/orgbd} \\\\\\\\ \\\\cline{2-3} \\n\\\\multicolumn{1}{|c|}{} & Sports-1M & \\\\url{https://cs.stanford.edu/people/karpathy/deepvideo/} \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n}\\n\\\\caption{A summary of datasets (from year 2014-2019) used in vision-based prediction papers and corresponding links.}\\n\\\\label{data_set_links1}\\n\\\\end{table*}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\resizebox{1\\\\textwidth}{!}{\\n\\\\begin{tabular}{|l|l|l|}\\n\\\\hline\\nYear & Dataset & Links \\\\\\\\ \\\\hline\\n\\\\multirow{7}{*}{2013} & 50 salads & \\\\url{https://cvip.computing.dundee.ac.uk/datasets/foodpreparation/50salads/} \\\\\\\\ \\\\cline{2-3} \\n & ATC & \\\\url{https://irc.atr.jp/crest2010_HRI/ATC_dataset/} \\\\\\\\ \\\\cline{2-3} \\n & CAD-120 & \\\\url{http://pr.cs.cornell.edu/humanactivities/data.php} \\\\\\\\ \\\\cline{2-3} \\n & CHUK Avenue & \\\\url{http://www.cse.cuhk.edu.hk/leojia/projects/detectabnormal/dataset.html} \\\\\\\\ \\\\cline{2-3} \\n & Daimler path & \\\\url{http://www.gavrila.net/Datasets/Daimler_Pedestrian_Benchmark_D/Pedestrian_Path_Predict_GCPR_1/pedestrian_path_predict_gcpr_1.html} \\\\\\\\ \\\\cline{2-3} \\n & JHMDB & \\\\url{http://jhmdb.is.tue.mpg.de/} \\\\\\\\ \\\\cline{2-3} \\n & Penn Action & \\\\url{http://dreamdragon.github.io/PennAction/} \\\\\\\\ \\\\hline\\n\\\\multirow{11}{*}{2012} & BIT & \\\\url{https://sites.google.com/site/alexkongy/software} \\\\\\\\ \\\\cline{2-3} \\n & GTEA Gaze & \\\\url{http://www.cbi.gatech.edu/fpv/} \\\\\\\\ \\\\cline{2-3} \\n & KITTI & \\\\url{http://www.cvlibs.net/datasets/kitti/} \\\\\\\\ \\\\cline{2-3} \\n & MANIAC & \\\\url{https://alexandria.physik3.uni-goettingen.de/cns-group/datasets/maniac/} \\\\\\\\ \\\\cline{2-3} \\n & MPII-cooking & \\\\url{https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/research/human-activity-recognition/mpii-cooking-activities-dataset/} \\\\\\\\ \\\\cline{2-3} \\n & MSRDA & \\\\url{https://documents.uow.edu.au/$\\\\sim$wanqing/\\\\#MSRAction3DDatasets} \\\\\\\\ \\\\cline{2-3} \\n & GC & \\\\url{http://www.ee.cuhk.edu.hk/$\\\\sim$xgwang/grandcentral.html} \\\\\\\\ \\\\cline{2-3} \\n & SBUKI & \\\\url{https://www3.cs.stonybrook.edu/$\\\\sim$kyun/research/kinect_interaction/index.html} \\\\\\\\ \\\\cline{2-3} \\n & UCF-101 & \\\\url{https://www.crcv.ucf.edu/data/UCF101.php} \\\\\\\\ \\\\cline{2-3} \\n & UTKA & \\\\url{http://cvrc.ece.utexas.edu/KinectDatasets/HOJ3D.html} \\\\\\\\ \\\\cline{2-3} \\n & UvA-NEMO & \\\\url{https://www.uva-nemo.org/} \\\\\\\\ \\\\hline\\n\\\\multirow{5}{*}{2011} & FCVL & \\\\url{http://robots.engin.umich.edu/SoftwareData/Ford} \\\\\\\\ \\\\cline{2-3} \\n & HMDB & \\\\url{http://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/} \\\\\\\\ \\\\cline{2-3} \\n & Stanford 40 & \\\\url{http://vision.stanford.edu/Datasets/40actions.html} \\\\\\\\ \\\\cline{2-3} \\n & Town Center & \\\\url{http://www.robots.ox.ac.uk/ActiveVision/Research/Projects/2009bbenfold_headpose/project.html\\\\#datasets} \\\\\\\\ \\\\cline{2-3} \\n & VIRAT & \\\\url{http://viratdata.org/} \\\\\\\\ \\\\hline\\n\\\\multirow{8}{*}{2010} & DISPLECS & \\\\url{https://cvssp.org/data/diplecs/} \\\\\\\\ \\\\cline{2-3} \\n & MSR & \\\\url{https://www.microsoft.com/en-us/download/details.aspx?id=52315} \\\\\\\\ \\\\cline{2-3} \\n & MUG Facial Expression & \\\\url{https://mug.ee.auth.gr/fed/} \\\\\\\\ \\\\cline{2-3} \\n & PROST & \\\\url{www.gpu4vision.com } \\\\\\\\ \\\\cline{2-3} \\n & THI & \\\\url{http://www.robots.ox.ac.uk/$\\\\sim$alonso/tv_human_interactions.html} \\\\\\\\ \\\\cline{2-3} \\n & UTI & \\\\url{http://cvrc.ece.utexas.edu/SDHA2010/Human_Interaction.html} \\\\\\\\ \\\\cline{2-3} \\n & VISOR & \\\\url{imagelab.ing.unimore.it/visor} \\\\\\\\ \\\\cline{2-3} \\n & Willow Action & \\\\url{https://www.di.ens.fr/willow/research/stillactions/} \\\\\\\\ \\\\hline\\n\\\\multirow{9}{*}{2009} & Caltech Pedestrian & \\\\url{http://www.vision.caltech.edu/Image_Datasets/CaltechPedestrians/} \\\\\\\\ \\\\cline{2-3} \\n & Collective Activity (CA) & \\\\url{http://www-personal.umich.edu/$\\\\sim$wgchoi/eccv12/wongun_eccv12.html} \\\\\\\\ \\\\cline{2-3} \\n & EIFP & \\\\url{http://homepages.inf.ed.ac.uk/rbf/FORUMTRACKING/} \\\\\\\\ \\\\cline{2-3} \\n & ETH & \\\\url{http://www.vision.ee.ethz.ch/en/datasets/} \\\\\\\\ \\\\cline{2-3} \\n & OSU & \\\\url{http://eecs.oregonstate.edu/football/tracking/dataset} \\\\\\\\ \\\\cline{2-3} \\n & PETS2009 & \\\\url{http://www.cvg.reading.ac.uk/PETS2009/a.html} \\\\\\\\ \\\\cline{2-3} \\n & QMUL & \\\\url{http://personal.ie.cuhk.edu.hk/$\\\\sim$ccloy/downloads_qmul_junction.html} \\\\\\\\ \\\\cline{2-3} \\n & TUM Kitchen & \\\\url{https://ias.in.tum.de/dokuwiki/software/kitchen-activity-data} \\\\\\\\ \\\\cline{2-3} \\n & YUV Videos & \\\\url{http://trace.kom.aau.dk/yuv/index.html} \\\\\\\\ \\\\hline\\n\\\\multirow{2}{*}{2008} & Daimler & \\\\url{http://www.gavrila.net/Datasets/Daimler_Pedestrian_Benchmark_D/Daimler_Mono_Ped__Detection_Be/daimler_mono_ped__detection_be.html} \\\\\\\\ \\\\cline{2-3} \\n & MITT & \\\\url{http://www.ee.cuhk.edu.hk/$\\\\sim$xgwang/MITtrajsingle.html} \\\\\\\\ \\\\hline\\n\\\\multirow{5}{*}{2007} & AMOS & \\\\url{http://amos.cse.wustl.edu/} \\\\\\\\ \\\\cline{2-3} \\n & ETH pedestrian & \\\\url{https://data.vision.ee.ethz.ch/cvl/aess/} \\\\\\\\ \\\\cline{2-3} \\n & Lankershim Boulevard & \\\\url{https://www.fhwa.dot.gov/publications/research/operations/07029/index.cfm} \\\\\\\\ \\\\cline{2-3} \\n & NGSIM & \\\\url{https://ops.fhwa.dot.gov/trafficanalysistools/ngsim.htm} \\\\\\\\ \\\\cline{2-3} \\n & UCY & \\\\url{https://graphics.cs.ucy.ac.cy/research/downloads/crowd-data} \\\\\\\\ \\\\hline\\n2006 & Tuscan Arizona & \\\\url{http://www.mmto.org/} \\\\\\\\ \\\\hline\\n2004 & KTH & \\\\url{http://www.nada.kth.se/cvap/actions/} \\\\\\\\ \\\\hline\\n1981 & Golden Colorado & \\\\url{https://www.osti.gov/dataexplorer/biblio/dataset/1052221} \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n}\\n\\\\caption{A summary of datasets (from year 2013 and earlier) used in vision-based prediction papers and corresponding links.}\\n\\\\label{data_set_links2}\\n\\\\end{table*}\\nLists of datasets with associated repository links can be found in Tables \\\\ref{data_set_links1} and \\\\ref{data_set_links2}.\\n\\\\clearpage\",\n \"id\": \"e82c2ffc-f5dc-4d01-9184-399825652bbe\",\n \"level\": \"section\",\n \"origin_cites_number\": 113,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Links to the datasets\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Links to the datasets\"\n },\n {\n \"cite_extract_rate\": 0.505813953488372,\n \"cites\": [\n 2829,\n 2871,\n 2844,\n 2838,\n 2808,\n 2807,\n 2869,\n 2841,\n 2822,\n 979,\n 2840,\n 2824,\n 2811,\n 2845,\n 2850,\n 2835,\n 2873,\n 2870,\n 2827,\n 2842,\n 2836,\n 2887,\n 9096,\n 8590,\n 2849,\n 2877,\n 2813,\n 2814,\n 7117,\n 2853,\n 2879,\n 2816,\n 2856,\n 2866,\n 2888,\n 2809,\n 2858,\n 2802,\n 2805,\n 2826,\n 2833,\n 2876,\n 7618,\n 2839,\n 2843,\n 2847,\n 2854,\n 8592,\n 2868,\n 2832,\n 2874,\n 2828,\n 2806,\n 8595,\n 2834,\n 2859,\n 2875,\n 2881,\n 2812,\n 2886,\n 2878,\n 2851,\n 2863,\n 7617,\n 7379,\n 2803,\n 2890,\n 8594,\n 2804,\n 2821,\n 2837,\n 2867,\n 2817,\n 2880,\n 2885,\n 2819,\n 2846,\n 2815,\n 988,\n 2818,\n 2820,\n 2810,\n 2831,\n 2884,\n 2861,\n 2872,\n 2864\n ],\n \"content\": \"\\\\label{papers_datasets}\\nLists of datasets and corresponding papers can be found in Tables \\\\ref{video_datasets_table}, \\\\ref{action_datasets_table}, \\\\ref{trajectory_datasets_table}, \\\\ref{motion_datasets_table}, and \\\\ref{other_datasets_table}. \\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\begin{tabular}{|l|l|}\\n\\\\hline\\n\\\\textbf{Datasets} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Atari}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{BAIR Push}} & ,, \\n \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Bouncing Ball}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CHUK Avenue}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Caltech Pedestrian}} & , , ,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Cityscapes}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Human 3.6M}} & ,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{JAAD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{JHMDB}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{KITTI}} & ,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{KTH}} & ,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MGIF}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MMNIST}} & ,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MSR}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MUG}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Own}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PROST}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Penn Action}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Penn action}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ShanghaiTech Campus}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ShapeStack}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Sports-1M}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{THUMOS}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{UCF-101}} & ,,,,,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{UvA-NEMO}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ViSOR}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{YUV}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Youtube-8M}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{pedestrian}} & \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Datasets used in \\\\textbf{video prediction} applications.}\\n\\\\label{video_datasets_table}\\n\\\\end{table*}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\begin{tabular}{|l|l|}\\n\\\\hline\\n\\\\textbf{Datasets} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{50Salad}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ANTICIPATE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{AVA}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ActEV/VIRAT}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{BIT}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{BU Action}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Brain4Cars}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Breakfast}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CAD-120}} & ,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CMU Panoptic}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CMU Mocap}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Caltech Pedestrian}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{DAD}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Daimler}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Daimler Path}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{EGTEA Gaze+}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ETH Pedestrian}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Epic-fail}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Epic-Kitchen}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{FPPA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{GTEA Gaze}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{GTEA Gaze+}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{HMDB}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Human 3.6M}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{JAAD}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{JHMDB}} & ,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Luggage}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MANIAC}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MPII Cooking}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MSRDA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{NGSIM}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{NTU RGB-D}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{OA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{OAD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ORGBD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PIE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PKU-MMD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Recipe1M}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SBUIK}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SYSU 3DHOI}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Stanford-40}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{THUMOS}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TV Human Interaction}} & , ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TV Series}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{UCF-101}} & ,,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{UTI}} & ,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{UTKA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{VIENA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{VIRAT}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{WIDER}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Willow Action}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{WnP}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{YouCook2}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Sports-1M}} & \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Datasets used in \\\\textbf{action prediction} applications.}\\n\\\\label{action_datasets_table}\\n\\\\end{table*}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\begin{tabular}{|l|p{0.8\\\\textwidth}|}\\n\\\\hline\\n\\\\textbf{Datasets} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ARGOVerse}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ATC}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ActEV/VIRAT}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CARLA}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CHUK}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CityPerson}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Cityscapes}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Daimler Path}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{ETH}} & ,,,,,,,,,,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Edinburgh (IFP)}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{FM}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{GC}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{INTEARCTION}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{JAAD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{KITTI}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{L-CAS}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Lankershim Boulevard}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MITT}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MOT}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{NGSIM}} & ,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{OSU}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Oxford}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PETS2009}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{PIE}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{QMUL}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SBUIK}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SD}} & ,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{STRANDS}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TRAF}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{TUM Kitchen}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Town Center}} & ,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{UCY}} & ,,,,,,,,,,,,,,,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{VIRAT}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{VPM}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{nuScenes}} & \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Datasets used in \\\\textbf{trajectory prediction} applications.}\\n\\\\label{trajectory_datasets_table}\\n\\\\end{table*}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\begin{tabular}{|l|l|}\\n\\\\hline\\n\\\\textbf{Datasets} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{3DPW}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CA}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CMU Mocap}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{CMU Panoptic}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Egopose}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Human 3.6M}} & ,,,,,,,,,,,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{InstaVariety}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MPII Human Pose}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Mouse Fish}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Own}} & ,,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Penn Action}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SBUIK}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{UCF-101}} & \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Datasets used in \\\\textbf{motion prediction} applications.}\\n\\\\label{motion_datasets_table}\\n\\\\end{table*}\\n\\\\begin{table*}[]\\n\\\\centering\\n\\\\begin{tabular}{|l|l|}\\n\\\\hline\\n\\\\textbf{Datasets} & \\\\textbf{Papers} \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{AMOS}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Amazon}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Cityscapes}} & ,,, \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{DIPLECS}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{FCVL}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Golden Colorado}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{JAAD}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{KITTI}} & , \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MBI-1M}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{MU}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{SUN RGB-D}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{Tuscan Arizona}} & \\\\\\\\ \\\\hline\\n\\\\textit{\\\\textbf{VIST}} & \\\\\\\\ \\\\hline\\n\\\\end{tabular}\\n\\\\caption{Datasets used in \\\\textbf{other prediction} applications.}\\n\\\\label{other_datasets_table}\\n\\\\end{table*}\\n\\\\end{document}\",\n \"id\": \"a33fb5e9-81a5-4a37-8f37-6dd6031d77ae\",\n \"level\": \"section\",\n \"origin_cites_number\": 172,\n \"parent_id\": \"54da09bf-16a7-43ca-8a25-4aafacb8f53e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Deep Learning for Vision-based Prediction: A Survey\"\n ],\n [\n \"section\",\n \"Datasets and corresponding papers\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Datasets and corresponding papers\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":97,"string":"97"},"all_cites":{"kind":"list like","value":[2804,2814,2817,2807,2808,2816,5680,2819,2815,2806,2809,2818,2811,2802,2820,2810,2805,2812,8590,7617,2803,2813,2073,2821,2823,2822,2824,2829,2837,7117,2838,2832,2828,2825,2834,2835,2827,2826,2831,2836,2833,7618,2839,2830,2841,2845,2844,2840,2843,2842,2867,2855,2847,2860,2854,2853,2868,2857,2869,2856,2846,2862,2852,2866,979,988,2858,2859,2848,2850,2865,2851,8591,1261,2861,2863,2849,7379,2864,2871,8592,2874,2875,2873,2870,2878,2876,2877,2872,2882,2879,2880,2885,2888,2883,2881,2887,2886,9096,2884,8593,2889,7619,6980,117,7620,491,2890,8594,8595,2899,2892,8596,2897,1517,2903,2894,2891,2900,2626,2908,2906,2896,2905,2898,2902,8597,1733,2904,2907,2893,2901,2895],"string":"[\n 2804,\n 2814,\n 2817,\n 2807,\n 2808,\n 2816,\n 5680,\n 2819,\n 2815,\n 2806,\n 2809,\n 2818,\n 2811,\n 2802,\n 2820,\n 2810,\n 2805,\n 2812,\n 8590,\n 7617,\n 2803,\n 2813,\n 2073,\n 2821,\n 2823,\n 2822,\n 2824,\n 2829,\n 2837,\n 7117,\n 2838,\n 2832,\n 2828,\n 2825,\n 2834,\n 2835,\n 2827,\n 2826,\n 2831,\n 2836,\n 2833,\n 7618,\n 2839,\n 2830,\n 2841,\n 2845,\n 2844,\n 2840,\n 2843,\n 2842,\n 2867,\n 2855,\n 2847,\n 2860,\n 2854,\n 2853,\n 2868,\n 2857,\n 2869,\n 2856,\n 2846,\n 2862,\n 2852,\n 2866,\n 979,\n 988,\n 2858,\n 2859,\n 2848,\n 2850,\n 2865,\n 2851,\n 8591,\n 1261,\n 2861,\n 2863,\n 2849,\n 7379,\n 2864,\n 2871,\n 8592,\n 2874,\n 2875,\n 2873,\n 2870,\n 2878,\n 2876,\n 2877,\n 2872,\n 2882,\n 2879,\n 2880,\n 2885,\n 2888,\n 2883,\n 2881,\n 2887,\n 2886,\n 9096,\n 2884,\n 8593,\n 2889,\n 7619,\n 6980,\n 117,\n 7620,\n 491,\n 2890,\n 8594,\n 8595,\n 2899,\n 2892,\n 8596,\n 2897,\n 1517,\n 2903,\n 2894,\n 2891,\n 2900,\n 2626,\n 2908,\n 2906,\n 2896,\n 2905,\n 2898,\n 2902,\n 8597,\n 1733,\n 2904,\n 2907,\n 2893,\n 2901,\n 2895\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":0.8649678061957291,"string":"0.864968"}}},{"rowIdx":65,"cells":{"authors":{"kind":"list like","value":["Nan Wu","Yuan Xie"],"string":"[\n \"Nan Wu\",\n \"Yuan Xie\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"A Survey of Machine Learning for Computer Architecture and Systems"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-02-16T04:09:57Z","string":"2021-02-16T04:09:57Z"},"category":{"kind":"string","value":"cs.LG"},"abstract":{"kind":"string","value":" It has been a long time that computer architecture and systems are optimized for efficient execution of machine learning (ML) models. Now, it is time to reconsider the relationship between ML and systems, and let ML transform the way that computer architecture and systems are designed. This embraces a twofold meaning: improvement of designers' productivity, and completion of the virtuous cycle. In this paper, we present a comprehensive review of the work that applies ML for computer architecture and system design. First, we perform a high-level taxonomy by considering the typical role that ML techniques take in architecture/system design, i.e., either for fast predictive modeling or as the design methodology. Then, we summarize the common problems in computer architecture/system design that can be solved by ML techniques, and the typical ML techniques employed to resolve each of them. In addition to emphasis on computer architecture in a narrow sense, we adopt the concept that data centers can be recognized as warehouse-scale computers; sketchy discussions are provided in adjacent computer systems, such as code generation and compiler; we also give attention to how ML techniques can aid and transform design automation. We further provide a future vision of opportunities and potential directions, and envision that applying ML for computer architecture and systems would thrive in the community. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"5631fef3-6389-407d-9101-745f45240141","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"]],"subsections":["5359b569-700e-4932-aef8-e103109cdfb1","06a8a978-bb13-405e-8614-2ae9a01d5c56","15e5a658-0b25-4890-8335-c1ec062ad3a2","87e1e057-9e86-4a5c-afa3-431c75b16a64","a80b2264-c6ef-4ec9-ae9a-838d036e281c","79dc37a5-e0c0-498b-b911-e6bc545700a1"],"title":"root"},{"cite_extract_rate":0,"cites":[],"content":"Machine learning (ML) has been doing wonders in many fields.\nAs people are seeking better artificial intelligence (AI), there is a trend towards larger, more expressive, and more complex models.\nAccording to the data reported by OpenAI , from 1959 to 2012, the amount of compute used in the largest AI training runs doubles every two years; since 2012, deep learning starts taking off, and the required amount of compute has been increasing exponentially with a 3.4-month doubling period. By comparison, Moore’s law , the principle that has powered the integrated-circuit revolution since 1960s, doubles the transistor density every 18 months.\nWhile Moore's law is approaching its end , more pressure is put on innovations of computer architecture and systems, so as to keep up with the compute demand of AI applications.\nConventionally, computer architecture/system designs are made by human experts based on intuitions and heuristics, which requires expertise in both ML and architecture/system.\nMeanwhile, these heuristic-based designs can not guarantee scalability and optimality, especially in the case of increasingly complicated systems.\nAs such, it seems natural to move towards more automated and powerful methodologies for computer architecture and system design, and the relationship between ML and system design is being reconsidered.\nOver the past decade, architecture and systems are optimized to accelerate the execution and improve the performance of ML models.\nRecently, there have been signs of emergence of applying ML for computer architecture and systems, which embraces a twofold meaning: \\textcircled{\\small{1}} the reduction of burdens on human experts designing systems manually, so as to improve designers' productivity, and \\textcircled{\\small{2}} the close of the positive feedback loop, i.e., architecture/systems for ML and simultaneously ML for architecture/systems, formulating a virtuous cycle to encourage improvements on both sides.\n\\begin{figure}[tbp]\n \\centering\n \\vspace{-12pt}\n\t\\includegraphics[width=0.95\\linewidth]{fig0.png}\n\t\\vspace{-8pt}\n \\caption{A comprehensive overview of applying ML for computer architecture and systems. Existing work roughly falls into two categories: ML for fast system modeling, and ML as design methodology.}\n \\label{fig:frame}\n \\vspace{-20pt}\n\\end{figure}\nExisting work related to applying ML for computer architecture and system design falls into two categories.\n\\textcircled{\\small{1}} ML techniques are employed for \\textbf{fast and accurate system modeling}, which involves performance metrics or some criteria of interest (e.g. power consumption, latency, throughput, etc.).\nDuring the process of designing systems, it is necessary to make fast and accurate predictions of system behaviors.\nTraditionally, system modeling is achieved through the forms of cycle-accurate or functional virtual platforms, and instruction set simulators (e.g. gem5 ).\nEven though these methods provide accurate estimations, they bring expensive computation costs associated with performance modeling, which limits the scalability to large-scale and complex systems; meanwhile, the long simulation time often dominates design iteration, making it impossible to fully explore the design space.\nBy contrast, ML-based modeling and performance prediction are capable to balance simulation cost and prediction accuracy.\n\\textcircled{\\small{2}} ML techniques are employed as \\textbf{a design methodology to directly enhance architecture/system design}.\nML techniques are skilled at extracting features that might be implicit to human experts, making decisions without explicit programming, and improving themselves automatically with accumulated experience.\nTherefore, applying ML techniques as design tools can explore design space proactively and intelligently, and manage resource through better understanding of the complicated and non-linear interactions between workloads and systems, making it possible to deliver truly optimal solutions.\nIn this paper, we present a comprehensive overview of applying ML for computer architecture and systems. \nAs depicted in Figure \\ref{fig:frame},\nwe first perform a high-level taxonomy by considering the typical role that ML techniques take in architecture/system design, i.e., either for fast predictive modeling or as the design methodology;\nthen, we summarize the common problems in architecture/system design that can be solved by ML techniques, and the typical ML techniques employed to resolve each of them.\nIn addition to emphasis on computer architecture in a narrow sense, we adopt the concept that data centers can be recognized as warehouse-scale computers , and review studies associated with data center management; we provide sketchy discussions on adjacent computer systems, such as code generation and compiler; we also give attention to how ML techniques can aid and transform design automation that involves both analog and digital circuits.\nAt the end of the paper, we discuss challenges and future prospects of applying ML for architecture/system design, aiming to convey insights of design considerations.","id":"5359b569-700e-4932-aef8-e103109cdfb1","level":"section","origin_cites_number":5,"parent_id":"5631fef3-6389-407d-9101-745f45240141","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0.210526315789473,"cites":[1390,553,2219,166],"content":"\\label{sec:ml techniques}\nThere are three general frameworks in ML: supervised learning, unsupervised learning and reinforcement learning.\nThese frameworks mainly differentiate on what data are sampled and how these sample data are used to build learning models.\nTable \\ref{table:ml_techniques} summarizes the commonly used ML techniques for computer architecture and system designs.\nSometimes, multiple learning models may work well for one given problem, and the appropriate selection can be made based on available hardware resource and data, implementation overheads, performance targets, etc.\n\\begin{table}[tbp]\n\\vspace{-12pt}\n\\caption{Machine learning techniques.}\n\\vspace{-12pt}\n\\label{table:ml_techniques}\n\\centering\n \\tiny\n \\renewcommand{\\arraystretch}{1}\n \\setlength{\\tabcolsep}{6pt}\n\\begin{tabular}{c|c|c|c}\n\\toprule\n\\textbf{Realm of ML} & \\textbf{Category} & \\makecell[c]{\\textbf{Classical ML}} & \\textbf{Deep Learning Counterpart } \\\\ \\midrule\n\\multirow{10}{*}{\\makecell*[c]{\\textbf{Supervised} \\\\ \\textbf{Learning}}} & \\multirow{1}{*}{Classification} &\n Logistic regression & \n \\multirow{8}{*}{\\makecell{CNN, RNN,\\\\ GNN , etc.}} \\\\ \\cline{2-3}\n & \\multirow{6}{*}{Working for both} & Support vector machines/regression & \\\\ \\cline{3-3}\n && K-nearest neighbors & \\\\ \\cline{3-3}\n && Decision tree, e.g., CART , MARS & \\\\ \\cline{3-3}\n && ANN & \\\\ \\cline{3-3}\n && Bayesian analysis & \\\\ \\cline{3-3}\n && \\makecell{Ensemble learning , e.g., gradient boosting, random forest} & \\\\ \\cline{2-3}\n & \\multirow{2}{*}{Regression} & \\makecell{Linear regression with variants , e.g., lasso (L1 regularization),\\\\ ridge (L2 regularization), elastic-net (hybrid L1/L2 regularization)} & \\\\ \\cline{3-3}\n && Non-linear regression & \\\\\n \\hline\n\\multirow{2}{*}{\\makecell[c]{\\textbf{Unsupervised} \\\\ \\textbf{Learning}}} \n& Clustering & K-means clustering & \\multirow{2}{*}{\\makecell[c]{Autoencoder, \\\\ GAN, etc.}} \\\\ \\cline{2-3}\n & Dimension reduction & Principal component analysis (PCA) &\\\\ \\hline\n\\multirow{3}{*}{\\makecell*[c]{\\textbf{Reinforcement} \\\\ \\textbf{Learning}}} &\n Value-based & Q-learning & DQN \\\\ \n \\cline{2-4} \n & \\multirow{2}{*}{Policy-based} & Actor-critic \n & A3C , DDPG \\\\ \\cline{3-4}\n & & Policy gradient, e.g., REINFORCE \n & PPO \\\\\n\\bottomrule\n\\end{tabular}\n\\vspace{-15pt}\n\\end{table}\n\\vspace{-5pt}","id":"06a8a978-bb13-405e-8614-2ae9a01d5c56","level":"section","origin_cites_number":19,"parent_id":"5631fef3-6389-407d-9101-745f45240141","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Different ML Techniques"]],"subsections":["f49480df-db26-480b-8149-671064eda13f","c3e4e81a-5169-40c5-91a3-4a2389c943b1","218b240c-27bf-4571-953a-c14101e51384"],"title":"Different ML Techniques"},{"cite_extract_rate":0.16666666666666602,"cites":[166],"content":"Supervised learning is the process of learning a set of rules able to map an input to an output based on labeled datasets. These learned rules can be generalized to make predictions for unseen inputs.\nWe briefly introduce several prevalent techniques in supervised learning, as shown in Figure \\ref{fig:sl}.\n\\begin{itemize}[leftmargin=*]\n \\item Regression is a process for estimating the relationships between a dependent variable and one or more independent variables. The most common form is linear regression , and some other forms include different types of non-linear regression . Regression techniques are primarily used for two purposes, prediction/forecasting, and inference of causal relationships.\n \\item Support vector machines (SVMs) try to find the best hyperplanes to separate data classes by maximizing margins. One variant is support vector regression (SVR), which is able to conduct regression tasks. Predictions or classifications of new inputs can be decided by their relative positions to these hyperplanes.\n \\item Decision tree is one representative of logical learning methods, which uses tree structures to build regression or classification models.\n The final result is a tree with decision nodes and leaf nodes. Each decision node represents a feature and branches of this node represent possible values of the corresponding feature. Starting from the root node, input instances are classified by sequentially passing through nodes and branches, until they reach leaf nodes that represent either classification results or numerical values.\n \\item Artificial neural networks (ANNs) are capable to approximate a broad family of functions: a single-layer perceptron is usually used for linear regression; complex DNNs consisting of multiple layers are able to approximate non-linear functions, such as the multi-layer perceptron (MLP); variants of DNNs that achieve excellent performance in specific fields benefit from the exploitation of certain computation operations, e.g., convolutional neural networks (CNNs) with convolution operations leveraging spatial features, and recurrent neural networks (RNNs) with recurrent connections enabling learning from sequences and histories.\n \\item Ensemble learning employs multiple models that are strategically designed to solve a particular problem, and the primary goal is to achieve better predictive performance than those could be obtained from any of the constituent models alone. Several common types of ensembles include random forest and gradient boosting.\n\\end{itemize}\n\\begin{figure}[tbp]\n \\centering\n \\vspace{-15pt}\n\t\\includegraphics[width=0.97\\linewidth]{fig1.png}\n\t\\vspace{-10pt}\n \\caption{Examples of supervised learning: (a) regression, (b) SVM, (c) decision tree, (d) MLP, and (e) ensemble learning.}\n \\vspace{-20pt}\n \\label{fig:sl}\n\\end{figure}\nDifferent learning models have different preference of input features: \nSVMs and ANNs generally perform much better with multi-dimension and continuous features, while logic-based systems tend to perform better when dealing with discrete/categorical features.\nIn system design, supervised learning is commonly used for performance modeling, configuration predictions, or predicting higher-level features/behaviors from lower-level features.\nOne thing worth noting is that supervised learning techniques need well labeled training data prior to the training phase, which usually require tremendous human expertise and engineering.\n\\vspace{-5pt}","id":"f49480df-db26-480b-8149-671064eda13f","level":"subsection","origin_cites_number":6,"parent_id":"06a8a978-bb13-405e-8614-2ae9a01d5c56","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Different ML Techniques"],["subsection","Supervised Learning"]],"subsections":[],"title":"Supervised Learning"},{"cite_extract_rate":0,"cites":[],"content":"Unsupervised learning is the process of finding previously unknown patterns based on unlabeled datasets.\nTwo prevailing methods are clustering analysis and principal component analysis (PCA) , as depicted in Figure \\ref{fig:usl}.\n\\begin{itemize}[leftmargin=*]\n \\item Clustering is a process of grouping data objects into disjoint clusters based on a measure of similarity, such that data objects in the same cluster are similar while data objects in different clusters share low similarities. The goal of clustering is to classify raw data reasonably and to find possibly existing hidden structures or patterns in datasets. One of the most popular and simple clustering algorithms is k-means clustering.\n \\item PCA is essentially a coordinate transformation leveraging information from data statistics. It aims to reduce the dimensionality of the high-dimensional variable space by representing it with a few orthogonal (linearly uncorrelated) variables that capture most of its variability.\n\\end{itemize}\nSince there is no label in unsupervised learning, it is difficult to simultaneously measure the performance of learning models and decide when to stop the learning process.\nOne potential workaround is \\textit{semi-supervised learning} , which uses a small amount of labeled data together with a large amount of unlabeled data. This approach stands between unsupervised and supervised learning, requiring less human effort and producing higher accuracy.\nThe unlabeled data are used to either finetune or re-prioritize hypotheses obtained from labeled data alone.\n\\begin{figure}[t]\n\\vspace{-15pt}\n\\flushleft\n\\begin{minipage}{.63\\textwidth}\n \\centering\n \\includegraphics[width=0.82\\linewidth]{fig2.png}\n \\vspace{-10pt}\n \\caption{Examples of unsupervised learning: (a) clustering, and (b) PCA.}\n \\label{fig:usl}\n\\end{minipage}\n\\hspace{.03\\textwidth}\n\\begin{minipage}{.33\\textwidth}\n \\centering\n \\includegraphics[width=\\linewidth]{fig3.png}\n \\caption{A typical framing of RL.}\n \\label{fig:rl}\n\\end{minipage}\n\\vspace{-10pt}\n\\end{figure}","id":"c3e4e81a-5169-40c5-91a3-4a2389c943b1","level":"subsection","origin_cites_number":3,"parent_id":"06a8a978-bb13-405e-8614-2ae9a01d5c56","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Different ML Techniques"],["subsection","Unsupervised Learning"]],"subsections":[],"title":"Unsupervised Learning"},{"cite_extract_rate":0,"cites":[],"content":"In standard reinforcement learning (RL) , an agent interacts with an environment $\\mathcal{E}$ over a number of discrete time steps, as shown in Figure \\ref{fig:rl}.\nAt each time step $t$, the agent receives a state $s_t$ from the \\textit{state space} $\\mathcal{S}$, and selects an action $a_t$ from the \\textit{action space} $\\mathcal{A}$ according to its policy $\\pi$, where $\\pi$ is a mapping from states $s_t$ to actions $a_t$. \nIn return, the agent receives the next state $s_{t+1}$ and a scalar reward $r_t: \\mathcal{S}\\times\\mathcal{A}\\to\\mathbb{R}$. \nThis process continues until the agent reaches a terminal state after which the process restarts.\nThe return $R_t={\\sum\\limits_{k=0}^\\infty \\gamma^kr_{t+k}}$ is the totally accumulated rewards at the time step $t$ with a \\textit{discount factor} $\\gamma\\in(0,1]$. \nThe goal of the agent is to maximize the expected return for each state $s$.\nThe state-action value $Q_{\\pi}(s,a)=\\mathbb{E_\\pi}\\lbrack R_t|s_t=s, a_t=a\\rbrack$ is the expected return of selecting action $a$ at state $s$ with policy $\\pi$.\nSimilarly, the state value $V_{\\pi}(s)=\\mathbb{E_\\pi}\\lbrack R_t|s_t=s\\rbrack$ is the expected return starting from state $s$ by following policy $\\pi$. There are two general types of methods in RL: value-based, and policy-based.\n\\begin{itemize}[leftmargin=*]\n \\item In value-based RL, the state-action value function $Q_{\\pi}(s,a)$ is approximated by either tabular approaches or function approximations. At each state $s_t$, the agent always selects the optimal action $a^*_t$ that could bring the maximal state-action value $Q_{\\pi}(s_t,a^*_t)$. One well-known example of value-based methods is Q-learning.\n \\item In policy-based RL, it directly parameterizes the policy $\\pi(a|s;\\theta)$ and updates the parameters $\\theta$ by performing gradient ascent on $\\mathbb{E}[R_t]$. One example is the REINFORCE algorithm. \n\\end{itemize}\nRL is modeled based on Markov decision process, and thus it is suitable to handle control problems or sequential decision-making processes.\nWith these characteristics, RL is able to explore design space proactively and intelligently, and learn how to achieve resource management or task scheduling in system designs through interactions with environments. The optimal behaviors can be found by embedding optimization goals into reward functions.","id":"218b240c-27bf-4571-953a-c14101e51384","level":"subsection","origin_cites_number":1,"parent_id":"06a8a978-bb13-405e-8614-2ae9a01d5c56","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Different ML Techniques"],["subsection","Reinforcement Learning"]],"subsections":[],"title":"Reinforcement Learning"},{"cite_extract_rate":0.04918032786885201,"cites":[4378,5390,5389],"content":"\\label{sec:system modeling}\nThis section reviews studies that employ ML techniques for fast and accurate system modeling, which involves predictions of performance metrics or some other criteria of interest.\nAlthough cycle-accurate simulators, which are commonly used for system performance prediction, can provide accurate estimations, they usually run multiple orders of magnitude slower than native executions.\nBy contrast, ML-based techniques can balance simulation costs and prediction accuracy, showing great potentials in exploring huge configuration spaces and learning non-linear impacts of configurations.\nMost of existing work applies supervised learning for either pure system modeling or efficient design space exploration (DSE) enabled by fast predictions.\nTable \\ref{table:model_sys} and Table \\ref{table:model_eda} summarize the studies for predictive modeling in computer architecture/system and design automation respectively, in terms of task domains, prediction targets, adopted ML techniques, and corresponding inputs.\n\\begin{table}[tbp]\n\\vspace{-10pt}\n\\caption{Summary of applying ML techniques for fast modeling in computer architecture and systems.}\n\\label{table:model_sys}\n\\centering\n \\tiny\n \\renewcommand{\\arraystretch}{0.9}\n \\setlength{\\tabcolsep}{1pt}\n\\vspace{-10pt}\n\\begin{tabular}{c|c|c|c}\n\\toprule\n\\textbf{Domain} & \\textbf{Prediction of} & \\textbf{Technique} & \\textbf{Input} \\\\ \\midrule\n\\multirow{4}{*}{\\makecell[c]{\\textbf{Memory} \\\\ \\textbf{System} \\\\ $( \\S$ \\textbf{\\ref{sec:model_memory}}$)$}} \n& Throughput/cache miss & ANN & Cache configurations \\\\ \\cline{2-4} \n& Throughput/lifetime/energy & \\makecell[c]{Gradient boosting/quadratic regression with lasso } & NVM configurations \\\\ \\cline{2-4} \n& Throughput & CNN & Memory controller placements \\\\ \\cline{2-4} \n& Disk block correlation & DNN & Data blocks in the context window \\\\ \\hline\n\\multirow{4}{*}{\\makecell[c]{\\textbf{NoC} \\\\ $( \\S$ \\textbf{\\ref{sec:model_noc}}$)$}} \n& Latency & SVR & Queue information \\\\ \\cline{2-4} \n& Hotspots & ANN & Buffer utilization \\\\ \\cline{2-4} \n& Buffer/link utilization & \\makecell[c]{Regression with ridge regularization , decision tree } & NoC configurations \\\\ \\cline{2-4} \n& Probability of errors & Decision tree & \\makecell[c]{Link utilization and transistor wearing-out} \\\\ \\hline\n\\multirow{6}{*}{\\makecell[c]{\\textbf{GPU} \\\\ $( \\S$ \\textbf{\\ref{sec:model_gpu}}$)$}} & \\makecell[c]{Speedup or execution time \\\\ by cross-platform inputs} & \\makecell[c]{Nearest neighbor/SVM , \\\\ ensemble of regression-based learners , \\\\ random forest } & \\makecell[c]{Static analysis and/or \\\\ dynamic profiling of \\\\ source CPU code} \\\\ \\cline{2-4} \n & Execution time & \\makecell[c]{Stepwise regression , \\\\ ensemble of linear regression and random forest } & \\multirow{3}{*}{\\makecell[c]{GPU configurations \\\\ and performance counters}} \\\\ \\cline{2-3}\n & Throughput/power & Ensemble of NNs & \\\\ \\cline{2-3}\n & Scaling behavior of GPGPU & ANN and K-means clustering & \\\\ \\cline{2-4} \n & Kernel affinity and execution time & Logistic regression and linear regression & Kernel characteristics \\\\ \\cline{2-4}\n & Traffic patterns in GPGPU & CNN & Grayscale heat maps \\\\ \\hline\n\\multirow{2}{*}{\\makecell[c]{\\textbf{Single-Core} \\\\ \\textbf{CPU} $( \\S$ \\textbf{\\ref{sec:model_single}}$)$}} \n& Throughput & \\makecell[c]{Linear regression , non-linear regression } & \\multirow{2}{*}{\\makecell[c]{Micro-architectural parameters \\\\ and performance counters}} \\\\ \\cline{2-3}\n & Program execution time & Linear regression & \\\\ \\hline\n\\multirow{8}{*}{\\makecell[c]{\\textbf{General}\\\\ \\textbf{Modeling} \\\\ $( \\S$ \\textbf{\\ref{sec:model_general}}$)$}} & \\multirow{6}{*}{Throughput/latency/power} & ANN & \\multirow{7}{*}{\\makecell[c]{Micro-architectural parameters \\\\ and performance counters}} \\\\ \\cline{3-3}\n & & \\makecell[c]{Non-linear regression } & \\\\ \\cline{3-3}\n & & Linear regression & \\\\ \\cline{3-3}\n & & Hierarchical Bayesian model & \\\\ \\cline{3-3}\n & & LSTM & \\\\ \\cline{3-3}\n & & Generative model & \\\\ \\cline{2-3}\n & \\makecell[c]{Slowdown caused by \\\\ application interference} & \\makecell[c]{Linear regression with elastic-net \\\\ regularization } & \\\\ \\cline{2-4} \n & \\makecell[c]{Speedup of multi-thread applications} & Gausian process regression & Profiling of single-thread execution \\\\ \\hline\n\\multirow{9}{*}{\\makecell[c]{\\textbf{Data} \\\\ \\textbf{Center} \\\\ $( \\S$ \\textbf{\\ref{sec:model_dc}}$)$}} & Job completion time & SVR & \\makecell[c]{Application characteristics \\\\ and cluster configurations} \\\\ \\cline{2-4} \n & Resource demand & \\makecell[c]{Statistical learning , linear regression/MLP } & \\multirow{3}{*}{Workload characterization} \\\\ \\cline{2-3}\n & Incoming workload & ARMA , ARIMA & \\\\ \\cline{2-3}\n & Workload pattern & Hidden Markov model & \\\\ \\cline{2-4} \n & Power usage effectiveness & MLP & Data center configurations \\\\ \\cline{2-4} \n & Disk Failure & \\makecell[c]{Bayesian methods , clustering , \\\\ SVM/MLP , random forest } & \\multirow{3}{*}{\\makecell[c]{SMART (Self-Monitoring, Analysis \\\\ and Reporting Technology) attributes \\\\ of data centers}} \\\\ \\cline{2-3}\n & Health assessment of drives & \\makecell[c]{CART , gradient boosted regressions tree , RNN } & \\\\ \\cline{2-3}\n & \\multirow{2}{*}{Partial drive failure} & \\makecell[c]{CART/random forest/SVM/ANN/logistic regression } & \\\\ \\cline{3-4} \n & & Gradient boosted regression trees & \\makecell[c]{SMART attributes and system-level signals} \\\\ \\bottomrule\n\\end{tabular}\n\\vspace{-15pt}\n\\end{table}","id":"15e5a658-0b25-4890-8335-c1ec062ad3a2","level":"section","origin_cites_number":61,"parent_id":"5631fef3-6389-407d-9101-745f45240141","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"]],"subsections":["cb797a5f-f6ce-4412-9791-ccdeb4f88886","9ac2d4c9-6bb8-44b3-aa23-38c568e05513","419e1fc4-5eb3-40c4-b251-cc170597bc2e"],"title":"ML for Fast System Modeling"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:model_sub_sys}","id":"cb797a5f-f6ce-4412-9791-ccdeb4f88886","level":"subsection","origin_cites_number":0,"parent_id":"15e5a658-0b25-4890-8335-c1ec062ad3a2","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","Sub-system Modeling and Performance Prediction"]],"subsections":["1e898be7-db20-4db0-b960-280b242ecf62","5ac93425-5d57-405e-be1e-d3cab4e56f18"],"title":"Sub-system Modeling and Performance Prediction"},{"cite_extract_rate":0.2,"cites":[5391],"content":"\\label{sec:model_memory}\nIn memory systems, ML-based performance models are exploited to help explore trade-offs among different objectives.\nTo explore non-volatile memory (NVM) based cache hierarchies, Dong \\textit{et al.} develop an ANN model to predict higher-level features (e.g. miss of cache read/write, and instruction-per-cycle (IPC)) from lower-level features (e.g. cache associativity, capacity and latency). \nTo adaptively select architectural techniques in NVMs for different applications, Memory Cocktail Therapy estimates lifetime, IPC, and energy consumption through lightweight online predictors by gradient boosting and quadratic regression with lasso. \nTo optimize memory controller placements in throughput processors, Lin \\textit{et al.} build a CNN model that takes memory controller placements as inputs to predict throughput, which accelerates the optimization process by two orders of magnitude.\nSome studies concentrate on learning efficient representations of memory access patterns.\nBlock2Vec tries to mine data block correlations by training a DNN to learn the best vector representation of each block and capturing block similarities via vector distances, which enables further optimization for caching and prefetching.\nShi \\textit{et al.} use a graph neural network (GNN) to learn fused representations of static code and its dynamic execution. This unified representation is capable to model both data flows (e.g., prefetching) and control flows (e.g., branch prediction).","id":"1e898be7-db20-4db0-b960-280b242ecf62","level":"subsubsection","origin_cites_number":5,"parent_id":"cb797a5f-f6ce-4412-9791-ccdeb4f88886","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","Sub-system Modeling and Performance Prediction"],["subsubsection","Memory System"]],"subsections":[],"title":"Memory System"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:model_noc}\nIn NoCs, several performance metrics of interest are latency, energy consumption, and reliability.\n\\textcircled{\\small{1}}\nRegarding latency predictions, Qian \\textit{et al.} use an SVR model to predict the traffic flow latency and the average channel waiting time in mesh-based NoCs, which relaxes some assumptions in the classical queuing theory. \nRather than explicitly predicting latency, a lightweight hardware-based ANN predicts existence of traffic hotspots, which are intensive network congestions significantly degrading the effective throughput and implicitly indicate the average communication latency in NoCs.\nThe input features are buffer utilization rates from neighboring NoC routers, and the trained predictor is combined with a proactive hotspot-preventive routing algorithm to avert hotspot formation, attaining significant improvements for synthetic workloads while modest melioration for real-world benchmarks.\n\\textcircled{\\small{2}}\nRegarding estimating energy consumption, the learned predictors are often leveraged for saving dynamic and/or static energy in NoCs.\nDiTomaso \\textit{et al.} use per-router decision trees to predict link utilization and traffic direction, which are combined with sleepy link storage units to power-gate links/routers and to change link directions. \nClark \\textit{et al.} use ridge regression models to predict buffer utilization, changes in buffer utilization, or a combined metric of energy and throughput,\nbased on which a router can select proper voltage/frequency.\nIn photonic NoCs, the ridge regression model is also applicable to predict the number of packets to be injected into each router in the following time window , based on which the number of wavelengths are scaled properly to reduce static energy consumed by photonic links.\n\\textcircled{\\small{3}}\nRegarding the reliability of NoCs, a per-link decision tree trained offline can predict the probability of timing faults on links during runtime , based on which a proactive fault-tolerant technique is developed to mitigate errors by using the strengthened cyclic redundancy check with error-correction code and relaxed transmission.","id":"5ac93425-5d57-405e-be1e-d3cab4e56f18","level":"subsubsection","origin_cites_number":6,"parent_id":"cb797a5f-f6ce-4412-9791-ccdeb4f88886","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","Sub-system Modeling and Performance Prediction"],["subsubsection","Network-on-Chip (NoC)"]],"subsections":[],"title":"Network-on-Chip (NoC)"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:model_sys}\nAccurate and fast performance estimation is a necessity for system optimization and design space exploration.\nWith the increasing complexity of systems and variety of workloads, ML-based techniques can provide highly accurate performance estimations with reasonable simulation costs, surpassing the capability of commonly-used cycle-accurate simulators that require highly computational costs and long simulation time.","id":"9ac2d4c9-6bb8-44b3-aa23-38c568e05513","level":"subsection","origin_cites_number":0,"parent_id":"15e5a658-0b25-4890-8335-c1ec062ad3a2","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","System Modeling and Performance Prediction"]],"subsections":["a2990a76-751b-4214-b5a9-77c66ab0dd6a","a74c5980-8f81-4447-be80-f6c683740060","3890381b-cc70-4717-9bbf-fa3c92dca1c8","37bd7027-5f70-4a17-9138-7b53f9c12df8"],"title":"System Modeling and Performance Prediction"},{"cite_extract_rate":0.11111111111111101,"cites":[5390],"content":"\\label{sec:model_gpu}\nThere are two types of predictions for GPU modeling: cross-platform predictions and GPU-specific predictions.\nCross-platform predictions are used to decide in advance whether to offload an application from a CPU to a GPU, since not every application benefits from GPU execution and the porting process requires considerably additional efforts;\nGPU-specific predictions are used to estimate metrics of interest and to assist GPU design space exploration, helpful to handle design space irregularities and complicated interactions among configurations.\nCross-platform predictions can be formulated as a binary classification problem that identifies whether the potential GPU speedup of an application would be greater than a given threshold.\nThis task can be solved by the nearest neighbor and SVMs using dynamic instruction profiles , or a random forest that composes of one thousand decision trees using static analysis of source CPU code (i.e., memory coalescing, branch divergence, kernel size available parallelism and instruction intensities) .\nWith both dynamic and static program properties from single-thread CPU code, an ensemble of one hundred regression-based learners can predict the GPU execution time .\nIn terms of GPU-specific predictions that take GPU configurations and performance counters as input features, the execution time can be predicted by stepwise linear regression, which recognizes the most important input features among many GPU parameters and thus achieves high accuracy even with sparse samples ; the power/throughput can be modeled by an ensemble of NN predictors .\nProvided with profiling results from earlier-generation GPUs, an ensemble of linear and non-linear regression models is capable to predict cross-generation GPU execution time for later/future-generation GPUs, which achieves more than 10,000 times speedup compared to cycle-accurate GPU simulators .\nFocusing on processing-in-memory (PIM) assisted GPU architectures, Pattnaik \\textit{et al.} classify GPU cores into two types: powerful GPU cores but far away from memory, and auxiliary/simple GPU cores but close to memory. They develop a logistic regression model that takes kernel characteristics as input features to predict architecture affinity of kernels, aiming to accurately identify which kernels would benefit from PIM and offload them accordingly to auxiliary GPU cores. They also build a linear regression model to predict the execution time of each kernel, so that a concurrent kernel management mechanism can be developed based on these two models and kernel dependency information. \nFocusing on general-purpose GPUs (GPGPUs), Wu \\textit{et al.} model kernel scaling behaviors with respect to the number of compute units, engine frequency, and memory frequency. During training, kernels with similar performance scaling behaviors are grouped by K-means clustering, and when encountering a new kernel, it is mapped to the cluster that best describes its scaling performance by an ANN-based classifier.\nLi \\textit{et al.} reassess prevailing assumptions of GPGPU traffic patterns, and combine a CNN with a t-distributed stochastic neighbor embedding to classify different traffic patterns.","id":"a2990a76-751b-4214-b5a9-77c66ab0dd6a","level":"subsubsection","origin_cites_number":9,"parent_id":"9ac2d4c9-6bb8-44b3-aa23-38c568e05513","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","System Modeling and Performance Prediction"],["subsubsection","Graphics Processing Unit (GPU)"]],"subsections":[],"title":"Graphics Processing Unit (GPU)"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:model_single}\nIn predictive performance modeling of single-core processors, early-stage work mostly targets superscalar processors. \nTo predict the application-specific cycle-per-instruction (CPI) of superscalar processors, Joseph \\textit{et al.} introduce an iterative procedure to build linear regression models using 26 key micro-architectural parameters. Later they construct predictive models by non-linear regression techniques (i.e., radial basis function networks generated from regression trees) with 9 key micro-architectural parameters .\nIn parallel with Joseph's work, Lee and Brooks use regression modeling with cubic splines to predict application-specific performance (billions of instructions per second) and power.\nLater work focuses on performance modeling for existing hardware (e.g., Intel, AMD, and ARM processors) by using micro-architectural parameters and performance counters.\nEyerman \\textit{et al.} construct a mechanistic-empirical model for CPI predictions of three Intel processors. The initially parameterized performance model is inspired by mechanistic modeling, where the unknown parameters inside the model are derived through regression, benefiting from both mechanistic modeling (i.e., interpretability) and empirical modeling (i.e., ease of implementation).\nZheng \\textit{et al.} explore two approaches to cross-platform predictions of program execution time, where profiling results on Intel Core i7 and AMD Phenom processors are used to estimate the execution time on a target ARM processor. The first approach relaxes the assumption of global linearity to local linearity in the feature space and applies constrained locally sparse linear regression; the other approach applies lasso linear regression with phase-level performance features.","id":"a74c5980-8f81-4447-be80-f6c683740060","level":"subsubsection","origin_cites_number":6,"parent_id":"9ac2d4c9-6bb8-44b3-aa23-38c568e05513","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","System Modeling and Performance Prediction"],["subsubsection","Single-Core Processor"]],"subsections":[],"title":"Single-Core Processor"},{"cite_extract_rate":0.13636363636363602,"cites":[4378,5392,5389],"content":"\\label{sec:model_general}\nRegression techniques are the mainstream to predict performance metrics from micro-architectural parameters or other features, which attributes to their capability to make high-accuracy estimations with reasonable training costs.\nFor conventional regression-based models, ANNs and non-linear regression with different designs are the common practice to predict throughput/latency and power/energy .\nConsequently, there are comparisons among different techniques.\nLee \\textit{et al.} compare piecewise polynomial regression with ANNs, with emphasis that piecewise polynomial regression offers better explainability while ANNs show better generalization ability.\nOzisikyilmaz \\textit{et al.} contrast several linear regression models and different ANNs, indicating that the pruned ANNs achieve best accuracy while requiring longer training time.\nAgarwal \\textit{et al.} estimate the parallel execution speedup of multi-threaded applications on a target hardware platform, and mention that Gaussian process regression performs the best among several explored methods in this case.\nMore recent work tends to take advantage of data-driven approaches.\nIthemal leverages a hierarchical multi-scale RNN with long short term memory (LSTM) to predict throughput of basic blocks (i.e., sequences of instructions with no branches or jumps), and evaluations demonstrate that Ithemal is more accurate and as fast as analytical throughput estimators.\nBy employing a variant of Ithemal as a differentiable surrogate to approximate CPU simulators, DiffTune is able to apply gradient-based optimization techniques to learn the parameters of x86 basic block CPU simulators such that simulators' error is minimized. The learned parameters finally are plugged back into the original simulator.\nDing \\textit{et al.} provide some insights in learning-based modeling methods: the improvement of prediction accuracy may receive diminishing returns; the consideration of domain knowledge will be helpful for system optimizations, even if the overall accuracy may not be improved.\nThus, they propose a generative model to handle data scarcity by generating more training data, and apply a multi-phase sampling to improve prediction accuracy of optimal configuration points.\nML-based predictive performance modeling enables efficient resource management and rapid design space exploration to improve throughput.\nEquipped with ANNs for IPC predictions, strategies for resource allocation and task scheduling can always select decisions that would bring the best predicted IPC.\nESP constructs a regression model with elastic-net regularization to predict application interference (i.e., slowdown), which is integrated with schedulers to increase throughput.\nMetaTune is a meta-learning based cost model for convolution operations, and when combined with search algorithms, it enables efficient auto-tuning of parameters during compilation.\nIn consideration of rapid design space exploration of the uncore (i.e., memory hierarchies and NoCs), Sangaiah \\textit{et al.} uses a regression-based model with restricted cubic splines to estimate CPI, reducing the exploration time by up to four orders of magnitude.\nML-based predictive performance modeling benefits adaptations between performance and power budgets.\nLeveraging off-line multivariate linear regression to predict IPC and/or power of different architecture configurations, Curtis-Maury \\textit{et al.} maximize performance of OpenMP applications by dynamic concurrency throttling and dynamic voltage and frequency scaling (DVFS); Bailey \\textit{et al.} apply hardware frequency-limiting techniques to select optimal hardware configurations under given power constraints.\nTo effectively apply DVFS towards various optimization goals, the designed strategy can adopt predictions for power consumption by a constrained-posynomial model or job execution time by a linear regression model .\nTo conduct smart power management in a more general manner, LEO employs hierarchical Bayesian models to predict performance and power, and when integrated for runtime energy optimization, it is capable to figure out the performance-power Pareto frontier and select the configuration satisfying performance constraints with minimized energy.\nCALOREE~ further breaks up the power management task into two abstractions: a learner for performance modeling and an adaptive controller leveraging predictions from the learner. These abstractions enable both the learner to use multiple ML techniques and the controller to maintain control-theoretic formal guarantees. Since no user-specified parameter except the goal is required, CALOREE is applicable even for non-experts.","id":"3890381b-cc70-4717-9bbf-fa3c92dca1c8","level":"subsubsection","origin_cites_number":22,"parent_id":"9ac2d4c9-6bb8-44b3-aa23-38c568e05513","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","System Modeling and Performance Prediction"],["subsubsection","General Modeling and Performance Prediction"]],"subsections":[],"title":"General Modeling and Performance Prediction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:model_dc}\nData centers have been in widespread use for both traditional enterprise applications and cloud services.\nMany studies employ ML techniques to predict workload/resource-related metrics, so as to enable elastic resource provision.\nCommon examples include but not limited to using SVR to predict job completion time , leveraging the autoregressive moving average (ARMA) model or the autoregressive integrated moving average (ARIMA) model to forecast incoming workloads, exploiting the hidden Markov modeling to characterize variations in workload patterns , and estimating dynamic resource demand of workloads by light-weight statistical learning algorithms or MLP .\nJim Gao builds an MLP model to predict power usage effectiveness of data centers, which is extensively tested and validated at Google data centers.\nCortez \\textit{et al.} predict virtual machine (VM) behaviors (including VM lifetimes, maximum deployment sizes, and workload classes) for a broader set of purposes (e.g., health/resource management and power capping), where the evaluated ML models are random forests and extreme gradient boosting trees.\nIn addition to workload/resource-related metrics, the availability in data centers or cloud services is also a topic of concern, where one of the key tasks is to predict disk failure in advance.\nLeveraging SMART (Self-Monitoring, Analysis and Reporting Technology) attributes, the disk failure prediction model can be built via various ML techniques, such as different Bayesian methods , unsupervised clustering , SVM and MLP .\nThe adoption of classification and regression trees (CART) , RNNs , or gradient boosted regression trees makes it possible to assess health status of drives.\nWhile all the aforementioned methods rely on offline training, online random forests can evolve with forthcoming data on-the-fly by generating new trees and forget old information by discarding outdated trees, consequently avoiding the model aging problem in disk failure predictions.\nTo predict partial drive failures (i.e., disk error or sector error), Mahdisoltani \\textit{et al.} explore five ML techniques (CART, random forests, SVM, NN and logistic regression), among which random forests consistently outperform others.\nXu \\textit{et al.} incorporate SMART attributes and system-level signals to train a gradient boosted regression tree, which is an online prediction model that ranks disks according to the degree of error-proneness in the near future.","id":"37bd7027-5f70-4a17-9138-7b53f9c12df8","level":"subsubsection","origin_cites_number":17,"parent_id":"9ac2d4c9-6bb8-44b3-aa23-38c568e05513","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","System Modeling and Performance Prediction"],["subsubsection","Data Center Performance Modeling and Prediction"]],"subsections":[],"title":"Data Center Performance Modeling and Prediction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:model_eda}","id":"419e1fc4-5eb3-40c4-b251-cc170597bc2e","level":"subsection","origin_cites_number":0,"parent_id":"15e5a658-0b25-4890-8335-c1ec062ad3a2","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","Performance Modeling in Chip Design and Design Automation"]],"subsections":["df0787ef-e9e6-48be-837b-abd7e6a37852","3e9b9f12-6464-4e9c-b39f-515e81853fdb","8ba558a9-ad3c-4f4c-a999-80fc17f11ebf"],"title":"Performance Modeling in Chip Design and Design Automation"},{"cite_extract_rate":0.15625,"cites":[5393,5395,5396,8916,5394],"content":"\\label{sec:model_analog}\nAnalog circuit design is usually a manual process that requires many trial-and-error iterations between pre-layout and post-layout phases.\nIn recent years, the discrepancy between schematic (i.e., pre-layout) performance estimations and post-layout simulation results is further enlarged.\nOn the one hand, the analytical performance estimations from schematics are no longer accurate with device scaling; on the other hand, even though post-layout simulations can provide high-accuracy estimations, they are extremely time-consuming and have become the major bottleneck of design iteration time.\nTo shrink the gap in performance modeling of integrated circuits (ICs), ML techniques are widely applied for fast circuit evaluation.\nWe discuss the studies based on whether their input features are extracted from pre-layout or post-layout information.\n\\textcircled{\\small{1}}\nGiven design schematics, parasitics in layouts can be predicted from pre-layout stage, which helps bridge the gap of performance difference between pre-layout and post-layout simulations.\nParaGraph builds a GNN model to predict layout-dependent parasitics and physical device parameters.\nMLParest shows that non-graph based methods (e.g., random forest) also work well for estimating interconnect parasitics, whereas the lack of placement information may cause large variations in predictions.\n\\textcircled{\\small{2}}\nGiven circuit schematics as well as device information as inputs, it is possible to directly model post-layout performance from pre-layout designs.\nAlawieh \\textit{et al.} propose a hierarchical method that combines the Bayesian co-learning framework and semi-supervised learning to predict power consumption.\nThe entire circuit schematic is partitioned into multiple blocks to build block-level performance models, upon which circuit-level performance models are built. By combining these two low-dimensional models with a large amount of unlabeled data, pseudo samples can be labeled with almost no cost. Finally, a high-dimensional performance model mapping low-level features to circuit-level metrics is trained with pseudo samples and a small amount of labeled samples, which demonstrates the feasibility of performance modeling with inadequate labeled samples.\nSeveral variants of Bayesian-based methods also perform well for estimating post-layout performance, e.g., combining Bayesian regression with SVM to predict circuit performance and using Bayesian DNNs to compare circuit designs .\n\\textcircled{\\small{3}}\nSince post-layout simulations with SPICE-like simulators is time-consuming, ML techniques are applied to quickly assess layout design performance .\nTo make better use of structural information inside layouts, intermediate layout placement results are represented as 3D images to feed a 3D CNN model , or encoded as graphs to train a customized GNN model , with the goal to predict whether a design specification is satisfied.\n\\begin{table}[tbp]\n\\vspace{-10pt}\n\\caption{Summary of applying ML techniques for performance modeling and prediction in design automation.}\n\\vspace{-10pt}\n\\label{table:model_eda}\n\\centering\n \\tiny\n \\renewcommand{\\arraystretch}{1}\n \\setlength{\\tabcolsep}{0.5pt}\n\\begin{tabular}{c|c|c|c}\n\\toprule\n\\textbf{Domain} & \\textbf{Prediction of} & \\textbf{Technique} & \\textbf{Input} \\\\ \\midrule\n\\multirow{5}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Analog \\\\ Circuit \\\\ $( \\S$ \\ref{sec:model_analog}$)$ \\end{tabular}}} & Parasitics & GNN , random forest & Circuit schematics \\\\ \\cline{2-4} \n & Power/area/bandwidth & \\begin{tabular}[c]{@{}c@{}}Bayesian co-learning and semi-supervised learning , \\\\ Bayesian regression and SVM \\end{tabular} & \\multirow{2}{*}{\\begin{tabular}[c]{@{}c@{}}Circuit schematics \\\\ and device information\\end{tabular}} \\\\ \\cline{2-3}\n & \\begin{tabular}[c]{@{}c@{}}Probability of superiority between designs\\end{tabular} & Bayesian DNN & \\\\ \\cline{2-4} \n & \\begin{tabular}[c]{@{}c@{}}Gain/unity gain frequency/bandwidth/phase margin\\end{tabular} & \\begin{tabular}[c]{@{}c@{}}SVM/ANN/random forest , 3D CNN , GNN \\end{tabular} & \\multirow{2}{*}{Circuit placement} \\\\ \\cline{2-3}\n & Electromagnetic properties & GNN & \\\\ \\hline\n\\multirow{9}{*}{\\makecell[c]{\\textbf{HLS} \\\\ $( \\S$ \\textbf{\\ref{sec:model_hls}}$)$}} & \\begin{tabular}[c]{@{}c@{}}Area/latency/throughput/logic utilization\\end{tabular} & \\begin{tabular}[c]{@{}c@{}}Random forest , transfer learning \\end{tabular} & Directives in HLS scripts \\\\ \\cline{2-4} \n & Resource utilization & ANN & \\multirow{3}{*}{IR graphs from HLS front-ends} \\\\ \\cline{2-3}\n & Resource mapping and clustering & GraphSAGE & \\\\ \\cline{2-3}\n & Routing congestion & \\begin{tabular}[c]{@{}c@{}}Linear regression/ANN/gradient boosted regression tree \\end{tabular} & \\\\ \\cline{2-4} \n & Power & \\begin{tabular}[c]{@{}c@{}}Linear regression/SVM/tree-based models/DNNs \\end{tabular} & IR graphs and HLS reports \\\\ \\cline{2-4} \n & \\begin{tabular}[c]{@{}c@{}}Throughput and throughput-to-area ratio\\end{tabular} & \\begin{tabular}[c]{@{}c@{}}Ensemble learning by stacked regression \\end{tabular} & \\multirow{2}{*}{HLS reports} \\\\ \\cline{2-3}\n & Resource utilization and timing & \\begin{tabular}[c]{@{}c@{}}Linear regression/ANN/gradient tree boosting \\end{tabular} & \\\\ \\cline{2-4} \n & Cross-platform latency and power & Random forest & CPU program counters \\\\ \\cline{2-4} \n & Speedup over an ARM processor & ANN & \\begin{tabular}[c]{@{}c@{}}Application characteristics, \\\\ HLS reports, FPGA configurations\\end{tabular} \\\\ \\hline\n\\multirow{8}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Logic and \\\\ Physical \\\\ Synthesis \\\\ $( \\S$ \\ref{sec:model_logic}$)$\\end{tabular}}} & Area/delay & CNN , LSTM & Synthesis flows \\\\ \\cline{2-4} \n & \\multirow{2}{*}{DRVs} & \\begin{tabular}[c]{@{}c@{}}Linear regression/ANN/decision tree , MARS \\end{tabular} & \\begin{tabular}[c]{@{}c@{}}Placement and GR information\\end{tabular} \\\\ \\cline{3-4} \n & & MARS/SVM , MLP & Placement \\\\ \\cline{2-4} \n & \\multirow{2}{*}{DRC hotspots} & FCN & \\begin{tabular}[c]{@{}c@{}}Placement and GR information\\end{tabular} \\\\ \\cline{3-4} \n & & Variant of FCN & \\multirow{3}{*}{Placement} \\\\ \\cline{2-3} \n & GR congestion map & FCN & \\\\ \\cline{2-3} \n & \\multirow{2}{*}{Routing congestion in FPGAs} & Linear regression & \\\\ \\cline{3-4} \n & & Conditional GAN & Post-placement images \\\\ \\bottomrule\n\\end{tabular}\n\\vspace{-15pt}\n\\end{table}","id":"df0787ef-e9e6-48be-837b-abd7e6a37852","level":"subsubsection","origin_cites_number":32,"parent_id":"419e1fc4-5eb3-40c4-b251-cc170597bc2e","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","Performance Modeling in Chip Design and Design Automation"],["subsubsection","Analog Circuit Analysis"]],"subsections":[],"title":"Analog Circuit Analysis"},{"cite_extract_rate":0.18181818181818102,"cites":[5393,5394],"content":"\\label{sec:model_hls}\nHLS is an automated transformation from behavioral languages (e.g., C/C++/SystemC) to register-transfer level (RTL) designs, which significantly expedites the development of hardware designs involving with field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs).\nSince HLS tools usually take considerable time to synthesize each design, it prevents designers from exploring design space sufficiently, which motivates the application of ML models for fast and accurate performance estimation.\nIn performance estimation of HLS designs, the input features to ML models are extracted from three major sources: HLS directives, IRs from HLS front-ends, and HLS reports.\n\\textcircled{\\small{1}} Taking the directives in an HLS script as input features, random forest is capable to forecast different design metrics, such as area and effective latency , and throughput and logic utilization .\nIn order to reuse knowledge from previous experiences, a transfer learning approach can transfer knowledge across different applications or synthesis options.\n\\textcircled{\\small{2}} Taking advantages of IR graphs generated by HLS front-ends, Koeplinger \\textit{et al.} count resource requirements of each node in graphs by using pre-characterized area models, which are then used as inputs to ANNs to predict the LUT routing usage, register duplication, and unavailable LUTs.\nThe exploitation of GNNs makes it possible to automatically predict the mapping from arithmetic operations in IR graphs to different resources on FPGAs . \nTo forecast post-implementation routing congestion, Zhao \\textit{et al.} build a dataset that connects the routing congestion metrics after RTL implementation with operations in IRs, with the goal to train ML models locating highly congested regions in source code.\n\\textcircled{\\small{3}} Taking the information that can be directly extracted from HLS reports, Dai \\textit{et al.} try several ML models (linear regression, ANN, and gradient tree boosting) to predict post-implementation resource utilization and timing. \nPyramid applies the ensemble learning by stacked regression to accurately estimate the throughput or the throughput-to-area ratio.\nHL-Pow employs features from both IR graphs and HLS reports to predict the power by a variety of ML models. \nThe surge of heterogeneous platforms with FPGA/AISC and CPU provides more possibility of hardware/software co-design, motivating cross-platform performance predictions.\nHLSPredict uses random forest to predict FPGA cycle counts and power consumption based on program counter measurements obtained from CPU execution. \nWhile HLSPredict targets the same FPGA platform in training and testing, XPPE considers different FPGA platforms, and uses ANNs to predict the speedup of an application on a target FPGA over an ARM processor.","id":"3e9b9f12-6464-4e9c-b39f-515e81853fdb","level":"subsubsection","origin_cites_number":11,"parent_id":"419e1fc4-5eb3-40c4-b251-cc170597bc2e","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","Performance Modeling in Chip Design and Design Automation"],["subsubsection","High-Level Synthesis (HLS)"]],"subsections":[],"title":"High-Level Synthesis (HLS)"},{"cite_extract_rate":0.15384615384615302,"cites":[5395,8916],"content":"\\label{sec:model_logic}\nIn digital design, logic synthesis converts RTL designs into optimized gate-level representations;\nphysical synthesis then transforms these design netlists into physical layouts.\nSince these two stages may take hours or days to generate final bitstreams/layouts, many problems benefit from the power of ML models for fast performance estimation.\nIn logic synthesis, CNN models or LSTM-based models can be leveraged to forecast the delay and area after applying different synthesis flows on specific designs, where the inputs are synthesis flows represented in either matrices or time series.\nIn physical synthesis, routing is a sophisticated problem subject to stringent constraints, and EDA tools typically utilize a two-step method: global routing (GR) and detailed routing (DR). GR tool allocates routing resource coarsely, and provides routing plans to guide DR tools to complete the entire routing.\nIn general, routing congestion can be figured out during or after GR; the routability of a design is confirmed after DR and design rule checking (DRC).\nEndeavors have been made to predict routability from early layout stages, so as to avoid excessive iterations back and forth between placement and routing.\nIn ASICs, some investigations predict routability by estimating the number of design rule violations (DRVs).\nTaking GR results as inputs, Li \\textit{et al.} explore several ML models (linear regression, ANN, and decision tree) to predict the number of DRVs, final hold slack, power, and area. \nQi \\textit{et al.} rely on placement data and congestion maps from GR as input features, and use a nonparametric regression technique, multivariate adaptive regression splines (MARS) , to predict the utilization of routing resource and the number of DRVs.\nBy merely leveraging placement information, it is possible to predict routability by MARS and SVM , or to detect DR short violations by an MLP .\nWhen representing placement information as images, fully convolutional networks (FCNs) are capable to predict locations of DRC hotspots by considering GR information as inputs , or to forecast GR congestion maps by formulating the prediction task as a pixel-wise binary classification using placement data .\nJ-Net is a customized FCN model, and takes both high-resolution pin patterns and low-resolution layout information from the placement stage as features to output a 2D array that indicates if the tile corresponding to each entry is a DRC hotspot.\nIn FPGAs, routing congestion maps can be directly estimated by linear regression using feature vectors coming from pin counts and wirelength per area of SLICEs.\nBy constructing the routing congestion prediction as an image translation problem, a conditional GAN is able to take post-placement images as inputs to predict congestion heat maps.","id":"8ba558a9-ad3c-4f4c-a999-80fc17f11ebf","level":"subsubsection","origin_cites_number":13,"parent_id":"419e1fc4-5eb3-40c4-b251-cc170597bc2e","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML for Fast System Modeling"],["subsection","Performance Modeling in Chip Design and Design Automation"],["subsubsection","Logic and Physical Synthesis"]],"subsections":[],"title":"Logic and Physical Synthesis"},{"cite_extract_rate":0.15384615384615302,"cites":[5397,5409,3579,5400,5399,5406,5398,5407,5403,3603,5408,5401,5405,4096,5402,5404],"content":"\\label{sec:methodology}\nThis section introduces studies that directly employ ML techniques as the design methodology for computer architecture/systems.\nComputer architecture and systems have been becoming increasingly complicated, making it expensive and inefficient for human efforts to design or optimize them.\nIn response, visionaries have argued that computer architecture and systems should be imbued with the capability to design and configure themselves, adjust their behaviors according to workloads' needs or user-specified constraints, diagnose failures, repair themselves from the detected failures, etc.\nWith strong learning and generalization capabilities, ML-based techniques are naturally suitable to resolve these considerations, which can adjust their policies during system designs according to long-term planning and dynamic workload behaviors.\nAs many problems in architecture/system design can be formulated as combinatorial optimization or sequential decision-making problems, RL is broadly explored and exploited.\nTable \\ref{table:dse_sys} and Table \\ref{table:dse_eda} recapitulate the studies that apply ML techniques as the design methodology for computer architecture/system and design automation respectively, in terms of target tasks and adopted ML techniques.\n\\begin{table}[tbp]\n\\vspace{-10pt}\n\\caption{Summary of applying ML techniques as the design methodology for computer architecture/ systems.}\n\\vspace{-10pt}\n\\label{table:dse_sys}\n\\centering\n \\tiny\n \\renewcommand{\\arraystretch}{0.9}\n \\setlength{\\tabcolsep}{3pt}\n\\begin{tabular}{c|c|c}\n\\toprule\n\\textbf{Domain} & \\textbf{Task} & \\textbf{Technique} \\\\ \\midrule\n\\multirow{4}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Memory \\\\ System\\\\ Design \\\\ $( \\S$ \\ref{sec:dse_mem}$)$ \\end{tabular}}} & Cache replacement policy & \\begin{tabular}[c]{@{}c@{}}Perceptron learning , Markov decision process , LSTM and SVM \\end{tabular} \\\\ \\cline{2-3} \n & Cache prefetching policy & \\begin{tabular}[c]{@{}c@{}}Perceptron learning , contextual bandit , LSTM \\end{tabular} \\\\ \\cline{2-3} \n & Memory controller policy & Q-learning \\\\ \\cline{2-3} \n & Garbage collection & Q-learning \\\\ \\hline\n\\textbf{\\begin{tabular}[c]{@{}c@{}}Branch\\\\ Prediction \\\\ $( \\S$ \\ref{sec:dse_bp}$)$\\end{tabular}} & Branch direction & \\begin{tabular}[c]{@{}c@{}}MLP , piecewise linear regression , \\\\ perceptron , CNN \\end{tabular} \\\\ \\hline\n\\multirow{8}{*}{\\makecell[c]{\\textbf{NoC} \\\\ $( \\S$ \\textbf{\\ref{NoC}}$)$}} & Link management & ANN \\\\ \\cline{2-3} \n & DVFS for routers & Q-learning \\\\ \\cline{2-3} \n & Routing & Q-learning \\\\ \\cline{2-3} \n & Arbitration policy & DQN \\\\ \\cline{2-3} \n & Adjusting injection rates & Q-learning , ANN \\\\ \\cline{2-3} \n & Selection of fault-tolerant modes & Q-learning \\\\ \\cline{2-3} \n & Link placement in 3D NoCs & STAGE algorithm \\\\ \\cline{2-3} \n & Loop placement in routerless NoCs & Advantage actor-critic with MCTS \\\\ \\hline\n\\multirow{6}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Power \\\\ Management \\\\ $( \\S$ \\ref{power}$)$ \\end{tabular}}} & DVFS and thread packing & Multinomial logistic regression \\\\ \\cline{2-3} \n & DVFS and power gating & MLP \\\\ \\cline{2-3} \n & \\begin{tabular}[c]{@{}c@{}}DVFS, socket allocation, and use of HyperThreads\\end{tabular} & Extra trees/gradient boosting/KNN/MLP/SVM \\\\ \\cline{2-3} \n & \\begin{tabular}[c]{@{}c@{}}DVFS for CPU cores/uncore/through-silicon interposers\\end{tabular} & Propositional rule , ANN , Q-learning \\\\ \\cline{2-3} \n & \\begin{tabular}[c]{@{}c@{}}DVFS for CPU-GPU heterogeneous platforms\\end{tabular} & Weighted majority algorithm \\\\ \\cline{2-3} \n & DVFS for multi-/many-core systems & \\begin{tabular}[c]{@{}c@{}}Q-learning , semi-supervised RL , hierarchical Q-learning \\end{tabular} \\\\ \\hline\n\\multirow{6}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Resource \\\\ Management \\\\ \\& Task \\\\ Allocation \\\\ $( \\S$ \\ref{resource management}$)$\\end{tabular}}} & Tuning architecture configurations & \\begin{tabular}[c]{@{}c@{}}Maximum likelihood , statistical machine learning \\end{tabular} \\\\ \\cline{2-3} \n & Dynamic cache partitioning & Enforced subpopulations , Q-learning \\\\ \\cline{2-3} \n & Task allocation in many-core systems & Q-learning , DDPG \\\\ \\cline{2-3} \n & Workflow management & SVM and random forest \\\\ \\cline{2-3} \n & Hardware resource assignment & REINFORCE , Bayesian optimization \\\\ \\cline{2-3} \n & Device placement & \\begin{tabular}[c]{@{}c@{}}REINFORCE , policy gradient , PPO \\end{tabular} \\\\ \\hline\n\\multirow{2}{*}{\\makecell[c]{\\textbf{Scheduling}\\\\ $( \\S$ \\textbf{\\ref{scheduling}}$)$}} & Scheduling jobs in single-core processors & Q-learning \\\\ \\cline{2-3} \n & Scheduling jobs in multi-processor systems & Value-based RL \\\\ \\hline\n\\multirow{9}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Data Center\\\\ Management \\\\ $( \\S$ \\ref{data center}$)$\\end{tabular}}} & Assignment of servers to applications & Value-based RL \\\\ \\cline{2-3} \n & Content allocation in CDNs & Fuzzy RL \\\\ \\cline{2-3} \n & \\begin{tabular}[c]{@{}c@{}}Placement of virtual machines onto physical machines\\end{tabular} & PPO \\\\ \\cline{2-3} \n & Traffic optimization & Policy gradient and DDPG \\\\ \\cline{2-3} \n & Scheduling jobs with complex dependency & REINFORCE \\\\ \\cline{2-3} \n & Straggler diagnosis & Statistical ML \\\\ \\cline{2-3} \n & Data-center-level caching policy & \\begin{tabular}[c]{@{}c@{}}Decision tree , LSTM , gradient boosting , DDPG \\end{tabular} \\\\ \\cline{2-3} \n & Bitrate selection for video chunks & A3C \\\\ \\cline{2-3} \n & \\begin{tabular}[c]{@{}c@{}}Scheduling video workloads in hybrid CPU-GPU clusters\\end{tabular} & DQN \\\\ \\hline\n \\multirow{3}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Code\\\\ Generation \\\\ $( \\S$ \\ref{code_generation}$)$\\end{tabular}}} & Code completion & N-gram model and RNN \\\\ \\cline{2-3} \n & Code generation & LSTM \\\\ \\cline{2-3} \n & Program translation & Tree-to-tree encoder-decoder , seq2seq , transformer \\\\ \\hline\n\\multirow{6}{*}{\\makecell[c]{\\textbf{Compiler}\\\\ $( \\S$ \\textbf{\\ref{sec:compiler}}$)$}} & Instruction scheduling & Temporal difference , projective reparameterization \\\\ \\cline{2-3} \n & Improving compiler heuristics & NEAT , LSTM \\\\ \\cline{2-3} \n & Ordering of optimizations & NEAT \\\\ \\cline{2-3} \n & Automatic vectorization & Imitation learning \\\\ \\cline{2-3} \n & \\begin{tabular}[c]{@{}c@{}}Program transformation for approximate computing\\end{tabular} & MLP \\\\ \\cline{2-3} \n & Compilation for DNN workloads & PPO , policy gradient \\\\ \\bottomrule\n\\end{tabular}\n\\end{table}","id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","level":"section","origin_cites_number":104,"parent_id":"5631fef3-6389-407d-9101-745f45240141","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"]],"subsections":["4ce0d49a-419f-4a57-9bf8-bac10689ff82","93695066-a3b6-43e6-a561-f289514d99b8","02b1496f-b229-4d07-b91a-238e79b26753","8b48b63d-94af-4ec0-a218-72746bce7952","de7c5699-73a2-48d8-9e75-ff1b60c4f5f5","68c1d9ca-e725-4975-ad78-356d7848d1d1","e25c32d8-64dc-4364-8501-8063014c67c6"],"title":"ML as Design Methodology"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:dse_mem}\nThe \"memory wall\" has been a performance bottleneck in von Neumann architectures, where computation is orders of magnitude faster than memory access.\nTo alleviate this problem, hierarchical memory systems are widely used and there arise optimizations for different levels of memory systems.\nAs both the variety and the size of modern workloads are drastically growing, conventional designs that are based on heuristics or intuitions may not catch up with the demand of the ever-growing workloads, leading to sharply degradation in performance.\nAs such, many studies resort to ML-based techniques to design smart and intelligent memory systems.","id":"4ce0d49a-419f-4a57-9bf8-bac10689ff82","level":"subsection","origin_cites_number":0,"parent_id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Memory System Design"]],"subsections":["c842f575-5671-48ac-aa2a-715cd98df8a6","dbd16b4a-1edb-4482-a515-db26bea31ee6","dc411e9f-ce19-4cdd-9e1c-1930d2c68aa6"],"title":"Memory System Design"},{"cite_extract_rate":0.18181818181818102,"cites":[5401,5398],"content":"\\label{cache}\nThe conspicuous disparity in latency and bandwidth between CPUs and memory systems motivates investigations for efficient cache management.\nThere are two major types of studies on cache optimization: improving cache replacement policies, and designing intelligent prefetching policies.\n\\textcircled{\\small{1}} \nTo develop cache replacement policies, perceptron learning is employed to predict whether to bypass or reuse a referenced block in the last-level cache (LLC) .\nInstead of using perceptrons, Beckmann \\textit{et al.} model the cache replacement problem as a Markov decision process and replace lines according to the difference between their expected hits and the average hits.\nShi \\textit{et al.}~ train an attention-based LSTM model offline to extract insights from history program counters, which are then used to build an online SVM-based hardware predictor to serve as the cache replacement policy.\n\\textcircled{\\small{2}} \nTo devise intelligent prefetchers, Wang \\textit{et al.} propose a prefetching mechanism that uses conventionally table-based prefetchers to provide prefetching suggestions and a perceptron trained by spatio-temporal locality to reject unnecessary prefetching decisions, ameliorating the cache pollution problem. \nSimilarly, Bhatia \\textit{et al.} integrate a perceptron-based prefetching filter with conventional prefetchers, increasing the coverage of prefetches without hurting accuracy.\nInstead of the commonly used spatio-temporal locality, a context-based memory prefetcher leverages the semantic locality that characterizes access correlations inherent to program semantics and data structures, which is approximated by a contextual bandit model in RL.\nInterpreting semantics in memory access patterns is analogous to sequence analysis in natural language processing (NLP), and thus several studies use LSTM-based models and treat the prefetching as either a regression problem or a classification problem .\nEven with better performance, especially for long access sequences and noise traces, LSTM-based prefetchers suffer from long warm-up and prediction latency, and considerable storage overheads.\nThe discussion of how hyperparameters impact LSTM-based prefetchers' performance highlights that the lookback size (i.e. memory access history window) and the LSTM model size strongly affect prefetchers' learning ability under different noise levels or workload patterns.\nTo accommodate the large memory space, Shi \\textit{et al.} introduce a hierarchical sequence model to decouple predictions of pages and offsets by using two separate attention-based LSTM layers, whereas the corresponding hardware implementation is impractical for actual processors.","id":"c842f575-5671-48ac-aa2a-715cd98df8a6","level":"subsubsection","origin_cites_number":11,"parent_id":"4ce0d49a-419f-4a57-9bf8-bac10689ff82","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Memory System Design"],["subsubsection","Cache"]],"subsections":[],"title":"Cache"},{"cite_extract_rate":0,"cites":[],"content":"Smart memory controllers can significantly improve memory bandwidth utilization.\nAiming at a self-optimizing memory controller adaptive to dynamically changing workloads, it can be modeled as an RL agent that always selects legal DRAM commands with the highest expected long-term performance benefits (i.e., Q-values) .\nTo allow optimizations toward various objectives, this memory controller is then improved in two major aspects . First, the rewards of different actions (i.e., legal DRAM commands) are automatically calibrated by genetic algorithms to serve different objective functions (e.g., energy, throughput, etc). Second, a multi-factor method that considers the first-order attribute interactions is employed to select proper attributes used for state representations.\nSince both of them use table-based Q-learning and select limited attributes to represent states, the scalability may be a concern and their performance could be improved with more informative representations.","id":"dbd16b4a-1edb-4482-a515-db26bea31ee6","level":"subsubsection","origin_cites_number":3,"parent_id":"4ce0d49a-419f-4a57-9bf8-bac10689ff82","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Memory System Design"],["subsubsection","Memory Controller"]],"subsections":[],"title":"Memory Controller"},{"cite_extract_rate":0.2,"cites":[3600],"content":"A variety of work targets different parts of the memory system.\nMargaritov \\textit{et al.} accelerate virtual address translation through learned index structures . The results are encouraging in terms of the accuracy, which reaches almost 100\\% for all tested virtual addresses; yet this method has unacceptably long inference latency, leaving practical hardware implementation as the future work.\nWang \\textit{et al.} reduce data movement energy in interconnects by exploiting asymmetric transmission costs of different bits, where data blocks to be transmitted are dynamically grouped by K-majority clustering to derive energy-efficient expressions for transmission.\nIn terms of garbage collection in NAND flash, Kang \\textit{et al.} propose an RL-based method to reduce the long-tail latency. The key idea is to exploit the inter-request interval (idle time) to dynamically decide the number of pages to be copied or whether to perform an erase operation, where decisions are made by table-based Q-learning. Their following work considers more fine-grained states, and introduces a Q-table cache to manage key states among enormous amount of states.","id":"dc411e9f-ce19-4cdd-9e1c-1930d2c68aa6","level":"subsubsection","origin_cites_number":5,"parent_id":"4ce0d49a-419f-4a57-9bf8-bac10689ff82","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Memory System Design"],["subsubsection","Others"]],"subsections":[],"title":"Others"},{"cite_extract_rate":0.11111111111111101,"cites":[5405],"content":"\\label{sec:dse_bp}\nBranch predictor is one of the mainstays of modern processors, significantly improving the instruction-level parallelism. As pipelines gradually deepen, the penalty of mis-prediction increases. Traditional branch predictors often consider limited history length, which may hurt the prediction accuracy.\nIn contrast, the perceptron/MLP-based predictors can handle long histories with reasonable hardware budgets, outperforming prior state-of-the-art non-ML-based predictors.\nStarting with a static branch predictor trained with static features from program corpus and control flow graphs, an MLP is used to predict the direction of a branch at compile time .\nLater, a dynamic branch predictor uses a perceptron-based method . It hashes the branch address to select the proper perceptron and computes the dot product accordingly to decide whether to take this branch, which shows great performance on linearly separable branches. \nIts latency and accuracy can be further improved by applying ahead pipelining and selecting perceptrons based on path history .\nTo attain high accuracy in non-linearly separable branches, the perceptron-based prediction is generalized as piecewise linear branch prediction .\nIn addition to the path history, multiple types of features from different organizations of branch histories can be leveraged to enhance the overall performance .\nWhen considering practical hardware implementation of branch predictors, SNAP leverages current-steering digital-to-analog converters to transfer digital weights into analog currents and replaces the costly digital dot-product computation to the current summation.\nIts optimized version equips several new techniques, such as the use of global and per-branch history, trainable scaling coefficients, dynamic training thresholds, etc.\nRather than making binary decisions of whether to take a certain branch, it is possible to directly predict the target address of an indirect branch at the bit level via perceptron-based predictors .\nWhile high accuracy is achieved by current perceptron/MLP-based predictors, Tarsa \\textit{et al.} notice that a small amount of static branch instructions are systematically mispredicted, referred to as hard-to-predict branches (H2Ps). Consequently, they propose a CNN helper predictor for pattern matching of history branches, ultimately improving accuracy for H2Ps in conditional branches.","id":"93695066-a3b6-43e6-a561-f289514d99b8","level":"subsection","origin_cites_number":9,"parent_id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Branch Prediction"]],"subsections":[],"title":"Branch Prediction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{NoC}\nThe aggressive transistor scaling has paved the way for integrating more cores in a single chip or processor.\nWith the increasing number of cores per chip, NoC plays a gradually crucial role, since it is responsible for inter-core communication and data movement between cores and memory hierarchies.\nSeveral problems attracting attention are as follows. \nFirst, communication energy scales slower than computation energy , implying necessity to improve power efficiency of NoCs. \nSecond, the complexity of routing or traffic control grows with the number of cores per chip and this problem is even exacerbated by the rising variety and irregularity of workloads.\nThird, with the continuous scaling down of transistors, NoCs are more vulnerable to different types of errors and thus reliability becomes a key concern.\nFourth, some non-conventional NoC architectures might bring promising potentials in the future, whereas they usually come with large design spaces and complex design constraints, which is nearly impossible for manually optimization.\nAmong all aforementioned fields, ML-based design techniques display their strength and charm.","id":"02b1496f-b229-4d07-b91a-238e79b26753","level":"subsection","origin_cites_number":1,"parent_id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","NoC Design"]],"subsections":["f70d9a91-45f6-4778-9893-1e1691350184","9cbfc7ad-3786-493c-a695-e8c786d70c4c","0f23b3c6-50c2-4cf8-a777-94590c806c36","5afafbcf-6c39-4a33-8c76-314585dc4284"],"title":"NoC Design"},{"cite_extract_rate":0,"cites":[],"content":"Power consumption is one crucial concern in NoCs, in which links usually consume a considerable portion of network power.\nWhile turning on/off links according to a static threshold of link utilization is a trivial way to reduce power consumption, it can not adapt to dynamically changing workloads.\nSavva \\textit{et al.} use multiple ANNs for dynamic link management. Each ANN is responsible for one region of the NoC, and dynamically computes a threshold for every time interval to turn on/off links given the link utilization of each region. Despite significant power savings with low hardware overheads, this approach causes long latency in routing.\nIn order to meet certain power and thermal budgets, hierarchical ANNs are used to predict optimal NoC configurations (i.e., link bandwidth, node voltage and task assignment to nodes), where the global ANN predicts globally optimal NoC configurations exploiting local optimal energy consumption predicted by local ANNs.\nTo save dynamic power, several investigations employ per-router based Q-learning agents, which are offline trained ANNs to select optimal voltage/frequency levels for each router.","id":"f70d9a91-45f6-4778-9893-1e1691350184","level":"subsubsection","origin_cites_number":4,"parent_id":"02b1496f-b229-4d07-b91a-238e79b26753","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","NoC Design"],["subsubsection","Link Management and DVFS"]],"subsections":[],"title":"Link Management and DVFS"},{"cite_extract_rate":0,"cites":[],"content":"With the increasing variety and irregularity of workloads and their traffic patterns, learning-based routing algorithms and traffic control approaches show superior performance due to their excellent adaptability.\n\\textcircled{\\small{1}} \nAs routing problems can be formulated as sequential decision-making processes, several studies apply Q-learning based approaches, namely the Q-routing algorithm , which uses local estimation of delivery time to minimize total packets delivery time, capable to handle irregular network topologies and keep a higher network load than the conventional shortest path routing.\nQ-routing is then extended to several other scenarios, such as combining with dual RL to improve learning speed and routing performance , resolving packets routing in dynamic NoCs whose network structures/topologies are dynamically changing during runtime , handling irregular faults in bufferless NoCs by the reconfigurable fault-tolerant Q-routing , and enhancing the capability to reroute messages around congested regions by the congestion-aware non-minimal Q-routing .\nIn addition to routing problems, deep Q-network is also promising for NoC arbitration policies , where the agent/arbiter grants a certain output port to the input buffer with the largest Q-value.\nEven displaying some improvements in latency and throughput, the direct hardware implementation is impractical due to the complexity of deep Q-networks, and thus insights are distilled to derive a relatively simple circuitry implementation.\n\\textcircled{\\small{2}} \nWith the goal to control congestion in NoCs, the SCEPTER NoC architecture , a bufferless NoC with single-cycle multi-hop traversals and a self-learning throttling mechanism, controls the injection of new flits into the network by Q-learning. Each node in the network independently selects whether to increase, decrease, or retain the throttle rate according to their Q-values, which conspicuously improves bandwidth allocation fairness and network throughput.\nWang \\textit{et al.}~ design an ANN-based admission controller to determine the appropriate injection rate and the control policy of each node in a standard NoC.","id":"9cbfc7ad-3786-493c-a695-e8c786d70c4c","level":"subsubsection","origin_cites_number":9,"parent_id":"02b1496f-b229-4d07-b91a-238e79b26753","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","NoC Design"],["subsubsection","Routing and Traffic Control"]],"subsections":[],"title":"Routing and Traffic Control"},{"cite_extract_rate":0,"cites":[],"content":"With the aggressive technology scaling down, transistors and links in NoCs are more prone to different types of errors, indicating that reliability is a crucial concern and proactive fault-tolerant techniques are required to guarantee performance.\nWang~\\textit{et al.}~ employ per-router-based Q-learning agents to independently select one of four fault-tolerant modes, which can minimize the end-to-end packet latency and power consumption. These agents are pre-trained and then fine-tuned during runtime.\nIn their following work , these error-correction modes are extended and combined with various multi-function adaptive channel configurations, retransmission settings, and power management strategies, significantly improving latency, energy efficiency, and mean-time-to-failure.","id":"0f23b3c6-50c2-4cf8-a777-94590c806c36","level":"subsubsection","origin_cites_number":2,"parent_id":"02b1496f-b229-4d07-b91a-238e79b26753","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","NoC Design"],["subsubsection","Reliability and Fault Tolerance"]],"subsections":[],"title":"Reliability and Fault Tolerance"},{"cite_extract_rate":0.4,"cites":[5400,5406],"content":"With the growing number of cores per chip/system, the increasing heterogeneity of cores, and various performance targets, it is complicated to simultaneously optimize copious design knobs in NoCs.\nOne attempt to automated NoC design is the MLNoC , which utilizes supervised learning to quickly find near-optimal NoC designs under multiple optimization goals. MLNoC is trained by data from thousands of real-world and synthetic SoC (system-on-chip) designs, and evaluated with real-world SoC designs. Despite disclosure of limited details and absence of comprehensive comparison with other design methods, it shows superior performance to manually optimized NoC designs, delivering encouraging results.\nApart from conventional 2D mesh NoCs, a series of investigations focuses on 3D NoC designs, where the STAGE algorithm is applied to optimize vertical and planar placement of communication links in small-world network based 3D NoCs .\nThe STAGE algorithm repeatedly alternates between two stages, the base search that tries to find the local optima based on the learned evaluation function, and the meta-search that uses SVR to learn evaluation functions.\nLater, the STAGE algorithm is extended for multi-objective optimization in heterogeneous 3D NoC systems , which jointly considers GPU throughput, average latency between CPUs and LLCs, temperature, and energy.\nIn terms of routerless NoCs that any two nodes are connected via at least one ring/loop, a deep RL framework that exploits Monte-Carlo tree search for efficient design space exploration is developed to optimize loop placements , and the design constraints can be strictly enforced by carefully devising the reward function.","id":"5afafbcf-6c39-4a33-8c76-314585dc4284","level":"subsubsection","origin_cites_number":5,"parent_id":"02b1496f-b229-4d07-b91a-238e79b26753","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","NoC Design"],["subsubsection","General Design"]],"subsections":[],"title":"General Design"},{"cite_extract_rate":0,"cites":[],"content":"\\label{resource}\nResource allocation or management is the coordination between computer architecture/systems and workloads. Consequently, its optimization difficulty occurs with the booming complexity from both sides and their intricate interactions. \nML-based approaches have blazed the trail to adjusting policies wisely and promptly pursuant to dynamic workloads or specified constraints.","id":"8b48b63d-94af-4ec0-a218-72746bce7952","level":"subsection","origin_cites_number":0,"parent_id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Resource Allocation or Management"]],"subsections":["2eb80446-13aa-40ce-a764-e98e86d2b650","35decbd9-a2aa-43b2-b9e5-0585c9140768","5a43dcc0-acd3-4daf-bbb5-f9e380b060a3"],"title":"Resource Allocation or Management"},{"cite_extract_rate":0,"cites":[],"content":"\\label{power}\nML-based techniques have been applied broadly to improve power management, due to two main reasons.\nFirst, power/energy consumption can be recognized as one metric of runtime costs.\nSecond, under certain circumstances there could be a hard or soft constraint/budget of power/energy, making power efficiency a necessity.\nIn consideration of power management for different parts of systems, PACSL uses the propositional rule to adjust dynamic voltage scaling (DVS) for CPU cores and on-chip L2 cache, which achieves an improvement in the energy-delay product by 22\\% on average (up to 46\\%) over independently applying DVS for each part.\nWon \\textit{et al.} coordinate an ANN controller with a proportional integral for uncore DVFS. The ANN controller can be either pre-trained offline by a prepared dataset or trained online by bootstrapped learning.\nManoj~\\textit{et al.} deploy Q-learning to adaptively adjust the level of output-voltage swing at transmitters of 2.5D through-silicon interposer I/Os, under constraints of communication power and bit error rate.\nFrom the system level, DVFS is one of the most prevalent techniques.\nPack \\& Cap builds a multinomial logistic regression classifier that is trained offline and queried during runtime, to accurately identify the optimal operating point for both thread packing and DVFS under an arbitrary power cap.\nGreenGPU focuses on heterogeneous systems with CPUs and GPUs, and applies the weighted majority algorithm to scale frequency levels for both GPU cores and memory in a coordinated manner.\nCHARSTAR targets joint optimization of power gating and DVFS within a single core, where frequencies and configurations are dynamically selected by a lightweight offline trained MLP predictor.\nTo minimize energy consumption, Imes~\\textit{et al.}~ use ML-based classifiers (e.g., extra trees, gradient boosting, KNN, MLP and SVM) to predict the most energy-efficient resource settings (specifically, tuning socket allocation, the use of HyperThreads, and processor DVFS) by using low-level hardware performance counters.\nBai~\\textit{et al.}~ consider the loss caused by on-chip regulator efficiency during DVFS, and try to minimize energy consumption under a parameterized performance constraint. The online control policy is implemented by a table-based Q-learning, which is portable across platforms without accurate modeling of a specific system.\nA series of studies leverages RL for dynamic power management in multi-/many-core systems.\nAs systems scale up, these RL-based methods often suffer from state space explosion, and two types of methods are introduced to resolve the scalability issue.\n\\textcircled{\\small{1}} \nBy combining RL with supervised learning, a semi-supervised RL-based approach achieves linear complexity with the number of cores, which is able to maximize throughput ensuring power constraints and cooperatively control cores and uncores in synergy.\n\\textcircled{\\small{2}} \nThe exploitation of hierarchical Q-learning reduces the time complexity to $O(n \\lg n)$, where $n$ denotes the number of cores.\nPan \\textit{et al.} introduce multi-level Q-learning to select target power modes, where Q-values are approximated by a generalized radial basis function.\nTable-based distributed Q-learning also performs well for DVFS , and there is one variant aware of the priorities of different applications.\nSome energy management policies target specific applications or platforms.\nJouleGuard~ is a runtime control system coordinating approximate computing applications with system resource under energy budgets.\nIt uses a multi-arm bandit approach to identifying the most energy efficient system configuration, upon which application configurations are determined to maximize compute accuracy within energy budgets.\nTargeting Intel SkyLake processors, a post-silicon CPU customization applies various ML models for dynamically clock-gating unused resource .","id":"2eb80446-13aa-40ce-a764-e98e86d2b650","level":"subsubsection","origin_cites_number":14,"parent_id":"8b48b63d-94af-4ec0-a218-72746bce7952","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Resource Allocation or Management"],["subsubsection","Power Management"]],"subsections":[],"title":"Power Management"},{"cite_extract_rate":0.23529411764705802,"cites":[5404,5403,5407,5409],"content":"\\label{resource management}\nModern architectures and systems have been becoming so sophisticated and diverse that it is non-trivial to either optimize performance or fully utilize system resource.\nThis rapidly evolving landscape is further complicated by various workloads with specific requirements or targets.\nIn order to keep the pace, one cure is to develop more efficient and automated methods for resource management and task allocation, where ML-based techniques are excelled to explore large design spaces and simultaneously optimize multiple objectives, and preserve better scalablility and portability after carefully designed.\nFor a single-core processor, a regularized maximum likelihood approach predicts the best hardware micro-architectural configuration for each phase of a program, based on runtime hardware counters. \nFor multi-core processors, a statistical machine learning (SML) based method can quickly find configurations that simultaneously optimize running time and energy efficiency. Since this method is agnostic to application and micro-architecture domain knowledge, it is a portable alternative to human expert optimization.\nSML can also be applied as a holistic method to design self-evolving systems that optimize performance hierarchically across circuit, platform, and application levels .\nIn addition to tuning architectural configurations, dynamic on-chip resource management is crucial for multi-core processors, where one example is dynamic cache partitioning.\nIn response to changing workload demands, an RNN evolved by the enforced subpopulations algorithm is introduced to partition L2 cache dynamically.\nWhen integrating dynamic partitioning of LLC with DVFS on cores and uncore, a co-optimization method using table-based Q-learning achieves much lower energy-delay products than any of the techniques applied individually .\nTo guarantee efficient and reliable execution in many-core systems, task allocation should consider several aspects, such as heat and communication issues.\nTargeting the heat interaction of processor cores and NoC routers, Lu \\textit{et al.} apply Q-learning to assign tasks to cores based on current temperatures of cores and routers, such that the maximum temperature in the future is minimized.\nTargeting the non-uniform and hierarchical on/off-chip communication capability in multi-chip many-core systems, core placement optimization leverages deep deterministic policy gradient (DDPG) to map computation onto physical cores, able to work in a manner agnostic to domain-specific information.\nSome studies pay attention to workflow management and general hardware resource assignment.\nSmartFlux focuses on the workflow of data-intensive and continuous processing. It intelligently guides asynchronous triggering of processing steps with the help of predictions made by multiple ML models (e.g., SVM, random forest), which indicate whether to execute certain steps and to decide corresponding configurations upon each wave of data.\nGiven target DNN models, deployment scenarios, platform constraints, and optimization objectives (latency/energy), ConfuciuX~ applies a hybrid two-step scheme for optimal hardware resource assignments (i.e., assigning the number of processing elements and the buffer sizes to each DNN layer), where REINFORCE performs a global coarse-grained search followed by a genetic algorithm for fine-grained tuning.\nApollo is a general architecture exploration framework for sample-efficient accelerator designs, which leverages ML-based black-box optimization techniques (e.g., Bayesian optimization) to optimize accelerator configurations to satisfy use-specified design constraints.\nIn heterogeneous systems with CPUs and GPUs, device placement refers to the process of mapping nodes in computational graphs of neural networks onto proper hardware devices.\nInitially, computational operations are grouped manually, and assigned to devices by REINFORCE that employs a sequence-to-sequence RNN model as the parameterized policy .\nLater, a hierarchical end-to-end model makes this manual grouping process automatic .\nThe training speed is further improved by introduction of proximal policy optimization (PPO) .\nDespite great advance brought by the above approaches, they are not transferable and a new policy should be trained from scratch specifically for each new computational graph.\nBy encode structure of computational graphs with static graph embeddings or learnable graph embeddings , the trained placement policy exhibits great generalizability to unseen neural networks.","id":"35decbd9-a2aa-43b2-b9e5-0585c9140768","level":"subsubsection","origin_cites_number":17,"parent_id":"8b48b63d-94af-4ec0-a218-72746bce7952","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Resource Allocation or Management"],["subsubsection","Resource Management and Task Allocation"]],"subsections":[],"title":"Resource Management and Task Allocation"},{"cite_extract_rate":0,"cites":[],"content":"\\label{scheduling}\nIn classical real-time scheduling problems, the key task is to decide the order, according to which the currently unscheduled jobs should be executed by a single processor, such that the overall performance is optimized.\nAs multi-core processors have been the mainstream, the scheduling is gradually perplexing.\nOne major reason is that multiple objectives besides the performance should be carefully considered, such as balanced assignments among various cores and response time fairness.\nEquipped with the capability to well understand the feedback provided by the environment and to dynamically adjust policies, RL is a common tool for real-time scheduling.\nTo optimize the execution order of jobs after they are routed to a single CPU core, Whiteson and Stone propose an adaptive scheduling policy that exploits Q-routing, where the scheduler utilizes the router’s Q-table to assess a job’s priority and decides jobs' ordering accordingly so as to maximize the overall utility.\nIn multi-core systems, Fedorova~\\textit{et al.} present a blueprint for a self-tuning scheduling algorithm based on the value-based temporal-difference method in RL, aiming to maximize a cost function that is an arbitrary weighted sum of metrics of interest.\nThis algorithm is then improved to be a general method for online scheduling of parallel jobs , where the value functions are approximated by a parameterized fuzzy rulebase. This scheduling policy always selects to execute jobs with the maximum value functions in the job queue, which possibly preempts currently running jobs and squeezes some jobs into fewer CPUs than they ideally require, with the goal of achieving optimized long-term utility.","id":"5a43dcc0-acd3-4daf-bbb5-f9e380b060a3","level":"subsubsection","origin_cites_number":3,"parent_id":"8b48b63d-94af-4ec0-a218-72746bce7952","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Resource Allocation or Management"],["subsubsection","Scheduling"]],"subsections":[],"title":"Scheduling"},{"cite_extract_rate":0.2,"cites":[3579,3603,1390],"content":"\\label{data center}\nWith the rapid scale expansion of data centers, issues that may be trivial in a single machine become increasingly challenging, let alone the inherently complicated problems.\nEarly work aims at a relatively simple scenario of resource allocation, i.e., to dynamically assign different numbers of servers to multiple applications. \nThis problem can be modeled as an RL problem with service-level utility functions as rewards:\nthe arbiter will select a joint action that would bring the maximum total return after consulting local value functions estimated via either table-based methods or function approximation . \nIn order to better model interactions among multiple agents, a multi-agent coordination algorithm with fuzzy RL can be used to solve the dynamic content allocation in content delivery networks (CDNs), in which each requested content is modeled as an agent, trying to move toward the area with a high demand while coordinating with other agents/contents.\nA recent innovation pays attention to the placement of virtual machines onto physical machines, so as to minimize the peak-to-average ratio of resource usage across physical machines, where PPO and hindsight imitation learning are evaluated.\nTo improve data center performance and quality of experience (QoE) for users, ML-based techniques have been explored in a few directions.\n\\textcircled{\\small{1}} It is important to efficiently schedule jobs and effectively diagnose stragglers within jobs.\nAiming at traffic optimization (e.g., flow scheduling, load balancing) in data centers, Chen~\\textit{et al.} develop a two-level RL system: peripheral systems, which are trained by DDPG, reside on end-hosts and locally make instant traffic optimization decisions for short flows; the central system, which is trained by policy gradient, aggregates global traffic information, guides behaviors of peripheral systems, and makes traffic optimization decisions for long flows.\nDecima exploits GNNs to represent cluster information and dependency among job stages, so that the RL-based scheduler can automatically learn workload-specific scheduling policies to schedule data processing jobs with complex dependency. \nHound combines statistical ML with meta-learning to diagnose causes of stragglers at data-center-scale jobs.\n\\textcircled{\\small{2}} It is essential to deploy an intelligent data-center-level cache.\nDeepCache employs an LSTM encoder-decoder model to predict future content popularity, which can be combined with existing cache policies to make smarter decisions.\nSong \\textit{et al.} apply gradient boosting machines to mimic a relaxed Belady algorithm that evicts an object whose next request is beyond a reuse distance threshold but not necessarily the farthest in the future.\nPhoebe is an online cache replacement framework leveraging DDPG to predict priorities of objects and to conduct eviction accordingly.\nConsidering non-history based features, Wang \\textit{et al.}~ build a decision tree to predict whether the requested file will be accessed only once in the future. These one-time-access files will be directly sent to users without getting into cache, to avoid cache pollution. \n\\textcircled{\\small{3}} From the workload perspective, video workloads on CDNs or clusters are prevalent but their optimization is quite challenging: first, network conditions fluctuate overtime and a variety of QoE goals should be balanced simultaneously; second, only coarse decisions are available and current decisions will have long-term effects on following decisions.\nThis scenario naturally matches the foundation of RL-based techniques.\nTo optimize users' QoE of streaming videos, adaptive bitrate algorithms have been recognized as the primary tool used by content providers, which are executed on client-side video players and dynamically choose a bitrate for each video chunk based on underlying network conditions.\nPensieve applies asynchronous advantage actor-critic to select proper bitrate for future video chunks based on resulting performance from past decisions.\nWhen considering large-scale video workloads in hybrid CPU-GPU clusters, performance degradation often comes from uncertainty and variability of workloads, and unbalanced use of heterogeneous resources. \nTo accommodate this, Zhang \\textit{et al.} use two deep Q-networks to build a two-level task scheduler, where the cluster-level scheduler selects proper execution nodes for mutually independent video tasks and the node-level scheduler assigns interrelated video subtasks to appropriate computing units. This scheme enables the scheduling model to adjust policies according to runtime status of cluster environments, characteristics of video tasks, and dependency among video tasks.","id":"de7c5699-73a2-48d8-9e75-ff1b60c4f5f5","level":"subsection","origin_cites_number":15,"parent_id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Data Center Management"]],"subsections":[],"title":"Data Center Management"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:dse_code_compiler}","id":"68c1d9ca-e725-4975-ad78-356d7848d1d1","level":"subsection","origin_cites_number":0,"parent_id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Code Generation and Compiler"]],"subsections":["21901bea-3ca4-4650-b39a-4ecbde92fab7","9d772652-3e90-4f57-98a3-ad3bb92cfe28"],"title":"Code Generation and Compiler"},{"cite_extract_rate":0.5714285714285711,"cites":[5397,5408,4096,5410],"content":"\\label{code_generation}\nDue to the similarities in syntax and semantics between programming languages and natural languages, the problem of code generation or translation is often modeled as an NLP problem or a neural machine translation (NMT) problem. \nHere, we would like to bring up a brief discussion. \nFor more reference, a comprehensive survey detailedly contrasts programming languages against natural languages, and discusses how these similarities and differences drive the design and application of different ML models in code.\nTargeting code completion, several statistical language models (N-gram model, RNN, and a combination of these two) are explored to select sentences that have the highest probability and satisfy constraints to fill up partial programs with holes.\nAs for code generation, CLgen trains LSTM models by a corpus of hand-written code to learn semantics and structures of OpenCL programs, and generates human-like programs via iteratively sampling from the learned model.\nTargeting program translation, NMT-based techniques are widely applied to migrate code from one language to another.\nFor example, a tree-to-tree model with the encoder-decoder structure effectively translates programs from Java to C\\# ;\nthe sequence-to-sequence (seq2seq) model can translate from CUDA to OpenCL .\nRather than translating between high-level programming languages, Coda translates binary executables to the corresponding high-level code, which employs a tree-to-tree encoder-decoder structure for code sketch generation and an ensembled RNN-based error predictor for iterative error correction on the generated code.\nNotably, these supervised NMT-based techniques may confront several issues: difficulty to generalize to programs longer than training ones, limited size of vocabulary sets, and scarcity of aligned input-output data.\nFully counting on unsupervised machine translation, TransCoder adopts a transformer architecture and uses monolingual source code to translate among C++, Java, and Python.","id":"21901bea-3ca4-4650-b39a-4ecbde92fab7","level":"subsubsection","origin_cites_number":7,"parent_id":"68c1d9ca-e725-4975-ad78-356d7848d1d1","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Code Generation and Compiler"],["subsubsection","Code Generation"]],"subsections":[],"title":"Code Generation"},{"cite_extract_rate":0.25,"cites":[5402,5399,5411],"content":"\\label{sec:compiler}\nThe complexity of compilers grows with the complexity of computer architectures and workloads. \nML-based techniques can optimize compilers from many perspectives, such as instruction scheduling, compiler heuristics, the order to apply optimizations, hot path identification, auto-vectorization, and compilation for specific applications.\n\\textcircled{\\small{1}} For instruction scheduling, the preference function of one scheduling over another can be computed by the temporal difference algorithm in RL .\nRegarding scheduling under highly-constrained code optimization, the projective reparameterization enables automatic instruction scheduling under constraints of data-dependent partial orders over instructions.\n\\textcircled{\\small{2}} For improving compiler heuristics, Neuro-Evolution of Augmenting Topologies (NEAT) improves instruction placement heuristics by tuning placement cost functions.\nTo avoid manual feature engineering, LSTM-based model automatically learns compiler heuristics from raw code, which constructs proper embeddings of programs and simultaneously learn the optimization process.\n\\textcircled{\\small{3}} For choosing the appropriate order to apply different optimizations, NEAT can automatically generate beneficial optimization orderings for each method in a program.\n\\textcircled{\\small{4}} For path profiling, CrystalBall uses an LSTM model to statically identify hot paths, the sequences of instructions that are frequently executed. As CrystalBall only relies on IRs, it avoids manual feature crafting and is independent of language or platform. \n\\textcircled{\\small{5}} For automatic vectorization, Mendis~\\textit{et al.} leverage imitation learning to mimic optimal solutions provided by superword-level-parallelism based vectorization .\n\\textcircled{\\small{6}} For compilation of specific applications, there are studies improving compilation for approximate computing or DNN applications.\nConsidering compilation for approximate computing, Esmaeilzadeh \\textit{et al.} propose a program transformation method, which trains MLPs to mimic regions of approximable code and eventually replaces the original code with trained MLPs. The following work extends this algorithmic transformation to GPUs.\nConsidering compilation for DNNs, RELEASE utilizes PPO to search optimal compilation configurations for DNNs.\nEGRL optimizes memory placement of DNN tensors during compilation, which combines GNNs, RL, and evolutionary search to figure out optimal mapping onto different on-board memory components (i.e., SRAM, LLC, and DRAM).","id":"9d772652-3e90-4f57-98a3-ad3bb92cfe28","level":"subsubsection","origin_cites_number":12,"parent_id":"68c1d9ca-e725-4975-ad78-356d7848d1d1","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Code Generation and Compiler"],["subsubsection","Compiler"]],"subsections":[],"title":"Compiler"},{"cite_extract_rate":0,"cites":[],"content":"\\label{chip_design}\nAs technology scales down, the increased design complexity comes with growing process variations and reduced design margins, making chip design an overwhelmingly complex problem for human designers.\nRecent advancements in ML create a chance to transform chip design workflows.","id":"e25c32d8-64dc-4364-8501-8063014c67c6","level":"subsection","origin_cites_number":0,"parent_id":"87e1e057-9e86-4a5c-afa3-431c75b16a64","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Chip Design and Design Automation"]],"subsections":["4b384d57-c0f5-44ec-ae75-2d05825729b5","75c9f4ed-ccff-4fa2-af01-38a78105e680"],"title":"Chip Design and Design Automation"},{"cite_extract_rate":0.30000000000000004,"cites":[5414,727,5412,5413,5415,166],"content":"\\label{sec:dse_analog}\nCompared with the highly automated digital design counterpart, analog design usually demands many manual efforts and domain expertise.\nFirst, analog circuits have large design spaces to search proper topology and device sizes. Second, there is an absence of a general framework to optimize or evaluate analog designs, and design specifications often vary case by case.\nRecently, ML techniques have been introduced to expedite analog design automation.\nWe discuss these studies following the top-down flow of analog design: in the circuit level, a proper circuit topology is selected to satisfy system specifications; then in the device level, device sizes are optimized subject to various objectives. These two steps compose of pre-layout designs.\nAfter circuit schematics are carefully designed, analog layouts in the physical level are generated.\n\\textcircled{\\small{1}} \nIn the circuit level, there is an attempt towards automatic circuit generation currently targeting two-port linear analog circuits .\nThe design specifications are encoded by a hypernetwork to generate weights for an RNN model, which is trained to select circuit components and their configurations.\n\\textcircled{\\small{2}} \nIn the device level, the combination of RL and GNNs enables automatic transistor sizing , which is able to generalize across different circuit topologies or different technology nodes.\nAutoCkt introduces transfer learning techniques into deep RL for automatic sizing, achieving 40$\\times$ speedup over a traditional genetic algorithm.\nRosa \\textit{et al.} provide comprehensive discussions of how to address automatic sizing and layout of analog ICs via deep learning and ANNs.\n\\textcircled{\\small{3}}\nIn the physical level, GeniusRoute automates analog routing through the guidance learned by a generative neural network.\nThe analog placements and routing are represented as images to pass through a variational autoencoder (VAE) to learn routing likelihoods of each region. GeniusRoute achieves competitive performance to manual layouts and is capable to generalize to circuits of different functionality.\nLiu \\textit{et al.} apply multi-objective Bayesian optimization to optimize combinations of net weighting parameters, which could significantly change floor plans and placement solutions, so as to improve analog layouts of building block circuits.\n\\begin{table}[tbp]\n\\vspace{-12pt}\n\\caption{Summary of applying ML techniques as the design methodology for design automation.}\n\\vspace{-10pt}\n\\label{table:dse_eda}\n\\centering\n \\tiny\n \\renewcommand{\\arraystretch}{0.9}\n \\setlength{\\tabcolsep}{12pt}\n\\begin{tabular}{c|c|c|c}\n\\toprule\n\\multicolumn{2}{c|}{\\textbf{Domain}} & \\textbf{Task} & \\textbf{Technique} \\\\ \\midrule\n\\multirow{4}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Analog \\\\ Design \\\\ $( \\S$ \\ref{sec:dse_analog}$)$ \\end{tabular}}} & \\textbf{Circuit Level} & Generating circuit topology & RNN and hypernetwork \\\\ \\cline{2-4} \n & \\textbf{Device Level} & Device sizing & Actor critic , ANN \\\\ \\cline{2-4} \n & \\multirow{2}{*}{\\textbf{Physical Level}} & Routing & VAE \\\\ \\cline{3-4} \n & & Optimizing layout configurations & Multi-objective Bayesian optimization \\\\ \\cline{2-4} \n \\hline\n\\multirow{12}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Digital \\\\ Design \\\\ $( \\S$ \\ref{sec:dse_digital}$)$\\end{tabular}}} & \\multirow{3}{*}{\\textbf{HLS}} & Optimizing loop unrolling pragma & Random forest \\\\ \\cline{3-4} \n & & Optimizing placement of multiple pragmas & Bayesian optimization \\\\ \\cline{3-4} \n & & Optimizing resource pragma & Actor-critic and GNN \\\\ \\cline{2-4} \n & \\multirow{3}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Logic \\\\ Synthesis\\end{tabular}}} & Selecting proper optimizers & MLP \\\\ \\cline{3-4} \n & & Logic optimization & Policy gradient , actor-critic \\\\ \\cline{3-4} \n & & Determining the maximum error of each node & Q-learning \\\\ \\cline{2-4} \n & \\multirow{5}{*}{\\textbf{\\begin{tabular}[c]{@{}c@{}}Physical \\\\ Synthesis\\end{tabular}}} & Optimizing flip-flop placement in clock networks & K-means clustering \\\\ \\cline{3-4} \n & & Optimizing clock tree synthesis & Conditional GAN \\\\ \\cline{3-4} \n & & Optimizing memory cell placement & PPO \\\\ \\cline{3-4} \n & & Optimizing standard cell placement & Cast as an NN training problem \\\\ \\cline{3-4} \n & & Fix design rule violations & PPO \\\\ \\bottomrule\n\\end{tabular}\n\\vspace{-12pt}\n\\end{table}","id":"4b384d57-c0f5-44ec-ae75-2d05825729b5","level":"subsubsection","origin_cites_number":20,"parent_id":"e25c32d8-64dc-4364-8501-8063014c67c6","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Chip Design and Design Automation"],["subsubsection","Analog Design"]],"subsections":[],"title":"Analog Design"},{"cite_extract_rate":0.21428571428571402,"cites":[5414,5415,5413],"content":"\\label{sec:dse_digital}\nFor the studies applying ML techniques to directly optimize digital designs, we organize them following a top-down flow, i.e., HLS, logic synthesis, and physical synthesis.\nThe design space exploration in HLS designs usually relates to properly assigning directives (pragmas) in high-level source code, since directives significantly impact the quality of HLS designs by controlling parallelism, scheduling, and resource usage.\nThe optimization goal is often to find Pareto solutions between different objectives or to satisfy pre-defined constraints. \nWith IR analysis, the employment of a random forest is able to select suitable loop unrolling factors to optimize a weighted sum of execution latency and resource usage .\nProspector uses Bayesian optimization to optimize placement of directives (loop unrolling/pipelining, array partitioning, function inlining, and allocation), aiming to find Pareto solutions between execution latency and resource utilization in FPGAs. \nIronMan targets Pareto solutions between different resources while keeping the latency unchanged.\nIt combines GNNs with RL to conduct a finer-grained design exploration in the operation level, pursuing optimal resource allocation strategies by optimizing assignments of the resource pragma.\nIn logic synthesis, RTL-designs or logic networks are represented by directed acyclic graphs (DAGs). The goal is to optimize logic networks subject to certain constraints.\nLSOracle employs an MLP to automatically decide which one of the two optimizers should be applied on different parts of circuits.\nThe logic optimization can be formulated as an RL problem solved by the policy gradient or the advantage actor-critic : the state is the current status of a design; the action is a transformation between two DAGs with equivalent I/O behaviors; the optimization objective is to minimize area or delay of designs.\nQ-ALS aims at approximate logic synthesis and embeds a Q-learning agent to determine the maximum tolerable error of each node in a DAG, such that the total error rates at primary outputs are bounded by pre-specified constraints.\nIn physical synthesis, placement optimization is a popular topic.\n\\textcircled{\\small{1}} \nTo optimize flip-flop placement in clock networks, Wu \\textit{et al.} apply a modified K-means clustering to group post-placement flip-flops, and relocate these clusters by reducing the distance between flip-flops and their drivers while minimizing disruption of original placement results. \nTo optimize clock tree synthesis (CTS), Lu \\textit{et al.} train a regression model that takes pre-CTS placement images and CTS configurations as inputs to predict post-CTS metrics (clock power, clock wirelength, and maximum skew), which is used as the supervisor to guide the training of a conditional GAN, such that the well-trained generator can recommend CTS configurations leading to optimized clock trees. \n\\textcircled{\\small{2}} \nAiming at cell placement, a deep RL approach is introduced to place macros (memory cells), after which standard cells are placed by a force-directed method. \nThis method is able to generalize to unseen netlists, and outperforms RePlAce yet several times slower.\nDREAMPlace casts the analytical standard cell placement optimization into a neural network training problem, achieving over 30$\\times$ speedup without quality degradation compared to RePlAce.\nNVCell is an automated layout generator for standard cells, which employs RL to fix DRVs after placement and routing.\n\\textcircled{\\small{3}} \nML-based techniques demonstrates their versatility in many design automation tasks, such as post-silicon variation extraction by sparse Bayesian learning, and post-silicon timing tuning to mitigate the effects caused by process variation .","id":"75c9f4ed-ccff-4fa2-af01-38a78105e680","level":"subsubsection","origin_cites_number":14,"parent_id":"e25c32d8-64dc-4364-8501-8063014c67c6","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","ML as Design Methodology"],["subsection","Chip Design and Design Automation"],["subsubsection","Digital Design"]],"subsections":[],"title":"Digital Design"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:dis}\nIn this section, we discuss limitations and potentials of ML techniques for computer architecture and systems, which span the entire development and deployment stack that involves data, algorithms, implementation, and targets.\nWe also envision that the application of ML techniques could be the propulsive force for \\textit{hardware agile development}.\n\\vspace{-5pt}","id":"a80b2264-c6ef-4ec9-ae9a-838d036e281c","level":"section","origin_cites_number":0,"parent_id":"5631fef3-6389-407d-9101-745f45240141","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Discussion and Potential Directions"]],"subsections":["4da21a0b-f62c-4e97-84e3-786911c6b5c8","bf55ec88-5f58-456f-b066-c5e92f556071","ac2615bc-0a08-4fa7-932a-166c6e497c28","56d8fa9d-3bba-4c5d-9a9d-e6d1744157e3","5afb9503-6a23-454b-90e9-dd19be81420a"],"title":"Discussion and Potential Directions"},{"cite_extract_rate":0.16666666666666602,"cites":[5416],"content":"Data are the backbone to ML, however, perfect datasets are sometimes non-available or prohibitively expensive to obtain in computer architecture and system domain.\nHere, we would like to scrutinize two points, the gap between small data and big data, and non-perfect data.\n\\textcircled{\\small{1}} \nIn some EDA problems, such as placement and routing in physical synthesis, the simulation or evaluation is extremely expensive , leading to data scarcity. As ML models usually require enough data to learn underlying statistics and make decisions, this gap between small data and big data often limits the capability of ML-based techniques. There have been different attempts to bridge this gap. From the algorithm side, algorithms that can work with small data await to be developed, where one current technique is Bayesian optimization that is effective in small parameter space ; active learning , which significantly improve sample efficiency, may also be a cure to this problem. From the data side, generative methods can be used to generate synthetic data , mitigating data scarcity.\n\\textcircled{\\small{2}} \nRegarding non-perfect data, even if some EDA tools produce a lot of data, they are not always properly labeled nor presented in the form suitable to ML models. In the absence of perfectly labeled training data, possible alternatives are to use unsupervised learning, self-supervised learning , or to combine supervised with unsupervised techniques . Meanwhile, RL could be a workaround where training data can be generated on-the-fly.\n\\vspace{-5pt}","id":"4da21a0b-f62c-4e97-84e3-786911c6b5c8","level":"subsection","origin_cites_number":6,"parent_id":"a80b2264-c6ef-4ec9-ae9a-838d036e281c","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Discussion and Potential Directions"],["subsection","Bridging Data Gaps"]],"subsections":[],"title":"Bridging Data Gaps"},{"cite_extract_rate":0.45454545454545403,"cites":[5202,5413,5416,5407,3588],"content":"Despite the current achieved accomplishments, we are still expecting novel ML algorithms or schemes to further improve system modeling and optimization, with respect scalability, domain knowledge interpretability, etc.\n\\textbf{New ML Schemes.}\nClassical analytic-based methods usually adopt a bottom-up or top-down procedure, encouraging ML-based techniques to distill hierarchical structures of systems/architecture.\nOne example is hierarchical RL that has flexible goal specifications and learns goal-directed behaviors in complex environments with sparse feedback. Such kind of models enables more flexible and effective multi-level design and control.\nAdditionally, many system optimizations involve participation of multiple agents, such as NoC routing, which are naturally suitable to the realm of multi-agent RL . These agents can be fully cooperative, fully competitive, or a mix of the two, enabling versatility of system optimization.\nAnother promising approach is self-supervised learning , beneficial in both improving model robustness and mitigating data scarcity.\nWhile applying a single ML method solely has led to powerful results, hybrid methods, i.e., combining different ML techniques or combining ML techniques with heuristics, unleash more opportunities. For example, RL can be combined with genetic algorithms for hardware resource assignment .\n\\textbf{Scalability.}\nThe system scaling-up poses challenges on scalability issues.\nFrom the algorithm side, multi-level techniques can help reduce the computation complexity, e.g., multi-level Q-learning for DVFS .\nOne implicit workaround is to leverage transfer learning:\nthe pre-training is a one-time cost, which can be amortized in each future use; the fine-tuning provides flexibility between a quick solution from the pre-trained model and a longer yet better one for a particular task. \nSeveral examples are discussed in Section \\ref{chip_design}.\n\\textbf{Domain Knowledge and Interpretability.}\nMaking better use of domain knowledge unveils possibilities to choose more proper models for different system problems and provide more intuitions or explanations of why and how these models work.\nBy making analogy of semantics between memory access patterns/program languages and natural languages, the prefetching or code generation problems can be modeled as NLP problems, as discussed in Section \\ref{cache} and Section \\ref{code_generation}. \nBy making analogy of graphical representations in many EDA problems, where data are intrinsically presented as graphs (e.g., circuits, logic netlists or IRs), GNNs are expected to be powerful in these fields . \nSeveral examples are provided in Section \\ref{sec:model_eda} and Section \\ref{chip_design}.","id":"bf55ec88-5f58-456f-b066-c5e92f556071","level":"subsection","origin_cites_number":11,"parent_id":"a80b2264-c6ef-4ec9-ae9a-838d036e281c","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Discussion and Potential Directions"],["subsection","Developing Algorithms"]],"subsections":[],"title":"Developing Algorithms"},{"cite_extract_rate":0.5714285714285711,"cites":[5418,7633,9139,5417],"content":"To fully benefit from ML-based methods, we need to consider practical implementations, appropriate selection of deployment scenarios, and post-deployment model maintenance.\n\\textbf{Better Implementations.}\nTo enable practical implementations of ML-based techniques, improvement can be made from either the model side or software/hardware co-design .\nFrom the model level, network pruning and model compression reduce the number of operations and model size ; weight quantization improves computation efficiency by reducing the precision of operations/operands .\nFrom the co-design level, strategies that have been used for DNN acceleration could also be used in applying ML for system.\n\\textbf{Appropriate Scenarios: online vs. offline.}\nWhen deploying ML-based techniques for system designs, it is crucial to deliberate design constraints under different scenarios.\nGenerally, existing work falls into two categories. \n\\textcircled{\\small{1}} \nML-based techniques are deployed online or during runtime, no matter the training phase is performed online or offline. Obviously, the model complexity and runtime overhead are often strictly limited by specific constraints, e.g., power/energy, timing/latency, area, etc. \nTo take one more step, if the online training/learning is further desired, the design constraint will be more stringent. One promising approach is to employ semi-online learning models, which have been applied to solve some classical combinatorial optimization problems, such as bipartite matching and caching . These models enable smooth interpolation between the best possible online and offline training algorithms.\n\\textcircled{\\small{2}} ML-based techniques are applied offline, which often refers to architectural design space exploration. Such problems leverage ML-based techniques to guide system implementation, and once the designing phase is completed, ML models will not be invoked again. Thus, the offline applications can tolerate relatively higher overheads.\n\\textbf{Model Maintenance.}\nIn the case of offline training and online deployment, ML models employed for computer architecture domain, as in other scenarios, require regular maintenance and updating to meet performance expectations,\nsince workload variations over time and hardware aging often cause data drift or concept drift .\nTo proactively circumvent performance degradation of ML models, some measures could be taken during post-deployment periods.\n\\textcircled{\\small{1}}\nML models can be retrained either at a regular interval or when key performance indicators are below certain thresholds. Retraining models regularly, regardless of their performance, is a more direct way, but it requires a clear understanding of how frequently a model should be updated under its own scenario. The model performance will decline if retraining intervals are too spaced out in the interim. Monitoring key performance indicators relies on a comprehensive panel of measurements that explicitly demonstrate model drift, whereas this may introduce additional hardware/software overhead and incorrect selection of measurements often defeats the intention of this method.\n\\textcircled{\\small{2}} During the retraining of ML models, there is often a trade-off between newly collected data and previous data. Properly assigning importance of input data would improve retraining efficacy .\n\\vspace{-5pt}","id":"ac2615bc-0a08-4fa7-932a-166c6e497c28","level":"subsection","origin_cites_number":7,"parent_id":"a80b2264-c6ef-4ec9-ae9a-838d036e281c","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Discussion and Potential Directions"],["subsection","Improving Implementations and Deployments"]],"subsections":[],"title":"Improving Implementations and Deployments"},{"cite_extract_rate":1,"cites":[1695,7739,7612],"content":"ML-based techniques are supposed to be applicable in both current architectures and emerging systems, leading to long-term advancement in computer architecture and systems.\n\\textbf{Non-homogeneous Components.}\nDesign and development for computer architectures are often based upon earlier-generation architectures of similar purpose, but commonly rely on next-generation hardware components that were not present in earlier generations. Examples include employment of new device nodes with technology scaling, and replacement of conventional constituents in memory systems with NVM- or PIM-based components.\nIn addition to the heterogeneity of components from different generations, one architecture or system usually consists of both standard parts from library and specialized/customized hardware components.\nThis provides the motivation that ML-assisted architectures/systems should have the flexibility to transfer among different-generation components, and to support standard and specialized parts simultaneously.\n\\textbf{Non-homogeneous Applications.}\nIn computer architecture and system design, some issues are universal, while others may arise with the advent of new architecture/systems and new workloads.\n\\textcircled{\\small{1}} For evergreen design areas, several examples include caching in hardware/software/data centers (Section \\ref{cache} and Section \\ref{data center}), resource management and task allocation in single/multi-/many-core CPUs and heterogeneous systems (Section \\ref{resource}), NoC design under various scenarios (Section \\ref{NoC}), etc.\n\\textcircled{\\small{2}} For problems aroused from new systems/workloads, transfer learning and meta-learning could be helpful in either exploring new heuristics or directly deriving design methodology. \nFor example, combining meta-learning with RL allows training a \"meta\" agent that is designed to adapt to a specific workload with only a few observations.\n\\vspace{-5pt}","id":"56d8fa9d-3bba-4c5d-9a9d-e6d1744157e3","level":"subsection","origin_cites_number":3,"parent_id":"a80b2264-c6ef-4ec9-ae9a-838d036e281c","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Discussion and Potential Directions"],["subsection","Supporting Non-homogeneous Tasks"]],"subsections":[],"title":"Supporting Non-homogeneous Tasks"},{"cite_extract_rate":1,"cites":[4334],"content":"Even though ML-based modeling significantly reduces the evaluation cost during design iteration, making great strides towards the landing of hardware agile development, there is still a long way to go in the ML-based design methodology prospective.\nOne ultimate goal might be the fully automated design, which should entangle two core capabilities: holistic optimization in system-wise, and easy migration across different systems, to enable rapid and agile hardware design.\n\\textbf{Holistic Optimization.}\nFueled by recent advancements, ML techniques have been increasingly explored and exploited in computer system design and optimization . \nThe target problems that await further endeavors could be multi-objective optimizations under highly constrained situations, or optimizing several components in a system simultaneously.\nWe envisage an ML-based system-wise and holistic framework with a panoramic vision: it should be able to leverage information from different levels of systems in synergy, so that it could thoroughly characterize system behaviors as well as their intrinsically hierarchical abstractions; it should also be able to make decisions in different granularity, so that it could control and improve systems precisely and comprehensively.\n\\textbf{Portable, Rapid, and Agile.}\nStriving for portable, rapid, and agile hardware design, there are two potential directions.\n\\textcircled{\\small{1}} The well-designed interfaces between systems/architectures and ML-based techniques would facilitate the portability across different platforms, since ML models can perform well without explicit descriptions of the target domain.\n\\textcircled{\\small{2}} The proliferation of ML-based techniques have more or less transformed the workflow of design automation, directly driving rapid and agile hardware design.\nWe expect GNNs make better use of naturally graphical data in EDA field; we expect deep RL be a powerful and general-purpose tool for many EDA optimization problems, especially when the exact heuristic or objective is obscure; we expect these ML-based design automation tools enhance designers' productivity and thrive in the community.\n\\vspace{-5pt}","id":"5afb9503-6a23-454b-90e9-dd19be81420a","level":"subsection","origin_cites_number":1,"parent_id":"a80b2264-c6ef-4ec9-ae9a-838d036e281c","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Discussion and Potential Directions"],["subsection","Facilitating General Tool Design and Hardware Agile Development"]],"subsections":[],"title":"Facilitating General Tool Design and Hardware Agile Development"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:con}\nThe flourishing of ML would be retarded without the great systems and powerful architectures supportive to run these algorithms at scale. \nNow, it is the time to return the favor and let ML transform the way that computer architecture and systems are designed.\nExisting work that applies ML for computer architecture/systems roughly falls into two categories: ML-based fast modeling that involves performance metrics or some other criteria of interest, and ML-based design methodology that directly leverages ML as the design tool.\nWe hope to see the virtuous cycle, in which ML-based techniques are efficiently running on the most powerful computers with the pursuit of designing the next generation computers. We hope ML-based techniques could be the impetus to the revolution of computer architecture and systems.\n\\vspace{-5pt}\n\\bibliographystyle{ACM-Reference-Format}\n\\bibliography{ref}\n\\end{document}\n\\endinput","id":"79dc37a5-e0c0-498b-b911-e6bc545700a1","level":"section","origin_cites_number":0,"parent_id":"5631fef3-6389-407d-9101-745f45240141","prefix_titles":[["title","A Survey of Machine Learning for Computer Architecture and Systems"],["section","Conclusion"]],"subsections":[],"title":"Conclusion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"5631fef3-6389-407d-9101-745f45240141\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ]\n ],\n \"subsections\": [\n \"5359b569-700e-4932-aef8-e103109cdfb1\",\n \"06a8a978-bb13-405e-8614-2ae9a01d5c56\",\n \"15e5a658-0b25-4890-8335-c1ec062ad3a2\",\n \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"a80b2264-c6ef-4ec9-ae9a-838d036e281c\",\n \"79dc37a5-e0c0-498b-b911-e6bc545700a1\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Machine learning (ML) has been doing wonders in many fields.\\nAs people are seeking better artificial intelligence (AI), there is a trend towards larger, more expressive, and more complex models.\\nAccording to the data reported by OpenAI , from 1959 to 2012, the amount of compute used in the largest AI training runs doubles every two years; since 2012, deep learning starts taking off, and the required amount of compute has been increasing exponentially with a 3.4-month doubling period. By comparison, Moore’s law , the principle that has powered the integrated-circuit revolution since 1960s, doubles the transistor density every 18 months.\\nWhile Moore's law is approaching its end , more pressure is put on innovations of computer architecture and systems, so as to keep up with the compute demand of AI applications.\\nConventionally, computer architecture/system designs are made by human experts based on intuitions and heuristics, which requires expertise in both ML and architecture/system.\\nMeanwhile, these heuristic-based designs can not guarantee scalability and optimality, especially in the case of increasingly complicated systems.\\nAs such, it seems natural to move towards more automated and powerful methodologies for computer architecture and system design, and the relationship between ML and system design is being reconsidered.\\nOver the past decade, architecture and systems are optimized to accelerate the execution and improve the performance of ML models.\\nRecently, there have been signs of emergence of applying ML for computer architecture and systems, which embraces a twofold meaning: \\\\textcircled{\\\\small{1}} the reduction of burdens on human experts designing systems manually, so as to improve designers' productivity, and \\\\textcircled{\\\\small{2}} the close of the positive feedback loop, i.e., architecture/systems for ML and simultaneously ML for architecture/systems, formulating a virtuous cycle to encourage improvements on both sides.\\n\\\\begin{figure}[tbp]\\n \\\\centering\\n \\\\vspace{-12pt}\\n\\t\\\\includegraphics[width=0.95\\\\linewidth]{fig0.png}\\n\\t\\\\vspace{-8pt}\\n \\\\caption{A comprehensive overview of applying ML for computer architecture and systems. Existing work roughly falls into two categories: ML for fast system modeling, and ML as design methodology.}\\n \\\\label{fig:frame}\\n \\\\vspace{-20pt}\\n\\\\end{figure}\\nExisting work related to applying ML for computer architecture and system design falls into two categories.\\n\\\\textcircled{\\\\small{1}} ML techniques are employed for \\\\textbf{fast and accurate system modeling}, which involves performance metrics or some criteria of interest (e.g. power consumption, latency, throughput, etc.).\\nDuring the process of designing systems, it is necessary to make fast and accurate predictions of system behaviors.\\nTraditionally, system modeling is achieved through the forms of cycle-accurate or functional virtual platforms, and instruction set simulators (e.g. gem5 ).\\nEven though these methods provide accurate estimations, they bring expensive computation costs associated with performance modeling, which limits the scalability to large-scale and complex systems; meanwhile, the long simulation time often dominates design iteration, making it impossible to fully explore the design space.\\nBy contrast, ML-based modeling and performance prediction are capable to balance simulation cost and prediction accuracy.\\n\\\\textcircled{\\\\small{2}} ML techniques are employed as \\\\textbf{a design methodology to directly enhance architecture/system design}.\\nML techniques are skilled at extracting features that might be implicit to human experts, making decisions without explicit programming, and improving themselves automatically with accumulated experience.\\nTherefore, applying ML techniques as design tools can explore design space proactively and intelligently, and manage resource through better understanding of the complicated and non-linear interactions between workloads and systems, making it possible to deliver truly optimal solutions.\\nIn this paper, we present a comprehensive overview of applying ML for computer architecture and systems. \\nAs depicted in Figure \\\\ref{fig:frame},\\nwe first perform a high-level taxonomy by considering the typical role that ML techniques take in architecture/system design, i.e., either for fast predictive modeling or as the design methodology;\\nthen, we summarize the common problems in architecture/system design that can be solved by ML techniques, and the typical ML techniques employed to resolve each of them.\\nIn addition to emphasis on computer architecture in a narrow sense, we adopt the concept that data centers can be recognized as warehouse-scale computers , and review studies associated with data center management; we provide sketchy discussions on adjacent computer systems, such as code generation and compiler; we also give attention to how ML techniques can aid and transform design automation that involves both analog and digital circuits.\\nAt the end of the paper, we discuss challenges and future prospects of applying ML for architecture/system design, aiming to convey insights of design considerations.\",\n \"id\": \"5359b569-700e-4932-aef8-e103109cdfb1\",\n \"level\": \"section\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"5631fef3-6389-407d-9101-745f45240141\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0.210526315789473,\n \"cites\": [\n 1390,\n 553,\n 2219,\n 166\n ],\n \"content\": \"\\\\label{sec:ml techniques}\\nThere are three general frameworks in ML: supervised learning, unsupervised learning and reinforcement learning.\\nThese frameworks mainly differentiate on what data are sampled and how these sample data are used to build learning models.\\nTable \\\\ref{table:ml_techniques} summarizes the commonly used ML techniques for computer architecture and system designs.\\nSometimes, multiple learning models may work well for one given problem, and the appropriate selection can be made based on available hardware resource and data, implementation overheads, performance targets, etc.\\n\\\\begin{table}[tbp]\\n\\\\vspace{-12pt}\\n\\\\caption{Machine learning techniques.}\\n\\\\vspace{-12pt}\\n\\\\label{table:ml_techniques}\\n\\\\centering\\n \\\\tiny\\n \\\\renewcommand{\\\\arraystretch}{1}\\n \\\\setlength{\\\\tabcolsep}{6pt}\\n\\\\begin{tabular}{c|c|c|c}\\n\\\\toprule\\n\\\\textbf{Realm of ML} & \\\\textbf{Category} & \\\\makecell[c]{\\\\textbf{Classical ML}} & \\\\textbf{Deep Learning Counterpart } \\\\\\\\ \\\\midrule\\n\\\\multirow{10}{*}{\\\\makecell*[c]{\\\\textbf{Supervised} \\\\\\\\ \\\\textbf{Learning}}} & \\\\multirow{1}{*}{Classification} &\\n Logistic regression & \\n \\\\multirow{8}{*}{\\\\makecell{CNN, RNN,\\\\\\\\ GNN , etc.}} \\\\\\\\ \\\\cline{2-3}\\n & \\\\multirow{6}{*}{Working for both} & Support vector machines/regression & \\\\\\\\ \\\\cline{3-3}\\n && K-nearest neighbors & \\\\\\\\ \\\\cline{3-3}\\n && Decision tree, e.g., CART , MARS & \\\\\\\\ \\\\cline{3-3}\\n && ANN & \\\\\\\\ \\\\cline{3-3}\\n && Bayesian analysis & \\\\\\\\ \\\\cline{3-3}\\n && \\\\makecell{Ensemble learning , e.g., gradient boosting, random forest} & \\\\\\\\ \\\\cline{2-3}\\n & \\\\multirow{2}{*}{Regression} & \\\\makecell{Linear regression with variants , e.g., lasso (L1 regularization),\\\\\\\\ ridge (L2 regularization), elastic-net (hybrid L1/L2 regularization)} & \\\\\\\\ \\\\cline{3-3}\\n && Non-linear regression & \\\\\\\\\\n \\\\hline\\n\\\\multirow{2}{*}{\\\\makecell[c]{\\\\textbf{Unsupervised} \\\\\\\\ \\\\textbf{Learning}}} \\n& Clustering & K-means clustering & \\\\multirow{2}{*}{\\\\makecell[c]{Autoencoder, \\\\\\\\ GAN, etc.}} \\\\\\\\ \\\\cline{2-3}\\n & Dimension reduction & Principal component analysis (PCA) &\\\\\\\\ \\\\hline\\n\\\\multirow{3}{*}{\\\\makecell*[c]{\\\\textbf{Reinforcement} \\\\\\\\ \\\\textbf{Learning}}} &\\n Value-based & Q-learning & DQN \\\\\\\\ \\n \\\\cline{2-4} \\n & \\\\multirow{2}{*}{Policy-based} & Actor-critic \\n & A3C , DDPG \\\\\\\\ \\\\cline{3-4}\\n & & Policy gradient, e.g., REINFORCE \\n & PPO \\\\\\\\\\n\\\\bottomrule\\n\\\\end{tabular}\\n\\\\vspace{-15pt}\\n\\\\end{table}\\n\\\\vspace{-5pt}\",\n \"id\": \"06a8a978-bb13-405e-8614-2ae9a01d5c56\",\n \"level\": \"section\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"5631fef3-6389-407d-9101-745f45240141\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Different ML Techniques\"\n ]\n ],\n \"subsections\": [\n \"f49480df-db26-480b-8149-671064eda13f\",\n \"c3e4e81a-5169-40c5-91a3-4a2389c943b1\",\n \"218b240c-27bf-4571-953a-c14101e51384\"\n ],\n \"title\": \"Different ML Techniques\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 166\n ],\n \"content\": \"Supervised learning is the process of learning a set of rules able to map an input to an output based on labeled datasets. These learned rules can be generalized to make predictions for unseen inputs.\\nWe briefly introduce several prevalent techniques in supervised learning, as shown in Figure \\\\ref{fig:sl}.\\n\\\\begin{itemize}[leftmargin=*]\\n \\\\item Regression is a process for estimating the relationships between a dependent variable and one or more independent variables. The most common form is linear regression , and some other forms include different types of non-linear regression . Regression techniques are primarily used for two purposes, prediction/forecasting, and inference of causal relationships.\\n \\\\item Support vector machines (SVMs) try to find the best hyperplanes to separate data classes by maximizing margins. One variant is support vector regression (SVR), which is able to conduct regression tasks. Predictions or classifications of new inputs can be decided by their relative positions to these hyperplanes.\\n \\\\item Decision tree is one representative of logical learning methods, which uses tree structures to build regression or classification models.\\n The final result is a tree with decision nodes and leaf nodes. Each decision node represents a feature and branches of this node represent possible values of the corresponding feature. Starting from the root node, input instances are classified by sequentially passing through nodes and branches, until they reach leaf nodes that represent either classification results or numerical values.\\n \\\\item Artificial neural networks (ANNs) are capable to approximate a broad family of functions: a single-layer perceptron is usually used for linear regression; complex DNNs consisting of multiple layers are able to approximate non-linear functions, such as the multi-layer perceptron (MLP); variants of DNNs that achieve excellent performance in specific fields benefit from the exploitation of certain computation operations, e.g., convolutional neural networks (CNNs) with convolution operations leveraging spatial features, and recurrent neural networks (RNNs) with recurrent connections enabling learning from sequences and histories.\\n \\\\item Ensemble learning employs multiple models that are strategically designed to solve a particular problem, and the primary goal is to achieve better predictive performance than those could be obtained from any of the constituent models alone. Several common types of ensembles include random forest and gradient boosting.\\n\\\\end{itemize}\\n\\\\begin{figure}[tbp]\\n \\\\centering\\n \\\\vspace{-15pt}\\n\\t\\\\includegraphics[width=0.97\\\\linewidth]{fig1.png}\\n\\t\\\\vspace{-10pt}\\n \\\\caption{Examples of supervised learning: (a) regression, (b) SVM, (c) decision tree, (d) MLP, and (e) ensemble learning.}\\n \\\\vspace{-20pt}\\n \\\\label{fig:sl}\\n\\\\end{figure}\\nDifferent learning models have different preference of input features: \\nSVMs and ANNs generally perform much better with multi-dimension and continuous features, while logic-based systems tend to perform better when dealing with discrete/categorical features.\\nIn system design, supervised learning is commonly used for performance modeling, configuration predictions, or predicting higher-level features/behaviors from lower-level features.\\nOne thing worth noting is that supervised learning techniques need well labeled training data prior to the training phase, which usually require tremendous human expertise and engineering.\\n\\\\vspace{-5pt}\",\n \"id\": \"f49480df-db26-480b-8149-671064eda13f\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"06a8a978-bb13-405e-8614-2ae9a01d5c56\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Different ML Techniques\"\n ],\n [\n \"subsection\",\n \"Supervised Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Supervised Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Unsupervised learning is the process of finding previously unknown patterns based on unlabeled datasets.\\nTwo prevailing methods are clustering analysis and principal component analysis (PCA) , as depicted in Figure \\\\ref{fig:usl}.\\n\\\\begin{itemize}[leftmargin=*]\\n \\\\item Clustering is a process of grouping data objects into disjoint clusters based on a measure of similarity, such that data objects in the same cluster are similar while data objects in different clusters share low similarities. The goal of clustering is to classify raw data reasonably and to find possibly existing hidden structures or patterns in datasets. One of the most popular and simple clustering algorithms is k-means clustering.\\n \\\\item PCA is essentially a coordinate transformation leveraging information from data statistics. It aims to reduce the dimensionality of the high-dimensional variable space by representing it with a few orthogonal (linearly uncorrelated) variables that capture most of its variability.\\n\\\\end{itemize}\\nSince there is no label in unsupervised learning, it is difficult to simultaneously measure the performance of learning models and decide when to stop the learning process.\\nOne potential workaround is \\\\textit{semi-supervised learning} , which uses a small amount of labeled data together with a large amount of unlabeled data. This approach stands between unsupervised and supervised learning, requiring less human effort and producing higher accuracy.\\nThe unlabeled data are used to either finetune or re-prioritize hypotheses obtained from labeled data alone.\\n\\\\begin{figure}[t]\\n\\\\vspace{-15pt}\\n\\\\flushleft\\n\\\\begin{minipage}{.63\\\\textwidth}\\n \\\\centering\\n \\\\includegraphics[width=0.82\\\\linewidth]{fig2.png}\\n \\\\vspace{-10pt}\\n \\\\caption{Examples of unsupervised learning: (a) clustering, and (b) PCA.}\\n \\\\label{fig:usl}\\n\\\\end{minipage}\\n\\\\hspace{.03\\\\textwidth}\\n\\\\begin{minipage}{.33\\\\textwidth}\\n \\\\centering\\n \\\\includegraphics[width=\\\\linewidth]{fig3.png}\\n \\\\caption{A typical framing of RL.}\\n \\\\label{fig:rl}\\n\\\\end{minipage}\\n\\\\vspace{-10pt}\\n\\\\end{figure}\",\n \"id\": \"c3e4e81a-5169-40c5-91a3-4a2389c943b1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"06a8a978-bb13-405e-8614-2ae9a01d5c56\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Different ML Techniques\"\n ],\n [\n \"subsection\",\n \"Unsupervised Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Unsupervised Learning\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"In standard reinforcement learning (RL) , an agent interacts with an environment $\\\\mathcal{E}$ over a number of discrete time steps, as shown in Figure \\\\ref{fig:rl}.\\nAt each time step $t$, the agent receives a state $s_t$ from the \\\\textit{state space} $\\\\mathcal{S}$, and selects an action $a_t$ from the \\\\textit{action space} $\\\\mathcal{A}$ according to its policy $\\\\pi$, where $\\\\pi$ is a mapping from states $s_t$ to actions $a_t$. \\nIn return, the agent receives the next state $s_{t+1}$ and a scalar reward $r_t: \\\\mathcal{S}\\\\times\\\\mathcal{A}\\\\to\\\\mathbb{R}$. \\nThis process continues until the agent reaches a terminal state after which the process restarts.\\nThe return $R_t={\\\\sum\\\\limits_{k=0}^\\\\infty \\\\gamma^kr_{t+k}}$ is the totally accumulated rewards at the time step $t$ with a \\\\textit{discount factor} $\\\\gamma\\\\in(0,1]$. \\nThe goal of the agent is to maximize the expected return for each state $s$.\\nThe state-action value $Q_{\\\\pi}(s,a)=\\\\mathbb{E_\\\\pi}\\\\lbrack R_t|s_t=s, a_t=a\\\\rbrack$ is the expected return of selecting action $a$ at state $s$ with policy $\\\\pi$.\\nSimilarly, the state value $V_{\\\\pi}(s)=\\\\mathbb{E_\\\\pi}\\\\lbrack R_t|s_t=s\\\\rbrack$ is the expected return starting from state $s$ by following policy $\\\\pi$. There are two general types of methods in RL: value-based, and policy-based.\\n\\\\begin{itemize}[leftmargin=*]\\n \\\\item In value-based RL, the state-action value function $Q_{\\\\pi}(s,a)$ is approximated by either tabular approaches or function approximations. At each state $s_t$, the agent always selects the optimal action $a^*_t$ that could bring the maximal state-action value $Q_{\\\\pi}(s_t,a^*_t)$. One well-known example of value-based methods is Q-learning.\\n \\\\item In policy-based RL, it directly parameterizes the policy $\\\\pi(a|s;\\\\theta)$ and updates the parameters $\\\\theta$ by performing gradient ascent on $\\\\mathbb{E}[R_t]$. One example is the REINFORCE algorithm. \\n\\\\end{itemize}\\nRL is modeled based on Markov decision process, and thus it is suitable to handle control problems or sequential decision-making processes.\\nWith these characteristics, RL is able to explore design space proactively and intelligently, and learn how to achieve resource management or task scheduling in system designs through interactions with environments. The optimal behaviors can be found by embedding optimization goals into reward functions.\",\n \"id\": \"218b240c-27bf-4571-953a-c14101e51384\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"06a8a978-bb13-405e-8614-2ae9a01d5c56\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Different ML Techniques\"\n ],\n [\n \"subsection\",\n \"Reinforcement Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Reinforcement Learning\"\n },\n {\n \"cite_extract_rate\": 0.04918032786885201,\n \"cites\": [\n 4378,\n 5390,\n 5389\n ],\n \"content\": \"\\\\label{sec:system modeling}\\nThis section reviews studies that employ ML techniques for fast and accurate system modeling, which involves predictions of performance metrics or some other criteria of interest.\\nAlthough cycle-accurate simulators, which are commonly used for system performance prediction, can provide accurate estimations, they usually run multiple orders of magnitude slower than native executions.\\nBy contrast, ML-based techniques can balance simulation costs and prediction accuracy, showing great potentials in exploring huge configuration spaces and learning non-linear impacts of configurations.\\nMost of existing work applies supervised learning for either pure system modeling or efficient design space exploration (DSE) enabled by fast predictions.\\nTable \\\\ref{table:model_sys} and Table \\\\ref{table:model_eda} summarize the studies for predictive modeling in computer architecture/system and design automation respectively, in terms of task domains, prediction targets, adopted ML techniques, and corresponding inputs.\\n\\\\begin{table}[tbp]\\n\\\\vspace{-10pt}\\n\\\\caption{Summary of applying ML techniques for fast modeling in computer architecture and systems.}\\n\\\\label{table:model_sys}\\n\\\\centering\\n \\\\tiny\\n \\\\renewcommand{\\\\arraystretch}{0.9}\\n \\\\setlength{\\\\tabcolsep}{1pt}\\n\\\\vspace{-10pt}\\n\\\\begin{tabular}{c|c|c|c}\\n\\\\toprule\\n\\\\textbf{Domain} & \\\\textbf{Prediction of} & \\\\textbf{Technique} & \\\\textbf{Input} \\\\\\\\ \\\\midrule\\n\\\\multirow{4}{*}{\\\\makecell[c]{\\\\textbf{Memory} \\\\\\\\ \\\\textbf{System} \\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{sec:model_memory}}$)$}} \\n& Throughput/cache miss & ANN & Cache configurations \\\\\\\\ \\\\cline{2-4} \\n& Throughput/lifetime/energy & \\\\makecell[c]{Gradient boosting/quadratic regression with lasso } & NVM configurations \\\\\\\\ \\\\cline{2-4} \\n& Throughput & CNN & Memory controller placements \\\\\\\\ \\\\cline{2-4} \\n& Disk block correlation & DNN & Data blocks in the context window \\\\\\\\ \\\\hline\\n\\\\multirow{4}{*}{\\\\makecell[c]{\\\\textbf{NoC} \\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{sec:model_noc}}$)$}} \\n& Latency & SVR & Queue information \\\\\\\\ \\\\cline{2-4} \\n& Hotspots & ANN & Buffer utilization \\\\\\\\ \\\\cline{2-4} \\n& Buffer/link utilization & \\\\makecell[c]{Regression with ridge regularization , decision tree } & NoC configurations \\\\\\\\ \\\\cline{2-4} \\n& Probability of errors & Decision tree & \\\\makecell[c]{Link utilization and transistor wearing-out} \\\\\\\\ \\\\hline\\n\\\\multirow{6}{*}{\\\\makecell[c]{\\\\textbf{GPU} \\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{sec:model_gpu}}$)$}} & \\\\makecell[c]{Speedup or execution time \\\\\\\\ by cross-platform inputs} & \\\\makecell[c]{Nearest neighbor/SVM , \\\\\\\\ ensemble of regression-based learners , \\\\\\\\ random forest } & \\\\makecell[c]{Static analysis and/or \\\\\\\\ dynamic profiling of \\\\\\\\ source CPU code} \\\\\\\\ \\\\cline{2-4} \\n & Execution time & \\\\makecell[c]{Stepwise regression , \\\\\\\\ ensemble of linear regression and random forest } & \\\\multirow{3}{*}{\\\\makecell[c]{GPU configurations \\\\\\\\ and performance counters}} \\\\\\\\ \\\\cline{2-3}\\n & Throughput/power & Ensemble of NNs & \\\\\\\\ \\\\cline{2-3}\\n & Scaling behavior of GPGPU & ANN and K-means clustering & \\\\\\\\ \\\\cline{2-4} \\n & Kernel affinity and execution time & Logistic regression and linear regression & Kernel characteristics \\\\\\\\ \\\\cline{2-4}\\n & Traffic patterns in GPGPU & CNN & Grayscale heat maps \\\\\\\\ \\\\hline\\n\\\\multirow{2}{*}{\\\\makecell[c]{\\\\textbf{Single-Core} \\\\\\\\ \\\\textbf{CPU} $( \\\\S$ \\\\textbf{\\\\ref{sec:model_single}}$)$}} \\n& Throughput & \\\\makecell[c]{Linear regression , non-linear regression } & \\\\multirow{2}{*}{\\\\makecell[c]{Micro-architectural parameters \\\\\\\\ and performance counters}} \\\\\\\\ \\\\cline{2-3}\\n & Program execution time & Linear regression & \\\\\\\\ \\\\hline\\n\\\\multirow{8}{*}{\\\\makecell[c]{\\\\textbf{General}\\\\\\\\ \\\\textbf{Modeling} \\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{sec:model_general}}$)$}} & \\\\multirow{6}{*}{Throughput/latency/power} & ANN & \\\\multirow{7}{*}{\\\\makecell[c]{Micro-architectural parameters \\\\\\\\ and performance counters}} \\\\\\\\ \\\\cline{3-3}\\n & & \\\\makecell[c]{Non-linear regression } & \\\\\\\\ \\\\cline{3-3}\\n & & Linear regression & \\\\\\\\ \\\\cline{3-3}\\n & & Hierarchical Bayesian model & \\\\\\\\ \\\\cline{3-3}\\n & & LSTM & \\\\\\\\ \\\\cline{3-3}\\n & & Generative model & \\\\\\\\ \\\\cline{2-3}\\n & \\\\makecell[c]{Slowdown caused by \\\\\\\\ application interference} & \\\\makecell[c]{Linear regression with elastic-net \\\\\\\\ regularization } & \\\\\\\\ \\\\cline{2-4} \\n & \\\\makecell[c]{Speedup of multi-thread applications} & Gausian process regression & Profiling of single-thread execution \\\\\\\\ \\\\hline\\n\\\\multirow{9}{*}{\\\\makecell[c]{\\\\textbf{Data} \\\\\\\\ \\\\textbf{Center} \\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{sec:model_dc}}$)$}} & Job completion time & SVR & \\\\makecell[c]{Application characteristics \\\\\\\\ and cluster configurations} \\\\\\\\ \\\\cline{2-4} \\n & Resource demand & \\\\makecell[c]{Statistical learning , linear regression/MLP } & \\\\multirow{3}{*}{Workload characterization} \\\\\\\\ \\\\cline{2-3}\\n & Incoming workload & ARMA , ARIMA & \\\\\\\\ \\\\cline{2-3}\\n & Workload pattern & Hidden Markov model & \\\\\\\\ \\\\cline{2-4} \\n & Power usage effectiveness & MLP & Data center configurations \\\\\\\\ \\\\cline{2-4} \\n & Disk Failure & \\\\makecell[c]{Bayesian methods , clustering , \\\\\\\\ SVM/MLP , random forest } & \\\\multirow{3}{*}{\\\\makecell[c]{SMART (Self-Monitoring, Analysis \\\\\\\\ and Reporting Technology) attributes \\\\\\\\ of data centers}} \\\\\\\\ \\\\cline{2-3}\\n & Health assessment of drives & \\\\makecell[c]{CART , gradient boosted regressions tree , RNN } & \\\\\\\\ \\\\cline{2-3}\\n & \\\\multirow{2}{*}{Partial drive failure} & \\\\makecell[c]{CART/random forest/SVM/ANN/logistic regression } & \\\\\\\\ \\\\cline{3-4} \\n & & Gradient boosted regression trees & \\\\makecell[c]{SMART attributes and system-level signals} \\\\\\\\ \\\\bottomrule\\n\\\\end{tabular}\\n\\\\vspace{-15pt}\\n\\\\end{table}\",\n \"id\": \"15e5a658-0b25-4890-8335-c1ec062ad3a2\",\n \"level\": \"section\",\n \"origin_cites_number\": 61,\n \"parent_id\": \"5631fef3-6389-407d-9101-745f45240141\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ]\n ],\n \"subsections\": [\n \"cb797a5f-f6ce-4412-9791-ccdeb4f88886\",\n \"9ac2d4c9-6bb8-44b3-aa23-38c568e05513\",\n \"419e1fc4-5eb3-40c4-b251-cc170597bc2e\"\n ],\n \"title\": \"ML for Fast System Modeling\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:model_sub_sys}\",\n \"id\": \"cb797a5f-f6ce-4412-9791-ccdeb4f88886\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"15e5a658-0b25-4890-8335-c1ec062ad3a2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"Sub-system Modeling and Performance Prediction\"\n ]\n ],\n \"subsections\": [\n \"1e898be7-db20-4db0-b960-280b242ecf62\",\n \"5ac93425-5d57-405e-be1e-d3cab4e56f18\"\n ],\n \"title\": \"Sub-system Modeling and Performance Prediction\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 5391\n ],\n \"content\": \"\\\\label{sec:model_memory}\\nIn memory systems, ML-based performance models are exploited to help explore trade-offs among different objectives.\\nTo explore non-volatile memory (NVM) based cache hierarchies, Dong \\\\textit{et al.} develop an ANN model to predict higher-level features (e.g. miss of cache read/write, and instruction-per-cycle (IPC)) from lower-level features (e.g. cache associativity, capacity and latency). \\nTo adaptively select architectural techniques in NVMs for different applications, Memory Cocktail Therapy estimates lifetime, IPC, and energy consumption through lightweight online predictors by gradient boosting and quadratic regression with lasso. \\nTo optimize memory controller placements in throughput processors, Lin \\\\textit{et al.} build a CNN model that takes memory controller placements as inputs to predict throughput, which accelerates the optimization process by two orders of magnitude.\\nSome studies concentrate on learning efficient representations of memory access patterns.\\nBlock2Vec tries to mine data block correlations by training a DNN to learn the best vector representation of each block and capturing block similarities via vector distances, which enables further optimization for caching and prefetching.\\nShi \\\\textit{et al.} use a graph neural network (GNN) to learn fused representations of static code and its dynamic execution. This unified representation is capable to model both data flows (e.g., prefetching) and control flows (e.g., branch prediction).\",\n \"id\": \"1e898be7-db20-4db0-b960-280b242ecf62\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"cb797a5f-f6ce-4412-9791-ccdeb4f88886\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"Sub-system Modeling and Performance Prediction\"\n ],\n [\n \"subsubsection\",\n \"Memory System\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Memory System\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:model_noc}\\nIn NoCs, several performance metrics of interest are latency, energy consumption, and reliability.\\n\\\\textcircled{\\\\small{1}}\\nRegarding latency predictions, Qian \\\\textit{et al.} use an SVR model to predict the traffic flow latency and the average channel waiting time in mesh-based NoCs, which relaxes some assumptions in the classical queuing theory. \\nRather than explicitly predicting latency, a lightweight hardware-based ANN predicts existence of traffic hotspots, which are intensive network congestions significantly degrading the effective throughput and implicitly indicate the average communication latency in NoCs.\\nThe input features are buffer utilization rates from neighboring NoC routers, and the trained predictor is combined with a proactive hotspot-preventive routing algorithm to avert hotspot formation, attaining significant improvements for synthetic workloads while modest melioration for real-world benchmarks.\\n\\\\textcircled{\\\\small{2}}\\nRegarding estimating energy consumption, the learned predictors are often leveraged for saving dynamic and/or static energy in NoCs.\\nDiTomaso \\\\textit{et al.} use per-router decision trees to predict link utilization and traffic direction, which are combined with sleepy link storage units to power-gate links/routers and to change link directions. \\nClark \\\\textit{et al.} use ridge regression models to predict buffer utilization, changes in buffer utilization, or a combined metric of energy and throughput,\\nbased on which a router can select proper voltage/frequency.\\nIn photonic NoCs, the ridge regression model is also applicable to predict the number of packets to be injected into each router in the following time window , based on which the number of wavelengths are scaled properly to reduce static energy consumed by photonic links.\\n\\\\textcircled{\\\\small{3}}\\nRegarding the reliability of NoCs, a per-link decision tree trained offline can predict the probability of timing faults on links during runtime , based on which a proactive fault-tolerant technique is developed to mitigate errors by using the strengthened cyclic redundancy check with error-correction code and relaxed transmission.\",\n \"id\": \"5ac93425-5d57-405e-be1e-d3cab4e56f18\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"cb797a5f-f6ce-4412-9791-ccdeb4f88886\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"Sub-system Modeling and Performance Prediction\"\n ],\n [\n \"subsubsection\",\n \"Network-on-Chip (NoC)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Network-on-Chip (NoC)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:model_sys}\\nAccurate and fast performance estimation is a necessity for system optimization and design space exploration.\\nWith the increasing complexity of systems and variety of workloads, ML-based techniques can provide highly accurate performance estimations with reasonable simulation costs, surpassing the capability of commonly-used cycle-accurate simulators that require highly computational costs and long simulation time.\",\n \"id\": \"9ac2d4c9-6bb8-44b3-aa23-38c568e05513\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"15e5a658-0b25-4890-8335-c1ec062ad3a2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"System Modeling and Performance Prediction\"\n ]\n ],\n \"subsections\": [\n \"a2990a76-751b-4214-b5a9-77c66ab0dd6a\",\n \"a74c5980-8f81-4447-be80-f6c683740060\",\n \"3890381b-cc70-4717-9bbf-fa3c92dca1c8\",\n \"37bd7027-5f70-4a17-9138-7b53f9c12df8\"\n ],\n \"title\": \"System Modeling and Performance Prediction\"\n },\n {\n \"cite_extract_rate\": 0.11111111111111101,\n \"cites\": [\n 5390\n ],\n \"content\": \"\\\\label{sec:model_gpu}\\nThere are two types of predictions for GPU modeling: cross-platform predictions and GPU-specific predictions.\\nCross-platform predictions are used to decide in advance whether to offload an application from a CPU to a GPU, since not every application benefits from GPU execution and the porting process requires considerably additional efforts;\\nGPU-specific predictions are used to estimate metrics of interest and to assist GPU design space exploration, helpful to handle design space irregularities and complicated interactions among configurations.\\nCross-platform predictions can be formulated as a binary classification problem that identifies whether the potential GPU speedup of an application would be greater than a given threshold.\\nThis task can be solved by the nearest neighbor and SVMs using dynamic instruction profiles , or a random forest that composes of one thousand decision trees using static analysis of source CPU code (i.e., memory coalescing, branch divergence, kernel size available parallelism and instruction intensities) .\\nWith both dynamic and static program properties from single-thread CPU code, an ensemble of one hundred regression-based learners can predict the GPU execution time .\\nIn terms of GPU-specific predictions that take GPU configurations and performance counters as input features, the execution time can be predicted by stepwise linear regression, which recognizes the most important input features among many GPU parameters and thus achieves high accuracy even with sparse samples ; the power/throughput can be modeled by an ensemble of NN predictors .\\nProvided with profiling results from earlier-generation GPUs, an ensemble of linear and non-linear regression models is capable to predict cross-generation GPU execution time for later/future-generation GPUs, which achieves more than 10,000 times speedup compared to cycle-accurate GPU simulators .\\nFocusing on processing-in-memory (PIM) assisted GPU architectures, Pattnaik \\\\textit{et al.} classify GPU cores into two types: powerful GPU cores but far away from memory, and auxiliary/simple GPU cores but close to memory. They develop a logistic regression model that takes kernel characteristics as input features to predict architecture affinity of kernels, aiming to accurately identify which kernels would benefit from PIM and offload them accordingly to auxiliary GPU cores. They also build a linear regression model to predict the execution time of each kernel, so that a concurrent kernel management mechanism can be developed based on these two models and kernel dependency information. \\nFocusing on general-purpose GPUs (GPGPUs), Wu \\\\textit{et al.} model kernel scaling behaviors with respect to the number of compute units, engine frequency, and memory frequency. During training, kernels with similar performance scaling behaviors are grouped by K-means clustering, and when encountering a new kernel, it is mapped to the cluster that best describes its scaling performance by an ANN-based classifier.\\nLi \\\\textit{et al.} reassess prevailing assumptions of GPGPU traffic patterns, and combine a CNN with a t-distributed stochastic neighbor embedding to classify different traffic patterns.\",\n \"id\": \"a2990a76-751b-4214-b5a9-77c66ab0dd6a\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"9ac2d4c9-6bb8-44b3-aa23-38c568e05513\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"System Modeling and Performance Prediction\"\n ],\n [\n \"subsubsection\",\n \"Graphics Processing Unit (GPU)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Graphics Processing Unit (GPU)\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:model_single}\\nIn predictive performance modeling of single-core processors, early-stage work mostly targets superscalar processors. \\nTo predict the application-specific cycle-per-instruction (CPI) of superscalar processors, Joseph \\\\textit{et al.} introduce an iterative procedure to build linear regression models using 26 key micro-architectural parameters. Later they construct predictive models by non-linear regression techniques (i.e., radial basis function networks generated from regression trees) with 9 key micro-architectural parameters .\\nIn parallel with Joseph's work, Lee and Brooks use regression modeling with cubic splines to predict application-specific performance (billions of instructions per second) and power.\\nLater work focuses on performance modeling for existing hardware (e.g., Intel, AMD, and ARM processors) by using micro-architectural parameters and performance counters.\\nEyerman \\\\textit{et al.} construct a mechanistic-empirical model for CPI predictions of three Intel processors. The initially parameterized performance model is inspired by mechanistic modeling, where the unknown parameters inside the model are derived through regression, benefiting from both mechanistic modeling (i.e., interpretability) and empirical modeling (i.e., ease of implementation).\\nZheng \\\\textit{et al.} explore two approaches to cross-platform predictions of program execution time, where profiling results on Intel Core i7 and AMD Phenom processors are used to estimate the execution time on a target ARM processor. The first approach relaxes the assumption of global linearity to local linearity in the feature space and applies constrained locally sparse linear regression; the other approach applies lasso linear regression with phase-level performance features.\",\n \"id\": \"a74c5980-8f81-4447-be80-f6c683740060\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"9ac2d4c9-6bb8-44b3-aa23-38c568e05513\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"System Modeling and Performance Prediction\"\n ],\n [\n \"subsubsection\",\n \"Single-Core Processor\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Single-Core Processor\"\n },\n {\n \"cite_extract_rate\": 0.13636363636363602,\n \"cites\": [\n 4378,\n 5392,\n 5389\n ],\n \"content\": \"\\\\label{sec:model_general}\\nRegression techniques are the mainstream to predict performance metrics from micro-architectural parameters or other features, which attributes to their capability to make high-accuracy estimations with reasonable training costs.\\nFor conventional regression-based models, ANNs and non-linear regression with different designs are the common practice to predict throughput/latency and power/energy .\\nConsequently, there are comparisons among different techniques.\\nLee \\\\textit{et al.} compare piecewise polynomial regression with ANNs, with emphasis that piecewise polynomial regression offers better explainability while ANNs show better generalization ability.\\nOzisikyilmaz \\\\textit{et al.} contrast several linear regression models and different ANNs, indicating that the pruned ANNs achieve best accuracy while requiring longer training time.\\nAgarwal \\\\textit{et al.} estimate the parallel execution speedup of multi-threaded applications on a target hardware platform, and mention that Gaussian process regression performs the best among several explored methods in this case.\\nMore recent work tends to take advantage of data-driven approaches.\\nIthemal leverages a hierarchical multi-scale RNN with long short term memory (LSTM) to predict throughput of basic blocks (i.e., sequences of instructions with no branches or jumps), and evaluations demonstrate that Ithemal is more accurate and as fast as analytical throughput estimators.\\nBy employing a variant of Ithemal as a differentiable surrogate to approximate CPU simulators, DiffTune is able to apply gradient-based optimization techniques to learn the parameters of x86 basic block CPU simulators such that simulators' error is minimized. The learned parameters finally are plugged back into the original simulator.\\nDing \\\\textit{et al.} provide some insights in learning-based modeling methods: the improvement of prediction accuracy may receive diminishing returns; the consideration of domain knowledge will be helpful for system optimizations, even if the overall accuracy may not be improved.\\nThus, they propose a generative model to handle data scarcity by generating more training data, and apply a multi-phase sampling to improve prediction accuracy of optimal configuration points.\\nML-based predictive performance modeling enables efficient resource management and rapid design space exploration to improve throughput.\\nEquipped with ANNs for IPC predictions, strategies for resource allocation and task scheduling can always select decisions that would bring the best predicted IPC.\\nESP constructs a regression model with elastic-net regularization to predict application interference (i.e., slowdown), which is integrated with schedulers to increase throughput.\\nMetaTune is a meta-learning based cost model for convolution operations, and when combined with search algorithms, it enables efficient auto-tuning of parameters during compilation.\\nIn consideration of rapid design space exploration of the uncore (i.e., memory hierarchies and NoCs), Sangaiah \\\\textit{et al.} uses a regression-based model with restricted cubic splines to estimate CPI, reducing the exploration time by up to four orders of magnitude.\\nML-based predictive performance modeling benefits adaptations between performance and power budgets.\\nLeveraging off-line multivariate linear regression to predict IPC and/or power of different architecture configurations, Curtis-Maury \\\\textit{et al.} maximize performance of OpenMP applications by dynamic concurrency throttling and dynamic voltage and frequency scaling (DVFS); Bailey \\\\textit{et al.} apply hardware frequency-limiting techniques to select optimal hardware configurations under given power constraints.\\nTo effectively apply DVFS towards various optimization goals, the designed strategy can adopt predictions for power consumption by a constrained-posynomial model or job execution time by a linear regression model .\\nTo conduct smart power management in a more general manner, LEO employs hierarchical Bayesian models to predict performance and power, and when integrated for runtime energy optimization, it is capable to figure out the performance-power Pareto frontier and select the configuration satisfying performance constraints with minimized energy.\\nCALOREE~ further breaks up the power management task into two abstractions: a learner for performance modeling and an adaptive controller leveraging predictions from the learner. These abstractions enable both the learner to use multiple ML techniques and the controller to maintain control-theoretic formal guarantees. Since no user-specified parameter except the goal is required, CALOREE is applicable even for non-experts.\",\n \"id\": \"3890381b-cc70-4717-9bbf-fa3c92dca1c8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 22,\n \"parent_id\": \"9ac2d4c9-6bb8-44b3-aa23-38c568e05513\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"System Modeling and Performance Prediction\"\n ],\n [\n \"subsubsection\",\n \"General Modeling and Performance Prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"General Modeling and Performance Prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:model_dc}\\nData centers have been in widespread use for both traditional enterprise applications and cloud services.\\nMany studies employ ML techniques to predict workload/resource-related metrics, so as to enable elastic resource provision.\\nCommon examples include but not limited to using SVR to predict job completion time , leveraging the autoregressive moving average (ARMA) model or the autoregressive integrated moving average (ARIMA) model to forecast incoming workloads, exploiting the hidden Markov modeling to characterize variations in workload patterns , and estimating dynamic resource demand of workloads by light-weight statistical learning algorithms or MLP .\\nJim Gao builds an MLP model to predict power usage effectiveness of data centers, which is extensively tested and validated at Google data centers.\\nCortez \\\\textit{et al.} predict virtual machine (VM) behaviors (including VM lifetimes, maximum deployment sizes, and workload classes) for a broader set of purposes (e.g., health/resource management and power capping), where the evaluated ML models are random forests and extreme gradient boosting trees.\\nIn addition to workload/resource-related metrics, the availability in data centers or cloud services is also a topic of concern, where one of the key tasks is to predict disk failure in advance.\\nLeveraging SMART (Self-Monitoring, Analysis and Reporting Technology) attributes, the disk failure prediction model can be built via various ML techniques, such as different Bayesian methods , unsupervised clustering , SVM and MLP .\\nThe adoption of classification and regression trees (CART) , RNNs , or gradient boosted regression trees makes it possible to assess health status of drives.\\nWhile all the aforementioned methods rely on offline training, online random forests can evolve with forthcoming data on-the-fly by generating new trees and forget old information by discarding outdated trees, consequently avoiding the model aging problem in disk failure predictions.\\nTo predict partial drive failures (i.e., disk error or sector error), Mahdisoltani \\\\textit{et al.} explore five ML techniques (CART, random forests, SVM, NN and logistic regression), among which random forests consistently outperform others.\\nXu \\\\textit{et al.} incorporate SMART attributes and system-level signals to train a gradient boosted regression tree, which is an online prediction model that ranks disks according to the degree of error-proneness in the near future.\",\n \"id\": \"37bd7027-5f70-4a17-9138-7b53f9c12df8\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"9ac2d4c9-6bb8-44b3-aa23-38c568e05513\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"System Modeling and Performance Prediction\"\n ],\n [\n \"subsubsection\",\n \"Data Center Performance Modeling and Prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Center Performance Modeling and Prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:model_eda}\",\n \"id\": \"419e1fc4-5eb3-40c4-b251-cc170597bc2e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"15e5a658-0b25-4890-8335-c1ec062ad3a2\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"Performance Modeling in Chip Design and Design Automation\"\n ]\n ],\n \"subsections\": [\n \"df0787ef-e9e6-48be-837b-abd7e6a37852\",\n \"3e9b9f12-6464-4e9c-b39f-515e81853fdb\",\n \"8ba558a9-ad3c-4f4c-a999-80fc17f11ebf\"\n ],\n \"title\": \"Performance Modeling in Chip Design and Design Automation\"\n },\n {\n \"cite_extract_rate\": 0.15625,\n \"cites\": [\n 5393,\n 5395,\n 5396,\n 8916,\n 5394\n ],\n \"content\": \"\\\\label{sec:model_analog}\\nAnalog circuit design is usually a manual process that requires many trial-and-error iterations between pre-layout and post-layout phases.\\nIn recent years, the discrepancy between schematic (i.e., pre-layout) performance estimations and post-layout simulation results is further enlarged.\\nOn the one hand, the analytical performance estimations from schematics are no longer accurate with device scaling; on the other hand, even though post-layout simulations can provide high-accuracy estimations, they are extremely time-consuming and have become the major bottleneck of design iteration time.\\nTo shrink the gap in performance modeling of integrated circuits (ICs), ML techniques are widely applied for fast circuit evaluation.\\nWe discuss the studies based on whether their input features are extracted from pre-layout or post-layout information.\\n\\\\textcircled{\\\\small{1}}\\nGiven design schematics, parasitics in layouts can be predicted from pre-layout stage, which helps bridge the gap of performance difference between pre-layout and post-layout simulations.\\nParaGraph builds a GNN model to predict layout-dependent parasitics and physical device parameters.\\nMLParest shows that non-graph based methods (e.g., random forest) also work well for estimating interconnect parasitics, whereas the lack of placement information may cause large variations in predictions.\\n\\\\textcircled{\\\\small{2}}\\nGiven circuit schematics as well as device information as inputs, it is possible to directly model post-layout performance from pre-layout designs.\\nAlawieh \\\\textit{et al.} propose a hierarchical method that combines the Bayesian co-learning framework and semi-supervised learning to predict power consumption.\\nThe entire circuit schematic is partitioned into multiple blocks to build block-level performance models, upon which circuit-level performance models are built. By combining these two low-dimensional models with a large amount of unlabeled data, pseudo samples can be labeled with almost no cost. Finally, a high-dimensional performance model mapping low-level features to circuit-level metrics is trained with pseudo samples and a small amount of labeled samples, which demonstrates the feasibility of performance modeling with inadequate labeled samples.\\nSeveral variants of Bayesian-based methods also perform well for estimating post-layout performance, e.g., combining Bayesian regression with SVM to predict circuit performance and using Bayesian DNNs to compare circuit designs .\\n\\\\textcircled{\\\\small{3}}\\nSince post-layout simulations with SPICE-like simulators is time-consuming, ML techniques are applied to quickly assess layout design performance .\\nTo make better use of structural information inside layouts, intermediate layout placement results are represented as 3D images to feed a 3D CNN model , or encoded as graphs to train a customized GNN model , with the goal to predict whether a design specification is satisfied.\\n\\\\begin{table}[tbp]\\n\\\\vspace{-10pt}\\n\\\\caption{Summary of applying ML techniques for performance modeling and prediction in design automation.}\\n\\\\vspace{-10pt}\\n\\\\label{table:model_eda}\\n\\\\centering\\n \\\\tiny\\n \\\\renewcommand{\\\\arraystretch}{1}\\n \\\\setlength{\\\\tabcolsep}{0.5pt}\\n\\\\begin{tabular}{c|c|c|c}\\n\\\\toprule\\n\\\\textbf{Domain} & \\\\textbf{Prediction of} & \\\\textbf{Technique} & \\\\textbf{Input} \\\\\\\\ \\\\midrule\\n\\\\multirow{5}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Analog \\\\\\\\ Circuit \\\\\\\\ $( \\\\S$ \\\\ref{sec:model_analog}$)$ \\\\end{tabular}}} & Parasitics & GNN , random forest & Circuit schematics \\\\\\\\ \\\\cline{2-4} \\n & Power/area/bandwidth & \\\\begin{tabular}[c]{@{}c@{}}Bayesian co-learning and semi-supervised learning , \\\\\\\\ Bayesian regression and SVM \\\\end{tabular} & \\\\multirow{2}{*}{\\\\begin{tabular}[c]{@{}c@{}}Circuit schematics \\\\\\\\ and device information\\\\end{tabular}} \\\\\\\\ \\\\cline{2-3}\\n & \\\\begin{tabular}[c]{@{}c@{}}Probability of superiority between designs\\\\end{tabular} & Bayesian DNN & \\\\\\\\ \\\\cline{2-4} \\n & \\\\begin{tabular}[c]{@{}c@{}}Gain/unity gain frequency/bandwidth/phase margin\\\\end{tabular} & \\\\begin{tabular}[c]{@{}c@{}}SVM/ANN/random forest , 3D CNN , GNN \\\\end{tabular} & \\\\multirow{2}{*}{Circuit placement} \\\\\\\\ \\\\cline{2-3}\\n & Electromagnetic properties & GNN & \\\\\\\\ \\\\hline\\n\\\\multirow{9}{*}{\\\\makecell[c]{\\\\textbf{HLS} \\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{sec:model_hls}}$)$}} & \\\\begin{tabular}[c]{@{}c@{}}Area/latency/throughput/logic utilization\\\\end{tabular} & \\\\begin{tabular}[c]{@{}c@{}}Random forest , transfer learning \\\\end{tabular} & Directives in HLS scripts \\\\\\\\ \\\\cline{2-4} \\n & Resource utilization & ANN & \\\\multirow{3}{*}{IR graphs from HLS front-ends} \\\\\\\\ \\\\cline{2-3}\\n & Resource mapping and clustering & GraphSAGE & \\\\\\\\ \\\\cline{2-3}\\n & Routing congestion & \\\\begin{tabular}[c]{@{}c@{}}Linear regression/ANN/gradient boosted regression tree \\\\end{tabular} & \\\\\\\\ \\\\cline{2-4} \\n & Power & \\\\begin{tabular}[c]{@{}c@{}}Linear regression/SVM/tree-based models/DNNs \\\\end{tabular} & IR graphs and HLS reports \\\\\\\\ \\\\cline{2-4} \\n & \\\\begin{tabular}[c]{@{}c@{}}Throughput and throughput-to-area ratio\\\\end{tabular} & \\\\begin{tabular}[c]{@{}c@{}}Ensemble learning by stacked regression \\\\end{tabular} & \\\\multirow{2}{*}{HLS reports} \\\\\\\\ \\\\cline{2-3}\\n & Resource utilization and timing & \\\\begin{tabular}[c]{@{}c@{}}Linear regression/ANN/gradient tree boosting \\\\end{tabular} & \\\\\\\\ \\\\cline{2-4} \\n & Cross-platform latency and power & Random forest & CPU program counters \\\\\\\\ \\\\cline{2-4} \\n & Speedup over an ARM processor & ANN & \\\\begin{tabular}[c]{@{}c@{}}Application characteristics, \\\\\\\\ HLS reports, FPGA configurations\\\\end{tabular} \\\\\\\\ \\\\hline\\n\\\\multirow{8}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Logic and \\\\\\\\ Physical \\\\\\\\ Synthesis \\\\\\\\ $( \\\\S$ \\\\ref{sec:model_logic}$)$\\\\end{tabular}}} & Area/delay & CNN , LSTM & Synthesis flows \\\\\\\\ \\\\cline{2-4} \\n & \\\\multirow{2}{*}{DRVs} & \\\\begin{tabular}[c]{@{}c@{}}Linear regression/ANN/decision tree , MARS \\\\end{tabular} & \\\\begin{tabular}[c]{@{}c@{}}Placement and GR information\\\\end{tabular} \\\\\\\\ \\\\cline{3-4} \\n & & MARS/SVM , MLP & Placement \\\\\\\\ \\\\cline{2-4} \\n & \\\\multirow{2}{*}{DRC hotspots} & FCN & \\\\begin{tabular}[c]{@{}c@{}}Placement and GR information\\\\end{tabular} \\\\\\\\ \\\\cline{3-4} \\n & & Variant of FCN & \\\\multirow{3}{*}{Placement} \\\\\\\\ \\\\cline{2-3} \\n & GR congestion map & FCN & \\\\\\\\ \\\\cline{2-3} \\n & \\\\multirow{2}{*}{Routing congestion in FPGAs} & Linear regression & \\\\\\\\ \\\\cline{3-4} \\n & & Conditional GAN & Post-placement images \\\\\\\\ \\\\bottomrule\\n\\\\end{tabular}\\n\\\\vspace{-15pt}\\n\\\\end{table}\",\n \"id\": \"df0787ef-e9e6-48be-837b-abd7e6a37852\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 32,\n \"parent_id\": \"419e1fc4-5eb3-40c4-b251-cc170597bc2e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"Performance Modeling in Chip Design and Design Automation\"\n ],\n [\n \"subsubsection\",\n \"Analog Circuit Analysis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Analog Circuit Analysis\"\n },\n {\n \"cite_extract_rate\": 0.18181818181818102,\n \"cites\": [\n 5393,\n 5394\n ],\n \"content\": \"\\\\label{sec:model_hls}\\nHLS is an automated transformation from behavioral languages (e.g., C/C++/SystemC) to register-transfer level (RTL) designs, which significantly expedites the development of hardware designs involving with field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs).\\nSince HLS tools usually take considerable time to synthesize each design, it prevents designers from exploring design space sufficiently, which motivates the application of ML models for fast and accurate performance estimation.\\nIn performance estimation of HLS designs, the input features to ML models are extracted from three major sources: HLS directives, IRs from HLS front-ends, and HLS reports.\\n\\\\textcircled{\\\\small{1}} Taking the directives in an HLS script as input features, random forest is capable to forecast different design metrics, such as area and effective latency , and throughput and logic utilization .\\nIn order to reuse knowledge from previous experiences, a transfer learning approach can transfer knowledge across different applications or synthesis options.\\n\\\\textcircled{\\\\small{2}} Taking advantages of IR graphs generated by HLS front-ends, Koeplinger \\\\textit{et al.} count resource requirements of each node in graphs by using pre-characterized area models, which are then used as inputs to ANNs to predict the LUT routing usage, register duplication, and unavailable LUTs.\\nThe exploitation of GNNs makes it possible to automatically predict the mapping from arithmetic operations in IR graphs to different resources on FPGAs . \\nTo forecast post-implementation routing congestion, Zhao \\\\textit{et al.} build a dataset that connects the routing congestion metrics after RTL implementation with operations in IRs, with the goal to train ML models locating highly congested regions in source code.\\n\\\\textcircled{\\\\small{3}} Taking the information that can be directly extracted from HLS reports, Dai \\\\textit{et al.} try several ML models (linear regression, ANN, and gradient tree boosting) to predict post-implementation resource utilization and timing. \\nPyramid applies the ensemble learning by stacked regression to accurately estimate the throughput or the throughput-to-area ratio.\\nHL-Pow employs features from both IR graphs and HLS reports to predict the power by a variety of ML models. \\nThe surge of heterogeneous platforms with FPGA/AISC and CPU provides more possibility of hardware/software co-design, motivating cross-platform performance predictions.\\nHLSPredict uses random forest to predict FPGA cycle counts and power consumption based on program counter measurements obtained from CPU execution. \\nWhile HLSPredict targets the same FPGA platform in training and testing, XPPE considers different FPGA platforms, and uses ANNs to predict the speedup of an application on a target FPGA over an ARM processor.\",\n \"id\": \"3e9b9f12-6464-4e9c-b39f-515e81853fdb\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"419e1fc4-5eb3-40c4-b251-cc170597bc2e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"Performance Modeling in Chip Design and Design Automation\"\n ],\n [\n \"subsubsection\",\n \"High-Level Synthesis (HLS)\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"High-Level Synthesis (HLS)\"\n },\n {\n \"cite_extract_rate\": 0.15384615384615302,\n \"cites\": [\n 5395,\n 8916\n ],\n \"content\": \"\\\\label{sec:model_logic}\\nIn digital design, logic synthesis converts RTL designs into optimized gate-level representations;\\nphysical synthesis then transforms these design netlists into physical layouts.\\nSince these two stages may take hours or days to generate final bitstreams/layouts, many problems benefit from the power of ML models for fast performance estimation.\\nIn logic synthesis, CNN models or LSTM-based models can be leveraged to forecast the delay and area after applying different synthesis flows on specific designs, where the inputs are synthesis flows represented in either matrices or time series.\\nIn physical synthesis, routing is a sophisticated problem subject to stringent constraints, and EDA tools typically utilize a two-step method: global routing (GR) and detailed routing (DR). GR tool allocates routing resource coarsely, and provides routing plans to guide DR tools to complete the entire routing.\\nIn general, routing congestion can be figured out during or after GR; the routability of a design is confirmed after DR and design rule checking (DRC).\\nEndeavors have been made to predict routability from early layout stages, so as to avoid excessive iterations back and forth between placement and routing.\\nIn ASICs, some investigations predict routability by estimating the number of design rule violations (DRVs).\\nTaking GR results as inputs, Li \\\\textit{et al.} explore several ML models (linear regression, ANN, and decision tree) to predict the number of DRVs, final hold slack, power, and area. \\nQi \\\\textit{et al.} rely on placement data and congestion maps from GR as input features, and use a nonparametric regression technique, multivariate adaptive regression splines (MARS) , to predict the utilization of routing resource and the number of DRVs.\\nBy merely leveraging placement information, it is possible to predict routability by MARS and SVM , or to detect DR short violations by an MLP .\\nWhen representing placement information as images, fully convolutional networks (FCNs) are capable to predict locations of DRC hotspots by considering GR information as inputs , or to forecast GR congestion maps by formulating the prediction task as a pixel-wise binary classification using placement data .\\nJ-Net is a customized FCN model, and takes both high-resolution pin patterns and low-resolution layout information from the placement stage as features to output a 2D array that indicates if the tile corresponding to each entry is a DRC hotspot.\\nIn FPGAs, routing congestion maps can be directly estimated by linear regression using feature vectors coming from pin counts and wirelength per area of SLICEs.\\nBy constructing the routing congestion prediction as an image translation problem, a conditional GAN is able to take post-placement images as inputs to predict congestion heat maps.\",\n \"id\": \"8ba558a9-ad3c-4f4c-a999-80fc17f11ebf\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"419e1fc4-5eb3-40c4-b251-cc170597bc2e\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML for Fast System Modeling\"\n ],\n [\n \"subsection\",\n \"Performance Modeling in Chip Design and Design Automation\"\n ],\n [\n \"subsubsection\",\n \"Logic and Physical Synthesis\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Logic and Physical Synthesis\"\n },\n {\n \"cite_extract_rate\": 0.15384615384615302,\n \"cites\": [\n 5397,\n 5409,\n 3579,\n 5400,\n 5399,\n 5406,\n 5398,\n 5407,\n 5403,\n 3603,\n 5408,\n 5401,\n 5405,\n 4096,\n 5402,\n 5404\n ],\n \"content\": \"\\\\label{sec:methodology}\\nThis section introduces studies that directly employ ML techniques as the design methodology for computer architecture/systems.\\nComputer architecture and systems have been becoming increasingly complicated, making it expensive and inefficient for human efforts to design or optimize them.\\nIn response, visionaries have argued that computer architecture and systems should be imbued with the capability to design and configure themselves, adjust their behaviors according to workloads' needs or user-specified constraints, diagnose failures, repair themselves from the detected failures, etc.\\nWith strong learning and generalization capabilities, ML-based techniques are naturally suitable to resolve these considerations, which can adjust their policies during system designs according to long-term planning and dynamic workload behaviors.\\nAs many problems in architecture/system design can be formulated as combinatorial optimization or sequential decision-making problems, RL is broadly explored and exploited.\\nTable \\\\ref{table:dse_sys} and Table \\\\ref{table:dse_eda} recapitulate the studies that apply ML techniques as the design methodology for computer architecture/system and design automation respectively, in terms of target tasks and adopted ML techniques.\\n\\\\begin{table}[tbp]\\n\\\\vspace{-10pt}\\n\\\\caption{Summary of applying ML techniques as the design methodology for computer architecture/ systems.}\\n\\\\vspace{-10pt}\\n\\\\label{table:dse_sys}\\n\\\\centering\\n \\\\tiny\\n \\\\renewcommand{\\\\arraystretch}{0.9}\\n \\\\setlength{\\\\tabcolsep}{3pt}\\n\\\\begin{tabular}{c|c|c}\\n\\\\toprule\\n\\\\textbf{Domain} & \\\\textbf{Task} & \\\\textbf{Technique} \\\\\\\\ \\\\midrule\\n\\\\multirow{4}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Memory \\\\\\\\ System\\\\\\\\ Design \\\\\\\\ $( \\\\S$ \\\\ref{sec:dse_mem}$)$ \\\\end{tabular}}} & Cache replacement policy & \\\\begin{tabular}[c]{@{}c@{}}Perceptron learning , Markov decision process , LSTM and SVM \\\\end{tabular} \\\\\\\\ \\\\cline{2-3} \\n & Cache prefetching policy & \\\\begin{tabular}[c]{@{}c@{}}Perceptron learning , contextual bandit , LSTM \\\\end{tabular} \\\\\\\\ \\\\cline{2-3} \\n & Memory controller policy & Q-learning \\\\\\\\ \\\\cline{2-3} \\n & Garbage collection & Q-learning \\\\\\\\ \\\\hline\\n\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Branch\\\\\\\\ Prediction \\\\\\\\ $( \\\\S$ \\\\ref{sec:dse_bp}$)$\\\\end{tabular}} & Branch direction & \\\\begin{tabular}[c]{@{}c@{}}MLP , piecewise linear regression , \\\\\\\\ perceptron , CNN \\\\end{tabular} \\\\\\\\ \\\\hline\\n\\\\multirow{8}{*}{\\\\makecell[c]{\\\\textbf{NoC} \\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{NoC}}$)$}} & Link management & ANN \\\\\\\\ \\\\cline{2-3} \\n & DVFS for routers & Q-learning \\\\\\\\ \\\\cline{2-3} \\n & Routing & Q-learning \\\\\\\\ \\\\cline{2-3} \\n & Arbitration policy & DQN \\\\\\\\ \\\\cline{2-3} \\n & Adjusting injection rates & Q-learning , ANN \\\\\\\\ \\\\cline{2-3} \\n & Selection of fault-tolerant modes & Q-learning \\\\\\\\ \\\\cline{2-3} \\n & Link placement in 3D NoCs & STAGE algorithm \\\\\\\\ \\\\cline{2-3} \\n & Loop placement in routerless NoCs & Advantage actor-critic with MCTS \\\\\\\\ \\\\hline\\n\\\\multirow{6}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Power \\\\\\\\ Management \\\\\\\\ $( \\\\S$ \\\\ref{power}$)$ \\\\end{tabular}}} & DVFS and thread packing & Multinomial logistic regression \\\\\\\\ \\\\cline{2-3} \\n & DVFS and power gating & MLP \\\\\\\\ \\\\cline{2-3} \\n & \\\\begin{tabular}[c]{@{}c@{}}DVFS, socket allocation, and use of HyperThreads\\\\end{tabular} & Extra trees/gradient boosting/KNN/MLP/SVM \\\\\\\\ \\\\cline{2-3} \\n & \\\\begin{tabular}[c]{@{}c@{}}DVFS for CPU cores/uncore/through-silicon interposers\\\\end{tabular} & Propositional rule , ANN , Q-learning \\\\\\\\ \\\\cline{2-3} \\n & \\\\begin{tabular}[c]{@{}c@{}}DVFS for CPU-GPU heterogeneous platforms\\\\end{tabular} & Weighted majority algorithm \\\\\\\\ \\\\cline{2-3} \\n & DVFS for multi-/many-core systems & \\\\begin{tabular}[c]{@{}c@{}}Q-learning , semi-supervised RL , hierarchical Q-learning \\\\end{tabular} \\\\\\\\ \\\\hline\\n\\\\multirow{6}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Resource \\\\\\\\ Management \\\\\\\\ \\\\& Task \\\\\\\\ Allocation \\\\\\\\ $( \\\\S$ \\\\ref{resource management}$)$\\\\end{tabular}}} & Tuning architecture configurations & \\\\begin{tabular}[c]{@{}c@{}}Maximum likelihood , statistical machine learning \\\\end{tabular} \\\\\\\\ \\\\cline{2-3} \\n & Dynamic cache partitioning & Enforced subpopulations , Q-learning \\\\\\\\ \\\\cline{2-3} \\n & Task allocation in many-core systems & Q-learning , DDPG \\\\\\\\ \\\\cline{2-3} \\n & Workflow management & SVM and random forest \\\\\\\\ \\\\cline{2-3} \\n & Hardware resource assignment & REINFORCE , Bayesian optimization \\\\\\\\ \\\\cline{2-3} \\n & Device placement & \\\\begin{tabular}[c]{@{}c@{}}REINFORCE , policy gradient , PPO \\\\end{tabular} \\\\\\\\ \\\\hline\\n\\\\multirow{2}{*}{\\\\makecell[c]{\\\\textbf{Scheduling}\\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{scheduling}}$)$}} & Scheduling jobs in single-core processors & Q-learning \\\\\\\\ \\\\cline{2-3} \\n & Scheduling jobs in multi-processor systems & Value-based RL \\\\\\\\ \\\\hline\\n\\\\multirow{9}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Data Center\\\\\\\\ Management \\\\\\\\ $( \\\\S$ \\\\ref{data center}$)$\\\\end{tabular}}} & Assignment of servers to applications & Value-based RL \\\\\\\\ \\\\cline{2-3} \\n & Content allocation in CDNs & Fuzzy RL \\\\\\\\ \\\\cline{2-3} \\n & \\\\begin{tabular}[c]{@{}c@{}}Placement of virtual machines onto physical machines\\\\end{tabular} & PPO \\\\\\\\ \\\\cline{2-3} \\n & Traffic optimization & Policy gradient and DDPG \\\\\\\\ \\\\cline{2-3} \\n & Scheduling jobs with complex dependency & REINFORCE \\\\\\\\ \\\\cline{2-3} \\n & Straggler diagnosis & Statistical ML \\\\\\\\ \\\\cline{2-3} \\n & Data-center-level caching policy & \\\\begin{tabular}[c]{@{}c@{}}Decision tree , LSTM , gradient boosting , DDPG \\\\end{tabular} \\\\\\\\ \\\\cline{2-3} \\n & Bitrate selection for video chunks & A3C \\\\\\\\ \\\\cline{2-3} \\n & \\\\begin{tabular}[c]{@{}c@{}}Scheduling video workloads in hybrid CPU-GPU clusters\\\\end{tabular} & DQN \\\\\\\\ \\\\hline\\n \\\\multirow{3}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Code\\\\\\\\ Generation \\\\\\\\ $( \\\\S$ \\\\ref{code_generation}$)$\\\\end{tabular}}} & Code completion & N-gram model and RNN \\\\\\\\ \\\\cline{2-3} \\n & Code generation & LSTM \\\\\\\\ \\\\cline{2-3} \\n & Program translation & Tree-to-tree encoder-decoder , seq2seq , transformer \\\\\\\\ \\\\hline\\n\\\\multirow{6}{*}{\\\\makecell[c]{\\\\textbf{Compiler}\\\\\\\\ $( \\\\S$ \\\\textbf{\\\\ref{sec:compiler}}$)$}} & Instruction scheduling & Temporal difference , projective reparameterization \\\\\\\\ \\\\cline{2-3} \\n & Improving compiler heuristics & NEAT , LSTM \\\\\\\\ \\\\cline{2-3} \\n & Ordering of optimizations & NEAT \\\\\\\\ \\\\cline{2-3} \\n & Automatic vectorization & Imitation learning \\\\\\\\ \\\\cline{2-3} \\n & \\\\begin{tabular}[c]{@{}c@{}}Program transformation for approximate computing\\\\end{tabular} & MLP \\\\\\\\ \\\\cline{2-3} \\n & Compilation for DNN workloads & PPO , policy gradient \\\\\\\\ \\\\bottomrule\\n\\\\end{tabular}\\n\\\\end{table}\",\n \"id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"level\": \"section\",\n \"origin_cites_number\": 104,\n \"parent_id\": \"5631fef3-6389-407d-9101-745f45240141\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ]\n ],\n \"subsections\": [\n \"4ce0d49a-419f-4a57-9bf8-bac10689ff82\",\n \"93695066-a3b6-43e6-a561-f289514d99b8\",\n \"02b1496f-b229-4d07-b91a-238e79b26753\",\n \"8b48b63d-94af-4ec0-a218-72746bce7952\",\n \"de7c5699-73a2-48d8-9e75-ff1b60c4f5f5\",\n \"68c1d9ca-e725-4975-ad78-356d7848d1d1\",\n \"e25c32d8-64dc-4364-8501-8063014c67c6\"\n ],\n \"title\": \"ML as Design Methodology\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:dse_mem}\\nThe \\\"memory wall\\\" has been a performance bottleneck in von Neumann architectures, where computation is orders of magnitude faster than memory access.\\nTo alleviate this problem, hierarchical memory systems are widely used and there arise optimizations for different levels of memory systems.\\nAs both the variety and the size of modern workloads are drastically growing, conventional designs that are based on heuristics or intuitions may not catch up with the demand of the ever-growing workloads, leading to sharply degradation in performance.\\nAs such, many studies resort to ML-based techniques to design smart and intelligent memory systems.\",\n \"id\": \"4ce0d49a-419f-4a57-9bf8-bac10689ff82\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Memory System Design\"\n ]\n ],\n \"subsections\": [\n \"c842f575-5671-48ac-aa2a-715cd98df8a6\",\n \"dbd16b4a-1edb-4482-a515-db26bea31ee6\",\n \"dc411e9f-ce19-4cdd-9e1c-1930d2c68aa6\"\n ],\n \"title\": \"Memory System Design\"\n },\n {\n \"cite_extract_rate\": 0.18181818181818102,\n \"cites\": [\n 5401,\n 5398\n ],\n \"content\": \"\\\\label{cache}\\nThe conspicuous disparity in latency and bandwidth between CPUs and memory systems motivates investigations for efficient cache management.\\nThere are two major types of studies on cache optimization: improving cache replacement policies, and designing intelligent prefetching policies.\\n\\\\textcircled{\\\\small{1}} \\nTo develop cache replacement policies, perceptron learning is employed to predict whether to bypass or reuse a referenced block in the last-level cache (LLC) .\\nInstead of using perceptrons, Beckmann \\\\textit{et al.} model the cache replacement problem as a Markov decision process and replace lines according to the difference between their expected hits and the average hits.\\nShi \\\\textit{et al.}~ train an attention-based LSTM model offline to extract insights from history program counters, which are then used to build an online SVM-based hardware predictor to serve as the cache replacement policy.\\n\\\\textcircled{\\\\small{2}} \\nTo devise intelligent prefetchers, Wang \\\\textit{et al.} propose a prefetching mechanism that uses conventionally table-based prefetchers to provide prefetching suggestions and a perceptron trained by spatio-temporal locality to reject unnecessary prefetching decisions, ameliorating the cache pollution problem. \\nSimilarly, Bhatia \\\\textit{et al.} integrate a perceptron-based prefetching filter with conventional prefetchers, increasing the coverage of prefetches without hurting accuracy.\\nInstead of the commonly used spatio-temporal locality, a context-based memory prefetcher leverages the semantic locality that characterizes access correlations inherent to program semantics and data structures, which is approximated by a contextual bandit model in RL.\\nInterpreting semantics in memory access patterns is analogous to sequence analysis in natural language processing (NLP), and thus several studies use LSTM-based models and treat the prefetching as either a regression problem or a classification problem .\\nEven with better performance, especially for long access sequences and noise traces, LSTM-based prefetchers suffer from long warm-up and prediction latency, and considerable storage overheads.\\nThe discussion of how hyperparameters impact LSTM-based prefetchers' performance highlights that the lookback size (i.e. memory access history window) and the LSTM model size strongly affect prefetchers' learning ability under different noise levels or workload patterns.\\nTo accommodate the large memory space, Shi \\\\textit{et al.} introduce a hierarchical sequence model to decouple predictions of pages and offsets by using two separate attention-based LSTM layers, whereas the corresponding hardware implementation is impractical for actual processors.\",\n \"id\": \"c842f575-5671-48ac-aa2a-715cd98df8a6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"4ce0d49a-419f-4a57-9bf8-bac10689ff82\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Memory System Design\"\n ],\n [\n \"subsubsection\",\n \"Cache\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Cache\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Smart memory controllers can significantly improve memory bandwidth utilization.\\nAiming at a self-optimizing memory controller adaptive to dynamically changing workloads, it can be modeled as an RL agent that always selects legal DRAM commands with the highest expected long-term performance benefits (i.e., Q-values) .\\nTo allow optimizations toward various objectives, this memory controller is then improved in two major aspects . First, the rewards of different actions (i.e., legal DRAM commands) are automatically calibrated by genetic algorithms to serve different objective functions (e.g., energy, throughput, etc). Second, a multi-factor method that considers the first-order attribute interactions is employed to select proper attributes used for state representations.\\nSince both of them use table-based Q-learning and select limited attributes to represent states, the scalability may be a concern and their performance could be improved with more informative representations.\",\n \"id\": \"dbd16b4a-1edb-4482-a515-db26bea31ee6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"4ce0d49a-419f-4a57-9bf8-bac10689ff82\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Memory System Design\"\n ],\n [\n \"subsubsection\",\n \"Memory Controller\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Memory Controller\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 3600\n ],\n \"content\": \"A variety of work targets different parts of the memory system.\\nMargaritov \\\\textit{et al.} accelerate virtual address translation through learned index structures . The results are encouraging in terms of the accuracy, which reaches almost 100\\\\% for all tested virtual addresses; yet this method has unacceptably long inference latency, leaving practical hardware implementation as the future work.\\nWang \\\\textit{et al.} reduce data movement energy in interconnects by exploiting asymmetric transmission costs of different bits, where data blocks to be transmitted are dynamically grouped by K-majority clustering to derive energy-efficient expressions for transmission.\\nIn terms of garbage collection in NAND flash, Kang \\\\textit{et al.} propose an RL-based method to reduce the long-tail latency. The key idea is to exploit the inter-request interval (idle time) to dynamically decide the number of pages to be copied or whether to perform an erase operation, where decisions are made by table-based Q-learning. Their following work considers more fine-grained states, and introduces a Q-table cache to manage key states among enormous amount of states.\",\n \"id\": \"dc411e9f-ce19-4cdd-9e1c-1930d2c68aa6\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"4ce0d49a-419f-4a57-9bf8-bac10689ff82\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Memory System Design\"\n ],\n [\n \"subsubsection\",\n \"Others\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Others\"\n },\n {\n \"cite_extract_rate\": 0.11111111111111101,\n \"cites\": [\n 5405\n ],\n \"content\": \"\\\\label{sec:dse_bp}\\nBranch predictor is one of the mainstays of modern processors, significantly improving the instruction-level parallelism. As pipelines gradually deepen, the penalty of mis-prediction increases. Traditional branch predictors often consider limited history length, which may hurt the prediction accuracy.\\nIn contrast, the perceptron/MLP-based predictors can handle long histories with reasonable hardware budgets, outperforming prior state-of-the-art non-ML-based predictors.\\nStarting with a static branch predictor trained with static features from program corpus and control flow graphs, an MLP is used to predict the direction of a branch at compile time .\\nLater, a dynamic branch predictor uses a perceptron-based method . It hashes the branch address to select the proper perceptron and computes the dot product accordingly to decide whether to take this branch, which shows great performance on linearly separable branches. \\nIts latency and accuracy can be further improved by applying ahead pipelining and selecting perceptrons based on path history .\\nTo attain high accuracy in non-linearly separable branches, the perceptron-based prediction is generalized as piecewise linear branch prediction .\\nIn addition to the path history, multiple types of features from different organizations of branch histories can be leveraged to enhance the overall performance .\\nWhen considering practical hardware implementation of branch predictors, SNAP leverages current-steering digital-to-analog converters to transfer digital weights into analog currents and replaces the costly digital dot-product computation to the current summation.\\nIts optimized version equips several new techniques, such as the use of global and per-branch history, trainable scaling coefficients, dynamic training thresholds, etc.\\nRather than making binary decisions of whether to take a certain branch, it is possible to directly predict the target address of an indirect branch at the bit level via perceptron-based predictors .\\nWhile high accuracy is achieved by current perceptron/MLP-based predictors, Tarsa \\\\textit{et al.} notice that a small amount of static branch instructions are systematically mispredicted, referred to as hard-to-predict branches (H2Ps). Consequently, they propose a CNN helper predictor for pattern matching of history branches, ultimately improving accuracy for H2Ps in conditional branches.\",\n \"id\": \"93695066-a3b6-43e6-a561-f289514d99b8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Branch Prediction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Branch Prediction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{NoC}\\nThe aggressive transistor scaling has paved the way for integrating more cores in a single chip or processor.\\nWith the increasing number of cores per chip, NoC plays a gradually crucial role, since it is responsible for inter-core communication and data movement between cores and memory hierarchies.\\nSeveral problems attracting attention are as follows. \\nFirst, communication energy scales slower than computation energy , implying necessity to improve power efficiency of NoCs. \\nSecond, the complexity of routing or traffic control grows with the number of cores per chip and this problem is even exacerbated by the rising variety and irregularity of workloads.\\nThird, with the continuous scaling down of transistors, NoCs are more vulnerable to different types of errors and thus reliability becomes a key concern.\\nFourth, some non-conventional NoC architectures might bring promising potentials in the future, whereas they usually come with large design spaces and complex design constraints, which is nearly impossible for manually optimization.\\nAmong all aforementioned fields, ML-based design techniques display their strength and charm.\",\n \"id\": \"02b1496f-b229-4d07-b91a-238e79b26753\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"NoC Design\"\n ]\n ],\n \"subsections\": [\n \"f70d9a91-45f6-4778-9893-1e1691350184\",\n \"9cbfc7ad-3786-493c-a695-e8c786d70c4c\",\n \"0f23b3c6-50c2-4cf8-a777-94590c806c36\",\n \"5afafbcf-6c39-4a33-8c76-314585dc4284\"\n ],\n \"title\": \"NoC Design\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"Power consumption is one crucial concern in NoCs, in which links usually consume a considerable portion of network power.\\nWhile turning on/off links according to a static threshold of link utilization is a trivial way to reduce power consumption, it can not adapt to dynamically changing workloads.\\nSavva \\\\textit{et al.} use multiple ANNs for dynamic link management. Each ANN is responsible for one region of the NoC, and dynamically computes a threshold for every time interval to turn on/off links given the link utilization of each region. Despite significant power savings with low hardware overheads, this approach causes long latency in routing.\\nIn order to meet certain power and thermal budgets, hierarchical ANNs are used to predict optimal NoC configurations (i.e., link bandwidth, node voltage and task assignment to nodes), where the global ANN predicts globally optimal NoC configurations exploiting local optimal energy consumption predicted by local ANNs.\\nTo save dynamic power, several investigations employ per-router based Q-learning agents, which are offline trained ANNs to select optimal voltage/frequency levels for each router.\",\n \"id\": \"f70d9a91-45f6-4778-9893-1e1691350184\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"02b1496f-b229-4d07-b91a-238e79b26753\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"NoC Design\"\n ],\n [\n \"subsubsection\",\n \"Link Management and DVFS\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Link Management and DVFS\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"With the increasing variety and irregularity of workloads and their traffic patterns, learning-based routing algorithms and traffic control approaches show superior performance due to their excellent adaptability.\\n\\\\textcircled{\\\\small{1}} \\nAs routing problems can be formulated as sequential decision-making processes, several studies apply Q-learning based approaches, namely the Q-routing algorithm , which uses local estimation of delivery time to minimize total packets delivery time, capable to handle irregular network topologies and keep a higher network load than the conventional shortest path routing.\\nQ-routing is then extended to several other scenarios, such as combining with dual RL to improve learning speed and routing performance , resolving packets routing in dynamic NoCs whose network structures/topologies are dynamically changing during runtime , handling irregular faults in bufferless NoCs by the reconfigurable fault-tolerant Q-routing , and enhancing the capability to reroute messages around congested regions by the congestion-aware non-minimal Q-routing .\\nIn addition to routing problems, deep Q-network is also promising for NoC arbitration policies , where the agent/arbiter grants a certain output port to the input buffer with the largest Q-value.\\nEven displaying some improvements in latency and throughput, the direct hardware implementation is impractical due to the complexity of deep Q-networks, and thus insights are distilled to derive a relatively simple circuitry implementation.\\n\\\\textcircled{\\\\small{2}} \\nWith the goal to control congestion in NoCs, the SCEPTER NoC architecture , a bufferless NoC with single-cycle multi-hop traversals and a self-learning throttling mechanism, controls the injection of new flits into the network by Q-learning. Each node in the network independently selects whether to increase, decrease, or retain the throttle rate according to their Q-values, which conspicuously improves bandwidth allocation fairness and network throughput.\\nWang \\\\textit{et al.}~ design an ANN-based admission controller to determine the appropriate injection rate and the control policy of each node in a standard NoC.\",\n \"id\": \"9cbfc7ad-3786-493c-a695-e8c786d70c4c\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"02b1496f-b229-4d07-b91a-238e79b26753\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"NoC Design\"\n ],\n [\n \"subsubsection\",\n \"Routing and Traffic Control\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Routing and Traffic Control\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"With the aggressive technology scaling down, transistors and links in NoCs are more prone to different types of errors, indicating that reliability is a crucial concern and proactive fault-tolerant techniques are required to guarantee performance.\\nWang~\\\\textit{et al.}~ employ per-router-based Q-learning agents to independently select one of four fault-tolerant modes, which can minimize the end-to-end packet latency and power consumption. These agents are pre-trained and then fine-tuned during runtime.\\nIn their following work , these error-correction modes are extended and combined with various multi-function adaptive channel configurations, retransmission settings, and power management strategies, significantly improving latency, energy efficiency, and mean-time-to-failure.\",\n \"id\": \"0f23b3c6-50c2-4cf8-a777-94590c806c36\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"02b1496f-b229-4d07-b91a-238e79b26753\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"NoC Design\"\n ],\n [\n \"subsubsection\",\n \"Reliability and Fault Tolerance\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Reliability and Fault Tolerance\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 5400,\n 5406\n ],\n \"content\": \"With the growing number of cores per chip/system, the increasing heterogeneity of cores, and various performance targets, it is complicated to simultaneously optimize copious design knobs in NoCs.\\nOne attempt to automated NoC design is the MLNoC , which utilizes supervised learning to quickly find near-optimal NoC designs under multiple optimization goals. MLNoC is trained by data from thousands of real-world and synthetic SoC (system-on-chip) designs, and evaluated with real-world SoC designs. Despite disclosure of limited details and absence of comprehensive comparison with other design methods, it shows superior performance to manually optimized NoC designs, delivering encouraging results.\\nApart from conventional 2D mesh NoCs, a series of investigations focuses on 3D NoC designs, where the STAGE algorithm is applied to optimize vertical and planar placement of communication links in small-world network based 3D NoCs .\\nThe STAGE algorithm repeatedly alternates between two stages, the base search that tries to find the local optima based on the learned evaluation function, and the meta-search that uses SVR to learn evaluation functions.\\nLater, the STAGE algorithm is extended for multi-objective optimization in heterogeneous 3D NoC systems , which jointly considers GPU throughput, average latency between CPUs and LLCs, temperature, and energy.\\nIn terms of routerless NoCs that any two nodes are connected via at least one ring/loop, a deep RL framework that exploits Monte-Carlo tree search for efficient design space exploration is developed to optimize loop placements , and the design constraints can be strictly enforced by carefully devising the reward function.\",\n \"id\": \"5afafbcf-6c39-4a33-8c76-314585dc4284\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 5,\n \"parent_id\": \"02b1496f-b229-4d07-b91a-238e79b26753\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"NoC Design\"\n ],\n [\n \"subsubsection\",\n \"General Design\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"General Design\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{resource}\\nResource allocation or management is the coordination between computer architecture/systems and workloads. Consequently, its optimization difficulty occurs with the booming complexity from both sides and their intricate interactions. \\nML-based approaches have blazed the trail to adjusting policies wisely and promptly pursuant to dynamic workloads or specified constraints.\",\n \"id\": \"8b48b63d-94af-4ec0-a218-72746bce7952\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Resource Allocation or Management\"\n ]\n ],\n \"subsections\": [\n \"2eb80446-13aa-40ce-a764-e98e86d2b650\",\n \"35decbd9-a2aa-43b2-b9e5-0585c9140768\",\n \"5a43dcc0-acd3-4daf-bbb5-f9e380b060a3\"\n ],\n \"title\": \"Resource Allocation or Management\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{power}\\nML-based techniques have been applied broadly to improve power management, due to two main reasons.\\nFirst, power/energy consumption can be recognized as one metric of runtime costs.\\nSecond, under certain circumstances there could be a hard or soft constraint/budget of power/energy, making power efficiency a necessity.\\nIn consideration of power management for different parts of systems, PACSL uses the propositional rule to adjust dynamic voltage scaling (DVS) for CPU cores and on-chip L2 cache, which achieves an improvement in the energy-delay product by 22\\\\% on average (up to 46\\\\%) over independently applying DVS for each part.\\nWon \\\\textit{et al.} coordinate an ANN controller with a proportional integral for uncore DVFS. The ANN controller can be either pre-trained offline by a prepared dataset or trained online by bootstrapped learning.\\nManoj~\\\\textit{et al.} deploy Q-learning to adaptively adjust the level of output-voltage swing at transmitters of 2.5D through-silicon interposer I/Os, under constraints of communication power and bit error rate.\\nFrom the system level, DVFS is one of the most prevalent techniques.\\nPack \\\\& Cap builds a multinomial logistic regression classifier that is trained offline and queried during runtime, to accurately identify the optimal operating point for both thread packing and DVFS under an arbitrary power cap.\\nGreenGPU focuses on heterogeneous systems with CPUs and GPUs, and applies the weighted majority algorithm to scale frequency levels for both GPU cores and memory in a coordinated manner.\\nCHARSTAR targets joint optimization of power gating and DVFS within a single core, where frequencies and configurations are dynamically selected by a lightweight offline trained MLP predictor.\\nTo minimize energy consumption, Imes~\\\\textit{et al.}~ use ML-based classifiers (e.g., extra trees, gradient boosting, KNN, MLP and SVM) to predict the most energy-efficient resource settings (specifically, tuning socket allocation, the use of HyperThreads, and processor DVFS) by using low-level hardware performance counters.\\nBai~\\\\textit{et al.}~ consider the loss caused by on-chip regulator efficiency during DVFS, and try to minimize energy consumption under a parameterized performance constraint. The online control policy is implemented by a table-based Q-learning, which is portable across platforms without accurate modeling of a specific system.\\nA series of studies leverages RL for dynamic power management in multi-/many-core systems.\\nAs systems scale up, these RL-based methods often suffer from state space explosion, and two types of methods are introduced to resolve the scalability issue.\\n\\\\textcircled{\\\\small{1}} \\nBy combining RL with supervised learning, a semi-supervised RL-based approach achieves linear complexity with the number of cores, which is able to maximize throughput ensuring power constraints and cooperatively control cores and uncores in synergy.\\n\\\\textcircled{\\\\small{2}} \\nThe exploitation of hierarchical Q-learning reduces the time complexity to $O(n \\\\lg n)$, where $n$ denotes the number of cores.\\nPan \\\\textit{et al.} introduce multi-level Q-learning to select target power modes, where Q-values are approximated by a generalized radial basis function.\\nTable-based distributed Q-learning also performs well for DVFS , and there is one variant aware of the priorities of different applications.\\nSome energy management policies target specific applications or platforms.\\nJouleGuard~ is a runtime control system coordinating approximate computing applications with system resource under energy budgets.\\nIt uses a multi-arm bandit approach to identifying the most energy efficient system configuration, upon which application configurations are determined to maximize compute accuracy within energy budgets.\\nTargeting Intel SkyLake processors, a post-silicon CPU customization applies various ML models for dynamically clock-gating unused resource .\",\n \"id\": \"2eb80446-13aa-40ce-a764-e98e86d2b650\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"8b48b63d-94af-4ec0-a218-72746bce7952\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Resource Allocation or Management\"\n ],\n [\n \"subsubsection\",\n \"Power Management\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Power Management\"\n },\n {\n \"cite_extract_rate\": 0.23529411764705802,\n \"cites\": [\n 5404,\n 5403,\n 5407,\n 5409\n ],\n \"content\": \"\\\\label{resource management}\\nModern architectures and systems have been becoming so sophisticated and diverse that it is non-trivial to either optimize performance or fully utilize system resource.\\nThis rapidly evolving landscape is further complicated by various workloads with specific requirements or targets.\\nIn order to keep the pace, one cure is to develop more efficient and automated methods for resource management and task allocation, where ML-based techniques are excelled to explore large design spaces and simultaneously optimize multiple objectives, and preserve better scalablility and portability after carefully designed.\\nFor a single-core processor, a regularized maximum likelihood approach predicts the best hardware micro-architectural configuration for each phase of a program, based on runtime hardware counters. \\nFor multi-core processors, a statistical machine learning (SML) based method can quickly find configurations that simultaneously optimize running time and energy efficiency. Since this method is agnostic to application and micro-architecture domain knowledge, it is a portable alternative to human expert optimization.\\nSML can also be applied as a holistic method to design self-evolving systems that optimize performance hierarchically across circuit, platform, and application levels .\\nIn addition to tuning architectural configurations, dynamic on-chip resource management is crucial for multi-core processors, where one example is dynamic cache partitioning.\\nIn response to changing workload demands, an RNN evolved by the enforced subpopulations algorithm is introduced to partition L2 cache dynamically.\\nWhen integrating dynamic partitioning of LLC with DVFS on cores and uncore, a co-optimization method using table-based Q-learning achieves much lower energy-delay products than any of the techniques applied individually .\\nTo guarantee efficient and reliable execution in many-core systems, task allocation should consider several aspects, such as heat and communication issues.\\nTargeting the heat interaction of processor cores and NoC routers, Lu \\\\textit{et al.} apply Q-learning to assign tasks to cores based on current temperatures of cores and routers, such that the maximum temperature in the future is minimized.\\nTargeting the non-uniform and hierarchical on/off-chip communication capability in multi-chip many-core systems, core placement optimization leverages deep deterministic policy gradient (DDPG) to map computation onto physical cores, able to work in a manner agnostic to domain-specific information.\\nSome studies pay attention to workflow management and general hardware resource assignment.\\nSmartFlux focuses on the workflow of data-intensive and continuous processing. It intelligently guides asynchronous triggering of processing steps with the help of predictions made by multiple ML models (e.g., SVM, random forest), which indicate whether to execute certain steps and to decide corresponding configurations upon each wave of data.\\nGiven target DNN models, deployment scenarios, platform constraints, and optimization objectives (latency/energy), ConfuciuX~ applies a hybrid two-step scheme for optimal hardware resource assignments (i.e., assigning the number of processing elements and the buffer sizes to each DNN layer), where REINFORCE performs a global coarse-grained search followed by a genetic algorithm for fine-grained tuning.\\nApollo is a general architecture exploration framework for sample-efficient accelerator designs, which leverages ML-based black-box optimization techniques (e.g., Bayesian optimization) to optimize accelerator configurations to satisfy use-specified design constraints.\\nIn heterogeneous systems with CPUs and GPUs, device placement refers to the process of mapping nodes in computational graphs of neural networks onto proper hardware devices.\\nInitially, computational operations are grouped manually, and assigned to devices by REINFORCE that employs a sequence-to-sequence RNN model as the parameterized policy .\\nLater, a hierarchical end-to-end model makes this manual grouping process automatic .\\nThe training speed is further improved by introduction of proximal policy optimization (PPO) .\\nDespite great advance brought by the above approaches, they are not transferable and a new policy should be trained from scratch specifically for each new computational graph.\\nBy encode structure of computational graphs with static graph embeddings or learnable graph embeddings , the trained placement policy exhibits great generalizability to unseen neural networks.\",\n \"id\": \"35decbd9-a2aa-43b2-b9e5-0585c9140768\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"8b48b63d-94af-4ec0-a218-72746bce7952\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Resource Allocation or Management\"\n ],\n [\n \"subsubsection\",\n \"Resource Management and Task Allocation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Resource Management and Task Allocation\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{scheduling}\\nIn classical real-time scheduling problems, the key task is to decide the order, according to which the currently unscheduled jobs should be executed by a single processor, such that the overall performance is optimized.\\nAs multi-core processors have been the mainstream, the scheduling is gradually perplexing.\\nOne major reason is that multiple objectives besides the performance should be carefully considered, such as balanced assignments among various cores and response time fairness.\\nEquipped with the capability to well understand the feedback provided by the environment and to dynamically adjust policies, RL is a common tool for real-time scheduling.\\nTo optimize the execution order of jobs after they are routed to a single CPU core, Whiteson and Stone propose an adaptive scheduling policy that exploits Q-routing, where the scheduler utilizes the router’s Q-table to assess a job’s priority and decides jobs' ordering accordingly so as to maximize the overall utility.\\nIn multi-core systems, Fedorova~\\\\textit{et al.} present a blueprint for a self-tuning scheduling algorithm based on the value-based temporal-difference method in RL, aiming to maximize a cost function that is an arbitrary weighted sum of metrics of interest.\\nThis algorithm is then improved to be a general method for online scheduling of parallel jobs , where the value functions are approximated by a parameterized fuzzy rulebase. This scheduling policy always selects to execute jobs with the maximum value functions in the job queue, which possibly preempts currently running jobs and squeezes some jobs into fewer CPUs than they ideally require, with the goal of achieving optimized long-term utility.\",\n \"id\": \"5a43dcc0-acd3-4daf-bbb5-f9e380b060a3\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"8b48b63d-94af-4ec0-a218-72746bce7952\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Resource Allocation or Management\"\n ],\n [\n \"subsubsection\",\n \"Scheduling\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Scheduling\"\n },\n {\n \"cite_extract_rate\": 0.2,\n \"cites\": [\n 3579,\n 3603,\n 1390\n ],\n \"content\": \"\\\\label{data center}\\nWith the rapid scale expansion of data centers, issues that may be trivial in a single machine become increasingly challenging, let alone the inherently complicated problems.\\nEarly work aims at a relatively simple scenario of resource allocation, i.e., to dynamically assign different numbers of servers to multiple applications. \\nThis problem can be modeled as an RL problem with service-level utility functions as rewards:\\nthe arbiter will select a joint action that would bring the maximum total return after consulting local value functions estimated via either table-based methods or function approximation . \\nIn order to better model interactions among multiple agents, a multi-agent coordination algorithm with fuzzy RL can be used to solve the dynamic content allocation in content delivery networks (CDNs), in which each requested content is modeled as an agent, trying to move toward the area with a high demand while coordinating with other agents/contents.\\nA recent innovation pays attention to the placement of virtual machines onto physical machines, so as to minimize the peak-to-average ratio of resource usage across physical machines, where PPO and hindsight imitation learning are evaluated.\\nTo improve data center performance and quality of experience (QoE) for users, ML-based techniques have been explored in a few directions.\\n\\\\textcircled{\\\\small{1}} It is important to efficiently schedule jobs and effectively diagnose stragglers within jobs.\\nAiming at traffic optimization (e.g., flow scheduling, load balancing) in data centers, Chen~\\\\textit{et al.} develop a two-level RL system: peripheral systems, which are trained by DDPG, reside on end-hosts and locally make instant traffic optimization decisions for short flows; the central system, which is trained by policy gradient, aggregates global traffic information, guides behaviors of peripheral systems, and makes traffic optimization decisions for long flows.\\nDecima exploits GNNs to represent cluster information and dependency among job stages, so that the RL-based scheduler can automatically learn workload-specific scheduling policies to schedule data processing jobs with complex dependency. \\nHound combines statistical ML with meta-learning to diagnose causes of stragglers at data-center-scale jobs.\\n\\\\textcircled{\\\\small{2}} It is essential to deploy an intelligent data-center-level cache.\\nDeepCache employs an LSTM encoder-decoder model to predict future content popularity, which can be combined with existing cache policies to make smarter decisions.\\nSong \\\\textit{et al.} apply gradient boosting machines to mimic a relaxed Belady algorithm that evicts an object whose next request is beyond a reuse distance threshold but not necessarily the farthest in the future.\\nPhoebe is an online cache replacement framework leveraging DDPG to predict priorities of objects and to conduct eviction accordingly.\\nConsidering non-history based features, Wang \\\\textit{et al.}~ build a decision tree to predict whether the requested file will be accessed only once in the future. These one-time-access files will be directly sent to users without getting into cache, to avoid cache pollution. \\n\\\\textcircled{\\\\small{3}} From the workload perspective, video workloads on CDNs or clusters are prevalent but their optimization is quite challenging: first, network conditions fluctuate overtime and a variety of QoE goals should be balanced simultaneously; second, only coarse decisions are available and current decisions will have long-term effects on following decisions.\\nThis scenario naturally matches the foundation of RL-based techniques.\\nTo optimize users' QoE of streaming videos, adaptive bitrate algorithms have been recognized as the primary tool used by content providers, which are executed on client-side video players and dynamically choose a bitrate for each video chunk based on underlying network conditions.\\nPensieve applies asynchronous advantage actor-critic to select proper bitrate for future video chunks based on resulting performance from past decisions.\\nWhen considering large-scale video workloads in hybrid CPU-GPU clusters, performance degradation often comes from uncertainty and variability of workloads, and unbalanced use of heterogeneous resources. \\nTo accommodate this, Zhang \\\\textit{et al.} use two deep Q-networks to build a two-level task scheduler, where the cluster-level scheduler selects proper execution nodes for mutually independent video tasks and the node-level scheduler assigns interrelated video subtasks to appropriate computing units. This scheme enables the scheduling model to adjust policies according to runtime status of cluster environments, characteristics of video tasks, and dependency among video tasks.\",\n \"id\": \"de7c5699-73a2-48d8-9e75-ff1b60c4f5f5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Data Center Management\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Data Center Management\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:dse_code_compiler}\",\n \"id\": \"68c1d9ca-e725-4975-ad78-356d7848d1d1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Code Generation and Compiler\"\n ]\n ],\n \"subsections\": [\n \"21901bea-3ca4-4650-b39a-4ecbde92fab7\",\n \"9d772652-3e90-4f57-98a3-ad3bb92cfe28\"\n ],\n \"title\": \"Code Generation and Compiler\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 5397,\n 5408,\n 4096,\n 5410\n ],\n \"content\": \"\\\\label{code_generation}\\nDue to the similarities in syntax and semantics between programming languages and natural languages, the problem of code generation or translation is often modeled as an NLP problem or a neural machine translation (NMT) problem. \\nHere, we would like to bring up a brief discussion. \\nFor more reference, a comprehensive survey detailedly contrasts programming languages against natural languages, and discusses how these similarities and differences drive the design and application of different ML models in code.\\nTargeting code completion, several statistical language models (N-gram model, RNN, and a combination of these two) are explored to select sentences that have the highest probability and satisfy constraints to fill up partial programs with holes.\\nAs for code generation, CLgen trains LSTM models by a corpus of hand-written code to learn semantics and structures of OpenCL programs, and generates human-like programs via iteratively sampling from the learned model.\\nTargeting program translation, NMT-based techniques are widely applied to migrate code from one language to another.\\nFor example, a tree-to-tree model with the encoder-decoder structure effectively translates programs from Java to C\\\\# ;\\nthe sequence-to-sequence (seq2seq) model can translate from CUDA to OpenCL .\\nRather than translating between high-level programming languages, Coda translates binary executables to the corresponding high-level code, which employs a tree-to-tree encoder-decoder structure for code sketch generation and an ensembled RNN-based error predictor for iterative error correction on the generated code.\\nNotably, these supervised NMT-based techniques may confront several issues: difficulty to generalize to programs longer than training ones, limited size of vocabulary sets, and scarcity of aligned input-output data.\\nFully counting on unsupervised machine translation, TransCoder adopts a transformer architecture and uses monolingual source code to translate among C++, Java, and Python.\",\n \"id\": \"21901bea-3ca4-4650-b39a-4ecbde92fab7\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"68c1d9ca-e725-4975-ad78-356d7848d1d1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Code Generation and Compiler\"\n ],\n [\n \"subsubsection\",\n \"Code Generation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Code Generation\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 5402,\n 5399,\n 5411\n ],\n \"content\": \"\\\\label{sec:compiler}\\nThe complexity of compilers grows with the complexity of computer architectures and workloads. \\nML-based techniques can optimize compilers from many perspectives, such as instruction scheduling, compiler heuristics, the order to apply optimizations, hot path identification, auto-vectorization, and compilation for specific applications.\\n\\\\textcircled{\\\\small{1}} For instruction scheduling, the preference function of one scheduling over another can be computed by the temporal difference algorithm in RL .\\nRegarding scheduling under highly-constrained code optimization, the projective reparameterization enables automatic instruction scheduling under constraints of data-dependent partial orders over instructions.\\n\\\\textcircled{\\\\small{2}} For improving compiler heuristics, Neuro-Evolution of Augmenting Topologies (NEAT) improves instruction placement heuristics by tuning placement cost functions.\\nTo avoid manual feature engineering, LSTM-based model automatically learns compiler heuristics from raw code, which constructs proper embeddings of programs and simultaneously learn the optimization process.\\n\\\\textcircled{\\\\small{3}} For choosing the appropriate order to apply different optimizations, NEAT can automatically generate beneficial optimization orderings for each method in a program.\\n\\\\textcircled{\\\\small{4}} For path profiling, CrystalBall uses an LSTM model to statically identify hot paths, the sequences of instructions that are frequently executed. As CrystalBall only relies on IRs, it avoids manual feature crafting and is independent of language or platform. \\n\\\\textcircled{\\\\small{5}} For automatic vectorization, Mendis~\\\\textit{et al.} leverage imitation learning to mimic optimal solutions provided by superword-level-parallelism based vectorization .\\n\\\\textcircled{\\\\small{6}} For compilation of specific applications, there are studies improving compilation for approximate computing or DNN applications.\\nConsidering compilation for approximate computing, Esmaeilzadeh \\\\textit{et al.} propose a program transformation method, which trains MLPs to mimic regions of approximable code and eventually replaces the original code with trained MLPs. The following work extends this algorithmic transformation to GPUs.\\nConsidering compilation for DNNs, RELEASE utilizes PPO to search optimal compilation configurations for DNNs.\\nEGRL optimizes memory placement of DNN tensors during compilation, which combines GNNs, RL, and evolutionary search to figure out optimal mapping onto different on-board memory components (i.e., SRAM, LLC, and DRAM).\",\n \"id\": \"9d772652-3e90-4f57-98a3-ad3bb92cfe28\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"68c1d9ca-e725-4975-ad78-356d7848d1d1\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Code Generation and Compiler\"\n ],\n [\n \"subsubsection\",\n \"Compiler\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Compiler\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{chip_design}\\nAs technology scales down, the increased design complexity comes with growing process variations and reduced design margins, making chip design an overwhelmingly complex problem for human designers.\\nRecent advancements in ML create a chance to transform chip design workflows.\",\n \"id\": \"e25c32d8-64dc-4364-8501-8063014c67c6\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"87e1e057-9e86-4a5c-afa3-431c75b16a64\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Chip Design and Design Automation\"\n ]\n ],\n \"subsections\": [\n \"4b384d57-c0f5-44ec-ae75-2d05825729b5\",\n \"75c9f4ed-ccff-4fa2-af01-38a78105e680\"\n ],\n \"title\": \"Chip Design and Design Automation\"\n },\n {\n \"cite_extract_rate\": 0.30000000000000004,\n \"cites\": [\n 5414,\n 727,\n 5412,\n 5413,\n 5415,\n 166\n ],\n \"content\": \"\\\\label{sec:dse_analog}\\nCompared with the highly automated digital design counterpart, analog design usually demands many manual efforts and domain expertise.\\nFirst, analog circuits have large design spaces to search proper topology and device sizes. Second, there is an absence of a general framework to optimize or evaluate analog designs, and design specifications often vary case by case.\\nRecently, ML techniques have been introduced to expedite analog design automation.\\nWe discuss these studies following the top-down flow of analog design: in the circuit level, a proper circuit topology is selected to satisfy system specifications; then in the device level, device sizes are optimized subject to various objectives. These two steps compose of pre-layout designs.\\nAfter circuit schematics are carefully designed, analog layouts in the physical level are generated.\\n\\\\textcircled{\\\\small{1}} \\nIn the circuit level, there is an attempt towards automatic circuit generation currently targeting two-port linear analog circuits .\\nThe design specifications are encoded by a hypernetwork to generate weights for an RNN model, which is trained to select circuit components and their configurations.\\n\\\\textcircled{\\\\small{2}} \\nIn the device level, the combination of RL and GNNs enables automatic transistor sizing , which is able to generalize across different circuit topologies or different technology nodes.\\nAutoCkt introduces transfer learning techniques into deep RL for automatic sizing, achieving 40$\\\\times$ speedup over a traditional genetic algorithm.\\nRosa \\\\textit{et al.} provide comprehensive discussions of how to address automatic sizing and layout of analog ICs via deep learning and ANNs.\\n\\\\textcircled{\\\\small{3}}\\nIn the physical level, GeniusRoute automates analog routing through the guidance learned by a generative neural network.\\nThe analog placements and routing are represented as images to pass through a variational autoencoder (VAE) to learn routing likelihoods of each region. GeniusRoute achieves competitive performance to manual layouts and is capable to generalize to circuits of different functionality.\\nLiu \\\\textit{et al.} apply multi-objective Bayesian optimization to optimize combinations of net weighting parameters, which could significantly change floor plans and placement solutions, so as to improve analog layouts of building block circuits.\\n\\\\begin{table}[tbp]\\n\\\\vspace{-12pt}\\n\\\\caption{Summary of applying ML techniques as the design methodology for design automation.}\\n\\\\vspace{-10pt}\\n\\\\label{table:dse_eda}\\n\\\\centering\\n \\\\tiny\\n \\\\renewcommand{\\\\arraystretch}{0.9}\\n \\\\setlength{\\\\tabcolsep}{12pt}\\n\\\\begin{tabular}{c|c|c|c}\\n\\\\toprule\\n\\\\multicolumn{2}{c|}{\\\\textbf{Domain}} & \\\\textbf{Task} & \\\\textbf{Technique} \\\\\\\\ \\\\midrule\\n\\\\multirow{4}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Analog \\\\\\\\ Design \\\\\\\\ $( \\\\S$ \\\\ref{sec:dse_analog}$)$ \\\\end{tabular}}} & \\\\textbf{Circuit Level} & Generating circuit topology & RNN and hypernetwork \\\\\\\\ \\\\cline{2-4} \\n & \\\\textbf{Device Level} & Device sizing & Actor critic , ANN \\\\\\\\ \\\\cline{2-4} \\n & \\\\multirow{2}{*}{\\\\textbf{Physical Level}} & Routing & VAE \\\\\\\\ \\\\cline{3-4} \\n & & Optimizing layout configurations & Multi-objective Bayesian optimization \\\\\\\\ \\\\cline{2-4} \\n \\\\hline\\n\\\\multirow{12}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Digital \\\\\\\\ Design \\\\\\\\ $( \\\\S$ \\\\ref{sec:dse_digital}$)$\\\\end{tabular}}} & \\\\multirow{3}{*}{\\\\textbf{HLS}} & Optimizing loop unrolling pragma & Random forest \\\\\\\\ \\\\cline{3-4} \\n & & Optimizing placement of multiple pragmas & Bayesian optimization \\\\\\\\ \\\\cline{3-4} \\n & & Optimizing resource pragma & Actor-critic and GNN \\\\\\\\ \\\\cline{2-4} \\n & \\\\multirow{3}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Logic \\\\\\\\ Synthesis\\\\end{tabular}}} & Selecting proper optimizers & MLP \\\\\\\\ \\\\cline{3-4} \\n & & Logic optimization & Policy gradient , actor-critic \\\\\\\\ \\\\cline{3-4} \\n & & Determining the maximum error of each node & Q-learning \\\\\\\\ \\\\cline{2-4} \\n & \\\\multirow{5}{*}{\\\\textbf{\\\\begin{tabular}[c]{@{}c@{}}Physical \\\\\\\\ Synthesis\\\\end{tabular}}} & Optimizing flip-flop placement in clock networks & K-means clustering \\\\\\\\ \\\\cline{3-4} \\n & & Optimizing clock tree synthesis & Conditional GAN \\\\\\\\ \\\\cline{3-4} \\n & & Optimizing memory cell placement & PPO \\\\\\\\ \\\\cline{3-4} \\n & & Optimizing standard cell placement & Cast as an NN training problem \\\\\\\\ \\\\cline{3-4} \\n & & Fix design rule violations & PPO \\\\\\\\ \\\\bottomrule\\n\\\\end{tabular}\\n\\\\vspace{-12pt}\\n\\\\end{table}\",\n \"id\": \"4b384d57-c0f5-44ec-ae75-2d05825729b5\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 20,\n \"parent_id\": \"e25c32d8-64dc-4364-8501-8063014c67c6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Chip Design and Design Automation\"\n ],\n [\n \"subsubsection\",\n \"Analog Design\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Analog Design\"\n },\n {\n \"cite_extract_rate\": 0.21428571428571402,\n \"cites\": [\n 5414,\n 5415,\n 5413\n ],\n \"content\": \"\\\\label{sec:dse_digital}\\nFor the studies applying ML techniques to directly optimize digital designs, we organize them following a top-down flow, i.e., HLS, logic synthesis, and physical synthesis.\\nThe design space exploration in HLS designs usually relates to properly assigning directives (pragmas) in high-level source code, since directives significantly impact the quality of HLS designs by controlling parallelism, scheduling, and resource usage.\\nThe optimization goal is often to find Pareto solutions between different objectives or to satisfy pre-defined constraints. \\nWith IR analysis, the employment of a random forest is able to select suitable loop unrolling factors to optimize a weighted sum of execution latency and resource usage .\\nProspector uses Bayesian optimization to optimize placement of directives (loop unrolling/pipelining, array partitioning, function inlining, and allocation), aiming to find Pareto solutions between execution latency and resource utilization in FPGAs. \\nIronMan targets Pareto solutions between different resources while keeping the latency unchanged.\\nIt combines GNNs with RL to conduct a finer-grained design exploration in the operation level, pursuing optimal resource allocation strategies by optimizing assignments of the resource pragma.\\nIn logic synthesis, RTL-designs or logic networks are represented by directed acyclic graphs (DAGs). The goal is to optimize logic networks subject to certain constraints.\\nLSOracle employs an MLP to automatically decide which one of the two optimizers should be applied on different parts of circuits.\\nThe logic optimization can be formulated as an RL problem solved by the policy gradient or the advantage actor-critic : the state is the current status of a design; the action is a transformation between two DAGs with equivalent I/O behaviors; the optimization objective is to minimize area or delay of designs.\\nQ-ALS aims at approximate logic synthesis and embeds a Q-learning agent to determine the maximum tolerable error of each node in a DAG, such that the total error rates at primary outputs are bounded by pre-specified constraints.\\nIn physical synthesis, placement optimization is a popular topic.\\n\\\\textcircled{\\\\small{1}} \\nTo optimize flip-flop placement in clock networks, Wu \\\\textit{et al.} apply a modified K-means clustering to group post-placement flip-flops, and relocate these clusters by reducing the distance between flip-flops and their drivers while minimizing disruption of original placement results. \\nTo optimize clock tree synthesis (CTS), Lu \\\\textit{et al.} train a regression model that takes pre-CTS placement images and CTS configurations as inputs to predict post-CTS metrics (clock power, clock wirelength, and maximum skew), which is used as the supervisor to guide the training of a conditional GAN, such that the well-trained generator can recommend CTS configurations leading to optimized clock trees. \\n\\\\textcircled{\\\\small{2}} \\nAiming at cell placement, a deep RL approach is introduced to place macros (memory cells), after which standard cells are placed by a force-directed method. \\nThis method is able to generalize to unseen netlists, and outperforms RePlAce yet several times slower.\\nDREAMPlace casts the analytical standard cell placement optimization into a neural network training problem, achieving over 30$\\\\times$ speedup without quality degradation compared to RePlAce.\\nNVCell is an automated layout generator for standard cells, which employs RL to fix DRVs after placement and routing.\\n\\\\textcircled{\\\\small{3}} \\nML-based techniques demonstrates their versatility in many design automation tasks, such as post-silicon variation extraction by sparse Bayesian learning, and post-silicon timing tuning to mitigate the effects caused by process variation .\",\n \"id\": \"75c9f4ed-ccff-4fa2-af01-38a78105e680\",\n \"level\": \"subsubsection\",\n \"origin_cites_number\": 14,\n \"parent_id\": \"e25c32d8-64dc-4364-8501-8063014c67c6\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"ML as Design Methodology\"\n ],\n [\n \"subsection\",\n \"Chip Design and Design Automation\"\n ],\n [\n \"subsubsection\",\n \"Digital Design\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Digital Design\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:dis}\\nIn this section, we discuss limitations and potentials of ML techniques for computer architecture and systems, which span the entire development and deployment stack that involves data, algorithms, implementation, and targets.\\nWe also envision that the application of ML techniques could be the propulsive force for \\\\textit{hardware agile development}.\\n\\\\vspace{-5pt}\",\n \"id\": \"a80b2264-c6ef-4ec9-ae9a-838d036e281c\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5631fef3-6389-407d-9101-745f45240141\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Discussion and Potential Directions\"\n ]\n ],\n \"subsections\": [\n \"4da21a0b-f62c-4e97-84e3-786911c6b5c8\",\n \"bf55ec88-5f58-456f-b066-c5e92f556071\",\n \"ac2615bc-0a08-4fa7-932a-166c6e497c28\",\n \"56d8fa9d-3bba-4c5d-9a9d-e6d1744157e3\",\n \"5afb9503-6a23-454b-90e9-dd19be81420a\"\n ],\n \"title\": \"Discussion and Potential Directions\"\n },\n {\n \"cite_extract_rate\": 0.16666666666666602,\n \"cites\": [\n 5416\n ],\n \"content\": \"Data are the backbone to ML, however, perfect datasets are sometimes non-available or prohibitively expensive to obtain in computer architecture and system domain.\\nHere, we would like to scrutinize two points, the gap between small data and big data, and non-perfect data.\\n\\\\textcircled{\\\\small{1}} \\nIn some EDA problems, such as placement and routing in physical synthesis, the simulation or evaluation is extremely expensive , leading to data scarcity. As ML models usually require enough data to learn underlying statistics and make decisions, this gap between small data and big data often limits the capability of ML-based techniques. There have been different attempts to bridge this gap. From the algorithm side, algorithms that can work with small data await to be developed, where one current technique is Bayesian optimization that is effective in small parameter space ; active learning , which significantly improve sample efficiency, may also be a cure to this problem. From the data side, generative methods can be used to generate synthetic data , mitigating data scarcity.\\n\\\\textcircled{\\\\small{2}} \\nRegarding non-perfect data, even if some EDA tools produce a lot of data, they are not always properly labeled nor presented in the form suitable to ML models. In the absence of perfectly labeled training data, possible alternatives are to use unsupervised learning, self-supervised learning , or to combine supervised with unsupervised techniques . Meanwhile, RL could be a workaround where training data can be generated on-the-fly.\\n\\\\vspace{-5pt}\",\n \"id\": \"4da21a0b-f62c-4e97-84e3-786911c6b5c8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"a80b2264-c6ef-4ec9-ae9a-838d036e281c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Discussion and Potential Directions\"\n ],\n [\n \"subsection\",\n \"Bridging Data Gaps\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Bridging Data Gaps\"\n },\n {\n \"cite_extract_rate\": 0.45454545454545403,\n \"cites\": [\n 5202,\n 5413,\n 5416,\n 5407,\n 3588\n ],\n \"content\": \"Despite the current achieved accomplishments, we are still expecting novel ML algorithms or schemes to further improve system modeling and optimization, with respect scalability, domain knowledge interpretability, etc.\\n\\\\textbf{New ML Schemes.}\\nClassical analytic-based methods usually adopt a bottom-up or top-down procedure, encouraging ML-based techniques to distill hierarchical structures of systems/architecture.\\nOne example is hierarchical RL that has flexible goal specifications and learns goal-directed behaviors in complex environments with sparse feedback. Such kind of models enables more flexible and effective multi-level design and control.\\nAdditionally, many system optimizations involve participation of multiple agents, such as NoC routing, which are naturally suitable to the realm of multi-agent RL . These agents can be fully cooperative, fully competitive, or a mix of the two, enabling versatility of system optimization.\\nAnother promising approach is self-supervised learning , beneficial in both improving model robustness and mitigating data scarcity.\\nWhile applying a single ML method solely has led to powerful results, hybrid methods, i.e., combining different ML techniques or combining ML techniques with heuristics, unleash more opportunities. For example, RL can be combined with genetic algorithms for hardware resource assignment .\\n\\\\textbf{Scalability.}\\nThe system scaling-up poses challenges on scalability issues.\\nFrom the algorithm side, multi-level techniques can help reduce the computation complexity, e.g., multi-level Q-learning for DVFS .\\nOne implicit workaround is to leverage transfer learning:\\nthe pre-training is a one-time cost, which can be amortized in each future use; the fine-tuning provides flexibility between a quick solution from the pre-trained model and a longer yet better one for a particular task. \\nSeveral examples are discussed in Section \\\\ref{chip_design}.\\n\\\\textbf{Domain Knowledge and Interpretability.}\\nMaking better use of domain knowledge unveils possibilities to choose more proper models for different system problems and provide more intuitions or explanations of why and how these models work.\\nBy making analogy of semantics between memory access patterns/program languages and natural languages, the prefetching or code generation problems can be modeled as NLP problems, as discussed in Section \\\\ref{cache} and Section \\\\ref{code_generation}. \\nBy making analogy of graphical representations in many EDA problems, where data are intrinsically presented as graphs (e.g., circuits, logic netlists or IRs), GNNs are expected to be powerful in these fields . \\nSeveral examples are provided in Section \\\\ref{sec:model_eda} and Section \\\\ref{chip_design}.\",\n \"id\": \"bf55ec88-5f58-456f-b066-c5e92f556071\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"a80b2264-c6ef-4ec9-ae9a-838d036e281c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Discussion and Potential Directions\"\n ],\n [\n \"subsection\",\n \"Developing Algorithms\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Developing Algorithms\"\n },\n {\n \"cite_extract_rate\": 0.5714285714285711,\n \"cites\": [\n 5418,\n 7633,\n 9139,\n 5417\n ],\n \"content\": \"To fully benefit from ML-based methods, we need to consider practical implementations, appropriate selection of deployment scenarios, and post-deployment model maintenance.\\n\\\\textbf{Better Implementations.}\\nTo enable practical implementations of ML-based techniques, improvement can be made from either the model side or software/hardware co-design .\\nFrom the model level, network pruning and model compression reduce the number of operations and model size ; weight quantization improves computation efficiency by reducing the precision of operations/operands .\\nFrom the co-design level, strategies that have been used for DNN acceleration could also be used in applying ML for system.\\n\\\\textbf{Appropriate Scenarios: online vs. offline.}\\nWhen deploying ML-based techniques for system designs, it is crucial to deliberate design constraints under different scenarios.\\nGenerally, existing work falls into two categories. \\n\\\\textcircled{\\\\small{1}} \\nML-based techniques are deployed online or during runtime, no matter the training phase is performed online or offline. Obviously, the model complexity and runtime overhead are often strictly limited by specific constraints, e.g., power/energy, timing/latency, area, etc. \\nTo take one more step, if the online training/learning is further desired, the design constraint will be more stringent. One promising approach is to employ semi-online learning models, which have been applied to solve some classical combinatorial optimization problems, such as bipartite matching and caching . These models enable smooth interpolation between the best possible online and offline training algorithms.\\n\\\\textcircled{\\\\small{2}} ML-based techniques are applied offline, which often refers to architectural design space exploration. Such problems leverage ML-based techniques to guide system implementation, and once the designing phase is completed, ML models will not be invoked again. Thus, the offline applications can tolerate relatively higher overheads.\\n\\\\textbf{Model Maintenance.}\\nIn the case of offline training and online deployment, ML models employed for computer architecture domain, as in other scenarios, require regular maintenance and updating to meet performance expectations,\\nsince workload variations over time and hardware aging often cause data drift or concept drift .\\nTo proactively circumvent performance degradation of ML models, some measures could be taken during post-deployment periods.\\n\\\\textcircled{\\\\small{1}}\\nML models can be retrained either at a regular interval or when key performance indicators are below certain thresholds. Retraining models regularly, regardless of their performance, is a more direct way, but it requires a clear understanding of how frequently a model should be updated under its own scenario. The model performance will decline if retraining intervals are too spaced out in the interim. Monitoring key performance indicators relies on a comprehensive panel of measurements that explicitly demonstrate model drift, whereas this may introduce additional hardware/software overhead and incorrect selection of measurements often defeats the intention of this method.\\n\\\\textcircled{\\\\small{2}} During the retraining of ML models, there is often a trade-off between newly collected data and previous data. Properly assigning importance of input data would improve retraining efficacy .\\n\\\\vspace{-5pt}\",\n \"id\": \"ac2615bc-0a08-4fa7-932a-166c6e497c28\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"a80b2264-c6ef-4ec9-ae9a-838d036e281c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Discussion and Potential Directions\"\n ],\n [\n \"subsection\",\n \"Improving Implementations and Deployments\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Improving Implementations and Deployments\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 1695,\n 7739,\n 7612\n ],\n \"content\": \"ML-based techniques are supposed to be applicable in both current architectures and emerging systems, leading to long-term advancement in computer architecture and systems.\\n\\\\textbf{Non-homogeneous Components.}\\nDesign and development for computer architectures are often based upon earlier-generation architectures of similar purpose, but commonly rely on next-generation hardware components that were not present in earlier generations. Examples include employment of new device nodes with technology scaling, and replacement of conventional constituents in memory systems with NVM- or PIM-based components.\\nIn addition to the heterogeneity of components from different generations, one architecture or system usually consists of both standard parts from library and specialized/customized hardware components.\\nThis provides the motivation that ML-assisted architectures/systems should have the flexibility to transfer among different-generation components, and to support standard and specialized parts simultaneously.\\n\\\\textbf{Non-homogeneous Applications.}\\nIn computer architecture and system design, some issues are universal, while others may arise with the advent of new architecture/systems and new workloads.\\n\\\\textcircled{\\\\small{1}} For evergreen design areas, several examples include caching in hardware/software/data centers (Section \\\\ref{cache} and Section \\\\ref{data center}), resource management and task allocation in single/multi-/many-core CPUs and heterogeneous systems (Section \\\\ref{resource}), NoC design under various scenarios (Section \\\\ref{NoC}), etc.\\n\\\\textcircled{\\\\small{2}} For problems aroused from new systems/workloads, transfer learning and meta-learning could be helpful in either exploring new heuristics or directly deriving design methodology. \\nFor example, combining meta-learning with RL allows training a \\\"meta\\\" agent that is designed to adapt to a specific workload with only a few observations.\\n\\\\vspace{-5pt}\",\n \"id\": \"56d8fa9d-3bba-4c5d-9a9d-e6d1744157e3\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"a80b2264-c6ef-4ec9-ae9a-838d036e281c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Discussion and Potential Directions\"\n ],\n [\n \"subsection\",\n \"Supporting Non-homogeneous Tasks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Supporting Non-homogeneous Tasks\"\n },\n {\n \"cite_extract_rate\": 1,\n \"cites\": [\n 4334\n ],\n \"content\": \"Even though ML-based modeling significantly reduces the evaluation cost during design iteration, making great strides towards the landing of hardware agile development, there is still a long way to go in the ML-based design methodology prospective.\\nOne ultimate goal might be the fully automated design, which should entangle two core capabilities: holistic optimization in system-wise, and easy migration across different systems, to enable rapid and agile hardware design.\\n\\\\textbf{Holistic Optimization.}\\nFueled by recent advancements, ML techniques have been increasingly explored and exploited in computer system design and optimization . \\nThe target problems that await further endeavors could be multi-objective optimizations under highly constrained situations, or optimizing several components in a system simultaneously.\\nWe envisage an ML-based system-wise and holistic framework with a panoramic vision: it should be able to leverage information from different levels of systems in synergy, so that it could thoroughly characterize system behaviors as well as their intrinsically hierarchical abstractions; it should also be able to make decisions in different granularity, so that it could control and improve systems precisely and comprehensively.\\n\\\\textbf{Portable, Rapid, and Agile.}\\nStriving for portable, rapid, and agile hardware design, there are two potential directions.\\n\\\\textcircled{\\\\small{1}} The well-designed interfaces between systems/architectures and ML-based techniques would facilitate the portability across different platforms, since ML models can perform well without explicit descriptions of the target domain.\\n\\\\textcircled{\\\\small{2}} The proliferation of ML-based techniques have more or less transformed the workflow of design automation, directly driving rapid and agile hardware design.\\nWe expect GNNs make better use of naturally graphical data in EDA field; we expect deep RL be a powerful and general-purpose tool for many EDA optimization problems, especially when the exact heuristic or objective is obscure; we expect these ML-based design automation tools enhance designers' productivity and thrive in the community.\\n\\\\vspace{-5pt}\",\n \"id\": \"5afb9503-6a23-454b-90e9-dd19be81420a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"a80b2264-c6ef-4ec9-ae9a-838d036e281c\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Discussion and Potential Directions\"\n ],\n [\n \"subsection\",\n \"Facilitating General Tool Design and Hardware Agile Development\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Facilitating General Tool Design and Hardware Agile Development\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:con}\\nThe flourishing of ML would be retarded without the great systems and powerful architectures supportive to run these algorithms at scale. \\nNow, it is the time to return the favor and let ML transform the way that computer architecture and systems are designed.\\nExisting work that applies ML for computer architecture/systems roughly falls into two categories: ML-based fast modeling that involves performance metrics or some other criteria of interest, and ML-based design methodology that directly leverages ML as the design tool.\\nWe hope to see the virtuous cycle, in which ML-based techniques are efficiently running on the most powerful computers with the pursuit of designing the next generation computers. We hope ML-based techniques could be the impetus to the revolution of computer architecture and systems.\\n\\\\vspace{-5pt}\\n\\\\bibliographystyle{ACM-Reference-Format}\\n\\\\bibliography{ref}\\n\\\\end{document}\\n\\\\endinput\",\n \"id\": \"79dc37a5-e0c0-498b-b911-e6bc545700a1\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"5631fef3-6389-407d-9101-745f45240141\",\n \"prefix_titles\": [\n [\n \"title\",\n \"A Survey of Machine Learning for Computer Architecture and Systems\"\n ],\n [\n \"section\",\n \"Conclusion\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Conclusion\"\n }\n]","isConversational":false},"survey_id":{"kind":"number","value":98,"string":"98"},"all_cites":{"kind":"list like","value":[1390,553,2219,166,4378,5390,5389,5391,5392,5393,5395,5396,8916,5394,5397,5409,3579,5400,5399,5406,5398,5407,5403,3603,5408,5401,5405,4096,5402,5404,3600,5410,5411,5414,727,5412,5413,5415,5416,5202,3588,5418,7633,9139,5417,1695,7739,7612,4334],"string":"[\n 1390,\n 553,\n 2219,\n 166,\n 4378,\n 5390,\n 5389,\n 5391,\n 5392,\n 5393,\n 5395,\n 5396,\n 8916,\n 5394,\n 5397,\n 5409,\n 3579,\n 5400,\n 5399,\n 5406,\n 5398,\n 5407,\n 5403,\n 3603,\n 5408,\n 5401,\n 5405,\n 4096,\n 5402,\n 5404,\n 3600,\n 5410,\n 5411,\n 5414,\n 727,\n 5412,\n 5413,\n 5415,\n 5416,\n 5202,\n 3588,\n 5418,\n 7633,\n 9139,\n 5417,\n 1695,\n 7739,\n 7612,\n 4334\n]","isConversational":false},"Bertopic_CD":{"kind":"number","value":1.2527642270767,"string":"1.252764"}}},{"rowIdx":66,"cells":{"authors":{"kind":"list like","value":["Sebastian Houben","Stephanie Abrecht","Maram Akila","Andreas Bär","Felix Brockherde","Patrick Feifel","Tim Fingscheidt","Sujan Sai Gannamaneni","Seyed Eghbal Ghobadi","Ahmed Hammam","Anselm Haselhoff","Felix Hauser","Christian Heinzemann","Marco Hoffmann","Nikhil Kapoor","Falk Kappel","Marvin Klingner","Jan Kronenberger","Fabian Küppers","Jonas Löhdefink","Michael Mlynarski","Michael Mock","Firas Mualla","Svetlana Pavlitskaya","Maximilian Poretschkin","Alexander Pohl","Varun Ravi-Kumar","Julia Rosenzweig","Matthias Rottmann","Stefan Rüping","Timo Sämann","Jan David Schneider","Elena Schulz","Gesina Schwalbe","Joachim Sicking","Toshika Srivastava","Serin Varghese","Michael Weber","Sebastian Wirkert","Tim Wirtz","Matthias Woehrle"],"string":"[\n \"Sebastian Houben\",\n \"Stephanie Abrecht\",\n \"Maram Akila\",\n \"Andreas Bär\",\n \"Felix Brockherde\",\n \"Patrick Feifel\",\n \"Tim Fingscheidt\",\n \"Sujan Sai Gannamaneni\",\n \"Seyed Eghbal Ghobadi\",\n \"Ahmed Hammam\",\n \"Anselm Haselhoff\",\n \"Felix Hauser\",\n \"Christian Heinzemann\",\n \"Marco Hoffmann\",\n \"Nikhil Kapoor\",\n \"Falk Kappel\",\n \"Marvin Klingner\",\n \"Jan Kronenberger\",\n \"Fabian Küppers\",\n \"Jonas Löhdefink\",\n \"Michael Mlynarski\",\n \"Michael Mock\",\n \"Firas Mualla\",\n \"Svetlana Pavlitskaya\",\n \"Maximilian Poretschkin\",\n \"Alexander Pohl\",\n \"Varun Ravi-Kumar\",\n \"Julia Rosenzweig\",\n \"Matthias Rottmann\",\n \"Stefan Rüping\",\n \"Timo Sämann\",\n \"Jan David Schneider\",\n \"Elena Schulz\",\n \"Gesina Schwalbe\",\n \"Joachim Sicking\",\n \"Toshika Srivastava\",\n \"Serin Varghese\",\n \"Michael Weber\",\n \"Sebastian Wirkert\",\n \"Tim Wirtz\",\n \"Matthias Woehrle\"\n]","isConversational":false},"survey_title":{"kind":"string","value":"Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"},"year":{"kind":"string","value":"2021"},"date":{"kind":"timestamp","value":"2021-04-29T09:54:54Z","string":"2021-04-29T09:54:54Z"},"category":{"kind":"string","value":"cs.LG"},"abstract":{"kind":"string","value":" The use of deep neural networks (DNNs) in safety-critical applications like mobile health and autonomous driving is challenging due to numerous model-inherent shortcomings. These shortcomings are diverse and range from a lack of generalization over insufficient interpretability to problems with malicious inputs. Cyber-physical systems employing DNNs are therefore likely to suffer from \\emph{safety concerns}. In recent years, a zoo of state-of-the-art techniques aiming to address these safety concerns has emerged. This work provides a structured and broad overview of them. We first identify categories of insufficiencies to then describe research activities aiming at their detection, quantification, or mitigation. Our paper addresses both machine learning experts and safety engineers: The former ones might profit from the broad range of machine learning (ML) topics covered and discussions on limitations of recent methods. The latter ones might gain insights into the specifics of modern ML methods. We moreover hope that our contribution fuels discussions on desiderata for ML systems and strategies on how to propel existing approaches accordingly. "},"structure":{"kind":"list like","value":[{"cite_extract_rate":0,"cites":[],"content":"","id":"569f1197-c628-4e9d-9db6-9757cb02d82a","level":"root","origin_cites_number":0,"parent_id":null,"prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"]],"subsections":["86d6cdb9-2750-4e63-865d-831683dd115b","1fe2887d-3162-404c-bb44-c07ac2772367","380e4e9a-2d10-49f6-add4-04990cddb902","4396c1c3-a3ad-4efa-9f10-99f90060c0b4","e47014ad-1aec-4b32-9241-6aab90a7f0dd","9f13c349-4fb4-458e-8f34-84cd3a93fca0","71d4e278-f67c-443c-84a9-ff0254362119","94174462-51f2-497a-b88f-1ceaadd52aaf","f0f7d7d0-2bd8-48ae-bdb1-126b96887d72","e23cf554-7193-4f8f-a54e-3da02ce3d7c2","3497f966-f76d-439f-b7b6-ad5a513d9e31"],"title":"root"},{"cite_extract_rate":0.5,"cites":[8089,6010,8157],"content":"\\label{sec:intro} \n\\sectionauthor{Sebastian Houben\\textsuperscript{1}, Michael Mock\\textsuperscript{1}, Timo Sämann\\textsuperscript{4}, Gesina Schwalbe\\textsuperscript{3}, Joachim Sicking\\textsuperscript{1}}\nIn barely a decade, deep neural networks (DNNs) have revolutionized the field of machine learning by reaching unprecedented, sometimes superhuman, performances on a growing variety of tasks.\nMany of these neural models have found their way into consumer applications like smart speakers, machine translation engines or content feeds. However, in safety-critical systems, where human life might be at risk, the use of recent DNNs is challenging as various model-immanent insufficiencies are yet difficult to address. \n\\par\nThis paper summarizes the promising lines of research in how to identify, address, and at least partly mitigate these DNN insufficiencies.\nWhile some of the reviewed works are theoretically grounded and foster the overall understanding of training and predictive power of DNNs, others provide practical tools to adapt their development, training or predictions. \nWe refer to any such method as a \\emph{safety mechanism} if it addresses one or several safety concerns in a feasible manner.\nTheir effectiveness in mitigating safety concerns is assessed by \\emph{safety metrics} . \nAs most safety mechanisms target only a particular insufficiency, we conclude that a \\emph{holistic safety argumentation} for a complex DNN-based systems will in many cases rely on a variety of safety mechanisms. \n\\par\nWe structure our review of these mechanisms as follows: Chapter \\ref{sec:dataset_optimization} focuses on \\emph{dataset optimization} for network training and evaluation. It is motivated by the well-known fact that, in comparison to humans, DNNs perform poorly on data that is structurally different from training data. Apart from insufficient generalization capabilities of these models, the data acquisition process and distributional data shifts over time play vital roles. We survey potential counter-measures, \\eg augmentation strategies and outlier detection techniques.\n\\par\nMechanisms that improve on \\emph{robustness} are described in Chapters \\ref{sec:robust_training} and \\ref{sec:adversarial_attacks}, respectively. They deserve attention as DNNs are generally not resilient to common perturbations and adversarial attacks.\n\\par\nChapter \\ref{sec:interpretability} addresses incomprehensible network behavior and reviews mechanisms that aim at \\emph{explainability}, \\eg a more transparent functioning of DNNs.\n\\par\nMoreover, DNNs tend to overestimate their prediction confidence, especially on unseen data. \nStraightforward ways to estimate prediction confidence yield mostly unsatisfying results.\nAmong others, this observation fuelled research on more sophisticated \\emph{uncertainty estimations} (see Chapter \\ref{sec:uncertainty}), \\emph{redundancy mechanisms} (see Chapter \\ref{sec:aggregation}) and attempts to reach \\emph{formal verification} as addressed in Chapter \\ref{sec:verification}.\n\\par\nAt last, many safety-critical applications require not only accurate but also near real-time decisions. This is covered by mechanisms on the DNN \\emph{architectural level} (see Chapter \\ref{sec:architecture}) and furthermore by \\emph{compression} and \\emph{quantization} methods (see Chapter \\ref{sec:compression}).\n\\par\nWe conclude this review of mechanism categories with an outlook on the steps to transfer a carefully arranged combination of safety mechanisms into an actual holistic safety argumentation.","id":"86d6cdb9-2750-4e63-865d-831683dd115b","level":"section","origin_cites_number":6,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Introduction"]],"subsections":[],"title":"Introduction"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:dataset_optimization}\n\\sectionauthor{Matthias Rottmann\\textsuperscript{5}}\nThe performance of a trained model inherently relies on the nature of the underlying dataset.\nFor instance, a dataset with poor variability will hardly result in a model ready for real-world applications.\nIn order to approach the latter, data selection processes such as corner case selection and active learning are of utmost importance.\nThese approaches can help to design datasets that contain the most important information, while preventing the so much desired information from getting lost in an ocean of data.\nFor a given dataset and active learning setups, data augmentation techniques are very common aiming at extracting as much model performance out of the dataset as possible.\nOn the other hand, safety arguments also require the analysis of how a model behaves on out-of-distribution data, data that contains concepts the model has not encountered during training.\nThis is quite likely to happen as our world is under constant change, in other words exposed to a constantly growing domain shift.\nTherefore, these fields are lately gaining interest, also with respect to perception in automated driving.","id":"1fe2887d-3162-404c-bb44-c07ac2772367","level":"section","origin_cites_number":0,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Dataset Optimization"]],"subsections":["36c23394-2c6f-46c4-b022-50c53af7e345","f767152d-e919-4dba-8417-ed1faf4da217","7b5fda9c-0631-4e3b-bf4f-fb2fb6b0d60d","de2e598e-f35b-4b65-945b-25aee539ea0d","e4ebc416-15f6-4517-a761-acd4315ad09b"],"title":"Dataset Optimization"},{"cite_extract_rate":0.6923076923076921,"cites":[1624,8160,3251,3237,3219,8159,8158,3248,8161],"content":"\\label{subsec:dataset_optimization:anomaly_detection}\n\\sectionauthor{Sujan Sai Gannamaneni\\textsuperscript{1}, Matthias Rottmann\\textsuperscript{5}}\nThe terms anomaly, outlier and \\emph{out-of-distribution} (OOD) data detection are often used interchangeably in literature and refer to task of identifying data samples that are not representative of training data distribution. Uncertainty evaluation (\\cf Chapter \\ref{sec:uncertainty}) is closely tied to this field as self-evaluation of models is one of the active areas of research for OOD detection. In particular, for image classification problems it has been reported that neural networks often produce high confidence predictions on OOD data . The detection of such OOD inputs can either be tackled by post-processing techniques that adjust the estimated confidence or by enforcing low confidence on OOD samples during training . Even guarantees that neural networks produce low confidence predictions for OOD samples can be provided under specific assumptions (\\cf ). More precisely, this work utilizes Gaussian mixture models that, however, may suffer from high-dimensional data and require strong assumptions on the distribution parameters.\nSome approaches use generative models like \\emph{GANs} and \\emph{autoencoders} for outlier detection. The models are trained to learn in-distribution data manifolds and will produce higher reconstruction loss for outliers.\n\\par\nFor OOD detection in semantic segmentation, only a few works have been presented so far.\nAngus \\etal present a comparative study of common OOD detection methods, which mostly deal with image-level classification. In addition, they provide a novel setup of relevant OOD datasets for this task. Another work trains a fully convolutional binary classifier that distinguishes image patches from a known set of classes from image patches stemming from an unknown class . The classifier output applied at every pixel will give the per-pixel confidence value for an OOD object. Both of these works perform at pixel level and without any sophisticated feature generation methods specifically tailored for the detection of entire OOD instances.\nUp to now, outlier detection has not been studied extensively for object detection tasks based on benchmark object detection datasets. In , two CNNs are used to perform object detection and binary classification (benign or anomaly) in a sequential fashion, where the second CNN takes the localized object within the image as input.\n\\par\nFrom a safety standpoint, detecting outliers or OOD samples is extremely important and beneficial as training data cannot realistically be large enough to capture all situations. Research in this area is heavily entwined with progress in uncertainty estimation (\\cf Chapter \\ref{sec:uncertainty}) and domain adaptation (\\cf \\refsec{subsec:dataset_optimization:domains}). Extending research works to segmentation and object detection tasks would be particularly significant for leveraging autonomous driving research. In addition to safety, OOD detection can be beneficial in other aspects like when using local expert models. For example, when using an expert model for segmentation of urban driving scenes and another expert model for segmentation of highway driving scenes, an OOD detector could act as trigger on which models can be switched. \n\\par\nWith respect to the approaches presented above, uncertainty-based and generative model-based OOD detection methods are currently promising directions of research. However, it remains an open question whether they can unfold their potential well on segmentation and object detection tasks.","id":"36c23394-2c6f-46c4-b022-50c53af7e345","level":"subsection","origin_cites_number":13,"parent_id":"1fe2887d-3162-404c-bb44-c07ac2772367","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Dataset Optimization"],["subsection","Outlier/Anomaly Detection"]],"subsections":[],"title":"Outlier/Anomaly Detection"},{"cite_extract_rate":0.636363636363636,"cites":[8162,1044,5527,3237,4620,1043,1049],"content":"\\label{subsec:dataset_optimization:active_learning}\n\\sectionauthor{Matthias Rottmann\\textsuperscript{5}}\nIt is widely known that, as a rule of thumb, for the training of any kind of artificial neural network, an increase of training data leads to increased performance.\nObtaining labeled training data, however, is often very costly and time consuming.\n\\emph{Active learning} provides one possible remedy to this problem: Instead of labeling every data point, active learning utilizes a \\emph{query strategy} to request labels from a teacher/an oracle which leverage the model performance most.\nThe survey paper by Settles provides a broad overview regarding query strategies for active learning methods. However, except for \\emph{uncertainty sampling} and \\emph{query by committee}, most of them seem to be infeasible in deep learning applications up to now. Hence, most of the research activities in active deep learning focus on these two query strategies, as we outline in the following.\n\\par\nIt has been shown for image classification that labels corresponding to uncertain samples can leverage the networks' performance significantly and that a combination with semi-supervised learning is promising. In both works, uncertainty of unlabeled samples is estimated via Monte Carlo (MC) dropout inference. MC dropout inference and a chosen number of training epochs are executed alternatingly, after performing MC dropout inference, the unlabeled samples' uncertainties are assessed by means of sample-wise dispersion measures. Samples for which the DNN model is very uncertain about its prediction are presented to an oracle and labeled.\n\\par\nWith respect to object detection, a moderate number of active learning methods has been introduced .\nThese approaches include uncertainty sampling and query-by-committee methods . In , additional algorithmic features specifically tailored for object detection networks are presented, \\ie separate treatment of the localization and classification loss , as well as weak and strong supervision schemes . For semantic segmentation, an uncertainty-sampling-based approach has been presented , which queries polygone masks for image sections of a fixed size ($128 \\times 128$). Queries are performed by means of accumulated entropy in combination with a cost estimation for each candidate image section.\n\\par\nRecently, new methods for estimating the quality of a prediction as well as new uncertainty quantification approaches, \\eg gradient-based ones , have been proposed. It remains an open question whether they are suitable for active learning. Since most of the conducted studies are rather of academic nature, also their applicability to real-life data acquisition is not yet demonstrated sufficiently. In particular, it is not clear whether the proposed active learning schemes, including the label acquistion, for instance in semantic segmentation, is suitable to be performed by human labelers. Therefore, labeling acquisition with a common understanding of the labelers' convenience and suitability for active learning are a promising direction for research and development.","id":"f767152d-e919-4dba-8417-ed1faf4da217","level":"subsection","origin_cites_number":11,"parent_id":"1fe2887d-3162-404c-bb44-c07ac2772367","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Dataset Optimization"],["subsection","Active Learning"]],"subsections":[],"title":"Active Learning"},{"cite_extract_rate":0.4,"cites":[330,8163,3267,5113,328,8495],"content":"\\label{subsec:dataset_optimization:domains}\n\\sectionauthor{Julia Rosenzweig\\textsuperscript{1}}\nThe classical assumption in machine learning is that the training and testing data sets are drawn from the same distribution, implying that the model is deployed under the same conditions as it was trained under. However, as mention, in real-world applications this assumption is often violated in the sense that the training and the testing set stem from different domains having different distributions. This poses difficulties for statistical models and the performance will mostly degrade when they are deployed on a domain $D^{\\mathrm{test}}$, having a different distribution than the training dataset (\\ie generalizing from the training to the testing domain is not possible). This makes the study of domains not only relevant from the machine learning perspective, but also from a safety point of view.\n\\par\nMore formally, there are differing notions of a 'domain' in literature. For , a domain $\\mathcal{D} = \\{\\mathcal{X}, P(X) \\}$ consists of a feature space $\\mathcal{X} \\subset \\mathbb{R}^d$ together with a marginal probability distribution $P(X)$ with $X\\in \\mathcal{X}$. In , a domain is a pair consisting of a distribution over the inputs together with a labeling function.\nHowever, instead of a sharp labeling function, it is also widely accepted to define a (training) domain $\\mathcal{D} = \\{(x_i,y_i)\\}_{i=1}^n$ to consist of $n$ (labeled) samples that are sampled from a joint distribution $P(x,y)$ (\\cf ). \n\\par\nThe reasons for \\emph{distributional shift} are diverse---as are the names to indicate a shift. For example, if the rate of (class) images of interest is different between training and testing set this can lead to a domain gap and, \\eg result in differing overall error rates. Moreover, as mentions, changing weather conditions and camera setups in cars lead to a domain mismatch in applications of autonomous driving.\nIn biomedical image analysis, different imaging protocols and diverse anatomical structures can hinder generalization of trained models (\\cf ).\nCommon terms to indicate distributional shift are \\emph{domain shift, dataset shift, covariate shift, concept drift, domain divergence, data fracture, changing environments} or \\emph{dataset bias}. References provide an overview. \n\\par\nMethods and measures to overcome the problem of domain mismatch between one or more (\\cf ) source domains and target domain(s) and the resulting poor model performance are studied in the field of transfer learning and in particular its subtopic domain adaptation (\\cf ).\nFor instance, adapting a model that is trained on synthetically generated data to work on real data is one of the core challenges, as can be seen .\nFurthermore, detecting when samples are out-of-domain or out-of-distribution is an active field of research (\\cf and the outlier/anomaly detection in \\refsec{subsec:dataset_optimization:anomaly_detection} as well as the topic of observers in the black-box methods in \\refsec{subsec:verification:black_box_methods} for further reference). This is particularly relevant for machine learning models that operate in the real world: If, \\eg an autonomous vehicle encounters some situation that deviates strongly from what was seen during training (\\eg due to some special event like a biking competition, carnival, etc.) this can lead to wrong predictions and thereby potential safety issues if not detected in time.","id":"7b5fda9c-0631-4e3b-bf4f-fb2fb6b0d60d","level":"subsection","origin_cites_number":15,"parent_id":"1fe2887d-3162-404c-bb44-c07ac2772367","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Dataset Optimization"],["subsection","Domains"]],"subsections":[],"title":"Domains"},{"cite_extract_rate":0.7333333333333331,"cites":[8164,858,8165,5980,5552,90,2083,5548,7022,5553,62],"content":"\\label{subsec:dataset_optimization:augmentation}\n\\sectionauthor{Falk Kappel\\textsuperscript{13}}\nGiven the need for big amounts of data to train neural networks, one often runs into a situation where data is lacking. This can lead to insufficient generalization and an overfitting to the training data. An overview over different techniques to tackle this challenge can be found in . One approach to try and overcome this issue is the augmentation of data. It aims at optimizing available data and increasing its amount, curating a dataset that represents a wide variety of possible inputs during deployment. Augmentation can as well be of help when having to work with a heavily unbalanced dataset by creating more samples of underrepresented classes. A broad survey on data augmentation is provided by . They distinguish between two general approaches to data augmentation with the first one being data warping augmentations that focus on taking existing data and transforming it in a way that does not effect labels. The other option are oversampling augmentations, which create synthetic data that can be used to increase the size of the dataset.\\newline\nExamples of some of the most basic augmentations are flipping, cropping, rotating, translating, shearing and zooming. These are affecting the geometric properties of the image and are easily implemented . The machine learning toolkit \\texttt{Keras}, for example, provides an easy way of applying them to data using their \\texttt{ImageDataGenerator} class . Other simple methods include adaptations in color space that affect properties such as lighting, contrast and tints, which are common variations within image data. Filters can be used to control increased blur or sharpness . In random erasing is introduced as a method with similar effect as cropping, aiming at gaining robustness against occlusions. An example for mixing images together as an augmentation technique can be found in . \\newline\nThe abovementioned methods have in common that they work on the input data but there are different approaches that make use of deep learning for augmentation. An example for making augmentations in feature space using autoencoders can be found in . They use the representation generated by the encoder and generate new samples by interpolation and extrapolation between existing samples of a class. The lack of interpretability of augmentations in feature space in combination with the tendency to perform worse than augmentations in image space present open challenges for those types of augmentations . Adversarial training is another method that can be used for augmentation. The goal of adversarial training is to discover cases that would lead to wrong predictions. That means the augmented images won't necessarily represent samples that could occur during deployment but that can help in achieving more robust decision boundaries . An example of such an approach can be found in . Generative modelling can be used to generate synthetic samples that enlarge the dataset in a useful way with GANs, variational autoencoders and the combination of both are important tools in this area . Examples for data augmentation in medical context using a CycleGAN can be found in and using a progressively growing GAN in . Next to neural style transfer that can be used to change the style of an image to a target style, AutoAugment and population based augmentation are two more interesting publications. In the latter two, the idea is to search a predefined search space of augmentations to gather the best selection.","id":"de2e598e-f35b-4b65-945b-25aee539ea0d","level":"subsection","origin_cites_number":15,"parent_id":"1fe2887d-3162-404c-bb44-c07ac2772367","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Dataset Optimization"],["subsection","Augmentation"]],"subsections":[],"title":"Augmentation"},{"cite_extract_rate":0.555555555555555,"cites":[8133,8166,7698,1636,3500],"content":"\\label{subsec:dataset_optimization:corner_case_detection}\n\\sectionauthor{Alexander Pohl\\textsuperscript{16}, Marco Hoffmann\\textsuperscript{16}, Michael Mlynarski\\textsuperscript{16}, Timo Sämann\\textsuperscript{4}}\nEnsuring that AI-based applications behave correctly and predictably even in unexpected or rare situations is a major concern that gains importance especially in safety-critical applications such as autonomous driving. In the pursuit of more robust AI corner cases play an important role. \n\\par\nThe meaning of the term corner case varies in the literature. Some consider mere erroneous or incorrect behavior as corner cases . For example, in corner cases are referred to as situations in which an object detector fails to detect relevant objects at relevant locations. Others characterize corner cases mainly as rare combinations of input parameter values . This project adopts the first definition: Inputs that result in unexpected or incorrect behaviour of the AI function are defined as corner cases. \n\\par\nContingent on the hardware, the AI architecture and the training data, the search space of corner cases quickly becomes incomprehensibly large. While manual creation of corner cases (\\eg constructing or re-enacting scenarios) might be more controllable, approaches that scale better and allow for a broader and more systematic search for corner cases require extensive automation.\n\\par\nOne approach to automatic corner case detection is based on transforming the input data. The \\emph{DeepTest} framework uses three types of image transformations: linear, affine and convolutional transformations. In addition to these transformations, metamorphic relations help detect undesirable behaviors of deep learning systems. They allow changing the input while asserting some characteristics of the result . For example, changing the contrast of input frames should not affect the steering angle of a car . Input-output pairs that violate those metamorphic relations can be considered as corner cases.\n\\par\nAmong other things, the white-box testing framework \\emph{DeepXplore} applies a method called \\emph{gradient ascent} to find corner cases (\\cf \\refsec{subsec:verification:formal_testing}). In the experimental evaluation of the framework, three variants of deep learning architectures were used to classify the same input image. The input image was then changed according to the gradient ascent of an objective function that reflected the difference in the resulting class probabilities of the three model variants. When the changed (now artificial) input resulted in different class label predictions by the model variants, the input was considered as a corner case. \n\\par\nIn , corner cases are detected on video sequences by comparing predicted with actual frames. The detector has three components: The first component, semantic segmentation, is used to detect and locate objects in the input frame. As the second component, an image predictor trained on frame sequences predicts the actual frame based on the sequence preceding that frame. An error is determined by comparing the actual with the predicted (\\ie expected) frame, following the idea that only situations that are unexpected for AI-based perception functions may be potentially dangerous and therefore a corner case. Both the segmentation and the prediction error are then fed into the third component of the detector, which determines a corner case score that reflects the extent to which unexpected relevant objects are at relevant locations.\n\\par\nIn , a corner case detector based on simulations in a \\emph{Carla} environment is presented. In the simulated world, AI agents control the vehicles. During simulations, state information of both the environment and the AI agents are fed into the corner case detector. While the environment provides the real vehicle states, the AI agents provide estimated and perceived state information. Both sources are then compared to detect conflicts (\\eg collisions). These conflicts are recorded for analysis. \n\\par\nSeveral ways of automatically generating and detecting corner cases exist. However, corner case detection is a task with challenges of its own: Depending on the operational domain including its boundaries, the space of possible inputs can be very large. Also, some types of corner cases are specific to the AI architecture, \\eg the network type or the network layout used. Thus, corner case detection has to assume a holistic point of view on both model and input, adding further complexity and reducing transferability of previous insights. \n\\par\nAlthough it can be argued that rarity does not necessarily characterize corner cases, rare input data might have the potential of challenging the AI functionality (\\cf \\refsec{subsec:dataset_optimization:anomaly_detection}).\nAnother research direction could investigate whether structuring the input space in a way suitable for the AI functionality supports the detection of corner cases. Provided that the operational domain is conceptualized as an ontology, ontology-based testing may support automatic detection. A properly adapted generator may specifically select promising combinations of extreme parameter values and, thus, provide valuable input for synthetic test data generation.","id":"e4ebc416-15f6-4517-a761-acd4315ad09b","level":"subsection","origin_cites_number":9,"parent_id":"1fe2887d-3162-404c-bb44-c07ac2772367","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Dataset Optimization"],["subsection","Corner Case Detection"]],"subsections":[],"title":"Corner Case Detection"},{"cite_extract_rate":0.6666666666666661,"cites":[314,509,1606,8429,1621,1759,8167,8168,8169,206,520,4169],"content":"\\label{sec:robust_training}\n\\sectionauthor{Nikhil Kapoor\\textsuperscript{7}}\nRecent works~ have shown that state-of-the-art deep neural networks (DNNs) performing a wide variety of computer vision tasks such as image classification~, object detection~ and semantic segmentation~ are not robust to small changes in the input.\n\\par\nRobustness of neural networks is an active and open research field that can be considered highly relevant for achieving safety in autonomous driving. Currently, most of the research is directed towards either improving adversarial robustness~ (robustness against carefully designed perturbations that aim at causing misclassifications with high confidence), or improving corruption robustness~ (robustness against commonly occurring augmentations such as weather changes, addition of Gaussian noise, photometric changes, etc.). While adversarial robustness might be more of a security issue than a safety issue, corruption robustness, on the other hand, can be considered highly safety-relevant.\n\\par\nEquipped with these definitions, we broadly term \\emph{robust training} here as methods or mechanisms that aim at improving either adversarial or corruption robustness of a DNN, by incorporating modifications into the architecture or into the training mechanism itself.","id":"380e4e9a-2d10-49f6-add4-04990cddb902","level":"section","origin_cites_number":18,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Robust Training"]],"subsections":["62315f40-21bf-40c7-afd5-00e8d80aa4e1","cf7b5f9f-caf9-44e0-b472-2124535b4d7a","e4ab90c2-e9e5-49ce-a3ae-b1dfdb4472d4"],"title":"Robust Training"},{"cite_extract_rate":0.7142857142857141,"cites":[868,8171,8173,8172,8170],"content":"\\label{subsec:robust_training:hyperparameter_optimization}\n\\sectionauthor{Seyed Eghbal Ghobadi\\textsuperscript{8}, Patrick Feifel\\textsuperscript{8}}\nThe final performance of a neural network depends highly on the learning process. The process includes the actual optimization and may additionally introduce training methods such as dropout, regularization, or parametrization of a multi-task loss. \n\\par\nThese methods adapt their behavior for predefined parameters. Hence, their optimal configuration is a priori unknown. We refer to them as \\emph{hyperparameters}. Important hyperparameters comprise, for instance, the initial learning rate, steps for learning rate reduction, learning rate decay, momentum, batch size, dropout rate and number of iterations. Their configuration has to be determined according to the architecture and task of the CNN . The search of an optimal hyperparameter configuration is called hyperparameter optimization (HO).\n\\par\nHO is usually described as an optimization problem . Thereby, the combined configuration space is defined as \n$\\boldsymbol{\\Lambda} = \\lambda_1 \\times \\lambda_2 \\times \\cdots \\lambda_N$,\naccording to each domain $\\lambda_n$. Their individual spaces can be continuous, discrete, categorical or binary. \n\\par\nHence, we aim to find an optimal hyperparameter configuration $\\lambda^{\\star}$ by minimizing an objective function $\\mathcal{O}\\left ( \\right )$, which evaluates a trained model $\\mathcal{M}$ on the validation dataset $\\mathcal{D}^{\\textrm{val}}$ with the loss $\\mathcal{L}$:\n\\begin{equation}\n\\lambda^{\\star} = \n\\text{arg min}_{\\lambda \\in \\boldsymbol{\\Lambda}} \\; \n\\mathcal{O} \\left( \\mathcal{L}, \\mathcal{M}_{\\lambda}, \\mathcal{D}^{\\textrm{train}}, \\mathcal{D}^{\\textrm{val}} \\right)\n\\end{equation}\nThis problem statement is widely regarded in traditional machine learning and primarily based on Bayesian optimization (BO) in combination with Gaussian processes. However, a straightforward application to deep neural networks encounters problems due to a \\emph{lack of scalability, flexibility and robustness} , . \n\\par\nTo exploit the benefits of BO, many authors proposed different combinations with other approaches. \nHyperband in combination with BO (BOHB) frames the optimization as ``... a pure exploration non-stochastic infinite-armed bandit problem ...''. \nThe method of BO for iterative learning (BOIL) internalizes iteratively collected information about the learning curve and the learning algorithm itself. \nThe authors of introduce the trace-aware knowledge gradient (taKG) as an acquisition function for BO (BO-taKG) which ``leverages both trace information and multiple fidelity controls''.\nThereby BOIL and BO-taKG achieve state-of-research performance regarding CNNs outperforming Hyperband.\n\\par\nOther approaches such as the orthogonal array tuning method (OATM) or HO by reinforcement learning (Hyp-RL) turn away from the Bayesian approaches and offer new research directions.\n\\par\nFinally, the insight that many authors include kernel sizes and number of kernels and layers in their hyperparameter configuration should be emphasized. More work should be spent on the distinct integration of HO in the performance estimation strategy of neural architecture search (\\cf \\refsec{subsec:architecture:neural_architecture_search}).","id":"62315f40-21bf-40c7-afd5-00e8d80aa4e1","level":"subsection","origin_cites_number":7,"parent_id":"380e4e9a-2d10-49f6-add4-04990cddb902","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Robust Training"],["subsection","Hyperparameter Optimization"]],"subsections":[],"title":"Hyperparameter Optimization"},{"cite_extract_rate":0.42857142857142805,"cites":[8179,8177,958,8176,8174,8167,8175,8180,8178],"content":"\\label{subsec:robust_training:modification_of_loss}\n\\sectionauthor{Nikhil Kapoor\\textsuperscript{7}}\nThere exist many approaches that aim at directly modifying the loss function with an objective of improving either adversarial or corruption robustness~. One of the earliest approaches for improving corruption robustness was introduced by Zheng \\etal called \\emph{stability training}, where they introduce a regularization term that penalizes the network prediction to a clean and an augmented image. However, their approach does not scale to many augmentations at the same time. Janocha \\etal then introduced a detailed analysis on the influence of multiple loss functions to model performance as well as robustness and suggested that expectation-based losses tend to work better with noisy data and squared-hinge losses tend to work better for clean data. Other well-known approaches are mainly based on variations of data augmentation~, which can be computationally quite expensive. \n\\par\nIn contrast to corruption robustness, there exist many more approaches based on adversarial examples. We highlight some of the most interesting and relevant ones here. Mustafa \\etal proposes to add a loss term that maximally separates class-wise feature map representations, hence increasing the distance from data points to the corresponding decision boundaries. Similarly, Pang \\etal proposed the Max-Mahalanobis center (MMC) loss to learn more structured representations and induce high-density regions in the feature space. Chen \\etal proposed a variation of the well-known cross entropy (CE) loss that not only maximizes the model probabilities of the correct class, but in addition, also minimizes model probabilities of incorrect classes. Cisse \\etal constraints the Lipschitz constant of different layers to be less than one which restricts the error propagation introduced by adversarial perturbations to a DNN. Dezfooli \\etal proposed to minimize the curvature of the loss surface locally around data points. They emphasize that there exists a strong correlation between locally small curvature and correspondingly high adversarial robustness. \n\\par\nAll of these methods highlighted above are evaluated mostly for image classification tasks on smaller datasets, namely CIFAR-10~, CIFAR-100~, SVHN~, and only sometimes on ImageNet~. Very few approaches have been tested rigorously on complex safety-relevant tasks such as \\emph{object detection} and \\emph{semantic segmentation}, etc. Moreover, methods that improve adversarial robustness are only tested on a small subset of attack types under differing attack specifications. This makes comparing multiple methods difficult. \n\\par\nIn addition, methods that improve corruption robustness are evaluated over a standard data set of various corruption types which may or may not be relevant to its application domain. In order to assess multiple methods for their effect on safety-related aspects, a thorough robustness evaluation methodology is needed, which is largely missing in the current literature. This evaluation would need to take into account relevant disturbances/corruption types present in the real world (application domain) and had to assess robustness towards such changes in a rigorous manner. Without such an evaluation, we run the risk of being overconfident in our network, thereby harming safety.","id":"cf7b5f9f-caf9-44e0-b472-2124535b4d7a","level":"subsection","origin_cites_number":21,"parent_id":"380e4e9a-2d10-49f6-add4-04990cddb902","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Robust Training"],["subsection","Modification of Loss"]],"subsections":[],"title":"Modification of Loss"},{"cite_extract_rate":0.5666666666666661,"cites":[2800,5073,7109,3624,2690,8589,2761,2686,8181,8182,3655,8588,2705,1695,1256,8183,2793],"content":"\\sectionauthor{Firas Mualla\\textsuperscript{13}}\nDomain generalization (DG) can be seen as an extreme case of \\emph{domain adaptation} (DA). The latter is a type of transfer learning, where the source and target tasks are the same (\\eg shared class labels) but the source and target domains are different (\\eg another image acquisition protocol or a different background) . \nThe DA can be either supervised (SDA), where there is little available labeled data in the target domain, or unsupervised (UDA), where data in the target domain is not labeled. The DG goes one step further by assuming that the target domain is entirely unknown. Thus, it seeks to solve the train-test domain shift in general. While DA is already an established line of research in the machine learning community, DG is relatively new , though with an extensive list of papers in the last few years. \n\\par\nProbably, the first intuitive solution that one may think of to implement DG is neutralizing the domain-specific features. It was shown in that the gray-level co-occurrence matrices (GLCM) tend to perform poorly in semantic classification (\\eg digit recognition) but yield good accuracy in textural classification compared to other feature sets such as SURF and LBP. DG was thus implemented by decorrelating the model's decision from the GLCM features of the input image even without the need of domain labels.\nBesides the aforementioned intensity-based statistics of an input image, it is known that characterizing image style can be done based on the correlations between the filter responses of a DNN layer (neural style transfer). In , the training images are enriched with stylized versions, where a style is defined either by an external style (\\eg cartoon or art) or by an image from another domain. Here, DG is addressed as a \\emph{data augmentation} problem.\nSome approaches try to learn generalizable latent representations by a kind of adversarial training. This is done by a generator or an encoder, which is trained to generate a hidden feature space that maximizes the error of a domain discriminator but at the same time minimizes the classification error of the task of concern. Another flavor of adversarial training can be seen in , where an adversarial autoencoder is trained to generate features, which a discriminator cannot distinguish from random samples drawn from a prior Laplace distribution. This regularization prevents the hidden space from overfitting to the source domains, in a similar spirit to how variational autoencoders do not leave gaps in the latent space. In , it is argued that the domain labels needed in such approaches are not always well-defined or easily available. Therefore they assume unknown latent domains which are learned by clustering in a space similar to the style-transfer features mentioned above. The pseudo labels resulting from clustering are then used in the adversarial training.\nAutoencoders have been employed for DG not only in an adversarial setup, but also in the sense of \\emph{multi-task learning} nets , where the classification task in such nets is replaced by a reconstruction one. In , an autoencoder is trained to reconstruct not only the input image but also the corresponding images in the other domains.\nIn the core of both DA and DG we are confronted with a distribution matching problem. However, estimating the probability density in high-dimensional spaces is intractable. The density-based metrics such as Kullback-Leibler divergence are thus not directly applicable. In statistics, the so-called \\emph{two-samples tests} are usually employed to measure the distance between two distributions in a point-wise manner, \\ie without density estimation. For deep learning applications, these metrics need not only to be point-wise but also differentiable. The two-samples tests were approached in the machine learning literature using (differentiable) K-NNs , classifier two-samples tests (C2ST) , or based on the theory of kernel methods . More specifically, the \\emph{maximum mean discrepancy} (MMD) , which belongs to the kernel methods, is widely used for DA but also for DG . Using the MMD, the distance between two samples is estimated based on pairwise kernel evaluations, \\eg the radial basis function (RBF) kernel.\nWhile the DG approaches generalize to domains from which zero shots are available, the so-called \\emph{zero shot learning} (ZSL) approaches generalize to tasks (\\eg new classes in the same source domains) for which zero shots are available. Typically, the input in ZSL is mapped to a semantic vector per class instead of a simple class label. This can be, for instance, a vector of visual attributes or a word embedding of the class name . A task (with zero shots at training time) can be then described by a vector in this space. In , there is an attempt to combine ZSL and DG in the same framework in order to generalize to new domains as well as new tasks, which is also referred to as\n \\emph{heterogeneous domain generalization}.\nNote that most discussed approaches for DG require non-standard handling, \\ie modifications to models, data, and/or the optimization procedure. This issue poses a serious challenge as it limits the practical applicability of these approaches. There is a line of research which tries to address this point by linking DG to other machine learning paradigms, especially the model-agnostic meta-learning (MAML) algorithm, in an attempt to apply DG in a model-agnostic way. Loosely speaking, a model can be exposed to simulated train-test domain shift by training on a small \\emph{support set} to minimize the classification error on a small \\emph{validation set}. This can be seen as an instance of a \\emph{few shot learning} (FSL) problem . Moreover, the procedure can be repeated on other (but related) FSL tasks (\\eg different classes) in what is known as \\emph{episodic training}. The model transfers its knowledge from one task to another task and learns how to learn fast for new tasks. This can be thus seen as a \\emph{meta-learning} objective (in a FSL setup). Since the goal of DG is to adapt to new domains rather than new tasks, several model-agnostic approaches try to recast this procedure in a DG setup.","id":"e4ab90c2-e9e5-49ce-a3ae-b1dfdb4472d4","level":"subsection","origin_cites_number":30,"parent_id":"380e4e9a-2d10-49f6-add4-04990cddb902","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Robust Training"],["subsection","Domain Generalization"]],"subsections":[],"title":"Domain Generalization"},{"cite_extract_rate":0.7272727272727271,"cites":[314,900,888,6146,8396,892,917,906],"content":"\\label{sec:adversarial_attacks}\n\\sectionauthor{Andreas Bär\\textsuperscript{15}}\nOver the last few years, deep neural networks (DNNs) consistently showed state-of-the-art performance across several vision-related tasks.\nWhile their superior performance on clean data is indisputable, they show a lack of robustness to certain input patterns, denoted as \\emph{adversarial examples} .\nIn general, an algorithm for creating adversarial examples is referred to as an \\emph{adversarial attack} and aims at fooling an underlying DNN, such that the output changes in a desired and malicious way.\nThis can be carried out without any knowledge about the DNN to be attacked (black-box attack) , or with full knowledge about the parameters, architecture, or even training data of the respective DNN (white-box attack) .\nWhile initially being applied on simple classification tasks, some approaches aim at finding more realistic attacks , which particularly pose a threat to safety-critical applications, such as DNN-based environment perception systems in autonomous vehicles.\nAltogether, this motivated the research in finding ways of defending against such adversarial attacks .\nIn this section, we introduce the current state of research regarding adversarial attacks in general, more realistic adversarial attacks closely related to the task of environment perception for autonomous driving, and strategies for detecting or defending adversarial attacks.\nWe conclude each section by clarifying current challenges and research directions.","id":"4396c1c3-a3ad-4efa-9f10-99f90060c0b4","level":"section","origin_cites_number":11,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Adversarial Attacks"]],"subsections":["945be913-0c94-4b59-92f5-5fa8a6f0bf30","7e154b91-68e0-47d3-8c11-92f8eb25b769"],"title":"Adversarial Attacks"},{"cite_extract_rate":0.6341463414634141,"cites":[923,912,8184,892,917,908,8185,890,914,7313,975,314,953,954,1624,920,8186,6140,919,899,902,3267,8396,3234,906,6148],"content":"\\label{subsec:adverarial_attacks:adversarial_attacks_defenses}\n\\sectionauthor{Andreas Bär\\textsuperscript{15}, Seyed Eghbal Ghobadi\\textsuperscript{8}, Ahmed Hammam\\textsuperscript{8}}\nThe term \\emph{adversarial example} was first introduced by Szegedy \\etal .\nFrom there on, many researchers tried to find new ways of crafting adversarial examples more effectively.\nHere, the fast gradient sign method (FGSM) , DeepFool , least-likely class method (LLCM) , C\\&W , momentum iterative fast gradient sign method (MI-FGSM) , and projected gradient descent (PGD) are a few of the most famous attacks so far.\nIn general, these attacks can be executed in an iterative fashion, where the underlying adversarial perturbation is usually bounded by some norm and is following additional optimization criteria, \\eg minimizing the number of changed pixels.\n\\par\nThe mentioned attacks can be further categorized as image-specific attacks, where for each image a new perturbation needs to be computed.\nOn the other hand, image-agnostic attacks aim at finding a perturbation, which is able to fool an underlying DNN on a set of images.\nSuch a perturbation is also referred to as a \\emph{universal adversarial perturbation} (UAP).\nHere, the respective algorithm UAP , fast feature fool (FFF) , and prior driven uncertainty approximation (PD-UA) are a few honorable mentions.\nAlthough the creation process of a universal adversarial perturbation typically relies on a white-box setting, they show a high \\emph{transferability} across models .\nThis allows black-box attacks, where one model is used to create a universal adversarial perturbation, and another model is being attacked with the beforehand-created perturbation.\nAnother way of designing black-box attacks is to create a surrogate DNN, which mimics the respective DNN to be attacked and thus can be used in the process of adversarial example creation .\nOn the contrary, some research has been done to create completely incoherent images (based on evolutionary algorithms or gradient ascent) to fool an underlying DNN .\nDifferent from that, another line of work has been proposed to alter only some pixels in images to attack a respective model.\nHere and have used optimization approaches to perturb some pixels in images to produce targeted attacks, aiming at a specific class output, or non-targeted attacks, aiming at outputting a class different from the network output or the ground truth.\nThis can be extended up to finding one pixel in the image to be exclusively perturbed to generate adversarial images .\nThe authors of proposed to train generative models to generate adversarial examples.\nGiven an input image and the target label, a generative model is trained to produce adversarial examples for DNNs.\nHowever, while the produced adversarial examples look rather unrealistic to a human, they are able to completely deceive a DNN.\n\\par\nThe existence of adversarial examples not only motivated research in finding new attacks, but also in finding \\emph{defense strategies} to effectively defend these attacks.\nEspecially for safety-critical applications, such as DNN-based environment perception for autonomous driving, the existence of adversarial examples needs to be handled accordingly.\nSimilar to adversarial attacks, one can categorize defense strategies into two types: \\emph{model-specific} defense strategies and \\emph{model-agnostic} defense strategies.\nThe former refers to defense strategies, where the model of interest is modified in certain ways.\nThe modification can be done on the architecture, training procedure, training data, or model weights.\nOn the other hand, model-agnostic defense strategies consider the model to be a black box.\nHere, only the input or the output is accessible.\nSome well-known model-specific defense strategies include adversarial training , the inclusion of robustness-oriented loss functions during training , removing adversarial patterns in features by denoising layers , and redundant teacher-student frameworks .\nThe majority of model-agnostic defense strategies primarily focuses on various kinds of (gradient masking) pre-processing strategies .\nThe idea is to remove the adversary from the respective image, such that the image is transformed from the adversarial space back into the clean space.\n\\par\nNonetheless, Athalye \\etal showed that gradient masking alone is not a sufficient criterion for a reliable defense strategy.\nIn addition, detection and out-of-distribution techniques have also been proposed as model-agnostic defense strategies against adversarial attacks.\nHere, the Mahalanobis distance or area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPR) are used to detect adversarial examples.\nThe authors of on the other hand proposed to train networks to detect, whether the input image is out-of-distribution or not. \n\\par\nMoreover, Feinman \\etal proved that adversarial attacks usually produce high uncertainty on the output of the DNN.\nAs a consequence, they proposed to use the dropout technique to estimate uncertainty on the output to identify a possible adversarial attack.\n\\par\nRegarding adversarial attacks, the majority of the listed attacks are designed for image classification.\nOnly a few adversarial attacks consider tasks that are closely related to autonomous driving, such as bounding box detection, semantic segmentation, instance segmentation, or even panoptic segmentation.\nAlso, the majority of the adversarial attacks rely on a white-box setting, which is usually not present for a potential attacker.\nEspecially universal adversarial perturbations have to be considered as a real threat due to their high model transferability.\nGenerally speaking, the existence of adversarial examples has not been thoroughly studied yet.\nAn analytical interpretation is still missing, but could help in designing more mature defense strategies.\n\\par\nRegarding defense strategies, adversarial training is still considered as one of the most effective ways of increasing the robustness of a DNN.\nNonetheless, while adversarial training is indeed effective, it is rather inefficient in terms of training time.\nIn addition, model-agnostic defenses should be favored as once being designed, they can be easily transferred to different models.\nMoreover, as most model-agnostic defense strategies rely on gradient-masking and it has been shown that gradient-masking is not a sufficient property for a defense strategy, new ways of designing model-agnostic defenses should be taken into account.\nFurthermore, out-of-distribution and adversarial attacks detection or even correction methods have been a new trend for identifying attacks.\nHowever, as the environment perception system of an autonomous driving vehicle could rely on various information sources, including LiDAR, optical flow, or depth from a stereo camera, techniques of information fusion should be further investigated to mitigate or even eliminate the effect of adversarial examples.","id":"945be913-0c94-4b59-92f5-5fa8a6f0bf30","level":"subsection","origin_cites_number":41,"parent_id":"4396c1c3-a3ad-4efa-9f10-99f90060c0b4","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Adversarial Attacks"],["subsection","Adversarial Attacks and Defenses"]],"subsections":[],"title":"Adversarial Attacks and Defenses"},{"cite_extract_rate":0.684210526315789,"cites":[900,7698,8187,917,890,8189,904,901,8188,8190,899,7308,6014],"content":"\\label{subsec:adverarial_attacks:more_realistic_attacks}\n\\sectionauthor{Svetlana Pavlitskaya\\textsuperscript{14}}\nWe consider the following two categories of realistic adversarial attacks: (1) image-level attacks, which not only fool a neural network but also pose a provable threat to autonomous vehicles, and (2) attacks which have been applied in a real world or in a simulation environment, such as car learning to act (CARLA) .\n\\par\nSome notable examples in the first category of attacks include attacks on semantic segmentation or person detection .\n\\par\nIn the second group of approaches, the attacks are specifically designed to survive real world distortions, including different distances, weather and lighting conditions, as well as camera angles. For this, adversarial perturbations are usually concentrated in a specific image area, called \\emph{adversarial patch}. Crafting an adversarial patch involves specifying a patch region in each training image, applying transformations to the patch, and iteratively changing the pixel values within this region to maximize the network prediction error. The latter step typically relies on an algorithm, proposed for standard adversarial attacks, which aim at crafting invisible perturbations while misleading neural networks, \\eg C\\&W , Jacobian-based saliency map attack (JSMA) , and PGD .\n\\par\nThe first printable adversarial patch for image classification was described by Brown \\etal . Expectation over transformations (EOT) is one of the influential updates to the original algorithm---it permits to robustify patch-based attacks to distortions and affine transformations. Localized and visible adversarial noise (LaVAN) is a further method to generate much smaller patches (up to 2\\% of the pixels in the image). In general, fooling image classification with a patch is a comparatively simple task, because adversarial noise can mimic an instance of another class and thus lower the prediction probability for a true class.\n\\par\nRecently, patch-based attacks for a more tricky task of object detection have been described . Also, Lee and Kolter generate a patch using PGD , followed by EOT applied to the patch. With this approach, all detections in an image can be successfully suppressed, even without any overlap of a patch with bounding boxes. Furthermore, several approaches for generating an adversarial T-shirt have been proposed, including .\n\\par\nDeepBillboard is the first attempt to attack end-to-end driving models with adversarial patches. The authors propose to generate a single patch for a sequence of input images to mislead four steering models, including DAVE-2 in a drive-by scenario. \n\\par\nApart from physical feasibility, inconspicuousness is crucial for a realistic attack. Whereas adversarial patches usually look like regions of noise, several works have explored attacks with an inconspicuous patch. In particular, Eykholt \\etal demonstrate the vulnerability of road sign classification to the adversarial perturbations in the form of only black and white stickers. In , an end-to-end driving model is attacked in CARLA by painting of black lines on the road. Also, Kong and Liu use a generative adversarial network to get a realistic billboard to attack an end-to-end driving model in a drive-by scenario. In , a method to hide visible adversarial perturbations with customized styles is proposed, which leads to adversarial traffic signs that look unsuspicious to a human.\n\\par\nCurrent research mostly focuses on attacking image-based perception of an autonomous vehicle. Adversarial vulnerability of further components of an autonomous vehicle, \\eg LiDAR-based perception, optical flow and depth estimation, has only recently gained attention. Furthermore, most attacks consider only a single component of an autonomous driving pipeline, the question whether the existing attacks are able to propagate to further pipeline stages has not been studied yet. The first work in this direction describes an attack on object detection and tracking. The evaluation is, however, limited to a few clips, where no experiments in the real world have been performed. Overall, the research on realistic adversarial attacks, especially combined with physical tests, is currently in the starting phase.","id":"7e154b91-68e0-47d3-8c11-92f8eb25b769","level":"subsection","origin_cites_number":19,"parent_id":"4396c1c3-a3ad-4efa-9f10-99f90060c0b4","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Adversarial Attacks"],["subsection","More Realistic Attacks"]],"subsections":[],"title":"More Realistic Attacks"},{"cite_extract_rate":0.33333333333333304,"cites":[7514],"content":"\\label{sec:interpretability}\n\\sectionauthor{Felix Brockherde\\textsuperscript{10}}\nNeural networks are, by their nature, black boxes and therefore intrinsically hard to interpret . Due to their unrivaled performance, they still remain first choice for advanced systems even in many safety-critical areas, such as level 4 automated driving. This is why the research community has invested considerable effort to unhinge the black-box character and make deep neural networks more transparent.\nWe can observe three strategies that provide different view points towards this goal in the state of the art. First is the most direct approach of opening up the black box and looking at intermediate representations (\\refsec{subsec:interpretability:intermediate_representations}). Being able to interpret individual layers of the system facilitates interpretation of the whole. The second approach tries to provide interpretability by explaining the network's decisions with pixel attributions (\\refsec{subsec:interpretability:pixel_attribution}). Aggregated explanations of decision can then lead to interpretability of the system itself. Third is the idea of approximating the network with interpretable proxies to benefit from the deep neural networks performance while allowing interpretation via surrogate models (\\refsec{subsec:interpretability:interpretable_proxies}). Underlying all aspects here is the area of visual analytics (\\refsec{subsec:interpretability:visual_analytics}).\nThere exists earlier research in the medical domain to help human experts understand and convince them of machine learning decisions . Legal requirements in the finance industry gave rise to interpretable systems that can justify their decisions. An additional driver for interpretability research was the concern for Clever Hans predictors .","id":"e47014ad-1aec-4b32-9241-6aab90a7f0dd","level":"section","origin_cites_number":3,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Interpretability"]],"subsections":["cf04c28f-300a-4cf5-bc98-3f4d9734078e","a3e2cd79-8634-4ce7-9519-acd205ba8d1d","110c6df1-6598-4052-8ee0-6595325dfebc","37f2baca-d82b-4285-9422-b80c37554845"],"title":"Interpretability"},{"cite_extract_rate":0.17647058823529402,"cites":[8192,8191,8193],"content":"\\label{subsec:interpretability:visual_analytics}\n\\sectionauthor{Elena Schulz\\textsuperscript{1}}\nTraditional data science has developed a huge tool set of automated analysis processes conducted by computers, which are applied to problems that are well-defined in the sense that the dimensionality of input and output as well as the size of the data set they rely on is manageable. \nFor those problems that in comparison are more complex, the automation of the analysis process is limited and/or might not lead to the desired outcome. \nThis is especially the case with unstructured data like image or video data in which the underlying information cannot directly be expressed by numbers. \nRather, it needs to be transformed to some structured form to enable computers to perform some task of analysis. \nAdditionally, with an ever increasing amount of various types of data being collected, this ``information overload'' cannot solely be analyzed by automatic methods . \n\\par\n\\emph{Visual analytics} addresses this challenge as ``the science of analytical reasoning facilitated by interactive visual interfaces'' . \nVisual analytics therefore does not only focus on either computationally processing data or visualizing results but coupling both tightly with interactive techniques. \nThus, it enables an integration of the human expert into the iterative visual analytics process: Through visual understanding and human reasoning, the knowledge of the human expert can be incorporated to effectively refine the analysis. This is of particular importance, where a stringent safety argumentation for complex models is required. With the help of visual analytics, the line of argumentation can be built upon arguments that are understandable for humans.\nTo include the human analyst efficiently into this process, a possible guideline is the \\emph{visual analytics mantra} by Keim: ``Analyze first, show the important, zoom, filter and analyze further, details on demand'' \\footnote{Extending the original \\emph{visualization mantra} by Shneiderman ``Overview first, filter and zoom, details on demand''.}.\n\\par\nThe core concepts of visual analytics therefore rely on well-designed interactive visualizations, which support the analyst in the tasks of, \\eg reviewing, understanding, comparing and inferring not only the initial phenomenon or data but also the computational model and its results itself with the goal of enhancing the analytical process. \n\\par\nDriven by various fields of application, visual analytics is a multidisciplinary field with a wide variety of task-oriented development and research. \nAs follows, recent work has been done in several areas: \ndepending on the task, there exist different pipeline approaches to create whole \\emph{visual analytics systems} ; \nthe injection of human expert knowledge into the process of determining trends and patterns from data is the focus of \\emph{predictive visual analytics} ; \nenabling the human to explore \\emph{high-dimensional data} interactively and visually (\\eg via dimensionality reduction ) is a major technique to enhance the understandability of complex models (\\eg neural networks); \nthe iterative improvement and the understanding of machine learning models is addressed by using interactive visualizations in the field of \\emph{general machine learning} or the other way round: using machine learning to improve visualizations and guidance based on user interactions .\nEven more focused on the loop of simultaneously developing and refining machine learning models is the area of \\emph{interactive machine learning}, where the topics of \\emph{interface design} and the \\emph{importance of users} are discussed. One of the current research directions is using visual analytics in the area of \\emph{deep learning} . However, due to the interdisciplinarity of visual analytics, there are still open directions and ongoing research opportunities. \n\\par\nEspecially in the domain of neural networks and deep learning, visual analytics is a relatively new approach in tackling the challenge of \\emph{explainability} and \\emph{interpretability} of those often called \\emph{black boxes}. \nTo enable the human to better interact with the models, research is done in enhancing the \\emph{understandability} of complex deep learning models and their outputs with the use of proper visualizations. \nOther research directions attempt to achieve improving the \\emph{trustability} of the models, giving the opportunity to inspect, diagnose and refine the model. \nFurther, possible areas for research are \\emph{online training processes} and the development of \\emph{interactive systems} covering the whole process of training, enhancing and monitoring machine learning models. \nHere, the approach of \\emph{mixed guidance}, where system-initiated guidance is combined with user-initiated guidance, is discussed among the visual analytics community as well. \nAnother challenge and open question is creating ways of \\emph{comparing models} to examine which model yields a better performance, given specific situations and selecting or combining the best models with the goal of increasing performance and overall safety.","id":"cf04c28f-300a-4cf5-bc98-3f4d9734078e","level":"subsection","origin_cites_number":17,"parent_id":"e47014ad-1aec-4b32-9241-6aab90a7f0dd","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Interpretability"],["subsection","Visual Analytics"]],"subsections":[],"title":"Visual Analytics"},{"cite_extract_rate":0.514285714285714,"cites":[1002,8198,6410,8201,8197,8200,8194,8199,5680,3314,4423,8202,2709,8196,306,318,8195,1256],"content":"\\label{subsec:interpretability:intermediate_representations}\n\\sectionauthor{Felix Hauser\\textsuperscript{11}, Jan Kronenberger\\textsuperscript{9}, Seyed Eghbal Ghobadi\\textsuperscript{8}}\nIn general, representation learning aims to extract lower dimensional features in latent space from higher dimensional inputs.\nThese features are then used as an effective representation for regression, classification, object detection and other machine learning tasks.\nPreferably, latent features should be disentangled, meaning that they represent separate factors found in the data that are statistically independent.\nDue to their importance in machine learning, finding meaningful intermediate representations has long been a primary research goal.\nDisentangled representations can be interpreted more easily by humans and \ncan for example be used to explain the reasoning of neural networks .\nAmong the longer known methods for extracting disentangled representations are principal component analysis (PCA) , independent component analysis , and nonnegative matrix factorization .\nPCA is highly sensitive to outliers and noise in the data.\nTherefore, more robust algorithms were proposed.\nIn already a small neural network was used as an encoder and the algorithm proposed in can deal with high-dimensional data.\nSome robust PCA algorithms are provided with analytical performance guarantees .\nA popular method for representation learning with deep networks is the variational autoencoder (VAE) .\nAn important generalization of the method is the $ \\beta $-VAE variant , which improved the disentanglement capability .\nLater analysis added to the theoretical understanding of $ \\beta $-VAE\n.\nCompared to standard autoencoders, VAEs map inputs to a distribution, instead of mapping them to a fixed vector.\nThis allows for additional regularization of the training to avoid overfitting and ensure good representations.\nIn $ \\beta $-VAEs the trade-off between reconstruction quality and disentanglement can be fine-tuned by the hyperparameter $ \\beta $.\nDifferent regularization schemes have been suggested to improve the VAE method.\nAmong them are Wasserstein autoencoders , attribute regularization and relational regularization .\nRecently, a connection between VAEs and nonlinear independent component analysis was established and then expanded .\nBesides VAEs, deep generative adversarial networks can be used to construct latent features .\nOther works suggest centroid encoders or conditional learning of Gaussian distributions as alternatives to VAEs.\nIn concept activation vectors are defined as being orthogonal to the decision boundary of a classifier.\nApart from deep learning, entirely new architectures, such as capsule networks , might be used to disassemble inputs. \n\\par\nWhile many different approaches for disentangling exist, the feasibility of the task is not clear yet and a better theoretical understanding is needed.\nThe disentangling performance is hard to quantify, which is only feasible with information about the latent ground truth .\nModels that overly rely on single directions, single neurons in fully connected networks or single feature maps in CNNs, have the tendency to overfit .\nAccording to , unsupervised learning does not produce good disentangling and even small latent spaces do not reduce the sample complexity for simple tasks.\nThis is in direct contrast to newer findings that show a decreased sample complexity for more complex visual downstream tasks .\nSo far, it is unclear if disentangling improves the performance of machine learning tasks.\nIn order to be interpretable, latent disentangled representations need to be aligned with human understandable concepts.\nIn training with adversarial examples was used and the learned representations were shown to be more aligned with human perception.\nFor explainable AI, disentangling alone might not be enough to generate interpretable output and additional regularization could be needed.","id":"a3e2cd79-8634-4ce7-9519-acd205ba8d1d","level":"subsection","origin_cites_number":35,"parent_id":"e47014ad-1aec-4b32-9241-6aab90a7f0dd","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Interpretability"],["subsection","Intermediate Representations"]],"subsections":[],"title":"Intermediate Representations"},{"cite_extract_rate":0.6875,"cites":[4875,499,8600,1812,1822,1824,1814,7517,1617,6080,8203],"content":"\\label{subsec:interpretability:pixel_attribution}\n\\sectionauthor{Stephanie Abrecht\\textsuperscript{2}, Felix Brockherde\\textsuperscript{10}, Toshika Srivastava\\textsuperscript{12}}\nThe non-linearity and complexity of DNNs allow them to solve perception problems, like detecting a pedestrian, that cannot be specified in detail. At the same time, the automatic extraction of features given in an input image and the mapping to the respective prediction is counterintuitive and incomprehensible for humans, which makes it hard to argue safety for a neural network-based perception task. Feature importance techniques are currently predominantly used to diagnose the causes of incorrect model behaviors . So-called \\emph{attribution maps} are a visual technique to express the relationship between relevant pixels in the input image and the network's prediction. Regions in an image that contain relevant features are highlighted accordingly. \nAttribution approaches mostly map to one of three categories.\n\\par\nGradient-based and activation-based approaches (such as amongst others) rely on the gradient of the prediction with respect to the input. Regions that were most relevant for the prediction are highlighted. Activation-based approaches relate the feature maps of the last convolutional layer to output classes. \n\\par\nPerturbation-based approaches suggest manipulating the input. If the prediction changes significantly, the input may hold a possible explanation at least.\n\\par\nWhile gradient-based approaches are oftentimes faster in computation, perturbation-based approaches are much easier to interpret.\n\\par\nAs many studies have shown , there is still a lot of research to be done before attribution methods are able to robustly provide explanations for model predictions, in particular erroneous behavior. One key difficulty is the lack of an agreed-upon definition of a good attribution map including important properties. Even between humans, it is hard to agree on what a good explanation is due to its subjective nature. This lack of ground truth makes it hard or even impossible to quantitatively evaluate an explanation method. Instead, this evaluation is done only implicitly. \nOne typical way to do this is the axiomatic approach. Here a set of desiderata of an attribution method are defined, on which different attribution methods are then evaluated. Alternatively, different attribution methods may be compared by perturbing the input features starting with the ones deemed most important and measuring the drop in accuracy of the perturbed models. The best method will result into the greatest overall loss in accuracy as the number of inputs are omitted .\nMoreover, for gradient-based methods it is hard to assess if an unexpected attribution is caused by a poorly performing network or a poorly performing attribution method . How to cope with negative evidence, \\ie the object was predicted because a contrary clue in the input image was missing, is an open research question. Additionally, most methods were shown on classification tasks until now. It remains to be seen how they can be transferred to object detection and semantic segmentation tasks. In the case of perturbation-based methods, the high computation time and single-image analysis inhibit wide-spread application.","id":"110c6df1-6598-4052-8ee0-6595325dfebc","level":"subsection","origin_cites_number":16,"parent_id":"e47014ad-1aec-4b32-9241-6aab90a7f0dd","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Interpretability"],["subsection","Pixel Attribution"]],"subsections":[],"title":"Pixel Attribution"},{"cite_extract_rate":0.33333333333333304,"cites":[8502,7507,1821,8204],"content":"\\label{subsec:interpretability:interpretable_proxies}\n\\sectionauthor{Gesina Schwalbe\\textsuperscript{3}}\nNeural networks are capable of capturing complicated logical\n(cor)relations. However, this knowledge is encoded on a\n\\emph{sub-symbolic} level in the form of learned weights and biases,\nmeaning that the reasoning behind the processing chain cannot be\ndirectly read out or interpreted by humans .\nTo explain the sub-symbolic processing, one can either use attribution\nmethods (\\cf \\refsec{subsec:interpretability:pixel_attribution}), or lift\nthis sub-symbolic representation to a \\emph{symbolic} one\n, meaning a more interpretable one.\nInterpretable proxies or surrogate models try to achieve the latter:\nThe DNN behavior is approximated by a model that\nuses symbolic knowledge representations.\nSymbolic representations can be\nlinear models like LIME (proportionality),\ndecision trees (if-then-chains) ,\nor loose sets of logical rules.\nLogical connectors can simply be AND and OR but also more general\nones like at-least-M-of-N .\nThe \\emph{expressiveness} of an approach refers to the logic that is\nused: Boolean-only versus first-order logic, and binary versus fuzzy\nlogic truth values .\nOther than attribution methods (\\cf \\refsec{subsec:interpretability:pixel_attribution}), these\nrepresentations can capture combinations of features and (spatial)\nrelations of objects and attributes. As an example consider\n\\enquote{eyes are closed} as explanation for \\enquote{person asleep}:\nAttribution methods only could mark the location of the eyes, dismissing\nthe relations of the attributes .\nAll mentioned surrogate model types (linear, set of rules) require\ninterpretable input features in order to be interpretable themselves.\nThese features must either be directly obtained from the DNN input or\n(intermediate) output, or automatically be extracted from the DNN\nrepresentation.\nExamples for extraction are the super-pixeling used in LIME for input\nfeature detection, or concept activation vectors\n for DNN representation decoding.\n\\par\nQuality criteria and goals for interpretable proxies are :\n\\emph{accuracy} of the standalone surrogate model on unseen examples,\n\\emph{fidelity} of the approximation by the proxy,\n\\emph{consistency} with respect to different training sessions, and\n\\emph{comprehensibility} measured by the complexity of the rule set\n(number of rules, number of hierarchical dependencies).\nThe criteria are usually in conflict and need to be balanced:\nBetter accuracy may require a more complex, thus less expressive sets of rules.\n\\par\nApproaches for interpretable proxies differ in the validity range of\nthe representations:\nSome aim for surrogates that are only valid \\emph{locally} around\nspecific samples, like in LIME~ or in\n via inductive logic programming.\nOther approaches try to more \\emph{globally} approximate aspects of\nthe model behavior.\nAnother categorization is defined by whether full access\n(\\emph{white-box}), some access (\\emph{gray-box}), or no access\n(\\emph{black-box}) to the DNN internals is needed.\nOne can further differentiate between \\emph{post-hoc} approaches that\nare applied to a trained model, and approaches that try to integrate\nor \\emph{enforce symbolic representations} during training.\nPost-hoc methods cover the wide field of rule extraction techniques\nfor DNNs. The reader may refer to\n.\nMost white- and gray-box methods try to turn the DNN connections into\nif-then rules that are then simplified, like done in\nDeepRED~.\nA black-box example is validity interval\nanalysis~, which refines or\ngeneralizes rules on input intervals, either starting from one sample\nor a general set of rules.\nEnforcement of symbolic representations can be achieved by enforcing\nan output structure that provides insights to the decision logic, such as\ntextual explanations, or a rich output structure allowing investigation of correlations\n.\nAn older discipline for enforcing symbolic representations\nis the field of neural-symbolic learning\n. The idea is based on a hybrid learning\ncycle \nin which a symbolic learner and a DNN iteratively update each other via\nrule insertion and extraction.\n\\par\nThe comprehensibility of global surrogate models suffers from the\ncomplexity and size of concurrent DNNs. Thus, stronger rule\nsimplification methods are required .\nThe alternative direction of local approximations mostly\nconcentrates on linear models instead of more expressive rules .\nFurthermore, balancing of the quality objectives is hard since\navailable indicators for interpretability may not be ideal.\nAnd lastly, applicability is heavily infringed by the requirement of\ninterpretable input features. These are usually not readily available\nfrom input (often pixel-level) or DNN output. Supervised extraction\napproaches vary in their fidelity, and unsupervised ones do not\nguarantee to yield meaningful or interpretable results, respectively, such\nas the super-pixel clusters of LIME.","id":"37f2baca-d82b-4285-9422-b80c37554845","level":"subsection","origin_cites_number":12,"parent_id":"e47014ad-1aec-4b32-9241-6aab90a7f0dd","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Interpretability"],["subsection","Interpretable Proxies"]],"subsections":[],"title":"Interpretable Proxies"},{"cite_extract_rate":0.75,"cites":[4616,4025,3288],"content":"\\label{sec:uncertainty} \n\\sectionauthor{Michael Mock\\textsuperscript{1}}\nUncertainty refers to the view that a neural network is not conceived as a deterministic function but as a probabilistic function or estimator, delivering a random distribution for each input point. Ideally, the mean value of the distribution should be as close as possible to the ground truth value of the function being approximated by the neural network and the uncertainty of the neural network refers to its variance when being considered as a random variable, thus allowing to derive a confidence with respect to the mean value. Regarding safety, the variance may lead to estimations about the confidence associated with a specific network output and opens the option for discarding network outputs with insufficient confidence. \n\\par\nThere are roughly two broad approaches for training neural networks as probabilistic functions: Parametric approaches and Bayesian neural networks on the one hand, such as , where the transitions along the network edges are modeled as probability distributions, and ensemble-based approaches on the other hand , where multiple networks are trained and considered as samples of a common output distribution. Apart from training as probabilistic function, uncertainty measures have been derived from single, standard neural networks by post-processing on the trained network logits, leading for example to calibration measures (\\cf \\eg ).","id":"9f13c349-4fb4-458e-8f34-84cd3a93fca0","level":"section","origin_cites_number":4,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Uncertainty"]],"subsections":["e3a9977a-3391-4691-bc3b-9d131807d8dd","d4df1917-2df5-4888-9dac-86e261e7d3d8","329b292f-42f5-496b-9775-5b5a3bd0f517","dcb846f5-368c-45a9-a56f-6ddb586a0a6b","81f04d24-1455-4244-a481-8a943c0061e4","09524fc1-6625-4adb-b57e-b6186b5d4ef0"],"title":"Uncertainty"},{"cite_extract_rate":0.23333333333333303,"cites":[1003,5680,8430,8207,8206,8205,8208],"content":"\\label{subsec:uncertainty:generative_models}\n\\sectionauthor{Sebastian Wirkert\\textsuperscript{6}, Tim Wirtz\\textsuperscript{1}}\n\\emph{Generative models} belong to the class of unsupervised machine learning models. From a theoretical perspective, these are particularly interesting, because they offer a way to analyze and model the density of data. Given a finite data set $\\mathcal{D}$ independently distributed according to some distribution $p(x)$, generative models aim to estimate or enable sampling from the underlying density $p(x)$ in a model $q_\\theta(x)$. The resulting model can be used for data indexing~, data retrieval~, for visual recognition~, speech recognition and generation~, language processing~ and robotics~.\nFollowing , we can group generative models into two main classes:\n\\begin{itemize}\n \\item Cost function-based models such as autoencoder~, deep belief networks~ and generative adversarial networks~.\n \\item Energy-based models~, where the joint probability density is modeled by an energy function.\n\\end{itemize}\nBeside these \\emph{deep learning} approaches, generative models have been studied in machine learning in general for quite some time (\\cf ). A very prominent example of generative networks are Gaussian processes~ and their deep learning extensions~ as generative models. \n\\par\nAn example of a generative model being employed for image segmentation uncertainty estimation is the probabilistic U-Net . Here a variational autoencoder (VAE) conditioned on the image is trained to model uncertainties. Samples from the VAE are fed into a segmentation U-Net which can thus give different results for the same image. This was tested in context of medical images, where inter-rater disagreements lead to uncertain segmentation results and Cityscapes segmentation. For the Cityscapes segmentation the investigated use case was label ambiguity (\\eg is a BMW X7 a car or a van) using artificially created, controlled ambiguities. Results showed that the probabilistic U-Net could reproduce the segmentation ambiguity modes more reliably than competing methods such as a dropout U-Net which is based on techniques elaborated in the next section.","id":"e3a9977a-3391-4691-bc3b-9d131807d8dd","level":"subsection","origin_cites_number":30,"parent_id":"9f13c349-4fb4-458e-8f34-84cd3a93fca0","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Uncertainty"],["subsection","Generative Models"]],"subsections":[],"title":"Generative Models"},{"cite_extract_rate":0.538461538461538,"cites":[4616,3288,4640,8205,4025,8209,8814],"content":"\\label{subsec:uncertainty:mc_dropout} \n\\sectionauthor{Joachim Sicking\\textsuperscript{1}}\nA widely used technique to estimate model uncertainty is Monte-Carlo (MC) dropout , that offers a Bayesian motivation, conceptual simplicity and scalability to application-size networks. This combination distinguishes MC dropout from competing Bayesian neural network (BNN) approximations (like ,, see \\refsec{subsec:uncertainty:bayesian_neural_networks}). However, these approaches and MC dropout share the same goal: to equip neural networks with a \\emph{self-assessment} mechanism that detects unknown input concepts and thus potential model insufficiencies.\nOn a technical level, MC dropout assumes prior distributions on network activations, usually independent and identically distributed (i.i.d.) Bernoulli distributions. Model training with iteratively drawn Bernoulli samples, the so-called \\emph{dropout masks}, then yields a data-conditioned posterior distribution within the chosen parametric family. It is interesting to note that this training scheme was used earlier---independent from an uncertainty context---for better model generalization . At inference, sampling provides estimates of the input-dependent output distributions. The spread of these distributions is then interpreted as the prediction uncertainty that originates from limited knowledge of model parameters. Borrowing ‘frequentist’ terms, MC dropout can be considered as an implicit network ensemble, \\ie as a set of networks that share (most of) their parameters.\nIn practice, MC dropout requires only a minor modification of the optimization objective during training and multiple, trivially parallelizable forward passes during inference. The loss modification is largely agnostic to network architecture and does not cause substantial overhead. This is in contrast to the sampling-based inference that increases the computational effort massively---by estimated factors of 20-100 compared to networks without MC dropout. A common practice is therefore the use of \\emph{last-layer dropout} that reduces computational overhead to estimated factors of 2-10. Alternatively, analytical moment propagation allows sampling-free MC-dropout inference at the price of additional approximations (\\eg). Further extensions of MC dropout target the integration of data-inherent (aleatoric) uncertainty and tuned performance by learning layer-specific dropout rate using concrete relaxations .\nThe quality of MC-dropout uncertainties is typically evaluated using negative log-likelihood (NLL), expected calibration error (ECE) and its variants (\\cf \\refsec{subsec:uncertainty:confidence_calibration}) and by considering correlations between uncertainty estimates and model errors (\\eg AUSE ). Moreover, it is common to study how useful uncertainty estimates are for solving auxiliary tasks like out-of-distribution classification or robustness \\wrt adversarial attacks. \nMC dropout is a working horse of safe ML, being used with various networks and for a multitude of applications (\\eg). However, several authors pointed out shortcomings and limitations of the method: MC dropout bears the risk of over-confident false predictions (), offers less diverse uncertainty estimates compared to (the equally simple and scalable) deep ensembles (, see \\refsec{subsec:aggregation:ensemble_methods}) and provides only rudimentary approximations of true posteriors. \nRelaxing these modelling assumptions and strengthening the Bayesian motivation of MC dropout is therefore an important research avenue. Further directions for future work are the development of \\emph{semantic uncertainty mechanisms} (\\eg ), improved local uncertainty calibrations and a better understanding of the outlined sampling-free schemes to uncertainty estimation.","id":"d4df1917-2df5-4888-9dac-86e261e7d3d8","level":"subsection","origin_cites_number":13,"parent_id":"9f13c349-4fb4-458e-8f34-84cd3a93fca0","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Uncertainty"],["subsection","Monte-Carlo Dropout"]],"subsections":[],"title":"Monte-Carlo Dropout"},{"cite_extract_rate":0.25,"cites":[8633,4025],"content":"\\label{subsec:uncertainty:bayesian_neural_networks}\n\\sectionauthor{Maram Akila\\textsuperscript{1}}\nAs the name suggests, Bayesian neural networks (BNNs) are inspired by a Bayesian interpretation of probability (for an introduction \\cf ).\nIn essence, it rests on Bayes' theorem,\n\\begin{equation}\n\tp(x|y) p(y) = p(x,y) = p(y | x) p(x)\n\t\\quad \\Rightarrow \\quad\n\tp(x| y) = \\frac{ p(y | x) p(x)}{ p(y) }\n\t\\,,\n\t\\label{eq:bnn:bayes}\n\\end{equation}\nstating that the conditional probability density function (PDF) $p(x|y)$ for $x$ given $y$ may be expressed in terms of the inverted conditional PDF $p(y | x)$.\nFor machine learning, where one intends to make predictions $y$ for unknown $x$ given some training data $\\mathcal{D}$, this can be reformulated into\n\\begin{equation}\n\ty = \\operatorname{NN}(x|W)\n\t\\quad\\text{with}\\quad\n\t p(W | \\mathcal{D}) = \\frac{p(\\mathcal{D} | W) p(W)}{p(\\mathcal{D})}\n\t\\,.\n\t\\label{eq:bnn:NNpdf}\n\\end{equation}\nTherein NN denotes a conventional (deep) neural network (DNN) with model parameters $W$, \\eg the set of weights and biases.\nIn contrast to a regular DNN, the weights are given in terms of a probability distribution $ p(W | \\mathcal{D}) $ turning also the output $y$ of a BNN into a distribution.\nThis allows to study the mean $\\mu=\\langle y^1 \\rangle$ of the DNN for a given $x$ as well as higher moments of the distribution, typically the resulting variance $\\sigma^2 = \\langle (y-\\mu)^2 \\rangle$ is of interest, where\n\\begin{equation}\n\t\\left\\langle y^k \\right\\rangle = \\int\\!\n\t\\operatorname{NN}(x|W)^k p(W|\\mathcal{D})\\,\\mathrm{d}W\\,.\n\t\\label{eq:bnn:moments}\n\\end{equation}\nWhile $\\mu$ yields the output of the network, the variance $\\sigma^2$ is a measure for the uncertainty of the model for the prediction at the given point.\nCentral to this approach is the probability of the data given the model, here denoted by $p(\\mathcal{D} | W) $, as it is the key component connecting model and training data.\nTypically, the prior distribution $p(\\mathcal{D})$ is ``ignored'' as it only appears as a normalization constant within the averages, see \\eqref{eq:bnn:moments}.\nIn the cases where the data $\\mathcal{D}$ is itself a distribution due to inherent uncertainty, \\ie presence of aleatoric risk , such a concept seems natural.\nHowever, Bayesian approaches are also applicable for all other cases.\nIn those, loosely speaking, the likelihood of $W$ is determined via the chosen loss function (for the connection between the two concepts \\cf ).\n\\par\nOn this general level, Bayesian approaches are broadly accepted and also find use for many other model classes besides neural networks.\nHowever, the loss surfaces of DNNs are known for their high dimensionality and strong non-convexity.\nTypically, there are abundant parameter combinations $W$ that lead to (almost) equally good approximations to the training data $\\mathcal{D}$ with respect to a chosen loss.\nThis makes an evaluation of $p(W|\\mathcal{D})$ for DNNs close to impossible in full generality.\nAt least no (exact) solutions for this case exist at the moment.\n\\par\nFinding suitable approximations to the posterior distribution $p(W|\\mathcal{D})$ is an ongoing challenge for the construction of BNNs.\nAt this point we only summarize two major research directions in the field.\nOne approach is to assume that the distribution factorizes.\nWhile the full solution would be a joined distribution implying correlations between different weights (etc.), possibly even across layers, this approximation takes each element $w_i$ of $W$ to be independent form the others.\nAlthough this is a strong assumption, it is often made, in this case parameters for the respective distributions of each element can be learned via training (\\cf ).\nThe second class of approaches focuses on the region of the loss surface around the minimum chosen for the DNN.\nAs discussed, the loss relates to the likelihood and quantities such as the curvature at the minimum, therefore directly connected to the distribution of $W$.\nUnfortunately, already using this type of quantities requires further approximations .\nAlternatively, the convergence of the training process may be altered to sample networks close to the minimum .\nWhile this approach contains information about correlations among the $w_i$, it is usually restricted to a specific minimum.\nFor a non-Bayesian ansatz taking into account several minima, see deep ensembles in \\refsec{subsec:aggregation:ensemble_methods}.\nBNNs also touch other concepts such as MC dropout (\\cf \\refsec{subsec:uncertainty:mc_dropout} or ), or prior networks, which are based on a Bayesian interpretation but use conventional DNNs with an additional (learned) $\\sigma$ output .","id":"329b292f-42f5-496b-9775-5b5a3bd0f517","level":"subsection","origin_cites_number":8,"parent_id":"9f13c349-4fb4-458e-8f34-84cd3a93fca0","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Uncertainty"],["subsection","Bayesian Neural Networks"]],"subsections":[],"title":"Bayesian Neural Networks"},{"cite_extract_rate":0.5,"cites":[1624,4620,8211,8210,1733,206],"content":"\\label{subsec:uncertainty:gradient_based_uncertainty}\n\\sectionauthor{Matthias Rottmann\\textsuperscript{5}}\nClassical uncertainty quantification methods in frequentist inference are mostly based on the outputs of statistical models. Their uncertainty is quantified and assessed for instance via dispersion measures in classification (such as entropy, probability margin or variation ratio), or confidence intervals in regression. However, the nature of DNN architectures and the cutting edge applications tackled by those (\\eg semantic segmentation, \\cf ) open the way towards more elaborate uncertainty quantification methods. Besides the mentioned classical approaches, intermediate feature representations within a DNN (\\cf ) or gradients according to self-affirmation that represent re-learning stress (see ) reveal additional information. In addition, in case of semantic segmentation, the geometry of a prediction may give access to further information, cf.\\ . By the computation of statistics of those quantities as well as low-dimensional representations thereof, we obtain more elaborate uncertainty quantification methods specifically designed for DNNs that can help us to detect misclassifications and out-of-distribution objects (\\cf ). \n\\par\nFeatures gripped during a fordward pass of a data point $x$ through a DNN $f$ can be considered layer-wise, \\ie $f^{(\\ell)}(x)$ after the $\\ell$-th layer. These can be translated into a handful of quantities per layer or further processed by another DNN that aims that detecting errors . While in particular presents a proof of concept on small scale classification problems, their applicability to large scale datasets and problems such as semantic segmentation and object detection remain open.\n\\par\nThe development for gradient-based uncertainty quantification methods is guided by one central question:\nIf the present prediction was true, how much re-learning would this require.\nThe corresponding hypothesis is that wrong predictions would be more in conflict with the knowledge encoded in the deep neural network than correct ones, therefore causing increased re-learning stress.\nGiven a predicted class\n\\begin{equation}\n\\hat{y} = \\mathop{\\mathrm{arg\\,max}}_{y} f_y(x) \n\\end{equation}\nwe compute the gradient of layer $\\ell$ corresponding to the predicted label. That is, given a loss function $\\mathcal{L}$, we compute\n\\begin{equation}\n\\nabla_{w_\\ell} \\mathcal{L}( \\hat{y}, x, w ) \n\\end{equation}\nvia backpropagation. The obtained quantities can be treated similarly to the case of forward pass features. While this concept seems to be prohibitively expensive for semantic segmentation (at least when calculating gradients for each pixel of $\\hat{y}$), its applicability to object detection might be feasible, in particular with respect to offline applications. Gradients are also of special interest in active learning with query by \\emph{expected model change} (\\cf ).\n\\par\nIn the context of semantic segmentation, geometrical information on segments shapes as well as neighborhood relations of predicted segments can be taken into account along side with dispersion measures. It has been demonstrated that the detection of errors in an in-distribution setting strongly benefits from geometrical information. Recently, this has also been considered in scenarios under moderate domain shift . However, its applicability to out-of-distribution examples and to other sensors than the camera remains subject to further research.","id":"dcb846f5-368c-45a9-a56f-6ddb586a0a6b","level":"subsection","origin_cites_number":12,"parent_id":"9f13c349-4fb4-458e-8f34-84cd3a93fca0","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Uncertainty"],["subsection","Uncertainty Metrics for DNNs in Frequentist Inference"]],"subsections":[],"title":"Uncertainty Metrics for DNNs in Frequentist Inference"},{"cite_extract_rate":0.23529411764705802,"cites":[8213,8212,1737,1745],"content":"\\label{subsec:uncertainty:markov_random_fields}\n\\sectionauthor{Seyed Eghbal Ghobadi\\textsuperscript{8}, Ahmed Hammam\\textsuperscript{8}}\nAlthough deep neural networks are currently the state of the art for almost all computer vision tasks, Markov random fields (MRF) remain one of the fundamental techniques used for many computer vision tasks, specifically image segmentation ,. MRFs hold its power in the essence of being able to model dependencies between pixels in an image. With the use of energy functions, MRFs integrate pixels into models relating between unary and pairwise pixels together . \nGiven the model, MRFs are used to infer the optimal configuration yielding the lowest energy using mainly maximum a posteriori (MAP) techniques. Several MAP inference approaches are used to yield the optimal configuration such as graph cuts and belief propagation algorithms . However, as with neural networks, MAP inference techniques result in deterministic point estimates of the optimal configuration without any sense of uncertainty in the output. To obtain uncertainties on results from MRFs, most of the work is directed towards modelling MRFs with Gaussian distributions. Getting uncertainties from MRFs with Gaussian distributions is possible by two typical methods: Either approximate models are inferred to the posterior, from which sampling is easy or the variances can be estimated analytically, or approximate sampling from the posterior is used.\nApproximate models include those inferred using variational Bayesian (VB) methods, like mean-field approximations, and using Gaussian process (GP) models enforcing a simplified prior model , . Examples of approximate sampling methods include traditional Markov chain Monte Carlo (MCMC) methods like Gibbs sampling . Some recent theoretical advances propose the perturb-and-MAP framework and a Gumbel perturbation model (GPM) , to exactly sample from MRF distributions. Another line of work has also been proposed, where MAP inference techniques are used to estimate the probability of the network output. With the use of graph cuts, try to estimate uncertainty using the min-marginals associated with the label assignments of a random field. Here, the work by Kohli and Torr was extended to show how this approach can be extended to techniques other than graph cuts or compute uncertainties on multi-label marginal distributions .\n\\par\nA current research direction is the incorporation of MRFs with deep neural networks, along with providing uncertainties on the output . This can also be extended to other forms of neural networks such as recurrent neural networks to provide uncertainties on segmentation of streams of videos with extending dependencies of pixels to previous frames , .","id":"81f04d24-1455-4244-a481-8a943c0061e4","level":"subsection","origin_cites_number":17,"parent_id":"9f13c349-4fb4-458e-8f34-84cd3a93fca0","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Uncertainty"],["subsection","Markov Random Fields"]],"subsections":[],"title":"Markov Random Fields"},{"cite_extract_rate":0.41176470588235203,"cites":[8214,8216,759,8215,4141,8217,4690],"content":"\\label{subsec:uncertainty:confidence_calibration} \n\\sectionauthor{Fabian Küppers\\textsuperscript{9}, Anselm Haselhoff\\textsuperscript{9}}\nNeural network classifiers output a label $\\hat{Y} \\in \\mathcal{Y}$ on a given input $X \\in \\mathcal{X}$ with an associated confidence $\\hat{P}$. This confidence can be interpreted as a probability of correctness that the predicted label matches the ground truth label $Y \\in \\mathcal{Y}$. Therefore, these probabilities should reflect the ”self-confidence” of the system. If the empirical accuracy for any confidence level matches the predicted confidence, a model is called \\emph{well calibrated}.\nTherefore, a \\emph{classification} model is perfectly calibrated if \n\\begin{align}\n\\label{eq:classification_calibration}\n& \\underbrace{\\mathbb{P}(\\hat{Y} = Y | \\hat{P} = p)}_{\\text{accuracy given } p} = \\underbrace{p}_{\\text{confidence}} \\quad \\forall p \\in [0,1]\n\\end{align}\nis fullfilled . For example, assume 100 predictions with confidence values of 0.9. We call the model well calibrated if 90 out of these 100 predictions are actually correct. However, recent work has shown that modern neural networks tend to be too overconfident in their predictions . The deviation of a model to the perfect calibration can be measured by the expected calibration error (ECE) . \nIt is possible to recalibrate models as a post-processing step after classification. One way to get a calibration mapping is to group all predictions into several bins by their confidence. Using such a binning scheme, it is possible to compute the empirical accuracy for certain confidence levels, as it is known for a long time already in reconstructing confidence outputs for Viterbi decoding . Common methods are histogram binning , isotonic regression or more advanced methods like Bayesian binning into quantiles (BBQ) and ensembles of near-isotonic regression (ENIR) . Another way to get a calibration mapping is to use scaling methods based on logistic regression like Platt scaling , temperature scaling and beta calibration .\n\\par\nIn the setting of probabilistic regression, a model is calibrated if, \\eg 95\\% of the true target values are below or equal to a credible level of 95\\% (so called \\emph{quantile-calibrated regression}) . A regression model is usually calibrated by fine-tuning its predicted CDF in a post-processing step to match the empirical frequency. Common approaches utilize isotonic regression , logistic and beta calibration , as well as Gaussian process models to build a calibration mapping. In contrast to quantile-calibrated regression, have recently introduced the concept of distribution calibration, where calibration is applied on a distribution level and naturally leads to calibrated quantiles.\n\\par\nRecent work has shown that miscalibration in the scope of \\emph{object detection} also depends on the position and scale of a detected object . The additional box regression output is denoted by $\\hat{R}$ with $J$ as the size of the used box encoding.\nFurthermore, if we have no knowledge about all anchors of a model (which is a common case in many applications), it is not possible to determine the accuracy. Therefore, K\\\"{u}ppers \\etal~ use the precision as a surrogate for accuracy and propose that an \\emph{object detection model} is perfectly calibrated if\n\\begin{align}\n\\label{eq:detection_calibration}\n& \\underbrace{\\mathbb{P}(M=1 | \\hat{P} = p, \\hat{Y} = y, \\hat{R} = r)}_{\\text{precision given } p, y, r} = \\underbrace{p}_{\\text{confidence}} \\quad \\forall p \\in [0,1], y \\in \\mathcal{Y}, r \\in \\mathbb{R}^J \n\\end{align}\nis fulfilled, where $M=1$ denotes a correct prediction that matches a ground-truth object with a chosen IoU threshold and $M=0$ denotes a mismatch, respectively. The authors propose the detection-expected calibration error (D-ECE) as the extension of the ECE to object detection tasks in order to measure miscalibration also by means of the position and scale of detected objects. Other approaches try to fine-tune the regression output in order to obtain more reliable object proposals or to add a regularization term to the training objective such that training yields models that are both well-performing and well-calibrated .","id":"09524fc1-6625-4adb-b57e-b6186b5d4ef0","level":"subsection","origin_cites_number":17,"parent_id":"9f13c349-4fb4-458e-8f34-84cd3a93fca0","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Uncertainty"],["subsection","Confidence Calibration"]],"subsections":[],"title":"Confidence Calibration"},{"cite_extract_rate":0.5,"cites":[6435],"content":"\\label{sec:aggregation}\n\\sectionauthor{Maram Akila\\textsuperscript{1}}\nFrom a high-level perspective, a \\emph{neural network} is based on processing inputs and coming to some output conclusion, \\eg mapping incoming image data onto class labels.\nAggregation or collection of non-independent information on either the input or output side of this network function can be used as a tool to leverage its performance and reliability.\nStarting with the input, any additional ``dimension'' to add data can be of use.\nFor example, in the context of autonomous vehicles this might be input from any further sensor measuring the same scene as the original one, \\eg stereo cameras or LiDAR.\nCombining those sensor sets for prediction is commonly referred to as \\emph{sensor fusion} .\nStaying with the example, the scene will be monitored consecutively providing a whole (temporally ordered) stream of input information.\nThis may be used either by adjusting the network for this kind of input \n \nor in terms of a post-processing step, in which the predictions are aggregated by some measure of temporal consistency. \n\\par\nAnother more implicit form of aggregation is training the neural network on several ``independent'' tasks, \\eg segmentation and depth regression. Although the individual task is executed on the same input, the overall performance can still benefit from the correlation among all given tasks. We refer to the discussion on \\emph{multi-task networks} in \\refsec{subsec:architecture:multi_task_networks}.\nBy extension, solving the same task in multiple different ways, can be beneficial for performance and provide a measure of redundancy.\nIn this survey, we focus on single-task systems and discuss \\emph{ensemble methods} in the next section and the use of \\emph{temporal consistency} in the one thereafter.","id":"71d4e278-f67c-443c-84a9-ff0254362119","level":"section","origin_cites_number":2,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Aggregation"]],"subsections":["d1f9f7cf-fc5d-426f-8fba-a38a5d51acf5","61379b41-7fce-4042-aff6-0ec085cd8afc"],"title":"Aggregation"},{"cite_extract_rate":0.6666666666666661,"cites":[4616,4678,582,3288,680,681,8966,8218],"content":"\\label{subsec:aggregation:ensemble_methods}\n\\sectionauthor{Joachim Sicking\\textsuperscript{1}}\nTraining a neural network is optimizing its parameters to fit a given training data set. The commonly used gradient-based optimization schemes cause convergence in a ‘nearby’ local minimum. As the loss landscapes of neural networks are notoriously non-convex , various locally optimal model parameter sets exist. These local optima differ in the degree of optimality (“deepness”), qualitative characteristics (“optimal for different parts of the training data”) and their generalizability to unseen data (commonly referred to by the geometrical terms of “sharpness” and “flatness” of minima ). \nA single trained network corresponds to one local minimum of such a loss landscape and thus captures only a small part of a potentially diverse set of solutions. \\emph{Network ensembles} are collections of models and therefore better suited to reflect this multi-modality. Various modelling choices shape a loss landscape: the selected model class and its meta-parameters (like architecture and layer width), the training data and the optimization objective. Accordingly, approaches to diversify ensemble components range from combinations of different model classes over varying training data (bagging) to methods that train and weight ensemble components to make up for the flaws of other ensemble members (boosting) .\nGiven the millions of parameters of application-size networks, ensembles of NNs are resource-demanding \\wrt computational load, storage and runtime during training and inference. This complexity increases linearly with ensemble size for naïve ensembling. Several approaches were put forward to reduce some dimensions of this complexity: \\emph{snapshot ensembles} require only one model optimization with a cyclical learning-rate schedule—leading to an optimized training runtime. The resulting training trajectory passes through several local minima. The corresponding models compose the ensemble. On the contrary, \\emph{model distillation} tackles runtime at inference. They ‘squeeze’ a NN ensemble into a single model that is optimized to capture the gist of the model set. However, such a compression goes along with reduced performance compared to the original ensemble.\nSeveral hybrids of single model and model ensemble exist: Multi-head networks share a backbone network that provides inputs to multiple prediction networks. Another variant are mixture-of-expert models that utilize a gating network to assign inputs to specialized expert networks . Multi-task networks (\\cf \\refsec{subsec:architecture:multi_task_networks}) and Bayesian approximations of NNs (\\cf \\refsec{subsec:uncertainty:bayesian_neural_networks} and \\refsec{subsec:uncertainty:mc_dropout}) can be seen as implicit ensembles. \nNN ensembles (or deep ensembles) are not only used to boost model quality. They pose the frequentist's approach to estimating NN uncertainties and are state-of-the-art in this regard . The emergent field of federated learning is concerned with the integration of decentrally trained ensemble components and safety-relevant applications of ensembling range from autonomous driving to medical diagnostics . \nTaking this safe-ML perspective, promising research directions comprise a more principled and efficient composition of model ensembles, \\eg by application-driven diversification, as well as improved techniques to miniaturize ensembles, \\eg by gaining a better understanding of methods like distillation. In the long run, better designed, more powerful learning systems might partially reduce the need for combining weaker models in a network ensemble.","id":"d1f9f7cf-fc5d-426f-8fba-a38a5d51acf5","level":"subsection","origin_cites_number":12,"parent_id":"71d4e278-f67c-443c-84a9-ff0254362119","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Aggregation"],["subsection","Ensemble Methods"]],"subsections":[],"title":"Ensemble Methods"},{"cite_extract_rate":0.4,"cites":[8219,1745,8220,784],"content":"\\label{subsec:aggregation:temporal_consistency}\n\\sectionauthor{Timo Sämann\\textsuperscript{4}}\nThe focus of previous DNN development for semantic segmentation has been on single image prediction. This means that the final and intermediate results of the DNN are discarded after each image. However, the application of a computer vision model often involves the processing of images in a sequence, \\ie there is a temporal consistency in the image content between consecutive frames (for a metric, \\cf, \\eg ). This consistency has been exploited in previous work to increase quality and reduce computing effort. Furthermore, this approach offers the potential to improve the robustness of DNN prediction by incorporating this consistency as a-priori knowledge into DNN development. The relevant work in the field of video prediction can be divided in two major approaches: \n\\begin{enumerate}\n\\item DNNs are specially designed for video prediction. This usually requires training from scratch and the availability of training data in a sequence.\n\\item A transformation from single prediction DNNs to video prediction DNNs takes place. Usually no training is required, \\ie the existing weights of the model can be used unaltered.\n\\end{enumerate}\nThe first set of approaches often involves \\emph{conditional random fields} (CRF) and its variants. \nCRFs are known for their use as postprocessing step in the prediction of semantic segmentation, in which their parameters are learned separately or jointly with the DNN~. \nAnother way to use spatiotemporal features is to include 3D convolutions, which add an additional dimension to the conventional 2D convolutional layer. Tran \\etal use 3D convolution layers for video recognition tasks such as action and object recognition. \nOne further approach to use spatial and temporal characteristics of the input data is to integrate \\emph{long short-term memory} (LSTM)~, a variant of the \\emph{recurrent neural network} (RNN). Fayyaz \\etal integrate LSTM layers between the encoder and decoder of their convolutional neural network for semantic segmentation. The significantly higher GPU memory requirements and computational effort are a disadvantage of this method. More recently, Nilsson and Sminchisescu deployed \\emph{gated recurrent units}, which generally requires significantly less memory.\nAn approach to improve temporal consistency of automatic speech recognition outputs is known as a posterior-in-posterior-out (PIPO) LSTM ``sequence enhancer'', a postfilter which could be applicable to video processing as well .\nA disadvantage of the described methods is that sequential data for training must be available, which may be limited or show a lack of diversity. \n\\par\nThe second class of approaches has the advantage that it is model-independent most of the time. Shelhamer \\etal found that the deep feature maps within the network change only slightly with temporal changes in video content.\nAccordingly, calculate the optical flow of the input images from time steps $t_0$ and $t_{-1}$ and convert it into the so-called \\emph{transform flow} which is used to transform the feature maps of the time step $t_{-1}$ so that an aligned representation to the feature map $t_0$ is achieved. S{\\\"a}mann \\etal use a confidence-based combination of feature maps from previous time steps based on the calculated optical flow.","id":"61379b41-7fce-4042-aff6-0ec085cd8afc","level":"subsection","origin_cites_number":10,"parent_id":"71d4e278-f67c-443c-84a9-ff0254362119","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Aggregation"],["subsection","Temporal Consistency"]],"subsections":[],"title":"Temporal Consistency"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:verification}\n\\sectionauthor{Gesina Schwalbe\\textsuperscript{3}}\nVerification and validation is an integral part of the safety\nassurance for any safety critical systems. As of the functional safety\nstandard for automotive systems ,\n\\emph{verification} means to determine whether given requirements are\nmet \\cite[3.180]{iso/tc_22/sc_32_iso_2018}, such as performance\ngoals.\n\\emph{Validation} on the other side tries to assess whether the given\nrequirements are sufficient and adequate to guarantee safety\n\\cite[3.148]{iso/tc_22/sc_32_iso_2018}, \\eg whether certain types of\nfailures or interactions simply were overlooked. The latter is usually\nachieved via extensive testing in real operation conditions of the\nintegrated product. This differs from the notion of validation used in\nthe machine learning community in which it usually refers to simple\nperformance tests on a selected dataset.\nIn this section, we want to concentrate on general verification\naspects for deep neural networks.\nVerification as in the safety domain encompasses (manual) inspection\nand analysis activities, and testing. However, the contribution\nof single processing steps within a neural network to the final\nbehavior can hardly be assessed manually (compare to the problem of\ninterpretability in \\refsec{sec:interpretability}). Therefore, we here will concentrate on different\napproaches to verification testing.\nSection~\\ref{subsec:verification:formal_testing} covers approaches\nto systematic test data selection. While the suggested methods assume\nfull access to the model internals for coverage measurement, this is\nnot in all cases available. Therefore,\n\\refsec{subsec:verification:black_box_methods} highlights\nassessment approaches that consider the DNN as a black-box component.\nAlso, the topic of verification activity during operation with the\nhelp of observers is discussed.","id":"94174462-51f2-497a-b88f-1ceaadd52aaf","level":"section","origin_cites_number":1,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Verification"]],"subsections":["5ad8ed4b-a3fa-43df-8241-7dd338b6350c","5a9669c2-3369-414d-a5ae-bc8b21dd4bcd"],"title":"Verification"},{"cite_extract_rate":0.28571428571428503,"cites":[8089,8971,1636,6024],"content":"\\label{subsec:verification:formal_testing}\n\\sectionauthor{Christian Heinzemann\\textsuperscript{2}, Gesina Schwalbe\\textsuperscript{3}, Matthias Woehrle\\textsuperscript{2}}\nFormal testing refers to testing methods that include formalized and\nformally verifiable steps, \\eg for test data acquisition, or for\nverification in the local vicinity of test samples.\nFor image data, local testing around valid samples is usually more\npractical than fully formal verification: (Safety) properties are not\nexpected to hold on the complete input space but only on the much\nsmaller unknown lower-dimensional manifold of real images .\nSources of such samples can be real ones or generated ones using an\ninput space formalization or a trained generative model.\n\\par\nCoverage criteria for the data samples are commonly used for two purposes:\n(a) deciding when to stop testing or\n(b) identifying missing tests.\nFor CNNs, there are at least three different approaches towards\ncoverage:\n(1) approaches that establish coverage based on a model with semantic\nfeatures of the input space~,\n(2) approaches trying to semantically cover the latent feature space\nof neural network or a proxy network (\\eg an\nautoencoder)~, and\n(3) approaches trying to cover neurons and their interactions,\ninspired by classical software white-box\nanalysis~.\n\\par\nTypical types of properties to verify are \nsimple test performance,\nlocal stability (robustness),\na specific structure of the latent spaces like\nembedding of semantic concepts ,\nand more complex logical constraints on inputs and outputs,\nwhich can be used for testing when fuzzified .\nMost of these properties require in-depth semantic information about\nthe DNN inner workings, which is often only available via interpreting\nintermediate representations ,\nor interpretable proxies / surrogates (\\cf \\refsec{subsec:interpretability:interpretable_proxies}), which do not guarantee fidelity.\n\\par\nThere exist different testing and formal verification methods from\nclassical software engineering that have already been applied to CNNs.\n\\emph{Differential testing} as used by\nDeepXPlore~ trains $n$ different CNNs for\nthe same task using independent data sets and compares the individual\nprediction results on a test set. This allows to identify\ninconsistencies between the CNNs but no common weak spots.\n\\emph{Data augmentation} techniques start from a given data set and\ngenerate additional transformed data. \\emph{Generic data augmentation}\nfor images like rotations and translation are state-of-the-art for\ntraining but may also be used for testing.\n\\emph{Concolic testing} approaches incrementally grow test suites with\nrespect to a coverage model to finally achieve completeness. Sun \\etal use an adversarial input model based on\nsome norm (\\cf \\refsec{subsec:adverarial_attacks:adversarial_attacks_defenses}),\n\\eg an $L_p$-norm, for generating additional images around a given image\nusing concolic testing. \\emph{Fuzzing} generates new test data constrained by\nan input model and tries to identify interesting test cases, \\eg by\noptimizing white-box coverage mentioned\nabove~. Fuzzing techniques may also be\ncombined with the differential testing approach discussed\nabove~.\nIn all these cases it needs to be ensured that the image as well as\nits meta-data remain valid for testing after transformation.\nFinally, \\emph{proving methods} surveyed by Liu \\etal try to formally prove properties on a\ntrained neural network, \\eg based on satisfiablity modulo theories (SMT). \nThese approaches require a formal characterization of an input space\nand the property to be checked, which is hard for non-trivial\nproperties like contents of an image. \n\\par\nExisting formal testing approaches can be quite costly to integrate\ninto testing workflows (\\cf ):\nDifferential testing and data augmentation require several inferences\nper initial test sample;\nconcolic and fuzzy testing apply an optimization to each given test\nsample, while convergence towards the coverage goals is not guaranteed;\nalso, the iterative approaches need tight integration into the testing\nworkflow;\nand lastly, proving methods usually have to balance computational\nefficiency against the precision or completeness of the result\n.\nAnother challenge of formal testing is that machine learning\napplications usually solve problems for which no (formal) specification\nis possible. This makes it hard to find useful requirements for\ntesting and properties that can be formally\nverified. \nEven partial requirements such as specification of useful input\nperturbations, specified corner cases, and valuable coverage goals are\ntypically difficult to identify .","id":"5ad8ed4b-a3fa-43df-8241-7dd338b6350c","level":"subsection","origin_cites_number":14,"parent_id":"94174462-51f2-497a-b88f-1ceaadd52aaf","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Verification"],["subsection","Formal Testing"]],"subsections":[],"title":"Formal Testing"},{"cite_extract_rate":0.533333333333333,"cites":[7507,7610,681,1818,8221,603,1822,2676],"content":"\\label{subsec:verification:black_box_methods}\n\\sectionauthor{Jonas Löhdefink\\textsuperscript{15}, Julia Rosenzweig\\textsuperscript{1}}\nIn machine learning literature, neural networks are often referred to as black boxes due to the fact that their internal operations and their decision making are not completely understood , hinting at a lack of interpretability and transparency. However, in this survey we consider a black box to be a machine learning model to which we only have oracle (query) access . That means we can query the model to get input-output pairs, but we do not have access to the specific architecture (or weights, in case of neural networks). As describes, black boxes are increasingly wide-spread, \\eg healthcare, autonomous driving or ML as a service in general, due to proprietary, privacy or security reasons. \n\\par\nAs deploying black boxes gains popularity, so do methods that aim to extract internal information such as architecture and parameters or to find out, whether a sample belongs to the training dataset. These include \n\\emph{model extraction attacks} , \\emph{membership inference attacks} , general attempts to reverse-engineer the model or to attack it adversarially . Protection and counter-measures are also actively researched: proposes a warning system that estimates how much information an attacker could have gained from queries. The authors of use watermarks for models to prevent illegal re-distribution and to identify intellectual property.\n\\par\nMany papers in these fields make use of so-called \\emph{surrogate}, \\emph{avatar} or \\emph{proxy models} that are trained on input-output pairs of the black box.\nIn case the black-box output is available in soft form (\\eg logits), distillation as first proposed by can be applied to train the surrogate (student) model. Then, any white-box analysis can be performed on the surrogates (\\cf \\eg ) to craft adversarial attacks targeted at the black box. More generally, (local) surrogates as for example in can be used to (locally) explain its decision-making. \nMoreover, these techniques are also of interest if one wants to compare or test black-box models (\\cf \\refsec{subsec:verification:formal_testing}, formal verification). \nThis is the case, among others, in ML marketplaces, where you wish to buy a pre-trained model , or if you want to verify or audit that a third-party black-box model obeys regulatory rules (\\cf ). \nAnother topic of active research are so-called \\emph{observers}.\nThe concept of observers is to evaluate the interface of a black-box module to determine if it behaves as expected within a given set of parameters.\nThe approaches can be divided into \\emph{model-explaining} and \\emph{anomaly-detecting} observers.\nFirst, model explanation methods answer the question of which input characteristic is responsible for changes at the output.\nThe observer is able to alter the inputs for this purpose.\nIf the input of the model under test evolves only slightly but the output changes drastically, this can be a signal that the neural network is mislead, which is also strongly related to adversarial examples (\\cf Chapter \\ref{sec:adversarial_attacks}).\nHence, the reason for changes in the classification result via the input can be very important.\nIn order to figure out in which region of an input image the main reason for the classification is located, ``delete'' information from the image by replacing regions with generated patches until the output changes.\nThis replaced region is likely responsible for the decision of the neural network.\nBuilding upon this, adapt the approach to medical images and generate ``deleted'' regions by a variational autoencoder (VAE).\nSecond, anomaly-detecting observers register input and output anomalies, either examining input and output independently or as an input-output pair, and predict the black-box performance in the current situation.\nIn contrast to model-explaining approaches, this set of approaches has high potential to be used in an online scenario since it does not need to modify the model input.\nThe maximum mean discrepancy (MMD)~ measures the domain gap between two data distributions independently from the application and can be used to raise a warning if input or output distributions during inference deviate too strongly from their respective training distributions.\nBy use of a GAN-based autoencoder perform a domain shift estimation using neural networks in conjunction with the Wasserstein distance as domain mismatch metric.\nThis metric can also be evaluated by use of a casual time-variant aggregation of distributions during inference time.","id":"5a9669c2-3369-414d-a5ae-bc8b21dd4bcd","level":"subsection","origin_cites_number":15,"parent_id":"94174462-51f2-497a-b88f-1ceaadd52aaf","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Verification"],["subsection","Black Box Methods"]],"subsections":[],"title":"Black Box Methods"},{"cite_extract_rate":0.4,"cites":[97,7588,1455,3769],"content":"\\label{sec:architecture}\n\\sectionauthor{Michael Weber\\textsuperscript{14}}\nIn order to solve a specific task, the architecture of a CNN and its building blocks play a significant role.\nSince the early days of using CNNs in image processing, when they were applied to handwriting recognition~ and the later breakthrough in general image classification~, the architecture of the networks has changed radically.\nDid the term of \\emph{deep learning} for these first convolutional neural networks imply a depth of approximately four layers, their depth increased significantly during the last years and new techniques had to be developed to successfully train and utilize these networks~.\nIn this context, new activation functions~ as well as new loss functions~ have been designed and new optimization algorithms~ were investigated.\nWith regard to the layer architecture, the initially alternating repetition of convolution and pooling layers as well as their characteristics have changed significantly.\nThe convolution layers made the transition from a few layers with often large filters to many layers with small filters.\nA further trend was then the definition of entire modules, which were used repeatedly within the overall architecture as so-called \\emph{network in network}~.\nIn areas such as autonomous driving, there is also a strong interest in the simultaneous execution of different tasks within one single convolutional neural network architecture.\nThis kind of architecture is called \\emph{multi-task learning}~(MTL)~ and can be utilized in order to save computational resources and at the same time to increase performance of each task .\nWithin such multi-task networks, usually one shared feature extraction part is followed by one separate so-called head per task~.\nIn each of these architectures, manual design using expert knowledge plays a major role. The role of the expert is the crucial point here.\nIn recent years, however, there have also been great efforts to automate the process of finding architectures for networks or, in the best case, to learn them.\nThis is known under the name \\emph{neural architecture search}~(NAS).","id":"f0f7d7d0-2bd8-48ae-bdb1-126b96887d72","level":"section","origin_cites_number":10,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Architecture"]],"subsections":["0c7b2cb4-747c-4997-8a70-14bcffc01831","9fd0032b-a9e0-4b5c-ae60-155c9130c8d4","19fdc458-c9c5-4f0b-a62d-6dbcbe8fef85"],"title":"Architecture"},{"cite_extract_rate":0.764705882352941,"cites":[502,97,810,7588,8224,8223,514,305,1210,1455,8222,3769,40],"content":"\\label{subsec:architecture:building_blocks}\n\\sectionauthor{Michael Weber\\textsuperscript{14}}\nDesigning a convolutional neural network typically includes a number of design choices.\nThe general architecture usually contains a number of convolutional and pooling layers which are arranged in a certain pattern. \nConvolutional layers are commonly followed by a non-linear activation function. \nThe learning process is based on a loss function which determines the current error and an optimization function that propagates the error back to the single convolution layers and its learnable parameters.\nWhen CNNs became state of the art in computer vision~, they were usually built using a few alternating convolutional and pooling layers having a few fully connected layers in the end.\nIt turned out that better results are achieved with deeper networks and so the number of layers increased~ over the years.\nTo deal with these deeper networks, new architectures had to be developed.\nIn a first step, to reduce the number of parameters, the convolutional layers with partly large filter kernels were replaced by several layers with small $3 \\times 3$ kernels.\nToday, most architectures are based on the \\emph{network in network} principle~, where more complex modules are used repeatedly. \nExamples of such modules are the \\emph{inception module} from GoogleNet~ or the \\emph{residual block} from ResNet~. \nWhile the inception module consists of multiple parallel strings of layers, the residual blocks are based on the \\emph{highway network}~, which means that they can bypass the original information and the layers in between are just learning residuals.\nWith ResNeXt~ and Inception-ResNet~ there already exist two networks that combine both approaches.\nFor most tasks, it turned out that replacing the fully connected layers by convolutional layers is much more convenient making the networks fully convolutional~.\nThese so-called \\emph{fully convolutional networks}~(FCN) are no longer bound to fixed input dimensions.\nNote that with the availability of convolutional long short-term memory (ConvLSTM) structures also fully convolutional recurrent neural networks (FCRNs) became available for fully scalable sequence-based tasks .\nInside the CNNs, the \\emph{rectified linear unit}~(ReLU) has been the most frequently used activation function for a long time.\nHowever, since this function suffers from problems related to the mapping of all negative values to zero like the vanishing gradient problem, new functions have been introduced in recent years.\nExamples are the \\emph{exponential linear unit}~(ELU), \\emph{swish}~ and the \\emph{non-parametric linearly scaled hyperbolic tangent}~(LiSHT)~.\nIn order to be able to train a network consisting of these different building blocks, \nthe loss function is the most crucial part. This function is responsible for how and what the network ultimately learns and how exactly the training data is applied during the training process to make the network train faster or perform better. So the different classes can be weighted in a classification network with fixed values or so-called $\\alpha$-balancing according to their probability of occurrence. Another interesting approach is weighting training examples according to their easiness for the current network~,~. For multi-task learning also weighting tasks based on their uncertainty~ or gradients~ can be done as further explained in Sec.~\\ref{subsec:architecture:multi_task_networks}.\nA closer look on how a modification of the loss function might affect safety-related aspects is given in Sec.~\\ref{subsec:robust_training:modification_of_loss}.","id":"0c7b2cb4-747c-4997-8a70-14bcffc01831","level":"subsection","origin_cites_number":17,"parent_id":"f0f7d7d0-2bd8-48ae-bdb1-126b96887d72","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Architecture"],["subsection","Building Blocks"]],"subsections":[],"title":"Building Blocks"},{"cite_extract_rate":0.25,"cites":[324,8226,8225,8227],"content":"\\label{subsec:architecture:multi_task_networks}\n\\sectionauthor{Marvin Klingner\\textsuperscript{15}, Varun Ravi-Kumar\\textsuperscript{4}, Timo Sämann\\textsuperscript{4}, Gesina Schwalbe\\textsuperscript{3}}\n\\emph{Multi-task learning} (MTL) in the context of neural networks describes the process of optimizing several tasks simultaneously by learning a unified feature representation and coupling the task-specific loss contributions, thereby enforcing cross-task consistency .\\par\nUnified feature representation is usually implemented by sharing the parameters of the initial layers inside the encoder (also called feature extractor). It not only improves the single tasks by more generalized learned features but also reduces the demand for computational resources at inference. Not an entirely new network has to be added for each task but only a task-specific decoder head. It is essential to consider the growing amount of visual perception tasks in autonomous driving, e.g., depth estimation, semantic segmentation, motion segmentation, and object detection. While the parameter sharing can be soft, as in \\emph{cross stitch} and \\emph{sluice networks} , or hard , meaning ultimately sharing the parameters, the latter is usually preferred due to its straightforward implementation and lower computational complexity during training and inference.\\par\nCompared to implicitly coupling tasks via a shared feature representation, there are often more direct ways to optimize the tasks inside cross-task losses jointly. It is only made possible as, during MTL, there are network predictions for several tasks, which can be enforced to be consistent. As an example, sharp depth edges should only be at class boundaries of semantic segmentation predictions. Often both approaches to MTL are applied simultaneously~ to improve a neural network's performance as well as to reduce its computational complexity at inference.\\par\nWhile the theoretical expectations for MTL are quite clear, it is often challenging to find a good weighting strategy for all the different loss contributions as there is no theoretical basis on which one could choose such a weighting with early approaches either involving heuristics or extensive hyperparameter tuning. The easiest way to balance the tasks is to use uniform weight across all tasks. However, the losses from different tasks usually have different scales, and uniformly averaging them suppresses the gradient from tasks with smaller losses. Addressing these problems, Kendall et al. propose to weigh the loss functions by the \\emph{homoscedastic uncertainty} of each task. One does not need to tune the weighting parameters of the loss functions by hand, but they are adapted automatically during the training process. Concurrently Chen et al.~ propose \\emph{GradNorm}, which does not explicitly weigh the loss functions of different tasks but automatically adapts the gradient magnitudes coming from the task-specific network parts on the backward pass. Liu et al. proposed dynamic weight average (DWA), which uses an average of task losses over time to weigh the task losses.","id":"9fd0032b-a9e0-4b5c-ae60-155c9130c8d4","level":"subsection","origin_cites_number":16,"parent_id":"f0f7d7d0-2bd8-48ae-bdb1-126b96887d72","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Architecture"],["subsection","Multi-Task Networks"]],"subsections":[],"title":"Multi-Task Networks"},{"cite_extract_rate":0.9090909090909091,"cites":[97,875,8229,871,305,870,872,543,544,8228],"content":"\\label{subsec:architecture:neural_architecture_search}\n\\sectionauthor{Patrick Feifel\\textsuperscript{8}, Seyed Eghbal Ghobadi\\textsuperscript{8}}\nIn the previous sections we saw manually engineered modifications of existing CNN architectures proposed by ResNet or Inception . They are results of human design and showed their ability to improve performance. ResNet introduces a \\emph{skip connection} in building blocks and Inception makes use of its specific \\emph{inception module}. Hereby, the intervention by an expert is crucial. The approach of \\emph{neural architecture search} (NAS) aims to automate this time-consuming and manual design of neural network architectures. \n\\par\nNAS is closely related to hyperparameter optimization (HO), which is described in \\refsec{subsec:robust_training:hyperparameter_optimization}. Originally, both tasks were solved simultaneously. Consequently, the kernel size or number of filters were seen as additional hyperparamters. Nowadays, the distinction between HO and NAS should be stressed. The concatenation of complex building blocks or modules cannot be accurately described with single parameters. This simplification is no longer suitable. \n\\par\nTo describe the NAS process, the authors of define three steps: (1) definition of search space, (2) search strategy and (3) performance estimation strategy. \n\\par\nThe majority of search strategies take advantage of the \\emph{NASNet search space} which arranges various operations, \\eg convolution, pooling within a single cell. However, other spaces based on a chain or multi-branch structure are possible . The search strategy comprises advanced methods from \nsequential model-based optimization (SMBO) ,\nBayesian optimization , \nevolutionary algorithms , \nreinforcement learning and \ngradient descent . \nFinally, the performance estimation describes approximation techniques due to the impracticability of multiple evaluation runs. For a comprehensive survey regarding the NAS process we refer to . \n\\par\nRecent research has shown that reinforcement learning approaches such as NASNet-A and ENAS are partly outperformed by evolutionary algorithms, \\eg AmoebaNet and gradient-based approaches, \\eg DARTS . \n\\par\nEach of these approaches focuses on different optimization aspects. Gradient-based methods are applied to a continuous search space and offer faster optimization. On the contrary, the evolutionary approach LEMONADE enables multi-object optimization by considering the conjunction of resource consumption and performance as the two main objectives. Furthermore, single-path NAS extends the multi-path approach of former gradient-based methods and proposes the integration of 'over-parameterized superkernels', which significantly reduces memory consumption.\n\\par\nThe focus of NAS is on the optimized combination of humanly predefined CNN elements with respect to objectives such as resource consumption and performance. NAS offers automation, however, the realization of the objectives is strongly limited by the potential of the CNN elements.","id":"19fdc458-c9c5-4f0b-a62d-6dbcbe8fef85","level":"subsection","origin_cites_number":11,"parent_id":"f0f7d7d0-2bd8-48ae-bdb1-126b96887d72","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Architecture"],["subsection","Neural Architecture Search"]],"subsections":[],"title":"Neural Architecture Search"},{"cite_extract_rate":0.8,"cites":[509,8429,1621,1759,8167,206,520,4169],"content":"\\label{sec:compression}\n\\sectionauthor{Serin Varghese\\textsuperscript{7}}\nRecent developments in CNNs have resulted in neural networks being the state-of-the-art in computer vision tasks like image classification~, object detection~ and semantic segmentation~. This is largely due to the increasing availability of hardware computational power and an increasing amount of training data. We also observe a general upwards trend of the complexities of the neural networks along with their improvement in state-of-the-art performance. These CNNs are largely trained on back-end servers with significantly higher computing capabilities. The use of these CNNs in real-time applications are inhibited due to the restrictions on hardware, model size, inference time, and energy consumption. \nThis led to an emergence of a new field in machine learning, commonly termed as model compression. Model compression basically implies reducing the memory requirements, inference times and model size of DNNs to eventually enable the use of neural networks on edge devices. This is tackled by different approaches such as \\emph{network pruning} (identifying weights or filters that are not critical for network performance), \\emph{weight quantizations} (reducing the precision of the weights used in the network), \\emph{knowledge distillation} (a smaller network is trained with the knowledge gained by a bigger network), and \\emph{low-rank factorization} (decomposing a tensor into multiple smaller tensors).\nIn this section, we introduce some of these methods for model compression and discuss in brief the current open challenges and possible research directions with respect to its use in automated driving applications.","id":"e23cf554-7193-4f8f-a54e-3da02ce3d7c2","level":"section","origin_cites_number":10,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Model Compression"]],"subsections":["5ed86e29-ba4e-4bd5-8fc0-724c41e2999e","28ea17f4-7eb4-44e0-85c5-640bd64d35c5"],"title":"Model Compression"},{"cite_extract_rate":0.7333333333333331,"cites":[8234,8236,945,8232,688,8233,8230,6318,8231,6319,8235],"content":"\\label{subsec:compression:pruning}\n\\sectionauthor{Falk Kappel\\textsuperscript{13}, Serin Varghese\\textsuperscript{7}}\nPruning has been used as a systematic tool to reduce the complexity of deep neural networks. The redundancy in DNNs may exist on various levels, such as the individual weights, filters, and even layers. All the different methods for pruning try to take advantage of these available redundancies on various levels. Two of the initial approaches for neural networks proposed weight pruning in the 1990s as a way of systematically damaging neural networks~. As these weight pruning approaches do not aim at changing the structure of the neural network, these approaches are called \\emph{unstructured pruning}. Although there is reduction in the size of the network when it is saved in sparse format, the acceleration depends on the availability of hardware that facilitate sparse multiplications. As pruning filters and complete layers aim at exploiting the available redundancy in the architecture or structure of neural networks, these pruning approaches are called \\emph{structured pruning}. \nPruning approaches can also be broadly classified into: data-dependent and data-independent methods. Data-dependent methods~ make use of the training data to identify filters to prune. Theis \\etal and Molchanov \\etal propose a greedy pruning strategy that identifies the importance of feature maps one at a time from the network and measures the effect of removal of the filters on the training loss. This means that filters corresponding to those feature maps that have least effect on training loss are removed from the network. Within data-independent methods~, the selection of CNN filters to be pruned are based on the statistics of the filter values. Li \\etal proposed a straightforward method to calculate the rank of filters in a CNN. The selection of filters are based on the $\\ell_1$-norm, where the filter with the lowest norm is pruned away. He \\etal employ a LASSO regression-based selection of filters to minimize the least squares reconstruction.\\par\nAlthough the above-mentioned approaches demonstrated that a neural network can be compressed without affecting the accuracy, the effect on robustness is largely unstudied. Dhillon \\etal proposed pruning a subset of activations and scaling up the survivors to show improved adversarial robustness of a network. Lin \\etal quantize the precision of the weights after controlling the Lipschitz constant of layers. This restricts the error propagation property of adversarial pertubations within the neural network. Ye \\etal evaluated the relationship between adversarial robustness and model compression in detail and show that naive compression has a negative effect on robustness. Gui \\etal co-optimize robustness and compression constraints during the training phase and demonstrate improvement in the robustness along with reduction in the model size. However, these approaches have mostly been tested on image classification tasks and on smaller datasets only. Their effectiveness on safety-relevant automated driving tasks such as object detection and semantic segmentation tasks are not studied and remains an open research challenge.","id":"5ed86e29-ba4e-4bd5-8fc0-724c41e2999e","level":"subsection","origin_cites_number":15,"parent_id":"e23cf554-7193-4f8f-a54e-3da02ce3d7c2","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Model Compression"],["subsection","Pruning"]],"subsections":[],"title":"Pruning"},{"cite_extract_rate":0.529411764705882,"cites":[8241,8375,8239,8240,8238,4496,8237,6488,9139],"content":"\\label{subsec:compression:quantization}\n\\sectionauthor{Firas Mualla\\textsuperscript{13}}\nQuantization of a random variable $x$ having a probability density function $f(x)$ is the process of dividing the range of $x$ into intervals, each is represented using a single value (also called reconstruction value), such that the following reconstruction error is minimized:\n\\begin{equation}\n\\label{eq:quantization}\n\\sum_{i=1}^{L} \\int_{b_i}^{b_{i+1}} (q_i - x)^2 f(x) dx,\n\\end{equation}\nwhere $b_i$ is the left-side border of the i-th interval, $q_i$ is its reconstruction value, and $L$ is the number of intervals, \\eg $L = 8$ for a 3-bit quantization. This definition can be extended to multiple dimensions as well. \nQuantization of neural networks has been around since the 1990s , however, with a focus in the early days on improving the hardware implementations of these networks. In the deep learning literature, a remarkable application of quantization combined with unstructured pruning can be found in the approach of \\textit{deep compression} , where 1-dimensional k-means is utilized to cluster the weights per layer and thus finding the $L$ cluster centers ($q_i$ values in (\\ref{eq:quantization})) iteratively. This procedure conforms to an implicit assumption that $f(x)$ has the same spread inside all clusters. Deep compression can reduce the network size needed for image classification by a factor of 35 for AlexNet and a factor of 49 for VGG-16 without any loss in accuracy. However, as pointed out in , these networks from the early deep learning days are over-parameterized and a less impressing compression factor is thus expected when the same technique is applied to lightweight architectures such as MobileNet and SequeezeNet. For instance, considering SqueezeNet (50 times smaller than AlexNet), the compression factor of deep compression without accuracy loss drops to about 10.\nCompared to scalar quantization used in deep compression, there were attempts to exploit the structural information by applying variants of vector quantization of the weights . Remarkably, in the latter (\\ie ), the reconstruction error of the activations (instead of the weights) is minimized in order to find an optimal codebook for the weights, as the ultimate goal of quantization is to approximate the network's output not the network itself. This is performed in a layer-by-layer fashion (as to prevent error accumulation) using activations generated from unlabeled data. \nOther techniques apply variants of so-called ``linear'' \\textit{quantization}, \\ie the quantization staircase has a fixed interval size. This paradigm conforms to an implicit assumption that $f(x)$ in (\\ref{eq:quantization}) is uniform and is thus also called \\textit{uniform quantization}. The uniform quantization is widely applied both in specialized software packages such as the \\texttt{Texas Instruments Deep Learning Library} (automotive boards) and in general-purpose libraries such as the \\texttt{Tensorflow Lite}. The linearity assumption enables practical implementations, as the quantization and dequantization can be implemented using a scaling factor and an intercept, whereas no codebook needs to be stored. In many situations, the intercept can be omitted by employing a symmetric quantization mapping. Moreover, for power of 2 ranges, the scaling ends up being a bitwise shift operator, where quantization and dequantization differ only in the shift direction. It is also straightforward to apply this scheme \\textit{dynamically}, \\ie for each tensor separately using a tensor-specific multiplicative factor. This can be easily applied not only to filters (weight tensors) but also to activation tensors (see for instance ).\nUnless the scale factor in the linear quantization is assumed constant by construction, it is computed based on the statistics of the relevant tensor and can be thus sensitive to outliers. This is known to result in a low precision quantization. In order to mitigate this issue, the original range can be \\textit{clipped} and thus reduced to the most relevant part of the signal. Several approaches are proposed in the literature for finding an optimal clipping threshold: simple percentile analysis of the original range (\\eg clipping 2\\% of the largest magnitude values), minimizing the mean square error between the quantized and original range in the spirit of (\\ref{eq:quantization}) , or minimizing the Kullback-Leibler divergence between the original and the quantized distributions . While the clipping methods trade off large quantization errors of outliers against small errors of inliers , other methods tackle the outliers problem using a different trade-off, see for instance the outlier channel splitting approach in .\nAn essential point to consider when deciding for a quantization approach for a given problem is the allowed or intended interaction with the training procedure. The so-called \\textit{post-training quantization}, \\ie quantization of a pre-trained network, seems to be attractive from a practical point of view: No access to training data is required and the quantization and training toolsets can be independent from each other. On the other hand, the training-aware quantization methods often yield higher inference accuracy and shorter training times. The latter is a serious issue for large complicated models which may need weeks to train on modern GPU clusters.\nThe training-aware quantization can be implemented by inserting fake quantization operators in the computational graph of the forward-pass during training (simulated quantization), whereas the backward pass is done as usual in floating-point resolution . Other approaches go a step further by quantizing the gradients as well. This leads to much lower training time, as the time of the often computationally expensive backward pass is reduced. The gradient's quantization, however, is not directly applicable as it requires the derivative of the quantization function (staircase-like), which is zero almost everywhere. Luckily, this issue can be handled by employing a \\textit{straight-through estimator} (approximating the quantization function by an identity mapping). There are also other techniques proposed recently to mitigate this problem .","id":"28ea17f4-7eb4-44e0-85c5-640bd64d35c5","level":"subsection","origin_cites_number":17,"parent_id":"e23cf554-7193-4f8f-a54e-3da02ce3d7c2","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Model Compression"],["subsection","Quantization"]],"subsections":[],"title":"Quantization"},{"cite_extract_rate":0,"cites":[],"content":"\\label{sec:discussion}\nWe have presented an extensive overview of approaches to effectively handle safety concerns accompanying deep learning: lack of generalization, robustness, explainability, plausibility, and efficiency. \nIt has been described which lines of research we deem prevalent, important, and promising for each of the individual topics and categories into which the presented methods fall. \n\\par\nThe reviewed methods alone will not provide safe ML systems as such -- and neither will their future extensions. This is due to the limitations of quantifying complex real-world contexts. A complete and plausible \\emph{safety argumentation} will, thus, require more than advances in methodology and theoretical understanding of neural network properties and training processes. \nApart from methodological progress, it will be necessary to gain practical experience in using the presented methods to gather evidence for overall secure behavior, using this evidence to construct a tight safety argument, and testing its validity in various situations.\n\\par\nIn particular, each autonomously acting robotic system with state-of-the-art deep-learning-based perception and non-negligible actuation may serve as an object of study and is, in fact, in need of this kind of systematic reasoning before being transferred to widespread use or even market entry. We strongly believe that novel scientific insights, the potential market volume, and public interest will drive the arrival of reliant and trustworthy AI technology.\n\\section*{Acknowledgment}\nThe research leading to these results is funded by the German Federal Ministry for Economic Affairs and Energy within the project “KI Absicherung – Safe AI for Automated Driving”. The authors would like to thank the consortium for the successful cooperation. Furthermore, this research has been funded by the Federal Ministry of Education and Research of Germany as part of the competence center for machine learning ML2R (01IS18038B).\n\\bibliography{main} \n\\bibliographystyle{alpha}\n\\end{document}","id":"3497f966-f76d-439f-b7b6-ad5a513d9e31","level":"section","origin_cites_number":0,"parent_id":"569f1197-c628-4e9d-9db6-9757cb02d82a","prefix_titles":[["title","Inspect, Understand, Overcome:\\\\A Survey of Practical Methods\\\\for AI Safety"],["section","Discussion"]],"subsections":[],"title":"Discussion"}],"string":"[\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\",\n \"id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"level\": \"root\",\n \"origin_cites_number\": 0,\n \"parent_id\": null,\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ]\n ],\n \"subsections\": [\n \"86d6cdb9-2750-4e63-865d-831683dd115b\",\n \"1fe2887d-3162-404c-bb44-c07ac2772367\",\n \"380e4e9a-2d10-49f6-add4-04990cddb902\",\n \"4396c1c3-a3ad-4efa-9f10-99f90060c0b4\",\n \"e47014ad-1aec-4b32-9241-6aab90a7f0dd\",\n \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"71d4e278-f67c-443c-84a9-ff0254362119\",\n \"94174462-51f2-497a-b88f-1ceaadd52aaf\",\n \"f0f7d7d0-2bd8-48ae-bdb1-126b96887d72\",\n \"e23cf554-7193-4f8f-a54e-3da02ce3d7c2\",\n \"3497f966-f76d-439f-b7b6-ad5a513d9e31\"\n ],\n \"title\": \"root\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 8089,\n 6010,\n 8157\n ],\n \"content\": \"\\\\label{sec:intro} \\n\\\\sectionauthor{Sebastian Houben\\\\textsuperscript{1}, Michael Mock\\\\textsuperscript{1}, Timo Sämann\\\\textsuperscript{4}, Gesina Schwalbe\\\\textsuperscript{3}, Joachim Sicking\\\\textsuperscript{1}}\\nIn barely a decade, deep neural networks (DNNs) have revolutionized the field of machine learning by reaching unprecedented, sometimes superhuman, performances on a growing variety of tasks.\\nMany of these neural models have found their way into consumer applications like smart speakers, machine translation engines or content feeds. However, in safety-critical systems, where human life might be at risk, the use of recent DNNs is challenging as various model-immanent insufficiencies are yet difficult to address. \\n\\\\par\\nThis paper summarizes the promising lines of research in how to identify, address, and at least partly mitigate these DNN insufficiencies.\\nWhile some of the reviewed works are theoretically grounded and foster the overall understanding of training and predictive power of DNNs, others provide practical tools to adapt their development, training or predictions. \\nWe refer to any such method as a \\\\emph{safety mechanism} if it addresses one or several safety concerns in a feasible manner.\\nTheir effectiveness in mitigating safety concerns is assessed by \\\\emph{safety metrics} . \\nAs most safety mechanisms target only a particular insufficiency, we conclude that a \\\\emph{holistic safety argumentation} for a complex DNN-based systems will in many cases rely on a variety of safety mechanisms. \\n\\\\par\\nWe structure our review of these mechanisms as follows: Chapter \\\\ref{sec:dataset_optimization} focuses on \\\\emph{dataset optimization} for network training and evaluation. It is motivated by the well-known fact that, in comparison to humans, DNNs perform poorly on data that is structurally different from training data. Apart from insufficient generalization capabilities of these models, the data acquisition process and distributional data shifts over time play vital roles. We survey potential counter-measures, \\\\eg augmentation strategies and outlier detection techniques.\\n\\\\par\\nMechanisms that improve on \\\\emph{robustness} are described in Chapters \\\\ref{sec:robust_training} and \\\\ref{sec:adversarial_attacks}, respectively. They deserve attention as DNNs are generally not resilient to common perturbations and adversarial attacks.\\n\\\\par\\nChapter \\\\ref{sec:interpretability} addresses incomprehensible network behavior and reviews mechanisms that aim at \\\\emph{explainability}, \\\\eg a more transparent functioning of DNNs.\\n\\\\par\\nMoreover, DNNs tend to overestimate their prediction confidence, especially on unseen data. \\nStraightforward ways to estimate prediction confidence yield mostly unsatisfying results.\\nAmong others, this observation fuelled research on more sophisticated \\\\emph{uncertainty estimations} (see Chapter \\\\ref{sec:uncertainty}), \\\\emph{redundancy mechanisms} (see Chapter \\\\ref{sec:aggregation}) and attempts to reach \\\\emph{formal verification} as addressed in Chapter \\\\ref{sec:verification}.\\n\\\\par\\nAt last, many safety-critical applications require not only accurate but also near real-time decisions. This is covered by mechanisms on the DNN \\\\emph{architectural level} (see Chapter \\\\ref{sec:architecture}) and furthermore by \\\\emph{compression} and \\\\emph{quantization} methods (see Chapter \\\\ref{sec:compression}).\\n\\\\par\\nWe conclude this review of mechanism categories with an outlook on the steps to transfer a carefully arranged combination of safety mechanisms into an actual holistic safety argumentation.\",\n \"id\": \"86d6cdb9-2750-4e63-865d-831683dd115b\",\n \"level\": \"section\",\n \"origin_cites_number\": 6,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Introduction\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Introduction\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:dataset_optimization}\\n\\\\sectionauthor{Matthias Rottmann\\\\textsuperscript{5}}\\nThe performance of a trained model inherently relies on the nature of the underlying dataset.\\nFor instance, a dataset with poor variability will hardly result in a model ready for real-world applications.\\nIn order to approach the latter, data selection processes such as corner case selection and active learning are of utmost importance.\\nThese approaches can help to design datasets that contain the most important information, while preventing the so much desired information from getting lost in an ocean of data.\\nFor a given dataset and active learning setups, data augmentation techniques are very common aiming at extracting as much model performance out of the dataset as possible.\\nOn the other hand, safety arguments also require the analysis of how a model behaves on out-of-distribution data, data that contains concepts the model has not encountered during training.\\nThis is quite likely to happen as our world is under constant change, in other words exposed to a constantly growing domain shift.\\nTherefore, these fields are lately gaining interest, also with respect to perception in automated driving.\",\n \"id\": \"1fe2887d-3162-404c-bb44-c07ac2772367\",\n \"level\": \"section\",\n \"origin_cites_number\": 0,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Dataset Optimization\"\n ]\n ],\n \"subsections\": [\n \"36c23394-2c6f-46c4-b022-50c53af7e345\",\n \"f767152d-e919-4dba-8417-ed1faf4da217\",\n \"7b5fda9c-0631-4e3b-bf4f-fb2fb6b0d60d\",\n \"de2e598e-f35b-4b65-945b-25aee539ea0d\",\n \"e4ebc416-15f6-4517-a761-acd4315ad09b\"\n ],\n \"title\": \"Dataset Optimization\"\n },\n {\n \"cite_extract_rate\": 0.6923076923076921,\n \"cites\": [\n 1624,\n 8160,\n 3251,\n 3237,\n 3219,\n 8159,\n 8158,\n 3248,\n 8161\n ],\n \"content\": \"\\\\label{subsec:dataset_optimization:anomaly_detection}\\n\\\\sectionauthor{Sujan Sai Gannamaneni\\\\textsuperscript{1}, Matthias Rottmann\\\\textsuperscript{5}}\\nThe terms anomaly, outlier and \\\\emph{out-of-distribution} (OOD) data detection are often used interchangeably in literature and refer to task of identifying data samples that are not representative of training data distribution. Uncertainty evaluation (\\\\cf Chapter \\\\ref{sec:uncertainty}) is closely tied to this field as self-evaluation of models is one of the active areas of research for OOD detection. In particular, for image classification problems it has been reported that neural networks often produce high confidence predictions on OOD data . The detection of such OOD inputs can either be tackled by post-processing techniques that adjust the estimated confidence or by enforcing low confidence on OOD samples during training . Even guarantees that neural networks produce low confidence predictions for OOD samples can be provided under specific assumptions (\\\\cf ). More precisely, this work utilizes Gaussian mixture models that, however, may suffer from high-dimensional data and require strong assumptions on the distribution parameters.\\nSome approaches use generative models like \\\\emph{GANs} and \\\\emph{autoencoders} for outlier detection. The models are trained to learn in-distribution data manifolds and will produce higher reconstruction loss for outliers.\\n\\\\par\\nFor OOD detection in semantic segmentation, only a few works have been presented so far.\\nAngus \\\\etal present a comparative study of common OOD detection methods, which mostly deal with image-level classification. In addition, they provide a novel setup of relevant OOD datasets for this task. Another work trains a fully convolutional binary classifier that distinguishes image patches from a known set of classes from image patches stemming from an unknown class . The classifier output applied at every pixel will give the per-pixel confidence value for an OOD object. Both of these works perform at pixel level and without any sophisticated feature generation methods specifically tailored for the detection of entire OOD instances.\\nUp to now, outlier detection has not been studied extensively for object detection tasks based on benchmark object detection datasets. In , two CNNs are used to perform object detection and binary classification (benign or anomaly) in a sequential fashion, where the second CNN takes the localized object within the image as input.\\n\\\\par\\nFrom a safety standpoint, detecting outliers or OOD samples is extremely important and beneficial as training data cannot realistically be large enough to capture all situations. Research in this area is heavily entwined with progress in uncertainty estimation (\\\\cf Chapter \\\\ref{sec:uncertainty}) and domain adaptation (\\\\cf \\\\refsec{subsec:dataset_optimization:domains}). Extending research works to segmentation and object detection tasks would be particularly significant for leveraging autonomous driving research. In addition to safety, OOD detection can be beneficial in other aspects like when using local expert models. For example, when using an expert model for segmentation of urban driving scenes and another expert model for segmentation of highway driving scenes, an OOD detector could act as trigger on which models can be switched. \\n\\\\par\\nWith respect to the approaches presented above, uncertainty-based and generative model-based OOD detection methods are currently promising directions of research. However, it remains an open question whether they can unfold their potential well on segmentation and object detection tasks.\",\n \"id\": \"36c23394-2c6f-46c4-b022-50c53af7e345\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"1fe2887d-3162-404c-bb44-c07ac2772367\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Dataset Optimization\"\n ],\n [\n \"subsection\",\n \"Outlier/Anomaly Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Outlier/Anomaly Detection\"\n },\n {\n \"cite_extract_rate\": 0.636363636363636,\n \"cites\": [\n 8162,\n 1044,\n 5527,\n 3237,\n 4620,\n 1043,\n 1049\n ],\n \"content\": \"\\\\label{subsec:dataset_optimization:active_learning}\\n\\\\sectionauthor{Matthias Rottmann\\\\textsuperscript{5}}\\nIt is widely known that, as a rule of thumb, for the training of any kind of artificial neural network, an increase of training data leads to increased performance.\\nObtaining labeled training data, however, is often very costly and time consuming.\\n\\\\emph{Active learning} provides one possible remedy to this problem: Instead of labeling every data point, active learning utilizes a \\\\emph{query strategy} to request labels from a teacher/an oracle which leverage the model performance most.\\nThe survey paper by Settles provides a broad overview regarding query strategies for active learning methods. However, except for \\\\emph{uncertainty sampling} and \\\\emph{query by committee}, most of them seem to be infeasible in deep learning applications up to now. Hence, most of the research activities in active deep learning focus on these two query strategies, as we outline in the following.\\n\\\\par\\nIt has been shown for image classification that labels corresponding to uncertain samples can leverage the networks' performance significantly and that a combination with semi-supervised learning is promising. In both works, uncertainty of unlabeled samples is estimated via Monte Carlo (MC) dropout inference. MC dropout inference and a chosen number of training epochs are executed alternatingly, after performing MC dropout inference, the unlabeled samples' uncertainties are assessed by means of sample-wise dispersion measures. Samples for which the DNN model is very uncertain about its prediction are presented to an oracle and labeled.\\n\\\\par\\nWith respect to object detection, a moderate number of active learning methods has been introduced .\\nThese approaches include uncertainty sampling and query-by-committee methods . In , additional algorithmic features specifically tailored for object detection networks are presented, \\\\ie separate treatment of the localization and classification loss , as well as weak and strong supervision schemes . For semantic segmentation, an uncertainty-sampling-based approach has been presented , which queries polygone masks for image sections of a fixed size ($128 \\\\times 128$). Queries are performed by means of accumulated entropy in combination with a cost estimation for each candidate image section.\\n\\\\par\\nRecently, new methods for estimating the quality of a prediction as well as new uncertainty quantification approaches, \\\\eg gradient-based ones , have been proposed. It remains an open question whether they are suitable for active learning. Since most of the conducted studies are rather of academic nature, also their applicability to real-life data acquisition is not yet demonstrated sufficiently. In particular, it is not clear whether the proposed active learning schemes, including the label acquistion, for instance in semantic segmentation, is suitable to be performed by human labelers. Therefore, labeling acquisition with a common understanding of the labelers' convenience and suitability for active learning are a promising direction for research and development.\",\n \"id\": \"f767152d-e919-4dba-8417-ed1faf4da217\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"1fe2887d-3162-404c-bb44-c07ac2772367\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Dataset Optimization\"\n ],\n [\n \"subsection\",\n \"Active Learning\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Active Learning\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 330,\n 8163,\n 3267,\n 5113,\n 328,\n 8495\n ],\n \"content\": \"\\\\label{subsec:dataset_optimization:domains}\\n\\\\sectionauthor{Julia Rosenzweig\\\\textsuperscript{1}}\\nThe classical assumption in machine learning is that the training and testing data sets are drawn from the same distribution, implying that the model is deployed under the same conditions as it was trained under. However, as mention, in real-world applications this assumption is often violated in the sense that the training and the testing set stem from different domains having different distributions. This poses difficulties for statistical models and the performance will mostly degrade when they are deployed on a domain $D^{\\\\mathrm{test}}$, having a different distribution than the training dataset (\\\\ie generalizing from the training to the testing domain is not possible). This makes the study of domains not only relevant from the machine learning perspective, but also from a safety point of view.\\n\\\\par\\nMore formally, there are differing notions of a 'domain' in literature. For , a domain $\\\\mathcal{D} = \\\\{\\\\mathcal{X}, P(X) \\\\}$ consists of a feature space $\\\\mathcal{X} \\\\subset \\\\mathbb{R}^d$ together with a marginal probability distribution $P(X)$ with $X\\\\in \\\\mathcal{X}$. In , a domain is a pair consisting of a distribution over the inputs together with a labeling function.\\nHowever, instead of a sharp labeling function, it is also widely accepted to define a (training) domain $\\\\mathcal{D} = \\\\{(x_i,y_i)\\\\}_{i=1}^n$ to consist of $n$ (labeled) samples that are sampled from a joint distribution $P(x,y)$ (\\\\cf ). \\n\\\\par\\nThe reasons for \\\\emph{distributional shift} are diverse---as are the names to indicate a shift. For example, if the rate of (class) images of interest is different between training and testing set this can lead to a domain gap and, \\\\eg result in differing overall error rates. Moreover, as mentions, changing weather conditions and camera setups in cars lead to a domain mismatch in applications of autonomous driving.\\nIn biomedical image analysis, different imaging protocols and diverse anatomical structures can hinder generalization of trained models (\\\\cf ).\\nCommon terms to indicate distributional shift are \\\\emph{domain shift, dataset shift, covariate shift, concept drift, domain divergence, data fracture, changing environments} or \\\\emph{dataset bias}. References provide an overview. \\n\\\\par\\nMethods and measures to overcome the problem of domain mismatch between one or more (\\\\cf ) source domains and target domain(s) and the resulting poor model performance are studied in the field of transfer learning and in particular its subtopic domain adaptation (\\\\cf ).\\nFor instance, adapting a model that is trained on synthetically generated data to work on real data is one of the core challenges, as can be seen .\\nFurthermore, detecting when samples are out-of-domain or out-of-distribution is an active field of research (\\\\cf and the outlier/anomaly detection in \\\\refsec{subsec:dataset_optimization:anomaly_detection} as well as the topic of observers in the black-box methods in \\\\refsec{subsec:verification:black_box_methods} for further reference). This is particularly relevant for machine learning models that operate in the real world: If, \\\\eg an autonomous vehicle encounters some situation that deviates strongly from what was seen during training (\\\\eg due to some special event like a biking competition, carnival, etc.) this can lead to wrong predictions and thereby potential safety issues if not detected in time.\",\n \"id\": \"7b5fda9c-0631-4e3b-bf4f-fb2fb6b0d60d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"1fe2887d-3162-404c-bb44-c07ac2772367\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Dataset Optimization\"\n ],\n [\n \"subsection\",\n \"Domains\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Domains\"\n },\n {\n \"cite_extract_rate\": 0.7333333333333331,\n \"cites\": [\n 8164,\n 858,\n 8165,\n 5980,\n 5552,\n 90,\n 2083,\n 5548,\n 7022,\n 5553,\n 62\n ],\n \"content\": \"\\\\label{subsec:dataset_optimization:augmentation}\\n\\\\sectionauthor{Falk Kappel\\\\textsuperscript{13}}\\nGiven the need for big amounts of data to train neural networks, one often runs into a situation where data is lacking. This can lead to insufficient generalization and an overfitting to the training data. An overview over different techniques to tackle this challenge can be found in . One approach to try and overcome this issue is the augmentation of data. It aims at optimizing available data and increasing its amount, curating a dataset that represents a wide variety of possible inputs during deployment. Augmentation can as well be of help when having to work with a heavily unbalanced dataset by creating more samples of underrepresented classes. A broad survey on data augmentation is provided by . They distinguish between two general approaches to data augmentation with the first one being data warping augmentations that focus on taking existing data and transforming it in a way that does not effect labels. The other option are oversampling augmentations, which create synthetic data that can be used to increase the size of the dataset.\\\\newline\\nExamples of some of the most basic augmentations are flipping, cropping, rotating, translating, shearing and zooming. These are affecting the geometric properties of the image and are easily implemented . The machine learning toolkit \\\\texttt{Keras}, for example, provides an easy way of applying them to data using their \\\\texttt{ImageDataGenerator} class . Other simple methods include adaptations in color space that affect properties such as lighting, contrast and tints, which are common variations within image data. Filters can be used to control increased blur or sharpness . In random erasing is introduced as a method with similar effect as cropping, aiming at gaining robustness against occlusions. An example for mixing images together as an augmentation technique can be found in . \\\\newline\\nThe abovementioned methods have in common that they work on the input data but there are different approaches that make use of deep learning for augmentation. An example for making augmentations in feature space using autoencoders can be found in . They use the representation generated by the encoder and generate new samples by interpolation and extrapolation between existing samples of a class. The lack of interpretability of augmentations in feature space in combination with the tendency to perform worse than augmentations in image space present open challenges for those types of augmentations . Adversarial training is another method that can be used for augmentation. The goal of adversarial training is to discover cases that would lead to wrong predictions. That means the augmented images won't necessarily represent samples that could occur during deployment but that can help in achieving more robust decision boundaries . An example of such an approach can be found in . Generative modelling can be used to generate synthetic samples that enlarge the dataset in a useful way with GANs, variational autoencoders and the combination of both are important tools in this area . Examples for data augmentation in medical context using a CycleGAN can be found in and using a progressively growing GAN in . Next to neural style transfer that can be used to change the style of an image to a target style, AutoAugment and population based augmentation are two more interesting publications. In the latter two, the idea is to search a predefined search space of augmentations to gather the best selection.\",\n \"id\": \"de2e598e-f35b-4b65-945b-25aee539ea0d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 15,\n \"parent_id\": \"1fe2887d-3162-404c-bb44-c07ac2772367\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Dataset Optimization\"\n ],\n [\n \"subsection\",\n \"Augmentation\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Augmentation\"\n },\n {\n \"cite_extract_rate\": 0.555555555555555,\n \"cites\": [\n 8133,\n 8166,\n 7698,\n 1636,\n 3500\n ],\n \"content\": \"\\\\label{subsec:dataset_optimization:corner_case_detection}\\n\\\\sectionauthor{Alexander Pohl\\\\textsuperscript{16}, Marco Hoffmann\\\\textsuperscript{16}, Michael Mlynarski\\\\textsuperscript{16}, Timo Sämann\\\\textsuperscript{4}}\\nEnsuring that AI-based applications behave correctly and predictably even in unexpected or rare situations is a major concern that gains importance especially in safety-critical applications such as autonomous driving. In the pursuit of more robust AI corner cases play an important role. \\n\\\\par\\nThe meaning of the term corner case varies in the literature. Some consider mere erroneous or incorrect behavior as corner cases . For example, in corner cases are referred to as situations in which an object detector fails to detect relevant objects at relevant locations. Others characterize corner cases mainly as rare combinations of input parameter values . This project adopts the first definition: Inputs that result in unexpected or incorrect behaviour of the AI function are defined as corner cases. \\n\\\\par\\nContingent on the hardware, the AI architecture and the training data, the search space of corner cases quickly becomes incomprehensibly large. While manual creation of corner cases (\\\\eg constructing or re-enacting scenarios) might be more controllable, approaches that scale better and allow for a broader and more systematic search for corner cases require extensive automation.\\n\\\\par\\nOne approach to automatic corner case detection is based on transforming the input data. The \\\\emph{DeepTest} framework uses three types of image transformations: linear, affine and convolutional transformations. In addition to these transformations, metamorphic relations help detect undesirable behaviors of deep learning systems. They allow changing the input while asserting some characteristics of the result . For example, changing the contrast of input frames should not affect the steering angle of a car . Input-output pairs that violate those metamorphic relations can be considered as corner cases.\\n\\\\par\\nAmong other things, the white-box testing framework \\\\emph{DeepXplore} applies a method called \\\\emph{gradient ascent} to find corner cases (\\\\cf \\\\refsec{subsec:verification:formal_testing}). In the experimental evaluation of the framework, three variants of deep learning architectures were used to classify the same input image. The input image was then changed according to the gradient ascent of an objective function that reflected the difference in the resulting class probabilities of the three model variants. When the changed (now artificial) input resulted in different class label predictions by the model variants, the input was considered as a corner case. \\n\\\\par\\nIn , corner cases are detected on video sequences by comparing predicted with actual frames. The detector has three components: The first component, semantic segmentation, is used to detect and locate objects in the input frame. As the second component, an image predictor trained on frame sequences predicts the actual frame based on the sequence preceding that frame. An error is determined by comparing the actual with the predicted (\\\\ie expected) frame, following the idea that only situations that are unexpected for AI-based perception functions may be potentially dangerous and therefore a corner case. Both the segmentation and the prediction error are then fed into the third component of the detector, which determines a corner case score that reflects the extent to which unexpected relevant objects are at relevant locations.\\n\\\\par\\nIn , a corner case detector based on simulations in a \\\\emph{Carla} environment is presented. In the simulated world, AI agents control the vehicles. During simulations, state information of both the environment and the AI agents are fed into the corner case detector. While the environment provides the real vehicle states, the AI agents provide estimated and perceived state information. Both sources are then compared to detect conflicts (\\\\eg collisions). These conflicts are recorded for analysis. \\n\\\\par\\nSeveral ways of automatically generating and detecting corner cases exist. However, corner case detection is a task with challenges of its own: Depending on the operational domain including its boundaries, the space of possible inputs can be very large. Also, some types of corner cases are specific to the AI architecture, \\\\eg the network type or the network layout used. Thus, corner case detection has to assume a holistic point of view on both model and input, adding further complexity and reducing transferability of previous insights. \\n\\\\par\\nAlthough it can be argued that rarity does not necessarily characterize corner cases, rare input data might have the potential of challenging the AI functionality (\\\\cf \\\\refsec{subsec:dataset_optimization:anomaly_detection}).\\nAnother research direction could investigate whether structuring the input space in a way suitable for the AI functionality supports the detection of corner cases. Provided that the operational domain is conceptualized as an ontology, ontology-based testing may support automatic detection. A properly adapted generator may specifically select promising combinations of extreme parameter values and, thus, provide valuable input for synthetic test data generation.\",\n \"id\": \"e4ebc416-15f6-4517-a761-acd4315ad09b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 9,\n \"parent_id\": \"1fe2887d-3162-404c-bb44-c07ac2772367\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Dataset Optimization\"\n ],\n [\n \"subsection\",\n \"Corner Case Detection\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Corner Case Detection\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 314,\n 509,\n 1606,\n 8429,\n 1621,\n 1759,\n 8167,\n 8168,\n 8169,\n 206,\n 520,\n 4169\n ],\n \"content\": \"\\\\label{sec:robust_training}\\n\\\\sectionauthor{Nikhil Kapoor\\\\textsuperscript{7}}\\nRecent works~ have shown that state-of-the-art deep neural networks (DNNs) performing a wide variety of computer vision tasks such as image classification~, object detection~ and semantic segmentation~ are not robust to small changes in the input.\\n\\\\par\\nRobustness of neural networks is an active and open research field that can be considered highly relevant for achieving safety in autonomous driving. Currently, most of the research is directed towards either improving adversarial robustness~ (robustness against carefully designed perturbations that aim at causing misclassifications with high confidence), or improving corruption robustness~ (robustness against commonly occurring augmentations such as weather changes, addition of Gaussian noise, photometric changes, etc.). While adversarial robustness might be more of a security issue than a safety issue, corruption robustness, on the other hand, can be considered highly safety-relevant.\\n\\\\par\\nEquipped with these definitions, we broadly term \\\\emph{robust training} here as methods or mechanisms that aim at improving either adversarial or corruption robustness of a DNN, by incorporating modifications into the architecture or into the training mechanism itself.\",\n \"id\": \"380e4e9a-2d10-49f6-add4-04990cddb902\",\n \"level\": \"section\",\n \"origin_cites_number\": 18,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Robust Training\"\n ]\n ],\n \"subsections\": [\n \"62315f40-21bf-40c7-afd5-00e8d80aa4e1\",\n \"cf7b5f9f-caf9-44e0-b472-2124535b4d7a\",\n \"e4ab90c2-e9e5-49ce-a3ae-b1dfdb4472d4\"\n ],\n \"title\": \"Robust Training\"\n },\n {\n \"cite_extract_rate\": 0.7142857142857141,\n \"cites\": [\n 868,\n 8171,\n 8173,\n 8172,\n 8170\n ],\n \"content\": \"\\\\label{subsec:robust_training:hyperparameter_optimization}\\n\\\\sectionauthor{Seyed Eghbal Ghobadi\\\\textsuperscript{8}, Patrick Feifel\\\\textsuperscript{8}}\\nThe final performance of a neural network depends highly on the learning process. The process includes the actual optimization and may additionally introduce training methods such as dropout, regularization, or parametrization of a multi-task loss. \\n\\\\par\\nThese methods adapt their behavior for predefined parameters. Hence, their optimal configuration is a priori unknown. We refer to them as \\\\emph{hyperparameters}. Important hyperparameters comprise, for instance, the initial learning rate, steps for learning rate reduction, learning rate decay, momentum, batch size, dropout rate and number of iterations. Their configuration has to be determined according to the architecture and task of the CNN . The search of an optimal hyperparameter configuration is called hyperparameter optimization (HO).\\n\\\\par\\nHO is usually described as an optimization problem . Thereby, the combined configuration space is defined as \\n$\\\\boldsymbol{\\\\Lambda} = \\\\lambda_1 \\\\times \\\\lambda_2 \\\\times \\\\cdots \\\\lambda_N$,\\naccording to each domain $\\\\lambda_n$. Their individual spaces can be continuous, discrete, categorical or binary. \\n\\\\par\\nHence, we aim to find an optimal hyperparameter configuration $\\\\lambda^{\\\\star}$ by minimizing an objective function $\\\\mathcal{O}\\\\left ( \\\\right )$, which evaluates a trained model $\\\\mathcal{M}$ on the validation dataset $\\\\mathcal{D}^{\\\\textrm{val}}$ with the loss $\\\\mathcal{L}$:\\n\\\\begin{equation}\\n\\\\lambda^{\\\\star} = \\n\\\\text{arg min}_{\\\\lambda \\\\in \\\\boldsymbol{\\\\Lambda}} \\\\; \\n\\\\mathcal{O} \\\\left( \\\\mathcal{L}, \\\\mathcal{M}_{\\\\lambda}, \\\\mathcal{D}^{\\\\textrm{train}}, \\\\mathcal{D}^{\\\\textrm{val}} \\\\right)\\n\\\\end{equation}\\nThis problem statement is widely regarded in traditional machine learning and primarily based on Bayesian optimization (BO) in combination with Gaussian processes. However, a straightforward application to deep neural networks encounters problems due to a \\\\emph{lack of scalability, flexibility and robustness} , . \\n\\\\par\\nTo exploit the benefits of BO, many authors proposed different combinations with other approaches. \\nHyperband in combination with BO (BOHB) frames the optimization as ``... a pure exploration non-stochastic infinite-armed bandit problem ...''. \\nThe method of BO for iterative learning (BOIL) internalizes iteratively collected information about the learning curve and the learning algorithm itself. \\nThe authors of introduce the trace-aware knowledge gradient (taKG) as an acquisition function for BO (BO-taKG) which ``leverages both trace information and multiple fidelity controls''.\\nThereby BOIL and BO-taKG achieve state-of-research performance regarding CNNs outperforming Hyperband.\\n\\\\par\\nOther approaches such as the orthogonal array tuning method (OATM) or HO by reinforcement learning (Hyp-RL) turn away from the Bayesian approaches and offer new research directions.\\n\\\\par\\nFinally, the insight that many authors include kernel sizes and number of kernels and layers in their hyperparameter configuration should be emphasized. More work should be spent on the distinct integration of HO in the performance estimation strategy of neural architecture search (\\\\cf \\\\refsec{subsec:architecture:neural_architecture_search}).\",\n \"id\": \"62315f40-21bf-40c7-afd5-00e8d80aa4e1\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 7,\n \"parent_id\": \"380e4e9a-2d10-49f6-add4-04990cddb902\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Robust Training\"\n ],\n [\n \"subsection\",\n \"Hyperparameter Optimization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Hyperparameter Optimization\"\n },\n {\n \"cite_extract_rate\": 0.42857142857142805,\n \"cites\": [\n 8179,\n 8177,\n 958,\n 8176,\n 8174,\n 8167,\n 8175,\n 8180,\n 8178\n ],\n \"content\": \"\\\\label{subsec:robust_training:modification_of_loss}\\n\\\\sectionauthor{Nikhil Kapoor\\\\textsuperscript{7}}\\nThere exist many approaches that aim at directly modifying the loss function with an objective of improving either adversarial or corruption robustness~. One of the earliest approaches for improving corruption robustness was introduced by Zheng \\\\etal called \\\\emph{stability training}, where they introduce a regularization term that penalizes the network prediction to a clean and an augmented image. However, their approach does not scale to many augmentations at the same time. Janocha \\\\etal then introduced a detailed analysis on the influence of multiple loss functions to model performance as well as robustness and suggested that expectation-based losses tend to work better with noisy data and squared-hinge losses tend to work better for clean data. Other well-known approaches are mainly based on variations of data augmentation~, which can be computationally quite expensive. \\n\\\\par\\nIn contrast to corruption robustness, there exist many more approaches based on adversarial examples. We highlight some of the most interesting and relevant ones here. Mustafa \\\\etal proposes to add a loss term that maximally separates class-wise feature map representations, hence increasing the distance from data points to the corresponding decision boundaries. Similarly, Pang \\\\etal proposed the Max-Mahalanobis center (MMC) loss to learn more structured representations and induce high-density regions in the feature space. Chen \\\\etal proposed a variation of the well-known cross entropy (CE) loss that not only maximizes the model probabilities of the correct class, but in addition, also minimizes model probabilities of incorrect classes. Cisse \\\\etal constraints the Lipschitz constant of different layers to be less than one which restricts the error propagation introduced by adversarial perturbations to a DNN. Dezfooli \\\\etal proposed to minimize the curvature of the loss surface locally around data points. They emphasize that there exists a strong correlation between locally small curvature and correspondingly high adversarial robustness. \\n\\\\par\\nAll of these methods highlighted above are evaluated mostly for image classification tasks on smaller datasets, namely CIFAR-10~, CIFAR-100~, SVHN~, and only sometimes on ImageNet~. Very few approaches have been tested rigorously on complex safety-relevant tasks such as \\\\emph{object detection} and \\\\emph{semantic segmentation}, etc. Moreover, methods that improve adversarial robustness are only tested on a small subset of attack types under differing attack specifications. This makes comparing multiple methods difficult. \\n\\\\par\\nIn addition, methods that improve corruption robustness are evaluated over a standard data set of various corruption types which may or may not be relevant to its application domain. In order to assess multiple methods for their effect on safety-related aspects, a thorough robustness evaluation methodology is needed, which is largely missing in the current literature. This evaluation would need to take into account relevant disturbances/corruption types present in the real world (application domain) and had to assess robustness towards such changes in a rigorous manner. Without such an evaluation, we run the risk of being overconfident in our network, thereby harming safety.\",\n \"id\": \"cf7b5f9f-caf9-44e0-b472-2124535b4d7a\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 21,\n \"parent_id\": \"380e4e9a-2d10-49f6-add4-04990cddb902\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Robust Training\"\n ],\n [\n \"subsection\",\n \"Modification of Loss\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Modification of Loss\"\n },\n {\n \"cite_extract_rate\": 0.5666666666666661,\n \"cites\": [\n 2800,\n 5073,\n 7109,\n 3624,\n 2690,\n 8589,\n 2761,\n 2686,\n 8181,\n 8182,\n 3655,\n 8588,\n 2705,\n 1695,\n 1256,\n 8183,\n 2793\n ],\n \"content\": \"\\\\sectionauthor{Firas Mualla\\\\textsuperscript{13}}\\nDomain generalization (DG) can be seen as an extreme case of \\\\emph{domain adaptation} (DA). The latter is a type of transfer learning, where the source and target tasks are the same (\\\\eg shared class labels) but the source and target domains are different (\\\\eg another image acquisition protocol or a different background) . \\nThe DA can be either supervised (SDA), where there is little available labeled data in the target domain, or unsupervised (UDA), where data in the target domain is not labeled. The DG goes one step further by assuming that the target domain is entirely unknown. Thus, it seeks to solve the train-test domain shift in general. While DA is already an established line of research in the machine learning community, DG is relatively new , though with an extensive list of papers in the last few years. \\n\\\\par\\nProbably, the first intuitive solution that one may think of to implement DG is neutralizing the domain-specific features. It was shown in that the gray-level co-occurrence matrices (GLCM) tend to perform poorly in semantic classification (\\\\eg digit recognition) but yield good accuracy in textural classification compared to other feature sets such as SURF and LBP. DG was thus implemented by decorrelating the model's decision from the GLCM features of the input image even without the need of domain labels.\\nBesides the aforementioned intensity-based statistics of an input image, it is known that characterizing image style can be done based on the correlations between the filter responses of a DNN layer (neural style transfer). In , the training images are enriched with stylized versions, where a style is defined either by an external style (\\\\eg cartoon or art) or by an image from another domain. Here, DG is addressed as a \\\\emph{data augmentation} problem.\\nSome approaches try to learn generalizable latent representations by a kind of adversarial training. This is done by a generator or an encoder, which is trained to generate a hidden feature space that maximizes the error of a domain discriminator but at the same time minimizes the classification error of the task of concern. Another flavor of adversarial training can be seen in , where an adversarial autoencoder is trained to generate features, which a discriminator cannot distinguish from random samples drawn from a prior Laplace distribution. This regularization prevents the hidden space from overfitting to the source domains, in a similar spirit to how variational autoencoders do not leave gaps in the latent space. In , it is argued that the domain labels needed in such approaches are not always well-defined or easily available. Therefore they assume unknown latent domains which are learned by clustering in a space similar to the style-transfer features mentioned above. The pseudo labels resulting from clustering are then used in the adversarial training.\\nAutoencoders have been employed for DG not only in an adversarial setup, but also in the sense of \\\\emph{multi-task learning} nets , where the classification task in such nets is replaced by a reconstruction one. In , an autoencoder is trained to reconstruct not only the input image but also the corresponding images in the other domains.\\nIn the core of both DA and DG we are confronted with a distribution matching problem. However, estimating the probability density in high-dimensional spaces is intractable. The density-based metrics such as Kullback-Leibler divergence are thus not directly applicable. In statistics, the so-called \\\\emph{two-samples tests} are usually employed to measure the distance between two distributions in a point-wise manner, \\\\ie without density estimation. For deep learning applications, these metrics need not only to be point-wise but also differentiable. The two-samples tests were approached in the machine learning literature using (differentiable) K-NNs , classifier two-samples tests (C2ST) , or based on the theory of kernel methods . More specifically, the \\\\emph{maximum mean discrepancy} (MMD) , which belongs to the kernel methods, is widely used for DA but also for DG . Using the MMD, the distance between two samples is estimated based on pairwise kernel evaluations, \\\\eg the radial basis function (RBF) kernel.\\nWhile the DG approaches generalize to domains from which zero shots are available, the so-called \\\\emph{zero shot learning} (ZSL) approaches generalize to tasks (\\\\eg new classes in the same source domains) for which zero shots are available. Typically, the input in ZSL is mapped to a semantic vector per class instead of a simple class label. This can be, for instance, a vector of visual attributes or a word embedding of the class name . A task (with zero shots at training time) can be then described by a vector in this space. In , there is an attempt to combine ZSL and DG in the same framework in order to generalize to new domains as well as new tasks, which is also referred to as\\n \\\\emph{heterogeneous domain generalization}.\\nNote that most discussed approaches for DG require non-standard handling, \\\\ie modifications to models, data, and/or the optimization procedure. This issue poses a serious challenge as it limits the practical applicability of these approaches. There is a line of research which tries to address this point by linking DG to other machine learning paradigms, especially the model-agnostic meta-learning (MAML) algorithm, in an attempt to apply DG in a model-agnostic way. Loosely speaking, a model can be exposed to simulated train-test domain shift by training on a small \\\\emph{support set} to minimize the classification error on a small \\\\emph{validation set}. This can be seen as an instance of a \\\\emph{few shot learning} (FSL) problem . Moreover, the procedure can be repeated on other (but related) FSL tasks (\\\\eg different classes) in what is known as \\\\emph{episodic training}. The model transfers its knowledge from one task to another task and learns how to learn fast for new tasks. This can be thus seen as a \\\\emph{meta-learning} objective (in a FSL setup). Since the goal of DG is to adapt to new domains rather than new tasks, several model-agnostic approaches try to recast this procedure in a DG setup.\",\n \"id\": \"e4ab90c2-e9e5-49ce-a3ae-b1dfdb4472d4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 30,\n \"parent_id\": \"380e4e9a-2d10-49f6-add4-04990cddb902\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Robust Training\"\n ],\n [\n \"subsection\",\n \"Domain Generalization\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Domain Generalization\"\n },\n {\n \"cite_extract_rate\": 0.7272727272727271,\n \"cites\": [\n 314,\n 900,\n 888,\n 6146,\n 8396,\n 892,\n 917,\n 906\n ],\n \"content\": \"\\\\label{sec:adversarial_attacks}\\n\\\\sectionauthor{Andreas Bär\\\\textsuperscript{15}}\\nOver the last few years, deep neural networks (DNNs) consistently showed state-of-the-art performance across several vision-related tasks.\\nWhile their superior performance on clean data is indisputable, they show a lack of robustness to certain input patterns, denoted as \\\\emph{adversarial examples} .\\nIn general, an algorithm for creating adversarial examples is referred to as an \\\\emph{adversarial attack} and aims at fooling an underlying DNN, such that the output changes in a desired and malicious way.\\nThis can be carried out without any knowledge about the DNN to be attacked (black-box attack) , or with full knowledge about the parameters, architecture, or even training data of the respective DNN (white-box attack) .\\nWhile initially being applied on simple classification tasks, some approaches aim at finding more realistic attacks , which particularly pose a threat to safety-critical applications, such as DNN-based environment perception systems in autonomous vehicles.\\nAltogether, this motivated the research in finding ways of defending against such adversarial attacks .\\nIn this section, we introduce the current state of research regarding adversarial attacks in general, more realistic adversarial attacks closely related to the task of environment perception for autonomous driving, and strategies for detecting or defending adversarial attacks.\\nWe conclude each section by clarifying current challenges and research directions.\",\n \"id\": \"4396c1c3-a3ad-4efa-9f10-99f90060c0b4\",\n \"level\": \"section\",\n \"origin_cites_number\": 11,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Adversarial Attacks\"\n ]\n ],\n \"subsections\": [\n \"945be913-0c94-4b59-92f5-5fa8a6f0bf30\",\n \"7e154b91-68e0-47d3-8c11-92f8eb25b769\"\n ],\n \"title\": \"Adversarial Attacks\"\n },\n {\n \"cite_extract_rate\": 0.6341463414634141,\n \"cites\": [\n 923,\n 912,\n 8184,\n 892,\n 917,\n 908,\n 8185,\n 890,\n 914,\n 7313,\n 975,\n 314,\n 953,\n 954,\n 1624,\n 920,\n 8186,\n 6140,\n 919,\n 899,\n 902,\n 3267,\n 8396,\n 3234,\n 906,\n 6148\n ],\n \"content\": \"\\\\label{subsec:adverarial_attacks:adversarial_attacks_defenses}\\n\\\\sectionauthor{Andreas Bär\\\\textsuperscript{15}, Seyed Eghbal Ghobadi\\\\textsuperscript{8}, Ahmed Hammam\\\\textsuperscript{8}}\\nThe term \\\\emph{adversarial example} was first introduced by Szegedy \\\\etal .\\nFrom there on, many researchers tried to find new ways of crafting adversarial examples more effectively.\\nHere, the fast gradient sign method (FGSM) , DeepFool , least-likely class method (LLCM) , C\\\\&W , momentum iterative fast gradient sign method (MI-FGSM) , and projected gradient descent (PGD) are a few of the most famous attacks so far.\\nIn general, these attacks can be executed in an iterative fashion, where the underlying adversarial perturbation is usually bounded by some norm and is following additional optimization criteria, \\\\eg minimizing the number of changed pixels.\\n\\\\par\\nThe mentioned attacks can be further categorized as image-specific attacks, where for each image a new perturbation needs to be computed.\\nOn the other hand, image-agnostic attacks aim at finding a perturbation, which is able to fool an underlying DNN on a set of images.\\nSuch a perturbation is also referred to as a \\\\emph{universal adversarial perturbation} (UAP).\\nHere, the respective algorithm UAP , fast feature fool (FFF) , and prior driven uncertainty approximation (PD-UA) are a few honorable mentions.\\nAlthough the creation process of a universal adversarial perturbation typically relies on a white-box setting, they show a high \\\\emph{transferability} across models .\\nThis allows black-box attacks, where one model is used to create a universal adversarial perturbation, and another model is being attacked with the beforehand-created perturbation.\\nAnother way of designing black-box attacks is to create a surrogate DNN, which mimics the respective DNN to be attacked and thus can be used in the process of adversarial example creation .\\nOn the contrary, some research has been done to create completely incoherent images (based on evolutionary algorithms or gradient ascent) to fool an underlying DNN .\\nDifferent from that, another line of work has been proposed to alter only some pixels in images to attack a respective model.\\nHere and have used optimization approaches to perturb some pixels in images to produce targeted attacks, aiming at a specific class output, or non-targeted attacks, aiming at outputting a class different from the network output or the ground truth.\\nThis can be extended up to finding one pixel in the image to be exclusively perturbed to generate adversarial images .\\nThe authors of proposed to train generative models to generate adversarial examples.\\nGiven an input image and the target label, a generative model is trained to produce adversarial examples for DNNs.\\nHowever, while the produced adversarial examples look rather unrealistic to a human, they are able to completely deceive a DNN.\\n\\\\par\\nThe existence of adversarial examples not only motivated research in finding new attacks, but also in finding \\\\emph{defense strategies} to effectively defend these attacks.\\nEspecially for safety-critical applications, such as DNN-based environment perception for autonomous driving, the existence of adversarial examples needs to be handled accordingly.\\nSimilar to adversarial attacks, one can categorize defense strategies into two types: \\\\emph{model-specific} defense strategies and \\\\emph{model-agnostic} defense strategies.\\nThe former refers to defense strategies, where the model of interest is modified in certain ways.\\nThe modification can be done on the architecture, training procedure, training data, or model weights.\\nOn the other hand, model-agnostic defense strategies consider the model to be a black box.\\nHere, only the input or the output is accessible.\\nSome well-known model-specific defense strategies include adversarial training , the inclusion of robustness-oriented loss functions during training , removing adversarial patterns in features by denoising layers , and redundant teacher-student frameworks .\\nThe majority of model-agnostic defense strategies primarily focuses on various kinds of (gradient masking) pre-processing strategies .\\nThe idea is to remove the adversary from the respective image, such that the image is transformed from the adversarial space back into the clean space.\\n\\\\par\\nNonetheless, Athalye \\\\etal showed that gradient masking alone is not a sufficient criterion for a reliable defense strategy.\\nIn addition, detection and out-of-distribution techniques have also been proposed as model-agnostic defense strategies against adversarial attacks.\\nHere, the Mahalanobis distance or area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPR) are used to detect adversarial examples.\\nThe authors of on the other hand proposed to train networks to detect, whether the input image is out-of-distribution or not. \\n\\\\par\\nMoreover, Feinman \\\\etal proved that adversarial attacks usually produce high uncertainty on the output of the DNN.\\nAs a consequence, they proposed to use the dropout technique to estimate uncertainty on the output to identify a possible adversarial attack.\\n\\\\par\\nRegarding adversarial attacks, the majority of the listed attacks are designed for image classification.\\nOnly a few adversarial attacks consider tasks that are closely related to autonomous driving, such as bounding box detection, semantic segmentation, instance segmentation, or even panoptic segmentation.\\nAlso, the majority of the adversarial attacks rely on a white-box setting, which is usually not present for a potential attacker.\\nEspecially universal adversarial perturbations have to be considered as a real threat due to their high model transferability.\\nGenerally speaking, the existence of adversarial examples has not been thoroughly studied yet.\\nAn analytical interpretation is still missing, but could help in designing more mature defense strategies.\\n\\\\par\\nRegarding defense strategies, adversarial training is still considered as one of the most effective ways of increasing the robustness of a DNN.\\nNonetheless, while adversarial training is indeed effective, it is rather inefficient in terms of training time.\\nIn addition, model-agnostic defenses should be favored as once being designed, they can be easily transferred to different models.\\nMoreover, as most model-agnostic defense strategies rely on gradient-masking and it has been shown that gradient-masking is not a sufficient property for a defense strategy, new ways of designing model-agnostic defenses should be taken into account.\\nFurthermore, out-of-distribution and adversarial attacks detection or even correction methods have been a new trend for identifying attacks.\\nHowever, as the environment perception system of an autonomous driving vehicle could rely on various information sources, including LiDAR, optical flow, or depth from a stereo camera, techniques of information fusion should be further investigated to mitigate or even eliminate the effect of adversarial examples.\",\n \"id\": \"945be913-0c94-4b59-92f5-5fa8a6f0bf30\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 41,\n \"parent_id\": \"4396c1c3-a3ad-4efa-9f10-99f90060c0b4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Adversarial Attacks\"\n ],\n [\n \"subsection\",\n \"Adversarial Attacks and Defenses\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Adversarial Attacks and Defenses\"\n },\n {\n \"cite_extract_rate\": 0.684210526315789,\n \"cites\": [\n 900,\n 7698,\n 8187,\n 917,\n 890,\n 8189,\n 904,\n 901,\n 8188,\n 8190,\n 899,\n 7308,\n 6014\n ],\n \"content\": \"\\\\label{subsec:adverarial_attacks:more_realistic_attacks}\\n\\\\sectionauthor{Svetlana Pavlitskaya\\\\textsuperscript{14}}\\nWe consider the following two categories of realistic adversarial attacks: (1) image-level attacks, which not only fool a neural network but also pose a provable threat to autonomous vehicles, and (2) attacks which have been applied in a real world or in a simulation environment, such as car learning to act (CARLA) .\\n\\\\par\\nSome notable examples in the first category of attacks include attacks on semantic segmentation or person detection .\\n\\\\par\\nIn the second group of approaches, the attacks are specifically designed to survive real world distortions, including different distances, weather and lighting conditions, as well as camera angles. For this, adversarial perturbations are usually concentrated in a specific image area, called \\\\emph{adversarial patch}. Crafting an adversarial patch involves specifying a patch region in each training image, applying transformations to the patch, and iteratively changing the pixel values within this region to maximize the network prediction error. The latter step typically relies on an algorithm, proposed for standard adversarial attacks, which aim at crafting invisible perturbations while misleading neural networks, \\\\eg C\\\\&W , Jacobian-based saliency map attack (JSMA) , and PGD .\\n\\\\par\\nThe first printable adversarial patch for image classification was described by Brown \\\\etal . Expectation over transformations (EOT) is one of the influential updates to the original algorithm---it permits to robustify patch-based attacks to distortions and affine transformations. Localized and visible adversarial noise (LaVAN) is a further method to generate much smaller patches (up to 2\\\\% of the pixels in the image). In general, fooling image classification with a patch is a comparatively simple task, because adversarial noise can mimic an instance of another class and thus lower the prediction probability for a true class.\\n\\\\par\\nRecently, patch-based attacks for a more tricky task of object detection have been described . Also, Lee and Kolter generate a patch using PGD , followed by EOT applied to the patch. With this approach, all detections in an image can be successfully suppressed, even without any overlap of a patch with bounding boxes. Furthermore, several approaches for generating an adversarial T-shirt have been proposed, including .\\n\\\\par\\nDeepBillboard is the first attempt to attack end-to-end driving models with adversarial patches. The authors propose to generate a single patch for a sequence of input images to mislead four steering models, including DAVE-2 in a drive-by scenario. \\n\\\\par\\nApart from physical feasibility, inconspicuousness is crucial for a realistic attack. Whereas adversarial patches usually look like regions of noise, several works have explored attacks with an inconspicuous patch. In particular, Eykholt \\\\etal demonstrate the vulnerability of road sign classification to the adversarial perturbations in the form of only black and white stickers. In , an end-to-end driving model is attacked in CARLA by painting of black lines on the road. Also, Kong and Liu use a generative adversarial network to get a realistic billboard to attack an end-to-end driving model in a drive-by scenario. In , a method to hide visible adversarial perturbations with customized styles is proposed, which leads to adversarial traffic signs that look unsuspicious to a human.\\n\\\\par\\nCurrent research mostly focuses on attacking image-based perception of an autonomous vehicle. Adversarial vulnerability of further components of an autonomous vehicle, \\\\eg LiDAR-based perception, optical flow and depth estimation, has only recently gained attention. Furthermore, most attacks consider only a single component of an autonomous driving pipeline, the question whether the existing attacks are able to propagate to further pipeline stages has not been studied yet. The first work in this direction describes an attack on object detection and tracking. The evaluation is, however, limited to a few clips, where no experiments in the real world have been performed. Overall, the research on realistic adversarial attacks, especially combined with physical tests, is currently in the starting phase.\",\n \"id\": \"7e154b91-68e0-47d3-8c11-92f8eb25b769\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 19,\n \"parent_id\": \"4396c1c3-a3ad-4efa-9f10-99f90060c0b4\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Adversarial Attacks\"\n ],\n [\n \"subsection\",\n \"More Realistic Attacks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"More Realistic Attacks\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 7514\n ],\n \"content\": \"\\\\label{sec:interpretability}\\n\\\\sectionauthor{Felix Brockherde\\\\textsuperscript{10}}\\nNeural networks are, by their nature, black boxes and therefore intrinsically hard to interpret . Due to their unrivaled performance, they still remain first choice for advanced systems even in many safety-critical areas, such as level 4 automated driving. This is why the research community has invested considerable effort to unhinge the black-box character and make deep neural networks more transparent.\\nWe can observe three strategies that provide different view points towards this goal in the state of the art. First is the most direct approach of opening up the black box and looking at intermediate representations (\\\\refsec{subsec:interpretability:intermediate_representations}). Being able to interpret individual layers of the system facilitates interpretation of the whole. The second approach tries to provide interpretability by explaining the network's decisions with pixel attributions (\\\\refsec{subsec:interpretability:pixel_attribution}). Aggregated explanations of decision can then lead to interpretability of the system itself. Third is the idea of approximating the network with interpretable proxies to benefit from the deep neural networks performance while allowing interpretation via surrogate models (\\\\refsec{subsec:interpretability:interpretable_proxies}). Underlying all aspects here is the area of visual analytics (\\\\refsec{subsec:interpretability:visual_analytics}).\\nThere exists earlier research in the medical domain to help human experts understand and convince them of machine learning decisions . Legal requirements in the finance industry gave rise to interpretable systems that can justify their decisions. An additional driver for interpretability research was the concern for Clever Hans predictors .\",\n \"id\": \"e47014ad-1aec-4b32-9241-6aab90a7f0dd\",\n \"level\": \"section\",\n \"origin_cites_number\": 3,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Interpretability\"\n ]\n ],\n \"subsections\": [\n \"cf04c28f-300a-4cf5-bc98-3f4d9734078e\",\n \"a3e2cd79-8634-4ce7-9519-acd205ba8d1d\",\n \"110c6df1-6598-4052-8ee0-6595325dfebc\",\n \"37f2baca-d82b-4285-9422-b80c37554845\"\n ],\n \"title\": \"Interpretability\"\n },\n {\n \"cite_extract_rate\": 0.17647058823529402,\n \"cites\": [\n 8192,\n 8191,\n 8193\n ],\n \"content\": \"\\\\label{subsec:interpretability:visual_analytics}\\n\\\\sectionauthor{Elena Schulz\\\\textsuperscript{1}}\\nTraditional data science has developed a huge tool set of automated analysis processes conducted by computers, which are applied to problems that are well-defined in the sense that the dimensionality of input and output as well as the size of the data set they rely on is manageable. \\nFor those problems that in comparison are more complex, the automation of the analysis process is limited and/or might not lead to the desired outcome. \\nThis is especially the case with unstructured data like image or video data in which the underlying information cannot directly be expressed by numbers. \\nRather, it needs to be transformed to some structured form to enable computers to perform some task of analysis. \\nAdditionally, with an ever increasing amount of various types of data being collected, this ``information overload'' cannot solely be analyzed by automatic methods . \\n\\\\par\\n\\\\emph{Visual analytics} addresses this challenge as ``the science of analytical reasoning facilitated by interactive visual interfaces'' . \\nVisual analytics therefore does not only focus on either computationally processing data or visualizing results but coupling both tightly with interactive techniques. \\nThus, it enables an integration of the human expert into the iterative visual analytics process: Through visual understanding and human reasoning, the knowledge of the human expert can be incorporated to effectively refine the analysis. This is of particular importance, where a stringent safety argumentation for complex models is required. With the help of visual analytics, the line of argumentation can be built upon arguments that are understandable for humans.\\nTo include the human analyst efficiently into this process, a possible guideline is the \\\\emph{visual analytics mantra} by Keim: ``Analyze first, show the important, zoom, filter and analyze further, details on demand'' \\\\footnote{Extending the original \\\\emph{visualization mantra} by Shneiderman ``Overview first, filter and zoom, details on demand''.}.\\n\\\\par\\nThe core concepts of visual analytics therefore rely on well-designed interactive visualizations, which support the analyst in the tasks of, \\\\eg reviewing, understanding, comparing and inferring not only the initial phenomenon or data but also the computational model and its results itself with the goal of enhancing the analytical process. \\n\\\\par\\nDriven by various fields of application, visual analytics is a multidisciplinary field with a wide variety of task-oriented development and research. \\nAs follows, recent work has been done in several areas: \\ndepending on the task, there exist different pipeline approaches to create whole \\\\emph{visual analytics systems} ; \\nthe injection of human expert knowledge into the process of determining trends and patterns from data is the focus of \\\\emph{predictive visual analytics} ; \\nenabling the human to explore \\\\emph{high-dimensional data} interactively and visually (\\\\eg via dimensionality reduction ) is a major technique to enhance the understandability of complex models (\\\\eg neural networks); \\nthe iterative improvement and the understanding of machine learning models is addressed by using interactive visualizations in the field of \\\\emph{general machine learning} or the other way round: using machine learning to improve visualizations and guidance based on user interactions .\\nEven more focused on the loop of simultaneously developing and refining machine learning models is the area of \\\\emph{interactive machine learning}, where the topics of \\\\emph{interface design} and the \\\\emph{importance of users} are discussed. One of the current research directions is using visual analytics in the area of \\\\emph{deep learning} . However, due to the interdisciplinarity of visual analytics, there are still open directions and ongoing research opportunities. \\n\\\\par\\nEspecially in the domain of neural networks and deep learning, visual analytics is a relatively new approach in tackling the challenge of \\\\emph{explainability} and \\\\emph{interpretability} of those often called \\\\emph{black boxes}. \\nTo enable the human to better interact with the models, research is done in enhancing the \\\\emph{understandability} of complex deep learning models and their outputs with the use of proper visualizations. \\nOther research directions attempt to achieve improving the \\\\emph{trustability} of the models, giving the opportunity to inspect, diagnose and refine the model. \\nFurther, possible areas for research are \\\\emph{online training processes} and the development of \\\\emph{interactive systems} covering the whole process of training, enhancing and monitoring machine learning models. \\nHere, the approach of \\\\emph{mixed guidance}, where system-initiated guidance is combined with user-initiated guidance, is discussed among the visual analytics community as well. \\nAnother challenge and open question is creating ways of \\\\emph{comparing models} to examine which model yields a better performance, given specific situations and selecting or combining the best models with the goal of increasing performance and overall safety.\",\n \"id\": \"cf04c28f-300a-4cf5-bc98-3f4d9734078e\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"e47014ad-1aec-4b32-9241-6aab90a7f0dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Interpretability\"\n ],\n [\n \"subsection\",\n \"Visual Analytics\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Visual Analytics\"\n },\n {\n \"cite_extract_rate\": 0.514285714285714,\n \"cites\": [\n 1002,\n 8198,\n 6410,\n 8201,\n 8197,\n 8200,\n 8194,\n 8199,\n 5680,\n 3314,\n 4423,\n 8202,\n 2709,\n 8196,\n 306,\n 318,\n 8195,\n 1256\n ],\n \"content\": \"\\\\label{subsec:interpretability:intermediate_representations}\\n\\\\sectionauthor{Felix Hauser\\\\textsuperscript{11}, Jan Kronenberger\\\\textsuperscript{9}, Seyed Eghbal Ghobadi\\\\textsuperscript{8}}\\nIn general, representation learning aims to extract lower dimensional features in latent space from higher dimensional inputs.\\nThese features are then used as an effective representation for regression, classification, object detection and other machine learning tasks.\\nPreferably, latent features should be disentangled, meaning that they represent separate factors found in the data that are statistically independent.\\nDue to their importance in machine learning, finding meaningful intermediate representations has long been a primary research goal.\\nDisentangled representations can be interpreted more easily by humans and \\ncan for example be used to explain the reasoning of neural networks .\\nAmong the longer known methods for extracting disentangled representations are principal component analysis (PCA) , independent component analysis , and nonnegative matrix factorization .\\nPCA is highly sensitive to outliers and noise in the data.\\nTherefore, more robust algorithms were proposed.\\nIn already a small neural network was used as an encoder and the algorithm proposed in can deal with high-dimensional data.\\nSome robust PCA algorithms are provided with analytical performance guarantees .\\nA popular method for representation learning with deep networks is the variational autoencoder (VAE) .\\nAn important generalization of the method is the $ \\\\beta $-VAE variant , which improved the disentanglement capability .\\nLater analysis added to the theoretical understanding of $ \\\\beta $-VAE\\n.\\nCompared to standard autoencoders, VAEs map inputs to a distribution, instead of mapping them to a fixed vector.\\nThis allows for additional regularization of the training to avoid overfitting and ensure good representations.\\nIn $ \\\\beta $-VAEs the trade-off between reconstruction quality and disentanglement can be fine-tuned by the hyperparameter $ \\\\beta $.\\nDifferent regularization schemes have been suggested to improve the VAE method.\\nAmong them are Wasserstein autoencoders , attribute regularization and relational regularization .\\nRecently, a connection between VAEs and nonlinear independent component analysis was established and then expanded .\\nBesides VAEs, deep generative adversarial networks can be used to construct latent features .\\nOther works suggest centroid encoders or conditional learning of Gaussian distributions as alternatives to VAEs.\\nIn concept activation vectors are defined as being orthogonal to the decision boundary of a classifier.\\nApart from deep learning, entirely new architectures, such as capsule networks , might be used to disassemble inputs. \\n\\\\par\\nWhile many different approaches for disentangling exist, the feasibility of the task is not clear yet and a better theoretical understanding is needed.\\nThe disentangling performance is hard to quantify, which is only feasible with information about the latent ground truth .\\nModels that overly rely on single directions, single neurons in fully connected networks or single feature maps in CNNs, have the tendency to overfit .\\nAccording to , unsupervised learning does not produce good disentangling and even small latent spaces do not reduce the sample complexity for simple tasks.\\nThis is in direct contrast to newer findings that show a decreased sample complexity for more complex visual downstream tasks .\\nSo far, it is unclear if disentangling improves the performance of machine learning tasks.\\nIn order to be interpretable, latent disentangled representations need to be aligned with human understandable concepts.\\nIn training with adversarial examples was used and the learned representations were shown to be more aligned with human perception.\\nFor explainable AI, disentangling alone might not be enough to generate interpretable output and additional regularization could be needed.\",\n \"id\": \"a3e2cd79-8634-4ce7-9519-acd205ba8d1d\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 35,\n \"parent_id\": \"e47014ad-1aec-4b32-9241-6aab90a7f0dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Interpretability\"\n ],\n [\n \"subsection\",\n \"Intermediate Representations\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Intermediate Representations\"\n },\n {\n \"cite_extract_rate\": 0.6875,\n \"cites\": [\n 4875,\n 499,\n 8600,\n 1812,\n 1822,\n 1824,\n 1814,\n 7517,\n 1617,\n 6080,\n 8203\n ],\n \"content\": \"\\\\label{subsec:interpretability:pixel_attribution}\\n\\\\sectionauthor{Stephanie Abrecht\\\\textsuperscript{2}, Felix Brockherde\\\\textsuperscript{10}, Toshika Srivastava\\\\textsuperscript{12}}\\nThe non-linearity and complexity of DNNs allow them to solve perception problems, like detecting a pedestrian, that cannot be specified in detail. At the same time, the automatic extraction of features given in an input image and the mapping to the respective prediction is counterintuitive and incomprehensible for humans, which makes it hard to argue safety for a neural network-based perception task. Feature importance techniques are currently predominantly used to diagnose the causes of incorrect model behaviors . So-called \\\\emph{attribution maps} are a visual technique to express the relationship between relevant pixels in the input image and the network's prediction. Regions in an image that contain relevant features are highlighted accordingly. \\nAttribution approaches mostly map to one of three categories.\\n\\\\par\\nGradient-based and activation-based approaches (such as amongst others) rely on the gradient of the prediction with respect to the input. Regions that were most relevant for the prediction are highlighted. Activation-based approaches relate the feature maps of the last convolutional layer to output classes. \\n\\\\par\\nPerturbation-based approaches suggest manipulating the input. If the prediction changes significantly, the input may hold a possible explanation at least.\\n\\\\par\\nWhile gradient-based approaches are oftentimes faster in computation, perturbation-based approaches are much easier to interpret.\\n\\\\par\\nAs many studies have shown , there is still a lot of research to be done before attribution methods are able to robustly provide explanations for model predictions, in particular erroneous behavior. One key difficulty is the lack of an agreed-upon definition of a good attribution map including important properties. Even between humans, it is hard to agree on what a good explanation is due to its subjective nature. This lack of ground truth makes it hard or even impossible to quantitatively evaluate an explanation method. Instead, this evaluation is done only implicitly. \\nOne typical way to do this is the axiomatic approach. Here a set of desiderata of an attribution method are defined, on which different attribution methods are then evaluated. Alternatively, different attribution methods may be compared by perturbing the input features starting with the ones deemed most important and measuring the drop in accuracy of the perturbed models. The best method will result into the greatest overall loss in accuracy as the number of inputs are omitted .\\nMoreover, for gradient-based methods it is hard to assess if an unexpected attribution is caused by a poorly performing network or a poorly performing attribution method . How to cope with negative evidence, \\\\ie the object was predicted because a contrary clue in the input image was missing, is an open research question. Additionally, most methods were shown on classification tasks until now. It remains to be seen how they can be transferred to object detection and semantic segmentation tasks. In the case of perturbation-based methods, the high computation time and single-image analysis inhibit wide-spread application.\",\n \"id\": \"110c6df1-6598-4052-8ee0-6595325dfebc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 16,\n \"parent_id\": \"e47014ad-1aec-4b32-9241-6aab90a7f0dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Interpretability\"\n ],\n [\n \"subsection\",\n \"Pixel Attribution\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Pixel Attribution\"\n },\n {\n \"cite_extract_rate\": 0.33333333333333304,\n \"cites\": [\n 8502,\n 7507,\n 1821,\n 8204\n ],\n \"content\": \"\\\\label{subsec:interpretability:interpretable_proxies}\\n\\\\sectionauthor{Gesina Schwalbe\\\\textsuperscript{3}}\\nNeural networks are capable of capturing complicated logical\\n(cor)relations. However, this knowledge is encoded on a\\n\\\\emph{sub-symbolic} level in the form of learned weights and biases,\\nmeaning that the reasoning behind the processing chain cannot be\\ndirectly read out or interpreted by humans .\\nTo explain the sub-symbolic processing, one can either use attribution\\nmethods (\\\\cf \\\\refsec{subsec:interpretability:pixel_attribution}), or lift\\nthis sub-symbolic representation to a \\\\emph{symbolic} one\\n, meaning a more interpretable one.\\nInterpretable proxies or surrogate models try to achieve the latter:\\nThe DNN behavior is approximated by a model that\\nuses symbolic knowledge representations.\\nSymbolic representations can be\\nlinear models like LIME (proportionality),\\ndecision trees (if-then-chains) ,\\nor loose sets of logical rules.\\nLogical connectors can simply be AND and OR but also more general\\nones like at-least-M-of-N .\\nThe \\\\emph{expressiveness} of an approach refers to the logic that is\\nused: Boolean-only versus first-order logic, and binary versus fuzzy\\nlogic truth values .\\nOther than attribution methods (\\\\cf \\\\refsec{subsec:interpretability:pixel_attribution}), these\\nrepresentations can capture combinations of features and (spatial)\\nrelations of objects and attributes. As an example consider\\n\\\\enquote{eyes are closed} as explanation for \\\\enquote{person asleep}:\\nAttribution methods only could mark the location of the eyes, dismissing\\nthe relations of the attributes .\\nAll mentioned surrogate model types (linear, set of rules) require\\ninterpretable input features in order to be interpretable themselves.\\nThese features must either be directly obtained from the DNN input or\\n(intermediate) output, or automatically be extracted from the DNN\\nrepresentation.\\nExamples for extraction are the super-pixeling used in LIME for input\\nfeature detection, or concept activation vectors\\n for DNN representation decoding.\\n\\\\par\\nQuality criteria and goals for interpretable proxies are :\\n\\\\emph{accuracy} of the standalone surrogate model on unseen examples,\\n\\\\emph{fidelity} of the approximation by the proxy,\\n\\\\emph{consistency} with respect to different training sessions, and\\n\\\\emph{comprehensibility} measured by the complexity of the rule set\\n(number of rules, number of hierarchical dependencies).\\nThe criteria are usually in conflict and need to be balanced:\\nBetter accuracy may require a more complex, thus less expressive sets of rules.\\n\\\\par\\nApproaches for interpretable proxies differ in the validity range of\\nthe representations:\\nSome aim for surrogates that are only valid \\\\emph{locally} around\\nspecific samples, like in LIME~ or in\\n via inductive logic programming.\\nOther approaches try to more \\\\emph{globally} approximate aspects of\\nthe model behavior.\\nAnother categorization is defined by whether full access\\n(\\\\emph{white-box}), some access (\\\\emph{gray-box}), or no access\\n(\\\\emph{black-box}) to the DNN internals is needed.\\nOne can further differentiate between \\\\emph{post-hoc} approaches that\\nare applied to a trained model, and approaches that try to integrate\\nor \\\\emph{enforce symbolic representations} during training.\\nPost-hoc methods cover the wide field of rule extraction techniques\\nfor DNNs. The reader may refer to\\n.\\nMost white- and gray-box methods try to turn the DNN connections into\\nif-then rules that are then simplified, like done in\\nDeepRED~.\\nA black-box example is validity interval\\nanalysis~, which refines or\\ngeneralizes rules on input intervals, either starting from one sample\\nor a general set of rules.\\nEnforcement of symbolic representations can be achieved by enforcing\\nan output structure that provides insights to the decision logic, such as\\ntextual explanations, or a rich output structure allowing investigation of correlations\\n.\\nAn older discipline for enforcing symbolic representations\\nis the field of neural-symbolic learning\\n. The idea is based on a hybrid learning\\ncycle \\nin which a symbolic learner and a DNN iteratively update each other via\\nrule insertion and extraction.\\n\\\\par\\nThe comprehensibility of global surrogate models suffers from the\\ncomplexity and size of concurrent DNNs. Thus, stronger rule\\nsimplification methods are required .\\nThe alternative direction of local approximations mostly\\nconcentrates on linear models instead of more expressive rules .\\nFurthermore, balancing of the quality objectives is hard since\\navailable indicators for interpretability may not be ideal.\\nAnd lastly, applicability is heavily infringed by the requirement of\\ninterpretable input features. These are usually not readily available\\nfrom input (often pixel-level) or DNN output. Supervised extraction\\napproaches vary in their fidelity, and unsupervised ones do not\\nguarantee to yield meaningful or interpretable results, respectively, such\\nas the super-pixel clusters of LIME.\",\n \"id\": \"37f2baca-d82b-4285-9422-b80c37554845\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"e47014ad-1aec-4b32-9241-6aab90a7f0dd\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Interpretability\"\n ],\n [\n \"subsection\",\n \"Interpretable Proxies\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Interpretable Proxies\"\n },\n {\n \"cite_extract_rate\": 0.75,\n \"cites\": [\n 4616,\n 4025,\n 3288\n ],\n \"content\": \"\\\\label{sec:uncertainty} \\n\\\\sectionauthor{Michael Mock\\\\textsuperscript{1}}\\nUncertainty refers to the view that a neural network is not conceived as a deterministic function but as a probabilistic function or estimator, delivering a random distribution for each input point. Ideally, the mean value of the distribution should be as close as possible to the ground truth value of the function being approximated by the neural network and the uncertainty of the neural network refers to its variance when being considered as a random variable, thus allowing to derive a confidence with respect to the mean value. Regarding safety, the variance may lead to estimations about the confidence associated with a specific network output and opens the option for discarding network outputs with insufficient confidence. \\n\\\\par\\nThere are roughly two broad approaches for training neural networks as probabilistic functions: Parametric approaches and Bayesian neural networks on the one hand, such as , where the transitions along the network edges are modeled as probability distributions, and ensemble-based approaches on the other hand , where multiple networks are trained and considered as samples of a common output distribution. Apart from training as probabilistic function, uncertainty measures have been derived from single, standard neural networks by post-processing on the trained network logits, leading for example to calibration measures (\\\\cf \\\\eg ).\",\n \"id\": \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"level\": \"section\",\n \"origin_cites_number\": 4,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Uncertainty\"\n ]\n ],\n \"subsections\": [\n \"e3a9977a-3391-4691-bc3b-9d131807d8dd\",\n \"d4df1917-2df5-4888-9dac-86e261e7d3d8\",\n \"329b292f-42f5-496b-9775-5b5a3bd0f517\",\n \"dcb846f5-368c-45a9-a56f-6ddb586a0a6b\",\n \"81f04d24-1455-4244-a481-8a943c0061e4\",\n \"09524fc1-6625-4adb-b57e-b6186b5d4ef0\"\n ],\n \"title\": \"Uncertainty\"\n },\n {\n \"cite_extract_rate\": 0.23333333333333303,\n \"cites\": [\n 1003,\n 5680,\n 8430,\n 8207,\n 8206,\n 8205,\n 8208\n ],\n \"content\": \"\\\\label{subsec:uncertainty:generative_models}\\n\\\\sectionauthor{Sebastian Wirkert\\\\textsuperscript{6}, Tim Wirtz\\\\textsuperscript{1}}\\n\\\\emph{Generative models} belong to the class of unsupervised machine learning models. From a theoretical perspective, these are particularly interesting, because they offer a way to analyze and model the density of data. Given a finite data set $\\\\mathcal{D}$ independently distributed according to some distribution $p(x)$, generative models aim to estimate or enable sampling from the underlying density $p(x)$ in a model $q_\\\\theta(x)$. The resulting model can be used for data indexing~, data retrieval~, for visual recognition~, speech recognition and generation~, language processing~ and robotics~.\\nFollowing , we can group generative models into two main classes:\\n\\\\begin{itemize}\\n \\\\item Cost function-based models such as autoencoder~, deep belief networks~ and generative adversarial networks~.\\n \\\\item Energy-based models~, where the joint probability density is modeled by an energy function.\\n\\\\end{itemize}\\nBeside these \\\\emph{deep learning} approaches, generative models have been studied in machine learning in general for quite some time (\\\\cf ). A very prominent example of generative networks are Gaussian processes~ and their deep learning extensions~ as generative models. \\n\\\\par\\nAn example of a generative model being employed for image segmentation uncertainty estimation is the probabilistic U-Net . Here a variational autoencoder (VAE) conditioned on the image is trained to model uncertainties. Samples from the VAE are fed into a segmentation U-Net which can thus give different results for the same image. This was tested in context of medical images, where inter-rater disagreements lead to uncertain segmentation results and Cityscapes segmentation. For the Cityscapes segmentation the investigated use case was label ambiguity (\\\\eg is a BMW X7 a car or a van) using artificially created, controlled ambiguities. Results showed that the probabilistic U-Net could reproduce the segmentation ambiguity modes more reliably than competing methods such as a dropout U-Net which is based on techniques elaborated in the next section.\",\n \"id\": \"e3a9977a-3391-4691-bc3b-9d131807d8dd\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 30,\n \"parent_id\": \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Uncertainty\"\n ],\n [\n \"subsection\",\n \"Generative Models\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Generative Models\"\n },\n {\n \"cite_extract_rate\": 0.538461538461538,\n \"cites\": [\n 4616,\n 3288,\n 4640,\n 8205,\n 4025,\n 8209,\n 8814\n ],\n \"content\": \"\\\\label{subsec:uncertainty:mc_dropout} \\n\\\\sectionauthor{Joachim Sicking\\\\textsuperscript{1}}\\nA widely used technique to estimate model uncertainty is Monte-Carlo (MC) dropout , that offers a Bayesian motivation, conceptual simplicity and scalability to application-size networks. This combination distinguishes MC dropout from competing Bayesian neural network (BNN) approximations (like ,, see \\\\refsec{subsec:uncertainty:bayesian_neural_networks}). However, these approaches and MC dropout share the same goal: to equip neural networks with a \\\\emph{self-assessment} mechanism that detects unknown input concepts and thus potential model insufficiencies.\\nOn a technical level, MC dropout assumes prior distributions on network activations, usually independent and identically distributed (i.i.d.) Bernoulli distributions. Model training with iteratively drawn Bernoulli samples, the so-called \\\\emph{dropout masks}, then yields a data-conditioned posterior distribution within the chosen parametric family. It is interesting to note that this training scheme was used earlier---independent from an uncertainty context---for better model generalization . At inference, sampling provides estimates of the input-dependent output distributions. The spread of these distributions is then interpreted as the prediction uncertainty that originates from limited knowledge of model parameters. Borrowing ‘frequentist’ terms, MC dropout can be considered as an implicit network ensemble, \\\\ie as a set of networks that share (most of) their parameters.\\nIn practice, MC dropout requires only a minor modification of the optimization objective during training and multiple, trivially parallelizable forward passes during inference. The loss modification is largely agnostic to network architecture and does not cause substantial overhead. This is in contrast to the sampling-based inference that increases the computational effort massively---by estimated factors of 20-100 compared to networks without MC dropout. A common practice is therefore the use of \\\\emph{last-layer dropout} that reduces computational overhead to estimated factors of 2-10. Alternatively, analytical moment propagation allows sampling-free MC-dropout inference at the price of additional approximations (\\\\eg). Further extensions of MC dropout target the integration of data-inherent (aleatoric) uncertainty and tuned performance by learning layer-specific dropout rate using concrete relaxations .\\nThe quality of MC-dropout uncertainties is typically evaluated using negative log-likelihood (NLL), expected calibration error (ECE) and its variants (\\\\cf \\\\refsec{subsec:uncertainty:confidence_calibration}) and by considering correlations between uncertainty estimates and model errors (\\\\eg AUSE ). Moreover, it is common to study how useful uncertainty estimates are for solving auxiliary tasks like out-of-distribution classification or robustness \\\\wrt adversarial attacks. \\nMC dropout is a working horse of safe ML, being used with various networks and for a multitude of applications (\\\\eg). However, several authors pointed out shortcomings and limitations of the method: MC dropout bears the risk of over-confident false predictions (), offers less diverse uncertainty estimates compared to (the equally simple and scalable) deep ensembles (, see \\\\refsec{subsec:aggregation:ensemble_methods}) and provides only rudimentary approximations of true posteriors. \\nRelaxing these modelling assumptions and strengthening the Bayesian motivation of MC dropout is therefore an important research avenue. Further directions for future work are the development of \\\\emph{semantic uncertainty mechanisms} (\\\\eg ), improved local uncertainty calibrations and a better understanding of the outlined sampling-free schemes to uncertainty estimation.\",\n \"id\": \"d4df1917-2df5-4888-9dac-86e261e7d3d8\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 13,\n \"parent_id\": \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Uncertainty\"\n ],\n [\n \"subsection\",\n \"Monte-Carlo Dropout\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Monte-Carlo Dropout\"\n },\n {\n \"cite_extract_rate\": 0.25,\n \"cites\": [\n 8633,\n 4025\n ],\n \"content\": \"\\\\label{subsec:uncertainty:bayesian_neural_networks}\\n\\\\sectionauthor{Maram Akila\\\\textsuperscript{1}}\\nAs the name suggests, Bayesian neural networks (BNNs) are inspired by a Bayesian interpretation of probability (for an introduction \\\\cf ).\\nIn essence, it rests on Bayes' theorem,\\n\\\\begin{equation}\\n\\tp(x|y) p(y) = p(x,y) = p(y | x) p(x)\\n\\t\\\\quad \\\\Rightarrow \\\\quad\\n\\tp(x| y) = \\\\frac{ p(y | x) p(x)}{ p(y) }\\n\\t\\\\,,\\n\\t\\\\label{eq:bnn:bayes}\\n\\\\end{equation}\\nstating that the conditional probability density function (PDF) $p(x|y)$ for $x$ given $y$ may be expressed in terms of the inverted conditional PDF $p(y | x)$.\\nFor machine learning, where one intends to make predictions $y$ for unknown $x$ given some training data $\\\\mathcal{D}$, this can be reformulated into\\n\\\\begin{equation}\\n\\ty = \\\\operatorname{NN}(x|W)\\n\\t\\\\quad\\\\text{with}\\\\quad\\n\\t p(W | \\\\mathcal{D}) = \\\\frac{p(\\\\mathcal{D} | W) p(W)}{p(\\\\mathcal{D})}\\n\\t\\\\,.\\n\\t\\\\label{eq:bnn:NNpdf}\\n\\\\end{equation}\\nTherein NN denotes a conventional (deep) neural network (DNN) with model parameters $W$, \\\\eg the set of weights and biases.\\nIn contrast to a regular DNN, the weights are given in terms of a probability distribution $ p(W | \\\\mathcal{D}) $ turning also the output $y$ of a BNN into a distribution.\\nThis allows to study the mean $\\\\mu=\\\\langle y^1 \\\\rangle$ of the DNN for a given $x$ as well as higher moments of the distribution, typically the resulting variance $\\\\sigma^2 = \\\\langle (y-\\\\mu)^2 \\\\rangle$ is of interest, where\\n\\\\begin{equation}\\n\\t\\\\left\\\\langle y^k \\\\right\\\\rangle = \\\\int\\\\!\\n\\t\\\\operatorname{NN}(x|W)^k p(W|\\\\mathcal{D})\\\\,\\\\mathrm{d}W\\\\,.\\n\\t\\\\label{eq:bnn:moments}\\n\\\\end{equation}\\nWhile $\\\\mu$ yields the output of the network, the variance $\\\\sigma^2$ is a measure for the uncertainty of the model for the prediction at the given point.\\nCentral to this approach is the probability of the data given the model, here denoted by $p(\\\\mathcal{D} | W) $, as it is the key component connecting model and training data.\\nTypically, the prior distribution $p(\\\\mathcal{D})$ is ``ignored'' as it only appears as a normalization constant within the averages, see \\\\eqref{eq:bnn:moments}.\\nIn the cases where the data $\\\\mathcal{D}$ is itself a distribution due to inherent uncertainty, \\\\ie presence of aleatoric risk , such a concept seems natural.\\nHowever, Bayesian approaches are also applicable for all other cases.\\nIn those, loosely speaking, the likelihood of $W$ is determined via the chosen loss function (for the connection between the two concepts \\\\cf ).\\n\\\\par\\nOn this general level, Bayesian approaches are broadly accepted and also find use for many other model classes besides neural networks.\\nHowever, the loss surfaces of DNNs are known for their high dimensionality and strong non-convexity.\\nTypically, there are abundant parameter combinations $W$ that lead to (almost) equally good approximations to the training data $\\\\mathcal{D}$ with respect to a chosen loss.\\nThis makes an evaluation of $p(W|\\\\mathcal{D})$ for DNNs close to impossible in full generality.\\nAt least no (exact) solutions for this case exist at the moment.\\n\\\\par\\nFinding suitable approximations to the posterior distribution $p(W|\\\\mathcal{D})$ is an ongoing challenge for the construction of BNNs.\\nAt this point we only summarize two major research directions in the field.\\nOne approach is to assume that the distribution factorizes.\\nWhile the full solution would be a joined distribution implying correlations between different weights (etc.), possibly even across layers, this approximation takes each element $w_i$ of $W$ to be independent form the others.\\nAlthough this is a strong assumption, it is often made, in this case parameters for the respective distributions of each element can be learned via training (\\\\cf ).\\nThe second class of approaches focuses on the region of the loss surface around the minimum chosen for the DNN.\\nAs discussed, the loss relates to the likelihood and quantities such as the curvature at the minimum, therefore directly connected to the distribution of $W$.\\nUnfortunately, already using this type of quantities requires further approximations .\\nAlternatively, the convergence of the training process may be altered to sample networks close to the minimum .\\nWhile this approach contains information about correlations among the $w_i$, it is usually restricted to a specific minimum.\\nFor a non-Bayesian ansatz taking into account several minima, see deep ensembles in \\\\refsec{subsec:aggregation:ensemble_methods}.\\nBNNs also touch other concepts such as MC dropout (\\\\cf \\\\refsec{subsec:uncertainty:mc_dropout} or ), or prior networks, which are based on a Bayesian interpretation but use conventional DNNs with an additional (learned) $\\\\sigma$ output .\",\n \"id\": \"329b292f-42f5-496b-9775-5b5a3bd0f517\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 8,\n \"parent_id\": \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Uncertainty\"\n ],\n [\n \"subsection\",\n \"Bayesian Neural Networks\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Bayesian Neural Networks\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 1624,\n 4620,\n 8211,\n 8210,\n 1733,\n 206\n ],\n \"content\": \"\\\\label{subsec:uncertainty:gradient_based_uncertainty}\\n\\\\sectionauthor{Matthias Rottmann\\\\textsuperscript{5}}\\nClassical uncertainty quantification methods in frequentist inference are mostly based on the outputs of statistical models. Their uncertainty is quantified and assessed for instance via dispersion measures in classification (such as entropy, probability margin or variation ratio), or confidence intervals in regression. However, the nature of DNN architectures and the cutting edge applications tackled by those (\\\\eg semantic segmentation, \\\\cf ) open the way towards more elaborate uncertainty quantification methods. Besides the mentioned classical approaches, intermediate feature representations within a DNN (\\\\cf ) or gradients according to self-affirmation that represent re-learning stress (see ) reveal additional information. In addition, in case of semantic segmentation, the geometry of a prediction may give access to further information, cf.\\\\ . By the computation of statistics of those quantities as well as low-dimensional representations thereof, we obtain more elaborate uncertainty quantification methods specifically designed for DNNs that can help us to detect misclassifications and out-of-distribution objects (\\\\cf ). \\n\\\\par\\nFeatures gripped during a fordward pass of a data point $x$ through a DNN $f$ can be considered layer-wise, \\\\ie $f^{(\\\\ell)}(x)$ after the $\\\\ell$-th layer. These can be translated into a handful of quantities per layer or further processed by another DNN that aims that detecting errors . While in particular presents a proof of concept on small scale classification problems, their applicability to large scale datasets and problems such as semantic segmentation and object detection remain open.\\n\\\\par\\nThe development for gradient-based uncertainty quantification methods is guided by one central question:\\nIf the present prediction was true, how much re-learning would this require.\\nThe corresponding hypothesis is that wrong predictions would be more in conflict with the knowledge encoded in the deep neural network than correct ones, therefore causing increased re-learning stress.\\nGiven a predicted class\\n\\\\begin{equation}\\n\\\\hat{y} = \\\\mathop{\\\\mathrm{arg\\\\,max}}_{y} f_y(x) \\n\\\\end{equation}\\nwe compute the gradient of layer $\\\\ell$ corresponding to the predicted label. That is, given a loss function $\\\\mathcal{L}$, we compute\\n\\\\begin{equation}\\n\\\\nabla_{w_\\\\ell} \\\\mathcal{L}( \\\\hat{y}, x, w ) \\n\\\\end{equation}\\nvia backpropagation. The obtained quantities can be treated similarly to the case of forward pass features. While this concept seems to be prohibitively expensive for semantic segmentation (at least when calculating gradients for each pixel of $\\\\hat{y}$), its applicability to object detection might be feasible, in particular with respect to offline applications. Gradients are also of special interest in active learning with query by \\\\emph{expected model change} (\\\\cf ).\\n\\\\par\\nIn the context of semantic segmentation, geometrical information on segments shapes as well as neighborhood relations of predicted segments can be taken into account along side with dispersion measures. It has been demonstrated that the detection of errors in an in-distribution setting strongly benefits from geometrical information. Recently, this has also been considered in scenarios under moderate domain shift . However, its applicability to out-of-distribution examples and to other sensors than the camera remains subject to further research.\",\n \"id\": \"dcb846f5-368c-45a9-a56f-6ddb586a0a6b\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Uncertainty\"\n ],\n [\n \"subsection\",\n \"Uncertainty Metrics for DNNs in Frequentist Inference\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Uncertainty Metrics for DNNs in Frequentist Inference\"\n },\n {\n \"cite_extract_rate\": 0.23529411764705802,\n \"cites\": [\n 8213,\n 8212,\n 1737,\n 1745\n ],\n \"content\": \"\\\\label{subsec:uncertainty:markov_random_fields}\\n\\\\sectionauthor{Seyed Eghbal Ghobadi\\\\textsuperscript{8}, Ahmed Hammam\\\\textsuperscript{8}}\\nAlthough deep neural networks are currently the state of the art for almost all computer vision tasks, Markov random fields (MRF) remain one of the fundamental techniques used for many computer vision tasks, specifically image segmentation ,. MRFs hold its power in the essence of being able to model dependencies between pixels in an image. With the use of energy functions, MRFs integrate pixels into models relating between unary and pairwise pixels together . \\nGiven the model, MRFs are used to infer the optimal configuration yielding the lowest energy using mainly maximum a posteriori (MAP) techniques. Several MAP inference approaches are used to yield the optimal configuration such as graph cuts and belief propagation algorithms . However, as with neural networks, MAP inference techniques result in deterministic point estimates of the optimal configuration without any sense of uncertainty in the output. To obtain uncertainties on results from MRFs, most of the work is directed towards modelling MRFs with Gaussian distributions. Getting uncertainties from MRFs with Gaussian distributions is possible by two typical methods: Either approximate models are inferred to the posterior, from which sampling is easy or the variances can be estimated analytically, or approximate sampling from the posterior is used.\\nApproximate models include those inferred using variational Bayesian (VB) methods, like mean-field approximations, and using Gaussian process (GP) models enforcing a simplified prior model , . Examples of approximate sampling methods include traditional Markov chain Monte Carlo (MCMC) methods like Gibbs sampling . Some recent theoretical advances propose the perturb-and-MAP framework and a Gumbel perturbation model (GPM) , to exactly sample from MRF distributions. Another line of work has also been proposed, where MAP inference techniques are used to estimate the probability of the network output. With the use of graph cuts, try to estimate uncertainty using the min-marginals associated with the label assignments of a random field. Here, the work by Kohli and Torr was extended to show how this approach can be extended to techniques other than graph cuts or compute uncertainties on multi-label marginal distributions .\\n\\\\par\\nA current research direction is the incorporation of MRFs with deep neural networks, along with providing uncertainties on the output . This can also be extended to other forms of neural networks such as recurrent neural networks to provide uncertainties on segmentation of streams of videos with extending dependencies of pixels to previous frames , .\",\n \"id\": \"81f04d24-1455-4244-a481-8a943c0061e4\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Uncertainty\"\n ],\n [\n \"subsection\",\n \"Markov Random Fields\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Markov Random Fields\"\n },\n {\n \"cite_extract_rate\": 0.41176470588235203,\n \"cites\": [\n 8214,\n 8216,\n 759,\n 8215,\n 4141,\n 8217,\n 4690\n ],\n \"content\": \"\\\\label{subsec:uncertainty:confidence_calibration} \\n\\\\sectionauthor{Fabian Küppers\\\\textsuperscript{9}, Anselm Haselhoff\\\\textsuperscript{9}}\\nNeural network classifiers output a label $\\\\hat{Y} \\\\in \\\\mathcal{Y}$ on a given input $X \\\\in \\\\mathcal{X}$ with an associated confidence $\\\\hat{P}$. This confidence can be interpreted as a probability of correctness that the predicted label matches the ground truth label $Y \\\\in \\\\mathcal{Y}$. Therefore, these probabilities should reflect the ”self-confidence” of the system. If the empirical accuracy for any confidence level matches the predicted confidence, a model is called \\\\emph{well calibrated}.\\nTherefore, a \\\\emph{classification} model is perfectly calibrated if \\n\\\\begin{align}\\n\\\\label{eq:classification_calibration}\\n& \\\\underbrace{\\\\mathbb{P}(\\\\hat{Y} = Y | \\\\hat{P} = p)}_{\\\\text{accuracy given } p} = \\\\underbrace{p}_{\\\\text{confidence}} \\\\quad \\\\forall p \\\\in [0,1]\\n\\\\end{align}\\nis fullfilled . For example, assume 100 predictions with confidence values of 0.9. We call the model well calibrated if 90 out of these 100 predictions are actually correct. However, recent work has shown that modern neural networks tend to be too overconfident in their predictions . The deviation of a model to the perfect calibration can be measured by the expected calibration error (ECE) . \\nIt is possible to recalibrate models as a post-processing step after classification. One way to get a calibration mapping is to group all predictions into several bins by their confidence. Using such a binning scheme, it is possible to compute the empirical accuracy for certain confidence levels, as it is known for a long time already in reconstructing confidence outputs for Viterbi decoding . Common methods are histogram binning , isotonic regression or more advanced methods like Bayesian binning into quantiles (BBQ) and ensembles of near-isotonic regression (ENIR) . Another way to get a calibration mapping is to use scaling methods based on logistic regression like Platt scaling , temperature scaling and beta calibration .\\n\\\\par\\nIn the setting of probabilistic regression, a model is calibrated if, \\\\eg 95\\\\% of the true target values are below or equal to a credible level of 95\\\\% (so called \\\\emph{quantile-calibrated regression}) . A regression model is usually calibrated by fine-tuning its predicted CDF in a post-processing step to match the empirical frequency. Common approaches utilize isotonic regression , logistic and beta calibration , as well as Gaussian process models to build a calibration mapping. In contrast to quantile-calibrated regression, have recently introduced the concept of distribution calibration, where calibration is applied on a distribution level and naturally leads to calibrated quantiles.\\n\\\\par\\nRecent work has shown that miscalibration in the scope of \\\\emph{object detection} also depends on the position and scale of a detected object . The additional box regression output is denoted by $\\\\hat{R}$ with $J$ as the size of the used box encoding.\\nFurthermore, if we have no knowledge about all anchors of a model (which is a common case in many applications), it is not possible to determine the accuracy. Therefore, K\\\\\\\"{u}ppers \\\\etal~ use the precision as a surrogate for accuracy and propose that an \\\\emph{object detection model} is perfectly calibrated if\\n\\\\begin{align}\\n\\\\label{eq:detection_calibration}\\n& \\\\underbrace{\\\\mathbb{P}(M=1 | \\\\hat{P} = p, \\\\hat{Y} = y, \\\\hat{R} = r)}_{\\\\text{precision given } p, y, r} = \\\\underbrace{p}_{\\\\text{confidence}} \\\\quad \\\\forall p \\\\in [0,1], y \\\\in \\\\mathcal{Y}, r \\\\in \\\\mathbb{R}^J \\n\\\\end{align}\\nis fulfilled, where $M=1$ denotes a correct prediction that matches a ground-truth object with a chosen IoU threshold and $M=0$ denotes a mismatch, respectively. The authors propose the detection-expected calibration error (D-ECE) as the extension of the ECE to object detection tasks in order to measure miscalibration also by means of the position and scale of detected objects. Other approaches try to fine-tune the regression output in order to obtain more reliable object proposals or to add a regularization term to the training objective such that training yields models that are both well-performing and well-calibrated .\",\n \"id\": \"09524fc1-6625-4adb-b57e-b6186b5d4ef0\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 17,\n \"parent_id\": \"9f13c349-4fb4-458e-8f34-84cd3a93fca0\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Uncertainty\"\n ],\n [\n \"subsection\",\n \"Confidence Calibration\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Confidence Calibration\"\n },\n {\n \"cite_extract_rate\": 0.5,\n \"cites\": [\n 6435\n ],\n \"content\": \"\\\\label{sec:aggregation}\\n\\\\sectionauthor{Maram Akila\\\\textsuperscript{1}}\\nFrom a high-level perspective, a \\\\emph{neural network} is based on processing inputs and coming to some output conclusion, \\\\eg mapping incoming image data onto class labels.\\nAggregation or collection of non-independent information on either the input or output side of this network function can be used as a tool to leverage its performance and reliability.\\nStarting with the input, any additional ``dimension'' to add data can be of use.\\nFor example, in the context of autonomous vehicles this might be input from any further sensor measuring the same scene as the original one, \\\\eg stereo cameras or LiDAR.\\nCombining those sensor sets for prediction is commonly referred to as \\\\emph{sensor fusion} .\\nStaying with the example, the scene will be monitored consecutively providing a whole (temporally ordered) stream of input information.\\nThis may be used either by adjusting the network for this kind of input \\n \\nor in terms of a post-processing step, in which the predictions are aggregated by some measure of temporal consistency. \\n\\\\par\\nAnother more implicit form of aggregation is training the neural network on several ``independent'' tasks, \\\\eg segmentation and depth regression. Although the individual task is executed on the same input, the overall performance can still benefit from the correlation among all given tasks. We refer to the discussion on \\\\emph{multi-task networks} in \\\\refsec{subsec:architecture:multi_task_networks}.\\nBy extension, solving the same task in multiple different ways, can be beneficial for performance and provide a measure of redundancy.\\nIn this survey, we focus on single-task systems and discuss \\\\emph{ensemble methods} in the next section and the use of \\\\emph{temporal consistency} in the one thereafter.\",\n \"id\": \"71d4e278-f67c-443c-84a9-ff0254362119\",\n \"level\": \"section\",\n \"origin_cites_number\": 2,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Aggregation\"\n ]\n ],\n \"subsections\": [\n \"d1f9f7cf-fc5d-426f-8fba-a38a5d51acf5\",\n \"61379b41-7fce-4042-aff6-0ec085cd8afc\"\n ],\n \"title\": \"Aggregation\"\n },\n {\n \"cite_extract_rate\": 0.6666666666666661,\n \"cites\": [\n 4616,\n 4678,\n 582,\n 3288,\n 680,\n 681,\n 8966,\n 8218\n ],\n \"content\": \"\\\\label{subsec:aggregation:ensemble_methods}\\n\\\\sectionauthor{Joachim Sicking\\\\textsuperscript{1}}\\nTraining a neural network is optimizing its parameters to fit a given training data set. The commonly used gradient-based optimization schemes cause convergence in a ‘nearby’ local minimum. As the loss landscapes of neural networks are notoriously non-convex , various locally optimal model parameter sets exist. These local optima differ in the degree of optimality (“deepness”), qualitative characteristics (“optimal for different parts of the training data”) and their generalizability to unseen data (commonly referred to by the geometrical terms of “sharpness” and “flatness” of minima ). \\nA single trained network corresponds to one local minimum of such a loss landscape and thus captures only a small part of a potentially diverse set of solutions. \\\\emph{Network ensembles} are collections of models and therefore better suited to reflect this multi-modality. Various modelling choices shape a loss landscape: the selected model class and its meta-parameters (like architecture and layer width), the training data and the optimization objective. Accordingly, approaches to diversify ensemble components range from combinations of different model classes over varying training data (bagging) to methods that train and weight ensemble components to make up for the flaws of other ensemble members (boosting) .\\nGiven the millions of parameters of application-size networks, ensembles of NNs are resource-demanding \\\\wrt computational load, storage and runtime during training and inference. This complexity increases linearly with ensemble size for naïve ensembling. Several approaches were put forward to reduce some dimensions of this complexity: \\\\emph{snapshot ensembles} require only one model optimization with a cyclical learning-rate schedule—leading to an optimized training runtime. The resulting training trajectory passes through several local minima. The corresponding models compose the ensemble. On the contrary, \\\\emph{model distillation} tackles runtime at inference. They ‘squeeze’ a NN ensemble into a single model that is optimized to capture the gist of the model set. However, such a compression goes along with reduced performance compared to the original ensemble.\\nSeveral hybrids of single model and model ensemble exist: Multi-head networks share a backbone network that provides inputs to multiple prediction networks. Another variant are mixture-of-expert models that utilize a gating network to assign inputs to specialized expert networks . Multi-task networks (\\\\cf \\\\refsec{subsec:architecture:multi_task_networks}) and Bayesian approximations of NNs (\\\\cf \\\\refsec{subsec:uncertainty:bayesian_neural_networks} and \\\\refsec{subsec:uncertainty:mc_dropout}) can be seen as implicit ensembles. \\nNN ensembles (or deep ensembles) are not only used to boost model quality. They pose the frequentist's approach to estimating NN uncertainties and are state-of-the-art in this regard . The emergent field of federated learning is concerned with the integration of decentrally trained ensemble components and safety-relevant applications of ensembling range from autonomous driving to medical diagnostics . \\nTaking this safe-ML perspective, promising research directions comprise a more principled and efficient composition of model ensembles, \\\\eg by application-driven diversification, as well as improved techniques to miniaturize ensembles, \\\\eg by gaining a better understanding of methods like distillation. In the long run, better designed, more powerful learning systems might partially reduce the need for combining weaker models in a network ensemble.\",\n \"id\": \"d1f9f7cf-fc5d-426f-8fba-a38a5d51acf5\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 12,\n \"parent_id\": \"71d4e278-f67c-443c-84a9-ff0254362119\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Aggregation\"\n ],\n [\n \"subsection\",\n \"Ensemble Methods\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Ensemble Methods\"\n },\n {\n \"cite_extract_rate\": 0.4,\n \"cites\": [\n 8219,\n 1745,\n 8220,\n 784\n ],\n \"content\": \"\\\\label{subsec:aggregation:temporal_consistency}\\n\\\\sectionauthor{Timo Sämann\\\\textsuperscript{4}}\\nThe focus of previous DNN development for semantic segmentation has been on single image prediction. This means that the final and intermediate results of the DNN are discarded after each image. However, the application of a computer vision model often involves the processing of images in a sequence, \\\\ie there is a temporal consistency in the image content between consecutive frames (for a metric, \\\\cf, \\\\eg ). This consistency has been exploited in previous work to increase quality and reduce computing effort. Furthermore, this approach offers the potential to improve the robustness of DNN prediction by incorporating this consistency as a-priori knowledge into DNN development. The relevant work in the field of video prediction can be divided in two major approaches: \\n\\\\begin{enumerate}\\n\\\\item DNNs are specially designed for video prediction. This usually requires training from scratch and the availability of training data in a sequence.\\n\\\\item A transformation from single prediction DNNs to video prediction DNNs takes place. Usually no training is required, \\\\ie the existing weights of the model can be used unaltered.\\n\\\\end{enumerate}\\nThe first set of approaches often involves \\\\emph{conditional random fields} (CRF) and its variants. \\nCRFs are known for their use as postprocessing step in the prediction of semantic segmentation, in which their parameters are learned separately or jointly with the DNN~. \\nAnother way to use spatiotemporal features is to include 3D convolutions, which add an additional dimension to the conventional 2D convolutional layer. Tran \\\\etal use 3D convolution layers for video recognition tasks such as action and object recognition. \\nOne further approach to use spatial and temporal characteristics of the input data is to integrate \\\\emph{long short-term memory} (LSTM)~, a variant of the \\\\emph{recurrent neural network} (RNN). Fayyaz \\\\etal integrate LSTM layers between the encoder and decoder of their convolutional neural network for semantic segmentation. The significantly higher GPU memory requirements and computational effort are a disadvantage of this method. More recently, Nilsson and Sminchisescu deployed \\\\emph{gated recurrent units}, which generally requires significantly less memory.\\nAn approach to improve temporal consistency of automatic speech recognition outputs is known as a posterior-in-posterior-out (PIPO) LSTM ``sequence enhancer'', a postfilter which could be applicable to video processing as well .\\nA disadvantage of the described methods is that sequential data for training must be available, which may be limited or show a lack of diversity. \\n\\\\par\\nThe second class of approaches has the advantage that it is model-independent most of the time. Shelhamer \\\\etal found that the deep feature maps within the network change only slightly with temporal changes in video content.\\nAccordingly, calculate the optical flow of the input images from time steps $t_0$ and $t_{-1}$ and convert it into the so-called \\\\emph{transform flow} which is used to transform the feature maps of the time step $t_{-1}$ so that an aligned representation to the feature map $t_0$ is achieved. S{\\\\\\\"a}mann \\\\etal use a confidence-based combination of feature maps from previous time steps based on the calculated optical flow.\",\n \"id\": \"61379b41-7fce-4042-aff6-0ec085cd8afc\",\n \"level\": \"subsection\",\n \"origin_cites_number\": 10,\n \"parent_id\": \"71d4e278-f67c-443c-84a9-ff0254362119\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Aggregation\"\n ],\n [\n \"subsection\",\n \"Temporal Consistency\"\n ]\n ],\n \"subsections\": [],\n \"title\": \"Temporal Consistency\"\n },\n {\n \"cite_extract_rate\": 0,\n \"cites\": [],\n \"content\": \"\\\\label{sec:verification}\\n\\\\sectionauthor{Gesina Schwalbe\\\\textsuperscript{3}}\\nVerification and validation is an integral part of the safety\\nassurance for any safety critical systems. As of the functional safety\\nstandard for automotive systems ,\\n\\\\emph{verification} means to determine whether given requirements are\\nmet \\\\cite[3.180]{iso/tc_22/sc_32_iso_2018}, such as performance\\ngoals.\\n\\\\emph{Validation} on the other side tries to assess whether the given\\nrequirements are sufficient and adequate to guarantee safety\\n\\\\cite[3.148]{iso/tc_22/sc_32_iso_2018}, \\\\eg whether certain types of\\nfailures or interactions simply were overlooked. The latter is usually\\nachieved via extensive testing in real operation conditions of the\\nintegrated product. This differs from the notion of validation used in\\nthe machine learning community in which it usually refers to simple\\nperformance tests on a selected dataset.\\nIn this section, we want to concentrate on general verification\\naspects for deep neural networks.\\nVerification as in the safety domain encompasses (manual) inspection\\nand analysis activities, and testing. However, the contribution\\nof single processing steps within a neural network to the final\\nbehavior can hardly be assessed manually (compare to the problem of\\ninterpretability in \\\\refsec{sec:interpretability}). Therefore, we here will concentrate on different\\napproaches to verification testing.\\nSection~\\\\ref{subsec:verification:formal_testing} covers approaches\\nto systematic test data selection. While the suggested methods assume\\nfull access to the model internals for coverage measurement, this is\\nnot in all cases available. Therefore,\\n\\\\refsec{subsec:verification:black_box_methods} highlights\\nassessment approaches that consider the DNN as a black-box component.\\nAlso, the topic of verification activity during operation with the\\nhelp of observers is discussed.\",\n \"id\": \"94174462-51f2-497a-b88f-1ceaadd52aaf\",\n \"level\": \"section\",\n \"origin_cites_number\": 1,\n \"parent_id\": \"569f1197-c628-4e9d-9db6-9757cb02d82a\",\n \"prefix_titles\": [\n [\n \"title\",\n \"Inspect, Understand, Overcome:\\\\\\\\A Survey of Practical Methods\\\\\\\\for AI Safety\"\n ],\n [\n \"section\",\n \"Verification\"\n ]\n ],\n \"subsections\": [\n \"5ad8ed4b-a3fa-43df-8241-7dd338b6350c\",\n \"5a9669c2-3369-414d-a5ae-bc8b21dd4bcd\"\n ],\n \"title\": \"Verification\"\n },\n {\n \"cite_extract_rate\": 0.28571428571428503,\n \"cites\": [\n 8089,\n 8971,\n 1636,\n 6024\n ],\n \"content\": \"